From 55ac7d333c55d808be33c590a4a2e6c965d5f9a8 Mon Sep 17 00:00:00 2001 From: Arjan Singh <1598641+arjansingh@users.noreply.github.com> Date: Wed, 15 Oct 2025 20:30:39 -0700 Subject: [PATCH 001/647] Bump frontend to 1.28.7 (#10364) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index a45057970..82457df54 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,4 +1,4 @@ -comfyui-frontend-package==1.28.6 +comfyui-frontend-package==1.28.7 comfyui-workflow-templates==0.1.95 comfyui-embedded-docs==0.3.0 torch From 4054b4bf38d11fc0c784c2d19f5fc0ed3bbc7ae4 Mon Sep 17 00:00:00 2001 From: Rizumu Ayaka Date: Thu, 16 Oct 2025 16:13:31 +0800 Subject: [PATCH 002/647] feat: deprecated API alert (#10366) --- server.py | 24 +++++++++++++++++++++++- 1 file changed, 23 insertions(+), 1 deletion(-) diff --git a/server.py b/server.py index 80e9d3fa7..a44f4f237 100644 --- a/server.py +++ b/server.py @@ -48,6 +48,28 @@ async def send_socket_catch_exception(function, message): except (aiohttp.ClientError, aiohttp.ClientPayloadError, ConnectionResetError, BrokenPipeError, ConnectionError) as err: logging.warning("send error: {}".format(err)) +# Track deprecated paths that have been warned about to only warn once per file +_deprecated_paths_warned = set() + +@web.middleware +async def deprecation_warning(request: web.Request, handler): + """Middleware to warn about deprecated frontend API paths""" + path = request.path + + if (path.startswith('/scripts/') or path.startswith('/extensions/core/')): + # Only warn once per unique file path + if path not in _deprecated_paths_warned: + _deprecated_paths_warned.add(path) + logging.warning( + f"[DEPRECATION WARNING] Detected import of deprecated legacy API: {path}. " + f"This is likely caused by a custom node extension using outdated APIs. " + f"Please update your extensions or contact the extension author for an updated version." + ) + + response: web.Response = await handler(request) + return response + + @web.middleware async def compress_body(request: web.Request, handler): accept_encoding = request.headers.get("Accept-Encoding", "") @@ -159,7 +181,7 @@ class PromptServer(): self.client_session:Optional[aiohttp.ClientSession] = None self.number = 0 - middlewares = [cache_control] + middlewares = [cache_control, deprecation_warning] if args.enable_compress_response_body: middlewares.append(compress_body) From bc0ad9bb49b642e081f99f92d239d634988d52bc Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Thu, 16 Oct 2025 20:12:50 +0300 Subject: [PATCH 003/647] fix(api-nodes): remove "veo2" model from Veo3 node (#10372) --- comfy_api_nodes/nodes_veo2.py | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/comfy_api_nodes/nodes_veo2.py b/comfy_api_nodes/nodes_veo2.py index 4ab5c5186..daeaa823e 100644 --- a/comfy_api_nodes/nodes_veo2.py +++ b/comfy_api_nodes/nodes_veo2.py @@ -393,7 +393,9 @@ class Veo3VideoGenerationNode(VeoVideoGenerationNode): ), IO.Combo.Input( "model", - options=list(MODELS_MAP.keys()), + options=[ + "veo-3.1-generate", "veo-3.1-fast-generate", "veo-3.0-generate-001", "veo-3.0-fast-generate-001" + ], default="veo-3.0-generate-001", tooltip="Veo 3 model to use for video generation", optional=True, From 19b466160c1cd43f707769adef6f8ed6e9fd50bf Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Thu, 16 Oct 2025 15:16:03 -0700 Subject: [PATCH 004/647] Workaround for nvidia issue where VAE uses 3x more memory on torch 2.9 (#10373) --- comfy/ops.py | 19 +++++++++++++++++++ 1 file changed, 19 insertions(+) diff --git a/comfy/ops.py b/comfy/ops.py index b2096b40e..893ceda98 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -52,6 +52,16 @@ try: except (ModuleNotFoundError, TypeError): logging.warning("Could not set sdpa backend priority.") +NVIDIA_MEMORY_CONV_BUG_WORKAROUND = False +try: + if comfy.model_management.is_nvidia(): + if torch.backends.cudnn.version() >= 91300 and comfy.model_management.torch_version_numeric >= (2, 9) and comfy.model_management.torch_version_numeric <= (2, 10): + #TODO: change upper bound version once it's fixed' + NVIDIA_MEMORY_CONV_BUG_WORKAROUND = True + logging.info("working around nvidia conv3d memory bug.") +except: + pass + cast_to = comfy.model_management.cast_to #TODO: remove once no more references if torch.cuda.is_available() and torch.backends.cudnn.is_available() and PerformanceFeature.AutoTune in args.fast: @@ -151,6 +161,15 @@ class disable_weight_init: def reset_parameters(self): return None + def _conv_forward(self, input, weight, bias, *args, **kwargs): + if NVIDIA_MEMORY_CONV_BUG_WORKAROUND and weight.dtype in (torch.float16, torch.bfloat16): + out = torch.cudnn_convolution(input, weight, self.padding, self.stride, self.dilation, self.groups, benchmark=False, deterministic=False, allow_tf32=True) + if bias is not None: + out += bias.reshape((1, -1) + (1,) * (out.ndim - 2)) + return out + else: + return super()._conv_forward(input, weight, bias, *args, **kwargs) + def forward_comfy_cast_weights(self, input): weight, bias = cast_bias_weight(self, input) return self._conv_forward(input, weight, bias) From b1293d50eff5f1ff2e54f73114fbe7c0f9aef8fe Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Thu, 16 Oct 2025 16:59:56 -0700 Subject: [PATCH 005/647] workaround also works on cudnn 91200 (#10375) --- comfy/ops.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy/ops.py b/comfy/ops.py index 893ceda98..56b07b44c 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -55,7 +55,7 @@ except (ModuleNotFoundError, TypeError): NVIDIA_MEMORY_CONV_BUG_WORKAROUND = False try: if comfy.model_management.is_nvidia(): - if torch.backends.cudnn.version() >= 91300 and comfy.model_management.torch_version_numeric >= (2, 9) and comfy.model_management.torch_version_numeric <= (2, 10): + if torch.backends.cudnn.version() >= 91200 and comfy.model_management.torch_version_numeric >= (2, 9) and comfy.model_management.torch_version_numeric <= (2, 10): #TODO: change upper bound version once it's fixed' NVIDIA_MEMORY_CONV_BUG_WORKAROUND = True logging.info("working around nvidia conv3d memory bug.") From d8d60b56093a15edc5d25486d387d3c5917dc3d3 Mon Sep 17 00:00:00 2001 From: Jedrzej Kosinski Date: Thu, 16 Oct 2025 21:39:37 -0700 Subject: [PATCH 006/647] Do batch_slice in EasyCache's apply_cache_diff (#10376) --- comfy_extras/nodes_easycache.py | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-) diff --git a/comfy_extras/nodes_easycache.py b/comfy_extras/nodes_easycache.py index c170e9fd9..1359e2f99 100644 --- a/comfy_extras/nodes_easycache.py +++ b/comfy_extras/nodes_easycache.py @@ -244,6 +244,8 @@ class EasyCacheHolder: self.total_steps_skipped += 1 batch_offset = x.shape[0] // len(uuids) for i, uuid in enumerate(uuids): + # slice out only what is relevant to this cond + batch_slice = [slice(i*batch_offset,(i+1)*batch_offset)] # if cached dims don't match x dims, cut off excess and hope for the best (cosmos world2video) if x.shape[1:] != self.uuid_cache_diffs[uuid].shape[1:]: if not self.allow_mismatch: @@ -261,9 +263,8 @@ class EasyCacheHolder: slicing.append(slice(None, dim_u)) else: slicing.append(slice(None)) - slicing = [slice(i*batch_offset,(i+1)*batch_offset)] + slicing - x = x[slicing] - x += self.uuid_cache_diffs[uuid].to(x.device) + batch_slice = batch_slice + slicing + x[batch_slice] += self.uuid_cache_diffs[uuid].to(x.device) return x def update_cache_diff(self, output: torch.Tensor, x: torch.Tensor, uuids: list[UUID]): From b1467da4803017a418c32c159525767f45871ca3 Mon Sep 17 00:00:00 2001 From: rattus128 <46076784+rattus128@users.noreply.github.com> Date: Sat, 18 Oct 2025 06:55:15 +1000 Subject: [PATCH 007/647] execution: fold in dependency aware caching / Fix --cache-none with loops/lazy etc (#10368) * execution: fold in dependency aware caching This makes --cache-none compatiable with lazy and expanded subgraphs. Currently the --cache-none option is powered by the DependencyAwareCache. The cache attempts to maintain a parallel copy of the execution list data structure, however it is only setup once at the start of execution and does not get meaninigful updates to the execution list. This causes multiple problems when --cache-none is used with lazy and expanded subgraphs as the DAC does not accurately update its copy of the execution data structure. DAC has an attempt to handle subgraphs ensure_subcache however this does not accurately connect to nodes outside the subgraph. The current semantics of DAC are to free a node ASAP after the dependent nodes are executed. This means that if a subgraph refs such a node it will be requed and re-executed by the execution_list but DAC wont see it in its to-free lists anymore and leak memory. Rather than try and cover all the cases where the execution list changes from inside the cache, move the while problem to the executor which maintains an always up-to-date copy of the wanted data-structure. The executor now has a fast-moving run-local cache of its own. Each _to node has its own mini cache, and the cache is unconditionally primed at the time of add_strong_link. add_strong_link is called for all of static workflows, lazy links and expanded subgraphs so its the singular source of truth for output dependendencies. In the case of a cache-hit, the executor cache will hold the non-none value (it will respect updates if they happen somehow as well). In the case of a cache-miss, the executor caches a None and will wait for a notification to update the value when the node completes. When a node completes execution, it simply releases its mini-cache and in turn its strong refs on its direct anscestor outputs, allowing for ASAP freeing (same as the DependencyAwareCache but a little more automatic). This now allows for re-implementation of --cache-none with no cache at all. The dependency aware cache was also observing the dependency sematics for the objects and UI cache which is not accurate (this entire logic was always outputs specific). This also prepares for more complex caching strategies (such as RAM pressure based caching), where a cache can implement any freeing strategy completely independently of the DepedancyAwareness requirement. * main: re-implement --cache-none as no cache at all The execution list now tracks the dependency aware caching more correctly that the DependancyAwareCache. Change it to a cache that does nothing. * test_execution: add --cache-none to the test suite --cache-none is now expected to work universally. Run it through the full unit test suite. Propagate the server parameterization for whether or not the server is capabale of caching, so that the minority of tests that specifically check for cache hits can if else. Hard assert NOT caching in the else to give some coverage of --cache-none expected behaviour to not acutally cache. --- comfy_execution/caching.py | 174 ++++-------------------------- comfy_execution/graph.py | 31 +++++- execution.py | 34 +++--- main.py | 2 +- tests/execution/test_execution.py | 50 +++++---- 5 files changed, 101 insertions(+), 190 deletions(-) diff --git a/comfy_execution/caching.py b/comfy_execution/caching.py index 41224ce3b..566bc3f9c 100644 --- a/comfy_execution/caching.py +++ b/comfy_execution/caching.py @@ -265,6 +265,26 @@ class HierarchicalCache(BasicCache): assert cache is not None return await cache._ensure_subcache(node_id, children_ids) +class NullCache: + + async def set_prompt(self, dynprompt, node_ids, is_changed_cache): + pass + + def all_node_ids(self): + return [] + + def clean_unused(self): + pass + + def get(self, node_id): + return None + + def set(self, node_id, value): + pass + + async def ensure_subcache_for(self, node_id, children_ids): + return self + class LRUCache(BasicCache): def __init__(self, key_class, max_size=100): super().__init__(key_class) @@ -316,157 +336,3 @@ class LRUCache(BasicCache): self._mark_used(child_id) self.children[cache_key].append(self.cache_key_set.get_data_key(child_id)) return self - - -class DependencyAwareCache(BasicCache): - """ - A cache implementation that tracks dependencies between nodes and manages - their execution and caching accordingly. It extends the BasicCache class. - Nodes are removed from this cache once all of their descendants have been - executed. - """ - - def __init__(self, key_class): - """ - Initialize the DependencyAwareCache. - - Args: - key_class: The class used for generating cache keys. - """ - super().__init__(key_class) - self.descendants = {} # Maps node_id -> set of descendant node_ids - self.ancestors = {} # Maps node_id -> set of ancestor node_ids - self.executed_nodes = set() # Tracks nodes that have been executed - - async def set_prompt(self, dynprompt, node_ids, is_changed_cache): - """ - Clear the entire cache and rebuild the dependency graph. - - Args: - dynprompt: The dynamic prompt object containing node information. - node_ids: List of node IDs to initialize the cache for. - is_changed_cache: Flag indicating if the cache has changed. - """ - # Clear all existing cache data - self.cache.clear() - self.subcaches.clear() - self.descendants.clear() - self.ancestors.clear() - self.executed_nodes.clear() - - # Call the parent method to initialize the cache with the new prompt - await super().set_prompt(dynprompt, node_ids, is_changed_cache) - - # Rebuild the dependency graph - self._build_dependency_graph(dynprompt, node_ids) - - def _build_dependency_graph(self, dynprompt, node_ids): - """ - Build the dependency graph for all nodes. - - Args: - dynprompt: The dynamic prompt object containing node information. - node_ids: List of node IDs to build the graph for. - """ - self.descendants.clear() - self.ancestors.clear() - for node_id in node_ids: - self.descendants[node_id] = set() - self.ancestors[node_id] = set() - - for node_id in node_ids: - inputs = dynprompt.get_node(node_id)["inputs"] - for input_data in inputs.values(): - if is_link(input_data): # Check if the input is a link to another node - ancestor_id = input_data[0] - self.descendants[ancestor_id].add(node_id) - self.ancestors[node_id].add(ancestor_id) - - def set(self, node_id, value): - """ - Mark a node as executed and store its value in the cache. - - Args: - node_id: The ID of the node to store. - value: The value to store for the node. - """ - self._set_immediate(node_id, value) - self.executed_nodes.add(node_id) - self._cleanup_ancestors(node_id) - - def get(self, node_id): - """ - Retrieve the cached value for a node. - - Args: - node_id: The ID of the node to retrieve. - - Returns: - The cached value for the node. - """ - return self._get_immediate(node_id) - - async def ensure_subcache_for(self, node_id, children_ids): - """ - Ensure a subcache exists for a node and update dependencies. - - Args: - node_id: The ID of the parent node. - children_ids: List of child node IDs to associate with the parent node. - - Returns: - The subcache object for the node. - """ - subcache = await super()._ensure_subcache(node_id, children_ids) - for child_id in children_ids: - self.descendants[node_id].add(child_id) - self.ancestors[child_id].add(node_id) - return subcache - - def _cleanup_ancestors(self, node_id): - """ - Check if ancestors of a node can be removed from the cache. - - Args: - node_id: The ID of the node whose ancestors are to be checked. - """ - for ancestor_id in self.ancestors.get(node_id, []): - if ancestor_id in self.executed_nodes: - # Remove ancestor if all its descendants have been executed - if all(descendant in self.executed_nodes for descendant in self.descendants[ancestor_id]): - self._remove_node(ancestor_id) - - def _remove_node(self, node_id): - """ - Remove a node from the cache. - - Args: - node_id: The ID of the node to remove. - """ - cache_key = self.cache_key_set.get_data_key(node_id) - if cache_key in self.cache: - del self.cache[cache_key] - subcache_key = self.cache_key_set.get_subcache_key(node_id) - if subcache_key in self.subcaches: - del self.subcaches[subcache_key] - - def clean_unused(self): - """ - Clean up unused nodes. This is a no-op for this cache implementation. - """ - pass - - def recursive_debug_dump(self): - """ - Dump the cache and dependency graph for debugging. - - Returns: - A list containing the cache state and dependency graph. - """ - result = super().recursive_debug_dump() - result.append({ - "descendants": self.descendants, - "ancestors": self.ancestors, - "executed_nodes": list(self.executed_nodes), - }) - return result diff --git a/comfy_execution/graph.py b/comfy_execution/graph.py index f4b427265..d5bbacde3 100644 --- a/comfy_execution/graph.py +++ b/comfy_execution/graph.py @@ -153,8 +153,9 @@ class TopologicalSort: continue _, _, input_info = self.get_input_info(unique_id, input_name) is_lazy = input_info is not None and "lazy" in input_info and input_info["lazy"] - if (include_lazy or not is_lazy) and not self.is_cached(from_node_id): - node_ids.append(from_node_id) + if (include_lazy or not is_lazy): + if not self.is_cached(from_node_id): + node_ids.append(from_node_id) links.append((from_node_id, from_socket, unique_id)) for link in links: @@ -194,10 +195,34 @@ class ExecutionList(TopologicalSort): super().__init__(dynprompt) self.output_cache = output_cache self.staged_node_id = None + self.execution_cache = {} + self.execution_cache_listeners = {} def is_cached(self, node_id): return self.output_cache.get(node_id) is not None + def cache_link(self, from_node_id, to_node_id): + if not to_node_id in self.execution_cache: + self.execution_cache[to_node_id] = {} + self.execution_cache[to_node_id][from_node_id] = self.output_cache.get(from_node_id) + if not from_node_id in self.execution_cache_listeners: + self.execution_cache_listeners[from_node_id] = set() + self.execution_cache_listeners[from_node_id].add(to_node_id) + + def get_output_cache(self, from_node_id, to_node_id): + if not to_node_id in self.execution_cache: + return None + return self.execution_cache[to_node_id].get(from_node_id) + + def cache_update(self, node_id, value): + if node_id in self.execution_cache_listeners: + for to_node_id in self.execution_cache_listeners[node_id]: + self.execution_cache[to_node_id][node_id] = value + + def add_strong_link(self, from_node_id, from_socket, to_node_id): + super().add_strong_link(from_node_id, from_socket, to_node_id) + self.cache_link(from_node_id, to_node_id) + async def stage_node_execution(self): assert self.staged_node_id is None if self.is_empty(): @@ -277,6 +302,8 @@ class ExecutionList(TopologicalSort): def complete_node_execution(self): node_id = self.staged_node_id self.pop_node(node_id) + self.execution_cache.pop(node_id, None) + self.execution_cache_listeners.pop(node_id, None) self.staged_node_id = None def get_nodes_in_cycle(self): diff --git a/execution.py b/execution.py index 1dc35738b..78c36a4b0 100644 --- a/execution.py +++ b/execution.py @@ -18,7 +18,7 @@ from comfy_execution.caching import ( BasicCache, CacheKeySetID, CacheKeySetInputSignature, - DependencyAwareCache, + NullCache, HierarchicalCache, LRUCache, ) @@ -91,13 +91,13 @@ class IsChangedCache: class CacheType(Enum): CLASSIC = 0 LRU = 1 - DEPENDENCY_AWARE = 2 + NONE = 2 class CacheSet: def __init__(self, cache_type=None, cache_size=None): - if cache_type == CacheType.DEPENDENCY_AWARE: - self.init_dependency_aware_cache() + if cache_type == CacheType.NONE: + self.init_null_cache() logging.info("Disabling intermediate node cache.") elif cache_type == CacheType.LRU: if cache_size is None: @@ -120,11 +120,12 @@ class CacheSet: self.ui = LRUCache(CacheKeySetInputSignature, max_size=cache_size) self.objects = HierarchicalCache(CacheKeySetID) - # only hold cached items while the decendents have not executed - def init_dependency_aware_cache(self): - self.outputs = DependencyAwareCache(CacheKeySetInputSignature) - self.ui = DependencyAwareCache(CacheKeySetInputSignature) - self.objects = DependencyAwareCache(CacheKeySetID) + def init_null_cache(self): + self.outputs = NullCache() + #The UI cache is expected to be iterable at the end of each workflow + #so it must cache at least a full workflow. Use Heirachical + self.ui = HierarchicalCache(CacheKeySetInputSignature) + self.objects = NullCache() def recursive_debug_dump(self): result = { @@ -135,7 +136,7 @@ class CacheSet: SENSITIVE_EXTRA_DATA_KEYS = ("auth_token_comfy_org", "api_key_comfy_org") -def get_input_data(inputs, class_def, unique_id, outputs=None, dynprompt=None, extra_data={}): +def get_input_data(inputs, class_def, unique_id, execution_list=None, dynprompt=None, extra_data={}): is_v3 = issubclass(class_def, _ComfyNodeInternal) if is_v3: valid_inputs, schema = class_def.INPUT_TYPES(include_hidden=False, return_schema=True) @@ -153,10 +154,10 @@ def get_input_data(inputs, class_def, unique_id, outputs=None, dynprompt=None, e if is_link(input_data) and (not input_info or not input_info.get("rawLink", False)): input_unique_id = input_data[0] output_index = input_data[1] - if outputs is None: + if execution_list is None: mark_missing() continue # This might be a lazily-evaluated input - cached_output = outputs.get(input_unique_id) + cached_output = execution_list.get_output_cache(input_unique_id, unique_id) if cached_output is None: mark_missing() continue @@ -405,6 +406,7 @@ async def execute(server, dynprompt, caches, current_item, extra_data, executed, cached_output = caches.ui.get(unique_id) or {} server.send_sync("executed", { "node": unique_id, "display_node": display_node_id, "output": cached_output.get("output",None), "prompt_id": prompt_id }, server.client_id) get_progress_state().finish_progress(unique_id) + execution_list.cache_update(unique_id, caches.outputs.get(unique_id)) return (ExecutionResult.SUCCESS, None, None) input_data_all = None @@ -434,7 +436,7 @@ async def execute(server, dynprompt, caches, current_item, extra_data, executed, for r in result: if is_link(r): source_node, source_output = r[0], r[1] - node_output = caches.outputs.get(source_node)[source_output] + node_output = execution_list.get_output_cache(source_node, unique_id)[source_output] for o in node_output: resolved_output.append(o) @@ -446,7 +448,7 @@ async def execute(server, dynprompt, caches, current_item, extra_data, executed, has_subgraph = False else: get_progress_state().start_progress(unique_id) - input_data_all, missing_keys, hidden_inputs = get_input_data(inputs, class_def, unique_id, caches.outputs, dynprompt, extra_data) + input_data_all, missing_keys, hidden_inputs = get_input_data(inputs, class_def, unique_id, execution_list, dynprompt, extra_data) if server.client_id is not None: server.last_node_id = display_node_id server.send_sync("executing", { "node": unique_id, "display_node": display_node_id, "prompt_id": prompt_id }, server.client_id) @@ -549,11 +551,15 @@ async def execute(server, dynprompt, caches, current_item, extra_data, executed, subcache.clean_unused() for node_id in new_output_ids: execution_list.add_node(node_id) + execution_list.cache_link(node_id, unique_id) for link in new_output_links: execution_list.add_strong_link(link[0], link[1], unique_id) pending_subgraph_results[unique_id] = cached_outputs return (ExecutionResult.PENDING, None, None) + caches.outputs.set(unique_id, output_data) + execution_list.cache_update(unique_id, output_data) + except comfy.model_management.InterruptProcessingException as iex: logging.info("Processing interrupted") diff --git a/main.py b/main.py index 35857dba8..4b4c5dcc4 100644 --- a/main.py +++ b/main.py @@ -173,7 +173,7 @@ def prompt_worker(q, server_instance): if args.cache_lru > 0: cache_type = execution.CacheType.LRU elif args.cache_none: - cache_type = execution.CacheType.DEPENDENCY_AWARE + cache_type = execution.CacheType.NONE e = execution.PromptExecutor(server_instance, cache_type=cache_type, cache_size=args.cache_lru) last_gc_collect = 0 diff --git a/tests/execution/test_execution.py b/tests/execution/test_execution.py index ef73ad9fd..ace0d2279 100644 --- a/tests/execution/test_execution.py +++ b/tests/execution/test_execution.py @@ -152,12 +152,12 @@ class TestExecution: # Initialize server and client # @fixture(scope="class", autouse=True, params=[ - # (use_lru, lru_size) - (False, 0), - (True, 0), - (True, 100), + { "extra_args" : [], "should_cache_results" : True }, + { "extra_args" : ["--cache-lru", 0], "should_cache_results" : True }, + { "extra_args" : ["--cache-lru", 100], "should_cache_results" : True }, + { "extra_args" : ["--cache-none"], "should_cache_results" : False }, ]) - def _server(self, args_pytest, request): + def server(self, args_pytest, request): # Start server pargs = [ 'python','main.py', @@ -167,12 +167,10 @@ class TestExecution: '--extra-model-paths-config', 'tests/execution/extra_model_paths.yaml', '--cpu', ] - use_lru, lru_size = request.param - if use_lru: - pargs += ['--cache-lru', str(lru_size)] + pargs += [ str(param) for param in request.param["extra_args"] ] print("Running server with args:", pargs) # noqa: T201 p = subprocess.Popen(pargs) - yield + yield request.param p.kill() torch.cuda.empty_cache() @@ -193,7 +191,7 @@ class TestExecution: return comfy_client @fixture(scope="class", autouse=True) - def shared_client(self, args_pytest, _server): + def shared_client(self, args_pytest, server): client = self.start_client(args_pytest["listen"], args_pytest["port"]) yield client del client @@ -225,7 +223,7 @@ class TestExecution: assert result.did_run(mask) assert result.did_run(lazy_mix) - def test_full_cache(self, client: ComfyClient, builder: GraphBuilder): + def test_full_cache(self, client: ComfyClient, builder: GraphBuilder, server): g = builder input1 = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1) input2 = g.node("StubImage", content="NOISE", height=512, width=512, batch_size=1) @@ -237,9 +235,12 @@ class TestExecution: client.run(g) result2 = client.run(g) for node_id, node in g.nodes.items(): - assert not result2.did_run(node), f"Node {node_id} ran, but should have been cached" + if server["should_cache_results"]: + assert not result2.did_run(node), f"Node {node_id} ran, but should have been cached" + else: + assert result2.did_run(node), f"Node {node_id} was cached, but should have been run" - def test_partial_cache(self, client: ComfyClient, builder: GraphBuilder): + def test_partial_cache(self, client: ComfyClient, builder: GraphBuilder, server): g = builder input1 = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1) input2 = g.node("StubImage", content="NOISE", height=512, width=512, batch_size=1) @@ -251,8 +252,12 @@ class TestExecution: client.run(g) mask.inputs['value'] = 0.4 result2 = client.run(g) - assert not result2.did_run(input1), "Input1 should have been cached" - assert not result2.did_run(input2), "Input2 should have been cached" + if server["should_cache_results"]: + assert not result2.did_run(input1), "Input1 should have been cached" + assert not result2.did_run(input2), "Input2 should have been cached" + else: + assert result2.did_run(input1), "Input1 should have been rerun" + assert result2.did_run(input2), "Input2 should have been rerun" def test_error(self, client: ComfyClient, builder: GraphBuilder): g = builder @@ -411,7 +416,7 @@ class TestExecution: input2 = g.node("StubImage", id="removeme", content="WHITE", height=512, width=512, batch_size=1) client.run(g) - def test_custom_is_changed(self, client: ComfyClient, builder: GraphBuilder): + def test_custom_is_changed(self, client: ComfyClient, builder: GraphBuilder, server): g = builder # Creating the nodes in this specific order previously caused a bug save = g.node("SaveImage") @@ -427,7 +432,10 @@ class TestExecution: result3 = client.run(g) result4 = client.run(g) assert result1.did_run(is_changed), "is_changed should have been run" - assert not result2.did_run(is_changed), "is_changed should have been cached" + if server["should_cache_results"]: + assert not result2.did_run(is_changed), "is_changed should have been cached" + else: + assert result2.did_run(is_changed), "is_changed should have been re-run" assert result3.did_run(is_changed), "is_changed should have been re-run" assert result4.did_run(is_changed), "is_changed should not have been cached" @@ -514,7 +522,7 @@ class TestExecution: assert len(images2) == 1, "Should have 1 image" # This tests that only constant outputs are used in the call to `IS_CHANGED` - def test_is_changed_with_outputs(self, client: ComfyClient, builder: GraphBuilder): + def test_is_changed_with_outputs(self, client: ComfyClient, builder: GraphBuilder, server): g = builder input1 = g.node("StubConstantImage", value=0.5, height=512, width=512, batch_size=1) test_node = g.node("TestIsChangedWithConstants", image=input1.out(0), value=0.5) @@ -530,7 +538,11 @@ class TestExecution: images = result.get_images(output) assert len(images) == 1, "Should have 1 image" assert numpy.array(images[0]).min() == 63 and numpy.array(images[0]).max() == 63, "Image should have value 0.25" - assert not result.did_run(test_node), "The execution should have been cached" + if server["should_cache_results"]: + assert not result.did_run(test_node), "The execution should have been cached" + else: + assert result.did_run(test_node), "The execution should have been re-run" + def test_parallel_sleep_nodes(self, client: ComfyClient, builder: GraphBuilder, skip_timing_checks): # Warmup execution to ensure server is fully initialized From 99ce2a1f66c4bcd500d76cc9a7430f7b2bf32776 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Sat, 18 Oct 2025 00:13:05 +0300 Subject: [PATCH 008/647] convert nodes_controlnet.py to V3 schema (#10202) --- comfy_extras/nodes_controlnet.py | 92 ++++++++++++++++++++------------ 1 file changed, 58 insertions(+), 34 deletions(-) diff --git a/comfy_extras/nodes_controlnet.py b/comfy_extras/nodes_controlnet.py index 2d20e1fed..e835feed7 100644 --- a/comfy_extras/nodes_controlnet.py +++ b/comfy_extras/nodes_controlnet.py @@ -1,20 +1,26 @@ from comfy.cldm.control_types import UNION_CONTROLNET_TYPES import nodes import comfy.utils +from typing_extensions import override +from comfy_api.latest import ComfyExtension, io -class SetUnionControlNetType: +class SetUnionControlNetType(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": {"control_net": ("CONTROL_NET", ), - "type": (["auto"] + list(UNION_CONTROLNET_TYPES.keys()),) - }} + def define_schema(cls): + return io.Schema( + node_id="SetUnionControlNetType", + category="conditioning/controlnet", + inputs=[ + io.ControlNet.Input("control_net"), + io.Combo.Input("type", options=["auto"] + list(UNION_CONTROLNET_TYPES.keys())), + ], + outputs=[ + io.ControlNet.Output(), + ], + ) - CATEGORY = "conditioning/controlnet" - RETURN_TYPES = ("CONTROL_NET",) - - FUNCTION = "set_controlnet_type" - - def set_controlnet_type(self, control_net, type): + @classmethod + def execute(cls, control_net, type) -> io.NodeOutput: control_net = control_net.copy() type_number = UNION_CONTROLNET_TYPES.get(type, -1) if type_number >= 0: @@ -22,27 +28,36 @@ class SetUnionControlNetType: else: control_net.set_extra_arg("control_type", []) - return (control_net,) + return io.NodeOutput(control_net) -class ControlNetInpaintingAliMamaApply(nodes.ControlNetApplyAdvanced): + set_controlnet_type = execute # TODO: remove + + +class ControlNetInpaintingAliMamaApply(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": {"positive": ("CONDITIONING", ), - "negative": ("CONDITIONING", ), - "control_net": ("CONTROL_NET", ), - "vae": ("VAE", ), - "image": ("IMAGE", ), - "mask": ("MASK", ), - "strength": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}), - "start_percent": ("FLOAT", {"default": 0.0, "min": 0.0, "max": 1.0, "step": 0.001}), - "end_percent": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0, "step": 0.001}) - }} + def define_schema(cls): + return io.Schema( + node_id="ControlNetInpaintingAliMamaApply", + category="conditioning/controlnet", + inputs=[ + io.Conditioning.Input("positive"), + io.Conditioning.Input("negative"), + io.ControlNet.Input("control_net"), + io.Vae.Input("vae"), + io.Image.Input("image"), + io.Mask.Input("mask"), + io.Float.Input("strength", default=1.0, min=0.0, max=10.0, step=0.01), + io.Float.Input("start_percent", default=0.0, min=0.0, max=1.0, step=0.001), + io.Float.Input("end_percent", default=1.0, min=0.0, max=1.0, step=0.001), + ], + outputs=[ + io.Conditioning.Output(display_name="positive"), + io.Conditioning.Output(display_name="negative"), + ], + ) - FUNCTION = "apply_inpaint_controlnet" - - CATEGORY = "conditioning/controlnet" - - def apply_inpaint_controlnet(self, positive, negative, control_net, vae, image, mask, strength, start_percent, end_percent): + @classmethod + def execute(cls, positive, negative, control_net, vae, image, mask, strength, start_percent, end_percent) -> io.NodeOutput: extra_concat = [] if control_net.concat_mask: mask = 1.0 - mask.reshape((-1, 1, mask.shape[-2], mask.shape[-1])) @@ -50,11 +65,20 @@ class ControlNetInpaintingAliMamaApply(nodes.ControlNetApplyAdvanced): image = image * mask_apply.movedim(1, -1).repeat(1, 1, 1, image.shape[3]) extra_concat = [mask] - return self.apply_controlnet(positive, negative, control_net, image, strength, start_percent, end_percent, vae=vae, extra_concat=extra_concat) + result = nodes.ControlNetApplyAdvanced().apply_controlnet(positive, negative, control_net, image, strength, start_percent, end_percent, vae=vae, extra_concat=extra_concat) + return io.NodeOutput(result[0], result[1]) + + apply_inpaint_controlnet = execute # TODO: remove +class ControlNetExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[io.ComfyNode]]: + return [ + SetUnionControlNetType, + ControlNetInpaintingAliMamaApply, + ] -NODE_CLASS_MAPPINGS = { - "SetUnionControlNetType": SetUnionControlNetType, - "ControlNetInpaintingAliMamaApply": ControlNetInpaintingAliMamaApply, -} + +async def comfy_entrypoint() -> ControlNetExtension: + return ControlNetExtension() From 92d97380bd02d9883295aeb2d29365cecd9a765e Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Fri, 17 Oct 2025 15:22:59 -0700 Subject: [PATCH 009/647] Update Python 3.14 installation instructions (#10385) Removed mention of installing pytorch nightly for Python 3.14. --- README.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/README.md b/README.md index b0731db33..c9a0644e3 100644 --- a/README.md +++ b/README.md @@ -197,7 +197,7 @@ comfy install ## Manual Install (Windows, Linux) -Python 3.14 will work if you comment out the `kornia` dependency in the requirements.txt file (breaks the canny node) and install pytorch nightly but it is not recommended. +Python 3.14 will work if you comment out the `kornia` dependency in the requirements.txt file (breaks the canny node) but it is not recommended. Python 3.13 is very well supported. If you have trouble with some custom node dependencies on 3.13 you can try 3.12 From 9da397ea2f271080406f0c14cf4f0db7221ddf70 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Fri, 17 Oct 2025 17:03:28 -0700 Subject: [PATCH 010/647] Disable torch compiler for cast_bias_weight function (#10384) * Disable torch compiler for cast_bias_weight function * Fix torch compile. --- comfy/ops.py | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/comfy/ops.py b/comfy/ops.py index 56b07b44c..5feeb3571 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -25,6 +25,9 @@ import comfy.rmsnorm import contextlib def run_every_op(): + if torch.compiler.is_compiling(): + return + comfy.model_management.throw_exception_if_processing_interrupted() def scaled_dot_product_attention(q, k, v, *args, **kwargs): @@ -70,6 +73,7 @@ if torch.cuda.is_available() and torch.backends.cudnn.is_available() and Perform def cast_to_input(weight, input, non_blocking=False, copy=True): return comfy.model_management.cast_to(weight, input.dtype, input.device, non_blocking=non_blocking, copy=copy) +@torch.compiler.disable() def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None): if input is not None: if dtype is None: From 5b80addafd24bda5b2f9f7a35e32dbd40823c3fd Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sat, 18 Oct 2025 19:35:46 -0700 Subject: [PATCH 011/647] Turn off cuda malloc by default when --fast autotune is turned on. (#10393) --- comfy/model_management.py | 3 +++ comfy/ops.py | 3 --- cuda_malloc.py | 7 ++++--- 3 files changed, 7 insertions(+), 6 deletions(-) diff --git a/comfy/model_management.py b/comfy/model_management.py index d82d5b8b0..7467391cd 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -371,6 +371,9 @@ try: except: pass +if torch.cuda.is_available() and torch.backends.cudnn.is_available() and PerformanceFeature.AutoTune in args.fast: + torch.backends.cudnn.benchmark = True + try: if torch_version_numeric >= (2, 5): torch.backends.cuda.allow_fp16_bf16_reduction_math_sdp(True) diff --git a/comfy/ops.py b/comfy/ops.py index 5feeb3571..967134f05 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -67,9 +67,6 @@ except: cast_to = comfy.model_management.cast_to #TODO: remove once no more references -if torch.cuda.is_available() and torch.backends.cudnn.is_available() and PerformanceFeature.AutoTune in args.fast: - torch.backends.cudnn.benchmark = True - def cast_to_input(weight, input, non_blocking=False, copy=True): return comfy.model_management.cast_to(weight, input.dtype, input.device, non_blocking=non_blocking, copy=copy) diff --git a/cuda_malloc.py b/cuda_malloc.py index c1d9ae3ca..6520d5123 100644 --- a/cuda_malloc.py +++ b/cuda_malloc.py @@ -1,6 +1,6 @@ import os import importlib.util -from comfy.cli_args import args +from comfy.cli_args import args, PerformanceFeature import subprocess #Can't use pytorch to get the GPU names because the cuda malloc has to be set before the first import. @@ -75,8 +75,9 @@ if not args.cuda_malloc: spec.loader.exec_module(module) version = module.__version__ - if int(version[0]) >= 2 and "+cu" in version: #enable by default for torch version 2.0 and up only on cuda torch - args.cuda_malloc = cuda_malloc_supported() + if int(version[0]) >= 2 and "+cu" in version: # enable by default for torch version 2.0 and up only on cuda torch + if PerformanceFeature.AutoTune not in args.fast: # Autotune has issues with cuda malloc + args.cuda_malloc = cuda_malloc_supported() except: pass From 0cf33953a7c951d163088cbfe36c55d1cdf8a718 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sat, 18 Oct 2025 20:15:34 -0700 Subject: [PATCH 012/647] Fix batch size above 1 giving bad output in chroma radiance. (#10394) --- comfy/ldm/chroma_radiance/model.py | 23 +++++++---------------- 1 file changed, 7 insertions(+), 16 deletions(-) diff --git a/comfy/ldm/chroma_radiance/model.py b/comfy/ldm/chroma_radiance/model.py index 47aa11b04..7d7be80f5 100644 --- a/comfy/ldm/chroma_radiance/model.py +++ b/comfy/ldm/chroma_radiance/model.py @@ -189,15 +189,15 @@ class ChromaRadiance(Chroma): nerf_pixels = nn.functional.unfold(img_orig, kernel_size=patch_size, stride=patch_size) nerf_pixels = nerf_pixels.transpose(1, 2) # -> [B, NumPatches, C * P * P] + # Reshape for per-patch processing + nerf_hidden = img_out.reshape(B * num_patches, params.hidden_size) + nerf_pixels = nerf_pixels.reshape(B * num_patches, C, patch_size**2).transpose(1, 2) + if params.nerf_tile_size > 0 and num_patches > params.nerf_tile_size: # Enable tiling if nerf_tile_size isn't 0 and we actually have more patches than # the tile size. - img_dct = self.forward_tiled_nerf(img_out, nerf_pixels, B, C, num_patches, patch_size, params) + img_dct = self.forward_tiled_nerf(nerf_hidden, nerf_pixels, B, C, num_patches, patch_size, params) else: - # Reshape for per-patch processing - nerf_hidden = img_out.reshape(B * num_patches, params.hidden_size) - nerf_pixels = nerf_pixels.reshape(B * num_patches, C, patch_size**2).transpose(1, 2) - # Get DCT-encoded pixel embeddings [pixel-dct] img_dct = self.nerf_image_embedder(nerf_pixels) @@ -240,17 +240,8 @@ class ChromaRadiance(Chroma): end = min(i + tile_size, num_patches) # Slice the current tile from the input tensors - nerf_hidden_tile = nerf_hidden[:, i:end, :] - nerf_pixels_tile = nerf_pixels[:, i:end, :] - - # Get the actual number of patches in this tile (can be smaller for the last tile) - num_patches_tile = nerf_hidden_tile.shape[1] - - # Reshape the tile for per-patch processing - # [B, NumPatches_tile, D] -> [B * NumPatches_tile, D] - nerf_hidden_tile = nerf_hidden_tile.reshape(batch * num_patches_tile, params.hidden_size) - # [B, NumPatches_tile, C*P*P] -> [B*NumPatches_tile, C, P*P] -> [B*NumPatches_tile, P*P, C] - nerf_pixels_tile = nerf_pixels_tile.reshape(batch * num_patches_tile, channels, patch_size**2).transpose(1, 2) + nerf_hidden_tile = nerf_hidden[i * batch:end * batch] + nerf_pixels_tile = nerf_pixels[i * batch:end * batch] # get DCT-encoded pixel embeddings [pixel-dct] img_dct_tile = self.nerf_image_embedder(nerf_pixels_tile) From dad076aee68ab676fb390d9663ab9e343824a080 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sat, 18 Oct 2025 20:19:52 -0700 Subject: [PATCH 013/647] Speed up chroma radiance. (#10395) --- comfy/model_detection.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy/model_detection.py b/comfy/model_detection.py index 7677617c0..141f1e164 100644 --- a/comfy/model_detection.py +++ b/comfy/model_detection.py @@ -213,7 +213,7 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): dit_config["nerf_mlp_ratio"] = 4 dit_config["nerf_depth"] = 4 dit_config["nerf_max_freqs"] = 8 - dit_config["nerf_tile_size"] = 32 + dit_config["nerf_tile_size"] = 512 dit_config["nerf_final_head_type"] = "conv" if f"{key_prefix}nerf_final_layer_conv.norm.scale" in state_dict_keys else "linear" dit_config["nerf_embedder_dtype"] = torch.float32 else: From b4f30bd4087a79b4c4fc89bb67b9889adb866294 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sat, 18 Oct 2025 22:25:35 -0700 Subject: [PATCH 014/647] Pytorch is stupid. (#10398) --- comfy/ops.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy/ops.py b/comfy/ops.py index 967134f05..934e21261 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -58,7 +58,7 @@ except (ModuleNotFoundError, TypeError): NVIDIA_MEMORY_CONV_BUG_WORKAROUND = False try: if comfy.model_management.is_nvidia(): - if torch.backends.cudnn.version() >= 91200 and comfy.model_management.torch_version_numeric >= (2, 9) and comfy.model_management.torch_version_numeric <= (2, 10): + if torch.backends.cudnn.version() >= 91002 and comfy.model_management.torch_version_numeric >= (2, 9) and comfy.model_management.torch_version_numeric <= (2, 10): #TODO: change upper bound version once it's fixed' NVIDIA_MEMORY_CONV_BUG_WORKAROUND = True logging.info("working around nvidia conv3d memory bug.") From b5c59b763c6b14e1362ec4274b09eca4f3f7091b Mon Sep 17 00:00:00 2001 From: Christian Byrne Date: Sun, 19 Oct 2025 13:05:46 -0700 Subject: [PATCH 015/647] Deprecation warning on unused files (#10387) * only warn for unused files * include internal extensions --- server.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/server.py b/server.py index a44f4f237..10c2698b5 100644 --- a/server.py +++ b/server.py @@ -56,7 +56,7 @@ async def deprecation_warning(request: web.Request, handler): """Middleware to warn about deprecated frontend API paths""" path = request.path - if (path.startswith('/scripts/') or path.startswith('/extensions/core/')): + if path.startswith("/scripts/ui") or path.startswith("/extensions/core/"): # Only warn once per unique file path if path not in _deprecated_paths_warned: _deprecated_paths_warned.add(path) From a4787ac83bf6c83eeb459ed80fc9b36f63d2a3a7 Mon Sep 17 00:00:00 2001 From: ComfyUI Wiki Date: Tue, 21 Oct 2025 03:28:36 +0800 Subject: [PATCH 016/647] Update template to 0.2.1 (#10413) * Update template to 0.1.97 * Update template to 0.2.1 --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 82457df54..dd2afcab0 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,5 +1,5 @@ comfyui-frontend-package==1.28.7 -comfyui-workflow-templates==0.1.95 +comfyui-workflow-templates==0.2.1 comfyui-embedded-docs==0.3.0 torch torchsde From 2c2aa409b01f513de88d2245931e5836ed1cd718 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 20 Oct 2025 12:43:24 -0700 Subject: [PATCH 017/647] Log message for cudnn disable on AMD. (#10418) --- comfy/model_management.py | 1 + 1 file changed, 1 insertion(+) diff --git a/comfy/model_management.py b/comfy/model_management.py index 7467391cd..a2c318ec3 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -333,6 +333,7 @@ SUPPORT_FP8_OPS = args.supports_fp8_compute try: if is_amd(): torch.backends.cudnn.enabled = False # Seems to improve things a lot on AMD + logging.info("Set: torch.backends.cudnn.enabled = False for better AMD performance.") try: rocm_version = tuple(map(int, str(torch.version.hip).split(".")[:2])) except: From b7992f871af38d89a459080caa57cc359ed93a46 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 20 Oct 2025 16:03:06 -0700 Subject: [PATCH 018/647] =?UTF-8?q?Revert=20"execution:=20fold=20in=20depe?= =?UTF-8?q?ndency=20aware=20caching=20/=20Fix=20--cache-none=20with=20l?= =?UTF-8?q?=E2=80=A6"=20(#10422)?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit This reverts commit b1467da4803017a418c32c159525767f45871ca3. --- comfy_execution/caching.py | 174 ++++++++++++++++++++++++++---- comfy_execution/graph.py | 31 +----- execution.py | 34 +++--- main.py | 2 +- tests/execution/test_execution.py | 50 ++++----- 5 files changed, 190 insertions(+), 101 deletions(-) diff --git a/comfy_execution/caching.py b/comfy_execution/caching.py index 566bc3f9c..41224ce3b 100644 --- a/comfy_execution/caching.py +++ b/comfy_execution/caching.py @@ -265,26 +265,6 @@ class HierarchicalCache(BasicCache): assert cache is not None return await cache._ensure_subcache(node_id, children_ids) -class NullCache: - - async def set_prompt(self, dynprompt, node_ids, is_changed_cache): - pass - - def all_node_ids(self): - return [] - - def clean_unused(self): - pass - - def get(self, node_id): - return None - - def set(self, node_id, value): - pass - - async def ensure_subcache_for(self, node_id, children_ids): - return self - class LRUCache(BasicCache): def __init__(self, key_class, max_size=100): super().__init__(key_class) @@ -336,3 +316,157 @@ class LRUCache(BasicCache): self._mark_used(child_id) self.children[cache_key].append(self.cache_key_set.get_data_key(child_id)) return self + + +class DependencyAwareCache(BasicCache): + """ + A cache implementation that tracks dependencies between nodes and manages + their execution and caching accordingly. It extends the BasicCache class. + Nodes are removed from this cache once all of their descendants have been + executed. + """ + + def __init__(self, key_class): + """ + Initialize the DependencyAwareCache. + + Args: + key_class: The class used for generating cache keys. + """ + super().__init__(key_class) + self.descendants = {} # Maps node_id -> set of descendant node_ids + self.ancestors = {} # Maps node_id -> set of ancestor node_ids + self.executed_nodes = set() # Tracks nodes that have been executed + + async def set_prompt(self, dynprompt, node_ids, is_changed_cache): + """ + Clear the entire cache and rebuild the dependency graph. + + Args: + dynprompt: The dynamic prompt object containing node information. + node_ids: List of node IDs to initialize the cache for. + is_changed_cache: Flag indicating if the cache has changed. + """ + # Clear all existing cache data + self.cache.clear() + self.subcaches.clear() + self.descendants.clear() + self.ancestors.clear() + self.executed_nodes.clear() + + # Call the parent method to initialize the cache with the new prompt + await super().set_prompt(dynprompt, node_ids, is_changed_cache) + + # Rebuild the dependency graph + self._build_dependency_graph(dynprompt, node_ids) + + def _build_dependency_graph(self, dynprompt, node_ids): + """ + Build the dependency graph for all nodes. + + Args: + dynprompt: The dynamic prompt object containing node information. + node_ids: List of node IDs to build the graph for. + """ + self.descendants.clear() + self.ancestors.clear() + for node_id in node_ids: + self.descendants[node_id] = set() + self.ancestors[node_id] = set() + + for node_id in node_ids: + inputs = dynprompt.get_node(node_id)["inputs"] + for input_data in inputs.values(): + if is_link(input_data): # Check if the input is a link to another node + ancestor_id = input_data[0] + self.descendants[ancestor_id].add(node_id) + self.ancestors[node_id].add(ancestor_id) + + def set(self, node_id, value): + """ + Mark a node as executed and store its value in the cache. + + Args: + node_id: The ID of the node to store. + value: The value to store for the node. + """ + self._set_immediate(node_id, value) + self.executed_nodes.add(node_id) + self._cleanup_ancestors(node_id) + + def get(self, node_id): + """ + Retrieve the cached value for a node. + + Args: + node_id: The ID of the node to retrieve. + + Returns: + The cached value for the node. + """ + return self._get_immediate(node_id) + + async def ensure_subcache_for(self, node_id, children_ids): + """ + Ensure a subcache exists for a node and update dependencies. + + Args: + node_id: The ID of the parent node. + children_ids: List of child node IDs to associate with the parent node. + + Returns: + The subcache object for the node. + """ + subcache = await super()._ensure_subcache(node_id, children_ids) + for child_id in children_ids: + self.descendants[node_id].add(child_id) + self.ancestors[child_id].add(node_id) + return subcache + + def _cleanup_ancestors(self, node_id): + """ + Check if ancestors of a node can be removed from the cache. + + Args: + node_id: The ID of the node whose ancestors are to be checked. + """ + for ancestor_id in self.ancestors.get(node_id, []): + if ancestor_id in self.executed_nodes: + # Remove ancestor if all its descendants have been executed + if all(descendant in self.executed_nodes for descendant in self.descendants[ancestor_id]): + self._remove_node(ancestor_id) + + def _remove_node(self, node_id): + """ + Remove a node from the cache. + + Args: + node_id: The ID of the node to remove. + """ + cache_key = self.cache_key_set.get_data_key(node_id) + if cache_key in self.cache: + del self.cache[cache_key] + subcache_key = self.cache_key_set.get_subcache_key(node_id) + if subcache_key in self.subcaches: + del self.subcaches[subcache_key] + + def clean_unused(self): + """ + Clean up unused nodes. This is a no-op for this cache implementation. + """ + pass + + def recursive_debug_dump(self): + """ + Dump the cache and dependency graph for debugging. + + Returns: + A list containing the cache state and dependency graph. + """ + result = super().recursive_debug_dump() + result.append({ + "descendants": self.descendants, + "ancestors": self.ancestors, + "executed_nodes": list(self.executed_nodes), + }) + return result diff --git a/comfy_execution/graph.py b/comfy_execution/graph.py index d5bbacde3..f4b427265 100644 --- a/comfy_execution/graph.py +++ b/comfy_execution/graph.py @@ -153,9 +153,8 @@ class TopologicalSort: continue _, _, input_info = self.get_input_info(unique_id, input_name) is_lazy = input_info is not None and "lazy" in input_info and input_info["lazy"] - if (include_lazy or not is_lazy): - if not self.is_cached(from_node_id): - node_ids.append(from_node_id) + if (include_lazy or not is_lazy) and not self.is_cached(from_node_id): + node_ids.append(from_node_id) links.append((from_node_id, from_socket, unique_id)) for link in links: @@ -195,34 +194,10 @@ class ExecutionList(TopologicalSort): super().__init__(dynprompt) self.output_cache = output_cache self.staged_node_id = None - self.execution_cache = {} - self.execution_cache_listeners = {} def is_cached(self, node_id): return self.output_cache.get(node_id) is not None - def cache_link(self, from_node_id, to_node_id): - if not to_node_id in self.execution_cache: - self.execution_cache[to_node_id] = {} - self.execution_cache[to_node_id][from_node_id] = self.output_cache.get(from_node_id) - if not from_node_id in self.execution_cache_listeners: - self.execution_cache_listeners[from_node_id] = set() - self.execution_cache_listeners[from_node_id].add(to_node_id) - - def get_output_cache(self, from_node_id, to_node_id): - if not to_node_id in self.execution_cache: - return None - return self.execution_cache[to_node_id].get(from_node_id) - - def cache_update(self, node_id, value): - if node_id in self.execution_cache_listeners: - for to_node_id in self.execution_cache_listeners[node_id]: - self.execution_cache[to_node_id][node_id] = value - - def add_strong_link(self, from_node_id, from_socket, to_node_id): - super().add_strong_link(from_node_id, from_socket, to_node_id) - self.cache_link(from_node_id, to_node_id) - async def stage_node_execution(self): assert self.staged_node_id is None if self.is_empty(): @@ -302,8 +277,6 @@ class ExecutionList(TopologicalSort): def complete_node_execution(self): node_id = self.staged_node_id self.pop_node(node_id) - self.execution_cache.pop(node_id, None) - self.execution_cache_listeners.pop(node_id, None) self.staged_node_id = None def get_nodes_in_cycle(self): diff --git a/execution.py b/execution.py index 78c36a4b0..1dc35738b 100644 --- a/execution.py +++ b/execution.py @@ -18,7 +18,7 @@ from comfy_execution.caching import ( BasicCache, CacheKeySetID, CacheKeySetInputSignature, - NullCache, + DependencyAwareCache, HierarchicalCache, LRUCache, ) @@ -91,13 +91,13 @@ class IsChangedCache: class CacheType(Enum): CLASSIC = 0 LRU = 1 - NONE = 2 + DEPENDENCY_AWARE = 2 class CacheSet: def __init__(self, cache_type=None, cache_size=None): - if cache_type == CacheType.NONE: - self.init_null_cache() + if cache_type == CacheType.DEPENDENCY_AWARE: + self.init_dependency_aware_cache() logging.info("Disabling intermediate node cache.") elif cache_type == CacheType.LRU: if cache_size is None: @@ -120,12 +120,11 @@ class CacheSet: self.ui = LRUCache(CacheKeySetInputSignature, max_size=cache_size) self.objects = HierarchicalCache(CacheKeySetID) - def init_null_cache(self): - self.outputs = NullCache() - #The UI cache is expected to be iterable at the end of each workflow - #so it must cache at least a full workflow. Use Heirachical - self.ui = HierarchicalCache(CacheKeySetInputSignature) - self.objects = NullCache() + # only hold cached items while the decendents have not executed + def init_dependency_aware_cache(self): + self.outputs = DependencyAwareCache(CacheKeySetInputSignature) + self.ui = DependencyAwareCache(CacheKeySetInputSignature) + self.objects = DependencyAwareCache(CacheKeySetID) def recursive_debug_dump(self): result = { @@ -136,7 +135,7 @@ class CacheSet: SENSITIVE_EXTRA_DATA_KEYS = ("auth_token_comfy_org", "api_key_comfy_org") -def get_input_data(inputs, class_def, unique_id, execution_list=None, dynprompt=None, extra_data={}): +def get_input_data(inputs, class_def, unique_id, outputs=None, dynprompt=None, extra_data={}): is_v3 = issubclass(class_def, _ComfyNodeInternal) if is_v3: valid_inputs, schema = class_def.INPUT_TYPES(include_hidden=False, return_schema=True) @@ -154,10 +153,10 @@ def get_input_data(inputs, class_def, unique_id, execution_list=None, dynprompt= if is_link(input_data) and (not input_info or not input_info.get("rawLink", False)): input_unique_id = input_data[0] output_index = input_data[1] - if execution_list is None: + if outputs is None: mark_missing() continue # This might be a lazily-evaluated input - cached_output = execution_list.get_output_cache(input_unique_id, unique_id) + cached_output = outputs.get(input_unique_id) if cached_output is None: mark_missing() continue @@ -406,7 +405,6 @@ async def execute(server, dynprompt, caches, current_item, extra_data, executed, cached_output = caches.ui.get(unique_id) or {} server.send_sync("executed", { "node": unique_id, "display_node": display_node_id, "output": cached_output.get("output",None), "prompt_id": prompt_id }, server.client_id) get_progress_state().finish_progress(unique_id) - execution_list.cache_update(unique_id, caches.outputs.get(unique_id)) return (ExecutionResult.SUCCESS, None, None) input_data_all = None @@ -436,7 +434,7 @@ async def execute(server, dynprompt, caches, current_item, extra_data, executed, for r in result: if is_link(r): source_node, source_output = r[0], r[1] - node_output = execution_list.get_output_cache(source_node, unique_id)[source_output] + node_output = caches.outputs.get(source_node)[source_output] for o in node_output: resolved_output.append(o) @@ -448,7 +446,7 @@ async def execute(server, dynprompt, caches, current_item, extra_data, executed, has_subgraph = False else: get_progress_state().start_progress(unique_id) - input_data_all, missing_keys, hidden_inputs = get_input_data(inputs, class_def, unique_id, execution_list, dynprompt, extra_data) + input_data_all, missing_keys, hidden_inputs = get_input_data(inputs, class_def, unique_id, caches.outputs, dynprompt, extra_data) if server.client_id is not None: server.last_node_id = display_node_id server.send_sync("executing", { "node": unique_id, "display_node": display_node_id, "prompt_id": prompt_id }, server.client_id) @@ -551,15 +549,11 @@ async def execute(server, dynprompt, caches, current_item, extra_data, executed, subcache.clean_unused() for node_id in new_output_ids: execution_list.add_node(node_id) - execution_list.cache_link(node_id, unique_id) for link in new_output_links: execution_list.add_strong_link(link[0], link[1], unique_id) pending_subgraph_results[unique_id] = cached_outputs return (ExecutionResult.PENDING, None, None) - caches.outputs.set(unique_id, output_data) - execution_list.cache_update(unique_id, output_data) - except comfy.model_management.InterruptProcessingException as iex: logging.info("Processing interrupted") diff --git a/main.py b/main.py index 4b4c5dcc4..35857dba8 100644 --- a/main.py +++ b/main.py @@ -173,7 +173,7 @@ def prompt_worker(q, server_instance): if args.cache_lru > 0: cache_type = execution.CacheType.LRU elif args.cache_none: - cache_type = execution.CacheType.NONE + cache_type = execution.CacheType.DEPENDENCY_AWARE e = execution.PromptExecutor(server_instance, cache_type=cache_type, cache_size=args.cache_lru) last_gc_collect = 0 diff --git a/tests/execution/test_execution.py b/tests/execution/test_execution.py index ace0d2279..ef73ad9fd 100644 --- a/tests/execution/test_execution.py +++ b/tests/execution/test_execution.py @@ -152,12 +152,12 @@ class TestExecution: # Initialize server and client # @fixture(scope="class", autouse=True, params=[ - { "extra_args" : [], "should_cache_results" : True }, - { "extra_args" : ["--cache-lru", 0], "should_cache_results" : True }, - { "extra_args" : ["--cache-lru", 100], "should_cache_results" : True }, - { "extra_args" : ["--cache-none"], "should_cache_results" : False }, + # (use_lru, lru_size) + (False, 0), + (True, 0), + (True, 100), ]) - def server(self, args_pytest, request): + def _server(self, args_pytest, request): # Start server pargs = [ 'python','main.py', @@ -167,10 +167,12 @@ class TestExecution: '--extra-model-paths-config', 'tests/execution/extra_model_paths.yaml', '--cpu', ] - pargs += [ str(param) for param in request.param["extra_args"] ] + use_lru, lru_size = request.param + if use_lru: + pargs += ['--cache-lru', str(lru_size)] print("Running server with args:", pargs) # noqa: T201 p = subprocess.Popen(pargs) - yield request.param + yield p.kill() torch.cuda.empty_cache() @@ -191,7 +193,7 @@ class TestExecution: return comfy_client @fixture(scope="class", autouse=True) - def shared_client(self, args_pytest, server): + def shared_client(self, args_pytest, _server): client = self.start_client(args_pytest["listen"], args_pytest["port"]) yield client del client @@ -223,7 +225,7 @@ class TestExecution: assert result.did_run(mask) assert result.did_run(lazy_mix) - def test_full_cache(self, client: ComfyClient, builder: GraphBuilder, server): + def test_full_cache(self, client: ComfyClient, builder: GraphBuilder): g = builder input1 = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1) input2 = g.node("StubImage", content="NOISE", height=512, width=512, batch_size=1) @@ -235,12 +237,9 @@ class TestExecution: client.run(g) result2 = client.run(g) for node_id, node in g.nodes.items(): - if server["should_cache_results"]: - assert not result2.did_run(node), f"Node {node_id} ran, but should have been cached" - else: - assert result2.did_run(node), f"Node {node_id} was cached, but should have been run" + assert not result2.did_run(node), f"Node {node_id} ran, but should have been cached" - def test_partial_cache(self, client: ComfyClient, builder: GraphBuilder, server): + def test_partial_cache(self, client: ComfyClient, builder: GraphBuilder): g = builder input1 = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1) input2 = g.node("StubImage", content="NOISE", height=512, width=512, batch_size=1) @@ -252,12 +251,8 @@ class TestExecution: client.run(g) mask.inputs['value'] = 0.4 result2 = client.run(g) - if server["should_cache_results"]: - assert not result2.did_run(input1), "Input1 should have been cached" - assert not result2.did_run(input2), "Input2 should have been cached" - else: - assert result2.did_run(input1), "Input1 should have been rerun" - assert result2.did_run(input2), "Input2 should have been rerun" + assert not result2.did_run(input1), "Input1 should have been cached" + assert not result2.did_run(input2), "Input2 should have been cached" def test_error(self, client: ComfyClient, builder: GraphBuilder): g = builder @@ -416,7 +411,7 @@ class TestExecution: input2 = g.node("StubImage", id="removeme", content="WHITE", height=512, width=512, batch_size=1) client.run(g) - def test_custom_is_changed(self, client: ComfyClient, builder: GraphBuilder, server): + def test_custom_is_changed(self, client: ComfyClient, builder: GraphBuilder): g = builder # Creating the nodes in this specific order previously caused a bug save = g.node("SaveImage") @@ -432,10 +427,7 @@ class TestExecution: result3 = client.run(g) result4 = client.run(g) assert result1.did_run(is_changed), "is_changed should have been run" - if server["should_cache_results"]: - assert not result2.did_run(is_changed), "is_changed should have been cached" - else: - assert result2.did_run(is_changed), "is_changed should have been re-run" + assert not result2.did_run(is_changed), "is_changed should have been cached" assert result3.did_run(is_changed), "is_changed should have been re-run" assert result4.did_run(is_changed), "is_changed should not have been cached" @@ -522,7 +514,7 @@ class TestExecution: assert len(images2) == 1, "Should have 1 image" # This tests that only constant outputs are used in the call to `IS_CHANGED` - def test_is_changed_with_outputs(self, client: ComfyClient, builder: GraphBuilder, server): + def test_is_changed_with_outputs(self, client: ComfyClient, builder: GraphBuilder): g = builder input1 = g.node("StubConstantImage", value=0.5, height=512, width=512, batch_size=1) test_node = g.node("TestIsChangedWithConstants", image=input1.out(0), value=0.5) @@ -538,11 +530,7 @@ class TestExecution: images = result.get_images(output) assert len(images) == 1, "Should have 1 image" assert numpy.array(images[0]).min() == 63 and numpy.array(images[0]).max() == 63, "Image should have value 0.25" - if server["should_cache_results"]: - assert not result.did_run(test_node), "The execution should have been cached" - else: - assert result.did_run(test_node), "The execution should have been re-run" - + assert not result.did_run(test_node), "The execution should have been cached" def test_parallel_sleep_nodes(self, client: ComfyClient, builder: GraphBuilder, skip_timing_checks): # Warmup execution to ensure server is fully initialized From 560b1bdfca77d9441ca2924fd9d6baa8dda05cd7 Mon Sep 17 00:00:00 2001 From: comfyanonymous Date: Mon, 20 Oct 2025 15:44:38 -0400 Subject: [PATCH 019/647] ComfyUI version v0.3.66 --- comfyui_version.py | 2 +- pyproject.toml | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/comfyui_version.py b/comfyui_version.py index d39c1fdc4..33a06bbb0 100644 --- a/comfyui_version.py +++ b/comfyui_version.py @@ -1,3 +1,3 @@ # This file is automatically generated by the build process when version is # updated in pyproject.toml. -__version__ = "0.3.65" +__version__ = "0.3.66" diff --git a/pyproject.toml b/pyproject.toml index 653604e24..0c6b23a25 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -1,6 +1,6 @@ [project] name = "ComfyUI" -version = "0.3.65" +version = "0.3.66" readme = "README.md" license = { file = "LICENSE" } requires-python = ">=3.9" From 9cdc64998f8990aed7688b0ebe89bc3b97733764 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 21 Oct 2025 16:15:23 -0700 Subject: [PATCH 020/647] Only disable cudnn on newer AMD GPUs. (#10437) --- comfy/model_management.py | 14 ++++++++++---- 1 file changed, 10 insertions(+), 4 deletions(-) diff --git a/comfy/model_management.py b/comfy/model_management.py index a2c318ec3..79d6ff9d4 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -330,15 +330,21 @@ except: SUPPORT_FP8_OPS = args.supports_fp8_compute + +AMD_RDNA2_AND_OLDER_ARCH = ["gfx1030", "gfx1031", "gfx1010", "gfx1011", "gfx1012", "gfx906", "gfx900", "gfx803"] + try: if is_amd(): - torch.backends.cudnn.enabled = False # Seems to improve things a lot on AMD - logging.info("Set: torch.backends.cudnn.enabled = False for better AMD performance.") + arch = torch.cuda.get_device_properties(get_torch_device()).gcnArchName + if not (any((a in arch) for a in AMD_RDNA2_AND_OLDER_ARCH)): + torch.backends.cudnn.enabled = False # Seems to improve things a lot on AMD + logging.info("Set: torch.backends.cudnn.enabled = False for better AMD performance.") + try: rocm_version = tuple(map(int, str(torch.version.hip).split(".")[:2])) except: rocm_version = (6, -1) - arch = torch.cuda.get_device_properties(get_torch_device()).gcnArchName + logging.info("AMD arch: {}".format(arch)) logging.info("ROCm version: {}".format(rocm_version)) if args.use_split_cross_attention == False and args.use_quad_cross_attention == False: @@ -1331,7 +1337,7 @@ def should_use_bf16(device=None, model_params=0, prioritize_performance=True, ma if is_amd(): arch = torch.cuda.get_device_properties(device).gcnArchName - if any((a in arch) for a in ["gfx1030", "gfx1031", "gfx1010", "gfx1011", "gfx1012", "gfx906", "gfx900", "gfx803"]): # RDNA2 and older don't support bf16 + if any((a in arch) for a in AMD_RDNA2_AND_OLDER_ARCH): # RDNA2 and older don't support bf16 if manual_cast: return True return False From f13cff0be65e35d34876b173bba2fec6bd94746b Mon Sep 17 00:00:00 2001 From: Jedrzej Kosinski Date: Tue, 21 Oct 2025 20:16:16 -0700 Subject: [PATCH 021/647] Add custom node published subgraphs endpoint (#10438) * Add get_subgraphs_dir to ComfyExtension and PUBLISHED_SUBGRAPH_DIRS to nodes.py * Created initial endpoints, although the returned paths are a bit off currently * Fix path and actually return real data * Sanitize returned /api/global_subgraphs entries * Remove leftover function from early prototyping * Remove added whitespace * Add None check for sanitize_entry --- app/subgraph_manager.py | 112 ++++++++++++++++++++++++++++++++++++++++ server.py | 3 ++ 2 files changed, 115 insertions(+) create mode 100644 app/subgraph_manager.py diff --git a/app/subgraph_manager.py b/app/subgraph_manager.py new file mode 100644 index 000000000..dbe404541 --- /dev/null +++ b/app/subgraph_manager.py @@ -0,0 +1,112 @@ +from __future__ import annotations + +from typing import TypedDict +import os +import folder_paths +import glob +from aiohttp import web +import hashlib + + +class Source: + custom_node = "custom_node" + +class SubgraphEntry(TypedDict): + source: str + """ + Source of subgraph - custom_nodes vs templates. + """ + path: str + """ + Relative path of the subgraph file. + For custom nodes, will be the relative directory like /subgraphs/.json + """ + name: str + """ + Name of subgraph file. + """ + info: CustomNodeSubgraphEntryInfo + """ + Additional info about subgraph; in the case of custom_nodes, will contain nodepack name + """ + data: str + +class CustomNodeSubgraphEntryInfo(TypedDict): + node_pack: str + """Node pack name.""" + +class SubgraphManager: + def __init__(self): + self.cached_custom_node_subgraphs: dict[SubgraphEntry] | None = None + + async def load_entry_data(self, entry: SubgraphEntry): + with open(entry['path'], 'r') as f: + entry['data'] = f.read() + return entry + + async def sanitize_entry(self, entry: SubgraphEntry | None, remove_data=False) -> SubgraphEntry | None: + if entry is None: + return None + entry = entry.copy() + entry.pop('path', None) + if remove_data: + entry.pop('data', None) + return entry + + async def sanitize_entries(self, entries: dict[str, SubgraphEntry], remove_data=False) -> dict[str, SubgraphEntry]: + entries = entries.copy() + for key in list(entries.keys()): + entries[key] = await self.sanitize_entry(entries[key], remove_data) + return entries + + async def get_custom_node_subgraphs(self, loadedModules, force_reload=False): + # if not forced to reload and cached, return cache + if not force_reload and self.cached_custom_node_subgraphs is not None: + return self.cached_custom_node_subgraphs + # Load subgraphs from custom nodes + subfolder = "subgraphs" + subgraphs_dict: dict[SubgraphEntry] = {} + + for folder in folder_paths.get_folder_paths("custom_nodes"): + pattern = os.path.join(folder, f"*/{subfolder}/*.json") + matched_files = glob.glob(pattern) + for file in matched_files: + # replace backslashes with forward slashes + file = file.replace('\\', '/') + info: CustomNodeSubgraphEntryInfo = { + "node_pack": "custom_nodes." + file.split('/')[-3] + } + source = Source.custom_node + # hash source + path to make sure id will be as unique as possible, but + # reproducible across backend reloads + id = hashlib.sha256(f"{source}{file}".encode()).hexdigest() + entry: SubgraphEntry = { + "source": Source.custom_node, + "name": os.path.splitext(os.path.basename(file))[0], + "path": file, + "info": info, + } + subgraphs_dict[id] = entry + self.cached_custom_node_subgraphs = subgraphs_dict + return subgraphs_dict + + async def get_custom_node_subgraph(self, id: str, loadedModules): + subgraphs = await self.get_custom_node_subgraphs(loadedModules) + entry: SubgraphEntry = subgraphs.get(id, None) + if entry is not None and entry.get('data', None) is None: + await self.load_entry_data(entry) + return entry + + def add_routes(self, routes, loadedModules): + @routes.get("/global_subgraphs") + async def get_global_subgraphs(request): + subgraphs_dict = await self.get_custom_node_subgraphs(loadedModules) + # NOTE: we may want to include other sources of global subgraphs such as templates in the future; + # that's the reasoning for the current implementation + return web.json_response(await self.sanitize_entries(subgraphs_dict, remove_data=True)) + + @routes.get("/global_subgraphs/{id}") + async def get_global_subgraph(request): + id = request.match_info.get("id", None) + subgraph = await self.get_custom_node_subgraph(id, loadedModules) + return web.json_response(await self.sanitize_entry(subgraph)) diff --git a/server.py b/server.py index 10c2698b5..fe58db286 100644 --- a/server.py +++ b/server.py @@ -35,6 +35,7 @@ from comfy_api.internal import _ComfyNodeInternal from app.user_manager import UserManager from app.model_manager import ModelFileManager from app.custom_node_manager import CustomNodeManager +from app.subgraph_manager import SubgraphManager from typing import Optional, Union from api_server.routes.internal.internal_routes import InternalRoutes from protocol import BinaryEventTypes @@ -173,6 +174,7 @@ class PromptServer(): self.user_manager = UserManager() self.model_file_manager = ModelFileManager() self.custom_node_manager = CustomNodeManager() + self.subgraph_manager = SubgraphManager() self.internal_routes = InternalRoutes(self) self.supports = ["custom_nodes_from_web"] self.prompt_queue = execution.PromptQueue(self) @@ -819,6 +821,7 @@ class PromptServer(): self.user_manager.add_routes(self.routes) self.model_file_manager.add_routes(self.routes) self.custom_node_manager.add_routes(self.routes, self.app, nodes.LOADED_MODULE_DIRS.items()) + self.subgraph_manager.add_routes(self.routes, nodes.LOADED_MODULE_DIRS.items()) self.app.add_subapp('/internal', self.internal_routes.get_app()) # Prefix every route with /api for easier matching for delegation. From 4739d7717fea56750d0ef98c64268d9c1e487d78 Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Thu, 23 Oct 2025 05:49:05 +1000 Subject: [PATCH 022/647] execution: fold in dependency aware caching / Fix --cache-none with loops/lazy etc (Resubmit) (#10440) * execution: fold in dependency aware caching This makes --cache-none compatiable with lazy and expanded subgraphs. Currently the --cache-none option is powered by the DependencyAwareCache. The cache attempts to maintain a parallel copy of the execution list data structure, however it is only setup once at the start of execution and does not get meaninigful updates to the execution list. This causes multiple problems when --cache-none is used with lazy and expanded subgraphs as the DAC does not accurately update its copy of the execution data structure. DAC has an attempt to handle subgraphs ensure_subcache however this does not accurately connect to nodes outside the subgraph. The current semantics of DAC are to free a node ASAP after the dependent nodes are executed. This means that if a subgraph refs such a node it will be requed and re-executed by the execution_list but DAC wont see it in its to-free lists anymore and leak memory. Rather than try and cover all the cases where the execution list changes from inside the cache, move the while problem to the executor which maintains an always up-to-date copy of the wanted data-structure. The executor now has a fast-moving run-local cache of its own. Each _to node has its own mini cache, and the cache is unconditionally primed at the time of add_strong_link. add_strong_link is called for all of static workflows, lazy links and expanded subgraphs so its the singular source of truth for output dependendencies. In the case of a cache-hit, the executor cache will hold the non-none value (it will respect updates if they happen somehow as well). In the case of a cache-miss, the executor caches a None and will wait for a notification to update the value when the node completes. When a node completes execution, it simply releases its mini-cache and in turn its strong refs on its direct anscestor outputs, allowing for ASAP freeing (same as the DependencyAwareCache but a little more automatic). This now allows for re-implementation of --cache-none with no cache at all. The dependency aware cache was also observing the dependency sematics for the objects and UI cache which is not accurate (this entire logic was always outputs specific). This also prepares for more complex caching strategies (such as RAM pressure based caching), where a cache can implement any freeing strategy completely independently of the DepedancyAwareness requirement. * main: re-implement --cache-none as no cache at all The execution list now tracks the dependency aware caching more correctly that the DependancyAwareCache. Change it to a cache that does nothing. * test_execution: add --cache-none to the test suite --cache-none is now expected to work universally. Run it through the full unit test suite. Propagate the server parameterization for whether or not the server is capabale of caching, so that the minority of tests that specifically check for cache hits can if else. Hard assert NOT caching in the else to give some coverage of --cache-none expected behaviour to not acutally cache. --- comfy_execution/caching.py | 174 ++++-------------------------- comfy_execution/graph.py | 32 +++++- execution.py | 34 +++--- main.py | 2 +- tests/execution/test_execution.py | 50 +++++---- 5 files changed, 102 insertions(+), 190 deletions(-) diff --git a/comfy_execution/caching.py b/comfy_execution/caching.py index 41224ce3b..566bc3f9c 100644 --- a/comfy_execution/caching.py +++ b/comfy_execution/caching.py @@ -265,6 +265,26 @@ class HierarchicalCache(BasicCache): assert cache is not None return await cache._ensure_subcache(node_id, children_ids) +class NullCache: + + async def set_prompt(self, dynprompt, node_ids, is_changed_cache): + pass + + def all_node_ids(self): + return [] + + def clean_unused(self): + pass + + def get(self, node_id): + return None + + def set(self, node_id, value): + pass + + async def ensure_subcache_for(self, node_id, children_ids): + return self + class LRUCache(BasicCache): def __init__(self, key_class, max_size=100): super().__init__(key_class) @@ -316,157 +336,3 @@ class LRUCache(BasicCache): self._mark_used(child_id) self.children[cache_key].append(self.cache_key_set.get_data_key(child_id)) return self - - -class DependencyAwareCache(BasicCache): - """ - A cache implementation that tracks dependencies between nodes and manages - their execution and caching accordingly. It extends the BasicCache class. - Nodes are removed from this cache once all of their descendants have been - executed. - """ - - def __init__(self, key_class): - """ - Initialize the DependencyAwareCache. - - Args: - key_class: The class used for generating cache keys. - """ - super().__init__(key_class) - self.descendants = {} # Maps node_id -> set of descendant node_ids - self.ancestors = {} # Maps node_id -> set of ancestor node_ids - self.executed_nodes = set() # Tracks nodes that have been executed - - async def set_prompt(self, dynprompt, node_ids, is_changed_cache): - """ - Clear the entire cache and rebuild the dependency graph. - - Args: - dynprompt: The dynamic prompt object containing node information. - node_ids: List of node IDs to initialize the cache for. - is_changed_cache: Flag indicating if the cache has changed. - """ - # Clear all existing cache data - self.cache.clear() - self.subcaches.clear() - self.descendants.clear() - self.ancestors.clear() - self.executed_nodes.clear() - - # Call the parent method to initialize the cache with the new prompt - await super().set_prompt(dynprompt, node_ids, is_changed_cache) - - # Rebuild the dependency graph - self._build_dependency_graph(dynprompt, node_ids) - - def _build_dependency_graph(self, dynprompt, node_ids): - """ - Build the dependency graph for all nodes. - - Args: - dynprompt: The dynamic prompt object containing node information. - node_ids: List of node IDs to build the graph for. - """ - self.descendants.clear() - self.ancestors.clear() - for node_id in node_ids: - self.descendants[node_id] = set() - self.ancestors[node_id] = set() - - for node_id in node_ids: - inputs = dynprompt.get_node(node_id)["inputs"] - for input_data in inputs.values(): - if is_link(input_data): # Check if the input is a link to another node - ancestor_id = input_data[0] - self.descendants[ancestor_id].add(node_id) - self.ancestors[node_id].add(ancestor_id) - - def set(self, node_id, value): - """ - Mark a node as executed and store its value in the cache. - - Args: - node_id: The ID of the node to store. - value: The value to store for the node. - """ - self._set_immediate(node_id, value) - self.executed_nodes.add(node_id) - self._cleanup_ancestors(node_id) - - def get(self, node_id): - """ - Retrieve the cached value for a node. - - Args: - node_id: The ID of the node to retrieve. - - Returns: - The cached value for the node. - """ - return self._get_immediate(node_id) - - async def ensure_subcache_for(self, node_id, children_ids): - """ - Ensure a subcache exists for a node and update dependencies. - - Args: - node_id: The ID of the parent node. - children_ids: List of child node IDs to associate with the parent node. - - Returns: - The subcache object for the node. - """ - subcache = await super()._ensure_subcache(node_id, children_ids) - for child_id in children_ids: - self.descendants[node_id].add(child_id) - self.ancestors[child_id].add(node_id) - return subcache - - def _cleanup_ancestors(self, node_id): - """ - Check if ancestors of a node can be removed from the cache. - - Args: - node_id: The ID of the node whose ancestors are to be checked. - """ - for ancestor_id in self.ancestors.get(node_id, []): - if ancestor_id in self.executed_nodes: - # Remove ancestor if all its descendants have been executed - if all(descendant in self.executed_nodes for descendant in self.descendants[ancestor_id]): - self._remove_node(ancestor_id) - - def _remove_node(self, node_id): - """ - Remove a node from the cache. - - Args: - node_id: The ID of the node to remove. - """ - cache_key = self.cache_key_set.get_data_key(node_id) - if cache_key in self.cache: - del self.cache[cache_key] - subcache_key = self.cache_key_set.get_subcache_key(node_id) - if subcache_key in self.subcaches: - del self.subcaches[subcache_key] - - def clean_unused(self): - """ - Clean up unused nodes. This is a no-op for this cache implementation. - """ - pass - - def recursive_debug_dump(self): - """ - Dump the cache and dependency graph for debugging. - - Returns: - A list containing the cache state and dependency graph. - """ - result = super().recursive_debug_dump() - result.append({ - "descendants": self.descendants, - "ancestors": self.ancestors, - "executed_nodes": list(self.executed_nodes), - }) - return result diff --git a/comfy_execution/graph.py b/comfy_execution/graph.py index f4b427265..341c9735d 100644 --- a/comfy_execution/graph.py +++ b/comfy_execution/graph.py @@ -153,8 +153,9 @@ class TopologicalSort: continue _, _, input_info = self.get_input_info(unique_id, input_name) is_lazy = input_info is not None and "lazy" in input_info and input_info["lazy"] - if (include_lazy or not is_lazy) and not self.is_cached(from_node_id): - node_ids.append(from_node_id) + if (include_lazy or not is_lazy): + if not self.is_cached(from_node_id): + node_ids.append(from_node_id) links.append((from_node_id, from_socket, unique_id)) for link in links: @@ -194,10 +195,35 @@ class ExecutionList(TopologicalSort): super().__init__(dynprompt) self.output_cache = output_cache self.staged_node_id = None + self.execution_cache = {} + self.execution_cache_listeners = {} def is_cached(self, node_id): return self.output_cache.get(node_id) is not None + def cache_link(self, from_node_id, to_node_id): + if not to_node_id in self.execution_cache: + self.execution_cache[to_node_id] = {} + self.execution_cache[to_node_id][from_node_id] = self.output_cache.get(from_node_id) + if not from_node_id in self.execution_cache_listeners: + self.execution_cache_listeners[from_node_id] = set() + self.execution_cache_listeners[from_node_id].add(to_node_id) + + def get_output_cache(self, from_node_id, to_node_id): + if not to_node_id in self.execution_cache: + return None + return self.execution_cache[to_node_id].get(from_node_id) + + def cache_update(self, node_id, value): + if node_id in self.execution_cache_listeners: + for to_node_id in self.execution_cache_listeners[node_id]: + if to_node_id in self.execution_cache: + self.execution_cache[to_node_id][node_id] = value + + def add_strong_link(self, from_node_id, from_socket, to_node_id): + super().add_strong_link(from_node_id, from_socket, to_node_id) + self.cache_link(from_node_id, to_node_id) + async def stage_node_execution(self): assert self.staged_node_id is None if self.is_empty(): @@ -277,6 +303,8 @@ class ExecutionList(TopologicalSort): def complete_node_execution(self): node_id = self.staged_node_id self.pop_node(node_id) + self.execution_cache.pop(node_id, None) + self.execution_cache_listeners.pop(node_id, None) self.staged_node_id = None def get_nodes_in_cycle(self): diff --git a/execution.py b/execution.py index 1dc35738b..78c36a4b0 100644 --- a/execution.py +++ b/execution.py @@ -18,7 +18,7 @@ from comfy_execution.caching import ( BasicCache, CacheKeySetID, CacheKeySetInputSignature, - DependencyAwareCache, + NullCache, HierarchicalCache, LRUCache, ) @@ -91,13 +91,13 @@ class IsChangedCache: class CacheType(Enum): CLASSIC = 0 LRU = 1 - DEPENDENCY_AWARE = 2 + NONE = 2 class CacheSet: def __init__(self, cache_type=None, cache_size=None): - if cache_type == CacheType.DEPENDENCY_AWARE: - self.init_dependency_aware_cache() + if cache_type == CacheType.NONE: + self.init_null_cache() logging.info("Disabling intermediate node cache.") elif cache_type == CacheType.LRU: if cache_size is None: @@ -120,11 +120,12 @@ class CacheSet: self.ui = LRUCache(CacheKeySetInputSignature, max_size=cache_size) self.objects = HierarchicalCache(CacheKeySetID) - # only hold cached items while the decendents have not executed - def init_dependency_aware_cache(self): - self.outputs = DependencyAwareCache(CacheKeySetInputSignature) - self.ui = DependencyAwareCache(CacheKeySetInputSignature) - self.objects = DependencyAwareCache(CacheKeySetID) + def init_null_cache(self): + self.outputs = NullCache() + #The UI cache is expected to be iterable at the end of each workflow + #so it must cache at least a full workflow. Use Heirachical + self.ui = HierarchicalCache(CacheKeySetInputSignature) + self.objects = NullCache() def recursive_debug_dump(self): result = { @@ -135,7 +136,7 @@ class CacheSet: SENSITIVE_EXTRA_DATA_KEYS = ("auth_token_comfy_org", "api_key_comfy_org") -def get_input_data(inputs, class_def, unique_id, outputs=None, dynprompt=None, extra_data={}): +def get_input_data(inputs, class_def, unique_id, execution_list=None, dynprompt=None, extra_data={}): is_v3 = issubclass(class_def, _ComfyNodeInternal) if is_v3: valid_inputs, schema = class_def.INPUT_TYPES(include_hidden=False, return_schema=True) @@ -153,10 +154,10 @@ def get_input_data(inputs, class_def, unique_id, outputs=None, dynprompt=None, e if is_link(input_data) and (not input_info or not input_info.get("rawLink", False)): input_unique_id = input_data[0] output_index = input_data[1] - if outputs is None: + if execution_list is None: mark_missing() continue # This might be a lazily-evaluated input - cached_output = outputs.get(input_unique_id) + cached_output = execution_list.get_output_cache(input_unique_id, unique_id) if cached_output is None: mark_missing() continue @@ -405,6 +406,7 @@ async def execute(server, dynprompt, caches, current_item, extra_data, executed, cached_output = caches.ui.get(unique_id) or {} server.send_sync("executed", { "node": unique_id, "display_node": display_node_id, "output": cached_output.get("output",None), "prompt_id": prompt_id }, server.client_id) get_progress_state().finish_progress(unique_id) + execution_list.cache_update(unique_id, caches.outputs.get(unique_id)) return (ExecutionResult.SUCCESS, None, None) input_data_all = None @@ -434,7 +436,7 @@ async def execute(server, dynprompt, caches, current_item, extra_data, executed, for r in result: if is_link(r): source_node, source_output = r[0], r[1] - node_output = caches.outputs.get(source_node)[source_output] + node_output = execution_list.get_output_cache(source_node, unique_id)[source_output] for o in node_output: resolved_output.append(o) @@ -446,7 +448,7 @@ async def execute(server, dynprompt, caches, current_item, extra_data, executed, has_subgraph = False else: get_progress_state().start_progress(unique_id) - input_data_all, missing_keys, hidden_inputs = get_input_data(inputs, class_def, unique_id, caches.outputs, dynprompt, extra_data) + input_data_all, missing_keys, hidden_inputs = get_input_data(inputs, class_def, unique_id, execution_list, dynprompt, extra_data) if server.client_id is not None: server.last_node_id = display_node_id server.send_sync("executing", { "node": unique_id, "display_node": display_node_id, "prompt_id": prompt_id }, server.client_id) @@ -549,11 +551,15 @@ async def execute(server, dynprompt, caches, current_item, extra_data, executed, subcache.clean_unused() for node_id in new_output_ids: execution_list.add_node(node_id) + execution_list.cache_link(node_id, unique_id) for link in new_output_links: execution_list.add_strong_link(link[0], link[1], unique_id) pending_subgraph_results[unique_id] = cached_outputs return (ExecutionResult.PENDING, None, None) + caches.outputs.set(unique_id, output_data) + execution_list.cache_update(unique_id, output_data) + except comfy.model_management.InterruptProcessingException as iex: logging.info("Processing interrupted") diff --git a/main.py b/main.py index 35857dba8..4b4c5dcc4 100644 --- a/main.py +++ b/main.py @@ -173,7 +173,7 @@ def prompt_worker(q, server_instance): if args.cache_lru > 0: cache_type = execution.CacheType.LRU elif args.cache_none: - cache_type = execution.CacheType.DEPENDENCY_AWARE + cache_type = execution.CacheType.NONE e = execution.PromptExecutor(server_instance, cache_type=cache_type, cache_size=args.cache_lru) last_gc_collect = 0 diff --git a/tests/execution/test_execution.py b/tests/execution/test_execution.py index ef73ad9fd..ace0d2279 100644 --- a/tests/execution/test_execution.py +++ b/tests/execution/test_execution.py @@ -152,12 +152,12 @@ class TestExecution: # Initialize server and client # @fixture(scope="class", autouse=True, params=[ - # (use_lru, lru_size) - (False, 0), - (True, 0), - (True, 100), + { "extra_args" : [], "should_cache_results" : True }, + { "extra_args" : ["--cache-lru", 0], "should_cache_results" : True }, + { "extra_args" : ["--cache-lru", 100], "should_cache_results" : True }, + { "extra_args" : ["--cache-none"], "should_cache_results" : False }, ]) - def _server(self, args_pytest, request): + def server(self, args_pytest, request): # Start server pargs = [ 'python','main.py', @@ -167,12 +167,10 @@ class TestExecution: '--extra-model-paths-config', 'tests/execution/extra_model_paths.yaml', '--cpu', ] - use_lru, lru_size = request.param - if use_lru: - pargs += ['--cache-lru', str(lru_size)] + pargs += [ str(param) for param in request.param["extra_args"] ] print("Running server with args:", pargs) # noqa: T201 p = subprocess.Popen(pargs) - yield + yield request.param p.kill() torch.cuda.empty_cache() @@ -193,7 +191,7 @@ class TestExecution: return comfy_client @fixture(scope="class", autouse=True) - def shared_client(self, args_pytest, _server): + def shared_client(self, args_pytest, server): client = self.start_client(args_pytest["listen"], args_pytest["port"]) yield client del client @@ -225,7 +223,7 @@ class TestExecution: assert result.did_run(mask) assert result.did_run(lazy_mix) - def test_full_cache(self, client: ComfyClient, builder: GraphBuilder): + def test_full_cache(self, client: ComfyClient, builder: GraphBuilder, server): g = builder input1 = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1) input2 = g.node("StubImage", content="NOISE", height=512, width=512, batch_size=1) @@ -237,9 +235,12 @@ class TestExecution: client.run(g) result2 = client.run(g) for node_id, node in g.nodes.items(): - assert not result2.did_run(node), f"Node {node_id} ran, but should have been cached" + if server["should_cache_results"]: + assert not result2.did_run(node), f"Node {node_id} ran, but should have been cached" + else: + assert result2.did_run(node), f"Node {node_id} was cached, but should have been run" - def test_partial_cache(self, client: ComfyClient, builder: GraphBuilder): + def test_partial_cache(self, client: ComfyClient, builder: GraphBuilder, server): g = builder input1 = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1) input2 = g.node("StubImage", content="NOISE", height=512, width=512, batch_size=1) @@ -251,8 +252,12 @@ class TestExecution: client.run(g) mask.inputs['value'] = 0.4 result2 = client.run(g) - assert not result2.did_run(input1), "Input1 should have been cached" - assert not result2.did_run(input2), "Input2 should have been cached" + if server["should_cache_results"]: + assert not result2.did_run(input1), "Input1 should have been cached" + assert not result2.did_run(input2), "Input2 should have been cached" + else: + assert result2.did_run(input1), "Input1 should have been rerun" + assert result2.did_run(input2), "Input2 should have been rerun" def test_error(self, client: ComfyClient, builder: GraphBuilder): g = builder @@ -411,7 +416,7 @@ class TestExecution: input2 = g.node("StubImage", id="removeme", content="WHITE", height=512, width=512, batch_size=1) client.run(g) - def test_custom_is_changed(self, client: ComfyClient, builder: GraphBuilder): + def test_custom_is_changed(self, client: ComfyClient, builder: GraphBuilder, server): g = builder # Creating the nodes in this specific order previously caused a bug save = g.node("SaveImage") @@ -427,7 +432,10 @@ class TestExecution: result3 = client.run(g) result4 = client.run(g) assert result1.did_run(is_changed), "is_changed should have been run" - assert not result2.did_run(is_changed), "is_changed should have been cached" + if server["should_cache_results"]: + assert not result2.did_run(is_changed), "is_changed should have been cached" + else: + assert result2.did_run(is_changed), "is_changed should have been re-run" assert result3.did_run(is_changed), "is_changed should have been re-run" assert result4.did_run(is_changed), "is_changed should not have been cached" @@ -514,7 +522,7 @@ class TestExecution: assert len(images2) == 1, "Should have 1 image" # This tests that only constant outputs are used in the call to `IS_CHANGED` - def test_is_changed_with_outputs(self, client: ComfyClient, builder: GraphBuilder): + def test_is_changed_with_outputs(self, client: ComfyClient, builder: GraphBuilder, server): g = builder input1 = g.node("StubConstantImage", value=0.5, height=512, width=512, batch_size=1) test_node = g.node("TestIsChangedWithConstants", image=input1.out(0), value=0.5) @@ -530,7 +538,11 @@ class TestExecution: images = result.get_images(output) assert len(images) == 1, "Should have 1 image" assert numpy.array(images[0]).min() == 63 and numpy.array(images[0]).max() == 63, "Image should have value 0.25" - assert not result.did_run(test_node), "The execution should have been cached" + if server["should_cache_results"]: + assert not result.did_run(test_node), "The execution should have been cached" + else: + assert result.did_run(test_node), "The execution should have been re-run" + def test_parallel_sleep_nodes(self, client: ComfyClient, builder: GraphBuilder, skip_timing_checks): # Warmup execution to ensure server is fully initialized From a1864c01f29cc43fe6bf823fc3fd46ba2781c2e0 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 22 Oct 2025 14:26:22 -0700 Subject: [PATCH 023/647] Small readme improvement. (#10442) --- README.md | 2 ++ 1 file changed, 2 insertions(+) diff --git a/README.md b/README.md index c9a0644e3..434d4ff06 100644 --- a/README.md +++ b/README.md @@ -201,6 +201,8 @@ Python 3.14 will work if you comment out the `kornia` dependency in the requirem Python 3.13 is very well supported. If you have trouble with some custom node dependencies on 3.13 you can try 3.12 +### Instructions: + Git clone this repo. Put your SD checkpoints (the huge ckpt/safetensors files) in: models/checkpoints From 1bcda6df987a6c92b39d8b6d29e0b029450d67d0 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Thu, 23 Oct 2025 18:21:14 -0700 Subject: [PATCH 024/647] WIP way to support multi multi dimensional latents. (#10456) --- comfy/model_base.py | 10 ++++- comfy/nested_tensor.py | 91 ++++++++++++++++++++++++++++++++++++++++++ comfy/sample.py | 27 ++++++++++--- comfy/samplers.py | 23 +++++++---- comfy/utils.py | 22 ++++++++++ 5 files changed, 158 insertions(+), 15 deletions(-) create mode 100644 comfy/nested_tensor.py diff --git a/comfy/model_base.py b/comfy/model_base.py index 8274c7dea..e877f19ac 100644 --- a/comfy/model_base.py +++ b/comfy/model_base.py @@ -197,8 +197,14 @@ class BaseModel(torch.nn.Module): extra_conds[o] = extra t = self.process_timestep(t, x=x, **extra_conds) - model_output = self.diffusion_model(xc, t, context=context, control=control, transformer_options=transformer_options, **extra_conds).float() - return self.model_sampling.calculate_denoised(sigma, model_output, x) + if "latent_shapes" in extra_conds: + xc = utils.unpack_latents(xc, extra_conds.pop("latent_shapes")) + + model_output = self.diffusion_model(xc, t, context=context, control=control, transformer_options=transformer_options, **extra_conds) + if len(model_output) > 1 and not torch.is_tensor(model_output): + model_output, _ = utils.pack_latents(model_output) + + return self.model_sampling.calculate_denoised(sigma, model_output.float(), x) def process_timestep(self, timestep, **kwargs): return timestep diff --git a/comfy/nested_tensor.py b/comfy/nested_tensor.py new file mode 100644 index 000000000..b700816fa --- /dev/null +++ b/comfy/nested_tensor.py @@ -0,0 +1,91 @@ +import torch + +class NestedTensor: + def __init__(self, tensors): + self.tensors = list(tensors) + self.is_nested = True + + def _copy(self): + return NestedTensor(self.tensors) + + def apply_operation(self, other, operation): + o = self._copy() + if isinstance(other, NestedTensor): + for i, t in enumerate(o.tensors): + o.tensors[i] = operation(t, other.tensors[i]) + else: + for i, t in enumerate(o.tensors): + o.tensors[i] = operation(t, other) + return o + + def __add__(self, b): + return self.apply_operation(b, lambda x, y: x + y) + + def __sub__(self, b): + return self.apply_operation(b, lambda x, y: x - y) + + def __mul__(self, b): + return self.apply_operation(b, lambda x, y: x * y) + + # def __itruediv__(self, b): + # return self.apply_operation(b, lambda x, y: x / y) + + def __truediv__(self, b): + return self.apply_operation(b, lambda x, y: x / y) + + def __getitem__(self, *args, **kwargs): + return self.apply_operation(None, lambda x, y: x.__getitem__(*args, **kwargs)) + + def unbind(self): + return self.tensors + + def to(self, *args, **kwargs): + o = self._copy() + for i, t in enumerate(o.tensors): + o.tensors[i] = t.to(*args, **kwargs) + return o + + def new_ones(self, *args, **kwargs): + return self.tensors[0].new_ones(*args, **kwargs) + + def float(self): + return self.to(dtype=torch.float) + + def chunk(self, *args, **kwargs): + return self.apply_operation(None, lambda x, y: x.chunk(*args, **kwargs)) + + def size(self): + return self.tensors[0].size() + + @property + def shape(self): + return self.tensors[0].shape + + @property + def ndim(self): + dims = 0 + for t in self.tensors: + dims = max(t.ndim, dims) + return dims + + @property + def device(self): + return self.tensors[0].device + + @property + def dtype(self): + return self.tensors[0].dtype + + @property + def layout(self): + return self.tensors[0].layout + + +def cat_nested(tensors, *args, **kwargs): + cated_tensors = [] + for i in range(len(tensors[0].tensors)): + tens = [] + for j in range(len(tensors)): + tens.append(tensors[j].tensors[i]) + cated_tensors.append(torch.cat(tens, *args, **kwargs)) + return NestedTensor(cated_tensors) diff --git a/comfy/sample.py b/comfy/sample.py index be5a7e246..b1395da84 100644 --- a/comfy/sample.py +++ b/comfy/sample.py @@ -4,13 +4,9 @@ import comfy.samplers import comfy.utils import numpy as np import logging +import comfy.nested_tensor -def prepare_noise(latent_image, seed, noise_inds=None): - """ - creates random noise given a latent image and a seed. - optional arg skip can be used to skip and discard x number of noise generations for a given seed - """ - generator = torch.manual_seed(seed) +def prepare_noise_inner(latent_image, generator, noise_inds=None): if noise_inds is None: return torch.randn(latent_image.size(), dtype=latent_image.dtype, layout=latent_image.layout, generator=generator, device="cpu") @@ -22,9 +18,28 @@ def prepare_noise(latent_image, seed, noise_inds=None): noises.append(noise) noises = [noises[i] for i in inverse] noises = torch.cat(noises, axis=0) + +def prepare_noise(latent_image, seed, noise_inds=None): + """ + creates random noise given a latent image and a seed. + optional arg skip can be used to skip and discard x number of noise generations for a given seed + """ + generator = torch.manual_seed(seed) + + if latent_image.is_nested: + tensors = latent_image.unbind() + noises = [] + for t in tensors: + noises.append(prepare_noise_inner(t, generator, noise_inds)) + noises = comfy.nested_tensor.NestedTensor(noises) + else: + noises = prepare_noise_inner(latent_image, generator, noise_inds) + return noises def fix_empty_latent_channels(model, latent_image): + if latent_image.is_nested: + return latent_image latent_format = model.get_model_object("latent_format") #Resize the empty latent image so it has the right number of channels if latent_format.latent_channels != latent_image.shape[1] and torch.count_nonzero(latent_image) == 0: latent_image = comfy.utils.repeat_to_batch_size(latent_image, latent_format.latent_channels, dim=1) diff --git a/comfy/samplers.py b/comfy/samplers.py index e7efaf470..fa4640842 100755 --- a/comfy/samplers.py +++ b/comfy/samplers.py @@ -782,7 +782,7 @@ def ksampler(sampler_name, extra_options={}, inpaint_options={}): return KSAMPLER(sampler_function, extra_options, inpaint_options) -def process_conds(model, noise, conds, device, latent_image=None, denoise_mask=None, seed=None): +def process_conds(model, noise, conds, device, latent_image=None, denoise_mask=None, seed=None, latent_shapes=None): for k in conds: conds[k] = conds[k][:] resolve_areas_and_cond_masks_multidim(conds[k], noise.shape[2:], device) @@ -792,7 +792,7 @@ def process_conds(model, noise, conds, device, latent_image=None, denoise_mask=N if hasattr(model, 'extra_conds'): for k in conds: - conds[k] = encode_model_conds(model.extra_conds, conds[k], noise, device, k, latent_image=latent_image, denoise_mask=denoise_mask, seed=seed) + conds[k] = encode_model_conds(model.extra_conds, conds[k], noise, device, k, latent_image=latent_image, denoise_mask=denoise_mask, seed=seed, latent_shapes=latent_shapes) #make sure each cond area has an opposite one with the same area for k in conds: @@ -962,11 +962,11 @@ class CFGGuider: def predict_noise(self, x, timestep, model_options={}, seed=None): return sampling_function(self.inner_model, x, timestep, self.conds.get("negative", None), self.conds.get("positive", None), self.cfg, model_options=model_options, seed=seed) - def inner_sample(self, noise, latent_image, device, sampler, sigmas, denoise_mask, callback, disable_pbar, seed): + def inner_sample(self, noise, latent_image, device, sampler, sigmas, denoise_mask, callback, disable_pbar, seed, latent_shapes=None): if latent_image is not None and torch.count_nonzero(latent_image) > 0: #Don't shift the empty latent image. latent_image = self.inner_model.process_latent_in(latent_image) - self.conds = process_conds(self.inner_model, noise, self.conds, device, latent_image, denoise_mask, seed) + self.conds = process_conds(self.inner_model, noise, self.conds, device, latent_image, denoise_mask, seed, latent_shapes=latent_shapes) extra_model_options = comfy.model_patcher.create_model_options_clone(self.model_options) extra_model_options.setdefault("transformer_options", {})["sample_sigmas"] = sigmas @@ -980,7 +980,7 @@ class CFGGuider: samples = executor.execute(self, sigmas, extra_args, callback, noise, latent_image, denoise_mask, disable_pbar) return self.inner_model.process_latent_out(samples.to(torch.float32)) - def outer_sample(self, noise, latent_image, sampler, sigmas, denoise_mask=None, callback=None, disable_pbar=False, seed=None): + def outer_sample(self, noise, latent_image, sampler, sigmas, denoise_mask=None, callback=None, disable_pbar=False, seed=None, latent_shapes=None): self.inner_model, self.conds, self.loaded_models = comfy.sampler_helpers.prepare_sampling(self.model_patcher, noise.shape, self.conds, self.model_options) device = self.model_patcher.load_device @@ -994,7 +994,7 @@ class CFGGuider: try: self.model_patcher.pre_run() - output = self.inner_sample(noise, latent_image, device, sampler, sigmas, denoise_mask, callback, disable_pbar, seed) + output = self.inner_sample(noise, latent_image, device, sampler, sigmas, denoise_mask, callback, disable_pbar, seed, latent_shapes=latent_shapes) finally: self.model_patcher.cleanup() @@ -1007,6 +1007,12 @@ class CFGGuider: if sigmas.shape[-1] == 0: return latent_image + if latent_image.is_nested: + latent_image, latent_shapes = comfy.utils.pack_latents(latent_image.unbind()) + noise, _ = comfy.utils.pack_latents(noise.unbind()) + else: + latent_shapes = [latent_image.shape] + self.conds = {} for k in self.original_conds: self.conds[k] = list(map(lambda a: a.copy(), self.original_conds[k])) @@ -1026,7 +1032,7 @@ class CFGGuider: self, comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.OUTER_SAMPLE, self.model_options, is_model_options=True) ) - output = executor.execute(noise, latent_image, sampler, sigmas, denoise_mask, callback, disable_pbar, seed) + output = executor.execute(noise, latent_image, sampler, sigmas, denoise_mask, callback, disable_pbar, seed, latent_shapes=latent_shapes) finally: cast_to_load_options(self.model_options, device=self.model_patcher.offload_device) self.model_options = orig_model_options @@ -1034,6 +1040,9 @@ class CFGGuider: self.model_patcher.restore_hook_patches() del self.conds + + if len(latent_shapes) > 1: + output = comfy.nested_tensor.NestedTensor(comfy.utils.unpack_latents(output, latent_shapes)) return output diff --git a/comfy/utils.py b/comfy/utils.py index 0fd03f165..4bd281057 100644 --- a/comfy/utils.py +++ b/comfy/utils.py @@ -1106,3 +1106,25 @@ def upscale_dit_mask(mask: torch.Tensor, img_size_in, img_size_out): dim=1 ) return out + +def pack_latents(latents): + latent_shapes = [] + tensors = [] + for tensor in latents: + latent_shapes.append(tensor.shape) + tensors.append(tensor.reshape(tensor.shape[0], 1, -1)) + + latent = torch.cat(tensors, dim=-1) + return latent, latent_shapes + +def unpack_latents(combined_latent, latent_shapes): + if len(latent_shapes) > 1: + output_tensors = [] + for shape in latent_shapes: + cut = math.prod(shape[1:]) + tens = combined_latent[:, :, :cut] + combined_latent = combined_latent[:, :, cut:] + output_tensors.append(tens.reshape([tens.shape[0]] + list(shape)[1:])) + else: + output_tensors = combined_latent + return output_tensors From 24188b3141aace272cb91b85578c76f5a8f70e1c Mon Sep 17 00:00:00 2001 From: ComfyUI Wiki Date: Fri, 24 Oct 2025 13:36:30 +0800 Subject: [PATCH 025/647] Update template to 0.2.2 (#10461) Fix template typo issue --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index dd2afcab0..8570c66b6 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,5 +1,5 @@ comfyui-frontend-package==1.28.7 -comfyui-workflow-templates==0.2.1 +comfyui-workflow-templates==0.2.2 comfyui-embedded-docs==0.3.0 torch torchsde From 388b306a2b48070737b092b51e76de933baee9ad Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Fri, 24 Oct 2025 08:37:16 +0300 Subject: [PATCH 026/647] feat(api-nodes): network client v2: async ops, cancellation, downloads, refactor (#10390) * feat(api-nodes): implement new API client for V3 nodes * feat(api-nodes): implement new API client for V3 nodes * feat(api-nodes): implement new API client for V3 nodes * converted WAN nodes to use new client; polishing * fix(auth): do not leak authentification for the absolute urls * convert BFL API nodes to use new API client; remove deprecated BFL nodes * converted Google Veo nodes * fix(Veo3.1 model): take into account "generate_audio" parameter --- comfy_api_nodes/apinode_utils.py | 435 +--------- comfy_api_nodes/apis/bfl_api.py | 51 +- comfy_api_nodes/apis/veo_api.py | 111 +++ comfy_api_nodes/nodes_bfl.py | 605 ++++---------- comfy_api_nodes/nodes_bytedance.py | 279 ++----- comfy_api_nodes/nodes_gemini.py | 5 +- comfy_api_nodes/nodes_kling.py | 350 +++----- comfy_api_nodes/nodes_luma.py | 2 +- comfy_api_nodes/nodes_minimax.py | 2 +- comfy_api_nodes/nodes_moonvalley.py | 366 ++------- comfy_api_nodes/nodes_openai.py | 4 +- comfy_api_nodes/nodes_pika.py | 6 +- comfy_api_nodes/nodes_pixverse.py | 13 +- comfy_api_nodes/nodes_recraft.py | 4 +- comfy_api_nodes/nodes_runway.py | 199 ++--- comfy_api_nodes/nodes_sora.py | 74 +- comfy_api_nodes/nodes_stability.py | 8 +- comfy_api_nodes/nodes_veo2.py | 176 ++-- comfy_api_nodes/nodes_vidu.py | 131 +-- comfy_api_nodes/nodes_wan.py | 245 +++--- comfy_api_nodes/util/__init__.py | 87 ++ comfy_api_nodes/util/_helpers.py | 71 ++ comfy_api_nodes/util/client.py | 941 ++++++++++++++++++++++ comfy_api_nodes/util/common_exceptions.py | 14 + comfy_api_nodes/util/conversions.py | 407 ++++++++++ comfy_api_nodes/util/download_helpers.py | 249 ++++++ comfy_api_nodes/util/upload_helpers.py | 338 ++++++++ comfy_api_nodes/util/validation_utils.py | 58 +- pyproject.toml | 2 + 29 files changed, 2935 insertions(+), 2298 deletions(-) create mode 100644 comfy_api_nodes/apis/veo_api.py create mode 100644 comfy_api_nodes/util/_helpers.py create mode 100644 comfy_api_nodes/util/client.py create mode 100644 comfy_api_nodes/util/common_exceptions.py create mode 100644 comfy_api_nodes/util/conversions.py create mode 100644 comfy_api_nodes/util/download_helpers.py create mode 100644 comfy_api_nodes/util/upload_helpers.py diff --git a/comfy_api_nodes/apinode_utils.py b/comfy_api_nodes/apinode_utils.py index bc3d2d07e..e3d282059 100644 --- a/comfy_api_nodes/apinode_utils.py +++ b/comfy_api_nodes/apinode_utils.py @@ -1,15 +1,10 @@ from __future__ import annotations import aiohttp -import io -import logging import mimetypes -import os from typing import Optional, Union from comfy.utils import common_upscale -from comfy_api.input_impl import VideoFromFile from comfy_api.util import VideoContainer, VideoCodec from comfy_api.input.video_types import VideoInput -from comfy_api.input.basic_types import AudioInput from comfy_api_nodes.apis.client import ( ApiClient, ApiEndpoint, @@ -26,43 +21,8 @@ from PIL import Image import torch import math import base64 -import uuid +from .util import tensor_to_bytesio, bytesio_to_image_tensor from io import BytesIO -import av - - -async def download_url_to_video_output( - video_url: str, timeout: int = None, auth_kwargs: Optional[dict[str, str]] = None -) -> VideoFromFile: - """Downloads a video from a URL and returns a `VIDEO` output. - - Args: - video_url: The URL of the video to download. - - Returns: - A Comfy node `VIDEO` output. - """ - video_io = await download_url_to_bytesio(video_url, timeout, auth_kwargs=auth_kwargs) - if video_io is None: - error_msg = f"Failed to download video from {video_url}" - logging.error(error_msg) - raise ValueError(error_msg) - return VideoFromFile(video_io) - - -def downscale_image_tensor(image, total_pixels=1536 * 1024) -> torch.Tensor: - """Downscale input image tensor to roughly the specified total pixels.""" - samples = image.movedim(-1, 1) - total = int(total_pixels) - scale_by = math.sqrt(total / (samples.shape[3] * samples.shape[2])) - if scale_by >= 1: - return image - width = round(samples.shape[3] * scale_by) - height = round(samples.shape[2] * scale_by) - - s = common_upscale(samples, width, height, "lanczos", "disabled") - s = s.movedim(1, -1) - return s async def validate_and_cast_response( @@ -162,11 +122,6 @@ def validate_aspect_ratio( return aspect_ratio -def mimetype_to_extension(mime_type: str) -> str: - """Converts a MIME type to a file extension.""" - return mime_type.split("/")[-1].lower() - - async def download_url_to_bytesio( url: str, timeout: int = None, auth_kwargs: Optional[dict[str, str]] = None ) -> BytesIO: @@ -195,136 +150,11 @@ async def download_url_to_bytesio( return BytesIO(await resp.read()) -def bytesio_to_image_tensor(image_bytesio: BytesIO, mode: str = "RGBA") -> torch.Tensor: - """Converts image data from BytesIO to a torch.Tensor. - - Args: - image_bytesio: BytesIO object containing the image data. - mode: The PIL mode to convert the image to (e.g., "RGB", "RGBA"). - - Returns: - A torch.Tensor representing the image (1, H, W, C). - - Raises: - PIL.UnidentifiedImageError: If the image data cannot be identified. - ValueError: If the specified mode is invalid. - """ - image = Image.open(image_bytesio) - image = image.convert(mode) - image_array = np.array(image).astype(np.float32) / 255.0 - return torch.from_numpy(image_array).unsqueeze(0) - - -async def download_url_to_image_tensor(url: str, timeout: int = None) -> torch.Tensor: - """Downloads an image from a URL and returns a [B, H, W, C] tensor.""" - image_bytesio = await download_url_to_bytesio(url, timeout) - return bytesio_to_image_tensor(image_bytesio) - - def process_image_response(response_content: bytes | str) -> torch.Tensor: """Uses content from a Response object and converts it to a torch.Tensor""" return bytesio_to_image_tensor(BytesIO(response_content)) -def _tensor_to_pil(image: torch.Tensor, total_pixels: int = 2048 * 2048) -> Image.Image: - """Converts a single torch.Tensor image [H, W, C] to a PIL Image, optionally downscaling.""" - if len(image.shape) > 3: - image = image[0] - # TODO: remove alpha if not allowed and present - input_tensor = image.cpu() - input_tensor = downscale_image_tensor( - input_tensor.unsqueeze(0), total_pixels=total_pixels - ).squeeze() - image_np = (input_tensor.numpy() * 255).astype(np.uint8) - img = Image.fromarray(image_np) - return img - - -def _pil_to_bytesio(img: Image.Image, mime_type: str = "image/png") -> BytesIO: - """Converts a PIL Image to a BytesIO object.""" - if not mime_type: - mime_type = "image/png" - - img_byte_arr = io.BytesIO() - # Derive PIL format from MIME type (e.g., 'image/png' -> 'PNG') - pil_format = mime_type.split("/")[-1].upper() - if pil_format == "JPG": - pil_format = "JPEG" - img.save(img_byte_arr, format=pil_format) - img_byte_arr.seek(0) - return img_byte_arr - - -def tensor_to_bytesio( - image: torch.Tensor, - name: Optional[str] = None, - total_pixels: int = 2048 * 2048, - mime_type: str = "image/png", -) -> BytesIO: - """Converts a torch.Tensor image to a named BytesIO object. - - Args: - image: Input torch.Tensor image. - name: Optional filename for the BytesIO object. - total_pixels: Maximum total pixels for potential downscaling. - mime_type: Target image MIME type (e.g., 'image/png', 'image/jpeg', 'image/webp', 'video/mp4'). - - Returns: - Named BytesIO object containing the image data, with pointer set to the start of buffer. - """ - if not mime_type: - mime_type = "image/png" - - pil_image = _tensor_to_pil(image, total_pixels=total_pixels) - img_binary = _pil_to_bytesio(pil_image, mime_type=mime_type) - img_binary.name = ( - f"{name if name else uuid.uuid4()}.{mimetype_to_extension(mime_type)}" - ) - return img_binary - - -def tensor_to_base64_string( - image_tensor: torch.Tensor, - total_pixels: int = 2048 * 2048, - mime_type: str = "image/png", -) -> str: - """Convert [B, H, W, C] or [H, W, C] tensor to a base64 string. - - Args: - image_tensor: Input torch.Tensor image. - total_pixels: Maximum total pixels for potential downscaling. - mime_type: Target image MIME type (e.g., 'image/png', 'image/jpeg', 'image/webp', 'video/mp4'). - - Returns: - Base64 encoded string of the image. - """ - pil_image = _tensor_to_pil(image_tensor, total_pixels=total_pixels) - img_byte_arr = _pil_to_bytesio(pil_image, mime_type=mime_type) - img_bytes = img_byte_arr.getvalue() - # Encode bytes to base64 string - base64_encoded_string = base64.b64encode(img_bytes).decode("utf-8") - return base64_encoded_string - - -def tensor_to_data_uri( - image_tensor: torch.Tensor, - total_pixels: int = 2048 * 2048, - mime_type: str = "image/png", -) -> str: - """Converts a tensor image to a Data URI string. - - Args: - image_tensor: Input torch.Tensor image. - total_pixels: Maximum total pixels for potential downscaling. - mime_type: Target image MIME type (e.g., 'image/png', 'image/jpeg', 'image/webp'). - - Returns: - Data URI string (e.g., 'data:image/png;base64,...'). - """ - base64_string = tensor_to_base64_string(image_tensor, total_pixels, mime_type) - return f"data:{mime_type};base64,{base64_string}" - - def text_filepath_to_base64_string(filepath: str) -> str: """Converts a text file to a base64 string.""" with open(filepath, "rb") as f: @@ -392,7 +222,7 @@ def video_to_base64_string( container_format: Optional container format to use (defaults to video.container if available) codec: Optional codec to use (defaults to video.codec if available) """ - video_bytes_io = io.BytesIO() + video_bytes_io = BytesIO() # Use provided format/codec if specified, otherwise use video's own if available format_to_use = container_format if container_format is not None else getattr(video, 'container', VideoContainer.MP4) @@ -403,214 +233,6 @@ def video_to_base64_string( return base64.b64encode(video_bytes_io.getvalue()).decode("utf-8") -async def upload_video_to_comfyapi( - video: VideoInput, - auth_kwargs: Optional[dict[str, str]] = None, - container: VideoContainer = VideoContainer.MP4, - codec: VideoCodec = VideoCodec.H264, - max_duration: Optional[int] = None, -) -> str: - """ - Uploads a single video to ComfyUI API and returns its download URL. - Uses the specified container and codec for saving the video before upload. - - Args: - video: VideoInput object (Comfy VIDEO type). - auth_kwargs: Optional authentication token(s). - container: The video container format to use (default: MP4). - codec: The video codec to use (default: H264). - max_duration: Optional maximum duration of the video in seconds. If the video is longer than this, an error will be raised. - - Returns: - The download URL for the uploaded video file. - """ - if max_duration is not None: - try: - actual_duration = video.duration_seconds - if actual_duration is not None and actual_duration > max_duration: - raise ValueError( - f"Video duration ({actual_duration:.2f}s) exceeds the maximum allowed ({max_duration}s)." - ) - except Exception as e: - logging.error("Error getting video duration: %s", str(e)) - raise ValueError(f"Could not verify video duration from source: {e}") from e - - upload_mime_type = f"video/{container.value.lower()}" - filename = f"uploaded_video.{container.value.lower()}" - - # Convert VideoInput to BytesIO using specified container/codec - video_bytes_io = io.BytesIO() - video.save_to(video_bytes_io, format=container, codec=codec) - video_bytes_io.seek(0) - - return await upload_file_to_comfyapi(video_bytes_io, filename, upload_mime_type, auth_kwargs) - - -def audio_tensor_to_contiguous_ndarray(waveform: torch.Tensor) -> np.ndarray: - """ - Prepares audio waveform for av library by converting to a contiguous numpy array. - - Args: - waveform: a tensor of shape (1, channels, samples) derived from a Comfy `AUDIO` type. - - Returns: - Contiguous numpy array of the audio waveform. If the audio was batched, - the first item is taken. - """ - if waveform.ndim != 3 or waveform.shape[0] != 1: - raise ValueError("Expected waveform tensor shape (1, channels, samples)") - - # If batch is > 1, take first item - if waveform.shape[0] > 1: - waveform = waveform[0] - - # Prepare for av: remove batch dim, move to CPU, make contiguous, convert to numpy array - audio_data_np = waveform.squeeze(0).cpu().contiguous().numpy() - if audio_data_np.dtype != np.float32: - audio_data_np = audio_data_np.astype(np.float32) - - return audio_data_np - - -def audio_ndarray_to_bytesio( - audio_data_np: np.ndarray, - sample_rate: int, - container_format: str = "mp4", - codec_name: str = "aac", -) -> BytesIO: - """ - Encodes a numpy array of audio data into a BytesIO object. - """ - audio_bytes_io = io.BytesIO() - with av.open(audio_bytes_io, mode="w", format=container_format) as output_container: - audio_stream = output_container.add_stream(codec_name, rate=sample_rate) - frame = av.AudioFrame.from_ndarray( - audio_data_np, - format="fltp", - layout="stereo" if audio_data_np.shape[0] > 1 else "mono", - ) - frame.sample_rate = sample_rate - frame.pts = 0 - - for packet in audio_stream.encode(frame): - output_container.mux(packet) - - # Flush stream - for packet in audio_stream.encode(None): - output_container.mux(packet) - - audio_bytes_io.seek(0) - return audio_bytes_io - - -async def upload_audio_to_comfyapi( - audio: AudioInput, - auth_kwargs: Optional[dict[str, str]] = None, - container_format: str = "mp4", - codec_name: str = "aac", - mime_type: str = "audio/mp4", - filename: str = "uploaded_audio.mp4", -) -> str: - """ - Uploads a single audio input to ComfyUI API and returns its download URL. - Encodes the raw waveform into the specified format before uploading. - - Args: - audio: a Comfy `AUDIO` type (contains waveform tensor and sample_rate) - auth_kwargs: Optional authentication token(s). - - Returns: - The download URL for the uploaded audio file. - """ - sample_rate: int = audio["sample_rate"] - waveform: torch.Tensor = audio["waveform"] - audio_data_np = audio_tensor_to_contiguous_ndarray(waveform) - audio_bytes_io = audio_ndarray_to_bytesio( - audio_data_np, sample_rate, container_format, codec_name - ) - - return await upload_file_to_comfyapi(audio_bytes_io, filename, mime_type, auth_kwargs) - - -def f32_pcm(wav: torch.Tensor) -> torch.Tensor: - """Convert audio to float 32 bits PCM format. Copy-paste from nodes_audio.py file.""" - if wav.dtype.is_floating_point: - return wav - elif wav.dtype == torch.int16: - return wav.float() / (2 ** 15) - elif wav.dtype == torch.int32: - return wav.float() / (2 ** 31) - raise ValueError(f"Unsupported wav dtype: {wav.dtype}") - - -def audio_bytes_to_audio_input(audio_bytes: bytes,) -> dict: - """ - Decode any common audio container from bytes using PyAV and return - a Comfy AUDIO dict: {"waveform": [1, C, T] float32, "sample_rate": int}. - """ - with av.open(io.BytesIO(audio_bytes)) as af: - if not af.streams.audio: - raise ValueError("No audio stream found in response.") - stream = af.streams.audio[0] - - in_sr = int(stream.codec_context.sample_rate) - out_sr = in_sr - - frames: list[torch.Tensor] = [] - n_channels = stream.channels or 1 - - for frame in af.decode(streams=stream.index): - arr = frame.to_ndarray() # shape can be [C, T] or [T, C] or [T] - buf = torch.from_numpy(arr) - if buf.ndim == 1: - buf = buf.unsqueeze(0) # [T] -> [1, T] - elif buf.shape[0] != n_channels and buf.shape[-1] == n_channels: - buf = buf.transpose(0, 1).contiguous() # [T, C] -> [C, T] - elif buf.shape[0] != n_channels: - buf = buf.reshape(-1, n_channels).t().contiguous() # fallback to [C, T] - frames.append(buf) - - if not frames: - raise ValueError("Decoded zero audio frames.") - - wav = torch.cat(frames, dim=1) # [C, T] - wav = f32_pcm(wav) - return {"waveform": wav.unsqueeze(0).contiguous(), "sample_rate": out_sr} - - -def audio_input_to_mp3(audio: AudioInput) -> io.BytesIO: - waveform = audio["waveform"].cpu() - - output_buffer = io.BytesIO() - output_container = av.open(output_buffer, mode='w', format="mp3") - - out_stream = output_container.add_stream("libmp3lame", rate=audio["sample_rate"]) - out_stream.bit_rate = 320000 - - frame = av.AudioFrame.from_ndarray(waveform.movedim(0, 1).reshape(1, -1).float().numpy(), format='flt', layout='mono' if waveform.shape[0] == 1 else 'stereo') - frame.sample_rate = audio["sample_rate"] - frame.pts = 0 - output_container.mux(out_stream.encode(frame)) - output_container.mux(out_stream.encode(None)) - output_container.close() - output_buffer.seek(0) - return output_buffer - - -def audio_to_base64_string( - audio: AudioInput, container_format: str = "mp4", codec_name: str = "aac" -) -> str: - """Converts an audio input to a base64 string.""" - sample_rate: int = audio["sample_rate"] - waveform: torch.Tensor = audio["waveform"] - audio_data_np = audio_tensor_to_contiguous_ndarray(waveform) - audio_bytes_io = audio_ndarray_to_bytesio( - audio_data_np, sample_rate, container_format, codec_name - ) - audio_bytes = audio_bytes_io.getvalue() - return base64.b64encode(audio_bytes).decode("utf-8") - - async def upload_images_to_comfyapi( image: torch.Tensor, max_images=8, @@ -663,56 +285,3 @@ def resize_mask_to_image( if not allow_gradient: mask = (mask > 0.5).float() return mask - - -def validate_string( - string: str, - strip_whitespace=True, - field_name="prompt", - min_length=None, - max_length=None, -): - if string is None: - raise Exception(f"Field '{field_name}' cannot be empty.") - if strip_whitespace: - string = string.strip() - if min_length and len(string) < min_length: - raise Exception( - f"Field '{field_name}' cannot be shorter than {min_length} characters; was {len(string)} characters long." - ) - if max_length and len(string) > max_length: - raise Exception( - f" Field '{field_name} cannot be longer than {max_length} characters; was {len(string)} characters long." - ) - - -def image_tensor_pair_to_batch( - image1: torch.Tensor, image2: torch.Tensor -) -> torch.Tensor: - """ - Converts a pair of image tensors to a batch tensor. - If the images are not the same size, the smaller image is resized to - match the larger image. - """ - if image1.shape[1:] != image2.shape[1:]: - image2 = common_upscale( - image2.movedim(-1, 1), - image1.shape[2], - image1.shape[1], - "bilinear", - "center", - ).movedim(1, -1) - return torch.cat((image1, image2), dim=0) - - -def get_size(path_or_object: Union[str, io.BytesIO]) -> int: - if isinstance(path_or_object, str): - return os.path.getsize(path_or_object) - return len(path_or_object.getvalue()) - - -def validate_container_format_is_mp4(video: VideoInput) -> None: - """Validates video container format is MP4.""" - container_format = video.get_container_format() - if container_format not in ["mp4", "mov,mp4,m4a,3gp,3g2,mj2"]: - raise ValueError(f"Only MP4 container format supported. Got: {container_format}") diff --git a/comfy_api_nodes/apis/bfl_api.py b/comfy_api_nodes/apis/bfl_api.py index 0e90aef7c..0fc8c0607 100644 --- a/comfy_api_nodes/apis/bfl_api.py +++ b/comfy_api_nodes/apis/bfl_api.py @@ -50,44 +50,6 @@ class BFLFluxFillImageRequest(BaseModel): mask: str = Field(None, description='A Base64-encoded string representing the mask of the areas you with to modify.') -class BFLFluxCannyImageRequest(BaseModel): - prompt: str = Field(..., description='Text prompt for image generation') - prompt_upsampling: Optional[bool] = Field( - None, description='Whether to perform upsampling on the prompt. If active, automatically modifies the prompt for more creative generation.' - ) - canny_low_threshold: Optional[int] = Field(None, description='Low threshold for Canny edge detection') - canny_high_threshold: Optional[int] = Field(None, description='High threshold for Canny edge detection') - seed: Optional[int] = Field(None, description='The seed value for reproducibility.') - steps: conint(ge=15, le=50) = Field(..., description='Number of steps for the image generation process') - guidance: confloat(ge=1, le=100) = Field(..., description='Guidance strength for the image generation process') - safety_tolerance: Optional[conint(ge=0, le=6)] = Field( - 6, description='Tolerance level for input and output moderation. Between 0 and 6, 0 being most strict, 6 being least strict. Defaults to 2.' - ) - output_format: Optional[BFLOutputFormat] = Field( - BFLOutputFormat.png, description="Output format for the generated image. Can be 'jpeg' or 'png'.", examples=['png'] - ) - control_image: Optional[str] = Field(None, description='Base64 encoded image to use as control input if no preprocessed image is provided') - preprocessed_image: Optional[str] = Field(None, description='Optional pre-processed image that will bypass the control preprocessing step') - - -class BFLFluxDepthImageRequest(BaseModel): - prompt: str = Field(..., description='Text prompt for image generation') - prompt_upsampling: Optional[bool] = Field( - None, description='Whether to perform upsampling on the prompt. If active, automatically modifies the prompt for more creative generation.' - ) - seed: Optional[int] = Field(None, description='The seed value for reproducibility.') - steps: conint(ge=15, le=50) = Field(..., description='Number of steps for the image generation process') - guidance: confloat(ge=1, le=100) = Field(..., description='Guidance strength for the image generation process') - safety_tolerance: Optional[conint(ge=0, le=6)] = Field( - 6, description='Tolerance level for input and output moderation. Between 0 and 6, 0 being most strict, 6 being least strict. Defaults to 2.' - ) - output_format: Optional[BFLOutputFormat] = Field( - BFLOutputFormat.png, description="Output format for the generated image. Can be 'jpeg' or 'png'.", examples=['png'] - ) - control_image: Optional[str] = Field(None, description='Base64 encoded image to use as control input if no preprocessed image is provided') - preprocessed_image: Optional[str] = Field(None, description='Optional pre-processed image that will bypass the control preprocessing step') - - class BFLFluxProGenerateRequest(BaseModel): prompt: str = Field(..., description='The text prompt for image generation.') prompt_upsampling: Optional[bool] = Field( @@ -160,15 +122,8 @@ class BFLStatus(str, Enum): error = "Error" -class BFLFluxProStatusResponse(BaseModel): +class BFLFluxStatusResponse(BaseModel): id: str = Field(..., description="The unique identifier for the generation task.") status: BFLStatus = Field(..., description="The status of the task.") - result: Optional[Dict[str, Any]] = Field( - None, description="The result of the task (null if not completed)." - ) - progress: confloat(ge=0.0, le=1.0) = Field( - ..., description="The progress of the task (0.0 to 1.0)." - ) - details: Optional[Dict[str, Any]] = Field( - None, description="Additional details about the task (null if not available)." - ) + result: Optional[Dict[str, Any]] = Field(None, description="The result of the task (null if not completed).") + progress: Optional[float] = Field(None, description="The progress of the task (0.0 to 1.0).", ge=0.0, le=1.0) diff --git a/comfy_api_nodes/apis/veo_api.py b/comfy_api_nodes/apis/veo_api.py new file mode 100644 index 000000000..a55137afb --- /dev/null +++ b/comfy_api_nodes/apis/veo_api.py @@ -0,0 +1,111 @@ +from typing import Optional, Union +from enum import Enum + +from pydantic import BaseModel, Field + + +class Image2(BaseModel): + bytesBase64Encoded: str + gcsUri: Optional[str] = None + mimeType: Optional[str] = None + + +class Image3(BaseModel): + bytesBase64Encoded: Optional[str] = None + gcsUri: str + mimeType: Optional[str] = None + + +class Instance1(BaseModel): + image: Optional[Union[Image2, Image3]] = Field( + None, description='Optional image to guide video generation' + ) + prompt: str = Field(..., description='Text description of the video') + + +class PersonGeneration1(str, Enum): + ALLOW = 'ALLOW' + BLOCK = 'BLOCK' + + +class Parameters1(BaseModel): + aspectRatio: Optional[str] = Field(None, examples=['16:9']) + durationSeconds: Optional[int] = None + enhancePrompt: Optional[bool] = None + generateAudio: Optional[bool] = Field( + None, + description='Generate audio for the video. Only supported by veo 3 models.', + ) + negativePrompt: Optional[str] = None + personGeneration: Optional[PersonGeneration1] = None + sampleCount: Optional[int] = None + seed: Optional[int] = None + storageUri: Optional[str] = Field( + None, description='Optional Cloud Storage URI to upload the video' + ) + + +class VeoGenVidRequest(BaseModel): + instances: Optional[list[Instance1]] = None + parameters: Optional[Parameters1] = None + + +class VeoGenVidResponse(BaseModel): + name: str = Field( + ..., + description='Operation resource name', + examples=[ + 'projects/PROJECT_ID/locations/us-central1/publishers/google/models/MODEL_ID/operations/a1b07c8e-7b5a-4aba-bb34-3e1ccb8afcc8' + ], + ) + + +class VeoGenVidPollRequest(BaseModel): + operationName: str = Field( + ..., + description='Full operation name (from predict response)', + examples=[ + 'projects/PROJECT_ID/locations/us-central1/publishers/google/models/MODEL_ID/operations/OPERATION_ID' + ], + ) + + +class Video(BaseModel): + bytesBase64Encoded: Optional[str] = Field( + None, description='Base64-encoded video content' + ) + gcsUri: Optional[str] = Field(None, description='Cloud Storage URI of the video') + mimeType: Optional[str] = Field(None, description='Video MIME type') + + +class Error1(BaseModel): + code: Optional[int] = Field(None, description='Error code') + message: Optional[str] = Field(None, description='Error message') + + +class Response1(BaseModel): + field_type: Optional[str] = Field( + None, + alias='@type', + examples=[ + 'type.googleapis.com/cloud.ai.large_models.vision.GenerateVideoResponse' + ], + ) + raiMediaFilteredCount: Optional[int] = Field( + None, description='Count of media filtered by responsible AI policies' + ) + raiMediaFilteredReasons: Optional[list[str]] = Field( + None, description='Reasons why media was filtered by responsible AI policies' + ) + videos: Optional[list[Video]] = None + + +class VeoGenVidPollResponse(BaseModel): + done: Optional[bool] = None + error: Optional[Error1] = Field( + None, description='Error details if operation failed' + ) + name: Optional[str] = None + response: Optional[Response1] = Field( + None, description='The actual prediction response if done is true' + ) diff --git a/comfy_api_nodes/nodes_bfl.py b/comfy_api_nodes/nodes_bfl.py index b6cc90f05..baa74fd52 100644 --- a/comfy_api_nodes/nodes_bfl.py +++ b/comfy_api_nodes/nodes_bfl.py @@ -1,136 +1,43 @@ -import asyncio -import io from inspect import cleandoc -from typing import Union, Optional +from typing import Optional + +import torch from typing_extensions import override -from comfy_api.latest import ComfyExtension, IO + +from comfy_api.latest import IO, ComfyExtension +from comfy_api_nodes.apinode_utils import ( + resize_mask_to_image, + validate_aspect_ratio, +) from comfy_api_nodes.apis.bfl_api import ( - BFLStatus, BFLFluxExpandImageRequest, BFLFluxFillImageRequest, - BFLFluxCannyImageRequest, - BFLFluxDepthImageRequest, - BFLFluxProGenerateRequest, BFLFluxKontextProGenerateRequest, - BFLFluxProUltraGenerateRequest, + BFLFluxProGenerateRequest, BFLFluxProGenerateResponse, + BFLFluxProUltraGenerateRequest, + BFLFluxStatusResponse, + BFLStatus, ) -from comfy_api_nodes.apis.client import ( +from comfy_api_nodes.util import ( ApiEndpoint, - HttpMethod, - SynchronousOperation, -) -from comfy_api_nodes.apinode_utils import ( - downscale_image_tensor, - validate_aspect_ratio, - process_image_response, - resize_mask_to_image, + download_url_to_image_tensor, + poll_op, + sync_op, + tensor_to_base64_string, validate_string, ) -import numpy as np -from PIL import Image -import aiohttp -import torch -import base64 -import time -from server import PromptServer - def convert_mask_to_image(mask: torch.Tensor): """ Make mask have the expected amount of dims (4) and channels (3) to be recognized as an image. """ mask = mask.unsqueeze(-1) - mask = torch.cat([mask]*3, dim=-1) + mask = torch.cat([mask] * 3, dim=-1) return mask -async def handle_bfl_synchronous_operation( - operation: SynchronousOperation, - timeout_bfl_calls=360, - node_id: Union[str, None] = None, -): - response_api: BFLFluxProGenerateResponse = await operation.execute() - return await _poll_until_generated( - response_api.polling_url, timeout=timeout_bfl_calls, node_id=node_id - ) - - -async def _poll_until_generated( - polling_url: str, timeout=360, node_id: Union[str, None] = None -): - # used bfl-comfy-nodes to verify code implementation: - # https://github.com/black-forest-labs/bfl-comfy-nodes/tree/main - start_time = time.time() - retries_404 = 0 - max_retries_404 = 5 - retry_404_seconds = 2 - retry_202_seconds = 2 - retry_pending_seconds = 1 - - async with aiohttp.ClientSession() as session: - # NOTE: should True loop be replaced with checking if workflow has been interrupted? - while True: - if node_id: - time_elapsed = time.time() - start_time - PromptServer.instance.send_progress_text( - f"Generating ({time_elapsed:.0f}s)", node_id - ) - - async with session.get(polling_url) as response: - if response.status == 200: - result = await response.json() - if result["status"] == BFLStatus.ready: - img_url = result["result"]["sample"] - if node_id: - PromptServer.instance.send_progress_text( - f"Result URL: {img_url}", node_id - ) - async with session.get(img_url) as img_resp: - return process_image_response(await img_resp.content.read()) - elif result["status"] in [ - BFLStatus.request_moderated, - BFLStatus.content_moderated, - ]: - status = result["status"] - raise Exception( - f"BFL API did not return an image due to: {status}." - ) - elif result["status"] == BFLStatus.error: - raise Exception(f"BFL API encountered an error: {result}.") - elif result["status"] == BFLStatus.pending: - await asyncio.sleep(retry_pending_seconds) - continue - elif response.status == 404: - if retries_404 < max_retries_404: - retries_404 += 1 - await asyncio.sleep(retry_404_seconds) - continue - raise Exception( - f"BFL API could not find task after {max_retries_404} tries." - ) - elif response.status == 202: - await asyncio.sleep(retry_202_seconds) - elif time.time() - start_time > timeout: - raise Exception( - f"BFL API experienced a timeout; could not return request under {timeout} seconds." - ) - else: - raise Exception(f"BFL API encountered an error: {response.json()}") - -def convert_image_to_base64(image: torch.Tensor): - scaled_image = downscale_image_tensor(image, total_pixels=2048 * 2048) - # remove batch dimension if present - if len(scaled_image.shape) > 3: - scaled_image = scaled_image[0] - image_np = (scaled_image.numpy() * 255).astype(np.uint8) - img = Image.fromarray(image_np) - img_byte_arr = io.BytesIO() - img.save(img_byte_arr, format="PNG") - return base64.b64encode(img_byte_arr.getvalue()).decode() - - class FluxProUltraImageNode(IO.ComfyNode): """ Generates images using Flux Pro 1.1 Ultra via api based on prompt and resolution. @@ -158,7 +65,9 @@ class FluxProUltraImageNode(IO.ComfyNode): IO.Boolean.Input( "prompt_upsampling", default=False, - tooltip="Whether to perform upsampling on the prompt. If active, automatically modifies the prompt for more creative generation, but results are nondeterministic (same seed will not produce exactly the same result).", + tooltip="Whether to perform upsampling on the prompt. " + "If active, automatically modifies the prompt for more creative generation, " + "but results are nondeterministic (same seed will not produce exactly the same result).", ), IO.Int.Input( "seed", @@ -220,22 +129,19 @@ class FluxProUltraImageNode(IO.ComfyNode): cls, prompt: str, aspect_ratio: str, - prompt_upsampling=False, - raw=False, - seed=0, - image_prompt=None, - image_prompt_strength=0.1, + prompt_upsampling: bool = False, + raw: bool = False, + seed: int = 0, + image_prompt: Optional[torch.Tensor] = None, + image_prompt_strength: float = 0.1, ) -> IO.NodeOutput: if image_prompt is None: validate_string(prompt, strip_whitespace=False) - operation = SynchronousOperation( - endpoint=ApiEndpoint( - path="/proxy/bfl/flux-pro-1.1-ultra/generate", - method=HttpMethod.POST, - request_model=BFLFluxProUltraGenerateRequest, - response_model=BFLFluxProGenerateResponse, - ), - request=BFLFluxProUltraGenerateRequest( + initial_response = await sync_op( + cls, + ApiEndpoint(path="/proxy/bfl/flux-pro-1.1-ultra/generate", method="POST"), + response_model=BFLFluxProGenerateResponse, + data=BFLFluxProUltraGenerateRequest( prompt=prompt, prompt_upsampling=prompt_upsampling, seed=seed, @@ -247,22 +153,26 @@ class FluxProUltraImageNode(IO.ComfyNode): maximum_ratio_str=cls.MAXIMUM_RATIO_STR, ), raw=raw, - image_prompt=( - image_prompt - if image_prompt is None - else convert_image_to_base64(image_prompt) - ), - image_prompt_strength=( - None if image_prompt is None else round(image_prompt_strength, 2) - ), + image_prompt=(image_prompt if image_prompt is None else tensor_to_base64_string(image_prompt)), + image_prompt_strength=(None if image_prompt is None else round(image_prompt_strength, 2)), ), - auth_kwargs={ - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - }, ) - output_image = await handle_bfl_synchronous_operation(operation, node_id=cls.hidden.unique_id) - return IO.NodeOutput(output_image) + response = await poll_op( + cls, + ApiEndpoint(initial_response.polling_url), + response_model=BFLFluxStatusResponse, + status_extractor=lambda r: r.status, + progress_extractor=lambda r: r.progress, + completed_statuses=[BFLStatus.ready], + failed_statuses=[ + BFLStatus.request_moderated, + BFLStatus.content_moderated, + BFLStatus.error, + BFLStatus.task_not_found, + ], + queued_statuses=[], + ) + return IO.NodeOutput(await download_url_to_image_tensor(response.result["sample"])) class FluxKontextProImageNode(IO.ComfyNode): @@ -347,7 +257,7 @@ class FluxKontextProImageNode(IO.ComfyNode): aspect_ratio: str, guidance: float, steps: int, - input_image: Optional[torch.Tensor]=None, + input_image: Optional[torch.Tensor] = None, seed=0, prompt_upsampling=False, ) -> IO.NodeOutput: @@ -360,33 +270,36 @@ class FluxKontextProImageNode(IO.ComfyNode): ) if input_image is None: validate_string(prompt, strip_whitespace=False) - operation = SynchronousOperation( - endpoint=ApiEndpoint( - path=cls.BFL_PATH, - method=HttpMethod.POST, - request_model=BFLFluxKontextProGenerateRequest, - response_model=BFLFluxProGenerateResponse, - ), - request=BFLFluxKontextProGenerateRequest( + initial_response = await sync_op( + cls, + ApiEndpoint(path=cls.BFL_PATH, method="POST"), + response_model=BFLFluxProGenerateResponse, + data=BFLFluxKontextProGenerateRequest( prompt=prompt, prompt_upsampling=prompt_upsampling, guidance=round(guidance, 1), steps=steps, seed=seed, aspect_ratio=aspect_ratio, - input_image=( - input_image - if input_image is None - else convert_image_to_base64(input_image) - ) + input_image=(input_image if input_image is None else tensor_to_base64_string(input_image)), ), - auth_kwargs={ - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - }, ) - output_image = await handle_bfl_synchronous_operation(operation, node_id=cls.hidden.unique_id) - return IO.NodeOutput(output_image) + response = await poll_op( + cls, + ApiEndpoint(initial_response.polling_url), + response_model=BFLFluxStatusResponse, + status_extractor=lambda r: r.status, + progress_extractor=lambda r: r.progress, + completed_statuses=[BFLStatus.ready], + failed_statuses=[ + BFLStatus.request_moderated, + BFLStatus.content_moderated, + BFLStatus.error, + BFLStatus.task_not_found, + ], + queued_statuses=[], + ) + return IO.NodeOutput(await download_url_to_image_tensor(response.result["sample"])) class FluxKontextMaxImageNode(FluxKontextProImageNode): @@ -422,7 +335,9 @@ class FluxProImageNode(IO.ComfyNode): IO.Boolean.Input( "prompt_upsampling", default=False, - tooltip="Whether to perform upsampling on the prompt. If active, automatically modifies the prompt for more creative generation, but results are nondeterministic (same seed will not produce exactly the same result).", + tooltip="Whether to perform upsampling on the prompt. " + "If active, automatically modifies the prompt for more creative generation, " + "but results are nondeterministic (same seed will not produce exactly the same result).", ), IO.Int.Input( "width", @@ -481,20 +396,15 @@ class FluxProImageNode(IO.ComfyNode): image_prompt=None, # image_prompt_strength=0.1, ) -> IO.NodeOutput: - image_prompt = ( - image_prompt - if image_prompt is None - else convert_image_to_base64(image_prompt) - ) - - operation = SynchronousOperation( - endpoint=ApiEndpoint( + image_prompt = image_prompt if image_prompt is None else tensor_to_base64_string(image_prompt) + initial_response = await sync_op( + cls, + ApiEndpoint( path="/proxy/bfl/flux-pro-1.1/generate", - method=HttpMethod.POST, - request_model=BFLFluxProGenerateRequest, - response_model=BFLFluxProGenerateResponse, + method="POST", ), - request=BFLFluxProGenerateRequest( + response_model=BFLFluxProGenerateResponse, + data=BFLFluxProGenerateRequest( prompt=prompt, prompt_upsampling=prompt_upsampling, width=width, @@ -502,13 +412,23 @@ class FluxProImageNode(IO.ComfyNode): seed=seed, image_prompt=image_prompt, ), - auth_kwargs={ - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - }, ) - output_image = await handle_bfl_synchronous_operation(operation, node_id=cls.hidden.unique_id) - return IO.NodeOutput(output_image) + response = await poll_op( + cls, + ApiEndpoint(initial_response.polling_url), + response_model=BFLFluxStatusResponse, + status_extractor=lambda r: r.status, + progress_extractor=lambda r: r.progress, + completed_statuses=[BFLStatus.ready], + failed_statuses=[ + BFLStatus.request_moderated, + BFLStatus.content_moderated, + BFLStatus.error, + BFLStatus.task_not_found, + ], + queued_statuses=[], + ) + return IO.NodeOutput(await download_url_to_image_tensor(response.result["sample"])) class FluxProExpandNode(IO.ComfyNode): @@ -534,7 +454,9 @@ class FluxProExpandNode(IO.ComfyNode): IO.Boolean.Input( "prompt_upsampling", default=False, - tooltip="Whether to perform upsampling on the prompt. If active, automatically modifies the prompt for more creative generation, but results are nondeterministic (same seed will not produce exactly the same result).", + tooltip="Whether to perform upsampling on the prompt. " + "If active, automatically modifies the prompt for more creative generation, " + "but results are nondeterministic (same seed will not produce exactly the same result).", ), IO.Int.Input( "top", @@ -610,16 +532,11 @@ class FluxProExpandNode(IO.ComfyNode): guidance: float, seed=0, ) -> IO.NodeOutput: - image = convert_image_to_base64(image) - - operation = SynchronousOperation( - endpoint=ApiEndpoint( - path="/proxy/bfl/flux-pro-1.0-expand/generate", - method=HttpMethod.POST, - request_model=BFLFluxExpandImageRequest, - response_model=BFLFluxProGenerateResponse, - ), - request=BFLFluxExpandImageRequest( + initial_response = await sync_op( + cls, + ApiEndpoint(path="/proxy/bfl/flux-pro-1.0-expand/generate", method="POST"), + response_model=BFLFluxProGenerateResponse, + data=BFLFluxExpandImageRequest( prompt=prompt, prompt_upsampling=prompt_upsampling, top=top, @@ -629,16 +546,25 @@ class FluxProExpandNode(IO.ComfyNode): steps=steps, guidance=guidance, seed=seed, - image=image, + image=tensor_to_base64_string(image), ), - auth_kwargs={ - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - }, ) - output_image = await handle_bfl_synchronous_operation(operation, node_id=cls.hidden.unique_id) - return IO.NodeOutput(output_image) - + response = await poll_op( + cls, + ApiEndpoint(initial_response.polling_url), + response_model=BFLFluxStatusResponse, + status_extractor=lambda r: r.status, + progress_extractor=lambda r: r.progress, + completed_statuses=[BFLStatus.ready], + failed_statuses=[ + BFLStatus.request_moderated, + BFLStatus.content_moderated, + BFLStatus.error, + BFLStatus.task_not_found, + ], + queued_statuses=[], + ) + return IO.NodeOutput(await download_url_to_image_tensor(response.result["sample"])) class FluxProFillNode(IO.ComfyNode): @@ -665,7 +591,9 @@ class FluxProFillNode(IO.ComfyNode): IO.Boolean.Input( "prompt_upsampling", default=False, - tooltip="Whether to perform upsampling on the prompt. If active, automatically modifies the prompt for more creative generation, but results are nondeterministic (same seed will not produce exactly the same result).", + tooltip="Whether to perform upsampling on the prompt. " + "If active, automatically modifies the prompt for more creative generation, " + "but results are nondeterministic (same seed will not produce exactly the same result).", ), IO.Float.Input( "guidance", @@ -712,272 +640,37 @@ class FluxProFillNode(IO.ComfyNode): ) -> IO.NodeOutput: # prepare mask mask = resize_mask_to_image(mask, image) - mask = convert_image_to_base64(convert_mask_to_image(mask)) - # make sure image will have alpha channel removed - image = convert_image_to_base64(image[:, :, :, :3]) - - operation = SynchronousOperation( - endpoint=ApiEndpoint( - path="/proxy/bfl/flux-pro-1.0-fill/generate", - method=HttpMethod.POST, - request_model=BFLFluxFillImageRequest, - response_model=BFLFluxProGenerateResponse, - ), - request=BFLFluxFillImageRequest( + mask = tensor_to_base64_string(convert_mask_to_image(mask)) + initial_response = await sync_op( + cls, + ApiEndpoint(path="/proxy/bfl/flux-pro-1.0-fill/generate", method="POST"), + response_model=BFLFluxProGenerateResponse, + data=BFLFluxFillImageRequest( prompt=prompt, prompt_upsampling=prompt_upsampling, steps=steps, guidance=guidance, seed=seed, - image=image, + image=tensor_to_base64_string(image[:, :, :, :3]), # make sure image will have alpha channel removed mask=mask, ), - auth_kwargs={ - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - }, ) - output_image = await handle_bfl_synchronous_operation(operation, node_id=cls.hidden.unique_id) - return IO.NodeOutput(output_image) - - -class FluxProCannyNode(IO.ComfyNode): - """ - Generate image using a control image (canny). - """ - - @classmethod - def define_schema(cls) -> IO.Schema: - return IO.Schema( - node_id="FluxProCannyNode", - display_name="Flux.1 Canny Control Image", - category="api node/image/BFL", - description=cleandoc(cls.__doc__ or ""), - inputs=[ - IO.Image.Input("control_image"), - IO.String.Input( - "prompt", - multiline=True, - default="", - tooltip="Prompt for the image generation", - ), - IO.Boolean.Input( - "prompt_upsampling", - default=False, - tooltip="Whether to perform upsampling on the prompt. If active, automatically modifies the prompt for more creative generation, but results are nondeterministic (same seed will not produce exactly the same result).", - ), - IO.Float.Input( - "canny_low_threshold", - default=0.1, - min=0.01, - max=0.99, - step=0.01, - tooltip="Low threshold for Canny edge detection; ignored if skip_processing is True", - ), - IO.Float.Input( - "canny_high_threshold", - default=0.4, - min=0.01, - max=0.99, - step=0.01, - tooltip="High threshold for Canny edge detection; ignored if skip_processing is True", - ), - IO.Boolean.Input( - "skip_preprocessing", - default=False, - tooltip="Whether to skip preprocessing; set to True if control_image already is canny-fied, False if it is a raw image.", - ), - IO.Float.Input( - "guidance", - default=30, - min=1, - max=100, - tooltip="Guidance strength for the image generation process", - ), - IO.Int.Input( - "steps", - default=50, - min=15, - max=50, - tooltip="Number of steps for the image generation process", - ), - IO.Int.Input( - "seed", - default=0, - min=0, - max=0xFFFFFFFFFFFFFFFF, - control_after_generate=True, - tooltip="The random seed used for creating the noise.", - ), + response = await poll_op( + cls, + ApiEndpoint(initial_response.polling_url), + response_model=BFLFluxStatusResponse, + status_extractor=lambda r: r.status, + progress_extractor=lambda r: r.progress, + completed_statuses=[BFLStatus.ready], + failed_statuses=[ + BFLStatus.request_moderated, + BFLStatus.content_moderated, + BFLStatus.error, + BFLStatus.task_not_found, ], - outputs=[IO.Image.Output()], - hidden=[ - IO.Hidden.auth_token_comfy_org, - IO.Hidden.api_key_comfy_org, - IO.Hidden.unique_id, - ], - is_api_node=True, + queued_statuses=[], ) - - @classmethod - async def execute( - cls, - control_image: torch.Tensor, - prompt: str, - prompt_upsampling: bool, - canny_low_threshold: float, - canny_high_threshold: float, - skip_preprocessing: bool, - steps: int, - guidance: float, - seed=0, - ) -> IO.NodeOutput: - control_image = convert_image_to_base64(control_image[:, :, :, :3]) - preprocessed_image = None - - # scale canny threshold between 0-500, to match BFL's API - def scale_value(value: float, min_val=0, max_val=500): - return min_val + value * (max_val - min_val) - canny_low_threshold = int(round(scale_value(canny_low_threshold))) - canny_high_threshold = int(round(scale_value(canny_high_threshold))) - - - if skip_preprocessing: - preprocessed_image = control_image - control_image = None - canny_low_threshold = None - canny_high_threshold = None - - operation = SynchronousOperation( - endpoint=ApiEndpoint( - path="/proxy/bfl/flux-pro-1.0-canny/generate", - method=HttpMethod.POST, - request_model=BFLFluxCannyImageRequest, - response_model=BFLFluxProGenerateResponse, - ), - request=BFLFluxCannyImageRequest( - prompt=prompt, - prompt_upsampling=prompt_upsampling, - steps=steps, - guidance=guidance, - seed=seed, - control_image=control_image, - canny_low_threshold=canny_low_threshold, - canny_high_threshold=canny_high_threshold, - preprocessed_image=preprocessed_image, - ), - auth_kwargs={ - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - }, - ) - output_image = await handle_bfl_synchronous_operation(operation, node_id=cls.hidden.unique_id) - return IO.NodeOutput(output_image) - - -class FluxProDepthNode(IO.ComfyNode): - """ - Generate image using a control image (depth). - """ - - @classmethod - def define_schema(cls) -> IO.Schema: - return IO.Schema( - node_id="FluxProDepthNode", - display_name="Flux.1 Depth Control Image", - category="api node/image/BFL", - description=cleandoc(cls.__doc__ or ""), - inputs=[ - IO.Image.Input("control_image"), - IO.String.Input( - "prompt", - multiline=True, - default="", - tooltip="Prompt for the image generation", - ), - IO.Boolean.Input( - "prompt_upsampling", - default=False, - tooltip="Whether to perform upsampling on the prompt. If active, automatically modifies the prompt for more creative generation, but results are nondeterministic (same seed will not produce exactly the same result).", - ), - IO.Boolean.Input( - "skip_preprocessing", - default=False, - tooltip="Whether to skip preprocessing; set to True if control_image already is depth-ified, False if it is a raw image.", - ), - IO.Float.Input( - "guidance", - default=15, - min=1, - max=100, - tooltip="Guidance strength for the image generation process", - ), - IO.Int.Input( - "steps", - default=50, - min=15, - max=50, - tooltip="Number of steps for the image generation process", - ), - IO.Int.Input( - "seed", - default=0, - min=0, - max=0xFFFFFFFFFFFFFFFF, - control_after_generate=True, - tooltip="The random seed used for creating the noise.", - ), - ], - outputs=[IO.Image.Output()], - hidden=[ - IO.Hidden.auth_token_comfy_org, - IO.Hidden.api_key_comfy_org, - IO.Hidden.unique_id, - ], - is_api_node=True, - ) - - @classmethod - async def execute( - cls, - control_image: torch.Tensor, - prompt: str, - prompt_upsampling: bool, - skip_preprocessing: bool, - steps: int, - guidance: float, - seed=0, - ) -> IO.NodeOutput: - control_image = convert_image_to_base64(control_image[:,:,:,:3]) - preprocessed_image = None - - if skip_preprocessing: - preprocessed_image = control_image - control_image = None - - operation = SynchronousOperation( - endpoint=ApiEndpoint( - path="/proxy/bfl/flux-pro-1.0-depth/generate", - method=HttpMethod.POST, - request_model=BFLFluxDepthImageRequest, - response_model=BFLFluxProGenerateResponse, - ), - request=BFLFluxDepthImageRequest( - prompt=prompt, - prompt_upsampling=prompt_upsampling, - steps=steps, - guidance=guidance, - seed=seed, - control_image=control_image, - preprocessed_image=preprocessed_image, - ), - auth_kwargs={ - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - }, - ) - output_image = await handle_bfl_synchronous_operation(operation, node_id=cls.hidden.unique_id) - return IO.NodeOutput(output_image) + return IO.NodeOutput(await download_url_to_image_tensor(response.result["sample"])) class BFLExtension(ComfyExtension): @@ -990,8 +683,6 @@ class BFLExtension(ComfyExtension): FluxKontextMaxImageNode, FluxProExpandNode, FluxProFillNode, - FluxProCannyNode, - FluxProDepthNode, ] diff --git a/comfy_api_nodes/nodes_bytedance.py b/comfy_api_nodes/nodes_bytedance.py index f3d3f8d3e..534af380d 100644 --- a/comfy_api_nodes/nodes_bytedance.py +++ b/comfy_api_nodes/nodes_bytedance.py @@ -1,35 +1,27 @@ import logging import math from enum import Enum -from typing import Literal, Optional, Type, Union -from typing_extensions import override +from typing import Literal, Optional, Union import torch from pydantic import BaseModel, Field +from typing_extensions import override -from comfy_api.latest import ComfyExtension, IO -from comfy_api_nodes.util.validation_utils import ( - validate_image_aspect_ratio_range, - get_number_of_images, - validate_image_dimensions, -) -from comfy_api_nodes.apis.client import ( +from comfy_api.latest import IO, ComfyExtension +from comfy_api_nodes.util import ( ApiEndpoint, - EmptyRequest, - HttpMethod, - SynchronousOperation, - PollingOperation, - T, -) -from comfy_api_nodes.apinode_utils import ( download_url_to_image_tensor, download_url_to_video_output, - upload_images_to_comfyapi, - validate_string, + get_number_of_images, image_tensor_pair_to_batch, + poll_op, + sync_op, + upload_images_to_comfyapi, + validate_image_aspect_ratio_range, + validate_image_dimensions, + validate_string, ) - BYTEPLUS_IMAGE_ENDPOINT = "/proxy/byteplus/api/v3/images/generations" # Long-running tasks endpoints(e.g., video) @@ -46,13 +38,14 @@ class Image2ImageModelName(str, Enum): class Text2VideoModelName(str, Enum): - seedance_1_pro = "seedance-1-0-pro-250528" + seedance_1_pro = "seedance-1-0-pro-250528" seedance_1_lite = "seedance-1-0-lite-t2v-250428" class Image2VideoModelName(str, Enum): """note(August 31): Pro model only supports FirstFrame: https://docs.byteplus.com/en/docs/ModelArk/1520757""" - seedance_1_pro = "seedance-1-0-pro-250528" + + seedance_1_pro = "seedance-1-0-pro-250528" seedance_1_lite = "seedance-1-0-lite-i2v-250428" @@ -208,35 +201,6 @@ def get_video_url_from_task_status(response: TaskStatusResponse) -> Union[str, N return None -async def poll_until_finished( - auth_kwargs: dict[str, str], - task_id: str, - estimated_duration: Optional[int] = None, - node_id: Optional[str] = None, -) -> TaskStatusResponse: - """Polls the ByteDance API endpoint until the task reaches a terminal state, then returns the response.""" - return await PollingOperation( - poll_endpoint=ApiEndpoint( - path=f"{BYTEPLUS_TASK_STATUS_ENDPOINT}/{task_id}", - method=HttpMethod.GET, - request_model=EmptyRequest, - response_model=TaskStatusResponse, - ), - completed_statuses=[ - "succeeded", - ], - failed_statuses=[ - "cancelled", - "failed", - ], - status_extractor=lambda response: response.status, - auth_kwargs=auth_kwargs, - result_url_extractor=get_video_url_from_task_status, - estimated_duration=estimated_duration, - node_id=node_id, - ).execute() - - class ByteDanceImageNode(IO.ComfyNode): @classmethod @@ -303,7 +267,7 @@ class ByteDanceImageNode(IO.ComfyNode): IO.Boolean.Input( "watermark", default=True, - tooltip="Whether to add an \"AI generated\" watermark to the image", + tooltip='Whether to add an "AI generated" watermark to the image', optional=True, ), ], @@ -341,8 +305,7 @@ class ByteDanceImageNode(IO.ComfyNode): w, h = width, height if not (512 <= w <= 2048) or not (512 <= h <= 2048): raise ValueError( - f"Custom size out of range: {w}x{h}. " - "Both width and height must be between 512 and 2048 pixels." + f"Custom size out of range: {w}x{h}. " "Both width and height must be between 512 and 2048 pixels." ) payload = Text2ImageTaskCreationRequest( @@ -353,20 +316,12 @@ class ByteDanceImageNode(IO.ComfyNode): guidance_scale=guidance_scale, watermark=watermark, ) - auth_kwargs = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } - response = await SynchronousOperation( - endpoint=ApiEndpoint( - path=BYTEPLUS_IMAGE_ENDPOINT, - method=HttpMethod.POST, - request_model=Text2ImageTaskCreationRequest, - response_model=ImageTaskCreationResponse, - ), - request=payload, - auth_kwargs=auth_kwargs, - ).execute() + response = await sync_op( + cls, + ApiEndpoint(path=BYTEPLUS_IMAGE_ENDPOINT, method="POST"), + data=payload, + response_model=ImageTaskCreationResponse, + ) return IO.NodeOutput(await download_url_to_image_tensor(get_image_url_from_response(response))) @@ -420,7 +375,7 @@ class ByteDanceImageEditNode(IO.ComfyNode): IO.Boolean.Input( "watermark", default=True, - tooltip="Whether to add an \"AI generated\" watermark to the image", + tooltip='Whether to add an "AI generated" watermark to the image', optional=True, ), ], @@ -449,16 +404,7 @@ class ByteDanceImageEditNode(IO.ComfyNode): if get_number_of_images(image) != 1: raise ValueError("Exactly one input image is required.") validate_image_aspect_ratio_range(image, (1, 3), (3, 1)) - auth_kwargs = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } - source_url = (await upload_images_to_comfyapi( - image, - max_images=1, - mime_type="image/png", - auth_kwargs=auth_kwargs, - ))[0] + source_url = (await upload_images_to_comfyapi(cls, image, max_images=1, mime_type="image/png"))[0] payload = Image2ImageTaskCreationRequest( model=model, prompt=prompt, @@ -467,16 +413,12 @@ class ByteDanceImageEditNode(IO.ComfyNode): guidance_scale=guidance_scale, watermark=watermark, ) - response = await SynchronousOperation( - endpoint=ApiEndpoint( - path=BYTEPLUS_IMAGE_ENDPOINT, - method=HttpMethod.POST, - request_model=Image2ImageTaskCreationRequest, - response_model=ImageTaskCreationResponse, - ), - request=payload, - auth_kwargs=auth_kwargs, - ).execute() + response = await sync_op( + cls, + ApiEndpoint(path=BYTEPLUS_IMAGE_ENDPOINT, method="POST"), + data=payload, + response_model=ImageTaskCreationResponse, + ) return IO.NodeOutput(await download_url_to_image_tensor(get_image_url_from_response(response))) @@ -504,7 +446,7 @@ class ByteDanceSeedreamNode(IO.ComfyNode): IO.Image.Input( "image", tooltip="Input image(s) for image-to-image generation. " - "List of 1-10 images for single or multi-reference generation.", + "List of 1-10 images for single or multi-reference generation.", optional=True, ), IO.Combo.Input( @@ -534,9 +476,9 @@ class ByteDanceSeedreamNode(IO.ComfyNode): "sequential_image_generation", options=["disabled", "auto"], tooltip="Group image generation mode. " - "'disabled' generates a single image. " - "'auto' lets the model decide whether to generate multiple related images " - "(e.g., story scenes, character variations).", + "'disabled' generates a single image. " + "'auto' lets the model decide whether to generate multiple related images " + "(e.g., story scenes, character variations).", optional=True, ), IO.Int.Input( @@ -547,7 +489,7 @@ class ByteDanceSeedreamNode(IO.ComfyNode): step=1, display_mode=IO.NumberDisplay.number, tooltip="Maximum number of images to generate when sequential_image_generation='auto'. " - "Total images (input + generated) cannot exceed 15.", + "Total images (input + generated) cannot exceed 15.", optional=True, ), IO.Int.Input( @@ -564,7 +506,7 @@ class ByteDanceSeedreamNode(IO.ComfyNode): IO.Boolean.Input( "watermark", default=True, - tooltip="Whether to add an \"AI generated\" watermark to the image.", + tooltip='Whether to add an "AI generated" watermark to the image.', optional=True, ), IO.Boolean.Input( @@ -611,8 +553,7 @@ class ByteDanceSeedreamNode(IO.ComfyNode): w, h = width, height if not (1024 <= w <= 4096) or not (1024 <= h <= 4096): raise ValueError( - f"Custom size out of range: {w}x{h}. " - "Both width and height must be between 1024 and 4096 pixels." + f"Custom size out of range: {w}x{h}. " "Both width and height must be between 1024 and 4096 pixels." ) n_input_images = get_number_of_images(image) if image is not None else 0 if n_input_images > 10: @@ -621,41 +562,31 @@ class ByteDanceSeedreamNode(IO.ComfyNode): raise ValueError( "The maximum number of generated images plus the number of reference images cannot exceed 15." ) - auth_kwargs = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } reference_images_urls = [] if n_input_images: for i in image: validate_image_aspect_ratio_range(i, (1, 3), (3, 1)) - reference_images_urls = (await upload_images_to_comfyapi( + reference_images_urls = await upload_images_to_comfyapi( + cls, image, max_images=n_input_images, mime_type="image/png", - auth_kwargs=auth_kwargs, - )) - payload = Seedream4TaskCreationRequest( - model=model, - prompt=prompt, - image=reference_images_urls, - size=f"{w}x{h}", - seed=seed, - sequential_image_generation=sequential_image_generation, - sequential_image_generation_options=Seedream4Options(max_images=max_images), - watermark=watermark, - ) - response = await SynchronousOperation( - endpoint=ApiEndpoint( - path=BYTEPLUS_IMAGE_ENDPOINT, - method=HttpMethod.POST, - request_model=Seedream4TaskCreationRequest, - response_model=ImageTaskCreationResponse, + ) + response = await sync_op( + cls, + ApiEndpoint(path=BYTEPLUS_IMAGE_ENDPOINT, method="POST"), + response_model=ImageTaskCreationResponse, + data=Seedream4TaskCreationRequest( + model=model, + prompt=prompt, + image=reference_images_urls, + size=f"{w}x{h}", + seed=seed, + sequential_image_generation=sequential_image_generation, + sequential_image_generation_options=Seedream4Options(max_images=max_images), + watermark=watermark, ), - request=payload, - auth_kwargs=auth_kwargs, - ).execute() - + ) if len(response.data) == 1: return IO.NodeOutput(await download_url_to_image_tensor(get_image_url_from_response(response))) urls = [str(d["url"]) for d in response.data if isinstance(d, dict) and "url" in d] @@ -719,13 +650,13 @@ class ByteDanceTextToVideoNode(IO.ComfyNode): "camera_fixed", default=False, tooltip="Specifies whether to fix the camera. The platform appends an instruction " - "to fix the camera to your prompt, but does not guarantee the actual effect.", + "to fix the camera to your prompt, but does not guarantee the actual effect.", optional=True, ), IO.Boolean.Input( "watermark", default=True, - tooltip="Whether to add an \"AI generated\" watermark to the video.", + tooltip='Whether to add an "AI generated" watermark to the video.', optional=True, ), ], @@ -764,19 +695,9 @@ class ByteDanceTextToVideoNode(IO.ComfyNode): f"--camerafixed {str(camera_fixed).lower()} " f"--watermark {str(watermark).lower()}" ) - - auth_kwargs = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } return await process_video_task( - request_model=Text2VideoTaskCreationRequest, - payload=Text2VideoTaskCreationRequest( - model=model, - content=[TaskTextContent(text=prompt)], - ), - auth_kwargs=auth_kwargs, - node_id=cls.hidden.unique_id, + cls, + payload=Text2VideoTaskCreationRequest(model=model, content=[TaskTextContent(text=prompt)]), estimated_duration=max(1, math.ceil(VIDEO_TASKS_EXECUTION_TIME[model][resolution] * (duration / 10.0))), ) @@ -840,13 +761,13 @@ class ByteDanceImageToVideoNode(IO.ComfyNode): "camera_fixed", default=False, tooltip="Specifies whether to fix the camera. The platform appends an instruction " - "to fix the camera to your prompt, but does not guarantee the actual effect.", + "to fix the camera to your prompt, but does not guarantee the actual effect.", optional=True, ), IO.Boolean.Input( "watermark", default=True, - tooltip="Whether to add an \"AI generated\" watermark to the video.", + tooltip='Whether to add an "AI generated" watermark to the video.', optional=True, ), ], @@ -879,13 +800,7 @@ class ByteDanceImageToVideoNode(IO.ComfyNode): validate_image_dimensions(image, min_width=300, min_height=300, max_width=6000, max_height=6000) validate_image_aspect_ratio_range(image, (2, 5), (5, 2), strict=False) # 0.4 to 2.5 - auth_kwargs = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } - - image_url = (await upload_images_to_comfyapi(image, max_images=1, auth_kwargs=auth_kwargs))[0] - + image_url = (await upload_images_to_comfyapi(cls, image, max_images=1))[0] prompt = ( f"{prompt} " f"--resolution {resolution} " @@ -897,13 +812,11 @@ class ByteDanceImageToVideoNode(IO.ComfyNode): ) return await process_video_task( - request_model=Image2VideoTaskCreationRequest, + cls, payload=Image2VideoTaskCreationRequest( model=model, content=[TaskTextContent(text=prompt), TaskImageContent(image_url=TaskImageContentUrl(url=image_url))], ), - auth_kwargs=auth_kwargs, - node_id=cls.hidden.unique_id, estimated_duration=max(1, math.ceil(VIDEO_TASKS_EXECUTION_TIME[model][resolution] * (duration / 10.0))), ) @@ -971,13 +884,13 @@ class ByteDanceFirstLastFrameNode(IO.ComfyNode): "camera_fixed", default=False, tooltip="Specifies whether to fix the camera. The platform appends an instruction " - "to fix the camera to your prompt, but does not guarantee the actual effect.", + "to fix the camera to your prompt, but does not guarantee the actual effect.", optional=True, ), IO.Boolean.Input( "watermark", default=True, - tooltip="Whether to add an \"AI generated\" watermark to the video.", + tooltip='Whether to add an "AI generated" watermark to the video.', optional=True, ), ], @@ -1012,16 +925,11 @@ class ByteDanceFirstLastFrameNode(IO.ComfyNode): validate_image_dimensions(i, min_width=300, min_height=300, max_width=6000, max_height=6000) validate_image_aspect_ratio_range(i, (2, 5), (5, 2), strict=False) # 0.4 to 2.5 - auth_kwargs = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } - download_urls = await upload_images_to_comfyapi( + cls, image_tensor_pair_to_batch(first_frame, last_frame), max_images=2, mime_type="image/png", - auth_kwargs=auth_kwargs, ) prompt = ( @@ -1035,7 +943,7 @@ class ByteDanceFirstLastFrameNode(IO.ComfyNode): ) return await process_video_task( - request_model=Image2VideoTaskCreationRequest, + cls, payload=Image2VideoTaskCreationRequest( model=model, content=[ @@ -1044,8 +952,6 @@ class ByteDanceFirstLastFrameNode(IO.ComfyNode): TaskImageContent(image_url=TaskImageContentUrl(url=str(download_urls[1])), role="last_frame"), ], ), - auth_kwargs=auth_kwargs, - node_id=cls.hidden.unique_id, estimated_duration=max(1, math.ceil(VIDEO_TASKS_EXECUTION_TIME[model][resolution] * (duration / 10.0))), ) @@ -1108,7 +1014,7 @@ class ByteDanceImageReferenceNode(IO.ComfyNode): IO.Boolean.Input( "watermark", default=True, - tooltip="Whether to add an \"AI generated\" watermark to the video.", + tooltip='Whether to add an "AI generated" watermark to the video.', optional=True, ), ], @@ -1141,15 +1047,7 @@ class ByteDanceImageReferenceNode(IO.ComfyNode): validate_image_dimensions(image, min_width=300, min_height=300, max_width=6000, max_height=6000) validate_image_aspect_ratio_range(image, (2, 5), (5, 2), strict=False) # 0.4 to 2.5 - auth_kwargs = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } - - image_urls = await upload_images_to_comfyapi( - images, max_images=4, mime_type="image/png", auth_kwargs=auth_kwargs - ) - + image_urls = await upload_images_to_comfyapi(cls, images, max_images=4, mime_type="image/png") prompt = ( f"{prompt} " f"--resolution {resolution} " @@ -1160,42 +1058,32 @@ class ByteDanceImageReferenceNode(IO.ComfyNode): ) x = [ TaskTextContent(text=prompt), - *[TaskImageContent(image_url=TaskImageContentUrl(url=str(i)), role="reference_image") for i in image_urls] + *[TaskImageContent(image_url=TaskImageContentUrl(url=str(i)), role="reference_image") for i in image_urls], ] return await process_video_task( - request_model=Image2VideoTaskCreationRequest, - payload=Image2VideoTaskCreationRequest( - model=model, - content=x, - ), - auth_kwargs=auth_kwargs, - node_id=cls.hidden.unique_id, + cls, + payload=Image2VideoTaskCreationRequest(model=model, content=x), estimated_duration=max(1, math.ceil(VIDEO_TASKS_EXECUTION_TIME[model][resolution] * (duration / 10.0))), ) async def process_video_task( - request_model: Type[T], + cls: type[IO.ComfyNode], payload: Union[Text2VideoTaskCreationRequest, Image2VideoTaskCreationRequest], - auth_kwargs: dict, - node_id: str, estimated_duration: Optional[int], ) -> IO.NodeOutput: - initial_response = await SynchronousOperation( - endpoint=ApiEndpoint( - path=BYTEPLUS_TASK_ENDPOINT, - method=HttpMethod.POST, - request_model=request_model, - response_model=TaskCreationResponse, - ), - request=payload, - auth_kwargs=auth_kwargs, - ).execute() - response = await poll_until_finished( - auth_kwargs, - initial_response.id, + initial_response = await sync_op( + cls, + ApiEndpoint(path=BYTEPLUS_TASK_ENDPOINT, method="POST"), + data=payload, + response_model=TaskCreationResponse, + ) + response = await poll_op( + cls, + ApiEndpoint(path=f"{BYTEPLUS_TASK_STATUS_ENDPOINT}/{initial_response.id}"), + status_extractor=lambda r: r.status, estimated_duration=estimated_duration, - node_id=node_id, + response_model=TaskStatusResponse, ) return IO.NodeOutput(await download_url_to_video_output(get_video_url_from_task_status(response))) @@ -1221,5 +1109,6 @@ class ByteDanceExtension(ComfyExtension): ByteDanceImageReferenceNode, ] + async def comfy_entrypoint() -> ByteDanceExtension: return ByteDanceExtension() diff --git a/comfy_api_nodes/nodes_gemini.py b/comfy_api_nodes/nodes_gemini.py index c1941cbe9..ca11b67ed 100644 --- a/comfy_api_nodes/nodes_gemini.py +++ b/comfy_api_nodes/nodes_gemini.py @@ -33,12 +33,9 @@ from comfy_api_nodes.apis.client import ( SynchronousOperation, ) from comfy_api_nodes.apinode_utils import ( - validate_string, - audio_to_base64_string, video_to_base64_string, - tensor_to_base64_string, - bytesio_to_image_tensor, ) +from comfy_api_nodes.util import validate_string, tensor_to_base64_string, bytesio_to_image_tensor, audio_to_base64_string from comfy_api.util import VideoContainer, VideoCodec diff --git a/comfy_api_nodes/nodes_kling.py b/comfy_api_nodes/nodes_kling.py index 67c8307c5..eea65c9ac 100644 --- a/comfy_api_nodes/nodes_kling.py +++ b/comfy_api_nodes/nodes_kling.py @@ -5,8 +5,7 @@ For source of truth on the allowed permutations of request fields, please refere """ from __future__ import annotations -from typing import Optional, TypeVar, Any -from collections.abc import Callable +from typing import Optional, TypeVar import math import logging @@ -15,7 +14,6 @@ from typing_extensions import override import torch from comfy_api_nodes.apis import ( - KlingTaskStatus, KlingCameraControl, KlingCameraConfig, KlingCameraControlType, @@ -52,26 +50,20 @@ from comfy_api_nodes.apis import ( KlingCharacterEffectModelName, KlingSingleImageEffectModelName, ) -from comfy_api_nodes.apis.client import ( - ApiEndpoint, - HttpMethod, - SynchronousOperation, - PollingOperation, - EmptyRequest, -) -from comfy_api_nodes.apinode_utils import ( - tensor_to_base64_string, - download_url_to_video_output, - upload_video_to_comfyapi, - upload_audio_to_comfyapi, - download_url_to_image_tensor, - validate_string, -) -from comfy_api_nodes.util.validation_utils import ( +from comfy_api_nodes.util import ( validate_image_dimensions, validate_image_aspect_ratio, validate_video_dimensions, validate_video_duration, + tensor_to_base64_string, + validate_string, + upload_audio_to_comfyapi, + download_url_to_image_tensor, + upload_video_to_comfyapi, + download_url_to_video_output, + sync_op, + ApiEndpoint, + poll_op, ) from comfy_api.input_impl import VideoFromFile from comfy_api.input.basic_types import AudioInput @@ -214,34 +206,6 @@ VOICES_CONFIG = { } -async def poll_until_finished( - auth_kwargs: dict[str, str], - api_endpoint: ApiEndpoint[Any, R], - result_url_extractor: Optional[Callable[[R], str]] = None, - estimated_duration: Optional[int] = None, - node_id: Optional[str] = None, -) -> R: - """Polls the Kling API endpoint until the task reaches a terminal state, then returns the response.""" - return await PollingOperation( - poll_endpoint=api_endpoint, - completed_statuses=[ - KlingTaskStatus.succeed.value, - ], - failed_statuses=[KlingTaskStatus.failed.value], - status_extractor=lambda response: ( - response.data.task_status.value - if response.data and response.data.task_status - else None - ), - auth_kwargs=auth_kwargs, - result_url_extractor=result_url_extractor, - estimated_duration=estimated_duration, - node_id=node_id, - poll_interval=16.0, - max_poll_attempts=256, - ).execute() - - def is_valid_camera_control_configs(configs: list[float]) -> bool: """Verifies that at least one camera control configuration is non-zero.""" return any(not math.isclose(value, 0.0) for value in configs) @@ -377,8 +341,7 @@ async def image_result_to_node_output( async def execute_text2video( - auth_kwargs: dict[str, str], - node_id: str, + cls: type[IO.ComfyNode], prompt: str, negative_prompt: str, cfg_scale: float, @@ -389,14 +352,11 @@ async def execute_text2video( camera_control: Optional[KlingCameraControl] = None, ) -> IO.NodeOutput: validate_prompts(prompt, negative_prompt, MAX_PROMPT_LENGTH_T2V) - initial_operation = SynchronousOperation( - endpoint=ApiEndpoint( - path=PATH_TEXT_TO_VIDEO, - method=HttpMethod.POST, - request_model=KlingText2VideoRequest, - response_model=KlingText2VideoResponse, - ), - request=KlingText2VideoRequest( + task_creation_response = await sync_op( + cls, + ApiEndpoint(path=PATH_TEXT_TO_VIDEO, method="POST"), + response_model=KlingText2VideoResponse, + data=KlingText2VideoRequest( prompt=prompt if prompt else None, negative_prompt=negative_prompt if negative_prompt else None, duration=KlingVideoGenDuration(duration), @@ -406,24 +366,17 @@ async def execute_text2video( aspect_ratio=KlingVideoGenAspectRatio(aspect_ratio), camera_control=camera_control, ), - auth_kwargs=auth_kwargs, ) - task_creation_response = await initial_operation.execute() validate_task_creation_response(task_creation_response) task_id = task_creation_response.data.task_id - final_response = await poll_until_finished( - auth_kwargs, - ApiEndpoint( - path=f"{PATH_TEXT_TO_VIDEO}/{task_id}", - method=HttpMethod.GET, - request_model=EmptyRequest, - response_model=KlingText2VideoResponse, - ), - result_url_extractor=get_video_url_from_response, + final_response = await poll_op( + cls, + ApiEndpoint(path=f"{PATH_TEXT_TO_VIDEO}/{task_id}"), + response_model=KlingText2VideoResponse, estimated_duration=AVERAGE_DURATION_T2V, - node_id=node_id, + status_extractor=lambda r: (r.data.task_status.value if r.data and r.data.task_status else None), ) validate_video_result_response(final_response) @@ -432,8 +385,7 @@ async def execute_text2video( async def execute_image2video( - auth_kwargs: dict[str, str], - node_id: str, + cls: type[IO.ComfyNode], start_frame: torch.Tensor, prompt: str, negative_prompt: str, @@ -455,14 +407,11 @@ async def execute_image2video( if model_mode == "std" and model_name == KlingVideoGenModelName.kling_v2_5_turbo.value: model_mode = "pro" # October 5: currently "std" mode is not supported for this model - initial_operation = SynchronousOperation( - endpoint=ApiEndpoint( - path=PATH_IMAGE_TO_VIDEO, - method=HttpMethod.POST, - request_model=KlingImage2VideoRequest, - response_model=KlingImage2VideoResponse, - ), - request=KlingImage2VideoRequest( + task_creation_response = await sync_op( + cls, + ApiEndpoint(path=PATH_IMAGE_TO_VIDEO, method="POST"), + response_model=KlingImage2VideoResponse, + data=KlingImage2VideoRequest( model_name=KlingVideoGenModelName(model_name), image=tensor_to_base64_string(start_frame), image_tail=( @@ -477,24 +426,17 @@ async def execute_image2video( duration=KlingVideoGenDuration(duration), camera_control=camera_control, ), - auth_kwargs=auth_kwargs, ) - task_creation_response = await initial_operation.execute() validate_task_creation_response(task_creation_response) task_id = task_creation_response.data.task_id - final_response = await poll_until_finished( - auth_kwargs, - ApiEndpoint( - path=f"{PATH_IMAGE_TO_VIDEO}/{task_id}", - method=HttpMethod.GET, - request_model=KlingImage2VideoRequest, - response_model=KlingImage2VideoResponse, - ), - result_url_extractor=get_video_url_from_response, + final_response = await poll_op( + cls, + ApiEndpoint(path=f"{PATH_IMAGE_TO_VIDEO}/{task_id}"), + response_model=KlingImage2VideoResponse, estimated_duration=AVERAGE_DURATION_I2V, - node_id=node_id, + status_extractor=lambda r: (r.data.task_status.value if r.data and r.data.task_status else None), ) validate_video_result_response(final_response) @@ -503,8 +445,7 @@ async def execute_image2video( async def execute_video_effect( - auth_kwargs: dict[str, str], - node_id: str, + cls: type[IO.ComfyNode], dual_character: bool, effect_scene: KlingDualCharacterEffectsScene | KlingSingleImageEffectsScene, model_name: str, @@ -530,35 +471,25 @@ async def execute_video_effect( duration=duration, ) - initial_operation = SynchronousOperation( - endpoint=ApiEndpoint( - path=PATH_VIDEO_EFFECTS, - method=HttpMethod.POST, - request_model=KlingVideoEffectsRequest, - response_model=KlingVideoEffectsResponse, - ), - request=KlingVideoEffectsRequest( + task_creation_response = await sync_op( + cls, + endpoint=ApiEndpoint(path=PATH_VIDEO_EFFECTS, method="POST"), + response_model=KlingVideoEffectsResponse, + data=KlingVideoEffectsRequest( effect_scene=effect_scene, input=request_input_field, ), - auth_kwargs=auth_kwargs, ) - task_creation_response = await initial_operation.execute() validate_task_creation_response(task_creation_response) task_id = task_creation_response.data.task_id - final_response = await poll_until_finished( - auth_kwargs, - ApiEndpoint( - path=f"{PATH_VIDEO_EFFECTS}/{task_id}", - method=HttpMethod.GET, - request_model=EmptyRequest, - response_model=KlingVideoEffectsResponse, - ), - result_url_extractor=get_video_url_from_response, + final_response = await poll_op( + cls, + ApiEndpoint(path=f"{PATH_VIDEO_EFFECTS}/{task_id}"), + response_model=KlingVideoEffectsResponse, estimated_duration=AVERAGE_DURATION_VIDEO_EFFECTS, - node_id=node_id, + status_extractor=lambda r: (r.data.task_status.value if r.data and r.data.task_status else None), ) validate_video_result_response(final_response) @@ -567,8 +498,7 @@ async def execute_video_effect( async def execute_lipsync( - auth_kwargs: dict[str, str], - node_id: str, + cls: type[IO.ComfyNode], video: VideoInput, audio: Optional[AudioInput] = None, voice_language: Optional[str] = None, @@ -583,24 +513,21 @@ async def execute_lipsync( validate_video_duration(video, 2, 10) # Upload video to Comfy API and get download URL - video_url = await upload_video_to_comfyapi(video, auth_kwargs=auth_kwargs) + video_url = await upload_video_to_comfyapi(cls, video) logging.info("Uploaded video to Comfy API. URL: %s", video_url) # Upload the audio file to Comfy API and get download URL if audio: - audio_url = await upload_audio_to_comfyapi(audio, auth_kwargs=auth_kwargs) + audio_url = await upload_audio_to_comfyapi(cls, audio) logging.info("Uploaded audio to Comfy API. URL: %s", audio_url) else: audio_url = None - initial_operation = SynchronousOperation( - endpoint=ApiEndpoint( - path=PATH_LIP_SYNC, - method=HttpMethod.POST, - request_model=KlingLipSyncRequest, - response_model=KlingLipSyncResponse, - ), - request=KlingLipSyncRequest( + task_creation_response = await sync_op( + cls, + ApiEndpoint(PATH_LIP_SYNC, "POST"), + response_model=KlingLipSyncResponse, + data=KlingLipSyncRequest( input=KlingLipSyncInputObject( video_url=video_url, mode=model_mode, @@ -612,24 +539,17 @@ async def execute_lipsync( voice_id=voice_id, ), ), - auth_kwargs=auth_kwargs, ) - task_creation_response = await initial_operation.execute() validate_task_creation_response(task_creation_response) task_id = task_creation_response.data.task_id - final_response = await poll_until_finished( - auth_kwargs, - ApiEndpoint( - path=f"{PATH_LIP_SYNC}/{task_id}", - method=HttpMethod.GET, - request_model=EmptyRequest, - response_model=KlingLipSyncResponse, - ), - result_url_extractor=get_video_url_from_response, + final_response = await poll_op( + cls, + ApiEndpoint(path=f"{PATH_LIP_SYNC}/{task_id}"), + response_model=KlingLipSyncResponse, estimated_duration=AVERAGE_DURATION_LIP_SYNC, - node_id=node_id, + status_extractor=lambda r: (r.data.task_status.value if r.data and r.data.task_status else None), ) validate_video_result_response(final_response) @@ -807,11 +727,7 @@ class KlingTextToVideoNode(IO.ComfyNode): ) -> IO.NodeOutput: model_mode, duration, model_name = MODE_TEXT2VIDEO[mode] return await execute_text2video( - auth_kwargs={ - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - }, - node_id=cls.hidden.unique_id, + cls, prompt=prompt, negative_prompt=negative_prompt, cfg_scale=cfg_scale, @@ -872,11 +788,7 @@ class KlingCameraControlT2VNode(IO.ComfyNode): camera_control: Optional[KlingCameraControl] = None, ) -> IO.NodeOutput: return await execute_text2video( - auth_kwargs={ - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - }, - node_id=cls.hidden.unique_id, + cls, model_name=KlingVideoGenModelName.kling_v1, cfg_scale=cfg_scale, model_mode=KlingVideoGenMode.std, @@ -944,11 +856,7 @@ class KlingImage2VideoNode(IO.ComfyNode): end_frame: Optional[torch.Tensor] = None, ) -> IO.NodeOutput: return await execute_image2video( - auth_kwargs={ - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - }, - node_id=cls.hidden.unique_id, + cls, start_frame=start_frame, prompt=prompt, negative_prompt=negative_prompt, @@ -1017,11 +925,7 @@ class KlingCameraControlI2VNode(IO.ComfyNode): camera_control: KlingCameraControl, ) -> IO.NodeOutput: return await execute_image2video( - auth_kwargs={ - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - }, - node_id=cls.hidden.unique_id, + cls, model_name=KlingVideoGenModelName.kling_v1_5, start_frame=start_frame, cfg_scale=cfg_scale, @@ -1097,11 +1001,7 @@ class KlingStartEndFrameNode(IO.ComfyNode): ) -> IO.NodeOutput: mode, duration, model_name = MODE_START_END_FRAME[mode] return await execute_image2video( - auth_kwargs={ - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - }, - node_id=cls.hidden.unique_id, + cls, prompt=prompt, negative_prompt=negative_prompt, model_name=model_name, @@ -1162,41 +1062,27 @@ class KlingVideoExtendNode(IO.ComfyNode): video_id: str, ) -> IO.NodeOutput: validate_prompts(prompt, negative_prompt, MAX_PROMPT_LENGTH_T2V) - auth = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } - initial_operation = SynchronousOperation( - endpoint=ApiEndpoint( - path=PATH_VIDEO_EXTEND, - method=HttpMethod.POST, - request_model=KlingVideoExtendRequest, - response_model=KlingVideoExtendResponse, - ), - request=KlingVideoExtendRequest( + task_creation_response = await sync_op( + cls, + ApiEndpoint(path=PATH_VIDEO_EXTEND, method="POST"), + response_model=KlingVideoExtendResponse, + data=KlingVideoExtendRequest( prompt=prompt if prompt else None, negative_prompt=negative_prompt if negative_prompt else None, cfg_scale=cfg_scale, video_id=video_id, ), - auth_kwargs=auth, ) - task_creation_response = await initial_operation.execute() validate_task_creation_response(task_creation_response) task_id = task_creation_response.data.task_id - final_response = await poll_until_finished( - auth, - ApiEndpoint( - path=f"{PATH_VIDEO_EXTEND}/{task_id}", - method=HttpMethod.GET, - request_model=EmptyRequest, - response_model=KlingVideoExtendResponse, - ), - result_url_extractor=get_video_url_from_response, + final_response = await poll_op( + cls, + ApiEndpoint(path=f"{PATH_VIDEO_EXTEND}/{task_id}"), + response_model=KlingVideoExtendResponse, estimated_duration=AVERAGE_DURATION_VIDEO_EXTEND, - node_id=cls.hidden.unique_id, + status_extractor=lambda r: (r.data.task_status.value if r.data and r.data.task_status else None), ) validate_video_result_response(final_response) @@ -1259,11 +1145,7 @@ class KlingDualCharacterVideoEffectNode(IO.ComfyNode): duration: KlingVideoGenDuration, ) -> IO.NodeOutput: video, _, duration = await execute_video_effect( - auth_kwargs={ - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - }, - node_id=cls.hidden.unique_id, + cls, dual_character=True, effect_scene=effect_scene, model_name=model_name, @@ -1324,11 +1206,7 @@ class KlingSingleImageVideoEffectNode(IO.ComfyNode): return IO.NodeOutput( *( await execute_video_effect( - auth_kwargs={ - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - }, - node_id=cls.hidden.unique_id, + cls, dual_character=False, effect_scene=effect_scene, model_name=model_name, @@ -1379,11 +1257,7 @@ class KlingLipSyncAudioToVideoNode(IO.ComfyNode): voice_language: str, ) -> IO.NodeOutput: return await execute_lipsync( - auth_kwargs={ - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - }, - node_id=cls.hidden.unique_id, + cls, video=video, audio=audio, voice_language=voice_language, @@ -1445,11 +1319,7 @@ class KlingLipSyncTextToVideoNode(IO.ComfyNode): ) -> IO.NodeOutput: voice_id, voice_language = VOICES_CONFIG[voice] return await execute_lipsync( - auth_kwargs={ - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - }, - node_id=cls.hidden.unique_id, + cls, video=video, text=text, voice_language=voice_language, @@ -1496,40 +1366,26 @@ class KlingVirtualTryOnNode(IO.ComfyNode): cloth_image: torch.Tensor, model_name: KlingVirtualTryOnModelName, ) -> IO.NodeOutput: - auth = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } - initial_operation = SynchronousOperation( - endpoint=ApiEndpoint( - path=PATH_VIRTUAL_TRY_ON, - method=HttpMethod.POST, - request_model=KlingVirtualTryOnRequest, - response_model=KlingVirtualTryOnResponse, - ), - request=KlingVirtualTryOnRequest( + task_creation_response = await sync_op( + cls, + ApiEndpoint(path=PATH_VIRTUAL_TRY_ON, method="POST"), + response_model=KlingVirtualTryOnResponse, + data=KlingVirtualTryOnRequest( human_image=tensor_to_base64_string(human_image), cloth_image=tensor_to_base64_string(cloth_image), model_name=model_name, ), - auth_kwargs=auth, ) - task_creation_response = await initial_operation.execute() validate_task_creation_response(task_creation_response) task_id = task_creation_response.data.task_id - final_response = await poll_until_finished( - auth, - ApiEndpoint( - path=f"{PATH_VIRTUAL_TRY_ON}/{task_id}", - method=HttpMethod.GET, - request_model=EmptyRequest, - response_model=KlingVirtualTryOnResponse, - ), - result_url_extractor=get_images_urls_from_response, + final_response = await poll_op( + cls, + ApiEndpoint(path=f"{PATH_VIRTUAL_TRY_ON}/{task_id}"), + response_model=KlingVirtualTryOnResponse, estimated_duration=AVERAGE_DURATION_VIRTUAL_TRY_ON, - node_id=cls.hidden.unique_id, + status_extractor=lambda r: (r.data.task_status.value if r.data and r.data.task_status else None), ) validate_image_result_response(final_response) @@ -1625,18 +1481,11 @@ class KlingImageGenerationNode(IO.ComfyNode): else: image = tensor_to_base64_string(image) - auth = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } - initial_operation = SynchronousOperation( - endpoint=ApiEndpoint( - path=PATH_IMAGE_GENERATIONS, - method=HttpMethod.POST, - request_model=KlingImageGenerationsRequest, - response_model=KlingImageGenerationsResponse, - ), - request=KlingImageGenerationsRequest( + task_creation_response = await sync_op( + cls, + ApiEndpoint(path=PATH_IMAGE_GENERATIONS, method="POST"), + response_model=KlingImageGenerationsResponse, + data=KlingImageGenerationsRequest( model_name=model_name, prompt=prompt, negative_prompt=negative_prompt, @@ -1647,24 +1496,17 @@ class KlingImageGenerationNode(IO.ComfyNode): n=n, aspect_ratio=aspect_ratio, ), - auth_kwargs=auth, ) - task_creation_response = await initial_operation.execute() validate_task_creation_response(task_creation_response) task_id = task_creation_response.data.task_id - final_response = await poll_until_finished( - auth, - ApiEndpoint( - path=f"{PATH_IMAGE_GENERATIONS}/{task_id}", - method=HttpMethod.GET, - request_model=EmptyRequest, - response_model=KlingImageGenerationsResponse, - ), - result_url_extractor=get_images_urls_from_response, + final_response = await poll_op( + cls, + ApiEndpoint(path=f"{PATH_IMAGE_GENERATIONS}/{task_id}"), + response_model=KlingImageGenerationsResponse, estimated_duration=AVERAGE_DURATION_IMAGE_GEN, - node_id=cls.hidden.unique_id, + status_extractor=lambda r: (r.data.task_status.value if r.data and r.data.task_status else None), ) validate_image_result_response(final_response) diff --git a/comfy_api_nodes/nodes_luma.py b/comfy_api_nodes/nodes_luma.py index 610d95a77..e74441e5e 100644 --- a/comfy_api_nodes/nodes_luma.py +++ b/comfy_api_nodes/nodes_luma.py @@ -35,9 +35,9 @@ from comfy_api_nodes.apis.client import ( from comfy_api_nodes.apinode_utils import ( upload_images_to_comfyapi, process_image_response, - validate_string, ) from server import PromptServer +from comfy_api_nodes.util import validate_string import aiohttp import torch diff --git a/comfy_api_nodes/nodes_minimax.py b/comfy_api_nodes/nodes_minimax.py index 23be1ae65..e3722e79b 100644 --- a/comfy_api_nodes/nodes_minimax.py +++ b/comfy_api_nodes/nodes_minimax.py @@ -24,8 +24,8 @@ from comfy_api_nodes.apis.client import ( from comfy_api_nodes.apinode_utils import ( download_url_to_bytesio, upload_images_to_comfyapi, - validate_string, ) +from comfy_api_nodes.util import validate_string from server import PromptServer diff --git a/comfy_api_nodes/nodes_moonvalley.py b/comfy_api_nodes/nodes_moonvalley.py index 7566188dd..7c31d95b3 100644 --- a/comfy_api_nodes/nodes_moonvalley.py +++ b/comfy_api_nodes/nodes_moonvalley.py @@ -1,35 +1,31 @@ import logging -from typing import Any, Callable, Optional, TypeVar +from typing import Optional + import torch from typing_extensions import override -from comfy_api_nodes.util.validation_utils import validate_image_dimensions +from comfy_api.input import VideoInput +from comfy_api.latest import IO, ComfyExtension from comfy_api_nodes.apis import ( - MoonvalleyTextToVideoRequest, + MoonvalleyPromptResponse, MoonvalleyTextToVideoInferenceParams, + MoonvalleyTextToVideoRequest, MoonvalleyVideoToVideoInferenceParams, MoonvalleyVideoToVideoRequest, - MoonvalleyPromptResponse, ) -from comfy_api_nodes.apis.client import ( +from comfy_api_nodes.util import ( ApiEndpoint, - HttpMethod, - SynchronousOperation, - PollingOperation, - EmptyRequest, -) -from comfy_api_nodes.apinode_utils import ( download_url_to_video_output, + poll_op, + sync_op, + trim_video, upload_images_to_comfyapi, upload_video_to_comfyapi, validate_container_format_is_mp4, + validate_image_dimensions, + validate_string, ) -from comfy_api.input import VideoInput -from comfy_api.latest import ComfyExtension, InputImpl, IO -import av -import io - API_UPLOADS_ENDPOINT = "/proxy/moonvalley/uploads" API_PROMPTS_ENDPOINT = "/proxy/moonvalley/prompts" API_VIDEO2VIDEO_ENDPOINT = "/proxy/moonvalley/prompts/video-to-video" @@ -51,13 +47,6 @@ MAX_VID_HEIGHT = 10000 MAX_VIDEO_SIZE = 1024 * 1024 * 1024 # 1 GB max for in-memory video processing MOONVALLEY_MAREY_MAX_PROMPT_LENGTH = 5000 -R = TypeVar("R") - - -class MoonvalleyApiError(Exception): - """Base exception for Moonvalley API errors.""" - - pass def is_valid_task_creation_response(response: MoonvalleyPromptResponse) -> bool: @@ -69,64 +58,7 @@ def validate_task_creation_response(response) -> None: if not is_valid_task_creation_response(response): error_msg = f"Moonvalley Marey API: Initial request failed. Code: {response.code}, Message: {response.message}, Data: {response}" logging.error(error_msg) - raise MoonvalleyApiError(error_msg) - - -def get_video_from_response(response): - video = response.output_url - logging.info( - "Moonvalley Marey API: Task %s succeeded. Video URL: %s", response.id, video - ) - return video - - -def get_video_url_from_response(response) -> Optional[str]: - """Returns the first video url from the Moonvalley video generation task result. - Will not raise an error if the response is not valid. - """ - if response: - return str(get_video_from_response(response)) - else: - return None - - -async def poll_until_finished( - auth_kwargs: dict[str, str], - api_endpoint: ApiEndpoint[Any, R], - result_url_extractor: Optional[Callable[[R], str]] = None, - node_id: Optional[str] = None, -) -> R: - """Polls the Moonvalley API endpoint until the task reaches a terminal state, then returns the response.""" - return await PollingOperation( - poll_endpoint=api_endpoint, - completed_statuses=[ - "completed", - ], - max_poll_attempts=240, # 64 minutes with 16s interval - poll_interval=16.0, - failed_statuses=["error"], - status_extractor=lambda response: ( - response.status if response and response.status else None - ), - auth_kwargs=auth_kwargs, - result_url_extractor=result_url_extractor, - node_id=node_id, - ).execute() - - -def validate_prompts( - prompt: str, negative_prompt: str, max_length=MOONVALLEY_MAREY_MAX_PROMPT_LENGTH -): - """Verifies that the prompt isn't empty and that neither prompt is too long.""" - if not prompt: - raise ValueError("Positive prompt is empty") - if len(prompt) > max_length: - raise ValueError(f"Positive prompt is too long: {len(prompt)} characters") - if negative_prompt and len(negative_prompt) > max_length: - raise ValueError( - f"Negative prompt is too long: {len(negative_prompt)} characters" - ) - return True + raise RuntimeError(error_msg) def validate_video_to_video_input(video: VideoInput) -> VideoInput: @@ -170,12 +102,8 @@ def _validate_video_dimensions(width: int, height: int) -> None: } if (width, height) not in supported_resolutions: - supported_list = ", ".join( - [f"{w}x{h}" for w, h in sorted(supported_resolutions)] - ) - raise ValueError( - f"Resolution {width}x{height} not supported. Supported: {supported_list}" - ) + supported_list = ", ".join([f"{w}x{h}" for w, h in sorted(supported_resolutions)]) + raise ValueError(f"Resolution {width}x{height} not supported. Supported: {supported_list}") def _validate_and_trim_duration(video: VideoInput) -> VideoInput: @@ -188,7 +116,7 @@ def _validate_and_trim_duration(video: VideoInput) -> VideoInput: def _validate_minimum_duration(duration: float) -> None: """Ensures video is at least 5 seconds long.""" if duration < 5: - raise MoonvalleyApiError("Input video must be at least 5 seconds long.") + raise ValueError("Input video must be at least 5 seconds long.") def _trim_if_too_long(video: VideoInput, duration: float) -> VideoInput: @@ -198,123 +126,6 @@ def _trim_if_too_long(video: VideoInput, duration: float) -> VideoInput: return video -def trim_video(video: VideoInput, duration_sec: float) -> VideoInput: - """ - Returns a new VideoInput object trimmed from the beginning to the specified duration, - using av to avoid loading entire video into memory. - - Args: - video: Input video to trim - duration_sec: Duration in seconds to keep from the beginning - - Returns: - VideoFromFile object that owns the output buffer - """ - output_buffer = io.BytesIO() - - input_container = None - output_container = None - - try: - # Get the stream source - this avoids loading entire video into memory - # when the source is already a file path - input_source = video.get_stream_source() - - # Open containers - input_container = av.open(input_source, mode="r") - output_container = av.open(output_buffer, mode="w", format="mp4") - - # Set up output streams for re-encoding - video_stream = None - audio_stream = None - - for stream in input_container.streams: - logging.info("Found stream: type=%s, class=%s", stream.type, type(stream)) - if isinstance(stream, av.VideoStream): - # Create output video stream with same parameters - video_stream = output_container.add_stream( - "h264", rate=stream.average_rate - ) - video_stream.width = stream.width - video_stream.height = stream.height - video_stream.pix_fmt = "yuv420p" - logging.info( - "Added video stream: %sx%s @ %sfps", stream.width, stream.height, stream.average_rate - ) - elif isinstance(stream, av.AudioStream): - # Create output audio stream with same parameters - audio_stream = output_container.add_stream( - "aac", rate=stream.sample_rate - ) - audio_stream.sample_rate = stream.sample_rate - audio_stream.layout = stream.layout - logging.info("Added audio stream: %sHz, %s channels", stream.sample_rate, stream.channels) - - # Calculate target frame count that's divisible by 16 - fps = input_container.streams.video[0].average_rate - estimated_frames = int(duration_sec * fps) - target_frames = ( - estimated_frames // 16 - ) * 16 # Round down to nearest multiple of 16 - - if target_frames == 0: - raise ValueError("Video too short: need at least 16 frames for Moonvalley") - - frame_count = 0 - audio_frame_count = 0 - - # Decode and re-encode video frames - if video_stream: - for frame in input_container.decode(video=0): - if frame_count >= target_frames: - break - - # Re-encode frame - for packet in video_stream.encode(frame): - output_container.mux(packet) - frame_count += 1 - - # Flush encoder - for packet in video_stream.encode(): - output_container.mux(packet) - - logging.info("Encoded %s video frames (target: %s)", frame_count, target_frames) - - # Decode and re-encode audio frames - if audio_stream: - input_container.seek(0) # Reset to beginning for audio - for frame in input_container.decode(audio=0): - if frame.time >= duration_sec: - break - - # Re-encode frame - for packet in audio_stream.encode(frame): - output_container.mux(packet) - audio_frame_count += 1 - - # Flush encoder - for packet in audio_stream.encode(): - output_container.mux(packet) - - logging.info("Encoded %s audio frames", audio_frame_count) - - # Close containers - output_container.close() - input_container.close() - - # Return as VideoFromFile using the buffer - output_buffer.seek(0) - return InputImpl.VideoFromFile(output_buffer) - - except Exception as e: - # Clean up on error - if input_container is not None: - input_container.close() - if output_container is not None: - output_container.close() - raise RuntimeError(f"Failed to trim video: {str(e)}") from e - - def parse_width_height_from_res(resolution: str): # Accepts a string like "16:9 (1920 x 1080)" and returns width, height as a dict res_map = { @@ -338,19 +149,14 @@ def parse_control_parameter(value): return control_map.get(value, control_map["Motion Transfer"]) -async def get_response( - task_id: str, auth_kwargs: dict[str, str], node_id: Optional[str] = None -) -> MoonvalleyPromptResponse: - return await poll_until_finished( - auth_kwargs, - ApiEndpoint( - path=f"{API_PROMPTS_ENDPOINT}/{task_id}", - method=HttpMethod.GET, - request_model=EmptyRequest, - response_model=MoonvalleyPromptResponse, - ), - result_url_extractor=get_video_url_from_response, - node_id=node_id, +async def get_response(cls: type[IO.ComfyNode], task_id: str) -> MoonvalleyPromptResponse: + return await poll_op( + cls, + ApiEndpoint(path=f"{API_PROMPTS_ENDPOINT}/{task_id}"), + response_model=MoonvalleyPromptResponse, + status_extractor=lambda r: (r.status if r and r.status else None), + poll_interval=16.0, + max_poll_attempts=240, ) @@ -444,14 +250,10 @@ class MoonvalleyImg2VideoNode(IO.ComfyNode): steps: int, ) -> IO.NodeOutput: validate_image_dimensions(image, min_width=300, min_height=300, max_height=MAX_HEIGHT, max_width=MAX_WIDTH) - validate_prompts(prompt, negative_prompt, MOONVALLEY_MAREY_MAX_PROMPT_LENGTH) + validate_string(prompt, min_length=1, max_length=MOONVALLEY_MAREY_MAX_PROMPT_LENGTH) + validate_string(negative_prompt, field_name="negative_prompt", max_length=MOONVALLEY_MAREY_MAX_PROMPT_LENGTH) width_height = parse_width_height_from_res(resolution) - auth = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } - inference_params = MoonvalleyTextToVideoInferenceParams( negative_prompt=negative_prompt, steps=steps, @@ -464,33 +266,17 @@ class MoonvalleyImg2VideoNode(IO.ComfyNode): # Get MIME type from tensor - assuming PNG format for image tensors mime_type = "image/png" - - image_url = ( - await upload_images_to_comfyapi( - image, max_images=1, auth_kwargs=auth, mime_type=mime_type - ) - )[0] - - request = MoonvalleyTextToVideoRequest( - image_url=image_url, prompt_text=prompt, inference_params=inference_params - ) - initial_operation = SynchronousOperation( - endpoint=ApiEndpoint( - path=API_IMG2VIDEO_ENDPOINT, - method=HttpMethod.POST, - request_model=MoonvalleyTextToVideoRequest, - response_model=MoonvalleyPromptResponse, + image_url = (await upload_images_to_comfyapi(cls, image, max_images=1, mime_type=mime_type))[0] + task_creation_response = await sync_op( + cls, + endpoint=ApiEndpoint(path=API_IMG2VIDEO_ENDPOINT, method="POST"), + response_model=MoonvalleyPromptResponse, + data=MoonvalleyTextToVideoRequest( + image_url=image_url, prompt_text=prompt, inference_params=inference_params ), - request=request, - auth_kwargs=auth, ) - task_creation_response = await initial_operation.execute() validate_task_creation_response(task_creation_response) - task_id = task_creation_response.id - - final_response = await get_response( - task_id, auth_kwargs=auth, node_id=cls.hidden.unique_id - ) + final_response = await get_response(cls, task_creation_response.id) video = await download_url_to_video_output(final_response.output_url) return IO.NodeOutput(video) @@ -582,15 +368,10 @@ class MoonvalleyVideo2VideoNode(IO.ComfyNode): steps=33, prompt_adherence=4.5, ) -> IO.NodeOutput: - auth = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } - validated_video = validate_video_to_video_input(video) - video_url = await upload_video_to_comfyapi(validated_video, auth_kwargs=auth) - - validate_prompts(prompt, negative_prompt) + video_url = await upload_video_to_comfyapi(cls, validated_video) + validate_string(prompt, min_length=1, max_length=MOONVALLEY_MAREY_MAX_PROMPT_LENGTH) + validate_string(negative_prompt, field_name="negative_prompt", max_length=MOONVALLEY_MAREY_MAX_PROMPT_LENGTH) # Only include motion_intensity for Motion Transfer control_params = {} @@ -605,35 +386,20 @@ class MoonvalleyVideo2VideoNode(IO.ComfyNode): guidance_scale=prompt_adherence, ) - control = parse_control_parameter(control_type) - - request = MoonvalleyVideoToVideoRequest( - control_type=control, - video_url=video_url, - prompt_text=prompt, - inference_params=inference_params, - ) - - initial_operation = SynchronousOperation( - endpoint=ApiEndpoint( - path=API_VIDEO2VIDEO_ENDPOINT, - method=HttpMethod.POST, - request_model=MoonvalleyVideoToVideoRequest, - response_model=MoonvalleyPromptResponse, + task_creation_response = await sync_op( + cls, + endpoint=ApiEndpoint(path=API_VIDEO2VIDEO_ENDPOINT, method="POST"), + response_model=MoonvalleyPromptResponse, + data=MoonvalleyVideoToVideoRequest( + control_type=parse_control_parameter(control_type), + video_url=video_url, + prompt_text=prompt, + inference_params=inference_params, ), - request=request, - auth_kwargs=auth, ) - task_creation_response = await initial_operation.execute() validate_task_creation_response(task_creation_response) - task_id = task_creation_response.id - - final_response = await get_response( - task_id, auth_kwargs=auth, node_id=cls.hidden.unique_id - ) - - video = await download_url_to_video_output(final_response.output_url) - return IO.NodeOutput(video) + final_response = await get_response(cls, task_creation_response.id) + return IO.NodeOutput(await download_url_to_video_output(final_response.output_url)) class MoonvalleyTxt2VideoNode(IO.ComfyNode): @@ -720,14 +486,10 @@ class MoonvalleyTxt2VideoNode(IO.ComfyNode): seed: int, steps: int, ) -> IO.NodeOutput: - validate_prompts(prompt, negative_prompt, MOONVALLEY_MAREY_MAX_PROMPT_LENGTH) + validate_string(prompt, min_length=1, max_length=MOONVALLEY_MAREY_MAX_PROMPT_LENGTH) + validate_string(negative_prompt, field_name="negative_prompt", max_length=MOONVALLEY_MAREY_MAX_PROMPT_LENGTH) width_height = parse_width_height_from_res(resolution) - auth = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } - inference_params = MoonvalleyTextToVideoInferenceParams( negative_prompt=negative_prompt, steps=steps, @@ -737,30 +499,16 @@ class MoonvalleyTxt2VideoNode(IO.ComfyNode): width=width_height["width"], height=width_height["height"], ) - request = MoonvalleyTextToVideoRequest( - prompt_text=prompt, inference_params=inference_params - ) - init_op = SynchronousOperation( - endpoint=ApiEndpoint( - path=API_TXT2VIDEO_ENDPOINT, - method=HttpMethod.POST, - request_model=MoonvalleyTextToVideoRequest, - response_model=MoonvalleyPromptResponse, - ), - request=request, - auth_kwargs=auth, + task_creation_response = await sync_op( + cls, + endpoint=ApiEndpoint(path=API_TXT2VIDEO_ENDPOINT, method="POST"), + response_model=MoonvalleyPromptResponse, + data=MoonvalleyTextToVideoRequest(prompt_text=prompt, inference_params=inference_params), ) - task_creation_response = await init_op.execute() validate_task_creation_response(task_creation_response) - task_id = task_creation_response.id - - final_response = await get_response( - task_id, auth_kwargs=auth, node_id=cls.hidden.unique_id - ) - - video = await download_url_to_video_output(final_response.output_url) - return IO.NodeOutput(video) + final_response = await get_response(cls, task_creation_response.id) + return IO.NodeOutput(await download_url_to_video_output(final_response.output_url)) class MoonvalleyExtension(ComfyExtension): diff --git a/comfy_api_nodes/nodes_openai.py b/comfy_api_nodes/nodes_openai.py index e3b81de75..c467e840c 100644 --- a/comfy_api_nodes/nodes_openai.py +++ b/comfy_api_nodes/nodes_openai.py @@ -43,13 +43,11 @@ from comfy_api_nodes.apis.client import ( ) from comfy_api_nodes.apinode_utils import ( - downscale_image_tensor, validate_and_cast_response, - validate_string, - tensor_to_base64_string, text_filepath_to_data_uri, ) from comfy_api_nodes.mapper_utils import model_field_to_node_input +from comfy_api_nodes.util import downscale_image_tensor, validate_string, tensor_to_base64_string RESPONSES_ENDPOINT = "/proxy/openai/v1/responses" diff --git a/comfy_api_nodes/nodes_pika.py b/comfy_api_nodes/nodes_pika.py index 27cb0067b..5bb406a3b 100644 --- a/comfy_api_nodes/nodes_pika.py +++ b/comfy_api_nodes/nodes_pika.py @@ -14,11 +14,6 @@ import torch from typing_extensions import override from comfy_api.latest import ComfyExtension, IO from comfy_api.input_impl.video_types import VideoCodec, VideoContainer, VideoInput -from comfy_api_nodes.apinode_utils import ( - download_url_to_video_output, - tensor_to_bytesio, - validate_string, -) from comfy_api_nodes.apis import pika_defs from comfy_api_nodes.apis.client import ( ApiEndpoint, @@ -27,6 +22,7 @@ from comfy_api_nodes.apis.client import ( PollingOperation, SynchronousOperation, ) +from comfy_api_nodes.util import validate_string, download_url_to_video_output, tensor_to_bytesio R = TypeVar("R") diff --git a/comfy_api_nodes/nodes_pixverse.py b/comfy_api_nodes/nodes_pixverse.py index 438a7f80b..b2b841be8 100644 --- a/comfy_api_nodes/nodes_pixverse.py +++ b/comfy_api_nodes/nodes_pixverse.py @@ -24,10 +24,7 @@ from comfy_api_nodes.apis.client import ( PollingOperation, EmptyRequest, ) -from comfy_api_nodes.apinode_utils import ( - tensor_to_bytesio, - validate_string, -) +from comfy_api_nodes.util import validate_string, tensor_to_bytesio from comfy_api.input_impl import VideoFromFile from comfy_api.latest import ComfyExtension, IO @@ -50,7 +47,6 @@ def get_video_url_from_response( async def upload_image_to_pixverse(image: torch.Tensor, auth_kwargs=None): # first, upload image to Pixverse and get image id to use in actual generation call - files = {"image": tensor_to_bytesio(image)} operation = SynchronousOperation( endpoint=ApiEndpoint( path="/proxy/pixverse/image/upload", @@ -59,16 +55,14 @@ async def upload_image_to_pixverse(image: torch.Tensor, auth_kwargs=None): response_model=PixverseImageUploadResponse, ), request=EmptyRequest(), - files=files, + files={"image": tensor_to_bytesio(image)}, content_type="multipart/form-data", auth_kwargs=auth_kwargs, ) response_upload: PixverseImageUploadResponse = await operation.execute() if response_upload.Resp is None: - raise Exception( - f"PixVerse image upload request failed: '{response_upload.ErrMsg}'" - ) + raise Exception(f"PixVerse image upload request failed: '{response_upload.ErrMsg}'") return response_upload.Resp.img_id @@ -95,7 +89,6 @@ class PixverseTemplateNode(IO.ComfyNode): template_id = pixverse_templates.get(template, None) if template_id is None: raise Exception(f"Template '{template}' is not recognized.") - # just return the integer return IO.NodeOutput(template_id) diff --git a/comfy_api_nodes/nodes_recraft.py b/comfy_api_nodes/nodes_recraft.py index 8beed5675..8ee7e55c4 100644 --- a/comfy_api_nodes/nodes_recraft.py +++ b/comfy_api_nodes/nodes_recraft.py @@ -24,12 +24,10 @@ from comfy_api_nodes.apis.client import ( EmptyRequest, ) from comfy_api_nodes.apinode_utils import ( - bytesio_to_image_tensor, download_url_to_bytesio, - tensor_to_bytesio, resize_mask_to_image, - validate_string, ) +from comfy_api_nodes.util import validate_string, tensor_to_bytesio, bytesio_to_image_tensor from server import PromptServer import torch diff --git a/comfy_api_nodes/nodes_runway.py b/comfy_api_nodes/nodes_runway.py index eb03a897d..0543d1d0e 100644 --- a/comfy_api_nodes/nodes_runway.py +++ b/comfy_api_nodes/nodes_runway.py @@ -11,7 +11,7 @@ User Guides: """ -from typing import Union, Optional, Any +from typing import Union, Optional from typing_extensions import override from enum import Enum @@ -21,7 +21,6 @@ from comfy_api_nodes.apis import ( RunwayImageToVideoRequest, RunwayImageToVideoResponse, RunwayTaskStatusResponse as TaskStatusResponse, - RunwayTaskStatusEnum as TaskStatus, RunwayModelEnum as Model, RunwayDurationEnum as Duration, RunwayAspectRatioEnum as AspectRatio, @@ -33,23 +32,20 @@ from comfy_api_nodes.apis import ( ReferenceImage, RunwayTextToImageAspectRatioEnum, ) -from comfy_api_nodes.apis.client import ( - ApiEndpoint, - HttpMethod, - SynchronousOperation, - PollingOperation, - EmptyRequest, -) -from comfy_api_nodes.apinode_utils import ( - upload_images_to_comfyapi, - download_url_to_video_output, +from comfy_api_nodes.util import ( image_tensor_pair_to_batch, validate_string, + validate_image_dimensions, + validate_image_aspect_ratio, + upload_images_to_comfyapi, + download_url_to_video_output, download_url_to_image_tensor, + ApiEndpoint, + sync_op, + poll_op, ) from comfy_api.input_impl import VideoFromFile from comfy_api.latest import ComfyExtension, IO -from comfy_api_nodes.util.validation_utils import validate_image_dimensions, validate_image_aspect_ratio PATH_IMAGE_TO_VIDEO = "/proxy/runway/image_to_video" PATH_TEXT_TO_IMAGE = "/proxy/runway/text_to_image" @@ -91,31 +87,6 @@ def get_video_url_from_task_status(response: TaskStatusResponse) -> Union[str, N return None -async def poll_until_finished( - auth_kwargs: dict[str, str], - api_endpoint: ApiEndpoint[Any, TaskStatusResponse], - estimated_duration: Optional[int] = None, - node_id: Optional[str] = None, -) -> TaskStatusResponse: - """Polls the Runway API endpoint until the task reaches a terminal state, then returns the response.""" - return await PollingOperation( - poll_endpoint=api_endpoint, - completed_statuses=[ - TaskStatus.SUCCEEDED.value, - ], - failed_statuses=[ - TaskStatus.FAILED.value, - TaskStatus.CANCELLED.value, - ], - status_extractor=lambda response: response.status.value, - auth_kwargs=auth_kwargs, - result_url_extractor=get_video_url_from_task_status, - estimated_duration=estimated_duration, - node_id=node_id, - progress_extractor=extract_progress_from_task_status, - ).execute() - - def extract_progress_from_task_status( response: TaskStatusResponse, ) -> Union[float, None]: @@ -132,42 +103,32 @@ def get_image_url_from_task_status(response: TaskStatusResponse) -> Union[str, N async def get_response( - task_id: str, auth_kwargs: dict[str, str], node_id: Optional[str] = None, estimated_duration: Optional[int] = None + cls: type[IO.ComfyNode], task_id: str, estimated_duration: Optional[int] = None ) -> TaskStatusResponse: """Poll the task status until it is finished then get the response.""" - return await poll_until_finished( - auth_kwargs, - ApiEndpoint( - path=f"{PATH_GET_TASK_STATUS}/{task_id}", - method=HttpMethod.GET, - request_model=EmptyRequest, - response_model=TaskStatusResponse, - ), + return await poll_op( + cls, + ApiEndpoint(path=f"{PATH_GET_TASK_STATUS}/{task_id}"), + response_model=TaskStatusResponse, + status_extractor=lambda r: r.status.value, estimated_duration=estimated_duration, - node_id=node_id, + progress_extractor=extract_progress_from_task_status, ) async def generate_video( + cls: type[IO.ComfyNode], request: RunwayImageToVideoRequest, - auth_kwargs: dict[str, str], - node_id: Optional[str] = None, estimated_duration: Optional[int] = None, ) -> VideoFromFile: - initial_operation = SynchronousOperation( - endpoint=ApiEndpoint( - path=PATH_IMAGE_TO_VIDEO, - method=HttpMethod.POST, - request_model=RunwayImageToVideoRequest, - response_model=RunwayImageToVideoResponse, - ), - request=request, - auth_kwargs=auth_kwargs, + initial_response = await sync_op( + cls, + endpoint=ApiEndpoint(path=PATH_IMAGE_TO_VIDEO, method="POST"), + response_model=RunwayImageToVideoResponse, + data=request, ) - initial_response = await initial_operation.execute() - - final_response = await get_response(initial_response.id, auth_kwargs, node_id, estimated_duration) + final_response = await get_response(cls, initial_response.id, estimated_duration) if not final_response.output: raise RunwayApiError("Runway task succeeded but no video data found in response.") @@ -184,9 +145,9 @@ class RunwayImageToVideoNodeGen3a(IO.ComfyNode): display_name="Runway Image to Video (Gen3a Turbo)", category="api node/video/Runway", description="Generate a video from a single starting frame using Gen3a Turbo model. " - "Before diving in, review these best practices to ensure that " - "your input selections will set your generation up for success: " - "https://help.runwayml.com/hc/en-us/articles/33927968552339-Creating-with-Act-One-on-Gen-3-Alpha-and-Turbo.", + "Before diving in, review these best practices to ensure that " + "your input selections will set your generation up for success: " + "https://help.runwayml.com/hc/en-us/articles/33927968552339-Creating-with-Act-One-on-Gen-3-Alpha-and-Turbo.", inputs=[ IO.String.Input( "prompt", @@ -241,20 +202,16 @@ class RunwayImageToVideoNodeGen3a(IO.ComfyNode): validate_image_dimensions(start_frame, max_width=7999, max_height=7999) validate_image_aspect_ratio(start_frame, min_aspect_ratio=0.5, max_aspect_ratio=2.0) - auth_kwargs = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } - download_urls = await upload_images_to_comfyapi( + cls, start_frame, max_images=1, mime_type="image/png", - auth_kwargs=auth_kwargs, ) return IO.NodeOutput( await generate_video( + cls, RunwayImageToVideoRequest( promptText=prompt, seed=seed, @@ -262,15 +219,9 @@ class RunwayImageToVideoNodeGen3a(IO.ComfyNode): duration=Duration(duration), ratio=AspectRatio(ratio), promptImage=RunwayPromptImageObject( - root=[ - RunwayPromptImageDetailedObject( - uri=str(download_urls[0]), position="first" - ) - ] + root=[RunwayPromptImageDetailedObject(uri=str(download_urls[0]), position="first")] ), ), - auth_kwargs=auth_kwargs, - node_id=cls.hidden.unique_id, ) ) @@ -284,9 +235,9 @@ class RunwayImageToVideoNodeGen4(IO.ComfyNode): display_name="Runway Image to Video (Gen4 Turbo)", category="api node/video/Runway", description="Generate a video from a single starting frame using Gen4 Turbo model. " - "Before diving in, review these best practices to ensure that " - "your input selections will set your generation up for success: " - "https://help.runwayml.com/hc/en-us/articles/37327109429011-Creating-with-Gen-4-Video.", + "Before diving in, review these best practices to ensure that " + "your input selections will set your generation up for success: " + "https://help.runwayml.com/hc/en-us/articles/37327109429011-Creating-with-Gen-4-Video.", inputs=[ IO.String.Input( "prompt", @@ -341,20 +292,16 @@ class RunwayImageToVideoNodeGen4(IO.ComfyNode): validate_image_dimensions(start_frame, max_width=7999, max_height=7999) validate_image_aspect_ratio(start_frame, min_aspect_ratio=0.5, max_aspect_ratio=2.0) - auth_kwargs = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } - download_urls = await upload_images_to_comfyapi( + cls, start_frame, max_images=1, mime_type="image/png", - auth_kwargs=auth_kwargs, ) return IO.NodeOutput( await generate_video( + cls, RunwayImageToVideoRequest( promptText=prompt, seed=seed, @@ -362,15 +309,9 @@ class RunwayImageToVideoNodeGen4(IO.ComfyNode): duration=Duration(duration), ratio=AspectRatio(ratio), promptImage=RunwayPromptImageObject( - root=[ - RunwayPromptImageDetailedObject( - uri=str(download_urls[0]), position="first" - ) - ] + root=[RunwayPromptImageDetailedObject(uri=str(download_urls[0]), position="first")] ), ), - auth_kwargs=auth_kwargs, - node_id=cls.hidden.unique_id, estimated_duration=AVERAGE_DURATION_FLF_SECONDS, ) ) @@ -385,12 +326,12 @@ class RunwayFirstLastFrameNode(IO.ComfyNode): display_name="Runway First-Last-Frame to Video", category="api node/video/Runway", description="Upload first and last keyframes, draft a prompt, and generate a video. " - "More complex transitions, such as cases where the Last frame is completely different " - "from the First frame, may benefit from the longer 10s duration. " - "This would give the generation more time to smoothly transition between the two inputs. " - "Before diving in, review these best practices to ensure that your input selections " - "will set your generation up for success: " - "https://help.runwayml.com/hc/en-us/articles/34170748696595-Creating-with-Keyframes-on-Gen-3.", + "More complex transitions, such as cases where the Last frame is completely different " + "from the First frame, may benefit from the longer 10s duration. " + "This would give the generation more time to smoothly transition between the two inputs. " + "Before diving in, review these best practices to ensure that your input selections " + "will set your generation up for success: " + "https://help.runwayml.com/hc/en-us/articles/34170748696595-Creating-with-Keyframes-on-Gen-3.", inputs=[ IO.String.Input( "prompt", @@ -452,23 +393,19 @@ class RunwayFirstLastFrameNode(IO.ComfyNode): validate_image_aspect_ratio(start_frame, min_aspect_ratio=0.5, max_aspect_ratio=2.0) validate_image_aspect_ratio(end_frame, min_aspect_ratio=0.5, max_aspect_ratio=2.0) - auth_kwargs = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } - stacked_input_images = image_tensor_pair_to_batch(start_frame, end_frame) download_urls = await upload_images_to_comfyapi( + cls, stacked_input_images, max_images=2, mime_type="image/png", - auth_kwargs=auth_kwargs, ) if len(download_urls) != 2: raise RunwayApiError("Failed to upload one or more images to comfy api.") return IO.NodeOutput( await generate_video( + cls, RunwayImageToVideoRequest( promptText=prompt, seed=seed, @@ -477,17 +414,11 @@ class RunwayFirstLastFrameNode(IO.ComfyNode): ratio=AspectRatio(ratio), promptImage=RunwayPromptImageObject( root=[ - RunwayPromptImageDetailedObject( - uri=str(download_urls[0]), position="first" - ), - RunwayPromptImageDetailedObject( - uri=str(download_urls[1]), position="last" - ), + RunwayPromptImageDetailedObject(uri=str(download_urls[0]), position="first"), + RunwayPromptImageDetailedObject(uri=str(download_urls[1]), position="last"), ] ), ), - auth_kwargs=auth_kwargs, - node_id=cls.hidden.unique_id, estimated_duration=AVERAGE_DURATION_FLF_SECONDS, ) ) @@ -502,7 +433,7 @@ class RunwayTextToImageNode(IO.ComfyNode): display_name="Runway Text to Image", category="api node/image/Runway", description="Generate an image from a text prompt using Runway's Gen 4 model. " - "You can also include reference image to guide the generation.", + "You can also include reference image to guide the generation.", inputs=[ IO.String.Input( "prompt", @@ -540,49 +471,34 @@ class RunwayTextToImageNode(IO.ComfyNode): ) -> IO.NodeOutput: validate_string(prompt, min_length=1) - auth_kwargs = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } - # Prepare reference images if provided reference_images = None if reference_image is not None: validate_image_dimensions(reference_image, max_width=7999, max_height=7999) validate_image_aspect_ratio(reference_image, min_aspect_ratio=0.5, max_aspect_ratio=2.0) download_urls = await upload_images_to_comfyapi( + cls, reference_image, max_images=1, mime_type="image/png", - auth_kwargs=auth_kwargs, ) reference_images = [ReferenceImage(uri=str(download_urls[0]))] - request = RunwayTextToImageRequest( - promptText=prompt, - model=Model4.gen4_image, - ratio=ratio, - referenceImages=reference_images, - ) - - initial_operation = SynchronousOperation( - endpoint=ApiEndpoint( - path=PATH_TEXT_TO_IMAGE, - method=HttpMethod.POST, - request_model=RunwayTextToImageRequest, - response_model=RunwayTextToImageResponse, + initial_response = await sync_op( + cls, + endpoint=ApiEndpoint(path=PATH_TEXT_TO_IMAGE, method="POST"), + response_model=RunwayTextToImageResponse, + data=RunwayTextToImageRequest( + promptText=prompt, + model=Model4.gen4_image, + ratio=ratio, + referenceImages=reference_images, ), - request=request, - auth_kwargs=auth_kwargs, ) - initial_response = await initial_operation.execute() - - # Poll for completion final_response = await get_response( + cls, initial_response.id, - auth_kwargs=auth_kwargs, - node_id=cls.hidden.unique_id, estimated_duration=AVERAGE_DURATION_T2I_SECONDS, ) if not final_response.output: @@ -601,5 +517,6 @@ class RunwayExtension(ComfyExtension): RunwayTextToImageNode, ] + async def comfy_entrypoint() -> RunwayExtension: return RunwayExtension() diff --git a/comfy_api_nodes/nodes_sora.py b/comfy_api_nodes/nodes_sora.py index efc954869..92b225d40 100644 --- a/comfy_api_nodes/nodes_sora.py +++ b/comfy_api_nodes/nodes_sora.py @@ -1,23 +1,20 @@ from typing import Optional -from typing_extensions import override import torch from pydantic import BaseModel, Field -from comfy_api.latest import ComfyExtension, IO -from comfy_api_nodes.apis.client import ( - ApiEndpoint, - HttpMethod, - SynchronousOperation, - PollingOperation, - EmptyRequest, -) -from comfy_api_nodes.util.validation_utils import get_number_of_images +from typing_extensions import override -from comfy_api_nodes.apinode_utils import ( +from comfy_api.latest import IO, ComfyExtension +from comfy_api_nodes.util import ( + ApiEndpoint, download_url_to_video_output, + get_number_of_images, + poll_op, + sync_op, tensor_to_bytesio, ) + class Sora2GenerationRequest(BaseModel): prompt: str = Field(...) model: str = Field(...) @@ -80,7 +77,7 @@ class OpenAIVideoSora2(IO.ComfyNode): control_after_generate=True, optional=True, tooltip="Seed to determine if node should re-run; " - "actual results are nondeterministic regardless of seed.", + "actual results are nondeterministic regardless of seed.", ), ], outputs=[ @@ -111,55 +108,34 @@ class OpenAIVideoSora2(IO.ComfyNode): if get_number_of_images(image) != 1: raise ValueError("Currently only one input image is supported.") files_input = {"input_reference": ("image.png", tensor_to_bytesio(image), "image/png")} - auth = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } - payload = Sora2GenerationRequest( - model=model, - prompt=prompt, - seconds=str(duration), - size=size, - ) - initial_operation = SynchronousOperation( - endpoint=ApiEndpoint( - path="/proxy/openai/v1/videos", - method=HttpMethod.POST, - request_model=Sora2GenerationRequest, - response_model=Sora2GenerationResponse + initial_response = await sync_op( + cls, + endpoint=ApiEndpoint(path="/proxy/openai/v1/videos", method="POST"), + data=Sora2GenerationRequest( + model=model, + prompt=prompt, + seconds=str(duration), + size=size, ), - request=payload, files=files_input, - auth_kwargs=auth, + response_model=Sora2GenerationResponse, content_type="multipart/form-data", ) - initial_response = await initial_operation.execute() if initial_response.error: - raise Exception(initial_response.error.message) + raise Exception(initial_response.error["message"]) model_time_multiplier = 1 if model == "sora-2" else 2 - poll_operation = PollingOperation( - poll_endpoint=ApiEndpoint( - path=f"/proxy/openai/v1/videos/{initial_response.id}", - method=HttpMethod.GET, - request_model=EmptyRequest, - response_model=Sora2GenerationResponse - ), - completed_statuses=["completed"], - failed_statuses=["failed"], + await poll_op( + cls, + poll_endpoint=ApiEndpoint(path=f"/proxy/openai/v1/videos/{initial_response.id}"), + response_model=Sora2GenerationResponse, status_extractor=lambda x: x.status, - auth_kwargs=auth, poll_interval=8.0, max_poll_attempts=160, - node_id=cls.hidden.unique_id, - estimated_duration=45 * (duration / 4) * model_time_multiplier, + estimated_duration=int(45 * (duration / 4) * model_time_multiplier), ) - await poll_operation.execute() return IO.NodeOutput( - await download_url_to_video_output( - f"/proxy/openai/v1/videos/{initial_response.id}/content", - auth_kwargs=auth, - ) + await download_url_to_video_output(f"/proxy/openai/v1/videos/{initial_response.id}/content", cls=cls), ) diff --git a/comfy_api_nodes/nodes_stability.py b/comfy_api_nodes/nodes_stability.py index 8af03cfd1..783666ddf 100644 --- a/comfy_api_nodes/nodes_stability.py +++ b/comfy_api_nodes/nodes_stability.py @@ -27,14 +27,14 @@ from comfy_api_nodes.apis.client import ( PollingOperation, EmptyRequest, ) -from comfy_api_nodes.apinode_utils import ( +from comfy_api_nodes.util import ( + validate_audio_duration, + validate_string, + audio_input_to_mp3, bytesio_to_image_tensor, tensor_to_bytesio, - validate_string, audio_bytes_to_audio_input, - audio_input_to_mp3, ) -from comfy_api_nodes.util.validation_utils import validate_audio_duration import torch import base64 diff --git a/comfy_api_nodes/nodes_veo2.py b/comfy_api_nodes/nodes_veo2.py index daeaa823e..d37e9e9b4 100644 --- a/comfy_api_nodes/nodes_veo2.py +++ b/comfy_api_nodes/nodes_veo2.py @@ -1,28 +1,21 @@ -import logging import base64 -import aiohttp -import torch from io import BytesIO -from typing import Optional + from typing_extensions import override -from comfy_api.latest import ComfyExtension, IO from comfy_api.input_impl.video_types import VideoFromFile -from comfy_api_nodes.apis import ( - VeoGenVidRequest, - VeoGenVidResponse, +from comfy_api.latest import IO, ComfyExtension +from comfy_api_nodes.apis.veo_api import ( VeoGenVidPollRequest, VeoGenVidPollResponse, + VeoGenVidRequest, + VeoGenVidResponse, ) -from comfy_api_nodes.apis.client import ( +from comfy_api_nodes.util import ( ApiEndpoint, - HttpMethod, - SynchronousOperation, - PollingOperation, -) - -from comfy_api_nodes.apinode_utils import ( - downscale_image_tensor, + download_url_to_video_output, + poll_op, + sync_op, tensor_to_base64_string, ) @@ -35,28 +28,6 @@ MODELS_MAP = { "veo-3.0-fast-generate-001": "veo-3.0-fast-generate-001", } -def convert_image_to_base64(image: torch.Tensor): - if image is None: - return None - - scaled_image = downscale_image_tensor(image, total_pixels=2048*2048) - return tensor_to_base64_string(scaled_image) - - -def get_video_url_from_response(poll_response: VeoGenVidPollResponse) -> Optional[str]: - if ( - poll_response.response - and hasattr(poll_response.response, "videos") - and poll_response.response.videos - and len(poll_response.response.videos) > 0 - ): - video = poll_response.response.videos[0] - else: - return None - if hasattr(video, "gcsUri") and video.gcsUri: - return str(video.gcsUri) - return None - class VeoVideoGenerationNode(IO.ComfyNode): """ @@ -169,18 +140,13 @@ class VeoVideoGenerationNode(IO.ComfyNode): # Prepare the instances for the request instances = [] - instance = { - "prompt": prompt - } + instance = {"prompt": prompt} # Add image if provided if image is not None: - image_base64 = convert_image_to_base64(image) + image_base64 = tensor_to_base64_string(image) if image_base64: - instance["image"] = { - "bytesBase64Encoded": image_base64, - "mimeType": "image/png" - } + instance["image"] = {"bytesBase64Encoded": image_base64, "mimeType": "image/png"} instances.append(instance) @@ -198,119 +164,77 @@ class VeoVideoGenerationNode(IO.ComfyNode): if seed > 0: parameters["seed"] = seed # Only add generateAudio for Veo 3 models - if "veo-3.0" in model: + if model.find("veo-2.0") == -1: parameters["generateAudio"] = generate_audio - auth = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } - # Initial request to start video generation - initial_operation = SynchronousOperation( - endpoint=ApiEndpoint( - path=f"/proxy/veo/{model}/generate", - method=HttpMethod.POST, - request_model=VeoGenVidRequest, - response_model=VeoGenVidResponse - ), - request=VeoGenVidRequest( + initial_response = await sync_op( + cls, + ApiEndpoint(path=f"/proxy/veo/{model}/generate", method="POST"), + response_model=VeoGenVidResponse, + data=VeoGenVidRequest( instances=instances, - parameters=parameters + parameters=parameters, ), - auth_kwargs=auth, ) - initial_response = await initial_operation.execute() - operation_name = initial_response.name - - logging.info("Veo generation started with operation name: %s", operation_name) - - # Define status extractor function def status_extractor(response): # Only return "completed" if the operation is done, regardless of success or failure # We'll check for errors after polling completes return "completed" if response.done else "pending" - # Define progress extractor function - def progress_extractor(response): - # Could be enhanced if the API provides progress information - return None - - # Define the polling operation - poll_operation = PollingOperation( - poll_endpoint=ApiEndpoint( - path=f"/proxy/veo/{model}/poll", - method=HttpMethod.POST, - request_model=VeoGenVidPollRequest, - response_model=VeoGenVidPollResponse - ), - completed_statuses=["completed"], - failed_statuses=[], # No failed statuses, we'll handle errors after polling + poll_response = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/veo/{model}/poll", method="POST"), + response_model=VeoGenVidPollResponse, status_extractor=status_extractor, - progress_extractor=progress_extractor, - request=VeoGenVidPollRequest( - operationName=operation_name + data=VeoGenVidPollRequest( + operationName=initial_response.name, ), - auth_kwargs=auth, poll_interval=5.0, - result_url_extractor=get_video_url_from_response, - node_id=cls.hidden.unique_id, estimated_duration=AVERAGE_DURATION_VIDEO_GEN, ) - # Execute the polling operation - poll_response = await poll_operation.execute() - # Now check for errors in the final response # Check for error in poll response - if hasattr(poll_response, 'error') and poll_response.error: - error_message = f"Veo API error: {poll_response.error.message} (code: {poll_response.error.code})" - logging.error(error_message) - raise Exception(error_message) + if poll_response.error: + raise Exception(f"Veo API error: {poll_response.error.message} (code: {poll_response.error.code})") # Check for RAI filtered content - if (hasattr(poll_response.response, 'raiMediaFilteredCount') and - poll_response.response.raiMediaFilteredCount > 0): + if ( + hasattr(poll_response.response, "raiMediaFilteredCount") + and poll_response.response.raiMediaFilteredCount > 0 + ): # Extract reason message if available - if (hasattr(poll_response.response, 'raiMediaFilteredReasons') and - poll_response.response.raiMediaFilteredReasons): + if ( + hasattr(poll_response.response, "raiMediaFilteredReasons") + and poll_response.response.raiMediaFilteredReasons + ): reason = poll_response.response.raiMediaFilteredReasons[0] error_message = f"Content filtered by Google's Responsible AI practices: {reason} ({poll_response.response.raiMediaFilteredCount} videos filtered.)" else: error_message = f"Content filtered by Google's Responsible AI practices ({poll_response.response.raiMediaFilteredCount} videos filtered.)" - logging.error(error_message) raise Exception(error_message) # Extract video data - if poll_response.response and hasattr(poll_response.response, 'videos') and poll_response.response.videos and len(poll_response.response.videos) > 0: + if ( + poll_response.response + and hasattr(poll_response.response, "videos") + and poll_response.response.videos + and len(poll_response.response.videos) > 0 + ): video = poll_response.response.videos[0] # Check if video is provided as base64 or URL - if hasattr(video, 'bytesBase64Encoded') and video.bytesBase64Encoded: - # Decode base64 string to bytes - video_data = base64.b64decode(video.bytesBase64Encoded) - elif hasattr(video, 'gcsUri') and video.gcsUri: - # Download from URL - async with aiohttp.ClientSession() as session: - async with session.get(video.gcsUri) as video_response: - video_data = await video_response.content.read() - else: - raise Exception("Video returned but no data or URL was provided") - else: - raise Exception("Video generation completed but no video was returned") + if hasattr(video, "bytesBase64Encoded") and video.bytesBase64Encoded: + return IO.NodeOutput(VideoFromFile(BytesIO(base64.b64decode(video.bytesBase64Encoded)))) - if not video_data: - raise Exception("No video data was returned") + if hasattr(video, "gcsUri") and video.gcsUri: + return IO.NodeOutput(await download_url_to_video_output(video.gcsUri)) - logging.info("Video generation completed successfully") - - # Convert video data to BytesIO object - video_io = BytesIO(video_data) - - # Return VideoFromFile object - return IO.NodeOutput(VideoFromFile(video_io)) + raise Exception("Video returned but no data or URL was provided") + raise Exception("Video generation completed but no video was returned") class Veo3VideoGenerationNode(VeoVideoGenerationNode): @@ -394,7 +318,10 @@ class Veo3VideoGenerationNode(VeoVideoGenerationNode): IO.Combo.Input( "model", options=[ - "veo-3.1-generate", "veo-3.1-fast-generate", "veo-3.0-generate-001", "veo-3.0-fast-generate-001" + "veo-3.1-generate", + "veo-3.1-fast-generate", + "veo-3.0-generate-001", + "veo-3.0-fast-generate-001", ], default="veo-3.0-generate-001", tooltip="Veo 3 model to use for video generation", @@ -427,5 +354,6 @@ class VeoExtension(ComfyExtension): Veo3VideoGenerationNode, ] + async def comfy_entrypoint() -> VeoExtension: return VeoExtension() diff --git a/comfy_api_nodes/nodes_vidu.py b/comfy_api_nodes/nodes_vidu.py index 639be4b2b..0e0572f8c 100644 --- a/comfy_api_nodes/nodes_vidu.py +++ b/comfy_api_nodes/nodes_vidu.py @@ -1,27 +1,23 @@ import logging from enum import Enum -from typing import Any, Callable, Optional, Literal, TypeVar -from typing_extensions import override +from typing import Literal, Optional, TypeVar import torch from pydantic import BaseModel, Field +from typing_extensions import override -from comfy_api.latest import ComfyExtension, IO -from comfy_api_nodes.util.validation_utils import ( - validate_aspect_ratio_closeness, - validate_image_dimensions, - validate_image_aspect_ratio_range, - get_number_of_images, -) -from comfy_api_nodes.apis.client import ( +from comfy_api.latest import IO, ComfyExtension +from comfy_api_nodes.util import ( ApiEndpoint, - HttpMethod, - SynchronousOperation, - PollingOperation, - EmptyRequest, + download_url_to_video_output, + get_number_of_images, + poll_op, + sync_op, + upload_images_to_comfyapi, + validate_aspect_ratio_closeness, + validate_image_aspect_ratio_range, + validate_image_dimensions, ) -from comfy_api_nodes.apinode_utils import download_url_to_video_output, upload_images_to_comfyapi - VIDU_TEXT_TO_VIDEO = "/proxy/vidu/text2video" VIDU_IMAGE_TO_VIDEO = "/proxy/vidu/img2video" @@ -31,8 +27,9 @@ VIDU_GET_GENERATION_STATUS = "/proxy/vidu/tasks/%s/creations" R = TypeVar("R") + class VideoModelName(str, Enum): - vidu_q1 = 'viduq1' + vidu_q1 = "viduq1" class AspectRatio(str, Enum): @@ -63,17 +60,9 @@ class TaskCreationRequest(BaseModel): images: Optional[list[str]] = Field(None, description="Base64 encoded string or image URL") -class TaskStatus(str, Enum): - created = "created" - queueing = "queueing" - processing = "processing" - success = "success" - failed = "failed" - - class TaskCreationResponse(BaseModel): task_id: str = Field(...) - state: TaskStatus = Field(...) + state: str = Field(...) created_at: str = Field(...) code: Optional[int] = Field(None, description="Error code") @@ -85,32 +74,11 @@ class TaskResult(BaseModel): class TaskStatusResponse(BaseModel): - state: TaskStatus = Field(...) + state: str = Field(...) err_code: Optional[str] = Field(None) creations: list[TaskResult] = Field(..., description="Generated results") -async def poll_until_finished( - auth_kwargs: dict[str, str], - api_endpoint: ApiEndpoint[Any, R], - result_url_extractor: Optional[Callable[[R], str]] = None, - estimated_duration: Optional[int] = None, - node_id: Optional[str] = None, -) -> R: - return await PollingOperation( - poll_endpoint=api_endpoint, - completed_statuses=[TaskStatus.success.value], - failed_statuses=[TaskStatus.failed.value], - status_extractor=lambda response: response.state.value, - auth_kwargs=auth_kwargs, - result_url_extractor=result_url_extractor, - estimated_duration=estimated_duration, - node_id=node_id, - poll_interval=16.0, - max_poll_attempts=256, - ).execute() - - def get_video_url_from_response(response) -> Optional[str]: if response.creations: return response.creations[0].url @@ -127,37 +95,27 @@ def get_video_from_response(response) -> TaskResult: async def execute_task( + cls: type[IO.ComfyNode], vidu_endpoint: str, - auth_kwargs: Optional[dict[str, str]], payload: TaskCreationRequest, estimated_duration: int, - node_id: str, ) -> R: - response = await SynchronousOperation( - endpoint=ApiEndpoint( - path=vidu_endpoint, - method=HttpMethod.POST, - request_model=TaskCreationRequest, - response_model=TaskCreationResponse, - ), - request=payload, - auth_kwargs=auth_kwargs, - ).execute() - if response.state == TaskStatus.failed: + response = await sync_op( + cls, + endpoint=ApiEndpoint(path=vidu_endpoint, method="POST"), + response_model=TaskCreationResponse, + data=payload, + ) + if response.state == "failed": error_msg = f"Vidu request failed. Code: {response.code}" logging.error(error_msg) raise RuntimeError(error_msg) - return await poll_until_finished( - auth_kwargs, - ApiEndpoint( - path=VIDU_GET_GENERATION_STATUS % response.task_id, - method=HttpMethod.GET, - request_model=EmptyRequest, - response_model=TaskStatusResponse, - ), - result_url_extractor=get_video_url_from_response, + return await poll_op( + cls, + ApiEndpoint(path=VIDU_GET_GENERATION_STATUS % response.task_id), + response_model=TaskStatusResponse, + status_extractor=lambda r: r.state.value, estimated_duration=estimated_duration, - node_id=node_id, ) @@ -258,11 +216,7 @@ class ViduTextToVideoNode(IO.ComfyNode): resolution=resolution, movement_amplitude=movement_amplitude, ) - auth = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } - results = await execute_task(VIDU_TEXT_TO_VIDEO, auth, payload, 320, cls.hidden.unique_id) + results = await execute_task(cls, VIDU_TEXT_TO_VIDEO, payload, 320) return IO.NodeOutput(await download_url_to_video_output(get_video_from_response(results).url)) @@ -362,17 +316,13 @@ class ViduImageToVideoNode(IO.ComfyNode): resolution=resolution, movement_amplitude=movement_amplitude, ) - auth = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } payload.images = await upload_images_to_comfyapi( + cls, image, max_images=1, mime_type="image/png", - auth_kwargs=auth, ) - results = await execute_task(VIDU_IMAGE_TO_VIDEO, auth, payload, 120, cls.hidden.unique_id) + results = await execute_task(cls, VIDU_IMAGE_TO_VIDEO, payload, 120) return IO.NodeOutput(await download_url_to_video_output(get_video_from_response(results).url)) @@ -484,17 +434,13 @@ class ViduReferenceVideoNode(IO.ComfyNode): resolution=resolution, movement_amplitude=movement_amplitude, ) - auth = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } payload.images = await upload_images_to_comfyapi( + cls, images, max_images=7, mime_type="image/png", - auth_kwargs=auth, ) - results = await execute_task(VIDU_REFERENCE_VIDEO, auth, payload, 120, cls.hidden.unique_id) + results = await execute_task(cls, VIDU_REFERENCE_VIDEO, payload, 120) return IO.NodeOutput(await download_url_to_video_output(get_video_from_response(results).url)) @@ -596,15 +542,11 @@ class ViduStartEndToVideoNode(IO.ComfyNode): resolution=resolution, movement_amplitude=movement_amplitude, ) - auth = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } payload.images = [ - (await upload_images_to_comfyapi(frame, max_images=1, mime_type="image/png", auth_kwargs=auth))[0] + (await upload_images_to_comfyapi(cls, frame, max_images=1, mime_type="image/png"))[0] for frame in (first_frame, end_frame) ] - results = await execute_task(VIDU_START_END_VIDEO, auth, payload, 96, cls.hidden.unique_id) + results = await execute_task(cls, VIDU_START_END_VIDEO, payload, 96) return IO.NodeOutput(await download_url_to_video_output(get_video_from_response(results).url)) @@ -618,5 +560,6 @@ class ViduExtension(ComfyExtension): ViduStartEndToVideoNode, ] + async def comfy_entrypoint() -> ViduExtension: return ViduExtension() diff --git a/comfy_api_nodes/nodes_wan.py b/comfy_api_nodes/nodes_wan.py index b089bd907..2aab3c2ff 100644 --- a/comfy_api_nodes/nodes_wan.py +++ b/comfy_api_nodes/nodes_wan.py @@ -1,28 +1,24 @@ import re -from typing import Optional, Type, Union -from typing_extensions import override +from typing import Optional import torch from pydantic import BaseModel, Field -from comfy_api.latest import ComfyExtension, Input, IO -from comfy_api_nodes.apis.client import ( - ApiEndpoint, - HttpMethod, - SynchronousOperation, - PollingOperation, - EmptyRequest, - R, - T, -) -from comfy_api_nodes.util.validation_utils import get_number_of_images, validate_audio_duration +from typing_extensions import override -from comfy_api_nodes.apinode_utils import ( +from comfy_api.latest import IO, ComfyExtension, Input +from comfy_api_nodes.util import ( + ApiEndpoint, + audio_to_base64_string, download_url_to_image_tensor, download_url_to_video_output, + get_number_of_images, + poll_op, + sync_op, tensor_to_base64_string, - audio_to_base64_string, + validate_audio_duration, ) + class Text2ImageInputField(BaseModel): prompt: str = Field(...) negative_prompt: Optional[str] = Field(None) @@ -146,53 +142,7 @@ class VideoTaskStatusResponse(BaseModel): request_id: str = Field(...) -RES_IN_PARENS = re.compile(r'\((\d+)\s*[x×]\s*(\d+)\)') - - -async def process_task( - auth_kwargs: dict[str, str], - url: str, - request_model: Type[T], - response_model: Type[R], - payload: Union[ - Text2ImageTaskCreationRequest, - Image2ImageTaskCreationRequest, - Text2VideoTaskCreationRequest, - Image2VideoTaskCreationRequest, - ], - node_id: str, - estimated_duration: int, - poll_interval: int, -) -> Type[R]: - initial_response = await SynchronousOperation( - endpoint=ApiEndpoint( - path=url, - method=HttpMethod.POST, - request_model=request_model, - response_model=TaskCreationResponse, - ), - request=payload, - auth_kwargs=auth_kwargs, - ).execute() - - if not initial_response.output: - raise Exception(f"Unknown error occurred: {initial_response.code} - {initial_response.message}") - - return await PollingOperation( - poll_endpoint=ApiEndpoint( - path=f"/proxy/wan/api/v1/tasks/{initial_response.output.task_id}", - method=HttpMethod.GET, - request_model=EmptyRequest, - response_model=response_model, - ), - completed_statuses=["SUCCEEDED"], - failed_statuses=["FAILED", "CANCELED", "UNKNOWN"], - status_extractor=lambda x: x.output.task_status, - estimated_duration=estimated_duration, - poll_interval=poll_interval, - node_id=node_id, - auth_kwargs=auth_kwargs, - ).execute() +RES_IN_PARENS = re.compile(r"\((\d+)\s*[x×]\s*(\d+)\)") class WanTextToImageApi(IO.ComfyNode): @@ -259,7 +209,7 @@ class WanTextToImageApi(IO.ComfyNode): IO.Boolean.Input( "watermark", default=True, - tooltip="Whether to add an \"AI generated\" watermark to the result.", + tooltip='Whether to add an "AI generated" watermark to the result.', optional=True, ), ], @@ -286,26 +236,28 @@ class WanTextToImageApi(IO.ComfyNode): prompt_extend: bool = True, watermark: bool = True, ): - payload = Text2ImageTaskCreationRequest( - model=model, - input=Text2ImageInputField(prompt=prompt, negative_prompt=negative_prompt), - parameters=Txt2ImageParametersField( - size=f"{width}*{height}", - seed=seed, - prompt_extend=prompt_extend, - watermark=watermark, + initial_response = await sync_op( + cls, + ApiEndpoint(path="/proxy/wan/api/v1/services/aigc/text2image/image-synthesis", method="POST"), + response_model=TaskCreationResponse, + data=Text2ImageTaskCreationRequest( + model=model, + input=Text2ImageInputField(prompt=prompt, negative_prompt=negative_prompt), + parameters=Txt2ImageParametersField( + size=f"{width}*{height}", + seed=seed, + prompt_extend=prompt_extend, + watermark=watermark, + ), ), ) - response = await process_task( - { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - }, - "/proxy/wan/api/v1/services/aigc/text2image/image-synthesis", - request_model=Text2ImageTaskCreationRequest, + if not initial_response.output: + raise Exception(f"Unknown error occurred: {initial_response.code} - {initial_response.message}") + response = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/wan/api/v1/tasks/{initial_response.output.task_id}"), response_model=ImageTaskStatusResponse, - payload=payload, - node_id=cls.hidden.unique_id, + status_extractor=lambda x: x.output.task_status, estimated_duration=9, poll_interval=3, ) @@ -320,7 +272,7 @@ class WanImageToImageApi(IO.ComfyNode): display_name="Wan Image to Image", category="api node/image/Wan", description="Generates an image from one or two input images and a text prompt. " - "The output image is currently fixed at 1.6 MP; its aspect ratio matches the input image(s).", + "The output image is currently fixed at 1.6 MP; its aspect ratio matches the input image(s).", inputs=[ IO.Combo.Input( "model", @@ -376,7 +328,7 @@ class WanImageToImageApi(IO.ComfyNode): IO.Boolean.Input( "watermark", default=True, - tooltip="Whether to add an \"AI generated\" watermark to the result.", + tooltip='Whether to add an "AI generated" watermark to the result.', optional=True, ), ], @@ -408,28 +360,30 @@ class WanImageToImageApi(IO.ComfyNode): raise ValueError(f"Expected 1 or 2 input images, got {n_images}.") images = [] for i in image: - images.append("data:image/png;base64," + tensor_to_base64_string(i, total_pixels=4096*4096)) - payload = Image2ImageTaskCreationRequest( - model=model, - input=Image2ImageInputField(prompt=prompt, negative_prompt=negative_prompt, images=images), - parameters=Image2ImageParametersField( - # size=f"{width}*{height}", - seed=seed, - watermark=watermark, + images.append("data:image/png;base64," + tensor_to_base64_string(i, total_pixels=4096 * 4096)) + initial_response = await sync_op( + cls, + ApiEndpoint(path="/proxy/wan/api/v1/services/aigc/image2image/image-synthesis", method="POST"), + response_model=TaskCreationResponse, + data=Image2ImageTaskCreationRequest( + model=model, + input=Image2ImageInputField(prompt=prompt, negative_prompt=negative_prompt, images=images), + parameters=Image2ImageParametersField( + # size=f"{width}*{height}", + seed=seed, + watermark=watermark, + ), ), ) - response = await process_task( - { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - }, - "/proxy/wan/api/v1/services/aigc/image2image/image-synthesis", - request_model=Image2ImageTaskCreationRequest, + if not initial_response.output: + raise Exception(f"Unknown error occurred: {initial_response.code} - {initial_response.message}") + response = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/wan/api/v1/tasks/{initial_response.output.task_id}"), response_model=ImageTaskStatusResponse, - payload=payload, - node_id=cls.hidden.unique_id, + status_extractor=lambda x: x.output.task_status, estimated_duration=42, - poll_interval=3, + poll_interval=4, ) return IO.NodeOutput(await download_url_to_image_tensor(str(response.output.results[0].url))) @@ -523,7 +477,7 @@ class WanTextToVideoApi(IO.ComfyNode): IO.Boolean.Input( "watermark", default=True, - tooltip="Whether to add an \"AI generated\" watermark to the result.", + tooltip='Whether to add an "AI generated" watermark to the result.', optional=True, ), ], @@ -557,28 +511,31 @@ class WanTextToVideoApi(IO.ComfyNode): if audio is not None: validate_audio_duration(audio, 3.0, 29.0) audio_url = "data:audio/mp3;base64," + audio_to_base64_string(audio, "mp3", "libmp3lame") - payload = Text2VideoTaskCreationRequest( - model=model, - input=Text2VideoInputField(prompt=prompt, negative_prompt=negative_prompt, audio_url=audio_url), - parameters=Text2VideoParametersField( - size=f"{width}*{height}", - duration=duration, - seed=seed, - audio=generate_audio, - prompt_extend=prompt_extend, - watermark=watermark, + + initial_response = await sync_op( + cls, + ApiEndpoint(path="/proxy/wan/api/v1/services/aigc/video-generation/video-synthesis", method="POST"), + response_model=TaskCreationResponse, + data=Text2VideoTaskCreationRequest( + model=model, + input=Text2VideoInputField(prompt=prompt, negative_prompt=negative_prompt, audio_url=audio_url), + parameters=Text2VideoParametersField( + size=f"{width}*{height}", + duration=duration, + seed=seed, + audio=generate_audio, + prompt_extend=prompt_extend, + watermark=watermark, + ), ), ) - response = await process_task( - { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - }, - "/proxy/wan/api/v1/services/aigc/video-generation/video-synthesis", - request_model=Text2VideoTaskCreationRequest, + if not initial_response.output: + raise Exception(f"Unknown error occurred: {initial_response.code} - {initial_response.message}") + response = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/wan/api/v1/tasks/{initial_response.output.task_id}"), response_model=VideoTaskStatusResponse, - payload=payload, - node_id=cls.hidden.unique_id, + status_extractor=lambda x: x.output.task_status, estimated_duration=120 * int(duration / 5), poll_interval=6, ) @@ -667,7 +624,7 @@ class WanImageToVideoApi(IO.ComfyNode): IO.Boolean.Input( "watermark", default=True, - tooltip="Whether to add an \"AI generated\" watermark to the result.", + tooltip='Whether to add an "AI generated" watermark to the result.', optional=True, ), ], @@ -699,35 +656,37 @@ class WanImageToVideoApi(IO.ComfyNode): ): if get_number_of_images(image) != 1: raise ValueError("Exactly one input image is required.") - image_url = "data:image/png;base64," + tensor_to_base64_string(image, total_pixels=2000*2000) + image_url = "data:image/png;base64," + tensor_to_base64_string(image, total_pixels=2000 * 2000) audio_url = None if audio is not None: validate_audio_duration(audio, 3.0, 29.0) audio_url = "data:audio/mp3;base64," + audio_to_base64_string(audio, "mp3", "libmp3lame") - payload = Image2VideoTaskCreationRequest( - model=model, - input=Image2VideoInputField( - prompt=prompt, negative_prompt=negative_prompt, img_url=image_url, audio_url=audio_url - ), - parameters=Image2VideoParametersField( - resolution=resolution, - duration=duration, - seed=seed, - audio=generate_audio, - prompt_extend=prompt_extend, - watermark=watermark, + initial_response = await sync_op( + cls, + ApiEndpoint(path="/proxy/wan/api/v1/services/aigc/video-generation/video-synthesis", method="POST"), + response_model=TaskCreationResponse, + data=Image2VideoTaskCreationRequest( + model=model, + input=Image2VideoInputField( + prompt=prompt, negative_prompt=negative_prompt, img_url=image_url, audio_url=audio_url + ), + parameters=Image2VideoParametersField( + resolution=resolution, + duration=duration, + seed=seed, + audio=generate_audio, + prompt_extend=prompt_extend, + watermark=watermark, + ), ), ) - response = await process_task( - { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - }, - "/proxy/wan/api/v1/services/aigc/video-generation/video-synthesis", - request_model=Image2VideoTaskCreationRequest, + if not initial_response.output: + raise Exception(f"Unknown error occurred: {initial_response.code} - {initial_response.message}") + response = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/wan/api/v1/tasks/{initial_response.output.task_id}"), response_model=VideoTaskStatusResponse, - payload=payload, - node_id=cls.hidden.unique_id, + status_extractor=lambda x: x.output.task_status, estimated_duration=120 * int(duration / 5), poll_interval=6, ) diff --git a/comfy_api_nodes/util/__init__.py b/comfy_api_nodes/util/__init__.py index e69de29bb..c2ec391aa 100644 --- a/comfy_api_nodes/util/__init__.py +++ b/comfy_api_nodes/util/__init__.py @@ -0,0 +1,87 @@ +from ._helpers import get_fs_object_size +from .client import ( + ApiEndpoint, + poll_op, + poll_op_raw, + sync_op, + sync_op_raw, +) +from .conversions import ( + audio_bytes_to_audio_input, + audio_input_to_mp3, + audio_to_base64_string, + bytesio_to_image_tensor, + downscale_image_tensor, + image_tensor_pair_to_batch, + pil_to_bytesio, + tensor_to_base64_string, + tensor_to_bytesio, + tensor_to_pil, + trim_video, +) +from .download_helpers import ( + download_url_to_bytesio, + download_url_to_image_tensor, + download_url_to_video_output, +) +from .upload_helpers import ( + upload_audio_to_comfyapi, + upload_file_to_comfyapi, + upload_images_to_comfyapi, + upload_video_to_comfyapi, +) +from .validation_utils import ( + get_number_of_images, + validate_aspect_ratio_closeness, + validate_audio_duration, + validate_container_format_is_mp4, + validate_image_aspect_ratio, + validate_image_aspect_ratio_range, + validate_image_dimensions, + validate_string, + validate_video_dimensions, + validate_video_duration, +) + +__all__ = [ + # API client + "ApiEndpoint", + "poll_op", + "poll_op_raw", + "sync_op", + "sync_op_raw", + # Upload helpers + "upload_audio_to_comfyapi", + "upload_file_to_comfyapi", + "upload_images_to_comfyapi", + "upload_video_to_comfyapi", + # Download helpers + "download_url_to_bytesio", + "download_url_to_image_tensor", + "download_url_to_video_output", + # Conversions + "audio_bytes_to_audio_input", + "audio_input_to_mp3", + "audio_to_base64_string", + "bytesio_to_image_tensor", + "downscale_image_tensor", + "image_tensor_pair_to_batch", + "pil_to_bytesio", + "tensor_to_base64_string", + "tensor_to_bytesio", + "tensor_to_pil", + "trim_video", + # Validation utilities + "get_number_of_images", + "validate_aspect_ratio_closeness", + "validate_audio_duration", + "validate_container_format_is_mp4", + "validate_image_aspect_ratio", + "validate_image_aspect_ratio_range", + "validate_image_dimensions", + "validate_string", + "validate_video_dimensions", + "validate_video_duration", + # Misc functions + "get_fs_object_size", +] diff --git a/comfy_api_nodes/util/_helpers.py b/comfy_api_nodes/util/_helpers.py new file mode 100644 index 000000000..328fe5227 --- /dev/null +++ b/comfy_api_nodes/util/_helpers.py @@ -0,0 +1,71 @@ +import asyncio +import contextlib +import os +import time +from io import BytesIO +from typing import Callable, Optional, Union + +from comfy.cli_args import args +from comfy.model_management import processing_interrupted +from comfy_api.latest import IO + +from .common_exceptions import ProcessingInterrupted + + +def is_processing_interrupted() -> bool: + """Return True if user/runtime requested interruption.""" + return processing_interrupted() + + +def get_node_id(node_cls: type[IO.ComfyNode]) -> str: + return node_cls.hidden.unique_id + + +def get_auth_header(node_cls: type[IO.ComfyNode]) -> dict[str, str]: + if node_cls.hidden.auth_token_comfy_org: + return {"Authorization": f"Bearer {node_cls.hidden.auth_token_comfy_org}"} + if node_cls.hidden.api_key_comfy_org: + return {"X-API-KEY": node_cls.hidden.api_key_comfy_org} + return {} + + +def default_base_url() -> str: + return getattr(args, "comfy_api_base", "https://api.comfy.org") + + +async def sleep_with_interrupt( + seconds: float, + node_cls: Optional[type[IO.ComfyNode]], + label: Optional[str] = None, + start_ts: Optional[float] = None, + estimated_total: Optional[int] = None, + *, + display_callback: Optional[Callable[[type[IO.ComfyNode], str, int, Optional[int]], None]] = None, +): + """ + Sleep in 1s slices while: + - Checking for interruption (raises ProcessingInterrupted). + - Optionally emitting time progress via display_callback (if provided). + """ + end = time.monotonic() + seconds + while True: + if is_processing_interrupted(): + raise ProcessingInterrupted("Task cancelled") + now = time.monotonic() + if start_ts is not None and label and display_callback: + with contextlib.suppress(Exception): + display_callback(node_cls, label, int(now - start_ts), estimated_total) + if now >= end: + break + await asyncio.sleep(min(1.0, end - now)) + + +def mimetype_to_extension(mime_type: str) -> str: + """Converts a MIME type to a file extension.""" + return mime_type.split("/")[-1].lower() + + +def get_fs_object_size(path_or_object: Union[str, BytesIO]) -> int: + if isinstance(path_or_object, str): + return os.path.getsize(path_or_object) + return len(path_or_object.getvalue()) diff --git a/comfy_api_nodes/util/client.py b/comfy_api_nodes/util/client.py new file mode 100644 index 000000000..5833b118f --- /dev/null +++ b/comfy_api_nodes/util/client.py @@ -0,0 +1,941 @@ +import asyncio +import contextlib +import json +import logging +import socket +import time +import uuid +from dataclasses import dataclass +from enum import Enum +from io import BytesIO +from typing import Any, Callable, Iterable, Literal, Optional, Type, TypeVar, Union +from urllib.parse import urljoin, urlparse + +import aiohttp +from aiohttp.client_exceptions import ClientError, ContentTypeError +from pydantic import BaseModel + +from comfy import utils +from comfy_api.latest import IO +from comfy_api_nodes.apis import request_logger +from server import PromptServer + +from ._helpers import ( + default_base_url, + get_auth_header, + get_node_id, + is_processing_interrupted, + sleep_with_interrupt, +) +from .common_exceptions import ApiServerError, LocalNetworkError, ProcessingInterrupted + +M = TypeVar("M", bound=BaseModel) + + +class ApiEndpoint: + def __init__( + self, + path: str, + method: Literal["GET", "POST", "PUT", "DELETE", "PATCH"] = "GET", + *, + query_params: Optional[dict[str, Any]] = None, + headers: Optional[dict[str, str]] = None, + ): + self.path = path + self.method = method + self.query_params = query_params or {} + self.headers = headers or {} + + +@dataclass +class _RequestConfig: + node_cls: type[IO.ComfyNode] + endpoint: ApiEndpoint + timeout: float + content_type: str + data: Optional[dict[str, Any]] + files: Optional[Union[dict[str, Any], list[tuple[str, Any]]]] + multipart_parser: Optional[Callable] + max_retries: int + retry_delay: float + retry_backoff: float + wait_label: str = "Waiting" + monitor_progress: bool = True + estimated_total: Optional[int] = None + final_label_on_success: Optional[str] = "Completed" + progress_origin_ts: Optional[float] = None + + +@dataclass +class _PollUIState: + started: float + status_label: str = "Queued" + is_queued: bool = True + price: Optional[float] = None + estimated_duration: Optional[int] = None + base_processing_elapsed: float = 0.0 # sum of completed active intervals + active_since: Optional[float] = None # start time of current active interval (None if queued) + + +_RETRY_STATUS = {408, 429, 500, 502, 503, 504} +COMPLETED_STATUSES = ["succeeded", "succeed", "success", "completed"] +FAILED_STATUSES = ["cancelled", "canceled", "failed", "error"] +QUEUED_STATUSES = ["created", "queued", "queueing", "submitted"] + + +async def sync_op( + cls: type[IO.ComfyNode], + endpoint: ApiEndpoint, + *, + response_model: Type[M], + data: Optional[BaseModel] = None, + files: Optional[Union[dict[str, Any], list[tuple[str, Any]]]] = None, + content_type: str = "application/json", + timeout: float = 3600.0, + multipart_parser: Optional[Callable] = None, + max_retries: int = 3, + retry_delay: float = 1.0, + retry_backoff: float = 2.0, + wait_label: str = "Waiting for server", + estimated_duration: Optional[int] = None, + final_label_on_success: Optional[str] = "Completed", + progress_origin_ts: Optional[float] = None, + monitor_progress: bool = True, +) -> M: + raw = await sync_op_raw( + cls, + endpoint, + data=data, + files=files, + content_type=content_type, + timeout=timeout, + multipart_parser=multipart_parser, + max_retries=max_retries, + retry_delay=retry_delay, + retry_backoff=retry_backoff, + wait_label=wait_label, + estimated_duration=estimated_duration, + as_binary=False, + final_label_on_success=final_label_on_success, + progress_origin_ts=progress_origin_ts, + monitor_progress=monitor_progress, + ) + if not isinstance(raw, dict): + raise Exception("Expected JSON response to validate into a Pydantic model, got non-JSON (binary or text).") + return _validate_or_raise(response_model, raw) + + +async def poll_op( + cls: type[IO.ComfyNode], + poll_endpoint: ApiEndpoint, + *, + response_model: Type[M], + status_extractor: Callable[[M], Optional[Union[str, int]]], + progress_extractor: Optional[Callable[[M], Optional[int]]] = None, + price_extractor: Optional[Callable[[M], Optional[float]]] = None, + completed_statuses: Optional[list[Union[str, int]]] = None, + failed_statuses: Optional[list[Union[str, int]]] = None, + queued_statuses: Optional[list[Union[str, int]]] = None, + data: Optional[BaseModel] = None, + poll_interval: float = 5.0, + max_poll_attempts: int = 120, + timeout_per_poll: float = 120.0, + max_retries_per_poll: int = 3, + retry_delay_per_poll: float = 1.0, + retry_backoff_per_poll: float = 2.0, + estimated_duration: Optional[int] = None, + cancel_endpoint: Optional[ApiEndpoint] = None, + cancel_timeout: float = 10.0, +) -> M: + raw = await poll_op_raw( + cls, + poll_endpoint=poll_endpoint, + status_extractor=_wrap_model_extractor(response_model, status_extractor), + progress_extractor=_wrap_model_extractor(response_model, progress_extractor), + price_extractor=_wrap_model_extractor(response_model, price_extractor), + completed_statuses=completed_statuses, + failed_statuses=failed_statuses, + queued_statuses=queued_statuses, + data=data, + poll_interval=poll_interval, + max_poll_attempts=max_poll_attempts, + timeout_per_poll=timeout_per_poll, + max_retries_per_poll=max_retries_per_poll, + retry_delay_per_poll=retry_delay_per_poll, + retry_backoff_per_poll=retry_backoff_per_poll, + estimated_duration=estimated_duration, + cancel_endpoint=cancel_endpoint, + cancel_timeout=cancel_timeout, + ) + if not isinstance(raw, dict): + raise Exception("Expected JSON response to validate into a Pydantic model, got non-JSON (binary or text).") + return _validate_or_raise(response_model, raw) + + +async def sync_op_raw( + cls: type[IO.ComfyNode], + endpoint: ApiEndpoint, + *, + data: Optional[Union[dict[str, Any], BaseModel]] = None, + files: Optional[Union[dict[str, Any], list[tuple[str, Any]]]] = None, + content_type: str = "application/json", + timeout: float = 3600.0, + multipart_parser: Optional[Callable] = None, + max_retries: int = 3, + retry_delay: float = 1.0, + retry_backoff: float = 2.0, + wait_label: str = "Waiting for server", + estimated_duration: Optional[int] = None, + as_binary: bool = False, + final_label_on_success: Optional[str] = "Completed", + progress_origin_ts: Optional[float] = None, + monitor_progress: bool = True, +) -> Union[dict[str, Any], bytes]: + """ + Make a single network request. + - If as_binary=False (default): returns JSON dict (or {'_raw': ''} if non-JSON). + - If as_binary=True: returns bytes. + """ + if isinstance(data, BaseModel): + data = data.model_dump(exclude_none=True) + for k, v in list(data.items()): + if isinstance(v, Enum): + data[k] = v.value + cfg = _RequestConfig( + node_cls=cls, + endpoint=endpoint, + timeout=timeout, + content_type=content_type, + data=data, + files=files, + multipart_parser=multipart_parser, + max_retries=max_retries, + retry_delay=retry_delay, + retry_backoff=retry_backoff, + wait_label=wait_label, + monitor_progress=monitor_progress, + estimated_total=estimated_duration, + final_label_on_success=final_label_on_success, + progress_origin_ts=progress_origin_ts, + ) + return await _request_base(cfg, expect_binary=as_binary) + + +async def poll_op_raw( + cls: type[IO.ComfyNode], + poll_endpoint: ApiEndpoint, + *, + status_extractor: Callable[[dict[str, Any]], Optional[Union[str, int]]], + progress_extractor: Optional[Callable[[dict[str, Any]], Optional[int]]] = None, + price_extractor: Optional[Callable[[dict[str, Any]], Optional[float]]] = None, + completed_statuses: Optional[list[Union[str, int]]] = None, + failed_statuses: Optional[list[Union[str, int]]] = None, + queued_statuses: Optional[list[Union[str, int]]] = None, + data: Optional[Union[dict[str, Any], BaseModel]] = None, + poll_interval: float = 5.0, + max_poll_attempts: int = 120, + timeout_per_poll: float = 120.0, + max_retries_per_poll: int = 3, + retry_delay_per_poll: float = 1.0, + retry_backoff_per_poll: float = 2.0, + estimated_duration: Optional[int] = None, + cancel_endpoint: Optional[ApiEndpoint] = None, + cancel_timeout: float = 10.0, +) -> dict[str, Any]: + """ + Polls an endpoint until the task reaches a terminal state. Displays time while queued/processing, + checks interruption every second, and calls Cancel endpoint (if provided) on interruption. + + Uses default complete, failed and queued states assumption. + + Returns the final JSON response from the poll endpoint. + """ + completed_states = _normalize_statuses(COMPLETED_STATUSES if completed_statuses is None else completed_statuses) + failed_states = _normalize_statuses(FAILED_STATUSES if failed_statuses is None else failed_statuses) + queued_states = _normalize_statuses(QUEUED_STATUSES if queued_statuses is None else queued_statuses) + started = time.monotonic() + consumed_attempts = 0 # counts only non-queued polls + + progress_bar = utils.ProgressBar(100) if progress_extractor else None + last_progress: Optional[int] = None + + state = _PollUIState(started=started, estimated_duration=estimated_duration) + stop_ticker = asyncio.Event() + + async def _ticker(): + """Emit a UI update every second while polling is in progress.""" + try: + while not stop_ticker.is_set(): + if is_processing_interrupted(): + break + now = time.monotonic() + proc_elapsed = state.base_processing_elapsed + ( + (now - state.active_since) if state.active_since is not None else 0.0 + ) + _display_time_progress( + cls, + status=state.status_label, + elapsed_seconds=int(now - state.started), + estimated_total=state.estimated_duration, + price=state.price, + is_queued=state.is_queued, + processing_elapsed_seconds=int(proc_elapsed), + ) + await asyncio.sleep(1.0) + except Exception as exc: + logging.debug("Polling ticker exited: %s", exc) + + ticker_task = asyncio.create_task(_ticker()) + try: + while consumed_attempts < max_poll_attempts: + try: + resp_json = await sync_op_raw( + cls, + poll_endpoint, + data=data, + timeout=timeout_per_poll, + max_retries=max_retries_per_poll, + retry_delay=retry_delay_per_poll, + retry_backoff=retry_backoff_per_poll, + wait_label="Checking", + estimated_duration=None, + as_binary=False, + final_label_on_success=None, + monitor_progress=False, + ) + if not isinstance(resp_json, dict): + raise Exception("Polling endpoint returned non-JSON response.") + except ProcessingInterrupted: + if cancel_endpoint: + with contextlib.suppress(Exception): + await sync_op_raw( + cls, + cancel_endpoint, + timeout=cancel_timeout, + max_retries=0, + wait_label="Cancelling task", + estimated_duration=None, + as_binary=False, + final_label_on_success=None, + monitor_progress=False, + ) + raise + + try: + status = _normalize_status_value(status_extractor(resp_json)) + except Exception as e: + logging.error("Status extraction failed: %s", e) + status = None + + if price_extractor: + new_price = price_extractor(resp_json) + if new_price is not None: + state.price = new_price + + if progress_extractor: + new_progress = progress_extractor(resp_json) + if new_progress is not None and last_progress != new_progress: + progress_bar.update_absolute(new_progress, total=100) + last_progress = new_progress + + now_ts = time.monotonic() + is_queued = status in queued_states + + if is_queued: + if state.active_since is not None: # If we just moved from active -> queued, close the active interval + state.base_processing_elapsed += now_ts - state.active_since + state.active_since = None + else: + if state.active_since is None: # If we just moved from queued -> active, open a new active interval + state.active_since = now_ts + + state.is_queued = is_queued + state.status_label = status or ("Queued" if is_queued else "Processing") + if status in completed_states: + if state.active_since is not None: + state.base_processing_elapsed += now_ts - state.active_since + state.active_since = None + stop_ticker.set() + with contextlib.suppress(Exception): + await ticker_task + + if progress_bar and last_progress != 100: + progress_bar.update_absolute(100, total=100) + + _display_time_progress( + cls, + status=status if status else "Completed", + elapsed_seconds=int(now_ts - started), + estimated_total=estimated_duration, + price=state.price, + is_queued=False, + processing_elapsed_seconds=int(state.base_processing_elapsed), + ) + return resp_json + + if status in failed_states: + msg = f"Task failed: {json.dumps(resp_json)}" + logging.error(msg) + raise Exception(msg) + + try: + await sleep_with_interrupt(poll_interval, cls, None, None, None) + except ProcessingInterrupted: + if cancel_endpoint: + with contextlib.suppress(Exception): + await sync_op_raw( + cls, + cancel_endpoint, + timeout=cancel_timeout, + max_retries=0, + wait_label="Cancelling task", + estimated_duration=None, + as_binary=False, + final_label_on_success=None, + monitor_progress=False, + ) + raise + if not is_queued: + consumed_attempts += 1 + + raise Exception( + f"Polling timed out after {max_poll_attempts} non-queued attempts " + f"(~{int(max_poll_attempts * poll_interval)}s of active polling)." + ) + except ProcessingInterrupted: + raise + except (LocalNetworkError, ApiServerError): + raise + except Exception as e: + raise Exception(f"Polling aborted due to error: {e}") from e + finally: + stop_ticker.set() + with contextlib.suppress(Exception): + await ticker_task + + +def _display_text( + node_cls: type[IO.ComfyNode], + text: Optional[str], + *, + status: Optional[Union[str, int]] = None, + price: Optional[float] = None, +) -> None: + display_lines: list[str] = [] + if status: + display_lines.append(f"Status: {status.capitalize() if isinstance(status, str) else status}") + if price is not None: + display_lines.append(f"Price: ${float(price):,.4f}") + if text is not None: + display_lines.append(text) + if display_lines: + PromptServer.instance.send_progress_text("\n".join(display_lines), get_node_id(node_cls)) + + +def _display_time_progress( + node_cls: type[IO.ComfyNode], + status: Optional[Union[str, int]], + elapsed_seconds: int, + estimated_total: Optional[int] = None, + *, + price: Optional[float] = None, + is_queued: Optional[bool] = None, + processing_elapsed_seconds: Optional[int] = None, +) -> None: + if estimated_total is not None and estimated_total > 0 and is_queued is False: + pe = processing_elapsed_seconds if processing_elapsed_seconds is not None else elapsed_seconds + remaining = max(0, int(estimated_total) - int(pe)) + time_line = f"Time elapsed: {int(elapsed_seconds)}s (~{remaining}s remaining)" + else: + time_line = f"Time elapsed: {int(elapsed_seconds)}s" + _display_text(node_cls, time_line, status=status, price=price) + + +async def _diagnose_connectivity() -> dict[str, bool]: + """Best-effort connectivity diagnostics to distinguish local vs. server issues.""" + results = { + "internet_accessible": False, + "api_accessible": False, + "is_local_issue": False, + "is_api_issue": False, + } + timeout = aiohttp.ClientTimeout(total=5.0) + async with aiohttp.ClientSession(timeout=timeout) as session: + try: + async with session.get("https://www.google.com") as resp: + results["internet_accessible"] = resp.status < 500 + except (ClientError, asyncio.TimeoutError, socket.gaierror): + results["is_local_issue"] = True + return results + + parsed = urlparse(default_base_url()) + health_url = f"{parsed.scheme}://{parsed.netloc}/health" + with contextlib.suppress(ClientError, asyncio.TimeoutError): + async with session.get(health_url) as resp: + results["api_accessible"] = resp.status < 500 + results["is_api_issue"] = results["internet_accessible"] and not results["api_accessible"] + return results + + +def _unpack_tuple(t: tuple) -> tuple[str, Any, str]: + """Normalize (filename, value, content_type).""" + if len(t) == 2: + return t[0], t[1], "application/octet-stream" + if len(t) == 3: + return t[0], t[1], t[2] + raise ValueError("files tuple must be (filename, file[, content_type])") + + +def _merge_params(endpoint_params: dict[str, Any], method: str, data: Optional[dict[str, Any]]) -> dict[str, Any]: + params = dict(endpoint_params or {}) + if method.upper() == "GET" and data: + for k, v in data.items(): + if v is not None: + params[k] = v + return params + + +def _friendly_http_message(status: int, body: Any) -> str: + if status == 401: + return "Unauthorized: Please login first to use this node." + if status == 402: + return "Payment Required: Please add credits to your account to use this node." + if status == 409: + return "There is a problem with your account. Please contact support@comfy.org." + if status == 429: + return "Rate Limit Exceeded: Please try again later." + try: + if isinstance(body, dict): + err = body.get("error") + if isinstance(err, dict): + msg = err.get("message") + typ = err.get("type") + if msg and typ: + return f"API Error: {msg} (Type: {typ})" + if msg: + return f"API Error: {msg}" + return f"API Error: {json.dumps(body)}" + else: + txt = str(body) + if len(txt) <= 200: + return f"API Error (raw): {txt}" + return f"API Error (status {status})" + except Exception: + return f"HTTP {status}: Unknown error" + + +def _generate_operation_id(method: str, path: str, attempt: int) -> str: + slug = path.strip("/").replace("/", "_") or "op" + return f"{method}_{slug}_try{attempt}_{uuid.uuid4().hex[:8]}" + + +def _snapshot_request_body_for_logging( + content_type: str, + method: str, + data: Optional[dict[str, Any]], + files: Optional[Union[dict[str, Any], list[tuple[str, Any]]]], +) -> Optional[Union[dict[str, Any], str]]: + if method.upper() == "GET": + return None + if content_type == "multipart/form-data": + form_fields = sorted([k for k, v in (data or {}).items() if v is not None]) + file_fields: list[dict[str, str]] = [] + if files: + file_iter = files if isinstance(files, list) else list(files.items()) + for field_name, file_obj in file_iter: + if file_obj is None: + continue + if isinstance(file_obj, tuple): + filename = file_obj[0] + else: + filename = getattr(file_obj, "name", field_name) + file_fields.append({"field": field_name, "filename": str(filename or "")}) + return {"_multipart": True, "form_fields": form_fields, "file_fields": file_fields} + if content_type == "application/x-www-form-urlencoded": + return data or {} + return data or {} + + +async def _request_base(cfg: _RequestConfig, expect_binary: bool): + """Core request with retries, per-second interruption monitoring, true cancellation, and friendly errors.""" + url = cfg.endpoint.path + parsed_url = urlparse(url) + if not parsed_url.scheme and not parsed_url.netloc: # is URL relative? + url = urljoin(default_base_url().rstrip("/") + "/", url.lstrip("/")) + + method = cfg.endpoint.method + params = _merge_params(cfg.endpoint.query_params, method, cfg.data if method == "GET" else None) + + async def _monitor(stop_evt: asyncio.Event, start_ts: float): + """Every second: update elapsed time and signal interruption.""" + try: + while not stop_evt.is_set(): + if is_processing_interrupted(): + return + if cfg.monitor_progress: + _display_time_progress( + cfg.node_cls, cfg.wait_label, int(time.monotonic() - start_ts), cfg.estimated_total + ) + await asyncio.sleep(1.0) + except asyncio.CancelledError: + return # normal shutdown + + start_time = cfg.progress_origin_ts if cfg.progress_origin_ts is not None else time.monotonic() + attempt = 0 + delay = cfg.retry_delay + operation_succeeded: bool = False + final_elapsed_seconds: Optional[int] = None + while True: + attempt += 1 + stop_event = asyncio.Event() + monitor_task: Optional[asyncio.Task] = None + sess: Optional[aiohttp.ClientSession] = None + + operation_id = _generate_operation_id(method, cfg.endpoint.path, attempt) + logging.debug("[DEBUG] HTTP %s %s (attempt %d)", method, url, attempt) + + payload_headers = {"Accept": "*/*"} + if not parsed_url.scheme and not parsed_url.netloc: # is URL relative? + payload_headers.update(get_auth_header(cfg.node_cls)) + if cfg.endpoint.headers: + payload_headers.update(cfg.endpoint.headers) + + payload_kw: dict[str, Any] = {"headers": payload_headers} + if method == "GET": + payload_headers.pop("Content-Type", None) + request_body_log = _snapshot_request_body_for_logging(cfg.content_type, method, cfg.data, cfg.files) + try: + if cfg.monitor_progress: + monitor_task = asyncio.create_task(_monitor(stop_event, start_time)) + + timeout = aiohttp.ClientTimeout(total=cfg.timeout) + sess = aiohttp.ClientSession(timeout=timeout) + + if cfg.content_type == "multipart/form-data" and method != "GET": + # aiohttp will set Content-Type boundary; remove any fixed Content-Type + payload_headers.pop("Content-Type", None) + if cfg.multipart_parser and cfg.data: + form = cfg.multipart_parser(cfg.data) + if not isinstance(form, aiohttp.FormData): + raise ValueError("multipart_parser must return aiohttp.FormData") + else: + form = aiohttp.FormData(default_to_multipart=True) + if cfg.data: + for k, v in cfg.data.items(): + if v is None: + continue + form.add_field(k, str(v) if not isinstance(v, (bytes, bytearray)) else v) + if cfg.files: + file_iter = cfg.files if isinstance(cfg.files, list) else cfg.files.items() + for field_name, file_obj in file_iter: + if file_obj is None: + continue + if isinstance(file_obj, tuple): + filename, file_value, content_type = _unpack_tuple(file_obj) + else: + filename = getattr(file_obj, "name", field_name) + file_value = file_obj + content_type = "application/octet-stream" + # Attempt to rewind BytesIO for retries + if isinstance(file_value, BytesIO): + with contextlib.suppress(Exception): + file_value.seek(0) + form.add_field(field_name, file_value, filename=filename, content_type=content_type) + payload_kw["data"] = form + elif cfg.content_type == "application/x-www-form-urlencoded" and method != "GET": + payload_headers["Content-Type"] = "application/x-www-form-urlencoded" + payload_kw["data"] = cfg.data or {} + elif method != "GET": + payload_headers["Content-Type"] = "application/json" + payload_kw["json"] = cfg.data or {} + + try: + request_logger.log_request_response( + operation_id=operation_id, + request_method=method, + request_url=url, + request_headers=dict(payload_headers) if payload_headers else None, + request_params=dict(params) if params else None, + request_data=request_body_log, + ) + except Exception as _log_e: + logging.debug("[DEBUG] request logging failed: %s", _log_e) + + req_coro = sess.request(method, url, params=params, **payload_kw) + req_task = asyncio.create_task(req_coro) + + # Race: request vs. monitor (interruption) + tasks = {req_task} + if monitor_task: + tasks.add(monitor_task) + done, pending = await asyncio.wait(tasks, return_when=asyncio.FIRST_COMPLETED) + + if monitor_task and monitor_task in done: + # Interrupted – cancel the request and abort + if req_task in pending: + req_task.cancel() + raise ProcessingInterrupted("Task cancelled") + + # Otherwise, request finished + resp = await req_task + async with resp: + if resp.status >= 400: + try: + body = await resp.json() + except (ContentTypeError, json.JSONDecodeError): + body = await resp.text() + if resp.status in _RETRY_STATUS and attempt <= cfg.max_retries: + logging.warning( + "HTTP %s %s -> %s. Retrying in %.2fs (retry %d of %d).", + method, + url, + resp.status, + delay, + attempt, + cfg.max_retries, + ) + try: + request_logger.log_request_response( + operation_id=operation_id, + request_method=method, + request_url=url, + response_status_code=resp.status, + response_headers=dict(resp.headers), + response_content=body, + error_message=_friendly_http_message(resp.status, body), + ) + except Exception as _log_e: + logging.debug("[DEBUG] response logging failed: %s", _log_e) + + await sleep_with_interrupt( + delay, + cfg.node_cls, + cfg.wait_label if cfg.monitor_progress else None, + start_time if cfg.monitor_progress else None, + cfg.estimated_total, + display_callback=_display_time_progress if cfg.monitor_progress else None, + ) + delay *= cfg.retry_backoff + continue + msg = _friendly_http_message(resp.status, body) + try: + request_logger.log_request_response( + operation_id=operation_id, + request_method=method, + request_url=url, + response_status_code=resp.status, + response_headers=dict(resp.headers), + response_content=body, + error_message=msg, + ) + except Exception as _log_e: + logging.debug("[DEBUG] response logging failed: %s", _log_e) + raise Exception(msg) + + if expect_binary: + buff = bytearray() + last_tick = time.monotonic() + async for chunk in resp.content.iter_chunked(64 * 1024): + buff.extend(chunk) + now = time.monotonic() + if now - last_tick >= 1.0: + last_tick = now + if is_processing_interrupted(): + raise ProcessingInterrupted("Task cancelled") + if cfg.monitor_progress: + _display_time_progress( + cfg.node_cls, cfg.wait_label, int(now - start_time), cfg.estimated_total + ) + bytes_payload = bytes(buff) + operation_succeeded = True + final_elapsed_seconds = int(time.monotonic() - start_time) + try: + request_logger.log_request_response( + operation_id=operation_id, + request_method=method, + request_url=url, + response_status_code=resp.status, + response_headers=dict(resp.headers), + response_content=bytes_payload, + ) + except Exception as _log_e: + logging.debug("[DEBUG] response logging failed: %s", _log_e) + return bytes_payload + else: + try: + payload = await resp.json() + response_content_to_log: Any = payload + except (ContentTypeError, json.JSONDecodeError): + text = await resp.text() + try: + payload = json.loads(text) if text else {} + except json.JSONDecodeError: + payload = {"_raw": text} + response_content_to_log = payload if isinstance(payload, dict) else text + operation_succeeded = True + final_elapsed_seconds = int(time.monotonic() - start_time) + try: + request_logger.log_request_response( + operation_id=operation_id, + request_method=method, + request_url=url, + response_status_code=resp.status, + response_headers=dict(resp.headers), + response_content=response_content_to_log, + ) + except Exception as _log_e: + logging.debug("[DEBUG] response logging failed: %s", _log_e) + return payload + + except ProcessingInterrupted: + logging.debug("Polling was interrupted by user") + raise + except (ClientError, asyncio.TimeoutError, socket.gaierror) as e: + if attempt <= cfg.max_retries: + logging.warning( + "Connection error calling %s %s. Retrying in %.2fs (%d/%d): %s", + method, + url, + delay, + attempt, + cfg.max_retries, + str(e), + ) + try: + request_logger.log_request_response( + operation_id=operation_id, + request_method=method, + request_url=url, + request_headers=dict(payload_headers) if payload_headers else None, + request_params=dict(params) if params else None, + request_data=request_body_log, + error_message=f"{type(e).__name__}: {str(e)} (will retry)", + ) + except Exception as _log_e: + logging.debug("[DEBUG] request error logging failed: %s", _log_e) + await sleep_with_interrupt( + delay, + cfg.node_cls, + cfg.wait_label if cfg.monitor_progress else None, + start_time if cfg.monitor_progress else None, + cfg.estimated_total, + display_callback=_display_time_progress if cfg.monitor_progress else None, + ) + delay *= cfg.retry_backoff + continue + diag = await _diagnose_connectivity() + if diag.get("is_local_issue"): + try: + request_logger.log_request_response( + operation_id=operation_id, + request_method=method, + request_url=url, + request_headers=dict(payload_headers) if payload_headers else None, + request_params=dict(params) if params else None, + request_data=request_body_log, + error_message=f"LocalNetworkError: {str(e)}", + ) + except Exception as _log_e: + logging.debug("[DEBUG] final error logging failed: %s", _log_e) + raise LocalNetworkError( + "Unable to connect to the API server due to local network issues. " + "Please check your internet connection and try again." + ) from e + try: + request_logger.log_request_response( + operation_id=operation_id, + request_method=method, + request_url=url, + request_headers=dict(payload_headers) if payload_headers else None, + request_params=dict(params) if params else None, + request_data=request_body_log, + error_message=f"ApiServerError: {str(e)}", + ) + except Exception as _log_e: + logging.debug("[DEBUG] final error logging failed: %s", _log_e) + raise ApiServerError( + f"The API server at {default_base_url()} is currently unreachable. " + f"The service may be experiencing issues." + ) from e + finally: + stop_event.set() + if monitor_task: + monitor_task.cancel() + with contextlib.suppress(Exception): + await monitor_task + if sess: + with contextlib.suppress(Exception): + await sess.close() + if operation_succeeded and cfg.monitor_progress and cfg.final_label_on_success: + _display_time_progress( + cfg.node_cls, + status=cfg.final_label_on_success, + elapsed_seconds=( + final_elapsed_seconds + if final_elapsed_seconds is not None + else int(time.monotonic() - start_time) + ), + estimated_total=cfg.estimated_total, + price=None, + is_queued=False, + processing_elapsed_seconds=final_elapsed_seconds, + ) + + +def _validate_or_raise(response_model: Type[M], payload: Any) -> M: + try: + return response_model.model_validate(payload) + except Exception as e: + logging.error( + "Response validation failed for %s: %s", + getattr(response_model, "__name__", response_model), + e, + ) + raise Exception( + f"Response validation failed for {getattr(response_model, '__name__', response_model)}: {e}" + ) from e + + +def _wrap_model_extractor( + response_model: Type[M], + extractor: Optional[Callable[[M], Any]], +) -> Optional[Callable[[dict[str, Any]], Any]]: + """Wrap a typed extractor so it can be used by the dict-based poller. + Validates the dict into `response_model` before invoking `extractor`. + Uses a small per-wrapper cache keyed by `id(dict)` to avoid re-validating + the same response for multiple extractors in a single poll attempt. + """ + if extractor is None: + return None + _cache: dict[int, M] = {} + + def _wrapped(d: dict[str, Any]) -> Any: + try: + key = id(d) + model = _cache.get(key) + if model is None: + model = response_model.model_validate(d) + _cache[key] = model + return extractor(model) + except Exception as e: + logging.error("Extractor failed (typed -> dict wrapper): %s", e) + raise + + return _wrapped + + +def _normalize_statuses(values: Optional[Iterable[Union[str, int]]]) -> set[Union[str, int]]: + if not values: + return set() + out: set[Union[str, int]] = set() + for v in values: + nv = _normalize_status_value(v) + if nv is not None: + out.add(nv) + return out + + +def _normalize_status_value(val: Union[str, int, None]) -> Union[str, int, None]: + if isinstance(val, str): + return val.strip().lower() + return val diff --git a/comfy_api_nodes/util/common_exceptions.py b/comfy_api_nodes/util/common_exceptions.py new file mode 100644 index 000000000..0606a4407 --- /dev/null +++ b/comfy_api_nodes/util/common_exceptions.py @@ -0,0 +1,14 @@ +class NetworkError(Exception): + """Base exception for network-related errors with diagnostic information.""" + + +class LocalNetworkError(NetworkError): + """Exception raised when local network connectivity issues are detected.""" + + +class ApiServerError(NetworkError): + """Exception raised when the API server is unreachable but internet is working.""" + + +class ProcessingInterrupted(Exception): + """Operation was interrupted by user/runtime via processing_interrupted().""" diff --git a/comfy_api_nodes/util/conversions.py b/comfy_api_nodes/util/conversions.py new file mode 100644 index 000000000..10cd1051b --- /dev/null +++ b/comfy_api_nodes/util/conversions.py @@ -0,0 +1,407 @@ +import base64 +import logging +import math +import uuid +from io import BytesIO +from typing import Optional + +import av +import numpy as np +import torch +from PIL import Image + +from comfy.utils import common_upscale +from comfy_api.latest import Input, InputImpl + +from ._helpers import mimetype_to_extension + + +def bytesio_to_image_tensor(image_bytesio: BytesIO, mode: str = "RGBA") -> torch.Tensor: + """Converts image data from BytesIO to a torch.Tensor. + + Args: + image_bytesio: BytesIO object containing the image data. + mode: The PIL mode to convert the image to (e.g., "RGB", "RGBA"). + + Returns: + A torch.Tensor representing the image (1, H, W, C). + + Raises: + PIL.UnidentifiedImageError: If the image data cannot be identified. + ValueError: If the specified mode is invalid. + """ + image = Image.open(image_bytesio) + image = image.convert(mode) + image_array = np.array(image).astype(np.float32) / 255.0 + return torch.from_numpy(image_array).unsqueeze(0) + + +def image_tensor_pair_to_batch(image1: torch.Tensor, image2: torch.Tensor) -> torch.Tensor: + """ + Converts a pair of image tensors to a batch tensor. + If the images are not the same size, the smaller image is resized to + match the larger image. + """ + if image1.shape[1:] != image2.shape[1:]: + image2 = common_upscale( + image2.movedim(-1, 1), + image1.shape[2], + image1.shape[1], + "bilinear", + "center", + ).movedim(1, -1) + return torch.cat((image1, image2), dim=0) + + +def tensor_to_bytesio( + image: torch.Tensor, + name: Optional[str] = None, + total_pixels: int = 2048 * 2048, + mime_type: str = "image/png", +) -> BytesIO: + """Converts a torch.Tensor image to a named BytesIO object. + + Args: + image: Input torch.Tensor image. + name: Optional filename for the BytesIO object. + total_pixels: Maximum total pixels for potential downscaling. + mime_type: Target image MIME type (e.g., 'image/png', 'image/jpeg', 'image/webp', 'video/mp4'). + + Returns: + Named BytesIO object containing the image data, with pointer set to the start of buffer. + """ + if not mime_type: + mime_type = "image/png" + + pil_image = tensor_to_pil(image, total_pixels=total_pixels) + img_binary = pil_to_bytesio(pil_image, mime_type=mime_type) + img_binary.name = f"{name if name else uuid.uuid4()}.{mimetype_to_extension(mime_type)}" + return img_binary + + +def tensor_to_pil(image: torch.Tensor, total_pixels: int = 2048 * 2048) -> Image.Image: + """Converts a single torch.Tensor image [H, W, C] to a PIL Image, optionally downscaling.""" + if len(image.shape) > 3: + image = image[0] + # TODO: remove alpha if not allowed and present + input_tensor = image.cpu() + input_tensor = downscale_image_tensor(input_tensor.unsqueeze(0), total_pixels=total_pixels).squeeze() + image_np = (input_tensor.numpy() * 255).astype(np.uint8) + img = Image.fromarray(image_np) + return img + + +def tensor_to_base64_string( + image_tensor: torch.Tensor, + total_pixels: int = 2048 * 2048, + mime_type: str = "image/png", +) -> str: + """Convert [B, H, W, C] or [H, W, C] tensor to a base64 string. + + Args: + image_tensor: Input torch.Tensor image. + total_pixels: Maximum total pixels for potential downscaling. + mime_type: Target image MIME type (e.g., 'image/png', 'image/jpeg', 'image/webp', 'video/mp4'). + + Returns: + Base64 encoded string of the image. + """ + pil_image = tensor_to_pil(image_tensor, total_pixels=total_pixels) + img_byte_arr = pil_to_bytesio(pil_image, mime_type=mime_type) + img_bytes = img_byte_arr.getvalue() + # Encode bytes to base64 string + base64_encoded_string = base64.b64encode(img_bytes).decode("utf-8") + return base64_encoded_string + + +def pil_to_bytesio(img: Image.Image, mime_type: str = "image/png") -> BytesIO: + """Converts a PIL Image to a BytesIO object.""" + if not mime_type: + mime_type = "image/png" + + img_byte_arr = BytesIO() + # Derive PIL format from MIME type (e.g., 'image/png' -> 'PNG') + pil_format = mime_type.split("/")[-1].upper() + if pil_format == "JPG": + pil_format = "JPEG" + img.save(img_byte_arr, format=pil_format) + img_byte_arr.seek(0) + return img_byte_arr + + +def downscale_image_tensor(image, total_pixels=1536 * 1024) -> torch.Tensor: + """Downscale input image tensor to roughly the specified total pixels.""" + samples = image.movedim(-1, 1) + total = int(total_pixels) + scale_by = math.sqrt(total / (samples.shape[3] * samples.shape[2])) + if scale_by >= 1: + return image + width = round(samples.shape[3] * scale_by) + height = round(samples.shape[2] * scale_by) + + s = common_upscale(samples, width, height, "lanczos", "disabled") + s = s.movedim(1, -1) + return s + + +def tensor_to_data_uri( + image_tensor: torch.Tensor, + total_pixels: int = 2048 * 2048, + mime_type: str = "image/png", +) -> str: + """Converts a tensor image to a Data URI string. + + Args: + image_tensor: Input torch.Tensor image. + total_pixels: Maximum total pixels for potential downscaling. + mime_type: Target image MIME type (e.g., 'image/png', 'image/jpeg', 'image/webp'). + + Returns: + Data URI string (e.g., 'data:image/png;base64,...'). + """ + base64_string = tensor_to_base64_string(image_tensor, total_pixels, mime_type) + return f"data:{mime_type};base64,{base64_string}" + + +def audio_to_base64_string(audio: Input.Audio, container_format: str = "mp4", codec_name: str = "aac") -> str: + """Converts an audio input to a base64 string.""" + sample_rate: int = audio["sample_rate"] + waveform: torch.Tensor = audio["waveform"] + audio_data_np = audio_tensor_to_contiguous_ndarray(waveform) + audio_bytes_io = audio_ndarray_to_bytesio(audio_data_np, sample_rate, container_format, codec_name) + audio_bytes = audio_bytes_io.getvalue() + return base64.b64encode(audio_bytes).decode("utf-8") + + +def audio_ndarray_to_bytesio( + audio_data_np: np.ndarray, + sample_rate: int, + container_format: str = "mp4", + codec_name: str = "aac", +) -> BytesIO: + """ + Encodes a numpy array of audio data into a BytesIO object. + """ + audio_bytes_io = BytesIO() + with av.open(audio_bytes_io, mode="w", format=container_format) as output_container: + audio_stream = output_container.add_stream(codec_name, rate=sample_rate) + frame = av.AudioFrame.from_ndarray( + audio_data_np, + format="fltp", + layout="stereo" if audio_data_np.shape[0] > 1 else "mono", + ) + frame.sample_rate = sample_rate + frame.pts = 0 + + for packet in audio_stream.encode(frame): + output_container.mux(packet) + + # Flush stream + for packet in audio_stream.encode(None): + output_container.mux(packet) + + audio_bytes_io.seek(0) + return audio_bytes_io + + +def audio_tensor_to_contiguous_ndarray(waveform: torch.Tensor) -> np.ndarray: + """ + Prepares audio waveform for av library by converting to a contiguous numpy array. + + Args: + waveform: a tensor of shape (1, channels, samples) derived from a Comfy `AUDIO` type. + + Returns: + Contiguous numpy array of the audio waveform. If the audio was batched, + the first item is taken. + """ + if waveform.ndim != 3 or waveform.shape[0] != 1: + raise ValueError("Expected waveform tensor shape (1, channels, samples)") + + # If batch is > 1, take first item + if waveform.shape[0] > 1: + waveform = waveform[0] + + # Prepare for av: remove batch dim, move to CPU, make contiguous, convert to numpy array + audio_data_np = waveform.squeeze(0).cpu().contiguous().numpy() + if audio_data_np.dtype != np.float32: + audio_data_np = audio_data_np.astype(np.float32) + + return audio_data_np + + +def audio_input_to_mp3(audio: Input.Audio) -> BytesIO: + waveform = audio["waveform"].cpu() + + output_buffer = BytesIO() + output_container = av.open(output_buffer, mode="w", format="mp3") + + out_stream = output_container.add_stream("libmp3lame", rate=audio["sample_rate"]) + out_stream.bit_rate = 320000 + + frame = av.AudioFrame.from_ndarray( + waveform.movedim(0, 1).reshape(1, -1).float().numpy(), + format="flt", + layout="mono" if waveform.shape[0] == 1 else "stereo", + ) + frame.sample_rate = audio["sample_rate"] + frame.pts = 0 + output_container.mux(out_stream.encode(frame)) + output_container.mux(out_stream.encode(None)) + output_container.close() + output_buffer.seek(0) + return output_buffer + + +def trim_video(video: Input.Video, duration_sec: float) -> Input.Video: + """ + Returns a new VideoInput object trimmed from the beginning to the specified duration, + using av to avoid loading entire video into memory. + + Args: + video: Input video to trim + duration_sec: Duration in seconds to keep from the beginning + + Returns: + VideoFromFile object that owns the output buffer + """ + output_buffer = BytesIO() + input_container = None + output_container = None + + try: + # Get the stream source - this avoids loading entire video into memory + # when the source is already a file path + input_source = video.get_stream_source() + + # Open containers + input_container = av.open(input_source, mode="r") + output_container = av.open(output_buffer, mode="w", format="mp4") + + # Set up output streams for re-encoding + video_stream = None + audio_stream = None + + for stream in input_container.streams: + logging.info("Found stream: type=%s, class=%s", stream.type, type(stream)) + if isinstance(stream, av.VideoStream): + # Create output video stream with same parameters + video_stream = output_container.add_stream("h264", rate=stream.average_rate) + video_stream.width = stream.width + video_stream.height = stream.height + video_stream.pix_fmt = "yuv420p" + logging.info("Added video stream: %sx%s @ %sfps", stream.width, stream.height, stream.average_rate) + elif isinstance(stream, av.AudioStream): + # Create output audio stream with same parameters + audio_stream = output_container.add_stream("aac", rate=stream.sample_rate) + audio_stream.sample_rate = stream.sample_rate + audio_stream.layout = stream.layout + logging.info("Added audio stream: %sHz, %s channels", stream.sample_rate, stream.channels) + + # Calculate target frame count that's divisible by 16 + fps = input_container.streams.video[0].average_rate + estimated_frames = int(duration_sec * fps) + target_frames = (estimated_frames // 16) * 16 # Round down to nearest multiple of 16 + + if target_frames == 0: + raise ValueError("Video too short: need at least 16 frames for Moonvalley") + + frame_count = 0 + audio_frame_count = 0 + + # Decode and re-encode video frames + if video_stream: + for frame in input_container.decode(video=0): + if frame_count >= target_frames: + break + + # Re-encode frame + for packet in video_stream.encode(frame): + output_container.mux(packet) + frame_count += 1 + + # Flush encoder + for packet in video_stream.encode(): + output_container.mux(packet) + + logging.info("Encoded %s video frames (target: %s)", frame_count, target_frames) + + # Decode and re-encode audio frames + if audio_stream: + input_container.seek(0) # Reset to beginning for audio + for frame in input_container.decode(audio=0): + if frame.time >= duration_sec: + break + + # Re-encode frame + for packet in audio_stream.encode(frame): + output_container.mux(packet) + audio_frame_count += 1 + + # Flush encoder + for packet in audio_stream.encode(): + output_container.mux(packet) + + logging.info("Encoded %s audio frames", audio_frame_count) + + # Close containers + output_container.close() + input_container.close() + + # Return as VideoFromFile using the buffer + output_buffer.seek(0) + return InputImpl.VideoFromFile(output_buffer) + + except Exception as e: + # Clean up on error + if input_container is not None: + input_container.close() + if output_container is not None: + output_container.close() + raise RuntimeError(f"Failed to trim video: {str(e)}") from e + + +def _f32_pcm(wav: torch.Tensor) -> torch.Tensor: + """Convert audio to float 32 bits PCM format. Copy-paste from nodes_audio.py file.""" + if wav.dtype.is_floating_point: + return wav + elif wav.dtype == torch.int16: + return wav.float() / (2**15) + elif wav.dtype == torch.int32: + return wav.float() / (2**31) + raise ValueError(f"Unsupported wav dtype: {wav.dtype}") + + +def audio_bytes_to_audio_input(audio_bytes: bytes) -> dict: + """ + Decode any common audio container from bytes using PyAV and return + a Comfy AUDIO dict: {"waveform": [1, C, T] float32, "sample_rate": int}. + """ + with av.open(BytesIO(audio_bytes)) as af: + if not af.streams.audio: + raise ValueError("No audio stream found in response.") + stream = af.streams.audio[0] + + in_sr = int(stream.codec_context.sample_rate) + out_sr = in_sr + + frames: list[torch.Tensor] = [] + n_channels = stream.channels or 1 + + for frame in af.decode(streams=stream.index): + arr = frame.to_ndarray() # shape can be [C, T] or [T, C] or [T] + buf = torch.from_numpy(arr) + if buf.ndim == 1: + buf = buf.unsqueeze(0) # [T] -> [1, T] + elif buf.shape[0] != n_channels and buf.shape[-1] == n_channels: + buf = buf.transpose(0, 1).contiguous() # [T, C] -> [C, T] + elif buf.shape[0] != n_channels: + buf = buf.reshape(-1, n_channels).t().contiguous() # fallback to [C, T] + frames.append(buf) + + if not frames: + raise ValueError("Decoded zero audio frames.") + + wav = torch.cat(frames, dim=1) # [C, T] + wav = _f32_pcm(wav) + return {"waveform": wav.unsqueeze(0).contiguous(), "sample_rate": out_sr} diff --git a/comfy_api_nodes/util/download_helpers.py b/comfy_api_nodes/util/download_helpers.py new file mode 100644 index 000000000..055e690de --- /dev/null +++ b/comfy_api_nodes/util/download_helpers.py @@ -0,0 +1,249 @@ +import asyncio +import contextlib +import uuid +from io import BytesIO +from pathlib import Path +from typing import IO, Optional, Union +from urllib.parse import urljoin, urlparse + +import aiohttp +import torch +from aiohttp.client_exceptions import ClientError, ContentTypeError + +from comfy_api.input_impl import VideoFromFile +from comfy_api.latest import IO as COMFY_IO +from comfy_api_nodes.apis import request_logger + +from ._helpers import ( + default_base_url, + get_auth_header, + is_processing_interrupted, + sleep_with_interrupt, +) +from .client import _diagnose_connectivity +from .common_exceptions import ApiServerError, LocalNetworkError, ProcessingInterrupted +from .conversions import bytesio_to_image_tensor + +_RETRY_STATUS = {408, 429, 500, 502, 503, 504} + + +async def download_url_to_bytesio( + url: str, + dest: Optional[Union[BytesIO, IO[bytes], str, Path]], + *, + timeout: Optional[float] = None, + max_retries: int = 3, + retry_delay: float = 1.0, + retry_backoff: float = 2.0, + cls: type[COMFY_IO.ComfyNode] = None, +) -> None: + """Stream-download a URL to `dest`. + + `dest` must be one of: + - a BytesIO (rewound to 0 after write), + - a file-like object opened in binary write mode (must implement .write()), + - a filesystem path (str | pathlib.Path), which will be opened with 'wb'. + + If `url` starts with `/proxy/`, `cls` must be provided so the URL can be expanded + to an absolute URL and authentication headers can be applied. + + Raises: + ProcessingInterrupted, LocalNetworkError, ApiServerError, Exception (HTTP and other errors) + """ + if not isinstance(dest, (str, Path)) and not hasattr(dest, "write"): + raise ValueError("dest must be a path (str|Path) or a binary-writable object providing .write().") + + attempt = 0 + delay = retry_delay + headers: dict[str, str] = {} + + parsed_url = urlparse(url) + if not parsed_url.scheme and not parsed_url.netloc: # is URL relative? + if cls is None: + raise ValueError("For relative 'cloud' paths, the `cls` parameter is required.") + url = urljoin(default_base_url().rstrip("/") + "/", url.lstrip("/")) + headers = get_auth_header(cls) + + while True: + attempt += 1 + op_id = _generate_operation_id("GET", url, attempt) + timeout_cfg = aiohttp.ClientTimeout(total=timeout) + + is_path_sink = isinstance(dest, (str, Path)) + fhandle = None + session: Optional[aiohttp.ClientSession] = None + stop_evt: Optional[asyncio.Event] = None + monitor_task: Optional[asyncio.Task] = None + req_task: Optional[asyncio.Task] = None + + try: + with contextlib.suppress(Exception): + request_logger.log_request_response(operation_id=op_id, request_method="GET", request_url=url) + + session = aiohttp.ClientSession(timeout=timeout_cfg) + stop_evt = asyncio.Event() + + async def _monitor(): + try: + while not stop_evt.is_set(): + if is_processing_interrupted(): + return + await asyncio.sleep(1.0) + except asyncio.CancelledError: + return + + monitor_task = asyncio.create_task(_monitor()) + + req_task = asyncio.create_task(session.get(url, headers=headers)) + done, pending = await asyncio.wait({req_task, monitor_task}, return_when=asyncio.FIRST_COMPLETED) + + if monitor_task in done and req_task in pending: + req_task.cancel() + with contextlib.suppress(Exception): + await req_task + raise ProcessingInterrupted("Task cancelled") + + try: + resp = await req_task + except asyncio.CancelledError: + raise ProcessingInterrupted("Task cancelled") from None + + async with resp: + if resp.status >= 400: + with contextlib.suppress(Exception): + try: + body = await resp.json() + except (ContentTypeError, ValueError): + text = await resp.text() + body = text if len(text) <= 4096 else f"[text {len(text)} bytes]" + request_logger.log_request_response( + operation_id=op_id, + request_method="GET", + request_url=url, + response_status_code=resp.status, + response_headers=dict(resp.headers), + response_content=body, + error_message=f"HTTP {resp.status}", + ) + + if resp.status in _RETRY_STATUS and attempt <= max_retries: + await sleep_with_interrupt(delay, cls, None, None, None) + delay *= retry_backoff + continue + raise Exception(f"Failed to download (HTTP {resp.status}).") + + if is_path_sink: + p = Path(str(dest)) + with contextlib.suppress(Exception): + p.parent.mkdir(parents=True, exist_ok=True) + fhandle = open(p, "wb") + sink = fhandle + else: + sink = dest # BytesIO or file-like + + written = 0 + while True: + try: + chunk = await asyncio.wait_for(resp.content.read(1024 * 1024), timeout=1.0) + except asyncio.TimeoutError: + chunk = b"" + except asyncio.CancelledError: + raise ProcessingInterrupted("Task cancelled") from None + + if is_processing_interrupted(): + raise ProcessingInterrupted("Task cancelled") + + if not chunk: + if resp.content.at_eof(): + break + continue + + sink.write(chunk) + written += len(chunk) + + if isinstance(dest, BytesIO): + with contextlib.suppress(Exception): + dest.seek(0) + + with contextlib.suppress(Exception): + request_logger.log_request_response( + operation_id=op_id, + request_method="GET", + request_url=url, + response_status_code=resp.status, + response_headers=dict(resp.headers), + response_content=f"[streamed {written} bytes to dest]", + ) + return + except asyncio.CancelledError: + raise ProcessingInterrupted("Task cancelled") from None + except (ClientError, asyncio.TimeoutError) as e: + if attempt <= max_retries: + with contextlib.suppress(Exception): + request_logger.log_request_response( + operation_id=op_id, + request_method="GET", + request_url=url, + error_message=f"{type(e).__name__}: {str(e)} (will retry)", + ) + await sleep_with_interrupt(delay, cls, None, None, None) + delay *= retry_backoff + continue + + diag = await _diagnose_connectivity() + if diag.get("is_local_issue"): + raise LocalNetworkError( + "Unable to connect to the network. Please check your internet connection and try again." + ) from e + raise ApiServerError("The remote service appears unreachable at this time.") from e + finally: + if stop_evt is not None: + stop_evt.set() + if monitor_task: + monitor_task.cancel() + with contextlib.suppress(Exception): + await monitor_task + if req_task and not req_task.done(): + req_task.cancel() + with contextlib.suppress(Exception): + await req_task + if session: + with contextlib.suppress(Exception): + await session.close() + if fhandle: + with contextlib.suppress(Exception): + fhandle.flush() + fhandle.close() + + +async def download_url_to_image_tensor( + url: str, + *, + timeout: float = None, + cls: type[COMFY_IO.ComfyNode] = None, +) -> torch.Tensor: + """Downloads an image from a URL and returns a [B, H, W, C] tensor.""" + result = BytesIO() + await download_url_to_bytesio(url, result, timeout=timeout, cls=cls) + return bytesio_to_image_tensor(result) + + +async def download_url_to_video_output( + video_url: str, + *, + timeout: float = None, + cls: type[COMFY_IO.ComfyNode] = None, +) -> VideoFromFile: + """Downloads a video from a URL and returns a `VIDEO` output.""" + result = BytesIO() + await download_url_to_bytesio(video_url, result, timeout=timeout, cls=cls) + return VideoFromFile(result) + + +def _generate_operation_id(method: str, url: str, attempt: int) -> str: + try: + parsed = urlparse(url) + slug = (parsed.path.rsplit("/", 1)[-1] or parsed.netloc or "download").strip("/").replace("/", "_") + except Exception: + slug = "download" + return f"{method}_{slug}_try{attempt}_{uuid.uuid4().hex[:8]}" diff --git a/comfy_api_nodes/util/upload_helpers.py b/comfy_api_nodes/util/upload_helpers.py new file mode 100644 index 000000000..a345d451d --- /dev/null +++ b/comfy_api_nodes/util/upload_helpers.py @@ -0,0 +1,338 @@ +import asyncio +import contextlib +import logging +import time +import uuid +from io import BytesIO +from typing import Optional, Union +from urllib.parse import urlparse + +import aiohttp +import torch +from pydantic import BaseModel, Field + +from comfy_api.latest import IO, Input +from comfy_api.util import VideoCodec, VideoContainer +from comfy_api_nodes.apis import request_logger + +from ._helpers import is_processing_interrupted, sleep_with_interrupt +from .client import ( + ApiEndpoint, + _diagnose_connectivity, + _display_time_progress, + sync_op, +) +from .common_exceptions import ApiServerError, LocalNetworkError, ProcessingInterrupted +from .conversions import ( + audio_ndarray_to_bytesio, + audio_tensor_to_contiguous_ndarray, + tensor_to_bytesio, +) + + +class UploadRequest(BaseModel): + file_name: str = Field(..., description="Filename to upload") + content_type: Optional[str] = Field( + None, + description="Mime type of the file. For example: image/png, image/jpeg, video/mp4, etc.", + ) + + +class UploadResponse(BaseModel): + download_url: str = Field(..., description="URL to GET uploaded file") + upload_url: str = Field(..., description="URL to PUT file to upload") + + +async def upload_images_to_comfyapi( + cls: type[IO.ComfyNode], + image: torch.Tensor, + *, + max_images: int = 8, + mime_type: Optional[str] = None, + wait_label: Optional[str] = "Uploading", +) -> list[str]: + """ + Uploads images to ComfyUI API and returns download URLs. + To upload multiple images, stack them in the batch dimension first. + """ + # if batch, try to upload each file if max_images is greater than 0 + download_urls: list[str] = [] + is_batch = len(image.shape) > 3 + batch_len = image.shape[0] if is_batch else 1 + + for idx in range(min(batch_len, max_images)): + tensor = image[idx] if is_batch else image + img_io = tensor_to_bytesio(tensor, mime_type=mime_type) + url = await upload_file_to_comfyapi(cls, img_io, img_io.name, mime_type, wait_label) + download_urls.append(url) + return download_urls + + +async def upload_audio_to_comfyapi( + cls: type[IO.ComfyNode], + audio: Input.Audio, + *, + container_format: str = "mp4", + codec_name: str = "aac", + mime_type: str = "audio/mp4", + filename: str = "uploaded_audio.mp4", +) -> str: + """ + Uploads a single audio input to ComfyUI API and returns its download URL. + Encodes the raw waveform into the specified format before uploading. + """ + sample_rate: int = audio["sample_rate"] + waveform: torch.Tensor = audio["waveform"] + audio_data_np = audio_tensor_to_contiguous_ndarray(waveform) + audio_bytes_io = audio_ndarray_to_bytesio(audio_data_np, sample_rate, container_format, codec_name) + return await upload_file_to_comfyapi(cls, audio_bytes_io, filename, mime_type) + + +async def upload_video_to_comfyapi( + cls: type[IO.ComfyNode], + video: Input.Video, + *, + container: VideoContainer = VideoContainer.MP4, + codec: VideoCodec = VideoCodec.H264, + max_duration: Optional[int] = None, +) -> str: + """ + Uploads a single video to ComfyUI API and returns its download URL. + Uses the specified container and codec for saving the video before upload. + """ + if max_duration is not None: + try: + actual_duration = video.get_duration() + if actual_duration > max_duration: + raise ValueError( + f"Video duration ({actual_duration:.2f}s) exceeds the maximum allowed ({max_duration}s)." + ) + except Exception as e: + logging.error("Error getting video duration: %s", str(e)) + raise ValueError(f"Could not verify video duration from source: {e}") from e + + upload_mime_type = f"video/{container.value.lower()}" + filename = f"uploaded_video.{container.value.lower()}" + + # Convert VideoInput to BytesIO using specified container/codec + video_bytes_io = BytesIO() + video.save_to(video_bytes_io, format=container, codec=codec) + video_bytes_io.seek(0) + + return await upload_file_to_comfyapi(cls, video_bytes_io, filename, upload_mime_type) + + +async def upload_file_to_comfyapi( + cls: type[IO.ComfyNode], + file_bytes_io: BytesIO, + filename: str, + upload_mime_type: Optional[str], + wait_label: Optional[str] = "Uploading", +) -> str: + """Uploads a single file to ComfyUI API and returns its download URL.""" + if upload_mime_type is None: + request_object = UploadRequest(file_name=filename) + else: + request_object = UploadRequest(file_name=filename, content_type=upload_mime_type) + create_resp = await sync_op( + cls, + endpoint=ApiEndpoint(path="/customers/storage", method="POST"), + data=request_object, + response_model=UploadResponse, + final_label_on_success=None, + monitor_progress=False, + ) + await upload_file( + cls, + create_resp.upload_url, + file_bytes_io, + content_type=upload_mime_type, + wait_label=wait_label, + ) + return create_resp.download_url + + +async def upload_file( + cls: type[IO.ComfyNode], + upload_url: str, + file: Union[BytesIO, str], + *, + content_type: Optional[str] = None, + max_retries: int = 3, + retry_delay: float = 1.0, + retry_backoff: float = 2.0, + wait_label: Optional[str] = None, +) -> None: + """ + Upload a file to a signed URL (e.g., S3 pre-signed PUT) with retries, Comfy progress display, and interruption. + + Args: + cls: Node class (provides auth context + UI progress hooks). + upload_url: Pre-signed PUT URL. + file: BytesIO or path string. + content_type: Explicit MIME type. If None, we *suppress* Content-Type. + max_retries: Maximum retry attempts. + retry_delay: Initial delay in seconds. + retry_backoff: Exponential backoff factor. + wait_label: Progress label shown in Comfy UI. + + Raises: + ProcessingInterrupted, LocalNetworkError, ApiServerError, Exception + """ + if isinstance(file, BytesIO): + with contextlib.suppress(Exception): + file.seek(0) + data = file.read() + elif isinstance(file, str): + with open(file, "rb") as f: + data = f.read() + else: + raise ValueError("file must be a BytesIO or a filesystem path string") + + headers: dict[str, str] = {} + skip_auto_headers: set[str] = set() + if content_type: + headers["Content-Type"] = content_type + else: + skip_auto_headers.add("Content-Type") # Don't let aiohttp add Content-Type, it can break the signed request + + attempt = 0 + delay = retry_delay + start_ts = time.monotonic() + op_uuid = uuid.uuid4().hex[:8] + while True: + attempt += 1 + operation_id = _generate_operation_id("PUT", upload_url, attempt, op_uuid) + timeout = aiohttp.ClientTimeout(total=None) + stop_evt = asyncio.Event() + + async def _monitor(): + try: + while not stop_evt.is_set(): + if is_processing_interrupted(): + return + if wait_label: + _display_time_progress(cls, wait_label, int(time.monotonic() - start_ts), None) + await asyncio.sleep(1.0) + except asyncio.CancelledError: + return + + monitor_task = asyncio.create_task(_monitor()) + sess: Optional[aiohttp.ClientSession] = None + try: + try: + request_logger.log_request_response( + operation_id=operation_id, + request_method="PUT", + request_url=upload_url, + request_headers=headers or None, + request_params=None, + request_data=f"[File data {len(data)} bytes]", + ) + except Exception as e: + logging.debug("[DEBUG] upload request logging failed: %s", e) + + sess = aiohttp.ClientSession(timeout=timeout) + req = sess.put(upload_url, data=data, headers=headers, skip_auto_headers=skip_auto_headers) + req_task = asyncio.create_task(req) + + done, pending = await asyncio.wait({req_task, monitor_task}, return_when=asyncio.FIRST_COMPLETED) + + if monitor_task in done and req_task in pending: + req_task.cancel() + raise ProcessingInterrupted("Upload cancelled") + + try: + resp = await req_task + except asyncio.CancelledError: + raise ProcessingInterrupted("Upload cancelled") from None + + async with resp: + if resp.status >= 400: + with contextlib.suppress(Exception): + try: + body = await resp.json() + except Exception: + body = await resp.text() + msg = f"Upload failed with status {resp.status}" + request_logger.log_request_response( + operation_id=operation_id, + request_method="PUT", + request_url=upload_url, + response_status_code=resp.status, + response_headers=dict(resp.headers), + response_content=body, + error_message=msg, + ) + if resp.status in {408, 429, 500, 502, 503, 504} and attempt <= max_retries: + await sleep_with_interrupt( + delay, + cls, + wait_label, + start_ts, + None, + display_callback=_display_time_progress if wait_label else None, + ) + delay *= retry_backoff + continue + raise Exception(f"Failed to upload (HTTP {resp.status}).") + try: + request_logger.log_request_response( + operation_id=operation_id, + request_method="PUT", + request_url=upload_url, + response_status_code=resp.status, + response_headers=dict(resp.headers), + response_content="File uploaded successfully.", + ) + except Exception as e: + logging.debug("[DEBUG] upload response logging failed: %s", e) + return + except asyncio.CancelledError: + raise ProcessingInterrupted("Task cancelled") from None + except (aiohttp.ClientError, asyncio.TimeoutError) as e: + if attempt <= max_retries: + with contextlib.suppress(Exception): + request_logger.log_request_response( + operation_id=operation_id, + request_method="PUT", + request_url=upload_url, + request_headers=headers or None, + request_data=f"[File data {len(data)} bytes]", + error_message=f"{type(e).__name__}: {str(e)} (will retry)", + ) + await sleep_with_interrupt( + delay, + cls, + wait_label, + start_ts, + None, + display_callback=_display_time_progress if wait_label else None, + ) + delay *= retry_backoff + continue + + diag = await _diagnose_connectivity() + if diag.get("is_local_issue"): + raise LocalNetworkError( + "Unable to connect to the network. Please check your internet connection and try again." + ) from e + raise ApiServerError("The API service appears unreachable at this time.") from e + finally: + stop_evt.set() + if monitor_task: + monitor_task.cancel() + with contextlib.suppress(Exception): + await monitor_task + if sess: + with contextlib.suppress(Exception): + await sess.close() + + +def _generate_operation_id(method: str, url: str, attempt: int, op_uuid: str) -> str: + try: + parsed = urlparse(url) + slug = (parsed.path.rsplit("/", 1)[-1] or parsed.netloc or "upload").strip("/").replace("/", "_") + except Exception: + slug = "upload" + return f"{method}_{slug}_{op_uuid}_try{attempt}" diff --git a/comfy_api_nodes/util/validation_utils.py b/comfy_api_nodes/util/validation_utils.py index ca913e9b3..22da05bc1 100644 --- a/comfy_api_nodes/util/validation_utils.py +++ b/comfy_api_nodes/util/validation_utils.py @@ -2,6 +2,8 @@ import logging from typing import Optional import torch + +from comfy_api.input.video_types import VideoInput from comfy_api.latest import Input @@ -28,9 +30,7 @@ def validate_image_dimensions( if max_width is not None and width > max_width: raise ValueError(f"Image width must be at most {max_width}px, got {width}px") if min_height is not None and height < min_height: - raise ValueError( - f"Image height must be at least {min_height}px, got {height}px" - ) + raise ValueError(f"Image height must be at least {min_height}px, got {height}px") if max_height is not None and height > max_height: raise ValueError(f"Image height must be at most {max_height}px, got {height}px") @@ -44,13 +44,9 @@ def validate_image_aspect_ratio( aspect_ratio = width / height if min_aspect_ratio is not None and aspect_ratio < min_aspect_ratio: - raise ValueError( - f"Image aspect ratio must be at least {min_aspect_ratio}, got {aspect_ratio}" - ) + raise ValueError(f"Image aspect ratio must be at least {min_aspect_ratio}, got {aspect_ratio}") if max_aspect_ratio is not None and aspect_ratio > max_aspect_ratio: - raise ValueError( - f"Image aspect ratio must be at most {max_aspect_ratio}, got {aspect_ratio}" - ) + raise ValueError(f"Image aspect ratio must be at most {max_aspect_ratio}, got {aspect_ratio}") def validate_image_aspect_ratio_range( @@ -58,7 +54,7 @@ def validate_image_aspect_ratio_range( min_ratio: tuple[float, float], # e.g. (1, 4) max_ratio: tuple[float, float], # e.g. (4, 1) *, - strict: bool = True, # True -> (min, max); False -> [min, max] + strict: bool = True, # True -> (min, max); False -> [min, max] ) -> float: a1, b1 = min_ratio a2, b2 = max_ratio @@ -85,7 +81,7 @@ def validate_aspect_ratio_closeness( min_rel: float, max_rel: float, *, - strict: bool = False, # True => exclusive, False => inclusive + strict: bool = False, # True => exclusive, False => inclusive ) -> None: w1, h1 = get_image_dimensions(start_img) w2, h2 = get_image_dimensions(end_img) @@ -118,9 +114,7 @@ def validate_video_dimensions( if max_width is not None and width > max_width: raise ValueError(f"Video width must be at most {max_width}px, got {width}px") if min_height is not None and height < min_height: - raise ValueError( - f"Video height must be at least {min_height}px, got {height}px" - ) + raise ValueError(f"Video height must be at least {min_height}px, got {height}px") if max_height is not None and height > max_height: raise ValueError(f"Video height must be at most {max_height}px, got {height}px") @@ -138,13 +132,9 @@ def validate_video_duration( epsilon = 0.0001 if min_duration is not None and min_duration - epsilon > duration: - raise ValueError( - f"Video duration must be at least {min_duration}s, got {duration}s" - ) + raise ValueError(f"Video duration must be at least {min_duration}s, got {duration}s") if max_duration is not None and duration > max_duration + epsilon: - raise ValueError( - f"Video duration must be at most {max_duration}s, got {duration}s" - ) + raise ValueError(f"Video duration must be at most {max_duration}s, got {duration}s") def get_number_of_images(images): @@ -165,3 +155,31 @@ def validate_audio_duration( raise ValueError(f"Audio duration must be at least {min_duration}s, got {dur + eps:.2f}s") if max_duration is not None and dur - eps > max_duration: raise ValueError(f"Audio duration must be at most {max_duration}s, got {dur - eps:.2f}s") + + +def validate_string( + string: str, + strip_whitespace=True, + field_name="prompt", + min_length=None, + max_length=None, +): + if string is None: + raise Exception(f"Field '{field_name}' cannot be empty.") + if strip_whitespace: + string = string.strip() + if min_length and len(string) < min_length: + raise Exception( + f"Field '{field_name}' cannot be shorter than {min_length} characters; was {len(string)} characters long." + ) + if max_length and len(string) > max_length: + raise Exception( + f" Field '{field_name} cannot be longer than {max_length} characters; was {len(string)} characters long." + ) + + +def validate_container_format_is_mp4(video: VideoInput) -> None: + """Validates video container format is MP4.""" + container_format = video.get_container_format() + if container_format not in ["mp4", "mov,mp4,m4a,3gp,3g2,mj2"]: + raise ValueError(f"Only MP4 container format supported. Got: {container_format}") diff --git a/pyproject.toml b/pyproject.toml index 0c6b23a25..fcc4854a5 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -50,6 +50,8 @@ messages_control.disable = [ "too-many-branches", "too-many-locals", "too-many-arguments", + "too-many-return-statements", + "too-many-nested-blocks", "duplicate-code", "abstract-method", "superfluous-parens", From dd5af0c5871376c377b2e30f9725b67a768eea6f Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Sat, 25 Oct 2025 01:48:34 +0300 Subject: [PATCH 027/647] convert Tripo API nodes to V3 schema (#10469) --- comfy_api_nodes/apis/tripo_api.py | 15 +- comfy_api_nodes/nodes_tripo.py | 892 ++++++++++++----------- comfy_api_nodes/util/__init__.py | 2 + comfy_api_nodes/util/download_helpers.py | 12 + 4 files changed, 503 insertions(+), 418 deletions(-) diff --git a/comfy_api_nodes/apis/tripo_api.py b/comfy_api_nodes/apis/tripo_api.py index 9f43d4d09..713260e2a 100644 --- a/comfy_api_nodes/apis/tripo_api.py +++ b/comfy_api_nodes/apis/tripo_api.py @@ -1,13 +1,20 @@ from __future__ import annotations -from comfy_api_nodes.apis import ( - TripoModelVersion, - TripoTextureQuality, -) from enum import Enum from typing import Optional, List, Dict, Any, Union from pydantic import BaseModel, Field, RootModel +class TripoModelVersion(str, Enum): + v2_5_20250123 = 'v2.5-20250123' + v2_0_20240919 = 'v2.0-20240919' + v1_4_20240625 = 'v1.4-20240625' + + +class TripoTextureQuality(str, Enum): + standard = 'standard' + detailed = 'detailed' + + class TripoStyle(str, Enum): PERSON_TO_CARTOON = "person:person2cartoon" ANIMAL_VENOM = "animal:venom" diff --git a/comfy_api_nodes/nodes_tripo.py b/comfy_api_nodes/nodes_tripo.py index d08cf9007..697100ff2 100644 --- a/comfy_api_nodes/nodes_tripo.py +++ b/comfy_api_nodes/nodes_tripo.py @@ -1,46 +1,39 @@ import os -from folder_paths import get_output_directory -from comfy_api_nodes.mapper_utils import model_field_to_node_input -from comfy.comfy_types.node_typing import IO -from comfy_api_nodes.apis import ( - TripoOrientation, - TripoModelVersion, -) +from typing import Optional + +import torch +from typing_extensions import override + +from comfy_api.latest import IO, ComfyExtension from comfy_api_nodes.apis.tripo_api import ( - TripoTaskType, - TripoStyle, - TripoFileReference, + TripoAnimateRetargetRequest, + TripoAnimateRigRequest, + TripoConvertModelRequest, TripoFileEmptyReference, - TripoUrlReference, + TripoFileReference, + TripoImageToModelRequest, + TripoModelVersion, + TripoMultiviewToModelRequest, + TripoOrientation, + TripoRefineModelRequest, + TripoStyle, TripoTaskResponse, TripoTaskStatus, + TripoTaskType, TripoTextToModelRequest, - TripoImageToModelRequest, - TripoMultiviewToModelRequest, TripoTextureModelRequest, - TripoRefineModelRequest, - TripoAnimateRigRequest, - TripoAnimateRetargetRequest, - TripoConvertModelRequest, + TripoUrlReference, ) - -from comfy_api_nodes.apis.client import ( +from comfy_api_nodes.util import ( ApiEndpoint, - HttpMethod, - SynchronousOperation, - PollingOperation, - EmptyRequest, -) -from comfy_api_nodes.apinode_utils import ( + download_url_as_bytesio, + poll_op, + sync_op, upload_images_to_comfyapi, - download_url_to_bytesio, ) +from folder_paths import get_output_directory -async def upload_image_to_tripo(image, **kwargs): - urls = await upload_images_to_comfyapi(image, max_images=1, auth_kwargs=kwargs) - return TripoFileReference(TripoUrlReference(url=urls[0], type="jpeg")) - def get_model_url_from_response(response: TripoTaskResponse) -> str: if response.data is not None: for key in ["pbr_model", "model", "base_model"]: @@ -50,20 +43,18 @@ def get_model_url_from_response(response: TripoTaskResponse) -> str: async def poll_until_finished( - kwargs: dict[str, str], + node_cls: type[IO.ComfyNode], response: TripoTaskResponse, -) -> tuple[str, str]: + average_duration: Optional[int] = None, +) -> IO.NodeOutput: """Polls the Tripo API endpoint until the task reaches a terminal state, then returns the response.""" if response.code != 0: raise RuntimeError(f"Failed to generate mesh: {response.error}") task_id = response.data.task_id - response_poll = await PollingOperation( - poll_endpoint=ApiEndpoint( - path=f"/proxy/tripo/v2/openapi/task/{task_id}", - method=HttpMethod.GET, - request_model=EmptyRequest, - response_model=TripoTaskResponse, - ), + response_poll = await poll_op( + node_cls, + poll_endpoint=ApiEndpoint(path=f"/proxy/tripo/v2/openapi/task/{task_id}"), + response_model=TripoTaskResponse, completed_statuses=[TripoTaskStatus.SUCCESS], failed_statuses=[ TripoTaskStatus.FAILED, @@ -73,72 +64,84 @@ async def poll_until_finished( TripoTaskStatus.EXPIRED, ], status_extractor=lambda x: x.data.status, - auth_kwargs=kwargs, - node_id=kwargs["unique_id"], - result_url_extractor=get_model_url_from_response, progress_extractor=lambda x: x.data.progress, - ).execute() + estimated_duration=average_duration, + ) if response_poll.data.status == TripoTaskStatus.SUCCESS: url = get_model_url_from_response(response_poll) - bytesio = await download_url_to_bytesio(url) + bytesio = await download_url_as_bytesio(url) # Save the downloaded model file model_file = f"tripo_model_{task_id}.glb" with open(os.path.join(get_output_directory(), model_file), "wb") as f: f.write(bytesio.getvalue()) - return model_file, task_id + return IO.NodeOutput(model_file, task_id) raise RuntimeError(f"Failed to generate mesh: {response_poll}") -class TripoTextToModelNode: +class TripoTextToModelNode(IO.ComfyNode): """ Generates 3D models synchronously based on a text prompt using Tripo's API. """ - AVERAGE_DURATION = 80 + @classmethod - def INPUT_TYPES(s): - return { - "required": { - "prompt": ("STRING", {"multiline": True}), - }, - "optional": { - "negative_prompt": ("STRING", {"multiline": True}), - "model_version": model_field_to_node_input(IO.COMBO, TripoTextToModelRequest, "model_version", enum_type=TripoModelVersion), - "style": model_field_to_node_input(IO.COMBO, TripoTextToModelRequest, "style", enum_type=TripoStyle, default="None"), - "texture": ("BOOLEAN", {"default": True}), - "pbr": ("BOOLEAN", {"default": True}), - "image_seed": ("INT", {"default": 42}), - "model_seed": ("INT", {"default": 42}), - "texture_seed": ("INT", {"default": 42}), - "texture_quality": (["standard", "detailed"], {"default": "standard"}), - "face_limit": ("INT", {"min": -1, "max": 500000, "default": -1}), - "quad": ("BOOLEAN", {"default": False}) - }, - "hidden": { - "auth_token": "AUTH_TOKEN_COMFY_ORG", - "comfy_api_key": "API_KEY_COMFY_ORG", - "unique_id": "UNIQUE_ID", - }, - } + def define_schema(cls): + return IO.Schema( + node_id="TripoTextToModelNode", + display_name="Tripo: Text to Model", + category="api node/3d/Tripo", + inputs=[ + IO.String.Input("prompt", multiline=True), + IO.String.Input("negative_prompt", multiline=True, optional=True), + IO.Combo.Input( + "model_version", options=TripoModelVersion, default=TripoModelVersion.v2_5_20250123, optional=True + ), + IO.Combo.Input("style", options=TripoStyle, default="None", optional=True), + IO.Boolean.Input("texture", default=True, optional=True), + IO.Boolean.Input("pbr", default=True, optional=True), + IO.Int.Input("image_seed", default=42, optional=True), + IO.Int.Input("model_seed", default=42, optional=True), + IO.Int.Input("texture_seed", default=42, optional=True), + IO.Combo.Input("texture_quality", default="standard", options=["standard", "detailed"], optional=True), + IO.Int.Input("face_limit", default=-1, min=-1, max=500000, optional=True), + IO.Boolean.Input("quad", default=False, optional=True), + ], + outputs=[ + IO.String.Output(display_name="model_file"), + IO.Custom("MODEL_TASK_ID").Output(display_name="model task_id"), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + is_output_node=True, + ) - RETURN_TYPES = ("STRING", "MODEL_TASK_ID",) - RETURN_NAMES = ("model_file", "model task_id") - FUNCTION = "generate_mesh" - CATEGORY = "api node/3d/Tripo" - API_NODE = True - OUTPUT_NODE = True - - async def generate_mesh(self, prompt, negative_prompt=None, model_version=None, style=None, texture=None, pbr=None, image_seed=None, model_seed=None, texture_seed=None, texture_quality=None, face_limit=None, quad=None, **kwargs): + @classmethod + async def execute( + cls, + prompt: str, + negative_prompt: Optional[str] = None, + model_version=None, + style: Optional[str] = None, + texture: Optional[bool] = None, + pbr: Optional[bool] = None, + image_seed: Optional[int] = None, + model_seed: Optional[int] = None, + texture_seed: Optional[int] = None, + texture_quality: Optional[str] = None, + face_limit: Optional[int] = None, + quad: Optional[bool] = None, + ) -> IO.NodeOutput: style_enum = None if style == "None" else style if not prompt: raise RuntimeError("Prompt is required") - response = await SynchronousOperation( - endpoint=ApiEndpoint( - path="/proxy/tripo/v2/openapi/task", - method=HttpMethod.POST, - request_model=TripoTextToModelRequest, - response_model=TripoTaskResponse, - ), - request=TripoTextToModelRequest( + response = await sync_op( + cls, + endpoint=ApiEndpoint(path="/proxy/tripo/v2/openapi/task", method="POST"), + response_model=TripoTaskResponse, + data=TripoTextToModelRequest( type=TripoTaskType.TEXT_TO_MODEL, prompt=prompt, negative_prompt=negative_prompt if negative_prompt else None, @@ -152,64 +155,89 @@ class TripoTextToModelNode: texture_quality=texture_quality, face_limit=face_limit, auto_size=True, - quad=quad + quad=quad, ), - auth_kwargs=kwargs, - ).execute() - return await poll_until_finished(kwargs, response) + ) + return await poll_until_finished(cls, response, average_duration=80) -class TripoImageToModelNode: +class TripoImageToModelNode(IO.ComfyNode): """ Generates 3D models synchronously based on a single image using Tripo's API. """ - AVERAGE_DURATION = 80 + @classmethod - def INPUT_TYPES(s): - return { - "required": { - "image": ("IMAGE",), - }, - "optional": { - "model_version": model_field_to_node_input(IO.COMBO, TripoImageToModelRequest, "model_version", enum_type=TripoModelVersion), - "style": model_field_to_node_input(IO.COMBO, TripoTextToModelRequest, "style", enum_type=TripoStyle, default="None"), - "texture": ("BOOLEAN", {"default": True}), - "pbr": ("BOOLEAN", {"default": True}), - "model_seed": ("INT", {"default": 42}), - "orientation": model_field_to_node_input(IO.COMBO, TripoImageToModelRequest, "orientation", enum_type=TripoOrientation), - "texture_seed": ("INT", {"default": 42}), - "texture_quality": (["standard", "detailed"], {"default": "standard"}), - "texture_alignment": (["original_image", "geometry"], {"default": "original_image"}), - "face_limit": ("INT", {"min": -1, "max": 500000, "default": -1}), - "quad": ("BOOLEAN", {"default": False}) - }, - "hidden": { - "auth_token": "AUTH_TOKEN_COMFY_ORG", - "comfy_api_key": "API_KEY_COMFY_ORG", - "unique_id": "UNIQUE_ID", - }, - } + def define_schema(cls): + return IO.Schema( + node_id="TripoImageToModelNode", + display_name="Tripo: Image to Model", + category="api node/3d/Tripo", + inputs=[ + IO.Image.Input("image"), + IO.Combo.Input( + "model_version", + options=TripoModelVersion, + tooltip="The model version to use for generation", + optional=True, + ), + IO.Combo.Input("style", options=TripoStyle, default="None", optional=True), + IO.Boolean.Input("texture", default=True, optional=True), + IO.Boolean.Input("pbr", default=True, optional=True), + IO.Int.Input("model_seed", default=42, optional=True), + IO.Combo.Input( + "orientation", options=TripoOrientation, default=TripoOrientation.DEFAULT, optional=True + ), + IO.Int.Input("texture_seed", default=42, optional=True), + IO.Combo.Input("texture_quality", default="standard", options=["standard", "detailed"], optional=True), + IO.Combo.Input( + "texture_alignment", default="original_image", options=["original_image", "geometry"], optional=True + ), + IO.Int.Input("face_limit", default=-1, min=-1, max=500000, optional=True), + IO.Boolean.Input("quad", default=False, optional=True), + ], + outputs=[ + IO.String.Output(display_name="model_file"), + IO.Custom("MODEL_TASK_ID").Output(display_name="model task_id"), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + is_output_node=True, + ) - RETURN_TYPES = ("STRING", "MODEL_TASK_ID",) - RETURN_NAMES = ("model_file", "model task_id") - FUNCTION = "generate_mesh" - CATEGORY = "api node/3d/Tripo" - API_NODE = True - OUTPUT_NODE = True - - async def generate_mesh(self, image, model_version=None, style=None, texture=None, pbr=None, model_seed=None, orientation=None, texture_alignment=None, texture_seed=None, texture_quality=None, face_limit=None, quad=None, **kwargs): + @classmethod + async def execute( + cls, + image: torch.Tensor, + model_version: Optional[str] = None, + style: Optional[str] = None, + texture: Optional[bool] = None, + pbr: Optional[bool] = None, + model_seed: Optional[int] = None, + orientation=None, + texture_seed: Optional[int] = None, + texture_quality: Optional[str] = None, + texture_alignment: Optional[str] = None, + face_limit: Optional[int] = None, + quad: Optional[bool] = None, + ) -> IO.NodeOutput: style_enum = None if style == "None" else style if image is None: raise RuntimeError("Image is required") - tripo_file = await upload_image_to_tripo(image, **kwargs) - response = await SynchronousOperation( - endpoint=ApiEndpoint( - path="/proxy/tripo/v2/openapi/task", - method=HttpMethod.POST, - request_model=TripoImageToModelRequest, - response_model=TripoTaskResponse, - ), - request=TripoImageToModelRequest( + tripo_file = TripoFileReference( + root=TripoUrlReference( + url=(await upload_images_to_comfyapi(cls, image, max_images=1))[0], + type="jpeg", + ) + ) + response = await sync_op( + cls, + endpoint=ApiEndpoint(path="/proxy/tripo/v2/openapi/task", method="POST"), + response_model=TripoTaskResponse, + data=TripoImageToModelRequest( type=TripoTaskType.IMAGE_TO_MODEL, file=tripo_file, model_version=model_version, @@ -223,80 +251,105 @@ class TripoImageToModelNode: texture_quality=texture_quality, face_limit=face_limit, auto_size=True, - quad=quad + quad=quad, ), - auth_kwargs=kwargs, - ).execute() - return await poll_until_finished(kwargs, response) + ) + return await poll_until_finished(cls, response, average_duration=80) -class TripoMultiviewToModelNode: +class TripoMultiviewToModelNode(IO.ComfyNode): """ Generates 3D models synchronously based on up to four images (front, left, back, right) using Tripo's API. """ - AVERAGE_DURATION = 80 + @classmethod - def INPUT_TYPES(s): - return { - "required": { - "image": ("IMAGE",), - }, - "optional": { - "image_left": ("IMAGE",), - "image_back": ("IMAGE",), - "image_right": ("IMAGE",), - "model_version": model_field_to_node_input(IO.COMBO, TripoMultiviewToModelRequest, "model_version", enum_type=TripoModelVersion), - "orientation": model_field_to_node_input(IO.COMBO, TripoImageToModelRequest, "orientation", enum_type=TripoOrientation), - "texture": ("BOOLEAN", {"default": True}), - "pbr": ("BOOLEAN", {"default": True}), - "model_seed": ("INT", {"default": 42}), - "texture_seed": ("INT", {"default": 42}), - "texture_quality": (["standard", "detailed"], {"default": "standard"}), - "texture_alignment": (["original_image", "geometry"], {"default": "original_image"}), - "face_limit": ("INT", {"min": -1, "max": 500000, "default": -1}), - "quad": ("BOOLEAN", {"default": False}) - }, - "hidden": { - "auth_token": "AUTH_TOKEN_COMFY_ORG", - "comfy_api_key": "API_KEY_COMFY_ORG", - "unique_id": "UNIQUE_ID", - }, - } + def define_schema(cls): + return IO.Schema( + node_id="TripoMultiviewToModelNode", + display_name="Tripo: Multiview to Model", + category="api node/3d/Tripo", + inputs=[ + IO.Image.Input("image"), + IO.Image.Input("image_left", optional=True), + IO.Image.Input("image_back", optional=True), + IO.Image.Input("image_right", optional=True), + IO.Combo.Input( + "model_version", + options=TripoModelVersion, + optional=True, + tooltip="The model version to use for generation", + ), + IO.Combo.Input( + "orientation", + options=TripoOrientation, + default=TripoOrientation.DEFAULT, + optional=True, + ), + IO.Boolean.Input("texture", default=True, optional=True), + IO.Boolean.Input("pbr", default=True, optional=True), + IO.Int.Input("model_seed", default=42, optional=True), + IO.Int.Input("texture_seed", default=42, optional=True), + IO.Combo.Input("texture_quality", default="standard", options=["standard", "detailed"], optional=True), + IO.Combo.Input( + "texture_alignment", default="original_image", options=["original_image", "geometry"], optional=True + ), + IO.Int.Input("face_limit", default=-1, min=-1, max=500000, optional=True), + IO.Boolean.Input("quad", default=False, optional=True), + ], + outputs=[ + IO.String.Output(display_name="model_file"), + IO.Custom("MODEL_TASK_ID").Output(display_name="model task_id"), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + is_output_node=True, + ) - RETURN_TYPES = ("STRING", "MODEL_TASK_ID",) - RETURN_NAMES = ("model_file", "model task_id") - FUNCTION = "generate_mesh" - CATEGORY = "api node/3d/Tripo" - API_NODE = True - OUTPUT_NODE = True - - async def generate_mesh(self, image, image_left=None, image_back=None, image_right=None, model_version=None, orientation=None, texture=None, pbr=None, model_seed=None, texture_seed=None, texture_quality=None, texture_alignment=None, face_limit=None, quad=None, **kwargs): + @classmethod + async def execute( + cls, + image: torch.Tensor, + image_left: Optional[torch.Tensor] = None, + image_back: Optional[torch.Tensor] = None, + image_right: Optional[torch.Tensor] = None, + model_version: Optional[str] = None, + orientation: Optional[str] = None, + texture: Optional[bool] = None, + pbr: Optional[bool] = None, + model_seed: Optional[int] = None, + texture_seed: Optional[int] = None, + texture_quality: Optional[str] = None, + texture_alignment: Optional[str] = None, + face_limit: Optional[int] = None, + quad: Optional[bool] = None, + ) -> IO.NodeOutput: if image is None: raise RuntimeError("front image for multiview is required") images = [] - image_dict = { - "image": image, - "image_left": image_left, - "image_back": image_back, - "image_right": image_right - } + image_dict = {"image": image, "image_left": image_left, "image_back": image_back, "image_right": image_right} if image_left is None and image_back is None and image_right is None: raise RuntimeError("At least one of left, back, or right image must be provided for multiview") for image_name in ["image", "image_left", "image_back", "image_right"]: image_ = image_dict[image_name] if image_ is not None: - tripo_file = await upload_image_to_tripo(image_, **kwargs) - images.append(tripo_file) + images.append( + TripoFileReference( + root=TripoUrlReference( + url=(await upload_images_to_comfyapi(cls, image_, max_images=1))[0], type="jpeg" + ) + ) + ) else: images.append(TripoFileEmptyReference()) - response = await SynchronousOperation( - endpoint=ApiEndpoint( - path="/proxy/tripo/v2/openapi/task", - method=HttpMethod.POST, - request_model=TripoMultiviewToModelRequest, - response_model=TripoTaskResponse, - ), - request=TripoMultiviewToModelRequest( + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/tripo/v2/openapi/task", method="POST"), + response_model=TripoTaskResponse, + data=TripoMultiviewToModelRequest( type=TripoTaskType.MULTIVIEW_TO_MODEL, files=images, model_version=model_version, @@ -310,272 +363,283 @@ class TripoMultiviewToModelNode: face_limit=face_limit, quad=quad, ), - auth_kwargs=kwargs, - ).execute() - return await poll_until_finished(kwargs, response) + ) + return await poll_until_finished(cls, response, average_duration=80) -class TripoTextureNode: +class TripoTextureNode(IO.ComfyNode): + @classmethod - def INPUT_TYPES(s): - return { - "required": { - "model_task_id": ("MODEL_TASK_ID",), - }, - "optional": { - "texture": ("BOOLEAN", {"default": True}), - "pbr": ("BOOLEAN", {"default": True}), - "texture_seed": ("INT", {"default": 42}), - "texture_quality": (["standard", "detailed"], {"default": "standard"}), - "texture_alignment": (["original_image", "geometry"], {"default": "original_image"}), - }, - "hidden": { - "auth_token": "AUTH_TOKEN_COMFY_ORG", - "comfy_api_key": "API_KEY_COMFY_ORG", - "unique_id": "UNIQUE_ID", - }, - } + def define_schema(cls): + return IO.Schema( + node_id="TripoTextureNode", + display_name="Tripo: Texture model", + category="api node/3d/Tripo", + inputs=[ + IO.Custom("MODEL_TASK_ID").Input("model_task_id"), + IO.Boolean.Input("texture", default=True, optional=True), + IO.Boolean.Input("pbr", default=True, optional=True), + IO.Int.Input("texture_seed", default=42, optional=True), + IO.Combo.Input("texture_quality", default="standard", options=["standard", "detailed"], optional=True), + IO.Combo.Input( + "texture_alignment", default="original_image", options=["original_image", "geometry"], optional=True + ), + ], + outputs=[ + IO.String.Output(display_name="model_file"), + IO.Custom("MODEL_TASK_ID").Output(display_name="model task_id"), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + is_output_node=True, + ) - RETURN_TYPES = ("STRING", "MODEL_TASK_ID",) - RETURN_NAMES = ("model_file", "model task_id") - FUNCTION = "generate_mesh" - CATEGORY = "api node/3d/Tripo" - API_NODE = True - OUTPUT_NODE = True - AVERAGE_DURATION = 80 - - async def generate_mesh(self, model_task_id, texture=None, pbr=None, texture_seed=None, texture_quality=None, texture_alignment=None, **kwargs): - response = await SynchronousOperation( - endpoint=ApiEndpoint( - path="/proxy/tripo/v2/openapi/task", - method=HttpMethod.POST, - request_model=TripoTextureModelRequest, - response_model=TripoTaskResponse, - ), - request=TripoTextureModelRequest( + @classmethod + async def execute( + cls, + model_task_id, + texture: Optional[bool] = None, + pbr: Optional[bool] = None, + texture_seed: Optional[int] = None, + texture_quality: Optional[str] = None, + texture_alignment: Optional[str] = None, + ) -> IO.NodeOutput: + response = await sync_op( + cls, + endpoint=ApiEndpoint(path="/proxy/tripo/v2/openapi/task", method="POST"), + response_model=TripoTaskResponse, + data=TripoTextureModelRequest( original_model_task_id=model_task_id, texture=texture, pbr=pbr, texture_seed=texture_seed, texture_quality=texture_quality, - texture_alignment=texture_alignment + texture_alignment=texture_alignment, ), - auth_kwargs=kwargs, - ).execute() - return await poll_until_finished(kwargs, response) + ) + return await poll_until_finished(cls, response, average_duration=80) -class TripoRefineNode: +class TripoRefineNode(IO.ComfyNode): + @classmethod - def INPUT_TYPES(s): - return { - "required": { - "model_task_id": ("MODEL_TASK_ID", { - "tooltip": "Must be a v1.4 Tripo model" - }), - }, - "hidden": { - "auth_token": "AUTH_TOKEN_COMFY_ORG", - "comfy_api_key": "API_KEY_COMFY_ORG", - "unique_id": "UNIQUE_ID", - }, - } + def define_schema(cls): + return IO.Schema( + node_id="TripoRefineNode", + display_name="Tripo: Refine Draft model", + category="api node/3d/Tripo", + description="Refine a draft model created by v1.4 Tripo models only.", + inputs=[ + IO.Custom("MODEL_TASK_ID").Input("model_task_id", tooltip="Must be a v1.4 Tripo model"), + ], + outputs=[ + IO.String.Output(display_name="model_file"), + IO.Custom("MODEL_TASK_ID").Output(display_name="model task_id"), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + is_output_node=True, + ) - DESCRIPTION = "Refine a draft model created by v1.4 Tripo models only." - - RETURN_TYPES = ("STRING", "MODEL_TASK_ID",) - RETURN_NAMES = ("model_file", "model task_id") - FUNCTION = "generate_mesh" - CATEGORY = "api node/3d/Tripo" - API_NODE = True - OUTPUT_NODE = True - AVERAGE_DURATION = 240 - - async def generate_mesh(self, model_task_id, **kwargs): - response = await SynchronousOperation( - endpoint=ApiEndpoint( - path="/proxy/tripo/v2/openapi/task", - method=HttpMethod.POST, - request_model=TripoRefineModelRequest, - response_model=TripoTaskResponse, - ), - request=TripoRefineModelRequest( - draft_model_task_id=model_task_id - ), - auth_kwargs=kwargs, - ).execute() - return await poll_until_finished(kwargs, response) - - -class TripoRigNode: @classmethod - def INPUT_TYPES(s): - return { - "required": { - "original_model_task_id": ("MODEL_TASK_ID",), - }, - "hidden": { - "auth_token": "AUTH_TOKEN_COMFY_ORG", - "comfy_api_key": "API_KEY_COMFY_ORG", - "unique_id": "UNIQUE_ID", - }, - } - - RETURN_TYPES = ("STRING", "RIG_TASK_ID") - RETURN_NAMES = ("model_file", "rig task_id") - FUNCTION = "generate_mesh" - CATEGORY = "api node/3d/Tripo" - API_NODE = True - OUTPUT_NODE = True - AVERAGE_DURATION = 180 - - async def generate_mesh(self, original_model_task_id, **kwargs): - response = await SynchronousOperation( - endpoint=ApiEndpoint( - path="/proxy/tripo/v2/openapi/task", - method=HttpMethod.POST, - request_model=TripoAnimateRigRequest, - response_model=TripoTaskResponse, - ), - request=TripoAnimateRigRequest( - original_model_task_id=original_model_task_id, - out_format="glb", - spec="tripo" - ), - auth_kwargs=kwargs, - ).execute() - return await poll_until_finished(kwargs, response) + async def execute(cls, model_task_id) -> IO.NodeOutput: + response = await sync_op( + cls, + endpoint=ApiEndpoint(path="/proxy/tripo/v2/openapi/task", method="POST"), + response_model=TripoTaskResponse, + data=TripoRefineModelRequest(draft_model_task_id=model_task_id), + ) + return await poll_until_finished(cls, response, average_duration=240) -class TripoRetargetNode: +class TripoRigNode(IO.ComfyNode): + @classmethod - def INPUT_TYPES(s): - return { - "required": { - "original_model_task_id": ("RIG_TASK_ID",), - "animation": ([ - "preset:idle", - "preset:walk", - "preset:climb", - "preset:jump", - "preset:slash", - "preset:shoot", - "preset:hurt", - "preset:fall", - "preset:turn", - ],), - }, - "hidden": { - "auth_token": "AUTH_TOKEN_COMFY_ORG", - "comfy_api_key": "API_KEY_COMFY_ORG", - "unique_id": "UNIQUE_ID", - }, - } + def define_schema(cls): + return IO.Schema( + node_id="TripoRigNode", + display_name="Tripo: Rig model", + category="api node/3d/Tripo", + inputs=[IO.Custom("MODEL_TASK_ID").Input("original_model_task_id")], + outputs=[ + IO.String.Output(display_name="model_file"), + IO.Custom("RIG_TASK_ID").Output(display_name="rig task_id"), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + is_output_node=True, + ) - RETURN_TYPES = ("STRING", "RETARGET_TASK_ID") - RETURN_NAMES = ("model_file", "retarget task_id") - FUNCTION = "generate_mesh" - CATEGORY = "api node/3d/Tripo" - API_NODE = True - OUTPUT_NODE = True - AVERAGE_DURATION = 30 + @classmethod + async def execute(cls, original_model_task_id) -> IO.NodeOutput: + response = await sync_op( + cls, + endpoint=ApiEndpoint(path="/proxy/tripo/v2/openapi/task", method="POST"), + response_model=TripoTaskResponse, + data=TripoAnimateRigRequest(original_model_task_id=original_model_task_id, out_format="glb", spec="tripo"), + ) + return await poll_until_finished(cls, response, average_duration=180) - async def generate_mesh(self, animation, original_model_task_id, **kwargs): - response = await SynchronousOperation( - endpoint=ApiEndpoint( - path="/proxy/tripo/v2/openapi/task", - method=HttpMethod.POST, - request_model=TripoAnimateRetargetRequest, - response_model=TripoTaskResponse, - ), - request=TripoAnimateRetargetRequest( + +class TripoRetargetNode(IO.ComfyNode): + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="TripoRetargetNode", + display_name="Tripo: Retarget rigged model", + category="api node/3d/Tripo", + inputs=[ + IO.Custom("RIG_TASK_ID").Input("original_model_task_id"), + IO.Combo.Input( + "animation", + options=[ + "preset:idle", + "preset:walk", + "preset:climb", + "preset:jump", + "preset:slash", + "preset:shoot", + "preset:hurt", + "preset:fall", + "preset:turn", + ], + ), + ], + outputs=[ + IO.String.Output(display_name="model_file"), + IO.Custom("RETARGET_TASK_ID").Output(display_name="retarget task_id"), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + is_output_node=True, + ) + + @classmethod + async def execute(cls, original_model_task_id, animation: str) -> IO.NodeOutput: + response = await sync_op( + cls, + endpoint=ApiEndpoint(path="/proxy/tripo/v2/openapi/task", method="POST"), + response_model=TripoTaskResponse, + data=TripoAnimateRetargetRequest( original_model_task_id=original_model_task_id, animation=animation, out_format="glb", - bake_animation=True + bake_animation=True, ), - auth_kwargs=kwargs, - ).execute() - return await poll_until_finished(kwargs, response) + ) + return await poll_until_finished(cls, response, average_duration=30) -class TripoConversionNode: - @classmethod - def INPUT_TYPES(s): - return { - "required": { - "original_model_task_id": ("MODEL_TASK_ID,RIG_TASK_ID,RETARGET_TASK_ID",), - "format": (["GLTF", "USDZ", "FBX", "OBJ", "STL", "3MF"],), - }, - "optional": { - "quad": ("BOOLEAN", {"default": False}), - "face_limit": ("INT", {"min": -1, "max": 500000, "default": -1}), - "texture_size": ("INT", {"min": 128, "max": 4096, "default": 4096}), - "texture_format": (["BMP", "DPX", "HDR", "JPEG", "OPEN_EXR", "PNG", "TARGA", "TIFF", "WEBP"], {"default": "JPEG"}) - }, - "hidden": { - "auth_token": "AUTH_TOKEN_COMFY_ORG", - "comfy_api_key": "API_KEY_COMFY_ORG", - "unique_id": "UNIQUE_ID", - }, - } +class TripoConversionNode(IO.ComfyNode): @classmethod - def VALIDATE_INPUTS(cls, input_types): + def define_schema(cls): + return IO.Schema( + node_id="TripoConversionNode", + display_name="Tripo: Convert model", + category="api node/3d/Tripo", + inputs=[ + IO.Custom("MODEL_TASK_ID,RIG_TASK_ID,RETARGET_TASK_ID").Input("original_model_task_id"), + IO.Combo.Input("format", options=["GLTF", "USDZ", "FBX", "OBJ", "STL", "3MF"]), + IO.Boolean.Input("quad", default=False, optional=True), + IO.Int.Input( + "face_limit", + default=-1, + min=-1, + max=500000, + optional=True, + ), + IO.Int.Input( + "texture_size", + default=4096, + min=128, + max=4096, + optional=True, + ), + IO.Combo.Input( + "texture_format", + options=["BMP", "DPX", "HDR", "JPEG", "OPEN_EXR", "PNG", "TARGA", "TIFF", "WEBP"], + default="JPEG", + optional=True, + ), + ], + outputs=[], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + is_output_node=True, + ) + + @classmethod + def validate_inputs(cls, input_types): # The min and max of input1 and input2 are still validated because # we didn't take `input1` or `input2` as arguments if input_types["original_model_task_id"] not in ("MODEL_TASK_ID", "RIG_TASK_ID", "RETARGET_TASK_ID"): return "original_model_task_id must be MODEL_TASK_ID, RIG_TASK_ID or RETARGET_TASK_ID type" return True - RETURN_TYPES = () - FUNCTION = "generate_mesh" - CATEGORY = "api node/3d/Tripo" - API_NODE = True - OUTPUT_NODE = True - AVERAGE_DURATION = 30 - - async def generate_mesh(self, original_model_task_id, format, quad, face_limit, texture_size, texture_format, **kwargs): + @classmethod + async def execute( + cls, + original_model_task_id, + format: str, + quad: bool, + face_limit: int, + texture_size: int, + texture_format: str, + ) -> IO.NodeOutput: if not original_model_task_id: raise RuntimeError("original_model_task_id is required") - response = await SynchronousOperation( - endpoint=ApiEndpoint( - path="/proxy/tripo/v2/openapi/task", - method=HttpMethod.POST, - request_model=TripoConvertModelRequest, - response_model=TripoTaskResponse, - ), - request=TripoConvertModelRequest( + response = await sync_op( + cls, + endpoint=ApiEndpoint(path="/proxy/tripo/v2/openapi/task", method="POST"), + response_model=TripoTaskResponse, + data=TripoConvertModelRequest( original_model_task_id=original_model_task_id, format=format, quad=quad if quad else None, face_limit=face_limit if face_limit != -1 else None, texture_size=texture_size if texture_size != 4096 else None, - texture_format=texture_format if texture_format != "JPEG" else None + texture_format=texture_format if texture_format != "JPEG" else None, ), - auth_kwargs=kwargs, - ).execute() - return await poll_until_finished(kwargs, response) + ) + return await poll_until_finished(cls, response, average_duration=30) -NODE_CLASS_MAPPINGS = { - "TripoTextToModelNode": TripoTextToModelNode, - "TripoImageToModelNode": TripoImageToModelNode, - "TripoMultiviewToModelNode": TripoMultiviewToModelNode, - "TripoTextureNode": TripoTextureNode, - "TripoRefineNode": TripoRefineNode, - "TripoRigNode": TripoRigNode, - "TripoRetargetNode": TripoRetargetNode, - "TripoConversionNode": TripoConversionNode, -} +class TripoExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[IO.ComfyNode]]: + return [ + TripoTextToModelNode, + TripoImageToModelNode, + TripoMultiviewToModelNode, + TripoTextureNode, + TripoRefineNode, + TripoRigNode, + TripoRetargetNode, + TripoConversionNode, + ] -NODE_DISPLAY_NAME_MAPPINGS = { - "TripoTextToModelNode": "Tripo: Text to Model", - "TripoImageToModelNode": "Tripo: Image to Model", - "TripoMultiviewToModelNode": "Tripo: Multiview to Model", - "TripoTextureNode": "Tripo: Texture model", - "TripoRefineNode": "Tripo: Refine Draft model", - "TripoRigNode": "Tripo: Rig model", - "TripoRetargetNode": "Tripo: Retarget rigged model", - "TripoConversionNode": "Tripo: Convert model", -} + +async def comfy_entrypoint() -> TripoExtension: + return TripoExtension() diff --git a/comfy_api_nodes/util/__init__.py b/comfy_api_nodes/util/__init__.py index c2ec391aa..ab96760cb 100644 --- a/comfy_api_nodes/util/__init__.py +++ b/comfy_api_nodes/util/__init__.py @@ -20,6 +20,7 @@ from .conversions import ( trim_video, ) from .download_helpers import ( + download_url_as_bytesio, download_url_to_bytesio, download_url_to_image_tensor, download_url_to_video_output, @@ -56,6 +57,7 @@ __all__ = [ "upload_images_to_comfyapi", "upload_video_to_comfyapi", # Download helpers + "download_url_as_bytesio", "download_url_to_bytesio", "download_url_to_image_tensor", "download_url_to_video_output", diff --git a/comfy_api_nodes/util/download_helpers.py b/comfy_api_nodes/util/download_helpers.py index 055e690de..791dd5a50 100644 --- a/comfy_api_nodes/util/download_helpers.py +++ b/comfy_api_nodes/util/download_helpers.py @@ -240,6 +240,18 @@ async def download_url_to_video_output( return VideoFromFile(result) +async def download_url_as_bytesio( + url: str, + *, + timeout: float = None, + cls: type[COMFY_IO.ComfyNode] = None, +) -> BytesIO: + """Downloads content from a URL and returns a new BytesIO (rewound to 0).""" + result = BytesIO() + await download_url_to_bytesio(url, result, timeout=timeout, cls=cls) + return result + + def _generate_operation_id(method: str, url: str, attempt: int) -> str: try: parsed = urlparse(url) From 426cde37f10dc391f9601ab938e02c0faa42db14 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Fri, 24 Oct 2025 16:56:51 -0700 Subject: [PATCH 028/647] Remove useless function (#10472) --- comfy/model_management.py | 6 ------ 1 file changed, 6 deletions(-) diff --git a/comfy/model_management.py b/comfy/model_management.py index 79d6ff9d4..cf015a29a 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -998,12 +998,6 @@ def device_supports_non_blocking(device): return False return True -def device_should_use_non_blocking(device): - if not device_supports_non_blocking(device): - return False - return False - # return True #TODO: figure out why this causes memory issues on Nvidia and possibly others - def force_channels_last(): if args.force_channels_last: return True From e86b79ab9ea7e740b80490353f3f5763840ede81 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Sun, 26 Oct 2025 00:35:30 +0300 Subject: [PATCH 029/647] convert Gemini API nodes to V3 schema (#10476) --- comfy_api_nodes/apinode_utils.py | 26 -- comfy_api_nodes/nodes_gemini.py | 629 +++++++++++----------------- comfy_api_nodes/util/__init__.py | 2 + comfy_api_nodes/util/conversions.py | 25 ++ 4 files changed, 282 insertions(+), 400 deletions(-) diff --git a/comfy_api_nodes/apinode_utils.py b/comfy_api_nodes/apinode_utils.py index e3d282059..4182c8f80 100644 --- a/comfy_api_nodes/apinode_utils.py +++ b/comfy_api_nodes/apinode_utils.py @@ -3,8 +3,6 @@ import aiohttp import mimetypes from typing import Optional, Union from comfy.utils import common_upscale -from comfy_api.util import VideoContainer, VideoCodec -from comfy_api.input.video_types import VideoInput from comfy_api_nodes.apis.client import ( ApiClient, ApiEndpoint, @@ -209,30 +207,6 @@ async def upload_file_to_comfyapi( return response.download_url -def video_to_base64_string( - video: VideoInput, - container_format: VideoContainer = None, - codec: VideoCodec = None -) -> str: - """ - Converts a video input to a base64 string. - - Args: - video: The video input to convert - container_format: Optional container format to use (defaults to video.container if available) - codec: Optional codec to use (defaults to video.codec if available) - """ - video_bytes_io = BytesIO() - - # Use provided format/codec if specified, otherwise use video's own if available - format_to_use = container_format if container_format is not None else getattr(video, 'container', VideoContainer.MP4) - codec_to_use = codec if codec is not None else getattr(video, 'codec', VideoCodec.H264) - - video.save_to(video_bytes_io, format=format_to_use, codec=codec_to_use) - video_bytes_io.seek(0) - return base64.b64encode(video_bytes_io.getvalue()).decode("utf-8") - - async def upload_images_to_comfyapi( image: torch.Tensor, max_images=8, diff --git a/comfy_api_nodes/nodes_gemini.py b/comfy_api_nodes/nodes_gemini.py index ca11b67ed..67f2469ad 100644 --- a/comfy_api_nodes/nodes_gemini.py +++ b/comfy_api_nodes/nodes_gemini.py @@ -2,42 +2,47 @@ API Nodes for Gemini Multimodal LLM Usage via Remote API See: https://cloud.google.com/vertex-ai/generative-ai/docs/model-reference/inference """ + from __future__ import annotations -import json -import time -import os -import uuid import base64 -from io import BytesIO +import json +import os +import time +import uuid from enum import Enum -from typing import Optional, Literal +from io import BytesIO +from typing import Literal, Optional import torch +from typing_extensions import override import folder_paths -from comfy.comfy_types.node_typing import IO, ComfyNodeABC, InputTypeDict -from server import PromptServer +from comfy_api.latest import IO, ComfyExtension, Input +from comfy_api.util import VideoCodec, VideoContainer from comfy_api_nodes.apis import ( GeminiContent, GeminiGenerateContentRequest, GeminiGenerateContentResponse, GeminiInlineData, - GeminiPart, GeminiMimeType, + GeminiPart, ) -from comfy_api_nodes.apis.gemini_api import GeminiImageGenerationConfig, GeminiImageGenerateContentRequest, GeminiImageConfig -from comfy_api_nodes.apis.client import ( +from comfy_api_nodes.apis.gemini_api import ( + GeminiImageConfig, + GeminiImageGenerateContentRequest, + GeminiImageGenerationConfig, +) +from comfy_api_nodes.util import ( ApiEndpoint, - HttpMethod, - SynchronousOperation, -) -from comfy_api_nodes.apinode_utils import ( + audio_to_base64_string, + bytesio_to_image_tensor, + sync_op, + tensor_to_base64_string, + validate_string, video_to_base64_string, ) -from comfy_api_nodes.util import validate_string, tensor_to_base64_string, bytesio_to_image_tensor, audio_to_base64_string -from comfy_api.util import VideoContainer, VideoCodec - +from server import PromptServer GEMINI_BASE_ENDPOINT = "/proxy/vertexai/gemini" GEMINI_MAX_INPUT_FILE_SIZE = 20 * 1024 * 1024 # 20 MB @@ -63,50 +68,6 @@ class GeminiImageModel(str, Enum): gemini_2_5_flash_image = "gemini-2.5-flash-image" -def get_gemini_endpoint( - model: GeminiModel, -) -> ApiEndpoint[GeminiGenerateContentRequest, GeminiGenerateContentResponse]: - """ - Get the API endpoint for a given Gemini model. - - Args: - model: The Gemini model to use, either as enum or string value. - - Returns: - ApiEndpoint configured for the specific Gemini model. - """ - if isinstance(model, str): - model = GeminiModel(model) - return ApiEndpoint( - path=f"{GEMINI_BASE_ENDPOINT}/{model.value}", - method=HttpMethod.POST, - request_model=GeminiGenerateContentRequest, - response_model=GeminiGenerateContentResponse, - ) - - -def get_gemini_image_endpoint( - model: GeminiImageModel, -) -> ApiEndpoint[GeminiGenerateContentRequest, GeminiGenerateContentResponse]: - """ - Get the API endpoint for a given Gemini model. - - Args: - model: The Gemini model to use, either as enum or string value. - - Returns: - ApiEndpoint configured for the specific Gemini model. - """ - if isinstance(model, str): - model = GeminiImageModel(model) - return ApiEndpoint( - path=f"{GEMINI_BASE_ENDPOINT}/{model.value}", - method=HttpMethod.POST, - request_model=GeminiImageGenerateContentRequest, - response_model=GeminiGenerateContentResponse, - ) - - def create_image_parts(image_input: torch.Tensor) -> list[GeminiPart]: """ Convert image tensor input to Gemini API compatible parts. @@ -119,9 +80,7 @@ def create_image_parts(image_input: torch.Tensor) -> list[GeminiPart]: """ image_parts: list[GeminiPart] = [] for image_index in range(image_input.shape[0]): - image_as_b64 = tensor_to_base64_string( - image_input[image_index].unsqueeze(0) - ) + image_as_b64 = tensor_to_base64_string(image_input[image_index].unsqueeze(0)) image_parts.append( GeminiPart( inlineData=GeminiInlineData( @@ -133,37 +92,7 @@ def create_image_parts(image_input: torch.Tensor) -> list[GeminiPart]: return image_parts -def create_text_part(text: str) -> GeminiPart: - """ - Create a text part for the Gemini API request. - - Args: - text: The text content to include in the request. - - Returns: - A GeminiPart object with the text content. - """ - return GeminiPart(text=text) - - -def get_parts_from_response( - response: GeminiGenerateContentResponse -) -> list[GeminiPart]: - """ - Extract all parts from the Gemini API response. - - Args: - response: The API response from Gemini. - - Returns: - List of response parts from the first candidate. - """ - return response.candidates[0].content.parts - - -def get_parts_by_type( - response: GeminiGenerateContentResponse, part_type: Literal["text"] | str -) -> list[GeminiPart]: +def get_parts_by_type(response: GeminiGenerateContentResponse, part_type: Literal["text"] | str) -> list[GeminiPart]: """ Filter response parts by their type. @@ -175,14 +104,10 @@ def get_parts_by_type( List of response parts matching the requested type. """ parts = [] - for part in get_parts_from_response(response): + for part in response.candidates[0].content.parts: if part_type == "text" and hasattr(part, "text") and part.text: parts.append(part) - elif ( - hasattr(part, "inlineData") - and part.inlineData - and part.inlineData.mimeType == part_type - ): + elif hasattr(part, "inlineData") and part.inlineData and part.inlineData.mimeType == part_type: parts.append(part) # Skip parts that don't match the requested type return parts @@ -210,11 +135,11 @@ def get_image_from_response(response: GeminiGenerateContentResponse) -> torch.Te returned_image = bytesio_to_image_tensor(BytesIO(image_data)) image_tensors.append(returned_image) if len(image_tensors) == 0: - return torch.zeros((1,1024,1024,4)) + return torch.zeros((1, 1024, 1024, 4)) return torch.cat(image_tensors, dim=0) -class GeminiNode(ComfyNodeABC): +class GeminiNode(IO.ComfyNode): """ Node to generate text responses from a Gemini model. @@ -225,96 +150,79 @@ class GeminiNode(ComfyNodeABC): """ @classmethod - def INPUT_TYPES(cls) -> InputTypeDict: - return { - "required": { - "prompt": ( - IO.STRING, - { - "multiline": True, - "default": "", - "tooltip": "Text inputs to the model, used to generate a response. You can include detailed instructions, questions, or context for the model.", - }, + def define_schema(cls): + return IO.Schema( + node_id="GeminiNode", + display_name="Google Gemini", + category="api node/text/Gemini", + description="Generate text responses with Google's Gemini AI model. " + "You can provide multiple types of inputs (text, images, audio, video) " + "as context for generating more relevant and meaningful responses.", + inputs=[ + IO.String.Input( + "prompt", + multiline=True, + default="", + tooltip="Text inputs to the model, used to generate a response. " + "You can include detailed instructions, questions, or context for the model.", ), - "model": ( - IO.COMBO, - { - "tooltip": "The Gemini model to use for generating responses.", - "options": [model.value for model in GeminiModel], - "default": GeminiModel.gemini_2_5_pro.value, - }, + IO.Combo.Input( + "model", + options=GeminiModel, + default=GeminiModel.gemini_2_5_pro, + tooltip="The Gemini model to use for generating responses.", ), - "seed": ( - IO.INT, - { - "default": 42, - "min": 0, - "max": 0xFFFFFFFFFFFFFFFF, - "control_after_generate": True, - "tooltip": "When seed is fixed to a specific value, the model makes a best effort to provide the same response for repeated requests. Deterministic output isn't guaranteed. Also, changing the model or parameter settings, such as the temperature, can cause variations in the response even when you use the same seed value. By default, a random seed value is used.", - }, + IO.Int.Input( + "seed", + default=42, + min=0, + max=0xFFFFFFFFFFFFFFFF, + control_after_generate=True, + tooltip="When seed is fixed to a specific value, the model makes a best effort to provide " + "the same response for repeated requests. Deterministic output isn't guaranteed. " + "Also, changing the model or parameter settings, such as the temperature, " + "can cause variations in the response even when you use the same seed value. " + "By default, a random seed value is used.", ), - }, - "optional": { - "images": ( - IO.IMAGE, - { - "default": None, - "tooltip": "Optional image(s) to use as context for the model. To include multiple images, you can use the Batch Images node.", - }, + IO.Image.Input( + "images", + optional=True, + tooltip="Optional image(s) to use as context for the model. " + "To include multiple images, you can use the Batch Images node.", ), - "audio": ( - IO.AUDIO, - { - "tooltip": "Optional audio to use as context for the model.", - "default": None, - }, + IO.Audio.Input( + "audio", + optional=True, + tooltip="Optional audio to use as context for the model.", ), - "video": ( - IO.VIDEO, - { - "tooltip": "Optional video to use as context for the model.", - "default": None, - }, + IO.Video.Input( + "video", + optional=True, + tooltip="Optional video to use as context for the model.", ), - "files": ( - "GEMINI_INPUT_FILES", - { - "default": None, - "tooltip": "Optional file(s) to use as context for the model. Accepts inputs from the Gemini Generate Content Input Files node.", - }, + IO.Custom("GEMINI_INPUT_FILES").Input( + "files", + optional=True, + tooltip="Optional file(s) to use as context for the model. " + "Accepts inputs from the Gemini Generate Content Input Files node.", ), - }, - "hidden": { - "auth_token": "AUTH_TOKEN_COMFY_ORG", - "comfy_api_key": "API_KEY_COMFY_ORG", - "unique_id": "UNIQUE_ID", - }, - } - - DESCRIPTION = "Generate text responses with Google's Gemini AI model. You can provide multiple types of inputs (text, images, audio, video) as context for generating more relevant and meaningful responses." - RETURN_TYPES = ("STRING",) - FUNCTION = "api_call" - CATEGORY = "api node/text/Gemini" - API_NODE = True - - def create_video_parts(self, video_input: IO.VIDEO, **kwargs) -> list[GeminiPart]: - """ - Convert video input to Gemini API compatible parts. - - Args: - video_input: Video tensor from ComfyUI. - **kwargs: Additional arguments to pass to the conversion function. - - Returns: - List of GeminiPart objects containing the encoded video. - """ - - base_64_string = video_to_base64_string( - video_input, - container_format=VideoContainer.MP4, - codec=VideoCodec.H264 + ], + outputs=[ + IO.String.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, ) + + @classmethod + def create_video_parts(cls, video_input: Input.Video) -> list[GeminiPart]: + """Convert video input to Gemini API compatible parts.""" + + base_64_string = video_to_base64_string(video_input, container_format=VideoContainer.MP4, codec=VideoCodec.H264) return [ GeminiPart( inlineData=GeminiInlineData( @@ -324,7 +232,8 @@ class GeminiNode(ComfyNodeABC): ) ] - def create_audio_parts(self, audio_input: IO.AUDIO) -> list[GeminiPart]: + @classmethod + def create_audio_parts(cls, audio_input: Input.Audio) -> list[GeminiPart]: """ Convert audio input to Gemini API compatible parts. @@ -337,10 +246,10 @@ class GeminiNode(ComfyNodeABC): audio_parts: list[GeminiPart] = [] for batch_index in range(audio_input["waveform"].shape[0]): # Recreate an IO.AUDIO object for the given batch dimension index - audio_at_index = { - "waveform": audio_input["waveform"][batch_index].unsqueeze(0), - "sample_rate": audio_input["sample_rate"], - } + audio_at_index = Input.Audio( + waveform=audio_input["waveform"][batch_index].unsqueeze(0), + sample_rate=audio_input["sample_rate"], + ) # Convert to MP3 format for compatibility with Gemini API audio_bytes = audio_to_base64_string( audio_at_index, @@ -357,38 +266,38 @@ class GeminiNode(ComfyNodeABC): ) return audio_parts - async def api_call( - self, + @classmethod + async def execute( + cls, prompt: str, - model: GeminiModel, - images: Optional[IO.IMAGE] = None, - audio: Optional[IO.AUDIO] = None, - video: Optional[IO.VIDEO] = None, + model: str, + seed: int, + images: Optional[torch.Tensor] = None, + audio: Optional[Input.Audio] = None, + video: Optional[Input.Video] = None, files: Optional[list[GeminiPart]] = None, - unique_id: Optional[str] = None, - **kwargs, - ) -> tuple[str]: - # Validate inputs + ) -> IO.NodeOutput: validate_string(prompt, strip_whitespace=False) # Create parts list with text prompt as the first part - parts: list[GeminiPart] = [create_text_part(prompt)] + parts: list[GeminiPart] = [GeminiPart(text=prompt)] # Add other modal parts if images is not None: image_parts = create_image_parts(images) parts.extend(image_parts) if audio is not None: - parts.extend(self.create_audio_parts(audio)) + parts.extend(cls.create_audio_parts(audio)) if video is not None: - parts.extend(self.create_video_parts(video)) + parts.extend(cls.create_video_parts(video)) if files is not None: parts.extend(files) # Create response - response = await SynchronousOperation( - endpoint=get_gemini_endpoint(model), - request=GeminiGenerateContentRequest( + response = await sync_op( + cls, + endpoint=ApiEndpoint(path=f"{GEMINI_BASE_ENDPOINT}/{model}", method="POST"), + data=GeminiGenerateContentRequest( contents=[ GeminiContent( role="user", @@ -396,15 +305,15 @@ class GeminiNode(ComfyNodeABC): ) ] ), - auth_kwargs=kwargs, - ).execute() + response_model=GeminiGenerateContentResponse, + ) # Get result output output_text = get_text_from_response(response) - if unique_id and output_text: + if output_text: # Not a true chat history like the OpenAI Chat node. It is emulated so the frontend can show a copy button. render_spec = { - "node_id": unique_id, + "node_id": cls.hidden.unique_id, "component": "ChatHistoryWidget", "props": { "history": json.dumps( @@ -424,10 +333,10 @@ class GeminiNode(ComfyNodeABC): render_spec, ) - return (output_text or "Empty response from Gemini model...",) + return IO.NodeOutput(output_text or "Empty response from Gemini model...") -class GeminiInputFiles(ComfyNodeABC): +class GeminiInputFiles(IO.ComfyNode): """ Loads and formats input files for use with the Gemini API. @@ -438,7 +347,7 @@ class GeminiInputFiles(ComfyNodeABC): """ @classmethod - def INPUT_TYPES(cls) -> InputTypeDict: + def define_schema(cls): """ For details about the supported file input types, see: https://cloud.google.com/vertex-ai/generative-ai/docs/model-reference/inference @@ -453,39 +362,37 @@ class GeminiInputFiles(ComfyNodeABC): ] input_files = sorted(input_files, key=lambda x: x.name) input_files = [f.name for f in input_files] - return { - "required": { - "file": ( - IO.COMBO, - { - "tooltip": "Input files to include as context for the model. Only accepts text (.txt) and PDF (.pdf) files for now.", - "options": input_files, - "default": input_files[0] if input_files else None, - }, + return IO.Schema( + node_id="GeminiInputFiles", + display_name="Gemini Input Files", + category="api node/text/Gemini", + description="Loads and prepares input files to include as inputs for Gemini LLM nodes. " + "The files will be read by the Gemini model when generating a response. " + "The contents of the text file count toward the token limit. " + "🛈 TIP: Can be chained together with other Gemini Input File nodes.", + inputs=[ + IO.Combo.Input( + "file", + options=input_files, + default=input_files[0] if input_files else None, + tooltip="Input files to include as context for the model. " + "Only accepts text (.txt) and PDF (.pdf) files for now.", ), - }, - "optional": { - "GEMINI_INPUT_FILES": ( + IO.Custom("GEMINI_INPUT_FILES").Input( "GEMINI_INPUT_FILES", - { - "tooltip": "An optional additional file(s) to batch together with the file loaded from this node. Allows chaining of input files so that a single message can include multiple input files.", - "default": None, - }, + optional=True, + tooltip="An optional additional file(s) to batch together with the file loaded from this node. " + "Allows chaining of input files so that a single message can include multiple input files.", ), - }, - } - - DESCRIPTION = "Loads and prepares input files to include as inputs for Gemini LLM nodes. The files will be read by the Gemini model when generating a response. The contents of the text file count toward the token limit. 🛈 TIP: Can be chained together with other Gemini Input File nodes." - RETURN_TYPES = ("GEMINI_INPUT_FILES",) - FUNCTION = "prepare_files" - CATEGORY = "api node/text/Gemini" - - def create_file_part(self, file_path: str) -> GeminiPart: - mime_type = ( - GeminiMimeType.application_pdf - if file_path.endswith(".pdf") - else GeminiMimeType.text_plain + ], + outputs=[ + IO.Custom("GEMINI_INPUT_FILES").Output(), + ], ) + + @classmethod + def create_file_part(cls, file_path: str) -> GeminiPart: + mime_type = GeminiMimeType.application_pdf if file_path.endswith(".pdf") else GeminiMimeType.text_plain # Use base64 string directly, not the data URI with open(file_path, "rb") as f: file_content = f.read() @@ -498,120 +405,95 @@ class GeminiInputFiles(ComfyNodeABC): ) ) - def prepare_files( - self, file: str, GEMINI_INPUT_FILES: list[GeminiPart] = [] - ) -> tuple[list[GeminiPart]]: - """ - Loads and formats input files for Gemini API. - """ - file_path = folder_paths.get_annotated_filepath(file) - input_file_content = self.create_file_part(file_path) - files = [input_file_content] + GEMINI_INPUT_FILES - return (files,) - - -class GeminiImage(ComfyNodeABC): - """ - Node to generate text and image responses from a Gemini model. - - This node allows users to interact with Google's Gemini AI models, providing - multimodal inputs (text, images, files) to generate coherent - text and image responses. The node works with the latest Gemini models, handling the - API communication and response parsing. - """ @classmethod - def INPUT_TYPES(cls) -> InputTypeDict: - return { - "required": { - "prompt": ( - IO.STRING, - { - "multiline": True, - "default": "", - "tooltip": "Text prompt for generation", - }, - ), - "model": ( - IO.COMBO, - { - "tooltip": "The Gemini model to use for generating responses.", - "options": [model.value for model in GeminiImageModel], - "default": GeminiImageModel.gemini_2_5_flash_image.value, - }, - ), - "seed": ( - IO.INT, - { - "default": 42, - "min": 0, - "max": 0xFFFFFFFFFFFFFFFF, - "control_after_generate": True, - "tooltip": "When seed is fixed to a specific value, the model makes a best effort to provide the same response for repeated requests. Deterministic output isn't guaranteed. Also, changing the model or parameter settings, such as the temperature, can cause variations in the response even when you use the same seed value. By default, a random seed value is used.", - }, - ), - }, - "optional": { - "images": ( - IO.IMAGE, - { - "default": None, - "tooltip": "Optional image(s) to use as context for the model. To include multiple images, you can use the Batch Images node.", - }, - ), - "files": ( - "GEMINI_INPUT_FILES", - { - "default": None, - "tooltip": "Optional file(s) to use as context for the model. Accepts inputs from the Gemini Generate Content Input Files node.", - }, - ), - # TODO: later we can add this parameter later - # "n": ( - # IO.INT, - # { - # "default": 1, - # "min": 1, - # "max": 8, - # "step": 1, - # "display": "number", - # "tooltip": "How many images to generate", - # }, - # ), - "aspect_ratio": ( - IO.COMBO, - { - "tooltip": "Defaults to matching the output image size to that of your input image, or otherwise generates 1:1 squares.", - "options": ["auto", "1:1", "2:3", "3:2", "3:4", "4:3", "4:5", "5:4", "9:16", "16:9", "21:9"], - "default": "auto", - }, - ), - }, - "hidden": { - "auth_token": "AUTH_TOKEN_COMFY_ORG", - "comfy_api_key": "API_KEY_COMFY_ORG", - "unique_id": "UNIQUE_ID", - }, - } + def execute(cls, file: str, GEMINI_INPUT_FILES: Optional[list[GeminiPart]] = None) -> IO.NodeOutput: + """Loads and formats input files for Gemini API.""" + if GEMINI_INPUT_FILES is None: + GEMINI_INPUT_FILES = [] + file_path = folder_paths.get_annotated_filepath(file) + input_file_content = cls.create_file_part(file_path) + return IO.NodeOutput([input_file_content] + GEMINI_INPUT_FILES) - RETURN_TYPES = (IO.IMAGE, IO.STRING) - FUNCTION = "api_call" - CATEGORY = "api node/image/Gemini" - DESCRIPTION = "Edit images synchronously via Google API." - API_NODE = True - async def api_call( - self, +class GeminiImage(IO.ComfyNode): + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="GeminiImageNode", + display_name="Google Gemini Image", + category="api node/image/Gemini", + description="Edit images synchronously via Google API.", + inputs=[ + IO.String.Input( + "prompt", + multiline=True, + tooltip="Text prompt for generation", + default="", + ), + IO.Combo.Input( + "model", + options=GeminiImageModel, + default=GeminiImageModel.gemini_2_5_flash_image, + tooltip="The Gemini model to use for generating responses.", + ), + IO.Int.Input( + "seed", + default=42, + min=0, + max=0xFFFFFFFFFFFFFFFF, + control_after_generate=True, + tooltip="When seed is fixed to a specific value, the model makes a best effort to provide " + "the same response for repeated requests. Deterministic output isn't guaranteed. " + "Also, changing the model or parameter settings, such as the temperature, " + "can cause variations in the response even when you use the same seed value. " + "By default, a random seed value is used.", + ), + IO.Image.Input( + "images", + optional=True, + tooltip="Optional image(s) to use as context for the model. " + "To include multiple images, you can use the Batch Images node.", + ), + IO.Custom("GEMINI_INPUT_FILES").Input( + "files", + optional=True, + tooltip="Optional file(s) to use as context for the model. " + "Accepts inputs from the Gemini Generate Content Input Files node.", + ), + IO.Combo.Input( + "aspect_ratio", + options=["auto", "1:1", "2:3", "3:2", "3:4", "4:3", "4:5", "5:4", "9:16", "16:9", "21:9"], + default="auto", + tooltip="Defaults to matching the output image size to that of your input image, " + "or otherwise generates 1:1 squares.", + optional=True, + ), + ], + outputs=[ + IO.Image.Output(), + IO.String.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + ) + + @classmethod + async def execute( + cls, prompt: str, - model: GeminiImageModel, - images: Optional[IO.IMAGE] = None, + model: str, + seed: int, + images: Optional[torch.Tensor] = None, files: Optional[list[GeminiPart]] = None, - n=1, aspect_ratio: str = "auto", - unique_id: Optional[str] = None, - **kwargs, - ): + ) -> IO.NodeOutput: validate_string(prompt, strip_whitespace=True, min_length=1) - parts: list[GeminiPart] = [create_text_part(prompt)] + parts: list[GeminiPart] = [GeminiPart(text=prompt)] if not aspect_ratio: aspect_ratio = "auto" # for backward compatability with old workflows; to-do remove this in December @@ -623,29 +505,27 @@ class GeminiImage(ComfyNodeABC): if files is not None: parts.extend(files) - response = await SynchronousOperation( - endpoint=get_gemini_image_endpoint(model), - request=GeminiImageGenerateContentRequest( + response = await sync_op( + cls, + endpoint=ApiEndpoint(path=f"{GEMINI_BASE_ENDPOINT}/{model}", method="POST"), + data=GeminiImageGenerateContentRequest( contents=[ - GeminiContent( - role="user", - parts=parts, - ), + GeminiContent(role="user", parts=parts), ], generationConfig=GeminiImageGenerationConfig( - responseModalities=["TEXT","IMAGE"], + responseModalities=["TEXT", "IMAGE"], imageConfig=None if aspect_ratio == "auto" else image_config, - ) + ), ), - auth_kwargs=kwargs, - ).execute() + response_model=GeminiGenerateContentResponse, + ) output_image = get_image_from_response(response) output_text = get_text_from_response(response) - if unique_id and output_text: + if output_text: # Not a true chat history like the OpenAI Chat node. It is emulated so the frontend can show a copy button. render_spec = { - "node_id": unique_id, + "node_id": cls.hidden.unique_id, "component": "ChatHistoryWidget", "props": { "history": json.dumps( @@ -666,17 +546,18 @@ class GeminiImage(ComfyNodeABC): ) output_text = output_text or "Empty response from Gemini model..." - return (output_image, output_text,) + return IO.NodeOutput(output_image, output_text) -NODE_CLASS_MAPPINGS = { - "GeminiNode": GeminiNode, - "GeminiImageNode": GeminiImage, - "GeminiInputFiles": GeminiInputFiles, -} +class GeminiExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[IO.ComfyNode]]: + return [ + GeminiNode, + GeminiImage, + GeminiInputFiles, + ] -NODE_DISPLAY_NAME_MAPPINGS = { - "GeminiNode": "Google Gemini", - "GeminiImageNode": "Google Gemini Image", - "GeminiInputFiles": "Gemini Input Files", -} + +async def comfy_entrypoint() -> GeminiExtension: + return GeminiExtension() diff --git a/comfy_api_nodes/util/__init__.py b/comfy_api_nodes/util/__init__.py index ab96760cb..0cca2b59b 100644 --- a/comfy_api_nodes/util/__init__.py +++ b/comfy_api_nodes/util/__init__.py @@ -18,6 +18,7 @@ from .conversions import ( tensor_to_bytesio, tensor_to_pil, trim_video, + video_to_base64_string, ) from .download_helpers import ( download_url_as_bytesio, @@ -73,6 +74,7 @@ __all__ = [ "tensor_to_bytesio", "tensor_to_pil", "trim_video", + "video_to_base64_string", # Validation utilities "get_number_of_images", "validate_aspect_ratio_closeness", diff --git a/comfy_api_nodes/util/conversions.py b/comfy_api_nodes/util/conversions.py index 10cd1051b..9f4c90c5c 100644 --- a/comfy_api_nodes/util/conversions.py +++ b/comfy_api_nodes/util/conversions.py @@ -12,6 +12,7 @@ from PIL import Image from comfy.utils import common_upscale from comfy_api.latest import Input, InputImpl +from comfy_api.util import VideoContainer, VideoCodec from ._helpers import mimetype_to_extension @@ -173,6 +174,30 @@ def audio_to_base64_string(audio: Input.Audio, container_format: str = "mp4", co return base64.b64encode(audio_bytes).decode("utf-8") +def video_to_base64_string( + video: Input.Video, + container_format: VideoContainer = None, + codec: VideoCodec = None +) -> str: + """ + Converts a video input to a base64 string. + + Args: + video: The video input to convert + container_format: Optional container format to use (defaults to video.container if available) + codec: Optional codec to use (defaults to video.codec if available) + """ + video_bytes_io = BytesIO() + + # Use provided format/codec if specified, otherwise use video's own if available + format_to_use = container_format if container_format is not None else getattr(video, 'container', VideoContainer.MP4) + codec_to_use = codec if codec is not None else getattr(video, 'codec', VideoCodec.H264) + + video.save_to(video_bytes_io, format=format_to_use, codec=codec_to_use) + video_bytes_io.seek(0) + return base64.b64encode(video_bytes_io.getvalue()).decode("utf-8") + + def audio_ndarray_to_bytesio( audio_data_np: np.ndarray, sample_rate: int, From 098a352f136c610071bcb74f13e5b0ca16e6e7b3 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sat, 25 Oct 2025 17:05:22 -0700 Subject: [PATCH 030/647] Add warning for torch-directml usage (#10482) Added a warning message about the state of torch-directml. --- comfy/model_management.py | 1 + 1 file changed, 1 insertion(+) diff --git a/comfy/model_management.py b/comfy/model_management.py index cf015a29a..afe78f36e 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -89,6 +89,7 @@ if args.deterministic: directml_enabled = False if args.directml is not None: + logging.warning("WARNING: torch-directml barely works, is very slow, has not been updated in over 1 year and might be removed soon, please don't use it, there are better options.") import torch_directml directml_enabled = True device_index = args.directml From f6bbc1ac846b7d9a73ae50c3a45cf5a41058c54d Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sat, 25 Oct 2025 20:07:29 -0700 Subject: [PATCH 031/647] Fix mistake. (#10484) --- comfy/sample.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy/sample.py b/comfy/sample.py index b1395da84..2f8f3a51c 100644 --- a/comfy/sample.py +++ b/comfy/sample.py @@ -17,7 +17,7 @@ def prepare_noise_inner(latent_image, generator, noise_inds=None): if i in unique_inds: noises.append(noise) noises = [noises[i] for i in inverse] - noises = torch.cat(noises, axis=0) + return torch.cat(noises, axis=0) def prepare_noise(latent_image, seed, noise_inds=None): """ From 9d529e53084bdec28f684f3886a26c93598e7338 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Sun, 26 Oct 2025 08:51:06 +0200 Subject: [PATCH 032/647] fix(api-nodes): random issues on Windows by capturing general OSError for retries (#10486) --- comfy_api_nodes/util/client.py | 15 +++++---------- comfy_api_nodes/util/download_helpers.py | 6 +++--- comfy_api_nodes/util/upload_helpers.py | 4 ++-- 3 files changed, 10 insertions(+), 15 deletions(-) diff --git a/comfy_api_nodes/util/client.py b/comfy_api_nodes/util/client.py index 5833b118f..9c036d64b 100644 --- a/comfy_api_nodes/util/client.py +++ b/comfy_api_nodes/util/client.py @@ -2,7 +2,6 @@ import asyncio import contextlib import json import logging -import socket import time import uuid from dataclasses import dataclass @@ -456,24 +455,20 @@ async def _diagnose_connectivity() -> dict[str, bool]: results = { "internet_accessible": False, "api_accessible": False, - "is_local_issue": False, - "is_api_issue": False, } timeout = aiohttp.ClientTimeout(total=5.0) async with aiohttp.ClientSession(timeout=timeout) as session: - try: + with contextlib.suppress(ClientError, OSError): async with session.get("https://www.google.com") as resp: results["internet_accessible"] = resp.status < 500 - except (ClientError, asyncio.TimeoutError, socket.gaierror): - results["is_local_issue"] = True + if not results["internet_accessible"]: return results parsed = urlparse(default_base_url()) health_url = f"{parsed.scheme}://{parsed.netloc}/health" - with contextlib.suppress(ClientError, asyncio.TimeoutError): + with contextlib.suppress(ClientError, OSError): async with session.get(health_url) as resp: results["api_accessible"] = resp.status < 500 - results["is_api_issue"] = results["internet_accessible"] and not results["api_accessible"] return results @@ -790,7 +785,7 @@ async def _request_base(cfg: _RequestConfig, expect_binary: bool): except ProcessingInterrupted: logging.debug("Polling was interrupted by user") raise - except (ClientError, asyncio.TimeoutError, socket.gaierror) as e: + except (ClientError, OSError) as e: if attempt <= cfg.max_retries: logging.warning( "Connection error calling %s %s. Retrying in %.2fs (%d/%d): %s", @@ -824,7 +819,7 @@ async def _request_base(cfg: _RequestConfig, expect_binary: bool): delay *= cfg.retry_backoff continue diag = await _diagnose_connectivity() - if diag.get("is_local_issue"): + if not diag["internet_accessible"]: try: request_logger.log_request_response( operation_id=operation_id, diff --git a/comfy_api_nodes/util/download_helpers.py b/comfy_api_nodes/util/download_helpers.py index 791dd5a50..f89045e12 100644 --- a/comfy_api_nodes/util/download_helpers.py +++ b/comfy_api_nodes/util/download_helpers.py @@ -32,7 +32,7 @@ async def download_url_to_bytesio( dest: Optional[Union[BytesIO, IO[bytes], str, Path]], *, timeout: Optional[float] = None, - max_retries: int = 3, + max_retries: int = 5, retry_delay: float = 1.0, retry_backoff: float = 2.0, cls: type[COMFY_IO.ComfyNode] = None, @@ -177,7 +177,7 @@ async def download_url_to_bytesio( return except asyncio.CancelledError: raise ProcessingInterrupted("Task cancelled") from None - except (ClientError, asyncio.TimeoutError) as e: + except (ClientError, OSError) as e: if attempt <= max_retries: with contextlib.suppress(Exception): request_logger.log_request_response( @@ -191,7 +191,7 @@ async def download_url_to_bytesio( continue diag = await _diagnose_connectivity() - if diag.get("is_local_issue"): + if not diag["internet_accessible"]: raise LocalNetworkError( "Unable to connect to the network. Please check your internet connection and try again." ) from e diff --git a/comfy_api_nodes/util/upload_helpers.py b/comfy_api_nodes/util/upload_helpers.py index a345d451d..7bfc61704 100644 --- a/comfy_api_nodes/util/upload_helpers.py +++ b/comfy_api_nodes/util/upload_helpers.py @@ -290,7 +290,7 @@ async def upload_file( return except asyncio.CancelledError: raise ProcessingInterrupted("Task cancelled") from None - except (aiohttp.ClientError, asyncio.TimeoutError) as e: + except (aiohttp.ClientError, OSError) as e: if attempt <= max_retries: with contextlib.suppress(Exception): request_logger.log_request_response( @@ -313,7 +313,7 @@ async def upload_file( continue diag = await _diagnose_connectivity() - if diag.get("is_local_issue"): + if not diag["internet_accessible"]: raise LocalNetworkError( "Unable to connect to the network. Please check your internet connection and try again." ) from e From c170fd2db598a0bdce56f80e22e83e10ad731421 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sun, 26 Oct 2025 17:23:01 -0700 Subject: [PATCH 033/647] Bump portable deps workflow to torch cu130 python 3.13.9 (#10493) --- .github/workflows/windows_release_dependencies.yml | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/.github/workflows/windows_release_dependencies.yml b/.github/workflows/windows_release_dependencies.yml index f1e2946e6..f61ee21a2 100644 --- a/.github/workflows/windows_release_dependencies.yml +++ b/.github/workflows/windows_release_dependencies.yml @@ -17,7 +17,7 @@ on: description: 'cuda version' required: true type: string - default: "129" + default: "130" python_minor: description: 'python minor version' @@ -29,7 +29,7 @@ on: description: 'python patch version' required: true type: string - default: "6" + default: "9" # push: # branches: # - master From 601ee1775a3c06c9b4de1fa7d808af8625b2fcd5 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 27 Oct 2025 20:54:00 -0700 Subject: [PATCH 034/647] Add a bat to run comfyui portable without api nodes. (#10504) --- .../advanced/run_nvidia_gpu_disable_api_nodes.bat | 2 ++ 1 file changed, 2 insertions(+) create mode 100644 .ci/windows_nvidia_base_files/advanced/run_nvidia_gpu_disable_api_nodes.bat diff --git a/.ci/windows_nvidia_base_files/advanced/run_nvidia_gpu_disable_api_nodes.bat b/.ci/windows_nvidia_base_files/advanced/run_nvidia_gpu_disable_api_nodes.bat new file mode 100644 index 000000000..cfe4b9f0e --- /dev/null +++ b/.ci/windows_nvidia_base_files/advanced/run_nvidia_gpu_disable_api_nodes.bat @@ -0,0 +1,2 @@ +..\python_embeded\python.exe -s ..\ComfyUI\main.py --windows-standalone-build --disable-api-nodes +pause From c305deed56a6ed259563b2047d9fcd51471e6590 Mon Sep 17 00:00:00 2001 From: ComfyUI Wiki Date: Tue, 28 Oct 2025 13:24:16 +0800 Subject: [PATCH 035/647] Update template to 0.2.3 (#10503) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 8570c66b6..121301669 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,5 +1,5 @@ comfyui-frontend-package==1.28.7 -comfyui-workflow-templates==0.2.2 +comfyui-workflow-templates==0.2.3 comfyui-embedded-docs==0.3.0 torch torchsde From 55bad303754eb60fa98f3ccf598e95502b819149 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Tue, 28 Oct 2025 07:25:29 +0200 Subject: [PATCH 036/647] feat(api-nodes): add LTXV API nodes (#10496) --- comfy_api_nodes/nodes_ltxv.py | 191 ++++++++++++++++++++++++++++++++++ nodes.py | 1 + 2 files changed, 192 insertions(+) create mode 100644 comfy_api_nodes/nodes_ltxv.py diff --git a/comfy_api_nodes/nodes_ltxv.py b/comfy_api_nodes/nodes_ltxv.py new file mode 100644 index 000000000..e6ad6e27a --- /dev/null +++ b/comfy_api_nodes/nodes_ltxv.py @@ -0,0 +1,191 @@ +from io import BytesIO +from typing import Optional + +import torch +from pydantic import BaseModel, Field +from typing_extensions import override + +from comfy_api.input_impl import VideoFromFile +from comfy_api.latest import IO, ComfyExtension +from comfy_api_nodes.util import ( + ApiEndpoint, + get_number_of_images, + sync_op_raw, + upload_images_to_comfyapi, + validate_string, +) + +MODELS_MAP = { + "LTX-2 (Pro)": "ltx-2-pro", + "LTX-2 (Fast)": "ltx-2-fast", +} + + +class ExecuteTaskRequest(BaseModel): + prompt: str = Field(...) + model: str = Field(...) + duration: int = Field(...) + resolution: str = Field(...) + fps: Optional[int] = Field(25) + generate_audio: Optional[bool] = Field(True) + image_uri: Optional[str] = Field(None) + + +class TextToVideoNode(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="LtxvApiTextToVideo", + display_name="LTXV Text To Video", + category="api node/video/LTXV", + description="Professional-quality videos with customizable duration and resolution.", + inputs=[ + IO.Combo.Input("model", options=list(MODELS_MAP.keys())), + IO.String.Input( + "prompt", + multiline=True, + default="", + ), + IO.Combo.Input("duration", options=[6, 8, 10], default=8), + IO.Combo.Input( + "resolution", + options=[ + "1920x1080", + "2560x1440", + "3840x2160", + ], + ), + IO.Combo.Input("fps", options=[25, 50], default=25), + IO.Boolean.Input( + "generate_audio", + default=False, + optional=True, + tooltip="When true, the generated video will include AI-generated audio matching the scene.", + ), + ], + outputs=[ + IO.Video.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + ) + + @classmethod + async def execute( + cls, + model: str, + prompt: str, + duration: int, + resolution: str, + fps: int = 25, + generate_audio: bool = False, + ) -> IO.NodeOutput: + validate_string(prompt, min_length=1, max_length=10000) + response = await sync_op_raw( + cls, + ApiEndpoint("/proxy/ltx/v1/text-to-video", "POST"), + data=ExecuteTaskRequest( + prompt=prompt, + model=MODELS_MAP[model], + duration=duration, + resolution=resolution, + fps=fps, + generate_audio=generate_audio, + ), + as_binary=True, + max_retries=1, + ) + return IO.NodeOutput(VideoFromFile(BytesIO(response))) + + +class ImageToVideoNode(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="LtxvApiImageToVideo", + display_name="LTXV Image To Video", + category="api node/video/LTXV", + description="Professional-quality videos with customizable duration and resolution based on start image.", + inputs=[ + IO.Image.Input("image", tooltip="First frame to be used for the video."), + IO.Combo.Input("model", options=list(MODELS_MAP.keys())), + IO.String.Input( + "prompt", + multiline=True, + default="", + ), + IO.Combo.Input("duration", options=[6, 8, 10], default=8), + IO.Combo.Input( + "resolution", + options=[ + "1920x1080", + "2560x1440", + "3840x2160", + ], + ), + IO.Combo.Input("fps", options=[25, 50], default=25), + IO.Boolean.Input( + "generate_audio", + default=False, + optional=True, + tooltip="When true, the generated video will include AI-generated audio matching the scene.", + ), + ], + outputs=[ + IO.Video.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + ) + + @classmethod + async def execute( + cls, + image: torch.Tensor, + model: str, + prompt: str, + duration: int, + resolution: str, + fps: int = 25, + generate_audio: bool = False, + ) -> IO.NodeOutput: + validate_string(prompt, min_length=1, max_length=10000) + if get_number_of_images(image) != 1: + raise ValueError("Currently only one input image is supported.") + response = await sync_op_raw( + cls, + ApiEndpoint("/proxy/ltx/v1/image-to-video", "POST"), + data=ExecuteTaskRequest( + image_uri=(await upload_images_to_comfyapi(cls, image, max_images=1, mime_type="image/png"))[0], + prompt=prompt, + model=MODELS_MAP[model], + duration=duration, + resolution=resolution, + fps=fps, + generate_audio=generate_audio, + ), + as_binary=True, + max_retries=1, + ) + return IO.NodeOutput(VideoFromFile(BytesIO(response))) + + +class LtxvApiExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[IO.ComfyNode]]: + return [ + TextToVideoNode, + ImageToVideoNode, + ] + + +async def comfy_entrypoint() -> LtxvApiExtension: + return LtxvApiExtension() diff --git a/nodes.py b/nodes.py index 7cfa8ca14..12e365ca9 100644 --- a/nodes.py +++ b/nodes.py @@ -2349,6 +2349,7 @@ async def init_builtin_api_nodes(): "nodes_kling.py", "nodes_bfl.py", "nodes_bytedance.py", + "nodes_ltxv.py", "nodes_luma.py", "nodes_recraft.py", "nodes_pixverse.py", From 6abc30aae9bd13f31dafd32552a365f2df2cf715 Mon Sep 17 00:00:00 2001 From: ComfyUI Wiki Date: Tue, 28 Oct 2025 13:56:30 +0800 Subject: [PATCH 037/647] Update template to 0.2.4 (#10505) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 121301669..cc3d4ca94 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,5 +1,5 @@ comfyui-frontend-package==1.28.7 -comfyui-workflow-templates==0.2.3 +comfyui-workflow-templates==0.2.4 comfyui-embedded-docs==0.3.0 torch torchsde From 614b8d3345424481d94a22fe7496d908c1a5c526 Mon Sep 17 00:00:00 2001 From: Jedrzej Kosinski Date: Tue, 28 Oct 2025 00:01:13 -0700 Subject: [PATCH 038/647] frontend bump to 1.28.8 (#10506) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index cc3d4ca94..4d84b0d3e 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,4 +1,4 @@ -comfyui-frontend-package==1.28.7 +comfyui-frontend-package==1.28.8 comfyui-workflow-templates==0.2.4 comfyui-embedded-docs==0.3.0 torch From f2bb3230b796f6a486894fc3b597db2c0b9538c9 Mon Sep 17 00:00:00 2001 From: comfyanonymous Date: Tue, 28 Oct 2025 03:03:59 -0400 Subject: [PATCH 039/647] ComfyUI version v0.3.67 --- comfyui_version.py | 2 +- pyproject.toml | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/comfyui_version.py b/comfyui_version.py index 33a06bbb0..db48b05c4 100644 --- a/comfyui_version.py +++ b/comfyui_version.py @@ -1,3 +1,3 @@ # This file is automatically generated by the build process when version is # updated in pyproject.toml. -__version__ = "0.3.66" +__version__ = "0.3.67" diff --git a/pyproject.toml b/pyproject.toml index fcc4854a5..ab054355c 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -1,6 +1,6 @@ [project] name = "ComfyUI" -version = "0.3.66" +version = "0.3.67" readme = "README.md" license = { file = "LICENSE" } requires-python = ">=3.9" From b61a40cbc9c2eb648b4d22bb513ed3ab2e2f0fd7 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 28 Oct 2025 00:21:45 -0700 Subject: [PATCH 040/647] Bump stable portable to cu130 python 3.13.9 (#10508) --- .github/workflows/release-stable-all.yml | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/.github/workflows/release-stable-all.yml b/.github/workflows/release-stable-all.yml index 5c1024599..7dca7277b 100644 --- a/.github/workflows/release-stable-all.yml +++ b/.github/workflows/release-stable-all.yml @@ -18,9 +18,9 @@ jobs: uses: ./.github/workflows/stable-release.yml with: git_tag: ${{ inputs.git_tag }} - cache_tag: "cu129" + cache_tag: "cu130" python_minor: "13" - python_patch: "6" + python_patch: "9" rel_name: "nvidia" rel_extra_name: "" test_release: true From 8cf2ba4ba64203551276513068ee81145e90f0bc Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 28 Oct 2025 00:23:52 -0700 Subject: [PATCH 041/647] Remove comfy api key from queue api. (#10502) --- execution.py | 8 +++----- main.py | 11 +++++++++-- server.py | 11 ++++++++--- 3 files changed, 20 insertions(+), 10 deletions(-) diff --git a/execution.py b/execution.py index 78c36a4b0..b14bb14c7 100644 --- a/execution.py +++ b/execution.py @@ -1116,7 +1116,7 @@ class PromptQueue: messages: List[str] def task_done(self, item_id, history_result, - status: Optional['PromptQueue.ExecutionStatus']): + status: Optional['PromptQueue.ExecutionStatus'], process_item=None): with self.mutex: prompt = self.currently_running.pop(item_id) if len(self.history) > MAXIMUM_HISTORY_SIZE: @@ -1126,10 +1126,8 @@ class PromptQueue: if status is not None: status_dict = copy.deepcopy(status._asdict()) - # Remove sensitive data from extra_data before storing in history - for sensitive_val in SENSITIVE_EXTRA_DATA_KEYS: - if sensitive_val in prompt[3]: - prompt[3].pop(sensitive_val) + if process_item is not None: + prompt = process_item(prompt) self.history[prompt[1]] = { "prompt": prompt, diff --git a/main.py b/main.py index 4b4c5dcc4..8d466d2eb 100644 --- a/main.py +++ b/main.py @@ -192,14 +192,21 @@ def prompt_worker(q, server_instance): prompt_id = item[1] server_instance.last_prompt_id = prompt_id - e.execute(item[2], prompt_id, item[3], item[4]) + sensitive = item[5] + extra_data = item[3].copy() + for k in sensitive: + extra_data[k] = sensitive[k] + + e.execute(item[2], prompt_id, extra_data, item[4]) need_gc = True + + remove_sensitive = lambda prompt: prompt[:5] + prompt[6:] q.task_done(item_id, e.history_result, status=execution.PromptQueue.ExecutionStatus( status_str='success' if e.success else 'error', completed=e.success, - messages=e.status_messages)) + messages=e.status_messages), process_item=remove_sensitive) if server_instance.client_id is not None: server_instance.send_sync("executing", {"node": None, "prompt_id": prompt_id}, server_instance.client_id) diff --git a/server.py b/server.py index fe58db286..5d773b10a 100644 --- a/server.py +++ b/server.py @@ -691,8 +691,9 @@ class PromptServer(): async def get_queue(request): queue_info = {} current_queue = self.prompt_queue.get_current_queue_volatile() - queue_info['queue_running'] = current_queue[0] - queue_info['queue_pending'] = current_queue[1] + remove_sensitive = lambda queue: [x[:5] for x in queue] + queue_info['queue_running'] = remove_sensitive(current_queue[0]) + queue_info['queue_pending'] = remove_sensitive(current_queue[1]) return web.json_response(queue_info) @routes.post("/prompt") @@ -728,7 +729,11 @@ class PromptServer(): extra_data["client_id"] = json_data["client_id"] if valid[0]: outputs_to_execute = valid[2] - self.prompt_queue.put((number, prompt_id, prompt, extra_data, outputs_to_execute)) + sensitive = {} + for sensitive_val in execution.SENSITIVE_EXTRA_DATA_KEYS: + if sensitive_val in extra_data: + sensitive[sensitive_val] = extra_data.pop(sensitive_val) + self.prompt_queue.put((number, prompt_id, prompt, extra_data, outputs_to_execute, sensitive)) response = {"prompt_id": prompt_id, "number": number, "node_errors": valid[3]} return web.json_response(response) else: From 3bea4efc6b23d76c6b0672cd90421a9024e13fdb Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 28 Oct 2025 01:45:45 -0700 Subject: [PATCH 042/647] Tell users to update nvidia drivers if problem with portable. (#10510) --- README.md | 2 ++ 1 file changed, 2 insertions(+) diff --git a/README.md b/README.md index 434d4ff06..4204777e9 100644 --- a/README.md +++ b/README.md @@ -176,6 +176,8 @@ Simply download, extract with [7-Zip](https://7-zip.org) and run. Make sure you If you have trouble extracting it, right click the file -> properties -> unblock +Update your Nvidia drivers if it doesn't start. + #### Alternative Downloads: [Experimental portable for AMD GPUs](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_amd.7z) From 22e40d2ace0f53da025b3a41cbe4b664ef807097 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 28 Oct 2025 12:08:08 -0700 Subject: [PATCH 043/647] Tell users to update their nvidia drivers if portable doesn't start. (#10518) --- .../advanced/run_nvidia_gpu_disable_api_nodes.bat | 1 + .ci/windows_nvidia_base_files/run_nvidia_gpu.bat | 1 + .../run_nvidia_gpu_fast_fp16_accumulation.bat | 1 + 3 files changed, 3 insertions(+) diff --git a/.ci/windows_nvidia_base_files/advanced/run_nvidia_gpu_disable_api_nodes.bat b/.ci/windows_nvidia_base_files/advanced/run_nvidia_gpu_disable_api_nodes.bat index cfe4b9f0e..ed00583b6 100644 --- a/.ci/windows_nvidia_base_files/advanced/run_nvidia_gpu_disable_api_nodes.bat +++ b/.ci/windows_nvidia_base_files/advanced/run_nvidia_gpu_disable_api_nodes.bat @@ -1,2 +1,3 @@ ..\python_embeded\python.exe -s ..\ComfyUI\main.py --windows-standalone-build --disable-api-nodes +echo If you see this and ComfyUI did not start try updating your Nvidia Drivers to the latest. pause diff --git a/.ci/windows_nvidia_base_files/run_nvidia_gpu.bat b/.ci/windows_nvidia_base_files/run_nvidia_gpu.bat index 274d7c948..4898a424f 100755 --- a/.ci/windows_nvidia_base_files/run_nvidia_gpu.bat +++ b/.ci/windows_nvidia_base_files/run_nvidia_gpu.bat @@ -1,2 +1,3 @@ .\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build +echo If you see this and ComfyUI did not start try updating your Nvidia Drivers to the latest. pause diff --git a/.ci/windows_nvidia_base_files/run_nvidia_gpu_fast_fp16_accumulation.bat b/.ci/windows_nvidia_base_files/run_nvidia_gpu_fast_fp16_accumulation.bat index 38f06ecb2..32611e4af 100644 --- a/.ci/windows_nvidia_base_files/run_nvidia_gpu_fast_fp16_accumulation.bat +++ b/.ci/windows_nvidia_base_files/run_nvidia_gpu_fast_fp16_accumulation.bat @@ -1,2 +1,3 @@ .\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --fast fp16_accumulation +echo If you see this and ComfyUI did not start try updating your Nvidia Drivers to the latest. pause From 8817f8fc148c5a63ffd3f854975df8e72c740540 Mon Sep 17 00:00:00 2001 From: contentis Date: Tue, 28 Oct 2025 21:20:53 +0100 Subject: [PATCH 044/647] Mixed Precision Quantization System (#10498) * Implement mixed precision operations with a registry design and metadate for quant spec in checkpoint. * Updated design using Tensor Subclasses * Fix FP8 MM * An actually functional POC * Remove CK reference and ensure correct compute dtype * Update unit tests * ruff lint * Implement mixed precision operations with a registry design and metadate for quant spec in checkpoint. * Updated design using Tensor Subclasses * Fix FP8 MM * An actually functional POC * Remove CK reference and ensure correct compute dtype * Update unit tests * ruff lint * Fix missing keys * Rename quant dtype parameter * Rename quant dtype parameter * Fix unittests for CPU build --- comfy/model_base.py | 10 +- comfy/model_detection.py | 20 + comfy/ops.py | 146 +++++- comfy/quant_ops.py | 437 ++++++++++++++++++ comfy/sd.py | 13 +- comfy/supported_models_base.py | 1 + .../comfy_quant/test_mixed_precision.py | 232 ++++++++++ tests-unit/comfy_quant/test_quant_registry.py | 190 ++++++++ 8 files changed, 1030 insertions(+), 19 deletions(-) create mode 100644 comfy/quant_ops.py create mode 100644 tests-unit/comfy_quant/test_mixed_precision.py create mode 100644 tests-unit/comfy_quant/test_quant_registry.py diff --git a/comfy/model_base.py b/comfy/model_base.py index e877f19ac..7c788d085 100644 --- a/comfy/model_base.py +++ b/comfy/model_base.py @@ -134,7 +134,7 @@ class BaseModel(torch.nn.Module): if not unet_config.get("disable_unet_model_creation", False): if model_config.custom_operations is None: fp8 = model_config.optimizations.get("fp8", False) - operations = comfy.ops.pick_operations(unet_config.get("dtype", None), self.manual_cast_dtype, fp8_optimizations=fp8, scaled_fp8=model_config.scaled_fp8) + operations = comfy.ops.pick_operations(unet_config.get("dtype", None), self.manual_cast_dtype, fp8_optimizations=fp8, scaled_fp8=model_config.scaled_fp8, model_config=model_config) else: operations = model_config.custom_operations self.diffusion_model = unet_model(**unet_config, device=device, operations=operations) @@ -333,6 +333,14 @@ class BaseModel(torch.nn.Module): if self.model_config.scaled_fp8 is not None: unet_state_dict["scaled_fp8"] = torch.tensor([], dtype=self.model_config.scaled_fp8) + # Save mixed precision metadata + if hasattr(self.model_config, 'layer_quant_config') and self.model_config.layer_quant_config: + metadata = { + "format_version": "1.0", + "layers": self.model_config.layer_quant_config + } + unet_state_dict["_quantization_metadata"] = metadata + unet_state_dict = self.model_config.process_unet_state_dict_for_saving(unet_state_dict) if self.model_type == ModelType.V_PREDICTION: diff --git a/comfy/model_detection.py b/comfy/model_detection.py index 141f1e164..3142a7fc3 100644 --- a/comfy/model_detection.py +++ b/comfy/model_detection.py @@ -6,6 +6,20 @@ import math import logging import torch + +def detect_layer_quantization(metadata): + quant_key = "_quantization_metadata" + if metadata is not None and quant_key in metadata: + quant_metadata = metadata.pop(quant_key) + quant_metadata = json.loads(quant_metadata) + if isinstance(quant_metadata, dict) and "layers" in quant_metadata: + logging.info(f"Found quantization metadata (version {quant_metadata.get('format_version', 'unknown')})") + return quant_metadata["layers"] + else: + raise ValueError("Invalid quantization metadata format") + return None + + def count_blocks(state_dict_keys, prefix_string): count = 0 while True: @@ -701,6 +715,12 @@ def model_config_from_unet(state_dict, unet_key_prefix, use_base_if_no_match=Fal else: model_config.optimizations["fp8"] = True + # Detect per-layer quantization (mixed precision) + layer_quant_config = detect_layer_quantization(metadata) + if layer_quant_config: + model_config.layer_quant_config = layer_quant_config + logging.info(f"Detected mixed precision quantization: {len(layer_quant_config)} layers quantized") + return model_config def unet_prefix_from_state_dict(state_dict): diff --git a/comfy/ops.py b/comfy/ops.py index 934e21261..93731eedf 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -344,6 +344,10 @@ class manual_cast(disable_weight_init): def fp8_linear(self, input): + """ + Legacy FP8 linear function for backward compatibility. + Uses QuantizedTensor subclass for dispatch. + """ dtype = self.weight.dtype if dtype not in [torch.float8_e4m3fn]: return None @@ -355,9 +359,9 @@ def fp8_linear(self, input): input_shape = input.shape input_dtype = input.dtype + if len(input.shape) == 3: w, bias = cast_bias_weight(self, input, dtype=dtype, bias_dtype=input_dtype) - w = w.t() scale_weight = self.scale_weight scale_input = self.scale_input @@ -368,23 +372,18 @@ def fp8_linear(self, input): if scale_input is None: scale_input = torch.ones((), device=input.device, dtype=torch.float32) - input = torch.clamp(input, min=-448, max=448, out=input) - input = input.reshape(-1, input_shape[2]).to(dtype).contiguous() else: scale_input = scale_input.to(input.device) - input = (input * (1.0 / scale_input).to(input_dtype)).reshape(-1, input_shape[2]).to(dtype).contiguous() - if bias is not None: - o = torch._scaled_mm(input, w, out_dtype=input_dtype, bias=bias, scale_a=scale_input, scale_b=scale_weight) - else: - o = torch._scaled_mm(input, w, out_dtype=input_dtype, scale_a=scale_input, scale_b=scale_weight) - - if isinstance(o, tuple): - o = o[0] + # Wrap weight in QuantizedTensor - this enables unified dispatch + # Call F.linear - __torch_dispatch__ routes to fp8_linear handler in quant_ops.py! + layout_params_weight = {'scale': scale_weight, 'orig_dtype': input_dtype} + quantized_weight = QuantizedTensor(w, TensorCoreFP8Layout, layout_params_weight) + quantized_input = QuantizedTensor.from_float(input.reshape(-1, input_shape[2]), TensorCoreFP8Layout, scale=scale_input, dtype=dtype) + o = torch.nn.functional.linear(quantized_input, quantized_weight, bias) if tensor_2d: return o.reshape(input_shape[0], -1) - return o.reshape((-1, input_shape[1], self.weight.shape[0])) return None @@ -478,7 +477,128 @@ if CUBLAS_IS_AVAILABLE: def forward(self, *args, **kwargs): return super().forward(*args, **kwargs) -def pick_operations(weight_dtype, compute_dtype, load_device=None, disable_fast_fp8=False, fp8_optimizations=False, scaled_fp8=None): + +# ============================================================================== +# Mixed Precision Operations +# ============================================================================== +from .quant_ops import QuantizedTensor, TensorCoreFP8Layout + +QUANT_FORMAT_MIXINS = { + "float8_e4m3fn": { + "dtype": torch.float8_e4m3fn, + "layout_type": TensorCoreFP8Layout, + "parameters": { + "weight_scale": torch.nn.Parameter(torch.zeros((), dtype=torch.float32), requires_grad=False), + "input_scale": torch.nn.Parameter(torch.zeros((), dtype=torch.float32), requires_grad=False), + } + } +} + +class MixedPrecisionOps(disable_weight_init): + _layer_quant_config = {} + _compute_dtype = torch.bfloat16 + + class Linear(torch.nn.Module, CastWeightBiasOp): + def __init__( + self, + in_features: int, + out_features: int, + bias: bool = True, + device=None, + dtype=None, + ) -> None: + super().__init__() + + self.factory_kwargs = {"device": device, "dtype": MixedPrecisionOps._compute_dtype} + # self.factory_kwargs = {"device": device, "dtype": dtype} + + self.in_features = in_features + self.out_features = out_features + if bias: + self.bias = torch.nn.Parameter(torch.empty(out_features, **self.factory_kwargs)) + else: + self.register_parameter("bias", None) + + self.tensor_class = None + + def reset_parameters(self): + return None + + def _load_from_state_dict(self, state_dict, prefix, local_metadata, + strict, missing_keys, unexpected_keys, error_msgs): + + device = self.factory_kwargs["device"] + layer_name = prefix.rstrip('.') + weight_key = f"{prefix}weight" + weight = state_dict.pop(weight_key, None) + if weight is None: + raise ValueError(f"Missing weight for layer {layer_name}") + + manually_loaded_keys = [weight_key] + + if layer_name not in MixedPrecisionOps._layer_quant_config: + self.weight = torch.nn.Parameter(weight.to(device=device, dtype=MixedPrecisionOps._compute_dtype), requires_grad=False) + else: + quant_format = MixedPrecisionOps._layer_quant_config[layer_name].get("format", None) + if quant_format is None: + raise ValueError(f"Unknown quantization format for layer {layer_name}") + + mixin = QUANT_FORMAT_MIXINS[quant_format] + self.layout_type = mixin["layout_type"] + + scale_key = f"{prefix}weight_scale" + layout_params = { + 'scale': state_dict.pop(scale_key, None), + 'orig_dtype': MixedPrecisionOps._compute_dtype + } + if layout_params['scale'] is not None: + manually_loaded_keys.append(scale_key) + + self.weight = torch.nn.Parameter( + QuantizedTensor(weight.to(device=device, dtype=mixin["dtype"]), self.layout_type, layout_params), + requires_grad=False + ) + + for param_name, param_value in mixin["parameters"].items(): + param_key = f"{prefix}{param_name}" + _v = state_dict.pop(param_key, None) + if _v is None: + continue + setattr(self, param_name, torch.nn.Parameter(_v.to(device=device), requires_grad=False)) + manually_loaded_keys.append(param_key) + + super()._load_from_state_dict(state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs) + + for key in manually_loaded_keys: + if key in missing_keys: + missing_keys.remove(key) + + def _forward(self, input, weight, bias): + return torch.nn.functional.linear(input, weight, bias) + + def forward_comfy_cast_weights(self, input): + weight, bias = cast_bias_weight(self, input) + return self._forward(input, weight, bias) + + def forward(self, input, *args, **kwargs): + run_every_op() + + if self.comfy_cast_weights or len(self.weight_function) > 0 or len(self.bias_function) > 0: + return self.forward_comfy_cast_weights(input, *args, **kwargs) + if (getattr(self, 'layout_type', None) is not None and + getattr(self, 'input_scale', None) is not None and + not isinstance(input, QuantizedTensor)): + input = QuantizedTensor.from_float(input, self.layout_type, scale=self.input_scale, fp8_dtype=self.weight.dtype) + return self._forward(input, self.weight, self.bias) + + +def pick_operations(weight_dtype, compute_dtype, load_device=None, disable_fast_fp8=False, fp8_optimizations=False, scaled_fp8=None, model_config=None): + if model_config and hasattr(model_config, 'layer_quant_config') and model_config.layer_quant_config: + MixedPrecisionOps._layer_quant_config = model_config.layer_quant_config + MixedPrecisionOps._compute_dtype = compute_dtype + logging.info(f"Using mixed precision operations: {len(model_config.layer_quant_config)} quantized layers") + return MixedPrecisionOps + fp8_compute = comfy.model_management.supports_fp8_compute(load_device) if scaled_fp8 is not None: return scaled_fp8_ops(fp8_matrix_mult=fp8_compute and fp8_optimizations, scale_input=fp8_optimizations, override_dtype=scaled_fp8) diff --git a/comfy/quant_ops.py b/comfy/quant_ops.py new file mode 100644 index 000000000..b14e03084 --- /dev/null +++ b/comfy/quant_ops.py @@ -0,0 +1,437 @@ +import torch +import logging +from typing import Tuple, Dict + +_LAYOUT_REGISTRY = {} +_GENERIC_UTILS = {} + + +def register_layout_op(torch_op, layout_type): + """ + Decorator to register a layout-specific operation handler. + Args: + torch_op: PyTorch operation (e.g., torch.ops.aten.linear.default) + layout_type: Layout class (e.g., TensorCoreFP8Layout) + Example: + @register_layout_op(torch.ops.aten.linear.default, TensorCoreFP8Layout) + def fp8_linear(func, args, kwargs): + # FP8-specific linear implementation + ... + """ + def decorator(handler_func): + if torch_op not in _LAYOUT_REGISTRY: + _LAYOUT_REGISTRY[torch_op] = {} + _LAYOUT_REGISTRY[torch_op][layout_type] = handler_func + return handler_func + return decorator + + +def register_generic_util(torch_op): + """ + Decorator to register a generic utility that works for all layouts. + Args: + torch_op: PyTorch operation (e.g., torch.ops.aten.detach.default) + + Example: + @register_generic_util(torch.ops.aten.detach.default) + def generic_detach(func, args, kwargs): + # Works for any layout + ... + """ + def decorator(handler_func): + _GENERIC_UTILS[torch_op] = handler_func + return handler_func + return decorator + + +def _get_layout_from_args(args): + for arg in args: + if isinstance(arg, QuantizedTensor): + return arg._layout_type + elif isinstance(arg, (list, tuple)): + for item in arg: + if isinstance(item, QuantizedTensor): + return item._layout_type + return None + + +def _move_layout_params_to_device(params, device): + new_params = {} + for k, v in params.items(): + if isinstance(v, torch.Tensor): + new_params[k] = v.to(device=device) + else: + new_params[k] = v + return new_params + + +def _copy_layout_params(params): + new_params = {} + for k, v in params.items(): + if isinstance(v, torch.Tensor): + new_params[k] = v.clone() + else: + new_params[k] = v + return new_params + + +class QuantizedLayout: + """ + Base class for quantization layouts. + + A layout encapsulates the format-specific logic for quantization/dequantization + and provides a uniform interface for extracting raw tensors needed for computation. + + New quantization formats should subclass this and implement the required methods. + """ + @classmethod + def quantize(cls, tensor, **kwargs) -> Tuple[torch.Tensor, Dict]: + raise NotImplementedError(f"{cls.__name__} must implement quantize()") + + @staticmethod + def dequantize(qdata, **layout_params) -> torch.Tensor: + raise NotImplementedError("TensorLayout must implement dequantize()") + + @classmethod + def get_plain_tensors(cls, qtensor) -> torch.Tensor: + raise NotImplementedError(f"{cls.__name__} must implement get_plain_tensors()") + + +class QuantizedTensor(torch.Tensor): + """ + Universal quantized tensor that works with any layout. + + This tensor subclass uses a pluggable layout system to support multiple + quantization formats (FP8, INT4, INT8, etc.) without code duplication. + + The layout_type determines format-specific behavior, while common operations + (detach, clone, to) are handled generically. + + Attributes: + _qdata: The quantized tensor data + _layout_type: Layout class (e.g., TensorCoreFP8Layout) + _layout_params: Dict with layout-specific params (scale, zero_point, etc.) + """ + + @staticmethod + def __new__(cls, qdata, layout_type, layout_params): + """ + Create a quantized tensor. + + Args: + qdata: The quantized data tensor + layout_type: Layout class (subclass of QuantizedLayout) + layout_params: Dict with layout-specific parameters + """ + return torch.Tensor._make_subclass(cls, qdata, require_grad=False) + + def __init__(self, qdata, layout_type, layout_params): + self._qdata = qdata.contiguous() + self._layout_type = layout_type + self._layout_params = layout_params + + def __repr__(self): + layout_name = self._layout_type.__name__ + param_str = ", ".join(f"{k}={v}" for k, v in list(self._layout_params.items())[:2]) + return f"QuantizedTensor(shape={self.shape}, layout={layout_name}, {param_str})" + + @property + def layout_type(self): + return self._layout_type + + def __tensor_flatten__(self): + """ + Tensor flattening protocol for proper device movement. + """ + inner_tensors = ["_qdata"] + ctx = { + "layout_type": self._layout_type, + } + + tensor_params = {} + non_tensor_params = {} + for k, v in self._layout_params.items(): + if isinstance(v, torch.Tensor): + tensor_params[k] = v + else: + non_tensor_params[k] = v + + ctx["tensor_param_keys"] = list(tensor_params.keys()) + ctx["non_tensor_params"] = non_tensor_params + + for k, v in tensor_params.items(): + attr_name = f"_layout_param_{k}" + object.__setattr__(self, attr_name, v) + inner_tensors.append(attr_name) + + return inner_tensors, ctx + + @staticmethod + def __tensor_unflatten__(inner_tensors, ctx, outer_size, outer_stride): + """ + Tensor unflattening protocol for proper device movement. + Reconstructs the QuantizedTensor after device movement. + """ + layout_type = ctx["layout_type"] + layout_params = dict(ctx["non_tensor_params"]) + + for key in ctx["tensor_param_keys"]: + attr_name = f"_layout_param_{key}" + layout_params[key] = inner_tensors[attr_name] + + return QuantizedTensor(inner_tensors["_q_data"], layout_type, layout_params) + + @classmethod + def from_float(cls, tensor, layout_type, **quantize_kwargs) -> 'QuantizedTensor': + qdata, layout_params = layout_type.quantize(tensor, **quantize_kwargs) + return cls(qdata, layout_type, layout_params) + + def dequantize(self) -> torch.Tensor: + return self._layout_type.dequantize(self._qdata, **self._layout_params) + + @classmethod + def __torch_dispatch__(cls, func, types, args=(), kwargs=None): + kwargs = kwargs or {} + + # Step 1: Check generic utilities first (detach, clone, to, etc.) + if func in _GENERIC_UTILS: + return _GENERIC_UTILS[func](func, args, kwargs) + + # Step 2: Check layout-specific handlers (linear, matmul, etc.) + layout_type = _get_layout_from_args(args) + if layout_type and func in _LAYOUT_REGISTRY: + handler = _LAYOUT_REGISTRY[func].get(layout_type) + if handler: + return handler(func, args, kwargs) + + # Step 3: Fallback to dequantization + if isinstance(args[0] if args else None, QuantizedTensor): + logging.info(f"QuantizedTensor: Unhandled operation {func}, falling back to dequantization. kwargs={kwargs}") + return cls._dequant_and_fallback(func, args, kwargs) + + @classmethod + def _dequant_and_fallback(cls, func, args, kwargs): + def dequant_arg(arg): + if isinstance(arg, QuantizedTensor): + return arg.dequantize() + elif isinstance(arg, (list, tuple)): + return type(arg)(dequant_arg(a) for a in arg) + return arg + + new_args = dequant_arg(args) + new_kwargs = dequant_arg(kwargs) + return func(*new_args, **new_kwargs) + + +# ============================================================================== +# Generic Utilities (Layout-Agnostic Operations) +# ============================================================================== + +def _create_transformed_qtensor(qt, transform_fn): + new_data = transform_fn(qt._qdata) + new_params = _copy_layout_params(qt._layout_params) + return QuantizedTensor(new_data, qt._layout_type, new_params) + + +def _handle_device_transfer(qt, target_device, target_dtype=None, target_layout=None, op_name="to"): + if target_dtype is not None and target_dtype != qt.dtype: + logging.warning( + f"QuantizedTensor: dtype conversion requested to {target_dtype}, " + f"but not supported for quantized tensors. Ignoring dtype." + ) + + if target_layout is not None and target_layout != torch.strided: + logging.warning( + f"QuantizedTensor: layout change requested to {target_layout}, " + f"but not supported. Ignoring layout." + ) + + # Handle device transfer + current_device = qt._qdata.device + if target_device is not None: + # Normalize device for comparison + if isinstance(target_device, str): + target_device = torch.device(target_device) + if isinstance(current_device, str): + current_device = torch.device(current_device) + + if target_device != current_device: + logging.debug(f"QuantizedTensor.{op_name}: Moving from {current_device} to {target_device}") + new_q_data = qt._qdata.to(device=target_device) + new_params = _move_layout_params_to_device(qt._layout_params, target_device) + new_qt = QuantizedTensor(new_q_data, qt._layout_type, new_params) + logging.debug(f"QuantizedTensor.{op_name}: Created new tensor on {target_device}") + return new_qt + + logging.debug(f"QuantizedTensor.{op_name}: No device change needed, returning original") + return qt + + +@register_generic_util(torch.ops.aten.detach.default) +def generic_detach(func, args, kwargs): + """Detach operation - creates a detached copy of the quantized tensor.""" + qt = args[0] + if isinstance(qt, QuantizedTensor): + return _create_transformed_qtensor(qt, lambda x: x.detach()) + return func(*args, **kwargs) + + +@register_generic_util(torch.ops.aten.clone.default) +def generic_clone(func, args, kwargs): + """Clone operation - creates a deep copy of the quantized tensor.""" + qt = args[0] + if isinstance(qt, QuantizedTensor): + return _create_transformed_qtensor(qt, lambda x: x.clone()) + return func(*args, **kwargs) + + +@register_generic_util(torch.ops.aten._to_copy.default) +def generic_to_copy(func, args, kwargs): + """Device/dtype transfer operation - handles .to(device) calls.""" + qt = args[0] + if isinstance(qt, QuantizedTensor): + return _handle_device_transfer( + qt, + target_device=kwargs.get('device', None), + target_dtype=kwargs.get('dtype', None), + op_name="_to_copy" + ) + return func(*args, **kwargs) + + +@register_generic_util(torch.ops.aten.to.dtype_layout) +def generic_to_dtype_layout(func, args, kwargs): + """Handle .to(device) calls using the dtype_layout variant.""" + qt = args[0] + if isinstance(qt, QuantizedTensor): + return _handle_device_transfer( + qt, + target_device=kwargs.get('device', None), + target_dtype=kwargs.get('dtype', None), + target_layout=kwargs.get('layout', None), + op_name="to" + ) + return func(*args, **kwargs) + + +@register_generic_util(torch.ops.aten.copy_.default) +def generic_copy_(func, args, kwargs): + qt_dest = args[0] + src = args[1] + + if isinstance(qt_dest, QuantizedTensor): + if isinstance(src, QuantizedTensor): + # Copy from another quantized tensor + qt_dest._qdata.copy_(src._qdata) + qt_dest._layout_type = src._layout_type + qt_dest._layout_params = _copy_layout_params(src._layout_params) + else: + # Copy from regular tensor - just copy raw data + qt_dest._qdata.copy_(src) + return qt_dest + return func(*args, **kwargs) + + +@register_generic_util(torch.ops.aten._has_compatible_shallow_copy_type.default) +def generic_has_compatible_shallow_copy_type(func, args, kwargs): + return True + +# ============================================================================== +# FP8 Layout + Operation Handlers +# ============================================================================== +class TensorCoreFP8Layout(QuantizedLayout): + """ + Storage format: + - qdata: FP8 tensor (torch.float8_e4m3fn or torch.float8_e5m2) + - scale: Scalar tensor (float32) for dequantization + - orig_dtype: Original dtype before quantization (for casting back) + """ + @classmethod + def quantize(cls, tensor, scale=None, dtype=torch.float8_e4m3fn): + orig_dtype = tensor.dtype + + if scale is None: + scale = torch.amax(tensor.abs()) / torch.finfo(dtype).max + + if not isinstance(scale, torch.Tensor): + scale = torch.tensor(scale) + scale = scale.to(device=tensor.device, dtype=torch.float32) + + lp_amax = torch.finfo(dtype).max + tensor_scaled = tensor.float() / scale + torch.clamp(tensor_scaled, min=-lp_amax, max=lp_amax, out=tensor_scaled) + qdata = tensor_scaled.to(dtype, memory_format=torch.contiguous_format) + + layout_params = { + 'scale': scale, + 'orig_dtype': orig_dtype + } + return qdata, layout_params + + @staticmethod + def dequantize(qdata, scale, orig_dtype, **kwargs): + plain_tensor = torch.ops.aten._to_copy.default(qdata, dtype=orig_dtype) + return plain_tensor * scale + + @classmethod + def get_plain_tensors(cls, qtensor): + return qtensor._qdata, qtensor._layout_params['scale'] + + +@register_layout_op(torch.ops.aten.linear.default, TensorCoreFP8Layout) +def fp8_linear(func, args, kwargs): + input_tensor = args[0] + weight = args[1] + bias = args[2] if len(args) > 2 else None + + if isinstance(input_tensor, QuantizedTensor) and isinstance(weight, QuantizedTensor): + plain_input, scale_a = TensorCoreFP8Layout.get_plain_tensors(input_tensor) + plain_weight, scale_b = TensorCoreFP8Layout.get_plain_tensors(weight) + + out_dtype = kwargs.get("out_dtype") + if out_dtype is None: + out_dtype = input_tensor._layout_params['orig_dtype'] + + weight_t = plain_weight.t() + + tensor_2d = False + if len(plain_input.shape) == 2: + tensor_2d = True + plain_input = plain_input.unsqueeze(1) + + input_shape = plain_input.shape + if len(input_shape) != 3: + return None + + try: + output = torch._scaled_mm( + plain_input.reshape(-1, input_shape[2]), + weight_t, + bias=bias, + scale_a=scale_a, + scale_b=scale_b, + out_dtype=out_dtype, + ) + if not tensor_2d: + output = output.reshape((-1, input_shape[1], weight.shape[0])) + + if output.dtype in [torch.float8_e4m3fn, torch.float8_e5m2]: + output_scale = scale_a * scale_b + output_params = { + 'scale': output_scale, + 'orig_dtype': input_tensor._layout_params['orig_dtype'] + } + return QuantizedTensor(output, TensorCoreFP8Layout, output_params) + else: + return output + + except Exception as e: + raise RuntimeError(f"FP8 _scaled_mm failed, falling back to dequantization: {e}") + + # Case 2: DQ Fallback + if isinstance(weight, QuantizedTensor): + weight = weight.dequantize() + if isinstance(input_tensor, QuantizedTensor): + input_tensor = input_tensor.dequantize() + + return torch.nn.functional.linear(input_tensor, weight, bias) diff --git a/comfy/sd.py b/comfy/sd.py index 28bee248d..6411bb27d 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -1262,7 +1262,7 @@ def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_c return (model_patcher, clip, vae, clipvision) -def load_diffusion_model_state_dict(sd, model_options={}): +def load_diffusion_model_state_dict(sd, model_options={}, metadata=None): """ Loads a UNet diffusion model from a state dictionary, supporting both diffusers and regular formats. @@ -1296,7 +1296,7 @@ def load_diffusion_model_state_dict(sd, model_options={}): weight_dtype = comfy.utils.weight_dtype(sd) load_device = model_management.get_torch_device() - model_config = model_detection.model_config_from_unet(sd, "") + model_config = model_detection.model_config_from_unet(sd, "", metadata=metadata) if model_config is not None: new_sd = sd @@ -1330,7 +1330,10 @@ def load_diffusion_model_state_dict(sd, model_options={}): else: unet_dtype = dtype - manual_cast_dtype = model_management.unet_manual_cast(unet_dtype, load_device, model_config.supported_inference_dtypes) + if hasattr(model_config, "layer_quant_config"): + manual_cast_dtype = model_management.unet_manual_cast(None, load_device, model_config.supported_inference_dtypes) + else: + manual_cast_dtype = model_management.unet_manual_cast(unet_dtype, load_device, model_config.supported_inference_dtypes) model_config.set_inference_dtype(unet_dtype, manual_cast_dtype) model_config.custom_operations = model_options.get("custom_operations", model_config.custom_operations) if model_options.get("fp8_optimizations", False): @@ -1346,8 +1349,8 @@ def load_diffusion_model_state_dict(sd, model_options={}): def load_diffusion_model(unet_path, model_options={}): - sd = comfy.utils.load_torch_file(unet_path) - model = load_diffusion_model_state_dict(sd, model_options=model_options) + sd, metadata = comfy.utils.load_torch_file(unet_path, return_metadata=True) + model = load_diffusion_model_state_dict(sd, model_options=model_options, metadata=metadata) if model is None: logging.error("ERROR UNSUPPORTED DIFFUSION MODEL {}".format(unet_path)) raise RuntimeError("ERROR: Could not detect model type of: {}\n{}".format(unet_path, model_detection_error_hint(unet_path, sd))) diff --git a/comfy/supported_models_base.py b/comfy/supported_models_base.py index 54573abb1..e4bd74514 100644 --- a/comfy/supported_models_base.py +++ b/comfy/supported_models_base.py @@ -50,6 +50,7 @@ class BASE: manual_cast_dtype = None custom_operations = None scaled_fp8 = None + layer_quant_config = None # Per-layer quantization configuration for mixed precision optimizations = {"fp8": False} @classmethod diff --git a/tests-unit/comfy_quant/test_mixed_precision.py b/tests-unit/comfy_quant/test_mixed_precision.py new file mode 100644 index 000000000..267bc177b --- /dev/null +++ b/tests-unit/comfy_quant/test_mixed_precision.py @@ -0,0 +1,232 @@ +import unittest +import torch +import sys +import os + +# Add comfy to path +sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", "..")) + +def has_gpu(): + return torch.cuda.is_available() + +from comfy.cli_args import args +if not has_gpu(): + args.cpu = True + +from comfy import ops +from comfy.quant_ops import QuantizedTensor, TensorCoreFP8Layout + + +class SimpleModel(torch.nn.Module): + def __init__(self, operations=ops.disable_weight_init): + super().__init__() + self.layer1 = operations.Linear(10, 20, device="cpu", dtype=torch.bfloat16) + self.layer2 = operations.Linear(20, 30, device="cpu", dtype=torch.bfloat16) + self.layer3 = operations.Linear(30, 40, device="cpu", dtype=torch.bfloat16) + + def forward(self, x): + x = self.layer1(x) + x = torch.nn.functional.relu(x) + x = self.layer2(x) + x = torch.nn.functional.relu(x) + x = self.layer3(x) + return x + + +class TestMixedPrecisionOps(unittest.TestCase): + + def test_all_layers_standard(self): + """Test that model with no quantization works normally""" + # Configure no quantization + ops.MixedPrecisionOps._layer_quant_config = {} + + # Create model + model = SimpleModel(operations=ops.MixedPrecisionOps) + + # Initialize weights manually + model.layer1.weight = torch.nn.Parameter(torch.randn(20, 10, dtype=torch.bfloat16)) + model.layer1.bias = torch.nn.Parameter(torch.randn(20, dtype=torch.bfloat16)) + model.layer2.weight = torch.nn.Parameter(torch.randn(30, 20, dtype=torch.bfloat16)) + model.layer2.bias = torch.nn.Parameter(torch.randn(30, dtype=torch.bfloat16)) + model.layer3.weight = torch.nn.Parameter(torch.randn(40, 30, dtype=torch.bfloat16)) + model.layer3.bias = torch.nn.Parameter(torch.randn(40, dtype=torch.bfloat16)) + + # Initialize weight_function and bias_function + for layer in [model.layer1, model.layer2, model.layer3]: + layer.weight_function = [] + layer.bias_function = [] + + # Forward pass + input_tensor = torch.randn(5, 10, dtype=torch.bfloat16) + output = model(input_tensor) + + self.assertEqual(output.shape, (5, 40)) + self.assertEqual(output.dtype, torch.bfloat16) + + def test_mixed_precision_load(self): + """Test loading a mixed precision model from state dict""" + # Configure mixed precision: layer1 is FP8, layer2 and layer3 are standard + layer_quant_config = { + "layer1": { + "format": "float8_e4m3fn", + "params": {} + }, + "layer3": { + "format": "float8_e4m3fn", + "params": {} + } + } + ops.MixedPrecisionOps._layer_quant_config = layer_quant_config + + # Create state dict with mixed precision + fp8_weight1 = torch.randn(20, 10, dtype=torch.float32).to(torch.float8_e4m3fn) + fp8_weight3 = torch.randn(40, 30, dtype=torch.float32).to(torch.float8_e4m3fn) + + state_dict = { + # Layer 1: FP8 E4M3FN + "layer1.weight": fp8_weight1, + "layer1.bias": torch.randn(20, dtype=torch.bfloat16), + "layer1.weight_scale": torch.tensor(2.0, dtype=torch.float32), + + # Layer 2: Standard BF16 + "layer2.weight": torch.randn(30, 20, dtype=torch.bfloat16), + "layer2.bias": torch.randn(30, dtype=torch.bfloat16), + + # Layer 3: FP8 E4M3FN + "layer3.weight": fp8_weight3, + "layer3.bias": torch.randn(40, dtype=torch.bfloat16), + "layer3.weight_scale": torch.tensor(1.5, dtype=torch.float32), + } + + # Create model and load state dict (strict=False because custom loading pops keys) + model = SimpleModel(operations=ops.MixedPrecisionOps) + model.load_state_dict(state_dict, strict=False) + + # Verify weights are wrapped in QuantizedTensor + self.assertIsInstance(model.layer1.weight, QuantizedTensor) + self.assertEqual(model.layer1.weight._layout_type, TensorCoreFP8Layout) + + # Layer 2 should NOT be quantized + self.assertNotIsInstance(model.layer2.weight, QuantizedTensor) + + # Layer 3 should be quantized + self.assertIsInstance(model.layer3.weight, QuantizedTensor) + self.assertEqual(model.layer3.weight._layout_type, TensorCoreFP8Layout) + + # Verify scales were loaded + self.assertEqual(model.layer1.weight._layout_params['scale'].item(), 2.0) + self.assertEqual(model.layer3.weight._layout_params['scale'].item(), 1.5) + + # Forward pass + input_tensor = torch.randn(5, 10, dtype=torch.bfloat16) + output = model(input_tensor) + + self.assertEqual(output.shape, (5, 40)) + + def test_state_dict_quantized_preserved(self): + """Test that quantized weights are preserved in state_dict()""" + # Configure mixed precision + layer_quant_config = { + "layer1": { + "format": "float8_e4m3fn", + "params": {} + } + } + ops.MixedPrecisionOps._layer_quant_config = layer_quant_config + + # Create and load model + fp8_weight = torch.randn(20, 10, dtype=torch.float32).to(torch.float8_e4m3fn) + state_dict1 = { + "layer1.weight": fp8_weight, + "layer1.bias": torch.randn(20, dtype=torch.bfloat16), + "layer1.weight_scale": torch.tensor(3.0, dtype=torch.float32), + "layer2.weight": torch.randn(30, 20, dtype=torch.bfloat16), + "layer2.bias": torch.randn(30, dtype=torch.bfloat16), + "layer3.weight": torch.randn(40, 30, dtype=torch.bfloat16), + "layer3.bias": torch.randn(40, dtype=torch.bfloat16), + } + + model = SimpleModel(operations=ops.MixedPrecisionOps) + model.load_state_dict(state_dict1, strict=False) + + # Save state dict + state_dict2 = model.state_dict() + + # Verify layer1.weight is a QuantizedTensor with scale preserved + self.assertIsInstance(state_dict2["layer1.weight"], QuantizedTensor) + self.assertEqual(state_dict2["layer1.weight"]._layout_params['scale'].item(), 3.0) + self.assertEqual(state_dict2["layer1.weight"]._layout_type, TensorCoreFP8Layout) + + # Verify non-quantized layers are standard tensors + self.assertNotIsInstance(state_dict2["layer2.weight"], QuantizedTensor) + self.assertNotIsInstance(state_dict2["layer3.weight"], QuantizedTensor) + + def test_weight_function_compatibility(self): + """Test that weight_function (LoRA) works with quantized layers""" + # Configure FP8 quantization + layer_quant_config = { + "layer1": { + "format": "float8_e4m3fn", + "params": {} + } + } + ops.MixedPrecisionOps._layer_quant_config = layer_quant_config + + # Create and load model + fp8_weight = torch.randn(20, 10, dtype=torch.float32).to(torch.float8_e4m3fn) + state_dict = { + "layer1.weight": fp8_weight, + "layer1.bias": torch.randn(20, dtype=torch.bfloat16), + "layer1.weight_scale": torch.tensor(2.0, dtype=torch.float32), + "layer2.weight": torch.randn(30, 20, dtype=torch.bfloat16), + "layer2.bias": torch.randn(30, dtype=torch.bfloat16), + "layer3.weight": torch.randn(40, 30, dtype=torch.bfloat16), + "layer3.bias": torch.randn(40, dtype=torch.bfloat16), + } + + model = SimpleModel(operations=ops.MixedPrecisionOps) + model.load_state_dict(state_dict, strict=False) + + # Add a weight function (simulating LoRA) + # This should trigger dequantization during forward pass + def apply_lora(weight): + lora_delta = torch.randn_like(weight) * 0.01 + return weight + lora_delta + + model.layer1.weight_function.append(apply_lora) + + # Forward pass should work with LoRA (triggers weight_function path) + input_tensor = torch.randn(5, 10, dtype=torch.bfloat16) + output = model(input_tensor) + + self.assertEqual(output.shape, (5, 40)) + + def test_error_handling_unknown_format(self): + """Test that unknown formats raise error""" + # Configure with unknown format + layer_quant_config = { + "layer1": { + "format": "unknown_format_xyz", + "params": {} + } + } + ops.MixedPrecisionOps._layer_quant_config = layer_quant_config + + # Create state dict + state_dict = { + "layer1.weight": torch.randn(20, 10, dtype=torch.bfloat16), + "layer1.bias": torch.randn(20, dtype=torch.bfloat16), + "layer2.weight": torch.randn(30, 20, dtype=torch.bfloat16), + "layer2.bias": torch.randn(30, dtype=torch.bfloat16), + "layer3.weight": torch.randn(40, 30, dtype=torch.bfloat16), + "layer3.bias": torch.randn(40, dtype=torch.bfloat16), + } + + # Load should raise KeyError for unknown format in QUANT_FORMAT_MIXINS + model = SimpleModel(operations=ops.MixedPrecisionOps) + with self.assertRaises(KeyError): + model.load_state_dict(state_dict, strict=False) + +if __name__ == "__main__": + unittest.main() + diff --git a/tests-unit/comfy_quant/test_quant_registry.py b/tests-unit/comfy_quant/test_quant_registry.py new file mode 100644 index 000000000..477811029 --- /dev/null +++ b/tests-unit/comfy_quant/test_quant_registry.py @@ -0,0 +1,190 @@ +import unittest +import torch +import sys +import os + +# Add comfy to path +sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", "..")) + +def has_gpu(): + return torch.cuda.is_available() + +from comfy.cli_args import args +if not has_gpu(): + args.cpu = True + +from comfy.quant_ops import QuantizedTensor, TensorCoreFP8Layout + + +class TestQuantizedTensor(unittest.TestCase): + """Test the QuantizedTensor subclass with FP8 layout""" + + def test_creation(self): + """Test creating a QuantizedTensor with TensorCoreFP8Layout""" + fp8_data = torch.randn(256, 128, dtype=torch.float32).to(torch.float8_e4m3fn) + scale = torch.tensor(2.0) + layout_params = {'scale': scale, 'orig_dtype': torch.bfloat16} + + qt = QuantizedTensor(fp8_data, TensorCoreFP8Layout, layout_params) + + self.assertIsInstance(qt, QuantizedTensor) + self.assertEqual(qt.shape, (256, 128)) + self.assertEqual(qt.dtype, torch.float8_e4m3fn) + self.assertEqual(qt._layout_params['scale'], scale) + self.assertEqual(qt._layout_params['orig_dtype'], torch.bfloat16) + self.assertEqual(qt._layout_type, TensorCoreFP8Layout) + + def test_dequantize(self): + """Test explicit dequantization""" + + fp8_data = torch.ones(10, 20, dtype=torch.float32).to(torch.float8_e4m3fn) + scale = torch.tensor(3.0) + layout_params = {'scale': scale, 'orig_dtype': torch.float32} + + qt = QuantizedTensor(fp8_data, TensorCoreFP8Layout, layout_params) + dequantized = qt.dequantize() + + self.assertEqual(dequantized.dtype, torch.float32) + self.assertTrue(torch.allclose(dequantized, torch.ones(10, 20) * 3.0, rtol=0.1)) + + def test_from_float(self): + """Test creating QuantizedTensor from float tensor""" + float_tensor = torch.randn(64, 32, dtype=torch.float32) + scale = torch.tensor(1.5) + + qt = QuantizedTensor.from_float( + float_tensor, + TensorCoreFP8Layout, + scale=scale, + dtype=torch.float8_e4m3fn + ) + + self.assertIsInstance(qt, QuantizedTensor) + self.assertEqual(qt.dtype, torch.float8_e4m3fn) + self.assertEqual(qt.shape, (64, 32)) + + # Verify dequantization gives approximately original values + dequantized = qt.dequantize() + mean_rel_error = ((dequantized - float_tensor).abs() / (float_tensor.abs() + 1e-6)).mean() + self.assertLess(mean_rel_error, 0.1) + + +class TestGenericUtilities(unittest.TestCase): + """Test generic utility operations""" + + def test_detach(self): + """Test detach operation on quantized tensor""" + fp8_data = torch.randn(10, 20, dtype=torch.float32).to(torch.float8_e4m3fn) + scale = torch.tensor(1.5) + layout_params = {'scale': scale, 'orig_dtype': torch.float32} + qt = QuantizedTensor(fp8_data, TensorCoreFP8Layout, layout_params) + + # Detach should return a new QuantizedTensor + qt_detached = qt.detach() + + self.assertIsInstance(qt_detached, QuantizedTensor) + self.assertEqual(qt_detached.shape, qt.shape) + self.assertEqual(qt_detached._layout_type, TensorCoreFP8Layout) + + def test_clone(self): + """Test clone operation on quantized tensor""" + fp8_data = torch.randn(10, 20, dtype=torch.float32).to(torch.float8_e4m3fn) + scale = torch.tensor(1.5) + layout_params = {'scale': scale, 'orig_dtype': torch.float32} + qt = QuantizedTensor(fp8_data, TensorCoreFP8Layout, layout_params) + + # Clone should return a new QuantizedTensor + qt_cloned = qt.clone() + + self.assertIsInstance(qt_cloned, QuantizedTensor) + self.assertEqual(qt_cloned.shape, qt.shape) + self.assertEqual(qt_cloned._layout_type, TensorCoreFP8Layout) + + # Verify it's a deep copy + self.assertIsNot(qt_cloned._qdata, qt._qdata) + + @unittest.skipUnless(has_gpu(), "GPU not available") + def test_to_device(self): + """Test device transfer""" + fp8_data = torch.randn(10, 20, dtype=torch.float32).to(torch.float8_e4m3fn) + scale = torch.tensor(1.5) + layout_params = {'scale': scale, 'orig_dtype': torch.float32} + qt = QuantizedTensor(fp8_data, TensorCoreFP8Layout, layout_params) + + # Moving to same device should work (CPU to CPU) + qt_cpu = qt.to('cpu') + + self.assertIsInstance(qt_cpu, QuantizedTensor) + self.assertEqual(qt_cpu.device.type, 'cpu') + self.assertEqual(qt_cpu._layout_params['scale'].device.type, 'cpu') + + +class TestTensorCoreFP8Layout(unittest.TestCase): + """Test the TensorCoreFP8Layout implementation""" + + def test_quantize(self): + """Test quantization method""" + float_tensor = torch.randn(32, 64, dtype=torch.float32) + scale = torch.tensor(1.5) + + qdata, layout_params = TensorCoreFP8Layout.quantize( + float_tensor, + scale=scale, + dtype=torch.float8_e4m3fn + ) + + self.assertEqual(qdata.dtype, torch.float8_e4m3fn) + self.assertEqual(qdata.shape, float_tensor.shape) + self.assertIn('scale', layout_params) + self.assertIn('orig_dtype', layout_params) + self.assertEqual(layout_params['orig_dtype'], torch.float32) + + def test_dequantize(self): + """Test dequantization method""" + float_tensor = torch.ones(10, 20, dtype=torch.float32) * 3.0 + scale = torch.tensor(1.0) + + qdata, layout_params = TensorCoreFP8Layout.quantize( + float_tensor, + scale=scale, + dtype=torch.float8_e4m3fn + ) + + dequantized = TensorCoreFP8Layout.dequantize(qdata, **layout_params) + + # Should approximately match original + self.assertTrue(torch.allclose(dequantized, float_tensor, rtol=0.1, atol=0.1)) + + +class TestFallbackMechanism(unittest.TestCase): + """Test fallback for unsupported operations""" + + def test_unsupported_op_dequantizes(self): + """Test that unsupported operations fall back to dequantization""" + # Set seed for reproducibility + torch.manual_seed(42) + + # Create quantized tensor + a_fp32 = torch.randn(10, 20, dtype=torch.float32) + scale = torch.tensor(1.0) + a_q = QuantizedTensor.from_float( + a_fp32, + TensorCoreFP8Layout, + scale=scale, + dtype=torch.float8_e4m3fn + ) + + # Call an operation that doesn't have a registered handler + # For example, torch.abs + result = torch.abs(a_q) + + # Should work via fallback (dequantize → abs → return) + self.assertNotIsInstance(result, QuantizedTensor) + expected = torch.abs(a_fp32) + # FP8 introduces quantization error, so use loose tolerance + mean_error = (result - expected).abs().mean() + self.assertLess(mean_error, 0.05, f"Mean error {mean_error:.4f} is too large") + + +if __name__ == "__main__": + unittest.main() From d202c2ba7404affd58a2199aeb514b3cc48e0ef3 Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Wed, 29 Oct 2025 06:22:08 +1000 Subject: [PATCH 045/647] execution: Allow a subgraph nodes to execute multiple times (#10499) In the case of --cache-none lazy and subgraph execution can cause anything to be run multiple times per workflow. If that rerun nodes is in itself a subgraph generator, this will crash for two reasons. pending_subgraph_results[] does not cleanup entries after their use. So when a pending_subgraph_result is consumed, remove it from the list so that if the corresponding node is fully re-executed this misses lookup and it fall through to execute the node as it should. Secondly, theres is an explicit enforcement against dups in the addition of subgraphs nodes as ephemerals to the dymprompt. Remove this enforcement as the use case is now valid. --- execution.py | 5 +---- 1 file changed, 1 insertion(+), 4 deletions(-) diff --git a/execution.py b/execution.py index b14bb14c7..20e106213 100644 --- a/execution.py +++ b/execution.py @@ -445,6 +445,7 @@ async def execute(server, dynprompt, caches, current_item, extra_data, executed, resolved_outputs.append(tuple(resolved_output)) output_data = merge_result_data(resolved_outputs, class_def) output_ui = [] + del pending_subgraph_results[unique_id] has_subgraph = False else: get_progress_state().start_progress(unique_id) @@ -527,10 +528,6 @@ async def execute(server, dynprompt, caches, current_item, extra_data, executed, if new_graph is None: cached_outputs.append((False, node_outputs)) else: - # Check for conflicts - for node_id in new_graph.keys(): - if dynprompt.has_node(node_id): - raise DuplicateNodeError(f"Attempt to add duplicate node {node_id}. Ensure node ids are unique and deterministic or use graph_utils.GraphBuilder.") for node_id, node_info in new_graph.items(): new_node_ids.append(node_id) display_id = node_info.get("override_display_id", unique_id) From 210f7a1ba580d57d817ca68346cb72b8d0a26ad2 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Tue, 28 Oct 2025 23:38:05 +0200 Subject: [PATCH 046/647] convert nodes_recraft.py to V3 schema (#10507) --- comfy_api_nodes/nodes_recraft.py | 1319 +++++++++++++----------------- 1 file changed, 585 insertions(+), 734 deletions(-) diff --git a/comfy_api_nodes/nodes_recraft.py b/comfy_api_nodes/nodes_recraft.py index 8ee7e55c4..dee186cd6 100644 --- a/comfy_api_nodes/nodes_recraft.py +++ b/comfy_api_nodes/nodes_recraft.py @@ -1,82 +1,71 @@ -from __future__ import annotations -from inspect import cleandoc -from typing import Optional +from io import BytesIO +from typing import Optional, Union + +import aiohttp +import torch +from PIL import UnidentifiedImageError +from typing_extensions import override + from comfy.utils import ProgressBar -from comfy_extras.nodes_images import SVG # Added -from comfy.comfy_types.node_typing import IO +from comfy_api.latest import IO, ComfyExtension +from comfy_api_nodes.apinode_utils import ( + resize_mask_to_image, +) from comfy_api_nodes.apis.recraft_api import ( - RecraftImageGenerationRequest, - RecraftImageGenerationResponse, - RecraftImageSize, - RecraftModel, - RecraftStyle, - RecraftStyleV3, RecraftColor, RecraftColorChain, RecraftControls, + RecraftImageGenerationRequest, + RecraftImageGenerationResponse, + RecraftImageSize, RecraftIO, + RecraftModel, + RecraftStyle, + RecraftStyleV3, get_v3_substyles, ) -from comfy_api_nodes.apis.client import ( +from comfy_api_nodes.util import ( ApiEndpoint, - HttpMethod, - SynchronousOperation, - EmptyRequest, + bytesio_to_image_tensor, + download_url_as_bytesio, + sync_op, + tensor_to_bytesio, + validate_string, ) -from comfy_api_nodes.apinode_utils import ( - download_url_to_bytesio, - resize_mask_to_image, -) -from comfy_api_nodes.util import validate_string, tensor_to_bytesio, bytesio_to_image_tensor -from server import PromptServer - -import torch -from io import BytesIO -from PIL import UnidentifiedImageError -import aiohttp +from comfy_extras.nodes_images import SVG async def handle_recraft_file_request( + cls: type[IO.ComfyNode], image: torch.Tensor, path: str, - mask: torch.Tensor=None, - total_pixels=4096*4096, - timeout=1024, + mask: Optional[torch.Tensor] = None, + total_pixels: int = 4096 * 4096, + timeout: int = 1024, request=None, - auth_kwargs: dict[str,str] = None, ) -> list[BytesIO]: - """ - Handle sending common Recraft file-only request to get back file bytes. - """ - if request is None: - request = EmptyRequest() + """Handle sending common Recraft file-only request to get back file bytes.""" - files = { - 'image': tensor_to_bytesio(image, total_pixels=total_pixels).read() - } + files = {"image": tensor_to_bytesio(image, total_pixels=total_pixels).read()} if mask is not None: - files['mask'] = tensor_to_bytesio(mask, total_pixels=total_pixels).read() + files["mask"] = tensor_to_bytesio(mask, total_pixels=total_pixels).read() - operation = SynchronousOperation( - endpoint=ApiEndpoint( - path=path, - method=HttpMethod.POST, - request_model=type(request), - response_model=RecraftImageGenerationResponse, - ), - request=request, + response = await sync_op( + cls, + endpoint=ApiEndpoint(path=path, method="POST"), + response_model=RecraftImageGenerationResponse, + data=request if request else None, files=files, content_type="multipart/form-data", - auth_kwargs=auth_kwargs, multipart_parser=recraft_multipart_parser, + max_retries=1, ) - response: RecraftImageGenerationResponse = await operation.execute() all_bytesio = [] if response.image is not None: - all_bytesio.append(await download_url_to_bytesio(response.image.url, timeout=timeout)) + all_bytesio.append(await download_url_as_bytesio(response.image.url, timeout=timeout)) else: for data in response.data: - all_bytesio.append(await download_url_to_bytesio(data.url, timeout=timeout)) + all_bytesio.append(await download_url_as_bytesio(data.url, timeout=timeout)) return all_bytesio @@ -84,11 +73,11 @@ async def handle_recraft_file_request( def recraft_multipart_parser( data, parent_key=None, - formatter: callable = None, - converted_to_check: list[list] = None, + formatter: Optional[type[callable]] = None, + converted_to_check: Optional[list[list]] = None, is_list: bool = False, - return_mode: str = "formdata" # "dict" | "formdata" -) -> dict | aiohttp.FormData: + return_mode: str = "formdata", # "dict" | "formdata" +) -> Union[dict, aiohttp.FormData]: """ Formats data such that multipart/form-data will work with aiohttp library when both files and data are present. @@ -108,8 +97,8 @@ def recraft_multipart_parser( # Modification of a function that handled a different type of multipart parsing, big ups: # https://gist.github.com/kazqvaizer/4cebebe5db654a414132809f9f88067b - def handle_converted_lists(item, parent_key, lists_to_check=tuple[list]): - # if list already exists exists, just extend list with data + def handle_converted_lists(item, parent_key, lists_to_check=list[list]): + # if list already exists, just extend list with data for check_list in lists_to_check: for conv_tuple in check_list: if conv_tuple[0] == parent_key and isinstance(conv_tuple[1], list): @@ -125,7 +114,7 @@ def recraft_multipart_parser( formatter = lambda v: v # Multipart representation of value if not isinstance(data, dict): - # if list already exists exists, just extend list with data + # if list already exists, just extend list with data added = handle_converted_lists(data, parent_key, converted_to_check) if added: return {} @@ -146,7 +135,9 @@ def recraft_multipart_parser( elif isinstance(value, list): for ind, list_value in enumerate(value): iter_key = f"{current_key}[]" - converted.extend(recraft_multipart_parser(list_value, iter_key, formatter, next_check, is_list=True).items()) + converted.extend( + recraft_multipart_parser(list_value, iter_key, formatter, next_check, is_list=True).items() + ) else: converted.append((current_key, formatter(value))) @@ -166,6 +157,7 @@ class handle_recraft_image_output: """ Catch an exception related to receiving SVG data instead of image, when Infinite Style Library style_id is in use. """ + def __init__(self): pass @@ -174,243 +166,225 @@ class handle_recraft_image_output: def __exit__(self, exc_type, exc_val, exc_tb): if exc_type is not None and exc_type is UnidentifiedImageError: - raise Exception("Received output data was not an image; likely an SVG. If you used style_id, make sure it is not a Vector art style.") + raise Exception( + "Received output data was not an image; likely an SVG. " + "If you used style_id, make sure it is not a Vector art style." + ) -class RecraftColorRGBNode: - """ - Create Recraft Color by choosing specific RGB values. - """ - - RETURN_TYPES = (RecraftIO.COLOR,) - DESCRIPTION = cleandoc(__doc__ or "") # Handle potential None value - RETURN_NAMES = ("recraft_color",) - FUNCTION = "create_color" - CATEGORY = "api node/image/Recraft" +class RecraftColorRGBNode(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="RecraftColorRGB", + display_name="Recraft Color RGB", + category="api node/image/Recraft", + description="Create Recraft Color by choosing specific RGB values.", + inputs=[ + IO.Int.Input("r", default=0, min=0, max=255, tooltip="Red value of color."), + IO.Int.Input("g", default=0, min=0, max=255, tooltip="Green value of color."), + IO.Int.Input("b", default=0, min=0, max=255, tooltip="Blue value of color."), + IO.Custom(RecraftIO.COLOR).Input("recraft_color", optional=True), + ], + outputs=[ + IO.Custom(RecraftIO.COLOR).Output(display_name="recraft_color"), + ], + ) @classmethod - def INPUT_TYPES(s): - return { - "required": { - "r": (IO.INT, { - "default": 0, - "min": 0, - "max": 255, - "tooltip": "Red value of color." - }), - "g": (IO.INT, { - "default": 0, - "min": 0, - "max": 255, - "tooltip": "Green value of color." - }), - "b": (IO.INT, { - "default": 0, - "min": 0, - "max": 255, - "tooltip": "Blue value of color." - }), - }, - "optional": { - "recraft_color": (RecraftIO.COLOR,), - } - } - - def create_color(self, r: int, g: int, b: int, recraft_color: RecraftColorChain=None): + def execute(cls, r: int, g: int, b: int, recraft_color: RecraftColorChain = None) -> IO.NodeOutput: recraft_color = recraft_color.clone() if recraft_color else RecraftColorChain() recraft_color.add(RecraftColor(r, g, b)) - return (recraft_color, ) + return IO.NodeOutput(recraft_color) -class RecraftControlsNode: - """ - Create Recraft Controls for customizing Recraft generation. - """ - - RETURN_TYPES = (RecraftIO.CONTROLS,) - RETURN_NAMES = ("recraft_controls",) - DESCRIPTION = cleandoc(__doc__ or "") # Handle potential None value - FUNCTION = "create_controls" - CATEGORY = "api node/image/Recraft" +class RecraftControlsNode(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="RecraftControls", + display_name="Recraft Controls", + category="api node/image/Recraft", + description="Create Recraft Controls for customizing Recraft generation.", + inputs=[ + IO.Custom(RecraftIO.COLOR).Input("colors", optional=True), + IO.Custom(RecraftIO.COLOR).Input("background_color", optional=True), + ], + outputs=[ + IO.Custom(RecraftIO.CONTROLS).Output(display_name="recraft_controls"), + ], + ) @classmethod - def INPUT_TYPES(s): - return { - "required": { - }, - "optional": { - "colors": (RecraftIO.COLOR,), - "background_color": (RecraftIO.COLOR,), - } - } - - def create_controls(self, colors: RecraftColorChain=None, background_color: RecraftColorChain=None): - return (RecraftControls(colors=colors, background_color=background_color), ) + def execute(cls, colors: RecraftColorChain = None, background_color: RecraftColorChain = None) -> IO.NodeOutput: + return IO.NodeOutput(RecraftControls(colors=colors, background_color=background_color)) -class RecraftStyleV3RealisticImageNode: - """ - Select realistic_image style and optional substyle. - """ - - RETURN_TYPES = (RecraftIO.STYLEV3,) - RETURN_NAMES = ("recraft_style",) - DESCRIPTION = cleandoc(__doc__ or "") # Handle potential None value - FUNCTION = "create_style" - CATEGORY = "api node/image/Recraft" - +class RecraftStyleV3RealisticImageNode(IO.ComfyNode): RECRAFT_STYLE = RecraftStyleV3.realistic_image @classmethod - def INPUT_TYPES(s): - return { - "required": { - "substyle": (get_v3_substyles(s.RECRAFT_STYLE),), - } - } + def define_schema(cls): + return IO.Schema( + node_id="RecraftStyleV3RealisticImage", + display_name="Recraft Style - Realistic Image", + category="api node/image/Recraft", + description="Select realistic_image style and optional substyle.", + inputs=[ + IO.Combo.Input("substyle", options=get_v3_substyles(cls.RECRAFT_STYLE)), + ], + outputs=[ + IO.Custom(RecraftIO.STYLEV3).Output(display_name="recraft_style"), + ], + ) - def create_style(self, substyle: str): + @classmethod + def execute(cls, substyle: str) -> IO.NodeOutput: if substyle == "None": substyle = None - return (RecraftStyle(self.RECRAFT_STYLE, substyle),) + return IO.NodeOutput(RecraftStyle(cls.RECRAFT_STYLE, substyle)) class RecraftStyleV3DigitalIllustrationNode(RecraftStyleV3RealisticImageNode): - """ - Select digital_illustration style and optional substyle. - """ - RECRAFT_STYLE = RecraftStyleV3.digital_illustration + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="RecraftStyleV3DigitalIllustration", + display_name="Recraft Style - Digital Illustration", + category="api node/image/Recraft", + description="Select realistic_image style and optional substyle.", + inputs=[ + IO.Combo.Input("substyle", options=get_v3_substyles(cls.RECRAFT_STYLE)), + ], + outputs=[ + IO.Custom(RecraftIO.STYLEV3).Output(display_name="recraft_style"), + ], + ) + class RecraftStyleV3VectorIllustrationNode(RecraftStyleV3RealisticImageNode): - """ - Select vector_illustration style and optional substyle. - """ - RECRAFT_STYLE = RecraftStyleV3.vector_illustration + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="RecraftStyleV3VectorIllustrationNode", + display_name="Recraft Style - Realistic Image", + category="api node/image/Recraft", + description="Select realistic_image style and optional substyle.", + inputs=[ + IO.Combo.Input("substyle", options=get_v3_substyles(cls.RECRAFT_STYLE)), + ], + outputs=[ + IO.Custom(RecraftIO.STYLEV3).Output(display_name="recraft_style"), + ], + ) + class RecraftStyleV3LogoRasterNode(RecraftStyleV3RealisticImageNode): - """ - Select vector_illustration style and optional substyle. - """ - - @classmethod - def INPUT_TYPES(s): - return { - "required": { - "substyle": (get_v3_substyles(s.RECRAFT_STYLE, include_none=False),), - } - } - RECRAFT_STYLE = RecraftStyleV3.logo_raster + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="RecraftStyleV3LogoRaster", + display_name="Recraft Style - Logo Raster", + category="api node/image/Recraft", + description="Select realistic_image style and optional substyle.", + inputs=[ + IO.Combo.Input("substyle", options=get_v3_substyles(cls.RECRAFT_STYLE, include_none=False)), + ], + outputs=[ + IO.Custom(RecraftIO.STYLEV3).Output(display_name="recraft_style"), + ], + ) -class RecraftStyleInfiniteStyleLibrary: - """ - Select style based on preexisting UUID from Recraft's Infinite Style Library. - """ - RETURN_TYPES = (RecraftIO.STYLEV3,) - RETURN_NAMES = ("recraft_style",) - DESCRIPTION = cleandoc(__doc__ or "") # Handle potential None value - FUNCTION = "create_style" - CATEGORY = "api node/image/Recraft" +class RecraftStyleInfiniteStyleLibrary(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="RecraftStyleV3InfiniteStyleLibrary", + display_name="Recraft Style - Infinite Style Library", + category="api node/image/Recraft", + description="Select style based on preexisting UUID from Recraft's Infinite Style Library.", + inputs=[ + IO.String.Input("style_id", default="", tooltip="UUID of style from Infinite Style Library."), + ], + outputs=[ + IO.Custom(RecraftIO.STYLEV3).Output(display_name="recraft_style"), + ], + ) @classmethod - def INPUT_TYPES(s): - return { - "required": { - "style_id": (IO.STRING, { - "default": "", - "tooltip": "UUID of style from Infinite Style Library.", - }) - } - } - - def create_style(self, style_id: str): + def execute(cls, style_id: str) -> IO.NodeOutput: if not style_id: raise Exception("The style_id input cannot be empty.") - return (RecraftStyle(style_id=style_id),) + return IO.NodeOutput(RecraftStyle(style_id=style_id)) -class RecraftTextToImageNode: - """ - Generates images synchronously based on prompt and resolution. - """ - - RETURN_TYPES = (IO.IMAGE,) - DESCRIPTION = cleandoc(__doc__ or "") # Handle potential None value - FUNCTION = "api_call" - API_NODE = True - CATEGORY = "api node/image/Recraft" +class RecraftTextToImageNode(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="RecraftTextToImageNode", + display_name="Recraft Text to Image", + category="api node/image/Recraft", + description="Generates images synchronously based on prompt and resolution.", + inputs=[ + IO.String.Input("prompt", multiline=True, default="", tooltip="Prompt for the image generation."), + IO.Combo.Input( + "size", + options=[res.value for res in RecraftImageSize], + default=RecraftImageSize.res_1024x1024, + tooltip="The size of the generated image.", + ), + IO.Int.Input( + "n", + default=1, + min=1, + max=6, + tooltip="The number of images to generate.", + ), + IO.Int.Input( + "seed", + default=0, + min=0, + max=0xFFFFFFFFFFFFFFFF, + control_after_generate=True, + tooltip="Seed to determine if node should re-run; " + "actual results are nondeterministic regardless of seed.", + ), + IO.Custom(RecraftIO.STYLEV3).Input("recraft_style", optional=True), + IO.String.Input( + "negative_prompt", + default="", + force_input=True, + tooltip="An optional text description of undesired elements on an image.", + optional=True, + ), + IO.Custom(RecraftIO.CONTROLS).Input( + "recraft_controls", + tooltip="Optional additional controls over the generation via the Recraft Controls node.", + optional=True, + ), + ], + outputs=[ + IO.Image.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + ) @classmethod - def INPUT_TYPES(s): - return { - "required": { - "prompt": ( - IO.STRING, - { - "multiline": True, - "default": "", - "tooltip": "Prompt for the image generation.", - }, - ), - "size": ( - [res.value for res in RecraftImageSize], - { - "default": RecraftImageSize.res_1024x1024, - "tooltip": "The size of the generated image.", - }, - ), - "n": ( - IO.INT, - { - "default": 1, - "min": 1, - "max": 6, - "tooltip": "The number of images to generate.", - }, - ), - "seed": ( - IO.INT, - { - "default": 0, - "min": 0, - "max": 0xFFFFFFFFFFFFFFFF, - "control_after_generate": True, - "tooltip": "Seed to determine if node should re-run; actual results are nondeterministic regardless of seed.", - }, - ), - }, - "optional": { - "recraft_style": (RecraftIO.STYLEV3,), - "negative_prompt": ( - IO.STRING, - { - "default": "", - "forceInput": True, - "tooltip": "An optional text description of undesired elements on an image.", - }, - ), - "recraft_controls": ( - RecraftIO.CONTROLS, - { - "tooltip": "Optional additional controls over the generation via the Recraft Controls node." - }, - ), - }, - "hidden": { - "auth_token": "AUTH_TOKEN_COMFY_ORG", - "comfy_api_key": "API_KEY_COMFY_ORG", - "unique_id": "UNIQUE_ID", - }, - } - - async def api_call( - self, + async def execute( + cls, prompt: str, size: str, n: int, @@ -418,9 +392,7 @@ class RecraftTextToImageNode: recraft_style: RecraftStyle = None, negative_prompt: str = None, recraft_controls: RecraftControls = None, - unique_id: Optional[str] = None, - **kwargs, - ): + ) -> IO.NodeOutput: validate_string(prompt, strip_whitespace=False, max_length=1000) default_style = RecraftStyle(RecraftStyleV3.realistic_image) if recraft_style is None: @@ -433,14 +405,11 @@ class RecraftTextToImageNode: if not negative_prompt: negative_prompt = None - operation = SynchronousOperation( - endpoint=ApiEndpoint( - path="/proxy/recraft/image_generation", - method=HttpMethod.POST, - request_model=RecraftImageGenerationRequest, - response_model=RecraftImageGenerationResponse, - ), - request=RecraftImageGenerationRequest( + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/recraft/image_generation", method="POST"), + response_model=RecraftImageGenerationResponse, + data=RecraftImageGenerationRequest( prompt=prompt, negative_prompt=negative_prompt, model=RecraftModel.recraftv3, @@ -451,109 +420,83 @@ class RecraftTextToImageNode: style_id=recraft_style.style_id, controls=controls_api, ), - auth_kwargs=kwargs, + max_retries=1, ) - response: RecraftImageGenerationResponse = await operation.execute() images = [] - urls = [] for data in response.data: with handle_recraft_image_output(): - if unique_id and data.url: - urls.append(data.url) - urls_string = '\n'.join(urls) - PromptServer.instance.send_progress_text( - f"Result URL: {urls_string}", unique_id - ) - image = bytesio_to_image_tensor( - await download_url_to_bytesio(data.url, timeout=1024) - ) + image = bytesio_to_image_tensor(await download_url_as_bytesio(data.url, timeout=1024)) if len(image.shape) < 4: image = image.unsqueeze(0) images.append(image) - output_image = torch.cat(images, dim=0) - return (output_image,) + return IO.NodeOutput(torch.cat(images, dim=0)) -class RecraftImageToImageNode: - """ - Modify image based on prompt and strength. - """ - - RETURN_TYPES = (IO.IMAGE,) - DESCRIPTION = cleandoc(__doc__ or "") # Handle potential None value - FUNCTION = "api_call" - API_NODE = True - CATEGORY = "api node/image/Recraft" +class RecraftImageToImageNode(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="RecraftImageToImageNode", + display_name="Recraft Image to Image", + category="api node/image/Recraft", + description="Modify image based on prompt and strength.", + inputs=[ + IO.Image.Input("image"), + IO.String.Input("prompt", multiline=True, default="", tooltip="Prompt for the image generation."), + IO.Int.Input( + "n", + default=1, + min=1, + max=6, + tooltip="The number of images to generate.", + ), + IO.Float.Input( + "strength", + default=0.5, + min=0.0, + max=1.0, + step=0.01, + tooltip="Defines the difference with the original image, should lie in [0, 1], " + "where 0 means almost identical, and 1 means miserable similarity.", + ), + IO.Int.Input( + "seed", + default=0, + min=0, + max=0xFFFFFFFFFFFFFFFF, + control_after_generate=True, + tooltip="Seed to determine if node should re-run; " + "actual results are nondeterministic regardless of seed.", + ), + IO.Custom(RecraftIO.STYLEV3).Input("recraft_style", optional=True), + IO.String.Input( + "negative_prompt", + default="", + force_input=True, + tooltip="An optional text description of undesired elements on an image.", + optional=True, + ), + IO.Custom(RecraftIO.CONTROLS).Input( + "recraft_controls", + tooltip="Optional additional controls over the generation via the Recraft Controls node.", + optional=True, + ), + ], + outputs=[ + IO.Image.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + ) @classmethod - def INPUT_TYPES(s): - return { - "required": { - "image": (IO.IMAGE, ), - "prompt": ( - IO.STRING, - { - "multiline": True, - "default": "", - "tooltip": "Prompt for the image generation.", - }, - ), - "n": ( - IO.INT, - { - "default": 1, - "min": 1, - "max": 6, - "tooltip": "The number of images to generate.", - }, - ), - "strength": ( - IO.FLOAT, - { - "default": 0.5, - "min": 0.0, - "max": 1.0, - "step": 0.01, - "tooltip": "Defines the difference with the original image, should lie in [0, 1], where 0 means almost identical, and 1 means miserable similarity." - } - ), - "seed": ( - IO.INT, - { - "default": 0, - "min": 0, - "max": 0xFFFFFFFFFFFFFFFF, - "control_after_generate": True, - "tooltip": "Seed to determine if node should re-run; actual results are nondeterministic regardless of seed.", - }, - ), - }, - "optional": { - "recraft_style": (RecraftIO.STYLEV3,), - "negative_prompt": ( - IO.STRING, - { - "default": "", - "forceInput": True, - "tooltip": "An optional text description of undesired elements on an image.", - }, - ), - "recraft_controls": ( - RecraftIO.CONTROLS, - { - "tooltip": "Optional additional controls over the generation via the Recraft Controls node." - }, - ), - }, - "hidden": { - "auth_token": "AUTH_TOKEN_COMFY_ORG", - "comfy_api_key": "API_KEY_COMFY_ORG", - }, - } - - async def api_call( - self, + async def execute( + cls, image: torch.Tensor, prompt: str, n: int, @@ -562,8 +505,7 @@ class RecraftImageToImageNode: recraft_style: RecraftStyle = None, negative_prompt: str = None, recraft_controls: RecraftControls = None, - **kwargs, - ): + ) -> IO.NodeOutput: validate_string(prompt, strip_whitespace=False, max_length=1000) default_style = RecraftStyle(RecraftStyleV3.realistic_image) if recraft_style is None: @@ -593,83 +535,69 @@ class RecraftImageToImageNode: pbar = ProgressBar(total) for i in range(total): sub_bytes = await handle_recraft_file_request( + cls, image=image[i], path="/proxy/recraft/images/imageToImage", request=request, - auth_kwargs=kwargs, ) with handle_recraft_image_output(): images.append(torch.cat([bytesio_to_image_tensor(x) for x in sub_bytes], dim=0)) pbar.update(1) - images_tensor = torch.cat(images, dim=0) - return (images_tensor, ) + return IO.NodeOutput(torch.cat(images, dim=0)) -class RecraftImageInpaintingNode: - """ - Modify image based on prompt and mask. - """ - - RETURN_TYPES = (IO.IMAGE,) - DESCRIPTION = cleandoc(__doc__ or "") # Handle potential None value - FUNCTION = "api_call" - API_NODE = True - CATEGORY = "api node/image/Recraft" +class RecraftImageInpaintingNode(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="RecraftImageInpaintingNode", + display_name="Recraft Image Inpainting", + category="api node/image/Recraft", + description="Modify image based on prompt and mask.", + inputs=[ + IO.Image.Input("image"), + IO.Mask.Input("mask"), + IO.String.Input("prompt", multiline=True, default="", tooltip="Prompt for the image generation."), + IO.Int.Input( + "n", + default=1, + min=1, + max=6, + tooltip="The number of images to generate.", + ), + IO.Int.Input( + "seed", + default=0, + min=0, + max=0xFFFFFFFFFFFFFFFF, + control_after_generate=True, + tooltip="Seed to determine if node should re-run; " + "actual results are nondeterministic regardless of seed.", + ), + IO.Custom(RecraftIO.STYLEV3).Input("recraft_style", optional=True), + IO.String.Input( + "negative_prompt", + default="", + force_input=True, + tooltip="An optional text description of undesired elements on an image.", + optional=True, + ), + ], + outputs=[ + IO.Image.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + ) @classmethod - def INPUT_TYPES(s): - return { - "required": { - "image": (IO.IMAGE, ), - "mask": (IO.MASK, ), - "prompt": ( - IO.STRING, - { - "multiline": True, - "default": "", - "tooltip": "Prompt for the image generation.", - }, - ), - "n": ( - IO.INT, - { - "default": 1, - "min": 1, - "max": 6, - "tooltip": "The number of images to generate.", - }, - ), - "seed": ( - IO.INT, - { - "default": 0, - "min": 0, - "max": 0xFFFFFFFFFFFFFFFF, - "control_after_generate": True, - "tooltip": "Seed to determine if node should re-run; actual results are nondeterministic regardless of seed.", - }, - ), - }, - "optional": { - "recraft_style": (RecraftIO.STYLEV3,), - "negative_prompt": ( - IO.STRING, - { - "default": "", - "forceInput": True, - "tooltip": "An optional text description of undesired elements on an image.", - }, - ), - }, - "hidden": { - "auth_token": "AUTH_TOKEN_COMFY_ORG", - "comfy_api_key": "API_KEY_COMFY_ORG", - }, - } - - async def api_call( - self, + async def execute( + cls, image: torch.Tensor, mask: torch.Tensor, prompt: str, @@ -677,8 +605,7 @@ class RecraftImageInpaintingNode: seed, recraft_style: RecraftStyle = None, negative_prompt: str = None, - **kwargs, - ): + ) -> IO.NodeOutput: validate_string(prompt, strip_whitespace=False, max_length=1000) default_style = RecraftStyle(RecraftStyleV3.realistic_image) if recraft_style is None: @@ -705,96 +632,73 @@ class RecraftImageInpaintingNode: pbar = ProgressBar(total) for i in range(total): sub_bytes = await handle_recraft_file_request( + cls, image=image[i], - mask=mask[i:i+1], + mask=mask[i : i + 1], path="/proxy/recraft/images/inpaint", request=request, - auth_kwargs=kwargs, ) with handle_recraft_image_output(): images.append(torch.cat([bytesio_to_image_tensor(x) for x in sub_bytes], dim=0)) pbar.update(1) - images_tensor = torch.cat(images, dim=0) - return (images_tensor, ) + return IO.NodeOutput(torch.cat(images, dim=0)) -class RecraftTextToVectorNode: - """ - Generates SVG synchronously based on prompt and resolution. - """ - - RETURN_TYPES = ("SVG",) # Changed - DESCRIPTION = cleandoc(__doc__ or "") if 'cleandoc' in globals() else __doc__ # Keep cleandoc if other nodes use it - FUNCTION = "api_call" - API_NODE = True - CATEGORY = "api node/image/Recraft" +class RecraftTextToVectorNode(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="RecraftTextToVectorNode", + display_name="Recraft Text to Vector", + category="api node/image/Recraft", + description="Generates SVG synchronously based on prompt and resolution.", + inputs=[ + IO.String.Input("prompt", default="", tooltip="Prompt for the image generation.", multiline=True), + IO.Combo.Input("substyle", options=get_v3_substyles(RecraftStyleV3.vector_illustration)), + IO.Combo.Input( + "size", + options=[res.value for res in RecraftImageSize], + default=RecraftImageSize.res_1024x1024, + tooltip="The size of the generated image.", + ), + IO.Int.Input("n", default=1, min=1, max=6, tooltip="The number of images to generate."), + IO.Int.Input( + "seed", + default=0, + min=0, + max=0xFFFFFFFFFFFFFFFF, + control_after_generate=True, + tooltip="Seed to determine if node should re-run; " + "actual results are nondeterministic regardless of seed.", + ), + IO.String.Input( + "negative_prompt", + default="", + force_input=True, + tooltip="An optional text description of undesired elements on an image.", + optional=True, + ), + IO.Custom(RecraftIO.CONTROLS).Input( + "recraft_controls", + tooltip="Optional additional controls over the generation via the Recraft Controls node.", + optional=True, + ), + ], + outputs=[ + IO.SVG.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + ) @classmethod - def INPUT_TYPES(s): - return { - "required": { - "prompt": ( - IO.STRING, - { - "multiline": True, - "default": "", - "tooltip": "Prompt for the image generation.", - }, - ), - "substyle": (get_v3_substyles(RecraftStyleV3.vector_illustration),), - "size": ( - [res.value for res in RecraftImageSize], - { - "default": RecraftImageSize.res_1024x1024, - "tooltip": "The size of the generated image.", - }, - ), - "n": ( - IO.INT, - { - "default": 1, - "min": 1, - "max": 6, - "tooltip": "The number of images to generate.", - }, - ), - "seed": ( - IO.INT, - { - "default": 0, - "min": 0, - "max": 0xFFFFFFFFFFFFFFFF, - "control_after_generate": True, - "tooltip": "Seed to determine if node should re-run; actual results are nondeterministic regardless of seed.", - }, - ), - }, - "optional": { - "negative_prompt": ( - IO.STRING, - { - "default": "", - "forceInput": True, - "tooltip": "An optional text description of undesired elements on an image.", - }, - ), - "recraft_controls": ( - RecraftIO.CONTROLS, - { - "tooltip": "Optional additional controls over the generation via the Recraft Controls node." - }, - ), - }, - "hidden": { - "auth_token": "AUTH_TOKEN_COMFY_ORG", - "comfy_api_key": "API_KEY_COMFY_ORG", - "unique_id": "UNIQUE_ID", - }, - } - - async def api_call( - self, + async def execute( + cls, prompt: str, substyle: str, size: str, @@ -802,9 +706,7 @@ class RecraftTextToVectorNode: seed, negative_prompt: str = None, recraft_controls: RecraftControls = None, - unique_id: Optional[str] = None, - **kwargs, - ): + ) -> IO.NodeOutput: validate_string(prompt, strip_whitespace=False, max_length=1000) # create RecraftStyle so strings will be formatted properly (i.e. "None" will become None) recraft_style = RecraftStyle(RecraftStyleV3.vector_illustration, substyle=substyle) @@ -816,14 +718,11 @@ class RecraftTextToVectorNode: if not negative_prompt: negative_prompt = None - operation = SynchronousOperation( - endpoint=ApiEndpoint( - path="/proxy/recraft/image_generation", - method=HttpMethod.POST, - request_model=RecraftImageGenerationRequest, - response_model=RecraftImageGenerationResponse, - ), - request=RecraftImageGenerationRequest( + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/recraft/image_generation", method="POST"), + response_model=RecraftImageGenerationResponse, + data=RecraftImageGenerationRequest( prompt=prompt, negative_prompt=negative_prompt, model=RecraftModel.recraftv3, @@ -833,139 +732,105 @@ class RecraftTextToVectorNode: substyle=recraft_style.substyle, controls=controls_api, ), - auth_kwargs=kwargs, + max_retries=1, ) - response: RecraftImageGenerationResponse = await operation.execute() svg_data = [] - urls = [] for data in response.data: - if unique_id and data.url: - urls.append(data.url) - # Print result on each iteration in case of error - PromptServer.instance.send_progress_text( - f"Result URL: {' '.join(urls)}", unique_id - ) - svg_data.append(await download_url_to_bytesio(data.url, timeout=1024)) + svg_data.append(await download_url_as_bytesio(data.url, timeout=1024)) - return (SVG(svg_data),) + return IO.NodeOutput(SVG(svg_data)) -class RecraftVectorizeImageNode: - """ - Generates SVG synchronously from an input image. - """ - - RETURN_TYPES = ("SVG",) # Changed - DESCRIPTION = cleandoc(__doc__ or "") if 'cleandoc' in globals() else __doc__ # Keep cleandoc if other nodes use it - FUNCTION = "api_call" - API_NODE = True - CATEGORY = "api node/image/Recraft" +class RecraftVectorizeImageNode(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="RecraftVectorizeImageNode", + display_name="Recraft Vectorize Image", + category="api node/image/Recraft", + description="Generates SVG synchronously from an input image.", + inputs=[ + IO.Image.Input("image"), + ], + outputs=[ + IO.SVG.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + ) @classmethod - def INPUT_TYPES(s): - return { - "required": { - "image": (IO.IMAGE, ), - }, - "optional": { - }, - "hidden": { - "auth_token": "AUTH_TOKEN_COMFY_ORG", - "comfy_api_key": "API_KEY_COMFY_ORG", - }, - } - - async def api_call( - self, - image: torch.Tensor, - **kwargs, - ): + async def execute(cls, image: torch.Tensor) -> IO.NodeOutput: svgs = [] total = image.shape[0] pbar = ProgressBar(total) for i in range(total): sub_bytes = await handle_recraft_file_request( + cls, image=image[i], path="/proxy/recraft/images/vectorize", - auth_kwargs=kwargs, ) svgs.append(SVG(sub_bytes)) pbar.update(1) - return (SVG.combine_all(svgs), ) + return IO.NodeOutput(SVG.combine_all(svgs)) -class RecraftReplaceBackgroundNode: - """ - Replace background on image, based on provided prompt. - """ - - RETURN_TYPES = (IO.IMAGE,) - DESCRIPTION = cleandoc(__doc__ or "") # Handle potential None value - FUNCTION = "api_call" - API_NODE = True - CATEGORY = "api node/image/Recraft" +class RecraftReplaceBackgroundNode(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="RecraftReplaceBackgroundNode", + display_name="Recraft Replace Background", + category="api node/image/Recraft", + description="Replace background on image, based on provided prompt.", + inputs=[ + IO.Image.Input("image"), + IO.String.Input("prompt", tooltip="Prompt for the image generation.", default="", multiline=True), + IO.Int.Input("n", default=1, min=1, max=6, tooltip="The number of images to generate."), + IO.Int.Input( + "seed", + default=0, + min=0, + max=0xFFFFFFFFFFFFFFFF, + control_after_generate=True, + tooltip="Seed to determine if node should re-run; " + "actual results are nondeterministic regardless of seed.", + ), + IO.Custom(RecraftIO.STYLEV3).Input("recraft_style", optional=True), + IO.String.Input( + "negative_prompt", + default="", + force_input=True, + tooltip="An optional text description of undesired elements on an image.", + optional=True, + ), + ], + outputs=[ + IO.Image.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + ) @classmethod - def INPUT_TYPES(s): - return { - "required": { - "image": (IO.IMAGE, ), - "prompt": ( - IO.STRING, - { - "multiline": True, - "default": "", - "tooltip": "Prompt for the image generation.", - }, - ), - "n": ( - IO.INT, - { - "default": 1, - "min": 1, - "max": 6, - "tooltip": "The number of images to generate.", - }, - ), - "seed": ( - IO.INT, - { - "default": 0, - "min": 0, - "max": 0xFFFFFFFFFFFFFFFF, - "control_after_generate": True, - "tooltip": "Seed to determine if node should re-run; actual results are nondeterministic regardless of seed.", - }, - ), - }, - "optional": { - "recraft_style": (RecraftIO.STYLEV3,), - "negative_prompt": ( - IO.STRING, - { - "default": "", - "forceInput": True, - "tooltip": "An optional text description of undesired elements on an image.", - }, - ), - }, - "hidden": { - "auth_token": "AUTH_TOKEN_COMFY_ORG", - "comfy_api_key": "API_KEY_COMFY_ORG", - }, - } - - async def api_call( - self, + async def execute( + cls, image: torch.Tensor, prompt: str, n: int, seed, recraft_style: RecraftStyle = None, negative_prompt: str = None, - **kwargs, - ): + ) -> IO.NodeOutput: default_style = RecraftStyle(RecraftStyleV3.realistic_image) if recraft_style is None: recraft_style = default_style @@ -988,165 +853,151 @@ class RecraftReplaceBackgroundNode: pbar = ProgressBar(total) for i in range(total): sub_bytes = await handle_recraft_file_request( + cls, image=image[i], path="/proxy/recraft/images/replaceBackground", request=request, - auth_kwargs=kwargs, ) images.append(torch.cat([bytesio_to_image_tensor(x) for x in sub_bytes], dim=0)) pbar.update(1) - images_tensor = torch.cat(images, dim=0) - return (images_tensor, ) + return IO.NodeOutput(torch.cat(images, dim=0)) -class RecraftRemoveBackgroundNode: - """ - Remove background from image, and return processed image and mask. - """ - - RETURN_TYPES = (IO.IMAGE, IO.MASK) - DESCRIPTION = cleandoc(__doc__ or "") # Handle potential None value - FUNCTION = "api_call" - API_NODE = True - CATEGORY = "api node/image/Recraft" +class RecraftRemoveBackgroundNode(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="RecraftRemoveBackgroundNode", + display_name="Recraft Remove Background", + category="api node/image/Recraft", + description="Remove background from image, and return processed image and mask.", + inputs=[ + IO.Image.Input("image"), + ], + outputs=[ + IO.Image.Output(), + IO.Mask.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + ) @classmethod - def INPUT_TYPES(s): - return { - "required": { - "image": (IO.IMAGE, ), - }, - "optional": { - }, - "hidden": { - "auth_token": "AUTH_TOKEN_COMFY_ORG", - "comfy_api_key": "API_KEY_COMFY_ORG", - }, - } - - async def api_call( - self, - image: torch.Tensor, - **kwargs, - ): + async def execute(cls, image: torch.Tensor) -> IO.NodeOutput: images = [] total = image.shape[0] pbar = ProgressBar(total) for i in range(total): sub_bytes = await handle_recraft_file_request( + cls, image=image[i], path="/proxy/recraft/images/removeBackground", - auth_kwargs=kwargs, ) images.append(torch.cat([bytesio_to_image_tensor(x) for x in sub_bytes], dim=0)) pbar.update(1) images_tensor = torch.cat(images, dim=0) # use alpha channel as masks, in B,H,W format - masks_tensor = images_tensor[:,:,:,-1:].squeeze(-1) - return (images_tensor, masks_tensor) + masks_tensor = images_tensor[:, :, :, -1:].squeeze(-1) + return IO.NodeOutput(images_tensor, masks_tensor) -class RecraftCrispUpscaleNode: - """ - Upscale image synchronously. - Enhances a given raster image using ‘crisp upscale’ tool, increasing image resolution, making the image sharper and cleaner. - """ - - RETURN_TYPES = (IO.IMAGE,) - DESCRIPTION = cleandoc(__doc__ or "") # Handle potential None value - FUNCTION = "api_call" - API_NODE = True - CATEGORY = "api node/image/Recraft" - +class RecraftCrispUpscaleNode(IO.ComfyNode): RECRAFT_PATH = "/proxy/recraft/images/crispUpscale" @classmethod - def INPUT_TYPES(s): - return { - "required": { - "image": (IO.IMAGE, ), - }, - "optional": { - }, - "hidden": { - "auth_token": "AUTH_TOKEN_COMFY_ORG", - "comfy_api_key": "API_KEY_COMFY_ORG", - }, - } + def define_schema(cls): + return IO.Schema( + node_id="RecraftCrispUpscaleNode", + display_name="Recraft Crisp Upscale Image", + category="api node/image/Recraft", + description="Upscale image synchronously.\n" + "Enhances a given raster image using ‘crisp upscale’ tool, " + "increasing image resolution, making the image sharper and cleaner.", + inputs=[ + IO.Image.Input("image"), + ], + outputs=[ + IO.Image.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + ) - async def api_call( - self, - image: torch.Tensor, - **kwargs, - ): + @classmethod + async def execute(cls, image: torch.Tensor) -> IO.NodeOutput: images = [] total = image.shape[0] pbar = ProgressBar(total) for i in range(total): sub_bytes = await handle_recraft_file_request( + cls, image=image[i], - path=self.RECRAFT_PATH, - auth_kwargs=kwargs, + path=cls.RECRAFT_PATH, ) images.append(torch.cat([bytesio_to_image_tensor(x) for x in sub_bytes], dim=0)) pbar.update(1) - images_tensor = torch.cat(images, dim=0) - return (images_tensor,) + return IO.NodeOutput(torch.cat(images, dim=0)) class RecraftCreativeUpscaleNode(RecraftCrispUpscaleNode): - """ - Upscale image synchronously. - Enhances a given raster image using ‘creative upscale’ tool, boosting resolution with a focus on refining small details and faces. - """ - - RETURN_TYPES = (IO.IMAGE,) - DESCRIPTION = cleandoc(__doc__ or "") # Handle potential None value - FUNCTION = "api_call" - API_NODE = True - CATEGORY = "api node/image/Recraft" - RECRAFT_PATH = "/proxy/recraft/images/creativeUpscale" + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="RecraftCreativeUpscaleNode", + display_name="Recraft Creative Upscale Image", + category="api node/image/Recraft", + description="Upscale image synchronously.\n" + "Enhances a given raster image using ‘creative upscale’ tool, " + "boosting resolution with a focus on refining small details and faces.", + inputs=[ + IO.Image.Input("image"), + ], + outputs=[ + IO.Image.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + ) -# A dictionary that contains all nodes you want to export with their names -# NOTE: names should be globally unique -NODE_CLASS_MAPPINGS = { - "RecraftTextToImageNode": RecraftTextToImageNode, - "RecraftImageToImageNode": RecraftImageToImageNode, - "RecraftImageInpaintingNode": RecraftImageInpaintingNode, - "RecraftTextToVectorNode": RecraftTextToVectorNode, - "RecraftVectorizeImageNode": RecraftVectorizeImageNode, - "RecraftRemoveBackgroundNode": RecraftRemoveBackgroundNode, - "RecraftReplaceBackgroundNode": RecraftReplaceBackgroundNode, - "RecraftCrispUpscaleNode": RecraftCrispUpscaleNode, - "RecraftCreativeUpscaleNode": RecraftCreativeUpscaleNode, - "RecraftStyleV3RealisticImage": RecraftStyleV3RealisticImageNode, - "RecraftStyleV3DigitalIllustration": RecraftStyleV3DigitalIllustrationNode, - "RecraftStyleV3LogoRaster": RecraftStyleV3LogoRasterNode, - "RecraftStyleV3InfiniteStyleLibrary": RecraftStyleInfiniteStyleLibrary, - "RecraftColorRGB": RecraftColorRGBNode, - "RecraftControls": RecraftControlsNode, -} -# A dictionary that contains the friendly/humanly readable titles for the nodes -NODE_DISPLAY_NAME_MAPPINGS = { - "RecraftTextToImageNode": "Recraft Text to Image", - "RecraftImageToImageNode": "Recraft Image to Image", - "RecraftImageInpaintingNode": "Recraft Image Inpainting", - "RecraftTextToVectorNode": "Recraft Text to Vector", - "RecraftVectorizeImageNode": "Recraft Vectorize Image", - "RecraftRemoveBackgroundNode": "Recraft Remove Background", - "RecraftReplaceBackgroundNode": "Recraft Replace Background", - "RecraftCrispUpscaleNode": "Recraft Crisp Upscale Image", - "RecraftCreativeUpscaleNode": "Recraft Creative Upscale Image", - "RecraftStyleV3RealisticImage": "Recraft Style - Realistic Image", - "RecraftStyleV3DigitalIllustration": "Recraft Style - Digital Illustration", - "RecraftStyleV3LogoRaster": "Recraft Style - Logo Raster", - "RecraftStyleV3InfiniteStyleLibrary": "Recraft Style - Infinite Style Library", - "RecraftColorRGB": "Recraft Color RGB", - "RecraftControls": "Recraft Controls", -} +class RecraftExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[IO.ComfyNode]]: + return [ + RecraftTextToImageNode, + RecraftImageToImageNode, + RecraftImageInpaintingNode, + RecraftTextToVectorNode, + RecraftVectorizeImageNode, + RecraftRemoveBackgroundNode, + RecraftReplaceBackgroundNode, + RecraftCrispUpscaleNode, + RecraftCreativeUpscaleNode, + RecraftStyleV3RealisticImageNode, + RecraftStyleV3DigitalIllustrationNode, + RecraftStyleV3LogoRasterNode, + RecraftStyleInfiniteStyleLibrary, + RecraftColorRGBNode, + RecraftControlsNode, + ] + + +async def comfy_entrypoint() -> RecraftExtension: + return RecraftExtension() From 3fa7a5c04ae69ad168a875e8d3d453783d60899d Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 28 Oct 2025 21:21:01 -0700 Subject: [PATCH 047/647] Speed up offloading using pinned memory. (#10526) To enable this feature use: --fast pinned_memory --- comfy/cli_args.py | 1 + comfy/model_management.py | 30 ++++++++++++++++++++++++++++++ comfy/model_patcher.py | 26 +++++++++++++++++++++++++- 3 files changed, 56 insertions(+), 1 deletion(-) diff --git a/comfy/cli_args.py b/comfy/cli_args.py index cc1f12482..001abd843 100644 --- a/comfy/cli_args.py +++ b/comfy/cli_args.py @@ -144,6 +144,7 @@ class PerformanceFeature(enum.Enum): Fp8MatrixMultiplication = "fp8_matrix_mult" CublasOps = "cublas_ops" AutoTune = "autotune" + PinnedMem = "pinned_memory" parser.add_argument("--fast", nargs="*", type=PerformanceFeature, help="Enable some untested and potentially quality deteriorating optimizations. --fast with no arguments enables everything. You can pass a list specific optimizations if you only want to enable specific ones. Current valid optimizations: {}".format(" ".join(map(lambda c: c.value, PerformanceFeature)))) diff --git a/comfy/model_management.py b/comfy/model_management.py index afe78f36e..3e5b977d4 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -1080,6 +1080,36 @@ def cast_to_device(tensor, device, dtype, copy=False): non_blocking = device_supports_non_blocking(device) return cast_to(tensor, dtype=dtype, device=device, non_blocking=non_blocking, copy=copy) +def pin_memory(tensor): + if PerformanceFeature.PinnedMem not in args.fast: + return False + + if not is_nvidia(): + return False + + if not is_device_cpu(tensor.device): + return False + + if torch.cuda.cudart().cudaHostRegister(tensor.data_ptr(), tensor.numel() * tensor.element_size(), 1) == 0: + return True + + return False + +def unpin_memory(tensor): + if PerformanceFeature.PinnedMem not in args.fast: + return False + + if not is_nvidia(): + return False + + if not is_device_cpu(tensor.device): + return False + + if torch.cuda.cudart().cudaHostUnregister(tensor.data_ptr()) == 0: + return True + + return False + def sage_attention_enabled(): return args.use_sage_attention diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index c0b68fb8c..aec73349c 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -238,6 +238,7 @@ class ModelPatcher: self.force_cast_weights = False self.patches_uuid = uuid.uuid4() self.parent = None + self.pinned = set() self.attachments: dict[str] = {} self.additional_models: dict[str, list[ModelPatcher]] = {} @@ -618,6 +619,21 @@ class ModelPatcher: else: set_func(out_weight, inplace_update=inplace_update, seed=string_to_seed(key)) + def pin_weight_to_device(self, key): + weight, set_func, convert_func = get_key_weight(self.model, key) + if comfy.model_management.pin_memory(weight): + self.pinned.add(key) + + def unpin_weight(self, key): + if key in self.pinned: + weight, set_func, convert_func = get_key_weight(self.model, key) + comfy.model_management.unpin_memory(weight) + self.pinned.remove(key) + + def unpin_all_weights(self): + for key in list(self.pinned): + self.unpin_weight(key) + def _load_list(self): loading = [] for n, m in self.model.named_modules(): @@ -683,6 +699,8 @@ class ModelPatcher: patch_counter += 1 cast_weight = True + for param in params: + self.pin_weight_to_device("{}.{}".format(n, param)) else: if hasattr(m, "comfy_cast_weights"): wipe_lowvram_weight(m) @@ -713,7 +731,9 @@ class ModelPatcher: continue for param in params: - self.patch_weight_to_device("{}.{}".format(n, param), device_to=device_to) + key = "{}.{}".format(n, param) + self.unpin_weight(key) + self.patch_weight_to_device(key, device_to=device_to) logging.debug("lowvram: loaded module regularly {} {}".format(n, m)) m.comfy_patched_weights = True @@ -762,6 +782,7 @@ class ModelPatcher: self.eject_model() if unpatch_weights: self.unpatch_hooks() + self.unpin_all_weights() if self.model.model_lowvram: for m in self.model.modules(): move_weight_functions(m, device_to) @@ -857,6 +878,9 @@ class ModelPatcher: memory_freed += module_mem logging.debug("freed {}".format(n)) + for param in params: + self.pin_weight_to_device("{}.{}".format(n, param)) + self.model.model_lowvram = True self.model.lowvram_patch_counter += patch_counter self.model.model_loaded_weight_memory -= memory_freed From e525673f7201b6c49af0fa0e6baf44e4e98bb10c Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 28 Oct 2025 21:37:00 -0700 Subject: [PATCH 048/647] Fix issue. (#10527) --- comfy/sd.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy/sd.py b/comfy/sd.py index 6411bb27d..de4eee96e 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -1330,7 +1330,7 @@ def load_diffusion_model_state_dict(sd, model_options={}, metadata=None): else: unet_dtype = dtype - if hasattr(model_config, "layer_quant_config"): + if model_config.layer_quant_config is not None: manual_cast_dtype = model_management.unet_manual_cast(None, load_device, model_config.supported_inference_dtypes) else: manual_cast_dtype = model_management.unet_manual_cast(unet_dtype, load_device, model_config.supported_inference_dtypes) From 6c14f3afac0ea28dba24fe8783e7c1f09c03b31f Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Wed, 29 Oct 2025 20:14:56 +0200 Subject: [PATCH 049/647] use new API client in Luma and Minimax nodes (#10528) --- comfy_api_nodes/apinode_utils.py | 81 ------ comfy_api_nodes/apis/minimax_api.py | 120 ++++++++ comfy_api_nodes/nodes_ideogram.py | 2 +- comfy_api_nodes/nodes_luma.py | 354 ++++++----------------- comfy_api_nodes/nodes_minimax.py | 237 +++++---------- comfy_api_nodes/util/client.py | 2 +- comfy_api_nodes/util/download_helpers.py | 3 +- 7 files changed, 283 insertions(+), 516 deletions(-) create mode 100644 comfy_api_nodes/apis/minimax_api.py diff --git a/comfy_api_nodes/apinode_utils.py b/comfy_api_nodes/apinode_utils.py index 4182c8f80..6a72b9d1d 100644 --- a/comfy_api_nodes/apinode_utils.py +++ b/comfy_api_nodes/apinode_utils.py @@ -3,14 +3,6 @@ import aiohttp import mimetypes from typing import Optional, Union from comfy.utils import common_upscale -from comfy_api_nodes.apis.client import ( - ApiClient, - ApiEndpoint, - HttpMethod, - SynchronousOperation, - UploadRequest, - UploadResponse, -) from server import PromptServer from comfy.cli_args import args @@ -19,7 +11,6 @@ from PIL import Image import torch import math import base64 -from .util import tensor_to_bytesio, bytesio_to_image_tensor from io import BytesIO @@ -148,11 +139,6 @@ async def download_url_to_bytesio( return BytesIO(await resp.read()) -def process_image_response(response_content: bytes | str) -> torch.Tensor: - """Uses content from a Response object and converts it to a torch.Tensor""" - return bytesio_to_image_tensor(BytesIO(response_content)) - - def text_filepath_to_base64_string(filepath: str) -> str: """Converts a text file to a base64 string.""" with open(filepath, "rb") as f: @@ -169,73 +155,6 @@ def text_filepath_to_data_uri(filepath: str) -> str: return f"data:{mime_type};base64,{base64_string}" -async def upload_file_to_comfyapi( - file_bytes_io: BytesIO, - filename: str, - upload_mime_type: Optional[str], - auth_kwargs: Optional[dict[str, str]] = None, -) -> str: - """ - Uploads a single file to ComfyUI API and returns its download URL. - - Args: - file_bytes_io: BytesIO object containing the file data. - filename: The filename of the file. - upload_mime_type: MIME type of the file. - auth_kwargs: Optional authentication token(s). - - Returns: - The download URL for the uploaded file. - """ - if upload_mime_type is None: - request_object = UploadRequest(file_name=filename) - else: - request_object = UploadRequest(file_name=filename, content_type=upload_mime_type) - operation = SynchronousOperation( - endpoint=ApiEndpoint( - path="/customers/storage", - method=HttpMethod.POST, - request_model=UploadRequest, - response_model=UploadResponse, - ), - request=request_object, - auth_kwargs=auth_kwargs, - ) - - response: UploadResponse = await operation.execute() - await ApiClient.upload_file(response.upload_url, file_bytes_io, content_type=upload_mime_type) - return response.download_url - - -async def upload_images_to_comfyapi( - image: torch.Tensor, - max_images=8, - auth_kwargs: Optional[dict[str, str]] = None, - mime_type: Optional[str] = None, -) -> list[str]: - """ - Uploads images to ComfyUI API and returns download URLs. - To upload multiple images, stack them in the batch dimension first. - - Args: - image: Input torch.Tensor image. - max_images: Maximum number of images to upload. - auth_kwargs: Optional authentication token(s). - mime_type: Optional MIME type for the image. - """ - # if batch, try to upload each file if max_images is greater than 0 - download_urls: list[str] = [] - is_batch = len(image.shape) > 3 - batch_len = image.shape[0] if is_batch else 1 - - for idx in range(min(batch_len, max_images)): - tensor = image[idx] if is_batch else image - img_io = tensor_to_bytesio(tensor, mime_type=mime_type) - url = await upload_file_to_comfyapi(img_io, img_io.name, mime_type, auth_kwargs) - download_urls.append(url) - return download_urls - - def resize_mask_to_image( mask: torch.Tensor, image: torch.Tensor, diff --git a/comfy_api_nodes/apis/minimax_api.py b/comfy_api_nodes/apis/minimax_api.py new file mode 100644 index 000000000..d747e177a --- /dev/null +++ b/comfy_api_nodes/apis/minimax_api.py @@ -0,0 +1,120 @@ +from enum import Enum +from typing import Optional + +from pydantic import BaseModel, Field + + +class MinimaxBaseResponse(BaseModel): + status_code: int = Field( + ..., + description='Status code. 0 indicates success, other values indicate errors.', + ) + status_msg: str = Field( + ..., description='Specific error details or success message.' + ) + + +class File(BaseModel): + bytes: Optional[int] = Field(None, description='File size in bytes') + created_at: Optional[int] = Field( + None, description='Unix timestamp when the file was created, in seconds' + ) + download_url: Optional[str] = Field( + None, description='The URL to download the video' + ) + backup_download_url: Optional[str] = Field( + None, description='The backup URL to download the video' + ) + + file_id: Optional[int] = Field(None, description='Unique identifier for the file') + filename: Optional[str] = Field(None, description='The name of the file') + purpose: Optional[str] = Field(None, description='The purpose of using the file') + + +class MinimaxFileRetrieveResponse(BaseModel): + base_resp: MinimaxBaseResponse + file: File + + +class MiniMaxModel(str, Enum): + T2V_01_Director = 'T2V-01-Director' + I2V_01_Director = 'I2V-01-Director' + S2V_01 = 'S2V-01' + I2V_01 = 'I2V-01' + I2V_01_live = 'I2V-01-live' + T2V_01 = 'T2V-01' + Hailuo_02 = 'MiniMax-Hailuo-02' + + +class Status6(str, Enum): + Queueing = 'Queueing' + Preparing = 'Preparing' + Processing = 'Processing' + Success = 'Success' + Fail = 'Fail' + + +class MinimaxTaskResultResponse(BaseModel): + base_resp: MinimaxBaseResponse + file_id: Optional[str] = Field( + None, + description='After the task status changes to Success, this field returns the file ID corresponding to the generated video.', + ) + status: Status6 = Field( + ..., + description="Task status: 'Queueing' (in queue), 'Preparing' (task is preparing), 'Processing' (generating), 'Success' (task completed successfully), or 'Fail' (task failed).", + ) + task_id: str = Field(..., description='The task ID being queried.') + + +class SubjectReferenceItem(BaseModel): + image: Optional[str] = Field( + None, description='URL or base64 encoding of the subject reference image.' + ) + mask: Optional[str] = Field( + None, + description='URL or base64 encoding of the mask for the subject reference image.', + ) + + +class MinimaxVideoGenerationRequest(BaseModel): + callback_url: Optional[str] = Field( + None, + description='Optional. URL to receive real-time status updates about the video generation task.', + ) + first_frame_image: Optional[str] = Field( + None, + description='URL or base64 encoding of the first frame image. Required when model is I2V-01, I2V-01-Director, or I2V-01-live.', + ) + model: MiniMaxModel = Field( + ..., + description='Required. ID of model. Options: T2V-01-Director, I2V-01-Director, S2V-01, I2V-01, I2V-01-live, T2V-01', + ) + prompt: Optional[str] = Field( + None, + description='Description of the video. Should be less than 2000 characters. Supports camera movement instructions in [brackets].', + max_length=2000, + ) + prompt_optimizer: Optional[bool] = Field( + True, + description='If true (default), the model will automatically optimize the prompt. Set to false for more precise control.', + ) + subject_reference: Optional[list[SubjectReferenceItem]] = Field( + None, + description='Only available when model is S2V-01. The model will generate a video based on the subject uploaded through this parameter.', + ) + duration: Optional[int] = Field( + None, + description="The length of the output video in seconds." + ) + resolution: Optional[str] = Field( + None, + description="The dimensions of the video display. 1080p corresponds to 1920 x 1080 pixels, 768p corresponds to 1366 x 768 pixels." + ) + + +class MinimaxVideoGenerationResponse(BaseModel): + base_resp: MinimaxBaseResponse + task_id: str = Field( + ..., description='The task ID for the asynchronous video generation task.' + ) diff --git a/comfy_api_nodes/nodes_ideogram.py b/comfy_api_nodes/nodes_ideogram.py index 9eae5f11a..d8fd3378b 100644 --- a/comfy_api_nodes/nodes_ideogram.py +++ b/comfy_api_nodes/nodes_ideogram.py @@ -20,9 +20,9 @@ from comfy_api_nodes.apis.client import ( from comfy_api_nodes.apinode_utils import ( download_url_to_bytesio, - bytesio_to_image_tensor, resize_mask_to_image, ) +from comfy_api_nodes.util import bytesio_to_image_tensor from server import PromptServer V1_V1_RES_MAP = { diff --git a/comfy_api_nodes/nodes_luma.py b/comfy_api_nodes/nodes_luma.py index e74441e5e..894f2b08c 100644 --- a/comfy_api_nodes/nodes_luma.py +++ b/comfy_api_nodes/nodes_luma.py @@ -1,69 +1,51 @@ -from __future__ import annotations -from inspect import cleandoc from typing import Optional + +import torch from typing_extensions import override -from comfy_api.latest import ComfyExtension, IO -from comfy_api.input_impl.video_types import VideoFromFile + +from comfy_api.latest import IO, ComfyExtension from comfy_api_nodes.apis.luma_api import ( - LumaImageModel, - LumaVideoModel, - LumaVideoOutputResolution, - LumaVideoModelOutputDuration, LumaAspectRatio, - LumaState, - LumaImageGenerationRequest, - LumaGenerationRequest, - LumaGeneration, LumaCharacterRef, - LumaModifyImageRef, + LumaConceptChain, + LumaGeneration, + LumaGenerationRequest, + LumaImageGenerationRequest, LumaImageIdentity, + LumaImageModel, + LumaImageReference, + LumaIO, + LumaKeyframes, + LumaModifyImageRef, LumaReference, LumaReferenceChain, - LumaImageReference, - LumaKeyframes, - LumaConceptChain, - LumaIO, + LumaVideoModel, + LumaVideoModelOutputDuration, + LumaVideoOutputResolution, get_luma_concepts, ) -from comfy_api_nodes.apis.client import ( +from comfy_api_nodes.util import ( ApiEndpoint, - HttpMethod, - SynchronousOperation, - PollingOperation, - EmptyRequest, -) -from comfy_api_nodes.apinode_utils import ( + download_url_to_image_tensor, + download_url_to_video_output, + poll_op, + sync_op, upload_images_to_comfyapi, - process_image_response, + validate_string, ) -from server import PromptServer -from comfy_api_nodes.util import validate_string - -import aiohttp -import torch -from io import BytesIO LUMA_T2V_AVERAGE_DURATION = 105 LUMA_I2V_AVERAGE_DURATION = 100 -def image_result_url_extractor(response: LumaGeneration): - return response.assets.image if hasattr(response, "assets") and hasattr(response.assets, "image") else None - -def video_result_url_extractor(response: LumaGeneration): - return response.assets.video if hasattr(response, "assets") and hasattr(response.assets, "video") else None class LumaReferenceNode(IO.ComfyNode): - """ - Holds an image and weight for use with Luma Generate Image node. - """ - @classmethod def define_schema(cls) -> IO.Schema: return IO.Schema( node_id="LumaReferenceNode", display_name="Luma Reference", category="api node/image/Luma", - description=cleandoc(cls.__doc__ or ""), + description="Holds an image and weight for use with Luma Generate Image node.", inputs=[ IO.Image.Input( "image", @@ -83,17 +65,10 @@ class LumaReferenceNode(IO.ComfyNode): ), ], outputs=[IO.Custom(LumaIO.LUMA_REF).Output(display_name="luma_ref")], - hidden=[ - IO.Hidden.auth_token_comfy_org, - IO.Hidden.api_key_comfy_org, - IO.Hidden.unique_id, - ], ) @classmethod - def execute( - cls, image: torch.Tensor, weight: float, luma_ref: LumaReferenceChain = None - ) -> IO.NodeOutput: + def execute(cls, image: torch.Tensor, weight: float, luma_ref: LumaReferenceChain = None) -> IO.NodeOutput: if luma_ref is not None: luma_ref = luma_ref.clone() else: @@ -103,17 +78,13 @@ class LumaReferenceNode(IO.ComfyNode): class LumaConceptsNode(IO.ComfyNode): - """ - Holds one or more Camera Concepts for use with Luma Text to Video and Luma Image to Video nodes. - """ - @classmethod def define_schema(cls) -> IO.Schema: return IO.Schema( node_id="LumaConceptsNode", display_name="Luma Concepts", category="api node/video/Luma", - description=cleandoc(cls.__doc__ or ""), + description="Camera Concepts for use with Luma Text to Video and Luma Image to Video nodes.", inputs=[ IO.Combo.Input( "concept1", @@ -138,11 +109,6 @@ class LumaConceptsNode(IO.ComfyNode): ), ], outputs=[IO.Custom(LumaIO.LUMA_CONCEPTS).Output(display_name="luma_concepts")], - hidden=[ - IO.Hidden.auth_token_comfy_org, - IO.Hidden.api_key_comfy_org, - IO.Hidden.unique_id, - ], ) @classmethod @@ -161,17 +127,13 @@ class LumaConceptsNode(IO.ComfyNode): class LumaImageGenerationNode(IO.ComfyNode): - """ - Generates images synchronously based on prompt and aspect ratio. - """ - @classmethod def define_schema(cls) -> IO.Schema: return IO.Schema( node_id="LumaImageNode", display_name="Luma Text to Image", category="api node/image/Luma", - description=cleandoc(cls.__doc__ or ""), + description="Generates images synchronously based on prompt and aspect ratio.", inputs=[ IO.String.Input( "prompt", @@ -237,45 +199,30 @@ class LumaImageGenerationNode(IO.ComfyNode): aspect_ratio: str, seed, style_image_weight: float, - image_luma_ref: LumaReferenceChain = None, - style_image: torch.Tensor = None, - character_image: torch.Tensor = None, + image_luma_ref: Optional[LumaReferenceChain] = None, + style_image: Optional[torch.Tensor] = None, + character_image: Optional[torch.Tensor] = None, ) -> IO.NodeOutput: validate_string(prompt, strip_whitespace=True, min_length=3) - auth_kwargs = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } # handle image_luma_ref api_image_ref = None if image_luma_ref is not None: - api_image_ref = await cls._convert_luma_refs( - image_luma_ref, max_refs=4, auth_kwargs=auth_kwargs, - ) + api_image_ref = await cls._convert_luma_refs(image_luma_ref, max_refs=4) # handle style_luma_ref api_style_ref = None if style_image is not None: - api_style_ref = await cls._convert_style_image( - style_image, weight=style_image_weight, auth_kwargs=auth_kwargs, - ) + api_style_ref = await cls._convert_style_image(style_image, weight=style_image_weight) # handle character_ref images character_ref = None if character_image is not None: - download_urls = await upload_images_to_comfyapi( - character_image, max_images=4, auth_kwargs=auth_kwargs, - ) - character_ref = LumaCharacterRef( - identity0=LumaImageIdentity(images=download_urls) - ) + download_urls = await upload_images_to_comfyapi(cls, character_image, max_images=4) + character_ref = LumaCharacterRef(identity0=LumaImageIdentity(images=download_urls)) - operation = SynchronousOperation( - endpoint=ApiEndpoint( - path="/proxy/luma/generations/image", - method=HttpMethod.POST, - request_model=LumaImageGenerationRequest, - response_model=LumaGeneration, - ), - request=LumaImageGenerationRequest( + response_api = await sync_op( + cls, + ApiEndpoint(path="/proxy/luma/generations/image", method="POST"), + response_model=LumaGeneration, + data=LumaImageGenerationRequest( prompt=prompt, model=model, aspect_ratio=aspect_ratio, @@ -283,41 +230,21 @@ class LumaImageGenerationNode(IO.ComfyNode): style_ref=api_style_ref, character_ref=character_ref, ), - auth_kwargs=auth_kwargs, ) - response_api: LumaGeneration = await operation.execute() - - operation = PollingOperation( - poll_endpoint=ApiEndpoint( - path=f"/proxy/luma/generations/{response_api.id}", - method=HttpMethod.GET, - request_model=EmptyRequest, - response_model=LumaGeneration, - ), - completed_statuses=[LumaState.completed], - failed_statuses=[LumaState.failed], + response_poll = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/luma/generations/{response_api.id}"), + response_model=LumaGeneration, status_extractor=lambda x: x.state, - result_url_extractor=image_result_url_extractor, - node_id=cls.hidden.unique_id, - auth_kwargs=auth_kwargs, ) - response_poll = await operation.execute() - - async with aiohttp.ClientSession() as session: - async with session.get(response_poll.assets.image) as img_response: - img = process_image_response(await img_response.content.read()) - return IO.NodeOutput(img) + return IO.NodeOutput(await download_url_to_image_tensor(response_poll.assets.image)) @classmethod - async def _convert_luma_refs( - cls, luma_ref: LumaReferenceChain, max_refs: int, auth_kwargs: Optional[dict[str,str]] = None - ): + async def _convert_luma_refs(cls, luma_ref: LumaReferenceChain, max_refs: int): luma_urls = [] ref_count = 0 for ref in luma_ref.refs: - download_urls = await upload_images_to_comfyapi( - ref.image, max_images=1, auth_kwargs=auth_kwargs - ) + download_urls = await upload_images_to_comfyapi(cls, ref.image, max_images=1) luma_urls.append(download_urls[0]) ref_count += 1 if ref_count >= max_refs: @@ -325,27 +252,19 @@ class LumaImageGenerationNode(IO.ComfyNode): return luma_ref.create_api_model(download_urls=luma_urls, max_refs=max_refs) @classmethod - async def _convert_style_image( - cls, style_image: torch.Tensor, weight: float, auth_kwargs: Optional[dict[str,str]] = None - ): - chain = LumaReferenceChain( - first_ref=LumaReference(image=style_image, weight=weight) - ) - return await cls._convert_luma_refs(chain, max_refs=1, auth_kwargs=auth_kwargs) + async def _convert_style_image(cls, style_image: torch.Tensor, weight: float): + chain = LumaReferenceChain(first_ref=LumaReference(image=style_image, weight=weight)) + return await cls._convert_luma_refs(chain, max_refs=1) class LumaImageModifyNode(IO.ComfyNode): - """ - Modifies images synchronously based on prompt and aspect ratio. - """ - @classmethod def define_schema(cls) -> IO.Schema: return IO.Schema( node_id="LumaImageModifyNode", display_name="Luma Image to Image", category="api node/image/Luma", - description=cleandoc(cls.__doc__ or ""), + description="Modifies images synchronously based on prompt and aspect ratio.", inputs=[ IO.Image.Input( "image", @@ -395,68 +314,37 @@ class LumaImageModifyNode(IO.ComfyNode): image_weight: float, seed, ) -> IO.NodeOutput: - auth_kwargs = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } - # first, upload image - download_urls = await upload_images_to_comfyapi( - image, max_images=1, auth_kwargs=auth_kwargs, - ) + download_urls = await upload_images_to_comfyapi(cls, image, max_images=1) image_url = download_urls[0] - # next, make Luma call with download url provided - operation = SynchronousOperation( - endpoint=ApiEndpoint( - path="/proxy/luma/generations/image", - method=HttpMethod.POST, - request_model=LumaImageGenerationRequest, - response_model=LumaGeneration, - ), - request=LumaImageGenerationRequest( + response_api = await sync_op( + cls, + ApiEndpoint(path="/proxy/luma/generations/image", method="POST"), + response_model=LumaGeneration, + data=LumaImageGenerationRequest( prompt=prompt, model=model, modify_image_ref=LumaModifyImageRef( - url=image_url, weight=round(max(min(1.0-image_weight, 0.98), 0.0), 2) + url=image_url, weight=round(max(min(1.0 - image_weight, 0.98), 0.0), 2) ), ), - auth_kwargs=auth_kwargs, ) - response_api: LumaGeneration = await operation.execute() - - operation = PollingOperation( - poll_endpoint=ApiEndpoint( - path=f"/proxy/luma/generations/{response_api.id}", - method=HttpMethod.GET, - request_model=EmptyRequest, - response_model=LumaGeneration, - ), - completed_statuses=[LumaState.completed], - failed_statuses=[LumaState.failed], + response_poll = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/luma/generations/{response_api.id}"), + response_model=LumaGeneration, status_extractor=lambda x: x.state, - result_url_extractor=image_result_url_extractor, - node_id=cls.hidden.unique_id, - auth_kwargs=auth_kwargs, ) - response_poll = await operation.execute() - - async with aiohttp.ClientSession() as session: - async with session.get(response_poll.assets.image) as img_response: - img = process_image_response(await img_response.content.read()) - return IO.NodeOutput(img) + return IO.NodeOutput(await download_url_to_image_tensor(response_poll.assets.image)) class LumaTextToVideoGenerationNode(IO.ComfyNode): - """ - Generates videos synchronously based on prompt and output_size. - """ - @classmethod def define_schema(cls) -> IO.Schema: return IO.Schema( node_id="LumaVideoNode", display_name="Luma Text to Video", category="api node/video/Luma", - description=cleandoc(cls.__doc__ or ""), + description="Generates videos synchronously based on prompt and output_size.", inputs=[ IO.String.Input( "prompt", @@ -498,7 +386,7 @@ class LumaTextToVideoGenerationNode(IO.ComfyNode): "luma_concepts", tooltip="Optional Camera Concepts to dictate camera motion via the Luma Concepts node.", optional=True, - ) + ), ], outputs=[IO.Video.Output()], hidden=[ @@ -519,24 +407,17 @@ class LumaTextToVideoGenerationNode(IO.ComfyNode): duration: str, loop: bool, seed, - luma_concepts: LumaConceptChain = None, + luma_concepts: Optional[LumaConceptChain] = None, ) -> IO.NodeOutput: validate_string(prompt, strip_whitespace=False, min_length=3) duration = duration if model != LumaVideoModel.ray_1_6 else None resolution = resolution if model != LumaVideoModel.ray_1_6 else None - auth_kwargs = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } - operation = SynchronousOperation( - endpoint=ApiEndpoint( - path="/proxy/luma/generations", - method=HttpMethod.POST, - request_model=LumaGenerationRequest, - response_model=LumaGeneration, - ), - request=LumaGenerationRequest( + response_api = await sync_op( + cls, + ApiEndpoint(path="/proxy/luma/generations", method="POST"), + response_model=LumaGeneration, + data=LumaGenerationRequest( prompt=prompt, model=model, resolution=resolution, @@ -545,47 +426,25 @@ class LumaTextToVideoGenerationNode(IO.ComfyNode): loop=loop, concepts=luma_concepts.create_api_model() if luma_concepts else None, ), - auth_kwargs=auth_kwargs, ) - response_api: LumaGeneration = await operation.execute() - - if cls.hidden.unique_id: - PromptServer.instance.send_progress_text(f"Luma video generation started: {response_api.id}", cls.hidden.unique_id) - - operation = PollingOperation( - poll_endpoint=ApiEndpoint( - path=f"/proxy/luma/generations/{response_api.id}", - method=HttpMethod.GET, - request_model=EmptyRequest, - response_model=LumaGeneration, - ), - completed_statuses=[LumaState.completed], - failed_statuses=[LumaState.failed], + response_poll = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/luma/generations/{response_api.id}"), + response_model=LumaGeneration, status_extractor=lambda x: x.state, - result_url_extractor=video_result_url_extractor, - node_id=cls.hidden.unique_id, estimated_duration=LUMA_T2V_AVERAGE_DURATION, - auth_kwargs=auth_kwargs, ) - response_poll = await operation.execute() - - async with aiohttp.ClientSession() as session: - async with session.get(response_poll.assets.video) as vid_response: - return IO.NodeOutput(VideoFromFile(BytesIO(await vid_response.content.read()))) + return IO.NodeOutput(await download_url_to_video_output(response_poll.assets.video)) class LumaImageToVideoGenerationNode(IO.ComfyNode): - """ - Generates videos synchronously based on prompt, input images, and output_size. - """ - @classmethod def define_schema(cls) -> IO.Schema: return IO.Schema( node_id="LumaImageToVideoNode", display_name="Luma Image to Video", category="api node/video/Luma", - description=cleandoc(cls.__doc__ or ""), + description="Generates videos synchronously based on prompt, input images, and output_size.", inputs=[ IO.String.Input( "prompt", @@ -637,7 +496,7 @@ class LumaImageToVideoGenerationNode(IO.ComfyNode): "luma_concepts", tooltip="Optional Camera Concepts to dictate camera motion via the Luma Concepts node.", optional=True, - ) + ), ], outputs=[IO.Video.Output()], hidden=[ @@ -662,25 +521,15 @@ class LumaImageToVideoGenerationNode(IO.ComfyNode): luma_concepts: LumaConceptChain = None, ) -> IO.NodeOutput: if first_image is None and last_image is None: - raise Exception( - "At least one of first_image and last_image requires an input." - ) - auth_kwargs = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } - keyframes = await cls._convert_to_keyframes(first_image, last_image, auth_kwargs=auth_kwargs) + raise Exception("At least one of first_image and last_image requires an input.") + keyframes = await cls._convert_to_keyframes(first_image, last_image) duration = duration if model != LumaVideoModel.ray_1_6 else None resolution = resolution if model != LumaVideoModel.ray_1_6 else None - - operation = SynchronousOperation( - endpoint=ApiEndpoint( - path="/proxy/luma/generations", - method=HttpMethod.POST, - request_model=LumaGenerationRequest, - response_model=LumaGeneration, - ), - request=LumaGenerationRequest( + response_api = await sync_op( + cls, + ApiEndpoint(path="/proxy/luma/generations", method="POST"), + response_model=LumaGeneration, + data=LumaGenerationRequest( prompt=prompt, model=model, aspect_ratio=LumaAspectRatio.ratio_16_9, # ignored, but still needed by the API for some reason @@ -690,54 +539,31 @@ class LumaImageToVideoGenerationNode(IO.ComfyNode): keyframes=keyframes, concepts=luma_concepts.create_api_model() if luma_concepts else None, ), - auth_kwargs=auth_kwargs, ) - response_api: LumaGeneration = await operation.execute() - - if cls.hidden.unique_id: - PromptServer.instance.send_progress_text(f"Luma video generation started: {response_api.id}", cls.hidden.unique_id) - - operation = PollingOperation( - poll_endpoint=ApiEndpoint( - path=f"/proxy/luma/generations/{response_api.id}", - method=HttpMethod.GET, - request_model=EmptyRequest, - response_model=LumaGeneration, - ), - completed_statuses=[LumaState.completed], - failed_statuses=[LumaState.failed], + response_poll = await poll_op( + cls, + poll_endpoint=ApiEndpoint(path=f"/proxy/luma/generations/{response_api.id}"), + response_model=LumaGeneration, status_extractor=lambda x: x.state, - result_url_extractor=video_result_url_extractor, - node_id=cls.hidden.unique_id, estimated_duration=LUMA_I2V_AVERAGE_DURATION, - auth_kwargs=auth_kwargs, ) - response_poll = await operation.execute() - - async with aiohttp.ClientSession() as session: - async with session.get(response_poll.assets.video) as vid_response: - return IO.NodeOutput(VideoFromFile(BytesIO(await vid_response.content.read()))) + return IO.NodeOutput(await download_url_to_video_output(response_poll.assets.video)) @classmethod async def _convert_to_keyframes( cls, first_image: torch.Tensor = None, last_image: torch.Tensor = None, - auth_kwargs: Optional[dict[str,str]] = None, ): if first_image is None and last_image is None: return None frame0 = None frame1 = None if first_image is not None: - download_urls = await upload_images_to_comfyapi( - first_image, max_images=1, auth_kwargs=auth_kwargs, - ) + download_urls = await upload_images_to_comfyapi(cls, first_image, max_images=1) frame0 = LumaImageReference(type="image", url=download_urls[0]) if last_image is not None: - download_urls = await upload_images_to_comfyapi( - last_image, max_images=1, auth_kwargs=auth_kwargs, - ) + download_urls = await upload_images_to_comfyapi(cls, last_image, max_images=1) frame1 = LumaImageReference(type="image", url=download_urls[0]) return LumaKeyframes(frame0=frame0, frame1=frame1) diff --git a/comfy_api_nodes/nodes_minimax.py b/comfy_api_nodes/nodes_minimax.py index e3722e79b..05cbb700f 100644 --- a/comfy_api_nodes/nodes_minimax.py +++ b/comfy_api_nodes/nodes_minimax.py @@ -1,71 +1,57 @@ -from inspect import cleandoc from typing import Optional -import logging -import torch +import torch from typing_extensions import override -from comfy_api.latest import ComfyExtension, IO -from comfy_api.input_impl.video_types import VideoFromFile -from comfy_api_nodes.apis import ( + +from comfy_api.latest import IO, ComfyExtension +from comfy_api_nodes.apis.minimax_api import ( + MinimaxFileRetrieveResponse, + MiniMaxModel, + MinimaxTaskResultResponse, MinimaxVideoGenerationRequest, MinimaxVideoGenerationResponse, - MinimaxFileRetrieveResponse, - MinimaxTaskResultResponse, SubjectReferenceItem, - MiniMaxModel, ) -from comfy_api_nodes.apis.client import ( +from comfy_api_nodes.util import ( ApiEndpoint, - HttpMethod, - SynchronousOperation, - PollingOperation, - EmptyRequest, -) -from comfy_api_nodes.apinode_utils import ( - download_url_to_bytesio, + download_url_to_video_output, + poll_op, + sync_op, upload_images_to_comfyapi, + validate_string, ) -from comfy_api_nodes.util import validate_string -from server import PromptServer - I2V_AVERAGE_DURATION = 114 T2V_AVERAGE_DURATION = 234 async def _generate_mm_video( + cls: type[IO.ComfyNode], *, - auth: dict[str, str], - node_id: str, prompt_text: str, seed: int, model: str, - image: Optional[torch.Tensor] = None, # used for ImageToVideo - subject: Optional[torch.Tensor] = None, # used for SubjectToVideo + image: Optional[torch.Tensor] = None, # used for ImageToVideo + subject: Optional[torch.Tensor] = None, # used for SubjectToVideo average_duration: Optional[int] = None, ) -> IO.NodeOutput: if image is None: validate_string(prompt_text, field_name="prompt_text") - # upload image, if passed in image_url = None if image is not None: - image_url = (await upload_images_to_comfyapi(image, max_images=1, auth_kwargs=auth))[0] + image_url = (await upload_images_to_comfyapi(cls, image, max_images=1))[0] # TODO: figure out how to deal with subject properly, API returns invalid params when using S2V-01 model subject_reference = None if subject is not None: - subject_url = (await upload_images_to_comfyapi(subject, max_images=1, auth_kwargs=auth))[0] + subject_url = (await upload_images_to_comfyapi(cls, subject, max_images=1))[0] subject_reference = [SubjectReferenceItem(image=subject_url)] - - video_generate_operation = SynchronousOperation( - endpoint=ApiEndpoint( - path="/proxy/minimax/video_generation", - method=HttpMethod.POST, - request_model=MinimaxVideoGenerationRequest, - response_model=MinimaxVideoGenerationResponse, - ), - request=MinimaxVideoGenerationRequest( + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/minimax/video_generation", method="POST"), + response_model=MinimaxVideoGenerationResponse, + data=MinimaxVideoGenerationRequest( model=MiniMaxModel(model), prompt=prompt_text, callback_url=None, @@ -73,81 +59,50 @@ async def _generate_mm_video( subject_reference=subject_reference, prompt_optimizer=None, ), - auth_kwargs=auth, ) - response = await video_generate_operation.execute() task_id = response.task_id if not task_id: raise Exception(f"MiniMax generation failed: {response.base_resp}") - video_generate_operation = PollingOperation( - poll_endpoint=ApiEndpoint( - path="/proxy/minimax/query/video_generation", - method=HttpMethod.GET, - request_model=EmptyRequest, - response_model=MinimaxTaskResultResponse, - query_params={"task_id": task_id}, - ), - completed_statuses=["Success"], - failed_statuses=["Fail"], + task_result = await poll_op( + cls, + ApiEndpoint(path="/proxy/minimax/query/video_generation", query_params={"task_id": task_id}), + response_model=MinimaxTaskResultResponse, status_extractor=lambda x: x.status.value, estimated_duration=average_duration, - node_id=node_id, - auth_kwargs=auth, ) - task_result = await video_generate_operation.execute() file_id = task_result.file_id if file_id is None: raise Exception("Request was not successful. Missing file ID.") - file_retrieve_operation = SynchronousOperation( - endpoint=ApiEndpoint( - path="/proxy/minimax/files/retrieve", - method=HttpMethod.GET, - request_model=EmptyRequest, - response_model=MinimaxFileRetrieveResponse, - query_params={"file_id": int(file_id)}, - ), - request=EmptyRequest(), - auth_kwargs=auth, + file_result = await sync_op( + cls, + ApiEndpoint(path="/proxy/minimax/files/retrieve", query_params={"file_id": int(file_id)}), + response_model=MinimaxFileRetrieveResponse, ) - file_result = await file_retrieve_operation.execute() file_url = file_result.file.download_url if file_url is None: - raise Exception( - f"No video was found in the response. Full response: {file_result.model_dump()}" - ) - logging.info("Generated video URL: %s", file_url) - if node_id: - if hasattr(file_result.file, "backup_download_url"): - message = f"Result URL: {file_url}\nBackup URL: {file_result.file.backup_download_url}" - else: - message = f"Result URL: {file_url}" - PromptServer.instance.send_progress_text(message, node_id) - - # Download and return as VideoFromFile - video_io = await download_url_to_bytesio(file_url) - if video_io is None: - error_msg = f"Failed to download video from {file_url}" - logging.error(error_msg) - raise Exception(error_msg) - return IO.NodeOutput(VideoFromFile(video_io)) + raise Exception(f"No video was found in the response. Full response: {file_result.model_dump()}") + if file_result.file.backup_download_url: + try: + return IO.NodeOutput(await download_url_to_video_output(file_url, timeout=10, max_retries=2)) + except Exception: # if we have a second URL to retrieve the result, try again using that one + return IO.NodeOutput( + await download_url_to_video_output(file_result.file.backup_download_url, max_retries=3) + ) + return IO.NodeOutput(await download_url_to_video_output(file_url)) class MinimaxTextToVideoNode(IO.ComfyNode): - """ - Generates videos synchronously based on a prompt, and optional parameters using MiniMax's API. - """ - @classmethod def define_schema(cls) -> IO.Schema: return IO.Schema( node_id="MinimaxTextToVideoNode", display_name="MiniMax Text to Video", category="api node/video/MiniMax", - description=cleandoc(cls.__doc__ or ""), + description="Generates videos synchronously based on a prompt, and optional parameters.", inputs=[ IO.String.Input( "prompt_text", @@ -189,11 +144,7 @@ class MinimaxTextToVideoNode(IO.ComfyNode): seed: int = 0, ) -> IO.NodeOutput: return await _generate_mm_video( - auth={ - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - }, - node_id=cls.hidden.unique_id, + cls, prompt_text=prompt_text, seed=seed, model=model, @@ -204,17 +155,13 @@ class MinimaxTextToVideoNode(IO.ComfyNode): class MinimaxImageToVideoNode(IO.ComfyNode): - """ - Generates videos synchronously based on an image and prompt, and optional parameters using MiniMax's API. - """ - @classmethod def define_schema(cls) -> IO.Schema: return IO.Schema( node_id="MinimaxImageToVideoNode", display_name="MiniMax Image to Video", category="api node/video/MiniMax", - description=cleandoc(cls.__doc__ or ""), + description="Generates videos synchronously based on an image and prompt, and optional parameters.", inputs=[ IO.Image.Input( "image", @@ -261,11 +208,7 @@ class MinimaxImageToVideoNode(IO.ComfyNode): seed: int = 0, ) -> IO.NodeOutput: return await _generate_mm_video( - auth={ - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - }, - node_id=cls.hidden.unique_id, + cls, prompt_text=prompt_text, seed=seed, model=model, @@ -276,17 +219,13 @@ class MinimaxImageToVideoNode(IO.ComfyNode): class MinimaxSubjectToVideoNode(IO.ComfyNode): - """ - Generates videos synchronously based on an image and prompt, and optional parameters using MiniMax's API. - """ - @classmethod def define_schema(cls) -> IO.Schema: return IO.Schema( node_id="MinimaxSubjectToVideoNode", display_name="MiniMax Subject to Video", category="api node/video/MiniMax", - description=cleandoc(cls.__doc__ or ""), + description="Generates videos synchronously based on an image and prompt, and optional parameters.", inputs=[ IO.Image.Input( "subject", @@ -333,11 +272,7 @@ class MinimaxSubjectToVideoNode(IO.ComfyNode): seed: int = 0, ) -> IO.NodeOutput: return await _generate_mm_video( - auth={ - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - }, - node_id=cls.hidden.unique_id, + cls, prompt_text=prompt_text, seed=seed, model=model, @@ -348,15 +283,13 @@ class MinimaxSubjectToVideoNode(IO.ComfyNode): class MinimaxHailuoVideoNode(IO.ComfyNode): - """Generates videos from prompt, with optional start frame using the new MiniMax Hailuo-02 model.""" - @classmethod def define_schema(cls) -> IO.Schema: return IO.Schema( node_id="MinimaxHailuoVideoNode", display_name="MiniMax Hailuo Video", category="api node/video/MiniMax", - description=cleandoc(cls.__doc__ or ""), + description="Generates videos from prompt, with optional start frame using the new MiniMax Hailuo-02 model.", inputs=[ IO.String.Input( "prompt_text", @@ -420,10 +353,6 @@ class MinimaxHailuoVideoNode(IO.ComfyNode): resolution: str = "768P", model: str = "MiniMax-Hailuo-02", ) -> IO.NodeOutput: - auth = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } if first_frame_image is None: validate_string(prompt_text, field_name="prompt_text") @@ -435,16 +364,13 @@ class MinimaxHailuoVideoNode(IO.ComfyNode): # upload image, if passed in image_url = None if first_frame_image is not None: - image_url = (await upload_images_to_comfyapi(first_frame_image, max_images=1, auth_kwargs=auth))[0] + image_url = (await upload_images_to_comfyapi(cls, first_frame_image, max_images=1))[0] - video_generate_operation = SynchronousOperation( - endpoint=ApiEndpoint( - path="/proxy/minimax/video_generation", - method=HttpMethod.POST, - request_model=MinimaxVideoGenerationRequest, - response_model=MinimaxVideoGenerationResponse, - ), - request=MinimaxVideoGenerationRequest( + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/minimax/video_generation", method="POST"), + response_model=MinimaxVideoGenerationResponse, + data=MinimaxVideoGenerationRequest( model=MiniMaxModel(model), prompt=prompt_text, callback_url=None, @@ -453,67 +379,42 @@ class MinimaxHailuoVideoNode(IO.ComfyNode): duration=duration, resolution=resolution, ), - auth_kwargs=auth, ) - response = await video_generate_operation.execute() task_id = response.task_id if not task_id: raise Exception(f"MiniMax generation failed: {response.base_resp}") average_duration = 120 if resolution == "768P" else 240 - video_generate_operation = PollingOperation( - poll_endpoint=ApiEndpoint( - path="/proxy/minimax/query/video_generation", - method=HttpMethod.GET, - request_model=EmptyRequest, - response_model=MinimaxTaskResultResponse, - query_params={"task_id": task_id}, - ), - completed_statuses=["Success"], - failed_statuses=["Fail"], + task_result = await poll_op( + cls, + ApiEndpoint(path="/proxy/minimax/query/video_generation", query_params={"task_id": task_id}), + response_model=MinimaxTaskResultResponse, status_extractor=lambda x: x.status.value, estimated_duration=average_duration, - node_id=cls.hidden.unique_id, - auth_kwargs=auth, ) - task_result = await video_generate_operation.execute() file_id = task_result.file_id if file_id is None: raise Exception("Request was not successful. Missing file ID.") - file_retrieve_operation = SynchronousOperation( - endpoint=ApiEndpoint( - path="/proxy/minimax/files/retrieve", - method=HttpMethod.GET, - request_model=EmptyRequest, - response_model=MinimaxFileRetrieveResponse, - query_params={"file_id": int(file_id)}, - ), - request=EmptyRequest(), - auth_kwargs=auth, + file_result = await sync_op( + cls, + ApiEndpoint(path="/proxy/minimax/files/retrieve", query_params={"file_id": int(file_id)}), + response_model=MinimaxFileRetrieveResponse, ) - file_result = await file_retrieve_operation.execute() file_url = file_result.file.download_url if file_url is None: - raise Exception( - f"No video was found in the response. Full response: {file_result.model_dump()}" - ) - logging.info("Generated video URL: %s", file_url) - if cls.hidden.unique_id: - if hasattr(file_result.file, "backup_download_url"): - message = f"Result URL: {file_url}\nBackup URL: {file_result.file.backup_download_url}" - else: - message = f"Result URL: {file_url}" - PromptServer.instance.send_progress_text(message, cls.hidden.unique_id) + raise Exception(f"No video was found in the response. Full response: {file_result.model_dump()}") - video_io = await download_url_to_bytesio(file_url) - if video_io is None: - error_msg = f"Failed to download video from {file_url}" - logging.error(error_msg) - raise Exception(error_msg) - return IO.NodeOutput(VideoFromFile(video_io)) + if file_result.file.backup_download_url: + try: + return IO.NodeOutput(await download_url_to_video_output(file_url, timeout=10, max_retries=2)) + except Exception: # if we have a second URL to retrieve the result, try again using that one + return IO.NodeOutput( + await download_url_to_video_output(file_result.file.backup_download_url, max_retries=3) + ) + return IO.NodeOutput(await download_url_to_video_output(file_url)) class MinimaxExtension(ComfyExtension): diff --git a/comfy_api_nodes/util/client.py b/comfy_api_nodes/util/client.py index 9c036d64b..9ae512fe5 100644 --- a/comfy_api_nodes/util/client.py +++ b/comfy_api_nodes/util/client.py @@ -78,7 +78,7 @@ class _PollUIState: _RETRY_STATUS = {408, 429, 500, 502, 503, 504} COMPLETED_STATUSES = ["succeeded", "succeed", "success", "completed"] -FAILED_STATUSES = ["cancelled", "canceled", "failed", "error"] +FAILED_STATUSES = ["cancelled", "canceled", "fail", "failed", "error"] QUEUED_STATUSES = ["created", "queued", "queueing", "submitted"] diff --git a/comfy_api_nodes/util/download_helpers.py b/comfy_api_nodes/util/download_helpers.py index f89045e12..364874bed 100644 --- a/comfy_api_nodes/util/download_helpers.py +++ b/comfy_api_nodes/util/download_helpers.py @@ -232,11 +232,12 @@ async def download_url_to_video_output( video_url: str, *, timeout: float = None, + max_retries: int = 5, cls: type[COMFY_IO.ComfyNode] = None, ) -> VideoFromFile: """Downloads a video from a URL and returns a `VIDEO` output.""" result = BytesIO() - await download_url_to_bytesio(video_url, result, timeout=timeout, cls=cls) + await download_url_to_bytesio(video_url, result, timeout=timeout, max_retries=max_retries, cls=cls) return VideoFromFile(result) From 1a58087ac2eb64be3645934d0025aafaa5bdce38 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 29 Oct 2025 12:43:51 -0700 Subject: [PATCH 050/647] Reduce memory usage for fp8 scaled op. (#10531) --- comfy/quant_ops.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy/quant_ops.py b/comfy/quant_ops.py index b14e03084..fb35a0d40 100644 --- a/comfy/quant_ops.py +++ b/comfy/quant_ops.py @@ -358,7 +358,7 @@ class TensorCoreFP8Layout(QuantizedLayout): scale = scale.to(device=tensor.device, dtype=torch.float32) lp_amax = torch.finfo(dtype).max - tensor_scaled = tensor.float() / scale + tensor_scaled = tensor * (1.0 / scale).to(tensor.dtype) torch.clamp(tensor_scaled, min=-lp_amax, max=lp_amax, out=tensor_scaled) qdata = tensor_scaled.to(dtype, memory_format=torch.contiguous_format) From ec4fc2a09a390d0d81500c51fb9e4d8a7a5ce1fc Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 29 Oct 2025 12:48:06 -0700 Subject: [PATCH 051/647] Fix case of weights not being unpinned. (#10533) --- comfy/model_patcher.py | 1 + 1 file changed, 1 insertion(+) diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index aec73349c..119119e96 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -1283,5 +1283,6 @@ class ModelPatcher: self.clear_cached_hook_weights() def __del__(self): + self.unpin_all_weights() self.detach(unpatch_all=False) From ab7ab5be23fb9b71d1790f424e7dcf91dc1fe0cc Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Thu, 30 Oct 2025 07:17:46 +1000 Subject: [PATCH 052/647] Fix Race condition in --async-offload that can cause corruption (#10501) * mm: factor out the current stream getter Make this a reusable function. * ops: sync the offload stream with the consumption of w&b This sync is nessacary as pytorch will queue cuda async frees on the same stream as created to tensor. In the case of async offload, this will be on the offload stream. Weights and biases can go out of scope in python which then triggers the pytorch garbage collector to queue the free operation on the offload stream possible before the compute stream has used the weight. This causes a use after free on weight data leading to total corruption of some workflows. So sync the offload stream with the compute stream after the weight has been used so the free has to wait for the weight to be used. The cast_bias_weight is extended in a backwards compatible way with the new behaviour opt-in on a defaulted parameter. This handles custom node packs calling cast_bias_weight and defeatures async-offload for them (as they do not handle the race). The pattern is now: cast_bias_weight(... , offloadable=True) #This might be offloaded thing(weight, bias, ...) uncast_bias_weight(...) * controlnet: adopt new cast_bias_weight synchronization scheme This is nessacary for safe async weight offloading. * mm: sync the last stream in the queue, not the next Currently this peeks ahead to sync the next stream in the queue of streams with the compute stream. This doesnt allow a lot of parallelization, as then end result is you can only get one weight load ahead regardless of how many streams you have. Rotate the loop logic here to synchronize the end of the queue before returning the next stream. This allows weights to be loaded ahead of the compute streams position. --- comfy/controlnet.py | 17 +++--- comfy/model_management.py | 28 +++++---- comfy/ops.py | 121 ++++++++++++++++++++++++++++---------- 3 files changed, 114 insertions(+), 52 deletions(-) diff --git a/comfy/controlnet.py b/comfy/controlnet.py index f08ff4b36..0b5e30f52 100644 --- a/comfy/controlnet.py +++ b/comfy/controlnet.py @@ -310,11 +310,13 @@ class ControlLoraOps: self.bias = None def forward(self, input): - weight, bias = comfy.ops.cast_bias_weight(self, input) + weight, bias, offload_stream = comfy.ops.cast_bias_weight(self, input, offloadable=True) if self.up is not None: - return torch.nn.functional.linear(input, weight + (torch.mm(self.up.flatten(start_dim=1), self.down.flatten(start_dim=1))).reshape(self.weight.shape).type(input.dtype), bias) + x = torch.nn.functional.linear(input, weight + (torch.mm(self.up.flatten(start_dim=1), self.down.flatten(start_dim=1))).reshape(self.weight.shape).type(input.dtype), bias) else: - return torch.nn.functional.linear(input, weight, bias) + x = torch.nn.functional.linear(input, weight, bias) + comfy.ops.uncast_bias_weight(self, weight, bias, offload_stream) + return x class Conv2d(torch.nn.Module, comfy.ops.CastWeightBiasOp): def __init__( @@ -350,12 +352,13 @@ class ControlLoraOps: def forward(self, input): - weight, bias = comfy.ops.cast_bias_weight(self, input) + weight, bias, offload_stream = comfy.ops.cast_bias_weight(self, input, offloadable=True) if self.up is not None: - return torch.nn.functional.conv2d(input, weight + (torch.mm(self.up.flatten(start_dim=1), self.down.flatten(start_dim=1))).reshape(self.weight.shape).type(input.dtype), bias, self.stride, self.padding, self.dilation, self.groups) + x = torch.nn.functional.conv2d(input, weight + (torch.mm(self.up.flatten(start_dim=1), self.down.flatten(start_dim=1))).reshape(self.weight.shape).type(input.dtype), bias, self.stride, self.padding, self.dilation, self.groups) else: - return torch.nn.functional.conv2d(input, weight, bias, self.stride, self.padding, self.dilation, self.groups) - + x = torch.nn.functional.conv2d(input, weight, bias, self.stride, self.padding, self.dilation, self.groups) + comfy.ops.uncast_bias_weight(self, weight, bias, offload_stream) + return x class ControlLora(ControlNet): def __init__(self, control_weights, global_average_pooling=False, model_options={}): #TODO? model_options diff --git a/comfy/model_management.py b/comfy/model_management.py index 3e5b977d4..79c0dfdb4 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -1013,6 +1013,16 @@ if args.async_offload: NUM_STREAMS = 2 logging.info("Using async weight offloading with {} streams".format(NUM_STREAMS)) +def current_stream(device): + if device is None: + return None + if is_device_cuda(device): + return torch.cuda.current_stream() + elif is_device_xpu(device): + return torch.xpu.current_stream() + else: + return None + stream_counters = {} def get_offload_stream(device): stream_counter = stream_counters.get(device, 0) @@ -1021,21 +1031,17 @@ def get_offload_stream(device): if device in STREAMS: ss = STREAMS[device] - s = ss[stream_counter] + #Sync the oldest stream in the queue with the current + ss[stream_counter].wait_stream(current_stream(device)) stream_counter = (stream_counter + 1) % len(ss) - if is_device_cuda(device): - ss[stream_counter].wait_stream(torch.cuda.current_stream()) - elif is_device_xpu(device): - ss[stream_counter].wait_stream(torch.xpu.current_stream()) stream_counters[device] = stream_counter - return s + return ss[stream_counter] elif is_device_cuda(device): ss = [] for k in range(NUM_STREAMS): ss.append(torch.cuda.Stream(device=device, priority=0)) STREAMS[device] = ss s = ss[stream_counter] - stream_counter = (stream_counter + 1) % len(ss) stream_counters[device] = stream_counter return s elif is_device_xpu(device): @@ -1044,18 +1050,14 @@ def get_offload_stream(device): ss.append(torch.xpu.Stream(device=device, priority=0)) STREAMS[device] = ss s = ss[stream_counter] - stream_counter = (stream_counter + 1) % len(ss) stream_counters[device] = stream_counter return s return None def sync_stream(device, stream): - if stream is None: + if stream is None or current_stream(device) is None: return - if is_device_cuda(device): - torch.cuda.current_stream().wait_stream(stream) - elif is_device_xpu(device): - torch.xpu.current_stream().wait_stream(stream) + current_stream(device).wait_stream(stream) def cast_to(weight, dtype=None, device=None, non_blocking=False, copy=False, stream=None): if device is None or weight.device == device: diff --git a/comfy/ops.py b/comfy/ops.py index 93731eedf..71ca7a2bd 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -70,8 +70,12 @@ cast_to = comfy.model_management.cast_to #TODO: remove once no more references def cast_to_input(weight, input, non_blocking=False, copy=True): return comfy.model_management.cast_to(weight, input.dtype, input.device, non_blocking=non_blocking, copy=copy) + @torch.compiler.disable() -def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None): +def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, offloadable=False): + # NOTE: offloadable=False is a a legacy and if you are a custom node author reading this please pass + # offloadable=True and call uncast_bias_weight() after your last usage of the weight/bias. This + # will add async-offload support to your cast and improve performance. if input is not None: if dtype is None: dtype = input.dtype @@ -80,7 +84,11 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None): if device is None: device = input.device - offload_stream = comfy.model_management.get_offload_stream(device) + if offloadable: + offload_stream = comfy.model_management.get_offload_stream(device) + else: + offload_stream = None + if offload_stream is not None: wf_context = offload_stream else: @@ -105,7 +113,24 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None): weight = f(weight) comfy.model_management.sync_stream(device, offload_stream) - return weight, bias + if offloadable: + return weight, bias, offload_stream + else: + #Legacy function signature + return weight, bias + + +def uncast_bias_weight(s, weight, bias, offload_stream): + if offload_stream is None: + return + if weight is not None: + device = weight.device + else: + if bias is None: + return + device = bias.device + offload_stream.wait_stream(comfy.model_management.current_stream(device)) + class CastWeightBiasOp: comfy_cast_weights = False @@ -118,8 +143,10 @@ class disable_weight_init: return None def forward_comfy_cast_weights(self, input): - weight, bias = cast_bias_weight(self, input) - return torch.nn.functional.linear(input, weight, bias) + weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True) + x = torch.nn.functional.linear(input, weight, bias) + uncast_bias_weight(self, weight, bias, offload_stream) + return x def forward(self, *args, **kwargs): run_every_op() @@ -133,8 +160,10 @@ class disable_weight_init: return None def forward_comfy_cast_weights(self, input): - weight, bias = cast_bias_weight(self, input) - return self._conv_forward(input, weight, bias) + weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True) + x = self._conv_forward(input, weight, bias) + uncast_bias_weight(self, weight, bias, offload_stream) + return x def forward(self, *args, **kwargs): run_every_op() @@ -148,8 +177,10 @@ class disable_weight_init: return None def forward_comfy_cast_weights(self, input): - weight, bias = cast_bias_weight(self, input) - return self._conv_forward(input, weight, bias) + weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True) + x = self._conv_forward(input, weight, bias) + uncast_bias_weight(self, weight, bias, offload_stream) + return x def forward(self, *args, **kwargs): run_every_op() @@ -172,8 +203,10 @@ class disable_weight_init: return super()._conv_forward(input, weight, bias, *args, **kwargs) def forward_comfy_cast_weights(self, input): - weight, bias = cast_bias_weight(self, input) - return self._conv_forward(input, weight, bias) + weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True) + x = self._conv_forward(input, weight, bias) + uncast_bias_weight(self, weight, bias, offload_stream) + return x def forward(self, *args, **kwargs): run_every_op() @@ -187,8 +220,10 @@ class disable_weight_init: return None def forward_comfy_cast_weights(self, input): - weight, bias = cast_bias_weight(self, input) - return torch.nn.functional.group_norm(input, self.num_groups, weight, bias, self.eps) + weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True) + x = torch.nn.functional.group_norm(input, self.num_groups, weight, bias, self.eps) + uncast_bias_weight(self, weight, bias, offload_stream) + return x def forward(self, *args, **kwargs): run_every_op() @@ -203,11 +238,14 @@ class disable_weight_init: def forward_comfy_cast_weights(self, input): if self.weight is not None: - weight, bias = cast_bias_weight(self, input) + weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True) else: weight = None bias = None - return torch.nn.functional.layer_norm(input, self.normalized_shape, weight, bias, self.eps) + offload_stream = None + x = torch.nn.functional.layer_norm(input, self.normalized_shape, weight, bias, self.eps) + uncast_bias_weight(self, weight, bias, offload_stream) + return x def forward(self, *args, **kwargs): run_every_op() @@ -223,11 +261,15 @@ class disable_weight_init: def forward_comfy_cast_weights(self, input): if self.weight is not None: - weight, bias = cast_bias_weight(self, input) + weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True) else: weight = None - return comfy.rmsnorm.rms_norm(input, weight, self.eps) # TODO: switch to commented out line when old torch is deprecated - # return torch.nn.functional.rms_norm(input, self.normalized_shape, weight, self.eps) + bias = None + offload_stream = None + x = comfy.rmsnorm.rms_norm(input, weight, self.eps) # TODO: switch to commented out line when old torch is deprecated + # x = torch.nn.functional.rms_norm(input, self.normalized_shape, weight, self.eps) + uncast_bias_weight(self, weight, bias, offload_stream) + return x def forward(self, *args, **kwargs): run_every_op() @@ -246,10 +288,12 @@ class disable_weight_init: input, output_size, self.stride, self.padding, self.kernel_size, num_spatial_dims, self.dilation) - weight, bias = cast_bias_weight(self, input) - return torch.nn.functional.conv_transpose2d( + weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True) + x = torch.nn.functional.conv_transpose2d( input, weight, bias, self.stride, self.padding, output_padding, self.groups, self.dilation) + uncast_bias_weight(self, weight, bias, offload_stream) + return x def forward(self, *args, **kwargs): run_every_op() @@ -268,10 +312,12 @@ class disable_weight_init: input, output_size, self.stride, self.padding, self.kernel_size, num_spatial_dims, self.dilation) - weight, bias = cast_bias_weight(self, input) - return torch.nn.functional.conv_transpose1d( + weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True) + x = torch.nn.functional.conv_transpose1d( input, weight, bias, self.stride, self.padding, output_padding, self.groups, self.dilation) + uncast_bias_weight(self, weight, bias, offload_stream) + return x def forward(self, *args, **kwargs): run_every_op() @@ -289,8 +335,11 @@ class disable_weight_init: output_dtype = out_dtype if self.weight.dtype == torch.float16 or self.weight.dtype == torch.bfloat16: out_dtype = None - weight, bias = cast_bias_weight(self, device=input.device, dtype=out_dtype) - return torch.nn.functional.embedding(input, weight, self.padding_idx, self.max_norm, self.norm_type, self.scale_grad_by_freq, self.sparse).to(dtype=output_dtype) + weight, bias, offload_stream = cast_bias_weight(self, device=input.device, dtype=out_dtype, offloadable=True) + x = torch.nn.functional.embedding(input, weight, self.padding_idx, self.max_norm, self.norm_type, self.scale_grad_by_freq, self.sparse).to(dtype=output_dtype) + uncast_bias_weight(self, weight, bias, offload_stream) + return x + def forward(self, *args, **kwargs): run_every_op() @@ -361,7 +410,7 @@ def fp8_linear(self, input): input_dtype = input.dtype if len(input.shape) == 3: - w, bias = cast_bias_weight(self, input, dtype=dtype, bias_dtype=input_dtype) + w, bias, offload_stream = cast_bias_weight(self, input, dtype=dtype, bias_dtype=input_dtype, offloadable=True) scale_weight = self.scale_weight scale_input = self.scale_input @@ -382,6 +431,8 @@ def fp8_linear(self, input): quantized_input = QuantizedTensor.from_float(input.reshape(-1, input_shape[2]), TensorCoreFP8Layout, scale=scale_input, dtype=dtype) o = torch.nn.functional.linear(quantized_input, quantized_weight, bias) + uncast_bias_weight(self, w, bias, offload_stream) + if tensor_2d: return o.reshape(input_shape[0], -1) return o.reshape((-1, input_shape[1], self.weight.shape[0])) @@ -404,8 +455,10 @@ class fp8_ops(manual_cast): except Exception as e: logging.info("Exception during fp8 op: {}".format(e)) - weight, bias = cast_bias_weight(self, input) - return torch.nn.functional.linear(input, weight, bias) + weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True) + x = torch.nn.functional.linear(input, weight, bias) + uncast_bias_weight(self, weight, bias, offload_stream) + return x def scaled_fp8_ops(fp8_matrix_mult=False, scale_input=False, override_dtype=None): logging.info("Using scaled fp8: fp8 matrix mult: {}, scale input: {}".format(fp8_matrix_mult, scale_input)) @@ -433,12 +486,14 @@ def scaled_fp8_ops(fp8_matrix_mult=False, scale_input=False, override_dtype=None if out is not None: return out - weight, bias = cast_bias_weight(self, input) + weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True) if weight.numel() < input.numel(): #TODO: optimize - return torch.nn.functional.linear(input, weight * self.scale_weight.to(device=weight.device, dtype=weight.dtype), bias) + x = torch.nn.functional.linear(input, weight * self.scale_weight.to(device=weight.device, dtype=weight.dtype), bias) else: - return torch.nn.functional.linear(input * self.scale_weight.to(device=weight.device, dtype=weight.dtype), weight, bias) + x = torch.nn.functional.linear(input * self.scale_weight.to(device=weight.device, dtype=weight.dtype), weight, bias) + uncast_bias_weight(self, weight, bias, offload_stream) + return x def convert_weight(self, weight, inplace=False, **kwargs): if inplace: @@ -577,8 +632,10 @@ class MixedPrecisionOps(disable_weight_init): return torch.nn.functional.linear(input, weight, bias) def forward_comfy_cast_weights(self, input): - weight, bias = cast_bias_weight(self, input) - return self._forward(input, weight, bias) + weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True) + x = self._forward(input, weight, bias) + uncast_bias_weight(self, weight, bias, offload_stream) + return x def forward(self, input, *args, **kwargs): run_every_op() From 25de7b1bfa22dd98922f047a1342cc97f8e46c5b Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 29 Oct 2025 14:20:27 -0700 Subject: [PATCH 053/647] Try to fix slow load issue on low ram hardware with pinned mem. (#10536) --- comfy/model_patcher.py | 10 ++++++++-- 1 file changed, 8 insertions(+), 2 deletions(-) diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index 119119e96..74b9e48bc 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -658,6 +658,7 @@ class ModelPatcher: loading = self._load_list() load_completely = [] + offloaded = [] loading.sort(reverse=True) for x in loading: n = x[1] @@ -699,8 +700,7 @@ class ModelPatcher: patch_counter += 1 cast_weight = True - for param in params: - self.pin_weight_to_device("{}.{}".format(n, param)) + offloaded.append((module_mem, n, m, params)) else: if hasattr(m, "comfy_cast_weights"): wipe_lowvram_weight(m) @@ -741,6 +741,12 @@ class ModelPatcher: for x in load_completely: x[2].to(device_to) + for x in offloaded: + n = x[1] + params = x[3] + for param in params: + self.pin_weight_to_device("{}.{}".format(n, param)) + if lowvram_counter > 0: logging.info("loaded partially {} {} {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (1024 * 1024), patch_counter)) self.model.model_lowvram = True From 906c0899575a83ac69bb095e835fdec748891da4 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 29 Oct 2025 16:29:01 -0700 Subject: [PATCH 054/647] Fix small performance regression with fp8 fast and scaled fp8. (#10537) --- comfy/ops.py | 6 +++++- comfy/quant_ops.py | 5 +++-- 2 files changed, 8 insertions(+), 3 deletions(-) diff --git a/comfy/ops.py b/comfy/ops.py index 71ca7a2bd..18f6b804b 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -421,14 +421,18 @@ def fp8_linear(self, input): if scale_input is None: scale_input = torch.ones((), device=input.device, dtype=torch.float32) + input = torch.clamp(input, min=-448, max=448, out=input) + input = input.reshape(-1, input_shape[2]).to(dtype).contiguous() + layout_params_weight = {'scale': scale_input, 'orig_dtype': input_dtype} + quantized_input = QuantizedTensor(input.reshape(-1, input_shape[2]).to(dtype).contiguous(), TensorCoreFP8Layout, layout_params_weight) else: scale_input = scale_input.to(input.device) + quantized_input = QuantizedTensor.from_float(input.reshape(-1, input_shape[2]), TensorCoreFP8Layout, scale=scale_input, dtype=dtype) # Wrap weight in QuantizedTensor - this enables unified dispatch # Call F.linear - __torch_dispatch__ routes to fp8_linear handler in quant_ops.py! layout_params_weight = {'scale': scale_weight, 'orig_dtype': input_dtype} quantized_weight = QuantizedTensor(w, TensorCoreFP8Layout, layout_params_weight) - quantized_input = QuantizedTensor.from_float(input.reshape(-1, input_shape[2]), TensorCoreFP8Layout, scale=scale_input, dtype=dtype) o = torch.nn.functional.linear(quantized_input, quantized_weight, bias) uncast_bias_weight(self, w, bias, offload_stream) diff --git a/comfy/quant_ops.py b/comfy/quant_ops.py index fb35a0d40..c822fe53c 100644 --- a/comfy/quant_ops.py +++ b/comfy/quant_ops.py @@ -357,9 +357,10 @@ class TensorCoreFP8Layout(QuantizedLayout): scale = torch.tensor(scale) scale = scale.to(device=tensor.device, dtype=torch.float32) - lp_amax = torch.finfo(dtype).max tensor_scaled = tensor * (1.0 / scale).to(tensor.dtype) - torch.clamp(tensor_scaled, min=-lp_amax, max=lp_amax, out=tensor_scaled) + # TODO: uncomment this if it's actually needed because the clamp has a small performance penality' + # lp_amax = torch.finfo(dtype).max + # torch.clamp(tensor_scaled, min=-lp_amax, max=lp_amax, out=tensor_scaled) qdata = tensor_scaled.to(dtype, memory_format=torch.contiguous_format) layout_params = { From 998bf60bebd03e57a55e106434657849342b733f Mon Sep 17 00:00:00 2001 From: Jedrzej Kosinski Date: Wed, 29 Oct 2025 16:37:06 -0700 Subject: [PATCH 055/647] Add units/info for the numbers displayed on 'load completely' and 'load partially' log messages (#10538) --- comfy/model_patcher.py | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index 74b9e48bc..ed3f3f5cb 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -655,6 +655,7 @@ class ModelPatcher: mem_counter = 0 patch_counter = 0 lowvram_counter = 0 + lowvram_mem_counter = 0 loading = self._load_list() load_completely = [] @@ -675,6 +676,7 @@ class ModelPatcher: if mem_counter + module_mem >= lowvram_model_memory: lowvram_weight = True lowvram_counter += 1 + lowvram_mem_counter += module_mem if hasattr(m, "prev_comfy_cast_weights"): #Already lowvramed continue @@ -748,10 +750,10 @@ class ModelPatcher: self.pin_weight_to_device("{}.{}".format(n, param)) if lowvram_counter > 0: - logging.info("loaded partially {} {} {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (1024 * 1024), patch_counter)) + logging.info("loaded partially; {:.2f} MB usable, {:.2f} MB loaded, {:.2f} MB offloaded, lowvram patches: {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (1024 * 1024), lowvram_mem_counter / (1024 * 1024), patch_counter)) self.model.model_lowvram = True else: - logging.info("loaded completely {} {} {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (1024 * 1024), full_load)) + logging.info("loaded completely; {:.2f} MB usable, {:.2f} MB loaded, full load: {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (1024 * 1024), full_load)) self.model.model_lowvram = False if full_load: self.model.to(device_to) From 163b629c70a349c7d1e91eebc5365713e770af8a Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Thu, 30 Oct 2025 08:49:03 +0200 Subject: [PATCH 056/647] use new API client in Pixverse and Ideogram nodes (#10543) --- comfy_api_nodes/apinode_utils.py | 109 +------------ comfy_api_nodes/nodes_bfl.py | 43 +---- comfy_api_nodes/nodes_bytedance.py | 12 +- comfy_api_nodes/nodes_ideogram.py | 137 ++++------------ comfy_api_nodes/nodes_kling.py | 2 +- comfy_api_nodes/nodes_pixverse.py | 194 ++++++----------------- comfy_api_nodes/nodes_recraft.py | 4 +- comfy_api_nodes/nodes_runway.py | 10 +- comfy_api_nodes/nodes_vidu.py | 12 +- comfy_api_nodes/util/__init__.py | 10 +- comfy_api_nodes/util/conversions.py | 21 +++ comfy_api_nodes/util/validation_utils.py | 125 ++++++++++----- 12 files changed, 220 insertions(+), 459 deletions(-) diff --git a/comfy_api_nodes/apinode_utils.py b/comfy_api_nodes/apinode_utils.py index 6a72b9d1d..ecd604ff8 100644 --- a/comfy_api_nodes/apinode_utils.py +++ b/comfy_api_nodes/apinode_utils.py @@ -1,15 +1,12 @@ from __future__ import annotations import aiohttp import mimetypes -from typing import Optional, Union -from comfy.utils import common_upscale +from typing import Union from server import PromptServer -from comfy.cli_args import args import numpy as np from PIL import Image import torch -import math import base64 from io import BytesIO @@ -60,85 +57,6 @@ async def validate_and_cast_response( return torch.stack(image_tensors, dim=0) -def validate_aspect_ratio( - aspect_ratio: str, - minimum_ratio: float, - maximum_ratio: float, - minimum_ratio_str: str, - maximum_ratio_str: str, -) -> float: - """Validates and casts an aspect ratio string to a float. - - Args: - aspect_ratio: The aspect ratio string to validate. - minimum_ratio: The minimum aspect ratio. - maximum_ratio: The maximum aspect ratio. - minimum_ratio_str: The minimum aspect ratio string. - maximum_ratio_str: The maximum aspect ratio string. - - Returns: - The validated and cast aspect ratio. - - Raises: - Exception: If the aspect ratio is not valid. - """ - # get ratio values - numbers = aspect_ratio.split(":") - if len(numbers) != 2: - raise TypeError( - f"Aspect ratio must be in the format X:Y, such as 16:9, but was {aspect_ratio}." - ) - try: - numerator = int(numbers[0]) - denominator = int(numbers[1]) - except ValueError as exc: - raise TypeError( - f"Aspect ratio must contain numbers separated by ':', such as 16:9, but was {aspect_ratio}." - ) from exc - calculated_ratio = numerator / denominator - # if not close to minimum and maximum, check bounds - if not math.isclose(calculated_ratio, minimum_ratio) or not math.isclose( - calculated_ratio, maximum_ratio - ): - if calculated_ratio < minimum_ratio: - raise TypeError( - f"Aspect ratio cannot reduce to any less than {minimum_ratio_str} ({minimum_ratio}), but was {aspect_ratio} ({calculated_ratio})." - ) - if calculated_ratio > maximum_ratio: - raise TypeError( - f"Aspect ratio cannot reduce to any greater than {maximum_ratio_str} ({maximum_ratio}), but was {aspect_ratio} ({calculated_ratio})." - ) - return aspect_ratio - - -async def download_url_to_bytesio( - url: str, timeout: int = None, auth_kwargs: Optional[dict[str, str]] = None -) -> BytesIO: - """Downloads content from a URL using requests and returns it as BytesIO. - - Args: - url: The URL to download. - timeout: Request timeout in seconds. Defaults to None (no timeout). - - Returns: - BytesIO object containing the downloaded content. - """ - headers = {} - if url.startswith("/proxy/"): - url = str(args.comfy_api_base).rstrip("/") + url - auth_token = auth_kwargs.get("auth_token") - comfy_api_key = auth_kwargs.get("comfy_api_key") - if auth_token: - headers["Authorization"] = f"Bearer {auth_token}" - elif comfy_api_key: - headers["X-API-KEY"] = comfy_api_key - timeout_cfg = aiohttp.ClientTimeout(total=timeout) if timeout else None - async with aiohttp.ClientSession(timeout=timeout_cfg) as session: - async with session.get(url, headers=headers) as resp: - resp.raise_for_status() # Raises HTTPError for bad responses (4XX or 5XX) - return BytesIO(await resp.read()) - - def text_filepath_to_base64_string(filepath: str) -> str: """Converts a text file to a base64 string.""" with open(filepath, "rb") as f: @@ -153,28 +71,3 @@ def text_filepath_to_data_uri(filepath: str) -> str: if mime_type is None: mime_type = "application/octet-stream" return f"data:{mime_type};base64,{base64_string}" - - -def resize_mask_to_image( - mask: torch.Tensor, - image: torch.Tensor, - upscale_method="nearest-exact", - crop="disabled", - allow_gradient=True, - add_channel_dim=False, -): - """ - Resize mask to be the same dimensions as an image, while maintaining proper format for API calls. - """ - _, H, W, _ = image.shape - mask = mask.unsqueeze(-1) - mask = mask.movedim(-1, 1) - mask = common_upscale( - mask, width=W, height=H, upscale_method=upscale_method, crop=crop - ) - mask = mask.movedim(1, -1) - if not add_channel_dim: - mask = mask.squeeze(-1) - if not allow_gradient: - mask = (mask > 0.5).float() - return mask diff --git a/comfy_api_nodes/nodes_bfl.py b/comfy_api_nodes/nodes_bfl.py index baa74fd52..1740fb377 100644 --- a/comfy_api_nodes/nodes_bfl.py +++ b/comfy_api_nodes/nodes_bfl.py @@ -5,10 +5,6 @@ import torch from typing_extensions import override from comfy_api.latest import IO, ComfyExtension -from comfy_api_nodes.apinode_utils import ( - resize_mask_to_image, - validate_aspect_ratio, -) from comfy_api_nodes.apis.bfl_api import ( BFLFluxExpandImageRequest, BFLFluxFillImageRequest, @@ -23,8 +19,10 @@ from comfy_api_nodes.util import ( ApiEndpoint, download_url_to_image_tensor, poll_op, + resize_mask_to_image, sync_op, tensor_to_base64_string, + validate_aspect_ratio_string, validate_string, ) @@ -43,11 +41,6 @@ class FluxProUltraImageNode(IO.ComfyNode): Generates images using Flux Pro 1.1 Ultra via api based on prompt and resolution. """ - MINIMUM_RATIO = 1 / 4 - MAXIMUM_RATIO = 4 / 1 - MINIMUM_RATIO_STR = "1:4" - MAXIMUM_RATIO_STR = "4:1" - @classmethod def define_schema(cls) -> IO.Schema: return IO.Schema( @@ -112,16 +105,7 @@ class FluxProUltraImageNode(IO.ComfyNode): @classmethod def validate_inputs(cls, aspect_ratio: str): - try: - validate_aspect_ratio( - aspect_ratio, - minimum_ratio=cls.MINIMUM_RATIO, - maximum_ratio=cls.MAXIMUM_RATIO, - minimum_ratio_str=cls.MINIMUM_RATIO_STR, - maximum_ratio_str=cls.MAXIMUM_RATIO_STR, - ) - except Exception as e: - return str(e) + validate_aspect_ratio_string(aspect_ratio, (1, 4), (4, 1)) return True @classmethod @@ -145,13 +129,7 @@ class FluxProUltraImageNode(IO.ComfyNode): prompt=prompt, prompt_upsampling=prompt_upsampling, seed=seed, - aspect_ratio=validate_aspect_ratio( - aspect_ratio, - minimum_ratio=cls.MINIMUM_RATIO, - maximum_ratio=cls.MAXIMUM_RATIO, - minimum_ratio_str=cls.MINIMUM_RATIO_STR, - maximum_ratio_str=cls.MAXIMUM_RATIO_STR, - ), + aspect_ratio=aspect_ratio, raw=raw, image_prompt=(image_prompt if image_prompt is None else tensor_to_base64_string(image_prompt)), image_prompt_strength=(None if image_prompt is None else round(image_prompt_strength, 2)), @@ -180,11 +158,6 @@ class FluxKontextProImageNode(IO.ComfyNode): Edits images using Flux.1 Kontext [pro] via api based on prompt and aspect ratio. """ - MINIMUM_RATIO = 1 / 4 - MAXIMUM_RATIO = 4 / 1 - MINIMUM_RATIO_STR = "1:4" - MAXIMUM_RATIO_STR = "4:1" - @classmethod def define_schema(cls) -> IO.Schema: return IO.Schema( @@ -261,13 +234,7 @@ class FluxKontextProImageNode(IO.ComfyNode): seed=0, prompt_upsampling=False, ) -> IO.NodeOutput: - aspect_ratio = validate_aspect_ratio( - aspect_ratio, - minimum_ratio=cls.MINIMUM_RATIO, - maximum_ratio=cls.MAXIMUM_RATIO, - minimum_ratio_str=cls.MINIMUM_RATIO_STR, - maximum_ratio_str=cls.MAXIMUM_RATIO_STR, - ) + validate_aspect_ratio_string(aspect_ratio, (1, 4), (4, 1)) if input_image is None: validate_string(prompt, strip_whitespace=False) initial_response = await sync_op( diff --git a/comfy_api_nodes/nodes_bytedance.py b/comfy_api_nodes/nodes_bytedance.py index 534af380d..caced471e 100644 --- a/comfy_api_nodes/nodes_bytedance.py +++ b/comfy_api_nodes/nodes_bytedance.py @@ -17,7 +17,7 @@ from comfy_api_nodes.util import ( poll_op, sync_op, upload_images_to_comfyapi, - validate_image_aspect_ratio_range, + validate_image_aspect_ratio, validate_image_dimensions, validate_string, ) @@ -403,7 +403,7 @@ class ByteDanceImageEditNode(IO.ComfyNode): validate_string(prompt, strip_whitespace=True, min_length=1) if get_number_of_images(image) != 1: raise ValueError("Exactly one input image is required.") - validate_image_aspect_ratio_range(image, (1, 3), (3, 1)) + validate_image_aspect_ratio(image, (1, 3), (3, 1)) source_url = (await upload_images_to_comfyapi(cls, image, max_images=1, mime_type="image/png"))[0] payload = Image2ImageTaskCreationRequest( model=model, @@ -565,7 +565,7 @@ class ByteDanceSeedreamNode(IO.ComfyNode): reference_images_urls = [] if n_input_images: for i in image: - validate_image_aspect_ratio_range(i, (1, 3), (3, 1)) + validate_image_aspect_ratio(i, (1, 3), (3, 1)) reference_images_urls = await upload_images_to_comfyapi( cls, image, @@ -798,7 +798,7 @@ class ByteDanceImageToVideoNode(IO.ComfyNode): validate_string(prompt, strip_whitespace=True, min_length=1) raise_if_text_params(prompt, ["resolution", "ratio", "duration", "seed", "camerafixed", "watermark"]) validate_image_dimensions(image, min_width=300, min_height=300, max_width=6000, max_height=6000) - validate_image_aspect_ratio_range(image, (2, 5), (5, 2), strict=False) # 0.4 to 2.5 + validate_image_aspect_ratio(image, (2, 5), (5, 2), strict=False) # 0.4 to 2.5 image_url = (await upload_images_to_comfyapi(cls, image, max_images=1))[0] prompt = ( @@ -923,7 +923,7 @@ class ByteDanceFirstLastFrameNode(IO.ComfyNode): raise_if_text_params(prompt, ["resolution", "ratio", "duration", "seed", "camerafixed", "watermark"]) for i in (first_frame, last_frame): validate_image_dimensions(i, min_width=300, min_height=300, max_width=6000, max_height=6000) - validate_image_aspect_ratio_range(i, (2, 5), (5, 2), strict=False) # 0.4 to 2.5 + validate_image_aspect_ratio(i, (2, 5), (5, 2), strict=False) # 0.4 to 2.5 download_urls = await upload_images_to_comfyapi( cls, @@ -1045,7 +1045,7 @@ class ByteDanceImageReferenceNode(IO.ComfyNode): raise_if_text_params(prompt, ["resolution", "ratio", "duration", "seed", "watermark"]) for image in images: validate_image_dimensions(image, min_width=300, min_height=300, max_width=6000, max_height=6000) - validate_image_aspect_ratio_range(image, (2, 5), (5, 2), strict=False) # 0.4 to 2.5 + validate_image_aspect_ratio(image, (2, 5), (5, 2), strict=False) # 0.4 to 2.5 image_urls = await upload_images_to_comfyapi(cls, images, max_images=4, mime_type="image/png") prompt = ( diff --git a/comfy_api_nodes/nodes_ideogram.py b/comfy_api_nodes/nodes_ideogram.py index d8fd3378b..48f94e612 100644 --- a/comfy_api_nodes/nodes_ideogram.py +++ b/comfy_api_nodes/nodes_ideogram.py @@ -1,6 +1,6 @@ from io import BytesIO from typing_extensions import override -from comfy_api.latest import ComfyExtension, IO +from comfy_api.latest import IO, ComfyExtension from PIL import Image import numpy as np import torch @@ -11,19 +11,13 @@ from comfy_api_nodes.apis import ( IdeogramV3Request, IdeogramV3EditRequest, ) - -from comfy_api_nodes.apis.client import ( +from comfy_api_nodes.util import ( ApiEndpoint, - HttpMethod, - SynchronousOperation, -) - -from comfy_api_nodes.apinode_utils import ( - download_url_to_bytesio, + bytesio_to_image_tensor, + download_url_as_bytesio, resize_mask_to_image, + sync_op, ) -from comfy_api_nodes.util import bytesio_to_image_tensor -from server import PromptServer V1_V1_RES_MAP = { "Auto":"AUTO", @@ -220,7 +214,7 @@ async def download_and_process_images(image_urls): for image_url in image_urls: # Using functions from apinode_utils.py to handle downloading and processing - image_bytesio = await download_url_to_bytesio(image_url) # Download image content to BytesIO + image_bytesio = await download_url_as_bytesio(image_url) # Download image content to BytesIO img_tensor = bytesio_to_image_tensor(image_bytesio, mode="RGB") # Convert to torch.Tensor with RGB mode image_tensors.append(img_tensor) @@ -233,19 +227,6 @@ async def download_and_process_images(image_urls): return stacked_tensors -def display_image_urls_on_node(image_urls, node_id): - if node_id and image_urls: - if len(image_urls) == 1: - PromptServer.instance.send_progress_text( - f"Generated Image URL:\n{image_urls[0]}", node_id - ) - else: - urls_text = "Generated Image URLs:\n" + "\n".join( - f"{i+1}. {url}" for i, url in enumerate(image_urls) - ) - PromptServer.instance.send_progress_text(urls_text, node_id) - - class IdeogramV1(IO.ComfyNode): @classmethod @@ -334,44 +315,30 @@ class IdeogramV1(IO.ComfyNode): aspect_ratio = V1_V2_RATIO_MAP.get(aspect_ratio, None) model = "V_1_TURBO" if turbo else "V_1" - auth = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } - operation = SynchronousOperation( - endpoint=ApiEndpoint( - path="/proxy/ideogram/generate", - method=HttpMethod.POST, - request_model=IdeogramGenerateRequest, - response_model=IdeogramGenerateResponse, - ), - request=IdeogramGenerateRequest( + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/ideogram/generate", method="POST"), + response_model=IdeogramGenerateResponse, + data=IdeogramGenerateRequest( image_request=ImageRequest( prompt=prompt, model=model, num_images=num_images, seed=seed, aspect_ratio=aspect_ratio if aspect_ratio != "ASPECT_1_1" else None, - magic_prompt_option=( - magic_prompt_option if magic_prompt_option != "AUTO" else None - ), + magic_prompt_option=(magic_prompt_option if magic_prompt_option != "AUTO" else None), negative_prompt=negative_prompt if negative_prompt else None, ) ), - auth_kwargs=auth, + max_retries=1, ) - response = await operation.execute() - if not response.data or len(response.data) == 0: raise Exception("No images were generated in the response") image_urls = [image_data.url for image_data in response.data if image_data.url] - if not image_urls: raise Exception("No image URLs were generated in the response") - - display_image_urls_on_node(image_urls, cls.hidden.unique_id) return IO.NodeOutput(await download_and_process_images(image_urls)) @@ -500,18 +467,11 @@ class IdeogramV2(IO.ComfyNode): else: final_aspect_ratio = aspect_ratio if aspect_ratio != "ASPECT_1_1" else None - auth = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } - operation = SynchronousOperation( - endpoint=ApiEndpoint( - path="/proxy/ideogram/generate", - method=HttpMethod.POST, - request_model=IdeogramGenerateRequest, - response_model=IdeogramGenerateResponse, - ), - request=IdeogramGenerateRequest( + response = await sync_op( + cls, + endpoint=ApiEndpoint(path="/proxy/ideogram/generate", method="POST"), + response_model=IdeogramGenerateResponse, + data=IdeogramGenerateRequest( image_request=ImageRequest( prompt=prompt, model=model, @@ -519,28 +479,20 @@ class IdeogramV2(IO.ComfyNode): seed=seed, aspect_ratio=final_aspect_ratio, resolution=final_resolution, - magic_prompt_option=( - magic_prompt_option if magic_prompt_option != "AUTO" else None - ), + magic_prompt_option=(magic_prompt_option if magic_prompt_option != "AUTO" else None), style_type=style_type if style_type != "NONE" else None, negative_prompt=negative_prompt if negative_prompt else None, color_palette=color_palette if color_palette else None, ) ), - auth_kwargs=auth, + max_retries=1, ) - - response = await operation.execute() - if not response.data or len(response.data) == 0: raise Exception("No images were generated in the response") image_urls = [image_data.url for image_data in response.data if image_data.url] - if not image_urls: raise Exception("No image URLs were generated in the response") - - display_image_urls_on_node(image_urls, cls.hidden.unique_id) return IO.NodeOutput(await download_and_process_images(image_urls)) @@ -656,10 +608,6 @@ class IdeogramV3(IO.ComfyNode): character_image=None, character_mask=None, ): - auth = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } if rendering_speed == "BALANCED": # for backward compatibility rendering_speed = "DEFAULT" @@ -694,9 +642,6 @@ class IdeogramV3(IO.ComfyNode): # Check if both image and mask are provided for editing mode if image is not None and mask is not None: - # Edit mode - path = "/proxy/ideogram/ideogram-v3/edit" - # Process image and mask input_tensor = image.squeeze().cpu() # Resize mask to match image dimension @@ -749,27 +694,20 @@ class IdeogramV3(IO.ComfyNode): if character_mask_binary: files["character_mask_binary"] = character_mask_binary - # Execute the operation for edit mode - operation = SynchronousOperation( - endpoint=ApiEndpoint( - path=path, - method=HttpMethod.POST, - request_model=IdeogramV3EditRequest, - response_model=IdeogramGenerateResponse, - ), - request=edit_request, + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/ideogram/ideogram-v3/edit", method="POST"), + response_model=IdeogramGenerateResponse, + data=edit_request, files=files, content_type="multipart/form-data", - auth_kwargs=auth, + max_retries=1, ) elif image is not None or mask is not None: # If only one of image or mask is provided, raise an error raise Exception("Ideogram V3 image editing requires both an image AND a mask") else: - # Generation mode - path = "/proxy/ideogram/ideogram-v3/generate" - # Create generation request gen_request = IdeogramV3Request( prompt=prompt, @@ -800,32 +738,22 @@ class IdeogramV3(IO.ComfyNode): if files: gen_request.style_type = "AUTO" - # Execute the operation for generation mode - operation = SynchronousOperation( - endpoint=ApiEndpoint( - path=path, - method=HttpMethod.POST, - request_model=IdeogramV3Request, - response_model=IdeogramGenerateResponse, - ), - request=gen_request, + response = await sync_op( + cls, + endpoint=ApiEndpoint(path="/proxy/ideogram/ideogram-v3/generate", method="POST"), + response_model=IdeogramGenerateResponse, + data=gen_request, files=files if files else None, content_type="multipart/form-data", - auth_kwargs=auth, + max_retries=1, ) - # Execute the operation and process response - response = await operation.execute() - if not response.data or len(response.data) == 0: raise Exception("No images were generated in the response") image_urls = [image_data.url for image_data in response.data if image_data.url] - if not image_urls: raise Exception("No image URLs were generated in the response") - - display_image_urls_on_node(image_urls, cls.hidden.unique_id) return IO.NodeOutput(await download_and_process_images(image_urls)) @@ -838,5 +766,6 @@ class IdeogramExtension(ComfyExtension): IdeogramV3, ] + async def comfy_entrypoint() -> IdeogramExtension: return IdeogramExtension() diff --git a/comfy_api_nodes/nodes_kling.py b/comfy_api_nodes/nodes_kling.py index eea65c9ac..7b23e9cf9 100644 --- a/comfy_api_nodes/nodes_kling.py +++ b/comfy_api_nodes/nodes_kling.py @@ -282,7 +282,7 @@ def validate_input_image(image: torch.Tensor) -> None: See: https://app.klingai.com/global/dev/document-api/apiReference/model/imageToVideo """ validate_image_dimensions(image, min_width=300, min_height=300) - validate_image_aspect_ratio(image, min_aspect_ratio=1 / 2.5, max_aspect_ratio=2.5) + validate_image_aspect_ratio(image, (1, 2.5), (2.5, 1)) def get_video_from_response(response) -> KlingVideoResult: diff --git a/comfy_api_nodes/nodes_pixverse.py b/comfy_api_nodes/nodes_pixverse.py index b2b841be8..6e1686af0 100644 --- a/comfy_api_nodes/nodes_pixverse.py +++ b/comfy_api_nodes/nodes_pixverse.py @@ -1,7 +1,6 @@ -from inspect import cleandoc -from typing import Optional +import torch from typing_extensions import override -from io import BytesIO +from comfy_api.latest import IO, ComfyExtension from comfy_api_nodes.apis.pixverse_api import ( PixverseTextVideoRequest, PixverseImageVideoRequest, @@ -17,53 +16,30 @@ from comfy_api_nodes.apis.pixverse_api import ( PixverseIO, pixverse_templates, ) -from comfy_api_nodes.apis.client import ( +from comfy_api_nodes.util import ( ApiEndpoint, - HttpMethod, - SynchronousOperation, - PollingOperation, - EmptyRequest, + download_url_to_video_output, + poll_op, + sync_op, + tensor_to_bytesio, + validate_string, ) -from comfy_api_nodes.util import validate_string, tensor_to_bytesio -from comfy_api.input_impl import VideoFromFile -from comfy_api.latest import ComfyExtension, IO - -import torch -import aiohttp - AVERAGE_DURATION_T2V = 32 AVERAGE_DURATION_I2V = 30 AVERAGE_DURATION_T2T = 52 -def get_video_url_from_response( - response: PixverseGenerationStatusResponse, -) -> Optional[str]: - if response.Resp is None or response.Resp.url is None: - return None - return str(response.Resp.url) - - -async def upload_image_to_pixverse(image: torch.Tensor, auth_kwargs=None): - # first, upload image to Pixverse and get image id to use in actual generation call - operation = SynchronousOperation( - endpoint=ApiEndpoint( - path="/proxy/pixverse/image/upload", - method=HttpMethod.POST, - request_model=EmptyRequest, - response_model=PixverseImageUploadResponse, - ), - request=EmptyRequest(), +async def upload_image_to_pixverse(cls: type[IO.ComfyNode], image: torch.Tensor): + response_upload = await sync_op( + cls, + ApiEndpoint(path="/proxy/pixverse/image/upload", method="POST"), + response_model=PixverseImageUploadResponse, files={"image": tensor_to_bytesio(image)}, content_type="multipart/form-data", - auth_kwargs=auth_kwargs, ) - response_upload: PixverseImageUploadResponse = await operation.execute() - if response_upload.Resp is None: raise Exception(f"PixVerse image upload request failed: '{response_upload.ErrMsg}'") - return response_upload.Resp.img_id @@ -93,17 +69,13 @@ class PixverseTemplateNode(IO.ComfyNode): class PixverseTextToVideoNode(IO.ComfyNode): - """ - Generates videos based on prompt and output_size. - """ - @classmethod def define_schema(cls) -> IO.Schema: return IO.Schema( node_id="PixverseTextToVideoNode", display_name="PixVerse Text to Video", category="api node/video/PixVerse", - description=cleandoc(cls.__doc__ or ""), + description="Generates videos based on prompt and output_size.", inputs=[ IO.String.Input( "prompt", @@ -170,7 +142,7 @@ class PixverseTextToVideoNode(IO.ComfyNode): negative_prompt: str = None, pixverse_template: int = None, ) -> IO.NodeOutput: - validate_string(prompt, strip_whitespace=False) + validate_string(prompt, strip_whitespace=False, min_length=1) # 1080p is limited to 5 seconds duration # only normal motion_mode supported for 1080p or for non-5 second duration if quality == PixverseQuality.res_1080p: @@ -179,18 +151,11 @@ class PixverseTextToVideoNode(IO.ComfyNode): elif duration_seconds != PixverseDuration.dur_5: motion_mode = PixverseMotionMode.normal - auth = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } - operation = SynchronousOperation( - endpoint=ApiEndpoint( - path="/proxy/pixverse/video/text/generate", - method=HttpMethod.POST, - request_model=PixverseTextVideoRequest, - response_model=PixverseVideoResponse, - ), - request=PixverseTextVideoRequest( + response_api = await sync_op( + cls, + ApiEndpoint(path="/proxy/pixverse/video/text/generate", method="POST"), + response_model=PixverseVideoResponse, + data=PixverseTextVideoRequest( prompt=prompt, aspect_ratio=aspect_ratio, quality=quality, @@ -200,20 +165,14 @@ class PixverseTextToVideoNode(IO.ComfyNode): template_id=pixverse_template, seed=seed, ), - auth_kwargs=auth, ) - response_api = await operation.execute() - if response_api.Resp is None: raise Exception(f"PixVerse request failed: '{response_api.ErrMsg}'") - operation = PollingOperation( - poll_endpoint=ApiEndpoint( - path=f"/proxy/pixverse/video/result/{response_api.Resp.video_id}", - method=HttpMethod.GET, - request_model=EmptyRequest, - response_model=PixverseGenerationStatusResponse, - ), + response_poll = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/pixverse/video/result/{response_api.Resp.video_id}"), + response_model=PixverseGenerationStatusResponse, completed_statuses=[PixverseStatus.successful], failed_statuses=[ PixverseStatus.contents_moderation, @@ -221,30 +180,19 @@ class PixverseTextToVideoNode(IO.ComfyNode): PixverseStatus.deleted, ], status_extractor=lambda x: x.Resp.status, - auth_kwargs=auth, - node_id=cls.hidden.unique_id, - result_url_extractor=get_video_url_from_response, estimated_duration=AVERAGE_DURATION_T2V, ) - response_poll = await operation.execute() - - async with aiohttp.ClientSession() as session: - async with session.get(response_poll.Resp.url) as vid_response: - return IO.NodeOutput(VideoFromFile(BytesIO(await vid_response.content.read()))) + return IO.NodeOutput(await download_url_to_video_output(response_poll.Resp.url)) class PixverseImageToVideoNode(IO.ComfyNode): - """ - Generates videos based on prompt and output_size. - """ - @classmethod def define_schema(cls) -> IO.Schema: return IO.Schema( node_id="PixverseImageToVideoNode", display_name="PixVerse Image to Video", category="api node/video/PixVerse", - description=cleandoc(cls.__doc__ or ""), + description="Generates videos based on prompt and output_size.", inputs=[ IO.Image.Input("image"), IO.String.Input( @@ -309,11 +257,7 @@ class PixverseImageToVideoNode(IO.ComfyNode): pixverse_template: int = None, ) -> IO.NodeOutput: validate_string(prompt, strip_whitespace=False) - auth = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } - img_id = await upload_image_to_pixverse(image, auth_kwargs=auth) + img_id = await upload_image_to_pixverse(cls, image) # 1080p is limited to 5 seconds duration # only normal motion_mode supported for 1080p or for non-5 second duration @@ -323,14 +267,11 @@ class PixverseImageToVideoNode(IO.ComfyNode): elif duration_seconds != PixverseDuration.dur_5: motion_mode = PixverseMotionMode.normal - operation = SynchronousOperation( - endpoint=ApiEndpoint( - path="/proxy/pixverse/video/img/generate", - method=HttpMethod.POST, - request_model=PixverseImageVideoRequest, - response_model=PixverseVideoResponse, - ), - request=PixverseImageVideoRequest( + response_api = await sync_op( + cls, + ApiEndpoint(path="/proxy/pixverse/video/img/generate", method="POST"), + response_model=PixverseVideoResponse, + data=PixverseImageVideoRequest( img_id=img_id, prompt=prompt, quality=quality, @@ -340,20 +281,15 @@ class PixverseImageToVideoNode(IO.ComfyNode): template_id=pixverse_template, seed=seed, ), - auth_kwargs=auth, ) - response_api = await operation.execute() if response_api.Resp is None: raise Exception(f"PixVerse request failed: '{response_api.ErrMsg}'") - operation = PollingOperation( - poll_endpoint=ApiEndpoint( - path=f"/proxy/pixverse/video/result/{response_api.Resp.video_id}", - method=HttpMethod.GET, - request_model=EmptyRequest, - response_model=PixverseGenerationStatusResponse, - ), + response_poll = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/pixverse/video/result/{response_api.Resp.video_id}"), + response_model=PixverseGenerationStatusResponse, completed_statuses=[PixverseStatus.successful], failed_statuses=[ PixverseStatus.contents_moderation, @@ -361,30 +297,19 @@ class PixverseImageToVideoNode(IO.ComfyNode): PixverseStatus.deleted, ], status_extractor=lambda x: x.Resp.status, - auth_kwargs=auth, - node_id=cls.hidden.unique_id, - result_url_extractor=get_video_url_from_response, estimated_duration=AVERAGE_DURATION_I2V, ) - response_poll = await operation.execute() - - async with aiohttp.ClientSession() as session: - async with session.get(response_poll.Resp.url) as vid_response: - return IO.NodeOutput(VideoFromFile(BytesIO(await vid_response.content.read()))) + return IO.NodeOutput(await download_url_to_video_output(response_poll.Resp.url)) class PixverseTransitionVideoNode(IO.ComfyNode): - """ - Generates videos based on prompt and output_size. - """ - @classmethod def define_schema(cls) -> IO.Schema: return IO.Schema( node_id="PixverseTransitionVideoNode", display_name="PixVerse Transition Video", category="api node/video/PixVerse", - description=cleandoc(cls.__doc__ or ""), + description="Generates videos based on prompt and output_size.", inputs=[ IO.Image.Input("first_frame"), IO.Image.Input("last_frame"), @@ -445,12 +370,8 @@ class PixverseTransitionVideoNode(IO.ComfyNode): negative_prompt: str = None, ) -> IO.NodeOutput: validate_string(prompt, strip_whitespace=False) - auth = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } - first_frame_id = await upload_image_to_pixverse(first_frame, auth_kwargs=auth) - last_frame_id = await upload_image_to_pixverse(last_frame, auth_kwargs=auth) + first_frame_id = await upload_image_to_pixverse(cls, first_frame) + last_frame_id = await upload_image_to_pixverse(cls, last_frame) # 1080p is limited to 5 seconds duration # only normal motion_mode supported for 1080p or for non-5 second duration @@ -460,14 +381,11 @@ class PixverseTransitionVideoNode(IO.ComfyNode): elif duration_seconds != PixverseDuration.dur_5: motion_mode = PixverseMotionMode.normal - operation = SynchronousOperation( - endpoint=ApiEndpoint( - path="/proxy/pixverse/video/transition/generate", - method=HttpMethod.POST, - request_model=PixverseTransitionVideoRequest, - response_model=PixverseVideoResponse, - ), - request=PixverseTransitionVideoRequest( + response_api = await sync_op( + cls, + ApiEndpoint(path="/proxy/pixverse/video/transition/generate", method="POST"), + response_model=PixverseVideoResponse, + data=PixverseTransitionVideoRequest( first_frame_img=first_frame_id, last_frame_img=last_frame_id, prompt=prompt, @@ -477,20 +395,15 @@ class PixverseTransitionVideoNode(IO.ComfyNode): negative_prompt=negative_prompt if negative_prompt else None, seed=seed, ), - auth_kwargs=auth, ) - response_api = await operation.execute() if response_api.Resp is None: raise Exception(f"PixVerse request failed: '{response_api.ErrMsg}'") - operation = PollingOperation( - poll_endpoint=ApiEndpoint( - path=f"/proxy/pixverse/video/result/{response_api.Resp.video_id}", - method=HttpMethod.GET, - request_model=EmptyRequest, - response_model=PixverseGenerationStatusResponse, - ), + response_poll = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/pixverse/video/result/{response_api.Resp.video_id}"), + response_model=PixverseGenerationStatusResponse, completed_statuses=[PixverseStatus.successful], failed_statuses=[ PixverseStatus.contents_moderation, @@ -498,16 +411,9 @@ class PixverseTransitionVideoNode(IO.ComfyNode): PixverseStatus.deleted, ], status_extractor=lambda x: x.Resp.status, - auth_kwargs=auth, - node_id=cls.hidden.unique_id, - result_url_extractor=get_video_url_from_response, estimated_duration=AVERAGE_DURATION_T2V, ) - response_poll = await operation.execute() - - async with aiohttp.ClientSession() as session: - async with session.get(response_poll.Resp.url) as vid_response: - return IO.NodeOutput(VideoFromFile(BytesIO(await vid_response.content.read()))) + return IO.NodeOutput(await download_url_to_video_output(response_poll.Resp.url)) class PixVerseExtension(ComfyExtension): diff --git a/comfy_api_nodes/nodes_recraft.py b/comfy_api_nodes/nodes_recraft.py index dee186cd6..e3440b946 100644 --- a/comfy_api_nodes/nodes_recraft.py +++ b/comfy_api_nodes/nodes_recraft.py @@ -8,9 +8,6 @@ from typing_extensions import override from comfy.utils import ProgressBar from comfy_api.latest import IO, ComfyExtension -from comfy_api_nodes.apinode_utils import ( - resize_mask_to_image, -) from comfy_api_nodes.apis.recraft_api import ( RecraftColor, RecraftColorChain, @@ -28,6 +25,7 @@ from comfy_api_nodes.util import ( ApiEndpoint, bytesio_to_image_tensor, download_url_as_bytesio, + resize_mask_to_image, sync_op, tensor_to_bytesio, validate_string, diff --git a/comfy_api_nodes/nodes_runway.py b/comfy_api_nodes/nodes_runway.py index 0543d1d0e..2fdafbbfe 100644 --- a/comfy_api_nodes/nodes_runway.py +++ b/comfy_api_nodes/nodes_runway.py @@ -200,7 +200,7 @@ class RunwayImageToVideoNodeGen3a(IO.ComfyNode): ) -> IO.NodeOutput: validate_string(prompt, min_length=1) validate_image_dimensions(start_frame, max_width=7999, max_height=7999) - validate_image_aspect_ratio(start_frame, min_aspect_ratio=0.5, max_aspect_ratio=2.0) + validate_image_aspect_ratio(start_frame, (1, 2), (2, 1)) download_urls = await upload_images_to_comfyapi( cls, @@ -290,7 +290,7 @@ class RunwayImageToVideoNodeGen4(IO.ComfyNode): ) -> IO.NodeOutput: validate_string(prompt, min_length=1) validate_image_dimensions(start_frame, max_width=7999, max_height=7999) - validate_image_aspect_ratio(start_frame, min_aspect_ratio=0.5, max_aspect_ratio=2.0) + validate_image_aspect_ratio(start_frame, (1, 2), (2, 1)) download_urls = await upload_images_to_comfyapi( cls, @@ -390,8 +390,8 @@ class RunwayFirstLastFrameNode(IO.ComfyNode): validate_string(prompt, min_length=1) validate_image_dimensions(start_frame, max_width=7999, max_height=7999) validate_image_dimensions(end_frame, max_width=7999, max_height=7999) - validate_image_aspect_ratio(start_frame, min_aspect_ratio=0.5, max_aspect_ratio=2.0) - validate_image_aspect_ratio(end_frame, min_aspect_ratio=0.5, max_aspect_ratio=2.0) + validate_image_aspect_ratio(start_frame, (1, 2), (2, 1)) + validate_image_aspect_ratio(end_frame, (1, 2), (2, 1)) stacked_input_images = image_tensor_pair_to_batch(start_frame, end_frame) download_urls = await upload_images_to_comfyapi( @@ -475,7 +475,7 @@ class RunwayTextToImageNode(IO.ComfyNode): reference_images = None if reference_image is not None: validate_image_dimensions(reference_image, max_width=7999, max_height=7999) - validate_image_aspect_ratio(reference_image, min_aspect_ratio=0.5, max_aspect_ratio=2.0) + validate_image_aspect_ratio(reference_image, (1, 2), (2, 1)) download_urls = await upload_images_to_comfyapi( cls, reference_image, diff --git a/comfy_api_nodes/nodes_vidu.py b/comfy_api_nodes/nodes_vidu.py index 0e0572f8c..7a679f0d9 100644 --- a/comfy_api_nodes/nodes_vidu.py +++ b/comfy_api_nodes/nodes_vidu.py @@ -14,9 +14,9 @@ from comfy_api_nodes.util import ( poll_op, sync_op, upload_images_to_comfyapi, - validate_aspect_ratio_closeness, - validate_image_aspect_ratio_range, + validate_image_aspect_ratio, validate_image_dimensions, + validate_images_aspect_ratio_closeness, ) VIDU_TEXT_TO_VIDEO = "/proxy/vidu/text2video" @@ -114,7 +114,7 @@ async def execute_task( cls, ApiEndpoint(path=VIDU_GET_GENERATION_STATUS % response.task_id), response_model=TaskStatusResponse, - status_extractor=lambda r: r.state.value, + status_extractor=lambda r: r.state, estimated_duration=estimated_duration, ) @@ -307,7 +307,7 @@ class ViduImageToVideoNode(IO.ComfyNode): ) -> IO.NodeOutput: if get_number_of_images(image) > 1: raise ValueError("Only one input image is allowed.") - validate_image_aspect_ratio_range(image, (1, 4), (4, 1)) + validate_image_aspect_ratio(image, (1, 4), (4, 1)) payload = TaskCreationRequest( model_name=model, prompt=prompt, @@ -423,7 +423,7 @@ class ViduReferenceVideoNode(IO.ComfyNode): if a > 7: raise ValueError("Too many images, maximum allowed is 7.") for image in images: - validate_image_aspect_ratio_range(image, (1, 4), (4, 1)) + validate_image_aspect_ratio(image, (1, 4), (4, 1)) validate_image_dimensions(image, min_width=128, min_height=128) payload = TaskCreationRequest( model_name=model, @@ -533,7 +533,7 @@ class ViduStartEndToVideoNode(IO.ComfyNode): resolution: str, movement_amplitude: str, ) -> IO.NodeOutput: - validate_aspect_ratio_closeness(first_frame, end_frame, min_rel=0.8, max_rel=1.25, strict=False) + validate_images_aspect_ratio_closeness(first_frame, end_frame, min_rel=0.8, max_rel=1.25, strict=False) payload = TaskCreationRequest( model_name=model, prompt=prompt, diff --git a/comfy_api_nodes/util/__init__.py b/comfy_api_nodes/util/__init__.py index 0cca2b59b..bbc71363a 100644 --- a/comfy_api_nodes/util/__init__.py +++ b/comfy_api_nodes/util/__init__.py @@ -14,6 +14,7 @@ from .conversions import ( downscale_image_tensor, image_tensor_pair_to_batch, pil_to_bytesio, + resize_mask_to_image, tensor_to_base64_string, tensor_to_bytesio, tensor_to_pil, @@ -34,12 +35,12 @@ from .upload_helpers import ( ) from .validation_utils import ( get_number_of_images, - validate_aspect_ratio_closeness, + validate_aspect_ratio_string, validate_audio_duration, validate_container_format_is_mp4, validate_image_aspect_ratio, - validate_image_aspect_ratio_range, validate_image_dimensions, + validate_images_aspect_ratio_closeness, validate_string, validate_video_dimensions, validate_video_duration, @@ -70,6 +71,7 @@ __all__ = [ "downscale_image_tensor", "image_tensor_pair_to_batch", "pil_to_bytesio", + "resize_mask_to_image", "tensor_to_base64_string", "tensor_to_bytesio", "tensor_to_pil", @@ -77,12 +79,12 @@ __all__ = [ "video_to_base64_string", # Validation utilities "get_number_of_images", - "validate_aspect_ratio_closeness", + "validate_aspect_ratio_string", "validate_audio_duration", "validate_container_format_is_mp4", "validate_image_aspect_ratio", - "validate_image_aspect_ratio_range", "validate_image_dimensions", + "validate_images_aspect_ratio_closeness", "validate_string", "validate_video_dimensions", "validate_video_duration", diff --git a/comfy_api_nodes/util/conversions.py b/comfy_api_nodes/util/conversions.py index 9f4c90c5c..b59c2bd84 100644 --- a/comfy_api_nodes/util/conversions.py +++ b/comfy_api_nodes/util/conversions.py @@ -430,3 +430,24 @@ def audio_bytes_to_audio_input(audio_bytes: bytes) -> dict: wav = torch.cat(frames, dim=1) # [C, T] wav = _f32_pcm(wav) return {"waveform": wav.unsqueeze(0).contiguous(), "sample_rate": out_sr} + + +def resize_mask_to_image( + mask: torch.Tensor, + image: torch.Tensor, + upscale_method="nearest-exact", + crop="disabled", + allow_gradient=True, + add_channel_dim=False, +): + """Resize mask to be the same dimensions as an image, while maintaining proper format for API calls.""" + _, height, width, _ = image.shape + mask = mask.unsqueeze(-1) + mask = mask.movedim(-1, 1) + mask = common_upscale(mask, width=width, height=height, upscale_method=upscale_method, crop=crop) + mask = mask.movedim(1, -1) + if not add_channel_dim: + mask = mask.squeeze(-1) + if not allow_gradient: + mask = (mask > 0.5).float() + return mask diff --git a/comfy_api_nodes/util/validation_utils.py b/comfy_api_nodes/util/validation_utils.py index 22da05bc1..ec7006aed 100644 --- a/comfy_api_nodes/util/validation_utils.py +++ b/comfy_api_nodes/util/validation_utils.py @@ -37,63 +37,62 @@ def validate_image_dimensions( def validate_image_aspect_ratio( image: torch.Tensor, - min_aspect_ratio: Optional[float] = None, - max_aspect_ratio: Optional[float] = None, -): - width, height = get_image_dimensions(image) - aspect_ratio = width / height - - if min_aspect_ratio is not None and aspect_ratio < min_aspect_ratio: - raise ValueError(f"Image aspect ratio must be at least {min_aspect_ratio}, got {aspect_ratio}") - if max_aspect_ratio is not None and aspect_ratio > max_aspect_ratio: - raise ValueError(f"Image aspect ratio must be at most {max_aspect_ratio}, got {aspect_ratio}") - - -def validate_image_aspect_ratio_range( - image: torch.Tensor, - min_ratio: tuple[float, float], # e.g. (1, 4) - max_ratio: tuple[float, float], # e.g. (4, 1) + min_ratio: Optional[tuple[float, float]] = None, # e.g. (1, 4) + max_ratio: Optional[tuple[float, float]] = None, # e.g. (4, 1) *, strict: bool = True, # True -> (min, max); False -> [min, max] ) -> float: - a1, b1 = min_ratio - a2, b2 = max_ratio - if a1 <= 0 or b1 <= 0 or a2 <= 0 or b2 <= 0: - raise ValueError("Ratios must be positive, like (1, 4) or (4, 1).") - lo, hi = (a1 / b1), (a2 / b2) - if lo > hi: - lo, hi = hi, lo - a1, b1, a2, b2 = a2, b2, a1, b1 # swap only for error text + """Validates that image aspect ratio is within min and max. If a bound is None, that side is not checked.""" w, h = get_image_dimensions(image) if w <= 0 or h <= 0: raise ValueError(f"Invalid image dimensions: {w}x{h}") ar = w / h - ok = (lo < ar < hi) if strict else (lo <= ar <= hi) - if not ok: - op = "<" if strict else "≤" - raise ValueError(f"Image aspect ratio {ar:.6g} is outside allowed range: {a1}:{b1} {op} ratio {op} {a2}:{b2}") + _assert_ratio_bounds(ar, min_ratio=min_ratio, max_ratio=max_ratio, strict=strict) return ar -def validate_aspect_ratio_closeness( - start_img, - end_img, - min_rel: float, - max_rel: float, +def validate_images_aspect_ratio_closeness( + first_image: torch.Tensor, + second_image: torch.Tensor, + min_rel: float, # e.g. 0.8 + max_rel: float, # e.g. 1.25 *, - strict: bool = False, # True => exclusive, False => inclusive -) -> None: - w1, h1 = get_image_dimensions(start_img) - w2, h2 = get_image_dimensions(end_img) + strict: bool = False, # True -> (min, max); False -> [min, max] +) -> float: + """ + Validates that the two images' aspect ratios are 'close'. + The closeness factor is C = max(ar1, ar2) / min(ar1, ar2) (C >= 1). + We require C <= limit, where limit = max(max_rel, 1.0 / min_rel). + + Returns the computed closeness factor C. + """ + w1, h1 = get_image_dimensions(first_image) + w2, h2 = get_image_dimensions(second_image) if min(w1, h1, w2, h2) <= 0: raise ValueError("Invalid image dimensions") ar1 = w1 / h1 ar2 = w2 / h2 - # Normalize so it is symmetric (no need to check both ar1/ar2 and ar2/ar1) closeness = max(ar1, ar2) / min(ar1, ar2) - limit = max(max_rel, 1.0 / min_rel) # for 0.8..1.25 this is 1.25 + limit = max(max_rel, 1.0 / min_rel) if (closeness >= limit) if strict else (closeness > limit): - raise ValueError(f"Aspect ratios must be close: start/end={ar1/ar2:.4f}, allowed range {min_rel}–{max_rel}.") + raise ValueError( + f"Aspect ratios must be close: ar1/ar2={ar1/ar2:.2g}, " + f"allowed range {min_rel}–{max_rel} (limit {limit:.2g})." + ) + return closeness + + +def validate_aspect_ratio_string( + aspect_ratio: str, + min_ratio: Optional[tuple[float, float]] = None, # e.g. (1, 4) + max_ratio: Optional[tuple[float, float]] = None, # e.g. (4, 1) + *, + strict: bool = False, # True -> (min, max); False -> [min, max] +) -> float: + """Parses 'X:Y' and validates it against optional bounds. Returns the numeric ratio.""" + ar = _parse_aspect_ratio_string(aspect_ratio) + _assert_ratio_bounds(ar, min_ratio=min_ratio, max_ratio=max_ratio, strict=strict) + return ar def validate_video_dimensions( @@ -183,3 +182,49 @@ def validate_container_format_is_mp4(video: VideoInput) -> None: container_format = video.get_container_format() if container_format not in ["mp4", "mov,mp4,m4a,3gp,3g2,mj2"]: raise ValueError(f"Only MP4 container format supported. Got: {container_format}") + + +def _ratio_from_tuple(r: tuple[float, float]) -> float: + a, b = r + if a <= 0 or b <= 0: + raise ValueError(f"Ratios must be positive, got {a}:{b}.") + return a / b + + +def _assert_ratio_bounds( + ar: float, + *, + min_ratio: Optional[tuple[float, float]] = None, + max_ratio: Optional[tuple[float, float]] = None, + strict: bool = True, +) -> None: + """Validate a numeric aspect ratio against optional min/max ratio bounds.""" + lo = _ratio_from_tuple(min_ratio) if min_ratio is not None else None + hi = _ratio_from_tuple(max_ratio) if max_ratio is not None else None + + if lo is not None and hi is not None and lo > hi: + lo, hi = hi, lo # normalize order if caller swapped them + + if lo is not None: + if (ar <= lo) if strict else (ar < lo): + op = "<" if strict else "≤" + raise ValueError(f"Aspect ratio `{ar:.2g}` must be {op} {lo:.2g}.") + if hi is not None: + if (ar >= hi) if strict else (ar > hi): + op = "<" if strict else "≤" + raise ValueError(f"Aspect ratio `{ar:.2g}` must be {op} {hi:.2g}.") + + +def _parse_aspect_ratio_string(ar_str: str) -> float: + """Parse 'X:Y' with integer parts into a positive float ratio X/Y.""" + parts = ar_str.split(":") + if len(parts) != 2: + raise ValueError(f"Aspect ratio must be 'X:Y' (e.g., 16:9), got '{ar_str}'.") + try: + a = int(parts[0].strip()) + b = int(parts[1].strip()) + except ValueError as exc: + raise ValueError(f"Aspect ratio must contain integers separated by ':', got '{ar_str}'.") from exc + if a <= 0 or b <= 0: + raise ValueError(f"Aspect ratio parts must be positive integers, got {a}:{b}.") + return a / b From dfac94695be95076d8028d04005a744f3ec0de8d Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Thu, 30 Oct 2025 19:22:35 +0200 Subject: [PATCH 057/647] fix img2img operation in Dall2 node (#10552) --- comfy_api_nodes/nodes_openai.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy_api_nodes/nodes_openai.py b/comfy_api_nodes/nodes_openai.py index c467e840c..b4568fc85 100644 --- a/comfy_api_nodes/nodes_openai.py +++ b/comfy_api_nodes/nodes_openai.py @@ -225,7 +225,7 @@ class OpenAIDalle2(ComfyNodeABC): ), files=( { - "image": img_binary, + "image": ("image.png", img_binary, "image/png"), } if img_binary else None From 513b0c46fba3bf40191d684ff81207ad935f1717 Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Fri, 31 Oct 2025 07:39:02 +1000 Subject: [PATCH 058/647] Add RAM Pressure cache mode (#10454) * execution: Roll the UI cache into the outputs Currently the UI cache is parallel to the output cache with expectations of being a content superset of the output cache. At the same time the UI and output cache are maintained completely seperately, making it awkward to free the output cache content without changing the behaviour of the UI cache. There are two actual users (getters) of the UI cache. The first is the case of a direct content hit on the output cache when executing a node. This case is very naturally handled by merging the UI and outputs cache. The second case is the history JSON generation at the end of the prompt. This currently works by asking the cache for all_node_ids and then pulling the cache contents for those nodes. all_node_ids is the nodes of the dynamic prompt. So fold the UI cache into the output cache. The current UI cache setter now writes to a prompt-scope dict. When the output cache is set, just get this value from the dict and tuple up with the outputs. When generating the history, simply iterate prompt-scope dict. This prepares support for more complex caching strategies (like RAM pressure caching) where less than 1 workflow will be cached and it will be desirable to keep the UI cache and output cache in sync. * sd: Implement RAM getter for VAE * model_patcher: Implement RAM getter for ModelPatcher * sd: Implement RAM getter for CLIP * Implement RAM Pressure cache Implement a cache sensitive to RAM pressure. When RAM headroom drops down below a certain threshold, evict RAM-expensive nodes from the cache. Models and tensors are measured directly for RAM usage. An OOM score is then computed based on the RAM usage of the node. Note the due to indirection through shared objects (like a model patcher), multiple nodes can account the same RAM as their individual usage. The intent is this will free chains of nodes particularly model loaders and associate loras as they all score similar and are sorted in close to each other. Has a bias towards unloading model nodes mid flow while being able to keep results like text encodings and VAE. * execution: Convert the cache entry to NamedTuple As commented in review. Convert this to a named tuple and abstract away the tuple type completely from graph.py. --- comfy/cli_args.py | 1 + comfy/model_patcher.py | 3 ++ comfy/sd.py | 14 +++++++ comfy_execution/caching.py | 83 ++++++++++++++++++++++++++++++++++++++ comfy_execution/graph.py | 9 ++++- execution.py | 81 +++++++++++++++++++++---------------- main.py | 4 +- 7 files changed, 157 insertions(+), 38 deletions(-) diff --git a/comfy/cli_args.py b/comfy/cli_args.py index 001abd843..3d5bc7c90 100644 --- a/comfy/cli_args.py +++ b/comfy/cli_args.py @@ -105,6 +105,7 @@ cache_group = parser.add_mutually_exclusive_group() cache_group.add_argument("--cache-classic", action="store_true", help="Use the old style (aggressive) caching.") cache_group.add_argument("--cache-lru", type=int, default=0, help="Use LRU caching with a maximum of N node results cached. May use more RAM/VRAM.") cache_group.add_argument("--cache-none", action="store_true", help="Reduced RAM/VRAM usage at the expense of executing every node for each run.") +cache_group.add_argument("--cache-ram", nargs='?', const=4.0, type=float, default=0, help="Use RAM pressure caching with the specified headroom threshold. If available RAM drops below the threhold the cache remove large items to free RAM. Default 4GB") attn_group = parser.add_mutually_exclusive_group() attn_group.add_argument("--use-split-cross-attention", action="store_true", help="Use the split cross attention optimization. Ignored when xformers is used.") diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index ed3f3f5cb..674a214ca 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -276,6 +276,9 @@ class ModelPatcher: self.size = comfy.model_management.module_size(self.model) return self.size + def get_ram_usage(self): + return self.model_size() + def loaded_size(self): return self.model.model_loaded_weight_memory diff --git a/comfy/sd.py b/comfy/sd.py index de4eee96e..9e5ebbf15 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -143,6 +143,9 @@ class CLIP: n.apply_hooks_to_conds = self.apply_hooks_to_conds return n + def get_ram_usage(self): + return self.patcher.get_ram_usage() + def add_patches(self, patches, strength_patch=1.0, strength_model=1.0): return self.patcher.add_patches(patches, strength_patch, strength_model) @@ -293,6 +296,7 @@ class VAE: self.working_dtypes = [torch.bfloat16, torch.float32] self.disable_offload = False self.not_video = False + self.size = None self.downscale_index_formula = None self.upscale_index_formula = None @@ -595,6 +599,16 @@ class VAE: self.patcher = comfy.model_patcher.ModelPatcher(self.first_stage_model, load_device=self.device, offload_device=offload_device) logging.info("VAE load device: {}, offload device: {}, dtype: {}".format(self.device, offload_device, self.vae_dtype)) + self.model_size() + + def model_size(self): + if self.size is not None: + return self.size + self.size = comfy.model_management.module_size(self.first_stage_model) + return self.size + + def get_ram_usage(self): + return self.model_size() def throw_exception_if_invalid(self): if self.first_stage_model is None: diff --git a/comfy_execution/caching.py b/comfy_execution/caching.py index 566bc3f9c..b498f43e7 100644 --- a/comfy_execution/caching.py +++ b/comfy_execution/caching.py @@ -1,4 +1,9 @@ +import bisect +import gc import itertools +import psutil +import time +import torch from typing import Sequence, Mapping, Dict from comfy_execution.graph import DynamicPrompt from abc import ABC, abstractmethod @@ -188,6 +193,9 @@ class BasicCache: self._clean_cache() self._clean_subcaches() + def poll(self, **kwargs): + pass + def _set_immediate(self, node_id, value): assert self.initialized cache_key = self.cache_key_set.get_data_key(node_id) @@ -276,6 +284,9 @@ class NullCache: def clean_unused(self): pass + def poll(self, **kwargs): + pass + def get(self, node_id): return None @@ -336,3 +347,75 @@ class LRUCache(BasicCache): self._mark_used(child_id) self.children[cache_key].append(self.cache_key_set.get_data_key(child_id)) return self + + +#Iterating the cache for usage analysis might be expensive, so if we trigger make sure +#to take a chunk out to give breathing space on high-node / low-ram-per-node flows. + +RAM_CACHE_HYSTERESIS = 1.1 + +#This is kinda in GB but not really. It needs to be non-zero for the below heuristic +#and as long as Multi GB models dwarf this it will approximate OOM scoring OK + +RAM_CACHE_DEFAULT_RAM_USAGE = 0.1 + +#Exponential bias towards evicting older workflows so garbage will be taken out +#in constantly changing setups. + +RAM_CACHE_OLD_WORKFLOW_OOM_MULTIPLIER = 1.3 + +class RAMPressureCache(LRUCache): + + def __init__(self, key_class): + super().__init__(key_class, 0) + self.timestamps = {} + + def clean_unused(self): + self._clean_subcaches() + + def set(self, node_id, value): + self.timestamps[self.cache_key_set.get_data_key(node_id)] = time.time() + super().set(node_id, value) + + def get(self, node_id): + self.timestamps[self.cache_key_set.get_data_key(node_id)] = time.time() + return super().get(node_id) + + def poll(self, ram_headroom): + def _ram_gb(): + return psutil.virtual_memory().available / (1024**3) + + if _ram_gb() > ram_headroom: + return + gc.collect() + if _ram_gb() > ram_headroom: + return + + clean_list = [] + + for key, (outputs, _), in self.cache.items(): + oom_score = RAM_CACHE_OLD_WORKFLOW_OOM_MULTIPLIER ** (self.generation - self.used_generation[key]) + + ram_usage = RAM_CACHE_DEFAULT_RAM_USAGE + def scan_list_for_ram_usage(outputs): + nonlocal ram_usage + for output in outputs: + if isinstance(output, list): + scan_list_for_ram_usage(output) + elif isinstance(output, torch.Tensor) and output.device.type == 'cpu': + #score Tensors at a 50% discount for RAM usage as they are likely to + #be high value intermediates + ram_usage += (output.numel() * output.element_size()) * 0.5 + elif hasattr(output, "get_ram_usage"): + ram_usage += output.get_ram_usage() + scan_list_for_ram_usage(outputs) + + oom_score *= ram_usage + #In the case where we have no information on the node ram usage at all, + #break OOM score ties on the last touch timestamp (pure LRU) + bisect.insort(clean_list, (oom_score, self.timestamps[key], key)) + + while _ram_gb() < ram_headroom * RAM_CACHE_HYSTERESIS and clean_list: + _, _, key = clean_list.pop() + del self.cache[key] + gc.collect() diff --git a/comfy_execution/graph.py b/comfy_execution/graph.py index 341c9735d..0d811e354 100644 --- a/comfy_execution/graph.py +++ b/comfy_execution/graph.py @@ -209,10 +209,15 @@ class ExecutionList(TopologicalSort): self.execution_cache_listeners[from_node_id] = set() self.execution_cache_listeners[from_node_id].add(to_node_id) - def get_output_cache(self, from_node_id, to_node_id): + def get_cache(self, from_node_id, to_node_id): if not to_node_id in self.execution_cache: return None - return self.execution_cache[to_node_id].get(from_node_id) + value = self.execution_cache[to_node_id].get(from_node_id) + if value is None: + return None + #Write back to the main cache on touch. + self.output_cache.set(from_node_id, value) + return value def cache_update(self, node_id, value): if node_id in self.execution_cache_listeners: diff --git a/execution.py b/execution.py index 20e106213..17c77beab 100644 --- a/execution.py +++ b/execution.py @@ -21,6 +21,7 @@ from comfy_execution.caching import ( NullCache, HierarchicalCache, LRUCache, + RAMPressureCache, ) from comfy_execution.graph import ( DynamicPrompt, @@ -88,49 +89,56 @@ class IsChangedCache: return self.is_changed[node_id] +class CacheEntry(NamedTuple): + ui: dict + outputs: list + + class CacheType(Enum): CLASSIC = 0 LRU = 1 NONE = 2 + RAM_PRESSURE = 3 class CacheSet: - def __init__(self, cache_type=None, cache_size=None): + def __init__(self, cache_type=None, cache_args={}): if cache_type == CacheType.NONE: self.init_null_cache() logging.info("Disabling intermediate node cache.") + elif cache_type == CacheType.RAM_PRESSURE: + cache_ram = cache_args.get("ram", 16.0) + self.init_ram_cache(cache_ram) + logging.info("Using RAM pressure cache.") elif cache_type == CacheType.LRU: - if cache_size is None: - cache_size = 0 + cache_size = cache_args.get("lru", 0) self.init_lru_cache(cache_size) logging.info("Using LRU cache") else: self.init_classic_cache() - self.all = [self.outputs, self.ui, self.objects] + self.all = [self.outputs, self.objects] # Performs like the old cache -- dump data ASAP def init_classic_cache(self): self.outputs = HierarchicalCache(CacheKeySetInputSignature) - self.ui = HierarchicalCache(CacheKeySetInputSignature) self.objects = HierarchicalCache(CacheKeySetID) def init_lru_cache(self, cache_size): self.outputs = LRUCache(CacheKeySetInputSignature, max_size=cache_size) - self.ui = LRUCache(CacheKeySetInputSignature, max_size=cache_size) + self.objects = HierarchicalCache(CacheKeySetID) + + def init_ram_cache(self, min_headroom): + self.outputs = RAMPressureCache(CacheKeySetInputSignature) self.objects = HierarchicalCache(CacheKeySetID) def init_null_cache(self): self.outputs = NullCache() - #The UI cache is expected to be iterable at the end of each workflow - #so it must cache at least a full workflow. Use Heirachical - self.ui = HierarchicalCache(CacheKeySetInputSignature) self.objects = NullCache() def recursive_debug_dump(self): result = { "outputs": self.outputs.recursive_debug_dump(), - "ui": self.ui.recursive_debug_dump(), } return result @@ -157,14 +165,14 @@ def get_input_data(inputs, class_def, unique_id, execution_list=None, dynprompt= if execution_list is None: mark_missing() continue # This might be a lazily-evaluated input - cached_output = execution_list.get_output_cache(input_unique_id, unique_id) - if cached_output is None: + cached = execution_list.get_cache(input_unique_id, unique_id) + if cached is None or cached.outputs is None: mark_missing() continue - if output_index >= len(cached_output): + if output_index >= len(cached.outputs): mark_missing() continue - obj = cached_output[output_index] + obj = cached.outputs[output_index] input_data_all[x] = obj elif input_category is not None: input_data_all[x] = [input_data] @@ -393,7 +401,7 @@ def format_value(x): else: return str(x) -async def execute(server, dynprompt, caches, current_item, extra_data, executed, prompt_id, execution_list, pending_subgraph_results, pending_async_nodes): +async def execute(server, dynprompt, caches, current_item, extra_data, executed, prompt_id, execution_list, pending_subgraph_results, pending_async_nodes, ui_outputs): unique_id = current_item real_node_id = dynprompt.get_real_node_id(unique_id) display_node_id = dynprompt.get_display_node_id(unique_id) @@ -401,12 +409,15 @@ async def execute(server, dynprompt, caches, current_item, extra_data, executed, inputs = dynprompt.get_node(unique_id)['inputs'] class_type = dynprompt.get_node(unique_id)['class_type'] class_def = nodes.NODE_CLASS_MAPPINGS[class_type] - if caches.outputs.get(unique_id) is not None: + cached = caches.outputs.get(unique_id) + if cached is not None: if server.client_id is not None: - cached_output = caches.ui.get(unique_id) or {} - server.send_sync("executed", { "node": unique_id, "display_node": display_node_id, "output": cached_output.get("output",None), "prompt_id": prompt_id }, server.client_id) + cached_ui = cached.ui or {} + server.send_sync("executed", { "node": unique_id, "display_node": display_node_id, "output": cached_ui.get("output",None), "prompt_id": prompt_id }, server.client_id) + if cached.ui is not None: + ui_outputs[unique_id] = cached.ui get_progress_state().finish_progress(unique_id) - execution_list.cache_update(unique_id, caches.outputs.get(unique_id)) + execution_list.cache_update(unique_id, cached) return (ExecutionResult.SUCCESS, None, None) input_data_all = None @@ -436,8 +447,8 @@ async def execute(server, dynprompt, caches, current_item, extra_data, executed, for r in result: if is_link(r): source_node, source_output = r[0], r[1] - node_output = execution_list.get_output_cache(source_node, unique_id)[source_output] - for o in node_output: + node_cached = execution_list.get_cache(source_node, unique_id) + for o in node_cached.outputs[source_output]: resolved_output.append(o) else: @@ -507,7 +518,7 @@ async def execute(server, dynprompt, caches, current_item, extra_data, executed, asyncio.create_task(await_completion()) return (ExecutionResult.PENDING, None, None) if len(output_ui) > 0: - caches.ui.set(unique_id, { + ui_outputs[unique_id] = { "meta": { "node_id": unique_id, "display_node": display_node_id, @@ -515,7 +526,7 @@ async def execute(server, dynprompt, caches, current_item, extra_data, executed, "real_node_id": real_node_id, }, "output": output_ui - }) + } if server.client_id is not None: server.send_sync("executed", { "node": unique_id, "display_node": display_node_id, "output": output_ui, "prompt_id": prompt_id }, server.client_id) if has_subgraph: @@ -554,8 +565,9 @@ async def execute(server, dynprompt, caches, current_item, extra_data, executed, pending_subgraph_results[unique_id] = cached_outputs return (ExecutionResult.PENDING, None, None) - caches.outputs.set(unique_id, output_data) - execution_list.cache_update(unique_id, output_data) + cache_entry = CacheEntry(ui=ui_outputs.get(unique_id), outputs=output_data) + execution_list.cache_update(unique_id, cache_entry) + caches.outputs.set(unique_id, cache_entry) except comfy.model_management.InterruptProcessingException as iex: logging.info("Processing interrupted") @@ -600,14 +612,14 @@ async def execute(server, dynprompt, caches, current_item, extra_data, executed, return (ExecutionResult.SUCCESS, None, None) class PromptExecutor: - def __init__(self, server, cache_type=False, cache_size=None): - self.cache_size = cache_size + def __init__(self, server, cache_type=False, cache_args=None): + self.cache_args = cache_args self.cache_type = cache_type self.server = server self.reset() def reset(self): - self.caches = CacheSet(cache_type=self.cache_type, cache_size=self.cache_size) + self.caches = CacheSet(cache_type=self.cache_type, cache_args=self.cache_args) self.status_messages = [] self.success = True @@ -682,6 +694,7 @@ class PromptExecutor: broadcast=False) pending_subgraph_results = {} pending_async_nodes = {} # TODO - Unify this with pending_subgraph_results + ui_node_outputs = {} executed = set() execution_list = ExecutionList(dynamic_prompt, self.caches.outputs) current_outputs = self.caches.outputs.all_node_ids() @@ -695,7 +708,7 @@ class PromptExecutor: break assert node_id is not None, "Node ID should not be None at this point" - result, error, ex = await execute(self.server, dynamic_prompt, self.caches, node_id, extra_data, executed, prompt_id, execution_list, pending_subgraph_results, pending_async_nodes) + result, error, ex = await execute(self.server, dynamic_prompt, self.caches, node_id, extra_data, executed, prompt_id, execution_list, pending_subgraph_results, pending_async_nodes, ui_node_outputs) self.success = result != ExecutionResult.FAILURE if result == ExecutionResult.FAILURE: self.handle_execution_error(prompt_id, dynamic_prompt.original_prompt, current_outputs, executed, error, ex) @@ -704,18 +717,16 @@ class PromptExecutor: execution_list.unstage_node_execution() else: # result == ExecutionResult.SUCCESS: execution_list.complete_node_execution() + self.caches.outputs.poll(ram_headroom=self.cache_args["ram"]) else: # Only execute when the while-loop ends without break self.add_message("execution_success", { "prompt_id": prompt_id }, broadcast=False) ui_outputs = {} meta_outputs = {} - all_node_ids = self.caches.ui.all_node_ids() - for node_id in all_node_ids: - ui_info = self.caches.ui.get(node_id) - if ui_info is not None: - ui_outputs[node_id] = ui_info["output"] - meta_outputs[node_id] = ui_info["meta"] + for node_id, ui_info in ui_node_outputs.items(): + ui_outputs[node_id] = ui_info["output"] + meta_outputs[node_id] = ui_info["meta"] self.history_result = { "outputs": ui_outputs, "meta": meta_outputs, diff --git a/main.py b/main.py index 8d466d2eb..e1b0f1620 100644 --- a/main.py +++ b/main.py @@ -172,10 +172,12 @@ def prompt_worker(q, server_instance): cache_type = execution.CacheType.CLASSIC if args.cache_lru > 0: cache_type = execution.CacheType.LRU + elif args.cache_ram > 0: + cache_type = execution.CacheType.RAM_PRESSURE elif args.cache_none: cache_type = execution.CacheType.NONE - e = execution.PromptExecutor(server_instance, cache_type=cache_type, cache_size=args.cache_lru) + e = execution.PromptExecutor(server_instance, cache_type=cache_type, cache_args={ "lru" : args.cache_lru, "ram" : args.cache_ram } ) last_gc_collect = 0 need_gc = False gc_collect_interval = 10.0 From 614cf9805e1056216487a2d1b1a07206d77f87e7 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Thu, 30 Oct 2025 19:11:38 -0700 Subject: [PATCH 059/647] Add a ScaleROPE node. Currently only works on WAN models. (#10559) --- comfy/ldm/wan/model.py | 20 ++++++++++++---- comfy/model_patcher.py | 13 +++++++++++ comfy_extras/nodes_rope.py | 47 ++++++++++++++++++++++++++++++++++++++ nodes.py | 1 + 4 files changed, 77 insertions(+), 4 deletions(-) create mode 100644 comfy_extras/nodes_rope.py diff --git a/comfy/ldm/wan/model.py b/comfy/ldm/wan/model.py index 90c347d3d..77876c2e7 100644 --- a/comfy/ldm/wan/model.py +++ b/comfy/ldm/wan/model.py @@ -588,7 +588,7 @@ class WanModel(torch.nn.Module): x = self.unpatchify(x, grid_sizes) return x - def rope_encode(self, t, h, w, t_start=0, steps_t=None, steps_h=None, steps_w=None, device=None, dtype=None): + def rope_encode(self, t, h, w, t_start=0, steps_t=None, steps_h=None, steps_w=None, device=None, dtype=None, transformer_options={}): patch_size = self.patch_size t_len = ((t + (patch_size[0] // 2)) // patch_size[0]) h_len = ((h + (patch_size[1] // 2)) // patch_size[1]) @@ -601,10 +601,22 @@ class WanModel(torch.nn.Module): if steps_w is None: steps_w = w_len + h_start = 0 + w_start = 0 + rope_options = transformer_options.get("rope_options", None) + if rope_options is not None: + t_len = t_len * rope_options.get("scale_t", 1.0) + h_len = h_len * rope_options.get("scale_y", 1.0) + w_len = w_len * rope_options.get("scale_x", 1.0) + + t_start += rope_options.get("shift_t", 0.0) + h_start += rope_options.get("shift_y", 0.0) + w_start += rope_options.get("shift_x", 0.0) + img_ids = torch.zeros((steps_t, steps_h, steps_w, 3), device=device, dtype=dtype) img_ids[:, :, :, 0] = img_ids[:, :, :, 0] + torch.linspace(t_start, t_start + (t_len - 1), steps=steps_t, device=device, dtype=dtype).reshape(-1, 1, 1) - img_ids[:, :, :, 1] = img_ids[:, :, :, 1] + torch.linspace(0, h_len - 1, steps=steps_h, device=device, dtype=dtype).reshape(1, -1, 1) - img_ids[:, :, :, 2] = img_ids[:, :, :, 2] + torch.linspace(0, w_len - 1, steps=steps_w, device=device, dtype=dtype).reshape(1, 1, -1) + img_ids[:, :, :, 1] = img_ids[:, :, :, 1] + torch.linspace(h_start, h_start + (h_len - 1), steps=steps_h, device=device, dtype=dtype).reshape(1, -1, 1) + img_ids[:, :, :, 2] = img_ids[:, :, :, 2] + torch.linspace(w_start, w_start + (w_len - 1), steps=steps_w, device=device, dtype=dtype).reshape(1, 1, -1) img_ids = img_ids.reshape(1, -1, img_ids.shape[-1]) freqs = self.rope_embedder(img_ids).movedim(1, 2) @@ -630,7 +642,7 @@ class WanModel(torch.nn.Module): if self.ref_conv is not None and "reference_latent" in kwargs: t_len += 1 - freqs = self.rope_encode(t_len, h, w, device=x.device, dtype=x.dtype) + freqs = self.rope_encode(t_len, h, w, device=x.device, dtype=x.dtype, transformer_options=transformer_options) return self.forward_orig(x, timestep, context, clip_fea=clip_fea, freqs=freqs, transformer_options=transformer_options, **kwargs)[:, :, :t, :h, :w] def unpatchify(self, x, grid_sizes): diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index 674a214ca..3e8799983 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -454,6 +454,19 @@ class ModelPatcher: def set_model_post_input_patch(self, patch): self.set_model_patch(patch, "post_input") + def set_model_rope_options(self, scale_x, shift_x, scale_y, shift_y, scale_t, shift_t, **kwargs): + rope_options = self.model_options["transformer_options"].get("rope_options", {}) + rope_options["scale_x"] = scale_x + rope_options["scale_y"] = scale_y + rope_options["scale_t"] = scale_t + + rope_options["shift_x"] = shift_x + rope_options["shift_y"] = shift_y + rope_options["shift_t"] = shift_t + + self.model_options["transformer_options"]["rope_options"] = rope_options + + def add_object_patch(self, name, obj): self.object_patches[name] = obj diff --git a/comfy_extras/nodes_rope.py b/comfy_extras/nodes_rope.py new file mode 100644 index 000000000..d1feb031e --- /dev/null +++ b/comfy_extras/nodes_rope.py @@ -0,0 +1,47 @@ +from comfy_api.latest import ComfyExtension, io +from typing_extensions import override + + +class ScaleROPE(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="ScaleROPE", + category="advanced/model_patches", + description="Scale and shift the ROPE of the model.", + is_experimental=True, + inputs=[ + io.Model.Input("model"), + io.Float.Input("scale_x", default=1.0, min=0.0, max=100.0, step=0.1), + io.Float.Input("shift_x", default=0.0, min=-256.0, max=256.0, step=0.1), + + io.Float.Input("scale_y", default=1.0, min=0.0, max=100.0, step=0.1), + io.Float.Input("shift_y", default=0.0, min=-256.0, max=256.0, step=0.1), + + io.Float.Input("scale_t", default=1.0, min=0.0, max=100.0, step=0.1), + io.Float.Input("shift_t", default=0.0, min=-256.0, max=256.0, step=0.1), + + + ], + outputs=[ + io.Model.Output(), + ], + ) + + @classmethod + def execute(cls, model, scale_x, shift_x, scale_y, shift_y, scale_t, shift_t) -> io.NodeOutput: + m = model.clone() + m.set_model_rope_options(scale_x, shift_x, scale_y, shift_y, scale_t, shift_t) + return io.NodeOutput(m) + + +class RopeExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[io.ComfyNode]]: + return [ + ScaleROPE + ] + + +async def comfy_entrypoint() -> RopeExtension: + return RopeExtension() diff --git a/nodes.py b/nodes.py index 12e365ca9..5689f6fe1 100644 --- a/nodes.py +++ b/nodes.py @@ -2329,6 +2329,7 @@ async def init_builtin_extra_nodes(): "nodes_model_patch.py", "nodes_easycache.py", "nodes_audio_encoder.py", + "nodes_rope.py", ] import_failed = [] From 27d1bd882925e3bbdffb405cea098ac52bb20ac5 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Thu, 30 Oct 2025 19:51:58 -0700 Subject: [PATCH 060/647] Fix rope scaling. (#10560) --- comfy/ldm/wan/model.py | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/comfy/ldm/wan/model.py b/comfy/ldm/wan/model.py index 77876c2e7..5ec1511ce 100644 --- a/comfy/ldm/wan/model.py +++ b/comfy/ldm/wan/model.py @@ -605,9 +605,9 @@ class WanModel(torch.nn.Module): w_start = 0 rope_options = transformer_options.get("rope_options", None) if rope_options is not None: - t_len = t_len * rope_options.get("scale_t", 1.0) - h_len = h_len * rope_options.get("scale_y", 1.0) - w_len = w_len * rope_options.get("scale_x", 1.0) + t_len = (t_len - 1.0) * rope_options.get("scale_t", 1.0) + 1.0 + h_len = (h_len - 1.0) * rope_options.get("scale_y", 1.0) + 1.0 + w_len = (w_len - 1.0) * rope_options.get("scale_x", 1.0) + 1.0 t_start += rope_options.get("shift_t", 0.0) h_start += rope_options.get("shift_y", 0.0) From 7f374e42c833c69c71605507b90f79cc26d14a71 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Fri, 31 Oct 2025 12:41:40 -0700 Subject: [PATCH 061/647] ScaleROPE now works on Lumina models. (#10578) --- comfy/ldm/lumina/model.py | 20 ++++++++++++++++---- 1 file changed, 16 insertions(+), 4 deletions(-) diff --git a/comfy/ldm/lumina/model.py b/comfy/ldm/lumina/model.py index f87d98ac0..b4494a51d 100644 --- a/comfy/ldm/lumina/model.py +++ b/comfy/ldm/lumina/model.py @@ -522,7 +522,7 @@ class NextDiT(nn.Module): max_cap_len = max(l_effective_cap_len) max_img_len = max(l_effective_img_len) - position_ids = torch.zeros(bsz, max_seq_len, 3, dtype=torch.int32, device=device) + position_ids = torch.zeros(bsz, max_seq_len, 3, dtype=torch.float32, device=device) for i in range(bsz): cap_len = l_effective_cap_len[i] @@ -531,10 +531,22 @@ class NextDiT(nn.Module): H_tokens, W_tokens = H // pH, W // pW assert H_tokens * W_tokens == img_len - position_ids[i, :cap_len, 0] = torch.arange(cap_len, dtype=torch.int32, device=device) + rope_options = transformer_options.get("rope_options", None) + h_scale = 1.0 + w_scale = 1.0 + h_start = 0 + w_start = 0 + if rope_options is not None: + h_scale = rope_options.get("scale_y", 1.0) + w_scale = rope_options.get("scale_x", 1.0) + + h_start = rope_options.get("shift_y", 0.0) + w_start = rope_options.get("shift_x", 0.0) + + position_ids[i, :cap_len, 0] = torch.arange(cap_len, dtype=torch.float32, device=device) position_ids[i, cap_len:cap_len+img_len, 0] = cap_len - row_ids = torch.arange(H_tokens, dtype=torch.int32, device=device).view(-1, 1).repeat(1, W_tokens).flatten() - col_ids = torch.arange(W_tokens, dtype=torch.int32, device=device).view(1, -1).repeat(H_tokens, 1).flatten() + row_ids = (torch.arange(H_tokens, dtype=torch.float32, device=device) * h_scale + h_start).view(-1, 1).repeat(1, W_tokens).flatten() + col_ids = (torch.arange(W_tokens, dtype=torch.float32, device=device) * w_scale + w_start).view(1, -1).repeat(H_tokens, 1).flatten() position_ids[i, cap_len:cap_len+img_len, 1] = row_ids position_ids[i, cap_len:cap_len+img_len, 2] = col_ids From c58c13b2bad6df0de93cc0cf107e96522a3cb5b3 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Fri, 31 Oct 2025 21:25:17 -0700 Subject: [PATCH 062/647] Fix torch compile regression on fp8 ops. (#10580) --- comfy/ops.py | 24 +++++------------ comfy/quant_ops.py | 27 +++++++++++++++---- .../comfy_quant/test_mixed_precision.py | 8 +++--- tests-unit/comfy_quant/test_quant_registry.py | 20 +++++++------- 4 files changed, 43 insertions(+), 36 deletions(-) diff --git a/comfy/ops.py b/comfy/ops.py index 18f6b804b..279f6b1a7 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -401,15 +401,9 @@ def fp8_linear(self, input): if dtype not in [torch.float8_e4m3fn]: return None - tensor_2d = False - if len(input.shape) == 2: - tensor_2d = True - input = input.unsqueeze(1) - - input_shape = input.shape input_dtype = input.dtype - if len(input.shape) == 3: + if input.ndim == 3 or input.ndim == 2: w, bias, offload_stream = cast_bias_weight(self, input, dtype=dtype, bias_dtype=input_dtype, offloadable=True) scale_weight = self.scale_weight @@ -422,24 +416,20 @@ def fp8_linear(self, input): if scale_input is None: scale_input = torch.ones((), device=input.device, dtype=torch.float32) input = torch.clamp(input, min=-448, max=448, out=input) - input = input.reshape(-1, input_shape[2]).to(dtype).contiguous() layout_params_weight = {'scale': scale_input, 'orig_dtype': input_dtype} - quantized_input = QuantizedTensor(input.reshape(-1, input_shape[2]).to(dtype).contiguous(), TensorCoreFP8Layout, layout_params_weight) + quantized_input = QuantizedTensor(input.to(dtype).contiguous(), "TensorCoreFP8Layout", layout_params_weight) else: scale_input = scale_input.to(input.device) - quantized_input = QuantizedTensor.from_float(input.reshape(-1, input_shape[2]), TensorCoreFP8Layout, scale=scale_input, dtype=dtype) + quantized_input = QuantizedTensor.from_float(input, "TensorCoreFP8Layout", scale=scale_input, dtype=dtype) # Wrap weight in QuantizedTensor - this enables unified dispatch # Call F.linear - __torch_dispatch__ routes to fp8_linear handler in quant_ops.py! layout_params_weight = {'scale': scale_weight, 'orig_dtype': input_dtype} - quantized_weight = QuantizedTensor(w, TensorCoreFP8Layout, layout_params_weight) + quantized_weight = QuantizedTensor(w, "TensorCoreFP8Layout", layout_params_weight) o = torch.nn.functional.linear(quantized_input, quantized_weight, bias) uncast_bias_weight(self, w, bias, offload_stream) - - if tensor_2d: - return o.reshape(input_shape[0], -1) - return o.reshape((-1, input_shape[1], self.weight.shape[0])) + return o return None @@ -540,12 +530,12 @@ if CUBLAS_IS_AVAILABLE: # ============================================================================== # Mixed Precision Operations # ============================================================================== -from .quant_ops import QuantizedTensor, TensorCoreFP8Layout +from .quant_ops import QuantizedTensor QUANT_FORMAT_MIXINS = { "float8_e4m3fn": { "dtype": torch.float8_e4m3fn, - "layout_type": TensorCoreFP8Layout, + "layout_type": "TensorCoreFP8Layout", "parameters": { "weight_scale": torch.nn.Parameter(torch.zeros((), dtype=torch.float32), requires_grad=False), "input_scale": torch.nn.Parameter(torch.zeros((), dtype=torch.float32), requires_grad=False), diff --git a/comfy/quant_ops.py b/comfy/quant_ops.py index c822fe53c..873f173ed 100644 --- a/comfy/quant_ops.py +++ b/comfy/quant_ops.py @@ -123,7 +123,7 @@ class QuantizedTensor(torch.Tensor): layout_type: Layout class (subclass of QuantizedLayout) layout_params: Dict with layout-specific parameters """ - return torch.Tensor._make_subclass(cls, qdata, require_grad=False) + return torch.Tensor._make_wrapper_subclass(cls, qdata.shape, device=qdata.device, dtype=qdata.dtype, requires_grad=False) def __init__(self, qdata, layout_type, layout_params): self._qdata = qdata.contiguous() @@ -183,11 +183,11 @@ class QuantizedTensor(torch.Tensor): @classmethod def from_float(cls, tensor, layout_type, **quantize_kwargs) -> 'QuantizedTensor': - qdata, layout_params = layout_type.quantize(tensor, **quantize_kwargs) + qdata, layout_params = LAYOUTS[layout_type].quantize(tensor, **quantize_kwargs) return cls(qdata, layout_type, layout_params) def dequantize(self) -> torch.Tensor: - return self._layout_type.dequantize(self._qdata, **self._layout_params) + return LAYOUTS[self._layout_type].dequantize(self._qdata, **self._layout_params) @classmethod def __torch_dispatch__(cls, func, types, args=(), kwargs=None): @@ -379,7 +379,12 @@ class TensorCoreFP8Layout(QuantizedLayout): return qtensor._qdata, qtensor._layout_params['scale'] -@register_layout_op(torch.ops.aten.linear.default, TensorCoreFP8Layout) +LAYOUTS = { + "TensorCoreFP8Layout": TensorCoreFP8Layout, +} + + +@register_layout_op(torch.ops.aten.linear.default, "TensorCoreFP8Layout") def fp8_linear(func, args, kwargs): input_tensor = args[0] weight = args[1] @@ -422,7 +427,7 @@ def fp8_linear(func, args, kwargs): 'scale': output_scale, 'orig_dtype': input_tensor._layout_params['orig_dtype'] } - return QuantizedTensor(output, TensorCoreFP8Layout, output_params) + return QuantizedTensor(output, "TensorCoreFP8Layout", output_params) else: return output @@ -436,3 +441,15 @@ def fp8_linear(func, args, kwargs): input_tensor = input_tensor.dequantize() return torch.nn.functional.linear(input_tensor, weight, bias) + + +@register_layout_op(torch.ops.aten.view.default, "TensorCoreFP8Layout") +@register_layout_op(torch.ops.aten.t.default, "TensorCoreFP8Layout") +def fp8_func(func, args, kwargs): + input_tensor = args[0] + if isinstance(input_tensor, QuantizedTensor): + plain_input, scale_a = TensorCoreFP8Layout.get_plain_tensors(input_tensor) + ar = list(args) + ar[0] = plain_input + return QuantizedTensor(func(*ar, **kwargs), "TensorCoreFP8Layout", input_tensor._layout_params) + return func(*args, **kwargs) diff --git a/tests-unit/comfy_quant/test_mixed_precision.py b/tests-unit/comfy_quant/test_mixed_precision.py index 267bc177b..f8d1fd04e 100644 --- a/tests-unit/comfy_quant/test_mixed_precision.py +++ b/tests-unit/comfy_quant/test_mixed_precision.py @@ -14,7 +14,7 @@ if not has_gpu(): args.cpu = True from comfy import ops -from comfy.quant_ops import QuantizedTensor, TensorCoreFP8Layout +from comfy.quant_ops import QuantizedTensor class SimpleModel(torch.nn.Module): @@ -104,14 +104,14 @@ class TestMixedPrecisionOps(unittest.TestCase): # Verify weights are wrapped in QuantizedTensor self.assertIsInstance(model.layer1.weight, QuantizedTensor) - self.assertEqual(model.layer1.weight._layout_type, TensorCoreFP8Layout) + self.assertEqual(model.layer1.weight._layout_type, "TensorCoreFP8Layout") # Layer 2 should NOT be quantized self.assertNotIsInstance(model.layer2.weight, QuantizedTensor) # Layer 3 should be quantized self.assertIsInstance(model.layer3.weight, QuantizedTensor) - self.assertEqual(model.layer3.weight._layout_type, TensorCoreFP8Layout) + self.assertEqual(model.layer3.weight._layout_type, "TensorCoreFP8Layout") # Verify scales were loaded self.assertEqual(model.layer1.weight._layout_params['scale'].item(), 2.0) @@ -155,7 +155,7 @@ class TestMixedPrecisionOps(unittest.TestCase): # Verify layer1.weight is a QuantizedTensor with scale preserved self.assertIsInstance(state_dict2["layer1.weight"], QuantizedTensor) self.assertEqual(state_dict2["layer1.weight"]._layout_params['scale'].item(), 3.0) - self.assertEqual(state_dict2["layer1.weight"]._layout_type, TensorCoreFP8Layout) + self.assertEqual(state_dict2["layer1.weight"]._layout_type, "TensorCoreFP8Layout") # Verify non-quantized layers are standard tensors self.assertNotIsInstance(state_dict2["layer2.weight"], QuantizedTensor) diff --git a/tests-unit/comfy_quant/test_quant_registry.py b/tests-unit/comfy_quant/test_quant_registry.py index 477811029..9cb54ede8 100644 --- a/tests-unit/comfy_quant/test_quant_registry.py +++ b/tests-unit/comfy_quant/test_quant_registry.py @@ -25,14 +25,14 @@ class TestQuantizedTensor(unittest.TestCase): scale = torch.tensor(2.0) layout_params = {'scale': scale, 'orig_dtype': torch.bfloat16} - qt = QuantizedTensor(fp8_data, TensorCoreFP8Layout, layout_params) + qt = QuantizedTensor(fp8_data, "TensorCoreFP8Layout", layout_params) self.assertIsInstance(qt, QuantizedTensor) self.assertEqual(qt.shape, (256, 128)) self.assertEqual(qt.dtype, torch.float8_e4m3fn) self.assertEqual(qt._layout_params['scale'], scale) self.assertEqual(qt._layout_params['orig_dtype'], torch.bfloat16) - self.assertEqual(qt._layout_type, TensorCoreFP8Layout) + self.assertEqual(qt._layout_type, "TensorCoreFP8Layout") def test_dequantize(self): """Test explicit dequantization""" @@ -41,7 +41,7 @@ class TestQuantizedTensor(unittest.TestCase): scale = torch.tensor(3.0) layout_params = {'scale': scale, 'orig_dtype': torch.float32} - qt = QuantizedTensor(fp8_data, TensorCoreFP8Layout, layout_params) + qt = QuantizedTensor(fp8_data, "TensorCoreFP8Layout", layout_params) dequantized = qt.dequantize() self.assertEqual(dequantized.dtype, torch.float32) @@ -54,7 +54,7 @@ class TestQuantizedTensor(unittest.TestCase): qt = QuantizedTensor.from_float( float_tensor, - TensorCoreFP8Layout, + "TensorCoreFP8Layout", scale=scale, dtype=torch.float8_e4m3fn ) @@ -77,28 +77,28 @@ class TestGenericUtilities(unittest.TestCase): fp8_data = torch.randn(10, 20, dtype=torch.float32).to(torch.float8_e4m3fn) scale = torch.tensor(1.5) layout_params = {'scale': scale, 'orig_dtype': torch.float32} - qt = QuantizedTensor(fp8_data, TensorCoreFP8Layout, layout_params) + qt = QuantizedTensor(fp8_data, "TensorCoreFP8Layout", layout_params) # Detach should return a new QuantizedTensor qt_detached = qt.detach() self.assertIsInstance(qt_detached, QuantizedTensor) self.assertEqual(qt_detached.shape, qt.shape) - self.assertEqual(qt_detached._layout_type, TensorCoreFP8Layout) + self.assertEqual(qt_detached._layout_type, "TensorCoreFP8Layout") def test_clone(self): """Test clone operation on quantized tensor""" fp8_data = torch.randn(10, 20, dtype=torch.float32).to(torch.float8_e4m3fn) scale = torch.tensor(1.5) layout_params = {'scale': scale, 'orig_dtype': torch.float32} - qt = QuantizedTensor(fp8_data, TensorCoreFP8Layout, layout_params) + qt = QuantizedTensor(fp8_data, "TensorCoreFP8Layout", layout_params) # Clone should return a new QuantizedTensor qt_cloned = qt.clone() self.assertIsInstance(qt_cloned, QuantizedTensor) self.assertEqual(qt_cloned.shape, qt.shape) - self.assertEqual(qt_cloned._layout_type, TensorCoreFP8Layout) + self.assertEqual(qt_cloned._layout_type, "TensorCoreFP8Layout") # Verify it's a deep copy self.assertIsNot(qt_cloned._qdata, qt._qdata) @@ -109,7 +109,7 @@ class TestGenericUtilities(unittest.TestCase): fp8_data = torch.randn(10, 20, dtype=torch.float32).to(torch.float8_e4m3fn) scale = torch.tensor(1.5) layout_params = {'scale': scale, 'orig_dtype': torch.float32} - qt = QuantizedTensor(fp8_data, TensorCoreFP8Layout, layout_params) + qt = QuantizedTensor(fp8_data, "TensorCoreFP8Layout", layout_params) # Moving to same device should work (CPU to CPU) qt_cpu = qt.to('cpu') @@ -169,7 +169,7 @@ class TestFallbackMechanism(unittest.TestCase): scale = torch.tensor(1.0) a_q = QuantizedTensor.from_float( a_fp32, - TensorCoreFP8Layout, + "TensorCoreFP8Layout", scale=scale, dtype=torch.float8_e4m3fn ) From 5f109fe6a06a3462b31a066bcfd650de67d66102 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Sat, 1 Nov 2025 21:13:39 +0200 Subject: [PATCH 063/647] added 12s-20s as available output durations for the LTXV API nodes (#10570) --- comfy_api_nodes/nodes_ltxv.py | 12 ++++++++++-- 1 file changed, 10 insertions(+), 2 deletions(-) diff --git a/comfy_api_nodes/nodes_ltxv.py b/comfy_api_nodes/nodes_ltxv.py index e6ad6e27a..0b757a62b 100644 --- a/comfy_api_nodes/nodes_ltxv.py +++ b/comfy_api_nodes/nodes_ltxv.py @@ -46,7 +46,7 @@ class TextToVideoNode(IO.ComfyNode): multiline=True, default="", ), - IO.Combo.Input("duration", options=[6, 8, 10], default=8), + IO.Combo.Input("duration", options=[6, 8, 10, 12, 14, 16, 18, 20], default=8), IO.Combo.Input( "resolution", options=[ @@ -85,6 +85,10 @@ class TextToVideoNode(IO.ComfyNode): generate_audio: bool = False, ) -> IO.NodeOutput: validate_string(prompt, min_length=1, max_length=10000) + if duration > 10 and (model != "LTX-2 (Fast)" or resolution != "1920x1080" or fps != 25): + raise ValueError( + "Durations over 10s are only available for the Fast model at 1920x1080 resolution and 25 FPS." + ) response = await sync_op_raw( cls, ApiEndpoint("/proxy/ltx/v1/text-to-video", "POST"), @@ -118,7 +122,7 @@ class ImageToVideoNode(IO.ComfyNode): multiline=True, default="", ), - IO.Combo.Input("duration", options=[6, 8, 10], default=8), + IO.Combo.Input("duration", options=[6, 8, 10, 12, 14, 16, 18, 20], default=8), IO.Combo.Input( "resolution", options=[ @@ -158,6 +162,10 @@ class ImageToVideoNode(IO.ComfyNode): generate_audio: bool = False, ) -> IO.NodeOutput: validate_string(prompt, min_length=1, max_length=10000) + if duration > 10 and (model != "LTX-2 (Fast)" or resolution != "1920x1080" or fps != 25): + raise ValueError( + "Durations over 10s are only available for the Fast model at 1920x1080 resolution and 25 FPS." + ) if get_number_of_images(image) != 1: raise ValueError("Currently only one input image is supported.") response = await sync_op_raw( From 20182a393f43ab1fdf798f8da6aac0ef6116e7e6 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Sat, 1 Nov 2025 21:14:06 +0200 Subject: [PATCH 064/647] convert StabilityAI to use new API client (#10582) --- comfy_api_nodes/nodes_stability.py | 179 ++++++++--------------------- comfy_api_nodes/util/client.py | 4 +- 2 files changed, 48 insertions(+), 135 deletions(-) diff --git a/comfy_api_nodes/nodes_stability.py b/comfy_api_nodes/nodes_stability.py index 783666ddf..bb7ceed78 100644 --- a/comfy_api_nodes/nodes_stability.py +++ b/comfy_api_nodes/nodes_stability.py @@ -20,13 +20,6 @@ from comfy_api_nodes.apis.stability_api import ( StabilityAudioInpaintRequest, StabilityAudioResponse, ) -from comfy_api_nodes.apis.client import ( - ApiEndpoint, - HttpMethod, - SynchronousOperation, - PollingOperation, - EmptyRequest, -) from comfy_api_nodes.util import ( validate_audio_duration, validate_string, @@ -34,6 +27,9 @@ from comfy_api_nodes.util import ( bytesio_to_image_tensor, tensor_to_bytesio, audio_bytes_to_audio_input, + sync_op, + poll_op, + ApiEndpoint, ) import torch @@ -161,19 +157,11 @@ class StabilityStableImageUltraNode(IO.ComfyNode): "image": image_binary } - auth = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } - - operation = SynchronousOperation( - endpoint=ApiEndpoint( - path="/proxy/stability/v2beta/stable-image/generate/ultra", - method=HttpMethod.POST, - request_model=StabilityStableUltraRequest, - response_model=StabilityStableUltraResponse, - ), - request=StabilityStableUltraRequest( + response_api = await sync_op( + cls, + ApiEndpoint(path="/proxy/stability/v2beta/stable-image/generate/ultra", method="POST"), + response_model=StabilityStableUltraResponse, + data=StabilityStableUltraRequest( prompt=prompt, negative_prompt=negative_prompt, aspect_ratio=aspect_ratio, @@ -183,9 +171,7 @@ class StabilityStableImageUltraNode(IO.ComfyNode): ), files=files, content_type="multipart/form-data", - auth_kwargs=auth, ) - response_api = await operation.execute() if response_api.finish_reason != "SUCCESS": raise Exception(f"Stable Image Ultra generation failed: {response_api.finish_reason}.") @@ -313,19 +299,11 @@ class StabilityStableImageSD_3_5Node(IO.ComfyNode): "image": image_binary } - auth = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } - - operation = SynchronousOperation( - endpoint=ApiEndpoint( - path="/proxy/stability/v2beta/stable-image/generate/sd3", - method=HttpMethod.POST, - request_model=StabilityStable3_5Request, - response_model=StabilityStableUltraResponse, - ), - request=StabilityStable3_5Request( + response_api = await sync_op( + cls, + ApiEndpoint(path="/proxy/stability/v2beta/stable-image/generate/sd3", method="POST"), + response_model=StabilityStableUltraResponse, + data=StabilityStable3_5Request( prompt=prompt, negative_prompt=negative_prompt, aspect_ratio=aspect_ratio, @@ -338,9 +316,7 @@ class StabilityStableImageSD_3_5Node(IO.ComfyNode): ), files=files, content_type="multipart/form-data", - auth_kwargs=auth, ) - response_api = await operation.execute() if response_api.finish_reason != "SUCCESS": raise Exception(f"Stable Diffusion 3.5 Image generation failed: {response_api.finish_reason}.") @@ -427,19 +403,11 @@ class StabilityUpscaleConservativeNode(IO.ComfyNode): "image": image_binary } - auth = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } - - operation = SynchronousOperation( - endpoint=ApiEndpoint( - path="/proxy/stability/v2beta/stable-image/upscale/conservative", - method=HttpMethod.POST, - request_model=StabilityUpscaleConservativeRequest, - response_model=StabilityStableUltraResponse, - ), - request=StabilityUpscaleConservativeRequest( + response_api = await sync_op( + cls, + ApiEndpoint(path="/proxy/stability/v2beta/stable-image/upscale/conservative", method="POST"), + response_model=StabilityStableUltraResponse, + data=StabilityUpscaleConservativeRequest( prompt=prompt, negative_prompt=negative_prompt, creativity=round(creativity,2), @@ -447,9 +415,7 @@ class StabilityUpscaleConservativeNode(IO.ComfyNode): ), files=files, content_type="multipart/form-data", - auth_kwargs=auth, ) - response_api = await operation.execute() if response_api.finish_reason != "SUCCESS": raise Exception(f"Stability Upscale Conservative generation failed: {response_api.finish_reason}.") @@ -544,19 +510,11 @@ class StabilityUpscaleCreativeNode(IO.ComfyNode): "image": image_binary } - auth = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } - - operation = SynchronousOperation( - endpoint=ApiEndpoint( - path="/proxy/stability/v2beta/stable-image/upscale/creative", - method=HttpMethod.POST, - request_model=StabilityUpscaleCreativeRequest, - response_model=StabilityAsyncResponse, - ), - request=StabilityUpscaleCreativeRequest( + response_api = await sync_op( + cls, + ApiEndpoint(path="/proxy/stability/v2beta/stable-image/upscale/creative", method="POST"), + response_model=StabilityAsyncResponse, + data=StabilityUpscaleCreativeRequest( prompt=prompt, negative_prompt=negative_prompt, creativity=round(creativity,2), @@ -565,25 +523,15 @@ class StabilityUpscaleCreativeNode(IO.ComfyNode): ), files=files, content_type="multipart/form-data", - auth_kwargs=auth, ) - response_api = await operation.execute() - operation = PollingOperation( - poll_endpoint=ApiEndpoint( - path=f"/proxy/stability/v2beta/results/{response_api.id}", - method=HttpMethod.GET, - request_model=EmptyRequest, - response_model=StabilityResultsGetResponse, - ), + response_poll = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/stability/v2beta/results/{response_api.id}"), + response_model=StabilityResultsGetResponse, poll_interval=3, - completed_statuses=[StabilityPollStatus.finished], - failed_statuses=[StabilityPollStatus.failed], status_extractor=lambda x: get_async_dummy_status(x), - auth_kwargs=auth, - node_id=cls.hidden.unique_id, ) - response_poll: StabilityResultsGetResponse = await operation.execute() if response_poll.finish_reason != "SUCCESS": raise Exception(f"Stability Upscale Creative generation failed: {response_poll.finish_reason}.") @@ -628,24 +576,13 @@ class StabilityUpscaleFastNode(IO.ComfyNode): "image": image_binary } - auth = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } - - operation = SynchronousOperation( - endpoint=ApiEndpoint( - path="/proxy/stability/v2beta/stable-image/upscale/fast", - method=HttpMethod.POST, - request_model=EmptyRequest, - response_model=StabilityStableUltraResponse, - ), - request=EmptyRequest(), + response_api = await sync_op( + cls, + ApiEndpoint(path="/proxy/stability/v2beta/stable-image/upscale/fast", method="POST"), + response_model=StabilityStableUltraResponse, files=files, content_type="multipart/form-data", - auth_kwargs=auth, ) - response_api = await operation.execute() if response_api.finish_reason != "SUCCESS": raise Exception(f"Stability Upscale Fast failed: {response_api.finish_reason}.") @@ -717,21 +654,13 @@ class StabilityTextToAudio(IO.ComfyNode): async def execute(cls, model: str, prompt: str, duration: int, seed: int, steps: int) -> IO.NodeOutput: validate_string(prompt, max_length=10000) payload = StabilityTextToAudioRequest(prompt=prompt, model=model, duration=duration, seed=seed, steps=steps) - operation = SynchronousOperation( - endpoint=ApiEndpoint( - path="/proxy/stability/v2beta/audio/stable-audio-2/text-to-audio", - method=HttpMethod.POST, - request_model=StabilityTextToAudioRequest, - response_model=StabilityAudioResponse, - ), - request=payload, + response_api = await sync_op( + cls, + ApiEndpoint(path="/proxy/stability/v2beta/audio/stable-audio-2/text-to-audio", method="POST"), + response_model=StabilityAudioResponse, + data=payload, content_type="multipart/form-data", - auth_kwargs= { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - }, ) - response_api = await operation.execute() if not response_api.audio: raise ValueError("No audio file was received in response.") return IO.NodeOutput(audio_bytes_to_audio_input(base64.b64decode(response_api.audio))) @@ -814,22 +743,14 @@ class StabilityAudioToAudio(IO.ComfyNode): payload = StabilityAudioToAudioRequest( prompt=prompt, model=model, duration=duration, seed=seed, steps=steps, strength=strength ) - operation = SynchronousOperation( - endpoint=ApiEndpoint( - path="/proxy/stability/v2beta/audio/stable-audio-2/audio-to-audio", - method=HttpMethod.POST, - request_model=StabilityAudioToAudioRequest, - response_model=StabilityAudioResponse, - ), - request=payload, + response_api = await sync_op( + cls, + ApiEndpoint(path="/proxy/stability/v2beta/audio/stable-audio-2/audio-to-audio", method="POST"), + response_model=StabilityAudioResponse, + data=payload, content_type="multipart/form-data", files={"audio": audio_input_to_mp3(audio)}, - auth_kwargs= { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - }, ) - response_api = await operation.execute() if not response_api.audio: raise ValueError("No audio file was received in response.") return IO.NodeOutput(audio_bytes_to_audio_input(base64.b64decode(response_api.audio))) @@ -935,22 +856,14 @@ class StabilityAudioInpaint(IO.ComfyNode): mask_start=mask_start, mask_end=mask_end, ) - operation = SynchronousOperation( - endpoint=ApiEndpoint( - path="/proxy/stability/v2beta/audio/stable-audio-2/inpaint", - method=HttpMethod.POST, - request_model=StabilityAudioInpaintRequest, - response_model=StabilityAudioResponse, - ), - request=payload, + response_api = await sync_op( + cls, + endpoint=ApiEndpoint(path="/proxy/stability/v2beta/audio/stable-audio-2/inpaint", method="POST"), + response_model=StabilityAudioResponse, + data=payload, content_type="multipart/form-data", files={"audio": audio_input_to_mp3(audio)}, - auth_kwargs={ - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - }, ) - response_api = await operation.execute() if not response_api.audio: raise ValueError("No audio file was received in response.") return IO.NodeOutput(audio_bytes_to_audio_input(base64.b64decode(response_api.audio))) diff --git a/comfy_api_nodes/util/client.py b/comfy_api_nodes/util/client.py index 9ae512fe5..65bb35f0f 100644 --- a/comfy_api_nodes/util/client.py +++ b/comfy_api_nodes/util/client.py @@ -77,7 +77,7 @@ class _PollUIState: _RETRY_STATUS = {408, 429, 500, 502, 503, 504} -COMPLETED_STATUSES = ["succeeded", "succeed", "success", "completed"] +COMPLETED_STATUSES = ["succeeded", "succeed", "success", "completed", "finished"] FAILED_STATUSES = ["cancelled", "canceled", "fail", "failed", "error"] QUEUED_STATUSES = ["created", "queued", "queueing", "submitted"] @@ -589,7 +589,7 @@ async def _request_base(cfg: _RequestConfig, expect_binary: bool): operation_id = _generate_operation_id(method, cfg.endpoint.path, attempt) logging.debug("[DEBUG] HTTP %s %s (attempt %d)", method, url, attempt) - payload_headers = {"Accept": "*/*"} + payload_headers = {"Accept": "*/*"} if expect_binary else {"Accept": "application/json"} if not parsed_url.scheme and not parsed_url.netloc: # is URL relative? payload_headers.update(get_auth_header(cfg.node_cls)) if cfg.endpoint.headers: From 44869ff786dc90b36172fd766c9a110e4c40c04b Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sat, 1 Nov 2025 14:25:59 -0700 Subject: [PATCH 065/647] Fix issue with pinned memory. (#10597) --- comfy/model_patcher.py | 1 + 1 file changed, 1 insertion(+) diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index 3e8799983..5a31a8734 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -298,6 +298,7 @@ class ModelPatcher: n.backup = self.backup n.object_patches_backup = self.object_patches_backup n.parent = self + n.pinned = self.pinned n.force_cast_weights = self.force_cast_weights From 135fa49ec23320834f774cf3def9e51ad3773f86 Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Sun, 2 Nov 2025 08:48:53 +1000 Subject: [PATCH 066/647] Small speed improvements to --async-offload (#10593) * ops: dont take an offload stream if you dont need one * ops: prioritize mem transfer The async offload streams reason for existence is to transfer from RAM to GPU. The post processing compute steps are a bonus on the side stream, but if the compute stream is running a long kernel, it can stall the side stream, as it wait to type-cast the bias before transferring the weight. So do a pure xfer of the weight straight up, then do everything bias, then go back to fix the weight type and do weight patches. --- comfy/ops.py | 23 ++++++++++++++--------- 1 file changed, 14 insertions(+), 9 deletions(-) diff --git a/comfy/ops.py b/comfy/ops.py index 279f6b1a7..0c8f23848 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -84,7 +84,8 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, of if device is None: device = input.device - if offloadable: + if offloadable and (device != s.weight.device or + (s.bias is not None and device != s.bias.device)): offload_stream = comfy.model_management.get_offload_stream(device) else: offload_stream = None @@ -94,20 +95,24 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, of else: wf_context = contextlib.nullcontext() - bias = None non_blocking = comfy.model_management.device_supports_non_blocking(device) - if s.bias is not None: - has_function = len(s.bias_function) > 0 - bias = comfy.model_management.cast_to(s.bias, bias_dtype, device, non_blocking=non_blocking, copy=has_function, stream=offload_stream) - if has_function: + weight_has_function = len(s.weight_function) > 0 + bias_has_function = len(s.bias_function) > 0 + + weight = comfy.model_management.cast_to(s.weight, None, device, non_blocking=non_blocking, copy=weight_has_function, stream=offload_stream) + + bias = None + if s.bias is not None: + bias = comfy.model_management.cast_to(s.bias, bias_dtype, device, non_blocking=non_blocking, copy=bias_has_function, stream=offload_stream) + + if bias_has_function: with wf_context: for f in s.bias_function: bias = f(bias) - has_function = len(s.weight_function) > 0 - weight = comfy.model_management.cast_to(s.weight, dtype, device, non_blocking=non_blocking, copy=has_function, stream=offload_stream) - if has_function: + weight = weight.to(dtype=dtype) + if weight_has_function: with wf_context: for f in s.weight_function: weight = f(weight) From 97ff9fae7e728cffdfc3aee6d72aa1e0d0b78702 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sun, 2 Nov 2025 10:14:04 -0800 Subject: [PATCH 067/647] Clarify help text for --fast argument (#10609) Updated help text for the --fast argument to clarify potential risks. --- comfy/cli_args.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy/cli_args.py b/comfy/cli_args.py index 3d5bc7c90..3947e62a8 100644 --- a/comfy/cli_args.py +++ b/comfy/cli_args.py @@ -147,7 +147,7 @@ class PerformanceFeature(enum.Enum): AutoTune = "autotune" PinnedMem = "pinned_memory" -parser.add_argument("--fast", nargs="*", type=PerformanceFeature, help="Enable some untested and potentially quality deteriorating optimizations. --fast with no arguments enables everything. You can pass a list specific optimizations if you only want to enable specific ones. Current valid optimizations: {}".format(" ".join(map(lambda c: c.value, PerformanceFeature)))) +parser.add_argument("--fast", nargs="*", type=PerformanceFeature, help="Enable some untested and potentially quality deteriorating optimizations. This is used to test new features so using it might crash your comfyui. --fast with no arguments enables everything. You can pass a list specific optimizations if you only want to enable specific ones. Current valid optimizations: {}".format(" ".join(map(lambda c: c.value, PerformanceFeature)))) parser.add_argument("--mmap-torch-files", action="store_true", help="Use mmap when loading ckpt/pt files.") parser.add_argument("--disable-mmap", action="store_true", help="Don't use mmap when loading safetensors.") From 6d6a18b0b730351416556eeb0990ab219ffec189 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Mon, 3 Nov 2025 10:04:56 +0200 Subject: [PATCH 068/647] fix(api-nodes-cloud): stop using sub-folder and absolute path for output of Rodin3D nodes (#10556) --- comfy_api_nodes/nodes_rodin.py | 11 +++++------ 1 file changed, 5 insertions(+), 6 deletions(-) diff --git a/comfy_api_nodes/nodes_rodin.py b/comfy_api_nodes/nodes_rodin.py index cf2172bd6..ad4029236 100644 --- a/comfy_api_nodes/nodes_rodin.py +++ b/comfy_api_nodes/nodes_rodin.py @@ -225,21 +225,20 @@ async def get_rodin_download_list(uuid, auth_kwargs: Optional[dict[str, str]] = async def download_files(url_list, task_uuid): - save_path = os.path.join(comfy_paths.get_output_directory(), f"Rodin3D_{task_uuid}") + result_folder_name = f"Rodin3D_{task_uuid}" + save_path = os.path.join(comfy_paths.get_output_directory(), result_folder_name) os.makedirs(save_path, exist_ok=True) model_file_path = None async with aiohttp.ClientSession() as session: for i in url_list.list: - url = i.url - file_name = i.name - file_path = os.path.join(save_path, file_name) + file_path = os.path.join(save_path, i.name) if file_path.endswith(".glb"): - model_file_path = file_path + model_file_path = os.path.join(result_folder_name, i.name) logging.info("[ Rodin3D API - download_files ] Downloading file: %s", file_path) max_retries = 5 for attempt in range(max_retries): try: - async with session.get(url) as resp: + async with session.get(i.url) as resp: resp.raise_for_status() with open(file_path, "wb") as f: async for chunk in resp.content.iter_chunked(32 * 1024): From 88df172790b8ed7b2e6ea7c0f0bd63ca3553921b Mon Sep 17 00:00:00 2001 From: EverNebula Date: Mon, 3 Nov 2025 16:16:40 +0800 Subject: [PATCH 069/647] fix(caching): treat bytes as hashable (#10567) --- comfy_execution/caching.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy_execution/caching.py b/comfy_execution/caching.py index b498f43e7..e077f78b0 100644 --- a/comfy_execution/caching.py +++ b/comfy_execution/caching.py @@ -53,7 +53,7 @@ class Unhashable: def to_hashable(obj): # So that we don't infinitely recurse since frozenset and tuples # are Sequences. - if isinstance(obj, (int, float, str, bool, type(None))): + if isinstance(obj, (int, float, str, bool, bytes, type(None))): return obj elif isinstance(obj, Mapping): return frozenset([(to_hashable(k), to_hashable(v)) for k, v in sorted(obj.items())]) From 1f3f7a2823017ad193e060d9221fb6a52f2eba3a Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Mon, 3 Nov 2025 10:21:47 +0200 Subject: [PATCH 070/647] convert nodes_hypernetwork.py to V3 schema (#10583) --- comfy_extras/nodes_hypernetwork.py | 48 ++++++++++++++++++++---------- 1 file changed, 33 insertions(+), 15 deletions(-) diff --git a/comfy_extras/nodes_hypernetwork.py b/comfy_extras/nodes_hypernetwork.py index 665632292..2a6a87a81 100644 --- a/comfy_extras/nodes_hypernetwork.py +++ b/comfy_extras/nodes_hypernetwork.py @@ -2,6 +2,9 @@ import comfy.utils import folder_paths import torch import logging +from comfy_api.latest import IO, ComfyExtension +from typing_extensions import override + def load_hypernetwork_patch(path, strength): sd = comfy.utils.load_torch_file(path, safe_load=True) @@ -94,27 +97,42 @@ def load_hypernetwork_patch(path, strength): return hypernetwork_patch(out, strength) -class HypernetworkLoader: +class HypernetworkLoader(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": { "model": ("MODEL",), - "hypernetwork_name": (folder_paths.get_filename_list("hypernetworks"), ), - "strength": ("FLOAT", {"default": 1.0, "min": -10.0, "max": 10.0, "step": 0.01}), - }} - RETURN_TYPES = ("MODEL",) - FUNCTION = "load_hypernetwork" + def define_schema(cls): + return IO.Schema( + node_id="HypernetworkLoader", + category="loaders", + inputs=[ + IO.Model.Input("model"), + IO.Combo.Input("hypernetwork_name", options=folder_paths.get_filename_list("hypernetworks")), + IO.Float.Input("strength", default=1.0, min=-10.0, max=10.0, step=0.01), + ], + outputs=[ + IO.Model.Output(), + ], + ) - CATEGORY = "loaders" - - def load_hypernetwork(self, model, hypernetwork_name, strength): + @classmethod + def execute(cls, model, hypernetwork_name, strength) -> IO.NodeOutput: hypernetwork_path = folder_paths.get_full_path_or_raise("hypernetworks", hypernetwork_name) model_hypernetwork = model.clone() patch = load_hypernetwork_patch(hypernetwork_path, strength) if patch is not None: model_hypernetwork.set_model_attn1_patch(patch) model_hypernetwork.set_model_attn2_patch(patch) - return (model_hypernetwork,) + return IO.NodeOutput(model_hypernetwork) -NODE_CLASS_MAPPINGS = { - "HypernetworkLoader": HypernetworkLoader -} + load_hypernetwork = execute # TODO: remove + + +class HyperNetworkExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[IO.ComfyNode]]: + return [ + HypernetworkLoader, + ] + + +async def comfy_entrypoint() -> HyperNetworkExtension: + return HyperNetworkExtension() From e617cddf244e4b789afba4b4ece01661b12fdde5 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Mon, 3 Nov 2025 10:28:13 +0200 Subject: [PATCH 071/647] convert nodes_openai.py to V3 schema (#10604) --- comfy_api_nodes/apinode_utils.py | 73 -- comfy_api_nodes/nodes_openai.py | 1094 +++++++++++---------------- comfy_api_nodes/util/__init__.py | 4 + comfy_api_nodes/util/conversions.py | 19 +- 4 files changed, 482 insertions(+), 708 deletions(-) delete mode 100644 comfy_api_nodes/apinode_utils.py diff --git a/comfy_api_nodes/apinode_utils.py b/comfy_api_nodes/apinode_utils.py deleted file mode 100644 index ecd604ff8..000000000 --- a/comfy_api_nodes/apinode_utils.py +++ /dev/null @@ -1,73 +0,0 @@ -from __future__ import annotations -import aiohttp -import mimetypes -from typing import Union -from server import PromptServer - -import numpy as np -from PIL import Image -import torch -import base64 -from io import BytesIO - - -async def validate_and_cast_response( - response, timeout: int = None, node_id: Union[str, None] = None -) -> torch.Tensor: - """Validates and casts a response to a torch.Tensor. - - Args: - response: The response to validate and cast. - timeout: Request timeout in seconds. Defaults to None (no timeout). - - Returns: - A torch.Tensor representing the image (1, H, W, C). - - Raises: - ValueError: If the response is not valid. - """ - # validate raw JSON response - data = response.data - if not data or len(data) == 0: - raise ValueError("No images returned from API endpoint") - - # Initialize list to store image tensors - image_tensors: list[torch.Tensor] = [] - - # Process each image in the data array - async with aiohttp.ClientSession(timeout=aiohttp.ClientTimeout(total=timeout)) as session: - for img_data in data: - img_bytes: bytes - if img_data.b64_json: - img_bytes = base64.b64decode(img_data.b64_json) - elif img_data.url: - if node_id: - PromptServer.instance.send_progress_text(f"Result URL: {img_data.url}", node_id) - async with session.get(img_data.url) as resp: - if resp.status != 200: - raise ValueError("Failed to download generated image") - img_bytes = await resp.read() - else: - raise ValueError("Invalid image payload – neither URL nor base64 data present.") - - pil_img = Image.open(BytesIO(img_bytes)).convert("RGBA") - arr = np.asarray(pil_img).astype(np.float32) / 255.0 - image_tensors.append(torch.from_numpy(arr)) - - return torch.stack(image_tensors, dim=0) - - -def text_filepath_to_base64_string(filepath: str) -> str: - """Converts a text file to a base64 string.""" - with open(filepath, "rb") as f: - file_content = f.read() - return base64.b64encode(file_content).decode("utf-8") - - -def text_filepath_to_data_uri(filepath: str) -> str: - """Converts a text file to a data URI.""" - base64_string = text_filepath_to_base64_string(filepath) - mime_type, _ = mimetypes.guess_type(filepath) - if mime_type is None: - mime_type = "application/octet-stream" - return f"data:{mime_type};base64,{base64_string}" diff --git a/comfy_api_nodes/nodes_openai.py b/comfy_api_nodes/nodes_openai.py index b4568fc85..acf35d276 100644 --- a/comfy_api_nodes/nodes_openai.py +++ b/comfy_api_nodes/nodes_openai.py @@ -1,18 +1,19 @@ -import io -from typing import TypedDict, Optional +from io import BytesIO +from typing import Optional, Union import json import os import time -import re import uuid from enum import Enum from inspect import cleandoc import numpy as np import torch from PIL import Image -from comfy.comfy_types.node_typing import IO, ComfyNodeABC, InputTypeDict from server import PromptServer import folder_paths +import base64 +from comfy_api.latest import IO, ComfyExtension +from typing_extensions import override from comfy_api_nodes.apis import ( @@ -23,7 +24,6 @@ from comfy_api_nodes.apis import ( OpenAIResponse, CreateModelResponseProperties, Item, - Includable, OutputContent, InputImageContent, Detail, @@ -34,41 +34,22 @@ from comfy_api_nodes.apis import ( InputFileContent, ) -from comfy_api_nodes.apis.client import ( +from comfy_api_nodes.util import ( + downscale_image_tensor, + download_url_to_bytesio, + validate_string, + tensor_to_base64_string, ApiEndpoint, - HttpMethod, - SynchronousOperation, - PollingOperation, - EmptyRequest, -) - -from comfy_api_nodes.apinode_utils import ( - validate_and_cast_response, + sync_op, + poll_op, text_filepath_to_data_uri, ) -from comfy_api_nodes.mapper_utils import model_field_to_node_input -from comfy_api_nodes.util import downscale_image_tensor, validate_string, tensor_to_base64_string RESPONSES_ENDPOINT = "/proxy/openai/v1/responses" STARTING_POINT_ID_PATTERN = r"" -class HistoryEntry(TypedDict): - """Type definition for a single history entry in the chat.""" - - prompt: str - response: str - response_id: str - timestamp: float - - -class ChatHistory(TypedDict): - """Type definition for the chat history dictionary.""" - - __annotations__: dict[str, list[HistoryEntry]] - - class SupportedOpenAIModel(str, Enum): o4_mini = "o4-mini" o1 = "o1" @@ -83,98 +64,123 @@ class SupportedOpenAIModel(str, Enum): gpt_5_nano = "gpt-5-nano" -class OpenAIDalle2(ComfyNodeABC): +async def validate_and_cast_response(response, timeout: int = None) -> torch.Tensor: + """Validates and casts a response to a torch.Tensor. + + Args: + response: The response to validate and cast. + timeout: Request timeout in seconds. Defaults to None (no timeout). + + Returns: + A torch.Tensor representing the image (1, H, W, C). + + Raises: + ValueError: If the response is not valid. + """ + # validate raw JSON response + data = response.data + if not data or len(data) == 0: + raise ValueError("No images returned from API endpoint") + + # Initialize list to store image tensors + image_tensors: list[torch.Tensor] = [] + + # Process each image in the data array + for img_data in data: + if img_data.b64_json: + img_io = BytesIO(base64.b64decode(img_data.b64_json)) + elif img_data.url: + img_io = BytesIO() + await download_url_to_bytesio(img_data.url, img_io, timeout=timeout) + else: + raise ValueError("Invalid image payload – neither URL nor base64 data present.") + + pil_img = Image.open(img_io).convert("RGBA") + arr = np.asarray(pil_img).astype(np.float32) / 255.0 + image_tensors.append(torch.from_numpy(arr)) + + return torch.stack(image_tensors, dim=0) + + +class OpenAIDalle2(IO.ComfyNode): """ Generates images synchronously via OpenAI's DALL·E 2 endpoint. """ - def __init__(self): - pass + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="OpenAIDalle2", + display_name="OpenAI DALL·E 2", + category="api node/image/OpenAI", + description=cleandoc(cls.__doc__ or ""), + inputs=[ + IO.String.Input( + "prompt", + default="", + multiline=True, + tooltip="Text prompt for DALL·E", + ), + IO.Int.Input( + "seed", + default=0, + min=0, + max=2**31 - 1, + step=1, + display_mode=IO.NumberDisplay.number, + control_after_generate=True, + tooltip="not implemented yet in backend", + optional=True, + ), + IO.Combo.Input( + "size", + default="1024x1024", + options=["256x256", "512x512", "1024x1024"], + tooltip="Image size", + optional=True, + ), + IO.Int.Input( + "n", + default=1, + min=1, + max=8, + step=1, + tooltip="How many images to generate", + display_mode=IO.NumberDisplay.number, + optional=True, + ), + IO.Image.Input( + "image", + tooltip="Optional reference image for image editing.", + optional=True, + ), + IO.Mask.Input( + "mask", + tooltip="Optional mask for inpainting (white areas will be replaced)", + optional=True, + ), + ], + outputs=[ + IO.Image.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + ) @classmethod - def INPUT_TYPES(cls) -> InputTypeDict: - return { - "required": { - "prompt": ( - IO.STRING, - { - "multiline": True, - "default": "", - "tooltip": "Text prompt for DALL·E", - }, - ), - }, - "optional": { - "seed": ( - IO.INT, - { - "default": 0, - "min": 0, - "max": 2**31 - 1, - "step": 1, - "display": "number", - "control_after_generate": True, - "tooltip": "not implemented yet in backend", - }, - ), - "size": ( - IO.COMBO, - { - "options": ["256x256", "512x512", "1024x1024"], - "default": "1024x1024", - "tooltip": "Image size", - }, - ), - "n": ( - IO.INT, - { - "default": 1, - "min": 1, - "max": 8, - "step": 1, - "display": "number", - "tooltip": "How many images to generate", - }, - ), - "image": ( - IO.IMAGE, - { - "default": None, - "tooltip": "Optional reference image for image editing.", - }, - ), - "mask": ( - IO.MASK, - { - "default": None, - "tooltip": "Optional mask for inpainting (white areas will be replaced)", - }, - ), - }, - "hidden": { - "auth_token": "AUTH_TOKEN_COMFY_ORG", - "comfy_api_key": "API_KEY_COMFY_ORG", - "unique_id": "UNIQUE_ID", - }, - } - - RETURN_TYPES = (IO.IMAGE,) - FUNCTION = "api_call" - CATEGORY = "api node/image/OpenAI" - DESCRIPTION = cleandoc(__doc__ or "") - API_NODE = True - - async def api_call( - self, + async def execute( + cls, prompt, seed=0, image=None, mask=None, n=1, size="1024x1024", - unique_id=None, - **kwargs, - ): + ) -> IO.NodeOutput: validate_string(prompt, strip_whitespace=False) model = "dall-e-2" path = "/proxy/openai/images/generations" @@ -200,7 +206,7 @@ class OpenAIDalle2(ComfyNodeABC): image_np = (rgba_tensor.numpy() * 255).astype(np.uint8) img = Image.fromarray(image_np) - img_byte_arr = io.BytesIO() + img_byte_arr = BytesIO() img.save(img_byte_arr, format="PNG") img_byte_arr.seek(0) img_binary = img_byte_arr # .getvalue() @@ -208,15 +214,11 @@ class OpenAIDalle2(ComfyNodeABC): elif image is not None or mask is not None: raise Exception("Dall-E 2 image editing requires an image AND a mask") - # Build the operation - operation = SynchronousOperation( - endpoint=ApiEndpoint( - path=path, - method=HttpMethod.POST, - request_model=request_class, - response_model=OpenAIImageGenerationResponse, - ), - request=request_class( + response = await sync_op( + cls, + ApiEndpoint(path=path, method="POST"), + response_model=OpenAIImageGenerationResponse, + data=request_class( model=model, prompt=prompt, n=n, @@ -231,109 +233,92 @@ class OpenAIDalle2(ComfyNodeABC): else None ), content_type=content_type, - auth_kwargs=kwargs, ) - response = await operation.execute() - - img_tensor = await validate_and_cast_response(response, node_id=unique_id) - return (img_tensor,) + return IO.NodeOutput(await validate_and_cast_response(response)) -class OpenAIDalle3(ComfyNodeABC): +class OpenAIDalle3(IO.ComfyNode): """ Generates images synchronously via OpenAI's DALL·E 3 endpoint. """ - def __init__(self): - pass + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="OpenAIDalle3", + display_name="OpenAI DALL·E 3", + category="api node/image/OpenAI", + description=cleandoc(cls.__doc__ or ""), + inputs=[ + IO.String.Input( + "prompt", + default="", + multiline=True, + tooltip="Text prompt for DALL·E", + ), + IO.Int.Input( + "seed", + default=0, + min=0, + max=2 ** 31 - 1, + step=1, + display_mode=IO.NumberDisplay.number, + control_after_generate=True, + tooltip="not implemented yet in backend", + optional=True, + ), + IO.Combo.Input( + "quality", + default="standard", + options=["standard", "hd"], + tooltip="Image quality", + optional=True, + ), + IO.Combo.Input( + "style", + default="natural", + options=["natural", "vivid"], + tooltip="Vivid causes the model to lean towards generating hyper-real and dramatic images. Natural causes the model to produce more natural, less hyper-real looking images.", + optional=True, + ), + IO.Combo.Input( + "size", + default="1024x1024", + options=["1024x1024", "1024x1792", "1792x1024"], + tooltip="Image size", + optional=True, + ), + ], + outputs=[ + IO.Image.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + ) @classmethod - def INPUT_TYPES(cls) -> InputTypeDict: - return { - "required": { - "prompt": ( - IO.STRING, - { - "multiline": True, - "default": "", - "tooltip": "Text prompt for DALL·E", - }, - ), - }, - "optional": { - "seed": ( - IO.INT, - { - "default": 0, - "min": 0, - "max": 2**31 - 1, - "step": 1, - "display": "number", - "control_after_generate": True, - "tooltip": "not implemented yet in backend", - }, - ), - "quality": ( - IO.COMBO, - { - "options": ["standard", "hd"], - "default": "standard", - "tooltip": "Image quality", - }, - ), - "style": ( - IO.COMBO, - { - "options": ["natural", "vivid"], - "default": "natural", - "tooltip": "Vivid causes the model to lean towards generating hyper-real and dramatic images. Natural causes the model to produce more natural, less hyper-real looking images.", - }, - ), - "size": ( - IO.COMBO, - { - "options": ["1024x1024", "1024x1792", "1792x1024"], - "default": "1024x1024", - "tooltip": "Image size", - }, - ), - }, - "hidden": { - "auth_token": "AUTH_TOKEN_COMFY_ORG", - "comfy_api_key": "API_KEY_COMFY_ORG", - "unique_id": "UNIQUE_ID", - }, - } - - RETURN_TYPES = (IO.IMAGE,) - FUNCTION = "api_call" - CATEGORY = "api node/image/OpenAI" - DESCRIPTION = cleandoc(__doc__ or "") - API_NODE = True - - async def api_call( - self, + async def execute( + cls, prompt, seed=0, style="natural", quality="standard", size="1024x1024", - unique_id=None, - **kwargs, - ): + ) -> IO.NodeOutput: validate_string(prompt, strip_whitespace=False) model = "dall-e-3" # build the operation - operation = SynchronousOperation( - endpoint=ApiEndpoint( - path="/proxy/openai/images/generations", - method=HttpMethod.POST, - request_model=OpenAIImageGenerationRequest, - response_model=OpenAIImageGenerationResponse, - ), - request=OpenAIImageGenerationRequest( + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/openai/images/generations", method="POST"), + response_model=OpenAIImageGenerationResponse, + data=OpenAIImageGenerationRequest( model=model, prompt=prompt, quality=quality, @@ -341,114 +326,97 @@ class OpenAIDalle3(ComfyNodeABC): style=style, seed=seed, ), - auth_kwargs=kwargs, ) - response = await operation.execute() - - img_tensor = await validate_and_cast_response(response, node_id=unique_id) - return (img_tensor,) + return IO.NodeOutput(await validate_and_cast_response(response)) -class OpenAIGPTImage1(ComfyNodeABC): +class OpenAIGPTImage1(IO.ComfyNode): """ Generates images synchronously via OpenAI's GPT Image 1 endpoint. """ - def __init__(self): - pass + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="OpenAIGPTImage1", + display_name="OpenAI GPT Image 1", + category="api node/image/OpenAI", + description=cleandoc(cls.__doc__ or ""), + inputs=[ + IO.String.Input( + "prompt", + default="", + multiline=True, + tooltip="Text prompt for GPT Image 1", + ), + IO.Int.Input( + "seed", + default=0, + min=0, + max=2 ** 31 - 1, + step=1, + display_mode=IO.NumberDisplay.number, + control_after_generate=True, + tooltip="not implemented yet in backend", + optional=True, + ), + IO.Combo.Input( + "quality", + default="low", + options=["low", "medium", "high"], + tooltip="Image quality, affects cost and generation time.", + optional=True, + ), + IO.Combo.Input( + "background", + default="opaque", + options=["opaque", "transparent"], + tooltip="Return image with or without background", + optional=True, + ), + IO.Combo.Input( + "size", + default="auto", + options=["auto", "1024x1024", "1024x1536", "1536x1024"], + tooltip="Image size", + optional=True, + ), + IO.Int.Input( + "n", + default=1, + min=1, + max=8, + step=1, + tooltip="How many images to generate", + display_mode=IO.NumberDisplay.number, + optional=True, + ), + IO.Image.Input( + "image", + tooltip="Optional reference image for image editing.", + optional=True, + ), + IO.Mask.Input( + "mask", + tooltip="Optional mask for inpainting (white areas will be replaced)", + optional=True, + ), + ], + outputs=[ + IO.Image.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + ) @classmethod - def INPUT_TYPES(cls) -> InputTypeDict: - return { - "required": { - "prompt": ( - IO.STRING, - { - "multiline": True, - "default": "", - "tooltip": "Text prompt for GPT Image 1", - }, - ), - }, - "optional": { - "seed": ( - IO.INT, - { - "default": 0, - "min": 0, - "max": 2**31 - 1, - "step": 1, - "display": "number", - "control_after_generate": True, - "tooltip": "not implemented yet in backend", - }, - ), - "quality": ( - IO.COMBO, - { - "options": ["low", "medium", "high"], - "default": "low", - "tooltip": "Image quality, affects cost and generation time.", - }, - ), - "background": ( - IO.COMBO, - { - "options": ["opaque", "transparent"], - "default": "opaque", - "tooltip": "Return image with or without background", - }, - ), - "size": ( - IO.COMBO, - { - "options": ["auto", "1024x1024", "1024x1536", "1536x1024"], - "default": "auto", - "tooltip": "Image size", - }, - ), - "n": ( - IO.INT, - { - "default": 1, - "min": 1, - "max": 8, - "step": 1, - "display": "number", - "tooltip": "How many images to generate", - }, - ), - "image": ( - IO.IMAGE, - { - "default": None, - "tooltip": "Optional reference image for image editing.", - }, - ), - "mask": ( - IO.MASK, - { - "default": None, - "tooltip": "Optional mask for inpainting (white areas will be replaced)", - }, - ), - }, - "hidden": { - "auth_token": "AUTH_TOKEN_COMFY_ORG", - "comfy_api_key": "API_KEY_COMFY_ORG", - "unique_id": "UNIQUE_ID", - }, - } - - RETURN_TYPES = (IO.IMAGE,) - FUNCTION = "api_call" - CATEGORY = "api node/image/OpenAI" - DESCRIPTION = cleandoc(__doc__ or "") - API_NODE = True - - async def api_call( - self, + async def execute( + cls, prompt, seed=0, quality="low", @@ -457,9 +425,7 @@ class OpenAIGPTImage1(ComfyNodeABC): mask=None, n=1, size="1024x1024", - unique_id=None, - **kwargs, - ): + ) -> IO.NodeOutput: validate_string(prompt, strip_whitespace=False) model = "gpt-image-1" path = "/proxy/openai/images/generations" @@ -480,7 +446,7 @@ class OpenAIGPTImage1(ComfyNodeABC): image_np = (scaled_image.numpy() * 255).astype(np.uint8) img = Image.fromarray(image_np) - img_byte_arr = io.BytesIO() + img_byte_arr = BytesIO() img.save(img_byte_arr, format="PNG") img_byte_arr.seek(0) @@ -504,20 +470,17 @@ class OpenAIGPTImage1(ComfyNodeABC): mask_np = (scaled_mask.numpy() * 255).astype(np.uint8) mask_img = Image.fromarray(mask_np) - mask_img_byte_arr = io.BytesIO() + mask_img_byte_arr = BytesIO() mask_img.save(mask_img_byte_arr, format="PNG") mask_img_byte_arr.seek(0) files.append(("mask", ("mask.png", mask_img_byte_arr, "image/png"))) # Build the operation - operation = SynchronousOperation( - endpoint=ApiEndpoint( - path=path, - method=HttpMethod.POST, - request_model=request_class, - response_model=OpenAIImageGenerationResponse, - ), - request=request_class( + response = await sync_op( + cls, + ApiEndpoint(path=path, method="POST"), + response_model=OpenAIImageGenerationResponse, + data=request_class( model=model, prompt=prompt, quality=quality, @@ -528,127 +491,70 @@ class OpenAIGPTImage1(ComfyNodeABC): ), files=files if files else None, content_type=content_type, - auth_kwargs=kwargs, ) - response = await operation.execute() - - img_tensor = await validate_and_cast_response(response, node_id=unique_id) - return (img_tensor,) + return IO.NodeOutput(await validate_and_cast_response(response)) -class OpenAITextNode(ComfyNodeABC): - """ - Base class for OpenAI text generation nodes. - """ - - RETURN_TYPES = (IO.STRING,) - FUNCTION = "api_call" - CATEGORY = "api node/text/OpenAI" - API_NODE = True - - -class OpenAIChatNode(OpenAITextNode): +class OpenAIChatNode(IO.ComfyNode): """ Node to generate text responses from an OpenAI model. """ - def __init__(self) -> None: - """Initialize the chat node with a new session ID and empty history.""" - self.current_session_id: str = str(uuid.uuid4()) - self.history: dict[str, list[HistoryEntry]] = {} - self.previous_response_id: Optional[str] = None + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="OpenAIChatNode", + display_name="OpenAI ChatGPT", + category="api node/text/OpenAI", + description="Generate text responses from an OpenAI model.", + inputs=[ + IO.String.Input( + "prompt", + default="", + multiline=True, + tooltip="Text inputs to the model, used to generate a response.", + ), + IO.Boolean.Input( + "persist_context", + default=False, + tooltip="This parameter is deprecated and has no effect.", + ), + IO.Combo.Input( + "model", + options=SupportedOpenAIModel, + tooltip="The model used to generate the response", + ), + IO.Image.Input( + "images", + tooltip="Optional image(s) to use as context for the model. To include multiple images, you can use the Batch Images node.", + optional=True, + ), + IO.Custom("OPENAI_INPUT_FILES").Input( + "files", + optional=True, + tooltip="Optional file(s) to use as context for the model. Accepts inputs from the OpenAI Chat Input Files node.", + ), + IO.Custom("OPENAI_CHAT_CONFIG").Input( + "advanced_options", + optional=True, + tooltip="Optional configuration for the model. Accepts inputs from the OpenAI Chat Advanced Options node.", + ), + ], + outputs=[ + IO.String.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + ) @classmethod - def INPUT_TYPES(cls) -> InputTypeDict: - return { - "required": { - "prompt": ( - IO.STRING, - { - "multiline": True, - "default": "", - "tooltip": "Text inputs to the model, used to generate a response.", - }, - ), - "persist_context": ( - IO.BOOLEAN, - { - "default": True, - "tooltip": "Persist chat context between calls (multi-turn conversation)", - }, - ), - "model": model_field_to_node_input( - IO.COMBO, - OpenAICreateResponse, - "model", - enum_type=SupportedOpenAIModel, - ), - }, - "optional": { - "images": ( - IO.IMAGE, - { - "default": None, - "tooltip": "Optional image(s) to use as context for the model. To include multiple images, you can use the Batch Images node.", - }, - ), - "files": ( - "OPENAI_INPUT_FILES", - { - "default": None, - "tooltip": "Optional file(s) to use as context for the model. Accepts inputs from the OpenAI Chat Input Files node.", - }, - ), - "advanced_options": ( - "OPENAI_CHAT_CONFIG", - { - "default": None, - "tooltip": "Optional configuration for the model. Accepts inputs from the OpenAI Chat Advanced Options node.", - }, - ), - }, - "hidden": { - "auth_token": "AUTH_TOKEN_COMFY_ORG", - "comfy_api_key": "API_KEY_COMFY_ORG", - "unique_id": "UNIQUE_ID", - }, - } - - DESCRIPTION = "Generate text responses from an OpenAI model." - - async def get_result_response( - self, - response_id: str, - include: Optional[list[Includable]] = None, - auth_kwargs: Optional[dict[str, str]] = None, - ) -> OpenAIResponse: - """ - Retrieve a model response with the given ID from the OpenAI API. - - Args: - response_id (str): The ID of the response to retrieve. - include (Optional[List[Includable]]): Additional fields to include - in the response. See the `include` parameter for Response - creation above for more information. - - """ - return await PollingOperation( - poll_endpoint=ApiEndpoint( - path=f"{RESPONSES_ENDPOINT}/{response_id}", - method=HttpMethod.GET, - request_model=EmptyRequest, - response_model=OpenAIResponse, - query_params={"include": include}, - ), - completed_statuses=["completed"], - failed_statuses=["failed"], - status_extractor=lambda response: response.status, - auth_kwargs=auth_kwargs, - ).execute() - def get_message_content_from_response( - self, response: OpenAIResponse + cls, response: OpenAIResponse ) -> list[OutputContent]: """Extract message content from the API response.""" for output in response.output: @@ -656,8 +562,9 @@ class OpenAIChatNode(OpenAITextNode): return output.root.content raise TypeError("No output message found in response") + @classmethod def get_text_from_message_content( - self, message_content: list[OutputContent] + cls, message_content: list[OutputContent] ) -> str: """Extract text content from message content.""" for content_item in message_content: @@ -665,58 +572,9 @@ class OpenAIChatNode(OpenAITextNode): return str(content_item.root.text) return "No text output found in response" - def get_history_text(self, session_id: str) -> str: - """Convert the entire history for a given session to JSON string.""" - return json.dumps(self.history[session_id]) - - def display_history_on_node(self, session_id: str, node_id: str) -> None: - """Display formatted chat history on the node UI.""" - render_spec = { - "node_id": node_id, - "component": "ChatHistoryWidget", - "props": { - "history": self.get_history_text(session_id), - }, - } - PromptServer.instance.send_sync( - "display_component", - render_spec, - ) - - def add_to_history( - self, session_id: str, prompt: str, output_text: str, response_id: str - ) -> None: - """Add a new entry to the chat history.""" - if session_id not in self.history: - self.history[session_id] = [] - self.history[session_id].append( - { - "prompt": prompt, - "response": output_text, - "response_id": response_id, - "timestamp": time.time(), - } - ) - - def parse_output_text_from_response(self, response: OpenAIResponse) -> str: - """Extract text output from the API response.""" - message_contents = self.get_message_content_from_response(response) - return self.get_text_from_message_content(message_contents) - - def generate_new_session_id(self) -> str: - """Generate a new unique session ID.""" - return str(uuid.uuid4()) - - def get_session_id(self, persist_context: bool) -> str: - """Get the current or generate a new session ID based on context persistence.""" - return ( - self.current_session_id - if persist_context - else self.generate_new_session_id() - ) - + @classmethod def tensor_to_input_image_content( - self, image: torch.Tensor, detail_level: Detail = "auto" + cls, image: torch.Tensor, detail_level: Detail = "auto" ) -> InputImageContent: """Convert a tensor to an input image content object.""" return InputImageContent( @@ -725,21 +583,27 @@ class OpenAIChatNode(OpenAITextNode): type="input_image", ) + @classmethod def create_input_message_contents( - self, + cls, prompt: str, image: Optional[torch.Tensor] = None, files: Optional[list[InputFileContent]] = None, ) -> InputMessageContentList: """Create a list of input message contents from prompt and optional image.""" - content_list: list[InputContent] = [ + content_list: list[Union[InputContent, InputTextContent, InputImageContent, InputFileContent]] = [ InputTextContent(text=prompt, type="input_text"), ] if image is not None: for i in range(image.shape[0]): content_list.append( - self.tensor_to_input_image_content(image[i].unsqueeze(0)) + InputImageContent( + detail="auto", + image_url=f"data:image/png;base64,{tensor_to_base64_string(image[i].unsqueeze(0))}", + type="input_image", + ) ) + if files is not None: content_list.extend(files) @@ -747,80 +611,28 @@ class OpenAIChatNode(OpenAITextNode): root=content_list, ) - def parse_response_id_from_prompt(self, prompt: str) -> Optional[str]: - """Extract response ID from prompt if it exists.""" - parsed_id = re.search(STARTING_POINT_ID_PATTERN, prompt) - return parsed_id.group(1) if parsed_id else None - - def strip_response_tag_from_prompt(self, prompt: str) -> str: - """Remove the response ID tag from the prompt.""" - return re.sub(STARTING_POINT_ID_PATTERN, "", prompt.strip()) - - def delete_history_after_response_id( - self, new_start_id: str, session_id: str - ) -> None: - """Delete history entries after a specific response ID.""" - if session_id not in self.history: - return - - new_history = [] - i = 0 - while ( - i < len(self.history[session_id]) - and self.history[session_id][i]["response_id"] != new_start_id - ): - new_history.append(self.history[session_id][i]) - i += 1 - - # Since it's the new starting point (not the response being edited), we include it as well - if i < len(self.history[session_id]): - new_history.append(self.history[session_id][i]) - - self.history[session_id] = new_history - - async def api_call( - self, + @classmethod + async def execute( + cls, prompt: str, - persist_context: bool, - model: SupportedOpenAIModel, - unique_id: Optional[str] = None, + persist_context: bool = False, + model: SupportedOpenAIModel = SupportedOpenAIModel.gpt_5.value, images: Optional[torch.Tensor] = None, files: Optional[list[InputFileContent]] = None, advanced_options: Optional[CreateModelResponseProperties] = None, - **kwargs, - ) -> tuple[str]: - # Validate inputs + ) -> IO.NodeOutput: validate_string(prompt, strip_whitespace=False) - session_id = self.get_session_id(persist_context) - response_id_override = self.parse_response_id_from_prompt(prompt) - if response_id_override: - is_starting_from_beginning = response_id_override == "start" - if is_starting_from_beginning: - self.history[session_id] = [] - previous_response_id = None - else: - previous_response_id = response_id_override - self.delete_history_after_response_id(response_id_override, session_id) - prompt = self.strip_response_tag_from_prompt(prompt) - elif persist_context: - previous_response_id = self.previous_response_id - else: - previous_response_id = None - # Create response - create_response = await SynchronousOperation( - endpoint=ApiEndpoint( - path=RESPONSES_ENDPOINT, - method=HttpMethod.POST, - request_model=OpenAICreateResponse, - response_model=OpenAIResponse, - ), - request=OpenAICreateResponse( + create_response = await sync_op( + cls, + ApiEndpoint(path=RESPONSES_ENDPOINT, method="POST"), + response_model=OpenAIResponse, + data=OpenAICreateResponse( input=[ Item( root=InputMessage( - content=self.create_input_message_contents( + content=cls.create_input_message_contents( prompt, images, files ), role="user", @@ -830,36 +642,57 @@ class OpenAIChatNode(OpenAITextNode): store=True, stream=False, model=model, - previous_response_id=previous_response_id, + previous_response_id=None, **( advanced_options.model_dump(exclude_none=True) if advanced_options else {} ), ), - auth_kwargs=kwargs, - ).execute() + ) response_id = create_response.id # Get result output - result_response = await self.get_result_response(response_id, auth_kwargs=kwargs) - output_text = self.parse_output_text_from_response(result_response) + result_response = await poll_op( + cls, + ApiEndpoint(path=f"{RESPONSES_ENDPOINT}/{response_id}"), + response_model=OpenAIResponse, + status_extractor=lambda response: response.status, + completed_statuses=["incomplete", "completed"] + ) + output_text = cls.get_text_from_message_content(cls.get_message_content_from_response(result_response)) # Update history - self.add_to_history(session_id, prompt, output_text, response_id) - self.display_history_on_node(session_id, unique_id) - self.previous_response_id = response_id - - return (output_text,) + render_spec = { + "node_id": cls.hidden.unique_id, + "component": "ChatHistoryWidget", + "props": { + "history": json.dumps( + [ + { + "prompt": prompt, + "response": output_text, + "response_id": str(uuid.uuid4()), + "timestamp": time.time(), + } + ] + ), + }, + } + PromptServer.instance.send_sync( + "display_component", + render_spec, + ) + return IO.NodeOutput(output_text) -class OpenAIInputFiles(ComfyNodeABC): +class OpenAIInputFiles(IO.ComfyNode): """ Loads and formats input files for OpenAI API. """ @classmethod - def INPUT_TYPES(cls) -> InputTypeDict: + def define_schema(cls): """ For details about the supported file input types, see: https://platform.openai.com/docs/guides/pdf-files?api-mode=responses @@ -874,97 +707,92 @@ class OpenAIInputFiles(ComfyNodeABC): ] input_files = sorted(input_files, key=lambda x: x.name) input_files = [f.name for f in input_files] - return { - "required": { - "file": ( - IO.COMBO, - { - "tooltip": "Input files to include as context for the model. Only accepts text (.txt) and PDF (.pdf) files for now.", - "options": input_files, - "default": input_files[0] if input_files else None, - }, + return IO.Schema( + node_id="OpenAIInputFiles", + display_name="OpenAI ChatGPT Input Files", + category="api node/text/OpenAI", + description="Loads and prepares input files (text, pdf, etc.) to include as inputs for the OpenAI Chat Node. The files will be read by the OpenAI model when generating a response. 🛈 TIP: Can be chained together with other OpenAI Input File nodes.", + inputs=[ + IO.Combo.Input( + "file", + options=input_files, + default=input_files[0] if input_files else None, + tooltip="Input files to include as context for the model. Only accepts text (.txt) and PDF (.pdf) files for now.", ), - }, - "optional": { - "OPENAI_INPUT_FILES": ( + IO.Custom("OPENAI_INPUT_FILES").Input( "OPENAI_INPUT_FILES", - { - "tooltip": "An optional additional file(s) to batch together with the file loaded from this node. Allows chaining of input files so that a single message can include multiple input files.", - "default": None, - }, + tooltip="An optional additional file(s) to batch together with the file loaded from this node. Allows chaining of input files so that a single message can include multiple input files.", + optional=True, ), - }, - } + ], + outputs=[ + IO.Custom("OPENAI_INPUT_FILES").Output(), + ], + ) - DESCRIPTION = "Loads and prepares input files (text, pdf, etc.) to include as inputs for the OpenAI Chat Node. The files will be read by the OpenAI model when generating a response. 🛈 TIP: Can be chained together with other OpenAI Input File nodes." - RETURN_TYPES = ("OPENAI_INPUT_FILES",) - FUNCTION = "prepare_files" - CATEGORY = "api node/text/OpenAI" - - def create_input_file_content(self, file_path: str) -> InputFileContent: + @classmethod + def create_input_file_content(cls, file_path: str) -> InputFileContent: return InputFileContent( file_data=text_filepath_to_data_uri(file_path), filename=os.path.basename(file_path), type="input_file", ) - def prepare_files( - self, file: str, OPENAI_INPUT_FILES: list[InputFileContent] = [] - ) -> tuple[list[InputFileContent]]: + @classmethod + def execute(cls, file: str, OPENAI_INPUT_FILES: list[InputFileContent] = []) -> IO.NodeOutput: """ Loads and formats input files for OpenAI API. """ file_path = folder_paths.get_annotated_filepath(file) - input_file_content = self.create_input_file_content(file_path) + input_file_content = cls.create_input_file_content(file_path) files = [input_file_content] + OPENAI_INPUT_FILES - return (files,) + return IO.NodeOutput(files) -class OpenAIChatConfig(ComfyNodeABC): +class OpenAIChatConfig(IO.ComfyNode): """Allows setting additional configuration for the OpenAI Chat Node.""" - RETURN_TYPES = ("OPENAI_CHAT_CONFIG",) - FUNCTION = "configure" - DESCRIPTION = ( - "Allows specifying advanced configuration options for the OpenAI Chat Nodes." - ) - CATEGORY = "api node/text/OpenAI" - @classmethod - def INPUT_TYPES(cls) -> InputTypeDict: - return { - "required": { - "truncation": ( - IO.COMBO, - { - "options": ["auto", "disabled"], - "default": "auto", - "tooltip": "The truncation strategy to use for the model response. auto: If the context of this response and previous ones exceeds the model's context window size, the model will truncate the response to fit the context window by dropping input items in the middle of the conversation.disabled: If a model response will exceed the context window size for a model, the request will fail with a 400 error", - }, + def define_schema(cls): + return IO.Schema( + node_id="OpenAIChatConfig", + display_name="OpenAI ChatGPT Advanced Options", + category="api node/text/OpenAI", + description="Allows specifying advanced configuration options for the OpenAI Chat Nodes.", + inputs=[ + IO.Combo.Input( + "truncation", + options=["auto", "disabled"], + default="auto", + tooltip="The truncation strategy to use for the model response. auto: If the context of this response and previous ones exceeds the model's context window size, the model will truncate the response to fit the context window by dropping input items in the middle of the conversation.disabled: If a model response will exceed the context window size for a model, the request will fail with a 400 error", ), - }, - "optional": { - "max_output_tokens": model_field_to_node_input( - IO.INT, - OpenAICreateResponse, + IO.Int.Input( "max_output_tokens", min=16, default=4096, max=16384, tooltip="An upper bound for the number of tokens that can be generated for a response, including visible output tokens", + optional=True, ), - "instructions": model_field_to_node_input( - IO.STRING, OpenAICreateResponse, "instructions", multiline=True + IO.String.Input( + "instructions", + multiline=True, + optional=True, + tooltip="Instructions for the model on how to generate the response", ), - }, - } + ], + outputs=[ + IO.Custom("OPENAI_CHAT_CONFIG").Output(), + ], + ) - def configure( - self, + @classmethod + def execute( + cls, truncation: bool, instructions: Optional[str] = None, max_output_tokens: Optional[int] = None, - ) -> tuple[CreateModelResponseProperties]: + ) -> IO.NodeOutput: """ Configure advanced options for the OpenAI Chat Node. @@ -974,29 +802,27 @@ class OpenAIChatConfig(ComfyNodeABC): They are not exposed as inputs at all to avoid having to manually remove depending on model choice. """ - return ( + return IO.NodeOutput( CreateModelResponseProperties( instructions=instructions, truncation=truncation, max_output_tokens=max_output_tokens, - ), + ) ) -NODE_CLASS_MAPPINGS = { - "OpenAIDalle2": OpenAIDalle2, - "OpenAIDalle3": OpenAIDalle3, - "OpenAIGPTImage1": OpenAIGPTImage1, - "OpenAIChatNode": OpenAIChatNode, - "OpenAIInputFiles": OpenAIInputFiles, - "OpenAIChatConfig": OpenAIChatConfig, -} +class OpenAIExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[IO.ComfyNode]]: + return [ + OpenAIDalle2, + OpenAIDalle3, + OpenAIGPTImage1, + OpenAIChatNode, + OpenAIInputFiles, + OpenAIChatConfig, + ] -NODE_DISPLAY_NAME_MAPPINGS = { - "OpenAIDalle2": "OpenAI DALL·E 2", - "OpenAIDalle3": "OpenAI DALL·E 3", - "OpenAIGPTImage1": "OpenAI GPT Image 1", - "OpenAIChatNode": "OpenAI ChatGPT", - "OpenAIInputFiles": "OpenAI ChatGPT Input Files", - "OpenAIChatConfig": "OpenAI ChatGPT Advanced Options", -} + +async def comfy_entrypoint() -> OpenAIExtension: + return OpenAIExtension() diff --git a/comfy_api_nodes/util/__init__.py b/comfy_api_nodes/util/__init__.py index bbc71363a..21013b591 100644 --- a/comfy_api_nodes/util/__init__.py +++ b/comfy_api_nodes/util/__init__.py @@ -18,6 +18,8 @@ from .conversions import ( tensor_to_base64_string, tensor_to_bytesio, tensor_to_pil, + text_filepath_to_base64_string, + text_filepath_to_data_uri, trim_video, video_to_base64_string, ) @@ -75,6 +77,8 @@ __all__ = [ "tensor_to_base64_string", "tensor_to_bytesio", "tensor_to_pil", + "text_filepath_to_base64_string", + "text_filepath_to_data_uri", "trim_video", "video_to_base64_string", # Validation utilities diff --git a/comfy_api_nodes/util/conversions.py b/comfy_api_nodes/util/conversions.py index b59c2bd84..971dc57de 100644 --- a/comfy_api_nodes/util/conversions.py +++ b/comfy_api_nodes/util/conversions.py @@ -1,6 +1,7 @@ import base64 import logging import math +import mimetypes import uuid from io import BytesIO from typing import Optional @@ -12,7 +13,7 @@ from PIL import Image from comfy.utils import common_upscale from comfy_api.latest import Input, InputImpl -from comfy_api.util import VideoContainer, VideoCodec +from comfy_api.util import VideoCodec, VideoContainer from ._helpers import mimetype_to_extension @@ -451,3 +452,19 @@ def resize_mask_to_image( if not allow_gradient: mask = (mask > 0.5).float() return mask + + +def text_filepath_to_base64_string(filepath: str) -> str: + """Converts a text file to a base64 string.""" + with open(filepath, "rb") as f: + file_content = f.read() + return base64.b64encode(file_content).decode("utf-8") + + +def text_filepath_to_data_uri(filepath: str) -> str: + """Converts a text file to a data URI.""" + base64_string = text_filepath_to_base64_string(filepath) + mime_type, _ = mimetypes.guess_type(filepath) + if mime_type is None: + mime_type = "application/octet-stream" + return f"data:{mime_type};base64,{base64_string}" From 4e2110c794b187c9326d604e7c0b0a4fad81148a Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Mon, 3 Nov 2025 10:29:08 +0200 Subject: [PATCH 072/647] feat(Pika-API-nodes): use new API client (#10608) --- .../apis/{pika_defs.py => pika_api.py} | 0 comfy_api_nodes/nodes_pika.py | 181 ++++++------------ 2 files changed, 57 insertions(+), 124 deletions(-) rename comfy_api_nodes/apis/{pika_defs.py => pika_api.py} (100%) diff --git a/comfy_api_nodes/apis/pika_defs.py b/comfy_api_nodes/apis/pika_api.py similarity index 100% rename from comfy_api_nodes/apis/pika_defs.py rename to comfy_api_nodes/apis/pika_api.py diff --git a/comfy_api_nodes/nodes_pika.py b/comfy_api_nodes/nodes_pika.py index 5bb406a3b..51148211b 100644 --- a/comfy_api_nodes/nodes_pika.py +++ b/comfy_api_nodes/nodes_pika.py @@ -7,24 +7,23 @@ from __future__ import annotations from io import BytesIO import logging -from typing import Optional, TypeVar +from typing import Optional import torch from typing_extensions import override from comfy_api.latest import ComfyExtension, IO from comfy_api.input_impl.video_types import VideoCodec, VideoContainer, VideoInput -from comfy_api_nodes.apis import pika_defs -from comfy_api_nodes.apis.client import ( +from comfy_api_nodes.apis import pika_api as pika_defs +from comfy_api_nodes.util import ( + validate_string, + download_url_to_video_output, + tensor_to_bytesio, ApiEndpoint, - EmptyRequest, - HttpMethod, - PollingOperation, - SynchronousOperation, + sync_op, + poll_op, ) -from comfy_api_nodes.util import validate_string, download_url_to_video_output, tensor_to_bytesio -R = TypeVar("R") PATH_PIKADDITIONS = "/proxy/pika/generate/pikadditions" PATH_PIKASWAPS = "/proxy/pika/generate/pikaswaps" @@ -40,28 +39,18 @@ PATH_VIDEO_GET = "/proxy/pika/videos" async def execute_task( - initial_operation: SynchronousOperation[R, pika_defs.PikaGenerateResponse], - auth_kwargs: Optional[dict[str, str]] = None, - node_id: Optional[str] = None, + task_id: str, + cls: type[IO.ComfyNode], ) -> IO.NodeOutput: - task_id = (await initial_operation.execute()).video_id - final_response: pika_defs.PikaVideoResponse = await PollingOperation( - poll_endpoint=ApiEndpoint( - path=f"{PATH_VIDEO_GET}/{task_id}", - method=HttpMethod.GET, - request_model=EmptyRequest, - response_model=pika_defs.PikaVideoResponse, - ), - completed_statuses=["finished"], - failed_statuses=["failed", "cancelled"], + final_response: pika_defs.PikaVideoResponse = await poll_op( + cls, + ApiEndpoint(path=f"{PATH_VIDEO_GET}/{task_id}"), + response_model=pika_defs.PikaVideoResponse, status_extractor=lambda response: (response.status.value if response.status else None), progress_extractor=lambda response: (response.progress if hasattr(response, "progress") else None), - auth_kwargs=auth_kwargs, - result_url_extractor=lambda response: (response.url if hasattr(response, "url") else None), - node_id=node_id, estimated_duration=60, max_poll_attempts=240, - ).execute() + ) if not final_response.url: error_msg = f"Pika task {task_id} succeeded but no video data found in response:\n{final_response}" logging.error(error_msg) @@ -124,23 +113,15 @@ class PikaImageToVideo(IO.ComfyNode): resolution=resolution, duration=duration, ) - auth = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } - initial_operation = SynchronousOperation( - endpoint=ApiEndpoint( - path=PATH_IMAGE_TO_VIDEO, - method=HttpMethod.POST, - request_model=pika_defs.PikaBodyGenerate22I2vGenerate22I2vPost, - response_model=pika_defs.PikaGenerateResponse, - ), - request=pika_request_data, + initial_operation = await sync_op( + cls, + ApiEndpoint(path=PATH_IMAGE_TO_VIDEO, method="POST"), + response_model=pika_defs.PikaGenerateResponse, + data=pika_request_data, files=pika_files, content_type="multipart/form-data", - auth_kwargs=auth, ) - return await execute_task(initial_operation, auth_kwargs=auth, node_id=cls.hidden.unique_id) + return await execute_task(initial_operation.video_id, cls) class PikaTextToVideoNode(IO.ComfyNode): @@ -183,18 +164,11 @@ class PikaTextToVideoNode(IO.ComfyNode): duration: int, aspect_ratio: float, ) -> IO.NodeOutput: - auth = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } - initial_operation = SynchronousOperation( - endpoint=ApiEndpoint( - path=PATH_TEXT_TO_VIDEO, - method=HttpMethod.POST, - request_model=pika_defs.PikaBodyGenerate22T2vGenerate22T2vPost, - response_model=pika_defs.PikaGenerateResponse, - ), - request=pika_defs.PikaBodyGenerate22T2vGenerate22T2vPost( + initial_operation = await sync_op( + cls, + ApiEndpoint(path=PATH_TEXT_TO_VIDEO, method="POST"), + response_model=pika_defs.PikaGenerateResponse, + data=pika_defs.PikaBodyGenerate22T2vGenerate22T2vPost( promptText=prompt_text, negativePrompt=negative_prompt, seed=seed, @@ -202,10 +176,9 @@ class PikaTextToVideoNode(IO.ComfyNode): duration=duration, aspectRatio=aspect_ratio, ), - auth_kwargs=auth, content_type="application/x-www-form-urlencoded", ) - return await execute_task(initial_operation, auth_kwargs=auth, node_id=cls.hidden.unique_id) + return await execute_task(initial_operation.video_id, cls) class PikaScenes(IO.ComfyNode): @@ -309,24 +282,16 @@ class PikaScenes(IO.ComfyNode): duration=duration, aspectRatio=aspect_ratio, ) - auth = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } - initial_operation = SynchronousOperation( - endpoint=ApiEndpoint( - path=PATH_PIKASCENES, - method=HttpMethod.POST, - request_model=pika_defs.PikaBodyGenerate22C2vGenerate22PikascenesPost, - response_model=pika_defs.PikaGenerateResponse, - ), - request=pika_request_data, + initial_operation = await sync_op( + cls, + ApiEndpoint(path=PATH_PIKASCENES, method="POST"), + response_model=pika_defs.PikaGenerateResponse, + data=pika_request_data, files=pika_files, content_type="multipart/form-data", - auth_kwargs=auth, ) - return await execute_task(initial_operation, auth_kwargs=auth, node_id=cls.hidden.unique_id) + return await execute_task(initial_operation.video_id, cls) class PikAdditionsNode(IO.ComfyNode): @@ -383,24 +348,16 @@ class PikAdditionsNode(IO.ComfyNode): negativePrompt=negative_prompt, seed=seed, ) - auth = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } - initial_operation = SynchronousOperation( - endpoint=ApiEndpoint( - path=PATH_PIKADDITIONS, - method=HttpMethod.POST, - request_model=pika_defs.PikaBodyGeneratePikadditionsGeneratePikadditionsPost, - response_model=pika_defs.PikaGenerateResponse, - ), - request=pika_request_data, + initial_operation = await sync_op( + cls, + ApiEndpoint(path=PATH_PIKADDITIONS, method="POST"), + response_model=pika_defs.PikaGenerateResponse, + data=pika_request_data, files=pika_files, content_type="multipart/form-data", - auth_kwargs=auth, ) - return await execute_task(initial_operation, auth_kwargs=auth, node_id=cls.hidden.unique_id) + return await execute_task(initial_operation.video_id, cls) class PikaSwapsNode(IO.ComfyNode): @@ -472,23 +429,15 @@ class PikaSwapsNode(IO.ComfyNode): seed=seed, modifyRegionRoi=region_to_modify if region_to_modify else None, ) - auth = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } - initial_operation = SynchronousOperation( - endpoint=ApiEndpoint( - path=PATH_PIKASWAPS, - method=HttpMethod.POST, - request_model=pika_defs.PikaBodyGeneratePikaswapsGeneratePikaswapsPost, - response_model=pika_defs.PikaGenerateResponse, - ), - request=pika_request_data, + initial_operation = await sync_op( + cls, + ApiEndpoint(path=PATH_PIKASWAPS, method="POST"), + response_model=pika_defs.PikaGenerateResponse, + data=pika_request_data, files=pika_files, content_type="multipart/form-data", - auth_kwargs=auth, ) - return await execute_task(initial_operation, auth_kwargs=auth, node_id=cls.hidden.unique_id) + return await execute_task(initial_operation.video_id, cls) class PikaffectsNode(IO.ComfyNode): @@ -528,18 +477,11 @@ class PikaffectsNode(IO.ComfyNode): negative_prompt: str, seed: int, ) -> IO.NodeOutput: - auth = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } - initial_operation = SynchronousOperation( - endpoint=ApiEndpoint( - path=PATH_PIKAFFECTS, - method=HttpMethod.POST, - request_model=pika_defs.PikaBodyGeneratePikaffectsGeneratePikaffectsPost, - response_model=pika_defs.PikaGenerateResponse, - ), - request=pika_defs.PikaBodyGeneratePikaffectsGeneratePikaffectsPost( + initial_operation = await sync_op( + cls, + ApiEndpoint(path=PATH_PIKAFFECTS, method="POST"), + response_model=pika_defs.PikaGenerateResponse, + data=pika_defs.PikaBodyGeneratePikaffectsGeneratePikaffectsPost( pikaffect=pikaffect, promptText=prompt_text, negativePrompt=negative_prompt, @@ -547,9 +489,8 @@ class PikaffectsNode(IO.ComfyNode): ), files={"image": ("image.png", tensor_to_bytesio(image), "image/png")}, content_type="multipart/form-data", - auth_kwargs=auth, ) - return await execute_task(initial_operation, auth_kwargs=auth, node_id=cls.hidden.unique_id) + return await execute_task(initial_operation.video_id, cls) class PikaStartEndFrameNode(IO.ComfyNode): @@ -592,18 +533,11 @@ class PikaStartEndFrameNode(IO.ComfyNode): ("keyFrames", ("image_start.png", tensor_to_bytesio(image_start), "image/png")), ("keyFrames", ("image_end.png", tensor_to_bytesio(image_end), "image/png")), ] - auth = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } - initial_operation = SynchronousOperation( - endpoint=ApiEndpoint( - path=PATH_PIKAFRAMES, - method=HttpMethod.POST, - request_model=pika_defs.PikaBodyGenerate22KeyframeGenerate22PikaframesPost, - response_model=pika_defs.PikaGenerateResponse, - ), - request=pika_defs.PikaBodyGenerate22KeyframeGenerate22PikaframesPost( + initial_operation = await sync_op( + cls, + ApiEndpoint(path=PATH_PIKAFRAMES, method="POST"), + response_model=pika_defs.PikaGenerateResponse, + data=pika_defs.PikaBodyGenerate22KeyframeGenerate22PikaframesPost( promptText=prompt_text, negativePrompt=negative_prompt, seed=seed, @@ -612,9 +546,8 @@ class PikaStartEndFrameNode(IO.ComfyNode): ), files=pika_files, content_type="multipart/form-data", - auth_kwargs=auth, ) - return await execute_task(initial_operation, auth_kwargs=auth, node_id=cls.hidden.unique_id) + return await execute_task(initial_operation.video_id, cls) class PikaApiNodesExtension(ComfyExtension): From e974e554ca23be505b72bc9c1614f4285c1db6e3 Mon Sep 17 00:00:00 2001 From: ComfyUI Wiki Date: Tue, 4 Nov 2025 02:59:44 +0800 Subject: [PATCH 073/647] chore: update embedded docs to v0.3.1 (#10614) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 4d84b0d3e..856e373de 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,6 +1,6 @@ comfyui-frontend-package==1.28.8 comfyui-workflow-templates==0.2.4 -comfyui-embedded-docs==0.3.0 +comfyui-embedded-docs==0.3.1 torch torchsde torchvision From 958a17199ac519504e390ea0d53295ceb8cbd2c1 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 3 Nov 2025 14:08:30 -0800 Subject: [PATCH 074/647] People should update their pytorch versions. (#10618) --- comfy/quant_ops.py | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/comfy/quant_ops.py b/comfy/quant_ops.py index 873f173ed..5af6f118e 100644 --- a/comfy/quant_ops.py +++ b/comfy/quant_ops.py @@ -418,6 +418,10 @@ def fp8_linear(func, args, kwargs): scale_b=scale_b, out_dtype=out_dtype, ) + + if isinstance(output, tuple): # TODO: remove when we drop support for torch 2.4 + output = output[0] + if not tensor_2d: output = output.reshape((-1, input_shape[1], weight.shape[0])) From 0652cb8e2d343f68e38285755835c77bda7f6389 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 3 Nov 2025 14:37:12 -0800 Subject: [PATCH 075/647] Speed up torch.compile (#10620) --- comfy/ops.py | 1 - 1 file changed, 1 deletion(-) diff --git a/comfy/ops.py b/comfy/ops.py index 0c8f23848..afe498caa 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -71,7 +71,6 @@ def cast_to_input(weight, input, non_blocking=False, copy=True): return comfy.model_management.cast_to(weight, input.dtype, input.device, non_blocking=non_blocking, copy=copy) -@torch.compiler.disable() def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, offloadable=False): # NOTE: offloadable=False is a a legacy and if you are a custom node author reading this please pass # offloadable=True and call uncast_bias_weight() after your last usage of the weight/bias. This From e199c8cc6758d388792fd66b99e8de832814ff91 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 3 Nov 2025 14:58:24 -0800 Subject: [PATCH 076/647] Fixes (#10621) --- comfy/quant_ops.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/comfy/quant_ops.py b/comfy/quant_ops.py index 5af6f118e..835fc4b8d 100644 --- a/comfy/quant_ops.py +++ b/comfy/quant_ops.py @@ -131,7 +131,7 @@ class QuantizedTensor(torch.Tensor): self._layout_params = layout_params def __repr__(self): - layout_name = self._layout_type.__name__ + layout_name = self._layout_type param_str = ", ".join(f"{k}={v}" for k, v in list(self._layout_params.items())[:2]) return f"QuantizedTensor(shape={self.shape}, layout={layout_name}, {param_str})" @@ -179,7 +179,7 @@ class QuantizedTensor(torch.Tensor): attr_name = f"_layout_param_{key}" layout_params[key] = inner_tensors[attr_name] - return QuantizedTensor(inner_tensors["_q_data"], layout_type, layout_params) + return QuantizedTensor(inner_tensors["_qdata"], layout_type, layout_params) @classmethod def from_float(cls, tensor, layout_type, **quantize_kwargs) -> 'QuantizedTensor': From 6b88478f9fe0874c0e17468c9fca3a0a84e6c781 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 3 Nov 2025 16:22:10 -0800 Subject: [PATCH 077/647] Bring back fp8 torch compile performance to what it should be. (#10622) --- comfy/quant_ops.py | 41 +++++++++++++++++++++++++++++++++++++++-- 1 file changed, 39 insertions(+), 2 deletions(-) diff --git a/comfy/quant_ops.py b/comfy/quant_ops.py index 835fc4b8d..ed7b29963 100644 --- a/comfy/quant_ops.py +++ b/comfy/quant_ops.py @@ -126,7 +126,7 @@ class QuantizedTensor(torch.Tensor): return torch.Tensor._make_wrapper_subclass(cls, qdata.shape, device=qdata.device, dtype=qdata.dtype, requires_grad=False) def __init__(self, qdata, layout_type, layout_params): - self._qdata = qdata.contiguous() + self._qdata = qdata self._layout_type = layout_type self._layout_params = layout_params @@ -411,7 +411,7 @@ def fp8_linear(func, args, kwargs): try: output = torch._scaled_mm( - plain_input.reshape(-1, input_shape[2]), + plain_input.reshape(-1, input_shape[2]).contiguous(), weight_t, bias=bias, scale_a=scale_a, @@ -447,6 +447,43 @@ def fp8_linear(func, args, kwargs): return torch.nn.functional.linear(input_tensor, weight, bias) +@register_layout_op(torch.ops.aten.addmm.default, "TensorCoreFP8Layout") +def fp8_addmm(func, args, kwargs): + input_tensor = args[1] + weight = args[2] + bias = args[0] + + if isinstance(input_tensor, QuantizedTensor) and isinstance(weight, QuantizedTensor): + out_dtype = kwargs.get("out_dtype") + if out_dtype is None: + out_dtype = input_tensor._layout_params['orig_dtype'] + + plain_input, scale_a = TensorCoreFP8Layout.get_plain_tensors(input_tensor) + plain_weight, scale_b = TensorCoreFP8Layout.get_plain_tensors(weight) + + output = torch._scaled_mm( + plain_input.contiguous(), + plain_weight, + bias=bias, + scale_a=scale_a, + scale_b=scale_b, + out_dtype=out_dtype, + ) + + if isinstance(output, tuple): # TODO: remove when we drop support for torch 2.4 + output = output[0] + return output + + a = list(args) + if isinstance(args[0], QuantizedTensor): + a[0] = args[0].dequantize() + if isinstance(args[1], QuantizedTensor): + a[1] = args[1].dequantize() + if isinstance(args[2], QuantizedTensor): + a[2] = args[2].dequantize() + + return func(*a, **kwargs) + @register_layout_op(torch.ops.aten.view.default, "TensorCoreFP8Layout") @register_layout_op(torch.ops.aten.t.default, "TensorCoreFP8Layout") def fp8_func(func, args, kwargs): From 0f4ef3afa0772ad11d6d72ad21fb1e089c2fcf5f Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 3 Nov 2025 18:47:14 -0800 Subject: [PATCH 078/647] This seems to slow things down slightly on Linux. (#10624) --- comfy/ops.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy/ops.py b/comfy/ops.py index afe498caa..733bff99d 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -35,7 +35,7 @@ def scaled_dot_product_attention(q, k, v, *args, **kwargs): try: - if torch.cuda.is_available(): + if torch.cuda.is_available() and comfy.model_management.WINDOWS: from torch.nn.attention import SDPBackend, sdpa_kernel import inspect if "set_priority" in inspect.signature(sdpa_kernel).parameters: From af4b7b5edb339a15aa443e32aefbceac1810baa0 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 3 Nov 2025 19:14:20 -0800 Subject: [PATCH 079/647] More fp8 torch.compile regressions fixed. (#10625) --- comfy/quant_ops.py | 54 ++++++++++++++++++++++++++++++---------------- 1 file changed, 35 insertions(+), 19 deletions(-) diff --git a/comfy/quant_ops.py b/comfy/quant_ops.py index ed7b29963..c56e32a73 100644 --- a/comfy/quant_ops.py +++ b/comfy/quant_ops.py @@ -446,6 +446,25 @@ def fp8_linear(func, args, kwargs): return torch.nn.functional.linear(input_tensor, weight, bias) +def fp8_mm_(input_tensor, weight, bias=None, out_dtype=None): + if out_dtype is None: + out_dtype = input_tensor._layout_params['orig_dtype'] + + plain_input, scale_a = TensorCoreFP8Layout.get_plain_tensors(input_tensor) + plain_weight, scale_b = TensorCoreFP8Layout.get_plain_tensors(weight) + + output = torch._scaled_mm( + plain_input.contiguous(), + plain_weight, + bias=bias, + scale_a=scale_a, + scale_b=scale_b, + out_dtype=out_dtype, + ) + + if isinstance(output, tuple): # TODO: remove when we drop support for torch 2.4 + output = output[0] + return output @register_layout_op(torch.ops.aten.addmm.default, "TensorCoreFP8Layout") def fp8_addmm(func, args, kwargs): @@ -454,25 +473,7 @@ def fp8_addmm(func, args, kwargs): bias = args[0] if isinstance(input_tensor, QuantizedTensor) and isinstance(weight, QuantizedTensor): - out_dtype = kwargs.get("out_dtype") - if out_dtype is None: - out_dtype = input_tensor._layout_params['orig_dtype'] - - plain_input, scale_a = TensorCoreFP8Layout.get_plain_tensors(input_tensor) - plain_weight, scale_b = TensorCoreFP8Layout.get_plain_tensors(weight) - - output = torch._scaled_mm( - plain_input.contiguous(), - plain_weight, - bias=bias, - scale_a=scale_a, - scale_b=scale_b, - out_dtype=out_dtype, - ) - - if isinstance(output, tuple): # TODO: remove when we drop support for torch 2.4 - output = output[0] - return output + return fp8_mm_(input_tensor, weight, bias=bias, out_dtype=kwargs.get("out_dtype", None)) a = list(args) if isinstance(args[0], QuantizedTensor): @@ -484,6 +485,21 @@ def fp8_addmm(func, args, kwargs): return func(*a, **kwargs) +@register_layout_op(torch.ops.aten.mm.default, "TensorCoreFP8Layout") +def fp8_mm(func, args, kwargs): + input_tensor = args[0] + weight = args[1] + + if isinstance(input_tensor, QuantizedTensor) and isinstance(weight, QuantizedTensor): + return fp8_mm_(input_tensor, weight, bias=None, out_dtype=kwargs.get("out_dtype", None)) + + a = list(args) + if isinstance(args[0], QuantizedTensor): + a[0] = args[0].dequantize() + if isinstance(args[1], QuantizedTensor): + a[1] = args[1].dequantize() + return func(*a, **kwargs) + @register_layout_op(torch.ops.aten.view.default, "TensorCoreFP8Layout") @register_layout_op(torch.ops.aten.t.default, "TensorCoreFP8Layout") def fp8_func(func, args, kwargs): From 9c71a667904a049975531f2a7dd55f4a8fc92652 Mon Sep 17 00:00:00 2001 From: ComfyUI Wiki Date: Wed, 5 Nov 2025 02:51:53 +0800 Subject: [PATCH 080/647] chore: update workflow templates to v0.2.11 (#10634) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 856e373de..249c36dee 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,5 +1,5 @@ comfyui-frontend-package==1.28.8 -comfyui-workflow-templates==0.2.4 +comfyui-workflow-templates==0.2.11 comfyui-embedded-docs==0.3.1 torch torchsde From a389ee01bb7ba5174729906a7f85bd08b5c2cb87 Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Wed, 5 Nov 2025 08:14:10 +1000 Subject: [PATCH 081/647] caching: Handle None outputs tuple case (#10637) --- comfy_execution/caching.py | 2 ++ 1 file changed, 2 insertions(+) diff --git a/comfy_execution/caching.py b/comfy_execution/caching.py index e077f78b0..326a279fc 100644 --- a/comfy_execution/caching.py +++ b/comfy_execution/caching.py @@ -399,6 +399,8 @@ class RAMPressureCache(LRUCache): ram_usage = RAM_CACHE_DEFAULT_RAM_USAGE def scan_list_for_ram_usage(outputs): nonlocal ram_usage + if outputs is None: + return for output in outputs: if isinstance(output, list): scan_list_for_ram_usage(output) From 7f3e4d486cd77c3ad30eb4714ec18bdaf29e2b5c Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 4 Nov 2025 14:37:50 -0800 Subject: [PATCH 082/647] Limit amount of pinned memory on windows to prevent issues. (#10638) --- comfy/model_management.py | 32 ++++++++++++++++++++++++++++---- 1 file changed, 28 insertions(+), 4 deletions(-) diff --git a/comfy/model_management.py b/comfy/model_management.py index 79c0dfdb4..0d040e55e 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -1082,8 +1082,20 @@ def cast_to_device(tensor, device, dtype, copy=False): non_blocking = device_supports_non_blocking(device) return cast_to(tensor, dtype=dtype, device=device, non_blocking=non_blocking, copy=copy) + +PINNED_MEMORY = {} +TOTAL_PINNED_MEMORY = 0 +if PerformanceFeature.PinnedMem in args.fast: + if WINDOWS: + MAX_PINNED_MEMORY = get_total_memory(torch.device("cpu")) * 0.45 # Windows limit is apparently 50% + else: + MAX_PINNED_MEMORY = get_total_memory(torch.device("cpu")) * 0.95 +else: + MAX_PINNED_MEMORY = -1 + def pin_memory(tensor): - if PerformanceFeature.PinnedMem not in args.fast: + global TOTAL_PINNED_MEMORY + if MAX_PINNED_MEMORY <= 0: return False if not is_nvidia(): @@ -1092,13 +1104,21 @@ def pin_memory(tensor): if not is_device_cpu(tensor.device): return False - if torch.cuda.cudart().cudaHostRegister(tensor.data_ptr(), tensor.numel() * tensor.element_size(), 1) == 0: + size = tensor.numel() * tensor.element_size() + if (TOTAL_PINNED_MEMORY + size) > MAX_PINNED_MEMORY: + return False + + ptr = tensor.data_ptr() + if torch.cuda.cudart().cudaHostRegister(ptr, size, 1) == 0: + PINNED_MEMORY[ptr] = size + TOTAL_PINNED_MEMORY += size return True return False def unpin_memory(tensor): - if PerformanceFeature.PinnedMem not in args.fast: + global TOTAL_PINNED_MEMORY + if MAX_PINNED_MEMORY <= 0: return False if not is_nvidia(): @@ -1107,7 +1127,11 @@ def unpin_memory(tensor): if not is_device_cpu(tensor.device): return False - if torch.cuda.cudart().cudaHostUnregister(tensor.data_ptr()) == 0: + ptr = tensor.data_ptr() + if torch.cuda.cudart().cudaHostUnregister(ptr) == 0: + TOTAL_PINNED_MEMORY -= PINNED_MEMORY.pop(ptr) + if len(PINNED_MEMORY) == 0: + TOTAL_PINNED_MEMORY = 0 return True return False From 265adad858e1f31b66cd3523a02b16f5d34ced52 Mon Sep 17 00:00:00 2001 From: comfyanonymous Date: Tue, 4 Nov 2025 19:42:23 -0500 Subject: [PATCH 083/647] ComfyUI version v0.3.68 --- comfyui_version.py | 2 +- pyproject.toml | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/comfyui_version.py b/comfyui_version.py index db48b05c4..25d1a4157 100644 --- a/comfyui_version.py +++ b/comfyui_version.py @@ -1,3 +1,3 @@ # This file is automatically generated by the build process when version is # updated in pyproject.toml. -__version__ = "0.3.67" +__version__ = "0.3.68" diff --git a/pyproject.toml b/pyproject.toml index ab054355c..79ff3f74a 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -1,6 +1,6 @@ [project] name = "ComfyUI" -version = "0.3.67" +version = "0.3.68" readme = "README.md" license = { file = "LICENSE" } requires-python = ">=3.9" From 4cd881866bad0cde70273cc123d725693c1f2759 Mon Sep 17 00:00:00 2001 From: contentis Date: Wed, 5 Nov 2025 02:10:11 +0100 Subject: [PATCH 084/647] Use single apply_rope function across models (#10547) --- comfy/ldm/flux/layers.py | 4 +- comfy/ldm/flux/math.py | 10 +--- comfy/ldm/lightricks/model.py | 88 ++++++++++++++--------------------- comfy/ldm/qwen_image/model.py | 36 +++++++------- comfy/ldm/wan/model.py | 1 + 5 files changed, 59 insertions(+), 80 deletions(-) diff --git a/comfy/ldm/flux/layers.py b/comfy/ldm/flux/layers.py index ef21b416b..a3eab0470 100644 --- a/comfy/ldm/flux/layers.py +++ b/comfy/ldm/flux/layers.py @@ -195,8 +195,8 @@ class DoubleStreamBlock(nn.Module): txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1]:] # calculate the img bloks - img = img + apply_mod(self.img_attn.proj(img_attn), img_mod1.gate, None, modulation_dims_img) - img = img + apply_mod(self.img_mlp(apply_mod(self.img_norm2(img), (1 + img_mod2.scale), img_mod2.shift, modulation_dims_img)), img_mod2.gate, None, modulation_dims_img) + img += apply_mod(self.img_attn.proj(img_attn), img_mod1.gate, None, modulation_dims_img) + img += apply_mod(self.img_mlp(apply_mod(self.img_norm2(img), (1 + img_mod2.scale), img_mod2.shift, modulation_dims_img)), img_mod2.gate, None, modulation_dims_img) # calculate the txt bloks txt += apply_mod(self.txt_attn.proj(txt_attn), txt_mod1.gate, None, modulation_dims_txt) diff --git a/comfy/ldm/flux/math.py b/comfy/ldm/flux/math.py index 8deda0d4a..158420290 100644 --- a/comfy/ldm/flux/math.py +++ b/comfy/ldm/flux/math.py @@ -7,15 +7,7 @@ import comfy.model_management def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor, mask=None, transformer_options={}) -> Tensor: - q_shape = q.shape - k_shape = k.shape - - if pe is not None: - q = q.to(dtype=pe.dtype).reshape(*q.shape[:-1], -1, 1, 2) - k = k.to(dtype=pe.dtype).reshape(*k.shape[:-1], -1, 1, 2) - q = (pe[..., 0] * q[..., 0] + pe[..., 1] * q[..., 1]).reshape(*q_shape).type_as(v) - k = (pe[..., 0] * k[..., 0] + pe[..., 1] * k[..., 1]).reshape(*k_shape).type_as(v) - + q, k = apply_rope(q, k, pe) heads = q.shape[1] x = optimized_attention(q, k, v, heads, skip_reshape=True, mask=mask, transformer_options=transformer_options) return x diff --git a/comfy/ldm/lightricks/model.py b/comfy/ldm/lightricks/model.py index def365ba7..5bcba998b 100644 --- a/comfy/ldm/lightricks/model.py +++ b/comfy/ldm/lightricks/model.py @@ -3,12 +3,11 @@ from torch import nn import comfy.patcher_extension import comfy.ldm.modules.attention import comfy.ldm.common_dit -from einops import rearrange import math from typing import Dict, Optional, Tuple from .symmetric_patchifier import SymmetricPatchifier, latent_to_pixel_coords - +from comfy.ldm.flux.math import apply_rope1 def get_timestep_embedding( timesteps: torch.Tensor, @@ -238,20 +237,6 @@ class FeedForward(nn.Module): return self.net(x) -def apply_rotary_emb(input_tensor, freqs_cis): #TODO: remove duplicate funcs and pick the best/fastest one - cos_freqs = freqs_cis[0] - sin_freqs = freqs_cis[1] - - t_dup = rearrange(input_tensor, "... (d r) -> ... d r", r=2) - t1, t2 = t_dup.unbind(dim=-1) - t_dup = torch.stack((-t2, t1), dim=-1) - input_tensor_rot = rearrange(t_dup, "... d r -> ... (d r)") - - out = input_tensor * cos_freqs + input_tensor_rot * sin_freqs - - return out - - class CrossAttention(nn.Module): def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0., attn_precision=None, dtype=None, device=None, operations=None): super().__init__() @@ -281,8 +266,8 @@ class CrossAttention(nn.Module): k = self.k_norm(k) if pe is not None: - q = apply_rotary_emb(q, pe) - k = apply_rotary_emb(k, pe) + q = apply_rope1(q.unsqueeze(1), pe).squeeze(1) + k = apply_rope1(k.unsqueeze(1), pe).squeeze(1) if mask is None: out = comfy.ldm.modules.attention.optimized_attention(q, k, v, self.heads, attn_precision=self.attn_precision, transformer_options=transformer_options) @@ -306,12 +291,17 @@ class BasicTransformerBlock(nn.Module): def forward(self, x, context=None, attention_mask=None, timestep=None, pe=None, transformer_options={}): shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (self.scale_shift_table[None, None].to(device=x.device, dtype=x.dtype) + timestep.reshape(x.shape[0], timestep.shape[1], self.scale_shift_table.shape[0], -1)).unbind(dim=2) - x += self.attn1(comfy.ldm.common_dit.rms_norm(x) * (1 + scale_msa) + shift_msa, pe=pe, transformer_options=transformer_options) * gate_msa + norm_x = comfy.ldm.common_dit.rms_norm(x) + attn1_input = torch.addcmul(norm_x, norm_x, scale_msa).add_(shift_msa) + attn1_result = self.attn1(attn1_input, pe=pe, transformer_options=transformer_options) + x.addcmul_(attn1_result, gate_msa) x += self.attn2(x, context=context, mask=attention_mask, transformer_options=transformer_options) - y = comfy.ldm.common_dit.rms_norm(x) * (1 + scale_mlp) + shift_mlp - x += self.ff(y) * gate_mlp + norm_x = comfy.ldm.common_dit.rms_norm(x) + y = torch.addcmul(norm_x, norm_x, scale_mlp).add_(shift_mlp) + ff_result = self.ff(y) + x.addcmul_(ff_result, gate_mlp) return x @@ -327,41 +317,35 @@ def get_fractional_positions(indices_grid, max_pos): def precompute_freqs_cis(indices_grid, dim, out_dtype, theta=10000.0, max_pos=[20, 2048, 2048]): - dtype = torch.float32 #self.dtype + dtype = torch.float32 + device = indices_grid.device + # Get fractional positions and compute frequency indices fractional_positions = get_fractional_positions(indices_grid, max_pos) + indices = theta ** torch.linspace(0, 1, dim // 6, device=device, dtype=dtype) * math.pi / 2 - start = 1 - end = theta - device = fractional_positions.device + # Compute frequencies and apply cos/sin + freqs = (indices * (fractional_positions.unsqueeze(-1) * 2 - 1)).transpose(-1, -2).flatten(2) + cos_vals = freqs.cos().repeat_interleave(2, dim=-1) + sin_vals = freqs.sin().repeat_interleave(2, dim=-1) - indices = theta ** ( - torch.linspace( - math.log(start, theta), - math.log(end, theta), - dim // 6, - device=device, - dtype=dtype, - ) - ) - indices = indices.to(dtype=dtype) - - indices = indices * math.pi / 2 - - freqs = ( - (indices * (fractional_positions.unsqueeze(-1) * 2 - 1)) - .transpose(-1, -2) - .flatten(2) - ) - - cos_freq = freqs.cos().repeat_interleave(2, dim=-1) - sin_freq = freqs.sin().repeat_interleave(2, dim=-1) + # Pad if dim is not divisible by 6 if dim % 6 != 0: - cos_padding = torch.ones_like(cos_freq[:, :, : dim % 6]) - sin_padding = torch.zeros_like(cos_freq[:, :, : dim % 6]) - cos_freq = torch.cat([cos_padding, cos_freq], dim=-1) - sin_freq = torch.cat([sin_padding, sin_freq], dim=-1) - return cos_freq.to(out_dtype), sin_freq.to(out_dtype) + padding_size = dim % 6 + cos_vals = torch.cat([torch.ones_like(cos_vals[:, :, :padding_size]), cos_vals], dim=-1) + sin_vals = torch.cat([torch.zeros_like(sin_vals[:, :, :padding_size]), sin_vals], dim=-1) + + # Reshape and extract one value per pair (since repeat_interleave duplicates each value) + cos_vals = cos_vals.reshape(*cos_vals.shape[:2], -1, 2)[..., 0] # [B, N, dim//2] + sin_vals = sin_vals.reshape(*sin_vals.shape[:2], -1, 2)[..., 0] # [B, N, dim//2] + + # Build rotation matrix [[cos, -sin], [sin, cos]] and add heads dimension + freqs_cis = torch.stack([ + torch.stack([cos_vals, -sin_vals], dim=-1), + torch.stack([sin_vals, cos_vals], dim=-1) + ], dim=-2).unsqueeze(1) # [B, 1, N, dim//2, 2, 2] + + return freqs_cis.to(out_dtype) class LTXVModel(torch.nn.Module): @@ -501,7 +485,7 @@ class LTXVModel(torch.nn.Module): shift, scale = scale_shift_values[:, :, 0], scale_shift_values[:, :, 1] x = self.norm_out(x) # Modulation - x = x * (1 + scale) + shift + x = torch.addcmul(x, x, scale).add_(shift) x = self.proj_out(x) x = self.patchifier.unpatchify( diff --git a/comfy/ldm/qwen_image/model.py b/comfy/ldm/qwen_image/model.py index b9f60c2b7..81d3ee7c0 100644 --- a/comfy/ldm/qwen_image/model.py +++ b/comfy/ldm/qwen_image/model.py @@ -10,6 +10,7 @@ from comfy.ldm.modules.attention import optimized_attention_masked from comfy.ldm.flux.layers import EmbedND import comfy.ldm.common_dit import comfy.patcher_extension +from comfy.ldm.flux.math import apply_rope1 class GELU(nn.Module): def __init__(self, dim_in: int, dim_out: int, approximate: str = "none", bias: bool = True, dtype=None, device=None, operations=None): @@ -134,33 +135,34 @@ class Attention(nn.Module): image_rotary_emb: Optional[torch.Tensor] = None, transformer_options={}, ) -> Tuple[torch.Tensor, torch.Tensor]: + batch_size = hidden_states.shape[0] + seq_img = hidden_states.shape[1] seq_txt = encoder_hidden_states.shape[1] - img_query = self.to_q(hidden_states).unflatten(-1, (self.heads, -1)) - img_key = self.to_k(hidden_states).unflatten(-1, (self.heads, -1)) - img_value = self.to_v(hidden_states).unflatten(-1, (self.heads, -1)) + # Project and reshape to BHND format (batch, heads, seq, dim) + img_query = self.to_q(hidden_states).view(batch_size, seq_img, self.heads, -1).transpose(1, 2).contiguous() + img_key = self.to_k(hidden_states).view(batch_size, seq_img, self.heads, -1).transpose(1, 2).contiguous() + img_value = self.to_v(hidden_states).view(batch_size, seq_img, self.heads, -1).transpose(1, 2) - txt_query = self.add_q_proj(encoder_hidden_states).unflatten(-1, (self.heads, -1)) - txt_key = self.add_k_proj(encoder_hidden_states).unflatten(-1, (self.heads, -1)) - txt_value = self.add_v_proj(encoder_hidden_states).unflatten(-1, (self.heads, -1)) + txt_query = self.add_q_proj(encoder_hidden_states).view(batch_size, seq_txt, self.heads, -1).transpose(1, 2).contiguous() + txt_key = self.add_k_proj(encoder_hidden_states).view(batch_size, seq_txt, self.heads, -1).transpose(1, 2).contiguous() + txt_value = self.add_v_proj(encoder_hidden_states).view(batch_size, seq_txt, self.heads, -1).transpose(1, 2) img_query = self.norm_q(img_query) img_key = self.norm_k(img_key) txt_query = self.norm_added_q(txt_query) txt_key = self.norm_added_k(txt_key) - joint_query = torch.cat([txt_query, img_query], dim=1) - joint_key = torch.cat([txt_key, img_key], dim=1) - joint_value = torch.cat([txt_value, img_value], dim=1) + joint_query = torch.cat([txt_query, img_query], dim=2) + joint_key = torch.cat([txt_key, img_key], dim=2) + joint_value = torch.cat([txt_value, img_value], dim=2) - joint_query = apply_rotary_emb(joint_query, image_rotary_emb) - joint_key = apply_rotary_emb(joint_key, image_rotary_emb) + joint_query = apply_rope1(joint_query, image_rotary_emb) + joint_key = apply_rope1(joint_key, image_rotary_emb) - joint_query = joint_query.flatten(start_dim=2) - joint_key = joint_key.flatten(start_dim=2) - joint_value = joint_value.flatten(start_dim=2) - - joint_hidden_states = optimized_attention_masked(joint_query, joint_key, joint_value, self.heads, attention_mask, transformer_options=transformer_options) + joint_hidden_states = optimized_attention_masked(joint_query, joint_key, joint_value, self.heads, + attention_mask, transformer_options=transformer_options, + skip_reshape=True) txt_attn_output = joint_hidden_states[:, :seq_txt, :] img_attn_output = joint_hidden_states[:, seq_txt:, :] @@ -413,7 +415,7 @@ class QwenImageTransformer2DModel(nn.Module): txt_start = round(max(((x.shape[-1] + (self.patch_size // 2)) // self.patch_size) // 2, ((x.shape[-2] + (self.patch_size // 2)) // self.patch_size) // 2)) txt_ids = torch.arange(txt_start, txt_start + context.shape[1], device=x.device).reshape(1, -1, 1).repeat(x.shape[0], 1, 3) ids = torch.cat((txt_ids, img_ids), dim=1) - image_rotary_emb = self.pe_embedder(ids).squeeze(1).unsqueeze(2).to(x.dtype) + image_rotary_emb = self.pe_embedder(ids).to(torch.float32).contiguous() del ids, txt_ids, img_ids hidden_states = self.img_in(hidden_states) diff --git a/comfy/ldm/wan/model.py b/comfy/ldm/wan/model.py index 5ec1511ce..a9d5e10d9 100644 --- a/comfy/ldm/wan/model.py +++ b/comfy/ldm/wan/model.py @@ -232,6 +232,7 @@ class WanAttentionBlock(nn.Module): # assert e[0].dtype == torch.float32 # self-attention + x = x.contiguous() # otherwise implicit in LayerNorm y = self.self_attn( torch.addcmul(repeat_e(e[0], x), self.norm1(x), 1 + repeat_e(e[1], x)), freqs, transformer_options=transformer_options) From c4a6b389de1014471a75a46ee57d2fdac4f8df93 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 4 Nov 2025 19:47:35 -0800 Subject: [PATCH 085/647] Lower ltxv mem usage to what it was before previous pr. (#10643) Bring back qwen behavior to what it was before previous pr. --- comfy/ldm/lightricks/model.py | 22 +++++++++++----------- comfy/ldm/qwen_image/model.py | 2 +- 2 files changed, 12 insertions(+), 12 deletions(-) diff --git a/comfy/ldm/lightricks/model.py b/comfy/ldm/lightricks/model.py index 5bcba998b..593f7940f 100644 --- a/comfy/ldm/lightricks/model.py +++ b/comfy/ldm/lightricks/model.py @@ -291,17 +291,17 @@ class BasicTransformerBlock(nn.Module): def forward(self, x, context=None, attention_mask=None, timestep=None, pe=None, transformer_options={}): shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (self.scale_shift_table[None, None].to(device=x.device, dtype=x.dtype) + timestep.reshape(x.shape[0], timestep.shape[1], self.scale_shift_table.shape[0], -1)).unbind(dim=2) - norm_x = comfy.ldm.common_dit.rms_norm(x) - attn1_input = torch.addcmul(norm_x, norm_x, scale_msa).add_(shift_msa) - attn1_result = self.attn1(attn1_input, pe=pe, transformer_options=transformer_options) - x.addcmul_(attn1_result, gate_msa) + attn1_input = comfy.ldm.common_dit.rms_norm(x) + attn1_input = torch.addcmul(attn1_input, attn1_input, scale_msa).add_(shift_msa) + attn1_input = self.attn1(attn1_input, pe=pe, transformer_options=transformer_options) + x.addcmul_(attn1_input, gate_msa) + del attn1_input x += self.attn2(x, context=context, mask=attention_mask, transformer_options=transformer_options) - norm_x = comfy.ldm.common_dit.rms_norm(x) - y = torch.addcmul(norm_x, norm_x, scale_mlp).add_(shift_mlp) - ff_result = self.ff(y) - x.addcmul_(ff_result, gate_mlp) + y = comfy.ldm.common_dit.rms_norm(x) + y = torch.addcmul(y, y, scale_mlp).add_(shift_mlp) + x.addcmul_(self.ff(y), gate_mlp) return x @@ -336,8 +336,8 @@ def precompute_freqs_cis(indices_grid, dim, out_dtype, theta=10000.0, max_pos=[2 sin_vals = torch.cat([torch.zeros_like(sin_vals[:, :, :padding_size]), sin_vals], dim=-1) # Reshape and extract one value per pair (since repeat_interleave duplicates each value) - cos_vals = cos_vals.reshape(*cos_vals.shape[:2], -1, 2)[..., 0] # [B, N, dim//2] - sin_vals = sin_vals.reshape(*sin_vals.shape[:2], -1, 2)[..., 0] # [B, N, dim//2] + cos_vals = cos_vals.reshape(*cos_vals.shape[:2], -1, 2)[..., 0].to(out_dtype) # [B, N, dim//2] + sin_vals = sin_vals.reshape(*sin_vals.shape[:2], -1, 2)[..., 0].to(out_dtype) # [B, N, dim//2] # Build rotation matrix [[cos, -sin], [sin, cos]] and add heads dimension freqs_cis = torch.stack([ @@ -345,7 +345,7 @@ def precompute_freqs_cis(indices_grid, dim, out_dtype, theta=10000.0, max_pos=[2 torch.stack([sin_vals, cos_vals], dim=-1) ], dim=-2).unsqueeze(1) # [B, 1, N, dim//2, 2, 2] - return freqs_cis.to(out_dtype) + return freqs_cis class LTXVModel(torch.nn.Module): diff --git a/comfy/ldm/qwen_image/model.py b/comfy/ldm/qwen_image/model.py index 81d3ee7c0..e5d0d17c1 100644 --- a/comfy/ldm/qwen_image/model.py +++ b/comfy/ldm/qwen_image/model.py @@ -415,7 +415,7 @@ class QwenImageTransformer2DModel(nn.Module): txt_start = round(max(((x.shape[-1] + (self.patch_size // 2)) // self.patch_size) // 2, ((x.shape[-2] + (self.patch_size // 2)) // self.patch_size) // 2)) txt_ids = torch.arange(txt_start, txt_start + context.shape[1], device=x.device).reshape(1, -1, 1).repeat(x.shape[0], 1, 3) ids = torch.cat((txt_ids, img_ids), dim=1) - image_rotary_emb = self.pe_embedder(ids).to(torch.float32).contiguous() + image_rotary_emb = self.pe_embedder(ids).to(x.dtype).contiguous() del ids, txt_ids, img_ids hidden_states = self.img_in(hidden_states) From bda0eb2448135797d5a72f7236ce26d07e555baf Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Wed, 5 Nov 2025 12:16:00 +0200 Subject: [PATCH 086/647] feat(API-nodes): move Rodin3D nodes to new client; removed old api client.py (#10645) --- comfy_api_nodes/apis/PixverseController.py | 17 - comfy_api_nodes/apis/PixverseDto.py | 57 - comfy_api_nodes/apis/client.py | 981 ------------------ comfy_api_nodes/nodes_rodin.py | 196 ++-- comfy_api_nodes/util/client.py | 4 +- comfy_api_nodes/util/download_helpers.py | 2 +- .../{apis => util}/request_logger.py | 4 +- comfy_api_nodes/util/upload_helpers.py | 2 +- 8 files changed, 75 insertions(+), 1188 deletions(-) delete mode 100644 comfy_api_nodes/apis/PixverseController.py delete mode 100644 comfy_api_nodes/apis/PixverseDto.py delete mode 100644 comfy_api_nodes/apis/client.py rename comfy_api_nodes/{apis => util}/request_logger.py (100%) diff --git a/comfy_api_nodes/apis/PixverseController.py b/comfy_api_nodes/apis/PixverseController.py deleted file mode 100644 index 310c0f546..000000000 --- a/comfy_api_nodes/apis/PixverseController.py +++ /dev/null @@ -1,17 +0,0 @@ -# generated by datamodel-codegen: -# filename: filtered-openapi.yaml -# timestamp: 2025-04-29T23:44:54+00:00 - -from __future__ import annotations - -from typing import Optional - -from pydantic import BaseModel - -from . import PixverseDto - - -class ResponseData(BaseModel): - ErrCode: Optional[int] = None - ErrMsg: Optional[str] = None - Resp: Optional[PixverseDto.V2OpenAPII2VResp] = None diff --git a/comfy_api_nodes/apis/PixverseDto.py b/comfy_api_nodes/apis/PixverseDto.py deleted file mode 100644 index 323c38e96..000000000 --- a/comfy_api_nodes/apis/PixverseDto.py +++ /dev/null @@ -1,57 +0,0 @@ -# generated by datamodel-codegen: -# filename: filtered-openapi.yaml -# timestamp: 2025-04-29T23:44:54+00:00 - -from __future__ import annotations - -from typing import Optional - -from pydantic import BaseModel, Field - - -class V2OpenAPII2VResp(BaseModel): - video_id: Optional[int] = Field(None, description='Video_id') - - -class V2OpenAPIT2VReq(BaseModel): - aspect_ratio: str = Field( - ..., description='Aspect ratio (16:9, 4:3, 1:1, 3:4, 9:16)', examples=['16:9'] - ) - duration: int = Field( - ..., - description='Video duration (5, 8 seconds, --model=v3.5 only allows 5,8; --quality=1080p does not support 8s)', - examples=[5], - ) - model: str = Field( - ..., description='Model version (only supports v3.5)', examples=['v3.5'] - ) - motion_mode: Optional[str] = Field( - 'normal', - description='Motion mode (normal, fast, --fast only available when duration=5; --quality=1080p does not support fast)', - examples=['normal'], - ) - negative_prompt: Optional[str] = Field( - None, description='Negative prompt\n', max_length=2048 - ) - prompt: str = Field(..., description='Prompt', max_length=2048) - quality: str = Field( - ..., - description='Video quality ("360p"(Turbo model), "540p", "720p", "1080p")', - examples=['540p'], - ) - seed: Optional[int] = Field(None, description='Random seed, range: 0 - 2147483647') - style: Optional[str] = Field( - None, - description='Style (effective when model=v3.5, "anime", "3d_animation", "clay", "comic", "cyberpunk") Do not include style parameter unless needed', - examples=['anime'], - ) - template_id: Optional[int] = Field( - None, - description='Template ID (template_id must be activated before use)', - examples=[302325299692608], - ) - water_mark: Optional[bool] = Field( - False, - description='Watermark (true: add watermark, false: no watermark)', - examples=[False], - ) diff --git a/comfy_api_nodes/apis/client.py b/comfy_api_nodes/apis/client.py deleted file mode 100644 index bdaddcc88..000000000 --- a/comfy_api_nodes/apis/client.py +++ /dev/null @@ -1,981 +0,0 @@ -""" -API Client Framework for api.comfy.org. - -This module provides a flexible framework for making API requests from ComfyUI nodes. -It supports both synchronous and asynchronous API operations with proper type validation. - -Key Components: --------------- -1. ApiClient - Handles HTTP requests with authentication and error handling -2. ApiEndpoint - Defines a single HTTP endpoint with its request/response models -3. ApiOperation - Executes a single synchronous API operation - -Usage Examples: --------------- - -# Example 1: Synchronous API Operation -# ------------------------------------ -# For a simple API call that returns the result immediately: - -# 1. Create the API client -api_client = ApiClient( - base_url="https://api.example.com", - auth_token="your_auth_token_here", - comfy_api_key="your_comfy_api_key_here", - timeout=30.0, - verify_ssl=True -) - -# 2. Define the endpoint -user_info_endpoint = ApiEndpoint( - path="/v1/users/me", - method=HttpMethod.GET, - request_model=EmptyRequest, # No request body needed - response_model=UserProfile, # Pydantic model for the response - query_params=None -) - -# 3. Create the request object -request = EmptyRequest() - -# 4. Create and execute the operation -operation = ApiOperation( - endpoint=user_info_endpoint, - request=request -) -user_profile = await operation.execute(client=api_client) # Returns immediately with the result - - -# Example 2: Asynchronous API Operation with Polling -# ------------------------------------------------- -# For an API that starts a task and requires polling for completion: - -# 1. Define the endpoints (initial request and polling) -generate_image_endpoint = ApiEndpoint( - path="/v1/images/generate", - method=HttpMethod.POST, - request_model=ImageGenerationRequest, - response_model=TaskCreatedResponse, - query_params=None -) - -check_task_endpoint = ApiEndpoint( - path="/v1/tasks/{task_id}", - method=HttpMethod.GET, - request_model=EmptyRequest, - response_model=ImageGenerationResult, - query_params=None -) - -# 2. Create the request object -request = ImageGenerationRequest( - prompt="a beautiful sunset over mountains", - width=1024, - height=1024, - num_images=1 -) - -# 3. Create and execute the polling operation -operation = PollingOperation( - initial_endpoint=generate_image_endpoint, - initial_request=request, - poll_endpoint=check_task_endpoint, - task_id_field="task_id", - status_field="status", - completed_statuses=["completed"], - failed_statuses=["failed", "error"] -) - -# This will make the initial request and then poll until completion -result = await operation.execute(client=api_client) # Returns the final ImageGenerationResult when done -""" - -from __future__ import annotations -import aiohttp -import asyncio -import logging -import io -import os -import socket -from aiohttp.client_exceptions import ClientError, ClientResponseError -from typing import Type, Optional, Any, TypeVar, Generic, Callable -from enum import Enum -import json -from urllib.parse import urljoin, urlparse -from pydantic import BaseModel, Field -import uuid # For generating unique operation IDs - -from server import PromptServer -from comfy.cli_args import args -from comfy import utils -from . import request_logger - -T = TypeVar("T", bound=BaseModel) -R = TypeVar("R", bound=BaseModel) -P = TypeVar("P", bound=BaseModel) # For poll response - -PROGRESS_BAR_MAX = 100 - - -class NetworkError(Exception): - """Base exception for network-related errors with diagnostic information.""" - pass - - -class LocalNetworkError(NetworkError): - """Exception raised when local network connectivity issues are detected.""" - pass - - -class ApiServerError(NetworkError): - """Exception raised when the API server is unreachable but internet is working.""" - pass - - -class EmptyRequest(BaseModel): - """Base class for empty request bodies. - For GET requests, fields will be sent as query parameters.""" - - pass - - -class UploadRequest(BaseModel): - file_name: str = Field(..., description="Filename to upload") - content_type: Optional[str] = Field( - None, - description="Mime type of the file. For example: image/png, image/jpeg, video/mp4, etc.", - ) - - -class UploadResponse(BaseModel): - download_url: str = Field(..., description="URL to GET uploaded file") - upload_url: str = Field(..., description="URL to PUT file to upload") - - -class HttpMethod(str, Enum): - GET = "GET" - POST = "POST" - PUT = "PUT" - DELETE = "DELETE" - PATCH = "PATCH" - - -class ApiClient: - """ - Client for making HTTP requests to an API with authentication, error handling, and retry logic. - """ - - def __init__( - self, - base_url: str, - auth_token: Optional[str] = None, - comfy_api_key: Optional[str] = None, - timeout: float = 3600.0, - verify_ssl: bool = True, - max_retries: int = 3, - retry_delay: float = 1.0, - retry_backoff_factor: float = 2.0, - retry_status_codes: Optional[tuple[int, ...]] = None, - session: Optional[aiohttp.ClientSession] = None, - ): - self.base_url = base_url - self.auth_token = auth_token - self.comfy_api_key = comfy_api_key - self.timeout = timeout - self.verify_ssl = verify_ssl - self.max_retries = max_retries - self.retry_delay = retry_delay - self.retry_backoff_factor = retry_backoff_factor - # Default retry status codes: 408 (Request Timeout), 429 (Too Many Requests), - # 500, 502, 503, 504 (Server Errors) - self.retry_status_codes = retry_status_codes or (408, 429, 500, 502, 503, 504) - self._session: Optional[aiohttp.ClientSession] = session - self._owns_session = session is None # Track if we have to close it - - @staticmethod - def _generate_operation_id(path: str) -> str: - """Generates a unique operation ID for logging.""" - return f"{path.strip('/').replace('/', '_')}_{uuid.uuid4().hex[:8]}" - - @staticmethod - def _create_json_payload_args( - data: Optional[dict[str, Any]] = None, - headers: Optional[dict[str, str]] = None, - ) -> dict[str, Any]: - return { - "json": data, - "headers": headers, - } - - def _create_form_data_args( - self, - data: dict[str, Any] | None, - files: dict[str, Any] | None, - headers: Optional[dict[str, str]] = None, - multipart_parser: Callable | None = None, - ) -> dict[str, Any]: - if headers and "Content-Type" in headers: - del headers["Content-Type"] - - if multipart_parser and data: - data = multipart_parser(data) - - if isinstance(data, aiohttp.FormData): - form = data # If the parser already returned a FormData, pass it through - else: - form = aiohttp.FormData(default_to_multipart=True) - if data: # regular text fields - for k, v in data.items(): - if v is None: - continue # aiohttp fails to serialize "None" values - # aiohttp expects strings or bytes; convert enums etc. - form.add_field(k, str(v) if not isinstance(v, (bytes, bytearray)) else v) - - if files: - file_iter = files if isinstance(files, list) else files.items() - for field_name, file_obj in file_iter: - if file_obj is None: - continue # aiohttp fails to serialize "None" values - # file_obj can be (filename, bytes/io.BytesIO, content_type) tuple - if isinstance(file_obj, tuple): - filename, file_value, content_type = self._unpack_tuple(file_obj) - else: - file_value = file_obj - filename = getattr(file_obj, "name", field_name) - content_type = "application/octet-stream" - - form.add_field( - name=field_name, - value=file_value, - filename=filename, - content_type=content_type, - ) - return {"data": form, "headers": headers or {}} - - @staticmethod - def _create_urlencoded_form_data_args( - data: dict[str, Any], - headers: Optional[dict[str, str]] = None, - ) -> dict[str, Any]: - headers = headers or {} - headers["Content-Type"] = "application/x-www-form-urlencoded" - return { - "data": data, - "headers": headers, - } - - def get_headers(self) -> dict[str, str]: - """Get headers for API requests, including authentication if available""" - headers = {"Content-Type": "application/json", "Accept": "application/json"} - - if self.auth_token: - headers["Authorization"] = f"Bearer {self.auth_token}" - elif self.comfy_api_key: - headers["X-API-KEY"] = self.comfy_api_key - - return headers - - async def _check_connectivity(self, target_url: str) -> dict[str, bool]: - """ - Check connectivity to determine if network issues are local or server-related. - - Args: - target_url: URL to check connectivity to - - Returns: - Dictionary with connectivity status details - """ - results = { - "internet_accessible": False, - "api_accessible": False, - "is_local_issue": False, - "is_api_issue": False, - } - timeout = aiohttp.ClientTimeout(total=5.0) - async with aiohttp.ClientSession(timeout=timeout) as session: - try: - async with session.get("https://www.google.com", ssl=self.verify_ssl) as resp: - results["internet_accessible"] = resp.status < 500 - except (ClientError, asyncio.TimeoutError, socket.gaierror): - results["is_local_issue"] = True - return results # cannot reach the internet – early exit - - # Now check API health endpoint - parsed = urlparse(target_url) - health_url = f"{parsed.scheme}://{parsed.netloc}/health" - try: - async with session.get(health_url, ssl=self.verify_ssl) as resp: - results["api_accessible"] = resp.status < 500 - except ClientError: - pass # leave as False - - results["is_api_issue"] = results["internet_accessible"] and not results["api_accessible"] - return results - - async def request( - self, - method: str, - path: str, - params: Optional[dict[str, Any]] = None, - data: Optional[dict[str, Any]] = None, - files: Optional[dict[str, Any] | list[tuple[str, Any]]] = None, - headers: Optional[dict[str, str]] = None, - content_type: str = "application/json", - multipart_parser: Callable | None = None, - retry_count: int = 0, # Used internally for tracking retries - ) -> dict[str, Any]: - """ - Make an HTTP request to the API with automatic retries for transient errors. - - Args: - method: HTTP method (GET, POST, etc.) - path: API endpoint path (will be joined with base_url) - params: Query parameters - data: body data - files: Files to upload - headers: Additional headers - content_type: Content type of the request. Defaults to application/json. - retry_count: Internal parameter for tracking retries, do not set manually - - Returns: - Parsed JSON response - - Raises: - LocalNetworkError: If local network connectivity issues are detected - ApiServerError: If the API server is unreachable but internet is working - Exception: For other request failures - """ - - # Build full URL and merge headers - relative_path = path.lstrip("/") - url = urljoin(self.base_url, relative_path) - self._check_auth(self.auth_token, self.comfy_api_key) - - request_headers = self.get_headers() - if headers: - request_headers.update(headers) - if files: - request_headers.pop("Content-Type", None) - if params: - params = {k: v for k, v in params.items() if v is not None} # aiohttp fails to serialize None values - - logging.debug("[DEBUG] Request Headers: %s", request_headers) - logging.debug("[DEBUG] Files: %s", files) - logging.debug("[DEBUG] Params: %s", params) - logging.debug("[DEBUG] Data: %s", data) - - if content_type == "application/x-www-form-urlencoded": - payload_args = self._create_urlencoded_form_data_args(data or {}, request_headers) - elif content_type == "multipart/form-data": - payload_args = self._create_form_data_args(data, files, request_headers, multipart_parser) - else: - payload_args = self._create_json_payload_args(data, request_headers) - - operation_id = self._generate_operation_id(path) - request_logger.log_request_response( - operation_id=operation_id, - request_method=method, - request_url=url, - request_headers=request_headers, - request_params=params, - request_data=data if content_type == "application/json" else "[form-data or other]", - ) - - session = await self._get_session() - try: - async with session.request( - method, - url, - params=params, - ssl=self.verify_ssl, - **payload_args, - ) as resp: - if resp.status >= 400: - try: - error_data = await resp.json() - except (aiohttp.ContentTypeError, json.JSONDecodeError): - error_data = await resp.text() - - return await self._handle_http_error( - ClientResponseError(resp.request_info, resp.history, status=resp.status, message=error_data), - operation_id, - method, - url, - params, - data, - files, - headers, - content_type, - multipart_parser, - retry_count=retry_count, - response_content=error_data, - ) - - # Success – parse JSON (safely) and log - try: - payload = await resp.json() - response_content_to_log = payload - except (aiohttp.ContentTypeError, json.JSONDecodeError): - payload = {} - response_content_to_log = await resp.text() - - request_logger.log_request_response( - operation_id=operation_id, - request_method=method, - request_url=url, - response_status_code=resp.status, - response_headers=dict(resp.headers), - response_content=response_content_to_log, - ) - return payload - - except (ClientError, asyncio.TimeoutError, socket.gaierror) as e: - # Treat as *connection* problem – optionally retry, else escalate - if retry_count < self.max_retries: - delay = self.retry_delay * (self.retry_backoff_factor ** retry_count) - logging.warning("Connection error. Retrying in %.2fs (%s/%s): %s", delay, retry_count + 1, - self.max_retries, str(e)) - await asyncio.sleep(delay) - return await self.request( - method, - path, - params=params, - data=data, - files=files, - headers=headers, - content_type=content_type, - multipart_parser=multipart_parser, - retry_count=retry_count + 1, - ) - # One final connectivity check for diagnostics - connectivity = await self._check_connectivity(self.base_url) - if connectivity["is_local_issue"]: - raise LocalNetworkError( - "Unable to connect to the API server due to local network issues. " - "Please check your internet connection and try again." - ) from e - raise ApiServerError( - f"The API server at {self.base_url} is currently unreachable. " - f"The service may be experiencing issues. Please try again later." - ) from e - - @staticmethod - def _check_auth(auth_token, comfy_api_key): - """Verify that an auth token is present or comfy_api_key is present""" - if auth_token is None and comfy_api_key is None: - raise Exception("Unauthorized: Please login first to use this node.") - return auth_token or comfy_api_key - - @staticmethod - async def upload_file( - upload_url: str, - file: io.BytesIO | str, - content_type: str | None = None, - max_retries: int = 3, - retry_delay: float = 1.0, - retry_backoff_factor: float = 2.0, - ) -> aiohttp.ClientResponse: - """Upload a file to the API with retry logic. - - Args: - upload_url: The URL to upload to - file: Either a file path string, BytesIO object, or tuple of (file_path, filename) - content_type: Optional mime type to set for the upload - max_retries: Maximum number of retry attempts - retry_delay: Initial delay between retries in seconds - retry_backoff_factor: Multiplier for the delay after each retry - """ - headers: dict[str, str] = {} - skip_auto_headers: set[str] = set() - if content_type: - headers["Content-Type"] = content_type - else: - # tell aiohttp not to add Content-Type that will break the request signature and result in a 403 status. - skip_auto_headers.add("Content-Type") - - # Extract file bytes - if isinstance(file, io.BytesIO): - file.seek(0) - data = file.read() - elif isinstance(file, str): - with open(file, "rb") as f: - data = f.read() - else: - raise ValueError("File must be BytesIO or str path") - - parsed = urlparse(upload_url) - basename = os.path.basename(parsed.path) or parsed.netloc or "upload" - operation_id = f"upload_{basename}_{uuid.uuid4().hex[:8]}" - request_logger.log_request_response( - operation_id=operation_id, - request_method="PUT", - request_url=upload_url, - request_headers=headers, - request_data=f"[File data {len(data)} bytes]", - ) - - delay = retry_delay - for attempt in range(max_retries + 1): - try: - timeout = aiohttp.ClientTimeout(total=None) # honour server side timeouts - async with aiohttp.ClientSession(timeout=timeout) as session: - async with session.put( - upload_url, data=data, headers=headers, skip_auto_headers=skip_auto_headers, - ) as resp: - resp.raise_for_status() - request_logger.log_request_response( - operation_id=operation_id, - request_method="PUT", - request_url=upload_url, - response_status_code=resp.status, - response_headers=dict(resp.headers), - response_content="File uploaded successfully.", - ) - return resp - except (ClientError, asyncio.TimeoutError) as e: - request_logger.log_request_response( - operation_id=operation_id, - request_method="PUT", - request_url=upload_url, - response_status_code=e.status if hasattr(e, "status") else None, - response_headers=dict(e.headers) if hasattr(e, "headers") else None, - response_content=None, - error_message=f"{type(e).__name__}: {str(e)}", - ) - if attempt < max_retries: - logging.warning( - "Upload failed (%s/%s). Retrying in %.2fs. %s", attempt + 1, max_retries, delay, str(e) - ) - await asyncio.sleep(delay) - delay *= retry_backoff_factor - else: - raise NetworkError(f"Failed to upload file after {max_retries + 1} attempts: {e}") from e - - async def _handle_http_error( - self, - exc: ClientResponseError, - operation_id: str, - *req_meta, - retry_count: int, - response_content: dict | str = "", - ) -> dict[str, Any]: - status_code = exc.status - if status_code == 401: - user_friendly = "Unauthorized: Please login first to use this node." - elif status_code == 402: - user_friendly = "Payment Required: Please add credits to your account to use this node." - elif status_code == 409: - user_friendly = "There is a problem with your account. Please contact support@comfy.org." - elif status_code == 429: - user_friendly = "Rate Limit Exceeded: Please try again later." - else: - if isinstance(response_content, dict): - if "error" in response_content and "message" in response_content["error"]: - user_friendly = f"API Error: {response_content['error']['message']}" - if "type" in response_content["error"]: - user_friendly += f" (Type: {response_content['error']['type']})" - else: # Handle cases where error is just a JSON dict with unknown format - user_friendly = f"API Error: {json.dumps(response_content)}" - else: - if len(response_content) < 200: # Arbitrary limit for display - user_friendly = f"API Error (raw): {response_content}" - else: - user_friendly = f"API Error (raw, status {response_content})" - - request_logger.log_request_response( - operation_id=operation_id, - request_method=req_meta[0], - request_url=req_meta[1], - response_status_code=exc.status, - response_headers=dict(req_meta[5]) if req_meta[5] else None, - response_content=response_content, - error_message=f"HTTP Error {exc.status}", - ) - - logging.debug("[DEBUG] API Error: %s (Status: %s)", user_friendly, status_code) - if response_content: - logging.debug("[DEBUG] Response content: %s", response_content) - - # Retry if eligible - if status_code in self.retry_status_codes and retry_count < self.max_retries: - delay = self.retry_delay * (self.retry_backoff_factor ** retry_count) - logging.warning( - "HTTP error %s. Retrying in %.2fs (%s/%s)", - status_code, - delay, - retry_count + 1, - self.max_retries, - ) - await asyncio.sleep(delay) - return await self.request( - req_meta[0], # method - req_meta[1].replace(self.base_url, ""), # path - params=req_meta[2], - data=req_meta[3], - files=req_meta[4], - headers=req_meta[5], - content_type=req_meta[6], - multipart_parser=req_meta[7], - retry_count=retry_count + 1, - ) - - raise Exception(user_friendly) from exc - - @staticmethod - def _unpack_tuple(t): - """Helper to normalise (filename, file, content_type) tuples.""" - if len(t) == 3: - return t - elif len(t) == 2: - return t[0], t[1], "application/octet-stream" - else: - raise ValueError("files tuple must be (filename, file[, content_type])") - - async def _get_session(self) -> aiohttp.ClientSession: - if self._session is None or self._session.closed: - timeout = aiohttp.ClientTimeout(total=self.timeout) - self._session = aiohttp.ClientSession(timeout=timeout) - self._owns_session = True - return self._session - - async def close(self) -> None: - if self._owns_session and self._session and not self._session.closed: - await self._session.close() - - async def __aenter__(self) -> "ApiClient": - """Allow usage as async‑context‑manager – ensures clean teardown""" - return self - - async def __aexit__(self, exc_type, exc, tb): - await self.close() - - -class ApiEndpoint(Generic[T, R]): - """Defines an API endpoint with its request and response types""" - - def __init__( - self, - path: str, - method: HttpMethod, - request_model: Type[T], - response_model: Type[R], - query_params: Optional[dict[str, Any]] = None, - ): - """Initialize an API endpoint definition. - - Args: - path: The URL path for this endpoint, can include placeholders like {id} - method: The HTTP method to use (GET, POST, etc.) - request_model: Pydantic model class that defines the structure and validation rules for API requests to this endpoint - response_model: Pydantic model class that defines the structure and validation rules for API responses from this endpoint - query_params: Optional dictionary of query parameters to include in the request - """ - self.path = path - self.method = method - self.request_model = request_model - self.response_model = response_model - self.query_params = query_params or {} - - -class SynchronousOperation(Generic[T, R]): - """Represents a single synchronous API operation.""" - - def __init__( - self, - endpoint: ApiEndpoint[T, R], - request: T, - files: Optional[dict[str, Any] | list[tuple[str, Any]]] = None, - api_base: str | None = None, - auth_token: Optional[str] = None, - comfy_api_key: Optional[str] = None, - auth_kwargs: Optional[dict[str, str]] = None, - timeout: float = 7200.0, - verify_ssl: bool = True, - content_type: str = "application/json", - multipart_parser: Callable | None = None, - max_retries: int = 3, - retry_delay: float = 1.0, - retry_backoff_factor: float = 2.0, - ) -> None: - self.endpoint = endpoint - self.request = request - self.files = files - self.api_base: str = api_base or args.comfy_api_base - self.auth_token = auth_token - self.comfy_api_key = comfy_api_key - if auth_kwargs is not None: - self.auth_token = auth_kwargs.get("auth_token", self.auth_token) - self.comfy_api_key = auth_kwargs.get("comfy_api_key", self.comfy_api_key) - self.timeout = timeout - self.verify_ssl = verify_ssl - self.content_type = content_type - self.multipart_parser = multipart_parser - self.max_retries = max_retries - self.retry_delay = retry_delay - self.retry_backoff_factor = retry_backoff_factor - - async def execute(self, client: Optional[ApiClient] = None) -> R: - owns_client = client is None - if owns_client: - client = ApiClient( - base_url=self.api_base, - auth_token=self.auth_token, - comfy_api_key=self.comfy_api_key, - timeout=self.timeout, - verify_ssl=self.verify_ssl, - max_retries=self.max_retries, - retry_delay=self.retry_delay, - retry_backoff_factor=self.retry_backoff_factor, - ) - - try: - request_dict: Optional[dict[str, Any]] - if isinstance(self.request, EmptyRequest): - request_dict = None - else: - request_dict = self.request.model_dump(exclude_none=True) - for k, v in list(request_dict.items()): - if isinstance(v, Enum): - request_dict[k] = v.value - - logging.debug("[DEBUG] API Request: %s %s", self.endpoint.method.value, self.endpoint.path) - logging.debug("[DEBUG] Request Data: %s", json.dumps(request_dict, indent=2)) - logging.debug("[DEBUG] Query Params: %s", self.endpoint.query_params) - - response_json = await client.request( - self.endpoint.method.value, - self.endpoint.path, - params=self.endpoint.query_params, - data=request_dict, - files=self.files, - content_type=self.content_type, - multipart_parser=self.multipart_parser, - ) - - logging.debug("=" * 50) - logging.debug("[DEBUG] RESPONSE DETAILS:") - logging.debug("[DEBUG] Status Code: 200 (Success)") - logging.debug("[DEBUG] Response Body: %s", json.dumps(response_json, indent=2)) - logging.debug("=" * 50) - - parsed_response = self.endpoint.response_model.model_validate(response_json) - logging.debug("[DEBUG] Parsed Response: %s", parsed_response) - return parsed_response - finally: - if owns_client: - await client.close() - - -class TaskStatus(str, Enum): - """Enum for task status values""" - - COMPLETED = "completed" - FAILED = "failed" - PENDING = "pending" - - -class PollingOperation(Generic[T, R]): - """Represents an asynchronous API operation that requires polling for completion.""" - - def __init__( - self, - poll_endpoint: ApiEndpoint[EmptyRequest, R], - completed_statuses: list[str], - failed_statuses: list[str], - *, - status_extractor: Callable[[R], Optional[str]], - progress_extractor: Callable[[R], Optional[float]] | None = None, - result_url_extractor: Callable[[R], Optional[str]] | None = None, - price_extractor: Callable[[R], Optional[float]] | None = None, - request: Optional[T] = None, - api_base: str | None = None, - auth_token: Optional[str] = None, - comfy_api_key: Optional[str] = None, - auth_kwargs: Optional[dict[str, str]] = None, - poll_interval: float = 5.0, - max_poll_attempts: int = 120, # Default max polling attempts (10 minutes with 5s interval) - max_retries: int = 3, # Max retries per individual API call - retry_delay: float = 1.0, - retry_backoff_factor: float = 2.0, - estimated_duration: Optional[float] = None, - node_id: Optional[str] = None, - ) -> None: - self.poll_endpoint = poll_endpoint - self.request = request - self.api_base: str = api_base or args.comfy_api_base - self.auth_token = auth_token - self.comfy_api_key = comfy_api_key - if auth_kwargs is not None: - self.auth_token = auth_kwargs.get("auth_token", self.auth_token) - self.comfy_api_key = auth_kwargs.get("comfy_api_key", self.comfy_api_key) - self.poll_interval = poll_interval - self.max_poll_attempts = max_poll_attempts - self.max_retries = max_retries - self.retry_delay = retry_delay - self.retry_backoff_factor = retry_backoff_factor - self.estimated_duration = estimated_duration - self.status_extractor = status_extractor or (lambda x: getattr(x, "status", None)) - self.progress_extractor = progress_extractor - self.result_url_extractor = result_url_extractor - self.price_extractor = price_extractor - self.node_id = node_id - self.completed_statuses = completed_statuses - self.failed_statuses = failed_statuses - self.final_response: Optional[R] = None - self.extracted_price: Optional[float] = None - - async def execute(self, client: Optional[ApiClient] = None) -> R: - owns_client = client is None - if owns_client: - client = ApiClient( - base_url=self.api_base, - auth_token=self.auth_token, - comfy_api_key=self.comfy_api_key, - max_retries=self.max_retries, - retry_delay=self.retry_delay, - retry_backoff_factor=self.retry_backoff_factor, - ) - try: - return await self._poll_until_complete(client) - finally: - if owns_client: - await client.close() - - def _display_text_on_node(self, text: str): - if not self.node_id: - return - if self.extracted_price is not None: - text = f"Price: ${self.extracted_price}\n{text}" - PromptServer.instance.send_progress_text(text, self.node_id) - - def _display_time_progress_on_node(self, time_completed: int | float): - if not self.node_id: - return - if self.estimated_duration is not None: - remaining = max(0, int(self.estimated_duration) - time_completed) - message = f"Task in progress: {time_completed}s (~{remaining}s remaining)" - else: - message = f"Task in progress: {time_completed}s" - self._display_text_on_node(message) - - def _check_task_status(self, response: R) -> TaskStatus: - try: - status = self.status_extractor(response) - if status in self.completed_statuses: - return TaskStatus.COMPLETED - if status in self.failed_statuses: - return TaskStatus.FAILED - return TaskStatus.PENDING - except Exception as e: - logging.error("Error extracting status: %s", e) - return TaskStatus.PENDING - - async def _poll_until_complete(self, client: ApiClient) -> R: - """Poll until the task is complete""" - consecutive_errors = 0 - max_consecutive_errors = min(5, self.max_retries * 2) # Limit consecutive errors - - if self.progress_extractor: - progress = utils.ProgressBar(PROGRESS_BAR_MAX) - - status = TaskStatus.PENDING - for poll_count in range(1, self.max_poll_attempts + 1): - try: - logging.debug("[DEBUG] Polling attempt #%s", poll_count) - - request_dict = None if self.request is None else self.request.model_dump(exclude_none=True) - - if poll_count == 1: - logging.debug( - "[DEBUG] Poll Request: %s %s", - self.poll_endpoint.method.value, - self.poll_endpoint.path, - ) - logging.debug( - "[DEBUG] Poll Request Data: %s", - json.dumps(request_dict, indent=2) if request_dict else "None", - ) - - # Query task status - resp = await client.request( - self.poll_endpoint.method.value, - self.poll_endpoint.path, - params=self.poll_endpoint.query_params, - data=request_dict, - ) - consecutive_errors = 0 # reset on success - response_obj: R = self.poll_endpoint.response_model.model_validate(resp) - - # Check if task is complete - status = self._check_task_status(response_obj) - logging.debug("[DEBUG] Task Status: %s", status) - - # If progress extractor is provided, extract progress - if self.progress_extractor: - new_progress = self.progress_extractor(response_obj) - if new_progress is not None: - progress.update_absolute(new_progress, total=PROGRESS_BAR_MAX) - - if self.price_extractor: - price = self.price_extractor(response_obj) - if price is not None: - self.extracted_price = price - - if status == TaskStatus.COMPLETED: - message = "Task completed successfully" - if self.result_url_extractor: - result_url = self.result_url_extractor(response_obj) - if result_url: - message = f"Result URL: {result_url}" - logging.debug("[DEBUG] %s", message) - self._display_text_on_node(message) - self.final_response = response_obj - if self.progress_extractor: - progress.update(100) - return self.final_response - if status == TaskStatus.FAILED: - message = f"Task failed: {json.dumps(resp)}" - logging.error("[DEBUG] %s", message) - raise Exception(message) - logging.debug("[DEBUG] Task still pending, continuing to poll...") - # Task pending – wait - for i in range(int(self.poll_interval)): - self._display_time_progress_on_node((poll_count - 1) * self.poll_interval + i) - await asyncio.sleep(1) - - except (LocalNetworkError, ApiServerError, NetworkError) as e: - consecutive_errors += 1 - if consecutive_errors >= max_consecutive_errors: - raise Exception( - f"Polling aborted after {consecutive_errors} network errors: {str(e)}" - ) from e - logging.warning( - "Network error (%s/%s): %s", - consecutive_errors, - max_consecutive_errors, - str(e), - ) - await asyncio.sleep(self.poll_interval) - except Exception as e: - # For other errors, increment count and potentially abort - consecutive_errors += 1 - if consecutive_errors >= max_consecutive_errors or status == TaskStatus.FAILED: - raise Exception( - f"Polling aborted after {consecutive_errors} consecutive errors: {str(e)}" - ) from e - - logging.error("[DEBUG] Polling error: %s", str(e)) - logging.warning( - "Error during polling (attempt %s/%s): %s. Will retry in %s seconds.", - poll_count, - self.max_poll_attempts, - str(e), - self.poll_interval, - ) - await asyncio.sleep(self.poll_interval) - - # If we've exhausted all polling attempts - raise Exception( - f"Polling timed out after {self.max_poll_attempts} attempts (" f"{self.max_poll_attempts * self.poll_interval} seconds). " - "The operation may still be running on the server but is taking longer than expected." - ) diff --git a/comfy_api_nodes/nodes_rodin.py b/comfy_api_nodes/nodes_rodin.py index ad4029236..e60e7a6d6 100644 --- a/comfy_api_nodes/nodes_rodin.py +++ b/comfy_api_nodes/nodes_rodin.py @@ -5,12 +5,9 @@ Rodin API docs: https://developer.hyper3d.ai/ """ -from __future__ import annotations from inspect import cleandoc import folder_paths as comfy_paths -import aiohttp import os -import asyncio import logging import math from typing import Optional @@ -26,11 +23,11 @@ from comfy_api_nodes.apis.rodin_api import ( Rodin3DDownloadResponse, JobStatus, ) -from comfy_api_nodes.apis.client import ( +from comfy_api_nodes.util import ( + sync_op, + poll_op, ApiEndpoint, - HttpMethod, - SynchronousOperation, - PollingOperation, + download_url_to_bytesio, ) from comfy_api.latest import ComfyExtension, IO @@ -121,35 +118,31 @@ def tensor_to_filelike(tensor, max_pixels: int = 2048*2048): async def create_generate_task( + cls: type[IO.ComfyNode], images=None, seed=1, material="PBR", quality_override=18000, tier="Regular", mesh_mode="Quad", - TAPose = False, - auth_kwargs: Optional[dict[str, str]] = None, + ta_pose: bool = False, ): if images is None: raise Exception("Rodin 3D generate requires at least 1 image.") if len(images) > 5: raise Exception("Rodin 3D generate requires up to 5 image.") - path = "/proxy/rodin/api/v2/rodin" - operation = SynchronousOperation( - endpoint=ApiEndpoint( - path=path, - method=HttpMethod.POST, - request_model=Rodin3DGenerateRequest, - response_model=Rodin3DGenerateResponse, - ), - request=Rodin3DGenerateRequest( + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/rodin/api/v2/rodin", method="POST"), + response_model=Rodin3DGenerateResponse, + data=Rodin3DGenerateRequest( seed=seed, tier=tier, material=material, quality_override=quality_override, mesh_mode=mesh_mode, - TAPose=TAPose, + TAPose=ta_pose, ), files=[ ( @@ -159,11 +152,8 @@ async def create_generate_task( for image in images if image is not None ], content_type="multipart/form-data", - auth_kwargs=auth_kwargs, ) - response = await operation.execute() - if hasattr(response, "error"): error_message = f"Rodin3D Create 3D generate Task Failed. Message: {response.message}, error: {response.error}" logging.error(error_message) @@ -187,74 +177,46 @@ def check_rodin_status(response: Rodin3DCheckStatusResponse) -> str: return "DONE" return "Generating" +def extract_progress(response: Rodin3DCheckStatusResponse) -> Optional[int]: + if not response.jobs: + return None + completed_count = sum(1 for job in response.jobs if job.status == JobStatus.Done) + return int((completed_count / len(response.jobs)) * 100) -async def poll_for_task_status( - subscription_key, auth_kwargs: Optional[dict[str, str]] = None, -) -> Rodin3DCheckStatusResponse: - poll_operation = PollingOperation( - poll_endpoint=ApiEndpoint( - path="/proxy/rodin/api/v2/status", - method=HttpMethod.POST, - request_model=Rodin3DCheckStatusRequest, - response_model=Rodin3DCheckStatusResponse, - ), - request=Rodin3DCheckStatusRequest(subscription_key=subscription_key), - completed_statuses=["DONE"], - failed_statuses=["FAILED"], - status_extractor=check_rodin_status, - poll_interval=3.0, - auth_kwargs=auth_kwargs, - ) + +async def poll_for_task_status(subscription_key: str, cls: type[IO.ComfyNode]) -> Rodin3DCheckStatusResponse: logging.info("[ Rodin3D API - CheckStatus ] Generate Start!") - return await poll_operation.execute() - - -async def get_rodin_download_list(uuid, auth_kwargs: Optional[dict[str, str]] = None) -> Rodin3DDownloadResponse: - logging.info("[ Rodin3D API - Downloading ] Generate Successfully!") - operation = SynchronousOperation( - endpoint=ApiEndpoint( - path="/proxy/rodin/api/v2/download", - method=HttpMethod.POST, - request_model=Rodin3DDownloadRequest, - response_model=Rodin3DDownloadResponse, - ), - request=Rodin3DDownloadRequest(task_uuid=uuid), - auth_kwargs=auth_kwargs, + return await poll_op( + cls, + ApiEndpoint(path="/proxy/rodin/api/v2/status", method="POST"), + response_model=Rodin3DCheckStatusResponse, + data=Rodin3DCheckStatusRequest(subscription_key=subscription_key), + status_extractor=check_rodin_status, + progress_extractor=extract_progress, ) - return await operation.execute() -async def download_files(url_list, task_uuid): +async def get_rodin_download_list(uuid: str, cls: type[IO.ComfyNode]) -> Rodin3DDownloadResponse: + logging.info("[ Rodin3D API - Downloading ] Generate Successfully!") + return await sync_op( + cls, + ApiEndpoint(path="/proxy/rodin/api/v2/download", method="POST"), + response_model=Rodin3DDownloadResponse, + data=Rodin3DDownloadRequest(task_uuid=uuid), + monitor_progress=False, + ) + + +async def download_files(url_list, task_uuid: str): result_folder_name = f"Rodin3D_{task_uuid}" save_path = os.path.join(comfy_paths.get_output_directory(), result_folder_name) os.makedirs(save_path, exist_ok=True) model_file_path = None - async with aiohttp.ClientSession() as session: - for i in url_list.list: - file_path = os.path.join(save_path, i.name) - if file_path.endswith(".glb"): - model_file_path = os.path.join(result_folder_name, i.name) - logging.info("[ Rodin3D API - download_files ] Downloading file: %s", file_path) - max_retries = 5 - for attempt in range(max_retries): - try: - async with session.get(i.url) as resp: - resp.raise_for_status() - with open(file_path, "wb") as f: - async for chunk in resp.content.iter_chunked(32 * 1024): - f.write(chunk) - break - except Exception as e: - logging.info("[ Rodin3D API - download_files ] Error downloading %s:%s", file_path, str(e)) - if attempt < max_retries - 1: - logging.info("Retrying...") - await asyncio.sleep(2) - else: - logging.info( - "[ Rodin3D API - download_files ] Failed to download %s after %s attempts.", - file_path, - max_retries, - ) + for i in url_list.list: + file_path = os.path.join(save_path, i.name) + if file_path.endswith(".glb"): + model_file_path = os.path.join(result_folder_name, i.name) + await download_url_to_bytesio(i.url, file_path) return model_file_path @@ -276,6 +238,7 @@ class Rodin3D_Regular(IO.ComfyNode): hidden=[ IO.Hidden.auth_token_comfy_org, IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, ], is_api_node=True, ) @@ -294,21 +257,17 @@ class Rodin3D_Regular(IO.ComfyNode): for i in range(num_images): m_images.append(Images[i]) mesh_mode, quality_override = get_quality_mode(Polygon_count) - auth = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } task_uuid, subscription_key = await create_generate_task( + cls, images=m_images, seed=Seed, material=Material_Type, quality_override=quality_override, tier=tier, mesh_mode=mesh_mode, - auth_kwargs=auth, ) - await poll_for_task_status(subscription_key, auth_kwargs=auth) - download_list = await get_rodin_download_list(task_uuid, auth_kwargs=auth) + await poll_for_task_status(subscription_key, cls) + download_list = await get_rodin_download_list(task_uuid, cls) model = await download_files(download_list, task_uuid) return IO.NodeOutput(model) @@ -332,6 +291,7 @@ class Rodin3D_Detail(IO.ComfyNode): hidden=[ IO.Hidden.auth_token_comfy_org, IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, ], is_api_node=True, ) @@ -350,21 +310,17 @@ class Rodin3D_Detail(IO.ComfyNode): for i in range(num_images): m_images.append(Images[i]) mesh_mode, quality_override = get_quality_mode(Polygon_count) - auth = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } task_uuid, subscription_key = await create_generate_task( + cls, images=m_images, seed=Seed, material=Material_Type, quality_override=quality_override, tier=tier, mesh_mode=mesh_mode, - auth_kwargs=auth, ) - await poll_for_task_status(subscription_key, auth_kwargs=auth) - download_list = await get_rodin_download_list(task_uuid, auth_kwargs=auth) + await poll_for_task_status(subscription_key, cls) + download_list = await get_rodin_download_list(task_uuid, cls) model = await download_files(download_list, task_uuid) return IO.NodeOutput(model) @@ -388,6 +344,7 @@ class Rodin3D_Smooth(IO.ComfyNode): hidden=[ IO.Hidden.auth_token_comfy_org, IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, ], is_api_node=True, ) @@ -400,27 +357,22 @@ class Rodin3D_Smooth(IO.ComfyNode): Material_Type, Polygon_count, ) -> IO.NodeOutput: - tier = "Smooth" num_images = Images.shape[0] m_images = [] for i in range(num_images): m_images.append(Images[i]) mesh_mode, quality_override = get_quality_mode(Polygon_count) - auth = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } task_uuid, subscription_key = await create_generate_task( + cls, images=m_images, seed=Seed, material=Material_Type, quality_override=quality_override, - tier=tier, + tier="Smooth", mesh_mode=mesh_mode, - auth_kwargs=auth, ) - await poll_for_task_status(subscription_key, auth_kwargs=auth) - download_list = await get_rodin_download_list(task_uuid, auth_kwargs=auth) + await poll_for_task_status(subscription_key, cls) + download_list = await get_rodin_download_list(task_uuid, cls) model = await download_files(download_list, task_uuid) return IO.NodeOutput(model) @@ -451,6 +403,7 @@ class Rodin3D_Sketch(IO.ComfyNode): hidden=[ IO.Hidden.auth_token_comfy_org, IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, ], is_api_node=True, ) @@ -461,29 +414,21 @@ class Rodin3D_Sketch(IO.ComfyNode): Images, Seed, ) -> IO.NodeOutput: - tier = "Sketch" num_images = Images.shape[0] m_images = [] for i in range(num_images): m_images.append(Images[i]) - material_type = "PBR" - quality_override = 18000 - mesh_mode = "Quad" - auth = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } task_uuid, subscription_key = await create_generate_task( + cls, images=m_images, seed=Seed, - material=material_type, - quality_override=quality_override, - tier=tier, - mesh_mode=mesh_mode, - auth_kwargs=auth, + material="PBR", + quality_override=18000, + tier="Sketch", + mesh_mode="Quad", ) - await poll_for_task_status(subscription_key, auth_kwargs=auth) - download_list = await get_rodin_download_list(task_uuid, auth_kwargs=auth) + await poll_for_task_status(subscription_key, cls) + download_list = await get_rodin_download_list(task_uuid, cls) model = await download_files(download_list, task_uuid) return IO.NodeOutput(model) @@ -522,6 +467,7 @@ class Rodin3D_Gen2(IO.ComfyNode): hidden=[ IO.Hidden.auth_token_comfy_org, IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, ], is_api_node=True, ) @@ -541,22 +487,18 @@ class Rodin3D_Gen2(IO.ComfyNode): for i in range(num_images): m_images.append(Images[i]) mesh_mode, quality_override = get_quality_mode(Polygon_count) - auth = { - "auth_token": cls.hidden.auth_token_comfy_org, - "comfy_api_key": cls.hidden.api_key_comfy_org, - } task_uuid, subscription_key = await create_generate_task( + cls, images=m_images, seed=Seed, material=Material_Type, quality_override=quality_override, tier=tier, mesh_mode=mesh_mode, - TAPose=TAPose, - auth_kwargs=auth, + ta_pose=TAPose, ) - await poll_for_task_status(subscription_key, auth_kwargs=auth) - download_list = await get_rodin_download_list(task_uuid, auth_kwargs=auth) + await poll_for_task_status(subscription_key, cls) + download_list = await get_rodin_download_list(task_uuid, cls) model = await download_files(download_list, task_uuid) return IO.NodeOutput(model) diff --git a/comfy_api_nodes/util/client.py b/comfy_api_nodes/util/client.py index 65bb35f0f..2d5dcd648 100644 --- a/comfy_api_nodes/util/client.py +++ b/comfy_api_nodes/util/client.py @@ -16,9 +16,9 @@ from pydantic import BaseModel from comfy import utils from comfy_api.latest import IO -from comfy_api_nodes.apis import request_logger from server import PromptServer +from . import request_logger from ._helpers import ( default_base_url, get_auth_header, @@ -77,7 +77,7 @@ class _PollUIState: _RETRY_STATUS = {408, 429, 500, 502, 503, 504} -COMPLETED_STATUSES = ["succeeded", "succeed", "success", "completed", "finished"] +COMPLETED_STATUSES = ["succeeded", "succeed", "success", "completed", "finished", "done"] FAILED_STATUSES = ["cancelled", "canceled", "fail", "failed", "error"] QUEUED_STATUSES = ["created", "queued", "queueing", "submitted"] diff --git a/comfy_api_nodes/util/download_helpers.py b/comfy_api_nodes/util/download_helpers.py index 364874bed..14207dc68 100644 --- a/comfy_api_nodes/util/download_helpers.py +++ b/comfy_api_nodes/util/download_helpers.py @@ -12,8 +12,8 @@ from aiohttp.client_exceptions import ClientError, ContentTypeError from comfy_api.input_impl import VideoFromFile from comfy_api.latest import IO as COMFY_IO -from comfy_api_nodes.apis import request_logger +from . import request_logger from ._helpers import ( default_base_url, get_auth_header, diff --git a/comfy_api_nodes/apis/request_logger.py b/comfy_api_nodes/util/request_logger.py similarity index 100% rename from comfy_api_nodes/apis/request_logger.py rename to comfy_api_nodes/util/request_logger.py index c6974d35c..ac52e2eab 100644 --- a/comfy_api_nodes/apis/request_logger.py +++ b/comfy_api_nodes/util/request_logger.py @@ -1,11 +1,11 @@ from __future__ import annotations -import os import datetime +import hashlib import json import logging +import os import re -import hashlib from typing import Any import folder_paths diff --git a/comfy_api_nodes/util/upload_helpers.py b/comfy_api_nodes/util/upload_helpers.py index 7bfc61704..632450d9b 100644 --- a/comfy_api_nodes/util/upload_helpers.py +++ b/comfy_api_nodes/util/upload_helpers.py @@ -13,8 +13,8 @@ from pydantic import BaseModel, Field from comfy_api.latest import IO, Input from comfy_api.util import VideoCodec, VideoContainer -from comfy_api_nodes.apis import request_logger +from . import request_logger from ._helpers import is_processing_interrupted, sleep_with_interrupt from .client import ( ApiEndpoint, From 97f198e4215680a83749ba95849f3cdcfa7aa64a Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 5 Nov 2025 15:07:35 -0800 Subject: [PATCH 087/647] Fix qwen controlnet regression. (#10657) --- comfy/ldm/qwen_image/controlnet.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy/ldm/qwen_image/controlnet.py b/comfy/ldm/qwen_image/controlnet.py index 92ac3cf0a..a6d408104 100644 --- a/comfy/ldm/qwen_image/controlnet.py +++ b/comfy/ldm/qwen_image/controlnet.py @@ -44,7 +44,7 @@ class QwenImageControlNetModel(QwenImageTransformer2DModel): txt_start = round(max(((x.shape[-1] + (self.patch_size // 2)) // self.patch_size) // 2, ((x.shape[-2] + (self.patch_size // 2)) // self.patch_size) // 2)) txt_ids = torch.arange(txt_start, txt_start + context.shape[1], device=x.device).reshape(1, -1, 1).repeat(x.shape[0], 1, 3) ids = torch.cat((txt_ids, img_ids), dim=1) - image_rotary_emb = self.pe_embedder(ids).squeeze(1).unsqueeze(2).to(x.dtype) + image_rotary_emb = self.pe_embedder(ids).to(x.dtype).contiguous() del ids, txt_ids, img_ids hidden_states = self.img_in(hidden_states) + self.controlnet_x_embedder(hint) From 1d69245981f9fb3861018613246042296d887dd3 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 5 Nov 2025 15:08:13 -0800 Subject: [PATCH 088/647] Enable pinned memory by default on Nvidia. (#10656) Removed the --fast pinned_memory flag. You can use --disable-pinned-memory to disable it. Please report if it causes any issues. --- comfy/cli_args.py | 3 ++- comfy/model_management.py | 22 +++++++++------------- 2 files changed, 11 insertions(+), 14 deletions(-) diff --git a/comfy/cli_args.py b/comfy/cli_args.py index 3947e62a8..2f30b72d2 100644 --- a/comfy/cli_args.py +++ b/comfy/cli_args.py @@ -145,10 +145,11 @@ class PerformanceFeature(enum.Enum): Fp8MatrixMultiplication = "fp8_matrix_mult" CublasOps = "cublas_ops" AutoTune = "autotune" - PinnedMem = "pinned_memory" parser.add_argument("--fast", nargs="*", type=PerformanceFeature, help="Enable some untested and potentially quality deteriorating optimizations. This is used to test new features so using it might crash your comfyui. --fast with no arguments enables everything. You can pass a list specific optimizations if you only want to enable specific ones. Current valid optimizations: {}".format(" ".join(map(lambda c: c.value, PerformanceFeature)))) +parser.add_argument("--disable-pinned-memory", action="store_true", help="Disable pinned memory use.") + parser.add_argument("--mmap-torch-files", action="store_true", help="Use mmap when loading ckpt/pt files.") parser.add_argument("--disable-mmap", action="store_true", help="Don't use mmap when loading safetensors.") diff --git a/comfy/model_management.py b/comfy/model_management.py index 0d040e55e..4d13c52c1 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -1085,22 +1085,21 @@ def cast_to_device(tensor, device, dtype, copy=False): PINNED_MEMORY = {} TOTAL_PINNED_MEMORY = 0 -if PerformanceFeature.PinnedMem in args.fast: - if WINDOWS: - MAX_PINNED_MEMORY = get_total_memory(torch.device("cpu")) * 0.45 # Windows limit is apparently 50% - else: - MAX_PINNED_MEMORY = get_total_memory(torch.device("cpu")) * 0.95 -else: - MAX_PINNED_MEMORY = -1 +MAX_PINNED_MEMORY = -1 +if not args.disable_pinned_memory: + if is_nvidia(): + if WINDOWS: + MAX_PINNED_MEMORY = get_total_memory(torch.device("cpu")) * 0.45 # Windows limit is apparently 50% + else: + MAX_PINNED_MEMORY = get_total_memory(torch.device("cpu")) * 0.95 + logging.info("Enabled pinned memory {}".format(MAX_PINNED_MEMORY // (1024 * 1024))) + def pin_memory(tensor): global TOTAL_PINNED_MEMORY if MAX_PINNED_MEMORY <= 0: return False - if not is_nvidia(): - return False - if not is_device_cpu(tensor.device): return False @@ -1121,9 +1120,6 @@ def unpin_memory(tensor): if MAX_PINNED_MEMORY <= 0: return False - if not is_nvidia(): - return False - if not is_device_cpu(tensor.device): return False From 09dc24c8a982776abd5cb2f71e3d041139e1d5b2 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 5 Nov 2025 16:11:15 -0800 Subject: [PATCH 089/647] Pinned mem also seems to work on AMD. (#10658) --- comfy/model_management.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy/model_management.py b/comfy/model_management.py index 4d13c52c1..7a30c4bec 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -1087,7 +1087,7 @@ PINNED_MEMORY = {} TOTAL_PINNED_MEMORY = 0 MAX_PINNED_MEMORY = -1 if not args.disable_pinned_memory: - if is_nvidia(): + if is_nvidia() or is_amd(): if WINDOWS: MAX_PINNED_MEMORY = get_total_memory(torch.device("cpu")) * 0.45 # Windows limit is apparently 50% else: From e05c90712670fa4a2ffebd44046fc78747193a36 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Thu, 6 Nov 2025 01:11:30 -0800 Subject: [PATCH 090/647] Clarify release cycle. (#10667) --- README.md | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-) diff --git a/README.md b/README.md index 4204777e9..8142f595b 100644 --- a/README.md +++ b/README.md @@ -112,10 +112,11 @@ Workflow examples can be found on the [Examples page](https://comfyanonymous.git ## Release Process -ComfyUI follows a weekly release cycle targeting Friday but this regularly changes because of model releases or large changes to the codebase. There are three interconnected repositories: +ComfyUI follows a weekly release cycle targeting Monday but this regularly changes because of model releases or large changes to the codebase. There are three interconnected repositories: 1. **[ComfyUI Core](https://github.com/comfyanonymous/ComfyUI)** - - Releases a new stable version (e.g., v0.7.0) + - Releases a new stable version (e.g., v0.7.0) roughly every week. + - Commits outside of the stable release tags may be very unstable and break many custom nodes. - Serves as the foundation for the desktop release 2. **[ComfyUI Desktop](https://github.com/Comfy-Org/desktop)** From eb1c42f6498ce44aef4dbed3bb665ac98a28254d Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Thu, 6 Nov 2025 17:24:28 -0800 Subject: [PATCH 091/647] Tell users they need to upload their logs in bug reports. (#10671) --- .github/ISSUE_TEMPLATE/bug-report.yml | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-) diff --git a/.github/ISSUE_TEMPLATE/bug-report.yml b/.github/ISSUE_TEMPLATE/bug-report.yml index 3cf2717b7..6556677e0 100644 --- a/.github/ISSUE_TEMPLATE/bug-report.yml +++ b/.github/ISSUE_TEMPLATE/bug-report.yml @@ -8,13 +8,15 @@ body: Before submitting a **Bug Report**, please ensure the following: - **1:** You are running the latest version of ComfyUI. - - **2:** You have looked at the existing bug reports and made sure this isn't already reported. + - **2:** You have your ComfyUI logs and relevant workflow on hand and will post them in this bug report. - **3:** You confirmed that the bug is not caused by a custom node. You can disable all custom nodes by passing - `--disable-all-custom-nodes` command line argument. + `--disable-all-custom-nodes` command line argument. If you have custom node try updating them to the latest version. - **4:** This is an actual bug in ComfyUI, not just a support question. A bug is when you can specify exact steps to replicate what went wrong and others will be able to repeat your steps and see the same issue happen. - If unsure, ask on the [ComfyUI Matrix Space](https://app.element.io/#/room/%23comfyui_space%3Amatrix.org) or the [Comfy Org Discord](https://discord.gg/comfyorg) first. + ## Very Important + + Please make sure that you post ALL your ComfyUI logs in the bug report. A bug report without logs will likely be ignored. - type: checkboxes id: custom-nodes-test attributes: From cf97b033ee80cf245b4592d42f89e6de67e409a4 Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Fri, 7 Nov 2025 12:20:48 +1000 Subject: [PATCH 092/647] mm: guard against double pin and unpin explicitly (#10672) As commented, if you let cuda be the one to detect double pin/unpinning it actually creates an asyc GPU error. --- comfy/model_management.py | 12 ++++++++++++ 1 file changed, 12 insertions(+) diff --git a/comfy/model_management.py b/comfy/model_management.py index 7a30c4bec..a13b24cea 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -1103,6 +1103,12 @@ def pin_memory(tensor): if not is_device_cpu(tensor.device): return False + if tensor.is_pinned(): + #NOTE: Cuda does detect when a tensor is already pinned and would + #error below, but there are proven cases where this also queues an error + #on the GPU async. So dont trust the CUDA API and guard here + return False + size = tensor.numel() * tensor.element_size() if (TOTAL_PINNED_MEMORY + size) > MAX_PINNED_MEMORY: return False @@ -1123,6 +1129,12 @@ def unpin_memory(tensor): if not is_device_cpu(tensor.device): return False + if not tensor.is_pinned(): + #NOTE: Cuda does detect when a tensor is already pinned and would + #error below, but there are proven cases where this also queues an error + #on the GPU async. So dont trust the CUDA API and guard here + return False + ptr = tensor.data_ptr() if torch.cuda.cudart().cudaHostUnregister(ptr) == 0: TOTAL_PINNED_MEMORY -= PINNED_MEMORY.pop(ptr) From a1a70362ca376cff05a0514e0ce771ab26d92fd9 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Fri, 7 Nov 2025 08:15:05 -0800 Subject: [PATCH 093/647] Only unpin tensor if it was pinned by ComfyUI (#10677) --- comfy/model_management.py | 15 ++++++++++----- 1 file changed, 10 insertions(+), 5 deletions(-) diff --git a/comfy/model_management.py b/comfy/model_management.py index a13b24cea..7012df858 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -1129,13 +1129,18 @@ def unpin_memory(tensor): if not is_device_cpu(tensor.device): return False - if not tensor.is_pinned(): - #NOTE: Cuda does detect when a tensor is already pinned and would - #error below, but there are proven cases where this also queues an error - #on the GPU async. So dont trust the CUDA API and guard here + ptr = tensor.data_ptr() + size = tensor.numel() * tensor.element_size() + + size_stored = PINNED_MEMORY.get(ptr, None) + if size_stored is None: + logging.warning("Tried to unpin tensor not pinned by ComfyUI") + return False + + if size != size_stored: + logging.warning("Size of pinned tensor changed") return False - ptr = tensor.data_ptr() if torch.cuda.cudart().cudaHostUnregister(ptr) == 0: TOTAL_PINNED_MEMORY -= PINNED_MEMORY.pop(ptr) if len(PINNED_MEMORY) == 0: From 2abd2b5c2049a9625b342bcb7decedd5d1645f66 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sat, 8 Nov 2025 12:52:02 -0800 Subject: [PATCH 094/647] Make ScaleROPE node work on Flux. (#10686) --- comfy/ldm/flux/model.py | 22 +++++++++++++++++----- 1 file changed, 17 insertions(+), 5 deletions(-) diff --git a/comfy/ldm/flux/model.py b/comfy/ldm/flux/model.py index 14f90cea5..b9d36f202 100644 --- a/comfy/ldm/flux/model.py +++ b/comfy/ldm/flux/model.py @@ -210,7 +210,7 @@ class Flux(nn.Module): img = self.final_layer(img, vec) # (N, T, patch_size ** 2 * out_channels) return img - def process_img(self, x, index=0, h_offset=0, w_offset=0): + def process_img(self, x, index=0, h_offset=0, w_offset=0, transformer_options={}): bs, c, h, w = x.shape patch_size = self.patch_size x = comfy.ldm.common_dit.pad_to_patch_size(x, (patch_size, patch_size)) @@ -222,10 +222,22 @@ class Flux(nn.Module): h_offset = ((h_offset + (patch_size // 2)) // patch_size) w_offset = ((w_offset + (patch_size // 2)) // patch_size) - img_ids = torch.zeros((h_len, w_len, 3), device=x.device, dtype=x.dtype) + steps_h = h_len + steps_w = w_len + + rope_options = transformer_options.get("rope_options", None) + if rope_options is not None: + h_len = (h_len - 1.0) * rope_options.get("scale_y", 1.0) + 1.0 + w_len = (w_len - 1.0) * rope_options.get("scale_x", 1.0) + 1.0 + + index += rope_options.get("shift_t", 0.0) + h_offset += rope_options.get("shift_y", 0.0) + w_offset += rope_options.get("shift_x", 0.0) + + img_ids = torch.zeros((steps_h, steps_w, 3), device=x.device, dtype=x.dtype) img_ids[:, :, 0] = img_ids[:, :, 1] + index - img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(h_offset, h_len - 1 + h_offset, steps=h_len, device=x.device, dtype=x.dtype).unsqueeze(1) - img_ids[:, :, 2] = img_ids[:, :, 2] + torch.linspace(w_offset, w_len - 1 + w_offset, steps=w_len, device=x.device, dtype=x.dtype).unsqueeze(0) + img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(h_offset, h_len - 1 + h_offset, steps=steps_h, device=x.device, dtype=x.dtype).unsqueeze(1) + img_ids[:, :, 2] = img_ids[:, :, 2] + torch.linspace(w_offset, w_len - 1 + w_offset, steps=steps_w, device=x.device, dtype=x.dtype).unsqueeze(0) return img, repeat(img_ids, "h w c -> b (h w) c", b=bs) def forward(self, x, timestep, context, y=None, guidance=None, ref_latents=None, control=None, transformer_options={}, **kwargs): @@ -241,7 +253,7 @@ class Flux(nn.Module): h_len = ((h_orig + (patch_size // 2)) // patch_size) w_len = ((w_orig + (patch_size // 2)) // patch_size) - img, img_ids = self.process_img(x) + img, img_ids = self.process_img(x, transformer_options=transformer_options) img_tokens = img.shape[1] if ref_latents is not None: h = 0 From e632e5de281b91dd7199636dd6d82126fbfb07d5 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sun, 9 Nov 2025 15:06:39 -0800 Subject: [PATCH 095/647] Add logging for model unloading. (#10692) --- comfy/model_patcher.py | 1 + 1 file changed, 1 insertion(+) diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index 5a31a8734..17e06a869 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -909,6 +909,7 @@ class ModelPatcher: self.model.model_lowvram = True self.model.lowvram_patch_counter += patch_counter self.model.model_loaded_weight_memory -= memory_freed + logging.info("loaded partially: {:.2f} MB loaded, lowvram patches: {}".format(self.model.model_loaded_weight_memory / (1024 * 1024), self.model.lowvram_patch_counter)) return memory_freed def partially_load(self, device_to, extra_memory=0, force_patch_weights=False): From dea899f22125d38a8b48147d6cce89a2b659fdeb Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sun, 9 Nov 2025 15:51:33 -0800 Subject: [PATCH 096/647] Unload weights if vram usage goes up between runs. (#10690) --- comfy/model_management.py | 11 +++++++++-- comfy/model_patcher.py | 20 +++++++++++++------- 2 files changed, 22 insertions(+), 9 deletions(-) diff --git a/comfy/model_management.py b/comfy/model_management.py index 7012df858..a4410f2ec 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -503,7 +503,11 @@ class LoadedModel: use_more_vram = lowvram_model_memory if use_more_vram == 0: use_more_vram = 1e32 - self.model_use_more_vram(use_more_vram, force_patch_weights=force_patch_weights) + if use_more_vram > 0: + self.model_use_more_vram(use_more_vram, force_patch_weights=force_patch_weights) + else: + self.model.partially_unload(self.model.offload_device, -use_more_vram, force_patch_weights=force_patch_weights) + real_model = self.model.model if is_intel_xpu() and not args.disable_ipex_optimize and 'ipex' in globals() and real_model is not None: @@ -689,7 +693,10 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimu current_free_mem = get_free_memory(torch_dev) + loaded_memory lowvram_model_memory = max(128 * 1024 * 1024, (current_free_mem - minimum_memory_required), min(current_free_mem * MIN_WEIGHT_MEMORY_RATIO, current_free_mem - minimum_inference_memory())) - lowvram_model_memory = max(0.1, lowvram_model_memory - loaded_memory) + lowvram_model_memory = lowvram_model_memory - loaded_memory + + if lowvram_model_memory == 0: + lowvram_model_memory = 0.1 if vram_set_state == VRAMState.NO_VRAM: lowvram_model_memory = 0.1 diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index 17e06a869..68b0a9192 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -843,7 +843,7 @@ class ModelPatcher: self.object_patches_backup.clear() - def partially_unload(self, device_to, memory_to_free=0): + def partially_unload(self, device_to, memory_to_free=0, force_patch_weights=False): with self.use_ejected(): hooks_unpatched = False memory_freed = 0 @@ -887,13 +887,19 @@ class ModelPatcher: module_mem += move_weight_functions(m, device_to) if lowvram_possible: if weight_key in self.patches: - _, set_func, convert_func = get_key_weight(self.model, weight_key) - m.weight_function.append(LowVramPatch(weight_key, self.patches, convert_func, set_func)) - patch_counter += 1 + if force_patch_weights: + self.patch_weight_to_device(weight_key) + else: + _, set_func, convert_func = get_key_weight(self.model, weight_key) + m.weight_function.append(LowVramPatch(weight_key, self.patches, convert_func, set_func)) + patch_counter += 1 if bias_key in self.patches: - _, set_func, convert_func = get_key_weight(self.model, bias_key) - m.bias_function.append(LowVramPatch(bias_key, self.patches, convert_func, set_func)) - patch_counter += 1 + if force_patch_weights: + self.patch_weight_to_device(bias_key) + else: + _, set_func, convert_func = get_key_weight(self.model, bias_key) + m.bias_function.append(LowVramPatch(bias_key, self.patches, convert_func, set_func)) + patch_counter += 1 cast_weight = True if cast_weight: From c350009236e5d172a3050c04043ea70a301378ca Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Mon, 10 Nov 2025 13:52:11 +1000 Subject: [PATCH 097/647] ops: Put weight cast on the offload stream (#10697) This needs to be on the offload stream. This reproduced a black screen with low resolution images on a slow bus when using FP8. --- comfy/ops.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/comfy/ops.py b/comfy/ops.py index 733bff99d..96dffa85d 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -110,9 +110,9 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, of for f in s.bias_function: bias = f(bias) - weight = weight.to(dtype=dtype) - if weight_has_function: + if weight_has_function or weight.dtype != dtype: with wf_context: + weight = weight.to(dtype=dtype) for f in s.weight_function: weight = f(weight) From 5ebcab3c7d974963a89cecd37296a22fdb73bd2b Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 10 Nov 2025 12:35:29 -0800 Subject: [PATCH 098/647] Update CI workflow to remove dead macOS runner. (#10704) * Update CI workflow to remove dead macOS runner. * revert * revert --- .github/workflows/test-ci.yml | 20 +++++++++++--------- 1 file changed, 11 insertions(+), 9 deletions(-) diff --git a/.github/workflows/test-ci.yml b/.github/workflows/test-ci.yml index 418dca0ab..1660ec8e3 100644 --- a/.github/workflows/test-ci.yml +++ b/.github/workflows/test-ci.yml @@ -21,14 +21,15 @@ jobs: fail-fast: false matrix: # os: [macos, linux, windows] - os: [macos, linux] - python_version: ["3.9", "3.10", "3.11", "3.12"] + # os: [macos, linux] + os: [linux] + python_version: ["3.10", "3.11", "3.12"] cuda_version: ["12.1"] torch_version: ["stable"] include: - - os: macos - runner_label: [self-hosted, macOS] - flags: "--use-pytorch-cross-attention" + # - os: macos + # runner_label: [self-hosted, macOS] + # flags: "--use-pytorch-cross-attention" - os: linux runner_label: [self-hosted, Linux] flags: "" @@ -73,14 +74,15 @@ jobs: strategy: fail-fast: false matrix: - os: [macos, linux] + # os: [macos, linux] + os: [linux] python_version: ["3.11"] cuda_version: ["12.1"] torch_version: ["nightly"] include: - - os: macos - runner_label: [self-hosted, macOS] - flags: "--use-pytorch-cross-attention" + # - os: macos + # runner_label: [self-hosted, macOS] + # flags: "--use-pytorch-cross-attention" - os: linux runner_label: [self-hosted, Linux] flags: "" From 119941174704081a16a4c3f303d99f2fb1e95cde Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 11 Nov 2025 16:33:30 -0800 Subject: [PATCH 099/647] Don't pin tensor if not a torch.nn.parameter.Parameter (#10718) --- comfy/model_management.py | 6 ++++++ 1 file changed, 6 insertions(+) diff --git a/comfy/model_management.py b/comfy/model_management.py index a4410f2ec..d8913082a 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -1107,6 +1107,9 @@ def pin_memory(tensor): if MAX_PINNED_MEMORY <= 0: return False + if type(tensor) is not torch.nn.parameter.Parameter: + return False + if not is_device_cpu(tensor.device): return False @@ -1116,6 +1119,9 @@ def pin_memory(tensor): #on the GPU async. So dont trust the CUDA API and guard here return False + if not tensor.is_contiguous(): + return False + size = tensor.numel() * tensor.element_size() if (TOTAL_PINNED_MEMORY + size) > MAX_PINNED_MEMORY: return False From e1d85e7577d8f6355bd4cb3449bcb0a7e5f80cb8 Mon Sep 17 00:00:00 2001 From: Qiacheng Li Date: Wed, 12 Nov 2025 12:21:05 -0800 Subject: [PATCH 100/647] Update README.md for Intel Arc GPU installation, remove IPEX (#10729) IPEX is no longer needed for Intel Arc GPUs. Removing instruction to setup ipex. --- README.md | 6 +----- 1 file changed, 1 insertion(+), 5 deletions(-) diff --git a/README.md b/README.md index 8142f595b..9e28803a2 100644 --- a/README.md +++ b/README.md @@ -242,7 +242,7 @@ RDNA 4 (RX 9000 series): ### Intel GPUs (Windows and Linux) -(Option 1) Intel Arc GPU users can install native PyTorch with torch.xpu support using pip. More information can be found [here](https://pytorch.org/docs/main/notes/get_start_xpu.html) +Intel Arc GPU users can install native PyTorch with torch.xpu support using pip. More information can be found [here](https://pytorch.org/docs/main/notes/get_start_xpu.html) 1. To install PyTorch xpu, use the following command: @@ -252,10 +252,6 @@ This is the command to install the Pytorch xpu nightly which might have some per ```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/xpu``` -(Option 2) Alternatively, Intel GPUs supported by Intel Extension for PyTorch (IPEX) can leverage IPEX for improved performance. - -1. visit [Installation](https://intel.github.io/intel-extension-for-pytorch/index.html#installation?platform=gpu) for more information. - ### NVIDIA Nvidia users should install stable pytorch using this command: From 18e7d6dba5f1012d4cf09e8f777dc85d56ff25c0 Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Thu, 13 Nov 2025 07:19:53 +1000 Subject: [PATCH 101/647] mm/mp: always unload re-used but modified models (#10724) The partial unloader path in model re-use flow skips straight to the actual unload without any check of the patching UUID. This means that if you do an upscale flow with a model patch on an existing model, it will not apply your patchings. Fix by delaying the partial_unload until after the uuid checks. This is done by making partial_unload a model of partial_load where extra_mem is -ve. --- comfy/model_management.py | 5 +---- comfy/model_patcher.py | 3 +++ 2 files changed, 4 insertions(+), 4 deletions(-) diff --git a/comfy/model_management.py b/comfy/model_management.py index d8913082a..a21df54b3 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -503,10 +503,7 @@ class LoadedModel: use_more_vram = lowvram_model_memory if use_more_vram == 0: use_more_vram = 1e32 - if use_more_vram > 0: - self.model_use_more_vram(use_more_vram, force_patch_weights=force_patch_weights) - else: - self.model.partially_unload(self.model.offload_device, -use_more_vram, force_patch_weights=force_patch_weights) + self.model_use_more_vram(use_more_vram, force_patch_weights=force_patch_weights) real_model = self.model.model diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index 68b0a9192..cf1b0d441 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -928,6 +928,9 @@ class ModelPatcher: extra_memory += (used - self.model.model_loaded_weight_memory) self.patch_model(load_weights=False) + if extra_memory < 0 and not unpatch_weights: + self.partially_unload(self.offload_device, -extra_memory, force_patch_weights=force_patch_weights) + return 0 full_load = False if self.model.model_lowvram == False and self.model.model_loaded_weight_memory > 0: self.apply_hooks(self.forced_hooks, force_apply=True) From 1c7eaeca1013e4315f36e0d4d274faa106001121 Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Thu, 13 Nov 2025 07:20:53 +1000 Subject: [PATCH 102/647] qwen: reduce VRAM usage (#10725) Clean up a bunch of stacked and no-longer-needed tensors on the QWEN VRAM peak (currently FFN). With this I go from OOMing at B=37x1328x1328 to being able to succesfully run B=47 (RTX5090). --- comfy/ldm/qwen_image/model.py | 20 ++++++++++++-------- 1 file changed, 12 insertions(+), 8 deletions(-) diff --git a/comfy/ldm/qwen_image/model.py b/comfy/ldm/qwen_image/model.py index e5d0d17c1..427ea19c1 100644 --- a/comfy/ldm/qwen_image/model.py +++ b/comfy/ldm/qwen_image/model.py @@ -236,10 +236,10 @@ class QwenImageTransformerBlock(nn.Module): img_mod1, img_mod2 = img_mod_params.chunk(2, dim=-1) txt_mod1, txt_mod2 = txt_mod_params.chunk(2, dim=-1) - img_normed = self.img_norm1(hidden_states) - img_modulated, img_gate1 = self._modulate(img_normed, img_mod1) - txt_normed = self.txt_norm1(encoder_hidden_states) - txt_modulated, txt_gate1 = self._modulate(txt_normed, txt_mod1) + img_modulated, img_gate1 = self._modulate(self.img_norm1(hidden_states), img_mod1) + del img_mod1 + txt_modulated, txt_gate1 = self._modulate(self.txt_norm1(encoder_hidden_states), txt_mod1) + del txt_mod1 img_attn_output, txt_attn_output = self.attn( hidden_states=img_modulated, @@ -248,16 +248,20 @@ class QwenImageTransformerBlock(nn.Module): image_rotary_emb=image_rotary_emb, transformer_options=transformer_options, ) + del img_modulated + del txt_modulated hidden_states = hidden_states + img_gate1 * img_attn_output encoder_hidden_states = encoder_hidden_states + txt_gate1 * txt_attn_output + del img_attn_output + del txt_attn_output + del img_gate1 + del txt_gate1 - img_normed2 = self.img_norm2(hidden_states) - img_modulated2, img_gate2 = self._modulate(img_normed2, img_mod2) + img_modulated2, img_gate2 = self._modulate(self.img_norm2(hidden_states), img_mod2) hidden_states = torch.addcmul(hidden_states, img_gate2, self.img_mlp(img_modulated2)) - txt_normed2 = self.txt_norm2(encoder_hidden_states) - txt_modulated2, txt_gate2 = self._modulate(txt_normed2, txt_mod2) + txt_modulated2, txt_gate2 = self._modulate(self.txt_norm2(encoder_hidden_states), txt_mod2) encoder_hidden_states = torch.addcmul(encoder_hidden_states, txt_gate2, self.txt_mlp(txt_modulated2)) return encoder_hidden_states, hidden_states From 8b0b93df51d04f08eb779cb84dc331fa18b43ae8 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 12 Nov 2025 14:04:41 -0800 Subject: [PATCH 103/647] Update Python 3.14 compatibility notes in README (#10730) --- README.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/README.md b/README.md index 9e28803a2..f51807ad5 100644 --- a/README.md +++ b/README.md @@ -200,7 +200,7 @@ comfy install ## Manual Install (Windows, Linux) -Python 3.14 will work if you comment out the `kornia` dependency in the requirements.txt file (breaks the canny node) but it is not recommended. +Python 3.14 works but you may encounter issues with the torch compile node. The free threaded variant is still missing some dependencies. Python 3.13 is very well supported. If you have trouble with some custom node dependencies on 3.13 you can try 3.12 From 3b3ef9a77ac03ed516a45063f9736f33085cecca Mon Sep 17 00:00:00 2001 From: contentis Date: Thu, 13 Nov 2025 00:26:52 +0100 Subject: [PATCH 104/647] Quantized Ops fixes (#10715) * offload support, bug fixes, remove mixins * add readme --- QUANTIZATION.md | 168 +++++++++++++++++++++++++++++++++++++++++++++ comfy/ops.py | 37 ++++------ comfy/quant_ops.py | 39 ++++++++++- 3 files changed, 219 insertions(+), 25 deletions(-) create mode 100644 QUANTIZATION.md diff --git a/QUANTIZATION.md b/QUANTIZATION.md new file mode 100644 index 000000000..1693e13f3 --- /dev/null +++ b/QUANTIZATION.md @@ -0,0 +1,168 @@ +# The Comfy guide to Quantization + + +## How does quantization work? + +Quantization aims to map a high-precision value x_f to a lower precision format with minimal loss in accuracy. These smaller formats then serve to reduce the models memory footprint and increase throughput by using specialized hardware. + +When simply converting a value from FP16 to FP8 using the round-nearest method we might hit two issues: +- The dynamic range of FP16 (-65,504, 65,504) far exceeds FP8 formats like E4M3 (-448, 448) or E5M2 (-57,344, 57,344), potentially resulting in clipped values +- The original values are concentrated in a small range (e.g. -1,1) leaving many FP8-bits "unused" + +By using a scaling factor, we aim to map these values into the quantized-dtype range, making use of the full spectrum. One of the easiest approaches, and common, is using per-tensor absolute-maximum scaling. + +``` +absmax = max(abs(tensor)) +scale = amax / max_dynamic_range_low_precision + +# Quantization +tensor_q = (tensor / scale).to(low_precision_dtype) + +# De-Quantization +tensor_dq = tensor_q.to(fp16) * scale + +tensor_dq ~ tensor +``` + +Given that additional information (scaling factor) is needed to "interpret" the quantized values, we describe those as derived datatypes. + + +## Quantization in Comfy + +``` +QuantizedTensor (torch.Tensor subclass) + ↓ __torch_dispatch__ +Two-Level Registry (generic + layout handlers) + ↓ +MixedPrecisionOps + Metadata Detection +``` + +### Representation + +To represent these derived datatypes, ComfyUI uses a subclass of torch.Tensor to implements these using the `QuantizedTensor` class found in `comfy/quant_ops.py` + +A `Layout` class defines how a specific quantization format behaves: +- Required parameters +- Quantize method +- De-Quantize method + +```python +from comfy.quant_ops import QuantizedLayout + +class MyLayout(QuantizedLayout): + @classmethod + def quantize(cls, tensor, **kwargs): + # Convert to quantized format + qdata = ... + params = {'scale': ..., 'orig_dtype': tensor.dtype} + return qdata, params + + @staticmethod + def dequantize(qdata, scale, orig_dtype, **kwargs): + return qdata.to(orig_dtype) * scale +``` + +To then run operations using these QuantizedTensors we use two registry systems to define supported operations. +The first is a **generic registry** that handles operations common to all quantized formats (e.g., `.to()`, `.clone()`, `.reshape()`). + +The second registry is layout-specific and allows to implement fast-paths like nn.Linear. +```python +from comfy.quant_ops import register_layout_op + +@register_layout_op(torch.ops.aten.linear.default, MyLayout) +def my_linear(func, args, kwargs): + # Extract tensors, call optimized kernel + ... +``` +When `torch.nn.functional.linear()` is called with QuantizedTensor arguments, `__torch_dispatch__` automatically routes to the registered implementation. +For any unsupported operation, QuantizedTensor will fallback to call `dequantize` and dispatch using the high-precision implementation. + + +### Mixed Precision + +The `MixedPrecisionOps` class (lines 542-648 in `comfy/ops.py`) enables per-layer quantization decisions, allowing different layers in a model to use different precisions. This is activated when a model config contains a `layer_quant_config` dictionary that specifies which layers should be quantized and how. + +**Architecture:** + +```python +class MixedPrecisionOps(disable_weight_init): + _layer_quant_config = {} # Maps layer names to quantization configs + _compute_dtype = torch.bfloat16 # Default compute / dequantize precision +``` + +**Key mechanism:** + +The custom `Linear._load_from_state_dict()` method inspects each layer during model loading: +- If the layer name is **not** in `_layer_quant_config`: load weight as regular tensor in `_compute_dtype` +- If the layer name **is** in `_layer_quant_config`: + - Load weight as `QuantizedTensor` with the specified layout (e.g., `TensorCoreFP8Layout`) + - Load associated quantization parameters (scales, block_size, etc.) + +**Why it's needed:** + +Not all layers tolerate quantization equally. Sensitive operations like final projections can be kept in higher precision, while compute-heavy matmuls are quantized. This provides most of the performance benefits while maintaining quality. + +The system is selected in `pick_operations()` when `model_config.layer_quant_config` is present, making it the highest-priority operation mode. + + +## Checkpoint Format + +Quantized checkpoints are stored as standard safetensors files with quantized weight tensors and associated scaling parameters, plus a `_quantization_metadata` JSON entry describing the quantization scheme. + +The quantized checkpoint will contain the same layers as the original checkpoint but: +- The weights are stored as quantized values, sometimes using a different storage datatype. E.g. uint8 container for fp8. +- For each quantized weight a number of additional scaling parameters are stored alongside depending on the recipe. +- We store a metadata.json in the metadata of the final safetensor containing the `_quantization_metadata` describing which layers are quantized and what layout has been used. + +### Scaling Parameters details +We define 4 possible scaling parameters that should cover most recipes in the near-future: +- **weight_scale**: quantization scalers for the weights +- **weight_scale_2**: global scalers in the context of double scaling +- **pre_quant_scale**: scalers used for smoothing salient weights +- **input_scale**: quantization scalers for the activations + +| Format | Storage dtype | weight_scale | weight_scale_2 | pre_quant_scale | input_scale | +|--------|---------------|--------------|----------------|-----------------|-------------| +| float8_e4m3fn | float32 | float32 (scalar) | - | - | float32 (scalar) | + +You can find the defined formats in `comfy/quant_ops.py` (QUANT_ALGOS). + +### Quantization Metadata + +The metadata stored alongside the checkpoint contains: +- **format_version**: String to define a version of the standard +- **layers**: A dictionary mapping layer names to their quantization format. The format string maps to the definitions found in `QUANT_ALGOS`. + +Example: +```json +{ + "_quantization_metadata": { + "format_version": "1.0", + "layers": { + "model.layers.0.mlp.up_proj": "float8_e4m3fn", + "model.layers.0.mlp.down_proj": "float8_e4m3fn", + "model.layers.1.mlp.up_proj": "float8_e4m3fn" + } + } +} +``` + + +## Creating Quantized Checkpoints + +To create compatible checkpoints, use any quantization tool provided the output follows the checkpoint format described above and uses a layout defined in `QUANT_ALGOS`. + +### Weight Quantization + +Weight quantization is straightforward - compute the scaling factor directly from the weight tensor using the absolute maximum method described earlier. Each layer's weights are quantized independently and stored with their corresponding `weight_scale` parameter. + +### Calibration (for Activation Quantization) + +Activation quantization (e.g., for FP8 Tensor Core operations) requires `input_scale` parameters that cannot be determined from static weights alone. Since activation values depend on actual inputs, we use **post-training calibration (PTQ)**: + +1. **Collect statistics**: Run inference on N representative samples +2. **Track activations**: Record the absolute maximum (`amax`) of inputs to each quantized layer +3. **Compute scales**: Derive `input_scale` from collected statistics +4. **Store in checkpoint**: Save `input_scale` parameters alongside weights + +The calibration dataset should be representative of your target use case. For diffusion models, this typically means a diverse set of prompts and generation parameters. \ No newline at end of file diff --git a/comfy/ops.py b/comfy/ops.py index 96dffa85d..2a90a5ba2 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -77,7 +77,10 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, of # will add async-offload support to your cast and improve performance. if input is not None: if dtype is None: - dtype = input.dtype + if isinstance(input, QuantizedTensor): + dtype = input._layout_params["orig_dtype"] + else: + dtype = input.dtype if bias_dtype is None: bias_dtype = dtype if device is None: @@ -534,18 +537,7 @@ if CUBLAS_IS_AVAILABLE: # ============================================================================== # Mixed Precision Operations # ============================================================================== -from .quant_ops import QuantizedTensor - -QUANT_FORMAT_MIXINS = { - "float8_e4m3fn": { - "dtype": torch.float8_e4m3fn, - "layout_type": "TensorCoreFP8Layout", - "parameters": { - "weight_scale": torch.nn.Parameter(torch.zeros((), dtype=torch.float32), requires_grad=False), - "input_scale": torch.nn.Parameter(torch.zeros((), dtype=torch.float32), requires_grad=False), - } - } -} +from .quant_ops import QuantizedTensor, QUANT_ALGOS class MixedPrecisionOps(disable_weight_init): _layer_quant_config = {} @@ -596,23 +588,24 @@ class MixedPrecisionOps(disable_weight_init): if quant_format is None: raise ValueError(f"Unknown quantization format for layer {layer_name}") - mixin = QUANT_FORMAT_MIXINS[quant_format] - self.layout_type = mixin["layout_type"] + qconfig = QUANT_ALGOS[quant_format] + self.layout_type = qconfig["comfy_tensor_layout"] - scale_key = f"{prefix}weight_scale" + weight_scale_key = f"{prefix}weight_scale" layout_params = { - 'scale': state_dict.pop(scale_key, None), - 'orig_dtype': MixedPrecisionOps._compute_dtype + 'scale': state_dict.pop(weight_scale_key, None), + 'orig_dtype': MixedPrecisionOps._compute_dtype, + 'block_size': qconfig.get("group_size", None), } if layout_params['scale'] is not None: - manually_loaded_keys.append(scale_key) + manually_loaded_keys.append(weight_scale_key) self.weight = torch.nn.Parameter( - QuantizedTensor(weight.to(device=device, dtype=mixin["dtype"]), self.layout_type, layout_params), + QuantizedTensor(weight.to(device=device), self.layout_type, layout_params), requires_grad=False ) - for param_name, param_value in mixin["parameters"].items(): + for param_name in qconfig["parameters"]: param_key = f"{prefix}{param_name}" _v = state_dict.pop(param_key, None) if _v is None: @@ -643,7 +636,7 @@ class MixedPrecisionOps(disable_weight_init): if (getattr(self, 'layout_type', None) is not None and getattr(self, 'input_scale', None) is not None and not isinstance(input, QuantizedTensor)): - input = QuantizedTensor.from_float(input, self.layout_type, scale=self.input_scale, fp8_dtype=self.weight.dtype) + input = QuantizedTensor.from_float(input, self.layout_type, scale=self.input_scale, dtype=self.weight.dtype) return self._forward(input, self.weight, self.bias) diff --git a/comfy/quant_ops.py b/comfy/quant_ops.py index c56e32a73..1d058bece 100644 --- a/comfy/quant_ops.py +++ b/comfy/quant_ops.py @@ -74,6 +74,12 @@ def _copy_layout_params(params): new_params[k] = v return new_params +def _copy_layout_params_inplace(src, dst, non_blocking=False): + for k, v in src.items(): + if isinstance(v, torch.Tensor): + dst[k].copy_(v, non_blocking=non_blocking) + else: + dst[k] = v class QuantizedLayout: """ @@ -318,13 +324,13 @@ def generic_to_dtype_layout(func, args, kwargs): def generic_copy_(func, args, kwargs): qt_dest = args[0] src = args[1] - + non_blocking = args[2] if len(args) > 2 else False if isinstance(qt_dest, QuantizedTensor): if isinstance(src, QuantizedTensor): # Copy from another quantized tensor - qt_dest._qdata.copy_(src._qdata) + qt_dest._qdata.copy_(src._qdata, non_blocking=non_blocking) qt_dest._layout_type = src._layout_type - qt_dest._layout_params = _copy_layout_params(src._layout_params) + _copy_layout_params_inplace(src._layout_params, qt_dest._layout_params, non_blocking=non_blocking) else: # Copy from regular tensor - just copy raw data qt_dest._qdata.copy_(src) @@ -336,6 +342,26 @@ def generic_copy_(func, args, kwargs): def generic_has_compatible_shallow_copy_type(func, args, kwargs): return True + +@register_generic_util(torch.ops.aten.empty_like.default) +def generic_empty_like(func, args, kwargs): + """Empty_like operation - creates an empty tensor with the same quantized structure.""" + qt = args[0] + if isinstance(qt, QuantizedTensor): + # Create empty tensor with same shape and dtype as the quantized data + hp_dtype = kwargs.pop('dtype', qt._layout_params["orig_dtype"]) + new_qdata = torch.empty_like(qt._qdata, **kwargs) + + # Handle device transfer for layout params + target_device = kwargs.get('device', new_qdata.device) + new_params = _move_layout_params_to_device(qt._layout_params, target_device) + + # Update orig_dtype if dtype is specified + new_params['orig_dtype'] = hp_dtype + + return QuantizedTensor(new_qdata, qt._layout_type, new_params) + return func(*args, **kwargs) + # ============================================================================== # FP8 Layout + Operation Handlers # ============================================================================== @@ -378,6 +404,13 @@ class TensorCoreFP8Layout(QuantizedLayout): def get_plain_tensors(cls, qtensor): return qtensor._qdata, qtensor._layout_params['scale'] +QUANT_ALGOS = { + "float8_e4m3fn": { + "storage_t": torch.float8_e4m3fn, + "parameters": {"weight_scale", "input_scale"}, + "comfy_tensor_layout": "TensorCoreFP8Layout", + }, +} LAYOUTS = { "TensorCoreFP8Layout": TensorCoreFP8Layout, From f91078b1ffa484c424f78814f54de4d5846e4daa Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Thu, 13 Nov 2025 20:05:26 +0200 Subject: [PATCH 105/647] add PR template for API-Nodes (#10736) --- .github/PULL_REQUEST_TEMPLATE/api-node.md | 21 ++++++++ .github/workflows/api-node-template.yml | 58 +++++++++++++++++++++++ 2 files changed, 79 insertions(+) create mode 100644 .github/PULL_REQUEST_TEMPLATE/api-node.md create mode 100644 .github/workflows/api-node-template.yml diff --git a/.github/PULL_REQUEST_TEMPLATE/api-node.md b/.github/PULL_REQUEST_TEMPLATE/api-node.md new file mode 100644 index 000000000..f62744878 --- /dev/null +++ b/.github/PULL_REQUEST_TEMPLATE/api-node.md @@ -0,0 +1,21 @@ + + +## API Node PR Checklist + +### Scope +- [ ] **Is API Node Change** + +### Pricing & Billing +- [ ] **Need pricing update** +- [ ] **No pricing update** + +If **Need pricing update**: +- [ ] Metronome rate cards updated +- [ ] Auto‑billing tests updated and passing + +### QA +- [ ] **QA done** +- [ ] **QA not required** + +### Comms +- [ ] Informed **@Kosinkadink** diff --git a/.github/workflows/api-node-template.yml b/.github/workflows/api-node-template.yml new file mode 100644 index 000000000..0775f9979 --- /dev/null +++ b/.github/workflows/api-node-template.yml @@ -0,0 +1,58 @@ +name: Append API Node PR template + +on: + pull_request_target: + types: [opened, reopened, synchronize, edited, ready_for_review] + paths: + - 'comfy_api_nodes/**' # only run if these files changed + +permissions: + contents: read + pull-requests: write + +jobs: + inject: + runs-on: ubuntu-latest + steps: + - name: Ensure template exists and append to PR body + uses: actions/github-script@v7 + with: + script: | + const { owner, repo } = context.repo; + const number = context.payload.pull_request.number; + const templatePath = '.github/PULL_REQUEST_TEMPLATE/api-node.md'; + const marker = ''; + + const { data: pr } = await github.rest.pulls.get({ owner, repo, pull_number: number }); + + let templateText; + try { + const res = await github.rest.repos.getContent({ + owner, + repo, + path: templatePath, + ref: pr.base.ref + }); + const buf = Buffer.from(res.data.content, res.data.encoding || 'base64'); + templateText = buf.toString('utf8'); + } catch (e) { + core.setFailed(`Required PR template not found at "${templatePath}" on ${pr.base.ref}. Please add it to the repo.`); + return; + } + + // Enforce the presence of the marker inside the template (for idempotence) + if (!templateText.includes(marker)) { + core.setFailed(`Template at "${templatePath}" does not contain the required marker:\n${marker}\nAdd it so we can detect duplicates safely.`); + return; + } + + // If the PR already contains the marker, do not append again. + const body = pr.body || ''; + if (body.includes(marker)) { + core.info('Template already present in PR body; nothing to inject.'); + return; + } + + const newBody = (body ? body + '\n\n' : '') + templateText + '\n'; + await github.rest.pulls.update({ owner, repo, pull_number: number, body: newBody }); + core.notice('API Node template appended to PR description.'); From 2fde9597f4b02c5f06c1a5ceb3ca2fa6d74966ec Mon Sep 17 00:00:00 2001 From: ric-yu Date: Thu, 13 Nov 2025 15:11:52 -0800 Subject: [PATCH 106/647] feat: add create_time dict to prompt field in /history and /queue (#10741) --- server.py | 2 ++ 1 file changed, 2 insertions(+) diff --git a/server.py b/server.py index 5d773b10a..d059d3dc9 100644 --- a/server.py +++ b/server.py @@ -2,6 +2,7 @@ import os import sys import asyncio import traceback +import time import nodes import folder_paths @@ -733,6 +734,7 @@ class PromptServer(): for sensitive_val in execution.SENSITIVE_EXTRA_DATA_KEYS: if sensitive_val in extra_data: sensitive[sensitive_val] = extra_data.pop(sensitive_val) + extra_data["create_time"] = int(time.time() * 1000) # timestamp in milliseconds self.prompt_queue.put((number, prompt_id, prompt, extra_data, outputs_to_execute, sensitive)) response = {"prompt_id": prompt_id, "number": number, "node_errors": valid[3]} return web.json_response(response) From 94c298f9625b0fd9af8ea07a73075fdefe0d9e57 Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Fri, 14 Nov 2025 10:02:03 +1000 Subject: [PATCH 107/647] flux: reduce VRAM usage (#10737) Cleanup a bunch of stack tensors on Flux. This take me from B=19 to B=22 for 1600x1600 on RTX5090. --- comfy/ldm/flux/layers.py | 33 ++++++++++++++++++++++++++------- 1 file changed, 26 insertions(+), 7 deletions(-) diff --git a/comfy/ldm/flux/layers.py b/comfy/ldm/flux/layers.py index a3eab0470..f4bf56e01 100644 --- a/comfy/ldm/flux/layers.py +++ b/comfy/ldm/flux/layers.py @@ -167,39 +167,55 @@ class DoubleStreamBlock(nn.Module): img_modulated = self.img_norm1(img) img_modulated = apply_mod(img_modulated, (1 + img_mod1.scale), img_mod1.shift, modulation_dims_img) img_qkv = self.img_attn.qkv(img_modulated) + del img_modulated img_q, img_k, img_v = img_qkv.view(img_qkv.shape[0], img_qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4) + del img_qkv img_q, img_k = self.img_attn.norm(img_q, img_k, img_v) # prepare txt for attention txt_modulated = self.txt_norm1(txt) txt_modulated = apply_mod(txt_modulated, (1 + txt_mod1.scale), txt_mod1.shift, modulation_dims_txt) txt_qkv = self.txt_attn.qkv(txt_modulated) + del txt_modulated txt_q, txt_k, txt_v = txt_qkv.view(txt_qkv.shape[0], txt_qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4) + del txt_qkv txt_q, txt_k = self.txt_attn.norm(txt_q, txt_k, txt_v) if self.flipped_img_txt: + q = torch.cat((img_q, txt_q), dim=2) + del img_q, txt_q + k = torch.cat((img_k, txt_k), dim=2) + del img_k, txt_k + v = torch.cat((img_v, txt_v), dim=2) + del img_v, txt_v # run actual attention - attn = attention(torch.cat((img_q, txt_q), dim=2), - torch.cat((img_k, txt_k), dim=2), - torch.cat((img_v, txt_v), dim=2), + attn = attention(q, k, v, pe=pe, mask=attn_mask, transformer_options=transformer_options) + del q, k, v img_attn, txt_attn = attn[:, : img.shape[1]], attn[:, img.shape[1]:] else: + q = torch.cat((txt_q, img_q), dim=2) + del txt_q, img_q + k = torch.cat((txt_k, img_k), dim=2) + del txt_k, img_k + v = torch.cat((txt_v, img_v), dim=2) + del txt_v, img_v # run actual attention - attn = attention(torch.cat((txt_q, img_q), dim=2), - torch.cat((txt_k, img_k), dim=2), - torch.cat((txt_v, img_v), dim=2), + attn = attention(q, k, v, pe=pe, mask=attn_mask, transformer_options=transformer_options) + del q, k, v txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1]:] # calculate the img bloks img += apply_mod(self.img_attn.proj(img_attn), img_mod1.gate, None, modulation_dims_img) + del img_attn img += apply_mod(self.img_mlp(apply_mod(self.img_norm2(img), (1 + img_mod2.scale), img_mod2.shift, modulation_dims_img)), img_mod2.gate, None, modulation_dims_img) # calculate the txt bloks txt += apply_mod(self.txt_attn.proj(txt_attn), txt_mod1.gate, None, modulation_dims_txt) + del txt_attn txt += apply_mod(self.txt_mlp(apply_mod(self.txt_norm2(txt), (1 + txt_mod2.scale), txt_mod2.shift, modulation_dims_txt)), txt_mod2.gate, None, modulation_dims_txt) if txt.dtype == torch.float16: @@ -249,12 +265,15 @@ class SingleStreamBlock(nn.Module): qkv, mlp = torch.split(self.linear1(apply_mod(self.pre_norm(x), (1 + mod.scale), mod.shift, modulation_dims)), [3 * self.hidden_size, self.mlp_hidden_dim], dim=-1) q, k, v = qkv.view(qkv.shape[0], qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4) + del qkv q, k = self.norm(q, k, v) # compute attention attn = attention(q, k, v, pe=pe, mask=attn_mask, transformer_options=transformer_options) + del q, k, v # compute activation in mlp stream, cat again and run second linear layer - output = self.linear2(torch.cat((attn, self.mlp_act(mlp)), 2)) + mlp = self.mlp_act(mlp) + output = self.linear2(torch.cat((attn, mlp), 2)) x += apply_mod(output, mod.gate, None, modulation_dims) if x.dtype == torch.float16: x = torch.nan_to_num(x, nan=0.0, posinf=65504, neginf=-65504) From 1ef328c007a419c2c429df0f80532cc11579dc97 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Thu, 13 Nov 2025 18:32:39 -0800 Subject: [PATCH 108/647] Better instructions for the portable. (#10743) --- README.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/README.md b/README.md index f51807ad5..cd8273b0d 100644 --- a/README.md +++ b/README.md @@ -173,7 +173,7 @@ There is a portable standalone build for Windows that should work for running on ### [Direct link to download](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_nvidia.7z) -Simply download, extract with [7-Zip](https://7-zip.org) and run. Make sure you put your Stable Diffusion checkpoints/models (the huge ckpt/safetensors files) in: ComfyUI\models\checkpoints +Simply download, extract with [7-Zip](https://7-zip.org) or with the windows explorer on recent windows versions and run. For smaller models you normally only need to put the checkpoints (the huge ckpt/safetensors files) in: ComfyUI\models\checkpoints but many of the larger models have multiple files. Make sure to follow the instructions to know which subfolder to put them in ComfyUI\models\ If you have trouble extracting it, right click the file -> properties -> unblock From f60923590c3f2fd05e166e2ec57968aaf7007dd0 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Thu, 13 Nov 2025 22:28:05 -0800 Subject: [PATCH 109/647] Use same code for chroma and flux blocks so that optimizations are shared. (#10746) --- comfy/ldm/chroma/layers.py | 121 ----------------------------- comfy/ldm/chroma/model.py | 7 +- comfy/ldm/chroma_radiance/model.py | 7 +- comfy/ldm/flux/layers.py | 31 ++++++-- 4 files changed, 31 insertions(+), 135 deletions(-) diff --git a/comfy/ldm/chroma/layers.py b/comfy/ldm/chroma/layers.py index fc7110cce..9f4ad5bd2 100644 --- a/comfy/ldm/chroma/layers.py +++ b/comfy/ldm/chroma/layers.py @@ -1,12 +1,9 @@ import torch from torch import Tensor, nn -from comfy.ldm.flux.math import attention from comfy.ldm.flux.layers import ( MLPEmbedder, RMSNorm, - QKNorm, - SelfAttention, ModulationOut, ) @@ -48,124 +45,6 @@ class Approximator(nn.Module): return x -class DoubleStreamBlock(nn.Module): - def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False, flipped_img_txt=False, dtype=None, device=None, operations=None): - super().__init__() - - mlp_hidden_dim = int(hidden_size * mlp_ratio) - self.num_heads = num_heads - self.hidden_size = hidden_size - self.img_norm1 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) - self.img_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias, dtype=dtype, device=device, operations=operations) - - self.img_norm2 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) - self.img_mlp = nn.Sequential( - operations.Linear(hidden_size, mlp_hidden_dim, bias=True, dtype=dtype, device=device), - nn.GELU(approximate="tanh"), - operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device), - ) - - self.txt_norm1 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) - self.txt_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias, dtype=dtype, device=device, operations=operations) - - self.txt_norm2 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) - self.txt_mlp = nn.Sequential( - operations.Linear(hidden_size, mlp_hidden_dim, bias=True, dtype=dtype, device=device), - nn.GELU(approximate="tanh"), - operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device), - ) - self.flipped_img_txt = flipped_img_txt - - def forward(self, img: Tensor, txt: Tensor, pe: Tensor, vec: Tensor, attn_mask=None, transformer_options={}): - (img_mod1, img_mod2), (txt_mod1, txt_mod2) = vec - - # prepare image for attention - img_modulated = torch.addcmul(img_mod1.shift, 1 + img_mod1.scale, self.img_norm1(img)) - img_qkv = self.img_attn.qkv(img_modulated) - img_q, img_k, img_v = img_qkv.view(img_qkv.shape[0], img_qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4) - img_q, img_k = self.img_attn.norm(img_q, img_k, img_v) - - # prepare txt for attention - txt_modulated = torch.addcmul(txt_mod1.shift, 1 + txt_mod1.scale, self.txt_norm1(txt)) - txt_qkv = self.txt_attn.qkv(txt_modulated) - txt_q, txt_k, txt_v = txt_qkv.view(txt_qkv.shape[0], txt_qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4) - txt_q, txt_k = self.txt_attn.norm(txt_q, txt_k, txt_v) - - # run actual attention - attn = attention(torch.cat((txt_q, img_q), dim=2), - torch.cat((txt_k, img_k), dim=2), - torch.cat((txt_v, img_v), dim=2), - pe=pe, mask=attn_mask, transformer_options=transformer_options) - - txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1] :] - - # calculate the img bloks - img.addcmul_(img_mod1.gate, self.img_attn.proj(img_attn)) - img.addcmul_(img_mod2.gate, self.img_mlp(torch.addcmul(img_mod2.shift, 1 + img_mod2.scale, self.img_norm2(img)))) - - # calculate the txt bloks - txt.addcmul_(txt_mod1.gate, self.txt_attn.proj(txt_attn)) - txt.addcmul_(txt_mod2.gate, self.txt_mlp(torch.addcmul(txt_mod2.shift, 1 + txt_mod2.scale, self.txt_norm2(txt)))) - - if txt.dtype == torch.float16: - txt = torch.nan_to_num(txt, nan=0.0, posinf=65504, neginf=-65504) - - return img, txt - - -class SingleStreamBlock(nn.Module): - """ - A DiT block with parallel linear layers as described in - https://arxiv.org/abs/2302.05442 and adapted modulation interface. - """ - - def __init__( - self, - hidden_size: int, - num_heads: int, - mlp_ratio: float = 4.0, - qk_scale: float = None, - dtype=None, - device=None, - operations=None - ): - super().__init__() - self.hidden_dim = hidden_size - self.num_heads = num_heads - head_dim = hidden_size // num_heads - self.scale = qk_scale or head_dim**-0.5 - - self.mlp_hidden_dim = int(hidden_size * mlp_ratio) - # qkv and mlp_in - self.linear1 = operations.Linear(hidden_size, hidden_size * 3 + self.mlp_hidden_dim, dtype=dtype, device=device) - # proj and mlp_out - self.linear2 = operations.Linear(hidden_size + self.mlp_hidden_dim, hidden_size, dtype=dtype, device=device) - - self.norm = QKNorm(head_dim, dtype=dtype, device=device, operations=operations) - - self.hidden_size = hidden_size - self.pre_norm = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) - - self.mlp_act = nn.GELU(approximate="tanh") - - def forward(self, x: Tensor, pe: Tensor, vec: Tensor, attn_mask=None, transformer_options={}) -> Tensor: - mod = vec - x_mod = torch.addcmul(mod.shift, 1 + mod.scale, self.pre_norm(x)) - qkv, mlp = torch.split(self.linear1(x_mod), [3 * self.hidden_size, self.mlp_hidden_dim], dim=-1) - - q, k, v = qkv.view(qkv.shape[0], qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4) - q, k = self.norm(q, k, v) - - # compute attention - attn = attention(q, k, v, pe=pe, mask=attn_mask, transformer_options=transformer_options) - # compute activation in mlp stream, cat again and run second linear layer - output = self.linear2(torch.cat((attn, self.mlp_act(mlp)), 2)) - x.addcmul_(mod.gate, output) - if x.dtype == torch.float16: - x = torch.nan_to_num(x, nan=0.0, posinf=65504, neginf=-65504) - return x - - class LastLayer(nn.Module): def __init__(self, hidden_size: int, patch_size: int, out_channels: int, dtype=None, device=None, operations=None): super().__init__() diff --git a/comfy/ldm/chroma/model.py b/comfy/ldm/chroma/model.py index ad1c523fe..67bf70eb1 100644 --- a/comfy/ldm/chroma/model.py +++ b/comfy/ldm/chroma/model.py @@ -11,12 +11,12 @@ import comfy.ldm.common_dit from comfy.ldm.flux.layers import ( EmbedND, timestep_embedding, + DoubleStreamBlock, + SingleStreamBlock, ) from .layers import ( - DoubleStreamBlock, LastLayer, - SingleStreamBlock, Approximator, ChromaModulationOut, ) @@ -90,6 +90,7 @@ class Chroma(nn.Module): self.num_heads, mlp_ratio=params.mlp_ratio, qkv_bias=params.qkv_bias, + modulation=False, dtype=dtype, device=device, operations=operations ) for _ in range(params.depth) @@ -98,7 +99,7 @@ class Chroma(nn.Module): self.single_blocks = nn.ModuleList( [ - SingleStreamBlock(self.hidden_size, self.num_heads, mlp_ratio=params.mlp_ratio, dtype=dtype, device=device, operations=operations) + SingleStreamBlock(self.hidden_size, self.num_heads, mlp_ratio=params.mlp_ratio, modulation=False, dtype=dtype, device=device, operations=operations) for _ in range(params.depth_single_blocks) ] ) diff --git a/comfy/ldm/chroma_radiance/model.py b/comfy/ldm/chroma_radiance/model.py index 7d7be80f5..e643b4414 100644 --- a/comfy/ldm/chroma_radiance/model.py +++ b/comfy/ldm/chroma_radiance/model.py @@ -10,12 +10,10 @@ from torch import Tensor, nn from einops import repeat import comfy.ldm.common_dit -from comfy.ldm.flux.layers import EmbedND +from comfy.ldm.flux.layers import EmbedND, DoubleStreamBlock, SingleStreamBlock from comfy.ldm.chroma.model import Chroma, ChromaParams from comfy.ldm.chroma.layers import ( - DoubleStreamBlock, - SingleStreamBlock, Approximator, ) from .layers import ( @@ -89,7 +87,6 @@ class ChromaRadiance(Chroma): dtype=dtype, device=device, operations=operations ) - self.double_blocks = nn.ModuleList( [ DoubleStreamBlock( @@ -97,6 +94,7 @@ class ChromaRadiance(Chroma): self.num_heads, mlp_ratio=params.mlp_ratio, qkv_bias=params.qkv_bias, + modulation=False, dtype=dtype, device=device, operations=operations ) for _ in range(params.depth) @@ -109,6 +107,7 @@ class ChromaRadiance(Chroma): self.hidden_size, self.num_heads, mlp_ratio=params.mlp_ratio, + modulation=False, dtype=dtype, device=device, operations=operations, ) for _ in range(params.depth_single_blocks) diff --git a/comfy/ldm/flux/layers.py b/comfy/ldm/flux/layers.py index f4bf56e01..23150a712 100644 --- a/comfy/ldm/flux/layers.py +++ b/comfy/ldm/flux/layers.py @@ -130,13 +130,17 @@ def apply_mod(tensor, m_mult, m_add=None, modulation_dims=None): class DoubleStreamBlock(nn.Module): - def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False, flipped_img_txt=False, dtype=None, device=None, operations=None): + def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False, flipped_img_txt=False, modulation=True, dtype=None, device=None, operations=None): super().__init__() mlp_hidden_dim = int(hidden_size * mlp_ratio) self.num_heads = num_heads self.hidden_size = hidden_size - self.img_mod = Modulation(hidden_size, double=True, dtype=dtype, device=device, operations=operations) + self.modulation = modulation + + if self.modulation: + self.img_mod = Modulation(hidden_size, double=True, dtype=dtype, device=device, operations=operations) + self.img_norm1 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) self.img_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias, dtype=dtype, device=device, operations=operations) @@ -147,7 +151,9 @@ class DoubleStreamBlock(nn.Module): operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device), ) - self.txt_mod = Modulation(hidden_size, double=True, dtype=dtype, device=device, operations=operations) + if self.modulation: + self.txt_mod = Modulation(hidden_size, double=True, dtype=dtype, device=device, operations=operations) + self.txt_norm1 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) self.txt_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias, dtype=dtype, device=device, operations=operations) @@ -160,8 +166,11 @@ class DoubleStreamBlock(nn.Module): self.flipped_img_txt = flipped_img_txt def forward(self, img: Tensor, txt: Tensor, vec: Tensor, pe: Tensor, attn_mask=None, modulation_dims_img=None, modulation_dims_txt=None, transformer_options={}): - img_mod1, img_mod2 = self.img_mod(vec) - txt_mod1, txt_mod2 = self.txt_mod(vec) + if self.modulation: + img_mod1, img_mod2 = self.img_mod(vec) + txt_mod1, txt_mod2 = self.txt_mod(vec) + else: + (img_mod1, img_mod2), (txt_mod1, txt_mod2) = vec # prepare image for attention img_modulated = self.img_norm1(img) @@ -236,6 +245,7 @@ class SingleStreamBlock(nn.Module): num_heads: int, mlp_ratio: float = 4.0, qk_scale: float = None, + modulation=True, dtype=None, device=None, operations=None @@ -258,10 +268,17 @@ class SingleStreamBlock(nn.Module): self.pre_norm = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) self.mlp_act = nn.GELU(approximate="tanh") - self.modulation = Modulation(hidden_size, double=False, dtype=dtype, device=device, operations=operations) + if modulation: + self.modulation = Modulation(hidden_size, double=False, dtype=dtype, device=device, operations=operations) + else: + self.modulation = None def forward(self, x: Tensor, vec: Tensor, pe: Tensor, attn_mask=None, modulation_dims=None, transformer_options={}) -> Tensor: - mod, _ = self.modulation(vec) + if self.modulation: + mod, _ = self.modulation(vec) + else: + mod = vec + qkv, mlp = torch.split(self.linear1(apply_mod(self.pre_norm(x), (1 + mod.scale), mod.shift, modulation_dims)), [3 * self.hidden_size, self.mlp_hidden_dim], dim=-1) q, k, v = qkv.view(qkv.shape[0], qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4) From 443056c401c53953bb8eee6da71b9ad29afe2581 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Fri, 14 Nov 2025 00:26:05 -0800 Subject: [PATCH 110/647] Fix custom nodes import error. (#10747) This should fix the import errors but will break if the custom nodes actually try to use the class. --- comfy/ldm/chroma/layers.py | 3 +++ 1 file changed, 3 insertions(+) diff --git a/comfy/ldm/chroma/layers.py b/comfy/ldm/chroma/layers.py index 9f4ad5bd2..2d5684348 100644 --- a/comfy/ldm/chroma/layers.py +++ b/comfy/ldm/chroma/layers.py @@ -7,6 +7,9 @@ from comfy.ldm.flux.layers import ( ModulationOut, ) +# TODO: remove this in a few months +SingleStreamBlock = None +DoubleStreamBlock = None class ChromaModulationOut(ModulationOut): From bd01d9f7fd241a45bd08b60dfedbe78577383cc4 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sat, 15 Nov 2025 03:54:40 -0800 Subject: [PATCH 111/647] Add left padding support to tokenizers. (#10753) --- comfy/sd1_clip.py | 17 ++++++++++++----- 1 file changed, 12 insertions(+), 5 deletions(-) diff --git a/comfy/sd1_clip.py b/comfy/sd1_clip.py index f8a7c2a1b..3066de2d7 100644 --- a/comfy/sd1_clip.py +++ b/comfy/sd1_clip.py @@ -460,7 +460,7 @@ def load_embed(embedding_name, embedding_directory, embedding_size, embed_key=No return embed_out class SDTokenizer: - def __init__(self, tokenizer_path=None, max_length=77, pad_with_end=True, embedding_directory=None, embedding_size=768, embedding_key='clip_l', tokenizer_class=CLIPTokenizer, has_start_token=True, has_end_token=True, pad_to_max_length=True, min_length=None, pad_token=None, end_token=None, min_padding=None, tokenizer_data={}, tokenizer_args={}): + def __init__(self, tokenizer_path=None, max_length=77, pad_with_end=True, embedding_directory=None, embedding_size=768, embedding_key='clip_l', tokenizer_class=CLIPTokenizer, has_start_token=True, has_end_token=True, pad_to_max_length=True, min_length=None, pad_token=None, end_token=None, min_padding=None, pad_left=False, tokenizer_data={}, tokenizer_args={}): if tokenizer_path is None: tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "sd1_tokenizer") self.tokenizer = tokenizer_class.from_pretrained(tokenizer_path, **tokenizer_args) @@ -468,6 +468,7 @@ class SDTokenizer: self.min_length = tokenizer_data.get("{}_min_length".format(embedding_key), min_length) self.end_token = None self.min_padding = min_padding + self.pad_left = pad_left empty = self.tokenizer('')["input_ids"] self.tokenizer_adds_end_token = has_end_token @@ -522,6 +523,12 @@ class SDTokenizer: return (embed, "{} {}".format(embedding_name[len(stripped):], leftover)) return (embed, leftover) + def pad_tokens(self, tokens, amount): + if self.pad_left: + for i in range(amount): + tokens.insert(0, (self.pad_token, 1.0, 0)) + else: + tokens.extend([(self.pad_token, 1.0, 0)] * amount) def tokenize_with_weights(self, text:str, return_word_ids=False, tokenizer_options={}, **kwargs): ''' @@ -600,7 +607,7 @@ class SDTokenizer: if self.end_token is not None: batch.append((self.end_token, 1.0, 0)) if self.pad_to_max_length: - batch.extend([(self.pad_token, 1.0, 0)] * (remaining_length)) + self.pad_tokens(batch, remaining_length) #start new batch batch = [] if self.start_token is not None: @@ -614,11 +621,11 @@ class SDTokenizer: if self.end_token is not None: batch.append((self.end_token, 1.0, 0)) if min_padding is not None: - batch.extend([(self.pad_token, 1.0, 0)] * min_padding) + self.pad_tokens(batch, min_padding) if self.pad_to_max_length and len(batch) < self.max_length: - batch.extend([(self.pad_token, 1.0, 0)] * (self.max_length - len(batch))) + self.pad_tokens(batch, self.max_length - len(batch)) if min_length is not None and len(batch) < min_length: - batch.extend([(self.pad_token, 1.0, 0)] * (min_length - len(batch))) + self.pad_tokens(batch, min_length - len(batch)) if not return_word_ids: batched_tokens = [[(t, w) for t, w,_ in x] for x in batched_tokens] From 9a0238256873711bd38ce0e0b1d15a617a1ee454 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Sat, 15 Nov 2025 21:18:49 +0200 Subject: [PATCH 112/647] chore(api-nodes): mark OpenAIDalle2 and OpenAIDalle3 nodes as deprecated (#10757) --- comfy_api_nodes/nodes_openai.py | 2 ++ 1 file changed, 2 insertions(+) diff --git a/comfy_api_nodes/nodes_openai.py b/comfy_api_nodes/nodes_openai.py index acf35d276..e08bec08c 100644 --- a/comfy_api_nodes/nodes_openai.py +++ b/comfy_api_nodes/nodes_openai.py @@ -169,6 +169,7 @@ class OpenAIDalle2(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + is_deprecated=True, ) @classmethod @@ -299,6 +300,7 @@ class OpenAIDalle3(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + is_deprecated=True, ) @classmethod From 2d4a08b717c492fa45e98bd70beb48d4e77cb464 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Sat, 15 Nov 2025 22:37:34 +0200 Subject: [PATCH 113/647] Revert "chore(api-nodes): mark OpenAIDalle2 and OpenAIDalle3 nodes as deprecated (#10757)" (#10759) This reverts commit 9a0238256873711bd38ce0e0b1d15a617a1ee454. --- comfy_api_nodes/nodes_openai.py | 2 -- 1 file changed, 2 deletions(-) diff --git a/comfy_api_nodes/nodes_openai.py b/comfy_api_nodes/nodes_openai.py index e08bec08c..acf35d276 100644 --- a/comfy_api_nodes/nodes_openai.py +++ b/comfy_api_nodes/nodes_openai.py @@ -169,7 +169,6 @@ class OpenAIDalle2(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, - is_deprecated=True, ) @classmethod @@ -300,7 +299,6 @@ class OpenAIDalle3(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, - is_deprecated=True, ) @classmethod From 7d6103325e1c97aa54f963253e3e7f1d6da6947f Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sun, 16 Nov 2025 00:01:14 -0800 Subject: [PATCH 114/647] Change ROCm nightly install command to 7.1 (#10764) --- README.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/README.md b/README.md index cd8273b0d..c0384099d 100644 --- a/README.md +++ b/README.md @@ -221,7 +221,7 @@ AMD users can install rocm and pytorch with pip if you don't have it already ins This is the command to install the nightly with ROCm 7.0 which might have some performance improvements: -```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm7.0``` +```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm7.1``` ### AMD GPUs (Experimental: Windows and Linux), RDNA 3, 3.5 and 4 only. From 3d0003c24c1aec9f0c021dbc70ffb7cd8cf0685c Mon Sep 17 00:00:00 2001 From: comfyanonymous Date: Mon, 17 Nov 2025 17:17:24 -0500 Subject: [PATCH 115/647] ComfyUI version 0.3.69 --- comfyui_version.py | 2 +- pyproject.toml | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/comfyui_version.py b/comfyui_version.py index 25d1a4157..1e554eb9f 100644 --- a/comfyui_version.py +++ b/comfyui_version.py @@ -1,3 +1,3 @@ # This file is automatically generated by the build process when version is # updated in pyproject.toml. -__version__ = "0.3.68" +__version__ = "0.3.69" diff --git a/pyproject.toml b/pyproject.toml index 79ff3f74a..63778286f 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -1,6 +1,6 @@ [project] name = "ComfyUI" -version = "0.3.68" +version = "0.3.69" readme = "README.md" license = { file = "LICENSE" } requires-python = ">=3.9" From 27cbac865ec226cfd9c1563327b0d62cf5dbd484 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 17 Nov 2025 16:04:04 -0800 Subject: [PATCH 116/647] Add release workflow for NVIDIA cu126 (#10777) --- .github/workflows/release-stable-all.yml | 17 +++++++++++++++++ 1 file changed, 17 insertions(+) diff --git a/.github/workflows/release-stable-all.yml b/.github/workflows/release-stable-all.yml index 7dca7277b..f7de3a7c3 100644 --- a/.github/workflows/release-stable-all.yml +++ b/.github/workflows/release-stable-all.yml @@ -43,6 +43,23 @@ jobs: test_release: true secrets: inherit + release_nvidia_cu126: + permissions: + contents: "write" + packages: "write" + pull-requests: "read" + name: "Release NVIDIA cu126" + uses: ./.github/workflows/stable-release.yml + with: + git_tag: ${{ inputs.git_tag }} + cache_tag: "cu126" + python_minor: "12" + python_patch: "10" + rel_name: "nvidia" + rel_extra_name: "_cu126" + test_release: true + secrets: inherit + release_amd_rocm: permissions: contents: "write" From f41e5f398d5d4059a3c87cf157bd932afcce3c0d Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 17 Nov 2025 16:59:19 -0800 Subject: [PATCH 117/647] Update README with new portable download link (#10778) --- README.md | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/README.md b/README.md index c0384099d..323dfc587 100644 --- a/README.md +++ b/README.md @@ -183,7 +183,9 @@ Update your Nvidia drivers if it doesn't start. [Experimental portable for AMD GPUs](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_amd.7z) -[Portable with pytorch cuda 12.8 and python 3.12](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_nvidia_cu128.7z) (Supports Nvidia 10 series and older GPUs). +[Portable with pytorch cuda 12.8 and python 3.12](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_nvidia_cu128.7z). + +[Portable with pytorch cuda 12.6 and python 3.12](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_nvidia_cu128.7z) (Supports Nvidia 10 series and older GPUs). #### How do I share models between another UI and ComfyUI? From fdf49a28617f742d746ad209e57ed7420b3535dc Mon Sep 17 00:00:00 2001 From: ComfyUI Wiki Date: Tue, 18 Nov 2025 11:04:06 +0800 Subject: [PATCH 118/647] Fix the portable download link for CUDA 12.6 (#10780) --- README.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/README.md b/README.md index 323dfc587..28beec427 100644 --- a/README.md +++ b/README.md @@ -185,7 +185,7 @@ Update your Nvidia drivers if it doesn't start. [Portable with pytorch cuda 12.8 and python 3.12](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_nvidia_cu128.7z). -[Portable with pytorch cuda 12.6 and python 3.12](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_nvidia_cu128.7z) (Supports Nvidia 10 series and older GPUs). +[Portable with pytorch cuda 12.6 and python 3.12](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_nvidia_cu126.7z) (Supports Nvidia 10 series and older GPUs). #### How do I share models between another UI and ComfyUI? From 47bfd5a33fa984a1102fc2bd7b25c91a69ace288 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 17 Nov 2025 21:26:44 -0800 Subject: [PATCH 119/647] Native block swap custom nodes considered harmful. (#10783) --- comfy_extras/nodes_nop.py | 39 +++++++++++++++++++++++++++++++++++++++ nodes.py | 1 + 2 files changed, 40 insertions(+) create mode 100644 comfy_extras/nodes_nop.py diff --git a/comfy_extras/nodes_nop.py b/comfy_extras/nodes_nop.py new file mode 100644 index 000000000..953061bcb --- /dev/null +++ b/comfy_extras/nodes_nop.py @@ -0,0 +1,39 @@ +from comfy_api.latest import ComfyExtension, io +from typing_extensions import override +# If you write a node that is so useless that it breaks ComfyUI it will be featured in this exclusive list + +# "native" block swap nodes are placebo at best and break the ComfyUI memory management system. +# They are also considered harmful because instead of users reporting issues with the built in +# memory management they install these stupid nodes and complain even harder. Now it completely +# breaks with some of the new ComfyUI memory optimizations so I have made the decision to NOP it +# out of all workflows. +class wanBlockSwap(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="wanBlockSwap", + category="", + description="NOP", + inputs=[ + io.Model.Input("model"), + ], + outputs=[ + io.Model.Output(), + ], + is_deprecated=True, + ) + + @classmethod + def execute(cls, model) -> io.NodeOutput: + return io.NodeOutput(model) + + +class NopExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[io.ComfyNode]]: + return [ + wanBlockSwap + ] + +async def comfy_entrypoint() -> NopExtension: + return NopExtension() diff --git a/nodes.py b/nodes.py index 5689f6fe1..f6aeedc78 100644 --- a/nodes.py +++ b/nodes.py @@ -2330,6 +2330,7 @@ async def init_builtin_extra_nodes(): "nodes_easycache.py", "nodes_audio_encoder.py", "nodes_rope.py", + "nodes_nop.py", ] import_failed = [] From 048f49adbd19ac2d9c7c87682c832b7827a4b29d Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Tue, 18 Nov 2025 13:59:27 +0200 Subject: [PATCH 120/647] chore(api-nodes): adjusted PR template; set min python version for pylint to 3.10 (#10787) --- .github/PULL_REQUEST_TEMPLATE/api-node.md | 2 +- .github/workflows/api-node-template.yml | 2 +- pyproject.toml | 2 +- 3 files changed, 3 insertions(+), 3 deletions(-) diff --git a/.github/PULL_REQUEST_TEMPLATE/api-node.md b/.github/PULL_REQUEST_TEMPLATE/api-node.md index f62744878..c1f1bafb1 100644 --- a/.github/PULL_REQUEST_TEMPLATE/api-node.md +++ b/.github/PULL_REQUEST_TEMPLATE/api-node.md @@ -18,4 +18,4 @@ If **Need pricing update**: - [ ] **QA not required** ### Comms -- [ ] Informed **@Kosinkadink** +- [ ] Informed **Kosinkadink** diff --git a/.github/workflows/api-node-template.yml b/.github/workflows/api-node-template.yml index 0775f9979..fdb81c0c5 100644 --- a/.github/workflows/api-node-template.yml +++ b/.github/workflows/api-node-template.yml @@ -2,7 +2,7 @@ name: Append API Node PR template on: pull_request_target: - types: [opened, reopened, synchronize, edited, ready_for_review] + types: [opened, reopened, synchronize, ready_for_review] paths: - 'comfy_api_nodes/**' # only run if these files changed diff --git a/pyproject.toml b/pyproject.toml index 63778286f..a14b383b3 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -24,7 +24,7 @@ lint.select = [ exclude = ["*.ipynb", "**/generated/*.pyi"] [tool.pylint] -master.py-version = "3.9" +master.py-version = "3.10" master.extension-pkg-allow-list = [ "pydantic", ] From e1ab6bb394b82fa654d5bc84043f97479d12f84c Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Jukka=20Sepp=C3=A4nen?= <40791699+kijai@users.noreply.github.com> Date: Tue, 18 Nov 2025 17:00:21 +0200 Subject: [PATCH 121/647] EasyCache: Fix for mismatch in input/output channels with some models (#10788) Slices model input with output channels so the caching tracks only the noise channels, resolves channel mismatch with models like WanVideo I2V Also fix for slicing deprecation in pytorch 2.9 --- comfy_extras/nodes_easycache.py | 22 ++++++++++++---------- 1 file changed, 12 insertions(+), 10 deletions(-) diff --git a/comfy_extras/nodes_easycache.py b/comfy_extras/nodes_easycache.py index 1359e2f99..11b23ffdb 100644 --- a/comfy_extras/nodes_easycache.py +++ b/comfy_extras/nodes_easycache.py @@ -11,13 +11,13 @@ if TYPE_CHECKING: def easycache_forward_wrapper(executor, *args, **kwargs): # get values from args - x: torch.Tensor = args[0] transformer_options: dict[str] = args[-1] if not isinstance(transformer_options, dict): transformer_options = kwargs.get("transformer_options") if not transformer_options: transformer_options = args[-2] easycache: EasyCacheHolder = transformer_options["easycache"] + x: torch.Tensor = args[0][:, :easycache.output_channels] sigmas = transformer_options["sigmas"] uuids = transformer_options["uuids"] if sigmas is not None and easycache.is_past_end_timestep(sigmas): @@ -82,13 +82,13 @@ def easycache_forward_wrapper(executor, *args, **kwargs): def lazycache_predict_noise_wrapper(executor, *args, **kwargs): # get values from args - x: torch.Tensor = args[0] timestep: float = args[1] model_options: dict[str] = args[2] easycache: LazyCacheHolder = model_options["transformer_options"]["easycache"] if easycache.is_past_end_timestep(timestep): return executor(*args, **kwargs) # prepare next x_prev + x: torch.Tensor = args[0][:, :easycache.output_channels] next_x_prev = x input_change = None do_easycache = easycache.should_do_easycache(timestep) @@ -173,7 +173,7 @@ def easycache_sample_wrapper(executor, *args, **kwargs): class EasyCacheHolder: - def __init__(self, reuse_threshold: float, start_percent: float, end_percent: float, subsample_factor: int, offload_cache_diff: bool, verbose: bool=False): + def __init__(self, reuse_threshold: float, start_percent: float, end_percent: float, subsample_factor: int, offload_cache_diff: bool, verbose: bool=False, output_channels: int=None): self.name = "EasyCache" self.reuse_threshold = reuse_threshold self.start_percent = start_percent @@ -202,6 +202,7 @@ class EasyCacheHolder: self.allow_mismatch = True self.cut_from_start = True self.state_metadata = None + self.output_channels = output_channels def is_past_end_timestep(self, timestep: float) -> bool: return not (timestep[0] > self.end_t).item() @@ -264,7 +265,7 @@ class EasyCacheHolder: else: slicing.append(slice(None)) batch_slice = batch_slice + slicing - x[batch_slice] += self.uuid_cache_diffs[uuid].to(x.device) + x[tuple(batch_slice)] += self.uuid_cache_diffs[uuid].to(x.device) return x def update_cache_diff(self, output: torch.Tensor, x: torch.Tensor, uuids: list[UUID]): @@ -283,7 +284,7 @@ class EasyCacheHolder: else: slicing.append(slice(None)) skip_dim = False - x = x[slicing] + x = x[tuple(slicing)] diff = output - x batch_offset = diff.shape[0] // len(uuids) for i, uuid in enumerate(uuids): @@ -323,7 +324,7 @@ class EasyCacheHolder: return self def clone(self): - return EasyCacheHolder(self.reuse_threshold, self.start_percent, self.end_percent, self.subsample_factor, self.offload_cache_diff, self.verbose) + return EasyCacheHolder(self.reuse_threshold, self.start_percent, self.end_percent, self.subsample_factor, self.offload_cache_diff, self.verbose, output_channels=self.output_channels) class EasyCacheNode(io.ComfyNode): @@ -350,7 +351,7 @@ class EasyCacheNode(io.ComfyNode): @classmethod def execute(cls, model: io.Model.Type, reuse_threshold: float, start_percent: float, end_percent: float, verbose: bool) -> io.NodeOutput: model = model.clone() - model.model_options["transformer_options"]["easycache"] = EasyCacheHolder(reuse_threshold, start_percent, end_percent, subsample_factor=8, offload_cache_diff=False, verbose=verbose) + model.model_options["transformer_options"]["easycache"] = EasyCacheHolder(reuse_threshold, start_percent, end_percent, subsample_factor=8, offload_cache_diff=False, verbose=verbose, output_channels=model.model.latent_format.latent_channels) model.add_wrapper_with_key(comfy.patcher_extension.WrappersMP.OUTER_SAMPLE, "easycache", easycache_sample_wrapper) model.add_wrapper_with_key(comfy.patcher_extension.WrappersMP.CALC_COND_BATCH, "easycache", easycache_calc_cond_batch_wrapper) model.add_wrapper_with_key(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, "easycache", easycache_forward_wrapper) @@ -358,7 +359,7 @@ class EasyCacheNode(io.ComfyNode): class LazyCacheHolder: - def __init__(self, reuse_threshold: float, start_percent: float, end_percent: float, subsample_factor: int, offload_cache_diff: bool, verbose: bool=False): + def __init__(self, reuse_threshold: float, start_percent: float, end_percent: float, subsample_factor: int, offload_cache_diff: bool, verbose: bool=False, output_channels: int=None): self.name = "LazyCache" self.reuse_threshold = reuse_threshold self.start_percent = start_percent @@ -382,6 +383,7 @@ class LazyCacheHolder: self.approx_output_change_rates = [] self.total_steps_skipped = 0 self.state_metadata = None + self.output_channels = output_channels def has_cache_diff(self) -> bool: return self.cache_diff is not None @@ -456,7 +458,7 @@ class LazyCacheHolder: return self def clone(self): - return LazyCacheHolder(self.reuse_threshold, self.start_percent, self.end_percent, self.subsample_factor, self.offload_cache_diff, self.verbose) + return LazyCacheHolder(self.reuse_threshold, self.start_percent, self.end_percent, self.subsample_factor, self.offload_cache_diff, self.verbose, output_channels=self.output_channels) class LazyCacheNode(io.ComfyNode): @classmethod @@ -482,7 +484,7 @@ class LazyCacheNode(io.ComfyNode): @classmethod def execute(cls, model: io.Model.Type, reuse_threshold: float, start_percent: float, end_percent: float, verbose: bool) -> io.NodeOutput: model = model.clone() - model.model_options["transformer_options"]["easycache"] = LazyCacheHolder(reuse_threshold, start_percent, end_percent, subsample_factor=8, offload_cache_diff=False, verbose=verbose) + model.model_options["transformer_options"]["easycache"] = LazyCacheHolder(reuse_threshold, start_percent, end_percent, subsample_factor=8, offload_cache_diff=False, verbose=verbose, output_channels=model.model.latent_format.latent_channels) model.add_wrapper_with_key(comfy.patcher_extension.WrappersMP.OUTER_SAMPLE, "lazycache", easycache_sample_wrapper) model.add_wrapper_with_key(comfy.patcher_extension.WrappersMP.PREDICT_NOISE, "lazycache", lazycache_predict_noise_wrapper) return io.NodeOutput(model) From d52697457608a045cafc3b6d6cb89f0a49ba0709 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 18 Nov 2025 13:46:19 -0800 Subject: [PATCH 122/647] Fix hunyuan 3d 2.0 (#10792) --- comfy/ldm/flux/math.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/comfy/ldm/flux/math.py b/comfy/ldm/flux/math.py index 158420290..6a22df8bc 100644 --- a/comfy/ldm/flux/math.py +++ b/comfy/ldm/flux/math.py @@ -7,7 +7,8 @@ import comfy.model_management def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor, mask=None, transformer_options={}) -> Tensor: - q, k = apply_rope(q, k, pe) + if pe is not None: + q, k = apply_rope(q, k, pe) heads = q.shape[1] x = optimized_attention(q, k, v, heads, skip_reshape=True, mask=mask, transformer_options=transformer_options) return x From 24fdb92edf2e96fe757c480aa7f12be5bdfa3a15 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Wed, 19 Nov 2025 00:26:44 +0200 Subject: [PATCH 123/647] feat(api-nodes): add new Gemini model (#10789) --- comfy_api_nodes/apis/gemini_api.py | 231 +++++++++++++++++++++++++++-- comfy_api_nodes/nodes_gemini.py | 47 +++--- 2 files changed, 246 insertions(+), 32 deletions(-) diff --git a/comfy_api_nodes/apis/gemini_api.py b/comfy_api_nodes/apis/gemini_api.py index 2bf28bf93..f63e02693 100644 --- a/comfy_api_nodes/apis/gemini_api.py +++ b/comfy_api_nodes/apis/gemini_api.py @@ -1,22 +1,229 @@ -from typing import Optional +from datetime import date +from enum import Enum +from typing import Any -from comfy_api_nodes.apis import GeminiGenerationConfig, GeminiContent, GeminiSafetySetting, GeminiSystemInstructionContent, GeminiTool, GeminiVideoMetadata -from pydantic import BaseModel +from pydantic import BaseModel, Field + + +class GeminiSafetyCategory(str, Enum): + HARM_CATEGORY_SEXUALLY_EXPLICIT = "HARM_CATEGORY_SEXUALLY_EXPLICIT" + HARM_CATEGORY_HATE_SPEECH = "HARM_CATEGORY_HATE_SPEECH" + HARM_CATEGORY_HARASSMENT = "HARM_CATEGORY_HARASSMENT" + HARM_CATEGORY_DANGEROUS_CONTENT = "HARM_CATEGORY_DANGEROUS_CONTENT" + + +class GeminiSafetyThreshold(str, Enum): + OFF = "OFF" + BLOCK_NONE = "BLOCK_NONE" + BLOCK_LOW_AND_ABOVE = "BLOCK_LOW_AND_ABOVE" + BLOCK_MEDIUM_AND_ABOVE = "BLOCK_MEDIUM_AND_ABOVE" + BLOCK_ONLY_HIGH = "BLOCK_ONLY_HIGH" + + +class GeminiSafetySetting(BaseModel): + category: GeminiSafetyCategory + threshold: GeminiSafetyThreshold + + +class GeminiRole(str, Enum): + user = "user" + model = "model" + + +class GeminiMimeType(str, Enum): + application_pdf = "application/pdf" + audio_mpeg = "audio/mpeg" + audio_mp3 = "audio/mp3" + audio_wav = "audio/wav" + image_png = "image/png" + image_jpeg = "image/jpeg" + image_webp = "image/webp" + text_plain = "text/plain" + video_mov = "video/mov" + video_mpeg = "video/mpeg" + video_mp4 = "video/mp4" + video_mpg = "video/mpg" + video_avi = "video/avi" + video_wmv = "video/wmv" + video_mpegps = "video/mpegps" + video_flv = "video/flv" + + +class GeminiInlineData(BaseModel): + data: str | None = Field( + None, + description="The base64 encoding of the image, PDF, or video to include inline in the prompt. " + "When including media inline, you must also specify the media type (mimeType) of the data. Size limit: 20MB", + ) + mimeType: GeminiMimeType | None = Field(None) + + +class GeminiPart(BaseModel): + inlineData: GeminiInlineData | None = Field(None) + text: str | None = Field(None) + + +class GeminiTextPart(BaseModel): + text: str | None = Field(None) + + +class GeminiContent(BaseModel): + parts: list[GeminiPart] = Field(...) + role: GeminiRole = Field(..., examples=["user"]) + + +class GeminiSystemInstructionContent(BaseModel): + parts: list[GeminiTextPart] = Field( + ..., + description="A list of ordered parts that make up a single message. " + "Different parts may have different IANA MIME types.", + ) + role: GeminiRole = Field( + ..., + description="The identity of the entity that creates the message. " + "The following values are supported: " + "user: This indicates that the message is sent by a real person, typically a user-generated message. " + "model: This indicates that the message is generated by the model. " + "The model value is used to insert messages from model into the conversation during multi-turn conversations. " + "For non-multi-turn conversations, this field can be left blank or unset.", + ) + + +class GeminiFunctionDeclaration(BaseModel): + description: str | None = Field(None) + name: str = Field(...) + parameters: dict[str, Any] = Field(..., description="JSON schema for the function parameters") + + +class GeminiTool(BaseModel): + functionDeclarations: list[GeminiFunctionDeclaration] | None = Field(None) + + +class GeminiOffset(BaseModel): + nanos: int | None = Field(None, ge=0, le=999999999) + seconds: int | None = Field(None, ge=-315576000000, le=315576000000) + + +class GeminiVideoMetadata(BaseModel): + endOffset: GeminiOffset | None = Field(None) + startOffset: GeminiOffset | None = Field(None) + + +class GeminiGenerationConfig(BaseModel): + maxOutputTokens: int | None = Field(None, ge=16, le=8192) + seed: int | None = Field(None) + stopSequences: list[str] | None = Field(None) + temperature: float | None = Field(1, ge=0.0, le=2.0) + topK: int | None = Field(40, ge=1) + topP: float | None = Field(0.95, ge=0.0, le=1.0) class GeminiImageConfig(BaseModel): - aspectRatio: Optional[str] = None + aspectRatio: str | None = Field(None) + resolution: str | None = Field(None) class GeminiImageGenerationConfig(GeminiGenerationConfig): - responseModalities: Optional[list[str]] = None - imageConfig: Optional[GeminiImageConfig] = None + responseModalities: list[str] | None = Field(None) + imageConfig: GeminiImageConfig | None = Field(None) class GeminiImageGenerateContentRequest(BaseModel): - contents: list[GeminiContent] - generationConfig: Optional[GeminiImageGenerationConfig] = None - safetySettings: Optional[list[GeminiSafetySetting]] = None - systemInstruction: Optional[GeminiSystemInstructionContent] = None - tools: Optional[list[GeminiTool]] = None - videoMetadata: Optional[GeminiVideoMetadata] = None + contents: list[GeminiContent] = Field(...) + generationConfig: GeminiImageGenerationConfig | None = Field(None) + safetySettings: list[GeminiSafetySetting] | None = Field(None) + systemInstruction: GeminiSystemInstructionContent | None = Field(None) + tools: list[GeminiTool] | None = Field(None) + videoMetadata: GeminiVideoMetadata | None = Field(None) + + +class GeminiGenerateContentRequest(BaseModel): + contents: list[GeminiContent] = Field(...) + generationConfig: GeminiGenerationConfig | None = Field(None) + safetySettings: list[GeminiSafetySetting] | None = Field(None) + systemInstruction: GeminiSystemInstructionContent | None = Field(None) + tools: list[GeminiTool] | None = Field(None) + videoMetadata: GeminiVideoMetadata | None = Field(None) + + +class Modality(str, Enum): + MODALITY_UNSPECIFIED = "MODALITY_UNSPECIFIED" + TEXT = "TEXT" + IMAGE = "IMAGE" + VIDEO = "VIDEO" + AUDIO = "AUDIO" + DOCUMENT = "DOCUMENT" + + +class ModalityTokenCount(BaseModel): + modality: Modality | None = None + tokenCount: int | None = Field(None, description="Number of tokens for the given modality.") + + +class Probability(str, Enum): + NEGLIGIBLE = "NEGLIGIBLE" + LOW = "LOW" + MEDIUM = "MEDIUM" + HIGH = "HIGH" + UNKNOWN = "UNKNOWN" + + +class GeminiSafetyRating(BaseModel): + category: GeminiSafetyCategory | None = None + probability: Probability | None = Field( + None, + description="The probability that the content violates the specified safety category", + ) + + +class GeminiCitation(BaseModel): + authors: list[str] | None = None + endIndex: int | None = None + license: str | None = None + publicationDate: date | None = None + startIndex: int | None = None + title: str | None = None + uri: str | None = None + + +class GeminiCitationMetadata(BaseModel): + citations: list[GeminiCitation] | None = None + + +class GeminiCandidate(BaseModel): + citationMetadata: GeminiCitationMetadata | None = None + content: GeminiContent | None = None + finishReason: str | None = None + safetyRatings: list[GeminiSafetyRating] | None = None + + +class GeminiPromptFeedback(BaseModel): + blockReason: str | None = None + blockReasonMessage: str | None = None + safetyRatings: list[GeminiSafetyRating] | None = None + + +class GeminiUsageMetadata(BaseModel): + cachedContentTokenCount: int | None = Field( + None, + description="Output only. Number of tokens in the cached part in the input (the cached content).", + ) + candidatesTokenCount: int | None = Field(None, description="Number of tokens in the response(s).") + candidatesTokensDetails: list[ModalityTokenCount] | None = Field( + None, description="Breakdown of candidate tokens by modality." + ) + promptTokenCount: int | None = Field( + None, + description="Number of tokens in the request. When cachedContent is set, this is still the total effective prompt size meaning this includes the number of tokens in the cached content.", + ) + promptTokensDetails: list[ModalityTokenCount] | None = Field( + None, description="Breakdown of prompt tokens by modality." + ) + thoughtsTokenCount: int | None = Field(None, description="Number of tokens present in thoughts output.") + toolUsePromptTokenCount: int | None = Field(None, description="Number of tokens present in tool-use prompt(s).") + + +class GeminiGenerateContentResponse(BaseModel): + candidates: list[GeminiCandidate] | None = Field(None) + promptFeedback: GeminiPromptFeedback | None = Field(None) + usageMetadata: GeminiUsageMetadata | None = Field(None) diff --git a/comfy_api_nodes/nodes_gemini.py b/comfy_api_nodes/nodes_gemini.py index 67f2469ad..6e746eebd 100644 --- a/comfy_api_nodes/nodes_gemini.py +++ b/comfy_api_nodes/nodes_gemini.py @@ -3,8 +3,6 @@ API Nodes for Gemini Multimodal LLM Usage via Remote API See: https://cloud.google.com/vertex-ai/generative-ai/docs/model-reference/inference """ -from __future__ import annotations - import base64 import json import os @@ -12,7 +10,7 @@ import time import uuid from enum import Enum from io import BytesIO -from typing import Literal, Optional +from typing import Literal import torch from typing_extensions import override @@ -20,18 +18,17 @@ from typing_extensions import override import folder_paths from comfy_api.latest import IO, ComfyExtension, Input from comfy_api.util import VideoCodec, VideoContainer -from comfy_api_nodes.apis import ( +from comfy_api_nodes.apis.gemini_api import ( GeminiContent, GeminiGenerateContentRequest, GeminiGenerateContentResponse, - GeminiInlineData, - GeminiMimeType, - GeminiPart, -) -from comfy_api_nodes.apis.gemini_api import ( GeminiImageConfig, GeminiImageGenerateContentRequest, GeminiImageGenerationConfig, + GeminiInlineData, + GeminiMimeType, + GeminiPart, + GeminiRole, ) from comfy_api_nodes.util import ( ApiEndpoint, @@ -57,6 +54,7 @@ class GeminiModel(str, Enum): gemini_2_5_flash_preview_04_17 = "gemini-2.5-flash-preview-04-17" gemini_2_5_pro = "gemini-2.5-pro" gemini_2_5_flash = "gemini-2.5-flash" + gemini_3_0_pro = "gemini-3-pro-preview" class GeminiImageModel(str, Enum): @@ -103,6 +101,16 @@ def get_parts_by_type(response: GeminiGenerateContentResponse, part_type: Litera Returns: List of response parts matching the requested type. """ + if response.candidates is None: + if response.promptFeedback.blockReason: + feedback = response.promptFeedback + raise ValueError( + f"Gemini API blocked the request. Reason: {feedback.blockReason} ({feedback.blockReasonMessage})" + ) + raise NotImplementedError( + "Gemini returned no response candidates. " + "Please report to ComfyUI repository with the example of workflow to reproduce this." + ) parts = [] for part in response.candidates[0].content.parts: if part_type == "text" and hasattr(part, "text") and part.text: @@ -272,10 +280,10 @@ class GeminiNode(IO.ComfyNode): prompt: str, model: str, seed: int, - images: Optional[torch.Tensor] = None, - audio: Optional[Input.Audio] = None, - video: Optional[Input.Video] = None, - files: Optional[list[GeminiPart]] = None, + images: torch.Tensor | None = None, + audio: Input.Audio | None = None, + video: Input.Video | None = None, + files: list[GeminiPart] | None = None, ) -> IO.NodeOutput: validate_string(prompt, strip_whitespace=False) @@ -300,7 +308,7 @@ class GeminiNode(IO.ComfyNode): data=GeminiGenerateContentRequest( contents=[ GeminiContent( - role="user", + role=GeminiRole.user, parts=parts, ) ] @@ -308,7 +316,6 @@ class GeminiNode(IO.ComfyNode): response_model=GeminiGenerateContentResponse, ) - # Get result output output_text = get_text_from_response(response) if output_text: # Not a true chat history like the OpenAI Chat node. It is emulated so the frontend can show a copy button. @@ -406,7 +413,7 @@ class GeminiInputFiles(IO.ComfyNode): ) @classmethod - def execute(cls, file: str, GEMINI_INPUT_FILES: Optional[list[GeminiPart]] = None) -> IO.NodeOutput: + def execute(cls, file: str, GEMINI_INPUT_FILES: list[GeminiPart] | None = None) -> IO.NodeOutput: """Loads and formats input files for Gemini API.""" if GEMINI_INPUT_FILES is None: GEMINI_INPUT_FILES = [] @@ -421,7 +428,7 @@ class GeminiImage(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="GeminiImageNode", - display_name="Google Gemini Image", + display_name="Nano Banana (Google Gemini Image)", category="api node/image/Gemini", description="Edit images synchronously via Google API.", inputs=[ @@ -488,8 +495,8 @@ class GeminiImage(IO.ComfyNode): prompt: str, model: str, seed: int, - images: Optional[torch.Tensor] = None, - files: Optional[list[GeminiPart]] = None, + images: torch.Tensor | None = None, + files: list[GeminiPart] | None = None, aspect_ratio: str = "auto", ) -> IO.NodeOutput: validate_string(prompt, strip_whitespace=True, min_length=1) @@ -510,7 +517,7 @@ class GeminiImage(IO.ComfyNode): endpoint=ApiEndpoint(path=f"{GEMINI_BASE_ENDPOINT}/{model}", method="POST"), data=GeminiImageGenerateContentRequest( contents=[ - GeminiContent(role="user", parts=parts), + GeminiContent(role=GeminiRole.user, parts=parts), ], generationConfig=GeminiImageGenerationConfig( responseModalities=["TEXT", "IMAGE"], From b5c8be8b1db44ded07cb1b437b9f33ebff5848c1 Mon Sep 17 00:00:00 2001 From: comfyanonymous Date: Tue, 18 Nov 2025 19:37:20 -0500 Subject: [PATCH 124/647] ComfyUI 0.3.70 --- comfyui_version.py | 2 +- pyproject.toml | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/comfyui_version.py b/comfyui_version.py index 1e554eb9f..9b77aabe9 100644 --- a/comfyui_version.py +++ b/comfyui_version.py @@ -1,3 +1,3 @@ # This file is automatically generated by the build process when version is # updated in pyproject.toml. -__version__ = "0.3.69" +__version__ = "0.3.70" diff --git a/pyproject.toml b/pyproject.toml index a14b383b3..289b7145b 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -1,6 +1,6 @@ [project] name = "ComfyUI" -version = "0.3.69" +version = "0.3.70" readme = "README.md" license = { file = "LICENSE" } requires-python = ">=3.9" From 17027f2a6a20a31e2c6f3be2b1a06f39ad3a68d9 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 18 Nov 2025 19:36:03 -0800 Subject: [PATCH 125/647] Add a way to disable the final norm in the llama based TE models. (#10794) --- comfy/text_encoders/llama.py | 18 +++++++++++++++--- 1 file changed, 15 insertions(+), 3 deletions(-) diff --git a/comfy/text_encoders/llama.py b/comfy/text_encoders/llama.py index c050759fe..feb44bbb0 100644 --- a/comfy/text_encoders/llama.py +++ b/comfy/text_encoders/llama.py @@ -32,6 +32,7 @@ class Llama2Config: q_norm = None k_norm = None rope_scale = None + final_norm: bool = True @dataclass class Qwen25_3BConfig: @@ -53,6 +54,7 @@ class Qwen25_3BConfig: q_norm = None k_norm = None rope_scale = None + final_norm: bool = True @dataclass class Qwen25_7BVLI_Config: @@ -74,6 +76,7 @@ class Qwen25_7BVLI_Config: q_norm = None k_norm = None rope_scale = None + final_norm: bool = True @dataclass class Gemma2_2B_Config: @@ -96,6 +99,7 @@ class Gemma2_2B_Config: k_norm = None sliding_attention = None rope_scale = None + final_norm: bool = True @dataclass class Gemma3_4B_Config: @@ -118,6 +122,7 @@ class Gemma3_4B_Config: k_norm = "gemma3" sliding_attention = [False, False, False, False, False, 1024] rope_scale = [1.0, 8.0] + final_norm: bool = True class RMSNorm(nn.Module): def __init__(self, dim: int, eps: float = 1e-5, add=False, device=None, dtype=None): @@ -366,7 +371,12 @@ class Llama2_(nn.Module): transformer(config, index=i, device=device, dtype=dtype, ops=ops) for i in range(config.num_hidden_layers) ]) - self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps, add=config.rms_norm_add, device=device, dtype=dtype) + + if config.final_norm: + self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps, add=config.rms_norm_add, device=device, dtype=dtype) + else: + self.norm = None + # self.lm_head = ops.Linear(config.hidden_size, config.vocab_size, bias=False, device=device, dtype=dtype) def forward(self, x, attention_mask=None, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None, position_ids=None, embeds_info=[]): @@ -421,14 +431,16 @@ class Llama2_(nn.Module): if i == intermediate_output: intermediate = x.clone() - x = self.norm(x) + if self.norm is not None: + x = self.norm(x) + if all_intermediate is not None: all_intermediate.append(x.unsqueeze(1).clone()) if all_intermediate is not None: intermediate = torch.cat(all_intermediate, dim=1) - if intermediate is not None and final_layer_norm_intermediate: + if intermediate is not None and final_layer_norm_intermediate and self.norm is not None: intermediate = self.norm(intermediate) return x, intermediate From 65ee24c9789b93660ebe978a3186486f105298c2 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Wed, 19 Nov 2025 11:25:28 +0200 Subject: [PATCH 126/647] change display name of PreviewAny node to "Preview as Text" (#10796) --- comfy_extras/nodes_preview_any.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy_extras/nodes_preview_any.py b/comfy_extras/nodes_preview_any.py index e749fa6ae..139b07c93 100644 --- a/comfy_extras/nodes_preview_any.py +++ b/comfy_extras/nodes_preview_any.py @@ -39,5 +39,5 @@ NODE_CLASS_MAPPINGS = { } NODE_DISPLAY_NAME_MAPPINGS = { - "PreviewAny": "Preview Any", + "PreviewAny": "Preview as Text", } From 6a1d3a1ae131f3fff7f45a7e835eb10e9d1338ee Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Thu, 20 Nov 2025 00:49:01 +0200 Subject: [PATCH 127/647] convert hunyuan3d.py to V3 schema (#10664) --- comfy_api/latest/__init__.py | 4 +- comfy_api/latest/_io.py | 5 +- comfy_api/latest/_util/__init__.py | 3 + comfy_api/latest/_util/geometry_types.py | 12 + comfy_extras/nodes_hunyuan3d.py | 274 +++++++++++++---------- 5 files changed, 178 insertions(+), 120 deletions(-) create mode 100644 comfy_api/latest/_util/geometry_types.py diff --git a/comfy_api/latest/__init__.py b/comfy_api/latest/__init__.py index b7a3fa9c1..176ae36e0 100644 --- a/comfy_api/latest/__init__.py +++ b/comfy_api/latest/__init__.py @@ -7,7 +7,7 @@ from comfy_api.internal.singleton import ProxiedSingleton from comfy_api.internal.async_to_sync import create_sync_class from comfy_api.latest._input import ImageInput, AudioInput, MaskInput, LatentInput, VideoInput from comfy_api.latest._input_impl import VideoFromFile, VideoFromComponents -from comfy_api.latest._util import VideoCodec, VideoContainer, VideoComponents +from comfy_api.latest._util import VideoCodec, VideoContainer, VideoComponents, MESH, VOXEL from . import _io as io from . import _ui as ui # from comfy_api.latest._resources import _RESOURCES as resources #noqa: F401 @@ -104,6 +104,8 @@ class Types: VideoCodec = VideoCodec VideoContainer = VideoContainer VideoComponents = VideoComponents + MESH = MESH + VOXEL = VOXEL ComfyAPI = ComfyAPI_latest diff --git a/comfy_api/latest/_io.py b/comfy_api/latest/_io.py index 0b701260f..863254ce7 100644 --- a/comfy_api/latest/_io.py +++ b/comfy_api/latest/_io.py @@ -27,6 +27,7 @@ from comfy_api.internal import (_ComfyNodeInternal, _NodeOutputInternal, classpr prune_dict, shallow_clone_class) from comfy_api.latest._resources import Resources, ResourcesLocal from comfy_execution.graph_utils import ExecutionBlocker +from ._util import MESH, VOXEL # from comfy_extras.nodes_images import SVG as SVG_ # NOTE: needs to be moved before can be imported due to circular reference @@ -656,11 +657,11 @@ class LossMap(ComfyTypeIO): @comfytype(io_type="VOXEL") class Voxel(ComfyTypeIO): - Type = Any # TODO: VOXEL class is defined in comfy_extras/nodes_hunyuan3d.py; should be moved to somewhere else before referenced directly in v3 + Type = VOXEL @comfytype(io_type="MESH") class Mesh(ComfyTypeIO): - Type = Any # TODO: MESH class is defined in comfy_extras/nodes_hunyuan3d.py; should be moved to somewhere else before referenced directly in v3 + Type = MESH @comfytype(io_type="HOOKS") class Hooks(ComfyTypeIO): diff --git a/comfy_api/latest/_util/__init__.py b/comfy_api/latest/_util/__init__.py index 9019c46db..fc5431dda 100644 --- a/comfy_api/latest/_util/__init__.py +++ b/comfy_api/latest/_util/__init__.py @@ -1,8 +1,11 @@ from .video_types import VideoContainer, VideoCodec, VideoComponents +from .geometry_types import VOXEL, MESH __all__ = [ # Utility Types "VideoContainer", "VideoCodec", "VideoComponents", + "VOXEL", + "MESH", ] diff --git a/comfy_api/latest/_util/geometry_types.py b/comfy_api/latest/_util/geometry_types.py new file mode 100644 index 000000000..385122778 --- /dev/null +++ b/comfy_api/latest/_util/geometry_types.py @@ -0,0 +1,12 @@ +import torch + + +class VOXEL: + def __init__(self, data: torch.Tensor): + self.data = data + + +class MESH: + def __init__(self, vertices: torch.Tensor, faces: torch.Tensor): + self.vertices = vertices + self.faces = faces diff --git a/comfy_extras/nodes_hunyuan3d.py b/comfy_extras/nodes_hunyuan3d.py index f6e71e0a8..adca14f62 100644 --- a/comfy_extras/nodes_hunyuan3d.py +++ b/comfy_extras/nodes_hunyuan3d.py @@ -7,63 +7,79 @@ from comfy.ldm.modules.diffusionmodules.mmdit import get_1d_sincos_pos_embed_fro import folder_paths import comfy.model_management from comfy.cli_args import args +from typing_extensions import override +from comfy_api.latest import ComfyExtension, IO, Types +from comfy_api.latest._util import MESH, VOXEL # only for backward compatibility if someone import it from this file (will be removed later) # noqa -class EmptyLatentHunyuan3Dv2: + +class EmptyLatentHunyuan3Dv2(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return { - "required": { - "resolution": ("INT", {"default": 3072, "min": 1, "max": 8192}), - "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096, "tooltip": "The number of latent images in the batch."}), - } - } + def define_schema(cls): + return IO.Schema( + node_id="EmptyLatentHunyuan3Dv2", + category="latent/3d", + inputs=[ + IO.Int.Input("resolution", default=3072, min=1, max=8192), + IO.Int.Input("batch_size", default=1, min=1, max=4096, tooltip="The number of latent images in the batch."), + ], + outputs=[ + IO.Latent.Output(), + ] + ) - RETURN_TYPES = ("LATENT",) - FUNCTION = "generate" - - CATEGORY = "latent/3d" - - def generate(self, resolution, batch_size): + @classmethod + def execute(cls, resolution, batch_size) -> IO.NodeOutput: latent = torch.zeros([batch_size, 64, resolution], device=comfy.model_management.intermediate_device()) - return ({"samples": latent, "type": "hunyuan3dv2"}, ) + return IO.NodeOutput({"samples": latent, "type": "hunyuan3dv2"}) -class Hunyuan3Dv2Conditioning: + generate = execute # TODO: remove + + +class Hunyuan3Dv2Conditioning(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": {"clip_vision_output": ("CLIP_VISION_OUTPUT",), - }} + def define_schema(cls): + return IO.Schema( + node_id="Hunyuan3Dv2Conditioning", + category="conditioning/video_models", + inputs=[ + IO.ClipVisionOutput.Input("clip_vision_output"), + ], + outputs=[ + IO.Conditioning.Output(display_name="positive"), + IO.Conditioning.Output(display_name="negative"), + ] + ) - RETURN_TYPES = ("CONDITIONING", "CONDITIONING") - RETURN_NAMES = ("positive", "negative") - - FUNCTION = "encode" - - CATEGORY = "conditioning/video_models" - - def encode(self, clip_vision_output): + @classmethod + def execute(cls, clip_vision_output) -> IO.NodeOutput: embeds = clip_vision_output.last_hidden_state positive = [[embeds, {}]] negative = [[torch.zeros_like(embeds), {}]] - return (positive, negative) + return IO.NodeOutput(positive, negative) + + encode = execute # TODO: remove -class Hunyuan3Dv2ConditioningMultiView: +class Hunyuan3Dv2ConditioningMultiView(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": {}, - "optional": {"front": ("CLIP_VISION_OUTPUT",), - "left": ("CLIP_VISION_OUTPUT",), - "back": ("CLIP_VISION_OUTPUT",), - "right": ("CLIP_VISION_OUTPUT",), }} + def define_schema(cls): + return IO.Schema( + node_id="Hunyuan3Dv2ConditioningMultiView", + category="conditioning/video_models", + inputs=[ + IO.ClipVisionOutput.Input("front", optional=True), + IO.ClipVisionOutput.Input("left", optional=True), + IO.ClipVisionOutput.Input("back", optional=True), + IO.ClipVisionOutput.Input("right", optional=True), + ], + outputs=[ + IO.Conditioning.Output(display_name="positive"), + IO.Conditioning.Output(display_name="negative"), + ] + ) - RETURN_TYPES = ("CONDITIONING", "CONDITIONING") - RETURN_NAMES = ("positive", "negative") - - FUNCTION = "encode" - - CATEGORY = "conditioning/video_models" - - def encode(self, front=None, left=None, back=None, right=None): + @classmethod + def execute(cls, front=None, left=None, back=None, right=None) -> IO.NodeOutput: all_embeds = [front, left, back, right] out = [] pos_embeds = None @@ -76,29 +92,35 @@ class Hunyuan3Dv2ConditioningMultiView: embeds = torch.cat(out, dim=1) positive = [[embeds, {}]] negative = [[torch.zeros_like(embeds), {}]] - return (positive, negative) + return IO.NodeOutput(positive, negative) + + encode = execute # TODO: remove -class VOXEL: - def __init__(self, data): - self.data = data - -class VAEDecodeHunyuan3D: +class VAEDecodeHunyuan3D(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": {"samples": ("LATENT", ), - "vae": ("VAE", ), - "num_chunks": ("INT", {"default": 8000, "min": 1000, "max": 500000}), - "octree_resolution": ("INT", {"default": 256, "min": 16, "max": 512}), - }} - RETURN_TYPES = ("VOXEL",) - FUNCTION = "decode" + def define_schema(cls): + return IO.Schema( + node_id="VAEDecodeHunyuan3D", + category="latent/3d", + inputs=[ + IO.Latent.Input("samples"), + IO.Vae.Input("vae"), + IO.Int.Input("num_chunks", default=8000, min=1000, max=500000), + IO.Int.Input("octree_resolution", default=256, min=16, max=512), + ], + outputs=[ + IO.Voxel.Output(), + ] + ) - CATEGORY = "latent/3d" + @classmethod + def execute(cls, vae, samples, num_chunks, octree_resolution) -> IO.NodeOutput: + voxels = Types.VOXEL(vae.decode(samples["samples"], vae_options={"num_chunks": num_chunks, "octree_resolution": octree_resolution})) + return IO.NodeOutput(voxels) + + decode = execute # TODO: remove - def decode(self, vae, samples, num_chunks, octree_resolution): - voxels = VOXEL(vae.decode(samples["samples"], vae_options={"num_chunks": num_chunks, "octree_resolution": octree_resolution})) - return (voxels, ) def voxel_to_mesh(voxels, threshold=0.5, device=None): if device is None: @@ -396,24 +418,24 @@ def voxel_to_mesh_surfnet(voxels, threshold=0.5, device=None): return final_vertices, faces -class MESH: - def __init__(self, vertices, faces): - self.vertices = vertices - self.faces = faces - -class VoxelToMeshBasic: +class VoxelToMeshBasic(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": {"voxel": ("VOXEL", ), - "threshold": ("FLOAT", {"default": 0.6, "min": -1.0, "max": 1.0, "step": 0.01}), - }} - RETURN_TYPES = ("MESH",) - FUNCTION = "decode" + def define_schema(cls): + return IO.Schema( + node_id="VoxelToMeshBasic", + category="3d", + inputs=[ + IO.Voxel.Input("voxel"), + IO.Float.Input("threshold", default=0.6, min=-1.0, max=1.0, step=0.01), + ], + outputs=[ + IO.Mesh.Output(), + ] + ) - CATEGORY = "3d" - - def decode(self, voxel, threshold): + @classmethod + def execute(cls, voxel, threshold) -> IO.NodeOutput: vertices = [] faces = [] for x in voxel.data: @@ -421,21 +443,29 @@ class VoxelToMeshBasic: vertices.append(v) faces.append(f) - return (MESH(torch.stack(vertices), torch.stack(faces)), ) + return IO.NodeOutput(Types.MESH(torch.stack(vertices), torch.stack(faces))) -class VoxelToMesh: + decode = execute # TODO: remove + + +class VoxelToMesh(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": {"voxel": ("VOXEL", ), - "algorithm": (["surface net", "basic"], ), - "threshold": ("FLOAT", {"default": 0.6, "min": -1.0, "max": 1.0, "step": 0.01}), - }} - RETURN_TYPES = ("MESH",) - FUNCTION = "decode" + def define_schema(cls): + return IO.Schema( + node_id="VoxelToMesh", + category="3d", + inputs=[ + IO.Voxel.Input("voxel"), + IO.Combo.Input("algorithm", options=["surface net", "basic"]), + IO.Float.Input("threshold", default=0.6, min=-1.0, max=1.0, step=0.01), + ], + outputs=[ + IO.Mesh.Output(), + ] + ) - CATEGORY = "3d" - - def decode(self, voxel, algorithm, threshold): + @classmethod + def execute(cls, voxel, algorithm, threshold) -> IO.NodeOutput: vertices = [] faces = [] @@ -449,7 +479,9 @@ class VoxelToMesh: vertices.append(v) faces.append(f) - return (MESH(torch.stack(vertices), torch.stack(faces)), ) + return IO.NodeOutput(Types.MESH(torch.stack(vertices), torch.stack(faces))) + + decode = execute # TODO: remove def save_glb(vertices, faces, filepath, metadata=None): @@ -581,31 +613,32 @@ def save_glb(vertices, faces, filepath, metadata=None): return filepath -class SaveGLB: +class SaveGLB(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": {"mesh": ("MESH", ), - "filename_prefix": ("STRING", {"default": "mesh/ComfyUI"}), }, - "hidden": {"prompt": "PROMPT", "extra_pnginfo": "EXTRA_PNGINFO"}, } + def define_schema(cls): + return IO.Schema( + node_id="SaveGLB", + category="3d", + is_output_node=True, + inputs=[ + IO.Mesh.Input("mesh"), + IO.String.Input("filename_prefix", default="mesh/ComfyUI"), + ], + hidden=[IO.Hidden.prompt, IO.Hidden.extra_pnginfo] + ) - RETURN_TYPES = () - FUNCTION = "save" - - OUTPUT_NODE = True - - CATEGORY = "3d" - - def save(self, mesh, filename_prefix, prompt=None, extra_pnginfo=None): + @classmethod + def execute(cls, mesh, filename_prefix) -> IO.NodeOutput: full_output_folder, filename, counter, subfolder, filename_prefix = folder_paths.get_save_image_path(filename_prefix, folder_paths.get_output_directory()) results = [] metadata = {} if not args.disable_metadata: - if prompt is not None: - metadata["prompt"] = json.dumps(prompt) - if extra_pnginfo is not None: - for x in extra_pnginfo: - metadata[x] = json.dumps(extra_pnginfo[x]) + if cls.hidden.prompt is not None: + metadata["prompt"] = json.dumps(cls.hidden.prompt) + if cls.hidden.extra_pnginfo is not None: + for x in cls.hidden.extra_pnginfo: + metadata[x] = json.dumps(cls.hidden.extra_pnginfo[x]) for i in range(mesh.vertices.shape[0]): f = f"{filename}_{counter:05}_.glb" @@ -616,15 +649,22 @@ class SaveGLB: "type": "output" }) counter += 1 - return {"ui": {"3d": results}} + return IO.NodeOutput(ui={"3d": results}) -NODE_CLASS_MAPPINGS = { - "EmptyLatentHunyuan3Dv2": EmptyLatentHunyuan3Dv2, - "Hunyuan3Dv2Conditioning": Hunyuan3Dv2Conditioning, - "Hunyuan3Dv2ConditioningMultiView": Hunyuan3Dv2ConditioningMultiView, - "VAEDecodeHunyuan3D": VAEDecodeHunyuan3D, - "VoxelToMeshBasic": VoxelToMeshBasic, - "VoxelToMesh": VoxelToMesh, - "SaveGLB": SaveGLB, -} +class Hunyuan3dExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[IO.ComfyNode]]: + return [ + EmptyLatentHunyuan3Dv2, + Hunyuan3Dv2Conditioning, + Hunyuan3Dv2ConditioningMultiView, + VAEDecodeHunyuan3D, + VoxelToMeshBasic, + VoxelToMesh, + SaveGLB, + ] + + +async def comfy_entrypoint() -> Hunyuan3dExtension: + return Hunyuan3dExtension() From 7601e89255cde24667d3b4e6022f1385d901748b Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 19 Nov 2025 17:17:15 -0800 Subject: [PATCH 128/647] Fix workflow name. (#10806) --- .github/workflows/release-stable-all.yml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/.github/workflows/release-stable-all.yml b/.github/workflows/release-stable-all.yml index f7de3a7c3..9274b4170 100644 --- a/.github/workflows/release-stable-all.yml +++ b/.github/workflows/release-stable-all.yml @@ -14,7 +14,7 @@ jobs: contents: "write" packages: "write" pull-requests: "read" - name: "Release NVIDIA Default (cu129)" + name: "Release NVIDIA Default (cu130)" uses: ./.github/workflows/stable-release.yml with: git_tag: ${{ inputs.git_tag }} From 394348f5caaa062eac11a57e2997aacccd4246eb Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Thu, 20 Nov 2025 03:44:04 +0200 Subject: [PATCH 129/647] feat(api-nodes): add Topaz API nodes (#10755) --- comfy_api_nodes/apis/topaz_api.py | 133 ++++++++++ comfy_api_nodes/nodes_topaz.py | 421 ++++++++++++++++++++++++++++++ comfy_api_nodes/util/client.py | 9 +- nodes.py | 1 + 4 files changed, 560 insertions(+), 4 deletions(-) create mode 100644 comfy_api_nodes/apis/topaz_api.py create mode 100644 comfy_api_nodes/nodes_topaz.py diff --git a/comfy_api_nodes/apis/topaz_api.py b/comfy_api_nodes/apis/topaz_api.py new file mode 100644 index 000000000..4d9e62e72 --- /dev/null +++ b/comfy_api_nodes/apis/topaz_api.py @@ -0,0 +1,133 @@ +from typing import Optional, Union + +from pydantic import BaseModel, Field + + +class ImageEnhanceRequest(BaseModel): + model: str = Field("Reimagine") + output_format: str = Field("jpeg") + subject_detection: str = Field("All") + face_enhancement: bool = Field(True) + face_enhancement_creativity: float = Field(0, description="Is ignored if face_enhancement is false") + face_enhancement_strength: float = Field(0.8, description="Is ignored if face_enhancement is false") + source_url: str = Field(...) + output_width: Optional[int] = Field(None) + output_height: Optional[int] = Field(None) + crop_to_fill: bool = Field(False) + prompt: Optional[str] = Field(None, description="Text prompt for creative upscaling guidance") + creativity: int = Field(3, description="Creativity settings range from 1 to 9") + face_preservation: str = Field("true", description="To preserve the identity of characters") + color_preservation: str = Field("true", description="To preserve the original color") + + +class ImageAsyncTaskResponse(BaseModel): + process_id: str = Field(...) + + +class ImageStatusResponse(BaseModel): + process_id: str = Field(...) + status: str = Field(...) + progress: Optional[int] = Field(None) + credits: int = Field(...) + + +class ImageDownloadResponse(BaseModel): + download_url: str = Field(...) + expiry: int = Field(...) + + +class Resolution(BaseModel): + width: int = Field(...) + height: int = Field(...) + + +class CreateCreateVideoRequestSource(BaseModel): + container: str = Field(...) + size: int = Field(..., description="Size of the video file in bytes") + duration: int = Field(..., description="Duration of the video file in seconds") + frameCount: int = Field(..., description="Total number of frames in the video") + frameRate: int = Field(...) + resolution: Resolution = Field(...) + + +class VideoFrameInterpolationFilter(BaseModel): + model: str = Field(...) + slowmo: Optional[int] = Field(None) + fps: int = Field(...) + duplicate: bool = Field(...) + duplicate_threshold: float = Field(...) + + +class VideoEnhancementFilter(BaseModel): + model: str = Field(...) + auto: Optional[str] = Field(None, description="Auto, Manual, Relative") + focusFixLevel: Optional[str] = Field(None, description="Downscales video input for correction of blurred subjects") + compression: Optional[float] = Field(None, description="Strength of compression recovery") + details: Optional[float] = Field(None, description="Amount of detail reconstruction") + prenoise: Optional[float] = Field(None, description="Amount of noise to add to input to reduce over-smoothing") + noise: Optional[float] = Field(None, description="Amount of noise reduction") + halo: Optional[float] = Field(None, description="Amount of halo reduction") + preblur: Optional[float] = Field(None, description="Anti-aliasing and deblurring strength") + blur: Optional[float] = Field(None, description="Amount of sharpness applied") + grain: Optional[float] = Field(None, description="Grain after AI model processing") + grainSize: Optional[float] = Field(None, description="Size of generated grain") + recoverOriginalDetailValue: Optional[float] = Field(None, description="Source details into the output video") + creativity: Optional[str] = Field(None, description="Creativity level(high, low) for slc-1 only") + isOptimizedMode: Optional[bool] = Field(None, description="Set to true for Starlight Creative (slc-1) only") + + +class OutputInformationVideo(BaseModel): + resolution: Resolution = Field(...) + frameRate: int = Field(...) + audioCodec: Optional[str] = Field(..., description="Required if audioTransfer is Copy or Convert") + audioTransfer: str = Field(..., description="Copy, Convert, None") + dynamicCompressionLevel: str = Field(..., description="Low, Mid, High") + + +class Overrides(BaseModel): + isPaidDiffusion: bool = Field(True) + + +class CreateVideoRequest(BaseModel): + source: CreateCreateVideoRequestSource = Field(...) + filters: list[Union[VideoFrameInterpolationFilter, VideoEnhancementFilter]] = Field(...) + output: OutputInformationVideo = Field(...) + overrides: Overrides = Field(Overrides(isPaidDiffusion=True)) + + +class CreateVideoResponse(BaseModel): + requestId: str = Field(...) + + +class VideoAcceptResponse(BaseModel): + uploadId: str = Field(...) + urls: list[str] = Field(...) + + +class VideoCompleteUploadRequestPart(BaseModel): + partNum: int = Field(...) + eTag: str = Field(...) + + +class VideoCompleteUploadRequest(BaseModel): + uploadResults: list[VideoCompleteUploadRequestPart] = Field(...) + + +class VideoCompleteUploadResponse(BaseModel): + message: str = Field(..., description="Confirmation message") + + +class VideoStatusResponseEstimates(BaseModel): + cost: list[int] = Field(...) + + +class VideoStatusResponseDownloadUrl(BaseModel): + url: str = Field(...) + + +class VideoStatusResponse(BaseModel): + status: str = Field(...) + estimates: Optional[VideoStatusResponseEstimates] = Field(None) + progress: Optional[float] = Field(None) + message: Optional[str] = Field("") + download: Optional[VideoStatusResponseDownloadUrl] = Field(None) diff --git a/comfy_api_nodes/nodes_topaz.py b/comfy_api_nodes/nodes_topaz.py new file mode 100644 index 000000000..79c7bf43d --- /dev/null +++ b/comfy_api_nodes/nodes_topaz.py @@ -0,0 +1,421 @@ +import builtins +from io import BytesIO + +import aiohttp +import torch +from typing_extensions import override + +from comfy_api.input.video_types import VideoInput +from comfy_api.latest import IO, ComfyExtension +from comfy_api_nodes.apis import topaz_api +from comfy_api_nodes.util import ( + ApiEndpoint, + download_url_to_image_tensor, + download_url_to_video_output, + get_fs_object_size, + get_number_of_images, + poll_op, + sync_op, + upload_images_to_comfyapi, + validate_container_format_is_mp4, +) + +UPSCALER_MODELS_MAP = { + "Starlight (Astra) Fast": "slf-1", + "Starlight (Astra) Creative": "slc-1", +} +UPSCALER_VALUES_MAP = { + "FullHD (1080p)": 1920, + "4K (2160p)": 3840, +} + + +class TopazImageEnhance(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="TopazImageEnhance", + display_name="Topaz Image Enhance", + category="api node/image/Topaz", + description="Industry-standard upscaling and image enhancement.", + inputs=[ + IO.Combo.Input("model", options=["Reimagine"]), + IO.Image.Input("image"), + IO.String.Input( + "prompt", + multiline=True, + default="", + tooltip="Optional text prompt for creative upscaling guidance.", + optional=True, + ), + IO.Combo.Input( + "subject_detection", + options=["All", "Foreground", "Background"], + optional=True, + ), + IO.Boolean.Input( + "face_enhancement", + default=True, + optional=True, + tooltip="Enhance faces (if present) during processing.", + ), + IO.Float.Input( + "face_enhancement_creativity", + default=0.0, + min=0.0, + max=1.0, + step=0.01, + display_mode=IO.NumberDisplay.number, + optional=True, + tooltip="Set the creativity level for face enhancement.", + ), + IO.Float.Input( + "face_enhancement_strength", + default=1.0, + min=0.0, + max=1.0, + step=0.01, + display_mode=IO.NumberDisplay.number, + optional=True, + tooltip="Controls how sharp enhanced faces are relative to the background.", + ), + IO.Boolean.Input( + "crop_to_fill", + default=False, + optional=True, + tooltip="By default, the image is letterboxed when the output aspect ratio differs. " + "Enable to crop the image to fill the output dimensions.", + ), + IO.Int.Input( + "output_width", + default=0, + min=0, + max=32000, + step=1, + display_mode=IO.NumberDisplay.number, + optional=True, + tooltip="Zero value means to calculate automatically (usually it will be original size or output_height if specified).", + ), + IO.Int.Input( + "output_height", + default=0, + min=0, + max=32000, + step=1, + display_mode=IO.NumberDisplay.number, + optional=True, + tooltip="Zero value means to output in the same height as original or output width.", + ), + IO.Int.Input( + "creativity", + default=3, + min=1, + max=9, + step=1, + display_mode=IO.NumberDisplay.slider, + optional=True, + ), + IO.Boolean.Input( + "face_preservation", + default=True, + optional=True, + tooltip="Preserve subjects' facial identity.", + ), + IO.Boolean.Input( + "color_preservation", + default=True, + optional=True, + tooltip="Preserve the original colors.", + ), + ], + outputs=[ + IO.Image.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + ) + + @classmethod + async def execute( + cls, + model: str, + image: torch.Tensor, + prompt: str = "", + subject_detection: str = "All", + face_enhancement: bool = True, + face_enhancement_creativity: float = 1.0, + face_enhancement_strength: float = 0.8, + crop_to_fill: bool = False, + output_width: int = 0, + output_height: int = 0, + creativity: int = 3, + face_preservation: bool = True, + color_preservation: bool = True, + ) -> IO.NodeOutput: + if get_number_of_images(image) != 1: + raise ValueError("Only one input image is supported.") + download_url = await upload_images_to_comfyapi(cls, image, max_images=1, mime_type="image/png") + initial_response = await sync_op( + cls, + ApiEndpoint(path="/proxy/topaz/image/v1/enhance-gen/async", method="POST"), + response_model=topaz_api.ImageAsyncTaskResponse, + data=topaz_api.ImageEnhanceRequest( + model=model, + prompt=prompt, + subject_detection=subject_detection, + face_enhancement=face_enhancement, + face_enhancement_creativity=face_enhancement_creativity, + face_enhancement_strength=face_enhancement_strength, + crop_to_fill=crop_to_fill, + output_width=output_width if output_width else None, + output_height=output_height if output_height else None, + creativity=creativity, + face_preservation=str(face_preservation).lower(), + color_preservation=str(color_preservation).lower(), + source_url=download_url[0], + output_format="png", + ), + content_type="multipart/form-data", + ) + + await poll_op( + cls, + poll_endpoint=ApiEndpoint(path=f"/proxy/topaz/image/v1/status/{initial_response.process_id}"), + response_model=topaz_api.ImageStatusResponse, + status_extractor=lambda x: x.status, + progress_extractor=lambda x: getattr(x, "progress", 0), + price_extractor=lambda x: x.credits * 0.08, + poll_interval=8.0, + max_poll_attempts=160, + estimated_duration=60, + ) + + results = await sync_op( + cls, + ApiEndpoint(path=f"/proxy/topaz/image/v1/download/{initial_response.process_id}"), + response_model=topaz_api.ImageDownloadResponse, + monitor_progress=False, + ) + return IO.NodeOutput(await download_url_to_image_tensor(results.download_url)) + + +class TopazVideoEnhance(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="TopazVideoEnhance", + display_name="Topaz Video Enhance", + category="api node/video/Topaz", + description="Breathe new life into video with powerful upscaling and recovery technology.", + inputs=[ + IO.Video.Input("video"), + IO.Boolean.Input("upscaler_enabled", default=True), + IO.Combo.Input("upscaler_model", options=list(UPSCALER_MODELS_MAP.keys())), + IO.Combo.Input("upscaler_resolution", options=list(UPSCALER_VALUES_MAP.keys())), + IO.Combo.Input( + "upscaler_creativity", + options=["low", "middle", "high"], + default="low", + tooltip="Creativity level (applies only to Starlight (Astra) Creative).", + optional=True, + ), + IO.Boolean.Input("interpolation_enabled", default=False, optional=True), + IO.Combo.Input("interpolation_model", options=["apo-8"], default="apo-8", optional=True), + IO.Int.Input( + "interpolation_slowmo", + default=1, + min=1, + max=16, + display_mode=IO.NumberDisplay.number, + tooltip="Slow-motion factor applied to the input video. " + "For example, 2 makes the output twice as slow and doubles the duration.", + optional=True, + ), + IO.Int.Input( + "interpolation_frame_rate", + default=60, + min=15, + max=240, + display_mode=IO.NumberDisplay.number, + tooltip="Output frame rate.", + optional=True, + ), + IO.Boolean.Input( + "interpolation_duplicate", + default=False, + tooltip="Analyze the input for duplicate frames and remove them.", + optional=True, + ), + IO.Float.Input( + "interpolation_duplicate_threshold", + default=0.01, + min=0.001, + max=0.1, + step=0.001, + display_mode=IO.NumberDisplay.number, + tooltip="Detection sensitivity for duplicate frames.", + optional=True, + ), + IO.Combo.Input( + "dynamic_compression_level", + options=["Low", "Mid", "High"], + default="Low", + tooltip="CQP level.", + optional=True, + ), + ], + outputs=[ + IO.Video.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + ) + + @classmethod + async def execute( + cls, + video: VideoInput, + upscaler_enabled: bool, + upscaler_model: str, + upscaler_resolution: str, + upscaler_creativity: str = "low", + interpolation_enabled: bool = False, + interpolation_model: str = "apo-8", + interpolation_slowmo: int = 1, + interpolation_frame_rate: int = 60, + interpolation_duplicate: bool = False, + interpolation_duplicate_threshold: float = 0.01, + dynamic_compression_level: str = "Low", + ) -> IO.NodeOutput: + if upscaler_enabled is False and interpolation_enabled is False: + raise ValueError("There is nothing to do: both upscaling and interpolation are disabled.") + src_width, src_height = video.get_dimensions() + video_components = video.get_components() + src_frame_rate = int(video_components.frame_rate) + duration_sec = video.get_duration() + estimated_frames = int(duration_sec * src_frame_rate) + validate_container_format_is_mp4(video) + src_video_stream = video.get_stream_source() + target_width = src_width + target_height = src_height + target_frame_rate = src_frame_rate + filters = [] + if upscaler_enabled: + target_width = UPSCALER_VALUES_MAP[upscaler_resolution] + target_height = UPSCALER_VALUES_MAP[upscaler_resolution] + filters.append( + topaz_api.VideoEnhancementFilter( + model=UPSCALER_MODELS_MAP[upscaler_model], + creativity=(upscaler_creativity if UPSCALER_MODELS_MAP[upscaler_model] == "slc-1" else None), + isOptimizedMode=(True if UPSCALER_MODELS_MAP[upscaler_model] == "slc-1" else None), + ), + ) + if interpolation_enabled: + target_frame_rate = interpolation_frame_rate + filters.append( + topaz_api.VideoFrameInterpolationFilter( + model=interpolation_model, + slowmo=interpolation_slowmo, + fps=interpolation_frame_rate, + duplicate=interpolation_duplicate, + duplicate_threshold=interpolation_duplicate_threshold, + ), + ) + initial_res = await sync_op( + cls, + ApiEndpoint(path="/proxy/topaz/video/", method="POST"), + response_model=topaz_api.CreateVideoResponse, + data=topaz_api.CreateVideoRequest( + source=topaz_api.CreateCreateVideoRequestSource( + container="mp4", + size=get_fs_object_size(src_video_stream), + duration=int(duration_sec), + frameCount=estimated_frames, + frameRate=src_frame_rate, + resolution=topaz_api.Resolution(width=src_width, height=src_height), + ), + filters=filters, + output=topaz_api.OutputInformationVideo( + resolution=topaz_api.Resolution(width=target_width, height=target_height), + frameRate=target_frame_rate, + audioCodec="AAC", + audioTransfer="Copy", + dynamicCompressionLevel=dynamic_compression_level, + ), + ), + wait_label="Creating task", + final_label_on_success="Task created", + ) + upload_res = await sync_op( + cls, + ApiEndpoint( + path=f"/proxy/topaz/video/{initial_res.requestId}/accept", + method="PATCH", + ), + response_model=topaz_api.VideoAcceptResponse, + wait_label="Preparing upload", + final_label_on_success="Upload started", + ) + if len(upload_res.urls) > 1: + raise NotImplementedError( + "Large files are not currently supported. Please open an issue in the ComfyUI repository." + ) + async with aiohttp.ClientSession(headers={"Content-Type": "video/mp4"}) as session: + if isinstance(src_video_stream, BytesIO): + src_video_stream.seek(0) + async with session.put(upload_res.urls[0], data=src_video_stream, raise_for_status=True) as res: + upload_etag = res.headers["Etag"] + else: + with builtins.open(src_video_stream, "rb") as video_file: + async with session.put(upload_res.urls[0], data=video_file, raise_for_status=True) as res: + upload_etag = res.headers["Etag"] + await sync_op( + cls, + ApiEndpoint( + path=f"/proxy/topaz/video/{initial_res.requestId}/complete-upload", + method="PATCH", + ), + response_model=topaz_api.VideoCompleteUploadResponse, + data=topaz_api.VideoCompleteUploadRequest( + uploadResults=[ + topaz_api.VideoCompleteUploadRequestPart( + partNum=1, + eTag=upload_etag, + ), + ], + ), + wait_label="Finalizing upload", + final_label_on_success="Upload completed", + ) + final_response = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/topaz/video/{initial_res.requestId}/status"), + response_model=topaz_api.VideoStatusResponse, + status_extractor=lambda x: x.status, + progress_extractor=lambda x: getattr(x, "progress", 0), + price_extractor=lambda x: (x.estimates.cost[0] * 0.08 if x.estimates and x.estimates.cost[0] else None), + poll_interval=10.0, + max_poll_attempts=320, + ) + return IO.NodeOutput(await download_url_to_video_output(final_response.download.url)) + + +class TopazExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[IO.ComfyNode]]: + return [ + TopazImageEnhance, + TopazVideoEnhance, + ] + + +async def comfy_entrypoint() -> TopazExtension: + return TopazExtension() diff --git a/comfy_api_nodes/util/client.py b/comfy_api_nodes/util/client.py index 2d5dcd648..ad6e3c0d0 100644 --- a/comfy_api_nodes/util/client.py +++ b/comfy_api_nodes/util/client.py @@ -77,9 +77,9 @@ class _PollUIState: _RETRY_STATUS = {408, 429, 500, 502, 503, 504} -COMPLETED_STATUSES = ["succeeded", "succeed", "success", "completed", "finished", "done"] -FAILED_STATUSES = ["cancelled", "canceled", "fail", "failed", "error"] -QUEUED_STATUSES = ["created", "queued", "queueing", "submitted"] +COMPLETED_STATUSES = ["succeeded", "succeed", "success", "completed", "finished", "done", "complete"] +FAILED_STATUSES = ["cancelled", "canceled", "canceling", "fail", "failed", "error"] +QUEUED_STATUSES = ["created", "queued", "queueing", "submitted", "initializing"] async def sync_op( @@ -424,7 +424,8 @@ def _display_text( if status: display_lines.append(f"Status: {status.capitalize() if isinstance(status, str) else status}") if price is not None: - display_lines.append(f"Price: ${float(price):,.4f}") + p = f"{float(price):,.4f}".rstrip("0").rstrip(".") + display_lines.append(f"Price: ${p}") if text is not None: display_lines.append(text) if display_lines: diff --git a/nodes.py b/nodes.py index f6aeedc78..ac14e39a7 100644 --- a/nodes.py +++ b/nodes.py @@ -2359,6 +2359,7 @@ async def init_builtin_api_nodes(): "nodes_pika.py", "nodes_runway.py", "nodes_sora.py", + "nodes_topaz.py", "nodes_tripo.py", "nodes_moonvalley.py", "nodes_rodin.py", From cb96d4d18c78ee09d5fd70954ffcb4ad2c7f0d7a Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 19 Nov 2025 20:56:23 -0800 Subject: [PATCH 130/647] Disable workaround on newer cudnn. (#10807) --- comfy/ops.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/comfy/ops.py b/comfy/ops.py index 2a90a5ba2..640622fd1 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -58,7 +58,8 @@ except (ModuleNotFoundError, TypeError): NVIDIA_MEMORY_CONV_BUG_WORKAROUND = False try: if comfy.model_management.is_nvidia(): - if torch.backends.cudnn.version() >= 91002 and comfy.model_management.torch_version_numeric >= (2, 9) and comfy.model_management.torch_version_numeric <= (2, 10): + cudnn_version = torch.backends.cudnn.version() + if (cudnn_version >= 91002 and cudnn_version < 91500) and comfy.model_management.torch_version_numeric >= (2, 9) and comfy.model_management.torch_version_numeric <= (2, 10): #TODO: change upper bound version once it's fixed' NVIDIA_MEMORY_CONV_BUG_WORKAROUND = True logging.info("working around nvidia conv3d memory bug.") From 87b0359392219841c2214e1eb06678840cae470e Mon Sep 17 00:00:00 2001 From: Christian Byrne Date: Wed, 19 Nov 2025 22:36:56 -0800 Subject: [PATCH 131/647] Update server templates handler to use new multi-package distribution (comfyui-workflow-templates versions >=0.3) (#10791) * update templates for monorepo * refactor --- app/frontend_management.py | 67 ++++++++++++++++++++++++++++++++++++-- requirements.txt | 2 +- server.py | 32 ++++++++++++++---- 3 files changed, 92 insertions(+), 9 deletions(-) diff --git a/app/frontend_management.py b/app/frontend_management.py index cce0c117d..bdaa85812 100644 --- a/app/frontend_management.py +++ b/app/frontend_management.py @@ -10,7 +10,8 @@ import importlib from dataclasses import dataclass from functools import cached_property from pathlib import Path -from typing import TypedDict, Optional +from typing import Dict, TypedDict, Optional +from aiohttp import web from importlib.metadata import version import requests @@ -257,7 +258,54 @@ comfyui-frontend-package is not installed. sys.exit(-1) @classmethod - def templates_path(cls) -> str: + def template_asset_map(cls) -> Optional[Dict[str, str]]: + """Return a mapping of template asset names to their absolute paths.""" + try: + from comfyui_workflow_templates import ( + get_asset_path, + iter_templates, + ) + except ImportError: + logging.error( + f""" +********** ERROR *********** + +comfyui-workflow-templates is not installed. + +{frontend_install_warning_message()} + +********** ERROR *********** +""".strip() + ) + return None + + try: + template_entries = list(iter_templates()) + except Exception as exc: + logging.error(f"Failed to enumerate workflow templates: {exc}") + return None + + asset_map: Dict[str, str] = {} + try: + for entry in template_entries: + for asset in entry.assets: + asset_map[asset.filename] = get_asset_path( + entry.template_id, asset.filename + ) + except Exception as exc: + logging.error(f"Failed to resolve template asset paths: {exc}") + return None + + if not asset_map: + logging.error("No workflow template assets found. Did the packages install correctly?") + return None + + return asset_map + + + @classmethod + def legacy_templates_path(cls) -> Optional[str]: + """Return the legacy templates directory shipped inside the meta package.""" try: import comfyui_workflow_templates @@ -276,6 +324,7 @@ comfyui-workflow-templates is not installed. ********** ERROR *********** """.strip() ) + return None @classmethod def embedded_docs_path(cls) -> str: @@ -392,3 +441,17 @@ comfyui-workflow-templates is not installed. logging.info("Falling back to the default frontend.") check_frontend_version() return cls.default_frontend_path() + @classmethod + def template_asset_handler(cls): + assets = cls.template_asset_map() + if not assets: + return None + + async def serve_template(request: web.Request) -> web.StreamResponse: + rel_path = request.match_info.get("path", "") + target = assets.get(rel_path) + if target is None: + raise web.HTTPNotFound() + return web.FileResponse(target) + + return serve_template diff --git a/requirements.txt b/requirements.txt index 249c36dee..36c39f338 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,5 +1,5 @@ comfyui-frontend-package==1.28.8 -comfyui-workflow-templates==0.2.11 +comfyui-workflow-templates==0.3.1 comfyui-embedded-docs==0.3.1 torch torchsde diff --git a/server.py b/server.py index d059d3dc9..d9d5c491f 100644 --- a/server.py +++ b/server.py @@ -30,7 +30,7 @@ import comfy.model_management from comfy_api import feature_flags import node_helpers from comfyui_version import __version__ -from app.frontend_management import FrontendManager +from app.frontend_management import FrontendManager, parse_version from comfy_api.internal import _ComfyNodeInternal from app.user_manager import UserManager @@ -849,11 +849,31 @@ class PromptServer(): for name, dir in nodes.EXTENSION_WEB_DIRS.items(): self.app.add_routes([web.static('/extensions/' + name, dir)]) - workflow_templates_path = FrontendManager.templates_path() - if workflow_templates_path: - self.app.add_routes([ - web.static('/templates', workflow_templates_path) - ]) + installed_templates_version = FrontendManager.get_installed_templates_version() + use_legacy_templates = True + if installed_templates_version: + try: + use_legacy_templates = ( + parse_version(installed_templates_version) + < parse_version("0.3.0") + ) + except Exception as exc: + logging.warning( + "Unable to parse templates version '%s': %s", + installed_templates_version, + exc, + ) + + if use_legacy_templates: + workflow_templates_path = FrontendManager.legacy_templates_path() + if workflow_templates_path: + self.app.add_routes([ + web.static('/templates', workflow_templates_path) + ]) + else: + handler = FrontendManager.template_asset_handler() + if handler: + self.app.router.add_get("/templates/{path:.*}", handler) # Serve embedded documentation from the package embedded_docs_path = FrontendManager.embedded_docs_path() From f5e66d5e47271253edad5c4eddd817b0d6a23340 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Thu, 20 Nov 2025 12:08:03 -0800 Subject: [PATCH 132/647] Fix ImageBatch with different channel count. (#10815) --- nodes.py | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/nodes.py b/nodes.py index ac14e39a7..75e820e66 100644 --- a/nodes.py +++ b/nodes.py @@ -1852,6 +1852,10 @@ class ImageBatch: CATEGORY = "image" def batch(self, image1, image2): + if image1.shape[-1] != image2.shape[-1]: + channels = min(image1.shape[-1], image2.shape[-1]) + image1 = image1[..., :channels] + image2 = image2[..., :channels] if image1.shape[1:] != image2.shape[1:]: image2 = comfy.utils.common_upscale(image2.movedim(-1,1), image1.shape[2], image1.shape[1], "bilinear", "center").movedim(1,-1) s = torch.cat((image1, image2), dim=0) From 9e00ce5b76ec04be37375310512a443605b95077 Mon Sep 17 00:00:00 2001 From: Jedrzej Kosinski Date: Thu, 20 Nov 2025 14:42:46 -0800 Subject: [PATCH 133/647] Make Batch Images node add alpha channel when one of the inputs has it (#10816) * When one Batch Image input has alpha and one does not, add empty alpha channel * Use torch.nn.functional.pad --- nodes.py | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-) diff --git a/nodes.py b/nodes.py index 75e820e66..030371633 100644 --- a/nodes.py +++ b/nodes.py @@ -1853,9 +1853,10 @@ class ImageBatch: def batch(self, image1, image2): if image1.shape[-1] != image2.shape[-1]: - channels = min(image1.shape[-1], image2.shape[-1]) - image1 = image1[..., :channels] - image2 = image2[..., :channels] + if image1.shape[-1] > image2.shape[-1]: + image2 = torch.nn.functional.pad(image2, (0,1), mode='constant', value=1.0) + else: + image1 = torch.nn.functional.pad(image1, (0,1), mode='constant', value=1.0) if image1.shape[1:] != image2.shape[1:]: image2 = comfy.utils.common_upscale(image2.movedim(-1,1), image1.shape[2], image1.shape[1], "bilinear", "center").movedim(1,-1) s = torch.cat((image1, image2), dim=0) From 7b8389578e88dcd13b1cf6aea5404047298c9183 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Fri, 21 Nov 2025 02:17:47 +0200 Subject: [PATCH 134/647] feat(api-nodes): add Nano Banana Pro (#10814) * feat(api-nodes): add Nano Banana Pro * frontend bump to 1.28.9 --- comfy_api_nodes/apis/gemini_api.py | 5 +- comfy_api_nodes/nodes_gemini.py | 205 ++++++++++++++++++++++++++++- comfy_api_nodes/util/client.py | 13 +- requirements.txt | 2 +- 4 files changed, 215 insertions(+), 10 deletions(-) diff --git a/comfy_api_nodes/apis/gemini_api.py b/comfy_api_nodes/apis/gemini_api.py index f63e02693..710f173f1 100644 --- a/comfy_api_nodes/apis/gemini_api.py +++ b/comfy_api_nodes/apis/gemini_api.py @@ -68,7 +68,7 @@ class GeminiTextPart(BaseModel): class GeminiContent(BaseModel): - parts: list[GeminiPart] = Field(...) + parts: list[GeminiPart] = Field([]) role: GeminiRole = Field(..., examples=["user"]) @@ -120,7 +120,7 @@ class GeminiGenerationConfig(BaseModel): class GeminiImageConfig(BaseModel): aspectRatio: str | None = Field(None) - resolution: str | None = Field(None) + imageSize: str | None = Field(None) class GeminiImageGenerationConfig(GeminiGenerationConfig): @@ -227,3 +227,4 @@ class GeminiGenerateContentResponse(BaseModel): candidates: list[GeminiCandidate] | None = Field(None) promptFeedback: GeminiPromptFeedback | None = Field(None) usageMetadata: GeminiUsageMetadata | None = Field(None) + modelVersion: str | None = Field(None) diff --git a/comfy_api_nodes/nodes_gemini.py b/comfy_api_nodes/nodes_gemini.py index 6e746eebd..be752c885 100644 --- a/comfy_api_nodes/nodes_gemini.py +++ b/comfy_api_nodes/nodes_gemini.py @@ -29,11 +29,13 @@ from comfy_api_nodes.apis.gemini_api import ( GeminiMimeType, GeminiPart, GeminiRole, + Modality, ) from comfy_api_nodes.util import ( ApiEndpoint, audio_to_base64_string, bytesio_to_image_tensor, + get_number_of_images, sync_op, tensor_to_base64_string, validate_string, @@ -147,6 +149,49 @@ def get_image_from_response(response: GeminiGenerateContentResponse) -> torch.Te return torch.cat(image_tensors, dim=0) +def calculate_tokens_price(response: GeminiGenerateContentResponse) -> float | None: + if not response.modelVersion: + return None + # Define prices (Cost per 1,000,000 tokens), see https://cloud.google.com/vertex-ai/generative-ai/pricing + if response.modelVersion in ("gemini-2.5-pro-preview-05-06", "gemini-2.5-pro"): + input_tokens_price = 1.25 + output_text_tokens_price = 10.0 + output_image_tokens_price = 0.0 + elif response.modelVersion in ( + "gemini-2.5-flash-preview-04-17", + "gemini-2.5-flash", + ): + input_tokens_price = 0.30 + output_text_tokens_price = 2.50 + output_image_tokens_price = 0.0 + elif response.modelVersion in ( + "gemini-2.5-flash-image-preview", + "gemini-2.5-flash-image", + ): + input_tokens_price = 0.30 + output_text_tokens_price = 2.50 + output_image_tokens_price = 30.0 + elif response.modelVersion == "gemini-3-pro-preview": + input_tokens_price = 2 + output_text_tokens_price = 12.0 + output_image_tokens_price = 0.0 + elif response.modelVersion == "gemini-3-pro-image-preview": + input_tokens_price = 2 + output_text_tokens_price = 12.0 + output_image_tokens_price = 120.0 + else: + return None + final_price = response.usageMetadata.promptTokenCount * input_tokens_price + for i in response.usageMetadata.candidatesTokensDetails: + if i.modality == Modality.IMAGE: + final_price += output_image_tokens_price * i.tokenCount # for Nano Banana models + else: + final_price += output_text_tokens_price * i.tokenCount + if response.usageMetadata.thoughtsTokenCount: + final_price += output_text_tokens_price * response.usageMetadata.thoughtsTokenCount + return final_price / 1_000_000.0 + + class GeminiNode(IO.ComfyNode): """ Node to generate text responses from a Gemini model. @@ -314,6 +359,7 @@ class GeminiNode(IO.ComfyNode): ] ), response_model=GeminiGenerateContentResponse, + price_extractor=calculate_tokens_price, ) output_text = get_text_from_response(response) @@ -476,6 +522,13 @@ class GeminiImage(IO.ComfyNode): "or otherwise generates 1:1 squares.", optional=True, ), + IO.Combo.Input( + "response_modalities", + options=["IMAGE+TEXT", "IMAGE"], + tooltip="Choose 'IMAGE' for image-only output, or " + "'IMAGE+TEXT' to return both the generated image and a text response.", + optional=True, + ), ], outputs=[ IO.Image.Output(), @@ -498,6 +551,7 @@ class GeminiImage(IO.ComfyNode): images: torch.Tensor | None = None, files: list[GeminiPart] | None = None, aspect_ratio: str = "auto", + response_modalities: str = "IMAGE+TEXT", ) -> IO.NodeOutput: validate_string(prompt, strip_whitespace=True, min_length=1) parts: list[GeminiPart] = [GeminiPart(text=prompt)] @@ -520,17 +574,16 @@ class GeminiImage(IO.ComfyNode): GeminiContent(role=GeminiRole.user, parts=parts), ], generationConfig=GeminiImageGenerationConfig( - responseModalities=["TEXT", "IMAGE"], + responseModalities=(["IMAGE"] if response_modalities == "IMAGE" else ["TEXT", "IMAGE"]), imageConfig=None if aspect_ratio == "auto" else image_config, ), ), response_model=GeminiGenerateContentResponse, + price_extractor=calculate_tokens_price, ) - output_image = get_image_from_response(response) output_text = get_text_from_response(response) if output_text: - # Not a true chat history like the OpenAI Chat node. It is emulated so the frontend can show a copy button. render_spec = { "node_id": cls.hidden.unique_id, "component": "ChatHistoryWidget", @@ -551,9 +604,150 @@ class GeminiImage(IO.ComfyNode): "display_component", render_spec, ) + return IO.NodeOutput(get_image_from_response(response), output_text) - output_text = output_text or "Empty response from Gemini model..." - return IO.NodeOutput(output_image, output_text) + +class GeminiImage2(IO.ComfyNode): + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="GeminiImage2Node", + display_name="Nano Banana Pro (Google Gemini Image)", + category="api node/image/Gemini", + description="Generate or edit images synchronously via Google Vertex API.", + inputs=[ + IO.String.Input( + "prompt", + multiline=True, + tooltip="Text prompt describing the image to generate or the edits to apply. " + "Include any constraints, styles, or details the model should follow.", + default="", + ), + IO.Combo.Input( + "model", + options=["gemini-3-pro-image-preview"], + ), + IO.Int.Input( + "seed", + default=42, + min=0, + max=0xFFFFFFFFFFFFFFFF, + control_after_generate=True, + tooltip="When the seed is fixed to a specific value, the model makes a best effort to provide " + "the same response for repeated requests. Deterministic output isn't guaranteed. " + "Also, changing the model or parameter settings, such as the temperature, " + "can cause variations in the response even when you use the same seed value. " + "By default, a random seed value is used.", + ), + IO.Combo.Input( + "aspect_ratio", + options=["auto", "1:1", "2:3", "3:2", "3:4", "4:3", "4:5", "5:4", "9:16", "16:9", "21:9"], + default="auto", + tooltip="If set to 'auto', matches your input image's aspect ratio; " + "if no image is provided, generates a 1:1 square.", + ), + IO.Combo.Input( + "resolution", + options=["1K", "2K", "4K"], + tooltip="Target output resolution. For 2K/4K the native Gemini upscaler is used.", + ), + IO.Combo.Input( + "response_modalities", + options=["IMAGE+TEXT", "IMAGE"], + tooltip="Choose 'IMAGE' for image-only output, or " + "'IMAGE+TEXT' to return both the generated image and a text response.", + ), + IO.Image.Input( + "images", + optional=True, + tooltip="Optional reference image(s). " + "To include multiple images, use the Batch Images node (up to 14).", + ), + IO.Custom("GEMINI_INPUT_FILES").Input( + "files", + optional=True, + tooltip="Optional file(s) to use as context for the model. " + "Accepts inputs from the Gemini Generate Content Input Files node.", + ), + ], + outputs=[ + IO.Image.Output(), + IO.String.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + ) + + @classmethod + async def execute( + cls, + prompt: str, + model: str, + seed: int, + aspect_ratio: str, + resolution: str, + response_modalities: str, + images: torch.Tensor | None = None, + files: list[GeminiPart] | None = None, + ) -> IO.NodeOutput: + validate_string(prompt, strip_whitespace=True, min_length=1) + + parts: list[GeminiPart] = [GeminiPart(text=prompt)] + if images is not None: + if get_number_of_images(images) > 14: + raise ValueError("The current maximum number of supported images is 14.") + parts.extend(create_image_parts(images)) + if files is not None: + parts.extend(files) + + image_config = GeminiImageConfig(imageSize=resolution) + if aspect_ratio != "auto": + image_config.aspectRatio = aspect_ratio + + response = await sync_op( + cls, + ApiEndpoint(path=f"{GEMINI_BASE_ENDPOINT}/{model}", method="POST"), + data=GeminiImageGenerateContentRequest( + contents=[ + GeminiContent(role=GeminiRole.user, parts=parts), + ], + generationConfig=GeminiImageGenerationConfig( + responseModalities=(["IMAGE"] if response_modalities == "IMAGE" else ["TEXT", "IMAGE"]), + imageConfig=image_config, + ), + ), + response_model=GeminiGenerateContentResponse, + price_extractor=calculate_tokens_price, + ) + + output_text = get_text_from_response(response) + if output_text: + render_spec = { + "node_id": cls.hidden.unique_id, + "component": "ChatHistoryWidget", + "props": { + "history": json.dumps( + [ + { + "prompt": prompt, + "response": output_text, + "response_id": str(uuid.uuid4()), + "timestamp": time.time(), + } + ] + ), + }, + } + PromptServer.instance.send_sync( + "display_component", + render_spec, + ) + return IO.NodeOutput(get_image_from_response(response), output_text) class GeminiExtension(ComfyExtension): @@ -562,6 +756,7 @@ class GeminiExtension(ComfyExtension): return [ GeminiNode, GeminiImage, + GeminiImage2, GeminiInputFiles, ] diff --git a/comfy_api_nodes/util/client.py b/comfy_api_nodes/util/client.py index ad6e3c0d0..bf01d7d36 100644 --- a/comfy_api_nodes/util/client.py +++ b/comfy_api_nodes/util/client.py @@ -63,6 +63,7 @@ class _RequestConfig: estimated_total: Optional[int] = None final_label_on_success: Optional[str] = "Completed" progress_origin_ts: Optional[float] = None + price_extractor: Optional[Callable[[dict[str, Any]], Optional[float]]] = None @dataclass @@ -87,6 +88,7 @@ async def sync_op( endpoint: ApiEndpoint, *, response_model: Type[M], + price_extractor: Optional[Callable[[M], Optional[float]]] = None, data: Optional[BaseModel] = None, files: Optional[Union[dict[str, Any], list[tuple[str, Any]]]] = None, content_type: str = "application/json", @@ -104,6 +106,7 @@ async def sync_op( raw = await sync_op_raw( cls, endpoint, + price_extractor=_wrap_model_extractor(response_model, price_extractor), data=data, files=files, content_type=content_type, @@ -175,6 +178,7 @@ async def sync_op_raw( cls: type[IO.ComfyNode], endpoint: ApiEndpoint, *, + price_extractor: Optional[Callable[[dict[str, Any]], Optional[float]]] = None, data: Optional[Union[dict[str, Any], BaseModel]] = None, files: Optional[Union[dict[str, Any], list[tuple[str, Any]]]] = None, content_type: str = "application/json", @@ -216,6 +220,7 @@ async def sync_op_raw( estimated_total=estimated_duration, final_label_on_success=final_label_on_success, progress_origin_ts=progress_origin_ts, + price_extractor=price_extractor, ) return await _request_base(cfg, expect_binary=as_binary) @@ -425,7 +430,8 @@ def _display_text( display_lines.append(f"Status: {status.capitalize() if isinstance(status, str) else status}") if price is not None: p = f"{float(price):,.4f}".rstrip("0").rstrip(".") - display_lines.append(f"Price: ${p}") + if p != "0": + display_lines.append(f"Price: ${p}") if text is not None: display_lines.append(text) if display_lines: @@ -581,6 +587,7 @@ async def _request_base(cfg: _RequestConfig, expect_binary: bool): delay = cfg.retry_delay operation_succeeded: bool = False final_elapsed_seconds: Optional[int] = None + extracted_price: Optional[float] = None while True: attempt += 1 stop_event = asyncio.Event() @@ -768,6 +775,8 @@ async def _request_base(cfg: _RequestConfig, expect_binary: bool): except json.JSONDecodeError: payload = {"_raw": text} response_content_to_log = payload if isinstance(payload, dict) else text + with contextlib.suppress(Exception): + extracted_price = cfg.price_extractor(payload) if cfg.price_extractor else None operation_succeeded = True final_elapsed_seconds = int(time.monotonic() - start_time) try: @@ -872,7 +881,7 @@ async def _request_base(cfg: _RequestConfig, expect_binary: bool): else int(time.monotonic() - start_time) ), estimated_total=cfg.estimated_total, - price=None, + price=extracted_price, is_queued=False, processing_elapsed_seconds=final_elapsed_seconds, ) diff --git a/requirements.txt b/requirements.txt index 36c39f338..8c1946f3d 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,4 +1,4 @@ -comfyui-frontend-package==1.28.8 +comfyui-frontend-package==1.28.9 comfyui-workflow-templates==0.3.1 comfyui-embedded-docs==0.3.1 torch From b75d349f25ccb702895c6f1b8af7aded63a7f7e2 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Fri, 21 Nov 2025 02:33:54 +0200 Subject: [PATCH 135/647] fix(KlingLipSyncAudioToVideoNode): convert audio to mp3 format (#10811) --- comfy_api_nodes/nodes_kling.py | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/comfy_api_nodes/nodes_kling.py b/comfy_api_nodes/nodes_kling.py index 7b23e9cf9..36852038b 100644 --- a/comfy_api_nodes/nodes_kling.py +++ b/comfy_api_nodes/nodes_kling.py @@ -518,7 +518,9 @@ async def execute_lipsync( # Upload the audio file to Comfy API and get download URL if audio: - audio_url = await upload_audio_to_comfyapi(cls, audio) + audio_url = await upload_audio_to_comfyapi( + cls, audio, container_format="mp3", codec_name="libmp3lame", mime_type="audio/mpeg", filename="output.mp3" + ) logging.info("Uploaded audio to Comfy API. URL: %s", audio_url) else: audio_url = None From 10e90a5757906ecdb71b84d41173813d7f62c140 Mon Sep 17 00:00:00 2001 From: Christian Byrne Date: Thu, 20 Nov 2025 18:20:52 -0800 Subject: [PATCH 136/647] bump comfyui-workflow-templates for nano banana 2 (#10818) * bump templates * bump templates --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 8c1946f3d..624aa7362 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,5 +1,5 @@ comfyui-frontend-package==1.28.9 -comfyui-workflow-templates==0.3.1 +comfyui-workflow-templates==0.6.0 comfyui-embedded-docs==0.3.1 torch torchsde From 943b3b615d40542ea19bc8ff8ad2950c0a094605 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Thu, 20 Nov 2025 19:44:43 -0800 Subject: [PATCH 137/647] HunyuanVideo 1.5 (#10819) * init * update * Update model.py * Update model.py * remove print * Fix text encoding * Prevent empty negative prompt Really doesn't work otherwise * fp16 works * I2V * Update model_base.py * Update nodes_hunyuan.py * Better latent rgb factors * Use the correct sigclip output... * Support HunyuanVideo1.5 SR model * whitespaces... * Proper latent channel count * SR model fixes This also still needs timesteps scheduling based on the noise scale, can be used with two samplers too already * vae_refiner: roll the convolution through temporal Work in progress. Roll the convolution through time using 2-latent-frame chunks and a FIFO queue for the convolution seams. * Support HunyuanVideo15 latent resampler * fix * Some cleanup Co-Authored-By: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> * Proper hyvid15 I2V channels Co-Authored-By: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> * Fix TokenRefiner for fp16 Otherwise x.sum has infs, just in case only casting if input is fp16, I don't know if necessary. * Bugfix for the HunyuanVideo15 SR model * vae_refiner: roll the convolution through temporal II Roll the convolution through time using 2-latent-frame chunks and a FIFO queue for the convolution seams. Added support for encoder, lowered to 1 latent frame to save more VRAM, made work for Hunyuan Image 3.0 (as code shared). Fixed names, cleaned up code. * Allow any number of input frames in VAE. * Better VAE encode mem estimation. * Lowvram fix. * Fix hunyuan image 2.1 refiner. * Fix mistake. * Name changes. * Rename. * Whitespace. * Fix. * Fix. --------- Co-authored-by: kijai <40791699+kijai@users.noreply.github.com> Co-authored-by: Rattus --- comfy/latent_formats.py | 60 ++++ comfy/ldm/hunyuan_video/model.py | 54 +++- comfy/ldm/hunyuan_video/upsampler.py | 120 ++++++++ comfy/ldm/hunyuan_video/vae_refiner.py | 284 +++++++++++------- comfy/model_base.py | 91 ++++++ comfy/model_detection.py | 10 + comfy/sd.py | 12 +- comfy/supported_models.py | 50 ++- comfy/text_encoders/hunyuan_video.py | 9 + comfy/text_encoders/qwen_image.py | 4 +- comfy_api/latest/_io.py | 4 + comfy_extras/nodes_hunyuan.py | 201 ++++++++++++- folder_paths.py | 2 + .../put_latent_upscale_models_here | 0 nodes.py | 2 +- 15 files changed, 777 insertions(+), 126 deletions(-) create mode 100644 comfy/ldm/hunyuan_video/upsampler.py create mode 100644 models/latent_upscale_models/put_latent_upscale_models_here diff --git a/comfy/latent_formats.py b/comfy/latent_formats.py index 77e642a94..204fc048d 100644 --- a/comfy/latent_formats.py +++ b/comfy/latent_formats.py @@ -611,6 +611,66 @@ class HunyuanImage21Refiner(LatentFormat): latent_dimensions = 3 scale_factor = 1.03682 + def process_in(self, latent): + out = latent * self.scale_factor + out = torch.cat((out[:, :, :1], out), dim=2) + out = out.permute(0, 2, 1, 3, 4) + b, f_times_2, c, h, w = out.shape + out = out.reshape(b, f_times_2 // 2, 2 * c, h, w) + out = out.permute(0, 2, 1, 3, 4).contiguous() + return out + + def process_out(self, latent): + z = latent / self.scale_factor + z = z.permute(0, 2, 1, 3, 4) + b, f, c, h, w = z.shape + z = z.reshape(b, f, 2, c // 2, h, w) + z = z.permute(0, 1, 2, 3, 4, 5).reshape(b, f * 2, c // 2, h, w) + z = z.permute(0, 2, 1, 3, 4) + z = z[:, :, 1:] + return z + +class HunyuanVideo15(LatentFormat): + latent_rgb_factors = [ + [ 0.0568, -0.0521, -0.0131], + [ 0.0014, 0.0735, 0.0326], + [ 0.0186, 0.0531, -0.0138], + [-0.0031, 0.0051, 0.0288], + [ 0.0110, 0.0556, 0.0432], + [-0.0041, -0.0023, -0.0485], + [ 0.0530, 0.0413, 0.0253], + [ 0.0283, 0.0251, 0.0339], + [ 0.0277, -0.0372, -0.0093], + [ 0.0393, 0.0944, 0.1131], + [ 0.0020, 0.0251, 0.0037], + [-0.0017, 0.0012, 0.0234], + [ 0.0468, 0.0436, 0.0203], + [ 0.0354, 0.0439, -0.0233], + [ 0.0090, 0.0123, 0.0346], + [ 0.0382, 0.0029, 0.0217], + [ 0.0261, -0.0300, 0.0030], + [-0.0088, -0.0220, -0.0283], + [-0.0272, -0.0121, -0.0363], + [-0.0664, -0.0622, 0.0144], + [ 0.0414, 0.0479, 0.0529], + [ 0.0355, 0.0612, -0.0247], + [ 0.0147, 0.0264, 0.0174], + [ 0.0438, 0.0038, 0.0542], + [ 0.0431, -0.0573, -0.0033], + [-0.0162, -0.0211, -0.0406], + [-0.0487, -0.0295, -0.0393], + [ 0.0005, -0.0109, 0.0253], + [ 0.0296, 0.0591, 0.0353], + [ 0.0119, 0.0181, -0.0306], + [-0.0085, -0.0362, 0.0229], + [ 0.0005, -0.0106, 0.0242] + ] + + latent_rgb_factors_bias = [ 0.0456, -0.0202, -0.0644] + latent_channels = 32 + latent_dimensions = 3 + scale_factor = 1.03682 + class Hunyuan3Dv2(LatentFormat): latent_channels = 64 latent_dimensions = 1 diff --git a/comfy/ldm/hunyuan_video/model.py b/comfy/ldm/hunyuan_video/model.py index 5132e6c07..f75c6e0e1 100644 --- a/comfy/ldm/hunyuan_video/model.py +++ b/comfy/ldm/hunyuan_video/model.py @@ -6,7 +6,6 @@ import comfy.ldm.flux.layers import comfy.ldm.modules.diffusionmodules.mmdit from comfy.ldm.modules.attention import optimized_attention - from dataclasses import dataclass from einops import repeat @@ -42,6 +41,8 @@ class HunyuanVideoParams: guidance_embed: bool byt5: bool meanflow: bool + use_cond_type_embedding: bool + vision_in_dim: int class SelfAttentionRef(nn.Module): @@ -157,7 +158,10 @@ class TokenRefiner(nn.Module): t = self.t_embedder(timestep_embedding(timesteps, 256, time_factor=1.0).to(x.dtype)) # m = mask.float().unsqueeze(-1) # c = (x.float() * m).sum(dim=1) / m.sum(dim=1) #TODO: the following works when the x.shape is the same length as the tokens but might break otherwise - c = x.sum(dim=1) / x.shape[1] + if x.dtype == torch.float16: + c = x.float().sum(dim=1) / x.shape[1] + else: + c = x.sum(dim=1) / x.shape[1] c = t + self.c_embedder(c.to(x.dtype)) x = self.input_embedder(x) @@ -196,11 +200,15 @@ class HunyuanVideo(nn.Module): def __init__(self, image_model=None, final_layer=True, dtype=None, device=None, operations=None, **kwargs): super().__init__() self.dtype = dtype + operation_settings = {"operations": operations, "device": device, "dtype": dtype} + params = HunyuanVideoParams(**kwargs) self.params = params self.patch_size = params.patch_size self.in_channels = params.in_channels self.out_channels = params.out_channels + self.use_cond_type_embedding = params.use_cond_type_embedding + self.vision_in_dim = params.vision_in_dim if params.hidden_size % params.num_heads != 0: raise ValueError( f"Hidden size {params.hidden_size} must be divisible by num_heads {params.num_heads}" @@ -266,6 +274,18 @@ class HunyuanVideo(nn.Module): if final_layer: self.final_layer = LastLayer(self.hidden_size, self.patch_size[-1], self.out_channels, dtype=dtype, device=device, operations=operations) + # HunyuanVideo 1.5 specific modules + if self.vision_in_dim is not None: + from comfy.ldm.wan.model import MLPProj + self.vision_in = MLPProj(in_dim=self.vision_in_dim, out_dim=self.hidden_size, operation_settings=operation_settings) + else: + self.vision_in = None + if self.use_cond_type_embedding: + # 0: text_encoder feature 1: byt5 feature 2: vision_encoder feature + self.cond_type_embedding = nn.Embedding(3, self.hidden_size) + else: + self.cond_type_embedding = None + def forward_orig( self, img: Tensor, @@ -276,6 +296,7 @@ class HunyuanVideo(nn.Module): timesteps: Tensor, y: Tensor = None, txt_byt5=None, + clip_fea=None, guidance: Tensor = None, guiding_frame_index=None, ref_latent=None, @@ -331,12 +352,31 @@ class HunyuanVideo(nn.Module): txt = self.txt_in(txt, timesteps, txt_mask, transformer_options=transformer_options) + if self.cond_type_embedding is not None: + self.cond_type_embedding.to(txt.device) + cond_emb = self.cond_type_embedding(torch.zeros_like(txt[:, :, 0], device=txt.device, dtype=torch.long)) + txt = txt + cond_emb.to(txt.dtype) + if self.byt5_in is not None and txt_byt5 is not None: txt_byt5 = self.byt5_in(txt_byt5) + if self.cond_type_embedding is not None: + cond_emb = self.cond_type_embedding(torch.ones_like(txt_byt5[:, :, 0], device=txt_byt5.device, dtype=torch.long)) + txt_byt5 = txt_byt5 + cond_emb.to(txt_byt5.dtype) + txt = torch.cat((txt_byt5, txt), dim=1) # byt5 first for HunyuanVideo1.5 + else: + txt = torch.cat((txt, txt_byt5), dim=1) txt_byt5_ids = torch.zeros((txt_ids.shape[0], txt_byt5.shape[1], txt_ids.shape[-1]), device=txt_ids.device, dtype=txt_ids.dtype) - txt = torch.cat((txt, txt_byt5), dim=1) txt_ids = torch.cat((txt_ids, txt_byt5_ids), dim=1) + if clip_fea is not None: + txt_vision_states = self.vision_in(clip_fea) + if self.cond_type_embedding is not None: + cond_emb = self.cond_type_embedding(2 * torch.ones_like(txt_vision_states[:, :, 0], dtype=torch.long, device=txt_vision_states.device)) + txt_vision_states = txt_vision_states + cond_emb + txt = torch.cat((txt_vision_states.to(txt.dtype), txt), dim=1) + extra_txt_ids = torch.zeros((txt_ids.shape[0], txt_vision_states.shape[1], txt_ids.shape[-1]), device=txt_ids.device, dtype=txt_ids.dtype) + txt_ids = torch.cat((txt_ids, extra_txt_ids), dim=1) + ids = torch.cat((img_ids, txt_ids), dim=1) pe = self.pe_embedder(ids) @@ -430,14 +470,14 @@ class HunyuanVideo(nn.Module): img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(0, w_len - 1, steps=w_len, device=x.device, dtype=x.dtype).unsqueeze(0) return repeat(img_ids, "h w c -> b (h w) c", b=bs) - def forward(self, x, timestep, context, y=None, txt_byt5=None, guidance=None, attention_mask=None, guiding_frame_index=None, ref_latent=None, disable_time_r=False, control=None, transformer_options={}, **kwargs): + def forward(self, x, timestep, context, y=None, txt_byt5=None, clip_fea=None, guidance=None, attention_mask=None, guiding_frame_index=None, ref_latent=None, disable_time_r=False, control=None, transformer_options={}, **kwargs): return comfy.patcher_extension.WrapperExecutor.new_class_executor( self._forward, self, comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options) - ).execute(x, timestep, context, y, txt_byt5, guidance, attention_mask, guiding_frame_index, ref_latent, disable_time_r, control, transformer_options, **kwargs) + ).execute(x, timestep, context, y, txt_byt5, clip_fea, guidance, attention_mask, guiding_frame_index, ref_latent, disable_time_r, control, transformer_options, **kwargs) - def _forward(self, x, timestep, context, y=None, txt_byt5=None, guidance=None, attention_mask=None, guiding_frame_index=None, ref_latent=None, disable_time_r=False, control=None, transformer_options={}, **kwargs): + def _forward(self, x, timestep, context, y=None, txt_byt5=None, clip_fea=None, guidance=None, attention_mask=None, guiding_frame_index=None, ref_latent=None, disable_time_r=False, control=None, transformer_options={}, **kwargs): bs = x.shape[0] if len(self.patch_size) == 3: img_ids = self.img_ids(x) @@ -445,5 +485,5 @@ class HunyuanVideo(nn.Module): else: img_ids = self.img_ids_2d(x) txt_ids = torch.zeros((bs, context.shape[1], 2), device=x.device, dtype=x.dtype) - out = self.forward_orig(x, img_ids, context, txt_ids, attention_mask, timestep, y, txt_byt5, guidance, guiding_frame_index, ref_latent, disable_time_r=disable_time_r, control=control, transformer_options=transformer_options) + out = self.forward_orig(x, img_ids, context, txt_ids, attention_mask, timestep, y, txt_byt5, clip_fea, guidance, guiding_frame_index, ref_latent, disable_time_r=disable_time_r, control=control, transformer_options=transformer_options) return out diff --git a/comfy/ldm/hunyuan_video/upsampler.py b/comfy/ldm/hunyuan_video/upsampler.py new file mode 100644 index 000000000..9f5e91a59 --- /dev/null +++ b/comfy/ldm/hunyuan_video/upsampler.py @@ -0,0 +1,120 @@ +import torch +import torch.nn as nn +import torch.nn.functional as F +from comfy.ldm.hunyuan_video.vae_refiner import RMS_norm, ResnetBlock, VideoConv3d +import model_management, model_patcher + +class SRResidualCausalBlock3D(nn.Module): + def __init__(self, channels: int): + super().__init__() + self.block = nn.Sequential( + VideoConv3d(channels, channels, kernel_size=3), + nn.SiLU(inplace=True), + VideoConv3d(channels, channels, kernel_size=3), + nn.SiLU(inplace=True), + VideoConv3d(channels, channels, kernel_size=3), + ) + + def forward(self, x: torch.Tensor) -> torch.Tensor: + return x + self.block(x) + +class SRModel3DV2(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + hidden_channels: int = 64, + num_blocks: int = 6, + global_residual: bool = False, + ): + super().__init__() + self.in_conv = VideoConv3d(in_channels, hidden_channels, kernel_size=3) + self.blocks = nn.ModuleList([SRResidualCausalBlock3D(hidden_channels) for _ in range(num_blocks)]) + self.out_conv = VideoConv3d(hidden_channels, out_channels, kernel_size=3) + self.global_residual = bool(global_residual) + + def forward(self, x: torch.Tensor) -> torch.Tensor: + residual = x + y = self.in_conv(x) + for blk in self.blocks: + y = blk(y) + y = self.out_conv(y) + if self.global_residual and (y.shape == residual.shape): + y = y + residual + return y + + +class Upsampler(nn.Module): + def __init__( + self, + z_channels: int, + out_channels: int, + block_out_channels: tuple[int, ...], + num_res_blocks: int = 2, + ): + super().__init__() + self.num_res_blocks = num_res_blocks + self.block_out_channels = block_out_channels + self.z_channels = z_channels + + ch = block_out_channels[0] + self.conv_in = VideoConv3d(z_channels, ch, kernel_size=3) + + self.up = nn.ModuleList() + + for i, tgt in enumerate(block_out_channels): + stage = nn.Module() + stage.block = nn.ModuleList([ResnetBlock(in_channels=ch if j == 0 else tgt, + out_channels=tgt, + temb_channels=0, + conv_shortcut=False, + conv_op=VideoConv3d, norm_op=RMS_norm) + for j in range(num_res_blocks + 1)]) + ch = tgt + self.up.append(stage) + + self.norm_out = RMS_norm(ch) + self.conv_out = VideoConv3d(ch, out_channels, kernel_size=3) + + def forward(self, z): + """ + Args: + z: (B, C, T, H, W) + target_shape: (H, W) + """ + # z to block_in + repeats = self.block_out_channels[0] // (self.z_channels) + x = self.conv_in(z) + z.repeat_interleave(repeats=repeats, dim=1) + + # upsampling + for stage in self.up: + for blk in stage.block: + x = blk(x) + + out = self.conv_out(F.silu(self.norm_out(x))) + return out + +UPSAMPLERS = { + "720p": SRModel3DV2, + "1080p": Upsampler, +} + +class HunyuanVideo15SRModel(): + def __init__(self, model_type, config): + self.load_device = model_management.vae_device() + offload_device = model_management.vae_offload_device() + self.dtype = model_management.vae_dtype(self.load_device) + self.model_class = UPSAMPLERS.get(model_type) + self.model = self.model_class(**config).eval() + + self.patcher = model_patcher.ModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device) + + def load_sd(self, sd): + return self.model.load_state_dict(sd, strict=True) + + def get_sd(self): + return self.model.state_dict() + + def resample_latent(self, latent): + model_management.load_model_gpu(self.patcher) + return self.model(latent.to(self.load_device)) diff --git a/comfy/ldm/hunyuan_video/vae_refiner.py b/comfy/ldm/hunyuan_video/vae_refiner.py index c2a0b507d..9f750dcc4 100644 --- a/comfy/ldm/hunyuan_video/vae_refiner.py +++ b/comfy/ldm/hunyuan_video/vae_refiner.py @@ -4,8 +4,40 @@ import torch.nn.functional as F from comfy.ldm.modules.diffusionmodules.model import ResnetBlock, AttnBlock, VideoConv3d, Normalize import comfy.ops import comfy.ldm.models.autoencoder +import comfy.model_management ops = comfy.ops.disable_weight_init +class NoPadConv3d(nn.Module): + def __init__(self, n_channels, out_channels, kernel_size, stride=1, dilation=1, padding=0, **kwargs): + super().__init__() + self.conv = ops.Conv3d(n_channels, out_channels, kernel_size, stride=stride, dilation=dilation, **kwargs) + + def forward(self, x): + return self.conv(x) + + +def conv_carry_causal_3d(xl, op, conv_carry_in=None, conv_carry_out=None): + + x = xl[0] + xl.clear() + + if conv_carry_out is not None: + to_push = x[:, :, -2:, :, :].clone() + conv_carry_out.append(to_push) + + if isinstance(op, NoPadConv3d): + if conv_carry_in is None: + x = torch.nn.functional.pad(x, (1, 1, 1, 1, 2, 0), mode = 'replicate') + else: + carry_len = conv_carry_in[0].shape[2] + x = torch.cat([conv_carry_in.pop(0), x], dim=2) + x = torch.nn.functional.pad(x, (1, 1, 1, 1, 2 - carry_len, 0), mode = 'replicate') + + out = op(x) + + return out + + class RMS_norm(nn.Module): def __init__(self, dim): super().__init__() @@ -14,7 +46,7 @@ class RMS_norm(nn.Module): self.gamma = nn.Parameter(torch.empty(shape)) def forward(self, x): - return F.normalize(x, dim=1) * self.scale * self.gamma + return F.normalize(x, dim=1) * self.scale * comfy.model_management.cast_to(self.gamma, dtype=x.dtype, device=x.device) class DnSmpl(nn.Module): def __init__(self, ic, oc, tds=True, refiner_vae=True, op=VideoConv3d): @@ -27,11 +59,12 @@ class DnSmpl(nn.Module): self.tds = tds self.gs = fct * ic // oc - def forward(self, x): + def forward(self, x, conv_carry_in=None, conv_carry_out=None): r1 = 2 if self.tds else 1 - h = self.conv(x) + h = conv_carry_causal_3d([x], self.conv, conv_carry_in, conv_carry_out) + + if self.tds and self.refiner_vae and conv_carry_in is None: - if self.tds and self.refiner_vae: hf = h[:, :, :1, :, :] b, c, f, ht, wd = hf.shape hf = hf.reshape(b, c, f, ht // 2, 2, wd // 2, 2) @@ -39,14 +72,7 @@ class DnSmpl(nn.Module): hf = hf.reshape(b, 2 * 2 * c, f, ht // 2, wd // 2) hf = torch.cat([hf, hf], dim=1) - hn = h[:, :, 1:, :, :] - b, c, frms, ht, wd = hn.shape - nf = frms // r1 - hn = hn.reshape(b, c, nf, r1, ht // 2, 2, wd // 2, 2) - hn = hn.permute(0, 3, 5, 7, 1, 2, 4, 6) - hn = hn.reshape(b, r1 * 2 * 2 * c, nf, ht // 2, wd // 2) - - h = torch.cat([hf, hn], dim=2) + h = h[:, :, 1:, :, :] xf = x[:, :, :1, :, :] b, ci, f, ht, wd = xf.shape @@ -54,34 +80,32 @@ class DnSmpl(nn.Module): xf = xf.permute(0, 4, 6, 1, 2, 3, 5) xf = xf.reshape(b, 2 * 2 * ci, f, ht // 2, wd // 2) B, C, T, H, W = xf.shape - xf = xf.view(B, h.shape[1], self.gs // 2, T, H, W).mean(dim=2) + xf = xf.view(B, hf.shape[1], self.gs // 2, T, H, W).mean(dim=2) - xn = x[:, :, 1:, :, :] - b, ci, frms, ht, wd = xn.shape - nf = frms // r1 - xn = xn.reshape(b, ci, nf, r1, ht // 2, 2, wd // 2, 2) - xn = xn.permute(0, 3, 5, 7, 1, 2, 4, 6) - xn = xn.reshape(b, r1 * 2 * 2 * ci, nf, ht // 2, wd // 2) - B, C, T, H, W = xn.shape - xn = xn.view(B, h.shape[1], self.gs, T, H, W).mean(dim=2) - sc = torch.cat([xf, xn], dim=2) - else: - b, c, frms, ht, wd = h.shape + x = x[:, :, 1:, :, :] - nf = frms // r1 - h = h.reshape(b, c, nf, r1, ht // 2, 2, wd // 2, 2) - h = h.permute(0, 3, 5, 7, 1, 2, 4, 6) - h = h.reshape(b, r1 * 2 * 2 * c, nf, ht // 2, wd // 2) + if h.shape[2] == 0: + return hf + xf - b, ci, frms, ht, wd = x.shape - nf = frms // r1 - sc = x.reshape(b, ci, nf, r1, ht // 2, 2, wd // 2, 2) - sc = sc.permute(0, 3, 5, 7, 1, 2, 4, 6) - sc = sc.reshape(b, r1 * 2 * 2 * ci, nf, ht // 2, wd // 2) - B, C, T, H, W = sc.shape - sc = sc.view(B, h.shape[1], self.gs, T, H, W).mean(dim=2) + b, c, frms, ht, wd = h.shape + nf = frms // r1 + h = h.reshape(b, c, nf, r1, ht // 2, 2, wd // 2, 2) + h = h.permute(0, 3, 5, 7, 1, 2, 4, 6) + h = h.reshape(b, r1 * 2 * 2 * c, nf, ht // 2, wd // 2) - return h + sc + b, ci, frms, ht, wd = x.shape + nf = frms // r1 + x = x.reshape(b, ci, nf, r1, ht // 2, 2, wd // 2, 2) + x = x.permute(0, 3, 5, 7, 1, 2, 4, 6) + x = x.reshape(b, r1 * 2 * 2 * ci, nf, ht // 2, wd // 2) + B, C, T, H, W = x.shape + x = x.view(B, h.shape[1], self.gs, T, H, W).mean(dim=2) + + if self.tds and self.refiner_vae and conv_carry_in is None: + h = torch.cat([hf, h], dim=2) + x = torch.cat([xf, x], dim=2) + + return h + x class UpSmpl(nn.Module): @@ -94,11 +118,11 @@ class UpSmpl(nn.Module): self.tus = tus self.rp = fct * oc // ic - def forward(self, x): + def forward(self, x, conv_carry_in=None, conv_carry_out=None): r1 = 2 if self.tus else 1 - h = self.conv(x) + h = conv_carry_causal_3d([x], self.conv, conv_carry_in, conv_carry_out) - if self.tus and self.refiner_vae: + if self.tus and self.refiner_vae and conv_carry_in is None: hf = h[:, :, :1, :, :] b, c, f, ht, wd = hf.shape nc = c // (2 * 2) @@ -107,14 +131,7 @@ class UpSmpl(nn.Module): hf = hf.reshape(b, nc, f, ht * 2, wd * 2) hf = hf[:, : hf.shape[1] // 2] - hn = h[:, :, 1:, :, :] - b, c, frms, ht, wd = hn.shape - nc = c // (r1 * 2 * 2) - hn = hn.reshape(b, r1, 2, 2, nc, frms, ht, wd) - hn = hn.permute(0, 4, 5, 1, 6, 2, 7, 3) - hn = hn.reshape(b, nc, frms * r1, ht * 2, wd * 2) - - h = torch.cat([hf, hn], dim=2) + h = h[:, :, 1:, :, :] xf = x[:, :, :1, :, :] b, ci, f, ht, wd = xf.shape @@ -125,29 +142,43 @@ class UpSmpl(nn.Module): xf = xf.permute(0, 3, 4, 5, 1, 6, 2) xf = xf.reshape(b, nc, f, ht * 2, wd * 2) - xn = x[:, :, 1:, :, :] - xn = xn.repeat_interleave(repeats=self.rp, dim=1) - b, c, frms, ht, wd = xn.shape - nc = c // (r1 * 2 * 2) - xn = xn.reshape(b, r1, 2, 2, nc, frms, ht, wd) - xn = xn.permute(0, 4, 5, 1, 6, 2, 7, 3) - xn = xn.reshape(b, nc, frms * r1, ht * 2, wd * 2) - sc = torch.cat([xf, xn], dim=2) - else: - b, c, frms, ht, wd = h.shape - nc = c // (r1 * 2 * 2) - h = h.reshape(b, r1, 2, 2, nc, frms, ht, wd) - h = h.permute(0, 4, 5, 1, 6, 2, 7, 3) - h = h.reshape(b, nc, frms * r1, ht * 2, wd * 2) + x = x[:, :, 1:, :, :] - sc = x.repeat_interleave(repeats=self.rp, dim=1) - b, c, frms, ht, wd = sc.shape - nc = c // (r1 * 2 * 2) - sc = sc.reshape(b, r1, 2, 2, nc, frms, ht, wd) - sc = sc.permute(0, 4, 5, 1, 6, 2, 7, 3) - sc = sc.reshape(b, nc, frms * r1, ht * 2, wd * 2) + b, c, frms, ht, wd = h.shape + nc = c // (r1 * 2 * 2) + h = h.reshape(b, r1, 2, 2, nc, frms, ht, wd) + h = h.permute(0, 4, 5, 1, 6, 2, 7, 3) + h = h.reshape(b, nc, frms * r1, ht * 2, wd * 2) - return h + sc + x = x.repeat_interleave(repeats=self.rp, dim=1) + b, c, frms, ht, wd = x.shape + nc = c // (r1 * 2 * 2) + x = x.reshape(b, r1, 2, 2, nc, frms, ht, wd) + x = x.permute(0, 4, 5, 1, 6, 2, 7, 3) + x = x.reshape(b, nc, frms * r1, ht * 2, wd * 2) + + if self.tus and self.refiner_vae and conv_carry_in is None: + h = torch.cat([hf, h], dim=2) + x = torch.cat([xf, x], dim=2) + + return h + x + +class HunyuanRefinerResnetBlock(ResnetBlock): + def __init__(self, in_channels, out_channels, conv_op=NoPadConv3d, norm_op=RMS_norm): + super().__init__(in_channels=in_channels, out_channels=out_channels, temb_channels=0, conv_op=conv_op, norm_op=norm_op) + + def forward(self, x, conv_carry_in=None, conv_carry_out=None): + h = x + h = [ self.swish(self.norm1(x)) ] + h = conv_carry_causal_3d(h, self.conv1, conv_carry_in=conv_carry_in, conv_carry_out=conv_carry_out) + + h = [ self.dropout(self.swish(self.norm2(h))) ] + h = conv_carry_causal_3d(h, self.conv2, conv_carry_in=conv_carry_in, conv_carry_out=conv_carry_out) + + if self.in_channels != self.out_channels: + x = self.nin_shortcut(x) + + return x+h class Encoder(nn.Module): def __init__(self, in_channels, z_channels, block_out_channels, num_res_blocks, @@ -160,7 +191,7 @@ class Encoder(nn.Module): self.refiner_vae = refiner_vae if self.refiner_vae: - conv_op = VideoConv3d + conv_op = NoPadConv3d norm_op = RMS_norm else: conv_op = ops.Conv3d @@ -175,10 +206,9 @@ class Encoder(nn.Module): for i, tgt in enumerate(block_out_channels): stage = nn.Module() - stage.block = nn.ModuleList([ResnetBlock(in_channels=ch if j == 0 else tgt, - out_channels=tgt, - temb_channels=0, - conv_op=conv_op, norm_op=norm_op) + stage.block = nn.ModuleList([HunyuanRefinerResnetBlock(in_channels=ch if j == 0 else tgt, + out_channels=tgt, + conv_op=conv_op, norm_op=norm_op) for j in range(num_res_blocks)]) ch = tgt if i < depth: @@ -188,9 +218,9 @@ class Encoder(nn.Module): self.down.append(stage) self.mid = nn.Module() - self.mid.block_1 = ResnetBlock(in_channels=ch, out_channels=ch, temb_channels=0, conv_op=conv_op, norm_op=norm_op) + self.mid.block_1 = HunyuanRefinerResnetBlock(in_channels=ch, out_channels=ch, conv_op=conv_op, norm_op=norm_op) self.mid.attn_1 = AttnBlock(ch, conv_op=ops.Conv3d, norm_op=norm_op) - self.mid.block_2 = ResnetBlock(in_channels=ch, out_channels=ch, temb_channels=0, conv_op=conv_op, norm_op=norm_op) + self.mid.block_2 = HunyuanRefinerResnetBlock(in_channels=ch, out_channels=ch, conv_op=conv_op, norm_op=norm_op) self.norm_out = norm_op(ch) self.conv_out = conv_op(ch, z_channels << 1, 3, 1, 1) @@ -201,31 +231,50 @@ class Encoder(nn.Module): if not self.refiner_vae and x.shape[2] == 1: x = x.expand(-1, -1, self.ffactor_temporal, -1, -1) - x = self.conv_in(x) + if self.refiner_vae: + xl = [x[:, :, :1, :, :]] + if x.shape[2] > self.ffactor_temporal: + xl += torch.split(x[:, :, 1: 1 + ((x.shape[2] - 1) // self.ffactor_temporal) * self.ffactor_temporal, :, :], self.ffactor_temporal * 2, dim=2) + x = xl + else: + x = [x] + out = [] - for stage in self.down: - for blk in stage.block: - x = blk(x) - if hasattr(stage, 'downsample'): - x = stage.downsample(x) + conv_carry_in = None - x = self.mid.block_2(self.mid.attn_1(self.mid.block_1(x))) + for i, x1 in enumerate(x): + conv_carry_out = [] + if i == len(x) - 1: + conv_carry_out = None + x1 = [ x1 ] + x1 = conv_carry_causal_3d(x1, self.conv_in, conv_carry_in, conv_carry_out) + + for stage in self.down: + for blk in stage.block: + x1 = blk(x1, conv_carry_in, conv_carry_out) + if hasattr(stage, 'downsample'): + x1 = stage.downsample(x1, conv_carry_in, conv_carry_out) + + out.append(x1) + conv_carry_in = conv_carry_out + + if len(out) > 1: + out = torch.cat(out, dim=2) + else: + out = out[0] + + x = self.mid.block_2(self.mid.attn_1(self.mid.block_1(out))) + del out b, c, t, h, w = x.shape grp = c // (self.z_channels << 1) skip = x.view(b, c // grp, grp, t, h, w).mean(2) - out = self.conv_out(F.silu(self.norm_out(x))) + skip + out = conv_carry_causal_3d([F.silu(self.norm_out(x))], self.conv_out) + skip if self.refiner_vae: out = self.regul(out)[0] - out = torch.cat((out[:, :, :1], out), dim=2) - out = out.permute(0, 2, 1, 3, 4) - b, f_times_2, c, h, w = out.shape - out = out.reshape(b, f_times_2 // 2, 2 * c, h, w) - out = out.permute(0, 2, 1, 3, 4).contiguous() - return out class Decoder(nn.Module): @@ -239,7 +288,7 @@ class Decoder(nn.Module): self.refiner_vae = refiner_vae if self.refiner_vae: - conv_op = VideoConv3d + conv_op = NoPadConv3d norm_op = RMS_norm else: conv_op = ops.Conv3d @@ -249,9 +298,9 @@ class Decoder(nn.Module): self.conv_in = conv_op(z_channels, ch, kernel_size=3, stride=1, padding=1) self.mid = nn.Module() - self.mid.block_1 = ResnetBlock(in_channels=ch, out_channels=ch, temb_channels=0, conv_op=conv_op, norm_op=norm_op) + self.mid.block_1 = HunyuanRefinerResnetBlock(in_channels=ch, out_channels=ch, conv_op=conv_op, norm_op=norm_op) self.mid.attn_1 = AttnBlock(ch, conv_op=ops.Conv3d, norm_op=norm_op) - self.mid.block_2 = ResnetBlock(in_channels=ch, out_channels=ch, temb_channels=0, conv_op=conv_op, norm_op=norm_op) + self.mid.block_2 = HunyuanRefinerResnetBlock(in_channels=ch, out_channels=ch, conv_op=conv_op, norm_op=norm_op) self.up = nn.ModuleList() depth = (ffactor_spatial >> 1).bit_length() @@ -259,10 +308,9 @@ class Decoder(nn.Module): for i, tgt in enumerate(block_out_channels): stage = nn.Module() - stage.block = nn.ModuleList([ResnetBlock(in_channels=ch if j == 0 else tgt, - out_channels=tgt, - temb_channels=0, - conv_op=conv_op, norm_op=norm_op) + stage.block = nn.ModuleList([HunyuanRefinerResnetBlock(in_channels=ch if j == 0 else tgt, + out_channels=tgt, + conv_op=conv_op, norm_op=norm_op) for j in range(num_res_blocks + 1)]) ch = tgt if i < depth: @@ -275,27 +323,41 @@ class Decoder(nn.Module): self.conv_out = conv_op(ch, out_channels, 3, stride=1, padding=1) def forward(self, z): - if self.refiner_vae: - z = z.permute(0, 2, 1, 3, 4) - b, f, c, h, w = z.shape - z = z.reshape(b, f, 2, c // 2, h, w) - z = z.permute(0, 1, 2, 3, 4, 5).reshape(b, f * 2, c // 2, h, w) - z = z.permute(0, 2, 1, 3, 4) - z = z[:, :, 1:] - - x = self.conv_in(z) + z.repeat_interleave(self.block_out_channels[0] // self.z_channels, 1) + x = conv_carry_causal_3d([z], self.conv_in) + z.repeat_interleave(self.block_out_channels[0] // self.z_channels, 1) x = self.mid.block_2(self.mid.attn_1(self.mid.block_1(x))) - for stage in self.up: - for blk in stage.block: - x = blk(x) - if hasattr(stage, 'upsample'): - x = stage.upsample(x) + if self.refiner_vae: + x = torch.split(x, 2, dim=2) + else: + x = [ x ] + out = [] - out = self.conv_out(F.silu(self.norm_out(x))) + conv_carry_in = None + + for i, x1 in enumerate(x): + conv_carry_out = [] + if i == len(x) - 1: + conv_carry_out = None + for stage in self.up: + for blk in stage.block: + x1 = blk(x1, conv_carry_in, conv_carry_out) + if hasattr(stage, 'upsample'): + x1 = stage.upsample(x1, conv_carry_in, conv_carry_out) + + x1 = [ F.silu(self.norm_out(x1)) ] + x1 = conv_carry_causal_3d(x1, self.conv_out, conv_carry_in, conv_carry_out) + out.append(x1) + conv_carry_in = conv_carry_out + del x + + if len(out) > 1: + out = torch.cat(out, dim=2) + else: + out = out[0] if not self.refiner_vae: if z.shape[-3] == 1: out = out[:, :, -1:] return out + diff --git a/comfy/model_base.py b/comfy/model_base.py index 7c788d085..e14b552c5 100644 --- a/comfy/model_base.py +++ b/comfy/model_base.py @@ -1536,3 +1536,94 @@ class HunyuanImage21Refiner(HunyuanImage21): out = super().extra_conds(**kwargs) out['disable_time_r'] = comfy.conds.CONDConstant(True) return out + +class HunyuanVideo15(HunyuanVideo): + def __init__(self, model_config, model_type=ModelType.FLOW, device=None): + super().__init__(model_config, model_type, device=device) + + def concat_cond(self, **kwargs): + noise = kwargs.get("noise", None) + extra_channels = self.diffusion_model.img_in.proj.weight.shape[1] - noise.shape[1] - 1 #noise 32 img cond 32 + mask 1 + if extra_channels == 0: + return None + + image = kwargs.get("concat_latent_image", None) + device = kwargs["device"] + + if image is None: + shape_image = list(noise.shape) + shape_image[1] = extra_channels + image = torch.zeros(shape_image, dtype=noise.dtype, layout=noise.layout, device=noise.device) + else: + latent_dim = self.latent_format.latent_channels + image = utils.common_upscale(image.to(device), noise.shape[-1], noise.shape[-2], "bilinear", "center") + for i in range(0, image.shape[1], latent_dim): + image[:, i: i + latent_dim] = self.process_latent_in(image[:, i: i + latent_dim]) + image = utils.resize_to_batch_size(image, noise.shape[0]) + + mask = kwargs.get("concat_mask", kwargs.get("denoise_mask", None)) + if mask is None: + mask = torch.zeros_like(noise)[:, :1] + else: + mask = 1.0 - mask + mask = utils.common_upscale(mask.to(device), noise.shape[-1], noise.shape[-2], "bilinear", "center") + if mask.shape[-3] < noise.shape[-3]: + mask = torch.nn.functional.pad(mask, (0, 0, 0, 0, 0, noise.shape[-3] - mask.shape[-3]), mode='constant', value=0) + mask = utils.resize_to_batch_size(mask, noise.shape[0]) + + return torch.cat((image, mask), dim=1) + + def extra_conds(self, **kwargs): + out = super().extra_conds(**kwargs) + attention_mask = kwargs.get("attention_mask", None) + if attention_mask is not None: + if torch.numel(attention_mask) != attention_mask.sum(): + out['attention_mask'] = comfy.conds.CONDRegular(attention_mask) + cross_attn = kwargs.get("cross_attn", None) + if cross_attn is not None: + out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn) + + conditioning_byt5small = kwargs.get("conditioning_byt5small", None) + if conditioning_byt5small is not None: + out['txt_byt5'] = comfy.conds.CONDRegular(conditioning_byt5small) + + guidance = kwargs.get("guidance", 6.0) + if guidance is not None: + out['guidance'] = comfy.conds.CONDRegular(torch.FloatTensor([guidance])) + + clip_vision_output = kwargs.get("clip_vision_output", None) + if clip_vision_output is not None: + out['clip_fea'] = comfy.conds.CONDRegular(clip_vision_output.last_hidden_state) + + return out + +class HunyuanVideo15_SR_Distilled(HunyuanVideo15): + def __init__(self, model_config, model_type=ModelType.FLOW, device=None): + super().__init__(model_config, model_type, device=device) + + def concat_cond(self, **kwargs): + noise = kwargs.get("noise", None) + image = kwargs.get("concat_latent_image", None) + noise_augmentation = kwargs.get("noise_augmentation", 0.0) + device = kwargs["device"] + + if image is None: + image = torch.zeros([noise.shape[0], noise.shape[1] * 2 + 2, noise.shape[-3], noise.shape[-2], noise.shape[-1]], device=comfy.model_management.intermediate_device()) + else: + image = utils.common_upscale(image.to(device), noise.shape[-1], noise.shape[-2], "bilinear", "center") + #image = self.process_latent_in(image) # scaling wasn't applied in reference code + image = utils.resize_to_batch_size(image, noise.shape[0]) + lq_image_slice = slice(noise.shape[1] + 1, 2 * noise.shape[1] + 1) + if noise_augmentation > 0: + generator = torch.Generator(device="cpu") + generator.manual_seed(kwargs.get("seed", 0) - 10) + noise = torch.randn(image[:, lq_image_slice].shape, generator=generator, dtype=image.dtype, device="cpu").to(image.device) + image[:, lq_image_slice] = noise_augmentation * noise + min(1.0 - noise_augmentation, 0.75) * image[:, lq_image_slice] + else: + image[:, lq_image_slice] = 0.75 * image[:, lq_image_slice] + return image + + def extra_conds(self, **kwargs): + out = super().extra_conds(**kwargs) + out['disable_time_r'] = comfy.conds.CONDConstant(False) + return out diff --git a/comfy/model_detection.py b/comfy/model_detection.py index 3142a7fc3..0131ca25a 100644 --- a/comfy/model_detection.py +++ b/comfy/model_detection.py @@ -186,6 +186,16 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): guidance_keys = list(filter(lambda a: a.startswith("{}guidance_in.".format(key_prefix)), state_dict_keys)) dit_config["guidance_embed"] = len(guidance_keys) > 0 + + # HunyuanVideo 1.5 + if '{}cond_type_embedding.weight'.format(key_prefix) in state_dict_keys: + dit_config["use_cond_type_embedding"] = True + else: + dit_config["use_cond_type_embedding"] = False + if '{}vision_in.proj.0.weight'.format(key_prefix) in state_dict_keys: + dit_config["vision_in_dim"] = state_dict['{}vision_in.proj.0.weight'.format(key_prefix)].shape[0] + else: + dit_config["vision_in_dim"] = None return dit_config if '{}double_blocks.0.img_attn.norm.key_norm.scale'.format(key_prefix) in state_dict_keys and ('{}img_in.weight'.format(key_prefix) in state_dict_keys or f"{key_prefix}distilled_guidance_layer.norms.0.scale" in state_dict_keys): #Flux, Chroma or Chroma Radiance (has no img_in.weight) diff --git a/comfy/sd.py b/comfy/sd.py index 9e5ebbf15..dc0905ada 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -441,20 +441,20 @@ class VAE: elif "decoder.conv_in.conv.weight" in sd and sd['decoder.conv_in.conv.weight'].shape[1] == 32: ddconfig = {"block_out_channels": [128, 256, 512, 1024, 1024], "in_channels": 3, "out_channels": 3, "num_res_blocks": 2, "ffactor_spatial": 16, "ffactor_temporal": 4, "downsample_match_channel": True, "upsample_match_channel": True} ddconfig['z_channels'] = sd["decoder.conv_in.conv.weight"].shape[1] - self.latent_channels = 64 + self.latent_channels = 32 self.upscale_ratio = (lambda a: max(0, a * 4 - 3), 16, 16) self.upscale_index_formula = (4, 16, 16) self.downscale_ratio = (lambda a: max(0, math.floor((a + 3) / 4)), 16, 16) self.downscale_index_formula = (4, 16, 16) self.latent_dim = 3 - self.not_video = True + self.not_video = False self.working_dtypes = [torch.float16, torch.bfloat16, torch.float32] self.first_stage_model = AutoencodingEngine(regularizer_config={'target': "comfy.ldm.models.autoencoder.EmptyRegularizer"}, encoder_config={'target': "comfy.ldm.hunyuan_video.vae_refiner.Encoder", 'params': ddconfig}, decoder_config={'target': "comfy.ldm.hunyuan_video.vae_refiner.Decoder", 'params': ddconfig}) - self.memory_used_encode = lambda shape, dtype: (1400 * shape[-2] * shape[-1]) * model_management.dtype_size(dtype) - self.memory_used_decode = lambda shape, dtype: (1400 * shape[-3] * shape[-2] * shape[-1] * 16 * 16) * model_management.dtype_size(dtype) + self.memory_used_encode = lambda shape, dtype: (1400 * 9 * shape[-2] * shape[-1]) * model_management.dtype_size(dtype) + self.memory_used_decode = lambda shape, dtype: (2800 * 4 * shape[-2] * shape[-1] * 16 * 16) * model_management.dtype_size(dtype) elif "decoder.conv_in.conv.weight" in sd: ddconfig = {'double_z': True, 'z_channels': 4, 'resolution': 256, 'in_channels': 3, 'out_ch': 3, 'ch': 128, 'ch_mult': [1, 2, 4, 4], 'num_res_blocks': 2, 'attn_resolutions': [], 'dropout': 0.0} ddconfig["conv3d"] = True @@ -911,6 +911,7 @@ class CLIPType(Enum): OMNIGEN2 = 17 QWEN_IMAGE = 18 HUNYUAN_IMAGE = 19 + HUNYUAN_VIDEO_15 = 20 def load_clip(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DIFFUSION, model_options={}): @@ -1126,6 +1127,9 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip elif clip_type == CLIPType.HUNYUAN_IMAGE: clip_target.clip = comfy.text_encoders.hunyuan_image.te(**llama_detect(clip_data)) clip_target.tokenizer = comfy.text_encoders.hunyuan_image.HunyuanImageTokenizer + elif clip_type == CLIPType.HUNYUAN_VIDEO_15: + clip_target.clip = comfy.text_encoders.hunyuan_image.te(**llama_detect(clip_data)) + clip_target.tokenizer = comfy.text_encoders.hunyuan_video.HunyuanVideo15Tokenizer else: clip_target.clip = sdxl_clip.SDXLClipModel clip_target.tokenizer = sdxl_clip.SDXLTokenizer diff --git a/comfy/supported_models.py b/comfy/supported_models.py index 4064bdae1..2e64b85e8 100644 --- a/comfy/supported_models.py +++ b/comfy/supported_models.py @@ -1374,6 +1374,54 @@ class HunyuanImage21Refiner(HunyuanVideo): out = model_base.HunyuanImage21Refiner(self, device=device) return out -models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, Omnigen2, QwenImage] +class HunyuanVideo15(HunyuanVideo): + unet_config = { + "image_model": "hunyuan_video", + "vision_in_dim": 1152, + } + + sampling_settings = { + "shift": 7.0, + } + memory_usage_factor = 4.0 #TODO + supported_inference_dtypes = [torch.float16, torch.bfloat16, torch.float32] + + latent_format = latent_formats.HunyuanVideo15 + + def get_model(self, state_dict, prefix="", device=None): + out = model_base.HunyuanVideo15(self, device=device) + return out + + def clip_target(self, state_dict={}): + pref = self.text_encoder_key_prefix[0] + hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen25_7b.transformer.".format(pref)) + return supported_models_base.ClipTarget(comfy.text_encoders.hunyuan_video.HunyuanVideo15Tokenizer, comfy.text_encoders.hunyuan_image.te(**hunyuan_detect)) + + +class HunyuanVideo15_SR_Distilled(HunyuanVideo): + unet_config = { + "image_model": "hunyuan_video", + "vision_in_dim": 1152, + "in_channels": 98, + } + + sampling_settings = { + "shift": 2.0, + } + memory_usage_factor = 4.0 #TODO + supported_inference_dtypes = [torch.float16, torch.bfloat16, torch.float32] + + latent_format = latent_formats.HunyuanVideo15 + + def get_model(self, state_dict, prefix="", device=None): + out = model_base.HunyuanVideo15_SR_Distilled(self, device=device) + return out + + def clip_target(self, state_dict={}): + pref = self.text_encoder_key_prefix[0] + hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen25_7b.transformer.".format(pref)) + return supported_models_base.ClipTarget(comfy.text_encoders.hunyuan_video.HunyuanVideo15Tokenizer, comfy.text_encoders.hunyuan_image.te(**hunyuan_detect)) + +models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanVideo15_SR_Distilled, HunyuanVideo15, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, Omnigen2, QwenImage] models += [SVD_img2vid] diff --git a/comfy/text_encoders/hunyuan_video.py b/comfy/text_encoders/hunyuan_video.py index b02148b33..557094f49 100644 --- a/comfy/text_encoders/hunyuan_video.py +++ b/comfy/text_encoders/hunyuan_video.py @@ -1,6 +1,7 @@ from comfy import sd1_clip import comfy.model_management import comfy.text_encoders.llama +from .hunyuan_image import HunyuanImageTokenizer from transformers import LlamaTokenizerFast import torch import os @@ -73,6 +74,14 @@ class HunyuanVideoTokenizer: return {} +class HunyuanVideo15Tokenizer(HunyuanImageTokenizer): + def __init__(self, embedding_directory=None, tokenizer_data={}): + super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data) + self.llama_template = "<|im_start|>system\nYou are a helpful assistant. Describe the video by detailing the following aspects:\n1. The main content and theme of the video.\n2. The color, shape, size, texture, quantity, text, and spatial relationships of the objects.\n3. Actions, events, behaviors temporal relationships, physical movement changes of the objects.\n4. background environment, light, style and atmosphere.\n5. camera angles, movements, and transitions used in the video.<|im_end|>\n<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n" + + def tokenize_with_weights(self, text:str, return_word_ids=False, **kwargs): + return super().tokenize_with_weights(text, return_word_ids, prevent_empty_text=True, **kwargs) + class HunyuanVideoClipModel(torch.nn.Module): def __init__(self, dtype_llama=None, device="cpu", dtype=None, model_options={}): super().__init__() diff --git a/comfy/text_encoders/qwen_image.py b/comfy/text_encoders/qwen_image.py index 40fa67937..c0d32a6ef 100644 --- a/comfy/text_encoders/qwen_image.py +++ b/comfy/text_encoders/qwen_image.py @@ -17,12 +17,14 @@ class QwenImageTokenizer(sd1_clip.SD1Tokenizer): self.llama_template = "<|im_start|>system\nDescribe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background:<|im_end|>\n<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n" self.llama_template_images = "<|im_start|>system\nDescribe the key features of the input image (color, shape, size, texture, objects, background), then explain how the user's text instruction should alter or modify the image. Generate a new image that meets the user's requirements while maintaining consistency with the original input where appropriate.<|im_end|>\n<|im_start|>user\n<|vision_start|><|image_pad|><|vision_end|>{}<|im_end|>\n<|im_start|>assistant\n" - def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None, images=[], **kwargs): + def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None, images=[], prevent_empty_text=False, **kwargs): skip_template = False if text.startswith('<|im_start|>'): skip_template = True if text.startswith('<|start_header_id|>'): skip_template = True + if prevent_empty_text and text == '': + text = ' ' if skip_template: llama_text = text diff --git a/comfy_api/latest/_io.py b/comfy_api/latest/_io.py index 863254ce7..79c0722a9 100644 --- a/comfy_api/latest/_io.py +++ b/comfy_api/latest/_io.py @@ -629,6 +629,10 @@ class UpscaleModel(ComfyTypeIO): if TYPE_CHECKING: Type = ImageModelDescriptor +@comfytype(io_type="LATENT_UPSCALE_MODEL") +class LatentUpscaleModel(ComfyTypeIO): + Type = Any + @comfytype(io_type="AUDIO") class Audio(ComfyTypeIO): class AudioDict(TypedDict): diff --git a/comfy_extras/nodes_hunyuan.py b/comfy_extras/nodes_hunyuan.py index f7c34d059..5a2e8cc61 100644 --- a/comfy_extras/nodes_hunyuan.py +++ b/comfy_extras/nodes_hunyuan.py @@ -4,7 +4,8 @@ import torch import comfy.model_management from typing_extensions import override from comfy_api.latest import ComfyExtension, io - +from comfy.ldm.hunyuan_video.upsampler import HunyuanVideo15SRModel +import folder_paths class CLIPTextEncodeHunyuanDiT(io.ComfyNode): @classmethod @@ -57,6 +58,199 @@ class EmptyHunyuanLatentVideo(io.ComfyNode): generate = execute # TODO: remove +class EmptyHunyuanVideo15Latent(EmptyHunyuanLatentVideo): + @classmethod + def define_schema(cls): + schema = super().define_schema() + schema.node_id = "EmptyHunyuanVideo15Latent" + return schema + + @classmethod + def execute(cls, width, height, length, batch_size=1) -> io.NodeOutput: + # Using scale factor of 16 instead of 8 + latent = torch.zeros([batch_size, 32, ((length - 1) // 4) + 1, height // 16, width // 16], device=comfy.model_management.intermediate_device()) + return io.NodeOutput({"samples": latent}) + + generate = execute # TODO: remove + + +class HunyuanVideo15ImageToVideo(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="HunyuanVideo15ImageToVideo", + category="conditioning/video_models", + inputs=[ + io.Conditioning.Input("positive"), + io.Conditioning.Input("negative"), + io.Vae.Input("vae"), + io.Int.Input("width", default=848, min=16, max=nodes.MAX_RESOLUTION, step=16), + io.Int.Input("height", default=480, min=16, max=nodes.MAX_RESOLUTION, step=16), + io.Int.Input("length", default=33, min=1, max=nodes.MAX_RESOLUTION, step=4), + io.Int.Input("batch_size", default=1, min=1, max=4096), + io.Image.Input("start_image", optional=True), + io.ClipVisionOutput.Input("clip_vision_output", optional=True), + ], + outputs=[ + io.Conditioning.Output(display_name="positive"), + io.Conditioning.Output(display_name="negative"), + io.Latent.Output(display_name="latent"), + ], + ) + + @classmethod + def execute(cls, positive, negative, vae, width, height, length, batch_size, start_image=None, clip_vision_output=None) -> io.NodeOutput: + latent = torch.zeros([batch_size, 32, ((length - 1) // 4) + 1, height // 16, width // 16], device=comfy.model_management.intermediate_device()) + + if start_image is not None: + start_image = comfy.utils.common_upscale(start_image[:length].movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1) + + encoded = vae.encode(start_image[:, :, :, :3]) + concat_latent_image = torch.zeros((latent.shape[0], 32, latent.shape[2], latent.shape[3], latent.shape[4]), device=comfy.model_management.intermediate_device()) + concat_latent_image[:, :, :encoded.shape[2], :, :] = encoded + + mask = torch.ones((1, 1, latent.shape[2], concat_latent_image.shape[-2], concat_latent_image.shape[-1]), device=start_image.device, dtype=start_image.dtype) + mask[:, :, :((start_image.shape[0] - 1) // 4) + 1] = 0.0 + + positive = node_helpers.conditioning_set_values(positive, {"concat_latent_image": concat_latent_image, "concat_mask": mask}) + negative = node_helpers.conditioning_set_values(negative, {"concat_latent_image": concat_latent_image, "concat_mask": mask}) + + if clip_vision_output is not None: + positive = node_helpers.conditioning_set_values(positive, {"clip_vision_output": clip_vision_output}) + negative = node_helpers.conditioning_set_values(negative, {"clip_vision_output": clip_vision_output}) + + out_latent = {} + out_latent["samples"] = latent + return io.NodeOutput(positive, negative, out_latent) + + +class HunyuanVideo15SuperResolution(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="HunyuanVideo15SuperResolution", + inputs=[ + io.Conditioning.Input("positive"), + io.Conditioning.Input("negative"), + io.Vae.Input("vae", optional=True), + io.Image.Input("start_image", optional=True), + io.ClipVisionOutput.Input("clip_vision_output", optional=True), + io.Latent.Input("latent"), + io.Float.Input("noise_augmentation", default=0.70, min=0.0, max=1.0, step=0.01), + + ], + outputs=[ + io.Conditioning.Output(display_name="positive"), + io.Conditioning.Output(display_name="negative"), + io.Latent.Output(display_name="latent"), + ], + ) + + @classmethod + def execute(cls, positive, negative, latent, noise_augmentation, vae=None, start_image=None, clip_vision_output=None) -> io.NodeOutput: + in_latent = latent["samples"] + in_channels = in_latent.shape[1] + cond_latent = torch.zeros([in_latent.shape[0], in_channels * 2 + 2, in_latent.shape[-3], in_latent.shape[-2], in_latent.shape[-1]], device=comfy.model_management.intermediate_device()) + cond_latent[:, in_channels + 1 : 2 * in_channels + 1] = in_latent + cond_latent[:, 2 * in_channels + 1] = 1 + if start_image is not None: + start_image = comfy.utils.common_upscale(start_image.movedim(-1, 1), in_latent.shape[-1] * 16, in_latent.shape[-2] * 16, "bilinear", "center").movedim(1, -1) + encoded = vae.encode(start_image[:, :, :, :3]) + cond_latent[:, :in_channels, :encoded.shape[2], :, :] = encoded + cond_latent[:, in_channels + 1, 0] = 1 + + positive = node_helpers.conditioning_set_values(positive, {"concat_latent_image": cond_latent, "noise_augmentation": noise_augmentation}) + negative = node_helpers.conditioning_set_values(negative, {"concat_latent_image": cond_latent, "noise_augmentation": noise_augmentation}) + if clip_vision_output is not None: + positive = node_helpers.conditioning_set_values(positive, {"clip_vision_output": clip_vision_output}) + negative = node_helpers.conditioning_set_values(negative, {"clip_vision_output": clip_vision_output}) + + return io.NodeOutput(positive, negative, latent) + + +class LatentUpscaleModelLoader(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="LatentUpscaleModelLoader", + display_name="Load Latent Upscale Model", + category="loaders", + inputs=[ + io.Combo.Input("model_name", options=folder_paths.get_filename_list("latent_upscale_models")), + ], + outputs=[ + io.LatentUpscaleModel.Output(), + ], + ) + + @classmethod + def execute(cls, model_name) -> io.NodeOutput: + model_path = folder_paths.get_full_path_or_raise("upscale_models", model_name) + sd = comfy.utils.load_torch_file(model_path, safe_load=True) + + if "blocks.0.block.0.conv.weight" in sd: + config = { + "in_channels": sd["in_conv.conv.weight"].shape[1], + "out_channels": sd["out_conv.conv.weight"].shape[0], + "hidden_channels": sd["in_conv.conv.weight"].shape[0], + "num_blocks": len([k for k in sd.keys() if k.startswith("blocks.") and k.endswith(".block.0.conv.weight")]), + "global_residual": False, + } + model_type = "720p" + elif "up.0.block.0.conv1.conv.weight" in sd: + sd = {key.replace("nin_shortcut", "nin_shortcut.conv", 1): value for key, value in sd.items()} + config = { + "z_channels": sd["conv_in.conv.weight"].shape[1], + "out_channels": sd["conv_out.conv.weight"].shape[0], + "block_out_channels": tuple(sd[f"up.{i}.block.0.conv1.conv.weight"].shape[0] for i in range(len([k for k in sd.keys() if k.startswith("up.") and k.endswith(".block.0.conv1.conv.weight")]))), + } + model_type = "1080p" + + model = HunyuanVideo15SRModel(model_type, config) + model.load_sd(sd) + + return io.NodeOutput(model) + + +class HunyuanVideo15LatentUpscaleWithModel(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="HunyuanVideo15LatentUpscaleWithModel", + display_name="Hunyuan Video 15 Latent Upscale With Model", + category="latent", + inputs=[ + io.LatentUpscaleModel.Input("model"), + io.Latent.Input("samples"), + io.Combo.Input("upscale_method", options=["nearest-exact", "bilinear", "area", "bicubic", "bislerp"], default="bilinear"), + io.Int.Input("width", default=1280, min=0, max=16384, step=8), + io.Int.Input("height", default=720, min=0, max=16384, step=8), + io.Combo.Input("crop", options=["disabled", "center"]), + ], + outputs=[ + io.Latent.Output(), + ], + ) + + @classmethod + def execute(cls, model, samples, upscale_method, width, height, crop) -> io.NodeOutput: + if width == 0 and height == 0: + return io.NodeOutput(samples) + else: + if width == 0: + height = max(64, height) + width = max(64, round(samples["samples"].shape[-1] * height / samples["samples"].shape[-2])) + elif height == 0: + width = max(64, width) + height = max(64, round(samples["samples"].shape[-2] * width / samples["samples"].shape[-1])) + else: + width = max(64, width) + height = max(64, height) + s = comfy.utils.common_upscale(samples["samples"], width // 16, height // 16, upscale_method, crop) + s = model.resample_latent(s) + return io.NodeOutput({"samples": s.cpu().float()}) + + PROMPT_TEMPLATE_ENCODE_VIDEO_I2V = ( "<|start_header_id|>system<|end_header_id|>\n\n\nDescribe the video by detailing the following aspects according to the reference image: " "1. The main content and theme of the video." @@ -210,6 +404,11 @@ class HunyuanExtension(ComfyExtension): CLIPTextEncodeHunyuanDiT, TextEncodeHunyuanVideo_ImageToVideo, EmptyHunyuanLatentVideo, + EmptyHunyuanVideo15Latent, + HunyuanVideo15ImageToVideo, + HunyuanVideo15SuperResolution, + HunyuanVideo15LatentUpscaleWithModel, + LatentUpscaleModelLoader, HunyuanImageToVideo, EmptyHunyuanImageLatent, HunyuanRefinerLatent, diff --git a/folder_paths.py b/folder_paths.py index f110d832b..ffdc4d020 100644 --- a/folder_paths.py +++ b/folder_paths.py @@ -38,6 +38,8 @@ folder_names_and_paths["gligen"] = ([os.path.join(models_dir, "gligen")], suppor folder_names_and_paths["upscale_models"] = ([os.path.join(models_dir, "upscale_models")], supported_pt_extensions) +folder_names_and_paths["latent_upscale_models"] = ([os.path.join(models_dir, "latent_upscale_models")], supported_pt_extensions) + folder_names_and_paths["custom_nodes"] = ([os.path.join(base_path, "custom_nodes")], set()) folder_names_and_paths["hypernetworks"] = ([os.path.join(models_dir, "hypernetworks")], supported_pt_extensions) diff --git a/models/latent_upscale_models/put_latent_upscale_models_here b/models/latent_upscale_models/put_latent_upscale_models_here new file mode 100644 index 000000000..e69de29bb diff --git a/nodes.py b/nodes.py index 030371633..f023ae3b6 100644 --- a/nodes.py +++ b/nodes.py @@ -957,7 +957,7 @@ class DualCLIPLoader: def INPUT_TYPES(s): return {"required": { "clip_name1": (folder_paths.get_filename_list("text_encoders"), ), "clip_name2": (folder_paths.get_filename_list("text_encoders"), ), - "type": (["sdxl", "sd3", "flux", "hunyuan_video", "hidream", "hunyuan_image"], ), + "type": (["sdxl", "sd3", "flux", "hunyuan_video", "hidream", "hunyuan_image", "hunyuan_video_15"], ), }, "optional": { "device": (["default", "cpu"], {"advanced": True}), From 33981237527a3d84d4e9c3b113f75d6dd37af6a4 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Thu, 20 Nov 2025 20:39:37 -0800 Subject: [PATCH 138/647] Fix wrong path. (#10821) --- comfy_extras/nodes_hunyuan.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy_extras/nodes_hunyuan.py b/comfy_extras/nodes_hunyuan.py index 5a2e8cc61..aa36a471f 100644 --- a/comfy_extras/nodes_hunyuan.py +++ b/comfy_extras/nodes_hunyuan.py @@ -185,7 +185,7 @@ class LatentUpscaleModelLoader(io.ComfyNode): @classmethod def execute(cls, model_name) -> io.NodeOutput: - model_path = folder_paths.get_full_path_or_raise("upscale_models", model_name) + model_path = folder_paths.get_full_path_or_raise("latent_upscale_models", model_name) sd = comfy.utils.load_torch_file(model_path, safe_load=True) if "blocks.0.block.0.conv.weight" in sd: From c55fd7481626d8bee8044ea7512ea996d13a1b90 Mon Sep 17 00:00:00 2001 From: comfyanonymous Date: Fri, 21 Nov 2025 00:49:13 -0500 Subject: [PATCH 139/647] ComfyUI 0.3.71 --- comfyui_version.py | 2 +- pyproject.toml | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/comfyui_version.py b/comfyui_version.py index 9b77aabe9..b4655d553 100644 --- a/comfyui_version.py +++ b/comfyui_version.py @@ -1,3 +1,3 @@ # This file is automatically generated by the build process when version is # updated in pyproject.toml. -__version__ = "0.3.70" +__version__ = "0.3.71" diff --git a/pyproject.toml b/pyproject.toml index 289b7145b..280dbaf53 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -1,6 +1,6 @@ [project] name = "ComfyUI" -version = "0.3.70" +version = "0.3.71" readme = "README.md" license = { file = "LICENSE" } requires-python = ">=3.9" From ecb683b057a19f1a05d18d6d0b0ee9a6c6c8f4a0 Mon Sep 17 00:00:00 2001 From: Christian Byrne Date: Fri, 21 Nov 2025 13:34:47 -0800 Subject: [PATCH 140/647] update frontend to 1.30 (#10793) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 624aa7362..f83d561c9 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,4 +1,4 @@ -comfyui-frontend-package==1.28.9 +comfyui-frontend-package==1.30.6 comfyui-workflow-templates==0.6.0 comfyui-embedded-docs==0.3.1 torch From 532938b16b544e4492ba0ffbe18b201b1a7bc55f Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Fri, 21 Nov 2025 14:51:55 -0800 Subject: [PATCH 141/647] --disable-api-nodes now sets CSP header to force frontend offline. (#10829) --- comfy/cli_args.py | 2 +- server.py | 19 +++++++++++++++++++ 2 files changed, 20 insertions(+), 1 deletion(-) diff --git a/comfy/cli_args.py b/comfy/cli_args.py index 2f30b72d2..d2b60e347 100644 --- a/comfy/cli_args.py +++ b/comfy/cli_args.py @@ -160,7 +160,7 @@ parser.add_argument("--windows-standalone-build", action="store_true", help="Win parser.add_argument("--disable-metadata", action="store_true", help="Disable saving prompt metadata in files.") parser.add_argument("--disable-all-custom-nodes", action="store_true", help="Disable loading all custom nodes.") parser.add_argument("--whitelist-custom-nodes", type=str, nargs='+', default=[], help="Specify custom node folders to load even when --disable-all-custom-nodes is enabled.") -parser.add_argument("--disable-api-nodes", action="store_true", help="Disable loading all api nodes.") +parser.add_argument("--disable-api-nodes", action="store_true", help="Disable loading all api nodes. Also prevents the frontend from communicating with the internet.") parser.add_argument("--multi-user", action="store_true", help="Enables per-user storage.") diff --git a/server.py b/server.py index d9d5c491f..0fd2e49e3 100644 --- a/server.py +++ b/server.py @@ -164,6 +164,22 @@ def create_origin_only_middleware(): return origin_only_middleware + +def create_block_external_middleware(): + @web.middleware + async def block_external_middleware(request: web.Request, handler): + if request.method == "OPTIONS": + # Pre-flight request. Reply successfully: + response = web.Response() + else: + response = await handler(request) + + response.headers['Content-Security-Policy'] = "default-src 'self'; script-src 'self' 'unsafe-inline' blob:; style-src 'self' 'unsafe-inline'; img-src 'self' data: blob:; font-src 'self'; connect-src 'self'; frame-src 'self'; object-src 'self';" + return response + + return block_external_middleware + + class PromptServer(): def __init__(self, loop): PromptServer.instance = self @@ -193,6 +209,9 @@ class PromptServer(): else: middlewares.append(create_origin_only_middleware()) + if args.disable_api_nodes: + middlewares.append(create_block_external_middleware()) + max_upload_size = round(args.max_upload_size * 1024 * 1024) self.app = web.Application(client_max_size=max_upload_size, middlewares=middlewares) self.sockets = dict() From a9c35256bccd4018fbe74bf1e857cc18bd1900ed Mon Sep 17 00:00:00 2001 From: Christian Byrne Date: Sat, 22 Nov 2025 02:28:29 -0800 Subject: [PATCH 142/647] Update requirements.txt (#10834) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index f83d561c9..8e308cd6c 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,5 +1,5 @@ comfyui-frontend-package==1.30.6 -comfyui-workflow-templates==0.6.0 +comfyui-workflow-templates==0.7.9 comfyui-embedded-docs==0.3.1 torch torchsde From d89c29f25992713ec3102017c189858a457f1215 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sat, 22 Nov 2025 19:51:53 -0800 Subject: [PATCH 143/647] Add display names to Hunyuan latent video nodes. (#10837) --- comfy_extras/nodes_hunyuan.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/comfy_extras/nodes_hunyuan.py b/comfy_extras/nodes_hunyuan.py index aa36a471f..32be182f1 100644 --- a/comfy_extras/nodes_hunyuan.py +++ b/comfy_extras/nodes_hunyuan.py @@ -38,6 +38,7 @@ class EmptyHunyuanLatentVideo(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="EmptyHunyuanLatentVideo", + display_name="Empty HunyuanVideo 1.0 Latent", category="latent/video", inputs=[ io.Int.Input("width", default=848, min=16, max=nodes.MAX_RESOLUTION, step=16), @@ -63,6 +64,7 @@ class EmptyHunyuanVideo15Latent(EmptyHunyuanLatentVideo): def define_schema(cls): schema = super().define_schema() schema.node_id = "EmptyHunyuanVideo15Latent" + schema.display_name = "Empty HunyuanVideo 1.5 Latent" return schema @classmethod @@ -71,8 +73,6 @@ class EmptyHunyuanVideo15Latent(EmptyHunyuanLatentVideo): latent = torch.zeros([batch_size, 32, ((length - 1) // 4) + 1, height // 16, width // 16], device=comfy.model_management.intermediate_device()) return io.NodeOutput({"samples": latent}) - generate = execute # TODO: remove - class HunyuanVideo15ImageToVideo(io.ComfyNode): @classmethod From cbd68e3d587a1b345bdc6ebcd8a8c6ba1a9d3af3 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sun, 23 Nov 2025 01:55:22 -0800 Subject: [PATCH 144/647] Add better error message for common error. (#10846) --- comfy/cldm/cldm.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/comfy/cldm/cldm.py b/comfy/cldm/cldm.py index ec01665e2..c93c2e909 100644 --- a/comfy/cldm/cldm.py +++ b/comfy/cldm/cldm.py @@ -413,7 +413,8 @@ class ControlNet(nn.Module): out_middle = [] if self.num_classes is not None: - assert y.shape[0] == x.shape[0] + if y is None: + raise ValueError("y is None, did you try using a controlnet for SDXL on SD1?") emb = emb + self.label_emb(y) h = x From f66183a54142be693ab858e9f1f06ed62439a92e Mon Sep 17 00:00:00 2001 From: guill Date: Sun, 23 Nov 2025 22:56:20 -0800 Subject: [PATCH 145/647] [fix] Fixes non-async public API access (#10857) It looks like the synchronous version of the public API broke due to an addition of `from __future__ import annotations`. This change updates the async-to-sync adapter to work with both types of type annotations. --- comfy_api/internal/async_to_sync.py | 47 ++++++--- tests/execution/test_public_api.py | 153 ++++++++++++++++++++++++++++ 2 files changed, 184 insertions(+), 16 deletions(-) create mode 100644 tests/execution/test_public_api.py diff --git a/comfy_api/internal/async_to_sync.py b/comfy_api/internal/async_to_sync.py index f5f805a62..257ade82e 100644 --- a/comfy_api/internal/async_to_sync.py +++ b/comfy_api/internal/async_to_sync.py @@ -8,7 +8,7 @@ import os import textwrap import threading from enum import Enum -from typing import Optional, Type, get_origin, get_args +from typing import Optional, Type, get_origin, get_args, get_type_hints class TypeTracker: @@ -220,11 +220,18 @@ class AsyncToSyncConverter: self._async_instance = async_class(*args, **kwargs) # Handle annotated class attributes (like execution: Execution) - # Get all annotations from the class hierarchy - all_annotations = {} - for base_class in reversed(inspect.getmro(async_class)): - if hasattr(base_class, "__annotations__"): - all_annotations.update(base_class.__annotations__) + # Get all annotations from the class hierarchy and resolve string annotations + try: + # get_type_hints resolves string annotations to actual type objects + # This handles classes using 'from __future__ import annotations' + all_annotations = get_type_hints(async_class) + except Exception: + # Fallback to raw annotations if get_type_hints fails + # (e.g., for undefined forward references) + all_annotations = {} + for base_class in reversed(inspect.getmro(async_class)): + if hasattr(base_class, "__annotations__"): + all_annotations.update(base_class.__annotations__) # For each annotated attribute, check if it needs to be created or wrapped for attr_name, attr_type in all_annotations.items(): @@ -625,15 +632,19 @@ class AsyncToSyncConverter: """Extract class attributes that are classes themselves.""" class_attributes = [] + # Get resolved type hints to handle string annotations + try: + type_hints = get_type_hints(async_class) + except Exception: + type_hints = {} + # Look for class attributes that are classes for name, attr in sorted(inspect.getmembers(async_class)): if isinstance(attr, type) and not name.startswith("_"): class_attributes.append((name, attr)) - elif ( - hasattr(async_class, "__annotations__") - and name in async_class.__annotations__ - ): - annotation = async_class.__annotations__[name] + elif name in type_hints: + # Use resolved type hint instead of raw annotation + annotation = type_hints[name] if isinstance(annotation, type): class_attributes.append((name, annotation)) @@ -908,11 +919,15 @@ class AsyncToSyncConverter: attribute_mappings = {} # First check annotations for typed attributes (including from parent classes) - # Collect all annotations from the class hierarchy - all_annotations = {} - for base_class in reversed(inspect.getmro(async_class)): - if hasattr(base_class, "__annotations__"): - all_annotations.update(base_class.__annotations__) + # Resolve string annotations to actual types + try: + all_annotations = get_type_hints(async_class) + except Exception: + # Fallback to raw annotations + all_annotations = {} + for base_class in reversed(inspect.getmro(async_class)): + if hasattr(base_class, "__annotations__"): + all_annotations.update(base_class.__annotations__) for attr_name, attr_type in sorted(all_annotations.items()): for class_name, class_type in class_attributes: diff --git a/tests/execution/test_public_api.py b/tests/execution/test_public_api.py new file mode 100644 index 000000000..52bc2fcd8 --- /dev/null +++ b/tests/execution/test_public_api.py @@ -0,0 +1,153 @@ +""" +Tests for public ComfyAPI and ComfyAPISync functions. + +These tests verify that the public API methods work correctly in both sync and async contexts, +ensuring that the sync wrapper generation (via get_type_hints() in async_to_sync.py) correctly +handles string annotations from 'from __future__ import annotations'. +""" + +import pytest +import time +import subprocess +import torch +from pytest import fixture +from comfy_execution.graph_utils import GraphBuilder +from tests.execution.test_execution import ComfyClient + + +@pytest.mark.execution +class TestPublicAPI: + """Test suite for public ComfyAPI and ComfyAPISync methods.""" + + @fixture(scope="class", autouse=True) + def _server(self, args_pytest): + """Start ComfyUI server for testing.""" + pargs = [ + 'python', 'main.py', + '--output-directory', args_pytest["output_dir"], + '--listen', args_pytest["listen"], + '--port', str(args_pytest["port"]), + '--extra-model-paths-config', 'tests/execution/extra_model_paths.yaml', + '--cpu', + ] + p = subprocess.Popen(pargs) + yield + p.kill() + torch.cuda.empty_cache() + + @fixture(scope="class", autouse=True) + def shared_client(self, args_pytest, _server): + """Create shared client with connection retry.""" + client = ComfyClient() + n_tries = 5 + for i in range(n_tries): + time.sleep(4) + try: + client.connect(listen=args_pytest["listen"], port=args_pytest["port"]) + break + except ConnectionRefusedError: + if i == n_tries - 1: + raise + yield client + del client + torch.cuda.empty_cache() + + @fixture + def client(self, shared_client, request): + """Set test name for each test.""" + shared_client.set_test_name(f"public_api[{request.node.name}]") + yield shared_client + + @fixture + def builder(self, request): + """Create GraphBuilder for each test.""" + yield GraphBuilder(prefix=request.node.name) + + def test_sync_progress_update_executes(self, client: ComfyClient, builder: GraphBuilder): + """Test that TestSyncProgressUpdate executes without errors. + + This test validates that api_sync.execution.set_progress() works correctly, + which is the primary code path fixed by adding get_type_hints() to async_to_sync.py. + """ + g = builder + image = g.node("StubImage", content="BLACK", height=256, width=256, batch_size=1) + + # Use TestSyncProgressUpdate with short sleep + progress_node = g.node("TestSyncProgressUpdate", + value=image.out(0), + sleep_seconds=0.5) + output = g.node("SaveImage", images=progress_node.out(0)) + + # Execute workflow + result = client.run(g) + + # Verify execution + assert result.did_run(progress_node), "Progress node should have executed" + assert result.did_run(output), "Output node should have executed" + + # Verify output + images = result.get_images(output) + assert len(images) == 1, "Should have produced 1 image" + + def test_async_progress_update_executes(self, client: ComfyClient, builder: GraphBuilder): + """Test that TestAsyncProgressUpdate executes without errors. + + This test validates that await api.execution.set_progress() works correctly + in async contexts. + """ + g = builder + image = g.node("StubImage", content="WHITE", height=256, width=256, batch_size=1) + + # Use TestAsyncProgressUpdate with short sleep + progress_node = g.node("TestAsyncProgressUpdate", + value=image.out(0), + sleep_seconds=0.5) + output = g.node("SaveImage", images=progress_node.out(0)) + + # Execute workflow + result = client.run(g) + + # Verify execution + assert result.did_run(progress_node), "Async progress node should have executed" + assert result.did_run(output), "Output node should have executed" + + # Verify output + images = result.get_images(output) + assert len(images) == 1, "Should have produced 1 image" + + def test_sync_and_async_progress_together(self, client: ComfyClient, builder: GraphBuilder): + """Test both sync and async progress updates in same workflow. + + This test ensures that both ComfyAPISync and ComfyAPI can coexist and work + correctly in the same workflow execution. + """ + g = builder + image1 = g.node("StubImage", content="BLACK", height=256, width=256, batch_size=1) + image2 = g.node("StubImage", content="WHITE", height=256, width=256, batch_size=1) + + # Use both types of progress nodes + sync_progress = g.node("TestSyncProgressUpdate", + value=image1.out(0), + sleep_seconds=0.3) + async_progress = g.node("TestAsyncProgressUpdate", + value=image2.out(0), + sleep_seconds=0.3) + + # Create outputs + output1 = g.node("SaveImage", images=sync_progress.out(0)) + output2 = g.node("SaveImage", images=async_progress.out(0)) + + # Execute workflow + result = client.run(g) + + # Both should execute successfully + assert result.did_run(sync_progress), "Sync progress node should have executed" + assert result.did_run(async_progress), "Async progress node should have executed" + assert result.did_run(output1), "First output node should have executed" + assert result.did_run(output2), "Second output node should have executed" + + # Verify outputs + images1 = result.get_images(output1) + images2 = result.get_images(output2) + assert len(images1) == 1, "Should have produced 1 image from sync node" + assert len(images2) == 1, "Should have produced 1 image from async node" From 3bd71554a2df14b862cc5e1e875df37ba24af1ac Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Mon, 24 Nov 2025 19:48:37 +0200 Subject: [PATCH 146/647] fix(api-nodes): edge cases in responses for Gemini models (#10860) --- comfy_api_nodes/apis/gemini_api.py | 6 +++--- comfy_api_nodes/nodes_gemini.py | 21 +++++++++++---------- 2 files changed, 14 insertions(+), 13 deletions(-) diff --git a/comfy_api_nodes/apis/gemini_api.py b/comfy_api_nodes/apis/gemini_api.py index 710f173f1..d34590d28 100644 --- a/comfy_api_nodes/apis/gemini_api.py +++ b/comfy_api_nodes/apis/gemini_api.py @@ -113,9 +113,9 @@ class GeminiGenerationConfig(BaseModel): maxOutputTokens: int | None = Field(None, ge=16, le=8192) seed: int | None = Field(None) stopSequences: list[str] | None = Field(None) - temperature: float | None = Field(1, ge=0.0, le=2.0) - topK: int | None = Field(40, ge=1) - topP: float | None = Field(0.95, ge=0.0, le=1.0) + temperature: float | None = Field(None, ge=0.0, le=2.0) + topK: int | None = Field(None, ge=1) + topP: float | None = Field(None, ge=0.0, le=1.0) class GeminiImageConfig(BaseModel): diff --git a/comfy_api_nodes/nodes_gemini.py b/comfy_api_nodes/nodes_gemini.py index be752c885..938a20f84 100644 --- a/comfy_api_nodes/nodes_gemini.py +++ b/comfy_api_nodes/nodes_gemini.py @@ -104,14 +104,14 @@ def get_parts_by_type(response: GeminiGenerateContentResponse, part_type: Litera List of response parts matching the requested type. """ if response.candidates is None: - if response.promptFeedback.blockReason: + if response.promptFeedback and response.promptFeedback.blockReason: feedback = response.promptFeedback raise ValueError( f"Gemini API blocked the request. Reason: {feedback.blockReason} ({feedback.blockReasonMessage})" ) - raise NotImplementedError( - "Gemini returned no response candidates. " - "Please report to ComfyUI repository with the example of workflow to reproduce this." + raise ValueError( + "Gemini API returned no response candidates. If you are using the `IMAGE` modality, " + "try changing it to `IMAGE+TEXT` to view the model's reasoning and understand why image generation failed." ) parts = [] for part in response.candidates[0].content.parts: @@ -182,11 +182,12 @@ def calculate_tokens_price(response: GeminiGenerateContentResponse) -> float | N else: return None final_price = response.usageMetadata.promptTokenCount * input_tokens_price - for i in response.usageMetadata.candidatesTokensDetails: - if i.modality == Modality.IMAGE: - final_price += output_image_tokens_price * i.tokenCount # for Nano Banana models - else: - final_price += output_text_tokens_price * i.tokenCount + if response.usageMetadata.candidatesTokensDetails: + for i in response.usageMetadata.candidatesTokensDetails: + if i.modality == Modality.IMAGE: + final_price += output_image_tokens_price * i.tokenCount # for Nano Banana models + else: + final_price += output_text_tokens_price * i.tokenCount if response.usageMetadata.thoughtsTokenCount: final_price += output_text_tokens_price * response.usageMetadata.thoughtsTokenCount return final_price / 1_000_000.0 @@ -645,7 +646,7 @@ class GeminiImage2(IO.ComfyNode): options=["auto", "1:1", "2:3", "3:2", "3:4", "4:3", "4:5", "5:4", "9:16", "16:9", "21:9"], default="auto", tooltip="If set to 'auto', matches your input image's aspect ratio; " - "if no image is provided, generates a 1:1 square.", + "if no image is provided, a 16:9 square is usually generated.", ), IO.Combo.Input( "resolution", From 1286fcfe40b98052e4edbe9a02f12ad89ac74924 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Mon, 24 Nov 2025 20:24:29 +0200 Subject: [PATCH 147/647] add get_frame_count and get_frame_rate methods to VideoInput class (#10851) --- comfy_api/latest/_input/video_types.py | 28 ++++++++ comfy_api/latest/_input_impl/video_types.py | 72 +++++++++++++++++++++ comfy_api_nodes/nodes_topaz.py | 15 ++--- 3 files changed, 106 insertions(+), 9 deletions(-) diff --git a/comfy_api/latest/_input/video_types.py b/comfy_api/latest/_input/video_types.py index a335df4d0..87c81d73a 100644 --- a/comfy_api/latest/_input/video_types.py +++ b/comfy_api/latest/_input/video_types.py @@ -1,5 +1,6 @@ from __future__ import annotations from abc import ABC, abstractmethod +from fractions import Fraction from typing import Optional, Union, IO import io import av @@ -72,6 +73,33 @@ class VideoInput(ABC): frame_count = components.images.shape[0] return float(frame_count / components.frame_rate) + def get_frame_count(self) -> int: + """ + Returns the number of frames in the video. + + Default implementation uses :meth:`get_components`, which may require + loading all frames into memory. File-based implementations should + override this method and use container/stream metadata instead. + + Returns: + Total number of frames as an integer. + """ + return int(self.get_components().images.shape[0]) + + def get_frame_rate(self) -> Fraction: + """ + Returns the frame rate of the video. + + Default implementation materializes the video into memory via + `get_components()`. Subclasses that can inspect the underlying + container (e.g. `VideoFromFile`) should override this with a more + efficient implementation. + + Returns: + Frame rate as a Fraction. + """ + return self.get_components().frame_rate + def get_container_format(self) -> str: """ Returns the container format of the video (e.g., 'mp4', 'mov', 'avi'). diff --git a/comfy_api/latest/_input_impl/video_types.py b/comfy_api/latest/_input_impl/video_types.py index f646504c8..bde37f90a 100644 --- a/comfy_api/latest/_input_impl/video_types.py +++ b/comfy_api/latest/_input_impl/video_types.py @@ -121,6 +121,71 @@ class VideoFromFile(VideoInput): raise ValueError(f"Could not determine duration for file '{self.__file}'") + def get_frame_count(self) -> int: + """ + Returns the number of frames in the video without materializing them as + torch tensors. + """ + if isinstance(self.__file, io.BytesIO): + self.__file.seek(0) + + with av.open(self.__file, mode="r") as container: + video_stream = self._get_first_video_stream(container) + # 1. Prefer the frames field if available + if video_stream.frames and video_stream.frames > 0: + return int(video_stream.frames) + + # 2. Try to estimate from duration and average_rate using only metadata + if container.duration is not None and video_stream.average_rate: + duration_seconds = float(container.duration / av.time_base) + estimated_frames = int(round(duration_seconds * float(video_stream.average_rate))) + if estimated_frames > 0: + return estimated_frames + + if ( + getattr(video_stream, "duration", None) is not None + and getattr(video_stream, "time_base", None) is not None + and video_stream.average_rate + ): + duration_seconds = float(video_stream.duration * video_stream.time_base) + estimated_frames = int(round(duration_seconds * float(video_stream.average_rate))) + if estimated_frames > 0: + return estimated_frames + + # 3. Last resort: decode frames and count them (streaming) + frame_count = 0 + container.seek(0) + for packet in container.demux(video_stream): + for _ in packet.decode(): + frame_count += 1 + + if frame_count == 0: + raise ValueError(f"Could not determine frame count for file '{self.__file}'") + return frame_count + + def get_frame_rate(self) -> Fraction: + """ + Returns the average frame rate of the video using container metadata + without decoding all frames. + """ + if isinstance(self.__file, io.BytesIO): + self.__file.seek(0) + + with av.open(self.__file, mode="r") as container: + video_stream = self._get_first_video_stream(container) + # Preferred: use PyAV's average_rate (usually already a Fraction-like) + if video_stream.average_rate: + return Fraction(video_stream.average_rate) + + # Fallback: estimate from frames + duration if available + if video_stream.frames and container.duration: + duration_seconds = float(container.duration / av.time_base) + if duration_seconds > 0: + return Fraction(video_stream.frames / duration_seconds).limit_denominator() + + # Last resort: match get_components_internal default + return Fraction(1) + def get_container_format(self) -> str: """ Returns the container format of the video (e.g., 'mp4', 'mov', 'avi'). @@ -238,6 +303,13 @@ class VideoFromFile(VideoInput): packet.stream = stream_map[packet.stream] output_container.mux(packet) + def _get_first_video_stream(self, container: InputContainer): + video_stream = next((s for s in container.streams if s.type == "video"), None) + if video_stream is None: + raise ValueError(f"No video stream found in file '{self.__file}'") + return video_stream + + class VideoFromComponents(VideoInput): """ Class representing video input from tensors. diff --git a/comfy_api_nodes/nodes_topaz.py b/comfy_api_nodes/nodes_topaz.py index 79c7bf43d..f522756e5 100644 --- a/comfy_api_nodes/nodes_topaz.py +++ b/comfy_api_nodes/nodes_topaz.py @@ -5,8 +5,7 @@ import aiohttp import torch from typing_extensions import override -from comfy_api.input.video_types import VideoInput -from comfy_api.latest import IO, ComfyExtension +from comfy_api.latest import IO, ComfyExtension, Input from comfy_api_nodes.apis import topaz_api from comfy_api_nodes.util import ( ApiEndpoint, @@ -282,7 +281,7 @@ class TopazVideoEnhance(IO.ComfyNode): @classmethod async def execute( cls, - video: VideoInput, + video: Input.Video, upscaler_enabled: bool, upscaler_model: str, upscaler_resolution: str, @@ -297,12 +296,10 @@ class TopazVideoEnhance(IO.ComfyNode): ) -> IO.NodeOutput: if upscaler_enabled is False and interpolation_enabled is False: raise ValueError("There is nothing to do: both upscaling and interpolation are disabled.") - src_width, src_height = video.get_dimensions() - video_components = video.get_components() - src_frame_rate = int(video_components.frame_rate) - duration_sec = video.get_duration() - estimated_frames = int(duration_sec * src_frame_rate) validate_container_format_is_mp4(video) + src_width, src_height = video.get_dimensions() + src_frame_rate = int(video.get_frame_rate()) + duration_sec = video.get_duration() src_video_stream = video.get_stream_source() target_width = src_width target_height = src_height @@ -338,7 +335,7 @@ class TopazVideoEnhance(IO.ComfyNode): container="mp4", size=get_fs_object_size(src_video_stream), duration=int(duration_sec), - frameCount=estimated_frames, + frameCount=video.get_frame_count(), frameRate=src_frame_rate, resolution=topaz_api.Resolution(width=src_width, height=src_height), ), From 3d1fdaf9f448b34e4eba68bfd8e8de373ec0d22d Mon Sep 17 00:00:00 2001 From: Haoming <73768377+Haoming02@users.noreply.github.com> Date: Tue, 25 Nov 2025 02:30:40 +0800 Subject: [PATCH 148/647] block info (#10843) --- comfy/ldm/chroma/model.py | 6 ++++++ 1 file changed, 6 insertions(+) diff --git a/comfy/ldm/chroma/model.py b/comfy/ldm/chroma/model.py index 67bf70eb1..a72f8cc47 100644 --- a/comfy/ldm/chroma/model.py +++ b/comfy/ldm/chroma/model.py @@ -179,7 +179,10 @@ class Chroma(nn.Module): pe = self.pe_embedder(ids) blocks_replace = patches_replace.get("dit", {}) + transformer_options["total_blocks"] = len(self.double_blocks) + transformer_options["block_type"] = "double" for i, block in enumerate(self.double_blocks): + transformer_options["block_index"] = i if i not in self.skip_mmdit: double_mod = ( self.get_modulations(mod_vectors, "double_img", idx=i), @@ -222,7 +225,10 @@ class Chroma(nn.Module): img = torch.cat((txt, img), 1) + transformer_options["total_blocks"] = len(self.single_blocks) + transformer_options["block_type"] = "single" for i, block in enumerate(self.single_blocks): + transformer_options["block_index"] = i if i not in self.skip_dit: single_mod = self.get_modulations(mod_vectors, "single", idx=i) if ("single_block", i) in blocks_replace: From 6a6d456c88723538e3d0e5e942f78109ece5b73d Mon Sep 17 00:00:00 2001 From: Haoming <73768377+Haoming02@users.noreply.github.com> Date: Tue, 25 Nov 2025 02:38:38 +0800 Subject: [PATCH 149/647] block info (#10842) --- comfy/ldm/qwen_image/model.py | 3 +++ 1 file changed, 3 insertions(+) diff --git a/comfy/ldm/qwen_image/model.py b/comfy/ldm/qwen_image/model.py index 427ea19c1..8c75670cd 100644 --- a/comfy/ldm/qwen_image/model.py +++ b/comfy/ldm/qwen_image/model.py @@ -439,7 +439,10 @@ class QwenImageTransformer2DModel(nn.Module): patches = transformer_options.get("patches", {}) blocks_replace = patches_replace.get("dit", {}) + transformer_options["total_blocks"] = len(self.transformer_blocks) + transformer_options["block_type"] = "double" for i, block in enumerate(self.transformer_blocks): + transformer_options["block_index"] = i if ("double_block", i) in blocks_replace: def block_wrap(args): out = {} From b2ef58e2b17e73ca8cd376a1cdc976518ebbc168 Mon Sep 17 00:00:00 2001 From: Haoming <73768377+Haoming02@users.noreply.github.com> Date: Tue, 25 Nov 2025 02:40:09 +0800 Subject: [PATCH 150/647] block info (#10844) --- comfy/ldm/hunyuan_video/model.py | 6 ++++++ 1 file changed, 6 insertions(+) diff --git a/comfy/ldm/hunyuan_video/model.py b/comfy/ldm/hunyuan_video/model.py index f75c6e0e1..2749c53f5 100644 --- a/comfy/ldm/hunyuan_video/model.py +++ b/comfy/ldm/hunyuan_video/model.py @@ -389,7 +389,10 @@ class HunyuanVideo(nn.Module): attn_mask = None blocks_replace = patches_replace.get("dit", {}) + transformer_options["total_blocks"] = len(self.double_blocks) + transformer_options["block_type"] = "double" for i, block in enumerate(self.double_blocks): + transformer_options["block_index"] = i if ("double_block", i) in blocks_replace: def block_wrap(args): out = {} @@ -411,7 +414,10 @@ class HunyuanVideo(nn.Module): img = torch.cat((img, txt), 1) + transformer_options["total_blocks"] = len(self.single_blocks) + transformer_options["block_type"] = "single" for i, block in enumerate(self.single_blocks): + transformer_options["block_index"] = i if ("single_block", i) in blocks_replace: def block_wrap(args): out = {} From 22a2644e57530ee40e13486ccd7c953b87072093 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 24 Nov 2025 16:45:54 -0800 Subject: [PATCH 151/647] Bump transformers version in requirements.txt (#10869) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 8e308cd6c..b7014f956 100644 --- a/requirements.txt +++ b/requirements.txt @@ -7,7 +7,7 @@ torchvision torchaudio numpy>=1.25.0 einops -transformers>=4.37.2 +transformers>=4.50.3 tokenizers>=0.13.3 sentencepiece safetensors>=0.4.2 From 25022e0b0965975b35bcaf28b153184d60a4f9de Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 24 Nov 2025 22:48:53 -0800 Subject: [PATCH 152/647] Cleanup and fix issues with text encoder quants. (#10872) --- comfy/model_patcher.py | 3 +- comfy/ops.py | 168 +++++++++--------- comfy/quant_ops.py | 12 ++ comfy/sd.py | 9 +- comfy/sd1_clip.py | 18 +- comfy/text_encoders/hunyuan_video.py | 3 + .../comfy_quant/test_mixed_precision.py | 17 +- 7 files changed, 128 insertions(+), 102 deletions(-) diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index cf1b0d441..6551ced5a 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -231,7 +231,6 @@ class ModelPatcher: self.object_patches_backup = {} self.weight_wrapper_patches = {} self.model_options = {"transformer_options":{}} - self.model_size() self.load_device = load_device self.offload_device = offload_device self.weight_inplace_update = weight_inplace_update @@ -286,7 +285,7 @@ class ModelPatcher: return self.model.lowvram_patch_counter def clone(self): - n = self.__class__(self.model, self.load_device, self.offload_device, self.size, weight_inplace_update=self.weight_inplace_update) + n = self.__class__(self.model, self.load_device, self.offload_device, self.model_size(), weight_inplace_update=self.weight_inplace_update) n.patches = {} for k in self.patches: n.patches[k] = self.patches[k][:] diff --git a/comfy/ops.py b/comfy/ops.py index 640622fd1..af185ec24 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -540,113 +540,115 @@ if CUBLAS_IS_AVAILABLE: # ============================================================================== from .quant_ops import QuantizedTensor, QUANT_ALGOS -class MixedPrecisionOps(disable_weight_init): - _layer_quant_config = {} - _compute_dtype = torch.bfloat16 - class Linear(torch.nn.Module, CastWeightBiasOp): - def __init__( - self, - in_features: int, - out_features: int, - bias: bool = True, - device=None, - dtype=None, - ) -> None: - super().__init__() +def mixed_precision_ops(layer_quant_config={}, compute_dtype=torch.bfloat16, full_precision_mm=False): + class MixedPrecisionOps(manual_cast): + _layer_quant_config = layer_quant_config + _compute_dtype = compute_dtype + _full_precision_mm = full_precision_mm - self.factory_kwargs = {"device": device, "dtype": MixedPrecisionOps._compute_dtype} - # self.factory_kwargs = {"device": device, "dtype": dtype} + class Linear(torch.nn.Module, CastWeightBiasOp): + def __init__( + self, + in_features: int, + out_features: int, + bias: bool = True, + device=None, + dtype=None, + ) -> None: + super().__init__() - self.in_features = in_features - self.out_features = out_features - if bias: - self.bias = torch.nn.Parameter(torch.empty(out_features, **self.factory_kwargs)) - else: - self.register_parameter("bias", None) + self.factory_kwargs = {"device": device, "dtype": MixedPrecisionOps._compute_dtype} + # self.factory_kwargs = {"device": device, "dtype": dtype} - self.tensor_class = None + self.in_features = in_features + self.out_features = out_features + if bias: + self.bias = torch.nn.Parameter(torch.empty(out_features, **self.factory_kwargs)) + else: + self.register_parameter("bias", None) - def reset_parameters(self): - return None + self.tensor_class = None + self._full_precision_mm = MixedPrecisionOps._full_precision_mm - def _load_from_state_dict(self, state_dict, prefix, local_metadata, - strict, missing_keys, unexpected_keys, error_msgs): + def reset_parameters(self): + return None - device = self.factory_kwargs["device"] - layer_name = prefix.rstrip('.') - weight_key = f"{prefix}weight" - weight = state_dict.pop(weight_key, None) - if weight is None: - raise ValueError(f"Missing weight for layer {layer_name}") + def _load_from_state_dict(self, state_dict, prefix, local_metadata, + strict, missing_keys, unexpected_keys, error_msgs): - manually_loaded_keys = [weight_key] + device = self.factory_kwargs["device"] + layer_name = prefix.rstrip('.') + weight_key = f"{prefix}weight" + weight = state_dict.pop(weight_key, None) + if weight is None: + raise ValueError(f"Missing weight for layer {layer_name}") - if layer_name not in MixedPrecisionOps._layer_quant_config: - self.weight = torch.nn.Parameter(weight.to(device=device, dtype=MixedPrecisionOps._compute_dtype), requires_grad=False) - else: - quant_format = MixedPrecisionOps._layer_quant_config[layer_name].get("format", None) - if quant_format is None: - raise ValueError(f"Unknown quantization format for layer {layer_name}") + manually_loaded_keys = [weight_key] - qconfig = QUANT_ALGOS[quant_format] - self.layout_type = qconfig["comfy_tensor_layout"] + if layer_name not in MixedPrecisionOps._layer_quant_config: + self.weight = torch.nn.Parameter(weight.to(device=device, dtype=MixedPrecisionOps._compute_dtype), requires_grad=False) + else: + quant_format = MixedPrecisionOps._layer_quant_config[layer_name].get("format", None) + if quant_format is None: + raise ValueError(f"Unknown quantization format for layer {layer_name}") - weight_scale_key = f"{prefix}weight_scale" - layout_params = { - 'scale': state_dict.pop(weight_scale_key, None), - 'orig_dtype': MixedPrecisionOps._compute_dtype, - 'block_size': qconfig.get("group_size", None), - } - if layout_params['scale'] is not None: - manually_loaded_keys.append(weight_scale_key) + qconfig = QUANT_ALGOS[quant_format] + self.layout_type = qconfig["comfy_tensor_layout"] - self.weight = torch.nn.Parameter( - QuantizedTensor(weight.to(device=device), self.layout_type, layout_params), - requires_grad=False - ) + weight_scale_key = f"{prefix}weight_scale" + layout_params = { + 'scale': state_dict.pop(weight_scale_key, None), + 'orig_dtype': MixedPrecisionOps._compute_dtype, + 'block_size': qconfig.get("group_size", None), + } + if layout_params['scale'] is not None: + manually_loaded_keys.append(weight_scale_key) - for param_name in qconfig["parameters"]: - param_key = f"{prefix}{param_name}" - _v = state_dict.pop(param_key, None) - if _v is None: - continue - setattr(self, param_name, torch.nn.Parameter(_v.to(device=device), requires_grad=False)) - manually_loaded_keys.append(param_key) + self.weight = torch.nn.Parameter( + QuantizedTensor(weight.to(device=device), self.layout_type, layout_params), + requires_grad=False + ) - super()._load_from_state_dict(state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs) + for param_name in qconfig["parameters"]: + param_key = f"{prefix}{param_name}" + _v = state_dict.pop(param_key, None) + if _v is None: + continue + setattr(self, param_name, torch.nn.Parameter(_v.to(device=device), requires_grad=False)) + manually_loaded_keys.append(param_key) - for key in manually_loaded_keys: - if key in missing_keys: - missing_keys.remove(key) + super()._load_from_state_dict(state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs) - def _forward(self, input, weight, bias): - return torch.nn.functional.linear(input, weight, bias) + for key in manually_loaded_keys: + if key in missing_keys: + missing_keys.remove(key) - def forward_comfy_cast_weights(self, input): - weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True) - x = self._forward(input, weight, bias) - uncast_bias_weight(self, weight, bias, offload_stream) - return x + def _forward(self, input, weight, bias): + return torch.nn.functional.linear(input, weight, bias) - def forward(self, input, *args, **kwargs): - run_every_op() + def forward_comfy_cast_weights(self, input): + weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True) + x = self._forward(input, weight, bias) + uncast_bias_weight(self, weight, bias, offload_stream) + return x - if self.comfy_cast_weights or len(self.weight_function) > 0 or len(self.bias_function) > 0: - return self.forward_comfy_cast_weights(input, *args, **kwargs) - if (getattr(self, 'layout_type', None) is not None and - getattr(self, 'input_scale', None) is not None and - not isinstance(input, QuantizedTensor)): - input = QuantizedTensor.from_float(input, self.layout_type, scale=self.input_scale, dtype=self.weight.dtype) - return self._forward(input, self.weight, self.bias) + def forward(self, input, *args, **kwargs): + run_every_op() + if self._full_precision_mm or self.comfy_cast_weights or len(self.weight_function) > 0 or len(self.bias_function) > 0: + return self.forward_comfy_cast_weights(input, *args, **kwargs) + if (getattr(self, 'layout_type', None) is not None and + getattr(self, 'input_scale', None) is not None and + not isinstance(input, QuantizedTensor)): + input = QuantizedTensor.from_float(input, self.layout_type, scale=self.input_scale, dtype=self.weight.dtype) + return self._forward(input, self.weight, self.bias) + return MixedPrecisionOps def pick_operations(weight_dtype, compute_dtype, load_device=None, disable_fast_fp8=False, fp8_optimizations=False, scaled_fp8=None, model_config=None): if model_config and hasattr(model_config, 'layer_quant_config') and model_config.layer_quant_config: - MixedPrecisionOps._layer_quant_config = model_config.layer_quant_config - MixedPrecisionOps._compute_dtype = compute_dtype logging.info(f"Using mixed precision operations: {len(model_config.layer_quant_config)} quantized layers") - return MixedPrecisionOps + return mixed_precision_ops(model_config.layer_quant_config, compute_dtype) fp8_compute = comfy.model_management.supports_fp8_compute(load_device) if scaled_fp8 is not None: diff --git a/comfy/quant_ops.py b/comfy/quant_ops.py index 1d058bece..905b4729e 100644 --- a/comfy/quant_ops.py +++ b/comfy/quant_ops.py @@ -338,6 +338,18 @@ def generic_copy_(func, args, kwargs): return func(*args, **kwargs) +@register_generic_util(torch.ops.aten.to.dtype) +def generic_to_dtype(func, args, kwargs): + """Handle .to(dtype) calls - dtype conversion only.""" + src = args[0] + if isinstance(src, QuantizedTensor): + # For dtype-only conversion, just change the orig_dtype, no real cast is needed + target_dtype = args[1] if len(args) > 1 else kwargs.get('dtype') + src._layout_params["orig_dtype"] = target_dtype + return src + return func(*args, **kwargs) + + @register_generic_util(torch.ops.aten._has_compatible_shallow_copy_type.default) def generic_has_compatible_shallow_copy_type(func, args, kwargs): return True diff --git a/comfy/sd.py b/comfy/sd.py index dc0905ada..b6df0bd61 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -917,7 +917,12 @@ class CLIPType(Enum): def load_clip(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DIFFUSION, model_options={}): clip_data = [] for p in ckpt_paths: - clip_data.append(comfy.utils.load_torch_file(p, safe_load=True)) + sd, metadata = comfy.utils.load_torch_file(p, safe_load=True, return_metadata=True) + if metadata is not None: + quant_metadata = metadata.get("_quantization_metadata", None) + if quant_metadata is not None: + sd["_quantization_metadata"] = quant_metadata + clip_data.append(sd) return load_text_encoder_state_dicts(clip_data, embedding_directory=embedding_directory, clip_type=clip_type, model_options=model_options) @@ -1142,6 +1147,8 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip parameters = 0 for c in clip_data: + if "_quantization_metadata" in c: + c.pop("_quantization_metadata") parameters += comfy.utils.calculate_parameters(c) tokenizer_data, model_options = comfy.text_encoders.long_clipl.model_options_long_clip(c, tokenizer_data, model_options) diff --git a/comfy/sd1_clip.py b/comfy/sd1_clip.py index 3066de2d7..8f509bab1 100644 --- a/comfy/sd1_clip.py +++ b/comfy/sd1_clip.py @@ -109,13 +109,23 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder): operations = model_options.get("custom_operations", None) scaled_fp8 = None + quantization_metadata = model_options.get("quantization_metadata", None) if operations is None: - scaled_fp8 = model_options.get("scaled_fp8", None) - if scaled_fp8 is not None: - operations = comfy.ops.scaled_fp8_ops(fp8_matrix_mult=False, override_dtype=scaled_fp8) + layer_quant_config = None + if quantization_metadata is not None: + layer_quant_config = json.loads(quantization_metadata).get("layers", None) + + if layer_quant_config is not None: + operations = comfy.ops.mixed_precision_ops(layer_quant_config, dtype, full_precision_mm=True) + logging.info(f"Using MixedPrecisionOps for text encoder: {len(layer_quant_config)} quantized layers") else: - operations = comfy.ops.manual_cast + # Fallback to scaled_fp8_ops for backward compatibility + scaled_fp8 = model_options.get("scaled_fp8", None) + if scaled_fp8 is not None: + operations = comfy.ops.scaled_fp8_ops(fp8_matrix_mult=False, override_dtype=scaled_fp8) + else: + operations = comfy.ops.manual_cast self.operations = operations self.transformer = model_class(config, dtype, device, self.operations) diff --git a/comfy/text_encoders/hunyuan_video.py b/comfy/text_encoders/hunyuan_video.py index 557094f49..0110517bb 100644 --- a/comfy/text_encoders/hunyuan_video.py +++ b/comfy/text_encoders/hunyuan_video.py @@ -18,6 +18,9 @@ def llama_detect(state_dict, prefix=""): if scaled_fp8_key in state_dict: out["llama_scaled_fp8"] = state_dict[scaled_fp8_key].dtype + if "_quantization_metadata" in state_dict: + out["llama_quantization_metadata"] = state_dict["_quantization_metadata"] + return out diff --git a/tests-unit/comfy_quant/test_mixed_precision.py b/tests-unit/comfy_quant/test_mixed_precision.py index f8d1fd04e..63361309f 100644 --- a/tests-unit/comfy_quant/test_mixed_precision.py +++ b/tests-unit/comfy_quant/test_mixed_precision.py @@ -37,11 +37,8 @@ class TestMixedPrecisionOps(unittest.TestCase): def test_all_layers_standard(self): """Test that model with no quantization works normally""" - # Configure no quantization - ops.MixedPrecisionOps._layer_quant_config = {} - # Create model - model = SimpleModel(operations=ops.MixedPrecisionOps) + model = SimpleModel(operations=ops.mixed_precision_ops({})) # Initialize weights manually model.layer1.weight = torch.nn.Parameter(torch.randn(20, 10, dtype=torch.bfloat16)) @@ -76,7 +73,6 @@ class TestMixedPrecisionOps(unittest.TestCase): "params": {} } } - ops.MixedPrecisionOps._layer_quant_config = layer_quant_config # Create state dict with mixed precision fp8_weight1 = torch.randn(20, 10, dtype=torch.float32).to(torch.float8_e4m3fn) @@ -99,7 +95,7 @@ class TestMixedPrecisionOps(unittest.TestCase): } # Create model and load state dict (strict=False because custom loading pops keys) - model = SimpleModel(operations=ops.MixedPrecisionOps) + model = SimpleModel(operations=ops.mixed_precision_ops(layer_quant_config)) model.load_state_dict(state_dict, strict=False) # Verify weights are wrapped in QuantizedTensor @@ -132,7 +128,6 @@ class TestMixedPrecisionOps(unittest.TestCase): "params": {} } } - ops.MixedPrecisionOps._layer_quant_config = layer_quant_config # Create and load model fp8_weight = torch.randn(20, 10, dtype=torch.float32).to(torch.float8_e4m3fn) @@ -146,7 +141,7 @@ class TestMixedPrecisionOps(unittest.TestCase): "layer3.bias": torch.randn(40, dtype=torch.bfloat16), } - model = SimpleModel(operations=ops.MixedPrecisionOps) + model = SimpleModel(operations=ops.mixed_precision_ops(layer_quant_config)) model.load_state_dict(state_dict1, strict=False) # Save state dict @@ -170,7 +165,6 @@ class TestMixedPrecisionOps(unittest.TestCase): "params": {} } } - ops.MixedPrecisionOps._layer_quant_config = layer_quant_config # Create and load model fp8_weight = torch.randn(20, 10, dtype=torch.float32).to(torch.float8_e4m3fn) @@ -184,7 +178,7 @@ class TestMixedPrecisionOps(unittest.TestCase): "layer3.bias": torch.randn(40, dtype=torch.bfloat16), } - model = SimpleModel(operations=ops.MixedPrecisionOps) + model = SimpleModel(operations=ops.mixed_precision_ops(layer_quant_config)) model.load_state_dict(state_dict, strict=False) # Add a weight function (simulating LoRA) @@ -210,7 +204,6 @@ class TestMixedPrecisionOps(unittest.TestCase): "params": {} } } - ops.MixedPrecisionOps._layer_quant_config = layer_quant_config # Create state dict state_dict = { @@ -223,7 +216,7 @@ class TestMixedPrecisionOps(unittest.TestCase): } # Load should raise KeyError for unknown format in QUANT_FORMAT_MIXINS - model = SimpleModel(operations=ops.MixedPrecisionOps) + model = SimpleModel(operations=ops.mixed_precision_ops(layer_quant_config)) with self.assertRaises(KeyError): model.load_state_dict(state_dict, strict=False) From b6805429b9c2f3aa919035bea849ecd1de3ac8e4 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 24 Nov 2025 23:48:20 -0800 Subject: [PATCH 153/647] Allow pinning quantized tensors. (#10873) --- comfy/model_management.py | 6 +++++- comfy/quant_ops.py | 8 ++++++++ 2 files changed, 13 insertions(+), 1 deletion(-) diff --git a/comfy/model_management.py b/comfy/model_management.py index a21df54b3..a9327ac80 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -1098,13 +1098,14 @@ if not args.disable_pinned_memory: MAX_PINNED_MEMORY = get_total_memory(torch.device("cpu")) * 0.95 logging.info("Enabled pinned memory {}".format(MAX_PINNED_MEMORY // (1024 * 1024))) +PINNING_ALLOWED_TYPES = set(["Parameter", "QuantizedTensor"]) def pin_memory(tensor): global TOTAL_PINNED_MEMORY if MAX_PINNED_MEMORY <= 0: return False - if type(tensor) is not torch.nn.parameter.Parameter: + if type(tensor).__name__ not in PINNING_ALLOWED_TYPES: return False if not is_device_cpu(tensor.device): @@ -1124,6 +1125,9 @@ def pin_memory(tensor): return False ptr = tensor.data_ptr() + if ptr == 0: + return False + if torch.cuda.cudart().cudaHostRegister(ptr, size, 1) == 0: PINNED_MEMORY[ptr] = size TOTAL_PINNED_MEMORY += size diff --git a/comfy/quant_ops.py b/comfy/quant_ops.py index 905b4729e..e938144a7 100644 --- a/comfy/quant_ops.py +++ b/comfy/quant_ops.py @@ -228,6 +228,14 @@ class QuantizedTensor(torch.Tensor): new_kwargs = dequant_arg(kwargs) return func(*new_args, **new_kwargs) + def data_ptr(self): + return self._qdata.data_ptr() + + def is_pinned(self): + return self._qdata.is_pinned() + + def is_contiguous(self): + return self._qdata.is_contiguous() # ============================================================================== # Generic Utilities (Layout-Agnostic Operations) From acfaa5c4a132e1c01bc9d94e76b0d667c899bfd1 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 24 Nov 2025 23:55:49 -0800 Subject: [PATCH 154/647] Don't try fp8 matrix mult in quantized ops if not supported by hardware. (#10874) --- comfy/ops.py | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-) diff --git a/comfy/ops.py b/comfy/ops.py index af185ec24..785aa1c9f 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -646,11 +646,12 @@ def mixed_precision_ops(layer_quant_config={}, compute_dtype=torch.bfloat16, ful return MixedPrecisionOps def pick_operations(weight_dtype, compute_dtype, load_device=None, disable_fast_fp8=False, fp8_optimizations=False, scaled_fp8=None, model_config=None): + fp8_compute = comfy.model_management.supports_fp8_compute(load_device) # TODO: if we support more ops this needs to be more granular + if model_config and hasattr(model_config, 'layer_quant_config') and model_config.layer_quant_config: logging.info(f"Using mixed precision operations: {len(model_config.layer_quant_config)} quantized layers") - return mixed_precision_ops(model_config.layer_quant_config, compute_dtype) + return mixed_precision_ops(model_config.layer_quant_config, compute_dtype, full_precision_mm=not fp8_compute) - fp8_compute = comfy.model_management.supports_fp8_compute(load_device) if scaled_fp8 is not None: return scaled_fp8_ops(fp8_matrix_mult=fp8_compute and fp8_optimizations, scale_input=fp8_optimizations, override_dtype=scaled_fp8) From 015a0599d08f1072155b9213d488b73e502fea3c Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 25 Nov 2025 00:23:19 -0800 Subject: [PATCH 155/647] I found a case where this is needed (#10875) --- comfy/quant_ops.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/comfy/quant_ops.py b/comfy/quant_ops.py index e938144a7..0c16bcf8d 100644 --- a/comfy/quant_ops.py +++ b/comfy/quant_ops.py @@ -405,8 +405,8 @@ class TensorCoreFP8Layout(QuantizedLayout): tensor_scaled = tensor * (1.0 / scale).to(tensor.dtype) # TODO: uncomment this if it's actually needed because the clamp has a small performance penality' - # lp_amax = torch.finfo(dtype).max - # torch.clamp(tensor_scaled, min=-lp_amax, max=lp_amax, out=tensor_scaled) + lp_amax = torch.finfo(dtype).max + torch.clamp(tensor_scaled, min=-lp_amax, max=lp_amax, out=tensor_scaled) qdata = tensor_scaled.to(dtype, memory_format=torch.contiguous_format) layout_params = { From 6b573ae0cb11000a0330a35d9e31917c22c874a4 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 25 Nov 2025 07:50:19 -0800 Subject: [PATCH 156/647] Flux 2 (#10879) --- comfy/latent_formats.py | 9 +++ comfy/ldm/flux/layers.py | 90 +++++++++++++++++++-------- comfy/ldm/flux/model.py | 80 ++++++++++++++++++------ comfy/ldm/models/autoencoder.py | 42 +++++++++++++ comfy/model_base.py | 23 +++++-- comfy/model_detection.py | 50 +++++++++++---- comfy/sd.py | 26 +++++++- comfy/supported_models.py | 34 +++++++++- comfy/text_encoders/flux.py | 107 +++++++++++++++++++++++++++++++- comfy/text_encoders/llama.py | 31 +++++++++ comfy_extras/nodes_flux.py | 80 +++++++++++++++++++++++- nodes.py | 2 +- 12 files changed, 506 insertions(+), 68 deletions(-) diff --git a/comfy/latent_formats.py b/comfy/latent_formats.py index 204fc048d..e98c7d6d8 100644 --- a/comfy/latent_formats.py +++ b/comfy/latent_formats.py @@ -178,6 +178,15 @@ class Flux(SD3): def process_out(self, latent): return (latent / self.scale_factor) + self.shift_factor +class Flux2(LatentFormat): + latent_channels = 128 + + def process_in(self, latent): + return latent + + def process_out(self, latent): + return latent + class Mochi(LatentFormat): latent_channels = 12 latent_dimensions = 3 diff --git a/comfy/ldm/flux/layers.py b/comfy/ldm/flux/layers.py index 23150a712..2472ab79c 100644 --- a/comfy/ldm/flux/layers.py +++ b/comfy/ldm/flux/layers.py @@ -48,11 +48,11 @@ def timestep_embedding(t: Tensor, dim, max_period=10000, time_factor: float = 10 return embedding class MLPEmbedder(nn.Module): - def __init__(self, in_dim: int, hidden_dim: int, dtype=None, device=None, operations=None): + def __init__(self, in_dim: int, hidden_dim: int, bias=True, dtype=None, device=None, operations=None): super().__init__() - self.in_layer = operations.Linear(in_dim, hidden_dim, bias=True, dtype=dtype, device=device) + self.in_layer = operations.Linear(in_dim, hidden_dim, bias=bias, dtype=dtype, device=device) self.silu = nn.SiLU() - self.out_layer = operations.Linear(hidden_dim, hidden_dim, bias=True, dtype=dtype, device=device) + self.out_layer = operations.Linear(hidden_dim, hidden_dim, bias=bias, dtype=dtype, device=device) def forward(self, x: Tensor) -> Tensor: return self.out_layer(self.silu(self.in_layer(x))) @@ -80,14 +80,14 @@ class QKNorm(torch.nn.Module): class SelfAttention(nn.Module): - def __init__(self, dim: int, num_heads: int = 8, qkv_bias: bool = False, dtype=None, device=None, operations=None): + def __init__(self, dim: int, num_heads: int = 8, qkv_bias: bool = False, proj_bias: bool = True, dtype=None, device=None, operations=None): super().__init__() self.num_heads = num_heads head_dim = dim // num_heads self.qkv = operations.Linear(dim, dim * 3, bias=qkv_bias, dtype=dtype, device=device) self.norm = QKNorm(head_dim, dtype=dtype, device=device, operations=operations) - self.proj = operations.Linear(dim, dim, dtype=dtype, device=device) + self.proj = operations.Linear(dim, dim, bias=proj_bias, dtype=dtype, device=device) @dataclass @@ -98,11 +98,11 @@ class ModulationOut: class Modulation(nn.Module): - def __init__(self, dim: int, double: bool, dtype=None, device=None, operations=None): + def __init__(self, dim: int, double: bool, bias=True, dtype=None, device=None, operations=None): super().__init__() self.is_double = double self.multiplier = 6 if double else 3 - self.lin = operations.Linear(dim, self.multiplier * dim, bias=True, dtype=dtype, device=device) + self.lin = operations.Linear(dim, self.multiplier * dim, bias=bias, dtype=dtype, device=device) def forward(self, vec: Tensor) -> tuple: if vec.ndim == 2: @@ -129,8 +129,18 @@ def apply_mod(tensor, m_mult, m_add=None, modulation_dims=None): return tensor +class SiLUActivation(nn.Module): + def __init__(self): + super().__init__() + self.gate_fn = nn.SiLU() + + def forward(self, x: Tensor) -> Tensor: + x1, x2 = x.chunk(2, dim=-1) + return self.gate_fn(x1) * x2 + + class DoubleStreamBlock(nn.Module): - def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False, flipped_img_txt=False, modulation=True, dtype=None, device=None, operations=None): + def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False, flipped_img_txt=False, modulation=True, mlp_silu_act=False, proj_bias=True, dtype=None, device=None, operations=None): super().__init__() mlp_hidden_dim = int(hidden_size * mlp_ratio) @@ -142,27 +152,44 @@ class DoubleStreamBlock(nn.Module): self.img_mod = Modulation(hidden_size, double=True, dtype=dtype, device=device, operations=operations) self.img_norm1 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) - self.img_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias, dtype=dtype, device=device, operations=operations) + self.img_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias, proj_bias=proj_bias, dtype=dtype, device=device, operations=operations) self.img_norm2 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) - self.img_mlp = nn.Sequential( - operations.Linear(hidden_size, mlp_hidden_dim, bias=True, dtype=dtype, device=device), - nn.GELU(approximate="tanh"), - operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device), - ) + + if mlp_silu_act: + self.img_mlp = nn.Sequential( + operations.Linear(hidden_size, mlp_hidden_dim * 2, bias=False, dtype=dtype, device=device), + SiLUActivation(), + operations.Linear(mlp_hidden_dim, hidden_size, bias=False, dtype=dtype, device=device), + ) + else: + self.img_mlp = nn.Sequential( + operations.Linear(hidden_size, mlp_hidden_dim, bias=True, dtype=dtype, device=device), + nn.GELU(approximate="tanh"), + operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device), + ) if self.modulation: self.txt_mod = Modulation(hidden_size, double=True, dtype=dtype, device=device, operations=operations) self.txt_norm1 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) - self.txt_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias, dtype=dtype, device=device, operations=operations) + self.txt_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias, proj_bias=proj_bias, dtype=dtype, device=device, operations=operations) self.txt_norm2 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) - self.txt_mlp = nn.Sequential( - operations.Linear(hidden_size, mlp_hidden_dim, bias=True, dtype=dtype, device=device), - nn.GELU(approximate="tanh"), - operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device), - ) + + if mlp_silu_act: + self.txt_mlp = nn.Sequential( + operations.Linear(hidden_size, mlp_hidden_dim * 2, bias=False, dtype=dtype, device=device), + SiLUActivation(), + operations.Linear(mlp_hidden_dim, hidden_size, bias=False, dtype=dtype, device=device), + ) + else: + self.txt_mlp = nn.Sequential( + operations.Linear(hidden_size, mlp_hidden_dim, bias=True, dtype=dtype, device=device), + nn.GELU(approximate="tanh"), + operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device), + ) + self.flipped_img_txt = flipped_img_txt def forward(self, img: Tensor, txt: Tensor, vec: Tensor, pe: Tensor, attn_mask=None, modulation_dims_img=None, modulation_dims_txt=None, transformer_options={}): @@ -246,6 +273,8 @@ class SingleStreamBlock(nn.Module): mlp_ratio: float = 4.0, qk_scale: float = None, modulation=True, + mlp_silu_act=False, + bias=True, dtype=None, device=None, operations=None @@ -257,17 +286,24 @@ class SingleStreamBlock(nn.Module): self.scale = qk_scale or head_dim**-0.5 self.mlp_hidden_dim = int(hidden_size * mlp_ratio) + + self.mlp_hidden_dim_first = self.mlp_hidden_dim + if mlp_silu_act: + self.mlp_hidden_dim_first = int(hidden_size * mlp_ratio * 2) + self.mlp_act = SiLUActivation() + else: + self.mlp_act = nn.GELU(approximate="tanh") + # qkv and mlp_in - self.linear1 = operations.Linear(hidden_size, hidden_size * 3 + self.mlp_hidden_dim, dtype=dtype, device=device) + self.linear1 = operations.Linear(hidden_size, hidden_size * 3 + self.mlp_hidden_dim_first, bias=bias, dtype=dtype, device=device) # proj and mlp_out - self.linear2 = operations.Linear(hidden_size + self.mlp_hidden_dim, hidden_size, dtype=dtype, device=device) + self.linear2 = operations.Linear(hidden_size + self.mlp_hidden_dim, hidden_size, bias=bias, dtype=dtype, device=device) self.norm = QKNorm(head_dim, dtype=dtype, device=device, operations=operations) self.hidden_size = hidden_size self.pre_norm = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) - self.mlp_act = nn.GELU(approximate="tanh") if modulation: self.modulation = Modulation(hidden_size, double=False, dtype=dtype, device=device, operations=operations) else: @@ -279,7 +315,7 @@ class SingleStreamBlock(nn.Module): else: mod = vec - qkv, mlp = torch.split(self.linear1(apply_mod(self.pre_norm(x), (1 + mod.scale), mod.shift, modulation_dims)), [3 * self.hidden_size, self.mlp_hidden_dim], dim=-1) + qkv, mlp = torch.split(self.linear1(apply_mod(self.pre_norm(x), (1 + mod.scale), mod.shift, modulation_dims)), [3 * self.hidden_size, self.mlp_hidden_dim_first], dim=-1) q, k, v = qkv.view(qkv.shape[0], qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4) del qkv @@ -298,11 +334,11 @@ class SingleStreamBlock(nn.Module): class LastLayer(nn.Module): - def __init__(self, hidden_size: int, patch_size: int, out_channels: int, dtype=None, device=None, operations=None): + def __init__(self, hidden_size: int, patch_size: int, out_channels: int, bias=True, dtype=None, device=None, operations=None): super().__init__() self.norm_final = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) - self.linear = operations.Linear(hidden_size, patch_size * patch_size * out_channels, bias=True, dtype=dtype, device=device) - self.adaLN_modulation = nn.Sequential(nn.SiLU(), operations.Linear(hidden_size, 2 * hidden_size, bias=True, dtype=dtype, device=device)) + self.linear = operations.Linear(hidden_size, patch_size * patch_size * out_channels, bias=bias, dtype=dtype, device=device) + self.adaLN_modulation = nn.Sequential(nn.SiLU(), operations.Linear(hidden_size, 2 * hidden_size, bias=bias, dtype=dtype, device=device)) def forward(self, x: Tensor, vec: Tensor, modulation_dims=None) -> Tensor: if vec.ndim == 2: diff --git a/comfy/ldm/flux/model.py b/comfy/ldm/flux/model.py index b9d36f202..1a24e6d95 100644 --- a/comfy/ldm/flux/model.py +++ b/comfy/ldm/flux/model.py @@ -15,6 +15,7 @@ from .layers import ( MLPEmbedder, SingleStreamBlock, timestep_embedding, + Modulation ) @dataclass @@ -33,6 +34,11 @@ class FluxParams: patch_size: int qkv_bias: bool guidance_embed: bool + global_modulation: bool = False + mlp_silu_act: bool = False + ops_bias: bool = True + default_ref_method: str = "offset" + ref_index_scale: float = 1.0 class Flux(nn.Module): @@ -58,13 +64,17 @@ class Flux(nn.Module): self.hidden_size = params.hidden_size self.num_heads = params.num_heads self.pe_embedder = EmbedND(dim=pe_dim, theta=params.theta, axes_dim=params.axes_dim) - self.img_in = operations.Linear(self.in_channels, self.hidden_size, bias=True, dtype=dtype, device=device) - self.time_in = MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size, dtype=dtype, device=device, operations=operations) - self.vector_in = MLPEmbedder(params.vec_in_dim, self.hidden_size, dtype=dtype, device=device, operations=operations) + self.img_in = operations.Linear(self.in_channels, self.hidden_size, bias=params.ops_bias, dtype=dtype, device=device) + self.time_in = MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size, bias=params.ops_bias, dtype=dtype, device=device, operations=operations) + if params.vec_in_dim is not None: + self.vector_in = MLPEmbedder(params.vec_in_dim, self.hidden_size, dtype=dtype, device=device, operations=operations) + else: + self.vector_in = None + self.guidance_in = ( - MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size, dtype=dtype, device=device, operations=operations) if params.guidance_embed else nn.Identity() + MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size, bias=params.ops_bias, dtype=dtype, device=device, operations=operations) if params.guidance_embed else nn.Identity() ) - self.txt_in = operations.Linear(params.context_in_dim, self.hidden_size, dtype=dtype, device=device) + self.txt_in = operations.Linear(params.context_in_dim, self.hidden_size, bias=params.ops_bias, dtype=dtype, device=device) self.double_blocks = nn.ModuleList( [ @@ -73,6 +83,9 @@ class Flux(nn.Module): self.num_heads, mlp_ratio=params.mlp_ratio, qkv_bias=params.qkv_bias, + modulation=params.global_modulation is False, + mlp_silu_act=params.mlp_silu_act, + proj_bias=params.ops_bias, dtype=dtype, device=device, operations=operations ) for _ in range(params.depth) @@ -81,13 +94,30 @@ class Flux(nn.Module): self.single_blocks = nn.ModuleList( [ - SingleStreamBlock(self.hidden_size, self.num_heads, mlp_ratio=params.mlp_ratio, dtype=dtype, device=device, operations=operations) + SingleStreamBlock(self.hidden_size, self.num_heads, mlp_ratio=params.mlp_ratio, modulation=params.global_modulation is False, mlp_silu_act=params.mlp_silu_act, bias=params.ops_bias, dtype=dtype, device=device, operations=operations) for _ in range(params.depth_single_blocks) ] ) if final_layer: - self.final_layer = LastLayer(self.hidden_size, 1, self.out_channels, dtype=dtype, device=device, operations=operations) + self.final_layer = LastLayer(self.hidden_size, 1, self.out_channels, bias=params.ops_bias, dtype=dtype, device=device, operations=operations) + + if params.global_modulation: + self.double_stream_modulation_img = Modulation( + self.hidden_size, + double=True, + bias=False, + dtype=dtype, device=device, operations=operations + ) + self.double_stream_modulation_txt = Modulation( + self.hidden_size, + double=True, + bias=False, + dtype=dtype, device=device, operations=operations + ) + self.single_stream_modulation = Modulation( + self.hidden_size, double=False, bias=False, dtype=dtype, device=device, operations=operations + ) def forward_orig( self, @@ -103,9 +133,6 @@ class Flux(nn.Module): attn_mask: Tensor = None, ) -> Tensor: - if y is None: - y = torch.zeros((img.shape[0], self.params.vec_in_dim), device=img.device, dtype=img.dtype) - patches = transformer_options.get("patches", {}) patches_replace = transformer_options.get("patches_replace", {}) if img.ndim != 3 or txt.ndim != 3: @@ -118,9 +145,17 @@ class Flux(nn.Module): if guidance is not None: vec = vec + self.guidance_in(timestep_embedding(guidance, 256).to(img.dtype)) - vec = vec + self.vector_in(y[:, :self.params.vec_in_dim]) + if self.vector_in is not None: + if y is None: + y = torch.zeros((img.shape[0], self.params.vec_in_dim), device=img.device, dtype=img.dtype) + vec = vec + self.vector_in(y[:, :self.params.vec_in_dim]) + txt = self.txt_in(txt) + vec_orig = vec + if self.params.global_modulation: + vec = (self.double_stream_modulation_img(vec_orig), self.double_stream_modulation_txt(vec_orig)) + if "post_input" in patches: for p in patches["post_input"]: out = p({"img": img, "txt": txt, "img_ids": img_ids, "txt_ids": txt_ids}) @@ -177,6 +212,9 @@ class Flux(nn.Module): img = torch.cat((txt, img), 1) + if self.params.global_modulation: + vec, _ = self.single_stream_modulation(vec_orig) + for i, block in enumerate(self.single_blocks): if ("single_block", i) in blocks_replace: def block_wrap(args): @@ -207,7 +245,7 @@ class Flux(nn.Module): img = img[:, txt.shape[1] :, ...] - img = self.final_layer(img, vec) # (N, T, patch_size ** 2 * out_channels) + img = self.final_layer(img, vec_orig) # (N, T, patch_size ** 2 * out_channels) return img def process_img(self, x, index=0, h_offset=0, w_offset=0, transformer_options={}): @@ -234,10 +272,10 @@ class Flux(nn.Module): h_offset += rope_options.get("shift_y", 0.0) w_offset += rope_options.get("shift_x", 0.0) - img_ids = torch.zeros((steps_h, steps_w, 3), device=x.device, dtype=x.dtype) + img_ids = torch.zeros((steps_h, steps_w, len(self.params.axes_dim)), device=x.device, dtype=torch.float32) img_ids[:, :, 0] = img_ids[:, :, 1] + index - img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(h_offset, h_len - 1 + h_offset, steps=steps_h, device=x.device, dtype=x.dtype).unsqueeze(1) - img_ids[:, :, 2] = img_ids[:, :, 2] + torch.linspace(w_offset, w_len - 1 + w_offset, steps=steps_w, device=x.device, dtype=x.dtype).unsqueeze(0) + img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(h_offset, h_len - 1 + h_offset, steps=steps_h, device=x.device, dtype=torch.float32).unsqueeze(1) + img_ids[:, :, 2] = img_ids[:, :, 2] + torch.linspace(w_offset, w_len - 1 + w_offset, steps=steps_w, device=x.device, dtype=torch.float32).unsqueeze(0) return img, repeat(img_ids, "h w c -> b (h w) c", b=bs) def forward(self, x, timestep, context, y=None, guidance=None, ref_latents=None, control=None, transformer_options={}, **kwargs): @@ -259,10 +297,10 @@ class Flux(nn.Module): h = 0 w = 0 index = 0 - ref_latents_method = kwargs.get("ref_latents_method", "offset") + ref_latents_method = kwargs.get("ref_latents_method", self.params.default_ref_method) for ref in ref_latents: if ref_latents_method == "index": - index += 1 + index += self.params.ref_index_scale h_offset = 0 w_offset = 0 elif ref_latents_method == "uxo": @@ -286,7 +324,11 @@ class Flux(nn.Module): img = torch.cat([img, kontext], dim=1) img_ids = torch.cat([img_ids, kontext_ids], dim=1) - txt_ids = torch.zeros((bs, context.shape[1], 3), device=x.device, dtype=x.dtype) + txt_ids = torch.zeros((bs, context.shape[1], len(self.params.axes_dim)), device=x.device, dtype=torch.float32) + + if len(self.params.axes_dim) == 4: # Flux 2 + txt_ids[:, :, 3] = torch.linspace(0, context.shape[1] - 1, steps=context.shape[1], device=x.device, dtype=torch.float32) + out = self.forward_orig(img, img_ids, context, txt_ids, timestep, y, guidance, control, transformer_options, attn_mask=kwargs.get("attention_mask", None)) out = out[:, :img_tokens] - return rearrange(out, "b (h w) (c ph pw) -> b c (h ph) (w pw)", h=h_len, w=w_len, ph=2, pw=2)[:,:,:h_orig,:w_orig] + return rearrange(out, "b (h w) (c ph pw) -> b c (h ph) (w pw)", h=h_len, w=w_len, ph=self.patch_size, pw=self.patch_size)[:,:,:h_orig,:w_orig] diff --git a/comfy/ldm/models/autoencoder.py b/comfy/ldm/models/autoencoder.py index 611d36a1b..4f50810dc 100644 --- a/comfy/ldm/models/autoencoder.py +++ b/comfy/ldm/models/autoencoder.py @@ -9,6 +9,8 @@ from comfy.ldm.modules.distributions.distributions import DiagonalGaussianDistri from comfy.ldm.util import get_obj_from_str, instantiate_from_config from comfy.ldm.modules.ema import LitEma import comfy.ops +from einops import rearrange +import comfy.model_management class DiagonalGaussianRegularizer(torch.nn.Module): def __init__(self, sample: bool = False): @@ -179,6 +181,21 @@ class AutoencodingEngineLegacy(AutoencodingEngine): self.post_quant_conv = conv_op(embed_dim, ddconfig["z_channels"], 1) self.embed_dim = embed_dim + if ddconfig.get("batch_norm_latent", False): + self.bn_eps = 1e-4 + self.bn_momentum = 0.1 + self.ps = [2, 2] + self.bn = torch.nn.BatchNorm2d(math.prod(self.ps) * ddconfig["z_channels"], + eps=self.bn_eps, + momentum=self.bn_momentum, + affine=False, + track_running_stats=True, + ) + self.bn.eval() + else: + self.bn = None + + def get_autoencoder_params(self) -> list: params = super().get_autoencoder_params() return params @@ -201,11 +218,36 @@ class AutoencodingEngineLegacy(AutoencodingEngine): z = torch.cat(z, 0) z, reg_log = self.regularization(z) + + if self.bn is not None: + z = rearrange(z, + "... c (i pi) (j pj) -> ... (c pi pj) i j", + pi=self.ps[0], + pj=self.ps[1], + ) + + z = torch.nn.functional.batch_norm(z, + comfy.model_management.cast_to(self.bn.running_mean, dtype=z.dtype, device=z.device), + comfy.model_management.cast_to(self.bn.running_var, dtype=z.dtype, device=z.device), + momentum=self.bn_momentum, + eps=self.bn_eps) + if return_reg_log: return z, reg_log return z def decode(self, z: torch.Tensor, **decoder_kwargs) -> torch.Tensor: + if self.bn is not None: + s = torch.sqrt(comfy.model_management.cast_to(self.bn.running_var.view(1, -1, 1, 1), dtype=z.dtype, device=z.device) + self.bn_eps) + m = comfy.model_management.cast_to(self.bn.running_mean.view(1, -1, 1, 1), dtype=z.dtype, device=z.device) + z = z * s + m + z = rearrange( + z, + "... (c pi pj) i j -> ... c (i pi) (j pj)", + pi=self.ps[0], + pj=self.ps[1], + ) + if self.max_batch_size is None: dec = self.post_quant_conv(z) dec = self.decoder(dec, **decoder_kwargs) diff --git a/comfy/model_base.py b/comfy/model_base.py index e14b552c5..cad79ecbd 100644 --- a/comfy/model_base.py +++ b/comfy/model_base.py @@ -898,12 +898,13 @@ class Flux(BaseModel): attention_mask = kwargs.get("attention_mask", None) if attention_mask is not None: shape = kwargs["noise"].shape - mask_ref_size = kwargs["attention_mask_img_shape"] - # the model will pad to the patch size, and then divide - # essentially dividing and rounding up - (h_tok, w_tok) = (math.ceil(shape[2] / self.diffusion_model.patch_size), math.ceil(shape[3] / self.diffusion_model.patch_size)) - attention_mask = utils.upscale_dit_mask(attention_mask, mask_ref_size, (h_tok, w_tok)) - out['attention_mask'] = comfy.conds.CONDRegular(attention_mask) + mask_ref_size = kwargs.get("attention_mask_img_shape", None) + if mask_ref_size is not None: + # the model will pad to the patch size, and then divide + # essentially dividing and rounding up + (h_tok, w_tok) = (math.ceil(shape[2] / self.diffusion_model.patch_size), math.ceil(shape[3] / self.diffusion_model.patch_size)) + attention_mask = utils.upscale_dit_mask(attention_mask, mask_ref_size, (h_tok, w_tok)) + out['attention_mask'] = comfy.conds.CONDRegular(attention_mask) guidance = kwargs.get("guidance", 3.5) if guidance is not None: @@ -928,6 +929,16 @@ class Flux(BaseModel): out['ref_latents'] = list([1, 16, sum(map(lambda a: math.prod(a.size()), ref_latents)) // 16]) return out +class Flux2(Flux): + def extra_conds(self, **kwargs): + out = super().extra_conds(**kwargs) + cross_attn = kwargs.get("cross_attn", None) + if cross_attn is not None: + target_text_len = 512 + if cross_attn.shape[1] < target_text_len: + cross_attn = torch.nn.functional.pad(cross_attn, (0, 0, target_text_len - cross_attn.shape[1], 0)) + out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn) + return out class GenmoMochi(BaseModel): def __init__(self, model_config, model_type=ModelType.FLOW, device=None): diff --git a/comfy/model_detection.py b/comfy/model_detection.py index 0131ca25a..b2ba1459d 100644 --- a/comfy/model_detection.py +++ b/comfy/model_detection.py @@ -200,26 +200,54 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): if '{}double_blocks.0.img_attn.norm.key_norm.scale'.format(key_prefix) in state_dict_keys and ('{}img_in.weight'.format(key_prefix) in state_dict_keys or f"{key_prefix}distilled_guidance_layer.norms.0.scale" in state_dict_keys): #Flux, Chroma or Chroma Radiance (has no img_in.weight) dit_config = {} - dit_config["image_model"] = "flux" + if '{}double_stream_modulation_img.lin.weight'.format(key_prefix) in state_dict_keys: + dit_config["image_model"] = "flux2" + dit_config["axes_dim"] = [32, 32, 32, 32] + dit_config["num_heads"] = 48 + dit_config["mlp_ratio"] = 3.0 + dit_config["theta"] = 2000 + dit_config["out_channels"] = 128 + dit_config["global_modulation"] = True + dit_config["vec_in_dim"] = None + dit_config["mlp_silu_act"] = True + dit_config["qkv_bias"] = False + dit_config["ops_bias"] = False + dit_config["default_ref_method"] = "index" + dit_config["ref_index_scale"] = 10.0 + patch_size = 1 + else: + dit_config["image_model"] = "flux" + dit_config["axes_dim"] = [16, 56, 56] + dit_config["num_heads"] = 24 + dit_config["mlp_ratio"] = 4.0 + dit_config["theta"] = 10000 + dit_config["out_channels"] = 16 + dit_config["qkv_bias"] = True + patch_size = 2 + dit_config["in_channels"] = 16 - patch_size = 2 + dit_config["hidden_size"] = 3072 + dit_config["context_in_dim"] = 4096 + dit_config["patch_size"] = patch_size in_key = "{}img_in.weight".format(key_prefix) if in_key in state_dict_keys: - dit_config["in_channels"] = state_dict[in_key].shape[1] // (patch_size * patch_size) - dit_config["out_channels"] = 16 + w = state_dict[in_key] + dit_config["in_channels"] = w.shape[1] // (patch_size * patch_size) + dit_config["hidden_size"] = w.shape[0] + + txt_in_key = "{}txt_in.weight".format(key_prefix) + if txt_in_key in state_dict_keys: + w = state_dict[txt_in_key] + dit_config["context_in_dim"] = w.shape[1] + dit_config["hidden_size"] = w.shape[0] + vec_in_key = '{}vector_in.in_layer.weight'.format(key_prefix) if vec_in_key in state_dict_keys: dit_config["vec_in_dim"] = state_dict[vec_in_key].shape[1] - dit_config["context_in_dim"] = 4096 - dit_config["hidden_size"] = 3072 - dit_config["mlp_ratio"] = 4.0 - dit_config["num_heads"] = 24 + dit_config["depth"] = count_blocks(state_dict_keys, '{}double_blocks.'.format(key_prefix) + '{}.') dit_config["depth_single_blocks"] = count_blocks(state_dict_keys, '{}single_blocks.'.format(key_prefix) + '{}.') - dit_config["axes_dim"] = [16, 56, 56] - dit_config["theta"] = 10000 - dit_config["qkv_bias"] = True if '{}distilled_guidance_layer.0.norms.0.scale'.format(key_prefix) in state_dict_keys or '{}distilled_guidance_layer.norms.0.scale'.format(key_prefix) in state_dict_keys: #Chroma dit_config["image_model"] = "chroma" dit_config["in_channels"] = 64 diff --git a/comfy/sd.py b/comfy/sd.py index b6df0bd61..14dd8944c 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -356,7 +356,7 @@ class VAE: self.memory_used_encode = lambda shape, dtype: (700 * shape[2] * shape[3]) * model_management.dtype_size(dtype) self.memory_used_decode = lambda shape, dtype: (700 * shape[2] * shape[3] * 32 * 32) * model_management.dtype_size(dtype) - elif sd['decoder.conv_in.weight'].shape[1] == 32: + elif sd['decoder.conv_in.weight'].shape[1] == 32 and sd['decoder.conv_in.weight'].ndim == 5: ddconfig = {"block_out_channels": [128, 256, 512, 1024, 1024], "in_channels": 3, "out_channels": 3, "num_res_blocks": 2, "ffactor_spatial": 16, "ffactor_temporal": 4, "downsample_match_channel": True, "upsample_match_channel": True, "refiner_vae": False} self.latent_channels = ddconfig['z_channels'] = sd["decoder.conv_in.weight"].shape[1] self.working_dtypes = [torch.float16, torch.bfloat16, torch.float32] @@ -382,6 +382,17 @@ class VAE: self.upscale_ratio = 4 self.latent_channels = ddconfig['z_channels'] = sd["decoder.conv_in.weight"].shape[1] + if 'decoder.post_quant_conv.weight' in sd: + sd = comfy.utils.state_dict_prefix_replace(sd, {"decoder.post_quant_conv.": "post_quant_conv.", "encoder.quant_conv.": "quant_conv."}) + + if 'bn.running_mean' in sd: + ddconfig["batch_norm_latent"] = True + self.downscale_ratio *= 2 + self.upscale_ratio *= 2 + self.latent_channels *= 4 + old_memory_used_decode = self.memory_used_decode + self.memory_used_decode = lambda shape, dtype: old_memory_used_decode(shape, dtype) * 4.0 + if 'post_quant_conv.weight' in sd: self.first_stage_model = AutoencoderKL(ddconfig=ddconfig, embed_dim=sd['post_quant_conv.weight'].shape[1]) else: @@ -940,6 +951,8 @@ class TEModel(Enum): QWEN25_7B = 11 BYT5_SMALL_GLYPH = 12 GEMMA_3_4B = 13 + MISTRAL3_24B = 14 + MISTRAL3_24B_PRUNED_FLUX2 = 15 def detect_te_model(sd): if "text_model.encoder.layers.30.mlp.fc1.weight" in sd: @@ -972,6 +985,13 @@ def detect_te_model(sd): if weight.shape[0] == 512: return TEModel.QWEN25_7B if "model.layers.0.post_attention_layernorm.weight" in sd: + weight = sd['model.layers.0.post_attention_layernorm.weight'] + if weight.shape[0] == 5120: + if "model.layers.39.post_attention_layernorm.weight" in sd: + return TEModel.MISTRAL3_24B + else: + return TEModel.MISTRAL3_24B_PRUNED_FLUX2 + return TEModel.LLAMA3_8 return None @@ -1086,6 +1106,10 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip else: clip_target.clip = comfy.text_encoders.qwen_image.te(**llama_detect(clip_data)) clip_target.tokenizer = comfy.text_encoders.qwen_image.QwenImageTokenizer + elif te_model == TEModel.MISTRAL3_24B or te_model == TEModel.MISTRAL3_24B_PRUNED_FLUX2: + clip_target.clip = comfy.text_encoders.flux.flux2_te(**llama_detect(clip_data), pruned=te_model == TEModel.MISTRAL3_24B_PRUNED_FLUX2) + clip_target.tokenizer = comfy.text_encoders.flux.Flux2Tokenizer + tokenizer_data["tekken_model"] = clip_data[0].get("tekken_model", None) else: # clip_l if clip_type == CLIPType.SD3: diff --git a/comfy/supported_models.py b/comfy/supported_models.py index 2e64b85e8..8fe8e63f6 100644 --- a/comfy/supported_models.py +++ b/comfy/supported_models.py @@ -741,6 +741,37 @@ class FluxSchnell(Flux): out = model_base.Flux(self, model_type=model_base.ModelType.FLOW, device=device) return out +class Flux2(Flux): + unet_config = { + "image_model": "flux2", + } + + sampling_settings = { + "shift": 2.02, + } + + unet_extra_config = {} + latent_format = latent_formats.Flux2 + + supported_inference_dtypes = [torch.bfloat16, torch.float16, torch.float32] + + vae_key_prefix = ["vae."] + text_encoder_key_prefix = ["text_encoders."] + + def __init__(self, unet_config): + super().__init__(unet_config) + self.memory_usage_factor = self.memory_usage_factor * (2.0 * 2.0) * 2.36 + + def get_model(self, state_dict, prefix="", device=None): + out = model_base.Flux2(self, device=device) + return out + + def clip_target(self, state_dict={}): + return None # TODO + pref = self.text_encoder_key_prefix[0] + t5_detect = comfy.text_encoders.sd3_clip.t5_xxl_detect(state_dict, "{}t5xxl.transformer.".format(pref)) + return supported_models_base.ClipTarget(comfy.text_encoders.flux.FluxTokenizer, comfy.text_encoders.flux.flux_clip(**t5_detect)) + class GenmoMochi(supported_models_base.BASE): unet_config = { "image_model": "mochi_preview", @@ -1422,6 +1453,7 @@ class HunyuanVideo15_SR_Distilled(HunyuanVideo): hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen25_7b.transformer.".format(pref)) return supported_models_base.ClipTarget(comfy.text_encoders.hunyuan_video.HunyuanVideo15Tokenizer, comfy.text_encoders.hunyuan_image.te(**hunyuan_detect)) -models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanVideo15_SR_Distilled, HunyuanVideo15, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, Omnigen2, QwenImage] +models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanVideo15_SR_Distilled, HunyuanVideo15, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, Omnigen2, QwenImage, Flux2] + models += [SVD_img2vid] diff --git a/comfy/text_encoders/flux.py b/comfy/text_encoders/flux.py index d61ef6668..8dbbca16e 100644 --- a/comfy/text_encoders/flux.py +++ b/comfy/text_encoders/flux.py @@ -1,10 +1,13 @@ from comfy import sd1_clip import comfy.text_encoders.t5 import comfy.text_encoders.sd3_clip +import comfy.text_encoders.llama import comfy.model_management -from transformers import T5TokenizerFast +from transformers import T5TokenizerFast, LlamaTokenizerFast import torch import os +import json +import base64 class T5XXLTokenizer(sd1_clip.SDTokenizer): def __init__(self, embedding_directory=None, tokenizer_data={}): @@ -68,3 +71,105 @@ def flux_clip(dtype_t5=None, t5xxl_scaled_fp8=None): model_options["t5xxl_scaled_fp8"] = t5xxl_scaled_fp8 super().__init__(dtype_t5=dtype_t5, device=device, dtype=dtype, model_options=model_options) return FluxClipModel_ + +def load_mistral_tokenizer(data): + if torch.is_tensor(data): + data = data.numpy().tobytes() + + try: + from transformers.integrations.mistral import MistralConverter + except ModuleNotFoundError: + from transformers.models.pixtral.convert_pixtral_weights_to_hf import MistralConverter + + mistral_vocab = json.loads(data) + + special_tokens = {} + vocab = {} + + max_vocab = mistral_vocab["config"]["default_vocab_size"] + + for w in mistral_vocab["vocab"]: + r = w["rank"] + if r >= max_vocab: + continue + + vocab[base64.b64decode(w["token_bytes"])] = r + + for w in mistral_vocab["special_tokens"]: + if "token_bytes" in w: + special_tokens[base64.b64decode(w["token_bytes"])] = w["rank"] + else: + special_tokens[w["token_str"]] = w["rank"] + + all_special = [] + for v in special_tokens: + all_special.append(v) + + special_tokens.update(vocab) + vocab = special_tokens + return {"tokenizer_object": MistralConverter(vocab=vocab, additional_special_tokens=all_special).converted(), "legacy": False} + +class MistralTokenizerClass: + @staticmethod + def from_pretrained(path, **kwargs): + return LlamaTokenizerFast(**kwargs) + +class Mistral3Tokenizer(sd1_clip.SDTokenizer): + def __init__(self, embedding_directory=None, tokenizer_data={}): + self.tekken_data = tokenizer_data.get("tekken_model", None) + super().__init__("", pad_with_end=False, embedding_size=5120, embedding_key='mistral3_24b', tokenizer_class=MistralTokenizerClass, has_end_token=False, pad_to_max_length=False, pad_token=11, max_length=99999999, min_length=1, pad_left=True, tokenizer_args=load_mistral_tokenizer(self.tekken_data), tokenizer_data=tokenizer_data) + + def state_dict(self): + return {"tekken_model": self.tekken_data} + +class Flux2Tokenizer(sd1_clip.SD1Tokenizer): + def __init__(self, embedding_directory=None, tokenizer_data={}): + super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, name="mistral3_24b", tokenizer=Mistral3Tokenizer) + self.llama_template = '[SYSTEM_PROMPT]You are an AI that reasons about image descriptions. You give structured responses focusing on object relationships, object\nattribution and actions without speculation.[/SYSTEM_PROMPT][INST]{}[/INST]' + + def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None, **kwargs): + if llama_template is None: + llama_text = self.llama_template.format(text) + else: + llama_text = llama_template.format(text) + + tokens = super().tokenize_with_weights(llama_text, return_word_ids=return_word_ids, disable_weights=True, **kwargs) + return tokens + +class Mistral3_24BModel(sd1_clip.SDClipModel): + def __init__(self, device="cpu", layer="all", layer_idx=None, dtype=None, attention_mask=True, model_options={}): + textmodel_json_config = {} + num_layers = model_options.get("num_layers", None) + if num_layers is not None: + textmodel_json_config["num_hidden_layers"] = num_layers + if num_layers < 40: + textmodel_json_config["final_norm"] = False + super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"start": 1, "pad": 0}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Mistral3Small24B, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options) + +class Flux2TEModel(sd1_clip.SD1ClipModel): + def __init__(self, device="cpu", dtype=None, model_options={}, name="mistral3_24b", clip_model=Mistral3_24BModel): + super().__init__(device=device, dtype=dtype, name=name, clip_model=clip_model, model_options=model_options) + + def encode_token_weights(self, token_weight_pairs): + out, pooled, extra = super().encode_token_weights(token_weight_pairs) + + out = torch.stack((out[:, 10], out[:, 20], out[:, 30]), dim=1) + out = out.movedim(1, 2) + out = out.reshape(out.shape[0], out.shape[1], -1) + return out, pooled, extra + +def flux2_te(dtype_llama=None, llama_scaled_fp8=None, llama_quantization_metadata=None, pruned=False): + class Flux2TEModel_(Flux2TEModel): + def __init__(self, device="cpu", dtype=None, model_options={}): + if llama_scaled_fp8 is not None and "scaled_fp8" not in model_options: + model_options = model_options.copy() + model_options["scaled_fp8"] = llama_scaled_fp8 + if dtype_llama is not None: + dtype = dtype_llama + if llama_quantization_metadata is not None: + model_options["quantization_metadata"] = llama_quantization_metadata + if pruned: + model_options = model_options.copy() + model_options["num_layers"] = 30 + super().__init__(device=device, dtype=dtype, model_options=model_options) + return Flux2TEModel_ diff --git a/comfy/text_encoders/llama.py b/comfy/text_encoders/llama.py index feb44bbb0..749ff581b 100644 --- a/comfy/text_encoders/llama.py +++ b/comfy/text_encoders/llama.py @@ -34,6 +34,28 @@ class Llama2Config: rope_scale = None final_norm: bool = True +@dataclass +class Mistral3Small24BConfig: + vocab_size: int = 131072 + hidden_size: int = 5120 + intermediate_size: int = 32768 + num_hidden_layers: int = 40 + num_attention_heads: int = 32 + num_key_value_heads: int = 8 + max_position_embeddings: int = 8192 + rms_norm_eps: float = 1e-5 + rope_theta: float = 1000000000.0 + transformer_type: str = "llama" + head_dim = 128 + rms_norm_add = False + mlp_activation = "silu" + qkv_bias = False + rope_dims = None + q_norm = None + k_norm = None + rope_scale = None + final_norm: bool = True + @dataclass class Qwen25_3BConfig: vocab_size: int = 151936 @@ -465,6 +487,15 @@ class Llama2(BaseLlama, torch.nn.Module): self.model = Llama2_(config, device=device, dtype=dtype, ops=operations) self.dtype = dtype +class Mistral3Small24B(BaseLlama, torch.nn.Module): + def __init__(self, config_dict, dtype, device, operations): + super().__init__() + config = Mistral3Small24BConfig(**config_dict) + self.num_layers = config.num_hidden_layers + + self.model = Llama2_(config, device=device, dtype=dtype, ops=operations) + self.dtype = dtype + class Qwen25_3B(BaseLlama, torch.nn.Module): def __init__(self, config_dict, dtype, device, operations): super().__init__() diff --git a/comfy_extras/nodes_flux.py b/comfy_extras/nodes_flux.py index ce1b2e89f..d9c4bba81 100644 --- a/comfy_extras/nodes_flux.py +++ b/comfy_extras/nodes_flux.py @@ -2,7 +2,10 @@ import node_helpers import comfy.utils from typing_extensions import override from comfy_api.latest import ComfyExtension, io - +import comfy.model_management +import torch +import math +import nodes class CLIPTextEncodeFlux(io.ComfyNode): @classmethod @@ -30,6 +33,27 @@ class CLIPTextEncodeFlux(io.ComfyNode): encode = execute # TODO: remove +class EmptyFlux2LatentImage(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="EmptyFlux2LatentImage", + display_name="Empty Flux 2 Latent", + category="latent", + inputs=[ + io.Int.Input("width", default=1024, min=16, max=nodes.MAX_RESOLUTION, step=16), + io.Int.Input("height", default=1024, min=16, max=nodes.MAX_RESOLUTION, step=16), + io.Int.Input("batch_size", default=1, min=1, max=4096), + ], + outputs=[ + io.Latent.Output(), + ], + ) + + @classmethod + def execute(cls, width, height, batch_size=1) -> io.NodeOutput: + latent = torch.zeros([batch_size, 128, height // 16, width // 16], device=comfy.model_management.intermediate_device()) + return io.NodeOutput({"samples": latent}) class FluxGuidance(io.ComfyNode): @classmethod @@ -154,6 +178,58 @@ class FluxKontextMultiReferenceLatentMethod(io.ComfyNode): append = execute # TODO: remove +def generalized_time_snr_shift(t, mu: float, sigma: float): + return math.exp(mu) / (math.exp(mu) + (1 / t - 1) ** sigma) + + +def compute_empirical_mu(image_seq_len: int, num_steps: int) -> float: + a1, b1 = 8.73809524e-05, 1.89833333 + a2, b2 = 0.00016927, 0.45666666 + + if image_seq_len > 4300: + mu = a2 * image_seq_len + b2 + return float(mu) + + m_200 = a2 * image_seq_len + b2 + m_10 = a1 * image_seq_len + b1 + + a = (m_200 - m_10) / 190.0 + b = m_200 - 200.0 * a + mu = a * num_steps + b + + return float(mu) + + +def get_schedule(num_steps: int, image_seq_len: int) -> list[float]: + mu = compute_empirical_mu(image_seq_len, num_steps) + timesteps = torch.linspace(1, 0, num_steps + 1) + timesteps = generalized_time_snr_shift(timesteps, mu, 1.0) + return timesteps + + +class Flux2Scheduler(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="Flux2Scheduler", + category="sampling/custom_sampling/schedulers", + inputs=[ + io.Int.Input("steps", default=20, min=1, max=4096), + io.Int.Input("width", default=1024, min=16, max=nodes.MAX_RESOLUTION, step=1), + io.Int.Input("height", default=1024, min=16, max=nodes.MAX_RESOLUTION, step=1), + ], + outputs=[ + io.Sigmas.Output(), + ], + ) + + @classmethod + def execute(cls, steps, width, height) -> io.NodeOutput: + seq_len = (width * height / (16 * 16)) + sigmas = get_schedule(steps, round(seq_len)) + return io.NodeOutput(sigmas) + + class FluxExtension(ComfyExtension): @override async def get_node_list(self) -> list[type[io.ComfyNode]]: @@ -163,6 +239,8 @@ class FluxExtension(ComfyExtension): FluxDisableGuidance, FluxKontextImageScale, FluxKontextMultiReferenceLatentMethod, + EmptyFlux2LatentImage, + Flux2Scheduler, ] diff --git a/nodes.py b/nodes.py index f023ae3b6..f4835c02e 100644 --- a/nodes.py +++ b/nodes.py @@ -929,7 +929,7 @@ class CLIPLoader: @classmethod def INPUT_TYPES(s): return {"required": { "clip_name": (folder_paths.get_filename_list("text_encoders"), ), - "type": (["stable_diffusion", "stable_cascade", "sd3", "stable_audio", "mochi", "ltxv", "pixart", "cosmos", "lumina2", "wan", "hidream", "chroma", "ace", "omnigen2", "qwen_image", "hunyuan_image"], ), + "type": (["stable_diffusion", "stable_cascade", "sd3", "stable_audio", "mochi", "ltxv", "pixart", "cosmos", "lumina2", "wan", "hidream", "chroma", "ace", "omnigen2", "qwen_image", "hunyuan_image", "flux2"], ), }, "optional": { "device": (["default", "cpu"], {"advanced": True}), From 5c7b08ca58f5412b3a814b374793cacdb5b5f0a7 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Tue, 25 Nov 2025 18:09:07 +0200 Subject: [PATCH 157/647] [API Nodes] add Flux.2 Pro node (#10880) --- comfy_api_nodes/apis/bfl_api.py | 28 +++- comfy_api_nodes/nodes_bfl.py | 238 +++++++++++++++---------------- comfy_api_nodes/util/__init__.py | 2 + 3 files changed, 143 insertions(+), 125 deletions(-) diff --git a/comfy_api_nodes/apis/bfl_api.py b/comfy_api_nodes/apis/bfl_api.py index 0fc8c0607..d8d3557b3 100644 --- a/comfy_api_nodes/apis/bfl_api.py +++ b/comfy_api_nodes/apis/bfl_api.py @@ -70,6 +70,29 @@ class BFLFluxProGenerateRequest(BaseModel): # ) +class Flux2ProGenerateRequest(BaseModel): + prompt: str = Field(...) + width: int = Field(1024, description="Must be a multiple of 32.") + height: int = Field(768, description="Must be a multiple of 32.") + seed: int | None = Field(None) + prompt_upsampling: bool | None = Field(None) + input_image: str | None = Field(None, description="Base64 encoded image for image-to-image generation") + input_image_2: str | None = Field(None, description="Base64 encoded image for image-to-image generation") + input_image_3: str | None = Field(None, description="Base64 encoded image for image-to-image generation") + input_image_4: str | None = Field(None, description="Base64 encoded image for image-to-image generation") + input_image_5: str | None = Field(None, description="Base64 encoded image for image-to-image generation") + input_image_6: str | None = Field(None, description="Base64 encoded image for image-to-image generation") + input_image_7: str | None = Field(None, description="Base64 encoded image for image-to-image generation") + input_image_8: str | None = Field(None, description="Base64 encoded image for image-to-image generation") + input_image_9: str | None = Field(None, description="Base64 encoded image for image-to-image generation") + safety_tolerance: int | None = Field( + 5, description="Tolerance level for input and output moderation. Value 0 being most strict.", ge=0, le=5 + ) + output_format: str | None = Field( + "png", description="Output format for the generated image. Can be 'jpeg' or 'png'." + ) + + class BFLFluxKontextProGenerateRequest(BaseModel): prompt: str = Field(..., description='The text prompt for what you wannt to edit.') input_image: Optional[str] = Field(None, description='Image to edit in base64 format') @@ -109,8 +132,9 @@ class BFLFluxProUltraGenerateRequest(BaseModel): class BFLFluxProGenerateResponse(BaseModel): - id: str = Field(..., description='The unique identifier for the generation task.') - polling_url: str = Field(..., description='URL to poll for the generation result.') + id: str = Field(..., description="The unique identifier for the generation task.") + polling_url: str = Field(..., description="URL to poll for the generation result.") + cost: float | None = Field(None, description="Price in cents") class BFLStatus(str, Enum): diff --git a/comfy_api_nodes/nodes_bfl.py b/comfy_api_nodes/nodes_bfl.py index 1740fb377..8826dea0c 100644 --- a/comfy_api_nodes/nodes_bfl.py +++ b/comfy_api_nodes/nodes_bfl.py @@ -1,7 +1,7 @@ from inspect import cleandoc -from typing import Optional import torch +from pydantic import BaseModel from typing_extensions import override from comfy_api.latest import IO, ComfyExtension @@ -9,15 +9,16 @@ from comfy_api_nodes.apis.bfl_api import ( BFLFluxExpandImageRequest, BFLFluxFillImageRequest, BFLFluxKontextProGenerateRequest, - BFLFluxProGenerateRequest, BFLFluxProGenerateResponse, BFLFluxProUltraGenerateRequest, BFLFluxStatusResponse, BFLStatus, + Flux2ProGenerateRequest, ) from comfy_api_nodes.util import ( ApiEndpoint, download_url_to_image_tensor, + get_number_of_images, poll_op, resize_mask_to_image, sync_op, @@ -116,7 +117,7 @@ class FluxProUltraImageNode(IO.ComfyNode): prompt_upsampling: bool = False, raw: bool = False, seed: int = 0, - image_prompt: Optional[torch.Tensor] = None, + image_prompt: torch.Tensor | None = None, image_prompt_strength: float = 0.1, ) -> IO.NodeOutput: if image_prompt is None: @@ -230,7 +231,7 @@ class FluxKontextProImageNode(IO.ComfyNode): aspect_ratio: str, guidance: float, steps: int, - input_image: Optional[torch.Tensor] = None, + input_image: torch.Tensor | None = None, seed=0, prompt_upsampling=False, ) -> IO.NodeOutput: @@ -280,124 +281,6 @@ class FluxKontextMaxImageNode(FluxKontextProImageNode): DISPLAY_NAME = "Flux.1 Kontext [max] Image" -class FluxProImageNode(IO.ComfyNode): - """ - Generates images synchronously based on prompt and resolution. - """ - - @classmethod - def define_schema(cls) -> IO.Schema: - return IO.Schema( - node_id="FluxProImageNode", - display_name="Flux 1.1 [pro] Image", - category="api node/image/BFL", - description=cleandoc(cls.__doc__ or ""), - inputs=[ - IO.String.Input( - "prompt", - multiline=True, - default="", - tooltip="Prompt for the image generation", - ), - IO.Boolean.Input( - "prompt_upsampling", - default=False, - tooltip="Whether to perform upsampling on the prompt. " - "If active, automatically modifies the prompt for more creative generation, " - "but results are nondeterministic (same seed will not produce exactly the same result).", - ), - IO.Int.Input( - "width", - default=1024, - min=256, - max=1440, - step=32, - ), - IO.Int.Input( - "height", - default=768, - min=256, - max=1440, - step=32, - ), - IO.Int.Input( - "seed", - default=0, - min=0, - max=0xFFFFFFFFFFFFFFFF, - control_after_generate=True, - tooltip="The random seed used for creating the noise.", - ), - IO.Image.Input( - "image_prompt", - optional=True, - ), - # "image_prompt_strength": ( - # IO.FLOAT, - # { - # "default": 0.1, - # "min": 0.0, - # "max": 1.0, - # "step": 0.01, - # "tooltip": "Blend between the prompt and the image prompt.", - # }, - # ), - ], - outputs=[IO.Image.Output()], - hidden=[ - IO.Hidden.auth_token_comfy_org, - IO.Hidden.api_key_comfy_org, - IO.Hidden.unique_id, - ], - is_api_node=True, - ) - - @classmethod - async def execute( - cls, - prompt: str, - prompt_upsampling, - width: int, - height: int, - seed=0, - image_prompt=None, - # image_prompt_strength=0.1, - ) -> IO.NodeOutput: - image_prompt = image_prompt if image_prompt is None else tensor_to_base64_string(image_prompt) - initial_response = await sync_op( - cls, - ApiEndpoint( - path="/proxy/bfl/flux-pro-1.1/generate", - method="POST", - ), - response_model=BFLFluxProGenerateResponse, - data=BFLFluxProGenerateRequest( - prompt=prompt, - prompt_upsampling=prompt_upsampling, - width=width, - height=height, - seed=seed, - image_prompt=image_prompt, - ), - ) - response = await poll_op( - cls, - ApiEndpoint(initial_response.polling_url), - response_model=BFLFluxStatusResponse, - status_extractor=lambda r: r.status, - progress_extractor=lambda r: r.progress, - completed_statuses=[BFLStatus.ready], - failed_statuses=[ - BFLStatus.request_moderated, - BFLStatus.content_moderated, - BFLStatus.error, - BFLStatus.task_not_found, - ], - queued_statuses=[], - ) - return IO.NodeOutput(await download_url_to_image_tensor(response.result["sample"])) - - class FluxProExpandNode(IO.ComfyNode): """ Outpaints image based on prompt. @@ -640,16 +523,125 @@ class FluxProFillNode(IO.ComfyNode): return IO.NodeOutput(await download_url_to_image_tensor(response.result["sample"])) +class Flux2ProImageNode(IO.ComfyNode): + + @classmethod + def define_schema(cls) -> IO.Schema: + return IO.Schema( + node_id="Flux2ProImageNode", + display_name="Flux.2 [pro] Image", + category="api node/image/BFL", + description="Generates images synchronously based on prompt and resolution.", + inputs=[ + IO.String.Input( + "prompt", + multiline=True, + default="", + tooltip="Prompt for the image generation or edit", + ), + IO.Int.Input( + "width", + default=1024, + min=256, + max=2048, + step=32, + ), + IO.Int.Input( + "height", + default=768, + min=256, + max=2048, + step=32, + ), + IO.Int.Input( + "seed", + default=0, + min=0, + max=0xFFFFFFFFFFFFFFFF, + control_after_generate=True, + tooltip="The random seed used for creating the noise.", + ), + IO.Boolean.Input( + "prompt_upsampling", + default=False, + tooltip="Whether to perform upsampling on the prompt. " + "If active, automatically modifies the prompt for more creative generation, " + "but results are nondeterministic (same seed will not produce exactly the same result).", + ), + IO.Image.Input("images", optional=True, tooltip="Up to 4 images to be used as references."), + ], + outputs=[IO.Image.Output()], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + ) + + @classmethod + async def execute( + cls, + prompt: str, + width: int, + height: int, + seed: int, + prompt_upsampling: bool, + images: torch.Tensor | None = None, + ) -> IO.NodeOutput: + reference_images = {} + if images is not None: + if get_number_of_images(images) > 9: + raise ValueError("The current maximum number of supported images is 9.") + for image_index in range(images.shape[0]): + key_name = f"input_image_{image_index + 1}" if image_index else "input_image" + reference_images[key_name] = tensor_to_base64_string(images[image_index], total_pixels=2048 * 2048) + initial_response = await sync_op( + cls, + ApiEndpoint(path="/proxy/bfl/flux-2-pro/generate", method="POST"), + response_model=BFLFluxProGenerateResponse, + data=Flux2ProGenerateRequest( + prompt=prompt, + width=width, + height=height, + seed=seed, + prompt_upsampling=prompt_upsampling, + **reference_images, + ), + ) + + def price_extractor(_r: BaseModel) -> float | None: + return None if initial_response.cost is None else initial_response.cost / 100 + + response = await poll_op( + cls, + ApiEndpoint(initial_response.polling_url), + response_model=BFLFluxStatusResponse, + status_extractor=lambda r: r.status, + progress_extractor=lambda r: r.progress, + price_extractor=price_extractor, + completed_statuses=[BFLStatus.ready], + failed_statuses=[ + BFLStatus.request_moderated, + BFLStatus.content_moderated, + BFLStatus.error, + BFLStatus.task_not_found, + ], + queued_statuses=[], + ) + return IO.NodeOutput(await download_url_to_image_tensor(response.result["sample"])) + + class BFLExtension(ComfyExtension): @override async def get_node_list(self) -> list[type[IO.ComfyNode]]: return [ FluxProUltraImageNode, - # FluxProImageNode, FluxKontextProImageNode, FluxKontextMaxImageNode, FluxProExpandNode, FluxProFillNode, + Flux2ProImageNode, ] diff --git a/comfy_api_nodes/util/__init__.py b/comfy_api_nodes/util/__init__.py index 21013b591..80292fb3c 100644 --- a/comfy_api_nodes/util/__init__.py +++ b/comfy_api_nodes/util/__init__.py @@ -36,6 +36,7 @@ from .upload_helpers import ( upload_video_to_comfyapi, ) from .validation_utils import ( + get_image_dimensions, get_number_of_images, validate_aspect_ratio_string, validate_audio_duration, @@ -82,6 +83,7 @@ __all__ = [ "trim_video", "video_to_base64_string", # Validation utilities + "get_image_dimensions", "get_number_of_images", "validate_aspect_ratio_string", "validate_audio_duration", From af81cb962d9dd283ddb551962cc223b5a186a1ce Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 25 Nov 2025 08:40:32 -0800 Subject: [PATCH 158/647] Add Flux 2 support to README. (#10882) --- README.md | 1 + 1 file changed, 1 insertion(+) diff --git a/README.md b/README.md index 28beec427..b9300ab07 100644 --- a/README.md +++ b/README.md @@ -67,6 +67,7 @@ See what ComfyUI can do with the [example workflows](https://comfyanonymous.gith - [HiDream](https://comfyanonymous.github.io/ComfyUI_examples/hidream/) - [Qwen Image](https://comfyanonymous.github.io/ComfyUI_examples/qwen_image/) - [Hunyuan Image 2.1](https://comfyanonymous.github.io/ComfyUI_examples/hunyuan_image/) + - [Flux 2](https://comfyanonymous.github.io/ComfyUI_examples/flux2/) - Image Editing Models - [Omnigen 2](https://comfyanonymous.github.io/ComfyUI_examples/omnigen/) - [Flux Kontext](https://comfyanonymous.github.io/ComfyUI_examples/flux/#flux-kontext-image-editing-model) From 828b1b9953175b6df79459f417d1032869d0b46a Mon Sep 17 00:00:00 2001 From: comfyanonymous Date: Tue, 25 Nov 2025 12:40:58 -0500 Subject: [PATCH 159/647] ComfyUI version v0.3.72 --- comfyui_version.py | 2 +- pyproject.toml | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/comfyui_version.py b/comfyui_version.py index b4655d553..dac038c26 100644 --- a/comfyui_version.py +++ b/comfyui_version.py @@ -1,3 +1,3 @@ # This file is automatically generated by the build process when version is # updated in pyproject.toml. -__version__ = "0.3.71" +__version__ = "0.3.72" diff --git a/pyproject.toml b/pyproject.toml index 280dbaf53..75df8fb7c 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -1,6 +1,6 @@ [project] name = "ComfyUI" -version = "0.3.71" +version = "0.3.72" readme = "README.md" license = { file = "LICENSE" } requires-python = ">=3.9" From dff996ca39d86265bbabf15e666484e051f0b3d5 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 25 Nov 2025 11:30:24 -0800 Subject: [PATCH 160/647] Fix crash. (#10885) --- comfy/text_encoders/flux.py | 1 + 1 file changed, 1 insertion(+) diff --git a/comfy/text_encoders/flux.py b/comfy/text_encoders/flux.py index 8dbbca16e..024504a5b 100644 --- a/comfy/text_encoders/flux.py +++ b/comfy/text_encoders/flux.py @@ -87,6 +87,7 @@ def load_mistral_tokenizer(data): vocab = {} max_vocab = mistral_vocab["config"]["default_vocab_size"] + max_vocab -= len(mistral_vocab["special_tokens"]) for w in mistral_vocab["vocab"]: r = w["rank"] From 18b79acba95d44b4ea00bbbfc1856bc71bd58841 Mon Sep 17 00:00:00 2001 From: ComfyUI Wiki Date: Wed, 26 Nov 2025 03:58:21 +0800 Subject: [PATCH 161/647] Update workflow templates to v0.7.20 (#10883) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index b7014f956..5f20816d6 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,5 +1,5 @@ comfyui-frontend-package==1.30.6 -comfyui-workflow-templates==0.7.9 +comfyui-workflow-templates==0.7.20 comfyui-embedded-docs==0.3.1 torch torchsde From d196a905bb379a6d800d0c13f9b4fdea3965311a Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 25 Nov 2025 11:58:39 -0800 Subject: [PATCH 162/647] Lower vram usage for flux 2 text encoder. (#10887) --- comfy/sd1_clip.py | 7 ++++--- comfy/text_encoders/flux.py | 4 ++-- comfy/text_encoders/llama.py | 12 +++++++++--- 3 files changed, 15 insertions(+), 8 deletions(-) diff --git a/comfy/sd1_clip.py b/comfy/sd1_clip.py index 8f509bab1..0fc9ab3db 100644 --- a/comfy/sd1_clip.py +++ b/comfy/sd1_clip.py @@ -90,7 +90,6 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder): special_tokens={"start": 49406, "end": 49407, "pad": 49407}, layer_norm_hidden_state=True, enable_attention_masks=False, zero_out_masked=False, return_projected_pooled=True, return_attention_masks=False, model_options={}): # clip-vit-base-patch32 super().__init__() - assert layer in self.LAYERS if textmodel_json_config is None: textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "sd1_clip_config.json") @@ -164,7 +163,7 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder): def set_clip_options(self, options): layer_idx = options.get("layer", self.layer_idx) self.return_projected_pooled = options.get("projected_pooled", self.return_projected_pooled) - if self.layer == "all": + if isinstance(self.layer, list) or self.layer == "all": pass elif layer_idx is None or abs(layer_idx) > self.num_layers: self.layer = "last" @@ -266,7 +265,9 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder): if self.enable_attention_masks: attention_mask_model = attention_mask - if self.layer == "all": + if isinstance(self.layer, list): + intermediate_output = self.layer + elif self.layer == "all": intermediate_output = "all" else: intermediate_output = self.layer_idx diff --git a/comfy/text_encoders/flux.py b/comfy/text_encoders/flux.py index 024504a5b..99f4812bb 100644 --- a/comfy/text_encoders/flux.py +++ b/comfy/text_encoders/flux.py @@ -138,7 +138,7 @@ class Flux2Tokenizer(sd1_clip.SD1Tokenizer): return tokens class Mistral3_24BModel(sd1_clip.SDClipModel): - def __init__(self, device="cpu", layer="all", layer_idx=None, dtype=None, attention_mask=True, model_options={}): + def __init__(self, device="cpu", layer=[10, 20, 30], layer_idx=None, dtype=None, attention_mask=True, model_options={}): textmodel_json_config = {} num_layers = model_options.get("num_layers", None) if num_layers is not None: @@ -154,7 +154,7 @@ class Flux2TEModel(sd1_clip.SD1ClipModel): def encode_token_weights(self, token_weight_pairs): out, pooled, extra = super().encode_token_weights(token_weight_pairs) - out = torch.stack((out[:, 10], out[:, 20], out[:, 30]), dim=1) + out = torch.stack((out[:, 0], out[:, 1], out[:, 2]), dim=1) out = out.movedim(1, 2) out = out.reshape(out.shape[0], out.shape[1], -1) return out, pooled, extra diff --git a/comfy/text_encoders/llama.py b/comfy/text_encoders/llama.py index 749ff581b..d47ed27bc 100644 --- a/comfy/text_encoders/llama.py +++ b/comfy/text_encoders/llama.py @@ -434,8 +434,12 @@ class Llama2_(nn.Module): intermediate = None all_intermediate = None + only_layers = None if intermediate_output is not None: - if intermediate_output == "all": + if isinstance(intermediate_output, list): + all_intermediate = [] + only_layers = set(intermediate_output) + elif intermediate_output == "all": all_intermediate = [] intermediate_output = None elif intermediate_output < 0: @@ -443,7 +447,8 @@ class Llama2_(nn.Module): for i, layer in enumerate(self.layers): if all_intermediate is not None: - all_intermediate.append(x.unsqueeze(1).clone()) + if only_layers is None or (i in only_layers): + all_intermediate.append(x.unsqueeze(1).clone()) x = layer( x=x, attention_mask=mask, @@ -457,7 +462,8 @@ class Llama2_(nn.Module): x = self.norm(x) if all_intermediate is not None: - all_intermediate.append(x.unsqueeze(1).clone()) + if only_layers is None or ((i + 1) in only_layers): + all_intermediate.append(x.unsqueeze(1).clone()) if all_intermediate is not None: intermediate = torch.cat(all_intermediate, dim=1) From 0c18842acbdf546883b08808dd9feea7605d7649 Mon Sep 17 00:00:00 2001 From: comfyanonymous Date: Tue, 25 Nov 2025 14:59:37 -0500 Subject: [PATCH 163/647] ComfyUI v0.3.73 --- comfyui_version.py | 2 +- pyproject.toml | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/comfyui_version.py b/comfyui_version.py index dac038c26..f8818838e 100644 --- a/comfyui_version.py +++ b/comfyui_version.py @@ -1,3 +1,3 @@ # This file is automatically generated by the build process when version is # updated in pyproject.toml. -__version__ = "0.3.72" +__version__ = "0.3.73" diff --git a/pyproject.toml b/pyproject.toml index 75df8fb7c..7e4bac12d 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -1,6 +1,6 @@ [project] name = "ComfyUI" -version = "0.3.72" +version = "0.3.73" readme = "README.md" license = { file = "LICENSE" } requires-python = ">=3.9" From e9aae31fa241a6a63a368800146ea91629d4e8c2 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 25 Nov 2025 15:41:45 -0800 Subject: [PATCH 164/647] Z Image model. (#10892) --- comfy/ldm/lumina/model.py | 219 +++++++------------- comfy/ldm/modules/diffusionmodules/mmdit.py | 6 +- comfy/model_base.py | 4 + comfy/model_detection.py | 29 ++- comfy/sd.py | 8 + comfy/text_encoders/llama.py | 31 +++ comfy/text_encoders/z_image.py | 48 +++++ 7 files changed, 196 insertions(+), 149 deletions(-) create mode 100644 comfy/text_encoders/z_image.py diff --git a/comfy/ldm/lumina/model.py b/comfy/ldm/lumina/model.py index b4494a51d..c8643eb82 100644 --- a/comfy/ldm/lumina/model.py +++ b/comfy/ldm/lumina/model.py @@ -11,6 +11,7 @@ import comfy.ldm.common_dit from comfy.ldm.modules.diffusionmodules.mmdit import TimestepEmbedder from comfy.ldm.modules.attention import optimized_attention_masked from comfy.ldm.flux.layers import EmbedND +from comfy.ldm.flux.math import apply_rope import comfy.patcher_extension @@ -31,6 +32,7 @@ class JointAttention(nn.Module): n_heads: int, n_kv_heads: Optional[int], qk_norm: bool, + out_bias: bool = False, operation_settings={}, ): """ @@ -59,7 +61,7 @@ class JointAttention(nn.Module): self.out = operation_settings.get("operations").Linear( n_heads * self.head_dim, dim, - bias=False, + bias=out_bias, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"), ) @@ -70,35 +72,6 @@ class JointAttention(nn.Module): else: self.q_norm = self.k_norm = nn.Identity() - @staticmethod - def apply_rotary_emb( - x_in: torch.Tensor, - freqs_cis: torch.Tensor, - ) -> torch.Tensor: - """ - Apply rotary embeddings to input tensors using the given frequency - tensor. - - This function applies rotary embeddings to the given query 'xq' and - key 'xk' tensors using the provided frequency tensor 'freqs_cis'. The - input tensors are reshaped as complex numbers, and the frequency tensor - is reshaped for broadcasting compatibility. The resulting tensors - contain rotary embeddings and are returned as real tensors. - - Args: - x_in (torch.Tensor): Query or Key tensor to apply rotary embeddings. - freqs_cis (torch.Tensor): Precomputed frequency tensor for complex - exponentials. - - Returns: - Tuple[torch.Tensor, torch.Tensor]: Tuple of modified query tensor - and key tensor with rotary embeddings. - """ - - t_ = x_in.reshape(*x_in.shape[:-1], -1, 1, 2) - t_out = freqs_cis[..., 0] * t_[..., 0] + freqs_cis[..., 1] * t_[..., 1] - return t_out.reshape(*x_in.shape) - def forward( self, x: torch.Tensor, @@ -134,8 +107,7 @@ class JointAttention(nn.Module): xq = self.q_norm(xq) xk = self.k_norm(xk) - xq = JointAttention.apply_rotary_emb(xq, freqs_cis=freqs_cis) - xk = JointAttention.apply_rotary_emb(xk, freqs_cis=freqs_cis) + xq, xk = apply_rope(xq, xk, freqs_cis) n_rep = self.n_local_heads // self.n_local_kv_heads if n_rep >= 1: @@ -215,6 +187,8 @@ class JointTransformerBlock(nn.Module): norm_eps: float, qk_norm: bool, modulation=True, + z_image_modulation=False, + attn_out_bias=False, operation_settings={}, ) -> None: """ @@ -235,10 +209,10 @@ class JointTransformerBlock(nn.Module): super().__init__() self.dim = dim self.head_dim = dim // n_heads - self.attention = JointAttention(dim, n_heads, n_kv_heads, qk_norm, operation_settings=operation_settings) + self.attention = JointAttention(dim, n_heads, n_kv_heads, qk_norm, out_bias=attn_out_bias, operation_settings=operation_settings) self.feed_forward = FeedForward( dim=dim, - hidden_dim=4 * dim, + hidden_dim=dim, multiple_of=multiple_of, ffn_dim_multiplier=ffn_dim_multiplier, operation_settings=operation_settings, @@ -252,16 +226,27 @@ class JointTransformerBlock(nn.Module): self.modulation = modulation if modulation: - self.adaLN_modulation = nn.Sequential( - nn.SiLU(), - operation_settings.get("operations").Linear( - min(dim, 1024), - 4 * dim, - bias=True, - device=operation_settings.get("device"), - dtype=operation_settings.get("dtype"), - ), - ) + if z_image_modulation: + self.adaLN_modulation = nn.Sequential( + operation_settings.get("operations").Linear( + min(dim, 256), + 4 * dim, + bias=True, + device=operation_settings.get("device"), + dtype=operation_settings.get("dtype"), + ), + ) + else: + self.adaLN_modulation = nn.Sequential( + nn.SiLU(), + operation_settings.get("operations").Linear( + min(dim, 1024), + 4 * dim, + bias=True, + device=operation_settings.get("device"), + dtype=operation_settings.get("dtype"), + ), + ) def forward( self, @@ -323,7 +308,7 @@ class FinalLayer(nn.Module): The final layer of NextDiT. """ - def __init__(self, hidden_size, patch_size, out_channels, operation_settings={}): + def __init__(self, hidden_size, patch_size, out_channels, z_image_modulation=False, operation_settings={}): super().__init__() self.norm_final = operation_settings.get("operations").LayerNorm( hidden_size, @@ -340,10 +325,15 @@ class FinalLayer(nn.Module): dtype=operation_settings.get("dtype"), ) + if z_image_modulation: + min_mod = 256 + else: + min_mod = 1024 + self.adaLN_modulation = nn.Sequential( nn.SiLU(), operation_settings.get("operations").Linear( - min(hidden_size, 1024), + min(hidden_size, min_mod), hidden_size, bias=True, device=operation_settings.get("device"), @@ -373,12 +363,16 @@ class NextDiT(nn.Module): n_heads: int = 32, n_kv_heads: Optional[int] = None, multiple_of: int = 256, - ffn_dim_multiplier: Optional[float] = None, + ffn_dim_multiplier: float = 4.0, norm_eps: float = 1e-5, qk_norm: bool = False, cap_feat_dim: int = 5120, axes_dims: List[int] = (16, 56, 56), axes_lens: List[int] = (1, 512, 512), + rope_theta=10000.0, + z_image_modulation=False, + time_scale=1.0, + pad_tokens_multiple=None, image_model=None, device=None, dtype=None, @@ -390,6 +384,8 @@ class NextDiT(nn.Module): self.in_channels = in_channels self.out_channels = in_channels self.patch_size = patch_size + self.time_scale = time_scale + self.pad_tokens_multiple = pad_tokens_multiple self.x_embedder = operation_settings.get("operations").Linear( in_features=patch_size * patch_size * in_channels, @@ -411,6 +407,7 @@ class NextDiT(nn.Module): norm_eps, qk_norm, modulation=True, + z_image_modulation=z_image_modulation, operation_settings=operation_settings, ) for layer_id in range(n_refiner_layers) @@ -434,7 +431,7 @@ class NextDiT(nn.Module): ] ) - self.t_embedder = TimestepEmbedder(min(dim, 1024), **operation_settings) + self.t_embedder = TimestepEmbedder(min(dim, 1024), output_size=256 if z_image_modulation else None, **operation_settings) self.cap_embedder = nn.Sequential( operation_settings.get("operations").RMSNorm(cap_feat_dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")), operation_settings.get("operations").Linear( @@ -457,18 +454,24 @@ class NextDiT(nn.Module): ffn_dim_multiplier, norm_eps, qk_norm, + z_image_modulation=z_image_modulation, + attn_out_bias=False, operation_settings=operation_settings, ) for layer_id in range(n_layers) ] ) self.norm_final = operation_settings.get("operations").RMSNorm(dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) - self.final_layer = FinalLayer(dim, patch_size, self.out_channels, operation_settings=operation_settings) + self.final_layer = FinalLayer(dim, patch_size, self.out_channels, z_image_modulation=z_image_modulation, operation_settings=operation_settings) + + if self.pad_tokens_multiple is not None: + self.x_pad_token = nn.Parameter(torch.empty((1, dim), device=device, dtype=dtype)) + self.cap_pad_token = nn.Parameter(torch.empty((1, dim), device=device, dtype=dtype)) assert (dim // n_heads) == sum(axes_dims) self.axes_dims = axes_dims self.axes_lens = axes_lens - self.rope_embedder = EmbedND(dim=dim // n_heads, theta=10000.0, axes_dim=axes_dims) + self.rope_embedder = EmbedND(dim=dim // n_heads, theta=rope_theta, axes_dim=axes_dims) self.dim = dim self.n_heads = n_heads @@ -503,108 +506,42 @@ class NextDiT(nn.Module): bsz = len(x) pH = pW = self.patch_size device = x[0].device - dtype = x[0].dtype - if cap_mask is not None: - l_effective_cap_len = cap_mask.sum(dim=1).tolist() - else: - l_effective_cap_len = [num_tokens] * bsz + if self.pad_tokens_multiple is not None: + pad_extra = (-cap_feats.shape[1]) % self.pad_tokens_multiple + cap_feats = torch.cat((cap_feats, self.cap_pad_token.to(device=cap_feats.device, dtype=cap_feats.dtype).unsqueeze(0).repeat(cap_feats.shape[0], pad_extra, 1)), dim=1) - if cap_mask is not None and not torch.is_floating_point(cap_mask): - cap_mask = (cap_mask - 1).to(dtype) * torch.finfo(dtype).max + cap_pos_ids = torch.zeros(bsz, cap_feats.shape[1], 3, dtype=torch.float32, device=device) + cap_pos_ids[:, :, 0] = torch.arange(cap_feats.shape[1], dtype=torch.float32, device=device) + 1.0 - img_sizes = [(img.size(1), img.size(2)) for img in x] - l_effective_img_len = [(H // pH) * (W // pW) for (H, W) in img_sizes] + B, C, H, W = x.shape + x = self.x_embedder(x.view(B, C, H // pH, pH, W // pW, pW).permute(0, 2, 4, 3, 5, 1).flatten(3).flatten(1, 2)) - max_seq_len = max( - (cap_len+img_len for cap_len, img_len in zip(l_effective_cap_len, l_effective_img_len)) - ) - max_cap_len = max(l_effective_cap_len) - max_img_len = max(l_effective_img_len) + H_tokens, W_tokens = H // pH, W // pW + x_pos_ids = torch.zeros((bsz, x.shape[1], 3), dtype=torch.float32, device=device) + x_pos_ids[:, :, 0] = cap_feats.shape[1] + 1 + x_pos_ids[:, :, 1] = torch.arange(H_tokens, dtype=torch.float32, device=device).view(-1, 1).repeat(1, W_tokens).flatten() + x_pos_ids[:, :, 2] = torch.arange(W_tokens, dtype=torch.float32, device=device).view(1, -1).repeat(H_tokens, 1).flatten() - position_ids = torch.zeros(bsz, max_seq_len, 3, dtype=torch.float32, device=device) + if self.pad_tokens_multiple is not None: + pad_extra = (-x.shape[1]) % self.pad_tokens_multiple + x = torch.cat((x, self.x_pad_token.to(device=x.device, dtype=x.dtype).unsqueeze(0).repeat(x.shape[0], pad_extra, 1)), dim=1) + x_pos_ids = torch.nn.functional.pad(x_pos_ids, (0, 0, 0, pad_extra)) - for i in range(bsz): - cap_len = l_effective_cap_len[i] - img_len = l_effective_img_len[i] - H, W = img_sizes[i] - H_tokens, W_tokens = H // pH, W // pW - assert H_tokens * W_tokens == img_len - - rope_options = transformer_options.get("rope_options", None) - h_scale = 1.0 - w_scale = 1.0 - h_start = 0 - w_start = 0 - if rope_options is not None: - h_scale = rope_options.get("scale_y", 1.0) - w_scale = rope_options.get("scale_x", 1.0) - - h_start = rope_options.get("shift_y", 0.0) - w_start = rope_options.get("shift_x", 0.0) - - position_ids[i, :cap_len, 0] = torch.arange(cap_len, dtype=torch.float32, device=device) - position_ids[i, cap_len:cap_len+img_len, 0] = cap_len - row_ids = (torch.arange(H_tokens, dtype=torch.float32, device=device) * h_scale + h_start).view(-1, 1).repeat(1, W_tokens).flatten() - col_ids = (torch.arange(W_tokens, dtype=torch.float32, device=device) * w_scale + w_start).view(1, -1).repeat(H_tokens, 1).flatten() - position_ids[i, cap_len:cap_len+img_len, 1] = row_ids - position_ids[i, cap_len:cap_len+img_len, 2] = col_ids - - freqs_cis = self.rope_embedder(position_ids).movedim(1, 2).to(dtype) - - # build freqs_cis for cap and image individually - cap_freqs_cis_shape = list(freqs_cis.shape) - # cap_freqs_cis_shape[1] = max_cap_len - cap_freqs_cis_shape[1] = cap_feats.shape[1] - cap_freqs_cis = torch.zeros(*cap_freqs_cis_shape, device=device, dtype=freqs_cis.dtype) - - img_freqs_cis_shape = list(freqs_cis.shape) - img_freqs_cis_shape[1] = max_img_len - img_freqs_cis = torch.zeros(*img_freqs_cis_shape, device=device, dtype=freqs_cis.dtype) - - for i in range(bsz): - cap_len = l_effective_cap_len[i] - img_len = l_effective_img_len[i] - cap_freqs_cis[i, :cap_len] = freqs_cis[i, :cap_len] - img_freqs_cis[i, :img_len] = freqs_cis[i, cap_len:cap_len+img_len] + freqs_cis = self.rope_embedder(torch.cat((cap_pos_ids, x_pos_ids), dim=1)).movedim(1, 2) # refine context for layer in self.context_refiner: - cap_feats = layer(cap_feats, cap_mask, cap_freqs_cis, transformer_options=transformer_options) + cap_feats = layer(cap_feats, cap_mask, freqs_cis[:, :cap_pos_ids.shape[1]], transformer_options=transformer_options) - # refine image - flat_x = [] - for i in range(bsz): - img = x[i] - C, H, W = img.size() - img = img.view(C, H // pH, pH, W // pW, pW).permute(1, 3, 2, 4, 0).flatten(2).flatten(0, 1) - flat_x.append(img) - x = flat_x - padded_img_embed = torch.zeros(bsz, max_img_len, x[0].shape[-1], device=device, dtype=x[0].dtype) - padded_img_mask = torch.zeros(bsz, max_img_len, dtype=dtype, device=device) - for i in range(bsz): - padded_img_embed[i, :l_effective_img_len[i]] = x[i] - padded_img_mask[i, l_effective_img_len[i]:] = -torch.finfo(dtype).max - - padded_img_embed = self.x_embedder(padded_img_embed) - padded_img_mask = padded_img_mask.unsqueeze(1) + padded_img_mask = None for layer in self.noise_refiner: - padded_img_embed = layer(padded_img_embed, padded_img_mask, img_freqs_cis, t, transformer_options=transformer_options) - - if cap_mask is not None: - mask = torch.zeros(bsz, max_seq_len, dtype=dtype, device=device) - mask[:, :max_cap_len] = cap_mask[:, :max_cap_len] - else: - mask = None - - padded_full_embed = torch.zeros(bsz, max_seq_len, self.dim, device=device, dtype=x[0].dtype) - for i in range(bsz): - cap_len = l_effective_cap_len[i] - img_len = l_effective_img_len[i] - - padded_full_embed[i, :cap_len] = cap_feats[i, :cap_len] - padded_full_embed[i, cap_len:cap_len+img_len] = padded_img_embed[i, :img_len] + x = layer(x, padded_img_mask, freqs_cis[:, cap_pos_ids.shape[1]:], t, transformer_options=transformer_options) + padded_full_embed = torch.cat((cap_feats, x), dim=1) + mask = None + img_sizes = [(H, W)] * bsz + l_effective_cap_len = [cap_feats.shape[1]] * bsz return padded_full_embed, mask, img_sizes, l_effective_cap_len, freqs_cis def forward(self, x, timesteps, context, num_tokens, attention_mask=None, **kwargs): @@ -627,7 +564,7 @@ class NextDiT(nn.Module): y: (N,) tensor of text tokens/features """ - t = self.t_embedder(t, dtype=x.dtype) # (N, D) + t = self.t_embedder(t * self.time_scale, dtype=x.dtype) # (N, D) adaln_input = t cap_feats = self.cap_embedder(cap_feats) # (N, L, D) # todo check if able to batchify w.o. redundant compute diff --git a/comfy/ldm/modules/diffusionmodules/mmdit.py b/comfy/ldm/modules/diffusionmodules/mmdit.py index 42f406f1a..0dc8fe789 100644 --- a/comfy/ldm/modules/diffusionmodules/mmdit.py +++ b/comfy/ldm/modules/diffusionmodules/mmdit.py @@ -211,12 +211,14 @@ class TimestepEmbedder(nn.Module): Embeds scalar timesteps into vector representations. """ - def __init__(self, hidden_size, frequency_embedding_size=256, dtype=None, device=None, operations=None): + def __init__(self, hidden_size, frequency_embedding_size=256, output_size=None, dtype=None, device=None, operations=None): super().__init__() + if output_size is None: + output_size = hidden_size self.mlp = nn.Sequential( operations.Linear(frequency_embedding_size, hidden_size, bias=True, dtype=dtype, device=device), nn.SiLU(), - operations.Linear(hidden_size, hidden_size, bias=True, dtype=dtype, device=device), + operations.Linear(hidden_size, output_size, bias=True, dtype=dtype, device=device), ) self.frequency_embedding_size = frequency_embedding_size diff --git a/comfy/model_base.py b/comfy/model_base.py index cad79ecbd..cc21b1de9 100644 --- a/comfy/model_base.py +++ b/comfy/model_base.py @@ -1114,9 +1114,13 @@ class Lumina2(BaseModel): if torch.numel(attention_mask) != attention_mask.sum(): out['attention_mask'] = comfy.conds.CONDRegular(attention_mask) out['num_tokens'] = comfy.conds.CONDConstant(max(1, torch.sum(attention_mask).item())) + cross_attn = kwargs.get("cross_attn", None) if cross_attn is not None: out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn) + if 'num_tokens' not in out: + out['num_tokens'] = comfy.conds.CONDConstant(cross_attn.shape[1]) + return out class WAN21(BaseModel): diff --git a/comfy/model_detection.py b/comfy/model_detection.py index b2ba1459d..7afe4a798 100644 --- a/comfy/model_detection.py +++ b/comfy/model_detection.py @@ -416,14 +416,31 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): dit_config["image_model"] = "lumina2" dit_config["patch_size"] = 2 dit_config["in_channels"] = 16 - dit_config["dim"] = 2304 - dit_config["cap_feat_dim"] = state_dict['{}cap_embedder.1.weight'.format(key_prefix)].shape[1] + w = state_dict['{}cap_embedder.1.weight'.format(key_prefix)] + dit_config["dim"] = w.shape[0] + dit_config["cap_feat_dim"] = w.shape[1] dit_config["n_layers"] = count_blocks(state_dict_keys, '{}layers.'.format(key_prefix) + '{}.') - dit_config["n_heads"] = 24 - dit_config["n_kv_heads"] = 8 dit_config["qk_norm"] = True - dit_config["axes_dims"] = [32, 32, 32] - dit_config["axes_lens"] = [300, 512, 512] + + if dit_config["dim"] == 2304: # Original Lumina 2 + dit_config["n_heads"] = 24 + dit_config["n_kv_heads"] = 8 + dit_config["axes_dims"] = [32, 32, 32] + dit_config["axes_lens"] = [300, 512, 512] + dit_config["rope_theta"] = 10000.0 + dit_config["ffn_dim_multiplier"] = 4.0 + elif dit_config["dim"] == 3840: # Z image + dit_config["n_heads"] = 30 + dit_config["n_kv_heads"] = 30 + dit_config["axes_dims"] = [32, 48, 48] + dit_config["axes_lens"] = [1536, 512, 512] + dit_config["rope_theta"] = 256.0 + dit_config["ffn_dim_multiplier"] = (8.0 / 3.0) + dit_config["z_image_modulation"] = True + dit_config["time_scale"] = 1000.0 + if '{}cap_pad_token'.format(key_prefix) in state_dict_keys: + dit_config["pad_tokens_multiple"] = 32 + return dit_config if '{}head.modulation'.format(key_prefix) in state_dict_keys: # Wan 2.1 diff --git a/comfy/sd.py b/comfy/sd.py index 14dd8944c..350fae92b 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -52,6 +52,7 @@ import comfy.text_encoders.ace import comfy.text_encoders.omnigen2 import comfy.text_encoders.qwen_image import comfy.text_encoders.hunyuan_image +import comfy.text_encoders.z_image import comfy.model_patcher import comfy.lora @@ -953,6 +954,8 @@ class TEModel(Enum): GEMMA_3_4B = 13 MISTRAL3_24B = 14 MISTRAL3_24B_PRUNED_FLUX2 = 15 + QWEN3_4B = 16 + def detect_te_model(sd): if "text_model.encoder.layers.30.mlp.fc1.weight" in sd: @@ -985,6 +988,8 @@ def detect_te_model(sd): if weight.shape[0] == 512: return TEModel.QWEN25_7B if "model.layers.0.post_attention_layernorm.weight" in sd: + if 'model.layers.0.self_attn.q_norm.weight' in sd: + return TEModel.QWEN3_4B weight = sd['model.layers.0.post_attention_layernorm.weight'] if weight.shape[0] == 5120: if "model.layers.39.post_attention_layernorm.weight" in sd: @@ -1110,6 +1115,9 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip clip_target.clip = comfy.text_encoders.flux.flux2_te(**llama_detect(clip_data), pruned=te_model == TEModel.MISTRAL3_24B_PRUNED_FLUX2) clip_target.tokenizer = comfy.text_encoders.flux.Flux2Tokenizer tokenizer_data["tekken_model"] = clip_data[0].get("tekken_model", None) + elif te_model == TEModel.QWEN3_4B: + clip_target.clip = comfy.text_encoders.z_image.te(**llama_detect(clip_data)) + clip_target.tokenizer = comfy.text_encoders.z_image.ZImageTokenizer else: # clip_l if clip_type == CLIPType.SD3: diff --git a/comfy/text_encoders/llama.py b/comfy/text_encoders/llama.py index d47ed27bc..cd4b5f76c 100644 --- a/comfy/text_encoders/llama.py +++ b/comfy/text_encoders/llama.py @@ -78,6 +78,28 @@ class Qwen25_3BConfig: rope_scale = None final_norm: bool = True +@dataclass +class Qwen3_4BConfig: + vocab_size: int = 151936 + hidden_size: int = 2560 + intermediate_size: int = 9728 + num_hidden_layers: int = 36 + num_attention_heads: int = 32 + num_key_value_heads: int = 8 + max_position_embeddings: int = 40960 + rms_norm_eps: float = 1e-6 + rope_theta: float = 1000000.0 + transformer_type: str = "llama" + head_dim = 128 + rms_norm_add = False + mlp_activation = "silu" + qkv_bias = False + rope_dims = None + q_norm = "gemma3" + k_norm = "gemma3" + rope_scale = None + final_norm: bool = True + @dataclass class Qwen25_7BVLI_Config: vocab_size: int = 152064 @@ -511,6 +533,15 @@ class Qwen25_3B(BaseLlama, torch.nn.Module): self.model = Llama2_(config, device=device, dtype=dtype, ops=operations) self.dtype = dtype +class Qwen3_4B(BaseLlama, torch.nn.Module): + def __init__(self, config_dict, dtype, device, operations): + super().__init__() + config = Qwen3_4BConfig(**config_dict) + self.num_layers = config.num_hidden_layers + + self.model = Llama2_(config, device=device, dtype=dtype, ops=operations) + self.dtype = dtype + class Qwen25_7BVLI(BaseLlama, torch.nn.Module): def __init__(self, config_dict, dtype, device, operations): super().__init__() diff --git a/comfy/text_encoders/z_image.py b/comfy/text_encoders/z_image.py new file mode 100644 index 000000000..bb9273b20 --- /dev/null +++ b/comfy/text_encoders/z_image.py @@ -0,0 +1,48 @@ +from transformers import Qwen2Tokenizer +import comfy.text_encoders.llama +from comfy import sd1_clip +import os + +class Qwen3Tokenizer(sd1_clip.SDTokenizer): + def __init__(self, embedding_directory=None, tokenizer_data={}): + tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "qwen25_tokenizer") + super().__init__(tokenizer_path, pad_with_end=False, embedding_size=2560, embedding_key='qwen3_4b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, pad_token=151643, tokenizer_data=tokenizer_data) + + +class ZImageTokenizer(sd1_clip.SD1Tokenizer): + def __init__(self, embedding_directory=None, tokenizer_data={}): + super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, name="qwen3_4b", tokenizer=Qwen3Tokenizer) + self.llama_template = "<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n" + + def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None, **kwargs): + if llama_template is None: + llama_text = self.llama_template.format(text) + else: + llama_text = llama_template.format(text) + + tokens = super().tokenize_with_weights(llama_text, return_word_ids=return_word_ids, disable_weights=True, **kwargs) + return tokens + + +class Qwen3_4BModel(sd1_clip.SDClipModel): + def __init__(self, device="cpu", layer="hidden", layer_idx=-2, dtype=None, attention_mask=True, model_options={}): + super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"pad": 151643}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Qwen3_4B, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options) + + +class ZImageTEModel(sd1_clip.SD1ClipModel): + def __init__(self, device="cpu", dtype=None, model_options={}): + super().__init__(device=device, dtype=dtype, name="qwen3_4b", clip_model=Qwen3_4BModel, model_options=model_options) + + +def te(dtype_llama=None, llama_scaled_fp8=None, llama_quantization_metadata=None): + class ZImageTEModel_(ZImageTEModel): + def __init__(self, device="cpu", dtype=None, model_options={}): + if llama_scaled_fp8 is not None and "scaled_fp8" not in model_options: + model_options = model_options.copy() + model_options["scaled_fp8"] = llama_scaled_fp8 + if dtype_llama is not None: + dtype = dtype_llama + if llama_quantization_metadata is not None: + model_options["quantization_metadata"] = llama_quantization_metadata + super().__init__(device=device, dtype=dtype, model_options=model_options) + return ZImageTEModel_ From 0e24dbb19f34f242edb77c550396cf6806f7b22f Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 25 Nov 2025 16:02:51 -0800 Subject: [PATCH 165/647] Adjustments to Z Image. (#10893) --- comfy/supported_models.py | 23 +++++++++++++++++++++-- 1 file changed, 21 insertions(+), 2 deletions(-) diff --git a/comfy/supported_models.py b/comfy/supported_models.py index 8fe8e63f6..af8120400 100644 --- a/comfy/supported_models.py +++ b/comfy/supported_models.py @@ -21,6 +21,7 @@ import comfy.text_encoders.ace import comfy.text_encoders.omnigen2 import comfy.text_encoders.qwen_image import comfy.text_encoders.hunyuan_image +import comfy.text_encoders.z_image from . import supported_models_base from . import latent_formats @@ -994,7 +995,7 @@ class Lumina2(supported_models_base.BASE): "shift": 6.0, } - memory_usage_factor = 1.2 + memory_usage_factor = 1.4 unet_extra_config = {} latent_format = latent_formats.Flux @@ -1013,6 +1014,24 @@ class Lumina2(supported_models_base.BASE): hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}gemma2_2b.transformer.".format(pref)) return supported_models_base.ClipTarget(comfy.text_encoders.lumina2.LuminaTokenizer, comfy.text_encoders.lumina2.te(**hunyuan_detect)) +class ZImage(Lumina2): + unet_config = { + "image_model": "lumina2", + "dim": 3840, + } + + sampling_settings = { + "multiplier": 1.0, + "shift": 3.0, + } + + memory_usage_factor = 1.7 + + def clip_target(self, state_dict={}): + pref = self.text_encoder_key_prefix[0] + hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen3_4b.transformer.".format(pref)) + return supported_models_base.ClipTarget(comfy.text_encoders.z_image.ZImageTokenizer, comfy.text_encoders.z_image.te(**hunyuan_detect)) + class WAN21_T2V(supported_models_base.BASE): unet_config = { "image_model": "wan2.1", @@ -1453,7 +1472,7 @@ class HunyuanVideo15_SR_Distilled(HunyuanVideo): hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen25_7b.transformer.".format(pref)) return supported_models_base.ClipTarget(comfy.text_encoders.hunyuan_video.HunyuanVideo15Tokenizer, comfy.text_encoders.hunyuan_image.te(**hunyuan_detect)) -models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanVideo15_SR_Distilled, HunyuanVideo15, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, Omnigen2, QwenImage, Flux2] +models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanVideo15_SR_Distilled, HunyuanVideo15, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, ZImage, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, Omnigen2, QwenImage, Flux2] models += [SVD_img2vid] From bdb10a583f1b1e495ee00dbd1674f11016a6a93e Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 25 Nov 2025 21:07:58 -0800 Subject: [PATCH 166/647] Fix loras not working on mixed fp8. (#10899) --- comfy/model_patcher.py | 2 +- comfy/ops.py | 22 +++++++++++++++++++++- comfy/quant_ops.py | 21 ++++++++++++++------- comfy/weight_adapter/lora.py | 1 + 4 files changed, 37 insertions(+), 9 deletions(-) diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index 6551ced5a..73adc7f70 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -132,7 +132,7 @@ class LowVramPatch: def __call__(self, weight): intermediate_dtype = weight.dtype if self.convert_func is not None: - weight = self.convert_func(weight.to(dtype=torch.float32, copy=True), inplace=True) + weight = self.convert_func(weight, inplace=False) if intermediate_dtype not in [torch.float32, torch.float16, torch.bfloat16]: #intermediate_dtype has to be one that is supported in math ops intermediate_dtype = torch.float32 diff --git a/comfy/ops.py b/comfy/ops.py index 785aa1c9f..a0ff4e8f1 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -117,6 +117,8 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, of if weight_has_function or weight.dtype != dtype: with wf_context: weight = weight.to(dtype=dtype) + if isinstance(weight, QuantizedTensor): + weight = weight.dequantize() for f in s.weight_function: weight = f(weight) @@ -502,7 +504,7 @@ def scaled_fp8_ops(fp8_matrix_mult=False, scale_input=False, override_dtype=None weight *= self.scale_weight.to(device=weight.device, dtype=weight.dtype) return weight else: - return weight * self.scale_weight.to(device=weight.device, dtype=weight.dtype) + return weight.to(dtype=torch.float32) * self.scale_weight.to(device=weight.device, dtype=torch.float32) def set_weight(self, weight, inplace_update=False, seed=None, return_weight=False, **kwargs): weight = comfy.float.stochastic_rounding(weight / self.scale_weight.to(device=weight.device, dtype=weight.dtype), self.weight.dtype, seed=seed) @@ -643,6 +645,24 @@ def mixed_precision_ops(layer_quant_config={}, compute_dtype=torch.bfloat16, ful not isinstance(input, QuantizedTensor)): input = QuantizedTensor.from_float(input, self.layout_type, scale=self.input_scale, dtype=self.weight.dtype) return self._forward(input, self.weight, self.bias) + + def convert_weight(self, weight, inplace=False, **kwargs): + if isinstance(weight, QuantizedTensor): + return weight.dequantize() + else: + return weight + + def set_weight(self, weight, inplace_update=False, seed=None, return_weight=False, **kwargs): + if getattr(self, 'layout_type', None) is not None: + weight = QuantizedTensor.from_float(weight, self.layout_type, scale=None, dtype=self.weight.dtype, stochastic_rounding=seed, inplace_ops=True) + else: + weight = weight.to(self.weight.dtype) + if return_weight: + return weight + + assert inplace_update is False # TODO: eventually remove the inplace_update stuff + self.weight = torch.nn.Parameter(weight, requires_grad=False) + return MixedPrecisionOps def pick_operations(weight_dtype, compute_dtype, load_device=None, disable_fast_fp8=False, fp8_optimizations=False, scaled_fp8=None, model_config=None): diff --git a/comfy/quant_ops.py b/comfy/quant_ops.py index 0c16bcf8d..d2f3e7397 100644 --- a/comfy/quant_ops.py +++ b/comfy/quant_ops.py @@ -1,6 +1,7 @@ import torch import logging from typing import Tuple, Dict +import comfy.float _LAYOUT_REGISTRY = {} _GENERIC_UTILS = {} @@ -393,7 +394,7 @@ class TensorCoreFP8Layout(QuantizedLayout): - orig_dtype: Original dtype before quantization (for casting back) """ @classmethod - def quantize(cls, tensor, scale=None, dtype=torch.float8_e4m3fn): + def quantize(cls, tensor, scale=None, dtype=torch.float8_e4m3fn, stochastic_rounding=0, inplace_ops=False): orig_dtype = tensor.dtype if scale is None: @@ -403,17 +404,23 @@ class TensorCoreFP8Layout(QuantizedLayout): scale = torch.tensor(scale) scale = scale.to(device=tensor.device, dtype=torch.float32) - tensor_scaled = tensor * (1.0 / scale).to(tensor.dtype) - # TODO: uncomment this if it's actually needed because the clamp has a small performance penality' - lp_amax = torch.finfo(dtype).max - torch.clamp(tensor_scaled, min=-lp_amax, max=lp_amax, out=tensor_scaled) - qdata = tensor_scaled.to(dtype, memory_format=torch.contiguous_format) + if inplace_ops: + tensor *= (1.0 / scale).to(tensor.dtype) + else: + tensor = tensor * (1.0 / scale).to(tensor.dtype) + + if stochastic_rounding > 0: + tensor = comfy.float.stochastic_rounding(tensor, dtype=dtype, seed=stochastic_rounding) + else: + lp_amax = torch.finfo(dtype).max + torch.clamp(tensor, min=-lp_amax, max=lp_amax, out=tensor) + tensor = tensor.to(dtype, memory_format=torch.contiguous_format) layout_params = { 'scale': scale, 'orig_dtype': orig_dtype } - return qdata, layout_params + return tensor, layout_params @staticmethod def dequantize(qdata, scale, orig_dtype, **kwargs): diff --git a/comfy/weight_adapter/lora.py b/comfy/weight_adapter/lora.py index 4db004e50..3cc60bb1b 100644 --- a/comfy/weight_adapter/lora.py +++ b/comfy/weight_adapter/lora.py @@ -194,6 +194,7 @@ class LoRAAdapter(WeightAdapterBase): lora_diff = torch.mm( mat1.flatten(start_dim=1), mat2.flatten(start_dim=1) ).reshape(weight.shape) + del mat1, mat2 if dora_scale is not None: weight = weight_decompose( dora_scale, From 90b3995ec842335e44d70e0521ff6ff6c3ff9aaa Mon Sep 17 00:00:00 2001 From: comfyanonymous Date: Wed, 26 Nov 2025 00:34:15 -0500 Subject: [PATCH 167/647] ComfyUI v0.3.74 --- comfyui_version.py | 2 +- pyproject.toml | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/comfyui_version.py b/comfyui_version.py index f8818838e..b565c7367 100644 --- a/comfyui_version.py +++ b/comfyui_version.py @@ -1,3 +1,3 @@ # This file is automatically generated by the build process when version is # updated in pyproject.toml. -__version__ = "0.3.73" +__version__ = "0.3.74" diff --git a/pyproject.toml b/pyproject.toml index 7e4bac12d..ccf0fcdb9 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -1,6 +1,6 @@ [project] name = "ComfyUI" -version = "0.3.73" +version = "0.3.74" readme = "README.md" license = { file = "LICENSE" } requires-python = ">=3.9" From 58b85746618e2bc2dd32024c89403926aad59f48 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 25 Nov 2025 23:36:19 -0800 Subject: [PATCH 168/647] Fix Flux2 reference image mem estimation. (#10905) --- comfy/model_base.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy/model_base.py b/comfy/model_base.py index cc21b1de9..9b76c285e 100644 --- a/comfy/model_base.py +++ b/comfy/model_base.py @@ -926,7 +926,7 @@ class Flux(BaseModel): out = {} ref_latents = kwargs.get("reference_latents", None) if ref_latents is not None: - out['ref_latents'] = list([1, 16, sum(map(lambda a: math.prod(a.size()), ref_latents)) // 16]) + out['ref_latents'] = list([1, 16, sum(map(lambda a: math.prod(a.size()[2:]), ref_latents))]) return out class Flux2(Flux): From 8402c8700a29a97bc5d706d6a0b14c41bc2c2d8a Mon Sep 17 00:00:00 2001 From: comfyanonymous Date: Wed, 26 Nov 2025 02:41:13 -0500 Subject: [PATCH 169/647] ComfyUI version v0.3.75 --- comfyui_version.py | 2 +- pyproject.toml | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/comfyui_version.py b/comfyui_version.py index b565c7367..fa4b4f4b0 100644 --- a/comfyui_version.py +++ b/comfyui_version.py @@ -1,3 +1,3 @@ # This file is automatically generated by the build process when version is # updated in pyproject.toml. -__version__ = "0.3.74" +__version__ = "0.3.75" diff --git a/pyproject.toml b/pyproject.toml index ccf0fcdb9..9009e65fe 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -1,6 +1,6 @@ [project] name = "ComfyUI" -version = "0.3.74" +version = "0.3.75" readme = "README.md" license = { file = "LICENSE" } requires-python = ">=3.9" From f16219e3aadcb7a301a1a313ab8989c3ebe53764 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 26 Nov 2025 01:00:43 -0800 Subject: [PATCH 170/647] Add cheap latent preview for flux 2. (#10907) Thank you to the person who calculated them. You saved me a percent of my time. --- comfy/latent_formats.py | 40 ++++++++++++++++++++++++++++++++++++++++ latent_preview.py | 7 +++++-- 2 files changed, 45 insertions(+), 2 deletions(-) diff --git a/comfy/latent_formats.py b/comfy/latent_formats.py index e98c7d6d8..8e110f45d 100644 --- a/comfy/latent_formats.py +++ b/comfy/latent_formats.py @@ -6,6 +6,7 @@ class LatentFormat: latent_dimensions = 2 latent_rgb_factors = None latent_rgb_factors_bias = None + latent_rgb_factors_reshape = None taesd_decoder_name = None def process_in(self, latent): @@ -181,6 +182,45 @@ class Flux(SD3): class Flux2(LatentFormat): latent_channels = 128 + def __init__(self): + self.latent_rgb_factors =[ + [0.0058, 0.0113, 0.0073], + [0.0495, 0.0443, 0.0836], + [-0.0099, 0.0096, 0.0644], + [0.2144, 0.3009, 0.3652], + [0.0166, -0.0039, -0.0054], + [0.0157, 0.0103, -0.0160], + [-0.0398, 0.0902, -0.0235], + [-0.0052, 0.0095, 0.0109], + [-0.3527, -0.2712, -0.1666], + [-0.0301, -0.0356, -0.0180], + [-0.0107, 0.0078, 0.0013], + [0.0746, 0.0090, -0.0941], + [0.0156, 0.0169, 0.0070], + [-0.0034, -0.0040, -0.0114], + [0.0032, 0.0181, 0.0080], + [-0.0939, -0.0008, 0.0186], + [0.0018, 0.0043, 0.0104], + [0.0284, 0.0056, -0.0127], + [-0.0024, -0.0022, -0.0030], + [0.1207, -0.0026, 0.0065], + [0.0128, 0.0101, 0.0142], + [0.0137, -0.0072, -0.0007], + [0.0095, 0.0092, -0.0059], + [0.0000, -0.0077, -0.0049], + [-0.0465, -0.0204, -0.0312], + [0.0095, 0.0012, -0.0066], + [0.0290, -0.0034, 0.0025], + [0.0220, 0.0169, -0.0048], + [-0.0332, -0.0457, -0.0468], + [-0.0085, 0.0389, 0.0609], + [-0.0076, 0.0003, -0.0043], + [-0.0111, -0.0460, -0.0614], + ] + + self.latent_rgb_factors_bias = [-0.0329, -0.0718, -0.0851] + self.latent_rgb_factors_reshape = lambda t: t.reshape(t.shape[0], 32, 2, 2, t.shape[-2], t.shape[-1]).permute(0, 1, 4, 2, 5, 3).reshape(t.shape[0], 32, t.shape[-2] * 2, t.shape[-1] * 2) + def process_in(self, latent): return latent diff --git a/latent_preview.py b/latent_preview.py index 95d3cb733..ddf6dcf49 100644 --- a/latent_preview.py +++ b/latent_preview.py @@ -37,13 +37,16 @@ class TAESDPreviewerImpl(LatentPreviewer): class Latent2RGBPreviewer(LatentPreviewer): - def __init__(self, latent_rgb_factors, latent_rgb_factors_bias=None): + def __init__(self, latent_rgb_factors, latent_rgb_factors_bias=None, latent_rgb_factors_reshape=None): self.latent_rgb_factors = torch.tensor(latent_rgb_factors, device="cpu").transpose(0, 1) self.latent_rgb_factors_bias = None if latent_rgb_factors_bias is not None: self.latent_rgb_factors_bias = torch.tensor(latent_rgb_factors_bias, device="cpu") + self.latent_rgb_factors_reshape = latent_rgb_factors_reshape def decode_latent_to_preview(self, x0): + if self.latent_rgb_factors_reshape is not None: + x0 = self.latent_rgb_factors_reshape(x0) self.latent_rgb_factors = self.latent_rgb_factors.to(dtype=x0.dtype, device=x0.device) if self.latent_rgb_factors_bias is not None: self.latent_rgb_factors_bias = self.latent_rgb_factors_bias.to(dtype=x0.dtype, device=x0.device) @@ -85,7 +88,7 @@ def get_previewer(device, latent_format): if previewer is None: if latent_format.latent_rgb_factors is not None: - previewer = Latent2RGBPreviewer(latent_format.latent_rgb_factors, latent_format.latent_rgb_factors_bias) + previewer = Latent2RGBPreviewer(latent_format.latent_rgb_factors, latent_format.latent_rgb_factors_bias, latent_format.latent_rgb_factors_reshape) return previewer def prepare_callback(model, steps, x0_output_dict=None): From 8938aa3f3064415758fa8f3a628476535a676183 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Wed, 26 Nov 2025 19:14:02 +0200 Subject: [PATCH 171/647] add Veo3 First-Last-Frame node (#10878) --- comfy_api_nodes/apis/veo_api.py | 38 +++----- comfy_api_nodes/nodes_veo2.py | 155 ++++++++++++++++++++++++++++++++ 2 files changed, 168 insertions(+), 25 deletions(-) diff --git a/comfy_api_nodes/apis/veo_api.py b/comfy_api_nodes/apis/veo_api.py index a55137afb..8328d1aa4 100644 --- a/comfy_api_nodes/apis/veo_api.py +++ b/comfy_api_nodes/apis/veo_api.py @@ -1,34 +1,21 @@ -from typing import Optional, Union -from enum import Enum +from typing import Optional from pydantic import BaseModel, Field -class Image2(BaseModel): - bytesBase64Encoded: str - gcsUri: Optional[str] = None - mimeType: Optional[str] = None +class VeoRequestInstanceImage(BaseModel): + bytesBase64Encoded: str | None = Field(None) + gcsUri: str | None = Field(None) + mimeType: str | None = Field(None) -class Image3(BaseModel): - bytesBase64Encoded: Optional[str] = None - gcsUri: str - mimeType: Optional[str] = None - - -class Instance1(BaseModel): - image: Optional[Union[Image2, Image3]] = Field( - None, description='Optional image to guide video generation' - ) +class VeoRequestInstance(BaseModel): + image: VeoRequestInstanceImage | None = Field(None) + lastFrame: VeoRequestInstanceImage | None = Field(None) prompt: str = Field(..., description='Text description of the video') -class PersonGeneration1(str, Enum): - ALLOW = 'ALLOW' - BLOCK = 'BLOCK' - - -class Parameters1(BaseModel): +class VeoRequestParameters(BaseModel): aspectRatio: Optional[str] = Field(None, examples=['16:9']) durationSeconds: Optional[int] = None enhancePrompt: Optional[bool] = None @@ -37,17 +24,18 @@ class Parameters1(BaseModel): description='Generate audio for the video. Only supported by veo 3 models.', ) negativePrompt: Optional[str] = None - personGeneration: Optional[PersonGeneration1] = None + personGeneration: str | None = Field(None, description="ALLOW or BLOCK") sampleCount: Optional[int] = None seed: Optional[int] = None storageUri: Optional[str] = Field( None, description='Optional Cloud Storage URI to upload the video' ) + resolution: str | None = Field(None) class VeoGenVidRequest(BaseModel): - instances: Optional[list[Instance1]] = None - parameters: Optional[Parameters1] = None + instances: list[VeoRequestInstance] | None = Field(None) + parameters: VeoRequestParameters | None = Field(None) class VeoGenVidResponse(BaseModel): diff --git a/comfy_api_nodes/nodes_veo2.py b/comfy_api_nodes/nodes_veo2.py index d37e9e9b4..a54dc13ab 100644 --- a/comfy_api_nodes/nodes_veo2.py +++ b/comfy_api_nodes/nodes_veo2.py @@ -1,6 +1,7 @@ import base64 from io import BytesIO +import torch from typing_extensions import override from comfy_api.input_impl.video_types import VideoFromFile @@ -10,6 +11,9 @@ from comfy_api_nodes.apis.veo_api import ( VeoGenVidPollResponse, VeoGenVidRequest, VeoGenVidResponse, + VeoRequestInstance, + VeoRequestInstanceImage, + VeoRequestParameters, ) from comfy_api_nodes.util import ( ApiEndpoint, @@ -346,12 +350,163 @@ class Veo3VideoGenerationNode(VeoVideoGenerationNode): ) +class Veo3FirstLastFrameNode(IO.ComfyNode): + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="Veo3FirstLastFrameNode", + display_name="Google Veo 3 First-Last-Frame to Video", + category="api node/video/Veo", + description="Generate video using prompt and first and last frames.", + inputs=[ + IO.String.Input( + "prompt", + multiline=True, + default="", + tooltip="Text description of the video", + ), + IO.String.Input( + "negative_prompt", + multiline=True, + default="", + tooltip="Negative text prompt to guide what to avoid in the video", + ), + IO.Combo.Input("resolution", options=["720p", "1080p"]), + IO.Combo.Input( + "aspect_ratio", + options=["16:9", "9:16"], + default="16:9", + tooltip="Aspect ratio of the output video", + ), + IO.Int.Input( + "duration", + default=8, + min=4, + max=8, + step=2, + display_mode=IO.NumberDisplay.slider, + tooltip="Duration of the output video in seconds", + ), + IO.Int.Input( + "seed", + default=0, + min=0, + max=0xFFFFFFFF, + step=1, + display_mode=IO.NumberDisplay.number, + control_after_generate=True, + tooltip="Seed for video generation", + ), + IO.Image.Input("first_frame", tooltip="Start frame"), + IO.Image.Input("last_frame", tooltip="End frame"), + IO.Combo.Input( + "model", + options=["veo-3.1-generate", "veo-3.1-fast-generate"], + default="veo-3.1-fast-generate", + ), + IO.Boolean.Input( + "generate_audio", + default=True, + tooltip="Generate audio for the video.", + ), + ], + outputs=[ + IO.Video.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + ) + + @classmethod + async def execute( + cls, + prompt: str, + negative_prompt: str, + resolution: str, + aspect_ratio: str, + duration: int, + seed: int, + first_frame: torch.Tensor, + last_frame: torch.Tensor, + model: str, + generate_audio: bool, + ): + model = MODELS_MAP[model] + initial_response = await sync_op( + cls, + ApiEndpoint(path=f"/proxy/veo/{model}/generate", method="POST"), + response_model=VeoGenVidResponse, + data=VeoGenVidRequest( + instances=[ + VeoRequestInstance( + prompt=prompt, + image=VeoRequestInstanceImage( + bytesBase64Encoded=tensor_to_base64_string(first_frame), mimeType="image/png" + ), + lastFrame=VeoRequestInstanceImage( + bytesBase64Encoded=tensor_to_base64_string(last_frame), mimeType="image/png" + ), + ), + ], + parameters=VeoRequestParameters( + aspectRatio=aspect_ratio, + personGeneration="ALLOW", + durationSeconds=duration, + enhancePrompt=True, # cannot be False for Veo3 + seed=seed, + generateAudio=generate_audio, + negativePrompt=negative_prompt, + resolution=resolution, + ), + ), + ) + poll_response = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/veo/{model}/poll", method="POST"), + response_model=VeoGenVidPollResponse, + status_extractor=lambda r: "completed" if r.done else "pending", + data=VeoGenVidPollRequest( + operationName=initial_response.name, + ), + poll_interval=5.0, + estimated_duration=AVERAGE_DURATION_VIDEO_GEN, + ) + + if poll_response.error: + raise Exception(f"Veo API error: {poll_response.error.message} (code: {poll_response.error.code})") + + response = poll_response.response + filtered_count = response.raiMediaFilteredCount + if filtered_count: + reasons = response.raiMediaFilteredReasons or [] + reason_part = f": {reasons[0]}" if reasons else "" + raise Exception( + f"Content blocked by Google's Responsible AI filters{reason_part} " + f"({filtered_count} video{'s' if filtered_count != 1 else ''} filtered)." + ) + + if response.videos: + video = response.videos[0] + if video.bytesBase64Encoded: + return IO.NodeOutput(VideoFromFile(BytesIO(base64.b64decode(video.bytesBase64Encoded)))) + if video.gcsUri: + return IO.NodeOutput(await download_url_to_video_output(video.gcsUri)) + raise Exception("Video returned but no data or URL was provided") + raise Exception("Video generation completed but no video was returned") + + class VeoExtension(ComfyExtension): @override async def get_node_list(self) -> list[type[IO.ComfyNode]]: return [ VeoVideoGenerationNode, Veo3VideoGenerationNode, + Veo3FirstLastFrameNode, ] From 1105e0d139001ad602d0f883406bfce41e54ae67 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Wed, 26 Nov 2025 19:23:14 +0200 Subject: [PATCH 172/647] improve UX for batch uploads in upload_images_to_comfyapi (#10913) --- comfy_api_nodes/util/upload_helpers.py | 43 +++++++++++++------------- 1 file changed, 22 insertions(+), 21 deletions(-) diff --git a/comfy_api_nodes/util/upload_helpers.py b/comfy_api_nodes/util/upload_helpers.py index 632450d9b..b9019841f 100644 --- a/comfy_api_nodes/util/upload_helpers.py +++ b/comfy_api_nodes/util/upload_helpers.py @@ -4,7 +4,7 @@ import logging import time import uuid from io import BytesIO -from typing import Optional, Union +from typing import Optional from urllib.parse import urlparse import aiohttp @@ -48,8 +48,9 @@ async def upload_images_to_comfyapi( image: torch.Tensor, *, max_images: int = 8, - mime_type: Optional[str] = None, - wait_label: Optional[str] = "Uploading", + mime_type: str | None = None, + wait_label: str | None = "Uploading", + show_batch_index: bool = True, ) -> list[str]: """ Uploads images to ComfyUI API and returns download URLs. @@ -59,11 +60,18 @@ async def upload_images_to_comfyapi( download_urls: list[str] = [] is_batch = len(image.shape) > 3 batch_len = image.shape[0] if is_batch else 1 + num_to_upload = min(batch_len, max_images) + batch_start_ts = time.monotonic() - for idx in range(min(batch_len, max_images)): + for idx in range(num_to_upload): tensor = image[idx] if is_batch else image img_io = tensor_to_bytesio(tensor, mime_type=mime_type) - url = await upload_file_to_comfyapi(cls, img_io, img_io.name, mime_type, wait_label) + + effective_label = wait_label + if wait_label and show_batch_index and num_to_upload > 1: + effective_label = f"{wait_label} ({idx + 1}/{num_to_upload})" + + url = await upload_file_to_comfyapi(cls, img_io, img_io.name, mime_type, effective_label, batch_start_ts) download_urls.append(url) return download_urls @@ -126,8 +134,9 @@ async def upload_file_to_comfyapi( cls: type[IO.ComfyNode], file_bytes_io: BytesIO, filename: str, - upload_mime_type: Optional[str], - wait_label: Optional[str] = "Uploading", + upload_mime_type: str | None, + wait_label: str | None = "Uploading", + progress_origin_ts: float | None = None, ) -> str: """Uploads a single file to ComfyUI API and returns its download URL.""" if upload_mime_type is None: @@ -148,6 +157,7 @@ async def upload_file_to_comfyapi( file_bytes_io, content_type=upload_mime_type, wait_label=wait_label, + progress_origin_ts=progress_origin_ts, ) return create_resp.download_url @@ -155,27 +165,18 @@ async def upload_file_to_comfyapi( async def upload_file( cls: type[IO.ComfyNode], upload_url: str, - file: Union[BytesIO, str], + file: BytesIO | str, *, - content_type: Optional[str] = None, + content_type: str | None = None, max_retries: int = 3, retry_delay: float = 1.0, retry_backoff: float = 2.0, - wait_label: Optional[str] = None, + wait_label: str | None = None, + progress_origin_ts: float | None = None, ) -> None: """ Upload a file to a signed URL (e.g., S3 pre-signed PUT) with retries, Comfy progress display, and interruption. - Args: - cls: Node class (provides auth context + UI progress hooks). - upload_url: Pre-signed PUT URL. - file: BytesIO or path string. - content_type: Explicit MIME type. If None, we *suppress* Content-Type. - max_retries: Maximum retry attempts. - retry_delay: Initial delay in seconds. - retry_backoff: Exponential backoff factor. - wait_label: Progress label shown in Comfy UI. - Raises: ProcessingInterrupted, LocalNetworkError, ApiServerError, Exception """ @@ -198,7 +199,7 @@ async def upload_file( attempt = 0 delay = retry_delay - start_ts = time.monotonic() + start_ts = progress_origin_ts if progress_origin_ts is not None else time.monotonic() op_uuid = uuid.uuid4().hex[:8] while True: attempt += 1 From 8908ee262862f1252d1363d55c59872fb3361066 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Wed, 26 Nov 2025 20:38:30 +0200 Subject: [PATCH 173/647] fix(gemini): use first 10 images as fileData (URLs) and remaining images as inline base64 (#10918) --- comfy_api_nodes/apis/gemini_api.py | 6 ++++ comfy_api_nodes/nodes_gemini.py | 55 ++++++++++++++++++++---------- 2 files changed, 43 insertions(+), 18 deletions(-) diff --git a/comfy_api_nodes/apis/gemini_api.py b/comfy_api_nodes/apis/gemini_api.py index d34590d28..a380ecc86 100644 --- a/comfy_api_nodes/apis/gemini_api.py +++ b/comfy_api_nodes/apis/gemini_api.py @@ -58,8 +58,14 @@ class GeminiInlineData(BaseModel): mimeType: GeminiMimeType | None = Field(None) +class GeminiFileData(BaseModel): + fileUri: str | None = Field(None) + mimeType: GeminiMimeType | None = Field(None) + + class GeminiPart(BaseModel): inlineData: GeminiInlineData | None = Field(None) + fileData: GeminiFileData | None = Field(None) text: str | None = Field(None) diff --git a/comfy_api_nodes/nodes_gemini.py b/comfy_api_nodes/nodes_gemini.py index 938a20f84..976d9c225 100644 --- a/comfy_api_nodes/nodes_gemini.py +++ b/comfy_api_nodes/nodes_gemini.py @@ -20,6 +20,7 @@ from comfy_api.latest import IO, ComfyExtension, Input from comfy_api.util import VideoCodec, VideoContainer from comfy_api_nodes.apis.gemini_api import ( GeminiContent, + GeminiFileData, GeminiGenerateContentRequest, GeminiGenerateContentResponse, GeminiImageConfig, @@ -38,6 +39,7 @@ from comfy_api_nodes.util import ( get_number_of_images, sync_op, tensor_to_base64_string, + upload_images_to_comfyapi, validate_string, video_to_base64_string, ) @@ -68,24 +70,43 @@ class GeminiImageModel(str, Enum): gemini_2_5_flash_image = "gemini-2.5-flash-image" -def create_image_parts(image_input: torch.Tensor) -> list[GeminiPart]: - """ - Convert image tensor input to Gemini API compatible parts. - - Args: - image_input: Batch of image tensors from ComfyUI. - - Returns: - List of GeminiPart objects containing the encoded images. - """ +async def create_image_parts( + cls: type[IO.ComfyNode], + images: torch.Tensor, + image_limit: int = 0, +) -> list[GeminiPart]: image_parts: list[GeminiPart] = [] - for image_index in range(image_input.shape[0]): - image_as_b64 = tensor_to_base64_string(image_input[image_index].unsqueeze(0)) + if image_limit < 0: + raise ValueError("image_limit must be greater than or equal to 0 when creating Gemini image parts.") + total_images = get_number_of_images(images) + if total_images <= 0: + raise ValueError("No images provided to create_image_parts; at least one image is required.") + + # If image_limit == 0 --> use all images; otherwise clamp to image_limit. + effective_max = total_images if image_limit == 0 else min(total_images, image_limit) + + # Number of images we'll send as URLs (fileData) + num_url_images = min(effective_max, 10) # Vertex API max number of image links + reference_images_urls = await upload_images_to_comfyapi( + cls, + images, + max_images=num_url_images, + ) + for reference_image_url in reference_images_urls: + image_parts.append( + GeminiPart( + fileData=GeminiFileData( + mimeType=GeminiMimeType.image_png, + fileUri=reference_image_url, + ) + ) + ) + for idx in range(num_url_images, effective_max): image_parts.append( GeminiPart( inlineData=GeminiInlineData( mimeType=GeminiMimeType.image_png, - data=image_as_b64, + data=tensor_to_base64_string(images[idx]), ) ) ) @@ -338,8 +359,7 @@ class GeminiNode(IO.ComfyNode): # Add other modal parts if images is not None: - image_parts = create_image_parts(images) - parts.extend(image_parts) + parts.extend(await create_image_parts(cls, images)) if audio is not None: parts.extend(cls.create_audio_parts(audio)) if video is not None: @@ -562,8 +582,7 @@ class GeminiImage(IO.ComfyNode): image_config = GeminiImageConfig(aspectRatio=aspect_ratio) if images is not None: - image_parts = create_image_parts(images) - parts.extend(image_parts) + parts.extend(await create_image_parts(cls, images)) if files is not None: parts.extend(files) @@ -702,7 +721,7 @@ class GeminiImage2(IO.ComfyNode): if images is not None: if get_number_of_images(images) > 14: raise ValueError("The current maximum number of supported images is 14.") - parts.extend(create_image_parts(images)) + parts.extend(await create_image_parts(cls, images)) if files is not None: parts.extend(files) From 234c3dc85f7e61a537bbf6d8999c5880c5e0b746 Mon Sep 17 00:00:00 2001 From: Christian Byrne Date: Wed, 26 Nov 2025 11:58:08 -0800 Subject: [PATCH 174/647] Bump frontend to 1.32.9 (#10867) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 5f20816d6..9291552d3 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,4 +1,4 @@ -comfyui-frontend-package==1.30.6 +comfyui-frontend-package==1.32.9 comfyui-workflow-templates==0.7.20 comfyui-embedded-docs==0.3.1 torch From 58c6ed541d5aaf6d9b12f63bc23c33164e1cf7a3 Mon Sep 17 00:00:00 2001 From: Terry Jia Date: Wed, 26 Nov 2025 14:58:27 -0500 Subject: [PATCH 175/647] Merge 3d animation node (#10025) --- comfy_extras/nodes_load_3d.py | 110 +++++++--------------------------- 1 file changed, 23 insertions(+), 87 deletions(-) diff --git a/comfy_extras/nodes_load_3d.py b/comfy_extras/nodes_load_3d.py index 899608149..54c66ef68 100644 --- a/comfy_extras/nodes_load_3d.py +++ b/comfy_extras/nodes_load_3d.py @@ -7,6 +7,10 @@ from comfy_api.input_impl import VideoFromFile from pathlib import Path +from PIL import Image +import numpy as np + +import uuid def normalize_path(path): return path.replace('\\', '/') @@ -34,58 +38,6 @@ class Load3D(): "height": ("INT", {"default": 1024, "min": 1, "max": 4096, "step": 1}), }} - RETURN_TYPES = ("IMAGE", "MASK", "STRING", "IMAGE", "IMAGE", "LOAD3D_CAMERA", IO.VIDEO) - RETURN_NAMES = ("image", "mask", "mesh_path", "normal", "lineart", "camera_info", "recording_video") - - FUNCTION = "process" - EXPERIMENTAL = True - - CATEGORY = "3d" - - def process(self, model_file, image, **kwargs): - image_path = folder_paths.get_annotated_filepath(image['image']) - mask_path = folder_paths.get_annotated_filepath(image['mask']) - normal_path = folder_paths.get_annotated_filepath(image['normal']) - lineart_path = folder_paths.get_annotated_filepath(image['lineart']) - - load_image_node = nodes.LoadImage() - output_image, ignore_mask = load_image_node.load_image(image=image_path) - ignore_image, output_mask = load_image_node.load_image(image=mask_path) - normal_image, ignore_mask2 = load_image_node.load_image(image=normal_path) - lineart_image, ignore_mask3 = load_image_node.load_image(image=lineart_path) - - video = None - - if image['recording'] != "": - recording_video_path = folder_paths.get_annotated_filepath(image['recording']) - - video = VideoFromFile(recording_video_path) - - return output_image, output_mask, model_file, normal_image, lineart_image, image['camera_info'], video - -class Load3DAnimation(): - @classmethod - def INPUT_TYPES(s): - input_dir = os.path.join(folder_paths.get_input_directory(), "3d") - - os.makedirs(input_dir, exist_ok=True) - - input_path = Path(input_dir) - base_path = Path(folder_paths.get_input_directory()) - - files = [ - normalize_path(str(file_path.relative_to(base_path))) - for file_path in input_path.rglob("*") - if file_path.suffix.lower() in {'.gltf', '.glb', '.fbx'} - ] - - return {"required": { - "model_file": (sorted(files), {"file_upload": True}), - "image": ("LOAD_3D_ANIMATION", {}), - "width": ("INT", {"default": 1024, "min": 1, "max": 4096, "step": 1}), - "height": ("INT", {"default": 1024, "min": 1, "max": 4096, "step": 1}), - }} - RETURN_TYPES = ("IMAGE", "MASK", "STRING", "IMAGE", "LOAD3D_CAMERA", IO.VIDEO) RETURN_NAMES = ("image", "mask", "mesh_path", "normal", "camera_info", "recording_video") @@ -120,7 +72,8 @@ class Preview3D(): "model_file": ("STRING", {"default": "", "multiline": False}), }, "optional": { - "camera_info": ("LOAD3D_CAMERA", {}) + "camera_info": ("LOAD3D_CAMERA", {}), + "bg_image": ("IMAGE", {}) }} OUTPUT_NODE = True @@ -133,50 +86,33 @@ class Preview3D(): def process(self, model_file, **kwargs): camera_info = kwargs.get("camera_info", None) + bg_image = kwargs.get("bg_image", None) + + bg_image_path = None + if bg_image is not None: + + img_array = (bg_image[0].cpu().numpy() * 255).astype(np.uint8) + img = Image.fromarray(img_array) + + temp_dir = folder_paths.get_temp_directory() + filename = f"bg_{uuid.uuid4().hex}.png" + bg_image_path = os.path.join(temp_dir, filename) + img.save(bg_image_path, compress_level=1) + + bg_image_path = f"temp/{filename}" return { "ui": { - "result": [model_file, camera_info] - } - } - -class Preview3DAnimation(): - @classmethod - def INPUT_TYPES(s): - return {"required": { - "model_file": ("STRING", {"default": "", "multiline": False}), - }, - "optional": { - "camera_info": ("LOAD3D_CAMERA", {}) - }} - - OUTPUT_NODE = True - RETURN_TYPES = () - - CATEGORY = "3d" - - FUNCTION = "process" - EXPERIMENTAL = True - - def process(self, model_file, **kwargs): - camera_info = kwargs.get("camera_info", None) - - return { - "ui": { - "result": [model_file, camera_info] + "result": [model_file, camera_info, bg_image_path] } } NODE_CLASS_MAPPINGS = { "Load3D": Load3D, - "Load3DAnimation": Load3DAnimation, "Preview3D": Preview3D, - "Preview3DAnimation": Preview3DAnimation } NODE_DISPLAY_NAME_MAPPINGS = { - "Load3D": "Load 3D", - "Load3DAnimation": "Load 3D - Animation", - "Preview3D": "Preview 3D", - "Preview3DAnimation": "Preview 3D - Animation" + "Load3D": "Load 3D & Animation", + "Preview3D": "Preview 3D & Animation", } From 55f654db3ddaf5a10ac6dbe79774c23c350d279d Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 26 Nov 2025 12:16:40 -0800 Subject: [PATCH 176/647] Fix the CSP offline feature. (#10923) --- server.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/server.py b/server.py index 0fd2e49e3..fca5050bd 100644 --- a/server.py +++ b/server.py @@ -174,7 +174,7 @@ def create_block_external_middleware(): else: response = await handler(request) - response.headers['Content-Security-Policy'] = "default-src 'self'; script-src 'self' 'unsafe-inline' blob:; style-src 'self' 'unsafe-inline'; img-src 'self' data: blob:; font-src 'self'; connect-src 'self'; frame-src 'self'; object-src 'self';" + response.headers['Content-Security-Policy'] = "default-src 'self'; script-src 'self' 'unsafe-inline' 'unsafe-eval' blob:; style-src 'self' 'unsafe-inline'; img-src 'self' data: blob:; font-src 'self'; connect-src 'self'; frame-src 'self'; object-src 'self';" return response return block_external_middleware From dd41b745497cdbbafb0bd745f590726b0e41f9f3 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 26 Nov 2025 12:36:38 -0800 Subject: [PATCH 177/647] Add Z Image to readme. (#10924) --- README.md | 1 + 1 file changed, 1 insertion(+) diff --git a/README.md b/README.md index b9300ab07..91fb510e1 100644 --- a/README.md +++ b/README.md @@ -68,6 +68,7 @@ See what ComfyUI can do with the [example workflows](https://comfyanonymous.gith - [Qwen Image](https://comfyanonymous.github.io/ComfyUI_examples/qwen_image/) - [Hunyuan Image 2.1](https://comfyanonymous.github.io/ComfyUI_examples/hunyuan_image/) - [Flux 2](https://comfyanonymous.github.io/ComfyUI_examples/flux2/) + - [Z Image](https://comfyanonymous.github.io/ComfyUI_examples/z_image/) - Image Editing Models - [Omnigen 2](https://comfyanonymous.github.io/ComfyUI_examples/omnigen/) - [Flux Kontext](https://comfyanonymous.github.io/ComfyUI_examples/flux/#flux-kontext-image-editing-model) From d8433c63fdacef24f40da401b02ebba272bf1fbb Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Thu, 27 Nov 2025 00:42:01 +0200 Subject: [PATCH 178/647] chore(api-nodes): remove chat widgets from OpenAI/Gemini nodes (#10861) --- comfy_api_nodes/nodes_gemini.py | 77 +-------------------------------- comfy_api_nodes/nodes_openai.py | 46 ++++---------------- 2 files changed, 11 insertions(+), 112 deletions(-) diff --git a/comfy_api_nodes/nodes_gemini.py b/comfy_api_nodes/nodes_gemini.py index 976d9c225..08f7b0f64 100644 --- a/comfy_api_nodes/nodes_gemini.py +++ b/comfy_api_nodes/nodes_gemini.py @@ -4,10 +4,7 @@ See: https://cloud.google.com/vertex-ai/generative-ai/docs/model-reference/infer """ import base64 -import json import os -import time -import uuid from enum import Enum from io import BytesIO from typing import Literal @@ -43,7 +40,6 @@ from comfy_api_nodes.util import ( validate_string, video_to_base64_string, ) -from server import PromptServer GEMINI_BASE_ENDPOINT = "/proxy/vertexai/gemini" GEMINI_MAX_INPUT_FILE_SIZE = 20 * 1024 * 1024 # 20 MB @@ -384,29 +380,6 @@ class GeminiNode(IO.ComfyNode): ) output_text = get_text_from_response(response) - if output_text: - # Not a true chat history like the OpenAI Chat node. It is emulated so the frontend can show a copy button. - render_spec = { - "node_id": cls.hidden.unique_id, - "component": "ChatHistoryWidget", - "props": { - "history": json.dumps( - [ - { - "prompt": prompt, - "response": output_text, - "response_id": str(uuid.uuid4()), - "timestamp": time.time(), - } - ] - ), - }, - } - PromptServer.instance.send_sync( - "display_component", - render_spec, - ) - return IO.NodeOutput(output_text or "Empty response from Gemini model...") @@ -601,30 +574,7 @@ class GeminiImage(IO.ComfyNode): response_model=GeminiGenerateContentResponse, price_extractor=calculate_tokens_price, ) - - output_text = get_text_from_response(response) - if output_text: - render_spec = { - "node_id": cls.hidden.unique_id, - "component": "ChatHistoryWidget", - "props": { - "history": json.dumps( - [ - { - "prompt": prompt, - "response": output_text, - "response_id": str(uuid.uuid4()), - "timestamp": time.time(), - } - ] - ), - }, - } - PromptServer.instance.send_sync( - "display_component", - render_spec, - ) - return IO.NodeOutput(get_image_from_response(response), output_text) + return IO.NodeOutput(get_image_from_response(response), get_text_from_response(response)) class GeminiImage2(IO.ComfyNode): @@ -744,30 +694,7 @@ class GeminiImage2(IO.ComfyNode): response_model=GeminiGenerateContentResponse, price_extractor=calculate_tokens_price, ) - - output_text = get_text_from_response(response) - if output_text: - render_spec = { - "node_id": cls.hidden.unique_id, - "component": "ChatHistoryWidget", - "props": { - "history": json.dumps( - [ - { - "prompt": prompt, - "response": output_text, - "response_id": str(uuid.uuid4()), - "timestamp": time.time(), - } - ] - ), - }, - } - PromptServer.instance.send_sync( - "display_component", - render_spec, - ) - return IO.NodeOutput(get_image_from_response(response), output_text) + return IO.NodeOutput(get_image_from_response(response), get_text_from_response(response)) class GeminiExtension(ComfyExtension): diff --git a/comfy_api_nodes/nodes_openai.py b/comfy_api_nodes/nodes_openai.py index acf35d276..c8da5464b 100644 --- a/comfy_api_nodes/nodes_openai.py +++ b/comfy_api_nodes/nodes_openai.py @@ -1,15 +1,10 @@ from io import BytesIO -from typing import Optional, Union -import json import os -import time -import uuid from enum import Enum from inspect import cleandoc import numpy as np import torch from PIL import Image -from server import PromptServer import folder_paths import base64 from comfy_api.latest import IO, ComfyExtension @@ -587,11 +582,11 @@ class OpenAIChatNode(IO.ComfyNode): def create_input_message_contents( cls, prompt: str, - image: Optional[torch.Tensor] = None, - files: Optional[list[InputFileContent]] = None, + image: torch.Tensor | None = None, + files: list[InputFileContent] | None = None, ) -> InputMessageContentList: """Create a list of input message contents from prompt and optional image.""" - content_list: list[Union[InputContent, InputTextContent, InputImageContent, InputFileContent]] = [ + content_list: list[InputContent | InputTextContent | InputImageContent | InputFileContent] = [ InputTextContent(text=prompt, type="input_text"), ] if image is not None: @@ -617,9 +612,9 @@ class OpenAIChatNode(IO.ComfyNode): prompt: str, persist_context: bool = False, model: SupportedOpenAIModel = SupportedOpenAIModel.gpt_5.value, - images: Optional[torch.Tensor] = None, - files: Optional[list[InputFileContent]] = None, - advanced_options: Optional[CreateModelResponseProperties] = None, + images: torch.Tensor | None = None, + files: list[InputFileContent] | None = None, + advanced_options: CreateModelResponseProperties | None = None, ) -> IO.NodeOutput: validate_string(prompt, strip_whitespace=False) @@ -660,30 +655,7 @@ class OpenAIChatNode(IO.ComfyNode): status_extractor=lambda response: response.status, completed_statuses=["incomplete", "completed"] ) - output_text = cls.get_text_from_message_content(cls.get_message_content_from_response(result_response)) - - # Update history - render_spec = { - "node_id": cls.hidden.unique_id, - "component": "ChatHistoryWidget", - "props": { - "history": json.dumps( - [ - { - "prompt": prompt, - "response": output_text, - "response_id": str(uuid.uuid4()), - "timestamp": time.time(), - } - ] - ), - }, - } - PromptServer.instance.send_sync( - "display_component", - render_spec, - ) - return IO.NodeOutput(output_text) + return IO.NodeOutput(cls.get_text_from_message_content(cls.get_message_content_from_response(result_response))) class OpenAIInputFiles(IO.ComfyNode): @@ -790,8 +762,8 @@ class OpenAIChatConfig(IO.ComfyNode): def execute( cls, truncation: bool, - instructions: Optional[str] = None, - max_output_tokens: Optional[int] = None, + instructions: str | None = None, + max_output_tokens: int | None = None, ) -> IO.NodeOutput: """ Configure advanced options for the OpenAI Chat Node. From a2d60aad0f8e03657d501842460123f6eaaf6791 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Thu, 27 Nov 2025 00:55:31 +0200 Subject: [PATCH 179/647] convert nodes_customer_sampler.py to V3 schema (#10206) --- comfy_extras/nodes_custom_sampler.py | 1182 ++++++++++++++------------ 1 file changed, 633 insertions(+), 549 deletions(-) diff --git a/comfy_extras/nodes_custom_sampler.py b/comfy_extras/nodes_custom_sampler.py index d011f433b..fbb080886 100644 --- a/comfy_extras/nodes_custom_sampler.py +++ b/comfy_extras/nodes_custom_sampler.py @@ -3,272 +3,312 @@ import comfy.samplers import comfy.sample from comfy.k_diffusion import sampling as k_diffusion_sampling from comfy.k_diffusion import sa_solver -from comfy.comfy_types import IO, ComfyNodeABC, InputTypeDict import latent_preview import torch import comfy.utils import node_helpers +from typing_extensions import override +from comfy_api.latest import ComfyExtension, io -class BasicScheduler: +class BasicScheduler(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"model": ("MODEL",), - "scheduler": (comfy.samplers.SCHEDULER_NAMES, ), - "steps": ("INT", {"default": 20, "min": 1, "max": 10000}), - "denoise": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0, "step": 0.01}), - } - } - RETURN_TYPES = ("SIGMAS",) - CATEGORY = "sampling/custom_sampling/schedulers" + def define_schema(cls): + return io.Schema( + node_id="BasicScheduler", + category="sampling/custom_sampling/schedulers", + inputs=[ + io.Model.Input("model"), + io.Combo.Input("scheduler", options=comfy.samplers.SCHEDULER_NAMES), + io.Int.Input("steps", default=20, min=1, max=10000), + io.Float.Input("denoise", default=1.0, min=0.0, max=1.0, step=0.01), + ], + outputs=[io.Sigmas.Output()] + ) - FUNCTION = "get_sigmas" - - def get_sigmas(self, model, scheduler, steps, denoise): + @classmethod + def execute(cls, model, scheduler, steps, denoise) -> io.NodeOutput: total_steps = steps if denoise < 1.0: if denoise <= 0.0: - return (torch.FloatTensor([]),) + return io.NodeOutput(torch.FloatTensor([])) total_steps = int(steps/denoise) sigmas = comfy.samplers.calculate_sigmas(model.get_model_object("model_sampling"), scheduler, total_steps).cpu() sigmas = sigmas[-(steps + 1):] - return (sigmas, ) + return io.NodeOutput(sigmas) + + get_sigmas = execute -class KarrasScheduler: +class KarrasScheduler(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"steps": ("INT", {"default": 20, "min": 1, "max": 10000}), - "sigma_max": ("FLOAT", {"default": 14.614642, "min": 0.0, "max": 5000.0, "step":0.01, "round": False}), - "sigma_min": ("FLOAT", {"default": 0.0291675, "min": 0.0, "max": 5000.0, "step":0.01, "round": False}), - "rho": ("FLOAT", {"default": 7.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), - } - } - RETURN_TYPES = ("SIGMAS",) - CATEGORY = "sampling/custom_sampling/schedulers" + def define_schema(cls): + return io.Schema( + node_id="KarrasScheduler", + category="sampling/custom_sampling/schedulers", + inputs=[ + io.Int.Input("steps", default=20, min=1, max=10000), + io.Float.Input("sigma_max", default=14.614642, min=0.0, max=5000.0, step=0.01, round=False), + io.Float.Input("sigma_min", default=0.0291675, min=0.0, max=5000.0, step=0.01, round=False), + io.Float.Input("rho", default=7.0, min=0.0, max=100.0, step=0.01, round=False), + ], + outputs=[io.Sigmas.Output()] + ) - FUNCTION = "get_sigmas" - - def get_sigmas(self, steps, sigma_max, sigma_min, rho): + @classmethod + def execute(cls, steps, sigma_max, sigma_min, rho) -> io.NodeOutput: sigmas = k_diffusion_sampling.get_sigmas_karras(n=steps, sigma_min=sigma_min, sigma_max=sigma_max, rho=rho) - return (sigmas, ) + return io.NodeOutput(sigmas) -class ExponentialScheduler: + get_sigmas = execute + +class ExponentialScheduler(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"steps": ("INT", {"default": 20, "min": 1, "max": 10000}), - "sigma_max": ("FLOAT", {"default": 14.614642, "min": 0.0, "max": 5000.0, "step":0.01, "round": False}), - "sigma_min": ("FLOAT", {"default": 0.0291675, "min": 0.0, "max": 5000.0, "step":0.01, "round": False}), - } - } - RETURN_TYPES = ("SIGMAS",) - CATEGORY = "sampling/custom_sampling/schedulers" + def define_schema(cls): + return io.Schema( + node_id="ExponentialScheduler", + category="sampling/custom_sampling/schedulers", + inputs=[ + io.Int.Input("steps", default=20, min=1, max=10000), + io.Float.Input("sigma_max", default=14.614642, min=0.0, max=5000.0, step=0.01, round=False), + io.Float.Input("sigma_min", default=0.0291675, min=0.0, max=5000.0, step=0.01, round=False), + ], + outputs=[io.Sigmas.Output()] + ) - FUNCTION = "get_sigmas" - - def get_sigmas(self, steps, sigma_max, sigma_min): + @classmethod + def execute(cls, steps, sigma_max, sigma_min) -> io.NodeOutput: sigmas = k_diffusion_sampling.get_sigmas_exponential(n=steps, sigma_min=sigma_min, sigma_max=sigma_max) - return (sigmas, ) + return io.NodeOutput(sigmas) -class PolyexponentialScheduler: + get_sigmas = execute + +class PolyexponentialScheduler(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"steps": ("INT", {"default": 20, "min": 1, "max": 10000}), - "sigma_max": ("FLOAT", {"default": 14.614642, "min": 0.0, "max": 5000.0, "step":0.01, "round": False}), - "sigma_min": ("FLOAT", {"default": 0.0291675, "min": 0.0, "max": 5000.0, "step":0.01, "round": False}), - "rho": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), - } - } - RETURN_TYPES = ("SIGMAS",) - CATEGORY = "sampling/custom_sampling/schedulers" + def define_schema(cls): + return io.Schema( + node_id="PolyexponentialScheduler", + category="sampling/custom_sampling/schedulers", + inputs=[ + io.Int.Input("steps", default=20, min=1, max=10000), + io.Float.Input("sigma_max", default=14.614642, min=0.0, max=5000.0, step=0.01, round=False), + io.Float.Input("sigma_min", default=0.0291675, min=0.0, max=5000.0, step=0.01, round=False), + io.Float.Input("rho", default=1.0, min=0.0, max=100.0, step=0.01, round=False), + ], + outputs=[io.Sigmas.Output()] + ) - FUNCTION = "get_sigmas" - - def get_sigmas(self, steps, sigma_max, sigma_min, rho): + @classmethod + def execute(cls, steps, sigma_max, sigma_min, rho) -> io.NodeOutput: sigmas = k_diffusion_sampling.get_sigmas_polyexponential(n=steps, sigma_min=sigma_min, sigma_max=sigma_max, rho=rho) - return (sigmas, ) + return io.NodeOutput(sigmas) -class LaplaceScheduler: + get_sigmas = execute + +class LaplaceScheduler(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"steps": ("INT", {"default": 20, "min": 1, "max": 10000}), - "sigma_max": ("FLOAT", {"default": 14.614642, "min": 0.0, "max": 5000.0, "step":0.01, "round": False}), - "sigma_min": ("FLOAT", {"default": 0.0291675, "min": 0.0, "max": 5000.0, "step":0.01, "round": False}), - "mu": ("FLOAT", {"default": 0.0, "min": -10.0, "max": 10.0, "step":0.1, "round": False}), - "beta": ("FLOAT", {"default": 0.5, "min": 0.0, "max": 10.0, "step":0.1, "round": False}), - } - } - RETURN_TYPES = ("SIGMAS",) - CATEGORY = "sampling/custom_sampling/schedulers" + def define_schema(cls): + return io.Schema( + node_id="LaplaceScheduler", + category="sampling/custom_sampling/schedulers", + inputs=[ + io.Int.Input("steps", default=20, min=1, max=10000), + io.Float.Input("sigma_max", default=14.614642, min=0.0, max=5000.0, step=0.01, round=False), + io.Float.Input("sigma_min", default=0.0291675, min=0.0, max=5000.0, step=0.01, round=False), + io.Float.Input("mu", default=0.0, min=-10.0, max=10.0, step=0.1, round=False), + io.Float.Input("beta", default=0.5, min=0.0, max=10.0, step=0.1, round=False), + ], + outputs=[io.Sigmas.Output()] + ) - FUNCTION = "get_sigmas" - - def get_sigmas(self, steps, sigma_max, sigma_min, mu, beta): + @classmethod + def execute(cls, steps, sigma_max, sigma_min, mu, beta) -> io.NodeOutput: sigmas = k_diffusion_sampling.get_sigmas_laplace(n=steps, sigma_min=sigma_min, sigma_max=sigma_max, mu=mu, beta=beta) - return (sigmas, ) + return io.NodeOutput(sigmas) + + get_sigmas = execute -class SDTurboScheduler: +class SDTurboScheduler(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"model": ("MODEL",), - "steps": ("INT", {"default": 1, "min": 1, "max": 10}), - "denoise": ("FLOAT", {"default": 1.0, "min": 0, "max": 1.0, "step": 0.01}), - } - } - RETURN_TYPES = ("SIGMAS",) - CATEGORY = "sampling/custom_sampling/schedulers" + def define_schema(cls): + return io.Schema( + node_id="SDTurboScheduler", + category="sampling/custom_sampling/schedulers", + inputs=[ + io.Model.Input("model"), + io.Int.Input("steps", default=1, min=1, max=10), + io.Float.Input("denoise", default=1.0, min=0, max=1.0, step=0.01), + ], + outputs=[io.Sigmas.Output()] + ) - FUNCTION = "get_sigmas" - - def get_sigmas(self, model, steps, denoise): + @classmethod + def execute(cls, model, steps, denoise) -> io.NodeOutput: start_step = 10 - int(10 * denoise) timesteps = torch.flip(torch.arange(1, 11) * 100 - 1, (0,))[start_step:start_step + steps] sigmas = model.get_model_object("model_sampling").sigma(timesteps) sigmas = torch.cat([sigmas, sigmas.new_zeros([1])]) - return (sigmas, ) + return io.NodeOutput(sigmas) -class BetaSamplingScheduler: + get_sigmas = execute + +class BetaSamplingScheduler(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"model": ("MODEL",), - "steps": ("INT", {"default": 20, "min": 1, "max": 10000}), - "alpha": ("FLOAT", {"default": 0.6, "min": 0.0, "max": 50.0, "step":0.01, "round": False}), - "beta": ("FLOAT", {"default": 0.6, "min": 0.0, "max": 50.0, "step":0.01, "round": False}), - } - } - RETURN_TYPES = ("SIGMAS",) - CATEGORY = "sampling/custom_sampling/schedulers" + def define_schema(cls): + return io.Schema( + node_id="BetaSamplingScheduler", + category="sampling/custom_sampling/schedulers", + inputs=[ + io.Model.Input("model"), + io.Int.Input("steps", default=20, min=1, max=10000), + io.Float.Input("alpha", default=0.6, min=0.0, max=50.0, step=0.01, round=False), + io.Float.Input("beta", default=0.6, min=0.0, max=50.0, step=0.01, round=False), + ], + outputs=[io.Sigmas.Output()] + ) - FUNCTION = "get_sigmas" - - def get_sigmas(self, model, steps, alpha, beta): + @classmethod + def execute(cls, model, steps, alpha, beta) -> io.NodeOutput: sigmas = comfy.samplers.beta_scheduler(model.get_model_object("model_sampling"), steps, alpha=alpha, beta=beta) - return (sigmas, ) + return io.NodeOutput(sigmas) -class VPScheduler: + get_sigmas = execute + +class VPScheduler(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"steps": ("INT", {"default": 20, "min": 1, "max": 10000}), - "beta_d": ("FLOAT", {"default": 19.9, "min": 0.0, "max": 5000.0, "step":0.01, "round": False}), #TODO: fix default values - "beta_min": ("FLOAT", {"default": 0.1, "min": 0.0, "max": 5000.0, "step":0.01, "round": False}), - "eps_s": ("FLOAT", {"default": 0.001, "min": 0.0, "max": 1.0, "step":0.0001, "round": False}), - } - } - RETURN_TYPES = ("SIGMAS",) - CATEGORY = "sampling/custom_sampling/schedulers" + def define_schema(cls): + return io.Schema( + node_id="VPScheduler", + category="sampling/custom_sampling/schedulers", + inputs=[ + io.Int.Input("steps", default=20, min=1, max=10000), + io.Float.Input("beta_d", default=19.9, min=0.0, max=5000.0, step=0.01, round=False), #TODO: fix default values + io.Float.Input("beta_min", default=0.1, min=0.0, max=5000.0, step=0.01, round=False), + io.Float.Input("eps_s", default=0.001, min=0.0, max=1.0, step=0.0001, round=False), + ], + outputs=[io.Sigmas.Output()] + ) - FUNCTION = "get_sigmas" - - def get_sigmas(self, steps, beta_d, beta_min, eps_s): + @classmethod + def execute(cls, steps, beta_d, beta_min, eps_s) -> io.NodeOutput: sigmas = k_diffusion_sampling.get_sigmas_vp(n=steps, beta_d=beta_d, beta_min=beta_min, eps_s=eps_s) - return (sigmas, ) + return io.NodeOutput(sigmas) -class SplitSigmas: + get_sigmas = execute + +class SplitSigmas(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"sigmas": ("SIGMAS", ), - "step": ("INT", {"default": 0, "min": 0, "max": 10000}), - } - } - RETURN_TYPES = ("SIGMAS","SIGMAS") - RETURN_NAMES = ("high_sigmas", "low_sigmas") - CATEGORY = "sampling/custom_sampling/sigmas" + def define_schema(cls): + return io.Schema( + node_id="SplitSigmas", + category="sampling/custom_sampling/sigmas", + inputs=[ + io.Sigmas.Input("sigmas"), + io.Int.Input("step", default=0, min=0, max=10000), + ], + outputs=[ + io.Sigmas.Output(display_name="high_sigmas"), + io.Sigmas.Output(display_name="low_sigmas"), + ] + ) - FUNCTION = "get_sigmas" - - def get_sigmas(self, sigmas, step): + @classmethod + def execute(cls, sigmas, step) -> io.NodeOutput: sigmas1 = sigmas[:step + 1] sigmas2 = sigmas[step:] - return (sigmas1, sigmas2) + return io.NodeOutput(sigmas1, sigmas2) -class SplitSigmasDenoise: + get_sigmas = execute + +class SplitSigmasDenoise(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"sigmas": ("SIGMAS", ), - "denoise": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0, "step": 0.01}), - } - } - RETURN_TYPES = ("SIGMAS","SIGMAS") - RETURN_NAMES = ("high_sigmas", "low_sigmas") - CATEGORY = "sampling/custom_sampling/sigmas" + def define_schema(cls): + return io.Schema( + node_id="SplitSigmasDenoise", + category="sampling/custom_sampling/sigmas", + inputs=[ + io.Sigmas.Input("sigmas"), + io.Float.Input("denoise", default=1.0, min=0.0, max=1.0, step=0.01), + ], + outputs=[ + io.Sigmas.Output(display_name="high_sigmas"), + io.Sigmas.Output(display_name="low_sigmas"), + ] + ) - FUNCTION = "get_sigmas" - - def get_sigmas(self, sigmas, denoise): + @classmethod + def execute(cls, sigmas, denoise) -> io.NodeOutput: steps = max(sigmas.shape[-1] - 1, 0) total_steps = round(steps * denoise) sigmas1 = sigmas[:-(total_steps)] sigmas2 = sigmas[-(total_steps + 1):] - return (sigmas1, sigmas2) + return io.NodeOutput(sigmas1, sigmas2) -class FlipSigmas: + get_sigmas = execute + +class FlipSigmas(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"sigmas": ("SIGMAS", ), - } - } - RETURN_TYPES = ("SIGMAS",) - CATEGORY = "sampling/custom_sampling/sigmas" + def define_schema(cls): + return io.Schema( + node_id="FlipSigmas", + category="sampling/custom_sampling/sigmas", + inputs=[io.Sigmas.Input("sigmas")], + outputs=[io.Sigmas.Output()] + ) - FUNCTION = "get_sigmas" - - def get_sigmas(self, sigmas): + @classmethod + def execute(cls, sigmas) -> io.NodeOutput: if len(sigmas) == 0: - return (sigmas,) + return io.NodeOutput(sigmas) sigmas = sigmas.flip(0) if sigmas[0] == 0: sigmas[0] = 0.0001 - return (sigmas,) + return io.NodeOutput(sigmas) -class SetFirstSigma: + get_sigmas = execute + +class SetFirstSigma(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"sigmas": ("SIGMAS", ), - "sigma": ("FLOAT", {"default": 136.0, "min": 0.0, "max": 20000.0, "step": 0.001, "round": False}), - } - } - RETURN_TYPES = ("SIGMAS",) - CATEGORY = "sampling/custom_sampling/sigmas" + def define_schema(cls): + return io.Schema( + node_id="SetFirstSigma", + category="sampling/custom_sampling/sigmas", + inputs=[ + io.Sigmas.Input("sigmas"), + io.Float.Input("sigma", default=136.0, min=0.0, max=20000.0, step=0.001, round=False), + ], + outputs=[io.Sigmas.Output()] + ) - FUNCTION = "set_first_sigma" - - def set_first_sigma(self, sigmas, sigma): + @classmethod + def execute(cls, sigmas, sigma) -> io.NodeOutput: sigmas = sigmas.clone() sigmas[0] = sigma - return (sigmas, ) + return io.NodeOutput(sigmas) -class ExtendIntermediateSigmas: + set_first_sigma = execute + +class ExtendIntermediateSigmas(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"sigmas": ("SIGMAS", ), - "steps": ("INT", {"default": 2, "min": 1, "max": 100}), - "start_at_sigma": ("FLOAT", {"default": -1.0, "min": -1.0, "max": 20000.0, "step": 0.01, "round": False}), - "end_at_sigma": ("FLOAT", {"default": 12.0, "min": 0.0, "max": 20000.0, "step": 0.01, "round": False}), - "spacing": (['linear', 'cosine', 'sine'],), - } - } - RETURN_TYPES = ("SIGMAS",) - CATEGORY = "sampling/custom_sampling/sigmas" + def define_schema(cls): + return io.Schema( + node_id="ExtendIntermediateSigmas", + category="sampling/custom_sampling/sigmas", + inputs=[ + io.Sigmas.Input("sigmas"), + io.Int.Input("steps", default=2, min=1, max=100), + io.Float.Input("start_at_sigma", default=-1.0, min=-1.0, max=20000.0, step=0.01, round=False), + io.Float.Input("end_at_sigma", default=12.0, min=0.0, max=20000.0, step=0.01, round=False), + io.Combo.Input("spacing", options=['linear', 'cosine', 'sine']), + ], + outputs=[io.Sigmas.Output()] + ) - FUNCTION = "extend" - - def extend(self, sigmas: torch.Tensor, steps: int, start_at_sigma: float, end_at_sigma: float, spacing: str): + @classmethod + def execute(cls, sigmas: torch.Tensor, steps: int, start_at_sigma: float, end_at_sigma: float, spacing: str) -> io.NodeOutput: if start_at_sigma < 0: start_at_sigma = float("inf") @@ -299,27 +339,27 @@ class ExtendIntermediateSigmas: extended_sigmas = torch.FloatTensor(extended_sigmas) - return (extended_sigmas,) + return io.NodeOutput(extended_sigmas) + + extend = execute -class SamplingPercentToSigma: +class SamplingPercentToSigma(io.ComfyNode): @classmethod - def INPUT_TYPES(cls) -> InputTypeDict: - return { - "required": { - "model": (IO.MODEL, {}), - "sampling_percent": (IO.FLOAT, {"default": 0.0, "min": 0.0, "max": 1.0, "step": 0.0001}), - "return_actual_sigma": (IO.BOOLEAN, {"default": False, "tooltip": "Return the actual sigma value instead of the value used for interval checks.\nThis only affects results at 0.0 and 1.0."}), - } - } + def define_schema(cls): + return io.Schema( + node_id="SamplingPercentToSigma", + category="sampling/custom_sampling/sigmas", + inputs=[ + io.Model.Input("model"), + io.Float.Input("sampling_percent", default=0.0, min=0.0, max=1.0, step=0.0001), + io.Boolean.Input("return_actual_sigma", default=False, tooltip="Return the actual sigma value instead of the value used for interval checks.\nThis only affects results at 0.0 and 1.0."), + ], + outputs=[io.Float.Output(display_name="sigma_value")] + ) - RETURN_TYPES = (IO.FLOAT,) - RETURN_NAMES = ("sigma_value",) - CATEGORY = "sampling/custom_sampling/sigmas" - - FUNCTION = "get_sigma" - - def get_sigma(self, model, sampling_percent, return_actual_sigma): + @classmethod + def execute(cls, model, sampling_percent, return_actual_sigma) -> io.NodeOutput: model_sampling = model.get_model_object("model_sampling") sigma_val = model_sampling.percent_to_sigma(sampling_percent) if return_actual_sigma: @@ -327,212 +367,234 @@ class SamplingPercentToSigma: sigma_val = model_sampling.sigma_max.item() elif sampling_percent == 1.0: sigma_val = model_sampling.sigma_min.item() - return (sigma_val,) + return io.NodeOutput(sigma_val) + + get_sigma = execute -class KSamplerSelect: +class KSamplerSelect(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"sampler_name": (comfy.samplers.SAMPLER_NAMES, ), - } - } - RETURN_TYPES = ("SAMPLER",) - CATEGORY = "sampling/custom_sampling/samplers" + def define_schema(cls): + return io.Schema( + node_id="KSamplerSelect", + category="sampling/custom_sampling/samplers", + inputs=[io.Combo.Input("sampler_name", options=comfy.samplers.SAMPLER_NAMES)], + outputs=[io.Sampler.Output()] + ) - FUNCTION = "get_sampler" - - def get_sampler(self, sampler_name): + @classmethod + def execute(cls, sampler_name) -> io.NodeOutput: sampler = comfy.samplers.sampler_object(sampler_name) - return (sampler, ) + return io.NodeOutput(sampler) -class SamplerDPMPP_3M_SDE: + get_sampler = execute + +class SamplerDPMPP_3M_SDE(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"eta": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), - "s_noise": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), - "noise_device": (['gpu', 'cpu'], ), - } - } - RETURN_TYPES = ("SAMPLER",) - CATEGORY = "sampling/custom_sampling/samplers" + def define_schema(cls): + return io.Schema( + node_id="SamplerDPMPP_3M_SDE", + category="sampling/custom_sampling/samplers", + inputs=[ + io.Float.Input("eta", default=1.0, min=0.0, max=100.0, step=0.01, round=False), + io.Float.Input("s_noise", default=1.0, min=0.0, max=100.0, step=0.01, round=False), + io.Combo.Input("noise_device", options=['gpu', 'cpu']), + ], + outputs=[io.Sampler.Output()] + ) - FUNCTION = "get_sampler" - - def get_sampler(self, eta, s_noise, noise_device): + @classmethod + def execute(cls, eta, s_noise, noise_device) -> io.NodeOutput: if noise_device == 'cpu': sampler_name = "dpmpp_3m_sde" else: sampler_name = "dpmpp_3m_sde_gpu" sampler = comfy.samplers.ksampler(sampler_name, {"eta": eta, "s_noise": s_noise}) - return (sampler, ) + return io.NodeOutput(sampler) -class SamplerDPMPP_2M_SDE: + get_sampler = execute + +class SamplerDPMPP_2M_SDE(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"solver_type": (['midpoint', 'heun'], ), - "eta": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), - "s_noise": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), - "noise_device": (['gpu', 'cpu'], ), - } - } - RETURN_TYPES = ("SAMPLER",) - CATEGORY = "sampling/custom_sampling/samplers" + def define_schema(cls): + return io.Schema( + node_id="SamplerDPMPP_2M_SDE", + category="sampling/custom_sampling/samplers", + inputs=[ + io.Combo.Input("solver_type", options=['midpoint', 'heun']), + io.Float.Input("eta", default=1.0, min=0.0, max=100.0, step=0.01, round=False), + io.Float.Input("s_noise", default=1.0, min=0.0, max=100.0, step=0.01, round=False), + io.Combo.Input("noise_device", options=['gpu', 'cpu']), + ], + outputs=[io.Sampler.Output()] + ) - FUNCTION = "get_sampler" - - def get_sampler(self, solver_type, eta, s_noise, noise_device): + @classmethod + def execute(cls, solver_type, eta, s_noise, noise_device) -> io.NodeOutput: if noise_device == 'cpu': sampler_name = "dpmpp_2m_sde" else: sampler_name = "dpmpp_2m_sde_gpu" sampler = comfy.samplers.ksampler(sampler_name, {"eta": eta, "s_noise": s_noise, "solver_type": solver_type}) - return (sampler, ) + return io.NodeOutput(sampler) + + get_sampler = execute -class SamplerDPMPP_SDE: +class SamplerDPMPP_SDE(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"eta": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), - "s_noise": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), - "r": ("FLOAT", {"default": 0.5, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), - "noise_device": (['gpu', 'cpu'], ), - } - } - RETURN_TYPES = ("SAMPLER",) - CATEGORY = "sampling/custom_sampling/samplers" + def define_schema(cls): + return io.Schema( + node_id="SamplerDPMPP_SDE", + category="sampling/custom_sampling/samplers", + inputs=[ + io.Float.Input("eta", default=1.0, min=0.0, max=100.0, step=0.01, round=False), + io.Float.Input("s_noise", default=1.0, min=0.0, max=100.0, step=0.01, round=False), + io.Float.Input("r", default=0.5, min=0.0, max=100.0, step=0.01, round=False), + io.Combo.Input("noise_device", options=['gpu', 'cpu']), + ], + outputs=[io.Sampler.Output()] + ) - FUNCTION = "get_sampler" - - def get_sampler(self, eta, s_noise, r, noise_device): + @classmethod + def execute(cls, eta, s_noise, r, noise_device) -> io.NodeOutput: if noise_device == 'cpu': sampler_name = "dpmpp_sde" else: sampler_name = "dpmpp_sde_gpu" sampler = comfy.samplers.ksampler(sampler_name, {"eta": eta, "s_noise": s_noise, "r": r}) - return (sampler, ) + return io.NodeOutput(sampler) -class SamplerDPMPP_2S_Ancestral: + get_sampler = execute + +class SamplerDPMPP_2S_Ancestral(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"eta": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), - "s_noise": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), - } - } - RETURN_TYPES = ("SAMPLER",) - CATEGORY = "sampling/custom_sampling/samplers" + def define_schema(cls): + return io.Schema( + node_id="SamplerDPMPP_2S_Ancestral", + category="sampling/custom_sampling/samplers", + inputs=[ + io.Float.Input("eta", default=1.0, min=0.0, max=100.0, step=0.01, round=False), + io.Float.Input("s_noise", default=1.0, min=0.0, max=100.0, step=0.01, round=False), + ], + outputs=[io.Sampler.Output()] + ) - FUNCTION = "get_sampler" - - def get_sampler(self, eta, s_noise): + @classmethod + def execute(cls, eta, s_noise) -> io.NodeOutput: sampler = comfy.samplers.ksampler("dpmpp_2s_ancestral", {"eta": eta, "s_noise": s_noise}) - return (sampler, ) + return io.NodeOutput(sampler) -class SamplerEulerAncestral: + get_sampler = execute + +class SamplerEulerAncestral(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"eta": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), - "s_noise": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), - } - } - RETURN_TYPES = ("SAMPLER",) - CATEGORY = "sampling/custom_sampling/samplers" + def define_schema(cls): + return io.Schema( + node_id="SamplerEulerAncestral", + category="sampling/custom_sampling/samplers", + inputs=[ + io.Float.Input("eta", default=1.0, min=0.0, max=100.0, step=0.01, round=False), + io.Float.Input("s_noise", default=1.0, min=0.0, max=100.0, step=0.01, round=False), + ], + outputs=[io.Sampler.Output()] + ) - FUNCTION = "get_sampler" - - def get_sampler(self, eta, s_noise): + @classmethod + def execute(cls, eta, s_noise) -> io.NodeOutput: sampler = comfy.samplers.ksampler("euler_ancestral", {"eta": eta, "s_noise": s_noise}) - return (sampler, ) + return io.NodeOutput(sampler) -class SamplerEulerAncestralCFGPP: + get_sampler = execute + +class SamplerEulerAncestralCFGPP(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return { - "required": { - "eta": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0, "step":0.01, "round": False}), - "s_noise": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step":0.01, "round": False}), - }} - RETURN_TYPES = ("SAMPLER",) - CATEGORY = "sampling/custom_sampling/samplers" + def define_schema(cls): + return io.Schema( + node_id="SamplerEulerAncestralCFGPP", + display_name="SamplerEulerAncestralCFG++", + category="sampling/custom_sampling/samplers", + inputs=[ + io.Float.Input("eta", default=1.0, min=0.0, max=1.0, step=0.01, round=False), + io.Float.Input("s_noise", default=1.0, min=0.0, max=10.0, step=0.01, round=False), + ], + outputs=[io.Sampler.Output()] + ) - FUNCTION = "get_sampler" - - def get_sampler(self, eta, s_noise): + @classmethod + def execute(cls, eta, s_noise) -> io.NodeOutput: sampler = comfy.samplers.ksampler( "euler_ancestral_cfg_pp", {"eta": eta, "s_noise": s_noise}) - return (sampler, ) + return io.NodeOutput(sampler) -class SamplerLMS: + get_sampler = execute + +class SamplerLMS(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"order": ("INT", {"default": 4, "min": 1, "max": 100}), - } - } - RETURN_TYPES = ("SAMPLER",) - CATEGORY = "sampling/custom_sampling/samplers" + def define_schema(cls): + return io.Schema( + node_id="SamplerLMS", + category="sampling/custom_sampling/samplers", + inputs=[io.Int.Input("order", default=4, min=1, max=100)], + outputs=[io.Sampler.Output()] + ) - FUNCTION = "get_sampler" - - def get_sampler(self, order): + @classmethod + def execute(cls, order) -> io.NodeOutput: sampler = comfy.samplers.ksampler("lms", {"order": order}) - return (sampler, ) + return io.NodeOutput(sampler) -class SamplerDPMAdaptative: + get_sampler = execute + +class SamplerDPMAdaptative(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"order": ("INT", {"default": 3, "min": 2, "max": 3}), - "rtol": ("FLOAT", {"default": 0.05, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), - "atol": ("FLOAT", {"default": 0.0078, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), - "h_init": ("FLOAT", {"default": 0.05, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), - "pcoeff": ("FLOAT", {"default": 0.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), - "icoeff": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), - "dcoeff": ("FLOAT", {"default": 0.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), - "accept_safety": ("FLOAT", {"default": 0.81, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), - "eta": ("FLOAT", {"default": 0.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), - "s_noise": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step":0.01, "round": False}), - } - } - RETURN_TYPES = ("SAMPLER",) - CATEGORY = "sampling/custom_sampling/samplers" + def define_schema(cls): + return io.Schema( + node_id="SamplerDPMAdaptative", + category="sampling/custom_sampling/samplers", + inputs=[ + io.Int.Input("order", default=3, min=2, max=3), + io.Float.Input("rtol", default=0.05, min=0.0, max=100.0, step=0.01, round=False), + io.Float.Input("atol", default=0.0078, min=0.0, max=100.0, step=0.01, round=False), + io.Float.Input("h_init", default=0.05, min=0.0, max=100.0, step=0.01, round=False), + io.Float.Input("pcoeff", default=0.0, min=0.0, max=100.0, step=0.01, round=False), + io.Float.Input("icoeff", default=1.0, min=0.0, max=100.0, step=0.01, round=False), + io.Float.Input("dcoeff", default=0.0, min=0.0, max=100.0, step=0.01, round=False), + io.Float.Input("accept_safety", default=0.81, min=0.0, max=100.0, step=0.01, round=False), + io.Float.Input("eta", default=0.0, min=0.0, max=100.0, step=0.01, round=False), + io.Float.Input("s_noise", default=1.0, min=0.0, max=100.0, step=0.01, round=False), + ], + outputs=[io.Sampler.Output()] + ) - FUNCTION = "get_sampler" - - def get_sampler(self, order, rtol, atol, h_init, pcoeff, icoeff, dcoeff, accept_safety, eta, s_noise): + @classmethod + def execute(cls, order, rtol, atol, h_init, pcoeff, icoeff, dcoeff, accept_safety, eta, s_noise) -> io.NodeOutput: sampler = comfy.samplers.ksampler("dpm_adaptive", {"order": order, "rtol": rtol, "atol": atol, "h_init": h_init, "pcoeff": pcoeff, "icoeff": icoeff, "dcoeff": dcoeff, "accept_safety": accept_safety, "eta": eta, "s_noise":s_noise }) - return (sampler, ) + return io.NodeOutput(sampler) + + get_sampler = execute -class SamplerER_SDE(ComfyNodeABC): +class SamplerER_SDE(io.ComfyNode): @classmethod - def INPUT_TYPES(cls) -> InputTypeDict: - return { - "required": { - "solver_type": (IO.COMBO, {"options": ["ER-SDE", "Reverse-time SDE", "ODE"]}), - "max_stage": (IO.INT, {"default": 3, "min": 1, "max": 3}), - "eta": ( - IO.FLOAT, - {"default": 1.0, "min": 0.0, "max": 100.0, "step": 0.01, "round": False, "tooltip": "Stochastic strength of reverse-time SDE.\nWhen eta=0, it reduces to deterministic ODE. This setting doesn't apply to ER-SDE solver type."}, - ), - "s_noise": (IO.FLOAT, {"default": 1.0, "min": 0.0, "max": 100.0, "step": 0.01, "round": False}), - } - } + def define_schema(cls): + return io.Schema( + node_id="SamplerER_SDE", + category="sampling/custom_sampling/samplers", + inputs=[ + io.Combo.Input("solver_type", options=["ER-SDE", "Reverse-time SDE", "ODE"]), + io.Int.Input("max_stage", default=3, min=1, max=3), + io.Float.Input("eta", default=1.0, min=0.0, max=100.0, step=0.01, round=False, tooltip="Stochastic strength of reverse-time SDE.\nWhen eta=0, it reduces to deterministic ODE. This setting doesn't apply to ER-SDE solver type."), + io.Float.Input("s_noise", default=1.0, min=0.0, max=100.0, step=0.01, round=False), + ], + outputs=[io.Sampler.Output()] + ) - RETURN_TYPES = (IO.SAMPLER,) - CATEGORY = "sampling/custom_sampling/samplers" - - FUNCTION = "get_sampler" - - def get_sampler(self, solver_type, max_stage, eta, s_noise): + @classmethod + def execute(cls, solver_type, max_stage, eta, s_noise) -> io.NodeOutput: if solver_type == "ODE" or (solver_type == "Reverse-time SDE" and eta == 0): eta = 0 s_noise = 0 @@ -548,32 +610,33 @@ class SamplerER_SDE(ComfyNodeABC): sampler_name = "er_sde" sampler = comfy.samplers.ksampler(sampler_name, {"s_noise": s_noise, "noise_scaler": noise_scaler, "max_stage": max_stage}) - return (sampler,) + return io.NodeOutput(sampler) + + get_sampler = execute -class SamplerSASolver(ComfyNodeABC): +class SamplerSASolver(io.ComfyNode): @classmethod - def INPUT_TYPES(cls) -> InputTypeDict: - return { - "required": { - "model": (IO.MODEL, {}), - "eta": (IO.FLOAT, {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01, "round": False},), - "sde_start_percent": (IO.FLOAT, {"default": 0.2, "min": 0.0, "max": 1.0, "step": 0.001},), - "sde_end_percent": (IO.FLOAT, {"default": 0.8, "min": 0.0, "max": 1.0, "step": 0.001},), - "s_noise": (IO.FLOAT, {"default": 1.0, "min": 0.0, "max": 100.0, "step": 0.01, "round": False},), - "predictor_order": (IO.INT, {"default": 3, "min": 1, "max": 6}), - "corrector_order": (IO.INT, {"default": 4, "min": 0, "max": 6}), - "use_pece": (IO.BOOLEAN, {}), - "simple_order_2": (IO.BOOLEAN, {}), - } - } + def define_schema(cls): + return io.Schema( + node_id="SamplerSASolver", + category="sampling/custom_sampling/samplers", + inputs=[ + io.Model.Input("model"), + io.Float.Input("eta", default=1.0, min=0.0, max=10.0, step=0.01, round=False), + io.Float.Input("sde_start_percent", default=0.2, min=0.0, max=1.0, step=0.001), + io.Float.Input("sde_end_percent", default=0.8, min=0.0, max=1.0, step=0.001), + io.Float.Input("s_noise", default=1.0, min=0.0, max=100.0, step=0.01, round=False), + io.Int.Input("predictor_order", default=3, min=1, max=6), + io.Int.Input("corrector_order", default=4, min=0, max=6), + io.Boolean.Input("use_pece"), + io.Boolean.Input("simple_order_2"), + ], + outputs=[io.Sampler.Output()] + ) - RETURN_TYPES = (IO.SAMPLER,) - CATEGORY = "sampling/custom_sampling/samplers" - - FUNCTION = "get_sampler" - - def get_sampler(self, model, eta, sde_start_percent, sde_end_percent, s_noise, predictor_order, corrector_order, use_pece, simple_order_2): + @classmethod + def execute(cls, model, eta, sde_start_percent, sde_end_percent, s_noise, predictor_order, corrector_order, use_pece, simple_order_2) -> io.NodeOutput: model_sampling = model.get_model_object("model_sampling") start_sigma = model_sampling.percent_to_sigma(sde_start_percent) end_sigma = model_sampling.percent_to_sigma(sde_end_percent) @@ -591,7 +654,9 @@ class SamplerSASolver(ComfyNodeABC): "simple_order_2": simple_order_2, }, ) - return (sampler,) + return io.NodeOutput(sampler) + + get_sampler = execute class Noise_EmptyNoise: @@ -612,30 +677,31 @@ class Noise_RandomNoise: batch_inds = input_latent["batch_index"] if "batch_index" in input_latent else None return comfy.sample.prepare_noise(latent_image, self.seed, batch_inds) -class SamplerCustom: +class SamplerCustom(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"model": ("MODEL",), - "add_noise": ("BOOLEAN", {"default": True}), - "noise_seed": ("INT", {"default": 0, "min": 0, "max": 0xffffffffffffffff, "control_after_generate": True}), - "cfg": ("FLOAT", {"default": 8.0, "min": 0.0, "max": 100.0, "step":0.1, "round": 0.01}), - "positive": ("CONDITIONING", ), - "negative": ("CONDITIONING", ), - "sampler": ("SAMPLER", ), - "sigmas": ("SIGMAS", ), - "latent_image": ("LATENT", ), - } - } + def define_schema(cls): + return io.Schema( + node_id="SamplerCustom", + category="sampling/custom_sampling", + inputs=[ + io.Model.Input("model"), + io.Boolean.Input("add_noise", default=True), + io.Int.Input("noise_seed", default=0, min=0, max=0xffffffffffffffff, control_after_generate=True), + io.Float.Input("cfg", default=8.0, min=0.0, max=100.0, step=0.1, round=0.01), + io.Conditioning.Input("positive"), + io.Conditioning.Input("negative"), + io.Sampler.Input("sampler"), + io.Sigmas.Input("sigmas"), + io.Latent.Input("latent_image"), + ], + outputs=[ + io.Latent.Output(display_name="output"), + io.Latent.Output(display_name="denoised_output"), + ] + ) - RETURN_TYPES = ("LATENT","LATENT") - RETURN_NAMES = ("output", "denoised_output") - - FUNCTION = "sample" - - CATEGORY = "sampling/custom_sampling" - - def sample(self, model, add_noise, noise_seed, cfg, positive, negative, sampler, sigmas, latent_image): + @classmethod + def execute(cls, model, add_noise, noise_seed, cfg, positive, negative, sampler, sigmas, latent_image) -> io.NodeOutput: latent = latent_image latent_image = latent["samples"] latent = latent.copy() @@ -664,52 +730,58 @@ class SamplerCustom: out_denoised["samples"] = model.model.process_latent_out(x0_output["x0"].cpu()) else: out_denoised = out - return (out, out_denoised) + return io.NodeOutput(out, out_denoised) + + sample = execute class Guider_Basic(comfy.samplers.CFGGuider): def set_conds(self, positive): self.inner_set_conds({"positive": positive}) -class BasicGuider: +class BasicGuider(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"model": ("MODEL",), - "conditioning": ("CONDITIONING", ), - } - } + def define_schema(cls): + return io.Schema( + node_id="BasicGuider", + category="sampling/custom_sampling/guiders", + inputs=[ + io.Model.Input("model"), + io.Conditioning.Input("conditioning"), + ], + outputs=[io.Guider.Output()] + ) - RETURN_TYPES = ("GUIDER",) - - FUNCTION = "get_guider" - CATEGORY = "sampling/custom_sampling/guiders" - - def get_guider(self, model, conditioning): + @classmethod + def execute(cls, model, conditioning) -> io.NodeOutput: guider = Guider_Basic(model) guider.set_conds(conditioning) - return (guider,) + return io.NodeOutput(guider) -class CFGGuider: + get_guider = execute + +class CFGGuider(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"model": ("MODEL",), - "positive": ("CONDITIONING", ), - "negative": ("CONDITIONING", ), - "cfg": ("FLOAT", {"default": 8.0, "min": 0.0, "max": 100.0, "step":0.1, "round": 0.01}), - } - } + def define_schema(cls): + return io.Schema( + node_id="CFGGuider", + category="sampling/custom_sampling/guiders", + inputs=[ + io.Model.Input("model"), + io.Conditioning.Input("positive"), + io.Conditioning.Input("negative"), + io.Float.Input("cfg", default=8.0, min=0.0, max=100.0, step=0.1, round=0.01), + ], + outputs=[io.Guider.Output()] + ) - RETURN_TYPES = ("GUIDER",) - - FUNCTION = "get_guider" - CATEGORY = "sampling/custom_sampling/guiders" - - def get_guider(self, model, positive, negative, cfg): + @classmethod + def execute(cls, model, positive, negative, cfg) -> io.NodeOutput: guider = comfy.samplers.CFGGuider(model) guider.set_conds(positive, negative) guider.set_cfg(cfg) - return (guider,) + return io.NodeOutput(guider) + + get_guider = execute class Guider_DualCFG(comfy.samplers.CFGGuider): def set_cfg(self, cfg1, cfg2, nested=False): @@ -740,84 +812,88 @@ class Guider_DualCFG(comfy.samplers.CFGGuider): out = comfy.samplers.calc_cond_batch(self.inner_model, [negative_cond, middle_cond, positive_cond], x, timestep, model_options) return comfy.samplers.cfg_function(self.inner_model, out[1], out[0], self.cfg2, x, timestep, model_options=model_options, cond=middle_cond, uncond=negative_cond) + (out[2] - out[1]) * self.cfg1 -class DualCFGGuider: +class DualCFGGuider(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"model": ("MODEL",), - "cond1": ("CONDITIONING", ), - "cond2": ("CONDITIONING", ), - "negative": ("CONDITIONING", ), - "cfg_conds": ("FLOAT", {"default": 8.0, "min": 0.0, "max": 100.0, "step":0.1, "round": 0.01}), - "cfg_cond2_negative": ("FLOAT", {"default": 8.0, "min": 0.0, "max": 100.0, "step":0.1, "round": 0.01}), - "style": (["regular", "nested"],), - } - } + def define_schema(cls): + return io.Schema( + node_id="DualCFGGuider", + category="sampling/custom_sampling/guiders", + inputs=[ + io.Model.Input("model"), + io.Conditioning.Input("cond1"), + io.Conditioning.Input("cond2"), + io.Conditioning.Input("negative"), + io.Float.Input("cfg_conds", default=8.0, min=0.0, max=100.0, step=0.1, round=0.01), + io.Float.Input("cfg_cond2_negative", default=8.0, min=0.0, max=100.0, step=0.1, round=0.01), + io.Combo.Input("style", options=["regular", "nested"]), + ], + outputs=[io.Guider.Output()] + ) - RETURN_TYPES = ("GUIDER",) - - FUNCTION = "get_guider" - CATEGORY = "sampling/custom_sampling/guiders" - - def get_guider(self, model, cond1, cond2, negative, cfg_conds, cfg_cond2_negative, style): + @classmethod + def execute(cls, model, cond1, cond2, negative, cfg_conds, cfg_cond2_negative, style) -> io.NodeOutput: guider = Guider_DualCFG(model) guider.set_conds(cond1, cond2, negative) guider.set_cfg(cfg_conds, cfg_cond2_negative, nested=(style == "nested")) - return (guider,) + return io.NodeOutput(guider) -class DisableNoise: + get_guider = execute + +class DisableNoise(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required":{ - } - } + def define_schema(cls): + return io.Schema( + node_id="DisableNoise", + category="sampling/custom_sampling/noise", + inputs=[], + outputs=[io.Noise.Output()] + ) - RETURN_TYPES = ("NOISE",) - FUNCTION = "get_noise" - CATEGORY = "sampling/custom_sampling/noise" - - def get_noise(self): - return (Noise_EmptyNoise(),) - - -class RandomNoise(DisableNoise): @classmethod - def INPUT_TYPES(s): - return { - "required": { - "noise_seed": ("INT", { - "default": 0, - "min": 0, - "max": 0xffffffffffffffff, - "control_after_generate": True, - }), - } - } + def execute(cls) -> io.NodeOutput: + return io.NodeOutput(Noise_EmptyNoise()) - def get_noise(self, noise_seed): - return (Noise_RandomNoise(noise_seed),) + get_noise = execute -class SamplerCustomAdvanced: +class RandomNoise(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"noise": ("NOISE", ), - "guider": ("GUIDER", ), - "sampler": ("SAMPLER", ), - "sigmas": ("SIGMAS", ), - "latent_image": ("LATENT", ), - } - } + def define_schema(cls): + return io.Schema( + node_id="RandomNoise", + category="sampling/custom_sampling/noise", + inputs=[io.Int.Input("noise_seed", default=0, min=0, max=0xffffffffffffffff, control_after_generate=True)], + outputs=[io.Noise.Output()] + ) - RETURN_TYPES = ("LATENT","LATENT") - RETURN_NAMES = ("output", "denoised_output") + @classmethod + def execute(cls, noise_seed) -> io.NodeOutput: + return io.NodeOutput(Noise_RandomNoise(noise_seed)) - FUNCTION = "sample" + get_noise = execute - CATEGORY = "sampling/custom_sampling" - def sample(self, noise, guider, sampler, sigmas, latent_image): +class SamplerCustomAdvanced(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="SamplerCustomAdvanced", + category="sampling/custom_sampling", + inputs=[ + io.Noise.Input("noise"), + io.Guider.Input("guider"), + io.Sampler.Input("sampler"), + io.Sigmas.Input("sigmas"), + io.Latent.Input("latent_image"), + ], + outputs=[ + io.Latent.Output(display_name="output"), + io.Latent.Output(display_name="denoised_output"), + ] + ) + + @classmethod + def execute(cls, noise, guider, sampler, sigmas, latent_image) -> io.NodeOutput: latent = latent_image latent_image = latent["samples"] latent = latent.copy() @@ -842,28 +918,32 @@ class SamplerCustomAdvanced: out_denoised["samples"] = guider.model_patcher.model.process_latent_out(x0_output["x0"].cpu()) else: out_denoised = out - return (out, out_denoised) + return io.NodeOutput(out, out_denoised) -class AddNoise: + sample = execute + +class AddNoise(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": - {"model": ("MODEL",), - "noise": ("NOISE", ), - "sigmas": ("SIGMAS", ), - "latent_image": ("LATENT", ), - } - } + def define_schema(cls): + return io.Schema( + node_id="AddNoise", + category="_for_testing/custom_sampling/noise", + is_experimental=True, + inputs=[ + io.Model.Input("model"), + io.Noise.Input("noise"), + io.Sigmas.Input("sigmas"), + io.Latent.Input("latent_image"), + ], + outputs=[ + io.Latent.Output(), + ] + ) - RETURN_TYPES = ("LATENT",) - - FUNCTION = "add_noise" - - CATEGORY = "_for_testing/custom_sampling/noise" - - def add_noise(self, model, noise, sigmas, latent_image): + @classmethod + def execute(cls, model, noise, sigmas, latent_image) -> io.NodeOutput: if len(sigmas) == 0: - return latent_image + return io.NodeOutput(latent_image) latent = latent_image latent_image = latent["samples"] @@ -887,46 +967,50 @@ class AddNoise: out = latent.copy() out["samples"] = noisy - return (out,) + return io.NodeOutput(out) + + add_noise = execute -NODE_CLASS_MAPPINGS = { - "SamplerCustom": SamplerCustom, - "BasicScheduler": BasicScheduler, - "KarrasScheduler": KarrasScheduler, - "ExponentialScheduler": ExponentialScheduler, - "PolyexponentialScheduler": PolyexponentialScheduler, - "LaplaceScheduler": LaplaceScheduler, - "VPScheduler": VPScheduler, - "BetaSamplingScheduler": BetaSamplingScheduler, - "SDTurboScheduler": SDTurboScheduler, - "KSamplerSelect": KSamplerSelect, - "SamplerEulerAncestral": SamplerEulerAncestral, - "SamplerEulerAncestralCFGPP": SamplerEulerAncestralCFGPP, - "SamplerLMS": SamplerLMS, - "SamplerDPMPP_3M_SDE": SamplerDPMPP_3M_SDE, - "SamplerDPMPP_2M_SDE": SamplerDPMPP_2M_SDE, - "SamplerDPMPP_SDE": SamplerDPMPP_SDE, - "SamplerDPMPP_2S_Ancestral": SamplerDPMPP_2S_Ancestral, - "SamplerDPMAdaptative": SamplerDPMAdaptative, - "SamplerER_SDE": SamplerER_SDE, - "SamplerSASolver": SamplerSASolver, - "SplitSigmas": SplitSigmas, - "SplitSigmasDenoise": SplitSigmasDenoise, - "FlipSigmas": FlipSigmas, - "SetFirstSigma": SetFirstSigma, - "ExtendIntermediateSigmas": ExtendIntermediateSigmas, - "SamplingPercentToSigma": SamplingPercentToSigma, +class CustomSamplersExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[io.ComfyNode]]: + return [ + SamplerCustom, + BasicScheduler, + KarrasScheduler, + ExponentialScheduler, + PolyexponentialScheduler, + LaplaceScheduler, + VPScheduler, + BetaSamplingScheduler, + SDTurboScheduler, + KSamplerSelect, + SamplerEulerAncestral, + SamplerEulerAncestralCFGPP, + SamplerLMS, + SamplerDPMPP_3M_SDE, + SamplerDPMPP_2M_SDE, + SamplerDPMPP_SDE, + SamplerDPMPP_2S_Ancestral, + SamplerDPMAdaptative, + SamplerER_SDE, + SamplerSASolver, + SplitSigmas, + SplitSigmasDenoise, + FlipSigmas, + SetFirstSigma, + ExtendIntermediateSigmas, + SamplingPercentToSigma, + CFGGuider, + DualCFGGuider, + BasicGuider, + RandomNoise, + DisableNoise, + AddNoise, + SamplerCustomAdvanced, + ] - "CFGGuider": CFGGuider, - "DualCFGGuider": DualCFGGuider, - "BasicGuider": BasicGuider, - "RandomNoise": RandomNoise, - "DisableNoise": DisableNoise, - "AddNoise": AddNoise, - "SamplerCustomAdvanced": SamplerCustomAdvanced, -} -NODE_DISPLAY_NAME_MAPPINGS = { - "SamplerEulerAncestralCFGPP": "SamplerEulerAncestralCFG++", -} +async def comfy_entrypoint() -> CustomSamplersExtension: + return CustomSamplersExtension() From cc6a8dcd1ad9cc9ef7602ee141174a0cea0ed4ce Mon Sep 17 00:00:00 2001 From: Kohaku-Blueleaf <59680068+KohakuBlueleaf@users.noreply.github.com> Date: Thu, 27 Nov 2025 08:18:08 +0800 Subject: [PATCH 180/647] Dataset Processing Nodes and Improved LoRA Trainer Nodes with multi resolution supports. (#10708) * Create nodes_dataset.py * Add encoded dataset caching mechanism * make training node to work with our dataset system * allow trainer node to get different resolution dataset * move all dataset related implementation to nodes_dataset * Rewrite dataset system with new io schema * Rewrite training system with new io schema * add ui pbar * Add outputs' id/name * Fix bad id/naming * use single process instead of input list when no need * fix wrong output_list flag * use torch.load/save and fix bad behaviors --- comfy_extras/nodes_dataset.py | 1532 +++++++++++++++++++++++++++++++++ comfy_extras/nodes_train.py | 967 ++++++++++----------- nodes.py | 1 + 3 files changed, 1980 insertions(+), 520 deletions(-) create mode 100644 comfy_extras/nodes_dataset.py diff --git a/comfy_extras/nodes_dataset.py b/comfy_extras/nodes_dataset.py new file mode 100644 index 000000000..b23867505 --- /dev/null +++ b/comfy_extras/nodes_dataset.py @@ -0,0 +1,1532 @@ +import logging +import os +import math +import json + +import numpy as np +import torch +from PIL import Image +from typing_extensions import override + +import folder_paths +import node_helpers +from comfy_api.latest import ComfyExtension, io + + +def load_and_process_images(image_files, input_dir): + """Utility function to load and process a list of images. + + Args: + image_files: List of image filenames + input_dir: Base directory containing the images + resize_method: How to handle images of different sizes ("None", "Stretch", "Crop", "Pad") + + Returns: + torch.Tensor: Batch of processed images + """ + if not image_files: + raise ValueError("No valid images found in input") + + output_images = [] + + for file in image_files: + image_path = os.path.join(input_dir, file) + img = node_helpers.pillow(Image.open, image_path) + + if img.mode == "I": + img = img.point(lambda i: i * (1 / 255)) + img = img.convert("RGB") + img_array = np.array(img).astype(np.float32) / 255.0 + img_tensor = torch.from_numpy(img_array)[None,] + output_images.append(img_tensor) + + return output_images + + +class LoadImageDataSetFromFolderNode(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="LoadImageDataSetFromFolder", + display_name="Load Image Dataset from Folder", + category="dataset", + is_experimental=True, + inputs=[ + io.Combo.Input( + "folder", + options=folder_paths.get_input_subfolders(), + tooltip="The folder to load images from.", + ) + ], + outputs=[ + io.Image.Output( + display_name="images", + is_output_list=True, + tooltip="List of loaded images", + ) + ], + ) + + @classmethod + def execute(cls, folder): + sub_input_dir = os.path.join(folder_paths.get_input_directory(), folder) + valid_extensions = [".png", ".jpg", ".jpeg", ".webp"] + image_files = [ + f + for f in os.listdir(sub_input_dir) + if any(f.lower().endswith(ext) for ext in valid_extensions) + ] + output_tensor = load_and_process_images(image_files, sub_input_dir) + return io.NodeOutput(output_tensor) + + +class LoadImageTextDataSetFromFolderNode(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="LoadImageTextDataSetFromFolder", + display_name="Load Image and Text Dataset from Folder", + category="dataset", + is_experimental=True, + inputs=[ + io.Combo.Input( + "folder", + options=folder_paths.get_input_subfolders(), + tooltip="The folder to load images from.", + ) + ], + outputs=[ + io.Image.Output( + display_name="images", + is_output_list=True, + tooltip="List of loaded images", + ), + io.String.Output( + display_name="texts", + is_output_list=True, + tooltip="List of text captions", + ), + ], + ) + + @classmethod + def execute(cls, folder): + logging.info(f"Loading images from folder: {folder}") + + sub_input_dir = os.path.join(folder_paths.get_input_directory(), folder) + valid_extensions = [".png", ".jpg", ".jpeg", ".webp"] + + image_files = [] + for item in os.listdir(sub_input_dir): + path = os.path.join(sub_input_dir, item) + if any(item.lower().endswith(ext) for ext in valid_extensions): + image_files.append(path) + elif os.path.isdir(path): + # Support kohya-ss/sd-scripts folder structure + repeat = 1 + if item.split("_")[0].isdigit(): + repeat = int(item.split("_")[0]) + image_files.extend( + [ + os.path.join(path, f) + for f in os.listdir(path) + if any(f.lower().endswith(ext) for ext in valid_extensions) + ] + * repeat + ) + + caption_file_path = [ + f.replace(os.path.splitext(f)[1], ".txt") for f in image_files + ] + captions = [] + for caption_file in caption_file_path: + caption_path = os.path.join(sub_input_dir, caption_file) + if os.path.exists(caption_path): + with open(caption_path, "r", encoding="utf-8") as f: + caption = f.read().strip() + captions.append(caption) + else: + captions.append("") + + output_tensor = load_and_process_images(image_files, sub_input_dir) + + logging.info(f"Loaded {len(output_tensor)} images from {sub_input_dir}.") + return io.NodeOutput(output_tensor, captions) + + +def save_images_to_folder(image_list, output_dir, prefix="image"): + """Utility function to save a list of image tensors to disk. + + Args: + image_list: List of image tensors (each [1, H, W, C] or [H, W, C] or [C, H, W]) + output_dir: Directory to save images to + prefix: Filename prefix + + Returns: + List of saved filenames + """ + os.makedirs(output_dir, exist_ok=True) + saved_files = [] + + for idx, img_tensor in enumerate(image_list): + # Handle different tensor shapes + if isinstance(img_tensor, torch.Tensor): + # Remove batch dimension if present [1, H, W, C] -> [H, W, C] + if img_tensor.dim() == 4 and img_tensor.shape[0] == 1: + img_tensor = img_tensor.squeeze(0) + + # If tensor is [C, H, W], permute to [H, W, C] + if img_tensor.dim() == 3 and img_tensor.shape[0] in [1, 3, 4]: + if ( + img_tensor.shape[0] <= 4 + and img_tensor.shape[1] > 4 + and img_tensor.shape[2] > 4 + ): + img_tensor = img_tensor.permute(1, 2, 0) + + # Convert to numpy and scale to 0-255 + img_array = img_tensor.cpu().numpy() + img_array = np.clip(img_array * 255.0, 0, 255).astype(np.uint8) + + # Convert to PIL Image + img = Image.fromarray(img_array) + else: + raise ValueError(f"Expected torch.Tensor, got {type(img_tensor)}") + + # Save image + filename = f"{prefix}_{idx:05d}.png" + filepath = os.path.join(output_dir, filename) + img.save(filepath) + saved_files.append(filename) + + return saved_files + + +class SaveImageDataSetToFolderNode(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="SaveImageDataSetToFolder", + display_name="Save Image Dataset to Folder", + category="dataset", + is_experimental=True, + is_output_node=True, + is_input_list=True, # Receive images as list + inputs=[ + io.Image.Input("images", tooltip="List of images to save."), + io.String.Input( + "folder_name", + default="dataset", + tooltip="Name of the folder to save images to (inside output directory).", + ), + io.String.Input( + "filename_prefix", + default="image", + tooltip="Prefix for saved image filenames.", + ), + ], + outputs=[], + ) + + @classmethod + def execute(cls, images, folder_name, filename_prefix): + # Extract scalar values + folder_name = folder_name[0] + filename_prefix = filename_prefix[0] + + output_dir = os.path.join(folder_paths.get_output_directory(), folder_name) + saved_files = save_images_to_folder(images, output_dir, filename_prefix) + + logging.info(f"Saved {len(saved_files)} images to {output_dir}.") + return io.NodeOutput() + + +class SaveImageTextDataSetToFolderNode(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="SaveImageTextDataSetToFolder", + display_name="Save Image and Text Dataset to Folder", + category="dataset", + is_experimental=True, + is_output_node=True, + is_input_list=True, # Receive both images and texts as lists + inputs=[ + io.Image.Input("images", tooltip="List of images to save."), + io.String.Input("texts", tooltip="List of text captions to save."), + io.String.Input( + "folder_name", + default="dataset", + tooltip="Name of the folder to save images to (inside output directory).", + ), + io.String.Input( + "filename_prefix", + default="image", + tooltip="Prefix for saved image filenames.", + ), + ], + outputs=[], + ) + + @classmethod + def execute(cls, images, texts, folder_name, filename_prefix): + # Extract scalar values + folder_name = folder_name[0] + filename_prefix = filename_prefix[0] + + output_dir = os.path.join(folder_paths.get_output_directory(), folder_name) + saved_files = save_images_to_folder(images, output_dir, filename_prefix) + + # Save captions + for idx, (filename, caption) in enumerate(zip(saved_files, texts)): + caption_filename = filename.replace(".png", ".txt") + caption_path = os.path.join(output_dir, caption_filename) + with open(caption_path, "w", encoding="utf-8") as f: + f.write(caption) + + logging.info(f"Saved {len(saved_files)} images and captions to {output_dir}.") + return io.NodeOutput() + + +# ========== Helper Functions for Transform Nodes ========== + + +def tensor_to_pil(img_tensor): + """Convert tensor to PIL Image.""" + if img_tensor.dim() == 4 and img_tensor.shape[0] == 1: + img_tensor = img_tensor.squeeze(0) + img_array = (img_tensor.cpu().numpy() * 255).clip(0, 255).astype(np.uint8) + return Image.fromarray(img_array) + + +def pil_to_tensor(img): + """Convert PIL Image to tensor.""" + img_array = np.array(img).astype(np.float32) / 255.0 + return torch.from_numpy(img_array)[None,] + + +# ========== Base Classes for Transform Nodes ========== + + +class ImageProcessingNode(io.ComfyNode): + """Base class for image processing nodes that operate on images. + + Child classes should set: + node_id: Unique node identifier (required) + display_name: Display name (optional, defaults to node_id) + description: Node description (optional) + extra_inputs: List of additional io.Input objects beyond "images" (optional) + is_group_process: None (auto-detect), True (group), or False (individual) (optional) + is_output_list: True (list output) or False (single output) (optional, default True) + + Child classes must implement ONE of: + _process(cls, image, **kwargs) -> tensor (for single-item processing) + _group_process(cls, images, **kwargs) -> list[tensor] (for group processing) + """ + + node_id = None + display_name = None + description = None + extra_inputs = [] + is_group_process = None # None = auto-detect, True/False = explicit + is_output_list = None # None = auto-detect based on processing mode + + @classmethod + def _detect_processing_mode(cls): + """Detect whether this node uses group or individual processing. + + Returns: + bool: True if group processing, False if individual processing + """ + # Explicit setting takes precedence + if cls.is_group_process is not None: + return cls.is_group_process + + # Check which method is overridden by looking at the defining class in MRO + base_class = ImageProcessingNode + + # Find which class in MRO defines _process + process_definer = None + for klass in cls.__mro__: + if "_process" in klass.__dict__: + process_definer = klass + break + + # Find which class in MRO defines _group_process + group_definer = None + for klass in cls.__mro__: + if "_group_process" in klass.__dict__: + group_definer = klass + break + + # Check what was overridden (not defined in base class) + has_process = process_definer is not None and process_definer is not base_class + has_group = group_definer is not None and group_definer is not base_class + + if has_process and has_group: + raise ValueError( + f"{cls.__name__}: Cannot override both _process and _group_process. " + "Override only one, or set is_group_process explicitly." + ) + if not has_process and not has_group: + raise ValueError( + f"{cls.__name__}: Must override either _process or _group_process" + ) + + return has_group + + @classmethod + def define_schema(cls): + if cls.node_id is None: + raise NotImplementedError(f"{cls.__name__} must set node_id class variable") + + is_group = cls._detect_processing_mode() + + # Auto-detect is_output_list if not explicitly set + # Single processing: False (backend collects results into list) + # Group processing: True by default (can be False for single-output nodes) + output_is_list = ( + cls.is_output_list if cls.is_output_list is not None else is_group + ) + + inputs = [ + io.Image.Input( + "images", + tooltip=( + "List of images to process." if is_group else "Image to process." + ), + ) + ] + inputs.extend(cls.extra_inputs) + + return io.Schema( + node_id=cls.node_id, + display_name=cls.display_name or cls.node_id, + category="dataset/image", + is_experimental=True, + is_input_list=is_group, # True for group, False for individual + inputs=inputs, + outputs=[ + io.Image.Output( + display_name="images", + is_output_list=output_is_list, + tooltip="Processed images", + ) + ], + ) + + @classmethod + def execute(cls, images, **kwargs): + """Execute the node. Routes to _process or _group_process based on mode.""" + is_group = cls._detect_processing_mode() + + # Extract scalar values from lists for parameters + params = {} + for k, v in kwargs.items(): + if isinstance(v, list) and len(v) == 1: + params[k] = v[0] + else: + params[k] = v + + if is_group: + # Group processing: images is list, call _group_process + result = cls._group_process(images, **params) + else: + # Individual processing: images is single item, call _process + result = cls._process(images, **params) + + return io.NodeOutput(result) + + @classmethod + def _process(cls, image, **kwargs): + """Override this method for single-item processing. + + Args: + image: tensor - Single image tensor + **kwargs: Additional parameters (already extracted from lists) + + Returns: + tensor - Processed image + """ + raise NotImplementedError(f"{cls.__name__} must implement _process method") + + @classmethod + def _group_process(cls, images, **kwargs): + """Override this method for group processing. + + Args: + images: list[tensor] - List of image tensors + **kwargs: Additional parameters (already extracted from lists) + + Returns: + list[tensor] - Processed images + """ + raise NotImplementedError( + f"{cls.__name__} must implement _group_process method" + ) + + +class TextProcessingNode(io.ComfyNode): + """Base class for text processing nodes that operate on texts. + + Child classes should set: + node_id: Unique node identifier (required) + display_name: Display name (optional, defaults to node_id) + description: Node description (optional) + extra_inputs: List of additional io.Input objects beyond "texts" (optional) + is_group_process: None (auto-detect), True (group), or False (individual) (optional) + is_output_list: True (list output) or False (single output) (optional, default True) + + Child classes must implement ONE of: + _process(cls, text, **kwargs) -> str (for single-item processing) + _group_process(cls, texts, **kwargs) -> list[str] (for group processing) + """ + + node_id = None + display_name = None + description = None + extra_inputs = [] + is_group_process = None # None = auto-detect, True/False = explicit + is_output_list = None # None = auto-detect based on processing mode + + @classmethod + def _detect_processing_mode(cls): + """Detect whether this node uses group or individual processing. + + Returns: + bool: True if group processing, False if individual processing + """ + # Explicit setting takes precedence + if cls.is_group_process is not None: + return cls.is_group_process + + # Check which method is overridden by looking at the defining class in MRO + base_class = TextProcessingNode + + # Find which class in MRO defines _process + process_definer = None + for klass in cls.__mro__: + if "_process" in klass.__dict__: + process_definer = klass + break + + # Find which class in MRO defines _group_process + group_definer = None + for klass in cls.__mro__: + if "_group_process" in klass.__dict__: + group_definer = klass + break + + # Check what was overridden (not defined in base class) + has_process = process_definer is not None and process_definer is not base_class + has_group = group_definer is not None and group_definer is not base_class + + if has_process and has_group: + raise ValueError( + f"{cls.__name__}: Cannot override both _process and _group_process. " + "Override only one, or set is_group_process explicitly." + ) + if not has_process and not has_group: + raise ValueError( + f"{cls.__name__}: Must override either _process or _group_process" + ) + + return has_group + + @classmethod + def define_schema(cls): + if cls.node_id is None: + raise NotImplementedError(f"{cls.__name__} must set node_id class variable") + + is_group = cls._detect_processing_mode() + + inputs = [ + io.String.Input( + "texts", + tooltip="List of texts to process." if is_group else "Text to process.", + ) + ] + inputs.extend(cls.extra_inputs) + + return io.Schema( + node_id=cls.node_id, + display_name=cls.display_name or cls.node_id, + category="dataset/text", + is_experimental=True, + is_input_list=is_group, # True for group, False for individual + inputs=inputs, + outputs=[ + io.String.Output( + display_name="texts", + is_output_list=cls.is_output_list, + tooltip="Processed texts", + ) + ], + ) + + @classmethod + def execute(cls, texts, **kwargs): + """Execute the node. Routes to _process or _group_process based on mode.""" + is_group = cls._detect_processing_mode() + + # Extract scalar values from lists for parameters + params = {} + for k, v in kwargs.items(): + if isinstance(v, list) and len(v) == 1: + params[k] = v[0] + else: + params[k] = v + + if is_group: + # Group processing: texts is list, call _group_process + result = cls._group_process(texts, **params) + else: + # Individual processing: texts is single item, call _process + result = cls._process(texts, **params) + + # Wrap result based on is_output_list + if cls.is_output_list: + # Result should already be a list (or will be for individual) + return io.NodeOutput(result if is_group else [result]) + else: + # Single output - wrap in list for NodeOutput + return io.NodeOutput([result]) + + @classmethod + def _process(cls, text, **kwargs): + """Override this method for single-item processing. + + Args: + text: str - Single text string + **kwargs: Additional parameters (already extracted from lists) + + Returns: + str - Processed text + """ + raise NotImplementedError(f"{cls.__name__} must implement _process method") + + @classmethod + def _group_process(cls, texts, **kwargs): + """Override this method for group processing. + + Args: + texts: list[str] - List of text strings + **kwargs: Additional parameters (already extracted from lists) + + Returns: + list[str] - Processed texts + """ + raise NotImplementedError( + f"{cls.__name__} must implement _group_process method" + ) + + +# ========== Image Transform Nodes ========== + + +class ResizeImagesToSameSizeNode(ImageProcessingNode): + node_id = "ResizeImagesToSameSize" + display_name = "Resize Images to Same Size" + description = "Resize all images to the same width and height." + extra_inputs = [ + io.Int.Input("width", default=512, min=1, max=8192, tooltip="Target width."), + io.Int.Input("height", default=512, min=1, max=8192, tooltip="Target height."), + io.Combo.Input( + "mode", + options=["stretch", "crop_center", "pad"], + default="stretch", + tooltip="Resize mode.", + ), + ] + + @classmethod + def _process(cls, image, width, height, mode): + img = tensor_to_pil(image) + + if mode == "stretch": + img = img.resize((width, height), Image.Resampling.LANCZOS) + elif mode == "crop_center": + left = max(0, (img.width - width) // 2) + top = max(0, (img.height - height) // 2) + right = min(img.width, left + width) + bottom = min(img.height, top + height) + img = img.crop((left, top, right, bottom)) + if img.width != width or img.height != height: + img = img.resize((width, height), Image.Resampling.LANCZOS) + elif mode == "pad": + img.thumbnail((width, height), Image.Resampling.LANCZOS) + new_img = Image.new("RGB", (width, height), (0, 0, 0)) + paste_x = (width - img.width) // 2 + paste_y = (height - img.height) // 2 + new_img.paste(img, (paste_x, paste_y)) + img = new_img + + return pil_to_tensor(img) + + +class ResizeImagesToPixelCountNode(ImageProcessingNode): + node_id = "ResizeImagesToPixelCount" + display_name = "Resize Images to Pixel Count" + description = "Resize images so that the total pixel count matches the specified number while preserving aspect ratio." + extra_inputs = [ + io.Int.Input( + "pixel_count", + default=512 * 512, + min=1, + max=8192 * 8192, + tooltip="Target pixel count.", + ), + io.Int.Input( + "steps", + default=64, + min=1, + max=128, + tooltip="The stepping for resize width/height.", + ), + ] + + @classmethod + def _process(cls, image, pixel_count, steps): + img = tensor_to_pil(image) + w, h = img.size + pixel_count_ratio = math.sqrt(pixel_count / (w * h)) + new_w = int(w * pixel_count_ratio / steps) * steps + new_h = int(h * pixel_count_ratio / steps) * steps + logging.info(f"Resizing from {w}x{h} to {new_w}x{new_h}") + img = img.resize((new_w, new_h), Image.Resampling.LANCZOS) + return pil_to_tensor(img) + + +class ResizeImagesByShorterEdgeNode(ImageProcessingNode): + node_id = "ResizeImagesByShorterEdge" + display_name = "Resize Images by Shorter Edge" + description = "Resize images so that the shorter edge matches the specified length while preserving aspect ratio." + extra_inputs = [ + io.Int.Input( + "shorter_edge", + default=512, + min=1, + max=8192, + tooltip="Target length for the shorter edge.", + ), + ] + + @classmethod + def _process(cls, image, shorter_edge): + img = tensor_to_pil(image) + w, h = img.size + if w < h: + new_w = shorter_edge + new_h = int(h * (shorter_edge / w)) + else: + new_h = shorter_edge + new_w = int(w * (shorter_edge / h)) + img = img.resize((new_w, new_h), Image.Resampling.LANCZOS) + return pil_to_tensor(img) + + +class ResizeImagesByLongerEdgeNode(ImageProcessingNode): + node_id = "ResizeImagesByLongerEdge" + display_name = "Resize Images by Longer Edge" + description = "Resize images so that the longer edge matches the specified length while preserving aspect ratio." + extra_inputs = [ + io.Int.Input( + "longer_edge", + default=1024, + min=1, + max=8192, + tooltip="Target length for the longer edge.", + ), + ] + + @classmethod + def _process(cls, image, longer_edge): + img = tensor_to_pil(image) + w, h = img.size + if w > h: + new_w = longer_edge + new_h = int(h * (longer_edge / w)) + else: + new_h = longer_edge + new_w = int(w * (longer_edge / h)) + img = img.resize((new_w, new_h), Image.Resampling.LANCZOS) + return pil_to_tensor(img) + + +class CenterCropImagesNode(ImageProcessingNode): + node_id = "CenterCropImages" + display_name = "Center Crop Images" + description = "Center crop all images to the specified dimensions." + extra_inputs = [ + io.Int.Input("width", default=512, min=1, max=8192, tooltip="Crop width."), + io.Int.Input("height", default=512, min=1, max=8192, tooltip="Crop height."), + ] + + @classmethod + def _process(cls, image, width, height): + img = tensor_to_pil(image) + left = max(0, (img.width - width) // 2) + top = max(0, (img.height - height) // 2) + right = min(img.width, left + width) + bottom = min(img.height, top + height) + img = img.crop((left, top, right, bottom)) + return pil_to_tensor(img) + + +class RandomCropImagesNode(ImageProcessingNode): + node_id = "RandomCropImages" + display_name = "Random Crop Images" + description = ( + "Randomly crop all images to the specified dimensions (for data augmentation)." + ) + extra_inputs = [ + io.Int.Input("width", default=512, min=1, max=8192, tooltip="Crop width."), + io.Int.Input("height", default=512, min=1, max=8192, tooltip="Crop height."), + io.Int.Input( + "seed", default=0, min=0, max=0xFFFFFFFFFFFFFFFF, tooltip="Random seed." + ), + ] + + @classmethod + def _process(cls, image, width, height, seed): + np.random.seed(seed % (2**32 - 1)) + img = tensor_to_pil(image) + max_left = max(0, img.width - width) + max_top = max(0, img.height - height) + left = np.random.randint(0, max_left + 1) if max_left > 0 else 0 + top = np.random.randint(0, max_top + 1) if max_top > 0 else 0 + right = min(img.width, left + width) + bottom = min(img.height, top + height) + img = img.crop((left, top, right, bottom)) + return pil_to_tensor(img) + + +class FlipImagesNode(ImageProcessingNode): + node_id = "FlipImages" + display_name = "Flip Images" + description = "Flip all images horizontally or vertically." + extra_inputs = [ + io.Combo.Input( + "direction", + options=["horizontal", "vertical"], + default="horizontal", + tooltip="Flip direction.", + ), + ] + + @classmethod + def _process(cls, image, direction): + img = tensor_to_pil(image) + if direction == "horizontal": + img = img.transpose(Image.FLIP_LEFT_RIGHT) + else: + img = img.transpose(Image.FLIP_TOP_BOTTOM) + return pil_to_tensor(img) + + +class NormalizeImagesNode(ImageProcessingNode): + node_id = "NormalizeImages" + display_name = "Normalize Images" + description = "Normalize images using mean and standard deviation." + extra_inputs = [ + io.Float.Input( + "mean", + default=0.5, + min=0.0, + max=1.0, + tooltip="Mean value for normalization.", + ), + io.Float.Input( + "std", + default=0.5, + min=0.001, + max=1.0, + tooltip="Standard deviation for normalization.", + ), + ] + + @classmethod + def _process(cls, image, mean, std): + return (image - mean) / std + + +class AdjustBrightnessNode(ImageProcessingNode): + node_id = "AdjustBrightness" + display_name = "Adjust Brightness" + description = "Adjust brightness of all images." + extra_inputs = [ + io.Float.Input( + "factor", + default=1.0, + min=0.0, + max=2.0, + tooltip="Brightness factor. 1.0 = no change, <1.0 = darker, >1.0 = brighter.", + ), + ] + + @classmethod + def _process(cls, image, factor): + return (image * factor).clamp(0.0, 1.0) + + +class AdjustContrastNode(ImageProcessingNode): + node_id = "AdjustContrast" + display_name = "Adjust Contrast" + description = "Adjust contrast of all images." + extra_inputs = [ + io.Float.Input( + "factor", + default=1.0, + min=0.0, + max=2.0, + tooltip="Contrast factor. 1.0 = no change, <1.0 = less contrast, >1.0 = more contrast.", + ), + ] + + @classmethod + def _process(cls, image, factor): + return ((image - 0.5) * factor + 0.5).clamp(0.0, 1.0) + + +class ShuffleDatasetNode(ImageProcessingNode): + node_id = "ShuffleDataset" + display_name = "Shuffle Image Dataset" + description = "Randomly shuffle the order of images in the dataset." + is_group_process = True # Requires full list to shuffle + extra_inputs = [ + io.Int.Input( + "seed", default=0, min=0, max=0xFFFFFFFFFFFFFFFF, tooltip="Random seed." + ), + ] + + @classmethod + def _group_process(cls, images, seed): + np.random.seed(seed % (2**32 - 1)) + indices = np.random.permutation(len(images)) + return [images[i] for i in indices] + + +class ShuffleImageTextDatasetNode(io.ComfyNode): + """Special node that shuffles both images and texts together.""" + + @classmethod + def define_schema(cls): + return io.Schema( + node_id="ShuffleImageTextDataset", + display_name="Shuffle Image-Text Dataset", + category="dataset/image", + is_experimental=True, + is_input_list=True, + inputs=[ + io.Image.Input("images", tooltip="List of images to shuffle."), + io.String.Input("texts", tooltip="List of texts to shuffle."), + io.Int.Input( + "seed", + default=0, + min=0, + max=0xFFFFFFFFFFFFFFFF, + tooltip="Random seed.", + ), + ], + outputs=[ + io.Image.Output( + display_name="images", + is_output_list=True, + tooltip="Shuffled images", + ), + io.String.Output( + display_name="texts", is_output_list=True, tooltip="Shuffled texts" + ), + ], + ) + + @classmethod + def execute(cls, images, texts, seed): + seed = seed[0] # Extract scalar + np.random.seed(seed % (2**32 - 1)) + indices = np.random.permutation(len(images)) + shuffled_images = [images[i] for i in indices] + shuffled_texts = [texts[i] for i in indices] + return io.NodeOutput(shuffled_images, shuffled_texts) + + +# ========== Text Transform Nodes ========== + + +class TextToLowercaseNode(TextProcessingNode): + node_id = "TextToLowercase" + display_name = "Text to Lowercase" + description = "Convert all texts to lowercase." + + @classmethod + def _process(cls, text): + return text.lower() + + +class TextToUppercaseNode(TextProcessingNode): + node_id = "TextToUppercase" + display_name = "Text to Uppercase" + description = "Convert all texts to uppercase." + + @classmethod + def _process(cls, text): + return text.upper() + + +class TruncateTextNode(TextProcessingNode): + node_id = "TruncateText" + display_name = "Truncate Text" + description = "Truncate all texts to a maximum length." + extra_inputs = [ + io.Int.Input( + "max_length", default=77, min=1, max=10000, tooltip="Maximum text length." + ), + ] + + @classmethod + def _process(cls, text, max_length): + return text[:max_length] + + +class AddTextPrefixNode(TextProcessingNode): + node_id = "AddTextPrefix" + display_name = "Add Text Prefix" + description = "Add a prefix to all texts." + extra_inputs = [ + io.String.Input("prefix", default="", tooltip="Prefix to add."), + ] + + @classmethod + def _process(cls, text, prefix): + return prefix + text + + +class AddTextSuffixNode(TextProcessingNode): + node_id = "AddTextSuffix" + display_name = "Add Text Suffix" + description = "Add a suffix to all texts." + extra_inputs = [ + io.String.Input("suffix", default="", tooltip="Suffix to add."), + ] + + @classmethod + def _process(cls, text, suffix): + return text + suffix + + +class ReplaceTextNode(TextProcessingNode): + node_id = "ReplaceText" + display_name = "Replace Text" + description = "Replace text in all texts." + extra_inputs = [ + io.String.Input("find", default="", tooltip="Text to find."), + io.String.Input("replace", default="", tooltip="Text to replace with."), + ] + + @classmethod + def _process(cls, text, find, replace): + return text.replace(find, replace) + + +class StripWhitespaceNode(TextProcessingNode): + node_id = "StripWhitespace" + display_name = "Strip Whitespace" + description = "Strip leading and trailing whitespace from all texts." + + @classmethod + def _process(cls, text): + return text.strip() + + +# ========== Group Processing Example Nodes ========== + + +class ImageDeduplicationNode(ImageProcessingNode): + """Remove duplicate or very similar images from the dataset using perceptual hashing.""" + + node_id = "ImageDeduplication" + display_name = "Image Deduplication" + description = "Remove duplicate or very similar images from the dataset." + is_group_process = True # Requires full list to compare images + extra_inputs = [ + io.Float.Input( + "similarity_threshold", + default=0.95, + min=0.0, + max=1.0, + tooltip="Similarity threshold (0-1). Higher means more similar. Images above this threshold are considered duplicates.", + ), + ] + + @classmethod + def _group_process(cls, images, similarity_threshold): + """Remove duplicate images using perceptual hashing.""" + if len(images) == 0: + return [] + + # Compute simple perceptual hash for each image + def compute_hash(img_tensor): + """Compute a simple perceptual hash by resizing to 8x8 and comparing to average.""" + img = tensor_to_pil(img_tensor) + # Resize to 8x8 + img_small = img.resize((8, 8), Image.Resampling.LANCZOS).convert("L") + # Get pixels + pixels = list(img_small.getdata()) + # Compute average + avg = sum(pixels) / len(pixels) + # Create hash (1 if above average, 0 otherwise) + hash_bits = "".join("1" if p > avg else "0" for p in pixels) + return hash_bits + + def hamming_distance(hash1, hash2): + """Compute Hamming distance between two hash strings.""" + return sum(c1 != c2 for c1, c2 in zip(hash1, hash2)) + + # Compute hashes for all images + hashes = [compute_hash(img) for img in images] + + # Find duplicates + keep_indices = [] + for i in range(len(images)): + is_duplicate = False + for j in keep_indices: + # Compare hashes + distance = hamming_distance(hashes[i], hashes[j]) + similarity = 1.0 - (distance / 64.0) # 64 bits total + if similarity >= similarity_threshold: + is_duplicate = True + logging.info( + f"Image {i} is similar to image {j} (similarity: {similarity:.3f}), skipping" + ) + break + + if not is_duplicate: + keep_indices.append(i) + + # Return only unique images + unique_images = [images[i] for i in keep_indices] + logging.info( + f"Deduplication: kept {len(unique_images)} out of {len(images)} images" + ) + return unique_images + + +class ImageGridNode(ImageProcessingNode): + """Combine multiple images into a single grid/collage.""" + + node_id = "ImageGrid" + display_name = "Image Grid" + description = "Arrange multiple images into a grid layout." + is_group_process = True # Requires full list to create grid + is_output_list = False # Outputs single grid image + extra_inputs = [ + io.Int.Input( + "columns", + default=4, + min=1, + max=20, + tooltip="Number of columns in the grid.", + ), + io.Int.Input( + "cell_width", + default=256, + min=32, + max=2048, + tooltip="Width of each cell in the grid.", + ), + io.Int.Input( + "cell_height", + default=256, + min=32, + max=2048, + tooltip="Height of each cell in the grid.", + ), + io.Int.Input( + "padding", default=4, min=0, max=50, tooltip="Padding between images." + ), + ] + + @classmethod + def _group_process(cls, images, columns, cell_width, cell_height, padding): + """Arrange images into a grid.""" + if len(images) == 0: + raise ValueError("Cannot create grid from empty image list") + + # Calculate grid dimensions + num_images = len(images) + rows = (num_images + columns - 1) // columns # Ceiling division + + # Calculate total grid size + grid_width = columns * cell_width + (columns - 1) * padding + grid_height = rows * cell_height + (rows - 1) * padding + + # Create blank grid + grid = Image.new("RGB", (grid_width, grid_height), (0, 0, 0)) + + # Place images + for idx, img_tensor in enumerate(images): + row = idx // columns + col = idx % columns + + # Convert to PIL and resize to cell size + img = tensor_to_pil(img_tensor) + img = img.resize((cell_width, cell_height), Image.Resampling.LANCZOS) + + # Calculate position + x = col * (cell_width + padding) + y = row * (cell_height + padding) + + # Paste into grid + grid.paste(img, (x, y)) + + logging.info( + f"Created {columns}x{rows} grid with {num_images} images ({grid_width}x{grid_height})" + ) + return pil_to_tensor(grid) + + +class MergeImageListsNode(ImageProcessingNode): + """Merge multiple image lists into a single list.""" + + node_id = "MergeImageLists" + display_name = "Merge Image Lists" + description = "Concatenate multiple image lists into one." + is_group_process = True # Receives images as list + + @classmethod + def _group_process(cls, images): + """Simply return the images list (already merged by input handling).""" + # When multiple list inputs are connected, they're concatenated + # For now, this is a simple pass-through + logging.info(f"Merged image list contains {len(images)} images") + return images + + +class MergeTextListsNode(TextProcessingNode): + """Merge multiple text lists into a single list.""" + + node_id = "MergeTextLists" + display_name = "Merge Text Lists" + description = "Concatenate multiple text lists into one." + is_group_process = True # Receives texts as list + + @classmethod + def _group_process(cls, texts): + """Simply return the texts list (already merged by input handling).""" + # When multiple list inputs are connected, they're concatenated + # For now, this is a simple pass-through + logging.info(f"Merged text list contains {len(texts)} texts") + return texts + + +# ========== Training Dataset Nodes ========== + + +class MakeTrainingDataset(io.ComfyNode): + """Encode images with VAE and texts with CLIP to create a training dataset.""" + + @classmethod + def define_schema(cls): + return io.Schema( + node_id="MakeTrainingDataset", + display_name="Make Training Dataset", + category="dataset", + is_experimental=True, + is_input_list=True, # images and texts as lists + inputs=[ + io.Image.Input("images", tooltip="List of images to encode."), + io.Vae.Input( + "vae", tooltip="VAE model for encoding images to latents." + ), + io.Clip.Input( + "clip", tooltip="CLIP model for encoding text to conditioning." + ), + io.String.Input( + "texts", + optional=True, + tooltip="List of text captions. Can be length n (matching images), 1 (repeated for all), or omitted (uses empty string).", + ), + ], + outputs=[ + io.Latent.Output( + display_name="latents", + is_output_list=True, + tooltip="List of latent dicts", + ), + io.Conditioning.Output( + display_name="conditioning", + is_output_list=True, + tooltip="List of conditioning lists", + ), + ], + ) + + @classmethod + def execute(cls, images, vae, clip, texts=None): + # Extract scalars (vae and clip are single values wrapped in lists) + vae = vae[0] + clip = clip[0] + + # Handle text list + num_images = len(images) + + if texts is None or len(texts) == 0: + # Treat as [""] for unconditional training + texts = [""] + + if len(texts) == 1 and num_images > 1: + # Repeat single text for all images + texts = texts * num_images + elif len(texts) != num_images: + raise ValueError( + f"Number of texts ({len(texts)}) does not match number of images ({num_images}). " + f"Text list should have length {num_images}, 1, or 0." + ) + + # Encode images with VAE + logging.info(f"Encoding {num_images} images with VAE...") + latents_list = [] # list[{"samples": tensor}] + for img_tensor in images: + # img_tensor is [1, H, W, 3] + latent_tensor = vae.encode(img_tensor[:, :, :, :3]) + latents_list.append({"samples": latent_tensor}) + + # Encode texts with CLIP + logging.info(f"Encoding {len(texts)} texts with CLIP...") + conditioning_list = [] # list[list[cond]] + for text in texts: + if text == "": + cond = clip.encode_from_tokens_scheduled(clip.tokenize("")) + else: + tokens = clip.tokenize(text) + cond = clip.encode_from_tokens_scheduled(tokens) + conditioning_list.append(cond) + + logging.info( + f"Created dataset with {len(latents_list)} latents and {len(conditioning_list)} conditioning." + ) + return io.NodeOutput(latents_list, conditioning_list) + + +class SaveTrainingDataset(io.ComfyNode): + """Save encoded training dataset (latents + conditioning) to disk.""" + + @classmethod + def define_schema(cls): + return io.Schema( + node_id="SaveTrainingDataset", + display_name="Save Training Dataset", + category="dataset", + is_experimental=True, + is_output_node=True, + is_input_list=True, # Receive lists + inputs=[ + io.Latent.Input( + "latents", + tooltip="List of latent dicts from MakeTrainingDataset.", + ), + io.Conditioning.Input( + "conditioning", + tooltip="List of conditioning lists from MakeTrainingDataset.", + ), + io.String.Input( + "folder_name", + default="training_dataset", + tooltip="Name of folder to save dataset (inside output directory).", + ), + io.Int.Input( + "shard_size", + default=1000, + min=1, + max=100000, + tooltip="Number of samples per shard file.", + ), + ], + outputs=[], + ) + + @classmethod + def execute(cls, latents, conditioning, folder_name, shard_size): + # Extract scalars + folder_name = folder_name[0] + shard_size = shard_size[0] + + # latents: list[{"samples": tensor}] + # conditioning: list[list[cond]] + + # Validate lengths match + if len(latents) != len(conditioning): + raise ValueError( + f"Number of latents ({len(latents)}) does not match number of conditions ({len(conditioning)}). " + f"Something went wrong in dataset preparation." + ) + + # Create output directory + output_dir = os.path.join(folder_paths.get_output_directory(), folder_name) + os.makedirs(output_dir, exist_ok=True) + + # Prepare data pairs + num_samples = len(latents) + num_shards = (num_samples + shard_size - 1) // shard_size # Ceiling division + + logging.info( + f"Saving {num_samples} samples to {num_shards} shards in {output_dir}..." + ) + + # Save data in shards + for shard_idx in range(num_shards): + start_idx = shard_idx * shard_size + end_idx = min(start_idx + shard_size, num_samples) + + # Get shard data (list of latent dicts and conditioning lists) + shard_data = { + "latents": latents[start_idx:end_idx], + "conditioning": conditioning[start_idx:end_idx], + } + + # Save shard + shard_filename = f"shard_{shard_idx:04d}.pkl" + shard_path = os.path.join(output_dir, shard_filename) + + with open(shard_path, "wb") as f: + torch.save(shard_data, f) + + logging.info( + f"Saved shard {shard_idx + 1}/{num_shards}: {shard_filename} ({end_idx - start_idx} samples)" + ) + + # Save metadata + metadata = { + "num_samples": num_samples, + "num_shards": num_shards, + "shard_size": shard_size, + } + metadata_path = os.path.join(output_dir, "metadata.json") + with open(metadata_path, "w") as f: + json.dump(metadata, f, indent=2) + + logging.info(f"Successfully saved {num_samples} samples to {output_dir}.") + return io.NodeOutput() + + +class LoadTrainingDataset(io.ComfyNode): + """Load encoded training dataset from disk.""" + + @classmethod + def define_schema(cls): + return io.Schema( + node_id="LoadTrainingDataset", + display_name="Load Training Dataset", + category="dataset", + is_experimental=True, + inputs=[ + io.String.Input( + "folder_name", + default="training_dataset", + tooltip="Name of folder containing the saved dataset (inside output directory).", + ), + ], + outputs=[ + io.Latent.Output( + display_name="latents", + is_output_list=True, + tooltip="List of latent dicts", + ), + io.Conditioning.Output( + display_name="conditioning", + is_output_list=True, + tooltip="List of conditioning lists", + ), + ], + ) + + @classmethod + def execute(cls, folder_name): + # Get dataset directory + dataset_dir = os.path.join(folder_paths.get_output_directory(), folder_name) + + if not os.path.exists(dataset_dir): + raise ValueError(f"Dataset directory not found: {dataset_dir}") + + # Find all shard files + shard_files = sorted( + [ + f + for f in os.listdir(dataset_dir) + if f.startswith("shard_") and f.endswith(".pkl") + ] + ) + + if not shard_files: + raise ValueError(f"No shard files found in {dataset_dir}") + + logging.info(f"Loading {len(shard_files)} shards from {dataset_dir}...") + + # Load all shards + all_latents = [] # list[{"samples": tensor}] + all_conditioning = [] # list[list[cond]] + + for shard_file in shard_files: + shard_path = os.path.join(dataset_dir, shard_file) + + with open(shard_path, "rb") as f: + shard_data = torch.load(f) + + all_latents.extend(shard_data["latents"]) + all_conditioning.extend(shard_data["conditioning"]) + + logging.info(f"Loaded {shard_file}: {len(shard_data['latents'])} samples") + + logging.info( + f"Successfully loaded {len(all_latents)} samples from {dataset_dir}." + ) + return io.NodeOutput(all_latents, all_conditioning) + + +# ========== Extension Setup ========== + + +class DatasetExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[io.ComfyNode]]: + return [ + # Data loading/saving nodes + LoadImageDataSetFromFolderNode, + LoadImageTextDataSetFromFolderNode, + SaveImageDataSetToFolderNode, + SaveImageTextDataSetToFolderNode, + # Image transform nodes + ResizeImagesToSameSizeNode, + ResizeImagesToPixelCountNode, + ResizeImagesByShorterEdgeNode, + ResizeImagesByLongerEdgeNode, + CenterCropImagesNode, + RandomCropImagesNode, + FlipImagesNode, + NormalizeImagesNode, + AdjustBrightnessNode, + AdjustContrastNode, + ShuffleDatasetNode, + ShuffleImageTextDatasetNode, + # Text transform nodes + TextToLowercaseNode, + TextToUppercaseNode, + TruncateTextNode, + AddTextPrefixNode, + AddTextSuffixNode, + ReplaceTextNode, + StripWhitespaceNode, + # Group processing examples + ImageDeduplicationNode, + ImageGridNode, + MergeImageListsNode, + MergeTextListsNode, + # Training dataset nodes + MakeTrainingDataset, + SaveTrainingDataset, + LoadTrainingDataset, + ] + + +async def comfy_entrypoint() -> DatasetExtension: + return DatasetExtension() diff --git a/comfy_extras/nodes_train.py b/comfy_extras/nodes_train.py index 9e6ec6780..cb24ab709 100644 --- a/comfy_extras/nodes_train.py +++ b/comfy_extras/nodes_train.py @@ -1,15 +1,13 @@ -import datetime -import json import logging import os import numpy as np import safetensors import torch -from PIL import Image, ImageDraw, ImageFont -from PIL.PngImagePlugin import PngInfo import torch.utils.checkpoint -import tqdm +from tqdm.auto import trange +from PIL import Image, ImageDraw, ImageFont +from typing_extensions import override import comfy.samplers import comfy.sd @@ -18,9 +16,9 @@ import comfy.model_management import comfy_extras.nodes_custom_sampler import folder_paths import node_helpers -from comfy.cli_args import args -from comfy.comfy_types.node_typing import IO from comfy.weight_adapter import adapters, adapter_maps +from comfy_api.latest import ComfyExtension, io, ui +from comfy.utils import ProgressBar def make_batch_extra_option_dict(d, indicies, full_size=None): @@ -56,7 +54,18 @@ def process_cond_list(d, prefix=""): class TrainSampler(comfy.samplers.Sampler): - def __init__(self, loss_fn, optimizer, loss_callback=None, batch_size=1, grad_acc=1, total_steps=1, seed=0, training_dtype=torch.bfloat16): + def __init__( + self, + loss_fn, + optimizer, + loss_callback=None, + batch_size=1, + grad_acc=1, + total_steps=1, + seed=0, + training_dtype=torch.bfloat16, + real_dataset=None, + ): self.loss_fn = loss_fn self.optimizer = optimizer self.loss_callback = loss_callback @@ -65,54 +74,138 @@ class TrainSampler(comfy.samplers.Sampler): self.grad_acc = grad_acc self.seed = seed self.training_dtype = training_dtype + self.real_dataset: list[torch.Tensor] | None = real_dataset - def sample(self, model_wrap, sigmas, extra_args, callback, noise, latent_image=None, denoise_mask=None, disable_pbar=False): + def fwd_bwd( + self, + model_wrap, + batch_sigmas, + batch_noise, + batch_latent, + cond, + indicies, + extra_args, + dataset_size, + bwd=True, + ): + xt = model_wrap.inner_model.model_sampling.noise_scaling( + batch_sigmas, batch_noise, batch_latent, False + ) + x0 = model_wrap.inner_model.model_sampling.noise_scaling( + torch.zeros_like(batch_sigmas), + torch.zeros_like(batch_noise), + batch_latent, + False, + ) + + model_wrap.conds["positive"] = [cond[i] for i in indicies] + batch_extra_args = make_batch_extra_option_dict( + extra_args, indicies, full_size=dataset_size + ) + + with torch.autocast(xt.device.type, dtype=self.training_dtype): + x0_pred = model_wrap( + xt.requires_grad_(True), + batch_sigmas.requires_grad_(True), + **batch_extra_args, + ) + loss = self.loss_fn(x0_pred, x0) + if bwd: + bwd_loss = loss / self.grad_acc + bwd_loss.backward() + return loss + + def sample( + self, + model_wrap, + sigmas, + extra_args, + callback, + noise, + latent_image=None, + denoise_mask=None, + disable_pbar=False, + ): model_wrap.conds = process_cond_list(model_wrap.conds) cond = model_wrap.conds["positive"] dataset_size = sigmas.size(0) torch.cuda.empty_cache() - for i in (pbar:=tqdm.trange(self.total_steps, desc="Training LoRA", smoothing=0.01, disable=not comfy.utils.PROGRESS_BAR_ENABLED)): - noisegen = comfy_extras.nodes_custom_sampler.Noise_RandomNoise(self.seed + i * 1000) - indicies = torch.randperm(dataset_size)[:self.batch_size].tolist() - - batch_latent = torch.stack([latent_image[i] for i in indicies]) - batch_noise = noisegen.generate_noise({"samples": batch_latent}).to(batch_latent.device) - batch_sigmas = [ - model_wrap.inner_model.model_sampling.percent_to_sigma( - torch.rand((1,)).item() - ) for _ in range(min(self.batch_size, dataset_size)) - ] - batch_sigmas = torch.tensor(batch_sigmas).to(batch_latent.device) - - xt = model_wrap.inner_model.model_sampling.noise_scaling( - batch_sigmas, - batch_noise, - batch_latent, - False + ui_pbar = ProgressBar(self.total_steps) + for i in ( + pbar := trange( + self.total_steps, + desc="Training LoRA", + smoothing=0.01, + disable=not comfy.utils.PROGRESS_BAR_ENABLED, ) - x0 = model_wrap.inner_model.model_sampling.noise_scaling( - torch.zeros_like(batch_sigmas), - torch.zeros_like(batch_noise), - batch_latent, - False + ): + noisegen = comfy_extras.nodes_custom_sampler.Noise_RandomNoise( + self.seed + i * 1000 ) + indicies = torch.randperm(dataset_size)[: self.batch_size].tolist() - model_wrap.conds["positive"] = [ - cond[i] for i in indicies - ] - batch_extra_args = make_batch_extra_option_dict(extra_args, indicies, full_size=dataset_size) + if self.real_dataset is None: + batch_latent = torch.stack([latent_image[i] for i in indicies]) + batch_noise = noisegen.generate_noise({"samples": batch_latent}).to( + batch_latent.device + ) + batch_sigmas = [ + model_wrap.inner_model.model_sampling.percent_to_sigma( + torch.rand((1,)).item() + ) + for _ in range(min(self.batch_size, dataset_size)) + ] + batch_sigmas = torch.tensor(batch_sigmas).to(batch_latent.device) - with torch.autocast(xt.device.type, dtype=self.training_dtype): - x0_pred = model_wrap(xt, batch_sigmas, **batch_extra_args) - loss = self.loss_fn(x0_pred, x0) - loss.backward() - if self.loss_callback: - self.loss_callback(loss.item()) - pbar.set_postfix({"loss": f"{loss.item():.4f}"}) + loss = self.fwd_bwd( + model_wrap, + batch_sigmas, + batch_noise, + batch_latent, + cond, + indicies, + extra_args, + dataset_size, + bwd=True, + ) + if self.loss_callback: + self.loss_callback(loss.item()) + pbar.set_postfix({"loss": f"{loss.item():.4f}"}) + else: + total_loss = 0 + for index in indicies: + single_latent = self.real_dataset[index].to(latent_image) + batch_noise = noisegen.generate_noise( + {"samples": single_latent} + ).to(single_latent.device) + batch_sigmas = ( + model_wrap.inner_model.model_sampling.percent_to_sigma( + torch.rand((1,)).item() + ) + ) + batch_sigmas = torch.tensor([batch_sigmas]).to(single_latent.device) + loss = self.fwd_bwd( + model_wrap, + batch_sigmas, + batch_noise, + single_latent, + cond, + [index], + extra_args, + dataset_size, + bwd=False, + ) + total_loss += loss + total_loss = total_loss / self.grad_acc / len(indicies) + total_loss.backward() + if self.loss_callback: + self.loss_callback(total_loss.item()) + pbar.set_postfix({"loss": f"{total_loss.item():.4f}"}) - if (i+1) % self.grad_acc == 0: + if (i + 1) % self.grad_acc == 0: self.optimizer.step() self.optimizer.zero_grad() + ui_pbar.update(1) torch.cuda.empty_cache() return torch.zeros_like(latent_image) @@ -134,233 +227,6 @@ class BiasDiff(torch.nn.Module): return self.passive_memory_usage() -def load_and_process_images(image_files, input_dir, resize_method="None", w=None, h=None): - """Utility function to load and process a list of images. - - Args: - image_files: List of image filenames - input_dir: Base directory containing the images - resize_method: How to handle images of different sizes ("None", "Stretch", "Crop", "Pad") - - Returns: - torch.Tensor: Batch of processed images - """ - if not image_files: - raise ValueError("No valid images found in input") - - output_images = [] - - for file in image_files: - image_path = os.path.join(input_dir, file) - img = node_helpers.pillow(Image.open, image_path) - - if img.mode == "I": - img = img.point(lambda i: i * (1 / 255)) - img = img.convert("RGB") - - if w is None and h is None: - w, h = img.size[0], img.size[1] - - # Resize image to first image - if img.size[0] != w or img.size[1] != h: - if resize_method == "Stretch": - img = img.resize((w, h), Image.Resampling.LANCZOS) - elif resize_method == "Crop": - img = img.crop((0, 0, w, h)) - elif resize_method == "Pad": - img = img.resize((w, h), Image.Resampling.LANCZOS) - elif resize_method == "None": - raise ValueError( - "Your input image size does not match the first image in the dataset. Either select a valid resize method or use the same size for all images." - ) - - img_array = np.array(img).astype(np.float32) / 255.0 - img_tensor = torch.from_numpy(img_array)[None,] - output_images.append(img_tensor) - - return torch.cat(output_images, dim=0) - - -class LoadImageSetNode: - @classmethod - def INPUT_TYPES(s): - return { - "required": { - "images": ( - [ - f - for f in os.listdir(folder_paths.get_input_directory()) - if f.endswith((".png", ".jpg", ".jpeg", ".webp", ".bmp", ".gif", ".jpe", ".apng", ".tif", ".tiff")) - ], - {"image_upload": True, "allow_batch": True}, - ) - }, - "optional": { - "resize_method": ( - ["None", "Stretch", "Crop", "Pad"], - {"default": "None"}, - ), - }, - } - - INPUT_IS_LIST = True - RETURN_TYPES = ("IMAGE",) - FUNCTION = "load_images" - CATEGORY = "loaders" - EXPERIMENTAL = True - DESCRIPTION = "Loads a batch of images from a directory for training." - - @classmethod - def VALIDATE_INPUTS(s, images, resize_method): - filenames = images[0] if isinstance(images[0], list) else images - - for image in filenames: - if not folder_paths.exists_annotated_filepath(image): - return "Invalid image file: {}".format(image) - return True - - def load_images(self, input_files, resize_method): - input_dir = folder_paths.get_input_directory() - valid_extensions = [".png", ".jpg", ".jpeg", ".webp", ".bmp", ".gif", ".jpe", ".apng", ".tif", ".tiff"] - image_files = [ - f - for f in input_files - if any(f.lower().endswith(ext) for ext in valid_extensions) - ] - output_tensor = load_and_process_images(image_files, input_dir, resize_method) - return (output_tensor,) - - -class LoadImageSetFromFolderNode: - @classmethod - def INPUT_TYPES(s): - return { - "required": { - "folder": (folder_paths.get_input_subfolders(), {"tooltip": "The folder to load images from."}) - }, - "optional": { - "resize_method": ( - ["None", "Stretch", "Crop", "Pad"], - {"default": "None"}, - ), - }, - } - - RETURN_TYPES = ("IMAGE",) - FUNCTION = "load_images" - CATEGORY = "loaders" - EXPERIMENTAL = True - DESCRIPTION = "Loads a batch of images from a directory for training." - - def load_images(self, folder, resize_method): - sub_input_dir = os.path.join(folder_paths.get_input_directory(), folder) - valid_extensions = [".png", ".jpg", ".jpeg", ".webp"] - image_files = [ - f - for f in os.listdir(sub_input_dir) - if any(f.lower().endswith(ext) for ext in valid_extensions) - ] - output_tensor = load_and_process_images(image_files, sub_input_dir, resize_method) - return (output_tensor,) - - -class LoadImageTextSetFromFolderNode: - @classmethod - def INPUT_TYPES(s): - return { - "required": { - "folder": (folder_paths.get_input_subfolders(), {"tooltip": "The folder to load images from."}), - "clip": (IO.CLIP, {"tooltip": "The CLIP model used for encoding the text."}), - }, - "optional": { - "resize_method": ( - ["None", "Stretch", "Crop", "Pad"], - {"default": "None"}, - ), - "width": ( - IO.INT, - { - "default": -1, - "min": -1, - "max": 10000, - "step": 1, - "tooltip": "The width to resize the images to. -1 means use the original width.", - }, - ), - "height": ( - IO.INT, - { - "default": -1, - "min": -1, - "max": 10000, - "step": 1, - "tooltip": "The height to resize the images to. -1 means use the original height.", - }, - ) - }, - } - - RETURN_TYPES = ("IMAGE", IO.CONDITIONING,) - FUNCTION = "load_images" - CATEGORY = "loaders" - EXPERIMENTAL = True - DESCRIPTION = "Loads a batch of images and caption from a directory for training." - - def load_images(self, folder, clip, resize_method, width=None, height=None): - if clip is None: - raise RuntimeError("ERROR: clip input is invalid: None\n\nIf the clip is from a checkpoint loader node your checkpoint does not contain a valid clip or text encoder model.") - - logging.info(f"Loading images from folder: {folder}") - - sub_input_dir = os.path.join(folder_paths.get_input_directory(), folder) - valid_extensions = [".png", ".jpg", ".jpeg", ".webp"] - - image_files = [] - for item in os.listdir(sub_input_dir): - path = os.path.join(sub_input_dir, item) - if any(item.lower().endswith(ext) for ext in valid_extensions): - image_files.append(path) - elif os.path.isdir(path): - # Support kohya-ss/sd-scripts folder structure - repeat = 1 - if item.split("_")[0].isdigit(): - repeat = int(item.split("_")[0]) - image_files.extend([ - os.path.join(path, f) for f in os.listdir(path) if any(f.lower().endswith(ext) for ext in valid_extensions) - ] * repeat) - - caption_file_path = [ - f.replace(os.path.splitext(f)[1], ".txt") - for f in image_files - ] - captions = [] - for caption_file in caption_file_path: - caption_path = os.path.join(sub_input_dir, caption_file) - if os.path.exists(caption_path): - with open(caption_path, "r", encoding="utf-8") as f: - caption = f.read().strip() - captions.append(caption) - else: - captions.append("") - - width = width if width != -1 else None - height = height if height != -1 else None - output_tensor = load_and_process_images(image_files, sub_input_dir, resize_method, width, height) - - logging.info(f"Loaded {len(output_tensor)} images from {sub_input_dir}.") - - logging.info(f"Encoding captions from {sub_input_dir}.") - conditions = [] - empty_cond = clip.encode_from_tokens_scheduled(clip.tokenize("")) - for text in captions: - if text == "": - conditions.append(empty_cond) - tokens = clip.tokenize(text) - conditions.extend(clip.encode_from_tokens_scheduled(tokens)) - logging.info(f"Encoded {len(conditions)} captions from {sub_input_dir}.") - return (output_tensor, conditions) - - def draw_loss_graph(loss_map, steps): width, height = 500, 300 img = Image.new("RGB", (width, height), "white") @@ -379,10 +245,14 @@ def draw_loss_graph(loss_map, steps): return img -def find_all_highest_child_module_with_forward(model: torch.nn.Module, result = None, name = None): +def find_all_highest_child_module_with_forward( + model: torch.nn.Module, result=None, name=None +): if result is None: result = [] - elif hasattr(model, "forward") and not isinstance(model, (torch.nn.ModuleList, torch.nn.Sequential, torch.nn.ModuleDict)): + elif hasattr(model, "forward") and not isinstance( + model, (torch.nn.ModuleList, torch.nn.Sequential, torch.nn.ModuleDict) + ): result.append(model) logging.debug(f"Found module with forward: {name} ({model.__class__.__name__})") return result @@ -396,12 +266,13 @@ def patch(m): if not hasattr(m, "forward"): return org_forward = m.forward + def fwd(args, kwargs): return org_forward(*args, **kwargs) + def checkpointing_fwd(*args, **kwargs): - return torch.utils.checkpoint.checkpoint( - fwd, args, kwargs, use_reentrant=False - ) + return torch.utils.checkpoint.checkpoint(fwd, args, kwargs, use_reentrant=False) + m.org_forward = org_forward m.forward = checkpointing_fwd @@ -412,130 +283,126 @@ def unpatch(m): del m.org_forward -class TrainLoraNode: +class TrainLoraNode(io.ComfyNode): @classmethod - def INPUT_TYPES(s): - return { - "required": { - "model": (IO.MODEL, {"tooltip": "The model to train the LoRA on."}), - "latents": ( - "LATENT", - { - "tooltip": "The Latents to use for training, serve as dataset/input of the model." - }, + def define_schema(cls): + return io.Schema( + node_id="TrainLoraNode", + display_name="Train LoRA", + category="training", + is_experimental=True, + is_input_list=True, # All inputs become lists + inputs=[ + io.Model.Input("model", tooltip="The model to train the LoRA on."), + io.Latent.Input( + "latents", + tooltip="The Latents to use for training, serve as dataset/input of the model.", ), - "positive": ( - IO.CONDITIONING, - {"tooltip": "The positive conditioning to use for training."}, + io.Conditioning.Input( + "positive", tooltip="The positive conditioning to use for training." ), - "batch_size": ( - IO.INT, - { - "default": 1, - "min": 1, - "max": 10000, - "step": 1, - "tooltip": "The batch size to use for training.", - }, + io.Int.Input( + "batch_size", + default=1, + min=1, + max=10000, + tooltip="The batch size to use for training.", ), - "grad_accumulation_steps": ( - IO.INT, - { - "default": 1, - "min": 1, - "max": 1024, - "step": 1, - "tooltip": "The number of gradient accumulation steps to use for training.", - } + io.Int.Input( + "grad_accumulation_steps", + default=1, + min=1, + max=1024, + tooltip="The number of gradient accumulation steps to use for training.", ), - "steps": ( - IO.INT, - { - "default": 16, - "min": 1, - "max": 100000, - "tooltip": "The number of steps to train the LoRA for.", - }, + io.Int.Input( + "steps", + default=16, + min=1, + max=100000, + tooltip="The number of steps to train the LoRA for.", ), - "learning_rate": ( - IO.FLOAT, - { - "default": 0.0005, - "min": 0.0000001, - "max": 1.0, - "step": 0.000001, - "tooltip": "The learning rate to use for training.", - }, + io.Float.Input( + "learning_rate", + default=0.0005, + min=0.0000001, + max=1.0, + step=0.0000001, + tooltip="The learning rate to use for training.", ), - "rank": ( - IO.INT, - { - "default": 8, - "min": 1, - "max": 128, - "tooltip": "The rank of the LoRA layers.", - }, + io.Int.Input( + "rank", + default=8, + min=1, + max=128, + tooltip="The rank of the LoRA layers.", ), - "optimizer": ( - ["AdamW", "Adam", "SGD", "RMSprop"], - { - "default": "AdamW", - "tooltip": "The optimizer to use for training.", - }, + io.Combo.Input( + "optimizer", + options=["AdamW", "Adam", "SGD", "RMSprop"], + default="AdamW", + tooltip="The optimizer to use for training.", ), - "loss_function": ( - ["MSE", "L1", "Huber", "SmoothL1"], - { - "default": "MSE", - "tooltip": "The loss function to use for training.", - }, + io.Combo.Input( + "loss_function", + options=["MSE", "L1", "Huber", "SmoothL1"], + default="MSE", + tooltip="The loss function to use for training.", ), - "seed": ( - IO.INT, - { - "default": 0, - "min": 0, - "max": 0xFFFFFFFFFFFFFFFF, - "tooltip": "The seed to use for training (used in generator for LoRA weight initialization and noise sampling)", - }, + io.Int.Input( + "seed", + default=0, + min=0, + max=0xFFFFFFFFFFFFFFFF, + tooltip="The seed to use for training (used in generator for LoRA weight initialization and noise sampling)", ), - "training_dtype": ( - ["bf16", "fp32"], - {"default": "bf16", "tooltip": "The dtype to use for training."}, + io.Combo.Input( + "training_dtype", + options=["bf16", "fp32"], + default="bf16", + tooltip="The dtype to use for training.", ), - "lora_dtype": ( - ["bf16", "fp32"], - {"default": "bf16", "tooltip": "The dtype to use for lora."}, + io.Combo.Input( + "lora_dtype", + options=["bf16", "fp32"], + default="bf16", + tooltip="The dtype to use for lora.", ), - "algorithm": ( - list(adapter_maps.keys()), - {"default": list(adapter_maps.keys())[0], "tooltip": "The algorithm to use for training."}, + io.Combo.Input( + "algorithm", + options=list(adapter_maps.keys()), + default=list(adapter_maps.keys())[0], + tooltip="The algorithm to use for training.", ), - "gradient_checkpointing": ( - IO.BOOLEAN, - { - "default": True, - "tooltip": "Use gradient checkpointing for training.", - } + io.Boolean.Input( + "gradient_checkpointing", + default=True, + tooltip="Use gradient checkpointing for training.", ), - "existing_lora": ( - folder_paths.get_filename_list("loras") + ["[None]"], - { - "default": "[None]", - "tooltip": "The existing LoRA to append to. Set to None for new LoRA.", - }, + io.Combo.Input( + "existing_lora", + options=folder_paths.get_filename_list("loras") + ["[None]"], + default="[None]", + tooltip="The existing LoRA to append to. Set to None for new LoRA.", ), - }, - } + ], + outputs=[ + io.Model.Output( + display_name="model", tooltip="Model with LoRA applied" + ), + io.Custom("LORA_MODEL").Output( + display_name="lora", tooltip="LoRA weights" + ), + io.Custom("LOSS_MAP").Output( + display_name="loss_map", tooltip="Loss history" + ), + io.Int.Output(display_name="steps", tooltip="Total training steps"), + ], + ) - RETURN_TYPES = (IO.MODEL, IO.LORA_MODEL, IO.LOSS_MAP, IO.INT) - RETURN_NAMES = ("model_with_lora", "lora", "loss", "steps") - FUNCTION = "train" - CATEGORY = "training" - EXPERIMENTAL = True - - def train( - self, + @classmethod + def execute( + cls, model, latents, positive, @@ -553,13 +420,74 @@ class TrainLoraNode: gradient_checkpointing, existing_lora, ): + # Extract scalars from lists (due to is_input_list=True) + model = model[0] + batch_size = batch_size[0] + steps = steps[0] + grad_accumulation_steps = grad_accumulation_steps[0] + learning_rate = learning_rate[0] + rank = rank[0] + optimizer = optimizer[0] + loss_function = loss_function[0] + seed = seed[0] + training_dtype = training_dtype[0] + lora_dtype = lora_dtype[0] + algorithm = algorithm[0] + gradient_checkpointing = gradient_checkpointing[0] + existing_lora = existing_lora[0] + + # Handle latents - either single dict or list of dicts + if len(latents) == 1: + latents = latents[0]["samples"] # Single latent dict + else: + latent_list = [] + for latent in latents: + latent = latent["samples"] + bs = latent.shape[0] + if bs != 1: + for sub_latent in latent: + latent_list.append(sub_latent[None]) + else: + latent_list.append(latent) + latents = latent_list + + # Handle conditioning - either single list or list of lists + if len(positive) == 1: + positive = positive[0] # Single conditioning list + else: + # Multiple conditioning lists - flatten + flat_positive = [] + for cond in positive: + if isinstance(cond, list): + flat_positive.extend(cond) + else: + flat_positive.append(cond) + positive = flat_positive + mp = model.clone() dtype = node_helpers.string_to_torch_dtype(training_dtype) lora_dtype = node_helpers.string_to_torch_dtype(lora_dtype) mp.set_model_compute_dtype(dtype) - latents = latents["samples"].to(dtype) - num_images = latents.shape[0] + # latents here can be list of different size latent or one large batch + if isinstance(latents, list): + all_shapes = set() + latents = [t.to(dtype) for t in latents] + for latent in latents: + all_shapes.add(latent.shape) + logging.info(f"Latent shapes: {all_shapes}") + if len(all_shapes) > 1: + multi_res = True + else: + multi_res = False + latents = torch.cat(latents, dim=0) + num_images = len(latents) + elif isinstance(latents, torch.Tensor): + latents = latents.to(dtype) + num_images = latents.shape[0] + else: + logging.error(f"Invalid latents type: {type(latents)}") + logging.info(f"Total Images: {num_images}, Total Captions: {len(positive)}") if len(positive) == 1 and num_images > 1: positive = positive * num_images @@ -591,9 +519,7 @@ class TrainLoraNode: shape = m.weight.shape if len(shape) >= 2: alpha = float(existing_weights.get(f"{key}.alpha", 1.0)) - dora_scale = existing_weights.get( - f"{key}.dora_scale", None - ) + dora_scale = existing_weights.get(f"{key}.dora_scale", None) for adapter_cls in adapters: existing_adapter = adapter_cls.load( n, existing_weights, alpha, dora_scale @@ -605,7 +531,9 @@ class TrainLoraNode: adapter_cls = adapter_maps[algorithm] if existing_adapter is not None: - train_adapter = existing_adapter.to_train().to(lora_dtype) + train_adapter = existing_adapter.to_train().to( + lora_dtype + ) else: # Use LoRA with alpha=1.0 by default train_adapter = adapter_cls.create_train( @@ -629,7 +557,9 @@ class TrainLoraNode: if hasattr(m, "bias") and m.bias is not None: key = "{}.bias".format(n) bias = torch.nn.Parameter( - torch.zeros(m.bias.shape, dtype=lora_dtype, requires_grad=True) + torch.zeros( + m.bias.shape, dtype=lora_dtype, requires_grad=True + ) ) bias_module = BiasDiff(bias) lora_sd["{}.diff_b".format(n)] = bias @@ -657,24 +587,31 @@ class TrainLoraNode: # setup models if gradient_checkpointing: - for m in find_all_highest_child_module_with_forward(mp.model.diffusion_model): + for m in find_all_highest_child_module_with_forward( + mp.model.diffusion_model + ): patch(m) mp.model.requires_grad_(False) - comfy.model_management.load_models_gpu([mp], memory_required=1e20, force_full_load=True) + comfy.model_management.load_models_gpu( + [mp], memory_required=1e20, force_full_load=True + ) # Setup sampler and guider like in test script loss_map = {"loss": []} + def loss_callback(loss): loss_map["loss"].append(loss) + train_sampler = TrainSampler( criterion, optimizer, loss_callback=loss_callback, batch_size=batch_size, grad_acc=grad_accumulation_steps, - total_steps=steps*grad_accumulation_steps, + total_steps=steps * grad_accumulation_steps, seed=seed, - training_dtype=dtype + training_dtype=dtype, + real_dataset=latents if multi_res else None, ) guider = comfy_extras.nodes_custom_sampler.Guider_Basic(mp) guider.set_conds(positive) # Set conditioning from input @@ -684,12 +621,15 @@ class TrainLoraNode: # Generate dummy sigmas and noise sigmas = torch.tensor(range(num_images)) noise = comfy_extras.nodes_custom_sampler.Noise_RandomNoise(seed) + if multi_res: + # use first latent as dummy latent if multi_res + latents = latents[0].repeat(num_images, 1, 1, 1) guider.sample( noise.generate_noise({"samples": latents}), latents, train_sampler, sigmas, - seed=noise.seed + seed=noise.seed, ) finally: for m in mp.model.modules(): @@ -702,111 +642,118 @@ class TrainLoraNode: for param in lora_sd: lora_sd[param] = lora_sd[param].to(lora_dtype) - return (mp, lora_sd, loss_map, steps + existing_steps) + return io.NodeOutput(mp, lora_sd, loss_map, steps + existing_steps) -class LoraModelLoader: - def __init__(self): - self.loaded_lora = None +class LoraModelLoader(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="LoraModelLoader", + display_name="Load LoRA Model", + category="loaders", + is_experimental=True, + inputs=[ + io.Model.Input( + "model", tooltip="The diffusion model the LoRA will be applied to." + ), + io.Custom("LORA_MODEL").Input( + "lora", tooltip="The LoRA model to apply to the diffusion model." + ), + io.Float.Input( + "strength_model", + default=1.0, + min=-100.0, + max=100.0, + tooltip="How strongly to modify the diffusion model. This value can be negative.", + ), + ], + outputs=[ + io.Model.Output( + display_name="model", tooltip="The modified diffusion model." + ), + ], + ) @classmethod - def INPUT_TYPES(s): - return { - "required": { - "model": ("MODEL", {"tooltip": "The diffusion model the LoRA will be applied to."}), - "lora": (IO.LORA_MODEL, {"tooltip": "The LoRA model to apply to the diffusion model."}), - "strength_model": ("FLOAT", {"default": 1.0, "min": -100.0, "max": 100.0, "step": 0.01, "tooltip": "How strongly to modify the diffusion model. This value can be negative."}), - } - } - - RETURN_TYPES = ("MODEL",) - OUTPUT_TOOLTIPS = ("The modified diffusion model.",) - FUNCTION = "load_lora_model" - - CATEGORY = "loaders" - DESCRIPTION = "Load Trained LoRA weights from Train LoRA node." - EXPERIMENTAL = True - - def load_lora_model(self, model, lora, strength_model): + def execute(cls, model, lora, strength_model): if strength_model == 0: - return (model, ) + return io.NodeOutput(model) - model_lora, _ = comfy.sd.load_lora_for_models(model, None, lora, strength_model, 0) - return (model_lora, ) + model_lora, _ = comfy.sd.load_lora_for_models( + model, None, lora, strength_model, 0 + ) + return io.NodeOutput(model_lora) -class SaveLoRA: - def __init__(self): - self.output_dir = folder_paths.get_output_directory() +class SaveLoRA(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="SaveLoRA", + display_name="Save LoRA Weights", + category="loaders", + is_experimental=True, + is_output_node=True, + inputs=[ + io.Custom("LORA_MODEL").Input( + "lora", + tooltip="The LoRA model to save. Do not use the model with LoRA layers.", + ), + io.String.Input( + "prefix", + default="loras/ComfyUI_trained_lora", + tooltip="The prefix to use for the saved LoRA file.", + ), + io.Int.Input( + "steps", + optional=True, + tooltip="Optional: The number of steps to LoRA has been trained for, used to name the saved file.", + ), + ], + outputs=[], + ) @classmethod - def INPUT_TYPES(s): - return { - "required": { - "lora": ( - IO.LORA_MODEL, - { - "tooltip": "The LoRA model to save. Do not use the model with LoRA layers." - }, - ), - "prefix": ( - "STRING", - { - "default": "loras/ComfyUI_trained_lora", - "tooltip": "The prefix to use for the saved LoRA file.", - }, - ), - }, - "optional": { - "steps": ( - IO.INT, - { - "forceInput": True, - "tooltip": "Optional: The number of steps to LoRA has been trained for, used to name the saved file.", - }, - ), - }, - } - - RETURN_TYPES = () - FUNCTION = "save" - CATEGORY = "loaders" - EXPERIMENTAL = True - OUTPUT_NODE = True - - def save(self, lora, prefix, steps=None): - full_output_folder, filename, counter, subfolder, filename_prefix = folder_paths.get_save_image_path(prefix, self.output_dir) + def execute(cls, lora, prefix, steps=None): + output_dir = folder_paths.get_output_directory() + full_output_folder, filename, counter, subfolder, filename_prefix = ( + folder_paths.get_save_image_path(prefix, output_dir) + ) if steps is None: output_checkpoint = f"{filename}_{counter:05}_.safetensors" else: output_checkpoint = f"{filename}_{steps}_steps_{counter:05}_.safetensors" output_checkpoint = os.path.join(full_output_folder, output_checkpoint) safetensors.torch.save_file(lora, output_checkpoint) - return {} + return io.NodeOutput() -class LossGraphNode: - def __init__(self): - self.output_dir = folder_paths.get_temp_directory() +class LossGraphNode(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="LossGraphNode", + display_name="Plot Loss Graph", + category="training", + is_experimental=True, + is_output_node=True, + inputs=[ + io.Custom("LOSS_MAP").Input( + "loss", tooltip="Loss map from training node." + ), + io.String.Input( + "filename_prefix", + default="loss_graph", + tooltip="Prefix for the saved loss graph image.", + ), + ], + outputs=[], + hidden=[io.Hidden.prompt, io.Hidden.extra_pnginfo], + ) @classmethod - def INPUT_TYPES(s): - return { - "required": { - "loss": (IO.LOSS_MAP, {"default": {}}), - "filename_prefix": (IO.STRING, {"default": "loss_graph"}), - }, - "hidden": {"prompt": "PROMPT", "extra_pnginfo": "EXTRA_PNGINFO"}, - } - - RETURN_TYPES = () - FUNCTION = "plot_loss" - OUTPUT_NODE = True - CATEGORY = "training" - EXPERIMENTAL = True - DESCRIPTION = "Plots the loss graph and saves it to the output directory." - - def plot_loss(self, loss, filename_prefix, prompt=None, extra_pnginfo=None): + def execute(cls, loss, filename_prefix, prompt=None, extra_pnginfo=None): loss_values = loss["loss"] width, height = 800, 480 margin = 40 @@ -849,47 +796,27 @@ class LossGraphNode: (margin - 30, height - 10), f"{min_loss:.2f}", font=font, fill="black" ) - metadata = None - if not args.disable_metadata: - metadata = PngInfo() - if prompt is not None: - metadata.add_text("prompt", json.dumps(prompt)) - if extra_pnginfo is not None: - for x in extra_pnginfo: - metadata.add_text(x, json.dumps(extra_pnginfo[x])) + # Convert PIL image to tensor for PreviewImage + img_array = np.array(img).astype(np.float32) / 255.0 + img_tensor = torch.from_numpy(img_array)[None,] # [1, H, W, 3] - date = datetime.datetime.now().strftime("%Y%m%d_%H%M%S") - img.save( - os.path.join(self.output_dir, f"{filename_prefix}_{date}.png"), - pnginfo=metadata, - ) - return { - "ui": { - "images": [ - { - "filename": f"{filename_prefix}_{date}.png", - "subfolder": "", - "type": "temp", - } - ] - } - } + # Return preview UI + return io.NodeOutput(ui=ui.PreviewImage(img_tensor, cls=cls)) -NODE_CLASS_MAPPINGS = { - "TrainLoraNode": TrainLoraNode, - "SaveLoRANode": SaveLoRA, - "LoraModelLoader": LoraModelLoader, - "LoadImageSetFromFolderNode": LoadImageSetFromFolderNode, - "LoadImageTextSetFromFolderNode": LoadImageTextSetFromFolderNode, - "LossGraphNode": LossGraphNode, -} +# ========== Extension Setup ========== -NODE_DISPLAY_NAME_MAPPINGS = { - "TrainLoraNode": "Train LoRA", - "SaveLoRANode": "Save LoRA Weights", - "LoraModelLoader": "Load LoRA Model", - "LoadImageSetFromFolderNode": "Load Image Dataset from Folder", - "LoadImageTextSetFromFolderNode": "Load Image and Text Dataset from Folder", - "LossGraphNode": "Plot Loss Graph", -} + +class TrainingExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[io.ComfyNode]]: + return [ + TrainLoraNode, + LoraModelLoader, + SaveLoRA, + LossGraphNode, + ] + + +async def comfy_entrypoint() -> TrainingExtension: + return TrainingExtension() diff --git a/nodes.py b/nodes.py index f4835c02e..bf73eb90e 100644 --- a/nodes.py +++ b/nodes.py @@ -2278,6 +2278,7 @@ async def init_builtin_extra_nodes(): "nodes_images.py", "nodes_video_model.py", "nodes_train.py", + "nodes_dataset.py", "nodes_sag.py", "nodes_perpneg.py", "nodes_stable3d.py", From eaf68c9b5bbfbcdac8988741f3948678c9465c1d Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 26 Nov 2025 16:25:32 -0800 Subject: [PATCH 181/647] Make lora training work on Z Image and remove some redundant nodes. (#10927) --- comfy/ldm/lumina/model.py | 4 +- comfy_extras/nodes_dataset.py | 102 +--------------------------------- 2 files changed, 3 insertions(+), 103 deletions(-) diff --git a/comfy/ldm/lumina/model.py b/comfy/ldm/lumina/model.py index c8643eb82..565400b54 100644 --- a/comfy/ldm/lumina/model.py +++ b/comfy/ldm/lumina/model.py @@ -509,7 +509,7 @@ class NextDiT(nn.Module): if self.pad_tokens_multiple is not None: pad_extra = (-cap_feats.shape[1]) % self.pad_tokens_multiple - cap_feats = torch.cat((cap_feats, self.cap_pad_token.to(device=cap_feats.device, dtype=cap_feats.dtype).unsqueeze(0).repeat(cap_feats.shape[0], pad_extra, 1)), dim=1) + cap_feats = torch.cat((cap_feats, self.cap_pad_token.to(device=cap_feats.device, dtype=cap_feats.dtype, copy=True).unsqueeze(0).repeat(cap_feats.shape[0], pad_extra, 1)), dim=1) cap_pos_ids = torch.zeros(bsz, cap_feats.shape[1], 3, dtype=torch.float32, device=device) cap_pos_ids[:, :, 0] = torch.arange(cap_feats.shape[1], dtype=torch.float32, device=device) + 1.0 @@ -525,7 +525,7 @@ class NextDiT(nn.Module): if self.pad_tokens_multiple is not None: pad_extra = (-x.shape[1]) % self.pad_tokens_multiple - x = torch.cat((x, self.x_pad_token.to(device=x.device, dtype=x.dtype).unsqueeze(0).repeat(x.shape[0], pad_extra, 1)), dim=1) + x = torch.cat((x, self.x_pad_token.to(device=x.device, dtype=x.dtype, copy=True).unsqueeze(0).repeat(x.shape[0], pad_extra, 1)), dim=1) x_pos_ids = torch.nn.functional.pad(x_pos_ids, (0, 0, 0, pad_extra)) freqs_cis = self.rope_embedder(torch.cat((cap_pos_ids, x_pos_ids), dim=1)).movedim(1, 2) diff --git a/comfy_extras/nodes_dataset.py b/comfy_extras/nodes_dataset.py index b23867505..4789d7d53 100644 --- a/comfy_extras/nodes_dataset.py +++ b/comfy_extras/nodes_dataset.py @@ -1,6 +1,5 @@ import logging import os -import math import json import numpy as np @@ -624,79 +623,6 @@ class TextProcessingNode(io.ComfyNode): # ========== Image Transform Nodes ========== -class ResizeImagesToSameSizeNode(ImageProcessingNode): - node_id = "ResizeImagesToSameSize" - display_name = "Resize Images to Same Size" - description = "Resize all images to the same width and height." - extra_inputs = [ - io.Int.Input("width", default=512, min=1, max=8192, tooltip="Target width."), - io.Int.Input("height", default=512, min=1, max=8192, tooltip="Target height."), - io.Combo.Input( - "mode", - options=["stretch", "crop_center", "pad"], - default="stretch", - tooltip="Resize mode.", - ), - ] - - @classmethod - def _process(cls, image, width, height, mode): - img = tensor_to_pil(image) - - if mode == "stretch": - img = img.resize((width, height), Image.Resampling.LANCZOS) - elif mode == "crop_center": - left = max(0, (img.width - width) // 2) - top = max(0, (img.height - height) // 2) - right = min(img.width, left + width) - bottom = min(img.height, top + height) - img = img.crop((left, top, right, bottom)) - if img.width != width or img.height != height: - img = img.resize((width, height), Image.Resampling.LANCZOS) - elif mode == "pad": - img.thumbnail((width, height), Image.Resampling.LANCZOS) - new_img = Image.new("RGB", (width, height), (0, 0, 0)) - paste_x = (width - img.width) // 2 - paste_y = (height - img.height) // 2 - new_img.paste(img, (paste_x, paste_y)) - img = new_img - - return pil_to_tensor(img) - - -class ResizeImagesToPixelCountNode(ImageProcessingNode): - node_id = "ResizeImagesToPixelCount" - display_name = "Resize Images to Pixel Count" - description = "Resize images so that the total pixel count matches the specified number while preserving aspect ratio." - extra_inputs = [ - io.Int.Input( - "pixel_count", - default=512 * 512, - min=1, - max=8192 * 8192, - tooltip="Target pixel count.", - ), - io.Int.Input( - "steps", - default=64, - min=1, - max=128, - tooltip="The stepping for resize width/height.", - ), - ] - - @classmethod - def _process(cls, image, pixel_count, steps): - img = tensor_to_pil(image) - w, h = img.size - pixel_count_ratio = math.sqrt(pixel_count / (w * h)) - new_w = int(w * pixel_count_ratio / steps) * steps - new_h = int(h * pixel_count_ratio / steps) * steps - logging.info(f"Resizing from {w}x{h} to {new_w}x{new_h}") - img = img.resize((new_w, new_h), Image.Resampling.LANCZOS) - return pil_to_tensor(img) - - class ResizeImagesByShorterEdgeNode(ImageProcessingNode): node_id = "ResizeImagesByShorterEdge" display_name = "Resize Images by Shorter Edge" @@ -801,29 +727,6 @@ class RandomCropImagesNode(ImageProcessingNode): return pil_to_tensor(img) -class FlipImagesNode(ImageProcessingNode): - node_id = "FlipImages" - display_name = "Flip Images" - description = "Flip all images horizontally or vertically." - extra_inputs = [ - io.Combo.Input( - "direction", - options=["horizontal", "vertical"], - default="horizontal", - tooltip="Flip direction.", - ), - ] - - @classmethod - def _process(cls, image, direction): - img = tensor_to_pil(image) - if direction == "horizontal": - img = img.transpose(Image.FLIP_LEFT_RIGHT) - else: - img = img.transpose(Image.FLIP_TOP_BOTTOM) - return pil_to_tensor(img) - - class NormalizeImagesNode(ImageProcessingNode): node_id = "NormalizeImages" display_name = "Normalize Images" @@ -1470,7 +1373,7 @@ class LoadTrainingDataset(io.ComfyNode): shard_path = os.path.join(dataset_dir, shard_file) with open(shard_path, "rb") as f: - shard_data = torch.load(f) + shard_data = torch.load(f, weights_only=True) all_latents.extend(shard_data["latents"]) all_conditioning.extend(shard_data["conditioning"]) @@ -1496,13 +1399,10 @@ class DatasetExtension(ComfyExtension): SaveImageDataSetToFolderNode, SaveImageTextDataSetToFolderNode, # Image transform nodes - ResizeImagesToSameSizeNode, - ResizeImagesToPixelCountNode, ResizeImagesByShorterEdgeNode, ResizeImagesByLongerEdgeNode, CenterCropImagesNode, RandomCropImagesNode, - FlipImagesNode, NormalizeImagesNode, AdjustBrightnessNode, AdjustContrastNode, From c38e7d6599be1bdce580ccfdbb20b928315af05e Mon Sep 17 00:00:00 2001 From: Haoming <73768377+Haoming02@users.noreply.github.com> Date: Thu, 27 Nov 2025 12:28:44 +0800 Subject: [PATCH 182/647] block info (#10841) --- comfy/ldm/flux/model.py | 6 ++++++ 1 file changed, 6 insertions(+) diff --git a/comfy/ldm/flux/model.py b/comfy/ldm/flux/model.py index 1a24e6d95..d5674dea6 100644 --- a/comfy/ldm/flux/model.py +++ b/comfy/ldm/flux/model.py @@ -171,7 +171,10 @@ class Flux(nn.Module): pe = None blocks_replace = patches_replace.get("dit", {}) + transformer_options["total_blocks"] = len(self.double_blocks) + transformer_options["block_type"] = "double" for i, block in enumerate(self.double_blocks): + transformer_options["block_index"] = i if ("double_block", i) in blocks_replace: def block_wrap(args): out = {} @@ -215,7 +218,10 @@ class Flux(nn.Module): if self.params.global_modulation: vec, _ = self.single_stream_modulation(vec_orig) + transformer_options["total_blocks"] = len(self.single_blocks) + transformer_options["block_type"] = "single" for i, block in enumerate(self.single_blocks): + transformer_options["block_index"] = i if ("single_block", i) in blocks_replace: def block_wrap(args): out = {} From f17251bec65b5760cfedec29eace7d77f4b35130 Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Thu, 27 Nov 2025 16:03:03 +1000 Subject: [PATCH 183/647] Account for the VRAM cost of weight offloading (#10733) * mm: default to 0 for NUM_STREAMS Dont count the compute stream as an offload stream. This makes async offload accounting easier. * mm: remove 128MB minimum This is from a previous offloading system requirement. Remove it to make behaviour of the loader and partial unloader consistent. * mp: order the module list by offload expense Calculate an approximate offloading temporary VRAM cost to offload a weight and primary order the module load list by that. In the simple case this is just the same as the module weight, but with Loras, a weight with a lora consumes considerably more VRAM to do the Lora application on-the-fly. This will slightly prioritize lora weights, but is really for proper VRAM offload accounting. * mp: Account for the VRAM cost of weight offloading when checking the VRAM headroom, assume that the weight needs to be offloaded, and only load if it has space for both the load and offload * the number of streams. As the weights are ordered from largest to smallest by offload cost this is guaranteed to fit in VRAM (tm), as all weights that follow will be smaller. Make the partial unload aware of this system as well by saving the budget for offload VRAM to the model state and accounting accordingly. Its possible that partial unload increases the size of the largest offloaded weights, and thus needs to unload a little bit more than asked to accomodate the bigger temp buffers. Honor the existing codes floor on model weight loading of 128MB by having the patcher honor this separately withough regard to offloading. Otherwise when MM specifies its 128MB minimum, MP will see the biggest weights, and budget that 128MB to only offload buffer and load nothing which isnt the intent of these minimums. The same clamp applies in case of partial offload of the currently loading model. --- comfy/model_management.py | 6 ++-- comfy/model_patcher.py | 59 +++++++++++++++++++++++++++++---------- 2 files changed, 48 insertions(+), 17 deletions(-) diff --git a/comfy/model_management.py b/comfy/model_management.py index a9327ac80..9c403d580 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -689,7 +689,7 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimu loaded_memory = loaded_model.model_loaded_memory() current_free_mem = get_free_memory(torch_dev) + loaded_memory - lowvram_model_memory = max(128 * 1024 * 1024, (current_free_mem - minimum_memory_required), min(current_free_mem * MIN_WEIGHT_MEMORY_RATIO, current_free_mem - minimum_inference_memory())) + lowvram_model_memory = max(0, (current_free_mem - minimum_memory_required), min(current_free_mem * MIN_WEIGHT_MEMORY_RATIO, current_free_mem - minimum_inference_memory())) lowvram_model_memory = lowvram_model_memory - loaded_memory if lowvram_model_memory == 0: @@ -1012,7 +1012,7 @@ def force_channels_last(): STREAMS = {} -NUM_STREAMS = 1 +NUM_STREAMS = 0 if args.async_offload: NUM_STREAMS = 2 logging.info("Using async weight offloading with {} streams".format(NUM_STREAMS)) @@ -1030,7 +1030,7 @@ def current_stream(device): stream_counters = {} def get_offload_stream(device): stream_counter = stream_counters.get(device, 0) - if NUM_STREAMS <= 1: + if NUM_STREAMS == 0: return None if device in STREAMS: diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index 73adc7f70..3eac77275 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -148,6 +148,15 @@ class LowVramPatch: else: return out +#The above patch logic may cast up the weight to fp32, and do math. Go with fp32 x 3 +LOWVRAM_PATCH_ESTIMATE_MATH_FACTOR = 3 + +def low_vram_patch_estimate_vram(model, key): + weight, set_func, convert_func = get_key_weight(model, key) + if weight is None: + return 0 + return weight.numel() * torch.float32.itemsize * LOWVRAM_PATCH_ESTIMATE_MATH_FACTOR + def get_key_weight(model, key): set_func = None convert_func = None @@ -269,6 +278,9 @@ class ModelPatcher: if not hasattr(self.model, 'current_weight_patches_uuid'): self.model.current_weight_patches_uuid = None + if not hasattr(self.model, 'model_offload_buffer_memory'): + self.model.model_offload_buffer_memory = 0 + def model_size(self): if self.size > 0: return self.size @@ -662,7 +674,16 @@ class ModelPatcher: skip = True # skip random weights in non leaf modules break if not skip and (hasattr(m, "comfy_cast_weights") or len(params) > 0): - loading.append((comfy.model_management.module_size(m), n, m, params)) + module_mem = comfy.model_management.module_size(m) + module_offload_mem = module_mem + if hasattr(m, "comfy_cast_weights"): + weight_key = "{}.weight".format(n) + bias_key = "{}.bias".format(n) + if weight_key in self.patches: + module_offload_mem += low_vram_patch_estimate_vram(self.model, weight_key) + if bias_key in self.patches: + module_offload_mem += low_vram_patch_estimate_vram(self.model, bias_key) + loading.append((module_offload_mem, module_mem, n, m, params)) return loading def load(self, device_to=None, lowvram_model_memory=0, force_patch_weights=False, full_load=False): @@ -676,20 +697,22 @@ class ModelPatcher: load_completely = [] offloaded = [] + offload_buffer = 0 loading.sort(reverse=True) for x in loading: - n = x[1] - m = x[2] - params = x[3] - module_mem = x[0] + module_offload_mem, module_mem, n, m, params = x lowvram_weight = False + potential_offload = max(offload_buffer, module_offload_mem * (comfy.model_management.NUM_STREAMS + 1)) + lowvram_fits = mem_counter + module_mem + potential_offload < lowvram_model_memory + weight_key = "{}.weight".format(n) bias_key = "{}.bias".format(n) if not full_load and hasattr(m, "comfy_cast_weights"): - if mem_counter + module_mem >= lowvram_model_memory: + if not lowvram_fits: + offload_buffer = potential_offload lowvram_weight = True lowvram_counter += 1 lowvram_mem_counter += module_mem @@ -723,9 +746,11 @@ class ModelPatcher: if hasattr(m, "comfy_cast_weights"): wipe_lowvram_weight(m) - if full_load or mem_counter + module_mem < lowvram_model_memory: + if full_load or lowvram_fits: mem_counter += module_mem load_completely.append((module_mem, n, m, params)) + else: + offload_buffer = potential_offload if cast_weight and hasattr(m, "comfy_cast_weights"): m.prev_comfy_cast_weights = m.comfy_cast_weights @@ -766,7 +791,7 @@ class ModelPatcher: self.pin_weight_to_device("{}.{}".format(n, param)) if lowvram_counter > 0: - logging.info("loaded partially; {:.2f} MB usable, {:.2f} MB loaded, {:.2f} MB offloaded, lowvram patches: {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (1024 * 1024), lowvram_mem_counter / (1024 * 1024), patch_counter)) + logging.info("loaded partially; {:.2f} MB usable, {:.2f} MB loaded, {:.2f} MB offloaded, {:.2f} MB buffer reserved, lowvram patches: {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (1024 * 1024), lowvram_mem_counter / (1024 * 1024), offload_buffer / (1024 * 1024), patch_counter)) self.model.model_lowvram = True else: logging.info("loaded completely; {:.2f} MB usable, {:.2f} MB loaded, full load: {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (1024 * 1024), full_load)) @@ -778,6 +803,7 @@ class ModelPatcher: self.model.lowvram_patch_counter += patch_counter self.model.device = device_to self.model.model_loaded_weight_memory = mem_counter + self.model.model_offload_buffer_memory = offload_buffer self.model.current_weight_patches_uuid = self.patches_uuid for callback in self.get_all_callbacks(CallbacksMP.ON_LOAD): @@ -831,6 +857,7 @@ class ModelPatcher: self.model.to(device_to) self.model.device = device_to self.model.model_loaded_weight_memory = 0 + self.model.model_offload_buffer_memory = 0 for m in self.model.modules(): if hasattr(m, "comfy_patched_weights"): @@ -849,13 +876,14 @@ class ModelPatcher: patch_counter = 0 unload_list = self._load_list() unload_list.sort() + offload_buffer = self.model.model_offload_buffer_memory + for unload in unload_list: - if memory_to_free < memory_freed: + if memory_to_free + offload_buffer - self.model.model_offload_buffer_memory < memory_freed: break - module_mem = unload[0] - n = unload[1] - m = unload[2] - params = unload[3] + module_offload_mem, module_mem, n, m, params = unload + + potential_offload = (comfy.model_management.NUM_STREAMS + 1) * module_offload_mem lowvram_possible = hasattr(m, "comfy_cast_weights") if hasattr(m, "comfy_patched_weights") and m.comfy_patched_weights == True: @@ -906,15 +934,18 @@ class ModelPatcher: m.comfy_cast_weights = True m.comfy_patched_weights = False memory_freed += module_mem + offload_buffer = max(offload_buffer, potential_offload) logging.debug("freed {}".format(n)) for param in params: self.pin_weight_to_device("{}.{}".format(n, param)) + self.model.model_lowvram = True self.model.lowvram_patch_counter += patch_counter self.model.model_loaded_weight_memory -= memory_freed - logging.info("loaded partially: {:.2f} MB loaded, lowvram patches: {}".format(self.model.model_loaded_weight_memory / (1024 * 1024), self.model.lowvram_patch_counter)) + self.model.model_offload_buffer_memory = offload_buffer + logging.info("Unloaded partially: {:.2f} MB freed, {:.2f} MB remains loaded, {:.2f} MB buffer reserved, lowvram patches: {}".format(memory_freed / (1024 * 1024), self.model.model_loaded_weight_memory / (1024 * 1024), offload_buffer / (1024 * 1024), self.model.lowvram_patch_counter)) return memory_freed def partially_load(self, device_to, extra_memory=0, force_patch_weights=False): From 3f382a4f9884f7b672557028adb9bb85d075820d Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Fri, 28 Nov 2025 02:06:30 +1000 Subject: [PATCH 184/647] quant ops: Dequantize weight in-place (#10935) In flux2 these weights are huge (200MB). As plain_tensor is a throw-away deep copy, do this multiplication in-place to save VRAM. --- comfy/quant_ops.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/comfy/quant_ops.py b/comfy/quant_ops.py index d2f3e7397..9b924560b 100644 --- a/comfy/quant_ops.py +++ b/comfy/quant_ops.py @@ -425,7 +425,8 @@ class TensorCoreFP8Layout(QuantizedLayout): @staticmethod def dequantize(qdata, scale, orig_dtype, **kwargs): plain_tensor = torch.ops.aten._to_copy.default(qdata, dtype=orig_dtype) - return plain_tensor * scale + plain_tensor.mul_(scale) + return plain_tensor @classmethod def get_plain_tensors(cls, qtensor): From b59750a86a4687056528f1d59470e207063a73a3 Mon Sep 17 00:00:00 2001 From: ComfyUI Wiki Date: Fri, 28 Nov 2025 06:12:56 +0800 Subject: [PATCH 185/647] Update template to 0.7.23 (#10949) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 9291552d3..e0b2c566b 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,5 +1,5 @@ comfyui-frontend-package==1.32.9 -comfyui-workflow-templates==0.7.20 +comfyui-workflow-templates==0.7.23 comfyui-embedded-docs==0.3.1 torch torchsde From 9d8a817985bb069685e440b38762f95dc834d242 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Thu, 27 Nov 2025 14:46:12 -0800 Subject: [PATCH 186/647] Enable async offloading by default on Nvidia. (#10953) Add --disable-async-offload to disable it. If this causes OOMs that go away when you --disable-async-offload please report it. --- comfy/cli_args.py | 3 ++- comfy/model_management.py | 13 +++++++++++-- 2 files changed, 13 insertions(+), 3 deletions(-) diff --git a/comfy/cli_args.py b/comfy/cli_args.py index d2b60e347..5f0dfaa10 100644 --- a/comfy/cli_args.py +++ b/comfy/cli_args.py @@ -131,7 +131,8 @@ vram_group.add_argument("--cpu", action="store_true", help="To use the CPU for e parser.add_argument("--reserve-vram", type=float, default=None, help="Set the amount of vram in GB you want to reserve for use by your OS/other software. By default some amount is reserved depending on your OS.") -parser.add_argument("--async-offload", action="store_true", help="Use async weight offloading.") +parser.add_argument("--async-offload", nargs='?', const=2, type=int, default=None, metavar="NUM_STREAMS", help="Use async weight offloading. An optional argument controls the amount of offload streams. Default is 2. Enabled by default on Nvidia.") +parser.add_argument("--disable-async-offload", action="store_true", help="Disable async weight offloading.") parser.add_argument("--force-non-blocking", action="store_true", help="Force ComfyUI to use non-blocking operations for all applicable tensors. This may improve performance on some non-Nvidia systems but can cause issues with some workflows.") diff --git a/comfy/model_management.py b/comfy/model_management.py index 9c403d580..38c506df5 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -1013,8 +1013,17 @@ def force_channels_last(): STREAMS = {} NUM_STREAMS = 0 -if args.async_offload: - NUM_STREAMS = 2 +if args.async_offload is not None: + NUM_STREAMS = args.async_offload +else: + # Enable by default on Nvidia + if is_nvidia(): + NUM_STREAMS = 2 + +if args.disable_async_offload: + NUM_STREAMS = 0 + +if NUM_STREAMS > 0: logging.info("Using async weight offloading with {} streams".format(NUM_STREAMS)) def current_stream(device): From 52e778fff3c1d6f32c8d14cba9864faddba8475d Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Fri, 28 Nov 2025 12:52:59 +0200 Subject: [PATCH 187/647] feat(Kling-API-Nodes): add v2-5-turbo model to FirstLastFrame node (#10938) --- comfy_api_nodes/nodes_kling.py | 60 +++++++++++++++------------------- 1 file changed, 26 insertions(+), 34 deletions(-) diff --git a/comfy_api_nodes/nodes_kling.py b/comfy_api_nodes/nodes_kling.py index 36852038b..23a7f55f1 100644 --- a/comfy_api_nodes/nodes_kling.py +++ b/comfy_api_nodes/nodes_kling.py @@ -4,8 +4,6 @@ For source of truth on the allowed permutations of request fields, please refere - [Compatibility Table](https://app.klingai.com/global/dev/document-api/apiReference/model/skillsMap) """ -from __future__ import annotations -from typing import Optional, TypeVar import math import logging @@ -66,9 +64,7 @@ from comfy_api_nodes.util import ( poll_op, ) from comfy_api.input_impl import VideoFromFile -from comfy_api.input.basic_types import AudioInput -from comfy_api.input.video_types import VideoInput -from comfy_api.latest import ComfyExtension, IO +from comfy_api.latest import ComfyExtension, IO, Input KLING_API_VERSION = "v1" PATH_TEXT_TO_VIDEO = f"/proxy/kling/{KLING_API_VERSION}/videos/text2video" @@ -94,8 +90,6 @@ AVERAGE_DURATION_IMAGE_GEN = 32 AVERAGE_DURATION_VIDEO_EFFECTS = 320 AVERAGE_DURATION_VIDEO_EXTEND = 320 -R = TypeVar("R") - MODE_TEXT2VIDEO = { "standard mode / 5s duration / kling-v1": ("std", "5", "kling-v1"), @@ -130,6 +124,8 @@ MODE_START_END_FRAME = { "pro mode / 10s duration / kling-v1-6": ("pro", "10", "kling-v1-6"), "pro mode / 5s duration / kling-v2-1": ("pro", "5", "kling-v2-1"), "pro mode / 10s duration / kling-v2-1": ("pro", "10", "kling-v2-1"), + "pro mode / 5s duration / kling-v2-5-turbo": ("pro", "5", "kling-v2-5-turbo"), + "pro mode / 10s duration / kling-v2-5-turbo": ("pro", "10", "kling-v2-5-turbo"), } """ Returns a mapping of mode strings to their corresponding (mode, duration, model_name) tuples. @@ -296,7 +292,7 @@ def get_video_from_response(response) -> KlingVideoResult: return video -def get_video_url_from_response(response) -> Optional[str]: +def get_video_url_from_response(response) -> str | None: """Returns the first video url from the Kling video generation task result. Will not raise an error if the response is not valid. """ @@ -315,7 +311,7 @@ def get_images_from_response(response) -> list[KlingImageResult]: return images -def get_images_urls_from_response(response) -> Optional[str]: +def get_images_urls_from_response(response) -> str | None: """Returns the list of image urls from the Kling image generation task result. Will not raise an error if the response is not valid. If there is only one image, returns the url as a string. If there are multiple images, returns a list of urls. """ @@ -349,7 +345,7 @@ async def execute_text2video( model_mode: str, duration: str, aspect_ratio: str, - camera_control: Optional[KlingCameraControl] = None, + camera_control: KlingCameraControl | None = None, ) -> IO.NodeOutput: validate_prompts(prompt, negative_prompt, MAX_PROMPT_LENGTH_T2V) task_creation_response = await sync_op( @@ -394,8 +390,8 @@ async def execute_image2video( model_mode: str, aspect_ratio: str, duration: str, - camera_control: Optional[KlingCameraControl] = None, - end_frame: Optional[torch.Tensor] = None, + camera_control: KlingCameraControl | None = None, + end_frame: torch.Tensor | None = None, ) -> IO.NodeOutput: validate_prompts(prompt, negative_prompt, MAX_PROMPT_LENGTH_I2V) validate_input_image(start_frame) @@ -451,8 +447,8 @@ async def execute_video_effect( model_name: str, duration: KlingVideoGenDuration, image_1: torch.Tensor, - image_2: Optional[torch.Tensor] = None, - model_mode: Optional[KlingVideoGenMode] = None, + image_2: torch.Tensor | None = None, + model_mode: KlingVideoGenMode | None = None, ) -> tuple[VideoFromFile, str, str]: if dual_character: request_input_field = KlingDualCharacterEffectInput( @@ -499,13 +495,13 @@ async def execute_video_effect( async def execute_lipsync( cls: type[IO.ComfyNode], - video: VideoInput, - audio: Optional[AudioInput] = None, - voice_language: Optional[str] = None, - model_mode: Optional[str] = None, - text: Optional[str] = None, - voice_speed: Optional[float] = None, - voice_id: Optional[str] = None, + video: Input.Video, + audio: Input.Audio | None = None, + voice_language: str | None = None, + model_mode: str | None = None, + text: str | None = None, + voice_speed: float | None = None, + voice_id: str | None = None, ) -> IO.NodeOutput: if text: validate_string(text, field_name="Text", max_length=MAX_PROMPT_LENGTH_LIP_SYNC) @@ -787,7 +783,7 @@ class KlingCameraControlT2VNode(IO.ComfyNode): negative_prompt: str, cfg_scale: float, aspect_ratio: str, - camera_control: Optional[KlingCameraControl] = None, + camera_control: KlingCameraControl | None = None, ) -> IO.NodeOutput: return await execute_text2video( cls, @@ -854,8 +850,8 @@ class KlingImage2VideoNode(IO.ComfyNode): mode: str, aspect_ratio: str, duration: str, - camera_control: Optional[KlingCameraControl] = None, - end_frame: Optional[torch.Tensor] = None, + camera_control: KlingCameraControl | None = None, + end_frame: torch.Tensor | None = None, ) -> IO.NodeOutput: return await execute_image2video( cls, @@ -965,15 +961,11 @@ class KlingStartEndFrameNode(IO.ComfyNode): IO.String.Input("prompt", multiline=True, tooltip="Positive text prompt"), IO.String.Input("negative_prompt", multiline=True, tooltip="Negative text prompt"), IO.Float.Input("cfg_scale", default=0.5, min=0.0, max=1.0), - IO.Combo.Input( - "aspect_ratio", - options=[i.value for i in KlingVideoGenAspectRatio], - default="16:9", - ), + IO.Combo.Input("aspect_ratio", options=["16:9", "9:16", "1:1"]), IO.Combo.Input( "mode", options=modes, - default=modes[2], + default=modes[8], tooltip="The configuration to use for the video generation following the format: mode / duration / model_name.", ), ], @@ -1254,8 +1246,8 @@ class KlingLipSyncAudioToVideoNode(IO.ComfyNode): @classmethod async def execute( cls, - video: VideoInput, - audio: AudioInput, + video: Input.Video, + audio: Input.Audio, voice_language: str, ) -> IO.NodeOutput: return await execute_lipsync( @@ -1314,7 +1306,7 @@ class KlingLipSyncTextToVideoNode(IO.ComfyNode): @classmethod async def execute( cls, - video: VideoInput, + video: Input.Video, text: str, voice: str, voice_speed: float, @@ -1471,7 +1463,7 @@ class KlingImageGenerationNode(IO.ComfyNode): human_fidelity: float, n: int, aspect_ratio: KlingImageGenAspectRatio, - image: Optional[torch.Tensor] = None, + image: torch.Tensor | None = None, ) -> IO.NodeOutput: validate_string(prompt, field_name="prompt", min_length=1, max_length=MAX_PROMPT_LENGTH_IMAGE_GEN) validate_string(negative_prompt, field_name="negative_prompt", max_length=MAX_PROMPT_LENGTH_IMAGE_GEN) From ca7808f240d4d53e594d3b95753240313864c992 Mon Sep 17 00:00:00 2001 From: "Dr.Lt.Data" <128333288+ltdrdata@users.noreply.github.com> Date: Sat, 29 Nov 2025 05:43:17 +0900 Subject: [PATCH 188/647] fix(user_manager): fix typo in move_userdata dest validation (#10967) Check `dest` instead of `source` when validating destination path in move_userdata endpoint. --- app/user_manager.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/app/user_manager.py b/app/user_manager.py index a2d376c0c..675f6c0c6 100644 --- a/app/user_manager.py +++ b/app/user_manager.py @@ -424,7 +424,7 @@ class UserManager(): return source dest = get_user_data_path(request, check_exists=False, param="dest") - if not isinstance(source, str): + if not isinstance(dest, str): return dest overwrite = request.query.get("overwrite", 'true') != "false" From f55c98a89f76fc06c435a728bc2e76b6b4051463 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Fri, 28 Nov 2025 13:16:46 -0800 Subject: [PATCH 189/647] Disable offload stream when torch compile. (#10961) --- comfy/model_management.py | 3 +++ 1 file changed, 3 insertions(+) diff --git a/comfy/model_management.py b/comfy/model_management.py index 38c506df5..d8ce80010 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -1042,6 +1042,9 @@ def get_offload_stream(device): if NUM_STREAMS == 0: return None + if torch.compiler.is_compiling(): + return None + if device in STREAMS: ss = STREAMS[device] #Sync the oldest stream in the queue with the current From 6484ac89dc683b178d9ef3f7406800f7132147ba Mon Sep 17 00:00:00 2001 From: Urle Sistiana <55231606+urlesistiana@users.noreply.github.com> Date: Sat, 29 Nov 2025 05:33:07 +0800 Subject: [PATCH 190/647] fix QuantizedTensor.is_contiguous (#10956) (#10959) --- comfy/quant_ops.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/comfy/quant_ops.py b/comfy/quant_ops.py index 9b924560b..bb1fb860c 100644 --- a/comfy/quant_ops.py +++ b/comfy/quant_ops.py @@ -235,8 +235,8 @@ class QuantizedTensor(torch.Tensor): def is_pinned(self): return self._qdata.is_pinned() - def is_contiguous(self): - return self._qdata.is_contiguous() + def is_contiguous(self, *arg, **kwargs): + return self._qdata.is_contiguous(*arg, **kwargs) # ============================================================================== # Generic Utilities (Layout-Agnostic Operations) From 0ff0457892467ef8a6ea235dcd0618c10ca44ee3 Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Sat, 29 Nov 2025 07:38:12 +1000 Subject: [PATCH 191/647] mm: wrap the raw stream in context manager (#10958) The documentation of torch.foo.Stream being usable with with: suggests it starts at version 2.7. Use the old API for backwards compatibility. --- comfy/model_management.py | 19 +++++++++++++++---- comfy/ops.py | 2 ++ 2 files changed, 17 insertions(+), 4 deletions(-) diff --git a/comfy/model_management.py b/comfy/model_management.py index d8ce80010..aeddbaefe 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -1055,7 +1055,9 @@ def get_offload_stream(device): elif is_device_cuda(device): ss = [] for k in range(NUM_STREAMS): - ss.append(torch.cuda.Stream(device=device, priority=0)) + s1 = torch.cuda.Stream(device=device, priority=0) + s1.as_context = torch.cuda.stream + ss.append(s1) STREAMS[device] = ss s = ss[stream_counter] stream_counters[device] = stream_counter @@ -1063,7 +1065,9 @@ def get_offload_stream(device): elif is_device_xpu(device): ss = [] for k in range(NUM_STREAMS): - ss.append(torch.xpu.Stream(device=device, priority=0)) + s1 = torch.xpu.Stream(device=device, priority=0) + s1.as_context = torch.xpu.stream + ss.append(s1) STREAMS[device] = ss s = ss[stream_counter] stream_counters[device] = stream_counter @@ -1081,12 +1085,19 @@ def cast_to(weight, dtype=None, device=None, non_blocking=False, copy=False, str if dtype is None or weight.dtype == dtype: return weight if stream is not None: - with stream: + wf_context = stream + if hasattr(wf_context, "as_context"): + wf_context = wf_context.as_context(stream) + with wf_context: return weight.to(dtype=dtype, copy=copy) return weight.to(dtype=dtype, copy=copy) + if stream is not None: - with stream: + wf_context = stream + if hasattr(wf_context, "as_context"): + wf_context = wf_context.as_context(stream) + with wf_context: r = torch.empty_like(weight, dtype=dtype, device=device) r.copy_(weight, non_blocking=non_blocking) else: diff --git a/comfy/ops.py b/comfy/ops.py index a0ff4e8f1..61a2f0754 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -95,6 +95,8 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, of if offload_stream is not None: wf_context = offload_stream + if hasattr(wf_context, "as_context"): + wf_context = wf_context.as_context(offload_stream) else: wf_context = contextlib.nullcontext() From 065a2fbbec6af5c8e19a3add29703f83faf672d6 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Fri, 28 Nov 2025 16:37:39 -0800 Subject: [PATCH 192/647] Update driver link in AMD portable README (#10974) --- .ci/windows_amd_base_files/README_VERY_IMPORTANT.txt | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-) diff --git a/.ci/windows_amd_base_files/README_VERY_IMPORTANT.txt b/.ci/windows_amd_base_files/README_VERY_IMPORTANT.txt index 96a500be2..2cbb00d99 100755 --- a/.ci/windows_amd_base_files/README_VERY_IMPORTANT.txt +++ b/.ci/windows_amd_base_files/README_VERY_IMPORTANT.txt @@ -1,5 +1,5 @@ -As of the time of writing this you need this preview driver for best results: -https://www.amd.com/en/resources/support-articles/release-notes/RN-AMDGPU-WINDOWS-PYTORCH-PREVIEW.html +As of the time of writing this you need this driver for best results: +https://www.amd.com/en/resources/support-articles/release-notes/RN-AMDGPU-WINDOWS-PYTORCH-7-1-1.html HOW TO RUN: @@ -25,3 +25,4 @@ In the ComfyUI directory you will find a file: extra_model_paths.yaml.example Rename this file to: extra_model_paths.yaml and edit it with your favorite text editor. + From b9070857092a78cc952d70025fdcc0ff540d96ec Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Jukka=20Sepp=C3=A4nen?= <40791699+kijai@users.noreply.github.com> Date: Sat, 29 Nov 2025 02:40:19 +0200 Subject: [PATCH 193/647] Support video tiny VAEs (#10884) * Support video tiny VAEs * lighttaew scaling fix * Also support video taes in previews Only first frame for now as live preview playback is currently only available through VHS custom nodes. * Support Wan 2.1 lightVAE * Relocate elif block and set Wan VAE dim directly without using pruning rate for lightvae --- comfy/latent_formats.py | 5 +- comfy/sd.py | 34 +++++++- comfy/taesd/taehv.py | 171 ++++++++++++++++++++++++++++++++++++++++ latent_preview.py | 28 +++++-- nodes.py | 18 ++++- 5 files changed, 244 insertions(+), 12 deletions(-) create mode 100644 comfy/taesd/taehv.py diff --git a/comfy/latent_formats.py b/comfy/latent_formats.py index 8e110f45d..f1ca0151e 100644 --- a/comfy/latent_formats.py +++ b/comfy/latent_formats.py @@ -431,6 +431,7 @@ class HunyuanVideo(LatentFormat): ] latent_rgb_factors_bias = [ 0.0259, -0.0192, -0.0761] + taesd_decoder_name = "taehv" class Cosmos1CV8x8x8(LatentFormat): latent_channels = 16 @@ -494,7 +495,7 @@ class Wan21(LatentFormat): ]).view(1, self.latent_channels, 1, 1, 1) - self.taesd_decoder_name = None #TODO + self.taesd_decoder_name = "lighttaew2_1" def process_in(self, latent): latents_mean = self.latents_mean.to(latent.device, latent.dtype) @@ -565,6 +566,7 @@ class Wan22(Wan21): def __init__(self): self.scale_factor = 1.0 + self.taesd_decoder_name = "lighttaew2_2" self.latents_mean = torch.tensor([ -0.2289, -0.0052, -0.1323, -0.2339, -0.2799, 0.0174, 0.1838, 0.1557, -0.1382, 0.0542, 0.2813, 0.0891, 0.1570, -0.0098, 0.0375, -0.1825, @@ -719,6 +721,7 @@ class HunyuanVideo15(LatentFormat): latent_channels = 32 latent_dimensions = 3 scale_factor = 1.03682 + taesd_decoder_name = "lighttaehy1_5" class Hunyuan3Dv2(LatentFormat): latent_channels = 64 diff --git a/comfy/sd.py b/comfy/sd.py index 350fae92b..9eeb0c45a 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -60,6 +60,8 @@ import comfy.lora_convert import comfy.hooks import comfy.t2i_adapter.adapter import comfy.taesd.taesd +import comfy.taesd.taehv +import comfy.latent_formats import comfy.ldm.flux.redux @@ -508,13 +510,14 @@ class VAE: self.memory_used_encode = lambda shape, dtype: 3300 * shape[3] * shape[4] * model_management.dtype_size(dtype) self.memory_used_decode = lambda shape, dtype: 8000 * shape[3] * shape[4] * (16 * 16) * model_management.dtype_size(dtype) else: # Wan 2.1 VAE + dim = sd["decoder.head.0.gamma"].shape[0] self.upscale_ratio = (lambda a: max(0, a * 4 - 3), 8, 8) self.upscale_index_formula = (4, 8, 8) self.downscale_ratio = (lambda a: max(0, math.floor((a + 3) / 4)), 8, 8) self.downscale_index_formula = (4, 8, 8) self.latent_dim = 3 self.latent_channels = 16 - ddconfig = {"dim": 96, "z_dim": self.latent_channels, "dim_mult": [1, 2, 4, 4], "num_res_blocks": 2, "attn_scales": [], "temperal_downsample": [False, True, True], "dropout": 0.0} + ddconfig = {"dim": dim, "z_dim": self.latent_channels, "dim_mult": [1, 2, 4, 4], "num_res_blocks": 2, "attn_scales": [], "temperal_downsample": [False, True, True], "dropout": 0.0} self.first_stage_model = comfy.ldm.wan.vae.WanVAE(**ddconfig) self.working_dtypes = [torch.bfloat16, torch.float16, torch.float32] self.memory_used_encode = lambda shape, dtype: 6000 * shape[3] * shape[4] * model_management.dtype_size(dtype) @@ -584,6 +587,35 @@ class VAE: self.process_input = lambda audio: audio self.working_dtypes = [torch.float32] self.crop_input = False + elif "decoder.22.bias" in sd: # taehv, taew and lighttae + self.latent_channels = sd["decoder.1.weight"].shape[1] + self.latent_dim = 3 + self.upscale_ratio = (lambda a: max(0, a * 4 - 3), 16, 16) + self.upscale_index_formula = (4, 16, 16) + self.downscale_ratio = (lambda a: max(0, math.floor((a + 3) / 4)), 16, 16) + self.downscale_index_formula = (4, 16, 16) + if self.latent_channels == 48: # Wan 2.2 + self.first_stage_model = comfy.taesd.taehv.TAEHV(latent_channels=self.latent_channels, latent_format=None) # taehv doesn't need scaling + self.process_input = lambda image: (_ for _ in ()).throw(NotImplementedError("This light tae doesn't support encoding currently")) + self.process_output = lambda image: image + self.memory_used_decode = lambda shape, dtype: (1800 * (max(1, (shape[-3] ** 0.7 * 0.1)) * shape[-2] * shape[-1] * 16 * 16) * model_management.dtype_size(dtype)) + elif self.latent_channels == 32 and sd["decoder.22.bias"].shape[0] == 12: # lighttae_hv15 + self.first_stage_model = comfy.taesd.taehv.TAEHV(latent_channels=self.latent_channels, latent_format=comfy.latent_formats.HunyuanVideo15) + self.process_input = lambda image: (_ for _ in ()).throw(NotImplementedError("This light tae doesn't support encoding currently")) + self.memory_used_decode = lambda shape, dtype: (1200 * (max(1, (shape[-3] ** 0.7 * 0.05)) * shape[-2] * shape[-1] * 32 * 32) * model_management.dtype_size(dtype)) + else: + if sd["decoder.1.weight"].dtype == torch.float16: # taehv currently only available in float16, so assume it's not lighttaew2_1 as otherwise state dicts are identical + latent_format=comfy.latent_formats.HunyuanVideo + else: + latent_format=None # lighttaew2_1 doesn't need scaling + self.first_stage_model = comfy.taesd.taehv.TAEHV(latent_channels=self.latent_channels, latent_format=latent_format) + self.process_input = self.process_output = lambda image: image + self.upscale_ratio = (lambda a: max(0, a * 4 - 3), 8, 8) + self.upscale_index_formula = (4, 8, 8) + self.downscale_ratio = (lambda a: max(0, math.floor((a + 3) / 4)), 8, 8) + self.downscale_index_formula = (4, 8, 8) + self.memory_used_encode = lambda shape, dtype: (700 * (max(1, (shape[-3] ** 0.66 * 0.11)) * shape[-2] * shape[-1]) * model_management.dtype_size(dtype)) + self.memory_used_decode = lambda shape, dtype: (50 * (max(1, (shape[-3] ** 0.65 * 0.26)) * shape[-2] * shape[-1] * 32 * 32) * model_management.dtype_size(dtype)) else: logging.warning("WARNING: No VAE weights detected, VAE not initalized.") self.first_stage_model = None diff --git a/comfy/taesd/taehv.py b/comfy/taesd/taehv.py new file mode 100644 index 000000000..3dfe1e4d4 --- /dev/null +++ b/comfy/taesd/taehv.py @@ -0,0 +1,171 @@ +# Tiny AutoEncoder for HunyuanVideo and WanVideo https://github.com/madebyollin/taehv + +import torch +import torch.nn as nn +import torch.nn.functional as F +from tqdm.auto import tqdm +from collections import namedtuple, deque + +import comfy.ops +operations=comfy.ops.disable_weight_init + +DecoderResult = namedtuple("DecoderResult", ("frame", "memory")) +TWorkItem = namedtuple("TWorkItem", ("input_tensor", "block_index")) + +def conv(n_in, n_out, **kwargs): + return operations.Conv2d(n_in, n_out, 3, padding=1, **kwargs) + +class Clamp(nn.Module): + def forward(self, x): + return torch.tanh(x / 3) * 3 + +class MemBlock(nn.Module): + def __init__(self, n_in, n_out, act_func): + super().__init__() + self.conv = nn.Sequential(conv(n_in * 2, n_out), act_func, conv(n_out, n_out), act_func, conv(n_out, n_out)) + self.skip = operations.Conv2d(n_in, n_out, 1, bias=False) if n_in != n_out else nn.Identity() + self.act = act_func + def forward(self, x, past): + return self.act(self.conv(torch.cat([x, past], 1)) + self.skip(x)) + +class TPool(nn.Module): + def __init__(self, n_f, stride): + super().__init__() + self.stride = stride + self.conv = operations.Conv2d(n_f*stride,n_f, 1, bias=False) + def forward(self, x): + _NT, C, H, W = x.shape + return self.conv(x.reshape(-1, self.stride * C, H, W)) + +class TGrow(nn.Module): + def __init__(self, n_f, stride): + super().__init__() + self.stride = stride + self.conv = operations.Conv2d(n_f, n_f*stride, 1, bias=False) + def forward(self, x): + _NT, C, H, W = x.shape + x = self.conv(x) + return x.reshape(-1, C, H, W) + +def apply_model_with_memblocks(model, x, parallel, show_progress_bar): + + B, T, C, H, W = x.shape + if parallel: + x = x.reshape(B*T, C, H, W) + # parallel over input timesteps, iterate over blocks + for b in tqdm(model, disable=not show_progress_bar): + if isinstance(b, MemBlock): + BT, C, H, W = x.shape + T = BT // B + _x = x.reshape(B, T, C, H, W) + mem = F.pad(_x, (0,0,0,0,0,0,1,0), value=0)[:,:T].reshape(x.shape) + x = b(x, mem) + else: + x = b(x) + BT, C, H, W = x.shape + T = BT // B + x = x.view(B, T, C, H, W) + else: + out = [] + work_queue = deque([TWorkItem(xt, 0) for t, xt in enumerate(x.reshape(B, T * C, H, W).chunk(T, dim=1))]) + progress_bar = tqdm(range(T), disable=not show_progress_bar) + mem = [None] * len(model) + while work_queue: + xt, i = work_queue.popleft() + if i == 0: + progress_bar.update(1) + if i == len(model): + out.append(xt) + del xt + else: + b = model[i] + if isinstance(b, MemBlock): + if mem[i] is None: + xt_new = b(xt, xt * 0) + mem[i] = xt.detach().clone() + else: + xt_new = b(xt, mem[i]) + mem[i] = xt.detach().clone() + del xt + work_queue.appendleft(TWorkItem(xt_new, i+1)) + elif isinstance(b, TPool): + if mem[i] is None: + mem[i] = [] + mem[i].append(xt.detach().clone()) + if len(mem[i]) == b.stride: + B, C, H, W = xt.shape + xt = b(torch.cat(mem[i], 1).view(B*b.stride, C, H, W)) + mem[i] = [] + work_queue.appendleft(TWorkItem(xt, i+1)) + elif isinstance(b, TGrow): + xt = b(xt) + NT, C, H, W = xt.shape + for xt_next in reversed(xt.view(B, b.stride*C, H, W).chunk(b.stride, 1)): + work_queue.appendleft(TWorkItem(xt_next, i+1)) + del xt + else: + xt = b(xt) + work_queue.appendleft(TWorkItem(xt, i+1)) + progress_bar.close() + x = torch.stack(out, 1) + return x + + +class TAEHV(nn.Module): + def __init__(self, latent_channels, parallel=False, decoder_time_upscale=(True, True), decoder_space_upscale=(True, True, True), latent_format=None, show_progress_bar=True): + super().__init__() + self.image_channels = 3 + self.patch_size = 1 + self.latent_channels = latent_channels + self.parallel = parallel + self.latent_format = latent_format + self.show_progress_bar = show_progress_bar + self.process_in = latent_format().process_in if latent_format is not None else (lambda x: x) + self.process_out = latent_format().process_out if latent_format is not None else (lambda x: x) + if self.latent_channels in [48, 32]: # Wan 2.2 and HunyuanVideo1.5 + self.patch_size = 2 + if self.latent_channels == 32: # HunyuanVideo1.5 + act_func = nn.LeakyReLU(0.2, inplace=True) + else: # HunyuanVideo, Wan 2.1 + act_func = nn.ReLU(inplace=True) + + self.encoder = nn.Sequential( + conv(self.image_channels*self.patch_size**2, 64), act_func, + TPool(64, 2), conv(64, 64, stride=2, bias=False), MemBlock(64, 64, act_func), MemBlock(64, 64, act_func), MemBlock(64, 64, act_func), + TPool(64, 2), conv(64, 64, stride=2, bias=False), MemBlock(64, 64, act_func), MemBlock(64, 64, act_func), MemBlock(64, 64, act_func), + TPool(64, 1), conv(64, 64, stride=2, bias=False), MemBlock(64, 64, act_func), MemBlock(64, 64, act_func), MemBlock(64, 64, act_func), + conv(64, self.latent_channels), + ) + n_f = [256, 128, 64, 64] + self.frames_to_trim = 2**sum(decoder_time_upscale) - 1 + self.decoder = nn.Sequential( + Clamp(), conv(self.latent_channels, n_f[0]), act_func, + MemBlock(n_f[0], n_f[0], act_func), MemBlock(n_f[0], n_f[0], act_func), MemBlock(n_f[0], n_f[0], act_func), nn.Upsample(scale_factor=2 if decoder_space_upscale[0] else 1), TGrow(n_f[0], 1), conv(n_f[0], n_f[1], bias=False), + MemBlock(n_f[1], n_f[1], act_func), MemBlock(n_f[1], n_f[1], act_func), MemBlock(n_f[1], n_f[1], act_func), nn.Upsample(scale_factor=2 if decoder_space_upscale[1] else 1), TGrow(n_f[1], 2 if decoder_time_upscale[0] else 1), conv(n_f[1], n_f[2], bias=False), + MemBlock(n_f[2], n_f[2], act_func), MemBlock(n_f[2], n_f[2], act_func), MemBlock(n_f[2], n_f[2], act_func), nn.Upsample(scale_factor=2 if decoder_space_upscale[2] else 1), TGrow(n_f[2], 2 if decoder_time_upscale[1] else 1), conv(n_f[2], n_f[3], bias=False), + act_func, conv(n_f[3], self.image_channels*self.patch_size**2), + ) + @property + def show_progress_bar(self): + return self._show_progress_bar + + @show_progress_bar.setter + def show_progress_bar(self, value): + self._show_progress_bar = value + + def encode(self, x, **kwargs): + if self.patch_size > 1: x = F.pixel_unshuffle(x, self.patch_size) + x = x.movedim(2, 1) # [B, C, T, H, W] -> [B, T, C, H, W] + if x.shape[1] % 4 != 0: + # pad at end to multiple of 4 + n_pad = 4 - x.shape[1] % 4 + padding = x[:, -1:].repeat_interleave(n_pad, dim=1) + x = torch.cat([x, padding], 1) + x = apply_model_with_memblocks(self.encoder, x, self.parallel, self.show_progress_bar).movedim(2, 1) + return self.process_out(x) + + def decode(self, x, **kwargs): + x = self.process_in(x).movedim(2, 1) # [B, C, T, H, W] -> [B, T, C, H, W] + x = apply_model_with_memblocks(self.decoder, x, self.parallel, self.show_progress_bar) + if self.patch_size > 1: x = F.pixel_shuffle(x, self.patch_size) + return x[:, self.frames_to_trim:].movedim(2, 1) diff --git a/latent_preview.py b/latent_preview.py index ddf6dcf49..66bded4b9 100644 --- a/latent_preview.py +++ b/latent_preview.py @@ -2,17 +2,24 @@ import torch from PIL import Image from comfy.cli_args import args, LatentPreviewMethod from comfy.taesd.taesd import TAESD +from comfy.sd import VAE import comfy.model_management import folder_paths import comfy.utils import logging MAX_PREVIEW_RESOLUTION = args.preview_size +VIDEO_TAES = ["taehv", "lighttaew2_2", "lighttaew2_1", "lighttaehy1_5"] -def preview_to_image(latent_image): - latents_ubyte = (((latent_image + 1.0) / 2.0).clamp(0, 1) # change scale from -1..1 to 0..1 - .mul(0xFF) # to 0..255 - ) +def preview_to_image(latent_image, do_scale=True): + if do_scale: + latents_ubyte = (((latent_image + 1.0) / 2.0).clamp(0, 1) # change scale from -1..1 to 0..1 + .mul(0xFF) # to 0..255 + ) + else: + latents_ubyte = (latent_image.clamp(0, 1) + .mul(0xFF) # to 0..255 + ) if comfy.model_management.directml_enabled: latents_ubyte = latents_ubyte.to(dtype=torch.uint8) latents_ubyte = latents_ubyte.to(device="cpu", dtype=torch.uint8, non_blocking=comfy.model_management.device_supports_non_blocking(latent_image.device)) @@ -35,6 +42,10 @@ class TAESDPreviewerImpl(LatentPreviewer): x_sample = self.taesd.decode(x0[:1])[0].movedim(0, 2) return preview_to_image(x_sample) +class TAEHVPreviewerImpl(TAESDPreviewerImpl): + def decode_latent_to_preview(self, x0): + x_sample = self.taesd.decode(x0[:1, :, :1])[0][0] + return preview_to_image(x_sample, do_scale=False) class Latent2RGBPreviewer(LatentPreviewer): def __init__(self, latent_rgb_factors, latent_rgb_factors_bias=None, latent_rgb_factors_reshape=None): @@ -81,8 +92,13 @@ def get_previewer(device, latent_format): if method == LatentPreviewMethod.TAESD: if taesd_decoder_path: - taesd = TAESD(None, taesd_decoder_path, latent_channels=latent_format.latent_channels).to(device) - previewer = TAESDPreviewerImpl(taesd) + if latent_format.taesd_decoder_name in VIDEO_TAES: + taesd = VAE(comfy.utils.load_torch_file(taesd_decoder_path)) + taesd.first_stage_model.show_progress_bar = False + previewer = TAEHVPreviewerImpl(taesd) + else: + taesd = TAESD(None, taesd_decoder_path, latent_channels=latent_format.latent_channels).to(device) + previewer = TAESDPreviewerImpl(taesd) else: logging.warning("Warning: TAESD previews enabled, but could not find models/vae_approx/{}".format(latent_format.taesd_decoder_name)) diff --git a/nodes.py b/nodes.py index bf73eb90e..495dec806 100644 --- a/nodes.py +++ b/nodes.py @@ -692,8 +692,10 @@ class LoraLoaderModelOnly(LoraLoader): return (self.load_lora(model, None, lora_name, strength_model, 0)[0],) class VAELoader: + video_taes = ["taehv", "lighttaew2_2", "lighttaew2_1", "lighttaehy1_5"] + image_taes = ["taesd", "taesdxl", "taesd3", "taef1"] @staticmethod - def vae_list(): + def vae_list(s): vaes = folder_paths.get_filename_list("vae") approx_vaes = folder_paths.get_filename_list("vae_approx") sdxl_taesd_enc = False @@ -722,6 +724,11 @@ class VAELoader: f1_taesd_dec = True elif v.startswith("taef1_decoder."): f1_taesd_enc = True + else: + for tae in s.video_taes: + if v.startswith(tae): + vaes.append(v) + if sd1_taesd_dec and sd1_taesd_enc: vaes.append("taesd") if sdxl_taesd_dec and sdxl_taesd_enc: @@ -765,7 +772,7 @@ class VAELoader: @classmethod def INPUT_TYPES(s): - return {"required": { "vae_name": (s.vae_list(), )}} + return {"required": { "vae_name": (s.vae_list(s), )}} RETURN_TYPES = ("VAE",) FUNCTION = "load_vae" @@ -776,10 +783,13 @@ class VAELoader: if vae_name == "pixel_space": sd = {} sd["pixel_space_vae"] = torch.tensor(1.0) - elif vae_name in ["taesd", "taesdxl", "taesd3", "taef1"]: + elif vae_name in self.image_taes: sd = self.load_taesd(vae_name) else: - vae_path = folder_paths.get_full_path_or_raise("vae", vae_name) + if os.path.splitext(vae_name)[0] in self.video_taes: + vae_path = folder_paths.get_full_path_or_raise("vae_approx", vae_name) + else: + vae_path = folder_paths.get_full_path_or_raise("vae", vae_name) sd = comfy.utils.load_torch_file(vae_path) vae = comfy.sd.VAE(sd=sd) vae.throw_exception_if_invalid() From 52a32e2b323b90295ab05a8c299590c890f2ecb6 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Fri, 28 Nov 2025 18:12:42 -0800 Subject: [PATCH 194/647] Support some z image lora formats. (#10978) --- comfy/lora.py | 8 ++++++++ comfy/utils.py | 45 +++++++++++++++++++++++++++++++++++++++++++++ 2 files changed, 53 insertions(+) diff --git a/comfy/lora.py b/comfy/lora.py index 36d26293a..360cd128f 100644 --- a/comfy/lora.py +++ b/comfy/lora.py @@ -313,6 +313,14 @@ def model_lora_keys_unet(model, key_map={}): key_map["transformer.{}".format(key_lora)] = k key_map["lycoris_{}".format(key_lora.replace(".", "_"))] = k #SimpleTuner lycoris format + if isinstance(model, comfy.model_base.Lumina2): + diffusers_keys = comfy.utils.z_image_to_diffusers(model.model_config.unet_config, output_prefix="diffusion_model.") + for k in diffusers_keys: + to = diffusers_keys[k] + key_lora = k[:-len(".weight")] + key_map["diffusion_model.{}".format(key_lora)] = to + key_map["lycoris_{}".format(key_lora.replace(".", "_"))] = to + return key_map diff --git a/comfy/utils.py b/comfy/utils.py index 4bd281057..21bd6e8cf 100644 --- a/comfy/utils.py +++ b/comfy/utils.py @@ -675,6 +675,51 @@ def flux_to_diffusers(mmdit_config, output_prefix=""): return key_map +def z_image_to_diffusers(mmdit_config, output_prefix=""): + n_layers = mmdit_config.get("n_layers", 0) + hidden_size = mmdit_config.get("dim", 0) + + key_map = {} + + for index in range(n_layers): + prefix_from = "layers.{}".format(index) + prefix_to = "{}layers.{}".format(output_prefix, index) + + for end in ("weight", "bias"): + k = "{}.attention.".format(prefix_from) + qkv = "{}.attention.qkv.{}".format(prefix_to, end) + + key_map["{}to_q.{}".format(k, end)] = (qkv, (0, 0, hidden_size)) + key_map["{}to_k.{}".format(k, end)] = (qkv, (0, hidden_size, hidden_size)) + key_map["{}to_v.{}".format(k, end)] = (qkv, (0, hidden_size * 2, hidden_size)) + + block_map = { + "attention.norm_q.weight": "attention.q_norm.weight", + "attention.norm_k.weight": "attention.k_norm.weight", + "attention.to_out.0.weight": "attention.out.weight", + "attention.to_out.0.bias": "attention.out.bias", + } + + for k in block_map: + key_map["{}.{}".format(prefix_from, k)] = "{}.{}".format(prefix_to, block_map[k]) + + MAP_BASIC = { + # Final layer + ("final_layer.linear.weight", "all_final_layer.2-1.linear.weight"), + ("final_layer.linear.bias", "all_final_layer.2-1.linear.bias"), + ("final_layer.adaLN_modulation.1.weight", "all_final_layer.2-1.adaLN_modulation.1.weight"), + ("final_layer.adaLN_modulation.1.bias", "all_final_layer.2-1.adaLN_modulation.1.bias"), + # X embedder + ("x_embedder.weight", "all_x_embedder.2-1.weight"), + ("x_embedder.bias", "all_x_embedder.2-1.bias"), + } + + for k in MAP_BASIC: + key_map[k[1]] = "{}{}".format(output_prefix, k[0]) + + return key_map + + def repeat_to_batch_size(tensor, batch_size, dim=0): if tensor.shape[dim] > batch_size: return tensor.narrow(dim, 0, batch_size) From af96d9812de3e420abd43275d9a5960535b6333c Mon Sep 17 00:00:00 2001 From: "Dr.Lt.Data" <128333288+ltdrdata@users.noreply.github.com> Date: Sat, 29 Nov 2025 11:28:42 +0900 Subject: [PATCH 195/647] feat(security): add System User protection with `__` prefix (#10966) * feat(security): add System User protection with `__` prefix Add protected namespace for custom nodes to store sensitive data (API keys, licenses) that cannot be accessed via HTTP endpoints. Key changes: - New API: get_system_user_directory() for internal access - New API: get_public_user_directory() with structural blocking - 3-layer defense: header validation, path blocking, creation prevention - 54 tests covering security, edge cases, and backward compatibility System Users use `__` prefix (e.g., __system, __cache) following Python's private member convention. They exist in user_directory/ but are completely blocked from /userdata HTTP endpoints. * style: remove unused imports --- app/user_manager.py | 21 +- folder_paths.py | 65 +++ .../app_test/user_manager_system_user_test.py | 193 +++++++++ .../folder_paths_test/system_user_test.py | 206 ++++++++++ .../system_user_endpoint_test.py | 375 ++++++++++++++++++ 5 files changed, 855 insertions(+), 5 deletions(-) create mode 100644 tests-unit/app_test/user_manager_system_user_test.py create mode 100644 tests-unit/folder_paths_test/system_user_test.py create mode 100644 tests-unit/prompt_server_test/system_user_endpoint_test.py diff --git a/app/user_manager.py b/app/user_manager.py index 675f6c0c6..e2c00dab2 100644 --- a/app/user_manager.py +++ b/app/user_manager.py @@ -59,6 +59,9 @@ class UserManager(): user = "default" if args.multi_user and "comfy-user" in request.headers: user = request.headers["comfy-user"] + # Block System Users (use same error message to prevent probing) + if user.startswith(folder_paths.SYSTEM_USER_PREFIX): + raise KeyError("Unknown user: " + user) if user not in self.users: raise KeyError("Unknown user: " + user) @@ -66,15 +69,16 @@ class UserManager(): return user def get_request_user_filepath(self, request, file, type="userdata", create_dir=True): - user_directory = folder_paths.get_user_directory() - if type == "userdata": - root_dir = user_directory + root_dir = folder_paths.get_user_directory() else: raise KeyError("Unknown filepath type:" + type) user = self.get_request_user_id(request) - path = user_root = os.path.abspath(os.path.join(root_dir, user)) + user_root = folder_paths.get_public_user_directory(user) + if user_root is None: + return None + path = user_root # prevent leaving /{type} if os.path.commonpath((root_dir, user_root)) != root_dir: @@ -101,7 +105,11 @@ class UserManager(): name = name.strip() if not name: raise ValueError("username not provided") + if name.startswith(folder_paths.SYSTEM_USER_PREFIX): + raise ValueError("System User prefix not allowed") user_id = re.sub("[^a-zA-Z0-9-_]+", '-', name) + if user_id.startswith(folder_paths.SYSTEM_USER_PREFIX): + raise ValueError("System User prefix not allowed") user_id = user_id + "_" + str(uuid.uuid4()) self.users[user_id] = name @@ -132,7 +140,10 @@ class UserManager(): if username in self.users.values(): return web.json_response({"error": "Duplicate username."}, status=400) - user_id = self.add_user(username) + try: + user_id = self.add_user(username) + except ValueError as e: + return web.json_response({"error": str(e)}, status=400) return web.json_response(user_id) @routes.get("/userdata") diff --git a/folder_paths.py b/folder_paths.py index ffdc4d020..9c96540e3 100644 --- a/folder_paths.py +++ b/folder_paths.py @@ -137,6 +137,71 @@ def set_user_directory(user_dir: str) -> None: user_directory = user_dir +# System User Protection - Protects system directories from HTTP endpoint access +# System Users are internal-only users that cannot be accessed via HTTP endpoints. +# They use the '__' prefix convention (similar to Python's private member convention). +SYSTEM_USER_PREFIX = "__" + + +def get_system_user_directory(name: str = "system") -> str: + """ + Get the path to a System User directory. + + System User directories (prefixed with '__') are only accessible via internal API, + not through HTTP endpoints. Use this for storing system-internal data that + should not be exposed to users. + + Args: + name: System user name (e.g., "system", "cache"). Must be alphanumeric + with underscores allowed, but cannot start with underscore. + + Returns: + Absolute path to the system user directory. + + Raises: + ValueError: If name is empty, invalid, or starts with underscore. + + Example: + >>> get_system_user_directory("cache") + '/path/to/user/__cache' + """ + if not name or not isinstance(name, str): + raise ValueError("System user name cannot be empty") + if not name.replace("_", "").isalnum(): + raise ValueError(f"Invalid system user name: '{name}'") + if name.startswith("_"): + raise ValueError("System user name should not start with underscore") + return os.path.join(get_user_directory(), f"{SYSTEM_USER_PREFIX}{name}") + + +def get_public_user_directory(user_id: str) -> str | None: + """ + Get the path to a Public User directory for HTTP endpoint access. + + This function provides structural security by returning None for any + System User (prefixed with '__'). All HTTP endpoints should use this + function instead of directly constructing user paths. + + Args: + user_id: User identifier from HTTP request. + + Returns: + Absolute path to the user directory, or None if user_id is invalid + or refers to a System User. + + Example: + >>> get_public_user_directory("default") + '/path/to/user/default' + >>> get_public_user_directory("__system") + None + """ + if not user_id or not isinstance(user_id, str): + return None + if user_id.startswith(SYSTEM_USER_PREFIX): + return None + return os.path.join(get_user_directory(), user_id) + + #NOTE: used in http server so don't put folders that should not be accessed remotely def get_directory_by_type(type_name: str) -> str | None: if type_name == "output": diff --git a/tests-unit/app_test/user_manager_system_user_test.py b/tests-unit/app_test/user_manager_system_user_test.py new file mode 100644 index 000000000..63b1ac5e5 --- /dev/null +++ b/tests-unit/app_test/user_manager_system_user_test.py @@ -0,0 +1,193 @@ +"""Tests for System User Protection in user_manager.py + +Tests cover: +- get_request_user_id(): 1st defense layer - blocks System Users from HTTP headers +- get_request_user_filepath(): 2nd defense layer - structural blocking via get_public_user_directory() +- add_user(): 3rd defense layer - prevents creation of System User names +- Defense layers integration tests +""" + +import pytest +from unittest.mock import MagicMock, patch +import tempfile + +import folder_paths +from app.user_manager import UserManager + + +@pytest.fixture +def mock_user_directory(): + """Create a temporary user directory.""" + with tempfile.TemporaryDirectory() as temp_dir: + original_dir = folder_paths.get_user_directory() + folder_paths.set_user_directory(temp_dir) + yield temp_dir + folder_paths.set_user_directory(original_dir) + + +@pytest.fixture +def user_manager(mock_user_directory): + """Create a UserManager instance for testing.""" + with patch('app.user_manager.args') as mock_args: + mock_args.multi_user = True + manager = UserManager() + # Add a default user for testing + manager.users = {"default": "default", "test_user_123": "Test User"} + yield manager + + +@pytest.fixture +def mock_request(): + """Create a mock request object.""" + request = MagicMock() + request.headers = {} + return request + + +class TestGetRequestUserId: + """Tests for get_request_user_id() - 1st defense layer. + + Verifies: + - System Users (__ prefix) in HTTP header are rejected with KeyError + - Public Users pass through successfully + """ + + def test_system_user_raises_error(self, user_manager, mock_request): + """Test System User in header raises KeyError.""" + mock_request.headers = {"comfy-user": "__system"} + + with patch('app.user_manager.args') as mock_args: + mock_args.multi_user = True + with pytest.raises(KeyError, match="Unknown user"): + user_manager.get_request_user_id(mock_request) + + def test_system_user_cache_raises_error(self, user_manager, mock_request): + """Test System User cache raises KeyError.""" + mock_request.headers = {"comfy-user": "__cache"} + + with patch('app.user_manager.args') as mock_args: + mock_args.multi_user = True + with pytest.raises(KeyError, match="Unknown user"): + user_manager.get_request_user_id(mock_request) + + def test_normal_user_works(self, user_manager, mock_request): + """Test normal user access works.""" + mock_request.headers = {"comfy-user": "default"} + + with patch('app.user_manager.args') as mock_args: + mock_args.multi_user = True + user_id = user_manager.get_request_user_id(mock_request) + assert user_id == "default" + + def test_unknown_user_raises_error(self, user_manager, mock_request): + """Test unknown user raises KeyError.""" + mock_request.headers = {"comfy-user": "unknown_user"} + + with patch('app.user_manager.args') as mock_args: + mock_args.multi_user = True + with pytest.raises(KeyError, match="Unknown user"): + user_manager.get_request_user_id(mock_request) + + +class TestGetRequestUserFilepath: + """Tests for get_request_user_filepath() - 2nd defense layer. + + Verifies: + - Returns None when get_public_user_directory() returns None (System User) + - Acts as backup defense if 1st layer is bypassed + """ + + def test_system_user_returns_none(self, user_manager, mock_request, mock_user_directory): + """Test System User returns None (structural blocking).""" + # First, we need to mock get_request_user_id to return System User + # But actually, get_request_user_id will raise KeyError first + # So we test via get_public_user_directory returning None + mock_request.headers = {"comfy-user": "default"} + + with patch('app.user_manager.args') as mock_args: + mock_args.multi_user = True + # Patch get_public_user_directory to return None for testing + with patch.object(folder_paths, 'get_public_user_directory', return_value=None): + result = user_manager.get_request_user_filepath(mock_request, "test.txt") + assert result is None + + def test_normal_user_gets_path(self, user_manager, mock_request, mock_user_directory): + """Test normal user gets valid filepath.""" + mock_request.headers = {"comfy-user": "default"} + + with patch('app.user_manager.args') as mock_args: + mock_args.multi_user = True + path = user_manager.get_request_user_filepath(mock_request, "test.txt") + assert path is not None + assert "default" in path + assert path.endswith("test.txt") + + +class TestAddUser: + """Tests for add_user() - 3rd defense layer (creation-time blocking). + + Verifies: + - System User name (__ prefix) creation is rejected with ValueError + - Sanitized usernames that become System User are also rejected + """ + + def test_system_user_prefix_name_raises(self, user_manager): + """Test System User prefix in name raises ValueError.""" + with pytest.raises(ValueError, match="System User prefix not allowed"): + user_manager.add_user("__system") + + def test_system_user_prefix_cache_raises(self, user_manager): + """Test System User cache prefix raises ValueError.""" + with pytest.raises(ValueError, match="System User prefix not allowed"): + user_manager.add_user("__cache") + + def test_sanitized_system_user_prefix_raises(self, user_manager): + """Test sanitized name becoming System User prefix raises ValueError (bypass prevention).""" + # "__test" directly starts with System User prefix + with pytest.raises(ValueError, match="System User prefix not allowed"): + user_manager.add_user("__test") + + def test_normal_user_creation(self, user_manager, mock_user_directory): + """Test normal user creation works.""" + user_id = user_manager.add_user("Normal User") + assert user_id is not None + assert not user_id.startswith("__") + assert "Normal-User" in user_id or "Normal_User" in user_id + + def test_empty_name_raises(self, user_manager): + """Test empty name raises ValueError.""" + with pytest.raises(ValueError, match="username not provided"): + user_manager.add_user("") + + def test_whitespace_only_raises(self, user_manager): + """Test whitespace-only name raises ValueError.""" + with pytest.raises(ValueError, match="username not provided"): + user_manager.add_user(" ") + + +class TestDefenseLayers: + """Integration tests for all three defense layers. + + Verifies: + - Each defense layer blocks System Users independently + - System User bypass is impossible through any layer + """ + + def test_layer1_get_request_user_id(self, user_manager, mock_request): + """Test 1st defense layer blocks System Users.""" + mock_request.headers = {"comfy-user": "__system"} + + with patch('app.user_manager.args') as mock_args: + mock_args.multi_user = True + with pytest.raises(KeyError): + user_manager.get_request_user_id(mock_request) + + def test_layer2_get_public_user_directory(self): + """Test 2nd defense layer blocks System Users.""" + result = folder_paths.get_public_user_directory("__system") + assert result is None + + def test_layer3_add_user(self, user_manager): + """Test 3rd defense layer blocks System User creation.""" + with pytest.raises(ValueError): + user_manager.add_user("__system") diff --git a/tests-unit/folder_paths_test/system_user_test.py b/tests-unit/folder_paths_test/system_user_test.py new file mode 100644 index 000000000..cd46459f1 --- /dev/null +++ b/tests-unit/folder_paths_test/system_user_test.py @@ -0,0 +1,206 @@ +"""Tests for System User Protection in folder_paths.py + +Tests cover: +- get_system_user_directory(): Internal API for custom nodes to access System User directories +- get_public_user_directory(): HTTP endpoint access with System User blocking +- Backward compatibility: Existing APIs unchanged +- Security: Path traversal and injection prevention +""" + +import pytest +import os +import tempfile + +from folder_paths import ( + get_system_user_directory, + get_public_user_directory, + get_user_directory, + set_user_directory, +) + + +@pytest.fixture(scope="module") +def mock_user_directory(): + """Create a temporary user directory for testing.""" + with tempfile.TemporaryDirectory() as temp_dir: + original_dir = get_user_directory() + set_user_directory(temp_dir) + yield temp_dir + set_user_directory(original_dir) + + +class TestGetSystemUserDirectory: + """Tests for get_system_user_directory() - internal API for System User directories. + + Verifies: + - Custom nodes can access System User directories via internal API + - Input validation prevents path traversal attacks + """ + + def test_default_name(self, mock_user_directory): + """Test default 'system' name.""" + path = get_system_user_directory() + assert path.endswith("__system") + assert mock_user_directory in path + + def test_custom_name(self, mock_user_directory): + """Test custom system user name.""" + path = get_system_user_directory("cache") + assert path.endswith("__cache") + assert "__cache" in path + + def test_name_with_underscore(self, mock_user_directory): + """Test name with underscore in middle.""" + path = get_system_user_directory("my_cache") + assert "__my_cache" in path + + def test_empty_name_raises(self): + """Test empty name raises ValueError.""" + with pytest.raises(ValueError, match="cannot be empty"): + get_system_user_directory("") + + def test_none_name_raises(self): + """Test None name raises ValueError.""" + with pytest.raises(ValueError, match="cannot be empty"): + get_system_user_directory(None) + + def test_name_starting_with_underscore_raises(self): + """Test name starting with underscore raises ValueError.""" + with pytest.raises(ValueError, match="should not start with underscore"): + get_system_user_directory("_system") + + def test_path_traversal_raises(self): + """Test path traversal attempt raises ValueError (security).""" + with pytest.raises(ValueError, match="Invalid system user name"): + get_system_user_directory("../escape") + + def test_path_traversal_middle_raises(self): + """Test path traversal in middle raises ValueError (security).""" + with pytest.raises(ValueError, match="Invalid system user name"): + get_system_user_directory("system/../other") + + def test_special_chars_raise(self): + """Test special characters raise ValueError (security).""" + with pytest.raises(ValueError, match="Invalid system user name"): + get_system_user_directory("system!") + + def test_returns_absolute_path(self, mock_user_directory): + """Test returned path is absolute.""" + path = get_system_user_directory("test") + assert os.path.isabs(path) + + +class TestGetPublicUserDirectory: + """Tests for get_public_user_directory() - HTTP endpoint access with System User blocking. + + Verifies: + - System Users (__ prefix) return None, blocking HTTP access + - Public Users get valid paths + - New endpoints using this function are automatically protected + """ + + def test_normal_user(self, mock_user_directory): + """Test normal user returns valid path.""" + path = get_public_user_directory("default") + assert path is not None + assert "default" in path + assert mock_user_directory in path + + def test_system_user_returns_none(self): + """Test System User (__ prefix) returns None - blocks HTTP access.""" + assert get_public_user_directory("__system") is None + + def test_system_user_cache_returns_none(self): + """Test System User cache returns None.""" + assert get_public_user_directory("__cache") is None + + def test_empty_user_returns_none(self): + """Test empty user returns None.""" + assert get_public_user_directory("") is None + + def test_none_user_returns_none(self): + """Test None user returns None.""" + assert get_public_user_directory(None) is None + + def test_header_injection_returns_none(self): + """Test header injection attempt returns None (security).""" + assert get_public_user_directory("__system\r\nX-Injected: true") is None + + def test_null_byte_injection_returns_none(self): + """Test null byte injection handling (security).""" + # Note: startswith check happens before any path operations + result = get_public_user_directory("user\x00__system") + # This should return a path since it doesn't start with __ + # The actual security comes from the path not being __* + assert result is not None or result is None # Depends on validation + + def test_path_traversal_attempt(self, mock_user_directory): + """Test path traversal attempt handling.""" + # This function doesn't validate paths, only reserved prefix + # Path traversal should be handled by the caller + path = get_public_user_directory("../../../etc/passwd") + # Returns path but doesn't start with __, so not None + # Actual path validation happens in user_manager + assert path is not None or "__" not in "../../../etc/passwd" + + def test_returns_absolute_path(self, mock_user_directory): + """Test returned path is absolute.""" + path = get_public_user_directory("testuser") + assert path is not None + assert os.path.isabs(path) + + +class TestBackwardCompatibility: + """Tests for backward compatibility with existing APIs. + + Verifies: + - get_user_directory() API unchanged + - Existing user data remains accessible + """ + + def test_get_user_directory_unchanged(self, mock_user_directory): + """Test get_user_directory() still works as before.""" + user_dir = get_user_directory() + assert user_dir is not None + assert os.path.isabs(user_dir) + assert user_dir == mock_user_directory + + def test_existing_user_accessible(self, mock_user_directory): + """Test existing users can access their directories.""" + path = get_public_user_directory("default") + assert path is not None + assert "default" in path + + +class TestEdgeCases: + """Tests for edge cases in System User detection. + + Verifies: + - Only __ prefix is blocked (not _, not middle __) + - Bypass attempts are prevented + """ + + def test_prefix_only(self): + """Test prefix-only string is blocked.""" + assert get_public_user_directory("__") is None + + def test_single_underscore_allowed(self): + """Test single underscore prefix is allowed (not System User).""" + path = get_public_user_directory("_system") + assert path is not None + assert "_system" in path + + def test_triple_underscore_blocked(self): + """Test triple underscore is blocked (starts with __).""" + assert get_public_user_directory("___system") is None + + def test_underscore_in_middle_allowed(self): + """Test underscore in middle is allowed.""" + path = get_public_user_directory("my__system") + assert path is not None + assert "my__system" in path + + def test_leading_space_allowed(self): + """Test leading space + prefix is allowed (doesn't start with __).""" + path = get_public_user_directory(" __system") + assert path is not None diff --git a/tests-unit/prompt_server_test/system_user_endpoint_test.py b/tests-unit/prompt_server_test/system_user_endpoint_test.py new file mode 100644 index 000000000..22ac00af9 --- /dev/null +++ b/tests-unit/prompt_server_test/system_user_endpoint_test.py @@ -0,0 +1,375 @@ +"""E2E Tests for System User Protection HTTP Endpoints + +Tests cover: +- HTTP endpoint blocking: System Users cannot access /userdata (GET, POST, DELETE, move) +- User creation blocking: System User names cannot be created via POST /users +- Backward compatibility: Public Users work as before +- Custom node scenario: Internal API works while HTTP is blocked +- Structural security: get_public_user_directory() provides automatic protection +""" + +import pytest +import os +from aiohttp import web +from app.user_manager import UserManager +from unittest.mock import patch +import folder_paths + + +@pytest.fixture +def mock_user_directory(tmp_path): + """Create a temporary user directory.""" + original_dir = folder_paths.get_user_directory() + folder_paths.set_user_directory(str(tmp_path)) + yield tmp_path + folder_paths.set_user_directory(original_dir) + + +@pytest.fixture +def user_manager_multi_user(mock_user_directory): + """Create UserManager in multi-user mode.""" + with patch('app.user_manager.args') as mock_args: + mock_args.multi_user = True + um = UserManager() + # Add test users + um.users = {"default": "default", "test_user_123": "Test User"} + yield um + + +@pytest.fixture +def app_multi_user(user_manager_multi_user): + """Create app with multi-user mode enabled.""" + app = web.Application() + routes = web.RouteTableDef() + user_manager_multi_user.add_routes(routes) + app.add_routes(routes) + return app + + +class TestSystemUserEndpointBlocking: + """E2E tests for System User blocking on all HTTP endpoints. + + Verifies: + - GET /userdata blocked for System Users + - POST /userdata blocked for System Users + - DELETE /userdata blocked for System Users + - POST /userdata/.../move/... blocked for System Users + """ + + @pytest.mark.asyncio + async def test_userdata_get_blocks_system_user( + self, aiohttp_client, app_multi_user, mock_user_directory + ): + """ + GET /userdata with System User header should be blocked. + """ + # Create test directory for System User (simulating internal creation) + system_user_dir = mock_user_directory / "__system" + system_user_dir.mkdir() + (system_user_dir / "secret.txt").write_text("sensitive data") + + client = await aiohttp_client(app_multi_user) + + with patch('app.user_manager.args') as mock_args: + mock_args.multi_user = True + # Attempt to access System User's data via HTTP + resp = await client.get( + "/userdata?dir=.", + headers={"comfy-user": "__system"} + ) + + # Should be blocked (403 Forbidden or similar error) + assert resp.status in [400, 403, 500], \ + f"System User access should be blocked, got {resp.status}" + + @pytest.mark.asyncio + async def test_userdata_post_blocks_system_user( + self, aiohttp_client, app_multi_user, mock_user_directory + ): + """ + POST /userdata with System User header should be blocked. + """ + client = await aiohttp_client(app_multi_user) + + with patch('app.user_manager.args') as mock_args: + mock_args.multi_user = True + resp = await client.post( + "/userdata/test.txt", + headers={"comfy-user": "__system"}, + data=b"malicious content" + ) + + assert resp.status in [400, 403, 500], \ + f"System User write should be blocked, got {resp.status}" + + # Verify no file was created + assert not (mock_user_directory / "__system" / "test.txt").exists() + + @pytest.mark.asyncio + async def test_userdata_delete_blocks_system_user( + self, aiohttp_client, app_multi_user, mock_user_directory + ): + """ + DELETE /userdata with System User header should be blocked. + """ + # Create a file in System User directory + system_user_dir = mock_user_directory / "__system" + system_user_dir.mkdir() + secret_file = system_user_dir / "secret.txt" + secret_file.write_text("do not delete") + + client = await aiohttp_client(app_multi_user) + + with patch('app.user_manager.args') as mock_args: + mock_args.multi_user = True + resp = await client.delete( + "/userdata/secret.txt", + headers={"comfy-user": "__system"} + ) + + assert resp.status in [400, 403, 500], \ + f"System User delete should be blocked, got {resp.status}" + + # Verify file still exists + assert secret_file.exists() + + @pytest.mark.asyncio + async def test_v2_userdata_blocks_system_user( + self, aiohttp_client, app_multi_user, mock_user_directory + ): + """ + GET /v2/userdata with System User header should be blocked. + """ + client = await aiohttp_client(app_multi_user) + + with patch('app.user_manager.args') as mock_args: + mock_args.multi_user = True + resp = await client.get( + "/v2/userdata", + headers={"comfy-user": "__system"} + ) + + assert resp.status in [400, 403, 500], \ + f"System User v2 access should be blocked, got {resp.status}" + + @pytest.mark.asyncio + async def test_move_userdata_blocks_system_user( + self, aiohttp_client, app_multi_user, mock_user_directory + ): + """ + POST /userdata/{file}/move/{dest} with System User header should be blocked. + """ + system_user_dir = mock_user_directory / "__system" + system_user_dir.mkdir() + (system_user_dir / "source.txt").write_text("sensitive data") + + client = await aiohttp_client(app_multi_user) + + with patch('app.user_manager.args') as mock_args: + mock_args.multi_user = True + resp = await client.post( + "/userdata/source.txt/move/dest.txt", + headers={"comfy-user": "__system"} + ) + + assert resp.status in [400, 403, 500], \ + f"System User move should be blocked, got {resp.status}" + + # Verify source file still exists (move was blocked) + assert (system_user_dir / "source.txt").exists() + + +class TestSystemUserCreationBlocking: + """E2E tests for blocking System User name creation via POST /users. + + Verifies: + - POST /users returns 400 for System User name (not 500) + """ + + @pytest.mark.asyncio + async def test_post_users_blocks_system_user_name( + self, aiohttp_client, app_multi_user + ): + """POST /users with System User name should return 400 Bad Request.""" + client = await aiohttp_client(app_multi_user) + + resp = await client.post( + "/users", + json={"username": "__system"} + ) + + assert resp.status == 400, \ + f"System User creation should return 400, got {resp.status}" + + @pytest.mark.asyncio + async def test_post_users_blocks_system_user_prefix_variations( + self, aiohttp_client, app_multi_user + ): + """POST /users with any System User prefix variation should return 400 Bad Request.""" + client = await aiohttp_client(app_multi_user) + + system_user_names = ["__system", "__cache", "__config", "__anything"] + + for name in system_user_names: + resp = await client.post("/users", json={"username": name}) + assert resp.status == 400, \ + f"System User name '{name}' should return 400, got {resp.status}" + + +class TestPublicUserStillWorks: + """E2E tests for backward compatibility - Public Users should work as before. + + Verifies: + - Public Users can access their data via HTTP + - Public Users can create files via HTTP + """ + + @pytest.mark.asyncio + async def test_public_user_can_access_userdata( + self, aiohttp_client, app_multi_user, mock_user_directory + ): + """ + Public Users should still be able to access their data. + """ + # Create test directory for Public User + user_dir = mock_user_directory / "default" + user_dir.mkdir() + test_dir = user_dir / "workflows" + test_dir.mkdir() + (test_dir / "test.json").write_text('{"test": true}') + + client = await aiohttp_client(app_multi_user) + + with patch('app.user_manager.args') as mock_args: + mock_args.multi_user = True + resp = await client.get( + "/userdata?dir=workflows", + headers={"comfy-user": "default"} + ) + + assert resp.status == 200 + data = await resp.json() + assert "test.json" in data + + @pytest.mark.asyncio + async def test_public_user_can_create_files( + self, aiohttp_client, app_multi_user, mock_user_directory + ): + """ + Public Users should still be able to create files. + """ + # Create user directory + user_dir = mock_user_directory / "default" + user_dir.mkdir() + + client = await aiohttp_client(app_multi_user) + + with patch('app.user_manager.args') as mock_args: + mock_args.multi_user = True + resp = await client.post( + "/userdata/newfile.txt", + headers={"comfy-user": "default"}, + data=b"user content" + ) + + assert resp.status == 200 + assert (user_dir / "newfile.txt").exists() + + +class TestCustomNodeScenario: + """Tests for custom node use case: internal API access vs HTTP blocking. + + Verifies: + - Internal API (get_system_user_directory) works for custom nodes + - HTTP endpoint cannot access data created via internal API + """ + + def test_internal_api_can_access_system_user(self, mock_user_directory): + """ + Internal API (get_system_user_directory) should work for custom nodes. + """ + # Custom node uses internal API + system_path = folder_paths.get_system_user_directory("mynode_config") + + assert system_path is not None + assert "__mynode_config" in system_path + + # Can create and write to System User directory + os.makedirs(system_path, exist_ok=True) + config_file = os.path.join(system_path, "settings.json") + with open(config_file, "w") as f: + f.write('{"api_key": "secret"}') + + assert os.path.exists(config_file) + + @pytest.mark.asyncio + async def test_http_cannot_access_internal_data( + self, aiohttp_client, app_multi_user, mock_user_directory + ): + """ + HTTP endpoint cannot access data created via internal API. + """ + # Custom node creates data via internal API + system_path = folder_paths.get_system_user_directory("mynode_config") + os.makedirs(system_path, exist_ok=True) + with open(os.path.join(system_path, "secret.json"), "w") as f: + f.write('{"api_key": "secret"}') + + client = await aiohttp_client(app_multi_user) + + # Attacker tries to access via HTTP + with patch('app.user_manager.args') as mock_args: + mock_args.multi_user = True + resp = await client.get( + "/userdata/secret.json", + headers={"comfy-user": "__mynode_config"} + ) + + # Should be blocked + assert resp.status in [400, 403, 500] + + +class TestStructuralSecurity: + """Tests for structural security pattern. + + Verifies: + - get_public_user_directory() automatically blocks System Users + - New endpoints using this function are automatically protected + """ + + def test_get_public_user_directory_blocks_system_user(self): + """ + Any code using get_public_user_directory() is automatically protected. + """ + # This is the structural security - any new endpoint using this function + # will automatically block System Users + assert folder_paths.get_public_user_directory("__system") is None + assert folder_paths.get_public_user_directory("__cache") is None + assert folder_paths.get_public_user_directory("__anything") is None + + # Public Users work + assert folder_paths.get_public_user_directory("default") is not None + assert folder_paths.get_public_user_directory("user123") is not None + + def test_structural_security_pattern(self, mock_user_directory): + """ + Demonstrate the structural security pattern for new endpoints. + + Any new endpoint should follow this pattern: + 1. Get user from request + 2. Use get_public_user_directory() - automatically blocks System Users + 3. If None, return error + """ + def new_endpoint_handler(user_id: str) -> str | None: + """Example of how new endpoints should be implemented.""" + user_path = folder_paths.get_public_user_directory(user_id) + if user_path is None: + return None # Blocked + return user_path + + # System Users are automatically blocked + assert new_endpoint_handler("__system") is None + assert new_endpoint_handler("__secret") is None + + # Public Users work + assert new_endpoint_handler("default") is not None From 5151cff293607c2191981fd16c62c1b1a6939695 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Fri, 28 Nov 2025 20:55:00 -0800 Subject: [PATCH 196/647] Add some missing z image lora layers. (#10980) --- comfy/lora.py | 9 +++++---- comfy/utils.py | 51 +++++++++++++++++++++++++++++++++++--------------- 2 files changed, 41 insertions(+), 19 deletions(-) diff --git a/comfy/lora.py b/comfy/lora.py index 360cd128f..3a9077869 100644 --- a/comfy/lora.py +++ b/comfy/lora.py @@ -316,10 +316,11 @@ def model_lora_keys_unet(model, key_map={}): if isinstance(model, comfy.model_base.Lumina2): diffusers_keys = comfy.utils.z_image_to_diffusers(model.model_config.unet_config, output_prefix="diffusion_model.") for k in diffusers_keys: - to = diffusers_keys[k] - key_lora = k[:-len(".weight")] - key_map["diffusion_model.{}".format(key_lora)] = to - key_map["lycoris_{}".format(key_lora.replace(".", "_"))] = to + if k.endswith(".weight"): + to = diffusers_keys[k] + key_lora = k[:-len(".weight")] + key_map["diffusion_model.{}".format(key_lora)] = to + key_map["lycoris_{}".format(key_lora.replace(".", "_"))] = to return key_map diff --git a/comfy/utils.py b/comfy/utils.py index 21bd6e8cf..37485e497 100644 --- a/comfy/utils.py +++ b/comfy/utils.py @@ -678,17 +678,14 @@ def flux_to_diffusers(mmdit_config, output_prefix=""): def z_image_to_diffusers(mmdit_config, output_prefix=""): n_layers = mmdit_config.get("n_layers", 0) hidden_size = mmdit_config.get("dim", 0) - + n_context_refiner = mmdit_config.get("n_refiner_layers", 2) + n_noise_refiner = mmdit_config.get("n_refiner_layers", 2) key_map = {} - for index in range(n_layers): - prefix_from = "layers.{}".format(index) - prefix_to = "{}layers.{}".format(output_prefix, index) - + def add_block_keys(prefix_from, prefix_to, has_adaln=True): for end in ("weight", "bias"): k = "{}.attention.".format(prefix_from) qkv = "{}.attention.qkv.{}".format(prefix_to, end) - key_map["{}to_q.{}".format(k, end)] = (qkv, (0, 0, hidden_size)) key_map["{}to_k.{}".format(k, end)] = (qkv, (0, hidden_size, hidden_size)) key_map["{}to_v.{}".format(k, end)] = (qkv, (0, hidden_size * 2, hidden_size)) @@ -698,28 +695,52 @@ def z_image_to_diffusers(mmdit_config, output_prefix=""): "attention.norm_k.weight": "attention.k_norm.weight", "attention.to_out.0.weight": "attention.out.weight", "attention.to_out.0.bias": "attention.out.bias", + "attention_norm1.weight": "attention_norm1.weight", + "attention_norm2.weight": "attention_norm2.weight", + "feed_forward.w1.weight": "feed_forward.w1.weight", + "feed_forward.w2.weight": "feed_forward.w2.weight", + "feed_forward.w3.weight": "feed_forward.w3.weight", + "ffn_norm1.weight": "ffn_norm1.weight", + "ffn_norm2.weight": "ffn_norm2.weight", } + if has_adaln: + block_map["adaLN_modulation.0.weight"] = "adaLN_modulation.0.weight" + block_map["adaLN_modulation.0.bias"] = "adaLN_modulation.0.bias" + for k, v in block_map.items(): + key_map["{}.{}".format(prefix_from, k)] = "{}.{}".format(prefix_to, v) - for k in block_map: - key_map["{}.{}".format(prefix_from, k)] = "{}.{}".format(prefix_to, block_map[k]) + for i in range(n_layers): + add_block_keys("layers.{}".format(i), "{}layers.{}".format(output_prefix, i)) - MAP_BASIC = { - # Final layer + for i in range(n_context_refiner): + add_block_keys("context_refiner.{}".format(i), "{}context_refiner.{}".format(output_prefix, i)) + + for i in range(n_noise_refiner): + add_block_keys("noise_refiner.{}".format(i), "{}noise_refiner.{}".format(output_prefix, i)) + + MAP_BASIC = [ ("final_layer.linear.weight", "all_final_layer.2-1.linear.weight"), ("final_layer.linear.bias", "all_final_layer.2-1.linear.bias"), ("final_layer.adaLN_modulation.1.weight", "all_final_layer.2-1.adaLN_modulation.1.weight"), ("final_layer.adaLN_modulation.1.bias", "all_final_layer.2-1.adaLN_modulation.1.bias"), - # X embedder ("x_embedder.weight", "all_x_embedder.2-1.weight"), ("x_embedder.bias", "all_x_embedder.2-1.bias"), - } + ("x_pad_token", "x_pad_token"), + ("cap_embedder.0.weight", "cap_embedder.0.weight"), + ("cap_embedder.1.weight", "cap_embedder.1.weight"), + ("cap_embedder.1.bias", "cap_embedder.1.bias"), + ("cap_pad_token", "cap_pad_token"), + ("t_embedder.mlp.0.weight", "t_embedder.mlp.0.weight"), + ("t_embedder.mlp.0.bias", "t_embedder.mlp.0.bias"), + ("t_embedder.mlp.2.weight", "t_embedder.mlp.2.weight"), + ("t_embedder.mlp.2.bias", "t_embedder.mlp.2.bias"), + ] - for k in MAP_BASIC: - key_map[k[1]] = "{}{}".format(output_prefix, k[0]) + for c, diffusers in MAP_BASIC: + key_map[diffusers] = "{}{}".format(output_prefix, c) return key_map - def repeat_to_batch_size(tensor, batch_size, dim=0): if tensor.shape[dim] > batch_size: return tensor.narrow(dim, 0, batch_size) From 0a6746898d6864d65e2fc7504e5e875f8c19c0ba Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sat, 29 Nov 2025 15:00:55 -0800 Subject: [PATCH 197/647] Make the ScaleRope node work on Z Image and Lumina. (#10994) --- comfy/ldm/lumina/model.py | 16 ++++++++++++++-- 1 file changed, 14 insertions(+), 2 deletions(-) diff --git a/comfy/ldm/lumina/model.py b/comfy/ldm/lumina/model.py index 565400b54..7d7e9112c 100644 --- a/comfy/ldm/lumina/model.py +++ b/comfy/ldm/lumina/model.py @@ -517,11 +517,23 @@ class NextDiT(nn.Module): B, C, H, W = x.shape x = self.x_embedder(x.view(B, C, H // pH, pH, W // pW, pW).permute(0, 2, 4, 3, 5, 1).flatten(3).flatten(1, 2)) + rope_options = transformer_options.get("rope_options", None) + h_scale = 1.0 + w_scale = 1.0 + h_start = 0 + w_start = 0 + if rope_options is not None: + h_scale = rope_options.get("scale_y", 1.0) + w_scale = rope_options.get("scale_x", 1.0) + + h_start = rope_options.get("shift_y", 0.0) + w_start = rope_options.get("shift_x", 0.0) + H_tokens, W_tokens = H // pH, W // pW x_pos_ids = torch.zeros((bsz, x.shape[1], 3), dtype=torch.float32, device=device) x_pos_ids[:, :, 0] = cap_feats.shape[1] + 1 - x_pos_ids[:, :, 1] = torch.arange(H_tokens, dtype=torch.float32, device=device).view(-1, 1).repeat(1, W_tokens).flatten() - x_pos_ids[:, :, 2] = torch.arange(W_tokens, dtype=torch.float32, device=device).view(1, -1).repeat(H_tokens, 1).flatten() + x_pos_ids[:, :, 1] = (torch.arange(H_tokens, dtype=torch.float32, device=device) * h_scale + h_start).view(-1, 1).repeat(1, W_tokens).flatten() + x_pos_ids[:, :, 2] = (torch.arange(W_tokens, dtype=torch.float32, device=device) * w_scale + w_start).view(1, -1).repeat(H_tokens, 1).flatten() if self.pad_tokens_multiple is not None: pad_extra = (-x.shape[1]) % self.pad_tokens_multiple From 4967f81778f84b41acc40ed03536dd71dd88e5f2 Mon Sep 17 00:00:00 2001 From: ComfyUI Wiki Date: Sun, 30 Nov 2025 10:07:26 +0800 Subject: [PATCH 198/647] update template to 0.7.25 (#10996) * update template to 0.7.24 * Update template to 0.7.25 --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index e0b2c566b..386477808 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,5 +1,5 @@ comfyui-frontend-package==1.32.9 -comfyui-workflow-templates==0.7.23 +comfyui-workflow-templates==0.7.25 comfyui-embedded-docs==0.3.1 torch torchsde From f8b981ae9a5676311624bbafa636a1874db79459 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sun, 30 Nov 2025 01:21:31 -0800 Subject: [PATCH 199/647] Next AMD portable will have pytorch with ROCm 7.1.1 (#11002) --- .github/workflows/release-stable-all.yml | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/.github/workflows/release-stable-all.yml b/.github/workflows/release-stable-all.yml index 9274b4170..d72ece2ce 100644 --- a/.github/workflows/release-stable-all.yml +++ b/.github/workflows/release-stable-all.yml @@ -65,11 +65,11 @@ jobs: contents: "write" packages: "write" pull-requests: "read" - name: "Release AMD ROCm 6.4.4" + name: "Release AMD ROCm 7.1.1" uses: ./.github/workflows/stable-release.yml with: git_tag: ${{ inputs.git_tag }} - cache_tag: "rocm644" + cache_tag: "rocm711" python_minor: "12" python_patch: "10" rel_name: "amd" From 7dbd5dfe91f057b83dcba0c127f712f6d71f7def Mon Sep 17 00:00:00 2001 From: Christian Byrne Date: Mon, 1 Dec 2025 10:27:17 -0800 Subject: [PATCH 200/647] bump comfyui-frontend-package to 1.32.10 (#11018) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 386477808..045b2ac54 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,4 +1,4 @@ -comfyui-frontend-package==1.32.9 +comfyui-frontend-package==1.32.10 comfyui-workflow-templates==0.7.25 comfyui-embedded-docs==0.3.1 torch From 2640acb31ccfddee57ba22d5245bf456e8dffe53 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 1 Dec 2025 14:13:48 -0800 Subject: [PATCH 201/647] Update qwen tokenizer to add qwen 3 tokens. (#11029) Doesn't actually change anything for current workflows because none of the current models have a template with the think tokens. --- .../qwen25_tokenizer/tokenizer_config.json | 16 ++++++++-------- 1 file changed, 8 insertions(+), 8 deletions(-) diff --git a/comfy/text_encoders/qwen25_tokenizer/tokenizer_config.json b/comfy/text_encoders/qwen25_tokenizer/tokenizer_config.json index 67688e82c..df5b5d7fe 100644 --- a/comfy/text_encoders/qwen25_tokenizer/tokenizer_config.json +++ b/comfy/text_encoders/qwen25_tokenizer/tokenizer_config.json @@ -179,36 +179,36 @@ "special": false }, "151665": { - "content": "<|img|>", + "content": "", "lstrip": false, "normalized": false, "rstrip": false, "single_word": false, - "special": true + "special": false }, "151666": { - "content": "<|endofimg|>", + "content": "", "lstrip": false, "normalized": false, "rstrip": false, "single_word": false, - "special": true + "special": false }, "151667": { - "content": "<|meta|>", + "content": "", "lstrip": false, "normalized": false, "rstrip": false, "single_word": false, - "special": true + "special": false }, "151668": { - "content": "<|endofmeta|>", + "content": "", "lstrip": false, "normalized": false, "rstrip": false, "single_word": false, - "special": true + "special": false } }, "additional_special_tokens": [ From 1cb7e22a95701f2619d1ddf5683ea221b58a0c13 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Tue, 2 Dec 2025 02:11:52 +0200 Subject: [PATCH 202/647] [API Nodes] add Kling O1 model support (#11025) * feat(api-nodes): add Kling O1 model support * fix: increase max allowed duration to 10.05 seconds * fix(VideoInput): respect "format" argument --- comfy_api/latest/_input_impl/video_types.py | 5 +- comfy_api_nodes/apis/kling_api.py | 66 +++ comfy_api_nodes/nodes_kling.py | 444 +++++++++++++++++++- comfy_api_nodes/util/upload_helpers.py | 3 +- 4 files changed, 499 insertions(+), 19 deletions(-) create mode 100644 comfy_api_nodes/apis/kling_api.py diff --git a/comfy_api/latest/_input_impl/video_types.py b/comfy_api/latest/_input_impl/video_types.py index bde37f90a..7231bf13c 100644 --- a/comfy_api/latest/_input_impl/video_types.py +++ b/comfy_api/latest/_input_impl/video_types.py @@ -336,7 +336,10 @@ class VideoFromComponents(VideoInput): raise ValueError("Only MP4 format is supported for now") if codec != VideoCodec.AUTO and codec != VideoCodec.H264: raise ValueError("Only H264 codec is supported for now") - with av.open(path, mode='w', options={'movflags': 'use_metadata_tags'}) as output: + extra_kwargs = {} + if format != VideoContainer.AUTO: + extra_kwargs["format"] = format.value + with av.open(path, mode='w', options={'movflags': 'use_metadata_tags'}, **extra_kwargs) as output: # Add metadata before writing any streams if metadata is not None: for key, value in metadata.items(): diff --git a/comfy_api_nodes/apis/kling_api.py b/comfy_api_nodes/apis/kling_api.py new file mode 100644 index 000000000..0a3b447c5 --- /dev/null +++ b/comfy_api_nodes/apis/kling_api.py @@ -0,0 +1,66 @@ +from pydantic import BaseModel, Field + + +class OmniProText2VideoRequest(BaseModel): + model_name: str = Field(..., description="kling-video-o1") + aspect_ratio: str = Field(..., description="'16:9', '9:16' or '1:1'") + duration: str = Field(..., description="'5' or '10'") + prompt: str = Field(...) + mode: str = Field("pro") + + +class OmniParamImage(BaseModel): + image_url: str = Field(...) + type: str | None = Field(None, description="Can be 'first_frame' or 'end_frame'") + + +class OmniParamVideo(BaseModel): + video_url: str = Field(...) + refer_type: str | None = Field(..., description="Can be 'base' or 'feature'") + keep_original_sound: str = Field(..., description="'yes' or 'no'") + + +class OmniProFirstLastFrameRequest(BaseModel): + model_name: str = Field(..., description="kling-video-o1") + image_list: list[OmniParamImage] = Field(..., min_length=1, max_length=7) + duration: str = Field(..., description="'5' or '10'") + prompt: str = Field(...) + mode: str = Field("pro") + + +class OmniProReferences2VideoRequest(BaseModel): + model_name: str = Field(..., description="kling-video-o1") + aspect_ratio: str | None = Field(..., description="'16:9', '9:16' or '1:1'") + image_list: list[OmniParamImage] | None = Field( + None, max_length=7, description="Max length 4 when video is present." + ) + video_list: list[OmniParamVideo] | None = Field(None, max_length=1) + duration: str | None = Field(..., description="From 3 to 10.") + prompt: str = Field(...) + mode: str = Field("pro") + + +class TaskStatusVideoResult(BaseModel): + duration: str | None = Field(None, description="Total video duration") + id: str | None = Field(None, description="Generated video ID") + url: str | None = Field(None, description="URL for generated video") + + +class TaskStatusVideoResults(BaseModel): + videos: list[TaskStatusVideoResult] | None = Field(None) + + +class TaskStatusVideoResponseData(BaseModel): + created_at: int | None = Field(None, description="Task creation time") + updated_at: int | None = Field(None, description="Task update time") + task_status: str | None = None + task_status_msg: str | None = Field(None, description="Additional failure reason. Only for polling endpoint.") + task_id: str | None = Field(None, description="Task ID") + task_result: TaskStatusVideoResults | None = Field(None) + + +class TaskStatusVideoResponse(BaseModel): + code: int | None = Field(None, description="Error code") + message: str | None = Field(None, description="Error message") + request_id: str | None = Field(None, description="Request ID") + data: TaskStatusVideoResponseData | None = Field(None) diff --git a/comfy_api_nodes/nodes_kling.py b/comfy_api_nodes/nodes_kling.py index 23a7f55f1..850c44db6 100644 --- a/comfy_api_nodes/nodes_kling.py +++ b/comfy_api_nodes/nodes_kling.py @@ -4,13 +4,13 @@ For source of truth on the allowed permutations of request fields, please refere - [Compatibility Table](https://app.klingai.com/global/dev/document-api/apiReference/model/skillsMap) """ -import math import logging - -from typing_extensions import override +import math import torch +from typing_extensions import override +from comfy_api.latest import IO, ComfyExtension, Input, InputImpl from comfy_api_nodes.apis import ( KlingCameraControl, KlingCameraConfig, @@ -48,23 +48,31 @@ from comfy_api_nodes.apis import ( KlingCharacterEffectModelName, KlingSingleImageEffectModelName, ) +from comfy_api_nodes.apis.kling_api import ( + OmniParamImage, + OmniParamVideo, + OmniProFirstLastFrameRequest, + OmniProReferences2VideoRequest, + OmniProText2VideoRequest, + TaskStatusVideoResponse, +) from comfy_api_nodes.util import ( - validate_image_dimensions, + ApiEndpoint, + download_url_to_image_tensor, + download_url_to_video_output, + get_number_of_images, + poll_op, + sync_op, + tensor_to_base64_string, + upload_audio_to_comfyapi, + upload_images_to_comfyapi, + upload_video_to_comfyapi, validate_image_aspect_ratio, + validate_image_dimensions, + validate_string, validate_video_dimensions, validate_video_duration, - tensor_to_base64_string, - validate_string, - upload_audio_to_comfyapi, - download_url_to_image_tensor, - upload_video_to_comfyapi, - download_url_to_video_output, - sync_op, - ApiEndpoint, - poll_op, ) -from comfy_api.input_impl import VideoFromFile -from comfy_api.latest import ComfyExtension, IO, Input KLING_API_VERSION = "v1" PATH_TEXT_TO_VIDEO = f"/proxy/kling/{KLING_API_VERSION}/videos/text2video" @@ -202,6 +210,20 @@ VOICES_CONFIG = { } +async def finish_omni_video_task(cls: type[IO.ComfyNode], response: TaskStatusVideoResponse) -> IO.NodeOutput: + if response.code: + raise RuntimeError( + f"Kling request failed. Code: {response.code}, Message: {response.message}, Data: {response.data}" + ) + final_response = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/kling/v1/videos/omni-video/{response.data.task_id}"), + response_model=TaskStatusVideoResponse, + status_extractor=lambda r: (r.data.task_status if r.data else None), + ) + return IO.NodeOutput(await download_url_to_video_output(final_response.data.task_result.videos[0].url)) + + def is_valid_camera_control_configs(configs: list[float]) -> bool: """Verifies that at least one camera control configuration is non-zero.""" return any(not math.isclose(value, 0.0) for value in configs) @@ -449,7 +471,7 @@ async def execute_video_effect( image_1: torch.Tensor, image_2: torch.Tensor | None = None, model_mode: KlingVideoGenMode | None = None, -) -> tuple[VideoFromFile, str, str]: +) -> tuple[InputImpl.VideoFromFile, str, str]: if dual_character: request_input_field = KlingDualCharacterEffectInput( model_name=model_name, @@ -736,6 +758,386 @@ class KlingTextToVideoNode(IO.ComfyNode): ) +class OmniProTextToVideoNode(IO.ComfyNode): + + @classmethod + def define_schema(cls) -> IO.Schema: + return IO.Schema( + node_id="KlingOmniProTextToVideoNode", + display_name="Kling Omni Text to Video (Pro)", + category="api node/video/Kling", + description="Use text prompts to generate videos with the latest Kling model.", + inputs=[ + IO.Combo.Input("model_name", options=["kling-video-o1"]), + IO.String.Input( + "prompt", + multiline=True, + tooltip="A text prompt describing the video content. " + "This can include both positive and negative descriptions.", + ), + IO.Combo.Input("aspect_ratio", options=["16:9", "9:16", "1:1"]), + IO.Combo.Input("duration", options=[5, 10]), + ], + outputs=[ + IO.Video.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + ) + + @classmethod + async def execute( + cls, + model_name: str, + prompt: str, + aspect_ratio: str, + duration: int, + ) -> IO.NodeOutput: + validate_string(prompt, min_length=1, max_length=2500) + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/kling/v1/videos/omni-video", method="POST"), + response_model=TaskStatusVideoResponse, + data=OmniProText2VideoRequest( + model_name=model_name, + prompt=prompt, + aspect_ratio=aspect_ratio, + duration=str(duration), + ), + ) + return await finish_omni_video_task(cls, response) + + +class OmniProFirstLastFrameNode(IO.ComfyNode): + + @classmethod + def define_schema(cls) -> IO.Schema: + return IO.Schema( + node_id="KlingOmniProFirstLastFrameNode", + display_name="Kling Omni First-Last-Frame to Video (Pro)", + category="api node/video/Kling", + description="Use a start frame, an optional end frame, or reference images with the latest Kling model.", + inputs=[ + IO.Combo.Input("model_name", options=["kling-video-o1"]), + IO.String.Input( + "prompt", + multiline=True, + tooltip="A text prompt describing the video content. " + "This can include both positive and negative descriptions.", + ), + IO.Combo.Input("duration", options=["5", "10"]), + IO.Image.Input("first_frame"), + IO.Image.Input( + "end_frame", + optional=True, + tooltip="An optional end frame for the video. " + "This cannot be used simultaneously with 'reference_images'.", + ), + IO.Image.Input( + "reference_images", + optional=True, + tooltip="Up to 6 additional reference images.", + ), + ], + outputs=[ + IO.Video.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + ) + + @classmethod + async def execute( + cls, + model_name: str, + prompt: str, + duration: int, + first_frame: Input.Image, + end_frame: Input.Image | None = None, + reference_images: Input.Image | None = None, + ) -> IO.NodeOutput: + validate_string(prompt, min_length=1, max_length=2500) + if end_frame is not None and reference_images is not None: + raise ValueError("The 'end_frame' input cannot be used simultaneously with 'reference_images'.") + validate_image_dimensions(first_frame, min_width=300, min_height=300) + validate_image_aspect_ratio(first_frame, (1, 2.5), (2.5, 1)) + image_list: list[OmniParamImage] = [ + OmniParamImage( + image_url=(await upload_images_to_comfyapi(cls, first_frame, wait_label="Uploading first frame"))[0], + type="first_frame", + ) + ] + if end_frame is not None: + validate_image_dimensions(end_frame, min_width=300, min_height=300) + validate_image_aspect_ratio(end_frame, (1, 2.5), (2.5, 1)) + image_list.append( + OmniParamImage( + image_url=(await upload_images_to_comfyapi(cls, end_frame, wait_label="Uploading end frame"))[0], + type="end_frame", + ) + ) + if reference_images is not None: + if get_number_of_images(reference_images) > 6: + raise ValueError("The maximum number of reference images allowed is 6.") + for i in reference_images: + validate_image_dimensions(i, min_width=300, min_height=300) + validate_image_aspect_ratio(i, (1, 2.5), (2.5, 1)) + for i in await upload_images_to_comfyapi(cls, reference_images, wait_label="Uploading reference frame(s)"): + image_list.append(OmniParamImage(image_url=i)) + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/kling/v1/videos/omni-video", method="POST"), + response_model=TaskStatusVideoResponse, + data=OmniProFirstLastFrameRequest( + model_name=model_name, + prompt=prompt, + duration=str(duration), + image_list=image_list, + ), + ) + return await finish_omni_video_task(cls, response) + + +class OmniProImageToVideoNode(IO.ComfyNode): + + @classmethod + def define_schema(cls) -> IO.Schema: + return IO.Schema( + node_id="KlingOmniProImageToVideoNode", + display_name="Kling Omni Image to Video (Pro)", + category="api node/video/Kling", + description="Use up to 7 reference images to generate a video with the latest Kling model.", + inputs=[ + IO.Combo.Input("model_name", options=["kling-video-o1"]), + IO.String.Input( + "prompt", + multiline=True, + tooltip="A text prompt describing the video content. " + "This can include both positive and negative descriptions.", + ), + IO.Combo.Input("aspect_ratio", options=["16:9", "9:16", "1:1"]), + IO.Int.Input("duration", default=3, min=3, max=10, display_mode=IO.NumberDisplay.slider), + IO.Image.Input( + "reference_images", + tooltip="Up to 7 reference images.", + ), + ], + outputs=[ + IO.Video.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + ) + + @classmethod + async def execute( + cls, + model_name: str, + prompt: str, + aspect_ratio: str, + duration: int, + reference_images: Input.Image, + ) -> IO.NodeOutput: + validate_string(prompt, min_length=1, max_length=2500) + if get_number_of_images(reference_images) > 7: + raise ValueError("The maximum number of reference images is 7.") + for i in reference_images: + validate_image_dimensions(i, min_width=300, min_height=300) + validate_image_aspect_ratio(i, (1, 2.5), (2.5, 1)) + image_list: list[OmniParamImage] = [] + for i in await upload_images_to_comfyapi(cls, reference_images, wait_label="Uploading reference image"): + image_list.append(OmniParamImage(image_url=i)) + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/kling/v1/videos/omni-video", method="POST"), + response_model=TaskStatusVideoResponse, + data=OmniProReferences2VideoRequest( + model_name=model_name, + prompt=prompt, + aspect_ratio=aspect_ratio, + duration=str(duration), + image_list=image_list, + ), + ) + return await finish_omni_video_task(cls, response) + + +class OmniProVideoToVideoNode(IO.ComfyNode): + + @classmethod + def define_schema(cls) -> IO.Schema: + return IO.Schema( + node_id="KlingOmniProVideoToVideoNode", + display_name="Kling Omni Video to Video (Pro)", + category="api node/video/Kling", + description="Use a video and up to 4 reference images to generate a video with the latest Kling model.", + inputs=[ + IO.Combo.Input("model_name", options=["kling-video-o1"]), + IO.String.Input( + "prompt", + multiline=True, + tooltip="A text prompt describing the video content. " + "This can include both positive and negative descriptions.", + ), + IO.Combo.Input("aspect_ratio", options=["16:9", "9:16", "1:1"]), + IO.Int.Input("duration", default=3, min=3, max=10, display_mode=IO.NumberDisplay.slider), + IO.Video.Input("reference_video", tooltip="Video to use as a reference."), + IO.Boolean.Input("keep_original_sound", default=True), + IO.Image.Input( + "reference_images", + tooltip="Up to 4 additional reference images.", + optional=True, + ), + ], + outputs=[ + IO.Video.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + ) + + @classmethod + async def execute( + cls, + model_name: str, + prompt: str, + aspect_ratio: str, + duration: int, + reference_video: Input.Video, + keep_original_sound: bool, + reference_images: Input.Image | None = None, + ) -> IO.NodeOutput: + validate_string(prompt, min_length=1, max_length=2500) + validate_video_duration(reference_video, min_duration=3.0, max_duration=10.05) + validate_video_dimensions(reference_video, min_width=720, min_height=720, max_width=2160, max_height=2160) + image_list: list[OmniParamImage] = [] + if reference_images is not None: + if get_number_of_images(reference_images) > 4: + raise ValueError("The maximum number of reference images allowed with a video input is 4.") + for i in reference_images: + validate_image_dimensions(i, min_width=300, min_height=300) + validate_image_aspect_ratio(i, (1, 2.5), (2.5, 1)) + for i in await upload_images_to_comfyapi(cls, reference_images, wait_label="Uploading reference image"): + image_list.append(OmniParamImage(image_url=i)) + video_list = [ + OmniParamVideo( + video_url=await upload_video_to_comfyapi(cls, reference_video, wait_label="Uploading reference video"), + refer_type="feature", + keep_original_sound="yes" if keep_original_sound else "no", + ) + ] + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/kling/v1/videos/omni-video", method="POST"), + response_model=TaskStatusVideoResponse, + data=OmniProReferences2VideoRequest( + model_name=model_name, + prompt=prompt, + aspect_ratio=aspect_ratio, + duration=str(duration), + image_list=image_list if image_list else None, + video_list=video_list, + ), + ) + return await finish_omni_video_task(cls, response) + + +class OmniProEditVideoNode(IO.ComfyNode): + + @classmethod + def define_schema(cls) -> IO.Schema: + return IO.Schema( + node_id="KlingOmniProEditVideoNode", + display_name="Kling Omni Edit Video (Pro)", + category="api node/video/Kling", + description="Edit an existing video with the latest model from Kling.", + inputs=[ + IO.Combo.Input("model_name", options=["kling-video-o1"]), + IO.String.Input( + "prompt", + multiline=True, + tooltip="A text prompt describing the video content. " + "This can include both positive and negative descriptions.", + ), + IO.Video.Input("video", tooltip="Video for editing. The output video length will be the same."), + IO.Boolean.Input("keep_original_sound", default=True), + IO.Image.Input( + "reference_images", + tooltip="Up to 4 additional reference images.", + optional=True, + ), + ], + outputs=[ + IO.Video.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + ) + + @classmethod + async def execute( + cls, + model_name: str, + prompt: str, + video: Input.Video, + keep_original_sound: bool, + reference_images: Input.Image | None = None, + ) -> IO.NodeOutput: + validate_string(prompt, min_length=1, max_length=2500) + validate_video_duration(video, min_duration=3.0, max_duration=10.05) + validate_video_dimensions(video, min_width=720, min_height=720, max_width=2160, max_height=2160) + image_list: list[OmniParamImage] = [] + if reference_images is not None: + if get_number_of_images(reference_images) > 4: + raise ValueError("The maximum number of reference images allowed with a video input is 4.") + for i in reference_images: + validate_image_dimensions(i, min_width=300, min_height=300) + validate_image_aspect_ratio(i, (1, 2.5), (2.5, 1)) + for i in await upload_images_to_comfyapi(cls, reference_images, wait_label="Uploading reference image"): + image_list.append(OmniParamImage(image_url=i)) + video_list = [ + OmniParamVideo( + video_url=await upload_video_to_comfyapi(cls, video, wait_label="Uploading base video"), + refer_type="base", + keep_original_sound="yes" if keep_original_sound else "no", + ) + ] + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/kling/v1/videos/omni-video", method="POST"), + response_model=TaskStatusVideoResponse, + data=OmniProReferences2VideoRequest( + model_name=model_name, + prompt=prompt, + aspect_ratio=None, + duration=None, + image_list=image_list if image_list else None, + video_list=video_list, + ), + ) + return await finish_omni_video_task(cls, response) + + class KlingCameraControlT2VNode(IO.ComfyNode): """ Kling Text to Video Camera Control Node. This node is a text to video node, but it supports controlling the camera. @@ -1162,7 +1564,10 @@ class KlingSingleImageVideoEffectNode(IO.ComfyNode): category="api node/video/Kling", description="Achieve different special effects when generating a video based on the effect_scene.", inputs=[ - IO.Image.Input("image", tooltip=" Reference Image. URL or Base64 encoded string (without data:image prefix). File size cannot exceed 10MB, resolution not less than 300*300px, aspect ratio between 1:2.5 ~ 2.5:1"), + IO.Image.Input( + "image", + tooltip=" Reference Image. URL or Base64 encoded string (without data:image prefix). File size cannot exceed 10MB, resolution not less than 300*300px, aspect ratio between 1:2.5 ~ 2.5:1", + ), IO.Combo.Input( "effect_scene", options=[i.value for i in KlingSingleImageEffectsScene], @@ -1525,6 +1930,11 @@ class KlingExtension(ComfyExtension): KlingImageGenerationNode, KlingSingleImageVideoEffectNode, KlingDualCharacterVideoEffectNode, + OmniProTextToVideoNode, + OmniProFirstLastFrameNode, + OmniProImageToVideoNode, + OmniProVideoToVideoNode, + OmniProEditVideoNode, ] diff --git a/comfy_api_nodes/util/upload_helpers.py b/comfy_api_nodes/util/upload_helpers.py index b9019841f..0532bea9a 100644 --- a/comfy_api_nodes/util/upload_helpers.py +++ b/comfy_api_nodes/util/upload_helpers.py @@ -103,6 +103,7 @@ async def upload_video_to_comfyapi( container: VideoContainer = VideoContainer.MP4, codec: VideoCodec = VideoCodec.H264, max_duration: Optional[int] = None, + wait_label: str | None = "Uploading", ) -> str: """ Uploads a single video to ComfyUI API and returns its download URL. @@ -127,7 +128,7 @@ async def upload_video_to_comfyapi( video.save_to(video_bytes_io, format=container, codec=codec) video_bytes_io.seek(0) - return await upload_file_to_comfyapi(cls, video_bytes_io, filename, upload_mime_type) + return await upload_file_to_comfyapi(cls, video_bytes_io, filename, upload_mime_type, wait_label) async def upload_file_to_comfyapi( From 30c259cac8c08ff8d015f9aff3151cb525c9b702 Mon Sep 17 00:00:00 2001 From: comfyanonymous Date: Mon, 1 Dec 2025 20:25:35 -0500 Subject: [PATCH 203/647] ComfyUI version v0.3.76 --- comfyui_version.py | 2 +- pyproject.toml | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/comfyui_version.py b/comfyui_version.py index fa4b4f4b0..4b039356e 100644 --- a/comfyui_version.py +++ b/comfyui_version.py @@ -1,3 +1,3 @@ # This file is automatically generated by the build process when version is # updated in pyproject.toml. -__version__ = "0.3.75" +__version__ = "0.3.76" diff --git a/pyproject.toml b/pyproject.toml index 9009e65fe..02b94a0ce 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -1,6 +1,6 @@ [project] name = "ComfyUI" -version = "0.3.75" +version = "0.3.76" readme = "README.md" license = { file = "LICENSE" } requires-python = ">=3.9" From 878db3a727c1c6049bc1c4959cdfabc35eaf3d56 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 1 Dec 2025 17:56:17 -0800 Subject: [PATCH 204/647] Implement the Ovis image model. (#11030) --- comfy/ldm/chroma/model.py | 3 +- comfy/ldm/flux/layers.py | 68 +++++++++++++++++++++-------------- comfy/ldm/flux/model.py | 21 ++++++++--- comfy/model_detection.py | 10 +++++- comfy/sd.py | 13 +++++-- comfy/text_encoders/llama.py | 31 ++++++++++++++++ comfy/text_encoders/ovis.py | 69 ++++++++++++++++++++++++++++++++++++ nodes.py | 2 +- 8 files changed, 182 insertions(+), 35 deletions(-) create mode 100644 comfy/text_encoders/ovis.py diff --git a/comfy/ldm/chroma/model.py b/comfy/ldm/chroma/model.py index a72f8cc47..2e8ef0687 100644 --- a/comfy/ldm/chroma/model.py +++ b/comfy/ldm/chroma/model.py @@ -40,7 +40,8 @@ class ChromaParams: out_dim: int hidden_dim: int n_layers: int - + txt_ids_dims: list + vec_in_dim: int diff --git a/comfy/ldm/flux/layers.py b/comfy/ldm/flux/layers.py index 2472ab79c..60f2bdae2 100644 --- a/comfy/ldm/flux/layers.py +++ b/comfy/ldm/flux/layers.py @@ -57,6 +57,35 @@ class MLPEmbedder(nn.Module): def forward(self, x: Tensor) -> Tensor: return self.out_layer(self.silu(self.in_layer(x))) +class YakMLP(nn.Module): + def __init__(self, hidden_size: int, intermediate_size: int, dtype=None, device=None, operations=None): + super().__init__() + self.hidden_size = hidden_size + self.intermediate_size = intermediate_size + self.gate_proj = operations.Linear(self.hidden_size, self.intermediate_size, bias=True, dtype=dtype, device=device) + self.up_proj = operations.Linear(self.hidden_size, self.intermediate_size, bias=True, dtype=dtype, device=device) + self.down_proj = operations.Linear(self.intermediate_size, self.hidden_size, bias=True, dtype=dtype, device=device) + self.act_fn = nn.SiLU() + + def forward(self, x: Tensor) -> Tensor: + down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x)) + return down_proj + +def build_mlp(hidden_size, mlp_hidden_dim, mlp_silu_act=False, yak_mlp=False, dtype=None, device=None, operations=None): + if yak_mlp: + return YakMLP(hidden_size, mlp_hidden_dim, dtype=dtype, device=device, operations=operations) + if mlp_silu_act: + return nn.Sequential( + operations.Linear(hidden_size, mlp_hidden_dim * 2, bias=False, dtype=dtype, device=device), + SiLUActivation(), + operations.Linear(mlp_hidden_dim, hidden_size, bias=False, dtype=dtype, device=device), + ) + else: + return nn.Sequential( + operations.Linear(hidden_size, mlp_hidden_dim, bias=True, dtype=dtype, device=device), + nn.GELU(approximate="tanh"), + operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device), + ) class RMSNorm(torch.nn.Module): def __init__(self, dim: int, dtype=None, device=None, operations=None): @@ -140,7 +169,7 @@ class SiLUActivation(nn.Module): class DoubleStreamBlock(nn.Module): - def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False, flipped_img_txt=False, modulation=True, mlp_silu_act=False, proj_bias=True, dtype=None, device=None, operations=None): + def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False, flipped_img_txt=False, modulation=True, mlp_silu_act=False, proj_bias=True, yak_mlp=False, dtype=None, device=None, operations=None): super().__init__() mlp_hidden_dim = int(hidden_size * mlp_ratio) @@ -156,18 +185,7 @@ class DoubleStreamBlock(nn.Module): self.img_norm2 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) - if mlp_silu_act: - self.img_mlp = nn.Sequential( - operations.Linear(hidden_size, mlp_hidden_dim * 2, bias=False, dtype=dtype, device=device), - SiLUActivation(), - operations.Linear(mlp_hidden_dim, hidden_size, bias=False, dtype=dtype, device=device), - ) - else: - self.img_mlp = nn.Sequential( - operations.Linear(hidden_size, mlp_hidden_dim, bias=True, dtype=dtype, device=device), - nn.GELU(approximate="tanh"), - operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device), - ) + self.img_mlp = build_mlp(hidden_size, mlp_hidden_dim, mlp_silu_act=mlp_silu_act, yak_mlp=yak_mlp, dtype=dtype, device=device, operations=operations) if self.modulation: self.txt_mod = Modulation(hidden_size, double=True, dtype=dtype, device=device, operations=operations) @@ -177,18 +195,7 @@ class DoubleStreamBlock(nn.Module): self.txt_norm2 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) - if mlp_silu_act: - self.txt_mlp = nn.Sequential( - operations.Linear(hidden_size, mlp_hidden_dim * 2, bias=False, dtype=dtype, device=device), - SiLUActivation(), - operations.Linear(mlp_hidden_dim, hidden_size, bias=False, dtype=dtype, device=device), - ) - else: - self.txt_mlp = nn.Sequential( - operations.Linear(hidden_size, mlp_hidden_dim, bias=True, dtype=dtype, device=device), - nn.GELU(approximate="tanh"), - operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device), - ) + self.txt_mlp = build_mlp(hidden_size, mlp_hidden_dim, mlp_silu_act=mlp_silu_act, yak_mlp=yak_mlp, dtype=dtype, device=device, operations=operations) self.flipped_img_txt = flipped_img_txt @@ -275,6 +282,7 @@ class SingleStreamBlock(nn.Module): modulation=True, mlp_silu_act=False, bias=True, + yak_mlp=False, dtype=None, device=None, operations=None @@ -288,12 +296,17 @@ class SingleStreamBlock(nn.Module): self.mlp_hidden_dim = int(hidden_size * mlp_ratio) self.mlp_hidden_dim_first = self.mlp_hidden_dim + self.yak_mlp = yak_mlp if mlp_silu_act: self.mlp_hidden_dim_first = int(hidden_size * mlp_ratio * 2) self.mlp_act = SiLUActivation() else: self.mlp_act = nn.GELU(approximate="tanh") + if self.yak_mlp: + self.mlp_hidden_dim_first *= 2 + self.mlp_act = nn.SiLU() + # qkv and mlp_in self.linear1 = operations.Linear(hidden_size, hidden_size * 3 + self.mlp_hidden_dim_first, bias=bias, dtype=dtype, device=device) # proj and mlp_out @@ -325,7 +338,10 @@ class SingleStreamBlock(nn.Module): attn = attention(q, k, v, pe=pe, mask=attn_mask, transformer_options=transformer_options) del q, k, v # compute activation in mlp stream, cat again and run second linear layer - mlp = self.mlp_act(mlp) + if self.yak_mlp: + mlp = self.mlp_act(mlp[..., self.mlp_hidden_dim_first // 2:]) * mlp[..., :self.mlp_hidden_dim_first // 2] + else: + mlp = self.mlp_act(mlp) output = self.linear2(torch.cat((attn, mlp), 2)) x += apply_mod(output, mod.gate, None, modulation_dims) if x.dtype == torch.float16: diff --git a/comfy/ldm/flux/model.py b/comfy/ldm/flux/model.py index d5674dea6..f40c2a7a9 100644 --- a/comfy/ldm/flux/model.py +++ b/comfy/ldm/flux/model.py @@ -15,7 +15,8 @@ from .layers import ( MLPEmbedder, SingleStreamBlock, timestep_embedding, - Modulation + Modulation, + RMSNorm ) @dataclass @@ -34,11 +35,14 @@ class FluxParams: patch_size: int qkv_bias: bool guidance_embed: bool + txt_ids_dims: list global_modulation: bool = False mlp_silu_act: bool = False ops_bias: bool = True default_ref_method: str = "offset" ref_index_scale: float = 1.0 + yak_mlp: bool = False + txt_norm: bool = False class Flux(nn.Module): @@ -76,6 +80,11 @@ class Flux(nn.Module): ) self.txt_in = operations.Linear(params.context_in_dim, self.hidden_size, bias=params.ops_bias, dtype=dtype, device=device) + if params.txt_norm: + self.txt_norm = RMSNorm(params.context_in_dim, dtype=dtype, device=device, operations=operations) + else: + self.txt_norm = None + self.double_blocks = nn.ModuleList( [ DoubleStreamBlock( @@ -86,6 +95,7 @@ class Flux(nn.Module): modulation=params.global_modulation is False, mlp_silu_act=params.mlp_silu_act, proj_bias=params.ops_bias, + yak_mlp=params.yak_mlp, dtype=dtype, device=device, operations=operations ) for _ in range(params.depth) @@ -94,7 +104,7 @@ class Flux(nn.Module): self.single_blocks = nn.ModuleList( [ - SingleStreamBlock(self.hidden_size, self.num_heads, mlp_ratio=params.mlp_ratio, modulation=params.global_modulation is False, mlp_silu_act=params.mlp_silu_act, bias=params.ops_bias, dtype=dtype, device=device, operations=operations) + SingleStreamBlock(self.hidden_size, self.num_heads, mlp_ratio=params.mlp_ratio, modulation=params.global_modulation is False, mlp_silu_act=params.mlp_silu_act, bias=params.ops_bias, yak_mlp=params.yak_mlp, dtype=dtype, device=device, operations=operations) for _ in range(params.depth_single_blocks) ] ) @@ -150,6 +160,8 @@ class Flux(nn.Module): y = torch.zeros((img.shape[0], self.params.vec_in_dim), device=img.device, dtype=img.dtype) vec = vec + self.vector_in(y[:, :self.params.vec_in_dim]) + if self.txt_norm is not None: + txt = self.txt_norm(txt) txt = self.txt_in(txt) vec_orig = vec @@ -332,8 +344,9 @@ class Flux(nn.Module): txt_ids = torch.zeros((bs, context.shape[1], len(self.params.axes_dim)), device=x.device, dtype=torch.float32) - if len(self.params.axes_dim) == 4: # Flux 2 - txt_ids[:, :, 3] = torch.linspace(0, context.shape[1] - 1, steps=context.shape[1], device=x.device, dtype=torch.float32) + if len(self.params.txt_ids_dims) > 0: + for i in self.params.txt_ids_dims: + txt_ids[:, :, i] = torch.linspace(0, context.shape[1] - 1, steps=context.shape[1], device=x.device, dtype=torch.float32) out = self.forward_orig(img, img_ids, context, txt_ids, timestep, y, guidance, control, transformer_options, attn_mask=kwargs.get("attention_mask", None)) out = out[:, :img_tokens] diff --git a/comfy/model_detection.py b/comfy/model_detection.py index 7afe4a798..7d0517e61 100644 --- a/comfy/model_detection.py +++ b/comfy/model_detection.py @@ -208,12 +208,12 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): dit_config["theta"] = 2000 dit_config["out_channels"] = 128 dit_config["global_modulation"] = True - dit_config["vec_in_dim"] = None dit_config["mlp_silu_act"] = True dit_config["qkv_bias"] = False dit_config["ops_bias"] = False dit_config["default_ref_method"] = "index" dit_config["ref_index_scale"] = 10.0 + dit_config["txt_ids_dims"] = [3] patch_size = 1 else: dit_config["image_model"] = "flux" @@ -223,6 +223,7 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): dit_config["theta"] = 10000 dit_config["out_channels"] = 16 dit_config["qkv_bias"] = True + dit_config["txt_ids_dims"] = [] patch_size = 2 dit_config["in_channels"] = 16 @@ -245,6 +246,8 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): vec_in_key = '{}vector_in.in_layer.weight'.format(key_prefix) if vec_in_key in state_dict_keys: dit_config["vec_in_dim"] = state_dict[vec_in_key].shape[1] + else: + dit_config["vec_in_dim"] = None dit_config["depth"] = count_blocks(state_dict_keys, '{}double_blocks.'.format(key_prefix) + '{}.') dit_config["depth_single_blocks"] = count_blocks(state_dict_keys, '{}single_blocks.'.format(key_prefix) + '{}.') @@ -270,6 +273,11 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): dit_config["nerf_embedder_dtype"] = torch.float32 else: dit_config["guidance_embed"] = "{}guidance_in.in_layer.weight".format(key_prefix) in state_dict_keys + dit_config["yak_mlp"] = '{}double_blocks.0.img_mlp.gate_proj.weight'.format(key_prefix) in state_dict_keys + dit_config["txt_norm"] = "{}txt_norm.scale".format(key_prefix) in state_dict_keys + if dit_config["yak_mlp"] and dit_config["txt_norm"]: # Ovis model + dit_config["txt_ids_dims"] = [1, 2] + return dit_config if '{}t5_yproj.weight'.format(key_prefix) in state_dict_keys: #Genmo mochi preview diff --git a/comfy/sd.py b/comfy/sd.py index 9eeb0c45a..f9e5efab5 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -53,6 +53,7 @@ import comfy.text_encoders.omnigen2 import comfy.text_encoders.qwen_image import comfy.text_encoders.hunyuan_image import comfy.text_encoders.z_image +import comfy.text_encoders.ovis import comfy.model_patcher import comfy.lora @@ -956,6 +957,7 @@ class CLIPType(Enum): QWEN_IMAGE = 18 HUNYUAN_IMAGE = 19 HUNYUAN_VIDEO_15 = 20 + OVIS = 21 def load_clip(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DIFFUSION, model_options={}): @@ -987,6 +989,7 @@ class TEModel(Enum): MISTRAL3_24B = 14 MISTRAL3_24B_PRUNED_FLUX2 = 15 QWEN3_4B = 16 + QWEN3_2B = 17 def detect_te_model(sd): @@ -1020,9 +1023,12 @@ def detect_te_model(sd): if weight.shape[0] == 512: return TEModel.QWEN25_7B if "model.layers.0.post_attention_layernorm.weight" in sd: - if 'model.layers.0.self_attn.q_norm.weight' in sd: - return TEModel.QWEN3_4B weight = sd['model.layers.0.post_attention_layernorm.weight'] + if 'model.layers.0.self_attn.q_norm.weight' in sd: + if weight.shape[0] == 2560: + return TEModel.QWEN3_4B + elif weight.shape[0] == 2048: + return TEModel.QWEN3_2B if weight.shape[0] == 5120: if "model.layers.39.post_attention_layernorm.weight" in sd: return TEModel.MISTRAL3_24B @@ -1150,6 +1156,9 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip elif te_model == TEModel.QWEN3_4B: clip_target.clip = comfy.text_encoders.z_image.te(**llama_detect(clip_data)) clip_target.tokenizer = comfy.text_encoders.z_image.ZImageTokenizer + elif te_model == TEModel.QWEN3_2B: + clip_target.clip = comfy.text_encoders.ovis.te(**llama_detect(clip_data)) + clip_target.tokenizer = comfy.text_encoders.ovis.OvisTokenizer else: # clip_l if clip_type == CLIPType.SD3: diff --git a/comfy/text_encoders/llama.py b/comfy/text_encoders/llama.py index cd4b5f76c..0d07ac8c6 100644 --- a/comfy/text_encoders/llama.py +++ b/comfy/text_encoders/llama.py @@ -100,6 +100,28 @@ class Qwen3_4BConfig: rope_scale = None final_norm: bool = True +@dataclass +class Ovis25_2BConfig: + vocab_size: int = 151936 + hidden_size: int = 2048 + intermediate_size: int = 6144 + num_hidden_layers: int = 28 + num_attention_heads: int = 16 + num_key_value_heads: int = 8 + max_position_embeddings: int = 40960 + rms_norm_eps: float = 1e-6 + rope_theta: float = 1000000.0 + transformer_type: str = "llama" + head_dim = 128 + rms_norm_add = False + mlp_activation = "silu" + qkv_bias = False + rope_dims = None + q_norm = "gemma3" + k_norm = "gemma3" + rope_scale = None + final_norm: bool = True + @dataclass class Qwen25_7BVLI_Config: vocab_size: int = 152064 @@ -542,6 +564,15 @@ class Qwen3_4B(BaseLlama, torch.nn.Module): self.model = Llama2_(config, device=device, dtype=dtype, ops=operations) self.dtype = dtype +class Ovis25_2B(BaseLlama, torch.nn.Module): + def __init__(self, config_dict, dtype, device, operations): + super().__init__() + config = Ovis25_2BConfig(**config_dict) + self.num_layers = config.num_hidden_layers + + self.model = Llama2_(config, device=device, dtype=dtype, ops=operations) + self.dtype = dtype + class Qwen25_7BVLI(BaseLlama, torch.nn.Module): def __init__(self, config_dict, dtype, device, operations): super().__init__() diff --git a/comfy/text_encoders/ovis.py b/comfy/text_encoders/ovis.py new file mode 100644 index 000000000..81c9bd51c --- /dev/null +++ b/comfy/text_encoders/ovis.py @@ -0,0 +1,69 @@ +from transformers import Qwen2Tokenizer +import comfy.text_encoders.llama +from comfy import sd1_clip +import os +import torch +import numbers + +class Qwen3Tokenizer(sd1_clip.SDTokenizer): + def __init__(self, embedding_directory=None, tokenizer_data={}): + tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "qwen25_tokenizer") + super().__init__(tokenizer_path, pad_with_end=False, embedding_size=2048, embedding_key='qwen3_2b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=284, pad_token=151643, tokenizer_data=tokenizer_data) + + +class OvisTokenizer(sd1_clip.SD1Tokenizer): + def __init__(self, embedding_directory=None, tokenizer_data={}): + super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, name="qwen3_2b", tokenizer=Qwen3Tokenizer) + self.llama_template = "<|im_start|>user\nDescribe the image by detailing the color, quantity, text, shape, size, texture, spatial relationships of the objects and background: {}<|im_end|>\n<|im_start|>assistant\n\n\n\n\n" + + def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None, **kwargs): + if llama_template is None: + llama_text = self.llama_template.format(text) + else: + llama_text = llama_template.format(text) + + tokens = super().tokenize_with_weights(llama_text, return_word_ids=return_word_ids, disable_weights=True, **kwargs) + return tokens + +class Ovis25_2BModel(sd1_clip.SDClipModel): + def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, attention_mask=True, model_options={}): + super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"pad": 151643}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Ovis25_2B, enable_attention_masks=attention_mask, return_attention_masks=False, zero_out_masked=True, model_options=model_options) + + +class OvisTEModel(sd1_clip.SD1ClipModel): + def __init__(self, device="cpu", dtype=None, model_options={}): + super().__init__(device=device, dtype=dtype, name="qwen3_2b", clip_model=Ovis25_2BModel, model_options=model_options) + + def encode_token_weights(self, token_weight_pairs, template_end=-1): + out, pooled = super().encode_token_weights(token_weight_pairs) + tok_pairs = token_weight_pairs["qwen3_2b"][0] + count_im_start = 0 + if template_end == -1: + for i, v in enumerate(tok_pairs): + elem = v[0] + if not torch.is_tensor(elem): + if isinstance(elem, numbers.Integral): + if elem == 4004 and count_im_start < 1: + template_end = i + count_im_start += 1 + + if out.shape[1] > (template_end + 1): + if tok_pairs[template_end + 1][0] == 25: + template_end += 1 + + out = out[:, template_end:] + return out, pooled, {} + + +def te(dtype_llama=None, llama_scaled_fp8=None, llama_quantization_metadata=None): + class OvisTEModel_(OvisTEModel): + def __init__(self, device="cpu", dtype=None, model_options={}): + if llama_scaled_fp8 is not None and "scaled_fp8" not in model_options: + model_options = model_options.copy() + model_options["scaled_fp8"] = llama_scaled_fp8 + if dtype_llama is not None: + dtype = dtype_llama + if llama_quantization_metadata is not None: + model_options["quantization_metadata"] = llama_quantization_metadata + super().__init__(device=device, dtype=dtype, model_options=model_options) + return OvisTEModel_ diff --git a/nodes.py b/nodes.py index 495dec806..d5e5dc228 100644 --- a/nodes.py +++ b/nodes.py @@ -939,7 +939,7 @@ class CLIPLoader: @classmethod def INPUT_TYPES(s): return {"required": { "clip_name": (folder_paths.get_filename_list("text_encoders"), ), - "type": (["stable_diffusion", "stable_cascade", "sd3", "stable_audio", "mochi", "ltxv", "pixart", "cosmos", "lumina2", "wan", "hidream", "chroma", "ace", "omnigen2", "qwen_image", "hunyuan_image", "flux2"], ), + "type": (["stable_diffusion", "stable_cascade", "sd3", "stable_audio", "mochi", "ltxv", "pixart", "cosmos", "lumina2", "wan", "hidream", "chroma", "ace", "omnigen2", "qwen_image", "hunyuan_image", "flux2", "ovis"], ), }, "optional": { "device": (["default", "cpu"], {"advanced": True}), From c55dc857d5da5af203caf720ed7056047d382544 Mon Sep 17 00:00:00 2001 From: Christian Byrne Date: Mon, 1 Dec 2025 17:56:38 -0800 Subject: [PATCH 205/647] bump comfyui-frontend-package to 1.33.10 (#11028) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 045b2ac54..f98848e20 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,4 +1,4 @@ -comfyui-frontend-package==1.32.10 +comfyui-frontend-package==1.33.10 comfyui-workflow-templates==0.7.25 comfyui-embedded-docs==0.3.1 torch From b4a20acc54b0b94dc05a1bd09dc0b54dd12203f1 Mon Sep 17 00:00:00 2001 From: "Dr.Lt.Data" <128333288+ltdrdata@users.noreply.github.com> Date: Tue, 2 Dec 2025 12:32:52 +0900 Subject: [PATCH 206/647] feat: Support ComfyUI-Manager for pip version (#7555) --- comfy/cli_args.py | 7 +++++++ comfy_api/feature_flags.py | 1 + main.py | 30 ++++++++++++++++++++++++++++++ manager_requirements.txt | 1 + nodes.py | 9 +++++++++ server.py | 8 +++++++- 6 files changed, 55 insertions(+), 1 deletion(-) create mode 100644 manager_requirements.txt diff --git a/comfy/cli_args.py b/comfy/cli_args.py index 5f0dfaa10..209fc185b 100644 --- a/comfy/cli_args.py +++ b/comfy/cli_args.py @@ -121,6 +121,12 @@ upcast.add_argument("--force-upcast-attention", action="store_true", help="Force upcast.add_argument("--dont-upcast-attention", action="store_true", help="Disable all upcasting of attention. Should be unnecessary except for debugging.") +parser.add_argument("--enable-manager", action="store_true", help="Enable the ComfyUI-Manager feature.") +manager_group = parser.add_mutually_exclusive_group() +manager_group.add_argument("--disable-manager-ui", action="store_true", help="Disables only the ComfyUI-Manager UI and endpoints. Scheduled installations and similar background tasks will still operate.") +manager_group.add_argument("--enable-manager-legacy-ui", action="store_true", help="Enables the legacy UI of ComfyUI-Manager") + + vram_group = parser.add_mutually_exclusive_group() vram_group.add_argument("--gpu-only", action="store_true", help="Store and run everything (text encoders/CLIP models, etc... on the GPU).") vram_group.add_argument("--highvram", action="store_true", help="By default models will be unloaded to CPU memory after being used. This option keeps them in GPU memory.") @@ -168,6 +174,7 @@ parser.add_argument("--multi-user", action="store_true", help="Enables per-user parser.add_argument("--verbose", default='INFO', const='DEBUG', nargs="?", choices=['DEBUG', 'INFO', 'WARNING', 'ERROR', 'CRITICAL'], help='Set the logging level') parser.add_argument("--log-stdout", action="store_true", help="Send normal process output to stdout instead of stderr (default).") + # The default built-in provider hosted under web/ DEFAULT_VERSION_STRING = "comfyanonymous/ComfyUI@latest" diff --git a/comfy_api/feature_flags.py b/comfy_api/feature_flags.py index 0d4389a6e..bfb77eb5f 100644 --- a/comfy_api/feature_flags.py +++ b/comfy_api/feature_flags.py @@ -13,6 +13,7 @@ from comfy.cli_args import args SERVER_FEATURE_FLAGS: Dict[str, Any] = { "supports_preview_metadata": True, "max_upload_size": args.max_upload_size * 1024 * 1024, # Convert MB to bytes + "extension": {"manager": {"supports_v4": True}}, } diff --git a/main.py b/main.py index e1b0f1620..0cd815d9e 100644 --- a/main.py +++ b/main.py @@ -15,6 +15,7 @@ from comfy_execution.progress import get_progress_state from comfy_execution.utils import get_executing_context from comfy_api import feature_flags + if __name__ == "__main__": #NOTE: These do not do anything on core ComfyUI, they are for custom nodes. os.environ['HF_HUB_DISABLE_TELEMETRY'] = '1' @@ -22,6 +23,23 @@ if __name__ == "__main__": setup_logger(log_level=args.verbose, use_stdout=args.log_stdout) + +def handle_comfyui_manager_unavailable(): + if not args.windows_standalone_build: + logging.warning(f"\n\nYou appear to be running comfyui-manager from source, this is not recommended. Please install comfyui-manager using the following command:\ncommand:\n\t{sys.executable} -m pip install --pre comfyui_manager\n") + args.enable_manager = False + + +if args.enable_manager: + if importlib.util.find_spec("comfyui_manager"): + import comfyui_manager + + if not comfyui_manager.__file__ or not comfyui_manager.__file__.endswith('__init__.py'): + handle_comfyui_manager_unavailable() + else: + handle_comfyui_manager_unavailable() + + def apply_custom_paths(): # extra model paths extra_model_paths_config_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "extra_model_paths.yaml") @@ -79,6 +97,11 @@ def execute_prestartup_script(): for possible_module in possible_modules: module_path = os.path.join(custom_node_path, possible_module) + + if args.enable_manager: + if comfyui_manager.should_be_disabled(module_path): + continue + if os.path.isfile(module_path) or module_path.endswith(".disabled") or module_path == "__pycache__": continue @@ -101,6 +124,10 @@ def execute_prestartup_script(): logging.info("") apply_custom_paths() + +if args.enable_manager: + comfyui_manager.prestartup() + execute_prestartup_script() @@ -323,6 +350,9 @@ def start_comfyui(asyncio_loop=None): asyncio.set_event_loop(asyncio_loop) prompt_server = server.PromptServer(asyncio_loop) + if args.enable_manager and not args.disable_manager_ui: + comfyui_manager.start() + hook_breaker_ac10a0.save_functions() asyncio_loop.run_until_complete(nodes.init_extra_nodes( init_custom_nodes=(not args.disable_all_custom_nodes) or len(args.whitelist_custom_nodes) > 0, diff --git a/manager_requirements.txt b/manager_requirements.txt new file mode 100644 index 000000000..52cc5389c --- /dev/null +++ b/manager_requirements.txt @@ -0,0 +1 @@ +comfyui_manager==4.0.3b3 diff --git a/nodes.py b/nodes.py index d5e5dc228..4c910a34b 100644 --- a/nodes.py +++ b/nodes.py @@ -43,6 +43,9 @@ import folder_paths import latent_preview import node_helpers +if args.enable_manager: + import comfyui_manager + def before_node_execution(): comfy.model_management.throw_exception_if_processing_interrupted() @@ -2243,6 +2246,12 @@ async def init_external_custom_nodes(): if args.disable_all_custom_nodes and possible_module not in args.whitelist_custom_nodes: logging.info(f"Skipping {possible_module} due to disable_all_custom_nodes and whitelist_custom_nodes") continue + + if args.enable_manager: + if comfyui_manager.should_be_disabled(module_path): + logging.info(f"Blocked by policy: {module_path}") + continue + time_before = time.perf_counter() success = await load_custom_node(module_path, base_node_names, module_parent="custom_nodes") node_import_times.append((time.perf_counter() - time_before, module_path, success)) diff --git a/server.py b/server.py index fca5050bd..e3bd056d9 100644 --- a/server.py +++ b/server.py @@ -44,6 +44,9 @@ from protocol import BinaryEventTypes # Import cache control middleware from middleware.cache_middleware import cache_control +if args.enable_manager: + import comfyui_manager + async def send_socket_catch_exception(function, message): try: await function(message) @@ -212,6 +215,9 @@ class PromptServer(): if args.disable_api_nodes: middlewares.append(create_block_external_middleware()) + if args.enable_manager: + middlewares.append(comfyui_manager.create_middleware()) + max_upload_size = round(args.max_upload_size * 1024 * 1024) self.app = web.Application(client_max_size=max_upload_size, middlewares=middlewares) self.sockets = dict() @@ -599,7 +605,7 @@ class PromptServer(): system_stats = { "system": { - "os": os.name, + "os": sys.platform, "ram_total": ram_total, "ram_free": ram_free, "comfyui_version": __version__, From a17cf1c3871ad582c85c2bb6fddb63ec9c6df0ce Mon Sep 17 00:00:00 2001 From: Yoland Yan <4950057+yoland68@users.noreply.github.com> Date: Mon, 1 Dec 2025 19:40:44 -0800 Subject: [PATCH 207/647] Add @guill as a code owner (#11031) --- CODEOWNERS | 1 + 1 file changed, 1 insertion(+) diff --git a/CODEOWNERS b/CODEOWNERS index b7aca9b26..51acc4986 100644 --- a/CODEOWNERS +++ b/CODEOWNERS @@ -1,3 +1,4 @@ # Admins * @comfyanonymous * @kosinkadink +* @guill From 44baa0b7f32dd0c2ff0a9898aeb6c7929d855cd3 Mon Sep 17 00:00:00 2001 From: Jedrzej Kosinski Date: Tue, 2 Dec 2025 11:46:29 -0800 Subject: [PATCH 208/647] Fix CODEOWNERS formatting to have all on the same line, otherwise only last line applies (#11053) --- CODEOWNERS | 4 +--- 1 file changed, 1 insertion(+), 3 deletions(-) diff --git a/CODEOWNERS b/CODEOWNERS index 51acc4986..4d5448636 100644 --- a/CODEOWNERS +++ b/CODEOWNERS @@ -1,4 +1,2 @@ # Admins -* @comfyanonymous -* @kosinkadink -* @guill +* @comfyanonymous @kosinkadink @guill From 33d6aec3b70bc6f3e5bba26c85bd8f3bb1380d08 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Tue, 2 Dec 2025 21:50:13 +0200 Subject: [PATCH 209/647] add check for the format arg type in VideoFromComponents.save_to function (#11046) * add check for the format var type in VideoFromComponents.save_to function * convert "format" to VideoContainer enum --- comfy_api/latest/_input_impl/video_types.py | 2 +- comfy_extras/nodes_video.py | 4 ++-- 2 files changed, 3 insertions(+), 3 deletions(-) diff --git a/comfy_api/latest/_input_impl/video_types.py b/comfy_api/latest/_input_impl/video_types.py index 7231bf13c..a4cd3737d 100644 --- a/comfy_api/latest/_input_impl/video_types.py +++ b/comfy_api/latest/_input_impl/video_types.py @@ -337,7 +337,7 @@ class VideoFromComponents(VideoInput): if codec != VideoCodec.AUTO and codec != VideoCodec.H264: raise ValueError("Only H264 codec is supported for now") extra_kwargs = {} - if format != VideoContainer.AUTO: + if isinstance(format, VideoContainer) and format != VideoContainer.AUTO: extra_kwargs["format"] = format.value with av.open(path, mode='w', options={'movflags': 'use_metadata_tags'}, **extra_kwargs) as output: # Add metadata before writing any streams diff --git a/comfy_extras/nodes_video.py b/comfy_extras/nodes_video.py index 69fabb12e..6cf6e39bf 100644 --- a/comfy_extras/nodes_video.py +++ b/comfy_extras/nodes_video.py @@ -88,7 +88,7 @@ class SaveVideo(io.ComfyNode): ) @classmethod - def execute(cls, video: VideoInput, filename_prefix, format, codec) -> io.NodeOutput: + def execute(cls, video: VideoInput, filename_prefix, format: str, codec) -> io.NodeOutput: width, height = video.get_dimensions() full_output_folder, filename, counter, subfolder, filename_prefix = folder_paths.get_save_image_path( filename_prefix, @@ -108,7 +108,7 @@ class SaveVideo(io.ComfyNode): file = f"{filename}_{counter:05}_.{VideoContainer.get_extension(format)}" video.save_to( os.path.join(full_output_folder, file), - format=format, + format=VideoContainer(format), codec=codec, metadata=saved_metadata ) From daaceac769a1355ab975758ede064317ea7514b4 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 2 Dec 2025 14:11:58 -0800 Subject: [PATCH 210/647] Hack to make zimage work in fp16. (#11057) --- comfy/ldm/lumina/model.py | 18 +++++++++++------- comfy/supported_models.py | 2 ++ 2 files changed, 13 insertions(+), 7 deletions(-) diff --git a/comfy/ldm/lumina/model.py b/comfy/ldm/lumina/model.py index 7d7e9112c..070b5da09 100644 --- a/comfy/ldm/lumina/model.py +++ b/comfy/ldm/lumina/model.py @@ -22,6 +22,10 @@ def modulate(x, scale): # Core NextDiT Model # ############################################################################# +def clamp_fp16(x): + if x.dtype == torch.float16: + return torch.nan_to_num(x, nan=0.0, posinf=65504, neginf=-65504) + return x class JointAttention(nn.Module): """Multi-head attention module.""" @@ -169,7 +173,7 @@ class FeedForward(nn.Module): # @torch.compile def _forward_silu_gating(self, x1, x3): - return F.silu(x1) * x3 + return clamp_fp16(F.silu(x1) * x3) def forward(self, x): return self.w2(self._forward_silu_gating(self.w1(x), self.w3(x))) @@ -273,27 +277,27 @@ class JointTransformerBlock(nn.Module): scale_msa, gate_msa, scale_mlp, gate_mlp = self.adaLN_modulation(adaln_input).chunk(4, dim=1) x = x + gate_msa.unsqueeze(1).tanh() * self.attention_norm2( - self.attention( + clamp_fp16(self.attention( modulate(self.attention_norm1(x), scale_msa), x_mask, freqs_cis, transformer_options=transformer_options, - ) + )) ) x = x + gate_mlp.unsqueeze(1).tanh() * self.ffn_norm2( - self.feed_forward( + clamp_fp16(self.feed_forward( modulate(self.ffn_norm1(x), scale_mlp), - ) + )) ) else: assert adaln_input is None x = x + self.attention_norm2( - self.attention( + clamp_fp16(self.attention( self.attention_norm1(x), x_mask, freqs_cis, transformer_options=transformer_options, - ) + )) ) x = x + self.ffn_norm2( self.feed_forward( diff --git a/comfy/supported_models.py b/comfy/supported_models.py index af8120400..afd97160b 100644 --- a/comfy/supported_models.py +++ b/comfy/supported_models.py @@ -1027,6 +1027,8 @@ class ZImage(Lumina2): memory_usage_factor = 1.7 + supported_inference_dtypes = [torch.bfloat16, torch.float16, torch.float32] + def clip_target(self, state_dict={}): pref = self.text_encoder_key_prefix[0] hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen3_4b.transformer.".format(pref)) From 277237ccc1499bac7fcd221a666dfe7a32ac4206 Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Wed, 3 Dec 2025 08:24:19 +1000 Subject: [PATCH 211/647] attention: use flag based OOM fallback (#11038) Exception ref all local variables for the lifetime of exception context. Just set a flag and then if to dump the exception before falling back. --- comfy/ldm/modules/attention.py | 3 +++ comfy/ldm/modules/diffusionmodules/model.py | 3 +++ 2 files changed, 6 insertions(+) diff --git a/comfy/ldm/modules/attention.py b/comfy/ldm/modules/attention.py index 7437e0567..a8800ded0 100644 --- a/comfy/ldm/modules/attention.py +++ b/comfy/ldm/modules/attention.py @@ -517,6 +517,7 @@ def attention_pytorch(q, k, v, heads, mask=None, attn_precision=None, skip_resha @wrap_attn def attention_sage(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False, **kwargs): + exception_fallback = False if skip_reshape: b, _, _, dim_head = q.shape tensor_layout = "HND" @@ -541,6 +542,8 @@ def attention_sage(q, k, v, heads, mask=None, attn_precision=None, skip_reshape= out = sageattn(q, k, v, attn_mask=mask, is_causal=False, tensor_layout=tensor_layout) except Exception as e: logging.error("Error running sage attention: {}, using pytorch attention instead.".format(e)) + exception_fallback = True + if exception_fallback: if tensor_layout == "NHD": q, k, v = map( lambda t: t.transpose(1, 2), diff --git a/comfy/ldm/modules/diffusionmodules/model.py b/comfy/ldm/modules/diffusionmodules/model.py index 4245eedca..de1e01cc8 100644 --- a/comfy/ldm/modules/diffusionmodules/model.py +++ b/comfy/ldm/modules/diffusionmodules/model.py @@ -279,6 +279,7 @@ def pytorch_attention(q, k, v): orig_shape = q.shape B = orig_shape[0] C = orig_shape[1] + oom_fallback = False q, k, v = map( lambda t: t.view(B, 1, C, -1).transpose(2, 3).contiguous(), (q, k, v), @@ -289,6 +290,8 @@ def pytorch_attention(q, k, v): out = out.transpose(2, 3).reshape(orig_shape) except model_management.OOM_EXCEPTION: logging.warning("scaled_dot_product_attention OOMed: switched to slice attention") + oom_fallback = True + if oom_fallback: out = slice_attention(q.view(B, -1, C), k.view(B, -1, C).transpose(1, 2), v.view(B, -1, C).transpose(1, 2)).reshape(orig_shape) return out From b94d394a64dd0af06bca44b96c66549bb463331d Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 2 Dec 2025 18:38:31 -0800 Subject: [PATCH 212/647] Support Z Image alibaba pai fun controlnets. (#11062) These are not actual controlnets so put it in the models/model_patches folder and use the ModelPatchLoader + QwenImageDiffsynthControlnet node to use it. --- comfy/ldm/lumina/controlnet.py | 113 ++++++++++++++++++++++++++++++ comfy/ldm/lumina/model.py | 24 ++++--- comfy_extras/nodes_model_patch.py | 101 +++++++++++++++++++++++++- 3 files changed, 229 insertions(+), 9 deletions(-) create mode 100644 comfy/ldm/lumina/controlnet.py diff --git a/comfy/ldm/lumina/controlnet.py b/comfy/ldm/lumina/controlnet.py new file mode 100644 index 000000000..fd7ce3b5c --- /dev/null +++ b/comfy/ldm/lumina/controlnet.py @@ -0,0 +1,113 @@ +import torch +from torch import nn + +from .model import JointTransformerBlock + +class ZImageControlTransformerBlock(JointTransformerBlock): + def __init__( + self, + layer_id: int, + dim: int, + n_heads: int, + n_kv_heads: int, + multiple_of: int, + ffn_dim_multiplier: float, + norm_eps: float, + qk_norm: bool, + modulation=True, + block_id=0, + operation_settings=None, + ): + super().__init__(layer_id, dim, n_heads, n_kv_heads, multiple_of, ffn_dim_multiplier, norm_eps, qk_norm, modulation, z_image_modulation=True, operation_settings=operation_settings) + self.block_id = block_id + if block_id == 0: + self.before_proj = operation_settings.get("operations").Linear(self.dim, self.dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + self.after_proj = operation_settings.get("operations").Linear(self.dim, self.dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + + def forward(self, c, x, **kwargs): + if self.block_id == 0: + c = self.before_proj(c) + x + c = super().forward(c, **kwargs) + c_skip = self.after_proj(c) + return c_skip, c + +class ZImage_Control(torch.nn.Module): + def __init__( + self, + dim: int = 3840, + n_heads: int = 30, + n_kv_heads: int = 30, + multiple_of: int = 256, + ffn_dim_multiplier: float = (8.0 / 3.0), + norm_eps: float = 1e-5, + qk_norm: bool = True, + dtype=None, + device=None, + operations=None, + **kwargs + ): + super().__init__() + operation_settings = {"operations": operations, "device": device, "dtype": dtype} + + self.additional_in_dim = 0 + self.control_in_dim = 16 + n_refiner_layers = 2 + self.n_control_layers = 6 + self.control_layers = nn.ModuleList( + [ + ZImageControlTransformerBlock( + i, + dim, + n_heads, + n_kv_heads, + multiple_of, + ffn_dim_multiplier, + norm_eps, + qk_norm, + block_id=i, + operation_settings=operation_settings, + ) + for i in range(self.n_control_layers) + ] + ) + + all_x_embedder = {} + patch_size = 2 + f_patch_size = 1 + x_embedder = operations.Linear(f_patch_size * patch_size * patch_size * self.control_in_dim, dim, bias=True, device=device, dtype=dtype) + all_x_embedder[f"{patch_size}-{f_patch_size}"] = x_embedder + + self.control_all_x_embedder = nn.ModuleDict(all_x_embedder) + self.control_noise_refiner = nn.ModuleList( + [ + JointTransformerBlock( + layer_id, + dim, + n_heads, + n_kv_heads, + multiple_of, + ffn_dim_multiplier, + norm_eps, + qk_norm, + modulation=True, + z_image_modulation=True, + operation_settings=operation_settings, + ) + for layer_id in range(n_refiner_layers) + ] + ) + + def forward(self, cap_feats, control_context, x_freqs_cis, adaln_input): + patch_size = 2 + f_patch_size = 1 + pH = pW = patch_size + B, C, H, W = control_context.shape + control_context = self.control_all_x_embedder[f"{patch_size}-{f_patch_size}"](control_context.view(B, C, H // pH, pH, W // pW, pW).permute(0, 2, 4, 3, 5, 1).flatten(3).flatten(1, 2)) + + x_attn_mask = None + for layer in self.control_noise_refiner: + control_context = layer(control_context, x_attn_mask, x_freqs_cis[:control_context.shape[0], :control_context.shape[1]], adaln_input) + return control_context + + def forward_control_block(self, layer_id, control_context, x, x_attn_mask, x_freqs_cis, adaln_input): + return self.control_layers[layer_id](control_context, x, x_mask=x_attn_mask, freqs_cis=x_freqs_cis[:control_context.shape[0], :control_context.shape[1]], adaln_input=adaln_input) diff --git a/comfy/ldm/lumina/model.py b/comfy/ldm/lumina/model.py index 070b5da09..f1c1a0ec3 100644 --- a/comfy/ldm/lumina/model.py +++ b/comfy/ldm/lumina/model.py @@ -568,7 +568,7 @@ class NextDiT(nn.Module): ).execute(x, timesteps, context, num_tokens, attention_mask, **kwargs) # def forward(self, x, t, cap_feats, cap_mask): - def _forward(self, x, timesteps, context, num_tokens, attention_mask=None, **kwargs): + def _forward(self, x, timesteps, context, num_tokens, attention_mask=None, transformer_options={}, **kwargs): t = 1.0 - timesteps cap_feats = context cap_mask = attention_mask @@ -585,16 +585,24 @@ class NextDiT(nn.Module): cap_feats = self.cap_embedder(cap_feats) # (N, L, D) # todo check if able to batchify w.o. redundant compute + patches = transformer_options.get("patches", {}) transformer_options = kwargs.get("transformer_options", {}) x_is_tensor = isinstance(x, torch.Tensor) - x, mask, img_size, cap_size, freqs_cis = self.patchify_and_embed(x, cap_feats, cap_mask, t, num_tokens, transformer_options=transformer_options) - freqs_cis = freqs_cis.to(x.device) + img, mask, img_size, cap_size, freqs_cis = self.patchify_and_embed(x, cap_feats, cap_mask, t, num_tokens, transformer_options=transformer_options) + freqs_cis = freqs_cis.to(img.device) - for layer in self.layers: - x = layer(x, mask, freqs_cis, adaln_input, transformer_options=transformer_options) + for i, layer in enumerate(self.layers): + img = layer(img, mask, freqs_cis, adaln_input, transformer_options=transformer_options) + if "double_block" in patches: + for p in patches["double_block"]: + out = p({"img": img[:, cap_size[0]:], "txt": img[:, :cap_size[0]], "pe": freqs_cis[:, cap_size[0]:], "vec": adaln_input, "x": x, "block_index": i, "transformer_options": transformer_options}) + if "img" in out: + img[:, cap_size[0]:] = out["img"] + if "txt" in out: + img[:, :cap_size[0]] = out["txt"] - x = self.final_layer(x, adaln_input) - x = self.unpatchify(x, img_size, cap_size, return_tensor=x_is_tensor)[:,:,:h,:w] + img = self.final_layer(img, adaln_input) + img = self.unpatchify(img, img_size, cap_size, return_tensor=x_is_tensor)[:, :, :h, :w] - return -x + return -img diff --git a/comfy_extras/nodes_model_patch.py b/comfy_extras/nodes_model_patch.py index 783c59b6b..c61810dbf 100644 --- a/comfy_extras/nodes_model_patch.py +++ b/comfy_extras/nodes_model_patch.py @@ -6,6 +6,7 @@ import comfy.ops import comfy.model_management import comfy.ldm.common_dit import comfy.latent_formats +import comfy.ldm.lumina.controlnet class BlockWiseControlBlock(torch.nn.Module): @@ -189,6 +190,35 @@ class SigLIPMultiFeatProjModel(torch.nn.Module): return embedding +def z_image_convert(sd): + replace_keys = {".attention.to_out.0.bias": ".attention.out.bias", + ".attention.norm_k.weight": ".attention.k_norm.weight", + ".attention.norm_q.weight": ".attention.q_norm.weight", + ".attention.to_out.0.weight": ".attention.out.weight" + } + + out_sd = {} + for k in sorted(sd.keys()): + w = sd[k] + + k_out = k + if k_out.endswith(".attention.to_k.weight"): + cc = [w] + continue + if k_out.endswith(".attention.to_q.weight"): + cc = [w] + cc + continue + if k_out.endswith(".attention.to_v.weight"): + cc = cc + [w] + w = torch.cat(cc, dim=0) + k_out = k_out.replace(".attention.to_v.weight", ".attention.qkv.weight") + + for r, rr in replace_keys.items(): + k_out = k_out.replace(r, rr) + out_sd[k_out] = w + + return out_sd + class ModelPatchLoader: @classmethod def INPUT_TYPES(s): @@ -211,6 +241,9 @@ class ModelPatchLoader: elif 'feature_embedder.mid_layer_norm.bias' in sd: sd = comfy.utils.state_dict_prefix_replace(sd, {"feature_embedder.": ""}, filter_keys=True) model = SigLIPMultiFeatProjModel(device=comfy.model_management.unet_offload_device(), dtype=dtype, operations=comfy.ops.manual_cast) + elif 'control_all_x_embedder.2-1.weight' in sd: # alipai z image fun controlnet + sd = z_image_convert(sd) + model = comfy.ldm.lumina.controlnet.ZImage_Control(device=comfy.model_management.unet_offload_device(), dtype=dtype, operations=comfy.ops.manual_cast) model.load_state_dict(sd) model = comfy.model_patcher.ModelPatcher(model, load_device=comfy.model_management.get_torch_device(), offload_device=comfy.model_management.unet_offload_device()) @@ -263,6 +296,69 @@ class DiffSynthCnetPatch: def models(self): return [self.model_patch] +class ZImageControlPatch: + def __init__(self, model_patch, vae, image, strength): + self.model_patch = model_patch + self.vae = vae + self.image = image + self.strength = strength + self.encoded_image = self.encode_latent_cond(image) + self.encoded_image_size = (image.shape[1], image.shape[2]) + self.temp_data = None + + def encode_latent_cond(self, image): + latent_image = comfy.latent_formats.Flux().process_in(self.vae.encode(image)) + return latent_image + + def __call__(self, kwargs): + x = kwargs.get("x") + img = kwargs.get("img") + txt = kwargs.get("txt") + pe = kwargs.get("pe") + vec = kwargs.get("vec") + block_index = kwargs.get("block_index") + spacial_compression = self.vae.spacial_compression_encode() + if self.encoded_image is None or self.encoded_image_size != (x.shape[-2] * spacial_compression, x.shape[-1] * spacial_compression): + image_scaled = comfy.utils.common_upscale(self.image.movedim(-1, 1), x.shape[-1] * spacial_compression, x.shape[-2] * spacial_compression, "area", "center") + loaded_models = comfy.model_management.loaded_models(only_currently_used=True) + self.encoded_image = self.encode_latent_cond(image_scaled.movedim(1, -1)) + self.encoded_image_size = (image_scaled.shape[-2], image_scaled.shape[-1]) + comfy.model_management.load_models_gpu(loaded_models) + + cnet_index = (block_index // 5) + cnet_index_float = (block_index / 5) + + kwargs.pop("img") # we do ops in place + kwargs.pop("txt") + + cnet_blocks = self.model_patch.model.n_control_layers + if cnet_index_float > (cnet_blocks - 1): + self.temp_data = None + return kwargs + + if self.temp_data is None or self.temp_data[0] > cnet_index: + self.temp_data = (-1, (None, self.model_patch.model(txt, self.encoded_image.to(img.dtype), pe, vec))) + + while self.temp_data[0] < cnet_index and (self.temp_data[0] + 1) < cnet_blocks: + next_layer = self.temp_data[0] + 1 + self.temp_data = (next_layer, self.model_patch.model.forward_control_block(next_layer, self.temp_data[1][1], img[:, :self.temp_data[1][1].shape[1]], None, pe, vec)) + + if cnet_index_float == self.temp_data[0]: + img[:, :self.temp_data[1][0].shape[1]] += (self.temp_data[1][0] * self.strength) + if cnet_blocks == self.temp_data[0] + 1: + self.temp_data = None + + return kwargs + + def to(self, device_or_dtype): + if isinstance(device_or_dtype, torch.device): + self.encoded_image = self.encoded_image.to(device_or_dtype) + self.temp_data = None + return self + + def models(self): + return [self.model_patch] + class QwenImageDiffsynthControlnet: @classmethod def INPUT_TYPES(s): @@ -289,7 +385,10 @@ class QwenImageDiffsynthControlnet: mask = mask.unsqueeze(2) mask = 1.0 - mask - model_patched.set_model_double_block_patch(DiffSynthCnetPatch(model_patch, vae, image, strength, mask)) + if isinstance(model_patch.model, comfy.ldm.lumina.controlnet.ZImage_Control): + model_patched.set_model_double_block_patch(ZImageControlPatch(model_patch, vae, image, strength)) + else: + model_patched.set_model_double_block_patch(DiffSynthCnetPatch(model_patch, vae, image, strength, mask)) return (model_patched,) From 3f512f5659cfbb3c53999cde6ff557591740252b Mon Sep 17 00:00:00 2001 From: Jim Heising Date: Tue, 2 Dec 2025 19:29:27 -0800 Subject: [PATCH 213/647] Added PATCH method to CORS headers (#11066) Added PATCH http method to access-control-allow-header-methods header because there are now PATCH endpoints exposed in the API. See https://github.com/comfyanonymous/ComfyUI/blob/277237ccc1499bac7fcd221a666dfe7a32ac4206/api_server/routes/internal/internal_routes.py#L34 for an example of an API endpoint that uses the PATCH method. --- server.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/server.py b/server.py index e3bd056d9..ac4f42222 100644 --- a/server.py +++ b/server.py @@ -98,7 +98,7 @@ def create_cors_middleware(allowed_origin: str): response = await handler(request) response.headers['Access-Control-Allow-Origin'] = allowed_origin - response.headers['Access-Control-Allow-Methods'] = 'POST, GET, DELETE, PUT, OPTIONS' + response.headers['Access-Control-Allow-Methods'] = 'POST, GET, DELETE, PUT, OPTIONS, PATCH' response.headers['Access-Control-Allow-Headers'] = 'Content-Type, Authorization' response.headers['Access-Control-Allow-Credentials'] = 'true' return response From 73f5649196f472d3719e2e7513e0a9d029cc3e38 Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Wed, 3 Dec 2025 13:49:29 +1000 Subject: [PATCH 214/647] Implement temporal rolling VAE (Major VRAM reductions in Hunyuan and Kandinsky) (#10995) * hunyuan upsampler: rework imports Remove the transitive import of VideoConv3d and Resnet and takes these from actual implementation source. * model: remove unused give_pre_end According to git grep, this is not used now, and was not used in the initial commit that introduced it (see below). This semantic is difficult to implement temporal roll VAE for (and would defeat the purpose). Rather than implement the complex if, just delete the unused feature. (venv) rattus@rattus-box2:~/ComfyUI$ git log --oneline 220afe33 (HEAD) Initial commit. (venv) rattus@rattus-box2:~/ComfyUI$ git grep give_pre comfy/ldm/modules/diffusionmodules/model.py: resolution, z_channels, give_pre_end=False, tanh_out=False, use_linear_attn=False, comfy/ldm/modules/diffusionmodules/model.py: self.give_pre_end = give_pre_end comfy/ldm/modules/diffusionmodules/model.py: if self.give_pre_end: (venv) rattus@rattus-box2:~/ComfyUI$ git co origin/master Previous HEAD position was 220afe33 Initial commit. HEAD is now at 9d8a8179 Enable async offloading by default on Nvidia. (#10953) (venv) rattus@rattus-box2:~/ComfyUI$ git grep give_pre comfy/ldm/modules/diffusionmodules/model.py: resolution, z_channels, give_pre_end=False, tanh_out=False, use_linear_attn=False, comfy/ldm/modules/diffusionmodules/model.py: self.give_pre_end = give_pre_end comfy/ldm/modules/diffusionmodules/model.py: if self.give_pre_end: * move refiner VAE temporal roller to core Move the carrying conv op to the common VAE code and give it a better name. Roll the carry implementation logic for Resnet into the base class and scrap the Hunyuan specific subclass. * model: Add temporal roll to main VAE decoder If there are no attention layers, its a standard resnet and VideoConv3d is asked for, substitute in the temporal rolloing VAE algorithm. This reduces VAE usage by the temporal dimension (can be huge VRAM savings). * model: Add temporal roll to main VAE encoder If there are no attention layers, its a standard resnet and VideoConv3d is asked for, substitute in the temporal rolling VAE algorithm. This reduces VAE usage by the temporal dimension (can be huge VRAM savings). --- comfy/ldm/hunyuan_video/upsampler.py | 3 +- comfy/ldm/hunyuan_video/vae_refiner.py | 94 +++------ comfy/ldm/modules/diffusionmodules/model.py | 207 ++++++++++++++------ 3 files changed, 174 insertions(+), 130 deletions(-) diff --git a/comfy/ldm/hunyuan_video/upsampler.py b/comfy/ldm/hunyuan_video/upsampler.py index 9f5e91a59..85f515f67 100644 --- a/comfy/ldm/hunyuan_video/upsampler.py +++ b/comfy/ldm/hunyuan_video/upsampler.py @@ -1,7 +1,8 @@ import torch import torch.nn as nn import torch.nn.functional as F -from comfy.ldm.hunyuan_video.vae_refiner import RMS_norm, ResnetBlock, VideoConv3d +from comfy.ldm.modules.diffusionmodules.model import ResnetBlock, VideoConv3d +from comfy.ldm.hunyuan_video.vae_refiner import RMS_norm import model_management, model_patcher class SRResidualCausalBlock3D(nn.Module): diff --git a/comfy/ldm/hunyuan_video/vae_refiner.py b/comfy/ldm/hunyuan_video/vae_refiner.py index 9f750dcc4..ddf77cd0e 100644 --- a/comfy/ldm/hunyuan_video/vae_refiner.py +++ b/comfy/ldm/hunyuan_video/vae_refiner.py @@ -1,42 +1,12 @@ import torch import torch.nn as nn import torch.nn.functional as F -from comfy.ldm.modules.diffusionmodules.model import ResnetBlock, AttnBlock, VideoConv3d, Normalize +from comfy.ldm.modules.diffusionmodules.model import ResnetBlock, AttnBlock, CarriedConv3d, Normalize, conv_carry_causal_3d, torch_cat_if_needed import comfy.ops import comfy.ldm.models.autoencoder import comfy.model_management ops = comfy.ops.disable_weight_init -class NoPadConv3d(nn.Module): - def __init__(self, n_channels, out_channels, kernel_size, stride=1, dilation=1, padding=0, **kwargs): - super().__init__() - self.conv = ops.Conv3d(n_channels, out_channels, kernel_size, stride=stride, dilation=dilation, **kwargs) - - def forward(self, x): - return self.conv(x) - - -def conv_carry_causal_3d(xl, op, conv_carry_in=None, conv_carry_out=None): - - x = xl[0] - xl.clear() - - if conv_carry_out is not None: - to_push = x[:, :, -2:, :, :].clone() - conv_carry_out.append(to_push) - - if isinstance(op, NoPadConv3d): - if conv_carry_in is None: - x = torch.nn.functional.pad(x, (1, 1, 1, 1, 2, 0), mode = 'replicate') - else: - carry_len = conv_carry_in[0].shape[2] - x = torch.cat([conv_carry_in.pop(0), x], dim=2) - x = torch.nn.functional.pad(x, (1, 1, 1, 1, 2 - carry_len, 0), mode = 'replicate') - - out = op(x) - - return out - class RMS_norm(nn.Module): def __init__(self, dim): @@ -49,7 +19,7 @@ class RMS_norm(nn.Module): return F.normalize(x, dim=1) * self.scale * comfy.model_management.cast_to(self.gamma, dtype=x.dtype, device=x.device) class DnSmpl(nn.Module): - def __init__(self, ic, oc, tds=True, refiner_vae=True, op=VideoConv3d): + def __init__(self, ic, oc, tds, refiner_vae, op): super().__init__() fct = 2 * 2 * 2 if tds else 1 * 2 * 2 assert oc % fct == 0 @@ -109,7 +79,7 @@ class DnSmpl(nn.Module): class UpSmpl(nn.Module): - def __init__(self, ic, oc, tus=True, refiner_vae=True, op=VideoConv3d): + def __init__(self, ic, oc, tus, refiner_vae, op): super().__init__() fct = 2 * 2 * 2 if tus else 1 * 2 * 2 self.conv = op(ic, oc * fct, kernel_size=3, stride=1, padding=1) @@ -163,23 +133,6 @@ class UpSmpl(nn.Module): return h + x -class HunyuanRefinerResnetBlock(ResnetBlock): - def __init__(self, in_channels, out_channels, conv_op=NoPadConv3d, norm_op=RMS_norm): - super().__init__(in_channels=in_channels, out_channels=out_channels, temb_channels=0, conv_op=conv_op, norm_op=norm_op) - - def forward(self, x, conv_carry_in=None, conv_carry_out=None): - h = x - h = [ self.swish(self.norm1(x)) ] - h = conv_carry_causal_3d(h, self.conv1, conv_carry_in=conv_carry_in, conv_carry_out=conv_carry_out) - - h = [ self.dropout(self.swish(self.norm2(h))) ] - h = conv_carry_causal_3d(h, self.conv2, conv_carry_in=conv_carry_in, conv_carry_out=conv_carry_out) - - if self.in_channels != self.out_channels: - x = self.nin_shortcut(x) - - return x+h - class Encoder(nn.Module): def __init__(self, in_channels, z_channels, block_out_channels, num_res_blocks, ffactor_spatial, ffactor_temporal, downsample_match_channel=True, refiner_vae=True, **_): @@ -191,7 +144,7 @@ class Encoder(nn.Module): self.refiner_vae = refiner_vae if self.refiner_vae: - conv_op = NoPadConv3d + conv_op = CarriedConv3d norm_op = RMS_norm else: conv_op = ops.Conv3d @@ -206,9 +159,10 @@ class Encoder(nn.Module): for i, tgt in enumerate(block_out_channels): stage = nn.Module() - stage.block = nn.ModuleList([HunyuanRefinerResnetBlock(in_channels=ch if j == 0 else tgt, - out_channels=tgt, - conv_op=conv_op, norm_op=norm_op) + stage.block = nn.ModuleList([ResnetBlock(in_channels=ch if j == 0 else tgt, + out_channels=tgt, + temb_channels=0, + conv_op=conv_op, norm_op=norm_op) for j in range(num_res_blocks)]) ch = tgt if i < depth: @@ -218,9 +172,9 @@ class Encoder(nn.Module): self.down.append(stage) self.mid = nn.Module() - self.mid.block_1 = HunyuanRefinerResnetBlock(in_channels=ch, out_channels=ch, conv_op=conv_op, norm_op=norm_op) + self.mid.block_1 = ResnetBlock(in_channels=ch, out_channels=ch, conv_op=conv_op, norm_op=norm_op) self.mid.attn_1 = AttnBlock(ch, conv_op=ops.Conv3d, norm_op=norm_op) - self.mid.block_2 = HunyuanRefinerResnetBlock(in_channels=ch, out_channels=ch, conv_op=conv_op, norm_op=norm_op) + self.mid.block_2 = ResnetBlock(in_channels=ch, out_channels=ch, conv_op=conv_op, norm_op=norm_op) self.norm_out = norm_op(ch) self.conv_out = conv_op(ch, z_channels << 1, 3, 1, 1) @@ -246,22 +200,20 @@ class Encoder(nn.Module): conv_carry_out = [] if i == len(x) - 1: conv_carry_out = None + x1 = [ x1 ] x1 = conv_carry_causal_3d(x1, self.conv_in, conv_carry_in, conv_carry_out) for stage in self.down: for blk in stage.block: - x1 = blk(x1, conv_carry_in, conv_carry_out) + x1 = blk(x1, None, conv_carry_in, conv_carry_out) if hasattr(stage, 'downsample'): x1 = stage.downsample(x1, conv_carry_in, conv_carry_out) out.append(x1) conv_carry_in = conv_carry_out - if len(out) > 1: - out = torch.cat(out, dim=2) - else: - out = out[0] + out = torch_cat_if_needed(out, dim=2) x = self.mid.block_2(self.mid.attn_1(self.mid.block_1(out))) del out @@ -288,7 +240,7 @@ class Decoder(nn.Module): self.refiner_vae = refiner_vae if self.refiner_vae: - conv_op = NoPadConv3d + conv_op = CarriedConv3d norm_op = RMS_norm else: conv_op = ops.Conv3d @@ -298,9 +250,9 @@ class Decoder(nn.Module): self.conv_in = conv_op(z_channels, ch, kernel_size=3, stride=1, padding=1) self.mid = nn.Module() - self.mid.block_1 = HunyuanRefinerResnetBlock(in_channels=ch, out_channels=ch, conv_op=conv_op, norm_op=norm_op) + self.mid.block_1 = ResnetBlock(in_channels=ch, out_channels=ch, conv_op=conv_op, norm_op=norm_op) self.mid.attn_1 = AttnBlock(ch, conv_op=ops.Conv3d, norm_op=norm_op) - self.mid.block_2 = HunyuanRefinerResnetBlock(in_channels=ch, out_channels=ch, conv_op=conv_op, norm_op=norm_op) + self.mid.block_2 = ResnetBlock(in_channels=ch, out_channels=ch, conv_op=conv_op, norm_op=norm_op) self.up = nn.ModuleList() depth = (ffactor_spatial >> 1).bit_length() @@ -308,9 +260,10 @@ class Decoder(nn.Module): for i, tgt in enumerate(block_out_channels): stage = nn.Module() - stage.block = nn.ModuleList([HunyuanRefinerResnetBlock(in_channels=ch if j == 0 else tgt, - out_channels=tgt, - conv_op=conv_op, norm_op=norm_op) + stage.block = nn.ModuleList([ResnetBlock(in_channels=ch if j == 0 else tgt, + out_channels=tgt, + temb_channels=0, + conv_op=conv_op, norm_op=norm_op) for j in range(num_res_blocks + 1)]) ch = tgt if i < depth: @@ -340,7 +293,7 @@ class Decoder(nn.Module): conv_carry_out = None for stage in self.up: for blk in stage.block: - x1 = blk(x1, conv_carry_in, conv_carry_out) + x1 = blk(x1, None, conv_carry_in, conv_carry_out) if hasattr(stage, 'upsample'): x1 = stage.upsample(x1, conv_carry_in, conv_carry_out) @@ -350,10 +303,7 @@ class Decoder(nn.Module): conv_carry_in = conv_carry_out del x - if len(out) > 1: - out = torch.cat(out, dim=2) - else: - out = out[0] + out = torch_cat_if_needed(out, dim=2) if not self.refiner_vae: if z.shape[-3] == 1: diff --git a/comfy/ldm/modules/diffusionmodules/model.py b/comfy/ldm/modules/diffusionmodules/model.py index de1e01cc8..681a55db5 100644 --- a/comfy/ldm/modules/diffusionmodules/model.py +++ b/comfy/ldm/modules/diffusionmodules/model.py @@ -13,6 +13,12 @@ if model_management.xformers_enabled_vae(): import xformers import xformers.ops +def torch_cat_if_needed(xl, dim): + if len(xl) > 1: + return torch.cat(xl, dim) + else: + return xl[0] + def get_timestep_embedding(timesteps, embedding_dim): """ This matches the implementation in Denoising Diffusion Probabilistic Models: @@ -43,6 +49,37 @@ def Normalize(in_channels, num_groups=32): return ops.GroupNorm(num_groups=num_groups, num_channels=in_channels, eps=1e-6, affine=True) +class CarriedConv3d(nn.Module): + def __init__(self, n_channels, out_channels, kernel_size, stride=1, dilation=1, padding=0, **kwargs): + super().__init__() + self.conv = ops.Conv3d(n_channels, out_channels, kernel_size, stride=stride, dilation=dilation, **kwargs) + + def forward(self, x): + return self.conv(x) + + +def conv_carry_causal_3d(xl, op, conv_carry_in=None, conv_carry_out=None): + + x = xl[0] + xl.clear() + + if isinstance(op, CarriedConv3d): + if conv_carry_in is None: + x = torch.nn.functional.pad(x, (1, 1, 1, 1, 2, 0), mode = 'replicate') + else: + carry_len = conv_carry_in[0].shape[2] + x = torch.nn.functional.pad(x, (1, 1, 1, 1, 2 - carry_len, 0), mode = 'replicate') + x = torch.cat([conv_carry_in.pop(0), x], dim=2) + + if conv_carry_out is not None: + to_push = x[:, :, -2:, :, :].clone() + conv_carry_out.append(to_push) + + out = op(x) + + return out + + class VideoConv3d(nn.Module): def __init__(self, n_channels, out_channels, kernel_size, stride=1, dilation=1, padding_mode='replicate', padding=1, **kwargs): super().__init__() @@ -89,29 +126,24 @@ class Upsample(nn.Module): stride=1, padding=1) - def forward(self, x): + def forward(self, x, conv_carry_in=None, conv_carry_out=None): scale_factor = self.scale_factor if isinstance(scale_factor, (int, float)): scale_factor = (scale_factor,) * (x.ndim - 2) if x.ndim == 5 and scale_factor[0] > 1.0: - t = x.shape[2] - if t > 1: - a, b = x.split((1, t - 1), dim=2) - del x - b = interpolate_up(b, scale_factor) - else: - a = x - - a = interpolate_up(a.squeeze(2), scale_factor=scale_factor[1:]).unsqueeze(2) - if t > 1: - x = torch.cat((a, b), dim=2) - else: - x = a + results = [] + if conv_carry_in is None: + first = x[:, :, :1, :, :] + results.append(interpolate_up(first.squeeze(2), scale_factor=scale_factor[1:]).unsqueeze(2)) + x = x[:, :, 1:, :, :] + if x.shape[2] > 0: + results.append(interpolate_up(x, scale_factor)) + x = torch_cat_if_needed(results, dim=2) else: x = interpolate_up(x, scale_factor) if self.with_conv: - x = self.conv(x) + x = conv_carry_causal_3d([x], self.conv, conv_carry_in, conv_carry_out) return x @@ -127,17 +159,20 @@ class Downsample(nn.Module): stride=stride, padding=0) - def forward(self, x): + def forward(self, x, conv_carry_in=None, conv_carry_out=None): if self.with_conv: - if x.ndim == 4: + if isinstance(self.conv, CarriedConv3d): + x = conv_carry_causal_3d([x], self.conv, conv_carry_in, conv_carry_out) + elif x.ndim == 4: pad = (0, 1, 0, 1) mode = "constant" x = torch.nn.functional.pad(x, pad, mode=mode, value=0) + x = self.conv(x) elif x.ndim == 5: pad = (1, 1, 1, 1, 2, 0) mode = "replicate" x = torch.nn.functional.pad(x, pad, mode=mode) - x = self.conv(x) + x = self.conv(x) else: x = torch.nn.functional.avg_pool2d(x, kernel_size=2, stride=2) return x @@ -183,23 +218,23 @@ class ResnetBlock(nn.Module): stride=1, padding=0) - def forward(self, x, temb=None): + def forward(self, x, temb=None, conv_carry_in=None, conv_carry_out=None): h = x h = self.norm1(h) - h = self.swish(h) - h = self.conv1(h) + h = [ self.swish(h) ] + h = conv_carry_causal_3d(h, self.conv1, conv_carry_in=conv_carry_in, conv_carry_out=conv_carry_out) if temb is not None: h = h + self.temb_proj(self.swish(temb))[:,:,None,None] h = self.norm2(h) h = self.swish(h) - h = self.dropout(h) - h = self.conv2(h) + h = [ self.dropout(h) ] + h = conv_carry_causal_3d(h, self.conv2, conv_carry_in=conv_carry_in, conv_carry_out=conv_carry_out) if self.in_channels != self.out_channels: if self.use_conv_shortcut: - x = self.conv_shortcut(x) + x = conv_carry_causal_3d([x], self.conv_shortcut, conv_carry_in=conv_carry_in, conv_carry_out=conv_carry_out) else: x = self.nin_shortcut(x) @@ -520,9 +555,14 @@ class Encoder(nn.Module): self.num_res_blocks = num_res_blocks self.resolution = resolution self.in_channels = in_channels + self.carried = False if conv3d: - conv_op = VideoConv3d + if not attn_resolutions: + conv_op = CarriedConv3d + self.carried = True + else: + conv_op = VideoConv3d mid_attn_conv_op = ops.Conv3d else: conv_op = ops.Conv2d @@ -535,6 +575,7 @@ class Encoder(nn.Module): stride=1, padding=1) + self.time_compress = 1 curr_res = resolution in_ch_mult = (1,)+tuple(ch_mult) self.in_ch_mult = in_ch_mult @@ -561,10 +602,15 @@ class Encoder(nn.Module): if time_compress is not None: if (self.num_resolutions - 1 - i_level) > math.log2(time_compress): stride = (1, 2, 2) + else: + self.time_compress *= 2 down.downsample = Downsample(block_in, resamp_with_conv, stride=stride, conv_op=conv_op) curr_res = curr_res // 2 self.down.append(down) + if time_compress is not None: + self.time_compress = time_compress + # middle self.mid = nn.Module() self.mid.block_1 = ResnetBlock(in_channels=block_in, @@ -590,15 +636,42 @@ class Encoder(nn.Module): def forward(self, x): # timestep embedding temb = None - # downsampling - h = self.conv_in(x) - for i_level in range(self.num_resolutions): - for i_block in range(self.num_res_blocks): - h = self.down[i_level].block[i_block](h, temb) - if len(self.down[i_level].attn) > 0: - h = self.down[i_level].attn[i_block](h) - if i_level != self.num_resolutions-1: - h = self.down[i_level].downsample(h) + + if self.carried: + xl = [x[:, :, :1, :, :]] + if x.shape[2] > self.time_compress: + tc = self.time_compress + xl += torch.split(x[:, :, 1: 1 + ((x.shape[2] - 1) // tc) * tc, :, :], tc * 2, dim = 2) + x = xl + else: + x = [x] + out = [] + + conv_carry_in = None + + for i, x1 in enumerate(x): + conv_carry_out = [] + if i == len(x) - 1: + conv_carry_out = None + + # downsampling + x1 = [ x1 ] + h1 = conv_carry_causal_3d(x1, self.conv_in, conv_carry_in, conv_carry_out) + + for i_level in range(self.num_resolutions): + for i_block in range(self.num_res_blocks): + h1 = self.down[i_level].block[i_block](h1, temb, conv_carry_in, conv_carry_out) + if len(self.down[i_level].attn) > 0: + assert i == 0 #carried should not happen if attn exists + h1 = self.down[i_level].attn[i_block](h1) + if i_level != self.num_resolutions-1: + h1 = self.down[i_level].downsample(h1, conv_carry_in, conv_carry_out) + + out.append(h1) + conv_carry_in = conv_carry_out + + h = torch_cat_if_needed(out, dim=2) + del out # middle h = self.mid.block_1(h, temb) @@ -607,15 +680,15 @@ class Encoder(nn.Module): # end h = self.norm_out(h) - h = nonlinearity(h) - h = self.conv_out(h) + h = [ nonlinearity(h) ] + h = conv_carry_causal_3d(h, self.conv_out) return h class Decoder(nn.Module): def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks, attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels, - resolution, z_channels, give_pre_end=False, tanh_out=False, use_linear_attn=False, + resolution, z_channels, tanh_out=False, use_linear_attn=False, conv_out_op=ops.Conv2d, resnet_op=ResnetBlock, attn_op=AttnBlock, @@ -629,12 +702,18 @@ class Decoder(nn.Module): self.num_res_blocks = num_res_blocks self.resolution = resolution self.in_channels = in_channels - self.give_pre_end = give_pre_end self.tanh_out = tanh_out + self.carried = False if conv3d: - conv_op = VideoConv3d - conv_out_op = VideoConv3d + if not attn_resolutions and resnet_op == ResnetBlock: + conv_op = CarriedConv3d + conv_out_op = CarriedConv3d + self.carried = True + else: + conv_op = VideoConv3d + conv_out_op = VideoConv3d + mid_attn_conv_op = ops.Conv3d else: conv_op = ops.Conv2d @@ -709,29 +788,43 @@ class Decoder(nn.Module): temb = None # z to block_in - h = self.conv_in(z) + h = conv_carry_causal_3d([z], self.conv_in) # middle h = self.mid.block_1(h, temb, **kwargs) h = self.mid.attn_1(h, **kwargs) h = self.mid.block_2(h, temb, **kwargs) + if self.carried: + h = torch.split(h, 2, dim=2) + else: + h = [ h ] + out = [] + + conv_carry_in = None + # upsampling - for i_level in reversed(range(self.num_resolutions)): - for i_block in range(self.num_res_blocks+1): - h = self.up[i_level].block[i_block](h, temb, **kwargs) - if len(self.up[i_level].attn) > 0: - h = self.up[i_level].attn[i_block](h, **kwargs) - if i_level != 0: - h = self.up[i_level].upsample(h) + for i, h1 in enumerate(h): + conv_carry_out = [] + if i == len(h) - 1: + conv_carry_out = None + for i_level in reversed(range(self.num_resolutions)): + for i_block in range(self.num_res_blocks+1): + h1 = self.up[i_level].block[i_block](h1, temb, conv_carry_in, conv_carry_out, **kwargs) + if len(self.up[i_level].attn) > 0: + assert i == 0 #carried should not happen if attn exists + h1 = self.up[i_level].attn[i_block](h1, **kwargs) + if i_level != 0: + h1 = self.up[i_level].upsample(h1, conv_carry_in, conv_carry_out) - # end - if self.give_pre_end: - return h + h1 = self.norm_out(h1) + h1 = [ nonlinearity(h1) ] + h1 = conv_carry_causal_3d(h1, self.conv_out, conv_carry_in, conv_carry_out) + if self.tanh_out: + h1 = torch.tanh(h1) + out.append(h1) + conv_carry_in = conv_carry_out - h = self.norm_out(h) - h = nonlinearity(h) - h = self.conv_out(h, **kwargs) - if self.tanh_out: - h = torch.tanh(h) - return h + out = torch_cat_if_needed(out, dim=2) + + return out From c120eee5bacca643062657d2a7efad83c7d4d828 Mon Sep 17 00:00:00 2001 From: Jedrzej Kosinski Date: Tue, 2 Dec 2025 21:17:13 -0800 Subject: [PATCH 215/647] Add MatchType, DynamicCombo, and Autogrow support to V3 Schema (#10832) * Added output_matchtypes to generated json for v3, initial backend support for MatchType, created nodes_logic.py and added SwitchNode * Fixed providing list of allowed_types * Add workaround in validation.py for V3 Combo outputs not working as Combo inputs * Make match type receive_type pass validation * Also add MatchType check to input_type in validation - will likely trigger when connecting to non-lazy stuff * Make sure this PR only has MatchType stuff * Initial work on DynamicCombo * Add get_dynamic function, not yet filled out correctly * Mark Switch node as Beta * Make sure other unfinished dynamic types are not accidentally used * Send DynamicCombo.Option inputs in the same format as normal v1 inputs * add dynamic combo test node * Support validation of inputs and outputs * Add missing input params to DynamicCombo.Input * Add get_all function to inputs for id validation purposes * Fix imports for v3 returning everything when doing io/ui/IO/UI instead of what is in __all__ of _io.py and _ui.py * Modifying behavior of get_dynamic in V3 + serialization so can be used in execution code * Fix v3 schema validation code after changes * Refactor hidden_values for v3 in execution.py to be more general v3_data, add helper functions for dynamic behavior, preparing for restructuring dynamic type into object (not finished yet) * Add nesting of inputs on DynamicCombo during execution * Work with latest frontend commits * Fix cringe arrows * frontend will no longer namespace dynamic inputs widgets so reflect that in code, refactor build_nested_inputs * Prepare Autogrow support for the love of the game * satisfy ruff * Create test nodes for Autogrow to collab with frontend development * Add nested combo to DCTestNode * Remove array support from build_nested_inputs, properly handle missing expected values * Make execution.validate_inputs properly validate required dynamic inputs, renamed dynamic_data to dynamic_paths for clarity * MatchType does not need any DynamicInput/Output features on backend; will increase compatibility with dynamic types * Probably need this for ruff check * Change MatchType to have template be the first and only required param; output id's do nothing right now, so no need * Fix merge regression with LatentUpscaleModel type not being put in __all__ for _io.py, fix invalid type hint for validate_inputs * Make Switch node inputs optional, disallow both inputs from being missing, and still work properly with lazy; when one input is missing, use the other no matter what the switch is set to * Satisfy ruff * Move MatchType code above the types that inherit from DynamicInput * Add DynamicSlot type, awaiting frontend support * Make curr_prefix creation happen in Autogrow, move curr_prefix in DynamicCombo to only be created if input exists in live_inputs * I was confused, fixing accidentally redundant curr_prefix addition in Autogrow * Make sure Autogrow inputs are force_input = True when WidgetInput, fix runtime validation by removing original input from expected inputs, fix min/max bounds, change test nodes slightly * Remove unnecessary id usage in Autogrow test node outputs * Commented out Switch node + test nodes * Remove commented out code from Autogrow * Make TemplatePrefix max more clear, allow max == 1 * Replace all dict[str] with dict[str, Any] * Renamed add_to_dict_live_inputs to expand_schema_for_dynamic * Fixed typo in DynamicSlot input code * note about live_inputs not being present soon in get_v1_info (internal function anyway) * For now, hide DynamicCombo and Autogrow from public interface * Removed comment --- comfy_api/latest/__init__.py | 4 +- comfy_api/latest/_io.py | 416 ++++++++++++++++++++++++++------- comfy_api/latest/_io_public.py | 1 + comfy_api/latest/_ui_public.py | 1 + comfy_api/v0_0_2/__init__.py | 6 +- comfy_execution/validation.py | 6 + comfy_extras/nodes_logic.py | 155 ++++++++++++ execution.py | 40 ++-- nodes.py | 1 + 9 files changed, 525 insertions(+), 105 deletions(-) create mode 100644 comfy_api/latest/_io_public.py create mode 100644 comfy_api/latest/_ui_public.py create mode 100644 comfy_extras/nodes_logic.py diff --git a/comfy_api/latest/__init__.py b/comfy_api/latest/__init__.py index 176ae36e0..0fa01d1e7 100644 --- a/comfy_api/latest/__init__.py +++ b/comfy_api/latest/__init__.py @@ -8,8 +8,8 @@ from comfy_api.internal.async_to_sync import create_sync_class from comfy_api.latest._input import ImageInput, AudioInput, MaskInput, LatentInput, VideoInput from comfy_api.latest._input_impl import VideoFromFile, VideoFromComponents from comfy_api.latest._util import VideoCodec, VideoContainer, VideoComponents, MESH, VOXEL -from . import _io as io -from . import _ui as ui +from . import _io_public as io +from . import _ui_public as ui # from comfy_api.latest._resources import _RESOURCES as resources #noqa: F401 from comfy_execution.utils import get_executing_context from comfy_execution.progress import get_progress_state, PreviewImageTuple diff --git a/comfy_api/latest/_io.py b/comfy_api/latest/_io.py index 79c0722a9..257f07c42 100644 --- a/comfy_api/latest/_io.py +++ b/comfy_api/latest/_io.py @@ -4,6 +4,7 @@ import copy import inspect from abc import ABC, abstractmethod from collections import Counter +from collections.abc import Iterable from dataclasses import asdict, dataclass from enum import Enum from typing import Any, Callable, Literal, TypedDict, TypeVar, TYPE_CHECKING @@ -150,6 +151,9 @@ class _IO_V3: def __init__(self): pass + def validate(self): + pass + @property def io_type(self): return self.Parent.io_type @@ -182,6 +186,9 @@ class Input(_IO_V3): def get_io_type(self): return _StringIOType(self.io_type) + def get_all(self) -> list[Input]: + return [self] + class WidgetInput(Input): ''' Base class for a V3 Input with widget. @@ -814,13 +821,61 @@ class MultiType: else: return super().as_dict() +@comfytype(io_type="COMFY_MATCHTYPE_V3") +class MatchType(ComfyTypeIO): + class Template: + def __init__(self, template_id: str, allowed_types: _ComfyType | list[_ComfyType] = AnyType): + self.template_id = template_id + # account for syntactic sugar + if not isinstance(allowed_types, Iterable): + allowed_types = [allowed_types] + for t in allowed_types: + if not isinstance(t, type): + if not isinstance(t, _ComfyType): + raise ValueError(f"Allowed types must be a ComfyType or a list of ComfyTypes, got {t.__class__.__name__}") + else: + if not issubclass(t, _ComfyType): + raise ValueError(f"Allowed types must be a ComfyType or a list of ComfyTypes, got {t.__name__}") + self.allowed_types = allowed_types + + def as_dict(self): + return { + "template_id": self.template_id, + "allowed_types": ",".join([t.io_type for t in self.allowed_types]), + } + + class Input(Input): + def __init__(self, id: str, template: MatchType.Template, + display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, extra_dict=None): + super().__init__(id, display_name, optional, tooltip, lazy, extra_dict) + self.template = template + + def as_dict(self): + return super().as_dict() | prune_dict({ + "template": self.template.as_dict(), + }) + + class Output(Output): + def __init__(self, template: MatchType.Template, id: str=None, display_name: str=None, tooltip: str=None, + is_output_list=False): + super().__init__(id, display_name, tooltip, is_output_list) + self.template = template + + def as_dict(self): + return super().as_dict() | prune_dict({ + "template": self.template.as_dict(), + }) + class DynamicInput(Input, ABC): ''' Abstract class for dynamic input registration. ''' - @abstractmethod def get_dynamic(self) -> list[Input]: - ... + return [] + + def expand_schema_for_dynamic(self, d: dict[str, Any], live_inputs: dict[str, Any], curr_prefix=''): + pass + class DynamicOutput(Output, ABC): ''' @@ -830,99 +885,223 @@ class DynamicOutput(Output, ABC): is_output_list=False): super().__init__(id, display_name, tooltip, is_output_list) - @abstractmethod def get_dynamic(self) -> list[Output]: - ... + return [] @comfytype(io_type="COMFY_AUTOGROW_V3") -class AutogrowDynamic(ComfyTypeI): - Type = list[Any] - class Input(DynamicInput): - def __init__(self, id: str, template_input: Input, min: int=1, max: int=None, - display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, extra_dict=None): - super().__init__(id, display_name, optional, tooltip, lazy, extra_dict) - self.template_input = template_input - if min is not None: - assert(min >= 1) - if max is not None: - assert(max >= 1) +class Autogrow(ComfyTypeI): + Type = dict[str, Any] + _MaxNames = 100 # NOTE: max 100 names for sanity + + class _AutogrowTemplate: + def __init__(self, input: Input): + # dynamic inputs are not allowed as the template input + assert(not isinstance(input, DynamicInput)) + self.input = copy.copy(input) + if isinstance(self.input, WidgetInput): + self.input.force_input = True + self.names: list[str] = [] + self.cached_inputs = {} + + def _create_input(self, input: Input, name: str): + new_input = copy.copy(self.input) + new_input.id = name + return new_input + + def _create_cached_inputs(self): + for name in self.names: + self.cached_inputs[name] = self._create_input(self.input, name) + + def get_all(self) -> list[Input]: + return list(self.cached_inputs.values()) + + def as_dict(self): + return prune_dict({ + "input": create_input_dict_v1([self.input]), + }) + + def validate(self): + self.input.validate() + + def expand_schema_for_dynamic(self, d: dict[str, Any], live_inputs: dict[str, Any], curr_prefix=''): + real_inputs = [] + for name, input in self.cached_inputs.items(): + if name in live_inputs: + real_inputs.append(input) + add_to_input_dict_v1(d, real_inputs, live_inputs, curr_prefix) + add_dynamic_id_mapping(d, real_inputs, curr_prefix) + + class TemplatePrefix(_AutogrowTemplate): + def __init__(self, input: Input, prefix: str, min: int=1, max: int=10): + super().__init__(input) + self.prefix = prefix + assert(min >= 0) + assert(max >= 1) + assert(max <= Autogrow._MaxNames) self.min = min self.max = max + self.names = [f"{self.prefix}{i}" for i in range(self.max)] + self._create_cached_inputs() + + def as_dict(self): + return super().as_dict() | prune_dict({ + "prefix": self.prefix, + "min": self.min, + "max": self.max, + }) + + class TemplateNames(_AutogrowTemplate): + def __init__(self, input: Input, names: list[str], min: int=1): + super().__init__(input) + self.names = names[:Autogrow._MaxNames] + assert(min >= 0) + self.min = min + self._create_cached_inputs() + + def as_dict(self): + return super().as_dict() | prune_dict({ + "names": self.names, + "min": self.min, + }) + + class Input(DynamicInput): + def __init__(self, id: str, template: Autogrow.TemplatePrefix | Autogrow.TemplateNames, + display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, extra_dict=None): + super().__init__(id, display_name, optional, tooltip, lazy, extra_dict) + self.template = template + + def as_dict(self): + return super().as_dict() | prune_dict({ + "template": self.template.as_dict(), + }) def get_dynamic(self) -> list[Input]: - curr_count = 1 - new_inputs = [] - for i in range(self.min): - new_input = copy.copy(self.template_input) - new_input.id = f"{new_input.id}{curr_count}_${self.id}_ag$" - if new_input.display_name is not None: - new_input.display_name = f"{new_input.display_name}{curr_count}" - new_input.optional = self.optional or new_input.optional - if isinstance(self.template_input, WidgetInput): - new_input.force_input = True - new_inputs.append(new_input) - curr_count += 1 - # pretend to expand up to max - for i in range(curr_count-1, self.max): - new_input = copy.copy(self.template_input) - new_input.id = f"{new_input.id}{curr_count}_${self.id}_ag$" - if new_input.display_name is not None: - new_input.display_name = f"{new_input.display_name}{curr_count}" - new_input.optional = True - if isinstance(self.template_input, WidgetInput): - new_input.force_input = True - new_inputs.append(new_input) - curr_count += 1 - return new_inputs + return self.template.get_all() -@comfytype(io_type="COMFY_COMBODYNAMIC_V3") -class ComboDynamic(ComfyTypeI): - class Input(DynamicInput): - def __init__(self, id: str): - pass + def get_all(self) -> list[Input]: + return [self] + self.template.get_all() -@comfytype(io_type="COMFY_MATCHTYPE_V3") -class MatchType(ComfyTypeIO): - class Template: - def __init__(self, template_id: str, allowed_types: _ComfyType | list[_ComfyType]): - self.template_id = template_id - self.allowed_types = [allowed_types] if isinstance(allowed_types, _ComfyType) else allowed_types + def validate(self): + self.template.validate() + + def expand_schema_for_dynamic(self, d: dict[str, Any], live_inputs: dict[str, Any], curr_prefix=''): + curr_prefix = f"{curr_prefix}{self.id}." + # need to remove self from expected inputs dictionary; replaced by template inputs in frontend + for inner_dict in d.values(): + if self.id in inner_dict: + del inner_dict[self.id] + self.template.expand_schema_for_dynamic(d, live_inputs, curr_prefix) + +@comfytype(io_type="COMFY_DYNAMICCOMBO_V3") +class DynamicCombo(ComfyTypeI): + Type = dict[str, Any] + + class Option: + def __init__(self, key: str, inputs: list[Input]): + self.key = key + self.inputs = inputs def as_dict(self): return { - "template_id": self.template_id, - "allowed_types": "".join(t.io_type for t in self.allowed_types), + "key": self.key, + "inputs": create_input_dict_v1(self.inputs), } class Input(DynamicInput): - def __init__(self, id: str, template: MatchType.Template, + def __init__(self, id: str, options: list[DynamicCombo.Option], display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, extra_dict=None): super().__init__(id, display_name, optional, tooltip, lazy, extra_dict) - self.template = template + self.options = options + + def expand_schema_for_dynamic(self, d: dict[str, Any], live_inputs: dict[str, Any], curr_prefix=''): + # check if dynamic input's id is in live_inputs + if self.id in live_inputs: + curr_prefix = f"{curr_prefix}{self.id}." + key = live_inputs[self.id] + selected_option = None + for option in self.options: + if option.key == key: + selected_option = option + break + if selected_option is not None: + add_to_input_dict_v1(d, selected_option.inputs, live_inputs, curr_prefix) + add_dynamic_id_mapping(d, selected_option.inputs, curr_prefix, self) def get_dynamic(self) -> list[Input]: - return [self] + return [input for option in self.options for input in option.inputs] + + def get_all(self) -> list[Input]: + return [self] + [input for option in self.options for input in option.inputs] def as_dict(self): return super().as_dict() | prune_dict({ - "template": self.template.as_dict(), + "options": [o.as_dict() for o in self.options], }) - class Output(DynamicOutput): - def __init__(self, id: str, template: MatchType.Template, display_name: str=None, tooltip: str=None, - is_output_list=False): - super().__init__(id, display_name, tooltip, is_output_list) - self.template = template + def validate(self): + # make sure all nested inputs are validated + for option in self.options: + for input in option.inputs: + input.validate() - def get_dynamic(self) -> list[Output]: - return [self] +@comfytype(io_type="COMFY_DYNAMICSLOT_V3") +class DynamicSlot(ComfyTypeI): + Type = dict[str, Any] + + class Input(DynamicInput): + def __init__(self, slot: Input, inputs: list[Input], + display_name: str=None, tooltip: str=None, lazy: bool=None, extra_dict=None): + assert(not isinstance(slot, DynamicInput)) + self.slot = copy.copy(slot) + self.slot.display_name = slot.display_name if slot.display_name is not None else display_name + optional = True + self.slot.tooltip = slot.tooltip if slot.tooltip is not None else tooltip + self.slot.lazy = slot.lazy if slot.lazy is not None else lazy + self.slot.extra_dict = slot.extra_dict if slot.extra_dict is not None else extra_dict + super().__init__(slot.id, self.slot.display_name, optional, self.slot.tooltip, self.slot.lazy, self.slot.extra_dict) + self.inputs = inputs + self.force_input = None + # force widget inputs to have no widgets, otherwise this would be awkward + if isinstance(self.slot, WidgetInput): + self.force_input = True + self.slot.force_input = True + + def expand_schema_for_dynamic(self, d: dict[str, Any], live_inputs: dict[str, Any], curr_prefix=''): + if self.id in live_inputs: + curr_prefix = f"{curr_prefix}{self.id}." + add_to_input_dict_v1(d, self.inputs, live_inputs, curr_prefix) + add_dynamic_id_mapping(d, [self.slot] + self.inputs, curr_prefix) + + def get_dynamic(self) -> list[Input]: + return [self.slot] + self.inputs + + def get_all(self) -> list[Input]: + return [self] + [self.slot] + self.inputs def as_dict(self): return super().as_dict() | prune_dict({ - "template": self.template.as_dict(), + "slotType": str(self.slot.get_io_type()), + "inputs": create_input_dict_v1(self.inputs), + "forceInput": self.force_input, }) + def validate(self): + self.slot.validate() + for input in self.inputs: + input.validate() + +def add_dynamic_id_mapping(d: dict[str, Any], inputs: list[Input], curr_prefix: str, self: DynamicInput=None): + dynamic = d.setdefault("dynamic_paths", {}) + if self is not None: + dynamic[self.id] = f"{curr_prefix}{self.id}" + for i in inputs: + if not isinstance(i, DynamicInput): + dynamic[f"{i.id}"] = f"{curr_prefix}{i.id}" + +class V3Data(TypedDict): + hidden_inputs: dict[str, Any] + dynamic_paths: dict[str, Any] class HiddenHolder: def __init__(self, unique_id: str, prompt: Any, @@ -984,6 +1163,7 @@ class NodeInfoV1: output_is_list: list[bool]=None output_name: list[str]=None output_tooltips: list[str]=None + output_matchtypes: list[str]=None name: str=None display_name: str=None description: str=None @@ -1061,7 +1241,11 @@ class Schema: '''Validate the schema: - verify ids on inputs and outputs are unique - both internally and in relation to each other ''' - input_ids = [i.id for i in self.inputs] if self.inputs is not None else [] + nested_inputs: list[Input] = [] + if self.inputs is not None: + for input in self.inputs: + nested_inputs.extend(input.get_all()) + input_ids = [i.id for i in nested_inputs] if nested_inputs is not None else [] output_ids = [o.id for o in self.outputs] if self.outputs is not None else [] input_set = set(input_ids) output_set = set(output_ids) @@ -1077,6 +1261,13 @@ class Schema: issues.append(f"Ids must be unique between inputs and outputs, but {intersection} are not.") if len(issues) > 0: raise ValueError("\n".join(issues)) + # validate inputs and outputs + if self.inputs is not None: + for input in self.inputs: + input.validate() + if self.outputs is not None: + for output in self.outputs: + output.validate() def finalize(self): """Add hidden based on selected schema options, and give outputs without ids default ids.""" @@ -1102,19 +1293,10 @@ class Schema: if output.id is None: output.id = f"_{i}_{output.io_type}_" - def get_v1_info(self, cls) -> NodeInfoV1: + def get_v1_info(self, cls, live_inputs: dict[str, Any]=None) -> NodeInfoV1: + # NOTE: live_inputs will not be used anymore very soon and this will be done another way # get V1 inputs - input = { - "required": {} - } - if self.inputs: - for i in self.inputs: - if isinstance(i, DynamicInput): - dynamic_inputs = i.get_dynamic() - for d in dynamic_inputs: - add_to_dict_v1(d, input) - else: - add_to_dict_v1(i, input) + input = create_input_dict_v1(self.inputs, live_inputs) if self.hidden: for hidden in self.hidden: input.setdefault("hidden", {})[hidden.name] = (hidden.value,) @@ -1123,12 +1305,24 @@ class Schema: output_is_list = [] output_name = [] output_tooltips = [] + output_matchtypes = [] + any_matchtypes = False if self.outputs: for o in self.outputs: output.append(o.io_type) output_is_list.append(o.is_output_list) output_name.append(o.display_name if o.display_name else o.io_type) output_tooltips.append(o.tooltip if o.tooltip else None) + # special handling for MatchType + if isinstance(o, MatchType.Output): + output_matchtypes.append(o.template.template_id) + any_matchtypes = True + else: + output_matchtypes.append(None) + + # clear out lists that are all None + if not any_matchtypes: + output_matchtypes = None info = NodeInfoV1( input=input, @@ -1137,6 +1331,7 @@ class Schema: output_is_list=output_is_list, output_name=output_name, output_tooltips=output_tooltips, + output_matchtypes=output_matchtypes, name=self.node_id, display_name=self.display_name, category=self.category, @@ -1182,16 +1377,57 @@ class Schema: return info -def add_to_dict_v1(i: Input, input: dict): +def create_input_dict_v1(inputs: list[Input], live_inputs: dict[str, Any]=None) -> dict: + input = { + "required": {} + } + add_to_input_dict_v1(input, inputs, live_inputs) + return input + +def add_to_input_dict_v1(d: dict[str, Any], inputs: list[Input], live_inputs: dict[str, Any]=None, curr_prefix=''): + for i in inputs: + if isinstance(i, DynamicInput): + add_to_dict_v1(i, d) + if live_inputs is not None: + i.expand_schema_for_dynamic(d, live_inputs, curr_prefix) + else: + add_to_dict_v1(i, d) + +def add_to_dict_v1(i: Input, d: dict, dynamic_dict: dict=None): key = "optional" if i.optional else "required" as_dict = i.as_dict() # for v1, we don't want to include the optional key as_dict.pop("optional", None) - input.setdefault(key, {})[i.id] = (i.get_io_type(), as_dict) + if dynamic_dict is None: + value = (i.get_io_type(), as_dict) + else: + value = (i.get_io_type(), as_dict, dynamic_dict) + d.setdefault(key, {})[i.id] = value def add_to_dict_v3(io: Input | Output, d: dict): d[io.id] = (io.get_io_type(), io.as_dict()) +def build_nested_inputs(values: dict[str, Any], v3_data: V3Data): + paths = v3_data.get("dynamic_paths", None) + if paths is None: + return values + values = values.copy() + result = {} + + for key, path in paths.items(): + parts = path.split(".") + current = result + + for i, p in enumerate(parts): + is_last = (i == len(parts) - 1) + + if is_last: + current[p] = values.pop(key, None) + else: + current = current.setdefault(p, {}) + + values.update(result) + return values class _ComfyNodeBaseInternal(_ComfyNodeInternal): @@ -1311,12 +1547,12 @@ class _ComfyNodeBaseInternal(_ComfyNodeInternal): @final @classmethod - def PREPARE_CLASS_CLONE(cls, hidden_inputs: dict) -> type[ComfyNode]: + def PREPARE_CLASS_CLONE(cls, v3_data: V3Data) -> type[ComfyNode]: """Creates clone of real node class to prevent monkey-patching.""" c_type: type[ComfyNode] = cls if is_class(cls) else type(cls) type_clone: type[ComfyNode] = shallow_clone_class(c_type) # set hidden - type_clone.hidden = HiddenHolder.from_dict(hidden_inputs) + type_clone.hidden = HiddenHolder.from_dict(v3_data["hidden_inputs"]) return type_clone @final @@ -1433,14 +1669,18 @@ class _ComfyNodeBaseInternal(_ComfyNodeInternal): @final @classmethod - def INPUT_TYPES(cls, include_hidden=True, return_schema=False) -> dict[str, dict] | tuple[dict[str, dict], Schema]: + def INPUT_TYPES(cls, include_hidden=True, return_schema=False, live_inputs=None) -> dict[str, dict] | tuple[dict[str, dict], Schema, V3Data]: schema = cls.FINALIZE_SCHEMA() - info = schema.get_v1_info(cls) + info = schema.get_v1_info(cls, live_inputs) input = info.input if not include_hidden: input.pop("hidden", None) if return_schema: - return input, schema + v3_data: V3Data = {} + dynamic = input.pop("dynamic_paths", None) + if dynamic is not None: + v3_data["dynamic_paths"] = dynamic + return input, schema, v3_data return input @final @@ -1513,7 +1753,7 @@ class ComfyNode(_ComfyNodeBaseInternal): raise NotImplementedError @classmethod - def validate_inputs(cls, **kwargs) -> bool: + def validate_inputs(cls, **kwargs) -> bool | str: """Optionally, define this function to validate inputs; equivalent to V1's VALIDATE_INPUTS.""" raise NotImplementedError @@ -1628,6 +1868,7 @@ __all__ = [ "StyleModel", "Gligen", "UpscaleModel", + "LatentUpscaleModel", "Audio", "Video", "SVG", @@ -1651,6 +1892,10 @@ __all__ = [ "SEGS", "AnyType", "MultiType", + # Dynamic Types + "MatchType", + # "DynamicCombo", + # "Autogrow", # Other classes "HiddenHolder", "Hidden", @@ -1661,4 +1906,5 @@ __all__ = [ "NodeOutput", "add_to_dict_v1", "add_to_dict_v3", + "V3Data", ] diff --git a/comfy_api/latest/_io_public.py b/comfy_api/latest/_io_public.py new file mode 100644 index 000000000..43c7680f3 --- /dev/null +++ b/comfy_api/latest/_io_public.py @@ -0,0 +1 @@ +from ._io import * # noqa: F403 diff --git a/comfy_api/latest/_ui_public.py b/comfy_api/latest/_ui_public.py new file mode 100644 index 000000000..85b11d78b --- /dev/null +++ b/comfy_api/latest/_ui_public.py @@ -0,0 +1 @@ +from ._ui import * # noqa: F403 diff --git a/comfy_api/v0_0_2/__init__.py b/comfy_api/v0_0_2/__init__.py index de0f95001..c4fa1d971 100644 --- a/comfy_api/v0_0_2/__init__.py +++ b/comfy_api/v0_0_2/__init__.py @@ -6,7 +6,7 @@ from comfy_api.latest import ( ) from typing import Type, TYPE_CHECKING from comfy_api.internal.async_to_sync import create_sync_class -from comfy_api.latest import io, ui, ComfyExtension #noqa: F401 +from comfy_api.latest import io, ui, IO, UI, ComfyExtension #noqa: F401 class ComfyAPIAdapter_v0_0_2(ComfyAPI_latest): @@ -42,4 +42,8 @@ __all__ = [ "InputImpl", "Types", "ComfyExtension", + "io", + "IO", + "ui", + "UI", ] diff --git a/comfy_execution/validation.py b/comfy_execution/validation.py index cec105fc9..24c0b4ed7 100644 --- a/comfy_execution/validation.py +++ b/comfy_execution/validation.py @@ -1,4 +1,5 @@ from __future__ import annotations +from comfy_api.latest import IO def validate_node_input( @@ -23,6 +24,11 @@ def validate_node_input( if not received_type != input_type: return True + # If the received type or input_type is a MatchType, we can return True immediately; + # validation for this is handled by the frontend + if received_type == IO.MatchType.io_type or input_type == IO.MatchType.io_type: + return True + # Not equal, and not strings if not isinstance(received_type, str) or not isinstance(input_type, str): return False diff --git a/comfy_extras/nodes_logic.py b/comfy_extras/nodes_logic.py new file mode 100644 index 000000000..95a6ba788 --- /dev/null +++ b/comfy_extras/nodes_logic.py @@ -0,0 +1,155 @@ +from typing import TypedDict +from typing_extensions import override +from comfy_api.latest import ComfyExtension, io +from comfy_api.latest import _io + + + +class SwitchNode(io.ComfyNode): + @classmethod + def define_schema(cls): + template = io.MatchType.Template("switch") + return io.Schema( + node_id="ComfySwitchNode", + display_name="Switch", + category="logic", + is_experimental=True, + inputs=[ + io.Boolean.Input("switch"), + io.MatchType.Input("on_false", template=template, lazy=True, optional=True), + io.MatchType.Input("on_true", template=template, lazy=True, optional=True), + ], + outputs=[ + io.MatchType.Output(template=template, display_name="output"), + ], + ) + + @classmethod + def check_lazy_status(cls, switch, on_false=..., on_true=...): + # We use ... instead of None, as None is passed for connected-but-unevaluated inputs. + # This trick allows us to ignore the value of the switch and still be able to run execute(). + + # One of the inputs may be missing, in which case we need to evaluate the other input + if on_false is ...: + return ["on_true"] + if on_true is ...: + return ["on_false"] + # Normal lazy switch operation + if switch and on_true is None: + return ["on_true"] + if not switch and on_false is None: + return ["on_false"] + + @classmethod + def validate_inputs(cls, switch, on_false=..., on_true=...): + # This check happens before check_lazy_status(), so we can eliminate the case where + # both inputs are missing. + if on_false is ... and on_true is ...: + return "At least one of on_false or on_true must be connected to Switch node" + return True + + @classmethod + def execute(cls, switch, on_true=..., on_false=...) -> io.NodeOutput: + if on_true is ...: + return io.NodeOutput(on_false) + if on_false is ...: + return io.NodeOutput(on_true) + return io.NodeOutput(on_true if switch else on_false) + + +class DCTestNode(io.ComfyNode): + class DCValues(TypedDict): + combo: str + string: str + integer: int + image: io.Image.Type + subcombo: dict[str] + + @classmethod + def define_schema(cls): + return io.Schema( + node_id="DCTestNode", + display_name="DCTest", + category="logic", + is_output_node=True, + inputs=[_io.DynamicCombo.Input("combo", options=[ + _io.DynamicCombo.Option("option1", [io.String.Input("string")]), + _io.DynamicCombo.Option("option2", [io.Int.Input("integer")]), + _io.DynamicCombo.Option("option3", [io.Image.Input("image")]), + _io.DynamicCombo.Option("option4", [ + _io.DynamicCombo.Input("subcombo", options=[ + _io.DynamicCombo.Option("opt1", [io.Float.Input("float_x"), io.Float.Input("float_y")]), + _io.DynamicCombo.Option("opt2", [io.Mask.Input("mask1", optional=True)]), + ]) + ])] + )], + outputs=[io.AnyType.Output()], + ) + + @classmethod + def execute(cls, combo: DCValues) -> io.NodeOutput: + combo_val = combo["combo"] + if combo_val == "option1": + return io.NodeOutput(combo["string"]) + elif combo_val == "option2": + return io.NodeOutput(combo["integer"]) + elif combo_val == "option3": + return io.NodeOutput(combo["image"]) + elif combo_val == "option4": + return io.NodeOutput(f"{combo['subcombo']}") + else: + raise ValueError(f"Invalid combo: {combo_val}") + + +class AutogrowNamesTestNode(io.ComfyNode): + @classmethod + def define_schema(cls): + template = _io.Autogrow.TemplateNames(input=io.Float.Input("float"), names=["a", "b", "c"]) + return io.Schema( + node_id="AutogrowNamesTestNode", + display_name="AutogrowNamesTest", + category="logic", + inputs=[ + _io.Autogrow.Input("autogrow", template=template) + ], + outputs=[io.String.Output()], + ) + + @classmethod + def execute(cls, autogrow: _io.Autogrow.Type) -> io.NodeOutput: + vals = list(autogrow.values()) + combined = ",".join([str(x) for x in vals]) + return io.NodeOutput(combined) + +class AutogrowPrefixTestNode(io.ComfyNode): + @classmethod + def define_schema(cls): + template = _io.Autogrow.TemplatePrefix(input=io.Float.Input("float"), prefix="float", min=1, max=10) + return io.Schema( + node_id="AutogrowPrefixTestNode", + display_name="AutogrowPrefixTest", + category="logic", + inputs=[ + _io.Autogrow.Input("autogrow", template=template) + ], + outputs=[io.String.Output()], + ) + + @classmethod + def execute(cls, autogrow: _io.Autogrow.Type) -> io.NodeOutput: + vals = list(autogrow.values()) + combined = ",".join([str(x) for x in vals]) + return io.NodeOutput(combined) + +class LogicExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[io.ComfyNode]]: + return [ + # SwitchNode, + # DCTestNode, + # AutogrowNamesTestNode, + # AutogrowPrefixTestNode, + ] + +async def comfy_entrypoint() -> LogicExtension: + return LogicExtension() diff --git a/execution.py b/execution.py index 17c77beab..c2186ac98 100644 --- a/execution.py +++ b/execution.py @@ -34,7 +34,7 @@ from comfy_execution.validation import validate_node_input from comfy_execution.progress import get_progress_state, reset_progress_state, add_progress_handler, WebUIProgressHandler from comfy_execution.utils import CurrentNodeContext from comfy_api.internal import _ComfyNodeInternal, _NodeOutputInternal, first_real_override, is_class, make_locked_method_func -from comfy_api.latest import io +from comfy_api.latest import io, _io class ExecutionResult(Enum): @@ -76,7 +76,7 @@ class IsChangedCache: return self.is_changed[node_id] # Intentionally do not use cached outputs here. We only want constants in IS_CHANGED - input_data_all, _, hidden_inputs = get_input_data(node["inputs"], class_def, node_id, None) + input_data_all, _, v3_data = get_input_data(node["inputs"], class_def, node_id, None) try: is_changed = await _async_map_node_over_list(self.prompt_id, node_id, class_def, input_data_all, is_changed_name) is_changed = await resolve_map_node_over_list_results(is_changed) @@ -146,8 +146,9 @@ SENSITIVE_EXTRA_DATA_KEYS = ("auth_token_comfy_org", "api_key_comfy_org") def get_input_data(inputs, class_def, unique_id, execution_list=None, dynprompt=None, extra_data={}): is_v3 = issubclass(class_def, _ComfyNodeInternal) + v3_data: io.V3Data = {} if is_v3: - valid_inputs, schema = class_def.INPUT_TYPES(include_hidden=False, return_schema=True) + valid_inputs, schema, v3_data = class_def.INPUT_TYPES(include_hidden=False, return_schema=True, live_inputs=inputs) else: valid_inputs = class_def.INPUT_TYPES() input_data_all = {} @@ -207,7 +208,8 @@ def get_input_data(inputs, class_def, unique_id, execution_list=None, dynprompt= input_data_all[x] = [extra_data.get("auth_token_comfy_org", None)] if h[x] == "API_KEY_COMFY_ORG": input_data_all[x] = [extra_data.get("api_key_comfy_org", None)] - return input_data_all, missing_keys, hidden_inputs_v3 + v3_data["hidden_inputs"] = hidden_inputs_v3 + return input_data_all, missing_keys, v3_data map_node_over_list = None #Don't hook this please @@ -223,7 +225,7 @@ async def resolve_map_node_over_list_results(results): raise exc return [x.result() if isinstance(x, asyncio.Task) else x for x in results] -async def _async_map_node_over_list(prompt_id, unique_id, obj, input_data_all, func, allow_interrupt=False, execution_block_cb=None, pre_execute_cb=None, hidden_inputs=None): +async def _async_map_node_over_list(prompt_id, unique_id, obj, input_data_all, func, allow_interrupt=False, execution_block_cb=None, pre_execute_cb=None, v3_data=None): # check if node wants the lists input_is_list = getattr(obj, "INPUT_IS_LIST", False) @@ -259,13 +261,16 @@ async def _async_map_node_over_list(prompt_id, unique_id, obj, input_data_all, f if is_class(obj): type_obj = obj obj.VALIDATE_CLASS() - class_clone = obj.PREPARE_CLASS_CLONE(hidden_inputs) + class_clone = obj.PREPARE_CLASS_CLONE(v3_data) # otherwise, use class instance to populate/reuse some fields else: type_obj = type(obj) type_obj.VALIDATE_CLASS() - class_clone = type_obj.PREPARE_CLASS_CLONE(hidden_inputs) + class_clone = type_obj.PREPARE_CLASS_CLONE(v3_data) f = make_locked_method_func(type_obj, func, class_clone) + # in case of dynamic inputs, restructure inputs to expected nested dict + if v3_data is not None: + inputs = _io.build_nested_inputs(inputs, v3_data) # V1 else: f = getattr(obj, func) @@ -320,8 +325,8 @@ def merge_result_data(results, obj): output.append([o[i] for o in results]) return output -async def get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=None, pre_execute_cb=None, hidden_inputs=None): - return_values = await _async_map_node_over_list(prompt_id, unique_id, obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, hidden_inputs=hidden_inputs) +async def get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=None, pre_execute_cb=None, v3_data=None): + return_values = await _async_map_node_over_list(prompt_id, unique_id, obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data) has_pending_task = any(isinstance(r, asyncio.Task) and not r.done() for r in return_values) if has_pending_task: return return_values, {}, False, has_pending_task @@ -460,7 +465,7 @@ async def execute(server, dynprompt, caches, current_item, extra_data, executed, has_subgraph = False else: get_progress_state().start_progress(unique_id) - input_data_all, missing_keys, hidden_inputs = get_input_data(inputs, class_def, unique_id, execution_list, dynprompt, extra_data) + input_data_all, missing_keys, v3_data = get_input_data(inputs, class_def, unique_id, execution_list, dynprompt, extra_data) if server.client_id is not None: server.last_node_id = display_node_id server.send_sync("executing", { "node": unique_id, "display_node": display_node_id, "prompt_id": prompt_id }, server.client_id) @@ -475,7 +480,7 @@ async def execute(server, dynprompt, caches, current_item, extra_data, executed, else: lazy_status_present = getattr(obj, "check_lazy_status", None) is not None if lazy_status_present: - required_inputs = await _async_map_node_over_list(prompt_id, unique_id, obj, input_data_all, "check_lazy_status", allow_interrupt=True, hidden_inputs=hidden_inputs) + required_inputs = await _async_map_node_over_list(prompt_id, unique_id, obj, input_data_all, "check_lazy_status", allow_interrupt=True, v3_data=v3_data) required_inputs = await resolve_map_node_over_list_results(required_inputs) required_inputs = set(sum([r for r in required_inputs if isinstance(r,list)], [])) required_inputs = [x for x in required_inputs if isinstance(x,str) and ( @@ -507,7 +512,7 @@ async def execute(server, dynprompt, caches, current_item, extra_data, executed, def pre_execute_cb(call_index): # TODO - How to handle this with async functions without contextvars (which requires Python 3.12)? GraphBuilder.set_default_prefix(unique_id, call_index, 0) - output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, hidden_inputs=hidden_inputs) + output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data) if has_pending_tasks: pending_async_nodes[unique_id] = output_data unblock = execution_list.add_external_block(unique_id) @@ -745,18 +750,17 @@ async def validate_inputs(prompt_id, prompt, item, validated): class_type = prompt[unique_id]['class_type'] obj_class = nodes.NODE_CLASS_MAPPINGS[class_type] - class_inputs = obj_class.INPUT_TYPES() - valid_inputs = set(class_inputs.get('required',{})).union(set(class_inputs.get('optional',{}))) - errors = [] valid = True validate_function_inputs = [] validate_has_kwargs = False if issubclass(obj_class, _ComfyNodeInternal): + class_inputs, _, _ = obj_class.INPUT_TYPES(include_hidden=False, return_schema=True, live_inputs=inputs) validate_function_name = "validate_inputs" validate_function = first_real_override(obj_class, validate_function_name) else: + class_inputs = obj_class.INPUT_TYPES() validate_function_name = "VALIDATE_INPUTS" validate_function = getattr(obj_class, validate_function_name, None) if validate_function is not None: @@ -765,6 +769,8 @@ async def validate_inputs(prompt_id, prompt, item, validated): validate_has_kwargs = argspec.varkw is not None received_types = {} + valid_inputs = set(class_inputs.get('required',{})).union(set(class_inputs.get('optional',{}))) + for x in valid_inputs: input_type, input_category, extra_info = get_input_info(obj_class, x, class_inputs) assert extra_info is not None @@ -935,7 +941,7 @@ async def validate_inputs(prompt_id, prompt, item, validated): continue if len(validate_function_inputs) > 0 or validate_has_kwargs: - input_data_all, _, hidden_inputs = get_input_data(inputs, obj_class, unique_id) + input_data_all, _, v3_data = get_input_data(inputs, obj_class, unique_id) input_filtered = {} for x in input_data_all: if x in validate_function_inputs or validate_has_kwargs: @@ -943,7 +949,7 @@ async def validate_inputs(prompt_id, prompt, item, validated): if 'input_types' in validate_function_inputs: input_filtered['input_types'] = [received_types] - ret = await _async_map_node_over_list(prompt_id, unique_id, obj_class, input_filtered, validate_function_name, hidden_inputs=hidden_inputs) + ret = await _async_map_node_over_list(prompt_id, unique_id, obj_class, input_filtered, validate_function_name, v3_data=v3_data) ret = await resolve_map_node_over_list_results(ret) for x in input_filtered: for i, r in enumerate(ret): diff --git a/nodes.py b/nodes.py index 4c910a34b..356aa63df 100644 --- a/nodes.py +++ b/nodes.py @@ -2355,6 +2355,7 @@ async def init_builtin_extra_nodes(): "nodes_easycache.py", "nodes_audio_encoder.py", "nodes_rope.py", + "nodes_logic.py", "nodes_nop.py", ] From 861817d22d2659099811b56005c9eaea18d64c73 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 2 Dec 2025 21:47:51 -0800 Subject: [PATCH 216/647] Fix issue with portable updater. (#11070) This should fix the problem with the portable updater not working with portables created from a separate branch on the repo. This does not affect any current portables who were all created on the master branch. --- .ci/update_windows/update.py | 7 +++++-- 1 file changed, 5 insertions(+), 2 deletions(-) diff --git a/.ci/update_windows/update.py b/.ci/update_windows/update.py index 51a263203..59ece5130 100755 --- a/.ci/update_windows/update.py +++ b/.ci/update_windows/update.py @@ -66,8 +66,10 @@ if branch is None: try: ref = repo.lookup_reference('refs/remotes/origin/master') except: - print("pulling.") # noqa: T201 - pull(repo) + print("fetching.") # noqa: T201 + for remote in repo.remotes: + if remote.name == "origin": + remote.fetch() ref = repo.lookup_reference('refs/remotes/origin/master') repo.checkout(ref) branch = repo.lookup_branch('master') @@ -149,3 +151,4 @@ try: shutil.copy(stable_update_script, stable_update_script_to) except: pass + From 519c9411653df99761053c30e101816e0ca3c24b Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Wed, 3 Dec 2025 17:28:45 +1000 Subject: [PATCH 217/647] Prs/lora reservations (reduce massive Lora reservations especially on Flux2) (#11069) * mp: only count the offload cost of math once This was previously bundling the combined weight storage and computation cost * ops: put all post async transfer compute on the main stream Some models have massive weights that need either complex dequantization or lora patching. Don't do these patchings on the offload stream, instead do them on the main stream to syncrhonize the potentially large vram spikes for these compute processes. This avoids having to assume a worst case scenario of multiple offload streams all spiking VRAM is parallel with whatever the main stream is doing. --- comfy/model_patcher.py | 4 ++-- comfy/ops.py | 39 ++++++++++++++++++++++----------------- 2 files changed, 24 insertions(+), 19 deletions(-) diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index 3eac77275..df2d8e827 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -704,7 +704,7 @@ class ModelPatcher: lowvram_weight = False - potential_offload = max(offload_buffer, module_offload_mem * (comfy.model_management.NUM_STREAMS + 1)) + potential_offload = max(offload_buffer, module_offload_mem + (comfy.model_management.NUM_STREAMS * module_mem)) lowvram_fits = mem_counter + module_mem + potential_offload < lowvram_model_memory weight_key = "{}.weight".format(n) @@ -883,7 +883,7 @@ class ModelPatcher: break module_offload_mem, module_mem, n, m, params = unload - potential_offload = (comfy.model_management.NUM_STREAMS + 1) * module_offload_mem + potential_offload = module_offload_mem + (comfy.model_management.NUM_STREAMS * module_mem) lowvram_possible = hasattr(m, "comfy_cast_weights") if hasattr(m, "comfy_patched_weights") and m.comfy_patched_weights == True: diff --git a/comfy/ops.py b/comfy/ops.py index 61a2f0754..eae434e68 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -111,22 +111,24 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, of if s.bias is not None: bias = comfy.model_management.cast_to(s.bias, bias_dtype, device, non_blocking=non_blocking, copy=bias_has_function, stream=offload_stream) - if bias_has_function: - with wf_context: - for f in s.bias_function: - bias = f(bias) + comfy.model_management.sync_stream(device, offload_stream) + + bias_a = bias + weight_a = weight + + if s.bias is not None: + for f in s.bias_function: + bias = f(bias) if weight_has_function or weight.dtype != dtype: - with wf_context: - weight = weight.to(dtype=dtype) - if isinstance(weight, QuantizedTensor): - weight = weight.dequantize() - for f in s.weight_function: - weight = f(weight) + weight = weight.to(dtype=dtype) + if isinstance(weight, QuantizedTensor): + weight = weight.dequantize() + for f in s.weight_function: + weight = f(weight) - comfy.model_management.sync_stream(device, offload_stream) if offloadable: - return weight, bias, offload_stream + return weight, bias, (offload_stream, weight_a, bias_a) else: #Legacy function signature return weight, bias @@ -135,13 +137,16 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, of def uncast_bias_weight(s, weight, bias, offload_stream): if offload_stream is None: return - if weight is not None: - device = weight.device + os, weight_a, bias_a = offload_stream + if os is None: + return + if weight_a is not None: + device = weight_a.device else: - if bias is None: + if bias_a is None: return - device = bias.device - offload_stream.wait_stream(comfy.model_management.current_stream(device)) + device = bias_a.device + os.wait_stream(comfy.model_management.current_stream(device)) class CastWeightBiasOp: From 19f2192d69d13445131b72ad1d87167f59b66fc4 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Wed, 3 Dec 2025 18:37:35 +0200 Subject: [PATCH 218/647] fix(V3-Schema): use empty list defaults for Schema.inputs/outputs/hidden to avoid None issues (#11083) --- comfy_api/latest/_io.py | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/comfy_api/latest/_io.py b/comfy_api/latest/_io.py index 257f07c42..866c3e0eb 100644 --- a/comfy_api/latest/_io.py +++ b/comfy_api/latest/_io.py @@ -5,7 +5,7 @@ import inspect from abc import ABC, abstractmethod from collections import Counter from collections.abc import Iterable -from dataclasses import asdict, dataclass +from dataclasses import asdict, dataclass, field from enum import Enum from typing import Any, Callable, Literal, TypedDict, TypeVar, TYPE_CHECKING from typing_extensions import NotRequired, final @@ -1199,9 +1199,9 @@ class Schema: """Display name of node.""" category: str = "sd" """The category of the node, as per the "Add Node" menu.""" - inputs: list[Input]=None - outputs: list[Output]=None - hidden: list[Hidden]=None + inputs: list[Input] = field(default_factory=list) + outputs: list[Output] = field(default_factory=list) + hidden: list[Hidden] = field(default_factory=list) description: str="" """Node description, shown as a tooltip when hovering over the node.""" is_input_list: bool = False From 87c104bfc1928f0b018a50f5867f425e10482929 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Wed, 3 Dec 2025 18:55:44 +0200 Subject: [PATCH 219/647] add support for "@image" reference format in Kling Omni API nodes (#11082) --- comfy_api_nodes/apis/kling_api.py | 30 +++++-- comfy_api_nodes/nodes_kling.py | 138 ++++++++++++++++++++++++++++-- 2 files changed, 155 insertions(+), 13 deletions(-) diff --git a/comfy_api_nodes/apis/kling_api.py b/comfy_api_nodes/apis/kling_api.py index 0a3b447c5..d8949f8ac 100644 --- a/comfy_api_nodes/apis/kling_api.py +++ b/comfy_api_nodes/apis/kling_api.py @@ -46,21 +46,41 @@ class TaskStatusVideoResult(BaseModel): url: str | None = Field(None, description="URL for generated video") -class TaskStatusVideoResults(BaseModel): +class TaskStatusImageResult(BaseModel): + index: int = Field(..., description="Image Number，0-9") + url: str = Field(..., description="URL for generated image") + + +class OmniTaskStatusResults(BaseModel): videos: list[TaskStatusVideoResult] | None = Field(None) + images: list[TaskStatusImageResult] | None = Field(None) -class TaskStatusVideoResponseData(BaseModel): +class OmniTaskStatusResponseData(BaseModel): created_at: int | None = Field(None, description="Task creation time") updated_at: int | None = Field(None, description="Task update time") task_status: str | None = None task_status_msg: str | None = Field(None, description="Additional failure reason. Only for polling endpoint.") task_id: str | None = Field(None, description="Task ID") - task_result: TaskStatusVideoResults | None = Field(None) + task_result: OmniTaskStatusResults | None = Field(None) -class TaskStatusVideoResponse(BaseModel): +class OmniTaskStatusResponse(BaseModel): code: int | None = Field(None, description="Error code") message: str | None = Field(None, description="Error message") request_id: str | None = Field(None, description="Request ID") - data: TaskStatusVideoResponseData | None = Field(None) + data: OmniTaskStatusResponseData | None = Field(None) + + +class OmniImageParamImage(BaseModel): + image: str = Field(...) + + +class OmniProImageRequest(BaseModel): + model_name: str = Field(..., description="kling-image-o1") + resolution: str = Field(..., description="'1k' or '2k'") + aspect_ratio: str | None = Field(...) + prompt: str = Field(...) + mode: str = Field("pro") + n: int | None = Field(1, le=9) + image_list: list[OmniImageParamImage] | None = Field(..., max_length=10) diff --git a/comfy_api_nodes/nodes_kling.py b/comfy_api_nodes/nodes_kling.py index 850c44db6..6c840dc47 100644 --- a/comfy_api_nodes/nodes_kling.py +++ b/comfy_api_nodes/nodes_kling.py @@ -6,6 +6,7 @@ For source of truth on the allowed permutations of request fields, please refere import logging import math +import re import torch from typing_extensions import override @@ -49,12 +50,14 @@ from comfy_api_nodes.apis import ( KlingSingleImageEffectModelName, ) from comfy_api_nodes.apis.kling_api import ( + OmniImageParamImage, OmniParamImage, OmniParamVideo, OmniProFirstLastFrameRequest, + OmniProImageRequest, OmniProReferences2VideoRequest, OmniProText2VideoRequest, - TaskStatusVideoResponse, + OmniTaskStatusResponse, ) from comfy_api_nodes.util import ( ApiEndpoint, @@ -210,7 +213,36 @@ VOICES_CONFIG = { } -async def finish_omni_video_task(cls: type[IO.ComfyNode], response: TaskStatusVideoResponse) -> IO.NodeOutput: +def normalize_omni_prompt_references(prompt: str) -> str: + """ + Rewrites Kling Omni-style placeholders used in the app, like: + + @image, @image1, @image2, ... @imageN + @video, @video1, @video2, ... @videoN + + into the API-compatible form: + + <<>>, <<>>, ... + <<>>, <<>>, ... + + This is a UX shim for ComfyUI so users can type the same syntax as in the Kling app. + """ + if not prompt: + return prompt + + def _image_repl(match): + return f"<<>>" + + def _video_repl(match): + return f"<<>>" + + # (? and not @imageFoo + prompt = re.sub(r"(?\d*)(?!\w)", _image_repl, prompt) + return re.sub(r"(?\d*)(?!\w)", _video_repl, prompt) + + +async def finish_omni_video_task(cls: type[IO.ComfyNode], response: OmniTaskStatusResponse) -> IO.NodeOutput: if response.code: raise RuntimeError( f"Kling request failed. Code: {response.code}, Message: {response.message}, Data: {response.data}" @@ -218,8 +250,9 @@ async def finish_omni_video_task(cls: type[IO.ComfyNode], response: TaskStatusVi final_response = await poll_op( cls, ApiEndpoint(path=f"/proxy/kling/v1/videos/omni-video/{response.data.task_id}"), - response_model=TaskStatusVideoResponse, + response_model=OmniTaskStatusResponse, status_extractor=lambda r: (r.data.task_status if r.data else None), + max_poll_attempts=160, ) return IO.NodeOutput(await download_url_to_video_output(final_response.data.task_result.videos[0].url)) @@ -801,7 +834,7 @@ class OmniProTextToVideoNode(IO.ComfyNode): response = await sync_op( cls, ApiEndpoint(path="/proxy/kling/v1/videos/omni-video", method="POST"), - response_model=TaskStatusVideoResponse, + response_model=OmniTaskStatusResponse, data=OmniProText2VideoRequest( model_name=model_name, prompt=prompt, @@ -864,6 +897,7 @@ class OmniProFirstLastFrameNode(IO.ComfyNode): end_frame: Input.Image | None = None, reference_images: Input.Image | None = None, ) -> IO.NodeOutput: + prompt = normalize_omni_prompt_references(prompt) validate_string(prompt, min_length=1, max_length=2500) if end_frame is not None and reference_images is not None: raise ValueError("The 'end_frame' input cannot be used simultaneously with 'reference_images'.") @@ -895,7 +929,7 @@ class OmniProFirstLastFrameNode(IO.ComfyNode): response = await sync_op( cls, ApiEndpoint(path="/proxy/kling/v1/videos/omni-video", method="POST"), - response_model=TaskStatusVideoResponse, + response_model=OmniTaskStatusResponse, data=OmniProFirstLastFrameRequest( model_name=model_name, prompt=prompt, @@ -950,6 +984,7 @@ class OmniProImageToVideoNode(IO.ComfyNode): duration: int, reference_images: Input.Image, ) -> IO.NodeOutput: + prompt = normalize_omni_prompt_references(prompt) validate_string(prompt, min_length=1, max_length=2500) if get_number_of_images(reference_images) > 7: raise ValueError("The maximum number of reference images is 7.") @@ -962,7 +997,7 @@ class OmniProImageToVideoNode(IO.ComfyNode): response = await sync_op( cls, ApiEndpoint(path="/proxy/kling/v1/videos/omni-video", method="POST"), - response_model=TaskStatusVideoResponse, + response_model=OmniTaskStatusResponse, data=OmniProReferences2VideoRequest( model_name=model_name, prompt=prompt, @@ -1023,6 +1058,7 @@ class OmniProVideoToVideoNode(IO.ComfyNode): keep_original_sound: bool, reference_images: Input.Image | None = None, ) -> IO.NodeOutput: + prompt = normalize_omni_prompt_references(prompt) validate_string(prompt, min_length=1, max_length=2500) validate_video_duration(reference_video, min_duration=3.0, max_duration=10.05) validate_video_dimensions(reference_video, min_width=720, min_height=720, max_width=2160, max_height=2160) @@ -1045,7 +1081,7 @@ class OmniProVideoToVideoNode(IO.ComfyNode): response = await sync_op( cls, ApiEndpoint(path="/proxy/kling/v1/videos/omni-video", method="POST"), - response_model=TaskStatusVideoResponse, + response_model=OmniTaskStatusResponse, data=OmniProReferences2VideoRequest( model_name=model_name, prompt=prompt, @@ -1103,6 +1139,7 @@ class OmniProEditVideoNode(IO.ComfyNode): keep_original_sound: bool, reference_images: Input.Image | None = None, ) -> IO.NodeOutput: + prompt = normalize_omni_prompt_references(prompt) validate_string(prompt, min_length=1, max_length=2500) validate_video_duration(video, min_duration=3.0, max_duration=10.05) validate_video_dimensions(video, min_width=720, min_height=720, max_width=2160, max_height=2160) @@ -1125,7 +1162,7 @@ class OmniProEditVideoNode(IO.ComfyNode): response = await sync_op( cls, ApiEndpoint(path="/proxy/kling/v1/videos/omni-video", method="POST"), - response_model=TaskStatusVideoResponse, + response_model=OmniTaskStatusResponse, data=OmniProReferences2VideoRequest( model_name=model_name, prompt=prompt, @@ -1138,6 +1175,90 @@ class OmniProEditVideoNode(IO.ComfyNode): return await finish_omni_video_task(cls, response) +class OmniProImageNode(IO.ComfyNode): + + @classmethod + def define_schema(cls) -> IO.Schema: + return IO.Schema( + node_id="KlingOmniProImageNode", + display_name="Kling Omni Image (Pro)", + category="api node/image/Kling", + description="Create or edit images with the latest model from Kling.", + inputs=[ + IO.Combo.Input("model_name", options=["kling-image-o1"]), + IO.String.Input( + "prompt", + multiline=True, + tooltip="A text prompt describing the image content. " + "This can include both positive and negative descriptions.", + ), + IO.Combo.Input("resolution", options=["1K", "2K"]), + IO.Combo.Input( + "aspect_ratio", + options=["16:9", "9:16", "1:1", "4:3", "3:4", "3:2", "2:3", "21:9"], + ), + IO.Image.Input( + "reference_images", + tooltip="Up to 10 additional reference images.", + optional=True, + ), + ], + outputs=[ + IO.Image.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + ) + + @classmethod + async def execute( + cls, + model_name: str, + prompt: str, + resolution: str, + aspect_ratio: str, + reference_images: Input.Image | None = None, + ) -> IO.NodeOutput: + prompt = normalize_omni_prompt_references(prompt) + validate_string(prompt, min_length=1, max_length=2500) + image_list: list[OmniImageParamImage] = [] + if reference_images is not None: + if get_number_of_images(reference_images) > 10: + raise ValueError("The maximum number of reference images is 10.") + for i in reference_images: + validate_image_dimensions(i, min_width=300, min_height=300) + validate_image_aspect_ratio(i, (1, 2.5), (2.5, 1)) + for i in await upload_images_to_comfyapi(cls, reference_images, wait_label="Uploading reference image"): + image_list.append(OmniImageParamImage(image=i)) + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/kling/v1/images/omni-image", method="POST"), + response_model=OmniTaskStatusResponse, + data=OmniProImageRequest( + model_name=model_name, + prompt=prompt, + resolution=resolution.lower(), + aspect_ratio=aspect_ratio, + image_list=image_list if image_list else None, + ), + ) + if response.code: + raise RuntimeError( + f"Kling request failed. Code: {response.code}, Message: {response.message}, Data: {response.data}" + ) + final_response = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/kling/v1/images/omni-image/{response.data.task_id}"), + response_model=OmniTaskStatusResponse, + status_extractor=lambda r: (r.data.task_status if r.data else None), + ) + return IO.NodeOutput(await download_url_to_image_tensor(final_response.data.task_result.images[0].url)) + + class KlingCameraControlT2VNode(IO.ComfyNode): """ Kling Text to Video Camera Control Node. This node is a text to video node, but it supports controlling the camera. @@ -1935,6 +2056,7 @@ class KlingExtension(ComfyExtension): OmniProImageToVideoNode, OmniProVideoToVideoNode, OmniProEditVideoNode, + # OmniProImageNode, # need support from backend ] From 440268d3940eb14a01595439bbc05c4aacde9c72 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Wed, 3 Dec 2025 23:52:31 +0200 Subject: [PATCH 220/647] convert nodes_load_3d.py to V3 schema (#10990) --- comfy_api/latest/_ui.py | 13 +++- comfy_extras/nodes_load_3d.py | 127 ++++++++++++++++------------------ 2 files changed, 71 insertions(+), 69 deletions(-) diff --git a/comfy_api/latest/_ui.py b/comfy_api/latest/_ui.py index b0bbabe2a..6d1bea599 100644 --- a/comfy_api/latest/_ui.py +++ b/comfy_api/latest/_ui.py @@ -3,6 +3,7 @@ from __future__ import annotations import json import os import random +import uuid from io import BytesIO from typing import Type @@ -436,9 +437,19 @@ class PreviewUI3D(_UIOutput): def __init__(self, model_file, camera_info, **kwargs): self.model_file = model_file self.camera_info = camera_info + self.bg_image_path = None + bg_image = kwargs.get("bg_image", None) + if bg_image is not None: + img_array = (bg_image[0].cpu().numpy() * 255).astype(np.uint8) + img = PILImage.fromarray(img_array) + temp_dir = folder_paths.get_temp_directory() + filename = f"bg_{uuid.uuid4().hex}.png" + bg_image_path = os.path.join(temp_dir, filename) + img.save(bg_image_path, compress_level=1) + self.bg_image_path = f"temp/{filename}" def as_dict(self): - return {"result": [self.model_file, self.camera_info]} + return {"result": [self.model_file, self.camera_info, self.bg_image_path]} class PreviewText(_UIOutput): diff --git a/comfy_extras/nodes_load_3d.py b/comfy_extras/nodes_load_3d.py index 54c66ef68..545588ef8 100644 --- a/comfy_extras/nodes_load_3d.py +++ b/comfy_extras/nodes_load_3d.py @@ -2,22 +2,18 @@ import nodes import folder_paths import os -from comfy.comfy_types import IO -from comfy_api.input_impl import VideoFromFile +from typing_extensions import override +from comfy_api.latest import IO, ComfyExtension, InputImpl, UI from pathlib import Path -from PIL import Image -import numpy as np - -import uuid def normalize_path(path): return path.replace('\\', '/') -class Load3D(): +class Load3D(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): + def define_schema(cls): input_dir = os.path.join(folder_paths.get_input_directory(), "3d") os.makedirs(input_dir, exist_ok=True) @@ -30,23 +26,29 @@ class Load3D(): for file_path in input_path.rglob("*") if file_path.suffix.lower() in {'.gltf', '.glb', '.obj', '.fbx', '.stl'} ] + return IO.Schema( + node_id="Load3D", + display_name="Load 3D & Animation", + category="3d", + is_experimental=True, + inputs=[ + IO.Combo.Input("model_file", options=sorted(files), upload=IO.UploadType.model), + IO.Load3D.Input("image"), + IO.Int.Input("width", default=1024, min=1, max=4096, step=1), + IO.Int.Input("height", default=1024, min=1, max=4096, step=1), + ], + outputs=[ + IO.Image.Output(display_name="image"), + IO.Mask.Output(display_name="mask"), + IO.String.Output(display_name="mesh_path"), + IO.Image.Output(display_name="normal"), + IO.Load3DCamera.Output(display_name="camera_info"), + IO.Video.Output(display_name="recording_video"), + ], + ) - return {"required": { - "model_file": (sorted(files), {"file_upload": True}), - "image": ("LOAD_3D", {}), - "width": ("INT", {"default": 1024, "min": 1, "max": 4096, "step": 1}), - "height": ("INT", {"default": 1024, "min": 1, "max": 4096, "step": 1}), - }} - - RETURN_TYPES = ("IMAGE", "MASK", "STRING", "IMAGE", "LOAD3D_CAMERA", IO.VIDEO) - RETURN_NAMES = ("image", "mask", "mesh_path", "normal", "camera_info", "recording_video") - - FUNCTION = "process" - EXPERIMENTAL = True - - CATEGORY = "3d" - - def process(self, model_file, image, **kwargs): + @classmethod + def execute(cls, model_file, image, **kwargs) -> IO.NodeOutput: image_path = folder_paths.get_annotated_filepath(image['image']) mask_path = folder_paths.get_annotated_filepath(image['mask']) normal_path = folder_paths.get_annotated_filepath(image['normal']) @@ -61,58 +63,47 @@ class Load3D(): if image['recording'] != "": recording_video_path = folder_paths.get_annotated_filepath(image['recording']) - video = VideoFromFile(recording_video_path) + video = InputImpl.VideoFromFile(recording_video_path) - return output_image, output_mask, model_file, normal_image, image['camera_info'], video + return IO.NodeOutput(output_image, output_mask, model_file, normal_image, image['camera_info'], video) -class Preview3D(): + process = execute # TODO: remove + + +class Preview3D(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": { - "model_file": ("STRING", {"default": "", "multiline": False}), - }, - "optional": { - "camera_info": ("LOAD3D_CAMERA", {}), - "bg_image": ("IMAGE", {}) - }} + def define_schema(cls): + return IO.Schema( + node_id="Preview3D", + display_name="Preview 3D & Animation", + category="3d", + is_experimental=True, + is_output_node=True, + inputs=[ + IO.String.Input("model_file", default="", multiline=False), + IO.Load3DCamera.Input("camera_info", optional=True), + IO.Image.Input("bg_image", optional=True), + ], + outputs=[], + ) - OUTPUT_NODE = True - RETURN_TYPES = () - - CATEGORY = "3d" - - FUNCTION = "process" - EXPERIMENTAL = True - - def process(self, model_file, **kwargs): + @classmethod + def execute(cls, model_file, **kwargs) -> IO.NodeOutput: camera_info = kwargs.get("camera_info", None) bg_image = kwargs.get("bg_image", None) + return IO.NodeOutput(ui=UI.PreviewUI3D(model_file, camera_info, bg_image=bg_image)) - bg_image_path = None - if bg_image is not None: + process = execute # TODO: remove - img_array = (bg_image[0].cpu().numpy() * 255).astype(np.uint8) - img = Image.fromarray(img_array) - temp_dir = folder_paths.get_temp_directory() - filename = f"bg_{uuid.uuid4().hex}.png" - bg_image_path = os.path.join(temp_dir, filename) - img.save(bg_image_path, compress_level=1) +class Load3DExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[IO.ComfyNode]]: + return [ + Load3D, + Preview3D, + ] - bg_image_path = f"temp/{filename}" - return { - "ui": { - "result": [model_file, camera_info, bg_image_path] - } - } - -NODE_CLASS_MAPPINGS = { - "Load3D": Load3D, - "Preview3D": Preview3D, -} - -NODE_DISPLAY_NAME_MAPPINGS = { - "Load3D": "Load 3D & Animation", - "Preview3D": "Preview 3D & Animation", -} +async def comfy_entrypoint() -> Load3DExtension: + return Load3DExtension() From dce518c2b4f99634b5fdde1924d9b0bd468fe1ce Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Thu, 4 Dec 2025 03:35:04 +0200 Subject: [PATCH 221/647] convert nodes_audio.py to V3 schema (#10798) --- comfy_api/latest/_ui.py | 9 +- comfy_extras/nodes_audio.py | 744 ++++++++++++++++++------------------ 2 files changed, 382 insertions(+), 371 deletions(-) diff --git a/comfy_api/latest/_ui.py b/comfy_api/latest/_ui.py index 6d1bea599..5a75a3aae 100644 --- a/comfy_api/latest/_ui.py +++ b/comfy_api/latest/_ui.py @@ -319,9 +319,10 @@ class AudioSaveHelper: for key, value in metadata.items(): output_container.metadata[key] = value + layout = "mono" if waveform.shape[0] == 1 else "stereo" # Set up the output stream with appropriate properties if format == "opus": - out_stream = output_container.add_stream("libopus", rate=sample_rate) + out_stream = output_container.add_stream("libopus", rate=sample_rate, layout=layout) if quality == "64k": out_stream.bit_rate = 64000 elif quality == "96k": @@ -333,7 +334,7 @@ class AudioSaveHelper: elif quality == "320k": out_stream.bit_rate = 320000 elif format == "mp3": - out_stream = output_container.add_stream("libmp3lame", rate=sample_rate) + out_stream = output_container.add_stream("libmp3lame", rate=sample_rate, layout=layout) if quality == "V0": # TODO i would really love to support V3 and V5 but there doesn't seem to be a way to set the qscale level, the property below is a bool out_stream.codec_context.qscale = 1 @@ -342,12 +343,12 @@ class AudioSaveHelper: elif quality == "320k": out_stream.bit_rate = 320000 else: # format == "flac": - out_stream = output_container.add_stream("flac", rate=sample_rate) + out_stream = output_container.add_stream("flac", rate=sample_rate, layout=layout) frame = av.AudioFrame.from_ndarray( waveform.movedim(0, 1).reshape(1, -1).float().numpy(), format="flt", - layout="mono" if waveform.shape[0] == 1 else "stereo", + layout=layout, ) frame.sample_rate = sample_rate frame.pts = 0 diff --git a/comfy_extras/nodes_audio.py b/comfy_extras/nodes_audio.py index 2ed7e0b22..812301fb7 100644 --- a/comfy_extras/nodes_audio.py +++ b/comfy_extras/nodes_audio.py @@ -6,65 +6,80 @@ import torch import comfy.model_management import folder_paths import os -import io -import json -import random import hashlib import node_helpers import logging -from comfy.cli_args import args -from comfy.comfy_types import FileLocator +from typing_extensions import override +from comfy_api.latest import ComfyExtension, IO, UI -class EmptyLatentAudio: - def __init__(self): - self.device = comfy.model_management.intermediate_device() +class EmptyLatentAudio(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="EmptyLatentAudio", + display_name="Empty Latent Audio", + category="latent/audio", + inputs=[ + IO.Float.Input("seconds", default=47.6, min=1.0, max=1000.0, step=0.1), + IO.Int.Input( + "batch_size", default=1, min=1, max=4096, tooltip="The number of latent images in the batch." + ), + ], + outputs=[IO.Latent.Output()], + ) @classmethod - def INPUT_TYPES(s): - return {"required": {"seconds": ("FLOAT", {"default": 47.6, "min": 1.0, "max": 1000.0, "step": 0.1}), - "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096, "tooltip": "The number of latent images in the batch."}), - }} - RETURN_TYPES = ("LATENT",) - FUNCTION = "generate" - - CATEGORY = "latent/audio" - - def generate(self, seconds, batch_size): + def execute(cls, seconds, batch_size) -> IO.NodeOutput: length = round((seconds * 44100 / 2048) / 2) * 2 - latent = torch.zeros([batch_size, 64, length], device=self.device) - return ({"samples":latent, "type": "audio"}, ) + latent = torch.zeros([batch_size, 64, length], device=comfy.model_management.intermediate_device()) + return IO.NodeOutput({"samples":latent, "type": "audio"}) -class ConditioningStableAudio: + generate = execute # TODO: remove + + +class ConditioningStableAudio(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": {"positive": ("CONDITIONING", ), - "negative": ("CONDITIONING", ), - "seconds_start": ("FLOAT", {"default": 0.0, "min": 0.0, "max": 1000.0, "step": 0.1}), - "seconds_total": ("FLOAT", {"default": 47.0, "min": 0.0, "max": 1000.0, "step": 0.1}), - }} + def define_schema(cls): + return IO.Schema( + node_id="ConditioningStableAudio", + category="conditioning", + inputs=[ + IO.Conditioning.Input("positive"), + IO.Conditioning.Input("negative"), + IO.Float.Input("seconds_start", default=0.0, min=0.0, max=1000.0, step=0.1), + IO.Float.Input("seconds_total", default=47.0, min=0.0, max=1000.0, step=0.1), + ], + outputs=[ + IO.Conditioning.Output(display_name="positive"), + IO.Conditioning.Output(display_name="negative"), + ], + ) - RETURN_TYPES = ("CONDITIONING","CONDITIONING") - RETURN_NAMES = ("positive", "negative") - - FUNCTION = "append" - - CATEGORY = "conditioning" - - def append(self, positive, negative, seconds_start, seconds_total): + @classmethod + def execute(cls, positive, negative, seconds_start, seconds_total) -> IO.NodeOutput: positive = node_helpers.conditioning_set_values(positive, {"seconds_start": seconds_start, "seconds_total": seconds_total}) negative = node_helpers.conditioning_set_values(negative, {"seconds_start": seconds_start, "seconds_total": seconds_total}) - return (positive, negative) + return IO.NodeOutput(positive, negative) -class VAEEncodeAudio: + append = execute # TODO: remove + + +class VAEEncodeAudio(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": { "audio": ("AUDIO", ), "vae": ("VAE", )}} - RETURN_TYPES = ("LATENT",) - FUNCTION = "encode" + def define_schema(cls): + return IO.Schema( + node_id="VAEEncodeAudio", + display_name="VAE Encode Audio", + category="latent/audio", + inputs=[ + IO.Audio.Input("audio"), + IO.Vae.Input("vae"), + ], + outputs=[IO.Latent.Output()], + ) - CATEGORY = "latent/audio" - - def encode(self, vae, audio): + @classmethod + def execute(cls, vae, audio) -> IO.NodeOutput: sample_rate = audio["sample_rate"] if 44100 != sample_rate: waveform = torchaudio.functional.resample(audio["waveform"], sample_rate, 44100) @@ -72,213 +87,134 @@ class VAEEncodeAudio: waveform = audio["waveform"] t = vae.encode(waveform.movedim(1, -1)) - return ({"samples":t}, ) + return IO.NodeOutput({"samples":t}) -class VAEDecodeAudio: + encode = execute # TODO: remove + + +class VAEDecodeAudio(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": { "samples": ("LATENT", ), "vae": ("VAE", )}} - RETURN_TYPES = ("AUDIO",) - FUNCTION = "decode" + def define_schema(cls): + return IO.Schema( + node_id="VAEDecodeAudio", + display_name="VAE Decode Audio", + category="latent/audio", + inputs=[ + IO.Latent.Input("samples"), + IO.Vae.Input("vae"), + ], + outputs=[IO.Audio.Output()], + ) - CATEGORY = "latent/audio" - - def decode(self, vae, samples): + @classmethod + def execute(cls, vae, samples) -> IO.NodeOutput: audio = vae.decode(samples["samples"]).movedim(-1, 1) std = torch.std(audio, dim=[1,2], keepdim=True) * 5.0 std[std < 1.0] = 1.0 audio /= std - return ({"waveform": audio, "sample_rate": 44100}, ) + return IO.NodeOutput({"waveform": audio, "sample_rate": 44100}) + + decode = execute # TODO: remove -def save_audio(self, audio, filename_prefix="ComfyUI", format="flac", prompt=None, extra_pnginfo=None, quality="128k"): - - filename_prefix += self.prefix_append - full_output_folder, filename, counter, subfolder, filename_prefix = folder_paths.get_save_image_path(filename_prefix, self.output_dir) - results: list[FileLocator] = [] - - # Prepare metadata dictionary - metadata = {} - if not args.disable_metadata: - if prompt is not None: - metadata["prompt"] = json.dumps(prompt) - if extra_pnginfo is not None: - for x in extra_pnginfo: - metadata[x] = json.dumps(extra_pnginfo[x]) - - # Opus supported sample rates - OPUS_RATES = [8000, 12000, 16000, 24000, 48000] - - for (batch_number, waveform) in enumerate(audio["waveform"].cpu()): - filename_with_batch_num = filename.replace("%batch_num%", str(batch_number)) - file = f"{filename_with_batch_num}_{counter:05}_.{format}" - output_path = os.path.join(full_output_folder, file) - - # Use original sample rate initially - sample_rate = audio["sample_rate"] - - # Handle Opus sample rate requirements - if format == "opus": - if sample_rate > 48000: - sample_rate = 48000 - elif sample_rate not in OPUS_RATES: - # Find the next highest supported rate - for rate in sorted(OPUS_RATES): - if rate > sample_rate: - sample_rate = rate - break - if sample_rate not in OPUS_RATES: # Fallback if still not supported - sample_rate = 48000 - - # Resample if necessary - if sample_rate != audio["sample_rate"]: - waveform = torchaudio.functional.resample(waveform, audio["sample_rate"], sample_rate) - - # Create output with specified format - output_buffer = io.BytesIO() - output_container = av.open(output_buffer, mode='w', format=format) - - # Set metadata on the container - for key, value in metadata.items(): - output_container.metadata[key] = value - - layout = 'mono' if waveform.shape[0] == 1 else 'stereo' - # Set up the output stream with appropriate properties - if format == "opus": - out_stream = output_container.add_stream("libopus", rate=sample_rate, layout=layout) - if quality == "64k": - out_stream.bit_rate = 64000 - elif quality == "96k": - out_stream.bit_rate = 96000 - elif quality == "128k": - out_stream.bit_rate = 128000 - elif quality == "192k": - out_stream.bit_rate = 192000 - elif quality == "320k": - out_stream.bit_rate = 320000 - elif format == "mp3": - out_stream = output_container.add_stream("libmp3lame", rate=sample_rate, layout=layout) - if quality == "V0": - #TODO i would really love to support V3 and V5 but there doesn't seem to be a way to set the qscale level, the property below is a bool - out_stream.codec_context.qscale = 1 - elif quality == "128k": - out_stream.bit_rate = 128000 - elif quality == "320k": - out_stream.bit_rate = 320000 - else: #format == "flac": - out_stream = output_container.add_stream("flac", rate=sample_rate, layout=layout) - - frame = av.AudioFrame.from_ndarray(waveform.movedim(0, 1).reshape(1, -1).float().numpy(), format='flt', layout=layout) - frame.sample_rate = sample_rate - frame.pts = 0 - output_container.mux(out_stream.encode(frame)) - - # Flush encoder - output_container.mux(out_stream.encode(None)) - - # Close containers - output_container.close() - - # Write the output to file - output_buffer.seek(0) - with open(output_path, 'wb') as f: - f.write(output_buffer.getbuffer()) - - results.append({ - "filename": file, - "subfolder": subfolder, - "type": self.type - }) - counter += 1 - - return { "ui": { "audio": results } } - -class SaveAudio: - def __init__(self): - self.output_dir = folder_paths.get_output_directory() - self.type = "output" - self.prefix_append = "" +class SaveAudio(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="SaveAudio", + display_name="Save Audio (FLAC)", + category="audio", + inputs=[ + IO.Audio.Input("audio"), + IO.String.Input("filename_prefix", default="audio/ComfyUI"), + ], + hidden=[IO.Hidden.prompt, IO.Hidden.extra_pnginfo], + is_output_node=True, + ) @classmethod - def INPUT_TYPES(s): - return {"required": { "audio": ("AUDIO", ), - "filename_prefix": ("STRING", {"default": "audio/ComfyUI"}), - }, - "hidden": {"prompt": "PROMPT", "extra_pnginfo": "EXTRA_PNGINFO"}, - } + def execute(cls, audio, filename_prefix="ComfyUI", format="flac") -> IO.NodeOutput: + return IO.NodeOutput( + ui=UI.AudioSaveHelper.get_save_audio_ui(audio, filename_prefix=filename_prefix, cls=cls, format=format) + ) - RETURN_TYPES = () - FUNCTION = "save_flac" + save_flac = execute # TODO: remove - OUTPUT_NODE = True - CATEGORY = "audio" - - def save_flac(self, audio, filename_prefix="ComfyUI", format="flac", prompt=None, extra_pnginfo=None): - return save_audio(self, audio, filename_prefix, format, prompt, extra_pnginfo) - -class SaveAudioMP3: - def __init__(self): - self.output_dir = folder_paths.get_output_directory() - self.type = "output" - self.prefix_append = "" +class SaveAudioMP3(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="SaveAudioMP3", + display_name="Save Audio (MP3)", + category="audio", + inputs=[ + IO.Audio.Input("audio"), + IO.String.Input("filename_prefix", default="audio/ComfyUI"), + IO.Combo.Input("quality", options=["V0", "128k", "320k"], default="V0"), + ], + hidden=[IO.Hidden.prompt, IO.Hidden.extra_pnginfo], + is_output_node=True, + ) @classmethod - def INPUT_TYPES(s): - return {"required": { "audio": ("AUDIO", ), - "filename_prefix": ("STRING", {"default": "audio/ComfyUI"}), - "quality": (["V0", "128k", "320k"], {"default": "V0"}), - }, - "hidden": {"prompt": "PROMPT", "extra_pnginfo": "EXTRA_PNGINFO"}, - } + def execute(cls, audio, filename_prefix="ComfyUI", format="mp3", quality="128k") -> IO.NodeOutput: + return IO.NodeOutput( + ui=UI.AudioSaveHelper.get_save_audio_ui( + audio, filename_prefix=filename_prefix, cls=cls, format=format, quality=quality + ) + ) - RETURN_TYPES = () - FUNCTION = "save_mp3" + save_mp3 = execute # TODO: remove - OUTPUT_NODE = True - CATEGORY = "audio" - - def save_mp3(self, audio, filename_prefix="ComfyUI", format="mp3", prompt=None, extra_pnginfo=None, quality="128k"): - return save_audio(self, audio, filename_prefix, format, prompt, extra_pnginfo, quality) - -class SaveAudioOpus: - def __init__(self): - self.output_dir = folder_paths.get_output_directory() - self.type = "output" - self.prefix_append = "" +class SaveAudioOpus(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="SaveAudioOpus", + display_name="Save Audio (Opus)", + category="audio", + inputs=[ + IO.Audio.Input("audio"), + IO.String.Input("filename_prefix", default="audio/ComfyUI"), + IO.Combo.Input("quality", options=["64k", "96k", "128k", "192k", "320k"], default="128k"), + ], + hidden=[IO.Hidden.prompt, IO.Hidden.extra_pnginfo], + is_output_node=True, + ) @classmethod - def INPUT_TYPES(s): - return {"required": { "audio": ("AUDIO", ), - "filename_prefix": ("STRING", {"default": "audio/ComfyUI"}), - "quality": (["64k", "96k", "128k", "192k", "320k"], {"default": "128k"}), - }, - "hidden": {"prompt": "PROMPT", "extra_pnginfo": "EXTRA_PNGINFO"}, - } + def execute(cls, audio, filename_prefix="ComfyUI", format="opus", quality="V3") -> IO.NodeOutput: + return IO.NodeOutput( + ui=UI.AudioSaveHelper.get_save_audio_ui( + audio, filename_prefix=filename_prefix, cls=cls, format=format, quality=quality + ) + ) - RETURN_TYPES = () - FUNCTION = "save_opus" + save_opus = execute # TODO: remove - OUTPUT_NODE = True - CATEGORY = "audio" - - def save_opus(self, audio, filename_prefix="ComfyUI", format="opus", prompt=None, extra_pnginfo=None, quality="V3"): - return save_audio(self, audio, filename_prefix, format, prompt, extra_pnginfo, quality) - -class PreviewAudio(SaveAudio): - def __init__(self): - self.output_dir = folder_paths.get_temp_directory() - self.type = "temp" - self.prefix_append = "_temp_" + ''.join(random.choice("abcdefghijklmnopqrstupvxyz") for x in range(5)) +class PreviewAudio(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="PreviewAudio", + display_name="Preview Audio", + category="audio", + inputs=[ + IO.Audio.Input("audio"), + ], + hidden=[IO.Hidden.prompt, IO.Hidden.extra_pnginfo], + is_output_node=True, + ) @classmethod - def INPUT_TYPES(s): - return {"required": - {"audio": ("AUDIO", ), }, - "hidden": {"prompt": "PROMPT", "extra_pnginfo": "EXTRA_PNGINFO"}, - } + def execute(cls, audio) -> IO.NodeOutput: + return IO.NodeOutput(ui=UI.PreviewAudio(audio, cls=cls)) + + save_flac = execute # TODO: remove + def f32_pcm(wav: torch.Tensor) -> torch.Tensor: """Convert audio to float 32 bits PCM format.""" @@ -316,26 +252,30 @@ def load(filepath: str) -> tuple[torch.Tensor, int]: wav = f32_pcm(wav) return wav, sr -class LoadAudio: +class LoadAudio(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): + def define_schema(cls): input_dir = folder_paths.get_input_directory() files = folder_paths.filter_files_content_types(os.listdir(input_dir), ["audio", "video"]) - return {"required": {"audio": (sorted(files), {"audio_upload": True})}} + return IO.Schema( + node_id="LoadAudio", + display_name="Load Audio", + category="audio", + inputs=[ + IO.Combo.Input("audio", upload=IO.UploadType.audio, options=sorted(files)), + ], + outputs=[IO.Audio.Output()], + ) - CATEGORY = "audio" - - RETURN_TYPES = ("AUDIO", ) - FUNCTION = "load" - - def load(self, audio): + @classmethod + def execute(cls, audio) -> IO.NodeOutput: audio_path = folder_paths.get_annotated_filepath(audio) waveform, sample_rate = load(audio_path) audio = {"waveform": waveform.unsqueeze(0), "sample_rate": sample_rate} - return (audio, ) + return IO.NodeOutput(audio) @classmethod - def IS_CHANGED(s, audio): + def fingerprint_inputs(cls, audio): image_path = folder_paths.get_annotated_filepath(audio) m = hashlib.sha256() with open(image_path, 'rb') as f: @@ -343,46 +283,69 @@ class LoadAudio: return m.digest().hex() @classmethod - def VALIDATE_INPUTS(s, audio): + def validate_inputs(cls, audio): if not folder_paths.exists_annotated_filepath(audio): return "Invalid audio file: {}".format(audio) return True -class RecordAudio: + load = execute # TODO: remove + + +class RecordAudio(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": {"audio": ("AUDIO_RECORD", {})}} + def define_schema(cls): + return IO.Schema( + node_id="RecordAudio", + display_name="Record Audio", + category="audio", + inputs=[ + IO.Custom("AUDIO_RECORD").Input("audio"), + ], + outputs=[IO.Audio.Output()], + ) - CATEGORY = "audio" - - RETURN_TYPES = ("AUDIO", ) - FUNCTION = "load" - - def load(self, audio): + @classmethod + def execute(cls, audio) -> IO.NodeOutput: audio_path = folder_paths.get_annotated_filepath(audio) waveform, sample_rate = load(audio_path) audio = {"waveform": waveform.unsqueeze(0), "sample_rate": sample_rate} - return (audio, ) + return IO.NodeOutput(audio) + + load = execute # TODO: remove -class TrimAudioDuration: +class TrimAudioDuration(IO.ComfyNode): @classmethod - def INPUT_TYPES(cls): - return { - "required": { - "audio": ("AUDIO",), - "start_index": ("FLOAT", {"default": 0.0, "min": -0xffffffffffffffff, "max": 0xffffffffffffffff, "step": 0.01, "tooltip": "Start time in seconds, can be negative to count from the end (supports sub-seconds)."}), - "duration": ("FLOAT", {"default": 60.0, "min": 0.0, "step": 0.01, "tooltip": "Duration in seconds"}), - }, - } + def define_schema(cls): + return IO.Schema( + node_id="TrimAudioDuration", + display_name="Trim Audio Duration", + description="Trim audio tensor into chosen time range.", + category="audio", + inputs=[ + IO.Audio.Input("audio"), + IO.Float.Input( + "start_index", + default=0.0, + min=-0xffffffffffffffff, + max=0xffffffffffffffff, + step=0.01, + tooltip="Start time in seconds, can be negative to count from the end (supports sub-seconds).", + ), + IO.Float.Input( + "duration", + default=60.0, + min=0.0, + step=0.01, + tooltip="Duration in seconds", + ), + ], + outputs=[IO.Audio.Output()], + ) - FUNCTION = "trim" - RETURN_TYPES = ("AUDIO",) - CATEGORY = "audio" - DESCRIPTION = "Trim audio tensor into chosen time range." - - def trim(self, audio, start_index, duration): + @classmethod + def execute(cls, audio, start_index, duration) -> IO.NodeOutput: waveform = audio["waveform"] sample_rate = audio["sample_rate"] audio_length = waveform.shape[-1] @@ -399,23 +362,30 @@ class TrimAudioDuration: if start_frame >= end_frame: raise ValueError("AudioTrim: Start time must be less than end time and be within the audio length.") - return ({"waveform": waveform[..., start_frame:end_frame], "sample_rate": sample_rate},) + return IO.NodeOutput({"waveform": waveform[..., start_frame:end_frame], "sample_rate": sample_rate}) + + trim = execute # TODO: remove -class SplitAudioChannels: +class SplitAudioChannels(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": { - "audio": ("AUDIO",), - }} + def define_schema(cls): + return IO.Schema( + node_id="SplitAudioChannels", + display_name="Split Audio Channels", + description="Separates the audio into left and right channels.", + category="audio", + inputs=[ + IO.Audio.Input("audio"), + ], + outputs=[ + IO.Audio.Output(display_name="left"), + IO.Audio.Output(display_name="right"), + ], + ) - RETURN_TYPES = ("AUDIO", "AUDIO") - RETURN_NAMES = ("left", "right") - FUNCTION = "separate" - CATEGORY = "audio" - DESCRIPTION = "Separates the audio into left and right channels." - - def separate(self, audio): + @classmethod + def execute(cls, audio) -> IO.NodeOutput: waveform = audio["waveform"] sample_rate = audio["sample_rate"] @@ -425,7 +395,9 @@ class SplitAudioChannels: left_channel = waveform[..., 0:1, :] right_channel = waveform[..., 1:2, :] - return ({"waveform": left_channel, "sample_rate": sample_rate}, {"waveform": right_channel, "sample_rate": sample_rate}) + return IO.NodeOutput({"waveform": left_channel, "sample_rate": sample_rate}, {"waveform": right_channel, "sample_rate": sample_rate}) + + separate = execute # TODO: remove def match_audio_sample_rates(waveform_1, sample_rate_1, waveform_2, sample_rate_2): @@ -443,21 +415,29 @@ def match_audio_sample_rates(waveform_1, sample_rate_1, waveform_2, sample_rate_ return waveform_1, waveform_2, output_sample_rate -class AudioConcat: +class AudioConcat(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": { - "audio1": ("AUDIO",), - "audio2": ("AUDIO",), - "direction": (['after', 'before'], {"default": 'after', "tooltip": "Whether to append audio2 after or before audio1."}), - }} + def define_schema(cls): + return IO.Schema( + node_id="AudioConcat", + display_name="Audio Concat", + description="Concatenates the audio1 to audio2 in the specified direction.", + category="audio", + inputs=[ + IO.Audio.Input("audio1"), + IO.Audio.Input("audio2"), + IO.Combo.Input( + "direction", + options=['after', 'before'], + default="after", + tooltip="Whether to append audio2 after or before audio1.", + ) + ], + outputs=[IO.Audio.Output()], + ) - RETURN_TYPES = ("AUDIO",) - FUNCTION = "concat" - CATEGORY = "audio" - DESCRIPTION = "Concatenates the audio1 to audio2 in the specified direction." - - def concat(self, audio1, audio2, direction): + @classmethod + def execute(cls, audio1, audio2, direction) -> IO.NodeOutput: waveform_1 = audio1["waveform"] waveform_2 = audio2["waveform"] sample_rate_1 = audio1["sample_rate"] @@ -477,26 +457,33 @@ class AudioConcat: elif direction == 'before': concatenated_audio = torch.cat((waveform_2, waveform_1), dim=2) - return ({"waveform": concatenated_audio, "sample_rate": output_sample_rate},) + return IO.NodeOutput({"waveform": concatenated_audio, "sample_rate": output_sample_rate}) + + concat = execute # TODO: remove -class AudioMerge: +class AudioMerge(IO.ComfyNode): @classmethod - def INPUT_TYPES(cls): - return { - "required": { - "audio1": ("AUDIO",), - "audio2": ("AUDIO",), - "merge_method": (["add", "mean", "subtract", "multiply"], {"tooltip": "The method used to combine the audio waveforms."}), - }, - } + def define_schema(cls): + return IO.Schema( + node_id="AudioMerge", + display_name="Audio Merge", + description="Combine two audio tracks by overlaying their waveforms.", + category="audio", + inputs=[ + IO.Audio.Input("audio1"), + IO.Audio.Input("audio2"), + IO.Combo.Input( + "merge_method", + options=["add", "mean", "subtract", "multiply"], + tooltip="The method used to combine the audio waveforms.", + ) + ], + outputs=[IO.Audio.Output()], + ) - FUNCTION = "merge" - RETURN_TYPES = ("AUDIO",) - CATEGORY = "audio" - DESCRIPTION = "Combine two audio tracks by overlaying their waveforms." - - def merge(self, audio1, audio2, merge_method): + @classmethod + def execute(cls, audio1, audio2, merge_method) -> IO.NodeOutput: waveform_1 = audio1["waveform"] waveform_2 = audio2["waveform"] sample_rate_1 = audio1["sample_rate"] @@ -530,85 +517,108 @@ class AudioMerge: if max_val > 1.0: waveform = waveform / max_val - return ({"waveform": waveform, "sample_rate": output_sample_rate},) + return IO.NodeOutput({"waveform": waveform, "sample_rate": output_sample_rate}) + + merge = execute # TODO: remove -class AudioAdjustVolume: +class AudioAdjustVolume(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": { - "audio": ("AUDIO",), - "volume": ("INT", {"default": 1.0, "min": -100, "max": 100, "tooltip": "Volume adjustment in decibels (dB). 0 = no change, +6 = double, -6 = half, etc"}), - }} + def define_schema(cls): + return IO.Schema( + node_id="AudioAdjustVolume", + display_name="Audio Adjust Volume", + category="audio", + inputs=[ + IO.Audio.Input("audio"), + IO.Int.Input( + "volume", + default=1, + min=-100, + max=100, + tooltip="Volume adjustment in decibels (dB). 0 = no change, +6 = double, -6 = half, etc", + ) + ], + outputs=[IO.Audio.Output()], + ) - RETURN_TYPES = ("AUDIO",) - FUNCTION = "adjust_volume" - CATEGORY = "audio" - - def adjust_volume(self, audio, volume): + @classmethod + def execute(cls, audio, volume) -> IO.NodeOutput: if volume == 0: - return (audio,) + return IO.NodeOutput(audio) waveform = audio["waveform"] sample_rate = audio["sample_rate"] gain = 10 ** (volume / 20) waveform = waveform * gain - return ({"waveform": waveform, "sample_rate": sample_rate},) + return IO.NodeOutput({"waveform": waveform, "sample_rate": sample_rate}) + + adjust_volume = execute # TODO: remove -class EmptyAudio: +class EmptyAudio(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": { - "duration": ("FLOAT", {"default": 60.0, "min": 0.0, "max": 0xffffffffffffffff, "step": 0.01, "tooltip": "Duration of the empty audio clip in seconds"}), - "sample_rate": ("INT", {"default": 44100, "tooltip": "Sample rate of the empty audio clip."}), - "channels": ("INT", {"default": 2, "min": 1, "max": 2, "tooltip": "Number of audio channels (1 for mono, 2 for stereo)."}), - }} + def define_schema(cls): + return IO.Schema( + node_id="EmptyAudio", + display_name="Empty Audio", + category="audio", + inputs=[ + IO.Float.Input( + "duration", + default=60.0, + min=0.0, + max=0xffffffffffffffff, + step=0.01, + tooltip="Duration of the empty audio clip in seconds", + ), + IO.Float.Input( + "sample_rate", + default=44100, + tooltip="Sample rate of the empty audio clip.", + ), + IO.Float.Input( + "channels", + default=2, + min=1, + max=2, + tooltip="Number of audio channels (1 for mono, 2 for stereo).", + ), + ], + outputs=[IO.Audio.Output()], + ) - RETURN_TYPES = ("AUDIO",) - FUNCTION = "create_empty_audio" - CATEGORY = "audio" - - def create_empty_audio(self, duration, sample_rate, channels): + @classmethod + def execute(cls, duration, sample_rate, channels) -> IO.NodeOutput: num_samples = int(round(duration * sample_rate)) waveform = torch.zeros((1, channels, num_samples), dtype=torch.float32) - return ({"waveform": waveform, "sample_rate": sample_rate},) + return IO.NodeOutput({"waveform": waveform, "sample_rate": sample_rate}) + + create_empty_audio = execute # TODO: remove -NODE_CLASS_MAPPINGS = { - "EmptyLatentAudio": EmptyLatentAudio, - "VAEEncodeAudio": VAEEncodeAudio, - "VAEDecodeAudio": VAEDecodeAudio, - "SaveAudio": SaveAudio, - "SaveAudioMP3": SaveAudioMP3, - "SaveAudioOpus": SaveAudioOpus, - "LoadAudio": LoadAudio, - "PreviewAudio": PreviewAudio, - "ConditioningStableAudio": ConditioningStableAudio, - "RecordAudio": RecordAudio, - "TrimAudioDuration": TrimAudioDuration, - "SplitAudioChannels": SplitAudioChannels, - "AudioConcat": AudioConcat, - "AudioMerge": AudioMerge, - "AudioAdjustVolume": AudioAdjustVolume, - "EmptyAudio": EmptyAudio, -} +class AudioExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[IO.ComfyNode]]: + return [ + EmptyLatentAudio, + VAEEncodeAudio, + VAEDecodeAudio, + SaveAudio, + SaveAudioMP3, + SaveAudioOpus, + LoadAudio, + PreviewAudio, + ConditioningStableAudio, + RecordAudio, + TrimAudioDuration, + SplitAudioChannels, + AudioConcat, + AudioMerge, + AudioAdjustVolume, + EmptyAudio, + ] -NODE_DISPLAY_NAME_MAPPINGS = { - "EmptyLatentAudio": "Empty Latent Audio", - "VAEEncodeAudio": "VAE Encode Audio", - "VAEDecodeAudio": "VAE Decode Audio", - "PreviewAudio": "Preview Audio", - "LoadAudio": "Load Audio", - "SaveAudio": "Save Audio (FLAC)", - "SaveAudioMP3": "Save Audio (MP3)", - "SaveAudioOpus": "Save Audio (Opus)", - "RecordAudio": "Record Audio", - "TrimAudioDuration": "Trim Audio Duration", - "SplitAudioChannels": "Split Audio Channels", - "AudioConcat": "Audio Concat", - "AudioMerge": "Audio Merge", - "AudioAdjustVolume": "Audio Adjust Volume", - "EmptyAudio": "Empty Audio", -} +async def comfy_entrypoint() -> AudioExtension: + return AudioExtension() From ecdc8697d53919a9178bf53ef327a110582db8ea Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 3 Dec 2025 19:49:28 -0800 Subject: [PATCH 222/647] Qwen Image Lora training fix from #11090 (#11094) --- comfy_extras/nodes_train.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy_extras/nodes_train.py b/comfy_extras/nodes_train.py index cb24ab709..19b8baaf4 100644 --- a/comfy_extras/nodes_train.py +++ b/comfy_extras/nodes_train.py @@ -623,7 +623,7 @@ class TrainLoraNode(io.ComfyNode): noise = comfy_extras.nodes_custom_sampler.Noise_RandomNoise(seed) if multi_res: # use first latent as dummy latent if multi_res - latents = latents[0].repeat(num_images, 1, 1, 1) + latents = latents[0].repeat((num_images,) + ((1,) * (latents[0].ndim - 1))) guider.sample( noise.generate_noise({"samples": latents}), latents, From ea17add3c62197b10fd0b71d9169d339adc55c47 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 3 Dec 2025 20:15:15 -0800 Subject: [PATCH 223/647] Fix case where text encoders where running on the CPU instead of GPU. (#11095) --- comfy/sd.py | 2 ++ comfy/sd1_clip.py | 9 ++++++++- 2 files changed, 10 insertions(+), 1 deletion(-) diff --git a/comfy/sd.py b/comfy/sd.py index f9e5efab5..734bd2845 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -193,6 +193,7 @@ class CLIP: self.cond_stage_model.set_clip_options({"projected_pooled": False}) self.load_model() + self.cond_stage_model.set_clip_options({"execution_device": self.patcher.load_device}) all_hooks.reset() self.patcher.patch_hooks(None) if show_pbar: @@ -240,6 +241,7 @@ class CLIP: self.cond_stage_model.set_clip_options({"projected_pooled": False}) self.load_model() + self.cond_stage_model.set_clip_options({"execution_device": self.patcher.load_device}) o = self.cond_stage_model.encode_token_weights(tokens) cond, pooled = o[:2] if return_dict: diff --git a/comfy/sd1_clip.py b/comfy/sd1_clip.py index 0fc9ab3db..503a51843 100644 --- a/comfy/sd1_clip.py +++ b/comfy/sd1_clip.py @@ -147,6 +147,7 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder): self.layer_norm_hidden_state = layer_norm_hidden_state self.return_projected_pooled = return_projected_pooled self.return_attention_masks = return_attention_masks + self.execution_device = None if layer == "hidden": assert layer_idx is not None @@ -163,6 +164,7 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder): def set_clip_options(self, options): layer_idx = options.get("layer", self.layer_idx) self.return_projected_pooled = options.get("projected_pooled", self.return_projected_pooled) + self.execution_device = options.get("execution_device", self.execution_device) if isinstance(self.layer, list) or self.layer == "all": pass elif layer_idx is None or abs(layer_idx) > self.num_layers: @@ -175,6 +177,7 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder): self.layer = self.options_default[0] self.layer_idx = self.options_default[1] self.return_projected_pooled = self.options_default[2] + self.execution_device = None def process_tokens(self, tokens, device): end_token = self.special_tokens.get("end", None) @@ -258,7 +261,11 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder): return torch.cat(embeds_out), torch.tensor(attention_masks, device=device, dtype=torch.long), num_tokens, embeds_info def forward(self, tokens): - device = self.transformer.get_input_embeddings().weight.device + if self.execution_device is None: + device = self.transformer.get_input_embeddings().weight.device + else: + device = self.execution_device + embeds, attention_mask, num_tokens, embeds_info = self.process_tokens(tokens, device) attention_mask_model = None From 6be85c7920224b45bbc6417e00147815e78c12a9 Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Thu, 4 Dec 2025 14:28:44 +1000 Subject: [PATCH 224/647] mp: use look-ahead actuals for stream offload VRAM calculation (#11096) TIL that the WAN TE has a 2GB weight followed by 16MB as the next size down. This means that team 8GB VRAM would fully offload the TE in async offload mode as it just multiplied this giant size my the num streams. Do the more complex logic of summing up the upcoming to-load weight sizes to avoid triple counting this massive weight. partial unload does the converse of recording the NS most recent unloads as they go. --- comfy/model_patcher.py | 12 +++++++++--- 1 file changed, 9 insertions(+), 3 deletions(-) diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index df2d8e827..3dcac3eef 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -699,12 +699,12 @@ class ModelPatcher: offloaded = [] offload_buffer = 0 loading.sort(reverse=True) - for x in loading: + for i, x in enumerate(loading): module_offload_mem, module_mem, n, m, params = x lowvram_weight = False - potential_offload = max(offload_buffer, module_offload_mem + (comfy.model_management.NUM_STREAMS * module_mem)) + potential_offload = max(offload_buffer, module_offload_mem + sum([ x1[1] for x1 in loading[i+1:i+1+comfy.model_management.NUM_STREAMS]])) lowvram_fits = mem_counter + module_mem + potential_offload < lowvram_model_memory weight_key = "{}.weight".format(n) @@ -876,14 +876,18 @@ class ModelPatcher: patch_counter = 0 unload_list = self._load_list() unload_list.sort() + offload_buffer = self.model.model_offload_buffer_memory + if len(unload_list) > 0: + NS = comfy.model_management.NUM_STREAMS + offload_weight_factor = [ min(offload_buffer / (NS + 1), unload_list[0][1]) ] * NS for unload in unload_list: if memory_to_free + offload_buffer - self.model.model_offload_buffer_memory < memory_freed: break module_offload_mem, module_mem, n, m, params = unload - potential_offload = module_offload_mem + (comfy.model_management.NUM_STREAMS * module_mem) + potential_offload = module_offload_mem + sum(offload_weight_factor) lowvram_possible = hasattr(m, "comfy_cast_weights") if hasattr(m, "comfy_patched_weights") and m.comfy_patched_weights == True: @@ -935,6 +939,8 @@ class ModelPatcher: m.comfy_patched_weights = False memory_freed += module_mem offload_buffer = max(offload_buffer, potential_offload) + offload_weight_factor.append(module_mem) + offload_weight_factor.pop(0) logging.debug("freed {}".format(n)) for param in params: From f4bdf5f8302ef10db99644a8672e614ddb29c473 Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Fri, 5 Dec 2025 03:50:04 +1000 Subject: [PATCH 225/647] sd: revise hy VAE VRAM (#11105) This was recently collapsed down to rolling VAE through temporal. Clamp The time dimension. --- comfy/sd.py | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) diff --git a/comfy/sd.py b/comfy/sd.py index 734bd2845..fe4dd65f8 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -483,8 +483,10 @@ class VAE: self.latent_dim = 3 self.latent_channels = ddconfig['z_channels'] = sd["decoder.conv_in.conv.weight"].shape[1] self.first_stage_model = AutoencoderKL(ddconfig=ddconfig, embed_dim=sd['post_quant_conv.weight'].shape[1]) - self.memory_used_decode = lambda shape, dtype: (1500 * shape[2] * shape[3] * shape[4] * (4 * 8 * 8)) * model_management.dtype_size(dtype) - self.memory_used_encode = lambda shape, dtype: (900 * max(shape[2], 2) * shape[3] * shape[4]) * model_management.dtype_size(dtype) + #This is likely to significantly over-estimate with single image or low frame counts as the + #implementation is able to completely skip caching. Rework if used as an image only VAE + self.memory_used_decode = lambda shape, dtype: (2800 * min(8, ((shape[2] - 1) * 4) + 1) * shape[3] * shape[4] * (8 * 8)) * model_management.dtype_size(dtype) + self.memory_used_encode = lambda shape, dtype: (1400 * min(9, shape[2]) * shape[3] * shape[4]) * model_management.dtype_size(dtype) self.working_dtypes = [torch.bfloat16, torch.float16, torch.float32] elif "decoder.unpatcher3d.wavelets" in sd: self.upscale_ratio = (lambda a: max(0, a * 8 - 7), 8, 8) From 9bc893c5bbd2838bdd15ebd40e3a3e548ce3e4f0 Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Fri, 5 Dec 2025 03:50:36 +1000 Subject: [PATCH 226/647] sd: bump HY1.5 VAE estimate (#11107) Im able to push vram above estimate on partial unload. Bump the estimate. This is experimentally determined with a 720P and 480P datapoint calibrating for 24GB VRAM total. --- comfy/sd.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy/sd.py b/comfy/sd.py index fe4dd65f8..03bdb33d5 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -471,7 +471,7 @@ class VAE: decoder_config={'target': "comfy.ldm.hunyuan_video.vae_refiner.Decoder", 'params': ddconfig}) self.memory_used_encode = lambda shape, dtype: (1400 * 9 * shape[-2] * shape[-1]) * model_management.dtype_size(dtype) - self.memory_used_decode = lambda shape, dtype: (2800 * 4 * shape[-2] * shape[-1] * 16 * 16) * model_management.dtype_size(dtype) + self.memory_used_decode = lambda shape, dtype: (3600 * 4 * shape[-2] * shape[-1] * 16 * 16) * model_management.dtype_size(dtype) elif "decoder.conv_in.conv.weight" in sd: ddconfig = {'double_z': True, 'z_channels': 4, 'resolution': 256, 'in_channels': 3, 'out_ch': 3, 'ch': 128, 'ch_mult': [1, 2, 4, 4], 'num_res_blocks': 2, 'attn_resolutions': [], 'dropout': 0.0} ddconfig["conv3d"] = True From 3c8456223c5f6a41af7d99219b391c8c58acb552 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Fri, 5 Dec 2025 00:05:28 +0200 Subject: [PATCH 227/647] [API Nodes]: fixes and refactor (#11104) * chore(api-nodes): applied ruff's pyupgrade(python3.10) to api-nodes client's to folder * chore(api-nodes): add validate_video_frame_count function from LTX PR * chore(api-nodes): replace deprecated V1 imports * fix(api-nodes): the types returned by the "poll_op" function are now correct. --- comfy_api_nodes/util/__init__.py | 2 + comfy_api_nodes/util/_helpers.py | 14 +-- comfy_api_nodes/util/client.py | 145 ++++++++++++----------- comfy_api_nodes/util/conversions.py | 21 ++-- comfy_api_nodes/util/download_helpers.py | 20 ++-- comfy_api_nodes/util/request_logger.py | 2 - comfy_api_nodes/util/upload_helpers.py | 16 ++- comfy_api_nodes/util/validation_utils.py | 61 ++++++---- 8 files changed, 146 insertions(+), 135 deletions(-) diff --git a/comfy_api_nodes/util/__init__.py b/comfy_api_nodes/util/__init__.py index 80292fb3c..4cc22abfb 100644 --- a/comfy_api_nodes/util/__init__.py +++ b/comfy_api_nodes/util/__init__.py @@ -47,6 +47,7 @@ from .validation_utils import ( validate_string, validate_video_dimensions, validate_video_duration, + validate_video_frame_count, ) __all__ = [ @@ -94,6 +95,7 @@ __all__ = [ "validate_string", "validate_video_dimensions", "validate_video_duration", + "validate_video_frame_count", # Misc functions "get_fs_object_size", ] diff --git a/comfy_api_nodes/util/_helpers.py b/comfy_api_nodes/util/_helpers.py index 328fe5227..491e6b6a8 100644 --- a/comfy_api_nodes/util/_helpers.py +++ b/comfy_api_nodes/util/_helpers.py @@ -2,8 +2,8 @@ import asyncio import contextlib import os import time +from collections.abc import Callable from io import BytesIO -from typing import Callable, Optional, Union from comfy.cli_args import args from comfy.model_management import processing_interrupted @@ -35,12 +35,12 @@ def default_base_url() -> str: async def sleep_with_interrupt( seconds: float, - node_cls: Optional[type[IO.ComfyNode]], - label: Optional[str] = None, - start_ts: Optional[float] = None, - estimated_total: Optional[int] = None, + node_cls: type[IO.ComfyNode] | None, + label: str | None = None, + start_ts: float | None = None, + estimated_total: int | None = None, *, - display_callback: Optional[Callable[[type[IO.ComfyNode], str, int, Optional[int]], None]] = None, + display_callback: Callable[[type[IO.ComfyNode], str, int, int | None], None] | None = None, ): """ Sleep in 1s slices while: @@ -65,7 +65,7 @@ def mimetype_to_extension(mime_type: str) -> str: return mime_type.split("/")[-1].lower() -def get_fs_object_size(path_or_object: Union[str, BytesIO]) -> int: +def get_fs_object_size(path_or_object: str | BytesIO) -> int: if isinstance(path_or_object, str): return os.path.getsize(path_or_object) return len(path_or_object.getvalue()) diff --git a/comfy_api_nodes/util/client.py b/comfy_api_nodes/util/client.py index bf01d7d36..bf37cba5f 100644 --- a/comfy_api_nodes/util/client.py +++ b/comfy_api_nodes/util/client.py @@ -4,10 +4,11 @@ import json import logging import time import uuid +from collections.abc import Callable, Iterable from dataclasses import dataclass from enum import Enum from io import BytesIO -from typing import Any, Callable, Iterable, Literal, Optional, Type, TypeVar, Union +from typing import Any, Literal, TypeVar from urllib.parse import urljoin, urlparse import aiohttp @@ -37,8 +38,8 @@ class ApiEndpoint: path: str, method: Literal["GET", "POST", "PUT", "DELETE", "PATCH"] = "GET", *, - query_params: Optional[dict[str, Any]] = None, - headers: Optional[dict[str, str]] = None, + query_params: dict[str, Any] | None = None, + headers: dict[str, str] | None = None, ): self.path = path self.method = method @@ -52,18 +53,18 @@ class _RequestConfig: endpoint: ApiEndpoint timeout: float content_type: str - data: Optional[dict[str, Any]] - files: Optional[Union[dict[str, Any], list[tuple[str, Any]]]] - multipart_parser: Optional[Callable] + data: dict[str, Any] | None + files: dict[str, Any] | list[tuple[str, Any]] | None + multipart_parser: Callable | None max_retries: int retry_delay: float retry_backoff: float wait_label: str = "Waiting" monitor_progress: bool = True - estimated_total: Optional[int] = None - final_label_on_success: Optional[str] = "Completed" - progress_origin_ts: Optional[float] = None - price_extractor: Optional[Callable[[dict[str, Any]], Optional[float]]] = None + estimated_total: int | None = None + final_label_on_success: str | None = "Completed" + progress_origin_ts: float | None = None + price_extractor: Callable[[dict[str, Any]], float | None] | None = None @dataclass @@ -71,10 +72,10 @@ class _PollUIState: started: float status_label: str = "Queued" is_queued: bool = True - price: Optional[float] = None - estimated_duration: Optional[int] = None + price: float | None = None + estimated_duration: int | None = None base_processing_elapsed: float = 0.0 # sum of completed active intervals - active_since: Optional[float] = None # start time of current active interval (None if queued) + active_since: float | None = None # start time of current active interval (None if queued) _RETRY_STATUS = {408, 429, 500, 502, 503, 504} @@ -87,20 +88,20 @@ async def sync_op( cls: type[IO.ComfyNode], endpoint: ApiEndpoint, *, - response_model: Type[M], - price_extractor: Optional[Callable[[M], Optional[float]]] = None, - data: Optional[BaseModel] = None, - files: Optional[Union[dict[str, Any], list[tuple[str, Any]]]] = None, + response_model: type[M], + price_extractor: Callable[[M | Any], float | None] | None = None, + data: BaseModel | None = None, + files: dict[str, Any] | list[tuple[str, Any]] | None = None, content_type: str = "application/json", timeout: float = 3600.0, - multipart_parser: Optional[Callable] = None, + multipart_parser: Callable | None = None, max_retries: int = 3, retry_delay: float = 1.0, retry_backoff: float = 2.0, wait_label: str = "Waiting for server", - estimated_duration: Optional[int] = None, - final_label_on_success: Optional[str] = "Completed", - progress_origin_ts: Optional[float] = None, + estimated_duration: int | None = None, + final_label_on_success: str | None = "Completed", + progress_origin_ts: float | None = None, monitor_progress: bool = True, ) -> M: raw = await sync_op_raw( @@ -131,22 +132,22 @@ async def poll_op( cls: type[IO.ComfyNode], poll_endpoint: ApiEndpoint, *, - response_model: Type[M], - status_extractor: Callable[[M], Optional[Union[str, int]]], - progress_extractor: Optional[Callable[[M], Optional[int]]] = None, - price_extractor: Optional[Callable[[M], Optional[float]]] = None, - completed_statuses: Optional[list[Union[str, int]]] = None, - failed_statuses: Optional[list[Union[str, int]]] = None, - queued_statuses: Optional[list[Union[str, int]]] = None, - data: Optional[BaseModel] = None, + response_model: type[M], + status_extractor: Callable[[M | Any], str | int | None], + progress_extractor: Callable[[M | Any], int | None] | None = None, + price_extractor: Callable[[M | Any], float | None] | None = None, + completed_statuses: list[str | int] | None = None, + failed_statuses: list[str | int] | None = None, + queued_statuses: list[str | int] | None = None, + data: BaseModel | None = None, poll_interval: float = 5.0, max_poll_attempts: int = 120, timeout_per_poll: float = 120.0, max_retries_per_poll: int = 3, retry_delay_per_poll: float = 1.0, retry_backoff_per_poll: float = 2.0, - estimated_duration: Optional[int] = None, - cancel_endpoint: Optional[ApiEndpoint] = None, + estimated_duration: int | None = None, + cancel_endpoint: ApiEndpoint | None = None, cancel_timeout: float = 10.0, ) -> M: raw = await poll_op_raw( @@ -178,22 +179,22 @@ async def sync_op_raw( cls: type[IO.ComfyNode], endpoint: ApiEndpoint, *, - price_extractor: Optional[Callable[[dict[str, Any]], Optional[float]]] = None, - data: Optional[Union[dict[str, Any], BaseModel]] = None, - files: Optional[Union[dict[str, Any], list[tuple[str, Any]]]] = None, + price_extractor: Callable[[dict[str, Any]], float | None] | None = None, + data: dict[str, Any] | BaseModel | None = None, + files: dict[str, Any] | list[tuple[str, Any]] | None = None, content_type: str = "application/json", timeout: float = 3600.0, - multipart_parser: Optional[Callable] = None, + multipart_parser: Callable | None = None, max_retries: int = 3, retry_delay: float = 1.0, retry_backoff: float = 2.0, wait_label: str = "Waiting for server", - estimated_duration: Optional[int] = None, + estimated_duration: int | None = None, as_binary: bool = False, - final_label_on_success: Optional[str] = "Completed", - progress_origin_ts: Optional[float] = None, + final_label_on_success: str | None = "Completed", + progress_origin_ts: float | None = None, monitor_progress: bool = True, -) -> Union[dict[str, Any], bytes]: +) -> dict[str, Any] | bytes: """ Make a single network request. - If as_binary=False (default): returns JSON dict (or {'_raw': ''} if non-JSON). @@ -229,21 +230,21 @@ async def poll_op_raw( cls: type[IO.ComfyNode], poll_endpoint: ApiEndpoint, *, - status_extractor: Callable[[dict[str, Any]], Optional[Union[str, int]]], - progress_extractor: Optional[Callable[[dict[str, Any]], Optional[int]]] = None, - price_extractor: Optional[Callable[[dict[str, Any]], Optional[float]]] = None, - completed_statuses: Optional[list[Union[str, int]]] = None, - failed_statuses: Optional[list[Union[str, int]]] = None, - queued_statuses: Optional[list[Union[str, int]]] = None, - data: Optional[Union[dict[str, Any], BaseModel]] = None, + status_extractor: Callable[[dict[str, Any]], str | int | None], + progress_extractor: Callable[[dict[str, Any]], int | None] | None = None, + price_extractor: Callable[[dict[str, Any]], float | None] | None = None, + completed_statuses: list[str | int] | None = None, + failed_statuses: list[str | int] | None = None, + queued_statuses: list[str | int] | None = None, + data: dict[str, Any] | BaseModel | None = None, poll_interval: float = 5.0, max_poll_attempts: int = 120, timeout_per_poll: float = 120.0, max_retries_per_poll: int = 3, retry_delay_per_poll: float = 1.0, retry_backoff_per_poll: float = 2.0, - estimated_duration: Optional[int] = None, - cancel_endpoint: Optional[ApiEndpoint] = None, + estimated_duration: int | None = None, + cancel_endpoint: ApiEndpoint | None = None, cancel_timeout: float = 10.0, ) -> dict[str, Any]: """ @@ -261,7 +262,7 @@ async def poll_op_raw( consumed_attempts = 0 # counts only non-queued polls progress_bar = utils.ProgressBar(100) if progress_extractor else None - last_progress: Optional[int] = None + last_progress: int | None = None state = _PollUIState(started=started, estimated_duration=estimated_duration) stop_ticker = asyncio.Event() @@ -420,10 +421,10 @@ async def poll_op_raw( def _display_text( node_cls: type[IO.ComfyNode], - text: Optional[str], + text: str | None, *, - status: Optional[Union[str, int]] = None, - price: Optional[float] = None, + status: str | int | None = None, + price: float | None = None, ) -> None: display_lines: list[str] = [] if status: @@ -440,13 +441,13 @@ def _display_text( def _display_time_progress( node_cls: type[IO.ComfyNode], - status: Optional[Union[str, int]], + status: str | int | None, elapsed_seconds: int, - estimated_total: Optional[int] = None, + estimated_total: int | None = None, *, - price: Optional[float] = None, - is_queued: Optional[bool] = None, - processing_elapsed_seconds: Optional[int] = None, + price: float | None = None, + is_queued: bool | None = None, + processing_elapsed_seconds: int | None = None, ) -> None: if estimated_total is not None and estimated_total > 0 and is_queued is False: pe = processing_elapsed_seconds if processing_elapsed_seconds is not None else elapsed_seconds @@ -488,7 +489,7 @@ def _unpack_tuple(t: tuple) -> tuple[str, Any, str]: raise ValueError("files tuple must be (filename, file[, content_type])") -def _merge_params(endpoint_params: dict[str, Any], method: str, data: Optional[dict[str, Any]]) -> dict[str, Any]: +def _merge_params(endpoint_params: dict[str, Any], method: str, data: dict[str, Any] | None) -> dict[str, Any]: params = dict(endpoint_params or {}) if method.upper() == "GET" and data: for k, v in data.items(): @@ -534,9 +535,9 @@ def _generate_operation_id(method: str, path: str, attempt: int) -> str: def _snapshot_request_body_for_logging( content_type: str, method: str, - data: Optional[dict[str, Any]], - files: Optional[Union[dict[str, Any], list[tuple[str, Any]]]], -) -> Optional[Union[dict[str, Any], str]]: + data: dict[str, Any] | None, + files: dict[str, Any] | list[tuple[str, Any]] | None, +) -> dict[str, Any] | str | None: if method.upper() == "GET": return None if content_type == "multipart/form-data": @@ -586,13 +587,13 @@ async def _request_base(cfg: _RequestConfig, expect_binary: bool): attempt = 0 delay = cfg.retry_delay operation_succeeded: bool = False - final_elapsed_seconds: Optional[int] = None - extracted_price: Optional[float] = None + final_elapsed_seconds: int | None = None + extracted_price: float | None = None while True: attempt += 1 stop_event = asyncio.Event() - monitor_task: Optional[asyncio.Task] = None - sess: Optional[aiohttp.ClientSession] = None + monitor_task: asyncio.Task | None = None + sess: aiohttp.ClientSession | None = None operation_id = _generate_operation_id(method, cfg.endpoint.path, attempt) logging.debug("[DEBUG] HTTP %s %s (attempt %d)", method, url, attempt) @@ -887,7 +888,7 @@ async def _request_base(cfg: _RequestConfig, expect_binary: bool): ) -def _validate_or_raise(response_model: Type[M], payload: Any) -> M: +def _validate_or_raise(response_model: type[M], payload: Any) -> M: try: return response_model.model_validate(payload) except Exception as e: @@ -902,9 +903,9 @@ def _validate_or_raise(response_model: Type[M], payload: Any) -> M: def _wrap_model_extractor( - response_model: Type[M], - extractor: Optional[Callable[[M], Any]], -) -> Optional[Callable[[dict[str, Any]], Any]]: + response_model: type[M], + extractor: Callable[[M], Any] | None, +) -> Callable[[dict[str, Any]], Any] | None: """Wrap a typed extractor so it can be used by the dict-based poller. Validates the dict into `response_model` before invoking `extractor`. Uses a small per-wrapper cache keyed by `id(dict)` to avoid re-validating @@ -929,10 +930,10 @@ def _wrap_model_extractor( return _wrapped -def _normalize_statuses(values: Optional[Iterable[Union[str, int]]]) -> set[Union[str, int]]: +def _normalize_statuses(values: Iterable[str | int] | None) -> set[str | int]: if not values: return set() - out: set[Union[str, int]] = set() + out: set[str | int] = set() for v in values: nv = _normalize_status_value(v) if nv is not None: @@ -940,7 +941,7 @@ def _normalize_statuses(values: Optional[Iterable[Union[str, int]]]) -> set[Unio return out -def _normalize_status_value(val: Union[str, int, None]) -> Union[str, int, None]: +def _normalize_status_value(val: str | int | None) -> str | int | None: if isinstance(val, str): return val.strip().lower() return val diff --git a/comfy_api_nodes/util/conversions.py b/comfy_api_nodes/util/conversions.py index 971dc57de..c57457580 100644 --- a/comfy_api_nodes/util/conversions.py +++ b/comfy_api_nodes/util/conversions.py @@ -4,7 +4,6 @@ import math import mimetypes import uuid from io import BytesIO -from typing import Optional import av import numpy as np @@ -12,8 +11,7 @@ import torch from PIL import Image from comfy.utils import common_upscale -from comfy_api.latest import Input, InputImpl -from comfy_api.util import VideoCodec, VideoContainer +from comfy_api.latest import Input, InputImpl, Types from ._helpers import mimetype_to_extension @@ -57,7 +55,7 @@ def image_tensor_pair_to_batch(image1: torch.Tensor, image2: torch.Tensor) -> to def tensor_to_bytesio( image: torch.Tensor, - name: Optional[str] = None, + name: str | None = None, total_pixels: int = 2048 * 2048, mime_type: str = "image/png", ) -> BytesIO: @@ -177,8 +175,8 @@ def audio_to_base64_string(audio: Input.Audio, container_format: str = "mp4", co def video_to_base64_string( video: Input.Video, - container_format: VideoContainer = None, - codec: VideoCodec = None + container_format: Types.VideoContainer | None = None, + codec: Types.VideoCodec | None = None, ) -> str: """ Converts a video input to a base64 string. @@ -189,12 +187,11 @@ def video_to_base64_string( codec: Optional codec to use (defaults to video.codec if available) """ video_bytes_io = BytesIO() - - # Use provided format/codec if specified, otherwise use video's own if available - format_to_use = container_format if container_format is not None else getattr(video, 'container', VideoContainer.MP4) - codec_to_use = codec if codec is not None else getattr(video, 'codec', VideoCodec.H264) - - video.save_to(video_bytes_io, format=format_to_use, codec=codec_to_use) + video.save_to( + video_bytes_io, + format=container_format or getattr(video, "container", Types.VideoContainer.MP4), + codec=codec or getattr(video, "codec", Types.VideoCodec.H264), + ) video_bytes_io.seek(0) return base64.b64encode(video_bytes_io.getvalue()).decode("utf-8") diff --git a/comfy_api_nodes/util/download_helpers.py b/comfy_api_nodes/util/download_helpers.py index 14207dc68..3e0d0352d 100644 --- a/comfy_api_nodes/util/download_helpers.py +++ b/comfy_api_nodes/util/download_helpers.py @@ -3,15 +3,15 @@ import contextlib import uuid from io import BytesIO from pathlib import Path -from typing import IO, Optional, Union +from typing import IO from urllib.parse import urljoin, urlparse import aiohttp import torch from aiohttp.client_exceptions import ClientError, ContentTypeError -from comfy_api.input_impl import VideoFromFile from comfy_api.latest import IO as COMFY_IO +from comfy_api.latest import InputImpl from . import request_logger from ._helpers import ( @@ -29,9 +29,9 @@ _RETRY_STATUS = {408, 429, 500, 502, 503, 504} async def download_url_to_bytesio( url: str, - dest: Optional[Union[BytesIO, IO[bytes], str, Path]], + dest: BytesIO | IO[bytes] | str | Path | None, *, - timeout: Optional[float] = None, + timeout: float | None = None, max_retries: int = 5, retry_delay: float = 1.0, retry_backoff: float = 2.0, @@ -71,10 +71,10 @@ async def download_url_to_bytesio( is_path_sink = isinstance(dest, (str, Path)) fhandle = None - session: Optional[aiohttp.ClientSession] = None - stop_evt: Optional[asyncio.Event] = None - monitor_task: Optional[asyncio.Task] = None - req_task: Optional[asyncio.Task] = None + session: aiohttp.ClientSession | None = None + stop_evt: asyncio.Event | None = None + monitor_task: asyncio.Task | None = None + req_task: asyncio.Task | None = None try: with contextlib.suppress(Exception): @@ -234,11 +234,11 @@ async def download_url_to_video_output( timeout: float = None, max_retries: int = 5, cls: type[COMFY_IO.ComfyNode] = None, -) -> VideoFromFile: +) -> InputImpl.VideoFromFile: """Downloads a video from a URL and returns a `VIDEO` output.""" result = BytesIO() await download_url_to_bytesio(video_url, result, timeout=timeout, max_retries=max_retries, cls=cls) - return VideoFromFile(result) + return InputImpl.VideoFromFile(result) async def download_url_as_bytesio( diff --git a/comfy_api_nodes/util/request_logger.py b/comfy_api_nodes/util/request_logger.py index ac52e2eab..e0cb4428d 100644 --- a/comfy_api_nodes/util/request_logger.py +++ b/comfy_api_nodes/util/request_logger.py @@ -1,5 +1,3 @@ -from __future__ import annotations - import datetime import hashlib import json diff --git a/comfy_api_nodes/util/upload_helpers.py b/comfy_api_nodes/util/upload_helpers.py index 0532bea9a..b8d33f4d1 100644 --- a/comfy_api_nodes/util/upload_helpers.py +++ b/comfy_api_nodes/util/upload_helpers.py @@ -4,15 +4,13 @@ import logging import time import uuid from io import BytesIO -from typing import Optional from urllib.parse import urlparse import aiohttp import torch from pydantic import BaseModel, Field -from comfy_api.latest import IO, Input -from comfy_api.util import VideoCodec, VideoContainer +from comfy_api.latest import IO, Input, Types from . import request_logger from ._helpers import is_processing_interrupted, sleep_with_interrupt @@ -32,7 +30,7 @@ from .conversions import ( class UploadRequest(BaseModel): file_name: str = Field(..., description="Filename to upload") - content_type: Optional[str] = Field( + content_type: str | None = Field( None, description="Mime type of the file. For example: image/png, image/jpeg, video/mp4, etc.", ) @@ -56,7 +54,7 @@ async def upload_images_to_comfyapi( Uploads images to ComfyUI API and returns download URLs. To upload multiple images, stack them in the batch dimension first. """ - # if batch, try to upload each file if max_images is greater than 0 + # if batched, try to upload each file if max_images is greater than 0 download_urls: list[str] = [] is_batch = len(image.shape) > 3 batch_len = image.shape[0] if is_batch else 1 @@ -100,9 +98,9 @@ async def upload_video_to_comfyapi( cls: type[IO.ComfyNode], video: Input.Video, *, - container: VideoContainer = VideoContainer.MP4, - codec: VideoCodec = VideoCodec.H264, - max_duration: Optional[int] = None, + container: Types.VideoContainer = Types.VideoContainer.MP4, + codec: Types.VideoCodec = Types.VideoCodec.H264, + max_duration: int | None = None, wait_label: str | None = "Uploading", ) -> str: """ @@ -220,7 +218,7 @@ async def upload_file( return monitor_task = asyncio.create_task(_monitor()) - sess: Optional[aiohttp.ClientSession] = None + sess: aiohttp.ClientSession | None = None try: try: request_logger.log_request_response( diff --git a/comfy_api_nodes/util/validation_utils.py b/comfy_api_nodes/util/validation_utils.py index ec7006aed..f01edea96 100644 --- a/comfy_api_nodes/util/validation_utils.py +++ b/comfy_api_nodes/util/validation_utils.py @@ -1,9 +1,7 @@ import logging -from typing import Optional import torch -from comfy_api.input.video_types import VideoInput from comfy_api.latest import Input @@ -18,10 +16,10 @@ def get_image_dimensions(image: torch.Tensor) -> tuple[int, int]: def validate_image_dimensions( image: torch.Tensor, - min_width: Optional[int] = None, - max_width: Optional[int] = None, - min_height: Optional[int] = None, - max_height: Optional[int] = None, + min_width: int | None = None, + max_width: int | None = None, + min_height: int | None = None, + max_height: int | None = None, ): height, width = get_image_dimensions(image) @@ -37,8 +35,8 @@ def validate_image_dimensions( def validate_image_aspect_ratio( image: torch.Tensor, - min_ratio: Optional[tuple[float, float]] = None, # e.g. (1, 4) - max_ratio: Optional[tuple[float, float]] = None, # e.g. (4, 1) + min_ratio: tuple[float, float] | None = None, # e.g. (1, 4) + max_ratio: tuple[float, float] | None = None, # e.g. (4, 1) *, strict: bool = True, # True -> (min, max); False -> [min, max] ) -> float: @@ -54,8 +52,8 @@ def validate_image_aspect_ratio( def validate_images_aspect_ratio_closeness( first_image: torch.Tensor, second_image: torch.Tensor, - min_rel: float, # e.g. 0.8 - max_rel: float, # e.g. 1.25 + min_rel: float, # e.g. 0.8 + max_rel: float, # e.g. 1.25 *, strict: bool = False, # True -> (min, max); False -> [min, max] ) -> float: @@ -84,8 +82,8 @@ def validate_images_aspect_ratio_closeness( def validate_aspect_ratio_string( aspect_ratio: str, - min_ratio: Optional[tuple[float, float]] = None, # e.g. (1, 4) - max_ratio: Optional[tuple[float, float]] = None, # e.g. (4, 1) + min_ratio: tuple[float, float] | None = None, # e.g. (1, 4) + max_ratio: tuple[float, float] | None = None, # e.g. (4, 1) *, strict: bool = False, # True -> (min, max); False -> [min, max] ) -> float: @@ -97,10 +95,10 @@ def validate_aspect_ratio_string( def validate_video_dimensions( video: Input.Video, - min_width: Optional[int] = None, - max_width: Optional[int] = None, - min_height: Optional[int] = None, - max_height: Optional[int] = None, + min_width: int | None = None, + max_width: int | None = None, + min_height: int | None = None, + max_height: int | None = None, ): try: width, height = video.get_dimensions() @@ -120,8 +118,8 @@ def validate_video_dimensions( def validate_video_duration( video: Input.Video, - min_duration: Optional[float] = None, - max_duration: Optional[float] = None, + min_duration: float | None = None, + max_duration: float | None = None, ): try: duration = video.get_duration() @@ -136,6 +134,23 @@ def validate_video_duration( raise ValueError(f"Video duration must be at most {max_duration}s, got {duration}s") +def validate_video_frame_count( + video: Input.Video, + min_frame_count: int | None = None, + max_frame_count: int | None = None, +): + try: + frame_count = video.get_frame_count() + except Exception as e: + logging.error("Error getting frame count of video: %s", e) + return + + if min_frame_count is not None and min_frame_count > frame_count: + raise ValueError(f"Video frame count must be at least {min_frame_count}, got {frame_count}") + if max_frame_count is not None and frame_count > max_frame_count: + raise ValueError(f"Video frame count must be at most {max_frame_count}, got {frame_count}") + + def get_number_of_images(images): if isinstance(images, torch.Tensor): return images.shape[0] if images.ndim >= 4 else 1 @@ -144,8 +159,8 @@ def get_number_of_images(images): def validate_audio_duration( audio: Input.Audio, - min_duration: Optional[float] = None, - max_duration: Optional[float] = None, + min_duration: float | None = None, + max_duration: float | None = None, ) -> None: sr = int(audio["sample_rate"]) dur = int(audio["waveform"].shape[-1]) / sr @@ -177,7 +192,7 @@ def validate_string( ) -def validate_container_format_is_mp4(video: VideoInput) -> None: +def validate_container_format_is_mp4(video: Input.Video) -> None: """Validates video container format is MP4.""" container_format = video.get_container_format() if container_format not in ["mp4", "mov,mp4,m4a,3gp,3g2,mj2"]: @@ -194,8 +209,8 @@ def _ratio_from_tuple(r: tuple[float, float]) -> float: def _assert_ratio_bounds( ar: float, *, - min_ratio: Optional[tuple[float, float]] = None, - max_ratio: Optional[tuple[float, float]] = None, + min_ratio: tuple[float, float] | None = None, + max_ratio: tuple[float, float] | None = None, strict: bool = True, ) -> None: """Validate a numeric aspect ratio against optional min/max ratio bounds.""" From 35fa091340c60612dfb71cb6822dc23b99a5dac2 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Thu, 4 Dec 2025 19:52:09 -0800 Subject: [PATCH 228/647] Forgot to put this in README. (#11112) --- README.md | 1 + 1 file changed, 1 insertion(+) diff --git a/README.md b/README.md index 91fb510e1..ed857df9f 100644 --- a/README.md +++ b/README.md @@ -81,6 +81,7 @@ See what ComfyUI can do with the [example workflows](https://comfyanonymous.gith - [Hunyuan Video](https://comfyanonymous.github.io/ComfyUI_examples/hunyuan_video/) - [Wan 2.1](https://comfyanonymous.github.io/ComfyUI_examples/wan/) - [Wan 2.2](https://comfyanonymous.github.io/ComfyUI_examples/wan22/) + - [Hunyuan Video 1.5](https://docs.comfy.org/tutorials/video/hunyuan/hunyuan-video-1-5) - Audio Models - [Stable Audio](https://comfyanonymous.github.io/ComfyUI_examples/audio/) - [ACE Step](https://comfyanonymous.github.io/ComfyUI_examples/audio/) From 0ec05b1481d12b299bc945dbd407b773cfb66483 Mon Sep 17 00:00:00 2001 From: Jedrzej Kosinski Date: Fri, 5 Dec 2025 11:05:38 -0800 Subject: [PATCH 229/647] Remove line made unnecessary (and wrong) after transformer_options was added to NextDiT's _forward definition (#11118) --- comfy/ldm/lumina/model.py | 1 - 1 file changed, 1 deletion(-) diff --git a/comfy/ldm/lumina/model.py b/comfy/ldm/lumina/model.py index f1c1a0ec3..6c24fed9b 100644 --- a/comfy/ldm/lumina/model.py +++ b/comfy/ldm/lumina/model.py @@ -586,7 +586,6 @@ class NextDiT(nn.Module): cap_feats = self.cap_embedder(cap_feats) # (N, L, D) # todo check if able to batchify w.o. redundant compute patches = transformer_options.get("patches", {}) - transformer_options = kwargs.get("transformer_options", {}) x_is_tensor = isinstance(x, torch.Tensor) img, mask, img_size, cap_size, freqs_cis = self.patchify_and_embed(x, cap_feats, cap_mask, t, num_tokens, transformer_options=transformer_options) freqs_cis = freqs_cis.to(img.device) From 43071e3de3780f984a46549e90935a0bf405e9df Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Fri, 5 Dec 2025 11:35:42 -0800 Subject: [PATCH 230/647] Make old scaled fp8 format use the new mixed quant ops system. (#11000) --- comfy/model_base.py | 14 +- comfy/model_detection.py | 33 +--- comfy/model_patcher.py | 20 +-- comfy/ops.py | 145 +++++++----------- comfy/quant_ops.py | 30 ++-- comfy/sd.py | 68 ++++++-- comfy/sd1_clip.py | 22 +-- comfy/supported_models_base.py | 3 +- comfy/text_encoders/cosmos.py | 12 +- comfy/text_encoders/flux.py | 12 +- comfy/text_encoders/genmo.py | 6 +- comfy/text_encoders/hidream.py | 10 +- comfy/text_encoders/hunyuan_image.py | 12 +- comfy/text_encoders/hunyuan_video.py | 23 ++- comfy/text_encoders/lumina2.py | 6 +- comfy/text_encoders/omnigen2.py | 6 +- comfy/text_encoders/ovis.py | 5 +- comfy/text_encoders/pixart_t5.py | 6 +- comfy/text_encoders/qwen_image.py | 6 +- comfy/text_encoders/sd3_clip.py | 19 +-- comfy/text_encoders/wan.py | 6 +- comfy/text_encoders/z_image.py | 5 +- comfy/utils.py | 66 ++++++++ .../comfy_quant/test_mixed_precision.py | 18 ++- 24 files changed, 278 insertions(+), 275 deletions(-) diff --git a/comfy/model_base.py b/comfy/model_base.py index 9b76c285e..3cedd4f31 100644 --- a/comfy/model_base.py +++ b/comfy/model_base.py @@ -134,7 +134,7 @@ class BaseModel(torch.nn.Module): if not unet_config.get("disable_unet_model_creation", False): if model_config.custom_operations is None: fp8 = model_config.optimizations.get("fp8", False) - operations = comfy.ops.pick_operations(unet_config.get("dtype", None), self.manual_cast_dtype, fp8_optimizations=fp8, scaled_fp8=model_config.scaled_fp8, model_config=model_config) + operations = comfy.ops.pick_operations(unet_config.get("dtype", None), self.manual_cast_dtype, fp8_optimizations=fp8, model_config=model_config) else: operations = model_config.custom_operations self.diffusion_model = unet_model(**unet_config, device=device, operations=operations) @@ -329,18 +329,6 @@ class BaseModel(torch.nn.Module): extra_sds.append(self.model_config.process_clip_vision_state_dict_for_saving(clip_vision_state_dict)) unet_state_dict = self.diffusion_model.state_dict() - - if self.model_config.scaled_fp8 is not None: - unet_state_dict["scaled_fp8"] = torch.tensor([], dtype=self.model_config.scaled_fp8) - - # Save mixed precision metadata - if hasattr(self.model_config, 'layer_quant_config') and self.model_config.layer_quant_config: - metadata = { - "format_version": "1.0", - "layers": self.model_config.layer_quant_config - } - unet_state_dict["_quantization_metadata"] = metadata - unet_state_dict = self.model_config.process_unet_state_dict_for_saving(unet_state_dict) if self.model_type == ModelType.V_PREDICTION: diff --git a/comfy/model_detection.py b/comfy/model_detection.py index 7d0517e61..fd1907627 100644 --- a/comfy/model_detection.py +++ b/comfy/model_detection.py @@ -6,20 +6,6 @@ import math import logging import torch - -def detect_layer_quantization(metadata): - quant_key = "_quantization_metadata" - if metadata is not None and quant_key in metadata: - quant_metadata = metadata.pop(quant_key) - quant_metadata = json.loads(quant_metadata) - if isinstance(quant_metadata, dict) and "layers" in quant_metadata: - logging.info(f"Found quantization metadata (version {quant_metadata.get('format_version', 'unknown')})") - return quant_metadata["layers"] - else: - raise ValueError("Invalid quantization metadata format") - return None - - def count_blocks(state_dict_keys, prefix_string): count = 0 while True: @@ -767,22 +753,11 @@ def model_config_from_unet(state_dict, unet_key_prefix, use_base_if_no_match=Fal if model_config is None and use_base_if_no_match: model_config = comfy.supported_models_base.BASE(unet_config) - scaled_fp8_key = "{}scaled_fp8".format(unet_key_prefix) - if scaled_fp8_key in state_dict: - scaled_fp8_weight = state_dict.pop(scaled_fp8_key) - model_config.scaled_fp8 = scaled_fp8_weight.dtype - if model_config.scaled_fp8 == torch.float32: - model_config.scaled_fp8 = torch.float8_e4m3fn - if scaled_fp8_weight.nelement() == 2: - model_config.optimizations["fp8"] = False - else: - model_config.optimizations["fp8"] = True - # Detect per-layer quantization (mixed precision) - layer_quant_config = detect_layer_quantization(metadata) - if layer_quant_config: - model_config.layer_quant_config = layer_quant_config - logging.info(f"Detected mixed precision quantization: {len(layer_quant_config)} layers quantized") + quant_config = comfy.utils.detect_layer_quantization(state_dict, unet_key_prefix) + if quant_config: + model_config.quant_config = quant_config + logging.info("Detected mixed precision quantization") return model_config diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index 3dcac3eef..215784874 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -126,27 +126,11 @@ class LowVramPatch: def __init__(self, key, patches, convert_func=None, set_func=None): self.key = key self.patches = patches - self.convert_func = convert_func + self.convert_func = convert_func # TODO: remove self.set_func = set_func def __call__(self, weight): - intermediate_dtype = weight.dtype - if self.convert_func is not None: - weight = self.convert_func(weight, inplace=False) - - if intermediate_dtype not in [torch.float32, torch.float16, torch.bfloat16]: #intermediate_dtype has to be one that is supported in math ops - intermediate_dtype = torch.float32 - out = comfy.lora.calculate_weight(self.patches[self.key], weight.to(intermediate_dtype), self.key, intermediate_dtype=intermediate_dtype) - if self.set_func is None: - return comfy.float.stochastic_rounding(out, weight.dtype, seed=string_to_seed(self.key)) - else: - return self.set_func(out, seed=string_to_seed(self.key), return_weight=True) - - out = comfy.lora.calculate_weight(self.patches[self.key], weight, self.key, intermediate_dtype=intermediate_dtype) - if self.set_func is not None: - return self.set_func(out, seed=string_to_seed(self.key), return_weight=True).to(dtype=intermediate_dtype) - else: - return out + return comfy.lora.calculate_weight(self.patches[self.key], weight, self.key, intermediate_dtype=weight.dtype) #The above patch logic may cast up the weight to fp32, and do math. Go with fp32 x 3 LOWVRAM_PATCH_ESTIMATE_MATH_FACTOR = 3 diff --git a/comfy/ops.py b/comfy/ops.py index eae434e68..dc06709a1 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -23,6 +23,7 @@ from comfy.cli_args import args, PerformanceFeature import comfy.float import comfy.rmsnorm import contextlib +import json def run_every_op(): if torch.compiler.is_compiling(): @@ -422,22 +423,12 @@ def fp8_linear(self, input): if input.ndim == 3 or input.ndim == 2: w, bias, offload_stream = cast_bias_weight(self, input, dtype=dtype, bias_dtype=input_dtype, offloadable=True) + scale_weight = torch.ones((), device=input.device, dtype=torch.float32) - scale_weight = self.scale_weight - scale_input = self.scale_input - if scale_weight is None: - scale_weight = torch.ones((), device=input.device, dtype=torch.float32) - else: - scale_weight = scale_weight.to(input.device) - - if scale_input is None: - scale_input = torch.ones((), device=input.device, dtype=torch.float32) - input = torch.clamp(input, min=-448, max=448, out=input) - layout_params_weight = {'scale': scale_input, 'orig_dtype': input_dtype} - quantized_input = QuantizedTensor(input.to(dtype).contiguous(), "TensorCoreFP8Layout", layout_params_weight) - else: - scale_input = scale_input.to(input.device) - quantized_input = QuantizedTensor.from_float(input, "TensorCoreFP8Layout", scale=scale_input, dtype=dtype) + scale_input = torch.ones((), device=input.device, dtype=torch.float32) + input = torch.clamp(input, min=-448, max=448, out=input) + layout_params_weight = {'scale': scale_input, 'orig_dtype': input_dtype} + quantized_input = QuantizedTensor(input.to(dtype).contiguous(), "TensorCoreFP8Layout", layout_params_weight) # Wrap weight in QuantizedTensor - this enables unified dispatch # Call F.linear - __torch_dispatch__ routes to fp8_linear handler in quant_ops.py! @@ -458,7 +449,7 @@ class fp8_ops(manual_cast): return None def forward_comfy_cast_weights(self, input): - if not self.training: + if len(self.weight_function) == 0 and len(self.bias_function) == 0: try: out = fp8_linear(self, input) if out is not None: @@ -471,59 +462,6 @@ class fp8_ops(manual_cast): uncast_bias_weight(self, weight, bias, offload_stream) return x -def scaled_fp8_ops(fp8_matrix_mult=False, scale_input=False, override_dtype=None): - logging.info("Using scaled fp8: fp8 matrix mult: {}, scale input: {}".format(fp8_matrix_mult, scale_input)) - class scaled_fp8_op(manual_cast): - class Linear(manual_cast.Linear): - def __init__(self, *args, **kwargs): - if override_dtype is not None: - kwargs['dtype'] = override_dtype - super().__init__(*args, **kwargs) - - def reset_parameters(self): - if not hasattr(self, 'scale_weight'): - self.scale_weight = torch.nn.parameter.Parameter(data=torch.ones((), device=self.weight.device, dtype=torch.float32), requires_grad=False) - - if not scale_input: - self.scale_input = None - - if not hasattr(self, 'scale_input'): - self.scale_input = torch.nn.parameter.Parameter(data=torch.ones((), device=self.weight.device, dtype=torch.float32), requires_grad=False) - return None - - def forward_comfy_cast_weights(self, input): - if fp8_matrix_mult: - out = fp8_linear(self, input) - if out is not None: - return out - - weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True) - - if weight.numel() < input.numel(): #TODO: optimize - x = torch.nn.functional.linear(input, weight * self.scale_weight.to(device=weight.device, dtype=weight.dtype), bias) - else: - x = torch.nn.functional.linear(input * self.scale_weight.to(device=weight.device, dtype=weight.dtype), weight, bias) - uncast_bias_weight(self, weight, bias, offload_stream) - return x - - def convert_weight(self, weight, inplace=False, **kwargs): - if inplace: - weight *= self.scale_weight.to(device=weight.device, dtype=weight.dtype) - return weight - else: - return weight.to(dtype=torch.float32) * self.scale_weight.to(device=weight.device, dtype=torch.float32) - - def set_weight(self, weight, inplace_update=False, seed=None, return_weight=False, **kwargs): - weight = comfy.float.stochastic_rounding(weight / self.scale_weight.to(device=weight.device, dtype=weight.dtype), self.weight.dtype, seed=seed) - if return_weight: - return weight - if inplace_update: - self.weight.data.copy_(weight) - else: - self.weight = torch.nn.Parameter(weight, requires_grad=False) - - return scaled_fp8_op - CUBLAS_IS_AVAILABLE = False try: from cublas_ops import CublasLinear @@ -550,9 +488,9 @@ if CUBLAS_IS_AVAILABLE: from .quant_ops import QuantizedTensor, QUANT_ALGOS -def mixed_precision_ops(layer_quant_config={}, compute_dtype=torch.bfloat16, full_precision_mm=False): +def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_precision_mm=False): class MixedPrecisionOps(manual_cast): - _layer_quant_config = layer_quant_config + _quant_config = quant_config _compute_dtype = compute_dtype _full_precision_mm = full_precision_mm @@ -595,27 +533,36 @@ def mixed_precision_ops(layer_quant_config={}, compute_dtype=torch.bfloat16, ful manually_loaded_keys = [weight_key] - if layer_name not in MixedPrecisionOps._layer_quant_config: + layer_conf = state_dict.pop(f"{prefix}comfy_quant", None) + if layer_conf is not None: + layer_conf = json.loads(layer_conf.numpy().tobytes()) + + if layer_conf is None: self.weight = torch.nn.Parameter(weight.to(device=device, dtype=MixedPrecisionOps._compute_dtype), requires_grad=False) else: - quant_format = MixedPrecisionOps._layer_quant_config[layer_name].get("format", None) - if quant_format is None: + self.quant_format = layer_conf.get("format", None) + if not self._full_precision_mm: + self._full_precision_mm = layer_conf.get("full_precision_matrix_mult", False) + + if self.quant_format is None: raise ValueError(f"Unknown quantization format for layer {layer_name}") - qconfig = QUANT_ALGOS[quant_format] + qconfig = QUANT_ALGOS[self.quant_format] self.layout_type = qconfig["comfy_tensor_layout"] weight_scale_key = f"{prefix}weight_scale" + scale = state_dict.pop(weight_scale_key, None) layout_params = { - 'scale': state_dict.pop(weight_scale_key, None), + 'scale': scale, 'orig_dtype': MixedPrecisionOps._compute_dtype, 'block_size': qconfig.get("group_size", None), } - if layout_params['scale'] is not None: + + if scale is not None: manually_loaded_keys.append(weight_scale_key) self.weight = torch.nn.Parameter( - QuantizedTensor(weight.to(device=device), self.layout_type, layout_params), + QuantizedTensor(weight.to(device=device, dtype=qconfig.get("storage_t", None)), self.layout_type, layout_params), requires_grad=False ) @@ -624,7 +571,7 @@ def mixed_precision_ops(layer_quant_config={}, compute_dtype=torch.bfloat16, ful _v = state_dict.pop(param_key, None) if _v is None: continue - setattr(self, param_name, torch.nn.Parameter(_v.to(device=device), requires_grad=False)) + self.register_parameter(param_name, torch.nn.Parameter(_v.to(device=device), requires_grad=False)) manually_loaded_keys.append(param_key) super()._load_from_state_dict(state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs) @@ -633,6 +580,16 @@ def mixed_precision_ops(layer_quant_config={}, compute_dtype=torch.bfloat16, ful if key in missing_keys: missing_keys.remove(key) + def state_dict(self, *args, destination=None, prefix="", **kwargs): + sd = super().state_dict(*args, destination=destination, prefix=prefix, **kwargs) + if isinstance(self.weight, QuantizedTensor): + sd["{}weight_scale".format(prefix)] = self.weight._layout_params['scale'] + quant_conf = {"format": self.quant_format} + if self._full_precision_mm: + quant_conf["full_precision_matrix_mult"] = True + sd["{}comfy_quant".format(prefix)] = torch.frombuffer(json.dumps(quant_conf).encode('utf-8'), dtype=torch.uint8) + return sd + def _forward(self, input, weight, bias): return torch.nn.functional.linear(input, weight, bias) @@ -648,9 +605,8 @@ def mixed_precision_ops(layer_quant_config={}, compute_dtype=torch.bfloat16, ful if self._full_precision_mm or self.comfy_cast_weights or len(self.weight_function) > 0 or len(self.bias_function) > 0: return self.forward_comfy_cast_weights(input, *args, **kwargs) if (getattr(self, 'layout_type', None) is not None and - getattr(self, 'input_scale', None) is not None and not isinstance(input, QuantizedTensor)): - input = QuantizedTensor.from_float(input, self.layout_type, scale=self.input_scale, dtype=self.weight.dtype) + input = QuantizedTensor.from_float(input, self.layout_type, scale=getattr(self, 'input_scale', None), dtype=self.weight.dtype) return self._forward(input, self.weight, self.bias) def convert_weight(self, weight, inplace=False, **kwargs): @@ -661,7 +617,7 @@ def mixed_precision_ops(layer_quant_config={}, compute_dtype=torch.bfloat16, ful def set_weight(self, weight, inplace_update=False, seed=None, return_weight=False, **kwargs): if getattr(self, 'layout_type', None) is not None: - weight = QuantizedTensor.from_float(weight, self.layout_type, scale=None, dtype=self.weight.dtype, stochastic_rounding=seed, inplace_ops=True) + weight = QuantizedTensor.from_float(weight, self.layout_type, scale="recalculate", dtype=self.weight.dtype, stochastic_rounding=seed, inplace_ops=True) else: weight = weight.to(self.weight.dtype) if return_weight: @@ -670,17 +626,28 @@ def mixed_precision_ops(layer_quant_config={}, compute_dtype=torch.bfloat16, ful assert inplace_update is False # TODO: eventually remove the inplace_update stuff self.weight = torch.nn.Parameter(weight, requires_grad=False) + def _apply(self, fn, recurse=True): # This is to get torch.compile + moving weights to another device working + if recurse: + for module in self.children(): + module._apply(fn) + + for key, param in self._parameters.items(): + if param is None: + continue + self.register_parameter(key, torch.nn.Parameter(fn(param), requires_grad=False)) + for key, buf in self._buffers.items(): + if buf is not None: + self._buffers[key] = fn(buf) + return self + return MixedPrecisionOps -def pick_operations(weight_dtype, compute_dtype, load_device=None, disable_fast_fp8=False, fp8_optimizations=False, scaled_fp8=None, model_config=None): +def pick_operations(weight_dtype, compute_dtype, load_device=None, disable_fast_fp8=False, fp8_optimizations=False, model_config=None): fp8_compute = comfy.model_management.supports_fp8_compute(load_device) # TODO: if we support more ops this needs to be more granular - if model_config and hasattr(model_config, 'layer_quant_config') and model_config.layer_quant_config: - logging.info(f"Using mixed precision operations: {len(model_config.layer_quant_config)} quantized layers") - return mixed_precision_ops(model_config.layer_quant_config, compute_dtype, full_precision_mm=not fp8_compute) - - if scaled_fp8 is not None: - return scaled_fp8_ops(fp8_matrix_mult=fp8_compute and fp8_optimizations, scale_input=fp8_optimizations, override_dtype=scaled_fp8) + if model_config and hasattr(model_config, 'quant_config') and model_config.quant_config: + logging.info("Using mixed precision operations") + return mixed_precision_ops(model_config.quant_config, compute_dtype, full_precision_mm=not fp8_compute) if ( fp8_compute and diff --git a/comfy/quant_ops.py b/comfy/quant_ops.py index bb1fb860c..571d3f760 100644 --- a/comfy/quant_ops.py +++ b/comfy/quant_ops.py @@ -238,6 +238,9 @@ class QuantizedTensor(torch.Tensor): def is_contiguous(self, *arg, **kwargs): return self._qdata.is_contiguous(*arg, **kwargs) + def storage(self): + return self._qdata.storage() + # ============================================================================== # Generic Utilities (Layout-Agnostic Operations) # ============================================================================== @@ -249,12 +252,6 @@ def _create_transformed_qtensor(qt, transform_fn): def _handle_device_transfer(qt, target_device, target_dtype=None, target_layout=None, op_name="to"): - if target_dtype is not None and target_dtype != qt.dtype: - logging.warning( - f"QuantizedTensor: dtype conversion requested to {target_dtype}, " - f"but not supported for quantized tensors. Ignoring dtype." - ) - if target_layout is not None and target_layout != torch.strided: logging.warning( f"QuantizedTensor: layout change requested to {target_layout}, " @@ -274,6 +271,8 @@ def _handle_device_transfer(qt, target_device, target_dtype=None, target_layout= logging.debug(f"QuantizedTensor.{op_name}: Moving from {current_device} to {target_device}") new_q_data = qt._qdata.to(device=target_device) new_params = _move_layout_params_to_device(qt._layout_params, target_device) + if target_dtype is not None: + new_params["orig_dtype"] = target_dtype new_qt = QuantizedTensor(new_q_data, qt._layout_type, new_params) logging.debug(f"QuantizedTensor.{op_name}: Created new tensor on {target_device}") return new_qt @@ -339,7 +338,9 @@ def generic_copy_(func, args, kwargs): # Copy from another quantized tensor qt_dest._qdata.copy_(src._qdata, non_blocking=non_blocking) qt_dest._layout_type = src._layout_type + orig_dtype = qt_dest._layout_params["orig_dtype"] _copy_layout_params_inplace(src._layout_params, qt_dest._layout_params, non_blocking=non_blocking) + qt_dest._layout_params["orig_dtype"] = orig_dtype else: # Copy from regular tensor - just copy raw data qt_dest._qdata.copy_(src) @@ -397,17 +398,20 @@ class TensorCoreFP8Layout(QuantizedLayout): def quantize(cls, tensor, scale=None, dtype=torch.float8_e4m3fn, stochastic_rounding=0, inplace_ops=False): orig_dtype = tensor.dtype - if scale is None: + if isinstance(scale, str) and scale == "recalculate": scale = torch.amax(tensor.abs()) / torch.finfo(dtype).max - if not isinstance(scale, torch.Tensor): - scale = torch.tensor(scale) - scale = scale.to(device=tensor.device, dtype=torch.float32) + if scale is not None: + if not isinstance(scale, torch.Tensor): + scale = torch.tensor(scale) + scale = scale.to(device=tensor.device, dtype=torch.float32) - if inplace_ops: - tensor *= (1.0 / scale).to(tensor.dtype) + if inplace_ops: + tensor *= (1.0 / scale).to(tensor.dtype) + else: + tensor = tensor * (1.0 / scale).to(tensor.dtype) else: - tensor = tensor * (1.0 / scale).to(tensor.dtype) + scale = torch.ones((), device=tensor.device, dtype=torch.float32) if stochastic_rounding > 0: tensor = comfy.float.stochastic_rounding(tensor, dtype=dtype, seed=stochastic_rounding) diff --git a/comfy/sd.py b/comfy/sd.py index 03bdb33d5..092715d79 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -968,10 +968,8 @@ def load_clip(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DI clip_data = [] for p in ckpt_paths: sd, metadata = comfy.utils.load_torch_file(p, safe_load=True, return_metadata=True) - if metadata is not None: - quant_metadata = metadata.get("_quantization_metadata", None) - if quant_metadata is not None: - sd["_quantization_metadata"] = quant_metadata + if model_options.get("custom_operations", None) is None: + sd, metadata = comfy.utils.convert_old_quants(sd, model_prefix="", metadata=metadata) clip_data.append(sd) return load_text_encoder_state_dicts(clip_data, embedding_directory=embedding_directory, clip_type=clip_type, model_options=model_options) @@ -1088,7 +1086,7 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip clip_target.clip = comfy.text_encoders.sd3_clip.sd3_clip(clip_l=False, clip_g=True, t5=False) clip_target.tokenizer = comfy.text_encoders.sd3_clip.SD3Tokenizer elif clip_type == CLIPType.HIDREAM: - clip_target.clip = comfy.text_encoders.hidream.hidream_clip(clip_l=False, clip_g=True, t5=False, llama=False, dtype_t5=None, dtype_llama=None, t5xxl_scaled_fp8=None, llama_scaled_fp8=None) + clip_target.clip = comfy.text_encoders.hidream.hidream_clip(clip_l=False, clip_g=True, t5=False, llama=False, dtype_t5=None, dtype_llama=None) clip_target.tokenizer = comfy.text_encoders.hidream.HiDreamTokenizer else: clip_target.clip = sdxl_clip.SDXLRefinerClipModel @@ -1112,7 +1110,7 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip tokenizer_data["spiece_model"] = clip_data[0].get("spiece_model", None) elif clip_type == CLIPType.HIDREAM: clip_target.clip = comfy.text_encoders.hidream.hidream_clip(**t5xxl_detect(clip_data), - clip_l=False, clip_g=False, t5=True, llama=False, dtype_llama=None, llama_scaled_fp8=None) + clip_l=False, clip_g=False, t5=True, llama=False, dtype_llama=None) clip_target.tokenizer = comfy.text_encoders.hidream.HiDreamTokenizer else: #CLIPType.MOCHI clip_target.clip = comfy.text_encoders.genmo.mochi_te(**t5xxl_detect(clip_data)) @@ -1141,7 +1139,7 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip tokenizer_data["spiece_model"] = clip_data[0].get("spiece_model", None) elif te_model == TEModel.LLAMA3_8: clip_target.clip = comfy.text_encoders.hidream.hidream_clip(**llama_detect(clip_data), - clip_l=False, clip_g=False, t5=False, llama=True, dtype_t5=None, t5xxl_scaled_fp8=None) + clip_l=False, clip_g=False, t5=False, llama=True, dtype_t5=None) clip_target.tokenizer = comfy.text_encoders.hidream.HiDreamTokenizer elif te_model == TEModel.QWEN25_3B: clip_target.clip = comfy.text_encoders.omnigen2.te(**llama_detect(clip_data)) @@ -1169,7 +1167,7 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip clip_target.clip = comfy.text_encoders.sd3_clip.sd3_clip(clip_l=True, clip_g=False, t5=False) clip_target.tokenizer = comfy.text_encoders.sd3_clip.SD3Tokenizer elif clip_type == CLIPType.HIDREAM: - clip_target.clip = comfy.text_encoders.hidream.hidream_clip(clip_l=True, clip_g=False, t5=False, llama=False, dtype_t5=None, dtype_llama=None, t5xxl_scaled_fp8=None, llama_scaled_fp8=None) + clip_target.clip = comfy.text_encoders.hidream.hidream_clip(clip_l=True, clip_g=False, t5=False, llama=False, dtype_t5=None, dtype_llama=None) clip_target.tokenizer = comfy.text_encoders.hidream.HiDreamTokenizer else: clip_target.clip = sd1_clip.SD1ClipModel @@ -1224,8 +1222,6 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip parameters = 0 for c in clip_data: - if "_quantization_metadata" in c: - c.pop("_quantization_metadata") parameters += comfy.utils.calculate_parameters(c) tokenizer_data, model_options = comfy.text_encoders.long_clipl.model_options_long_clip(c, tokenizer_data, model_options) @@ -1295,6 +1291,10 @@ def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_c weight_dtype = comfy.utils.weight_dtype(sd, diffusion_model_prefix) load_device = model_management.get_torch_device() + custom_operations = model_options.get("custom_operations", None) + if custom_operations is None: + sd, metadata = comfy.utils.convert_old_quants(sd, diffusion_model_prefix, metadata=metadata) + model_config = model_detection.model_config_from_unet(sd, diffusion_model_prefix, metadata=metadata) if model_config is None: logging.warning("Warning, This is not a checkpoint file, trying to load it as a diffusion model only.") @@ -1303,18 +1303,22 @@ def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_c return None return (diffusion_model, None, VAE(sd={}), None) # The VAE object is there to throw an exception if it's actually used' - unet_weight_dtype = list(model_config.supported_inference_dtypes) - if model_config.scaled_fp8 is not None: + if model_config.quant_config is not None: weight_dtype = None - model_config.custom_operations = model_options.get("custom_operations", None) + if custom_operations is not None: + model_config.custom_operations = custom_operations + unet_dtype = model_options.get("dtype", model_options.get("weight_dtype", None)) if unet_dtype is None: unet_dtype = model_management.unet_dtype(model_params=parameters, supported_dtypes=unet_weight_dtype, weight_dtype=weight_dtype) - manual_cast_dtype = model_management.unet_manual_cast(unet_dtype, load_device, model_config.supported_inference_dtypes) + if model_config.quant_config is not None: + manual_cast_dtype = model_management.unet_manual_cast(None, load_device, model_config.supported_inference_dtypes) + else: + manual_cast_dtype = model_management.unet_manual_cast(unet_dtype, load_device, model_config.supported_inference_dtypes) model_config.set_inference_dtype(unet_dtype, manual_cast_dtype) if model_config.clip_vision_prefix is not None: @@ -1332,6 +1336,27 @@ def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_c vae = VAE(sd=vae_sd, metadata=metadata) if output_clip: + if te_model_options.get("custom_operations", None) is None: + scaled_fp8_list = [] + for k in list(sd.keys()): # Convert scaled fp8 to mixed ops + if k.endswith(".scaled_fp8"): + scaled_fp8_list.append(k[:-len("scaled_fp8")]) + + if len(scaled_fp8_list) > 0: + out_sd = {} + for k in sd: + skip = False + for pref in scaled_fp8_list: + skip = skip or k.startswith(pref) + if not skip: + out_sd[k] = sd[k] + + for pref in scaled_fp8_list: + quant_sd, qmetadata = comfy.utils.convert_old_quants(sd, pref, metadata={}) + for k in quant_sd: + out_sd[k] = quant_sd[k] + sd = out_sd + clip_target = model_config.clip_target(state_dict=sd) if clip_target is not None: clip_sd = model_config.process_clip_state_dict(sd) @@ -1394,6 +1419,9 @@ def load_diffusion_model_state_dict(sd, model_options={}, metadata=None): if len(temp_sd) > 0: sd = temp_sd + custom_operations = model_options.get("custom_operations", None) + if custom_operations is None: + sd, metadata = comfy.utils.convert_old_quants(sd, "", metadata=metadata) parameters = comfy.utils.calculate_parameters(sd) weight_dtype = comfy.utils.weight_dtype(sd) @@ -1424,7 +1452,7 @@ def load_diffusion_model_state_dict(sd, model_options={}, metadata=None): offload_device = model_management.unet_offload_device() unet_weight_dtype = list(model_config.supported_inference_dtypes) - if model_config.scaled_fp8 is not None: + if model_config.quant_config is not None: weight_dtype = None if dtype is None: @@ -1432,12 +1460,15 @@ def load_diffusion_model_state_dict(sd, model_options={}, metadata=None): else: unet_dtype = dtype - if model_config.layer_quant_config is not None: + if model_config.quant_config is not None: manual_cast_dtype = model_management.unet_manual_cast(None, load_device, model_config.supported_inference_dtypes) else: manual_cast_dtype = model_management.unet_manual_cast(unet_dtype, load_device, model_config.supported_inference_dtypes) model_config.set_inference_dtype(unet_dtype, manual_cast_dtype) - model_config.custom_operations = model_options.get("custom_operations", model_config.custom_operations) + + if custom_operations is not None: + model_config.custom_operations = custom_operations + if model_options.get("fp8_optimizations", False): model_config.optimizations["fp8"] = True @@ -1476,6 +1507,9 @@ def save_checkpoint(output_path, model, clip=None, vae=None, clip_vision=None, m if vae is not None: vae_sd = vae.get_sd() + if metadata is None: + metadata = {} + model_management.load_models_gpu(load_models, force_patch_weights=True) clip_vision_sd = clip_vision.get_sd() if clip_vision is not None else None sd = model.model.state_dict_for_saving(clip_sd, vae_sd, clip_vision_sd) diff --git a/comfy/sd1_clip.py b/comfy/sd1_clip.py index 503a51843..962948dae 100644 --- a/comfy/sd1_clip.py +++ b/comfy/sd1_clip.py @@ -107,29 +107,17 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder): config[k] = v operations = model_options.get("custom_operations", None) - scaled_fp8 = None - quantization_metadata = model_options.get("quantization_metadata", None) + quant_config = model_options.get("quantization_metadata", None) if operations is None: - layer_quant_config = None - if quantization_metadata is not None: - layer_quant_config = json.loads(quantization_metadata).get("layers", None) - - if layer_quant_config is not None: - operations = comfy.ops.mixed_precision_ops(layer_quant_config, dtype, full_precision_mm=True) - logging.info(f"Using MixedPrecisionOps for text encoder: {len(layer_quant_config)} quantized layers") + if quant_config is not None: + operations = comfy.ops.mixed_precision_ops(quant_config, dtype, full_precision_mm=True) + logging.info("Using MixedPrecisionOps for text encoder") else: - # Fallback to scaled_fp8_ops for backward compatibility - scaled_fp8 = model_options.get("scaled_fp8", None) - if scaled_fp8 is not None: - operations = comfy.ops.scaled_fp8_ops(fp8_matrix_mult=False, override_dtype=scaled_fp8) - else: - operations = comfy.ops.manual_cast + operations = comfy.ops.manual_cast self.operations = operations self.transformer = model_class(config, dtype, device, self.operations) - if scaled_fp8 is not None: - self.transformer.scaled_fp8 = torch.nn.Parameter(torch.tensor([], dtype=scaled_fp8)) self.num_layers = self.transformer.num_layers diff --git a/comfy/supported_models_base.py b/comfy/supported_models_base.py index e4bd74514..9fd84d329 100644 --- a/comfy/supported_models_base.py +++ b/comfy/supported_models_base.py @@ -49,8 +49,7 @@ class BASE: manual_cast_dtype = None custom_operations = None - scaled_fp8 = None - layer_quant_config = None # Per-layer quantization configuration for mixed precision + quant_config = None # quantization configuration for mixed precision optimizations = {"fp8": False} @classmethod diff --git a/comfy/text_encoders/cosmos.py b/comfy/text_encoders/cosmos.py index a1adb5242..448381fa9 100644 --- a/comfy/text_encoders/cosmos.py +++ b/comfy/text_encoders/cosmos.py @@ -7,10 +7,10 @@ from transformers import T5TokenizerFast class T5XXLModel(sd1_clip.SDClipModel): def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, attention_mask=True, model_options={}): textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_old_config_xxl.json") - t5xxl_scaled_fp8 = model_options.get("t5xxl_scaled_fp8", None) - if t5xxl_scaled_fp8 is not None: + t5xxl_quantization_metadata = model_options.get("t5xxl_quantization_metadata", None) + if t5xxl_quantization_metadata is not None: model_options = model_options.copy() - model_options["scaled_fp8"] = t5xxl_scaled_fp8 + model_options["quantization_metadata"] = t5xxl_quantization_metadata super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 1, "pad": 0}, model_class=comfy.text_encoders.t5.T5, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, zero_out_masked=attention_mask, model_options=model_options) @@ -30,12 +30,12 @@ class CosmosT5Tokenizer(sd1_clip.SD1Tokenizer): super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, clip_name="t5xxl", tokenizer=T5XXLTokenizer) -def te(dtype_t5=None, t5xxl_scaled_fp8=None): +def te(dtype_t5=None, t5_quantization_metadata=None): class CosmosTEModel_(CosmosT5XXL): def __init__(self, device="cpu", dtype=None, model_options={}): - if t5xxl_scaled_fp8 is not None and "t5xxl_scaled_fp8" not in model_options: + if t5_quantization_metadata is not None: model_options = model_options.copy() - model_options["t5xxl_scaled_fp8"] = t5xxl_scaled_fp8 + model_options["t5xxl_quantization_metadata"] = t5_quantization_metadata if dtype is None: dtype = dtype_t5 super().__init__(device=device, dtype=dtype, model_options=model_options) diff --git a/comfy/text_encoders/flux.py b/comfy/text_encoders/flux.py index 99f4812bb..21d93d757 100644 --- a/comfy/text_encoders/flux.py +++ b/comfy/text_encoders/flux.py @@ -63,12 +63,12 @@ class FluxClipModel(torch.nn.Module): else: return self.t5xxl.load_sd(sd) -def flux_clip(dtype_t5=None, t5xxl_scaled_fp8=None): +def flux_clip(dtype_t5=None, t5_quantization_metadata=None): class FluxClipModel_(FluxClipModel): def __init__(self, device="cpu", dtype=None, model_options={}): - if t5xxl_scaled_fp8 is not None and "t5xxl_scaled_fp8" not in model_options: + if t5_quantization_metadata is not None: model_options = model_options.copy() - model_options["t5xxl_scaled_fp8"] = t5xxl_scaled_fp8 + model_options["t5xxl_quantization_metadata"] = t5_quantization_metadata super().__init__(dtype_t5=dtype_t5, device=device, dtype=dtype, model_options=model_options) return FluxClipModel_ @@ -159,15 +159,13 @@ class Flux2TEModel(sd1_clip.SD1ClipModel): out = out.reshape(out.shape[0], out.shape[1], -1) return out, pooled, extra -def flux2_te(dtype_llama=None, llama_scaled_fp8=None, llama_quantization_metadata=None, pruned=False): +def flux2_te(dtype_llama=None, llama_quantization_metadata=None, pruned=False): class Flux2TEModel_(Flux2TEModel): def __init__(self, device="cpu", dtype=None, model_options={}): - if llama_scaled_fp8 is not None and "scaled_fp8" not in model_options: - model_options = model_options.copy() - model_options["scaled_fp8"] = llama_scaled_fp8 if dtype_llama is not None: dtype = dtype_llama if llama_quantization_metadata is not None: + model_options = model_options.copy() model_options["quantization_metadata"] = llama_quantization_metadata if pruned: model_options = model_options.copy() diff --git a/comfy/text_encoders/genmo.py b/comfy/text_encoders/genmo.py index 9dcf190a2..5daea8135 100644 --- a/comfy/text_encoders/genmo.py +++ b/comfy/text_encoders/genmo.py @@ -26,12 +26,12 @@ class MochiT5Tokenizer(sd1_clip.SD1Tokenizer): super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, clip_name="t5xxl", tokenizer=T5XXLTokenizer) -def mochi_te(dtype_t5=None, t5xxl_scaled_fp8=None): +def mochi_te(dtype_t5=None, t5_quantization_metadata=None): class MochiTEModel_(MochiT5XXL): def __init__(self, device="cpu", dtype=None, model_options={}): - if t5xxl_scaled_fp8 is not None and "t5xxl_scaled_fp8" not in model_options: + if t5_quantization_metadata is not None: model_options = model_options.copy() - model_options["t5xxl_scaled_fp8"] = t5xxl_scaled_fp8 + model_options["t5xxl_quantization_metadata"] = t5_quantization_metadata if dtype is None: dtype = dtype_t5 super().__init__(device=device, dtype=dtype, model_options=model_options) diff --git a/comfy/text_encoders/hidream.py b/comfy/text_encoders/hidream.py index dbcf52784..600b34480 100644 --- a/comfy/text_encoders/hidream.py +++ b/comfy/text_encoders/hidream.py @@ -142,14 +142,14 @@ class HiDreamTEModel(torch.nn.Module): return self.llama.load_sd(sd) -def hidream_clip(clip_l=True, clip_g=True, t5=True, llama=True, dtype_t5=None, dtype_llama=None, t5xxl_scaled_fp8=None, llama_scaled_fp8=None): +def hidream_clip(clip_l=True, clip_g=True, t5=True, llama=True, dtype_t5=None, dtype_llama=None, t5_quantization_metadata=None, llama_quantization_metadata=None): class HiDreamTEModel_(HiDreamTEModel): def __init__(self, device="cpu", dtype=None, model_options={}): - if t5xxl_scaled_fp8 is not None and "t5xxl_scaled_fp8" not in model_options: + if t5_quantization_metadata is not None: model_options = model_options.copy() - model_options["t5xxl_scaled_fp8"] = t5xxl_scaled_fp8 - if llama_scaled_fp8 is not None and "llama_scaled_fp8" not in model_options: + model_options["t5xxl_quantization_metadata"] = t5_quantization_metadata + if llama_quantization_metadata is not None: model_options = model_options.copy() - model_options["llama_scaled_fp8"] = llama_scaled_fp8 + model_options["llama_quantization_metadata"] = llama_quantization_metadata super().__init__(clip_l=clip_l, clip_g=clip_g, t5=t5, llama=llama, dtype_t5=dtype_t5, dtype_llama=dtype_llama, device=device, dtype=dtype, model_options=model_options) return HiDreamTEModel_ diff --git a/comfy/text_encoders/hunyuan_image.py b/comfy/text_encoders/hunyuan_image.py index ff04726e1..cd198036c 100644 --- a/comfy/text_encoders/hunyuan_image.py +++ b/comfy/text_encoders/hunyuan_image.py @@ -40,10 +40,10 @@ class HunyuanImageTokenizer(QwenImageTokenizer): class Qwen25_7BVLIModel(sd1_clip.SDClipModel): def __init__(self, device="cpu", layer="hidden", layer_idx=-3, dtype=None, attention_mask=True, model_options={}): - llama_scaled_fp8 = model_options.get("qwen_scaled_fp8", None) - if llama_scaled_fp8 is not None: + llama_quantization_metadata = model_options.get("llama_quantization_metadata", None) + if llama_quantization_metadata is not None: model_options = model_options.copy() - model_options["scaled_fp8"] = llama_scaled_fp8 + model_options["quantization_metadata"] = llama_quantization_metadata super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"pad": 151643}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Qwen25_7BVLI, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options) @@ -91,12 +91,12 @@ class HunyuanImageTEModel(QwenImageTEModel): else: return super().load_sd(sd) -def te(byt5=True, dtype_llama=None, llama_scaled_fp8=None): +def te(byt5=True, dtype_llama=None, llama_quantization_metadata=None): class QwenImageTEModel_(HunyuanImageTEModel): def __init__(self, device="cpu", dtype=None, model_options={}): - if llama_scaled_fp8 is not None and "scaled_fp8" not in model_options: + if llama_quantization_metadata is not None: model_options = model_options.copy() - model_options["qwen_scaled_fp8"] = llama_scaled_fp8 + model_options["llama_quantization_metadata"] = llama_quantization_metadata if dtype_llama is not None: dtype = dtype_llama super().__init__(byt5=byt5, device=device, dtype=dtype, model_options=model_options) diff --git a/comfy/text_encoders/hunyuan_video.py b/comfy/text_encoders/hunyuan_video.py index 0110517bb..a9a6c525e 100644 --- a/comfy/text_encoders/hunyuan_video.py +++ b/comfy/text_encoders/hunyuan_video.py @@ -6,7 +6,7 @@ from transformers import LlamaTokenizerFast import torch import os import numbers - +import comfy.utils def llama_detect(state_dict, prefix=""): out = {} @@ -14,12 +14,9 @@ def llama_detect(state_dict, prefix=""): if t5_key in state_dict: out["dtype_llama"] = state_dict[t5_key].dtype - scaled_fp8_key = "{}scaled_fp8".format(prefix) - if scaled_fp8_key in state_dict: - out["llama_scaled_fp8"] = state_dict[scaled_fp8_key].dtype - - if "_quantization_metadata" in state_dict: - out["llama_quantization_metadata"] = state_dict["_quantization_metadata"] + quant = comfy.utils.detect_layer_quantization(state_dict, prefix) + if quant is not None: + out["llama_quantization_metadata"] = quant return out @@ -31,10 +28,10 @@ class LLAMA3Tokenizer(sd1_clip.SDTokenizer): class LLAMAModel(sd1_clip.SDClipModel): def __init__(self, device="cpu", layer="hidden", layer_idx=-3, dtype=None, attention_mask=True, model_options={}, special_tokens={"start": 128000, "pad": 128258}): - llama_scaled_fp8 = model_options.get("llama_scaled_fp8", None) - if llama_scaled_fp8 is not None: + llama_quantization_metadata = model_options.get("llama_quantization_metadata", None) + if llama_quantization_metadata is not None: model_options = model_options.copy() - model_options["scaled_fp8"] = llama_scaled_fp8 + model_options["quantization_metadata"] = llama_quantization_metadata textmodel_json_config = {} vocab_size = model_options.get("vocab_size", None) @@ -161,11 +158,11 @@ class HunyuanVideoClipModel(torch.nn.Module): return self.llama.load_sd(sd) -def hunyuan_video_clip(dtype_llama=None, llama_scaled_fp8=None): +def hunyuan_video_clip(dtype_llama=None, llama_quantization_metadata=None): class HunyuanVideoClipModel_(HunyuanVideoClipModel): def __init__(self, device="cpu", dtype=None, model_options={}): - if llama_scaled_fp8 is not None and "llama_scaled_fp8" not in model_options: + if llama_quantization_metadata is not None: model_options = model_options.copy() - model_options["llama_scaled_fp8"] = llama_scaled_fp8 + model_options["llama_quantization_metadata"] = llama_quantization_metadata super().__init__(dtype_llama=dtype_llama, device=device, dtype=dtype, model_options=model_options) return HunyuanVideoClipModel_ diff --git a/comfy/text_encoders/lumina2.py b/comfy/text_encoders/lumina2.py index fd986e2c1..7a6cfdab2 100644 --- a/comfy/text_encoders/lumina2.py +++ b/comfy/text_encoders/lumina2.py @@ -40,7 +40,7 @@ class LuminaModel(sd1_clip.SD1ClipModel): super().__init__(device=device, dtype=dtype, name=name, clip_model=clip_model, model_options=model_options) -def te(dtype_llama=None, llama_scaled_fp8=None, model_type="gemma2_2b"): +def te(dtype_llama=None, llama_quantization_metadata=None, model_type="gemma2_2b"): if model_type == "gemma2_2b": model = Gemma2_2BModel elif model_type == "gemma3_4b": @@ -48,9 +48,9 @@ def te(dtype_llama=None, llama_scaled_fp8=None, model_type="gemma2_2b"): class LuminaTEModel_(LuminaModel): def __init__(self, device="cpu", dtype=None, model_options={}): - if llama_scaled_fp8 is not None and "scaled_fp8" not in model_options: + if llama_quantization_metadata is not None: model_options = model_options.copy() - model_options["scaled_fp8"] = llama_scaled_fp8 + model_options["quantization_metadata"] = llama_quantization_metadata if dtype_llama is not None: dtype = dtype_llama super().__init__(device=device, dtype=dtype, name=model_type, model_options=model_options, clip_model=model) diff --git a/comfy/text_encoders/omnigen2.py b/comfy/text_encoders/omnigen2.py index 1a01b2dd4..50aa4121f 100644 --- a/comfy/text_encoders/omnigen2.py +++ b/comfy/text_encoders/omnigen2.py @@ -32,12 +32,12 @@ class Omnigen2Model(sd1_clip.SD1ClipModel): super().__init__(device=device, dtype=dtype, name="qwen25_3b", clip_model=Qwen25_3BModel, model_options=model_options) -def te(dtype_llama=None, llama_scaled_fp8=None): +def te(dtype_llama=None, llama_quantization_metadata=None): class Omnigen2TEModel_(Omnigen2Model): def __init__(self, device="cpu", dtype=None, model_options={}): - if llama_scaled_fp8 is not None and "scaled_fp8" not in model_options: + if llama_quantization_metadata is not None: model_options = model_options.copy() - model_options["scaled_fp8"] = llama_scaled_fp8 + model_options["quantization_metadata"] = llama_quantization_metadata if dtype_llama is not None: dtype = dtype_llama super().__init__(device=device, dtype=dtype, model_options=model_options) diff --git a/comfy/text_encoders/ovis.py b/comfy/text_encoders/ovis.py index 81c9bd51c..5754424d2 100644 --- a/comfy/text_encoders/ovis.py +++ b/comfy/text_encoders/ovis.py @@ -55,12 +55,9 @@ class OvisTEModel(sd1_clip.SD1ClipModel): return out, pooled, {} -def te(dtype_llama=None, llama_scaled_fp8=None, llama_quantization_metadata=None): +def te(dtype_llama=None, llama_quantization_metadata=None): class OvisTEModel_(OvisTEModel): def __init__(self, device="cpu", dtype=None, model_options={}): - if llama_scaled_fp8 is not None and "scaled_fp8" not in model_options: - model_options = model_options.copy() - model_options["scaled_fp8"] = llama_scaled_fp8 if dtype_llama is not None: dtype = dtype_llama if llama_quantization_metadata is not None: diff --git a/comfy/text_encoders/pixart_t5.py b/comfy/text_encoders/pixart_t5.py index 5f383de07..e5e5f18be 100644 --- a/comfy/text_encoders/pixart_t5.py +++ b/comfy/text_encoders/pixart_t5.py @@ -30,12 +30,12 @@ class PixArtTokenizer(sd1_clip.SD1Tokenizer): def __init__(self, embedding_directory=None, tokenizer_data={}): super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, clip_name="t5xxl", tokenizer=T5XXLTokenizer) -def pixart_te(dtype_t5=None, t5xxl_scaled_fp8=None): +def pixart_te(dtype_t5=None, t5_quantization_metadata=None): class PixArtTEModel_(PixArtT5XXL): def __init__(self, device="cpu", dtype=None, model_options={}): - if t5xxl_scaled_fp8 is not None and "t5xxl_scaled_fp8" not in model_options: + if t5_quantization_metadata is not None: model_options = model_options.copy() - model_options["t5xxl_scaled_fp8"] = t5xxl_scaled_fp8 + model_options["t5xxl_quantization_metadata"] = t5_quantization_metadata if dtype is None: dtype = dtype_t5 super().__init__(device=device, dtype=dtype, model_options=model_options) diff --git a/comfy/text_encoders/qwen_image.py b/comfy/text_encoders/qwen_image.py index c0d32a6ef..5c14dec23 100644 --- a/comfy/text_encoders/qwen_image.py +++ b/comfy/text_encoders/qwen_image.py @@ -85,12 +85,12 @@ class QwenImageTEModel(sd1_clip.SD1ClipModel): return out, pooled, extra -def te(dtype_llama=None, llama_scaled_fp8=None): +def te(dtype_llama=None, llama_quantization_metadata=None): class QwenImageTEModel_(QwenImageTEModel): def __init__(self, device="cpu", dtype=None, model_options={}): - if llama_scaled_fp8 is not None and "scaled_fp8" not in model_options: + if llama_quantization_metadata is not None: model_options = model_options.copy() - model_options["scaled_fp8"] = llama_scaled_fp8 + model_options["quantization_metadata"] = llama_quantization_metadata if dtype_llama is not None: dtype = dtype_llama super().__init__(device=device, dtype=dtype, model_options=model_options) diff --git a/comfy/text_encoders/sd3_clip.py b/comfy/text_encoders/sd3_clip.py index ff5d412db..8b153c72b 100644 --- a/comfy/text_encoders/sd3_clip.py +++ b/comfy/text_encoders/sd3_clip.py @@ -6,14 +6,15 @@ import torch import os import comfy.model_management import logging +import comfy.utils class T5XXLModel(sd1_clip.SDClipModel): def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, attention_mask=False, model_options={}): textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_config_xxl.json") - t5xxl_scaled_fp8 = model_options.get("t5xxl_scaled_fp8", None) - if t5xxl_scaled_fp8 is not None: + t5xxl_quantization_metadata = model_options.get("t5xxl_quantization_metadata", None) + if t5xxl_quantization_metadata is not None: model_options = model_options.copy() - model_options["scaled_fp8"] = t5xxl_scaled_fp8 + model_options["quantization_metadata"] = t5xxl_quantization_metadata model_options = {**model_options, "model_name": "t5xxl"} super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 1, "pad": 0}, model_class=comfy.text_encoders.t5.T5, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options) @@ -25,9 +26,9 @@ def t5_xxl_detect(state_dict, prefix=""): if t5_key in state_dict: out["dtype_t5"] = state_dict[t5_key].dtype - scaled_fp8_key = "{}scaled_fp8".format(prefix) - if scaled_fp8_key in state_dict: - out["t5xxl_scaled_fp8"] = state_dict[scaled_fp8_key].dtype + quant = comfy.utils.detect_layer_quantization(state_dict, prefix) + if quant is not None: + out["t5_quantization_metadata"] = quant return out @@ -156,11 +157,11 @@ class SD3ClipModel(torch.nn.Module): else: return self.t5xxl.load_sd(sd) -def sd3_clip(clip_l=True, clip_g=True, t5=True, dtype_t5=None, t5xxl_scaled_fp8=None, t5_attention_mask=False): +def sd3_clip(clip_l=True, clip_g=True, t5=True, dtype_t5=None, t5_quantization_metadata=None, t5_attention_mask=False): class SD3ClipModel_(SD3ClipModel): def __init__(self, device="cpu", dtype=None, model_options={}): - if t5xxl_scaled_fp8 is not None and "t5xxl_scaled_fp8" not in model_options: + if t5_quantization_metadata is not None: model_options = model_options.copy() - model_options["t5xxl_scaled_fp8"] = t5xxl_scaled_fp8 + model_options["t5xxl_quantization_metadata"] = t5_quantization_metadata super().__init__(clip_l=clip_l, clip_g=clip_g, t5=t5, dtype_t5=dtype_t5, t5_attention_mask=t5_attention_mask, device=device, dtype=dtype, model_options=model_options) return SD3ClipModel_ diff --git a/comfy/text_encoders/wan.py b/comfy/text_encoders/wan.py index d50fa4b28..164a57edd 100644 --- a/comfy/text_encoders/wan.py +++ b/comfy/text_encoders/wan.py @@ -25,12 +25,12 @@ class WanT5Model(sd1_clip.SD1ClipModel): def __init__(self, device="cpu", dtype=None, model_options={}, **kwargs): super().__init__(device=device, dtype=dtype, model_options=model_options, name="umt5xxl", clip_model=UMT5XXlModel, **kwargs) -def te(dtype_t5=None, t5xxl_scaled_fp8=None): +def te(dtype_t5=None, t5_quantization_metadata=None): class WanTEModel(WanT5Model): def __init__(self, device="cpu", dtype=None, model_options={}): - if t5xxl_scaled_fp8 is not None and "scaled_fp8" not in model_options: + if t5_quantization_metadata is not None: model_options = model_options.copy() - model_options["scaled_fp8"] = t5xxl_scaled_fp8 + model_options["quantization_metadata"] = t5_quantization_metadata if dtype_t5 is not None: dtype = dtype_t5 super().__init__(device=device, dtype=dtype, model_options=model_options) diff --git a/comfy/text_encoders/z_image.py b/comfy/text_encoders/z_image.py index bb9273b20..19adde0b7 100644 --- a/comfy/text_encoders/z_image.py +++ b/comfy/text_encoders/z_image.py @@ -34,12 +34,9 @@ class ZImageTEModel(sd1_clip.SD1ClipModel): super().__init__(device=device, dtype=dtype, name="qwen3_4b", clip_model=Qwen3_4BModel, model_options=model_options) -def te(dtype_llama=None, llama_scaled_fp8=None, llama_quantization_metadata=None): +def te(dtype_llama=None, llama_quantization_metadata=None): class ZImageTEModel_(ZImageTEModel): def __init__(self, device="cpu", dtype=None, model_options={}): - if llama_scaled_fp8 is not None and "scaled_fp8" not in model_options: - model_options = model_options.copy() - model_options["scaled_fp8"] = llama_scaled_fp8 if dtype_llama is not None: dtype = dtype_llama if llama_quantization_metadata is not None: diff --git a/comfy/utils.py b/comfy/utils.py index 37485e497..89846bc95 100644 --- a/comfy/utils.py +++ b/comfy/utils.py @@ -29,6 +29,7 @@ import itertools from torch.nn.functional import interpolate from einops import rearrange from comfy.cli_args import args +import json MMAP_TORCH_FILES = args.mmap_torch_files DISABLE_MMAP = args.disable_mmap @@ -1194,3 +1195,68 @@ def unpack_latents(combined_latent, latent_shapes): else: output_tensors = combined_latent return output_tensors + +def detect_layer_quantization(state_dict, prefix): + for k in state_dict: + if k.startswith(prefix) and k.endswith(".comfy_quant"): + logging.info("Found quantization metadata version 1") + return {"mixed_ops": True} + return None + +def convert_old_quants(state_dict, model_prefix="", metadata={}): + if metadata is None: + metadata = {} + + quant_metadata = None + if "_quantization_metadata" not in metadata: + scaled_fp8_key = "{}scaled_fp8".format(model_prefix) + + if scaled_fp8_key in state_dict: + scaled_fp8_weight = state_dict[scaled_fp8_key] + scaled_fp8_dtype = scaled_fp8_weight.dtype + if scaled_fp8_dtype == torch.float32: + scaled_fp8_dtype = torch.float8_e4m3fn + + if scaled_fp8_weight.nelement() == 2: + full_precision_matrix_mult = True + else: + full_precision_matrix_mult = False + + out_sd = {} + layers = {} + for k in list(state_dict.keys()): + if not k.startswith(model_prefix): + out_sd[k] = state_dict[k] + continue + k_out = k + w = state_dict.pop(k) + layer = None + if k_out.endswith(".scale_weight"): + layer = k_out[:-len(".scale_weight")] + k_out = "{}.weight_scale".format(layer) + + if layer is not None: + layer_conf = {"format": "float8_e4m3fn"} # TODO: check if anyone did some non e4m3fn scaled checkpoints + if full_precision_matrix_mult: + layer_conf["full_precision_matrix_mult"] = full_precision_matrix_mult + layers[layer] = layer_conf + + if k_out.endswith(".scale_input"): + layer = k_out[:-len(".scale_input")] + k_out = "{}.input_scale".format(layer) + if w.item() == 1.0: + continue + + out_sd[k_out] = w + + state_dict = out_sd + quant_metadata = {"layers": layers} + else: + quant_metadata = json.loads(metadata["_quantization_metadata"]) + + if quant_metadata is not None: + layers = quant_metadata["layers"] + for k, v in layers.items(): + state_dict["{}.comfy_quant".format(k)] = torch.frombuffer(json.dumps(v).encode('utf-8'), dtype=torch.uint8) + + return state_dict, metadata diff --git a/tests-unit/comfy_quant/test_mixed_precision.py b/tests-unit/comfy_quant/test_mixed_precision.py index 63361309f..3a54941e6 100644 --- a/tests-unit/comfy_quant/test_mixed_precision.py +++ b/tests-unit/comfy_quant/test_mixed_precision.py @@ -2,6 +2,7 @@ import unittest import torch import sys import os +import json # Add comfy to path sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", "..")) @@ -15,6 +16,7 @@ if not has_gpu(): from comfy import ops from comfy.quant_ops import QuantizedTensor +import comfy.utils class SimpleModel(torch.nn.Module): @@ -94,8 +96,9 @@ class TestMixedPrecisionOps(unittest.TestCase): "layer3.weight_scale": torch.tensor(1.5, dtype=torch.float32), } + state_dict, _ = comfy.utils.convert_old_quants(state_dict, metadata={"_quantization_metadata": json.dumps({"layers": layer_quant_config})}) # Create model and load state dict (strict=False because custom loading pops keys) - model = SimpleModel(operations=ops.mixed_precision_ops(layer_quant_config)) + model = SimpleModel(operations=ops.mixed_precision_ops({})) model.load_state_dict(state_dict, strict=False) # Verify weights are wrapped in QuantizedTensor @@ -115,7 +118,8 @@ class TestMixedPrecisionOps(unittest.TestCase): # Forward pass input_tensor = torch.randn(5, 10, dtype=torch.bfloat16) - output = model(input_tensor) + with torch.inference_mode(): + output = model(input_tensor) self.assertEqual(output.shape, (5, 40)) @@ -141,7 +145,8 @@ class TestMixedPrecisionOps(unittest.TestCase): "layer3.bias": torch.randn(40, dtype=torch.bfloat16), } - model = SimpleModel(operations=ops.mixed_precision_ops(layer_quant_config)) + state_dict1, _ = comfy.utils.convert_old_quants(state_dict1, metadata={"_quantization_metadata": json.dumps({"layers": layer_quant_config})}) + model = SimpleModel(operations=ops.mixed_precision_ops({})) model.load_state_dict(state_dict1, strict=False) # Save state dict @@ -178,7 +183,8 @@ class TestMixedPrecisionOps(unittest.TestCase): "layer3.bias": torch.randn(40, dtype=torch.bfloat16), } - model = SimpleModel(operations=ops.mixed_precision_ops(layer_quant_config)) + state_dict, _ = comfy.utils.convert_old_quants(state_dict, metadata={"_quantization_metadata": json.dumps({"layers": layer_quant_config})}) + model = SimpleModel(operations=ops.mixed_precision_ops({})) model.load_state_dict(state_dict, strict=False) # Add a weight function (simulating LoRA) @@ -215,8 +221,10 @@ class TestMixedPrecisionOps(unittest.TestCase): "layer3.bias": torch.randn(40, dtype=torch.bfloat16), } + state_dict, _ = comfy.utils.convert_old_quants(state_dict, metadata={"_quantization_metadata": json.dumps({"layers": layer_quant_config})}) + # Load should raise KeyError for unknown format in QUANT_FORMAT_MIXINS - model = SimpleModel(operations=ops.mixed_precision_ops(layer_quant_config)) + model = SimpleModel(operations=ops.mixed_precision_ops({})) with self.assertRaises(KeyError): model.load_state_dict(state_dict, strict=False) From 6fd463aec958f02be79a264eafd6c8fe7e52762a Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Fri, 5 Dec 2025 12:33:16 -0800 Subject: [PATCH 231/647] Fix regression when text encoder loaded directly on GPU. (#11129) --- comfy/ops.py | 2 ++ comfy/sd.py | 44 ++++++++++++++++++++++++-------------------- 2 files changed, 26 insertions(+), 20 deletions(-) diff --git a/comfy/ops.py b/comfy/ops.py index dc06709a1..35237c9f7 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -552,6 +552,8 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec weight_scale_key = f"{prefix}weight_scale" scale = state_dict.pop(weight_scale_key, None) + if scale is not None: + scale = scale.to(device) layout_params = { 'scale': scale, 'orig_dtype': MixedPrecisionOps._compute_dtype, diff --git a/comfy/sd.py b/comfy/sd.py index 092715d79..c350322f8 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -98,7 +98,7 @@ def load_lora_for_models(model, clip, lora, strength_model, strength_clip): class CLIP: - def __init__(self, target=None, embedding_directory=None, no_init=False, tokenizer_data={}, parameters=0, model_options={}): + def __init__(self, target=None, embedding_directory=None, no_init=False, tokenizer_data={}, parameters=0, state_dict=[], model_options={}): if no_init: return params = target.params.copy() @@ -129,6 +129,27 @@ class CLIP: self.patcher.hook_mode = comfy.hooks.EnumHookMode.MinVram self.patcher.is_clip = True self.apply_hooks_to_conds = None + if len(state_dict) > 0: + if isinstance(state_dict, list): + for c in state_dict: + m, u = self.load_sd(c) + if len(m) > 0: + logging.warning("clip missing: {}".format(m)) + + if len(u) > 0: + logging.debug("clip unexpected: {}".format(u)) + else: + m, u = self.load_sd(state_dict, full_model=True) + if len(m) > 0: + m_filter = list(filter(lambda a: ".logit_scale" not in a and ".transformer.text_projection.weight" not in a, m)) + if len(m_filter) > 0: + logging.warning("clip missing: {}".format(m)) + else: + logging.debug("clip missing: {}".format(m)) + + if len(u) > 0: + logging.debug("clip unexpected {}:".format(u)) + if params['device'] == load_device: model_management.load_models_gpu([self.patcher], force_full_load=True) self.layer_idx = None @@ -1225,14 +1246,7 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip parameters += comfy.utils.calculate_parameters(c) tokenizer_data, model_options = comfy.text_encoders.long_clipl.model_options_long_clip(c, tokenizer_data, model_options) - clip = CLIP(clip_target, embedding_directory=embedding_directory, parameters=parameters, tokenizer_data=tokenizer_data, model_options=model_options) - for c in clip_data: - m, u = clip.load_sd(c) - if len(m) > 0: - logging.warning("clip missing: {}".format(m)) - - if len(u) > 0: - logging.debug("clip unexpected: {}".format(u)) + clip = CLIP(clip_target, embedding_directory=embedding_directory, parameters=parameters, tokenizer_data=tokenizer_data, state_dict=clip_data, model_options=model_options) return clip def load_gligen(ckpt_path): @@ -1362,17 +1376,7 @@ def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_c clip_sd = model_config.process_clip_state_dict(sd) if len(clip_sd) > 0: parameters = comfy.utils.calculate_parameters(clip_sd) - clip = CLIP(clip_target, embedding_directory=embedding_directory, tokenizer_data=clip_sd, parameters=parameters, model_options=te_model_options) - m, u = clip.load_sd(clip_sd, full_model=True) - if len(m) > 0: - m_filter = list(filter(lambda a: ".logit_scale" not in a and ".transformer.text_projection.weight" not in a, m)) - if len(m_filter) > 0: - logging.warning("clip missing: {}".format(m)) - else: - logging.debug("clip missing: {}".format(m)) - - if len(u) > 0: - logging.debug("clip unexpected {}:".format(u)) + clip = CLIP(clip_target, embedding_directory=embedding_directory, tokenizer_data=clip_sd, parameters=parameters, state_dict=clip_sd, model_options=te_model_options) else: logging.warning("no CLIP/text encoder weights in checkpoint, the text encoder model will not be loaded.") From 79d17ba2339aaf4f3422673b3dad24ba4dbd7552 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Jukka=20Sepp=C3=A4nen?= <40791699+kijai@users.noreply.github.com> Date: Fri, 5 Dec 2025 22:42:46 +0200 Subject: [PATCH 232/647] Context windows fixes and features (#10975) * Apply cond slice fix * Add FreeNoise * Update context_windows.py * Add option to retain condition by indexes for each window This allows for example Wan/HunyuanVideo image to video to "work" by using the initial start frame for each window, otherwise windows beyond first will be pure T2V generations. * Update context_windows.py * Allow splitting multiple conds into different windows * Add handling for audio_embed * whitespace * Allow freenoise to work on other dims, handle 4D batch timestep Refactor Freenoise function. And fix batch handling as timesteps seem to be expanded to batch size now. * Disable experimental options for now So that the Freenoise and bugfixes can be merged first --------- Co-authored-by: Jedrzej Kosinski Co-authored-by: ozbayb <17261091+ozbayb@users.noreply.github.com> --- comfy/context_windows.py | 104 ++++++++++++++++++++++---- comfy_extras/nodes_context_windows.py | 22 +++++- 2 files changed, 108 insertions(+), 18 deletions(-) diff --git a/comfy/context_windows.py b/comfy/context_windows.py index 041f380f9..5c412d1c2 100644 --- a/comfy/context_windows.py +++ b/comfy/context_windows.py @@ -51,26 +51,36 @@ class ContextHandlerABC(ABC): class IndexListContextWindow(ContextWindowABC): - def __init__(self, index_list: list[int], dim: int=0): + def __init__(self, index_list: list[int], dim: int=0, total_frames: int=0): self.index_list = index_list self.context_length = len(index_list) self.dim = dim + self.total_frames = total_frames + self.center_ratio = (min(index_list) + max(index_list)) / (2 * total_frames) - def get_tensor(self, full: torch.Tensor, device=None, dim=None) -> torch.Tensor: + def get_tensor(self, full: torch.Tensor, device=None, dim=None, retain_index_list=[]) -> torch.Tensor: if dim is None: dim = self.dim if dim == 0 and full.shape[dim] == 1: return full - idx = [slice(None)] * dim + [self.index_list] - return full[idx].to(device) + idx = tuple([slice(None)] * dim + [self.index_list]) + window = full[idx] + if retain_index_list: + idx = tuple([slice(None)] * dim + [retain_index_list]) + window[idx] = full[idx] + return window.to(device) def add_window(self, full: torch.Tensor, to_add: torch.Tensor, dim=None) -> torch.Tensor: if dim is None: dim = self.dim - idx = [slice(None)] * dim + [self.index_list] + idx = tuple([slice(None)] * dim + [self.index_list]) full[idx] += to_add return full + def get_region_index(self, num_regions: int) -> int: + region_idx = int(self.center_ratio * num_regions) + return min(max(region_idx, 0), num_regions - 1) + class IndexListCallbacks: EVALUATE_CONTEXT_WINDOWS = "evaluate_context_windows" @@ -94,7 +104,8 @@ class ContextFuseMethod: ContextResults = collections.namedtuple("ContextResults", ['window_idx', 'sub_conds_out', 'sub_conds', 'window']) class IndexListContextHandler(ContextHandlerABC): - def __init__(self, context_schedule: ContextSchedule, fuse_method: ContextFuseMethod, context_length: int=1, context_overlap: int=0, context_stride: int=1, closed_loop=False, dim=0): + def __init__(self, context_schedule: ContextSchedule, fuse_method: ContextFuseMethod, context_length: int=1, context_overlap: int=0, context_stride: int=1, + closed_loop: bool=False, dim:int=0, freenoise: bool=False, cond_retain_index_list: list[int]=[], split_conds_to_windows: bool=False): self.context_schedule = context_schedule self.fuse_method = fuse_method self.context_length = context_length @@ -103,13 +114,18 @@ class IndexListContextHandler(ContextHandlerABC): self.closed_loop = closed_loop self.dim = dim self._step = 0 + self.freenoise = freenoise + self.cond_retain_index_list = [int(x.strip()) for x in cond_retain_index_list.split(",")] if cond_retain_index_list else [] + self.split_conds_to_windows = split_conds_to_windows self.callbacks = {} def should_use_context(self, model: BaseModel, conds: list[list[dict]], x_in: torch.Tensor, timestep: torch.Tensor, model_options: dict[str]) -> bool: # for now, assume first dim is batch - should have stored on BaseModel in actual implementation if x_in.size(self.dim) > self.context_length: - logging.info(f"Using context windows {self.context_length} for {x_in.size(self.dim)} frames.") + logging.info(f"Using context windows {self.context_length} with overlap {self.context_overlap} for {x_in.size(self.dim)} frames.") + if self.cond_retain_index_list: + logging.info(f"Retaining original cond for indexes: {self.cond_retain_index_list}") return True return False @@ -123,6 +139,11 @@ class IndexListContextHandler(ContextHandlerABC): return None # reuse or resize cond items to match context requirements resized_cond = [] + # if multiple conds, split based on primary region + if self.split_conds_to_windows and len(cond_in) > 1: + region = window.get_region_index(len(cond_in)) + logging.info(f"Splitting conds to windows; using region {region} for window {window[0]}-{window[-1]} with center ratio {window.center_ratio:.3f}") + cond_in = [cond_in[region]] # cond object is a list containing a dict - outer list is irrelevant, so just loop through it for actual_cond in cond_in: resized_actual_cond = actual_cond.copy() @@ -146,12 +167,19 @@ class IndexListContextHandler(ContextHandlerABC): # when in dictionary, look for tensors and CONDCrossAttn [comfy/conds.py] (has cond attr that is a tensor) for cond_key, cond_value in new_cond_item.items(): if isinstance(cond_value, torch.Tensor): - if cond_value.ndim < self.dim and cond_value.size(0) == x_in.size(self.dim): + if (self.dim < cond_value.ndim and cond_value(self.dim) == x_in.size(self.dim)) or \ + (cond_value.ndim < self.dim and cond_value.size(0) == x_in.size(self.dim)): new_cond_item[cond_key] = window.get_tensor(cond_value, device) + # Handle audio_embed (temporal dim is 1) + elif cond_key == "audio_embed" and hasattr(cond_value, "cond") and isinstance(cond_value.cond, torch.Tensor): + audio_cond = cond_value.cond + if audio_cond.ndim > 1 and audio_cond.size(1) == x_in.size(self.dim): + new_cond_item[cond_key] = cond_value._copy_with(window.get_tensor(audio_cond, device, dim=1)) # if has cond that is a Tensor, check if needs to be subset elif hasattr(cond_value, "cond") and isinstance(cond_value.cond, torch.Tensor): - if cond_value.cond.ndim < self.dim and cond_value.cond.size(0) == x_in.size(self.dim): - new_cond_item[cond_key] = cond_value._copy_with(window.get_tensor(cond_value.cond, device)) + if (self.dim < cond_value.cond.ndim and cond_value.cond.size(self.dim) == x_in.size(self.dim)) or \ + (cond_value.cond.ndim < self.dim and cond_value.cond.size(0) == x_in.size(self.dim)): + new_cond_item[cond_key] = cond_value._copy_with(window.get_tensor(cond_value.cond, device, retain_index_list=self.cond_retain_index_list)) elif cond_key == "num_video_frames": # for SVD new_cond_item[cond_key] = cond_value._copy_with(cond_value.cond) new_cond_item[cond_key].cond = window.context_length @@ -164,7 +192,7 @@ class IndexListContextHandler(ContextHandlerABC): return resized_cond def set_step(self, timestep: torch.Tensor, model_options: dict[str]): - mask = torch.isclose(model_options["transformer_options"]["sample_sigmas"], timestep, rtol=0.0001) + mask = torch.isclose(model_options["transformer_options"]["sample_sigmas"], timestep[0], rtol=0.0001) matches = torch.nonzero(mask) if torch.numel(matches) == 0: raise Exception("No sample_sigmas matched current timestep; something went wrong.") @@ -173,7 +201,7 @@ class IndexListContextHandler(ContextHandlerABC): def get_context_windows(self, model: BaseModel, x_in: torch.Tensor, model_options: dict[str]) -> list[IndexListContextWindow]: full_length = x_in.size(self.dim) # TODO: choose dim based on model context_windows = self.context_schedule.func(full_length, self, model_options) - context_windows = [IndexListContextWindow(window, dim=self.dim) for window in context_windows] + context_windows = [IndexListContextWindow(window, dim=self.dim, total_frames=full_length) for window in context_windows] return context_windows def execute(self, calc_cond_batch: Callable, model: BaseModel, conds: list[list[dict]], x_in: torch.Tensor, timestep: torch.Tensor, model_options: dict[str]): @@ -250,8 +278,8 @@ class IndexListContextHandler(ContextHandlerABC): prev_weight = (bias_total / (bias_total + bias)) new_weight = (bias / (bias_total + bias)) # account for dims of tensors - idx_window = [slice(None)] * self.dim + [idx] - pos_window = [slice(None)] * self.dim + [pos] + idx_window = tuple([slice(None)] * self.dim + [idx]) + pos_window = tuple([slice(None)] * self.dim + [pos]) # apply new values conds_final[i][idx_window] = conds_final[i][idx_window] * prev_weight + sub_conds_out[i][pos_window] * new_weight biases_final[i][idx] = bias_total + bias @@ -287,6 +315,28 @@ def create_prepare_sampling_wrapper(model: ModelPatcher): ) +def _sampler_sample_wrapper(executor, guider, sigmas, extra_args, callback, noise, *args, **kwargs): + model_options = extra_args.get("model_options", None) + if model_options is None: + raise Exception("model_options not found in sampler_sample_wrapper; this should never happen, something went wrong.") + handler: IndexListContextHandler = model_options.get("context_handler", None) + if handler is None: + raise Exception("context_handler not found in sampler_sample_wrapper; this should never happen, something went wrong.") + if not handler.freenoise: + return executor(guider, sigmas, extra_args, callback, noise, *args, **kwargs) + noise = apply_freenoise(noise, handler.dim, handler.context_length, handler.context_overlap, extra_args["seed"]) + + return executor(guider, sigmas, extra_args, callback, noise, *args, **kwargs) + + +def create_sampler_sample_wrapper(model: ModelPatcher): + model.add_wrapper_with_key( + comfy.patcher_extension.WrappersMP.SAMPLER_SAMPLE, + "ContextWindows_sampler_sample", + _sampler_sample_wrapper + ) + + def match_weights_to_dim(weights: list[float], x_in: torch.Tensor, dim: int, device=None) -> torch.Tensor: total_dims = len(x_in.shape) weights_tensor = torch.Tensor(weights).to(device=device) @@ -538,3 +588,29 @@ def shift_window_to_end(window: list[int], num_frames: int): for i in range(len(window)): # 2) add end_delta to each val to slide windows to end window[i] = window[i] + end_delta + + +# https://github.com/Kosinkadink/ComfyUI-AnimateDiff-Evolved/blob/90fb1331201a4b29488089e4fbffc0d82cc6d0a9/animatediff/sample_settings.py#L465 +def apply_freenoise(noise: torch.Tensor, dim: int, context_length: int, context_overlap: int, seed: int): + logging.info("Context windows: Applying FreeNoise") + generator = torch.Generator(device='cpu').manual_seed(seed) + latent_video_length = noise.shape[dim] + delta = context_length - context_overlap + + for start_idx in range(0, latent_video_length - context_length, delta): + place_idx = start_idx + context_length + + actual_delta = min(delta, latent_video_length - place_idx) + if actual_delta <= 0: + break + + list_idx = torch.randperm(actual_delta, generator=generator, device='cpu') + start_idx + + source_slice = [slice(None)] * noise.ndim + source_slice[dim] = list_idx + target_slice = [slice(None)] * noise.ndim + target_slice[dim] = slice(place_idx, place_idx + actual_delta) + + noise[tuple(target_slice)] = noise[tuple(source_slice)] + + return noise diff --git a/comfy_extras/nodes_context_windows.py b/comfy_extras/nodes_context_windows.py index 1c3d9e697..3799a9004 100644 --- a/comfy_extras/nodes_context_windows.py +++ b/comfy_extras/nodes_context_windows.py @@ -26,6 +26,9 @@ class ContextWindowsManualNode(io.ComfyNode): io.Boolean.Input("closed_loop", default=False, tooltip="Whether to close the context window loop; only applicable to looped schedules."), io.Combo.Input("fuse_method", options=comfy.context_windows.ContextFuseMethods.LIST_STATIC, default=comfy.context_windows.ContextFuseMethods.PYRAMID, tooltip="The method to use to fuse the context windows."), io.Int.Input("dim", min=0, max=5, default=0, tooltip="The dimension to apply the context windows to."), + io.Boolean.Input("freenoise", default=False, tooltip="Whether to apply FreeNoise noise shuffling, improves window blending."), + #io.String.Input("cond_retain_index_list", default="", tooltip="List of latent indices to retain in the conditioning tensors for each window, for example setting this to '0' will use the initial start image for each window."), + #io.Boolean.Input("split_conds_to_windows", default=False, tooltip="Whether to split multiple conditionings (created by ConditionCombine) to each window based on region index."), ], outputs=[ io.Model.Output(tooltip="The model with context windows applied during sampling."), @@ -34,7 +37,8 @@ class ContextWindowsManualNode(io.ComfyNode): ) @classmethod - def execute(cls, model: io.Model.Type, context_length: int, context_overlap: int, context_schedule: str, context_stride: int, closed_loop: bool, fuse_method: str, dim: int) -> io.Model: + def execute(cls, model: io.Model.Type, context_length: int, context_overlap: int, context_schedule: str, context_stride: int, closed_loop: bool, fuse_method: str, dim: int, freenoise: bool, + cond_retain_index_list: list[int]=[], split_conds_to_windows: bool=False) -> io.Model: model = model.clone() model.model_options["context_handler"] = comfy.context_windows.IndexListContextHandler( context_schedule=comfy.context_windows.get_matching_context_schedule(context_schedule), @@ -43,9 +47,15 @@ class ContextWindowsManualNode(io.ComfyNode): context_overlap=context_overlap, context_stride=context_stride, closed_loop=closed_loop, - dim=dim) + dim=dim, + freenoise=freenoise, + cond_retain_index_list=cond_retain_index_list, + split_conds_to_windows=split_conds_to_windows + ) # make memory usage calculation only take into account the context window latents comfy.context_windows.create_prepare_sampling_wrapper(model) + if freenoise: # no other use for this wrapper at this time + comfy.context_windows.create_sampler_sample_wrapper(model) return io.NodeOutput(model) class WanContextWindowsManualNode(ContextWindowsManualNode): @@ -68,14 +78,18 @@ class WanContextWindowsManualNode(ContextWindowsManualNode): io.Int.Input("context_stride", min=1, default=1, tooltip="The stride of the context window; only applicable to uniform schedules."), io.Boolean.Input("closed_loop", default=False, tooltip="Whether to close the context window loop; only applicable to looped schedules."), io.Combo.Input("fuse_method", options=comfy.context_windows.ContextFuseMethods.LIST_STATIC, default=comfy.context_windows.ContextFuseMethods.PYRAMID, tooltip="The method to use to fuse the context windows."), + io.Boolean.Input("freenoise", default=False, tooltip="Whether to apply FreeNoise noise shuffling, improves window blending."), + #io.String.Input("cond_retain_index_list", default="", tooltip="List of latent indices to retain in the conditioning tensors for each window, for example setting this to '0' will use the initial start image for each window."), + #io.Boolean.Input("split_conds_to_windows", default=False, tooltip="Whether to split multiple conditionings (created by ConditionCombine) to each window based on region index."), ] return schema @classmethod - def execute(cls, model: io.Model.Type, context_length: int, context_overlap: int, context_schedule: str, context_stride: int, closed_loop: bool, fuse_method: str) -> io.Model: + def execute(cls, model: io.Model.Type, context_length: int, context_overlap: int, context_schedule: str, context_stride: int, closed_loop: bool, fuse_method: str, freenoise: bool, + cond_retain_index_list: list[int]=[], split_conds_to_windows: bool=False) -> io.Model: context_length = max(((context_length - 1) // 4) + 1, 1) # at least length 1 context_overlap = max(((context_overlap - 1) // 4) + 1, 0) # at least overlap 0 - return super().execute(model, context_length, context_overlap, context_schedule, context_stride, closed_loop, fuse_method, dim=2) + return super().execute(model, context_length, context_overlap, context_schedule, context_stride, closed_loop, fuse_method, dim=2, freenoise=freenoise, cond_retain_index_list=cond_retain_index_list, split_conds_to_windows=split_conds_to_windows) class ContextWindowsExtension(ComfyExtension): From 092ee8a5008c8d558b0a72cc7961a31d9cc5400b Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Fri, 5 Dec 2025 15:25:31 -0800 Subject: [PATCH 233/647] Fix some custom nodes. (#11134) --- comfy/supported_models_base.py | 5 +++++ 1 file changed, 5 insertions(+) diff --git a/comfy/supported_models_base.py b/comfy/supported_models_base.py index 9fd84d329..0e7a829ba 100644 --- a/comfy/supported_models_base.py +++ b/comfy/supported_models_base.py @@ -17,6 +17,7 @@ """ import torch +import logging from . import model_base from . import utils from . import latent_formats @@ -117,3 +118,7 @@ class BASE: def set_inference_dtype(self, dtype, manual_cast_dtype): self.unet_config['dtype'] = dtype self.manual_cast_dtype = manual_cast_dtype + + def __getattr__(self, name): + logging.warning("\nWARNING, you accessed {} from the model config object which doesn't exist. Please fix your code.\n".format(name)) + return None From bed12674a1d2c4bfdfbdd098686390f807996c90 Mon Sep 17 00:00:00 2001 From: "Dr.Lt.Data" <128333288+ltdrdata@users.noreply.github.com> Date: Sat, 6 Dec 2025 08:45:38 +0900 Subject: [PATCH 234/647] docs: add ComfyUI-Manager documentation and update to v4.0.3b4 (#11133) - Add manager setup instructions and command line options to README - Document --enable-manager, --enable-manager-legacy-ui, and --disable-manager-ui flags - Bump comfyui_manager version from 4.0.3b3 to 4.0.3b4 --- README.md | 26 ++++++++++++++++++++++++++ manager_requirements.txt | 2 +- 2 files changed, 27 insertions(+), 1 deletion(-) diff --git a/README.md b/README.md index ed857df9f..bae955b1b 100644 --- a/README.md +++ b/README.md @@ -320,6 +320,32 @@ For models compatible with Iluvatar Extension for PyTorch. Here's a step-by-step 1. Install the Iluvatar Corex Toolkit by adhering to the platform-specific instructions on the [Installation](https://support.iluvatar.com/#/DocumentCentre?id=1&nameCenter=2&productId=520117912052801536) 2. Launch ComfyUI by running `python main.py` + +## [ComfyUI-Manager](https://github.com/Comfy-Org/ComfyUI-Manager/tree/manager-v4) + +**ComfyUI-Manager** is an extension that allows you to easily install, update, and manage custom nodes for ComfyUI. + +### Setup + +1. Install the manager dependencies: + ```bash + pip install -r manager_requirements.txt + ``` + +2. Enable the manager with the `--enable-manager` flag when running ComfyUI: + ```bash + python main.py --enable-manager + ``` + +### Command Line Options + +| Flag | Description | +|------|-------------| +| `--enable-manager` | Enable ComfyUI-Manager | +| `--enable-manager-legacy-ui` | Use the legacy manager UI instead of the new UI (requires `--enable-manager`) | +| `--disable-manager-ui` | Disable the manager UI and endpoints while keeping background features like security checks and scheduled installation completion (requires `--enable-manager`) | + + # Running ```python main.py``` diff --git a/manager_requirements.txt b/manager_requirements.txt index 52cc5389c..b95cefb74 100644 --- a/manager_requirements.txt +++ b/manager_requirements.txt @@ -1 +1 @@ -comfyui_manager==4.0.3b3 +comfyui_manager==4.0.3b4 From fd109325db7126f92c2dfb7e6b25310eded8c1f8 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Jukka=20Sepp=C3=A4nen?= <40791699+kijai@users.noreply.github.com> Date: Sat, 6 Dec 2025 05:20:22 +0200 Subject: [PATCH 235/647] Kandinsky5 model support (#10988) * Add Kandinsky5 model support lite and pro T2V tested to work * Update kandinsky5.py * Fix fp8 * Fix fp8_scaled text encoder * Add transformer_options for attention * Code cleanup, optimizations, use fp32 for all layers originally at fp32 * ImageToVideo -node * Fix I2V, add necessary latent post process nodes * Support text to image model * Support block replace patches (SLG mostly) * Support official LoRAs * Don't scale RoPE for lite model as that just doesn't work... * Update supported_models.py * Rever RoPE scaling to simpler one * Fix typo * Handle latent dim difference for image model in the VAE instead * Add node to use different prompts for clip_l and qwen25_7b * Reduce peak VRAM usage a bit * Further reduce peak VRAM consumption by chunking ffn * Update chunking * Update memory_usage_factor * Code cleanup, don't force the fp32 layers as it has minimal effect * Allow for stronger changes with first frames normalization Default values are too weak for any meaningful changes, these should probably be exposed as advanced node options when that's available. * Add image model's own chat template, remove unused image2video template * Remove hard error in ReplaceVideoLatentFrames -node * Update kandinsky5.py * Update supported_models.py * Fix typos in prompt template They were now fixed in the original repository as well * Update ReplaceVideoLatentFrames Add tooltips Make source optional Better handle negative index * Rename NormalizeVideoLatentFrames -node For bit better clarity what it does * Fix NormalizeVideoLatentStart node out on non-op --- comfy/ldm/kandinsky5/model.py | 407 ++++++++++++++++++++++++++++++ comfy/lora.py | 7 + comfy/model_base.py | 47 ++++ comfy/model_detection.py | 18 ++ comfy/sd.py | 11 + comfy/supported_models.py | 56 +++- comfy/text_encoders/kandinsky5.py | 68 +++++ comfy_api/latest/_io.py | 2 + comfy_extras/nodes_kandinsky5.py | 136 ++++++++++ comfy_extras/nodes_latent.py | 39 ++- nodes.py | 3 +- 11 files changed, 791 insertions(+), 3 deletions(-) create mode 100644 comfy/ldm/kandinsky5/model.py create mode 100644 comfy/text_encoders/kandinsky5.py create mode 100644 comfy_extras/nodes_kandinsky5.py diff --git a/comfy/ldm/kandinsky5/model.py b/comfy/ldm/kandinsky5/model.py new file mode 100644 index 000000000..a653e02fc --- /dev/null +++ b/comfy/ldm/kandinsky5/model.py @@ -0,0 +1,407 @@ +import torch +from torch import nn +import math + +import comfy.ldm.common_dit +from comfy.ldm.modules.attention import optimized_attention +from comfy.ldm.flux.math import apply_rope1 +from comfy.ldm.flux.layers import EmbedND + +def attention(q, k, v, heads, transformer_options={}): + return optimized_attention( + q.transpose(1, 2), + k.transpose(1, 2), + v.transpose(1, 2), + heads=heads, + skip_reshape=True, + transformer_options=transformer_options + ) + +def apply_scale_shift_norm(norm, x, scale, shift): + return torch.addcmul(shift, norm(x), scale + 1.0) + +def apply_gate_sum(x, out, gate): + return torch.addcmul(x, gate, out) + +def get_shift_scale_gate(params): + shift, scale, gate = torch.chunk(params, 3, dim=-1) + return tuple(x.unsqueeze(1) for x in (shift, scale, gate)) + +def get_freqs(dim, max_period=10000.0): + return torch.exp(-math.log(max_period) * torch.arange(start=0, end=dim, dtype=torch.float32) / dim) + + +class TimeEmbeddings(nn.Module): + def __init__(self, model_dim, time_dim, max_period=10000.0, operation_settings=None): + super().__init__() + assert model_dim % 2 == 0 + self.model_dim = model_dim + self.max_period = max_period + self.register_buffer("freqs", get_freqs(model_dim // 2, max_period), persistent=False) + operations = operation_settings.get("operations") + self.in_layer = operations.Linear(model_dim, time_dim, bias=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + self.activation = nn.SiLU() + self.out_layer = operations.Linear(time_dim, time_dim, bias=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + + def forward(self, timestep, dtype): + args = torch.outer(timestep, self.freqs.to(device=timestep.device)) + time_embed = torch.cat([torch.cos(args), torch.sin(args)], dim=-1).to(dtype) + time_embed = self.out_layer(self.activation(self.in_layer(time_embed))) + return time_embed + + +class TextEmbeddings(nn.Module): + def __init__(self, text_dim, model_dim, operation_settings=None): + super().__init__() + operations = operation_settings.get("operations") + self.in_layer = operations.Linear(text_dim, model_dim, bias=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + self.norm = operations.LayerNorm(model_dim, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + + def forward(self, text_embed): + text_embed = self.in_layer(text_embed) + return self.norm(text_embed).type_as(text_embed) + + +class VisualEmbeddings(nn.Module): + def __init__(self, visual_dim, model_dim, patch_size, operation_settings=None): + super().__init__() + self.patch_size = patch_size + operations = operation_settings.get("operations") + self.in_layer = operations.Linear(visual_dim, model_dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + + def forward(self, x): + x = x.movedim(1, -1) # B C T H W -> B T H W C + B, T, H, W, dim = x.shape + pt, ph, pw = self.patch_size + + x = x.view( + B, + T // pt, pt, + H // ph, ph, + W // pw, pw, + dim, + ).permute(0, 1, 3, 5, 2, 4, 6, 7).flatten(4, 7) + + return self.in_layer(x) + + +class Modulation(nn.Module): + def __init__(self, time_dim, model_dim, num_params, operation_settings=None): + super().__init__() + self.activation = nn.SiLU() + self.out_layer = operation_settings.get("operations").Linear(time_dim, num_params * model_dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + + def forward(self, x): + return self.out_layer(self.activation(x)) + + +class SelfAttention(nn.Module): + def __init__(self, num_channels, head_dim, operation_settings=None): + super().__init__() + assert num_channels % head_dim == 0 + self.num_heads = num_channels // head_dim + self.head_dim = head_dim + + operations = operation_settings.get("operations") + self.to_query = operations.Linear(num_channels, num_channels, bias=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + self.to_key = operations.Linear(num_channels, num_channels, bias=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + self.to_value = operations.Linear(num_channels, num_channels, bias=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + self.query_norm = operations.RMSNorm(head_dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + self.key_norm = operations.RMSNorm(head_dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + + self.out_layer = operations.Linear(num_channels, num_channels, bias=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + self.num_chunks = 2 + + def _compute_qk(self, x, freqs, proj_fn, norm_fn): + result = proj_fn(x).view(*x.shape[:-1], self.num_heads, -1) + return apply_rope1(norm_fn(result), freqs) + + def _forward(self, x, freqs, transformer_options={}): + q = self._compute_qk(x, freqs, self.to_query, self.query_norm) + k = self._compute_qk(x, freqs, self.to_key, self.key_norm) + v = self.to_value(x).view(*x.shape[:-1], self.num_heads, -1) + out = attention(q, k, v, self.num_heads, transformer_options=transformer_options) + return self.out_layer(out) + + def _forward_chunked(self, x, freqs, transformer_options={}): + def process_chunks(proj_fn, norm_fn): + x_chunks = torch.chunk(x, self.num_chunks, dim=1) + freqs_chunks = torch.chunk(freqs, self.num_chunks, dim=1) + chunks = [] + for x_chunk, freqs_chunk in zip(x_chunks, freqs_chunks): + chunks.append(self._compute_qk(x_chunk, freqs_chunk, proj_fn, norm_fn)) + return torch.cat(chunks, dim=1) + + q = process_chunks(self.to_query, self.query_norm) + k = process_chunks(self.to_key, self.key_norm) + v = self.to_value(x).view(*x.shape[:-1], self.num_heads, -1) + out = attention(q, k, v, self.num_heads, transformer_options=transformer_options) + return self.out_layer(out) + + def forward(self, x, freqs, transformer_options={}): + if x.shape[1] > 8192: + return self._forward_chunked(x, freqs, transformer_options=transformer_options) + else: + return self._forward(x, freqs, transformer_options=transformer_options) + + +class CrossAttention(SelfAttention): + def get_qkv(self, x, context): + q = self.to_query(x).view(*x.shape[:-1], self.num_heads, -1) + k = self.to_key(context).view(*context.shape[:-1], self.num_heads, -1) + v = self.to_value(context).view(*context.shape[:-1], self.num_heads, -1) + return q, k, v + + def forward(self, x, context, transformer_options={}): + q, k, v = self.get_qkv(x, context) + out = attention(self.query_norm(q), self.key_norm(k), v, self.num_heads, transformer_options=transformer_options) + return self.out_layer(out) + + +class FeedForward(nn.Module): + def __init__(self, dim, ff_dim, operation_settings=None): + super().__init__() + operations = operation_settings.get("operations") + self.in_layer = operations.Linear(dim, ff_dim, bias=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + self.activation = nn.GELU() + self.out_layer = operations.Linear(ff_dim, dim, bias=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + self.num_chunks = 4 + + def _forward(self, x): + return self.out_layer(self.activation(self.in_layer(x))) + + def _forward_chunked(self, x): + chunks = torch.chunk(x, self.num_chunks, dim=1) + output_chunks = [] + for chunk in chunks: + output_chunks.append(self._forward(chunk)) + return torch.cat(output_chunks, dim=1) + + def forward(self, x): + if x.shape[1] > 8192: + return self._forward_chunked(x) + else: + return self._forward(x) + + +class OutLayer(nn.Module): + def __init__(self, model_dim, time_dim, visual_dim, patch_size, operation_settings=None): + super().__init__() + self.patch_size = patch_size + self.modulation = Modulation(time_dim, model_dim, 2, operation_settings=operation_settings) + operations = operation_settings.get("operations") + self.norm = operations.LayerNorm(model_dim, elementwise_affine=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + self.out_layer = operations.Linear(model_dim, math.prod(patch_size) * visual_dim, bias=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + + def forward(self, visual_embed, time_embed): + B, T, H, W, _ = visual_embed.shape + shift, scale = torch.chunk(self.modulation(time_embed), 2, dim=-1) + scale = scale[:, None, None, None, :] + shift = shift[:, None, None, None, :] + visual_embed = apply_scale_shift_norm(self.norm, visual_embed, scale, shift) + x = self.out_layer(visual_embed) + + out_dim = x.shape[-1] // (self.patch_size[0] * self.patch_size[1] * self.patch_size[2]) + x = x.view( + B, T, H, W, + out_dim, + self.patch_size[0], self.patch_size[1], self.patch_size[2] + ) + return x.permute(0, 4, 1, 5, 2, 6, 3, 7).flatten(2, 3).flatten(3, 4).flatten(4, 5) + + +class TransformerEncoderBlock(nn.Module): + def __init__(self, model_dim, time_dim, ff_dim, head_dim, operation_settings=None): + super().__init__() + self.text_modulation = Modulation(time_dim, model_dim, 6, operation_settings=operation_settings) + operations = operation_settings.get("operations") + + self.self_attention_norm = operations.LayerNorm(model_dim, elementwise_affine=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + self.self_attention = SelfAttention(model_dim, head_dim, operation_settings=operation_settings) + + self.feed_forward_norm = operations.LayerNorm(model_dim, elementwise_affine=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + self.feed_forward = FeedForward(model_dim, ff_dim, operation_settings=operation_settings) + + def forward(self, x, time_embed, freqs, transformer_options={}): + self_attn_params, ff_params = torch.chunk(self.text_modulation(time_embed), 2, dim=-1) + shift, scale, gate = get_shift_scale_gate(self_attn_params) + out = apply_scale_shift_norm(self.self_attention_norm, x, scale, shift) + out = self.self_attention(out, freqs, transformer_options=transformer_options) + x = apply_gate_sum(x, out, gate) + + shift, scale, gate = get_shift_scale_gate(ff_params) + out = apply_scale_shift_norm(self.feed_forward_norm, x, scale, shift) + out = self.feed_forward(out) + x = apply_gate_sum(x, out, gate) + return x + + +class TransformerDecoderBlock(nn.Module): + def __init__(self, model_dim, time_dim, ff_dim, head_dim, operation_settings=None): + super().__init__() + self.visual_modulation = Modulation(time_dim, model_dim, 9, operation_settings=operation_settings) + + operations = operation_settings.get("operations") + self.self_attention_norm = operations.LayerNorm(model_dim, elementwise_affine=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + self.self_attention = SelfAttention(model_dim, head_dim, operation_settings=operation_settings) + + self.cross_attention_norm = operations.LayerNorm(model_dim, elementwise_affine=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + self.cross_attention = CrossAttention(model_dim, head_dim, operation_settings=operation_settings) + + self.feed_forward_norm = operations.LayerNorm(model_dim, elementwise_affine=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + self.feed_forward = FeedForward(model_dim, ff_dim, operation_settings=operation_settings) + + def forward(self, visual_embed, text_embed, time_embed, freqs, transformer_options={}): + self_attn_params, cross_attn_params, ff_params = torch.chunk(self.visual_modulation(time_embed), 3, dim=-1) + # self attention + shift, scale, gate = get_shift_scale_gate(self_attn_params) + visual_out = apply_scale_shift_norm(self.self_attention_norm, visual_embed, scale, shift) + visual_out = self.self_attention(visual_out, freqs, transformer_options=transformer_options) + visual_embed = apply_gate_sum(visual_embed, visual_out, gate) + # cross attention + shift, scale, gate = get_shift_scale_gate(cross_attn_params) + visual_out = apply_scale_shift_norm(self.cross_attention_norm, visual_embed, scale, shift) + visual_out = self.cross_attention(visual_out, text_embed, transformer_options=transformer_options) + visual_embed = apply_gate_sum(visual_embed, visual_out, gate) + # feed forward + shift, scale, gate = get_shift_scale_gate(ff_params) + visual_out = apply_scale_shift_norm(self.feed_forward_norm, visual_embed, scale, shift) + visual_out = self.feed_forward(visual_out) + visual_embed = apply_gate_sum(visual_embed, visual_out, gate) + return visual_embed + + +class Kandinsky5(nn.Module): + def __init__( + self, + in_visual_dim=16, out_visual_dim=16, in_text_dim=3584, in_text_dim2=768, time_dim=512, + model_dim=1792, ff_dim=7168, visual_embed_dim=132, patch_size=(1, 2, 2), num_text_blocks=2, num_visual_blocks=32, + axes_dims=(16, 24, 24), rope_scale_factor=(1.0, 2.0, 2.0), + dtype=None, device=None, operations=None, **kwargs + ): + super().__init__() + head_dim = sum(axes_dims) + self.rope_scale_factor = rope_scale_factor + self.in_visual_dim = in_visual_dim + self.model_dim = model_dim + self.patch_size = patch_size + self.visual_embed_dim = visual_embed_dim + self.dtype = dtype + self.device = device + operation_settings = {"operations": operations, "device": device, "dtype": dtype} + + self.time_embeddings = TimeEmbeddings(model_dim, time_dim, operation_settings=operation_settings) + self.text_embeddings = TextEmbeddings(in_text_dim, model_dim, operation_settings=operation_settings) + self.pooled_text_embeddings = TextEmbeddings(in_text_dim2, time_dim, operation_settings=operation_settings) + self.visual_embeddings = VisualEmbeddings(visual_embed_dim, model_dim, patch_size, operation_settings=operation_settings) + + self.text_transformer_blocks = nn.ModuleList( + [TransformerEncoderBlock(model_dim, time_dim, ff_dim, head_dim, operation_settings=operation_settings) for _ in range(num_text_blocks)] + ) + + self.visual_transformer_blocks = nn.ModuleList( + [TransformerDecoderBlock(model_dim, time_dim, ff_dim, head_dim, operation_settings=operation_settings) for _ in range(num_visual_blocks)] + ) + + self.out_layer = OutLayer(model_dim, time_dim, out_visual_dim, patch_size, operation_settings=operation_settings) + + self.rope_embedder_3d = EmbedND(dim=head_dim, theta=10000.0, axes_dim=axes_dims) + self.rope_embedder_1d = EmbedND(dim=head_dim, theta=10000.0, axes_dim=[head_dim]) + + def rope_encode_1d(self, seq_len, seq_start=0, steps=None, device=None, dtype=None, transformer_options={}): + steps = seq_len if steps is None else steps + seq_ids = torch.linspace(seq_start, seq_start + (seq_len - 1), steps=steps, device=device, dtype=dtype) + seq_ids = seq_ids.reshape(-1, 1).unsqueeze(0) # Shape: (1, steps, 1) + freqs = self.rope_embedder_1d(seq_ids).movedim(1, 2) + return freqs + + def rope_encode_3d(self, t, h, w, t_start=0, steps_t=None, steps_h=None, steps_w=None, device=None, dtype=None, transformer_options={}): + + patch_size = self.patch_size + t_len = ((t + (patch_size[0] // 2)) // patch_size[0]) + h_len = ((h + (patch_size[1] // 2)) // patch_size[1]) + w_len = ((w + (patch_size[2] // 2)) // patch_size[2]) + + if steps_t is None: + steps_t = t_len + if steps_h is None: + steps_h = h_len + if steps_w is None: + steps_w = w_len + + h_start = 0 + w_start = 0 + rope_options = transformer_options.get("rope_options", None) + if rope_options is not None: + t_len = (t_len - 1.0) * rope_options.get("scale_t", 1.0) + 1.0 + h_len = (h_len - 1.0) * rope_options.get("scale_y", 1.0) + 1.0 + w_len = (w_len - 1.0) * rope_options.get("scale_x", 1.0) + 1.0 + + t_start += rope_options.get("shift_t", 0.0) + h_start += rope_options.get("shift_y", 0.0) + w_start += rope_options.get("shift_x", 0.0) + else: + rope_scale_factor = self.rope_scale_factor + if self.model_dim == 4096: # pro video model uses different rope scaling at higher resolutions + if h * w >= 14080: + rope_scale_factor = (1.0, 3.16, 3.16) + + t_len = (t_len - 1.0) / rope_scale_factor[0] + 1.0 + h_len = (h_len - 1.0) / rope_scale_factor[1] + 1.0 + w_len = (w_len - 1.0) / rope_scale_factor[2] + 1.0 + + img_ids = torch.zeros((steps_t, steps_h, steps_w, 3), device=device, dtype=dtype) + img_ids[:, :, :, 0] = img_ids[:, :, :, 0] + torch.linspace(t_start, t_start + (t_len - 1), steps=steps_t, device=device, dtype=dtype).reshape(-1, 1, 1) + img_ids[:, :, :, 1] = img_ids[:, :, :, 1] + torch.linspace(h_start, h_start + (h_len - 1), steps=steps_h, device=device, dtype=dtype).reshape(1, -1, 1) + img_ids[:, :, :, 2] = img_ids[:, :, :, 2] + torch.linspace(w_start, w_start + (w_len - 1), steps=steps_w, device=device, dtype=dtype).reshape(1, 1, -1) + img_ids = img_ids.reshape(1, -1, img_ids.shape[-1]) + + freqs = self.rope_embedder_3d(img_ids).movedim(1, 2) + return freqs + + def forward_orig(self, x, timestep, context, y, freqs, freqs_text, transformer_options={}, **kwargs): + patches_replace = transformer_options.get("patches_replace", {}) + context = self.text_embeddings(context) + time_embed = self.time_embeddings(timestep, x.dtype) + self.pooled_text_embeddings(y) + + for block in self.text_transformer_blocks: + context = block(context, time_embed, freqs_text, transformer_options=transformer_options) + + visual_embed = self.visual_embeddings(x) + visual_shape = visual_embed.shape[:-1] + visual_embed = visual_embed.flatten(1, -2) + + blocks_replace = patches_replace.get("dit", {}) + transformer_options["total_blocks"] = len(self.visual_transformer_blocks) + transformer_options["block_type"] = "double" + for i, block in enumerate(self.visual_transformer_blocks): + transformer_options["block_index"] = i + if ("double_block", i) in blocks_replace: + def block_wrap(args): + return block(x=args["x"], context=args["context"], time_embed=args["time_embed"], freqs=args["freqs"], transformer_options=args.get("transformer_options")) + visual_embed = blocks_replace[("double_block", i)]({"x": visual_embed, "context": context, "time_embed": time_embed, "freqs": freqs, "transformer_options": transformer_options}, {"original_block": block_wrap})["x"] + else: + visual_embed = block(visual_embed, context, time_embed, freqs=freqs, transformer_options=transformer_options) + + visual_embed = visual_embed.reshape(*visual_shape, -1) + return self.out_layer(visual_embed, time_embed) + + def _forward(self, x, timestep, context, y, time_dim_replace=None, transformer_options={}, **kwargs): + bs, c, t_len, h, w = x.shape + x = comfy.ldm.common_dit.pad_to_patch_size(x, self.patch_size) + + if time_dim_replace is not None: + time_dim_replace = comfy.ldm.common_dit.pad_to_patch_size(time_dim_replace, self.patch_size) + x[:, :time_dim_replace.shape[1], :time_dim_replace.shape[2]] = time_dim_replace + + freqs = self.rope_encode_3d(t_len, h, w, device=x.device, dtype=x.dtype, transformer_options=transformer_options) + freqs_text = self.rope_encode_1d(context.shape[1], device=x.device, dtype=x.dtype, transformer_options=transformer_options) + + return self.forward_orig(x, timestep, context, y, freqs, freqs_text, transformer_options=transformer_options, **kwargs) + + def forward(self, x, timestep, context, y, time_dim_replace=None, transformer_options={}, **kwargs): + return comfy.patcher_extension.WrapperExecutor.new_class_executor( + self._forward, + self, + comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options) + ).execute(x, timestep, context, y, time_dim_replace=time_dim_replace, transformer_options=transformer_options, **kwargs) diff --git a/comfy/lora.py b/comfy/lora.py index 3a9077869..e7202ce97 100644 --- a/comfy/lora.py +++ b/comfy/lora.py @@ -322,6 +322,13 @@ def model_lora_keys_unet(model, key_map={}): key_map["diffusion_model.{}".format(key_lora)] = to key_map["lycoris_{}".format(key_lora.replace(".", "_"))] = to + if isinstance(model, comfy.model_base.Kandinsky5): + for k in sdk: + if k.startswith("diffusion_model.") and k.endswith(".weight"): + key_lora = k[len("diffusion_model."):-len(".weight")] + key_map["{}".format(key_lora)] = k + key_map["transformer.{}".format(key_lora)] = k + return key_map diff --git a/comfy/model_base.py b/comfy/model_base.py index 3cedd4f31..0be006cc2 100644 --- a/comfy/model_base.py +++ b/comfy/model_base.py @@ -47,6 +47,7 @@ import comfy.ldm.chroma_radiance.model import comfy.ldm.ace.model import comfy.ldm.omnigen.omnigen2 import comfy.ldm.qwen_image.model +import comfy.ldm.kandinsky5.model import comfy.model_management import comfy.patcher_extension @@ -1630,3 +1631,49 @@ class HunyuanVideo15_SR_Distilled(HunyuanVideo15): out = super().extra_conds(**kwargs) out['disable_time_r'] = comfy.conds.CONDConstant(False) return out + +class Kandinsky5(BaseModel): + def __init__(self, model_config, model_type=ModelType.FLOW, device=None): + super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.kandinsky5.model.Kandinsky5) + + def encode_adm(self, **kwargs): + return kwargs["pooled_output"] + + def concat_cond(self, **kwargs): + noise = kwargs.get("noise", None) + device = kwargs["device"] + image = torch.zeros_like(noise) + + mask = kwargs.get("concat_mask", kwargs.get("denoise_mask", None)) + if mask is None: + mask = torch.zeros_like(noise)[:, :1] + else: + mask = 1.0 - mask + mask = utils.common_upscale(mask.to(device), noise.shape[-1], noise.shape[-2], "bilinear", "center") + if mask.shape[-3] < noise.shape[-3]: + mask = torch.nn.functional.pad(mask, (0, 0, 0, 0, 0, noise.shape[-3] - mask.shape[-3]), mode='constant', value=0) + mask = utils.resize_to_batch_size(mask, noise.shape[0]) + + return torch.cat((image, mask), dim=1) + + def extra_conds(self, **kwargs): + out = super().extra_conds(**kwargs) + attention_mask = kwargs.get("attention_mask", None) + if attention_mask is not None: + out['attention_mask'] = comfy.conds.CONDRegular(attention_mask) + cross_attn = kwargs.get("cross_attn", None) + if cross_attn is not None: + out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn) + + time_dim_replace = kwargs.get("time_dim_replace", None) + if time_dim_replace is not None: + out['time_dim_replace'] = comfy.conds.CONDRegular(self.process_latent_in(time_dim_replace)) + + return out + +class Kandinsky5Image(Kandinsky5): + def __init__(self, model_config, model_type=ModelType.FLOW, device=None): + super().__init__(model_config, model_type, device=device) + + def concat_cond(self, **kwargs): + return None diff --git a/comfy/model_detection.py b/comfy/model_detection.py index fd1907627..30b33a486 100644 --- a/comfy/model_detection.py +++ b/comfy/model_detection.py @@ -611,6 +611,24 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): dit_config["num_layers"] = count_blocks(state_dict_keys, '{}transformer_blocks.'.format(key_prefix) + '{}.') return dit_config + if '{}visual_transformer_blocks.0.cross_attention.key_norm.weight'.format(key_prefix) in state_dict_keys: # Kandinsky 5 + dit_config = {} + model_dim = state_dict['{}visual_embeddings.in_layer.bias'.format(key_prefix)].shape[0] + dit_config["model_dim"] = model_dim + if model_dim in [4096, 2560]: # pro video and lite image + dit_config["axes_dims"] = (32, 48, 48) + if model_dim == 2560: # lite image + dit_config["rope_scale_factor"] = (1.0, 1.0, 1.0) + elif model_dim == 1792: # lite video + dit_config["axes_dims"] = (16, 24, 24) + dit_config["time_dim"] = state_dict['{}time_embeddings.in_layer.bias'.format(key_prefix)].shape[0] + dit_config["image_model"] = "kandinsky5" + dit_config["ff_dim"] = state_dict['{}visual_transformer_blocks.0.feed_forward.in_layer.weight'.format(key_prefix)].shape[0] + dit_config["visual_embed_dim"] = state_dict['{}visual_embeddings.in_layer.weight'.format(key_prefix)].shape[1] + dit_config["num_text_blocks"] = count_blocks(state_dict_keys, '{}text_transformer_blocks.'.format(key_prefix) + '{}.') + dit_config["num_visual_blocks"] = count_blocks(state_dict_keys, '{}visual_transformer_blocks.'.format(key_prefix) + '{}.') + return dit_config + if '{}input_blocks.0.0.weight'.format(key_prefix) not in state_dict_keys: return None diff --git a/comfy/sd.py b/comfy/sd.py index c350322f8..754b1703d 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -54,6 +54,7 @@ import comfy.text_encoders.qwen_image import comfy.text_encoders.hunyuan_image import comfy.text_encoders.z_image import comfy.text_encoders.ovis +import comfy.text_encoders.kandinsky5 import comfy.model_patcher import comfy.lora @@ -766,6 +767,8 @@ class VAE: self.throw_exception_if_invalid() pixel_samples = None do_tile = False + if self.latent_dim == 2 and samples_in.ndim == 5: + samples_in = samples_in[:, :, 0] try: memory_used = self.memory_used_decode(samples_in.shape, self.vae_dtype) model_management.load_models_gpu([self.patcher], memory_required=memory_used, force_full_load=self.disable_offload) @@ -983,6 +986,8 @@ class CLIPType(Enum): HUNYUAN_IMAGE = 19 HUNYUAN_VIDEO_15 = 20 OVIS = 21 + KANDINSKY5 = 22 + KANDINSKY5_IMAGE = 23 def load_clip(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DIFFUSION, model_options={}): @@ -1231,6 +1236,12 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip elif clip_type == CLIPType.HUNYUAN_VIDEO_15: clip_target.clip = comfy.text_encoders.hunyuan_image.te(**llama_detect(clip_data)) clip_target.tokenizer = comfy.text_encoders.hunyuan_video.HunyuanVideo15Tokenizer + elif clip_type == CLIPType.KANDINSKY5: + clip_target.clip = comfy.text_encoders.kandinsky5.te(**llama_detect(clip_data)) + clip_target.tokenizer = comfy.text_encoders.kandinsky5.Kandinsky5Tokenizer + elif clip_type == CLIPType.KANDINSKY5_IMAGE: + clip_target.clip = comfy.text_encoders.kandinsky5.te(**llama_detect(clip_data)) + clip_target.tokenizer = comfy.text_encoders.kandinsky5.Kandinsky5TokenizerImage else: clip_target.clip = sdxl_clip.SDXLClipModel clip_target.tokenizer = sdxl_clip.SDXLTokenizer diff --git a/comfy/supported_models.py b/comfy/supported_models.py index afd97160b..91cc4ef08 100644 --- a/comfy/supported_models.py +++ b/comfy/supported_models.py @@ -21,6 +21,7 @@ import comfy.text_encoders.ace import comfy.text_encoders.omnigen2 import comfy.text_encoders.qwen_image import comfy.text_encoders.hunyuan_image +import comfy.text_encoders.kandinsky5 import comfy.text_encoders.z_image from . import supported_models_base @@ -1474,7 +1475,60 @@ class HunyuanVideo15_SR_Distilled(HunyuanVideo): hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen25_7b.transformer.".format(pref)) return supported_models_base.ClipTarget(comfy.text_encoders.hunyuan_video.HunyuanVideo15Tokenizer, comfy.text_encoders.hunyuan_image.te(**hunyuan_detect)) -models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanVideo15_SR_Distilled, HunyuanVideo15, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, ZImage, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, Omnigen2, QwenImage, Flux2] +class Kandinsky5(supported_models_base.BASE): + unet_config = { + "image_model": "kandinsky5", + } + + sampling_settings = { + "shift": 10.0, + } + + unet_extra_config = {} + latent_format = latent_formats.HunyuanVideo + + memory_usage_factor = 1.1 #TODO + + supported_inference_dtypes = [torch.bfloat16, torch.float32] + + vae_key_prefix = ["vae."] + text_encoder_key_prefix = ["text_encoders."] + + def get_model(self, state_dict, prefix="", device=None): + out = model_base.Kandinsky5(self, device=device) + return out + + def clip_target(self, state_dict={}): + pref = self.text_encoder_key_prefix[0] + hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen25_7b.transformer.".format(pref)) + return supported_models_base.ClipTarget(comfy.text_encoders.kandinsky5.Kandinsky5Tokenizer, comfy.text_encoders.kandinsky5.te(**hunyuan_detect)) + + +class Kandinsky5Image(Kandinsky5): + unet_config = { + "image_model": "kandinsky5", + "model_dim": 2560, + "visual_embed_dim": 64, + } + + sampling_settings = { + "shift": 3.0, + } + + latent_format = latent_formats.Flux + memory_usage_factor = 1.1 #TODO + + def get_model(self, state_dict, prefix="", device=None): + out = model_base.Kandinsky5Image(self, device=device) + return out + + def clip_target(self, state_dict={}): + pref = self.text_encoder_key_prefix[0] + hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen25_7b.transformer.".format(pref)) + return supported_models_base.ClipTarget(comfy.text_encoders.kandinsky5.Kandinsky5TokenizerImage, comfy.text_encoders.kandinsky5.te(**hunyuan_detect)) + + +models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanVideo15_SR_Distilled, HunyuanVideo15, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, ZImage, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, Omnigen2, QwenImage, Kandinsky5Image, Kandinsky5] models += [SVD_img2vid] diff --git a/comfy/text_encoders/kandinsky5.py b/comfy/text_encoders/kandinsky5.py new file mode 100644 index 000000000..22f991c36 --- /dev/null +++ b/comfy/text_encoders/kandinsky5.py @@ -0,0 +1,68 @@ +from comfy import sd1_clip +from .qwen_image import QwenImageTokenizer, QwenImageTEModel +from .llama import Qwen25_7BVLI + + +class Kandinsky5Tokenizer(QwenImageTokenizer): + def __init__(self, embedding_directory=None, tokenizer_data={}): + super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data) + self.llama_template = "<|im_start|>system\nYou are a prompt engineer. Describe the video in detail.\nDescribe how the camera moves or shakes, describe the zoom and view angle, whether it follows the objects.\nDescribe the location of the video, main characters or objects and their action.\nDescribe the dynamism of the video and presented actions.\nName the visual style of the video: whether it is a professional footage, user generated content, some kind of animation, video game or screen content.\nDescribe the visual effects, postprocessing and transitions if they are presented in the video.\nPay attention to the order of key actions shown in the scene.<|im_end|>\n<|im_start|>user\n{}<|im_end|>" + self.clip_l = sd1_clip.SDTokenizer(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data) + + def tokenize_with_weights(self, text:str, return_word_ids=False, **kwargs): + out = super().tokenize_with_weights(text, return_word_ids, **kwargs) + out["l"] = self.clip_l.tokenize_with_weights(text, return_word_ids, **kwargs) + + return out + + +class Kandinsky5TokenizerImage(Kandinsky5Tokenizer): + def __init__(self, embedding_directory=None, tokenizer_data={}): + super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data) + self.llama_template = "<|im_start|>system\nYou are a promt engineer. Describe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background:<|im_end|>\n<|im_start|>user\n{}<|im_end|>" + + +class Qwen25_7BVLIModel(sd1_clip.SDClipModel): + def __init__(self, device="cpu", layer="hidden", layer_idx=-1, dtype=None, attention_mask=True, model_options={}): + llama_scaled_fp8 = model_options.get("qwen_scaled_fp8", None) + if llama_scaled_fp8 is not None: + model_options = model_options.copy() + model_options["scaled_fp8"] = llama_scaled_fp8 + super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"pad": 151643}, layer_norm_hidden_state=False, model_class=Qwen25_7BVLI, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options) + + +class Kandinsky5TEModel(QwenImageTEModel): + def __init__(self, device="cpu", dtype=None, model_options={}): + super(QwenImageTEModel, self).__init__(device=device, dtype=dtype, name="qwen25_7b", clip_model=Qwen25_7BVLIModel, model_options=model_options) + self.clip_l = sd1_clip.SDClipModel(device=device, dtype=dtype, return_projected_pooled=False, model_options=model_options) + + def encode_token_weights(self, token_weight_pairs): + cond, p, extra = super().encode_token_weights(token_weight_pairs, template_end=-1) + l_out, l_pooled = self.clip_l.encode_token_weights(token_weight_pairs["l"]) + + return cond, l_pooled, extra + + def set_clip_options(self, options): + super().set_clip_options(options) + self.clip_l.set_clip_options(options) + + def reset_clip_options(self): + super().reset_clip_options() + self.clip_l.reset_clip_options() + + def load_sd(self, sd): + if "text_model.encoder.layers.1.mlp.fc1.weight" in sd: + return self.clip_l.load_sd(sd) + else: + return super().load_sd(sd) + +def te(dtype_llama=None, llama_scaled_fp8=None): + class Kandinsky5TEModel_(Kandinsky5TEModel): + def __init__(self, device="cpu", dtype=None, model_options={}): + if llama_scaled_fp8 is not None and "scaled_fp8" not in model_options: + model_options = model_options.copy() + model_options["qwen_scaled_fp8"] = llama_scaled_fp8 + if dtype_llama is not None: + dtype = dtype_llama + super().__init__(device=device, dtype=dtype, model_options=model_options) + return Kandinsky5TEModel_ diff --git a/comfy_api/latest/_io.py b/comfy_api/latest/_io.py index 866c3e0eb..d7cbe68cf 100644 --- a/comfy_api/latest/_io.py +++ b/comfy_api/latest/_io.py @@ -568,6 +568,8 @@ class Conditioning(ComfyTypeIO): '''Used by WAN Camera.''' time_dim_concat: NotRequired[torch.Tensor] '''Used by WAN Phantom Subject.''' + time_dim_replace: NotRequired[torch.Tensor] + '''Used by Kandinsky5 I2V.''' CondList = list[tuple[torch.Tensor, PooledDict]] Type = CondList diff --git a/comfy_extras/nodes_kandinsky5.py b/comfy_extras/nodes_kandinsky5.py new file mode 100644 index 000000000..9cb234be1 --- /dev/null +++ b/comfy_extras/nodes_kandinsky5.py @@ -0,0 +1,136 @@ +import nodes +import node_helpers +import torch +import comfy.model_management +import comfy.utils + +from typing_extensions import override +from comfy_api.latest import ComfyExtension, io + + +class Kandinsky5ImageToVideo(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="Kandinsky5ImageToVideo", + category="conditioning/video_models", + inputs=[ + io.Conditioning.Input("positive"), + io.Conditioning.Input("negative"), + io.Vae.Input("vae"), + io.Int.Input("width", default=768, min=16, max=nodes.MAX_RESOLUTION, step=16), + io.Int.Input("height", default=512, min=16, max=nodes.MAX_RESOLUTION, step=16), + io.Int.Input("length", default=121, min=1, max=nodes.MAX_RESOLUTION, step=4), + io.Int.Input("batch_size", default=1, min=1, max=4096), + io.Image.Input("start_image", optional=True), + ], + outputs=[ + io.Conditioning.Output(display_name="positive"), + io.Conditioning.Output(display_name="negative"), + io.Latent.Output(display_name="latent", tooltip="Empty video latent"), + io.Latent.Output(display_name="cond_latent", tooltip="Clean encoded start images, used to replace the noisy start of the model output latents"), + ], + ) + + @classmethod + def execute(cls, positive, negative, vae, width, height, length, batch_size, start_image=None) -> io.NodeOutput: + latent = torch.zeros([batch_size, 16, ((length - 1) // 4) + 1, height // 8, width // 8], device=comfy.model_management.intermediate_device()) + cond_latent_out = {} + if start_image is not None: + start_image = comfy.utils.common_upscale(start_image[:length].movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1) + encoded = vae.encode(start_image[:, :, :, :3]) + cond_latent_out["samples"] = encoded + + mask = torch.ones((1, 1, latent.shape[2], latent.shape[-2], latent.shape[-1]), device=start_image.device, dtype=start_image.dtype) + mask[:, :, :((start_image.shape[0] - 1) // 4) + 1] = 0.0 + + positive = node_helpers.conditioning_set_values(positive, {"time_dim_replace": encoded, "concat_mask": mask}) + negative = node_helpers.conditioning_set_values(negative, {"time_dim_replace": encoded, "concat_mask": mask}) + + out_latent = {} + out_latent["samples"] = latent + return io.NodeOutput(positive, negative, out_latent, cond_latent_out) + + +def adaptive_mean_std_normalization(source, reference, clump_mean_low=0.3, clump_mean_high=0.35, clump_std_low=0.35, clump_std_high=0.5): + source_mean = source.mean(dim=(1, 3, 4), keepdim=True) # mean over C, H, W + source_std = source.std(dim=(1, 3, 4), keepdim=True) # std over C, H, W + + reference_mean = torch.clamp(reference.mean(), source_mean - clump_mean_low, source_mean + clump_mean_high) + reference_std = torch.clamp(reference.std(), source_std - clump_std_low, source_std + clump_std_high) + + # normalization + normalized = (source - source_mean) / (source_std + 1e-8) + normalized = normalized * reference_std + reference_mean + + return normalized + + +class NormalizeVideoLatentStart(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="NormalizeVideoLatentStart", + category="conditioning/video_models", + description="Normalizes the initial frames of a video latent to match the mean and standard deviation of subsequent reference frames. Helps reduce differences between the starting frames and the rest of the video.", + inputs=[ + io.Latent.Input("latent"), + io.Int.Input("start_frame_count", default=4, min=1, max=nodes.MAX_RESOLUTION, step=1, tooltip="Number of latent frames to normalize, counted from the start"), + io.Int.Input("reference_frame_count", default=5, min=1, max=nodes.MAX_RESOLUTION, step=1, tooltip="Number of latent frames after the start frames to use as reference"), + ], + outputs=[ + io.Latent.Output(display_name="latent"), + ], + ) + + @classmethod + def execute(cls, latent, start_frame_count, reference_frame_count) -> io.NodeOutput: + if latent["samples"].shape[2] <= 1: + return io.NodeOutput(latent) + s = latent.copy() + samples = latent["samples"].clone() + + first_frames = samples[:, :, :start_frame_count] + reference_frames_data = samples[:, :, start_frame_count:start_frame_count+min(reference_frame_count, samples.shape[2]-1)] + normalized_first_frames = adaptive_mean_std_normalization(first_frames, reference_frames_data) + + samples[:, :, :start_frame_count] = normalized_first_frames + s["samples"] = samples + return io.NodeOutput(s) + + +class CLIPTextEncodeKandinsky5(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="CLIPTextEncodeKandinsky5", + category="advanced/conditioning/kandinsky5", + inputs=[ + io.Clip.Input("clip"), + io.String.Input("clip_l", multiline=True, dynamic_prompts=True), + io.String.Input("qwen25_7b", multiline=True, dynamic_prompts=True), + ], + outputs=[ + io.Conditioning.Output(), + ], + ) + + @classmethod + def execute(cls, clip, clip_l, qwen25_7b) -> io.NodeOutput: + tokens = clip.tokenize(clip_l) + tokens["qwen25_7b"] = clip.tokenize(qwen25_7b)["qwen25_7b"] + + return io.NodeOutput(clip.encode_from_tokens_scheduled(tokens)) + + +class Kandinsky5Extension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[io.ComfyNode]]: + return [ + Kandinsky5ImageToVideo, + NormalizeVideoLatentStart, + CLIPTextEncodeKandinsky5, + ] + +async def comfy_entrypoint() -> Kandinsky5Extension: + return Kandinsky5Extension() diff --git a/comfy_extras/nodes_latent.py b/comfy_extras/nodes_latent.py index d2df07ff9..e439b18ef 100644 --- a/comfy_extras/nodes_latent.py +++ b/comfy_extras/nodes_latent.py @@ -4,7 +4,7 @@ import torch import nodes from typing_extensions import override from comfy_api.latest import ComfyExtension, io - +import logging def reshape_latent_to(target_shape, latent, repeat_batch=True): if latent.shape[1:] != target_shape[1:]: @@ -388,6 +388,42 @@ class LatentOperationSharpen(io.ComfyNode): return luminance * sharpened return io.NodeOutput(sharpen) +class ReplaceVideoLatentFrames(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="ReplaceVideoLatentFrames", + category="latent/batch", + inputs=[ + io.Latent.Input("destination", tooltip="The destination latent where frames will be replaced."), + io.Latent.Input("source", optional=True, tooltip="The source latent providing frames to insert into the destination latent. If not provided, the destination latent is returned unchanged."), + io.Int.Input("index", default=0, min=-nodes.MAX_RESOLUTION, max=nodes.MAX_RESOLUTION, step=1, tooltip="The starting latent frame index in the destination latent where the source latent frames will be placed. Negative values count from the end."), + ], + outputs=[ + io.Latent.Output(), + ], + ) + + @classmethod + def execute(cls, destination, index, source=None) -> io.NodeOutput: + if source is None: + return io.NodeOutput(destination) + dest_frames = destination["samples"].shape[2] + source_frames = source["samples"].shape[2] + if index < 0: + index = dest_frames + index + if index > dest_frames: + logging.warning(f"ReplaceVideoLatentFrames: Index {index} is out of bounds for destination latent frames {dest_frames}.") + return io.NodeOutput(destination) + if index + source_frames > dest_frames: + logging.warning(f"ReplaceVideoLatentFrames: Source latent frames {source_frames} do not fit within destination latent frames {dest_frames} at the specified index {index}.") + return io.NodeOutput(destination) + s = source.copy() + s_source = source["samples"] + s_destination = destination["samples"].clone() + s_destination[:, :, index:index + s_source.shape[2]] = s_source + s["samples"] = s_destination + return io.NodeOutput(s) class LatentExtension(ComfyExtension): @override @@ -405,6 +441,7 @@ class LatentExtension(ComfyExtension): LatentApplyOperationCFG, LatentOperationTonemapReinhard, LatentOperationSharpen, + ReplaceVideoLatentFrames ] diff --git a/nodes.py b/nodes.py index 356aa63df..8d28a725d 100644 --- a/nodes.py +++ b/nodes.py @@ -970,7 +970,7 @@ class DualCLIPLoader: def INPUT_TYPES(s): return {"required": { "clip_name1": (folder_paths.get_filename_list("text_encoders"), ), "clip_name2": (folder_paths.get_filename_list("text_encoders"), ), - "type": (["sdxl", "sd3", "flux", "hunyuan_video", "hidream", "hunyuan_image", "hunyuan_video_15"], ), + "type": (["sdxl", "sd3", "flux", "hunyuan_video", "hidream", "hunyuan_image", "hunyuan_video_15", "kandinsky5", "kandinsky5_image"], ), }, "optional": { "device": (["default", "cpu"], {"advanced": True}), @@ -2357,6 +2357,7 @@ async def init_builtin_extra_nodes(): "nodes_rope.py", "nodes_logic.py", "nodes_nop.py", + "nodes_kandinsky5.py", ] import_failed = [] From ae676ed105663bb225153c8bca406f00edf738b4 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Fri, 5 Dec 2025 20:01:19 -0800 Subject: [PATCH 236/647] Fix regression. (#11137) --- comfy/supported_models.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy/supported_models.py b/comfy/supported_models.py index 91cc4ef08..383c82c3e 100644 --- a/comfy/supported_models.py +++ b/comfy/supported_models.py @@ -1529,6 +1529,6 @@ class Kandinsky5Image(Kandinsky5): return supported_models_base.ClipTarget(comfy.text_encoders.kandinsky5.Kandinsky5TokenizerImage, comfy.text_encoders.kandinsky5.te(**hunyuan_detect)) -models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanVideo15_SR_Distilled, HunyuanVideo15, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, ZImage, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, Omnigen2, QwenImage, Kandinsky5Image, Kandinsky5] +models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanVideo15_SR_Distilled, HunyuanVideo15, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, ZImage, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, Omnigen2, QwenImage, Flux2, Kandinsky5Image, Kandinsky5] models += [SVD_img2vid] From 117bf3f2bd9235cb5942a1de10a534c9869c7444 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Sat, 6 Dec 2025 06:22:02 +0200 Subject: [PATCH 237/647] convert nodes_freelunch.py to the V3 schema (#10904) --- comfy_extras/nodes_freelunch.py | 89 +++++++++++++++++---------- comfy_extras/nodes_model_downscale.py | 5 -- 2 files changed, 57 insertions(+), 37 deletions(-) diff --git a/comfy_extras/nodes_freelunch.py b/comfy_extras/nodes_freelunch.py index e3ac58447..3429b731e 100644 --- a/comfy_extras/nodes_freelunch.py +++ b/comfy_extras/nodes_freelunch.py @@ -2,6 +2,8 @@ import torch import logging +from typing_extensions import override +from comfy_api.latest import ComfyExtension, IO def Fourier_filter(x, threshold, scale): # FFT @@ -22,21 +24,26 @@ def Fourier_filter(x, threshold, scale): return x_filtered.to(x.dtype) -class FreeU: +class FreeU(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": { "model": ("MODEL",), - "b1": ("FLOAT", {"default": 1.1, "min": 0.0, "max": 10.0, "step": 0.01}), - "b2": ("FLOAT", {"default": 1.2, "min": 0.0, "max": 10.0, "step": 0.01}), - "s1": ("FLOAT", {"default": 0.9, "min": 0.0, "max": 10.0, "step": 0.01}), - "s2": ("FLOAT", {"default": 0.2, "min": 0.0, "max": 10.0, "step": 0.01}), - }} - RETURN_TYPES = ("MODEL",) - FUNCTION = "patch" + def define_schema(cls): + return IO.Schema( + node_id="FreeU", + category="model_patches/unet", + inputs=[ + IO.Model.Input("model"), + IO.Float.Input("b1", default=1.1, min=0.0, max=10.0, step=0.01), + IO.Float.Input("b2", default=1.2, min=0.0, max=10.0, step=0.01), + IO.Float.Input("s1", default=0.9, min=0.0, max=10.0, step=0.01), + IO.Float.Input("s2", default=0.2, min=0.0, max=10.0, step=0.01), + ], + outputs=[ + IO.Model.Output(), + ], + ) - CATEGORY = "model_patches/unet" - - def patch(self, model, b1, b2, s1, s2): + @classmethod + def execute(cls, model, b1, b2, s1, s2) -> IO.NodeOutput: model_channels = model.model.model_config.unet_config["model_channels"] scale_dict = {model_channels * 4: (b1, s1), model_channels * 2: (b2, s2)} on_cpu_devices = {} @@ -59,23 +66,31 @@ class FreeU: m = model.clone() m.set_model_output_block_patch(output_block_patch) - return (m, ) + return IO.NodeOutput(m) -class FreeU_V2: + patch = execute # TODO: remove + + +class FreeU_V2(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": { "model": ("MODEL",), - "b1": ("FLOAT", {"default": 1.3, "min": 0.0, "max": 10.0, "step": 0.01}), - "b2": ("FLOAT", {"default": 1.4, "min": 0.0, "max": 10.0, "step": 0.01}), - "s1": ("FLOAT", {"default": 0.9, "min": 0.0, "max": 10.0, "step": 0.01}), - "s2": ("FLOAT", {"default": 0.2, "min": 0.0, "max": 10.0, "step": 0.01}), - }} - RETURN_TYPES = ("MODEL",) - FUNCTION = "patch" + def define_schema(cls): + return IO.Schema( + node_id="FreeU_V2", + category="model_patches/unet", + inputs=[ + IO.Model.Input("model"), + IO.Float.Input("b1", default=1.3, min=0.0, max=10.0, step=0.01), + IO.Float.Input("b2", default=1.4, min=0.0, max=10.0, step=0.01), + IO.Float.Input("s1", default=0.9, min=0.0, max=10.0, step=0.01), + IO.Float.Input("s2", default=0.2, min=0.0, max=10.0, step=0.01), + ], + outputs=[ + IO.Model.Output(), + ], + ) - CATEGORY = "model_patches/unet" - - def patch(self, model, b1, b2, s1, s2): + @classmethod + def execute(cls, model, b1, b2, s1, s2) -> IO.NodeOutput: model_channels = model.model.model_config.unet_config["model_channels"] scale_dict = {model_channels * 4: (b1, s1), model_channels * 2: (b2, s2)} on_cpu_devices = {} @@ -105,9 +120,19 @@ class FreeU_V2: m = model.clone() m.set_model_output_block_patch(output_block_patch) - return (m, ) + return IO.NodeOutput(m) -NODE_CLASS_MAPPINGS = { - "FreeU": FreeU, - "FreeU_V2": FreeU_V2, -} + patch = execute # TODO: remove + + +class FreelunchExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[IO.ComfyNode]]: + return [ + FreeU, + FreeU_V2, + ] + + +async def comfy_entrypoint() -> FreelunchExtension: + return FreelunchExtension() diff --git a/comfy_extras/nodes_model_downscale.py b/comfy_extras/nodes_model_downscale.py index f7ca9699d..dec2ae841 100644 --- a/comfy_extras/nodes_model_downscale.py +++ b/comfy_extras/nodes_model_downscale.py @@ -53,11 +53,6 @@ class PatchModelAddDownscale(io.ComfyNode): return io.NodeOutput(m) -NODE_DISPLAY_NAME_MAPPINGS = { - # Sampling - "PatchModelAddDownscale": "", -} - class ModelDownscaleExtension(ComfyExtension): @override async def get_node_list(self) -> list[type[io.ComfyNode]]: From 913f86b72740f84f759786a698108840a09b6498 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Sat, 6 Dec 2025 06:24:10 +0200 Subject: [PATCH 238/647] [V3] convert nodes_mask.py to V3 schema (#10669) * convert nodes_mask.py to V3 schema * set "Preview Mask" as display name for MaskPreview --- comfy_extras/nodes_mask.py | 508 +++++++++++++++++++------------------ 1 file changed, 263 insertions(+), 245 deletions(-) diff --git a/comfy_extras/nodes_mask.py b/comfy_extras/nodes_mask.py index a5e405008..290e6f55e 100644 --- a/comfy_extras/nodes_mask.py +++ b/comfy_extras/nodes_mask.py @@ -3,11 +3,10 @@ import scipy.ndimage import torch import comfy.utils import node_helpers -import folder_paths -import random +from typing_extensions import override +from comfy_api.latest import ComfyExtension, IO, UI import nodes -from nodes import MAX_RESOLUTION def composite(destination, source, x, y, mask = None, multiplier = 8, resize_source = False): source = source.to(destination.device) @@ -46,202 +45,213 @@ def composite(destination, source, x, y, mask = None, multiplier = 8, resize_sou destination[..., top:bottom, left:right] = source_portion + destination_portion return destination -class LatentCompositeMasked: +class LatentCompositeMasked(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return { - "required": { - "destination": ("LATENT",), - "source": ("LATENT",), - "x": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 8}), - "y": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 8}), - "resize_source": ("BOOLEAN", {"default": False}), - }, - "optional": { - "mask": ("MASK",), - } - } - RETURN_TYPES = ("LATENT",) - FUNCTION = "composite" + def define_schema(cls): + return IO.Schema( + node_id="LatentCompositeMasked", + category="latent", + inputs=[ + IO.Latent.Input("destination"), + IO.Latent.Input("source"), + IO.Int.Input("x", default=0, min=0, max=nodes.MAX_RESOLUTION, step=8), + IO.Int.Input("y", default=0, min=0, max=nodes.MAX_RESOLUTION, step=8), + IO.Boolean.Input("resize_source", default=False), + IO.Mask.Input("mask", optional=True), + ], + outputs=[IO.Latent.Output()], + ) - CATEGORY = "latent" - - def composite(self, destination, source, x, y, resize_source, mask = None): + @classmethod + def execute(cls, destination, source, x, y, resize_source, mask = None) -> IO.NodeOutput: output = destination.copy() destination = destination["samples"].clone() source = source["samples"] output["samples"] = composite(destination, source, x, y, mask, 8, resize_source) - return (output,) + return IO.NodeOutput(output) -class ImageCompositeMasked: + composite = execute # TODO: remove + + +class ImageCompositeMasked(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return { - "required": { - "destination": ("IMAGE",), - "source": ("IMAGE",), - "x": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 1}), - "y": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 1}), - "resize_source": ("BOOLEAN", {"default": False}), - }, - "optional": { - "mask": ("MASK",), - } - } - RETURN_TYPES = ("IMAGE",) - FUNCTION = "composite" + def define_schema(cls): + return IO.Schema( + node_id="ImageCompositeMasked", + category="image", + inputs=[ + IO.Image.Input("destination"), + IO.Image.Input("source"), + IO.Int.Input("x", default=0, min=0, max=nodes.MAX_RESOLUTION, step=1), + IO.Int.Input("y", default=0, min=0, max=nodes.MAX_RESOLUTION, step=1), + IO.Boolean.Input("resize_source", default=False), + IO.Mask.Input("mask", optional=True), + ], + outputs=[IO.Image.Output()], + ) - CATEGORY = "image" - - def composite(self, destination, source, x, y, resize_source, mask = None): + @classmethod + def execute(cls, destination, source, x, y, resize_source, mask = None) -> IO.NodeOutput: destination, source = node_helpers.image_alpha_fix(destination, source) destination = destination.clone().movedim(-1, 1) output = composite(destination, source.movedim(-1, 1), x, y, mask, 1, resize_source).movedim(1, -1) - return (output,) + return IO.NodeOutput(output) -class MaskToImage: + composite = execute # TODO: remove + + +class MaskToImage(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return { - "required": { - "mask": ("MASK",), - } - } + def define_schema(cls): + return IO.Schema( + node_id="MaskToImage", + display_name="Convert Mask to Image", + category="mask", + inputs=[ + IO.Mask.Input("mask"), + ], + outputs=[IO.Image.Output()], + ) - CATEGORY = "mask" - - RETURN_TYPES = ("IMAGE",) - FUNCTION = "mask_to_image" - - def mask_to_image(self, mask): + @classmethod + def execute(cls, mask) -> IO.NodeOutput: result = mask.reshape((-1, 1, mask.shape[-2], mask.shape[-1])).movedim(1, -1).expand(-1, -1, -1, 3) - return (result,) + return IO.NodeOutput(result) -class ImageToMask: + mask_to_image = execute # TODO: remove + + +class ImageToMask(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return { - "required": { - "image": ("IMAGE",), - "channel": (["red", "green", "blue", "alpha"],), - } - } + def define_schema(cls): + return IO.Schema( + node_id="ImageToMask", + display_name="Convert Image to Mask", + category="mask", + inputs=[ + IO.Image.Input("image"), + IO.Combo.Input("channel", options=["red", "green", "blue", "alpha"]), + ], + outputs=[IO.Mask.Output()], + ) - CATEGORY = "mask" - - RETURN_TYPES = ("MASK",) - FUNCTION = "image_to_mask" - - def image_to_mask(self, image, channel): + @classmethod + def execute(cls, image, channel) -> IO.NodeOutput: channels = ["red", "green", "blue", "alpha"] mask = image[:, :, :, channels.index(channel)] - return (mask,) + return IO.NodeOutput(mask) -class ImageColorToMask: + image_to_mask = execute # TODO: remove + + +class ImageColorToMask(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return { - "required": { - "image": ("IMAGE",), - "color": ("INT", {"default": 0, "min": 0, "max": 0xFFFFFF, "step": 1, "display": "color"}), - } - } + def define_schema(cls): + return IO.Schema( + node_id="ImageColorToMask", + category="mask", + inputs=[ + IO.Image.Input("image"), + IO.Int.Input("color", default=0, min=0, max=0xFFFFFF, step=1, display_mode=IO.NumberDisplay.number), + ], + outputs=[IO.Mask.Output()], + ) - CATEGORY = "mask" - - RETURN_TYPES = ("MASK",) - FUNCTION = "image_to_mask" - - def image_to_mask(self, image, color): + @classmethod + def execute(cls, image, color) -> IO.NodeOutput: temp = (torch.clamp(image, 0, 1.0) * 255.0).round().to(torch.int) temp = torch.bitwise_left_shift(temp[:,:,:,0], 16) + torch.bitwise_left_shift(temp[:,:,:,1], 8) + temp[:,:,:,2] mask = torch.where(temp == color, 1.0, 0).float() - return (mask,) + return IO.NodeOutput(mask) -class SolidMask: + image_to_mask = execute # TODO: remove + + +class SolidMask(IO.ComfyNode): @classmethod - def INPUT_TYPES(cls): - return { - "required": { - "value": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0, "step": 0.01}), - "width": ("INT", {"default": 512, "min": 1, "max": MAX_RESOLUTION, "step": 1}), - "height": ("INT", {"default": 512, "min": 1, "max": MAX_RESOLUTION, "step": 1}), - } - } + def define_schema(cls): + return IO.Schema( + node_id="SolidMask", + category="mask", + inputs=[ + IO.Float.Input("value", default=1.0, min=0.0, max=1.0, step=0.01), + IO.Int.Input("width", default=512, min=1, max=nodes.MAX_RESOLUTION, step=1), + IO.Int.Input("height", default=512, min=1, max=nodes.MAX_RESOLUTION, step=1), + ], + outputs=[IO.Mask.Output()], + ) - CATEGORY = "mask" - - RETURN_TYPES = ("MASK",) - - FUNCTION = "solid" - - def solid(self, value, width, height): + @classmethod + def execute(cls, value, width, height) -> IO.NodeOutput: out = torch.full((1, height, width), value, dtype=torch.float32, device="cpu") - return (out,) + return IO.NodeOutput(out) -class InvertMask: + solid = execute # TODO: remove + + +class InvertMask(IO.ComfyNode): @classmethod - def INPUT_TYPES(cls): - return { - "required": { - "mask": ("MASK",), - } - } + def define_schema(cls): + return IO.Schema( + node_id="InvertMask", + category="mask", + inputs=[ + IO.Mask.Input("mask"), + ], + outputs=[IO.Mask.Output()], + ) - CATEGORY = "mask" - - RETURN_TYPES = ("MASK",) - - FUNCTION = "invert" - - def invert(self, mask): + @classmethod + def execute(cls, mask) -> IO.NodeOutput: out = 1.0 - mask - return (out,) + return IO.NodeOutput(out) -class CropMask: + invert = execute # TODO: remove + + +class CropMask(IO.ComfyNode): @classmethod - def INPUT_TYPES(cls): - return { - "required": { - "mask": ("MASK",), - "x": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 1}), - "y": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 1}), - "width": ("INT", {"default": 512, "min": 1, "max": MAX_RESOLUTION, "step": 1}), - "height": ("INT", {"default": 512, "min": 1, "max": MAX_RESOLUTION, "step": 1}), - } - } + def define_schema(cls): + return IO.Schema( + node_id="CropMask", + category="mask", + inputs=[ + IO.Mask.Input("mask"), + IO.Int.Input("x", default=0, min=0, max=nodes.MAX_RESOLUTION, step=1), + IO.Int.Input("y", default=0, min=0, max=nodes.MAX_RESOLUTION, step=1), + IO.Int.Input("width", default=512, min=1, max=nodes.MAX_RESOLUTION, step=1), + IO.Int.Input("height", default=512, min=1, max=nodes.MAX_RESOLUTION, step=1), + ], + outputs=[IO.Mask.Output()], + ) - CATEGORY = "mask" - - RETURN_TYPES = ("MASK",) - - FUNCTION = "crop" - - def crop(self, mask, x, y, width, height): + @classmethod + def execute(cls, mask, x, y, width, height) -> IO.NodeOutput: mask = mask.reshape((-1, mask.shape[-2], mask.shape[-1])) out = mask[:, y:y + height, x:x + width] - return (out,) + return IO.NodeOutput(out) -class MaskComposite: + crop = execute # TODO: remove + + +class MaskComposite(IO.ComfyNode): @classmethod - def INPUT_TYPES(cls): - return { - "required": { - "destination": ("MASK",), - "source": ("MASK",), - "x": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 1}), - "y": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 1}), - "operation": (["multiply", "add", "subtract", "and", "or", "xor"],), - } - } + def define_schema(cls): + return IO.Schema( + node_id="MaskComposite", + category="mask", + inputs=[ + IO.Mask.Input("destination"), + IO.Mask.Input("source"), + IO.Int.Input("x", default=0, min=0, max=nodes.MAX_RESOLUTION, step=1), + IO.Int.Input("y", default=0, min=0, max=nodes.MAX_RESOLUTION, step=1), + IO.Combo.Input("operation", options=["multiply", "add", "subtract", "and", "or", "xor"]), + ], + outputs=[IO.Mask.Output()], + ) - CATEGORY = "mask" - - RETURN_TYPES = ("MASK",) - - FUNCTION = "combine" - - def combine(self, destination, source, x, y, operation): + @classmethod + def execute(cls, destination, source, x, y, operation) -> IO.NodeOutput: output = destination.reshape((-1, destination.shape[-2], destination.shape[-1])).clone() source = source.reshape((-1, source.shape[-2], source.shape[-1])) @@ -267,28 +277,29 @@ class MaskComposite: output = torch.clamp(output, 0.0, 1.0) - return (output,) + return IO.NodeOutput(output) -class FeatherMask: + combine = execute # TODO: remove + + +class FeatherMask(IO.ComfyNode): @classmethod - def INPUT_TYPES(cls): - return { - "required": { - "mask": ("MASK",), - "left": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 1}), - "top": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 1}), - "right": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 1}), - "bottom": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 1}), - } - } + def define_schema(cls): + return IO.Schema( + node_id="FeatherMask", + category="mask", + inputs=[ + IO.Mask.Input("mask"), + IO.Int.Input("left", default=0, min=0, max=nodes.MAX_RESOLUTION, step=1), + IO.Int.Input("top", default=0, min=0, max=nodes.MAX_RESOLUTION, step=1), + IO.Int.Input("right", default=0, min=0, max=nodes.MAX_RESOLUTION, step=1), + IO.Int.Input("bottom", default=0, min=0, max=nodes.MAX_RESOLUTION, step=1), + ], + outputs=[IO.Mask.Output()], + ) - CATEGORY = "mask" - - RETURN_TYPES = ("MASK",) - - FUNCTION = "feather" - - def feather(self, mask, left, top, right, bottom): + @classmethod + def execute(cls, mask, left, top, right, bottom) -> IO.NodeOutput: output = mask.reshape((-1, mask.shape[-2], mask.shape[-1])).clone() left = min(left, output.shape[-1]) @@ -312,26 +323,28 @@ class FeatherMask: feather_rate = (y + 1) / bottom output[:, -y, :] *= feather_rate - return (output,) + return IO.NodeOutput(output) -class GrowMask: + feather = execute # TODO: remove + + +class GrowMask(IO.ComfyNode): @classmethod - def INPUT_TYPES(cls): - return { - "required": { - "mask": ("MASK",), - "expand": ("INT", {"default": 0, "min": -MAX_RESOLUTION, "max": MAX_RESOLUTION, "step": 1}), - "tapered_corners": ("BOOLEAN", {"default": True}), - }, - } + def define_schema(cls): + return IO.Schema( + node_id="GrowMask", + display_name="Grow Mask", + category="mask", + inputs=[ + IO.Mask.Input("mask"), + IO.Int.Input("expand", default=0, min=-nodes.MAX_RESOLUTION, max=nodes.MAX_RESOLUTION, step=1), + IO.Boolean.Input("tapered_corners", default=True), + ], + outputs=[IO.Mask.Output()], + ) - CATEGORY = "mask" - - RETURN_TYPES = ("MASK",) - - FUNCTION = "expand_mask" - - def expand_mask(self, mask, expand, tapered_corners): + @classmethod + def execute(cls, mask, expand, tapered_corners) -> IO.NodeOutput: c = 0 if tapered_corners else 1 kernel = np.array([[c, 1, c], [1, 1, 1], @@ -347,69 +360,74 @@ class GrowMask: output = scipy.ndimage.grey_dilation(output, footprint=kernel) output = torch.from_numpy(output) out.append(output) - return (torch.stack(out, dim=0),) + return IO.NodeOutput(torch.stack(out, dim=0)) -class ThresholdMask: + expand_mask = execute # TODO: remove + + +class ThresholdMask(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return { - "required": { - "mask": ("MASK",), - "value": ("FLOAT", {"default": 0.5, "min": 0.0, "max": 1.0, "step": 0.01}), - } - } + def define_schema(cls): + return IO.Schema( + node_id="ThresholdMask", + category="mask", + inputs=[ + IO.Mask.Input("mask"), + IO.Float.Input("value", default=0.5, min=0.0, max=1.0, step=0.01), + ], + outputs=[IO.Mask.Output()], + ) - CATEGORY = "mask" - - RETURN_TYPES = ("MASK",) - FUNCTION = "image_to_mask" - - def image_to_mask(self, mask, value): + @classmethod + def execute(cls, mask, value) -> IO.NodeOutput: mask = (mask > value).float() - return (mask,) + return IO.NodeOutput(mask) + + image_to_mask = execute # TODO: remove + # Mask Preview - original implement from # https://github.com/cubiq/ComfyUI_essentials/blob/9d9f4bedfc9f0321c19faf71855e228c93bd0dc9/mask.py#L81 # upstream requested in https://github.com/Kosinkadink/rfcs/blob/main/rfcs/0000-corenodes.md#preview-nodes -class MaskPreview(nodes.SaveImage): - def __init__(self): - self.output_dir = folder_paths.get_temp_directory() - self.type = "temp" - self.prefix_append = "_temp_" + ''.join(random.choice("abcdefghijklmnopqrstupvxyz") for x in range(5)) - self.compress_level = 4 +class MaskPreview(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="MaskPreview", + display_name="Preview Mask", + category="mask", + description="Saves the input images to your ComfyUI output directory.", + inputs=[ + IO.Mask.Input("mask"), + ], + hidden=[IO.Hidden.prompt, IO.Hidden.extra_pnginfo], + is_output_node=True, + ) @classmethod - def INPUT_TYPES(s): - return { - "required": {"mask": ("MASK",), }, - "hidden": {"prompt": "PROMPT", "extra_pnginfo": "EXTRA_PNGINFO"}, - } - - FUNCTION = "execute" - CATEGORY = "mask" - - def execute(self, mask, filename_prefix="ComfyUI", prompt=None, extra_pnginfo=None): - preview = mask.reshape((-1, 1, mask.shape[-2], mask.shape[-1])).movedim(1, -1).expand(-1, -1, -1, 3) - return self.save_images(preview, filename_prefix, prompt, extra_pnginfo) + def execute(cls, mask, filename_prefix="ComfyUI") -> IO.NodeOutput: + return IO.NodeOutput(ui=UI.PreviewMask(mask)) -NODE_CLASS_MAPPINGS = { - "LatentCompositeMasked": LatentCompositeMasked, - "ImageCompositeMasked": ImageCompositeMasked, - "MaskToImage": MaskToImage, - "ImageToMask": ImageToMask, - "ImageColorToMask": ImageColorToMask, - "SolidMask": SolidMask, - "InvertMask": InvertMask, - "CropMask": CropMask, - "MaskComposite": MaskComposite, - "FeatherMask": FeatherMask, - "GrowMask": GrowMask, - "ThresholdMask": ThresholdMask, - "MaskPreview": MaskPreview -} +class MaskExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[IO.ComfyNode]]: + return [ + LatentCompositeMasked, + ImageCompositeMasked, + MaskToImage, + ImageToMask, + ImageColorToMask, + SolidMask, + InvertMask, + CropMask, + MaskComposite, + FeatherMask, + GrowMask, + ThresholdMask, + MaskPreview, + ] -NODE_DISPLAY_NAME_MAPPINGS = { - "ImageToMask": "Convert Image to Mask", - "MaskToImage": "Convert Mask to Image", -} + +async def comfy_entrypoint() -> MaskExtension: + return MaskExtension() From d7a0aef65033bf0fe56e521577a44fac1830b8b3 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Fri, 5 Dec 2025 21:15:21 -0800 Subject: [PATCH 239/647] Set OCL_SET_SVM_SIZE on AMD. (#11139) --- cuda_malloc.py | 27 +++++++++++++++++---------- main.py | 3 +++ 2 files changed, 20 insertions(+), 10 deletions(-) diff --git a/cuda_malloc.py b/cuda_malloc.py index 6520d5123..ee2bc4b69 100644 --- a/cuda_malloc.py +++ b/cuda_malloc.py @@ -63,18 +63,22 @@ def cuda_malloc_supported(): return True +version = "" + +try: + torch_spec = importlib.util.find_spec("torch") + for folder in torch_spec.submodule_search_locations: + ver_file = os.path.join(folder, "version.py") + if os.path.isfile(ver_file): + spec = importlib.util.spec_from_file_location("torch_version_import", ver_file) + module = importlib.util.module_from_spec(spec) + spec.loader.exec_module(module) + version = module.__version__ +except: + pass + if not args.cuda_malloc: try: - version = "" - torch_spec = importlib.util.find_spec("torch") - for folder in torch_spec.submodule_search_locations: - ver_file = os.path.join(folder, "version.py") - if os.path.isfile(ver_file): - spec = importlib.util.spec_from_file_location("torch_version_import", ver_file) - module = importlib.util.module_from_spec(spec) - spec.loader.exec_module(module) - version = module.__version__ - if int(version[0]) >= 2 and "+cu" in version: # enable by default for torch version 2.0 and up only on cuda torch if PerformanceFeature.AutoTune not in args.fast: # Autotune has issues with cuda malloc args.cuda_malloc = cuda_malloc_supported() @@ -90,3 +94,6 @@ if args.cuda_malloc and not args.disable_cuda_malloc: env_var += ",backend:cudaMallocAsync" os.environ['PYTORCH_CUDA_ALLOC_CONF'] = env_var + +def get_torch_version_noimport(): + return str(version) diff --git a/main.py b/main.py index 0cd815d9e..0d02a087b 100644 --- a/main.py +++ b/main.py @@ -167,6 +167,9 @@ if __name__ == "__main__": os.environ['CUBLAS_WORKSPACE_CONFIG'] = ":4096:8" import cuda_malloc + if "rocm" in cuda_malloc.get_torch_version_noimport(): + os.environ['OCL_SET_SVM_SIZE'] = '262144' # set at the request of AMD + if 'torch' in sys.modules: logging.warning("WARNING: Potential Error in code: Torch already imported, torch should never be imported before this point.") From 76f18e955dcbc88ed13d6802194fd897927f93e5 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Sat, 6 Dec 2025 13:28:08 +0200 Subject: [PATCH 240/647] marked all Pika API nodes a deprecated (#11146) --- comfy_api_nodes/nodes_pika.py | 7 +++++++ 1 file changed, 7 insertions(+) diff --git a/comfy_api_nodes/nodes_pika.py b/comfy_api_nodes/nodes_pika.py index 51148211b..acd88c391 100644 --- a/comfy_api_nodes/nodes_pika.py +++ b/comfy_api_nodes/nodes_pika.py @@ -92,6 +92,7 @@ class PikaImageToVideo(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + is_deprecated=True, ) @classmethod @@ -152,6 +153,7 @@ class PikaTextToVideoNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + is_deprecated=True, ) @classmethod @@ -239,6 +241,7 @@ class PikaScenes(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + is_deprecated=True, ) @classmethod @@ -323,6 +326,7 @@ class PikAdditionsNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + is_deprecated=True, ) @classmethod @@ -399,6 +403,7 @@ class PikaSwapsNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + is_deprecated=True, ) @classmethod @@ -466,6 +471,7 @@ class PikaffectsNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + is_deprecated=True, ) @classmethod @@ -515,6 +521,7 @@ class PikaStartEndFrameNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + is_deprecated=True, ) @classmethod From 7ac7d69d948e75c3a230d1262daab84d75aff895 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Jukka=20Sepp=C3=A4nen?= <40791699+kijai@users.noreply.github.com> Date: Sat, 6 Dec 2025 20:09:44 +0200 Subject: [PATCH 241/647] Fix EmptyAudio node input types (#11149) --- comfy_extras/nodes_audio.py | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) diff --git a/comfy_extras/nodes_audio.py b/comfy_extras/nodes_audio.py index 812301fb7..c7916443c 100644 --- a/comfy_extras/nodes_audio.py +++ b/comfy_extras/nodes_audio.py @@ -573,12 +573,14 @@ class EmptyAudio(IO.ComfyNode): step=0.01, tooltip="Duration of the empty audio clip in seconds", ), - IO.Float.Input( + IO.Int.Input( "sample_rate", default=44100, tooltip="Sample rate of the empty audio clip.", + min=1, + max=192000, ), - IO.Float.Input( + IO.Int.Input( "channels", default=2, min=1, From 50ca97e7765d9bbdbeec31a75f1f6c747d76948c Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sat, 6 Dec 2025 15:36:20 -0800 Subject: [PATCH 242/647] Speed up lora compute and lower memory usage by doing it in fp16. (#11161) --- comfy/model_management.py | 14 ++++++++++++++ comfy/model_patcher.py | 5 +++-- 2 files changed, 17 insertions(+), 2 deletions(-) diff --git a/comfy/model_management.py b/comfy/model_management.py index aeddbaefe..40717b1e4 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -1492,6 +1492,20 @@ def extended_fp16_support(): return True +LORA_COMPUTE_DTYPES = {} +def lora_compute_dtype(device): + dtype = LORA_COMPUTE_DTYPES.get(device, None) + if dtype is not None: + return dtype + + if should_use_fp16(device): + dtype = torch.float16 + else: + dtype = torch.float32 + + LORA_COMPUTE_DTYPES[device] = dtype + return dtype + def soft_empty_cache(force=False): global cpu_state if cpu_state == CPUState.MPS: diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index 215784874..4f076a6aa 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -614,10 +614,11 @@ class ModelPatcher: if key not in self.backup: self.backup[key] = collections.namedtuple('Dimension', ['weight', 'inplace_update'])(weight.to(device=self.offload_device, copy=inplace_update), inplace_update) + temp_dtype = comfy.model_management.lora_compute_dtype(device_to) if device_to is not None: - temp_weight = comfy.model_management.cast_to_device(weight, device_to, torch.float32, copy=True) + temp_weight = comfy.model_management.cast_to_device(weight, device_to, temp_dtype, copy=True) else: - temp_weight = weight.to(torch.float32, copy=True) + temp_weight = weight.to(temp_dtype, copy=True) if convert_func is not None: temp_weight = convert_func(temp_weight, inplace=True) From 4086acf3c2f0ca3a8861b04f6179fa9f908e3e25 Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Sun, 7 Dec 2025 09:42:09 +1000 Subject: [PATCH 243/647] Fix on-load VRAM OOM (#11144) slow down the CPU on model load to not run ahead. This fixes a VRAM on flux 2 load. I went to try and debug this with the memory trace pickles, which needs --disable-cuda-malloc which made the bug go away. So I tried this synchronize and it worked. The has some very complex interactions with the cuda malloc async and I dont have solid theory on this one yet. Still debugging but this gets us over the OOM for the moment. --- comfy/model_patcher.py | 2 ++ 1 file changed, 2 insertions(+) diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index 4f076a6aa..5b1ccb824 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -762,6 +762,8 @@ class ModelPatcher: key = "{}.{}".format(n, param) self.unpin_weight(key) self.patch_weight_to_device(key, device_to=device_to) + if comfy.model_management.is_device_cuda(device_to): + torch.cuda.synchronize() logging.debug("lowvram: loaded module regularly {} {}".format(n, m)) m.comfy_patched_weights = True From 329480da5ab32949a411548f821ea60ab3e90dc7 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sat, 6 Dec 2025 17:50:10 -0800 Subject: [PATCH 244/647] Fix qwen scaled fp8 not working with kandinsky. Make basic t2i wf work. (#11162) --- comfy/ldm/kandinsky5/model.py | 8 +++++++- comfy/text_encoders/kandinsky5.py | 12 ++++++------ 2 files changed, 13 insertions(+), 7 deletions(-) diff --git a/comfy/ldm/kandinsky5/model.py b/comfy/ldm/kandinsky5/model.py index a653e02fc..1509de2f8 100644 --- a/comfy/ldm/kandinsky5/model.py +++ b/comfy/ldm/kandinsky5/model.py @@ -387,6 +387,9 @@ class Kandinsky5(nn.Module): return self.out_layer(visual_embed, time_embed) def _forward(self, x, timestep, context, y, time_dim_replace=None, transformer_options={}, **kwargs): + original_dims = x.ndim + if original_dims == 4: + x = x.unsqueeze(2) bs, c, t_len, h, w = x.shape x = comfy.ldm.common_dit.pad_to_patch_size(x, self.patch_size) @@ -397,7 +400,10 @@ class Kandinsky5(nn.Module): freqs = self.rope_encode_3d(t_len, h, w, device=x.device, dtype=x.dtype, transformer_options=transformer_options) freqs_text = self.rope_encode_1d(context.shape[1], device=x.device, dtype=x.dtype, transformer_options=transformer_options) - return self.forward_orig(x, timestep, context, y, freqs, freqs_text, transformer_options=transformer_options, **kwargs) + out = self.forward_orig(x, timestep, context, y, freqs, freqs_text, transformer_options=transformer_options, **kwargs) + if original_dims == 4: + out = out.squeeze(2) + return out def forward(self, x, timestep, context, y, time_dim_replace=None, transformer_options={}, **kwargs): return comfy.patcher_extension.WrapperExecutor.new_class_executor( diff --git a/comfy/text_encoders/kandinsky5.py b/comfy/text_encoders/kandinsky5.py index 22f991c36..be086458c 100644 --- a/comfy/text_encoders/kandinsky5.py +++ b/comfy/text_encoders/kandinsky5.py @@ -24,10 +24,10 @@ class Kandinsky5TokenizerImage(Kandinsky5Tokenizer): class Qwen25_7BVLIModel(sd1_clip.SDClipModel): def __init__(self, device="cpu", layer="hidden", layer_idx=-1, dtype=None, attention_mask=True, model_options={}): - llama_scaled_fp8 = model_options.get("qwen_scaled_fp8", None) - if llama_scaled_fp8 is not None: + llama_quantization_metadata = model_options.get("llama_quantization_metadata", None) + if llama_quantization_metadata is not None: model_options = model_options.copy() - model_options["scaled_fp8"] = llama_scaled_fp8 + model_options["quantization_metadata"] = llama_quantization_metadata super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"pad": 151643}, layer_norm_hidden_state=False, model_class=Qwen25_7BVLI, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options) @@ -56,12 +56,12 @@ class Kandinsky5TEModel(QwenImageTEModel): else: return super().load_sd(sd) -def te(dtype_llama=None, llama_scaled_fp8=None): +def te(dtype_llama=None, llama_quantization_metadata=None): class Kandinsky5TEModel_(Kandinsky5TEModel): def __init__(self, device="cpu", dtype=None, model_options={}): - if llama_scaled_fp8 is not None and "scaled_fp8" not in model_options: + if llama_quantization_metadata is not None: model_options = model_options.copy() - model_options["qwen_scaled_fp8"] = llama_scaled_fp8 + model_options["llama_quantization_metadata"] = llama_quantization_metadata if dtype_llama is not None: dtype = dtype_llama super().__init__(device=device, dtype=dtype, model_options=model_options) From 56fa7dbe380cb5591c5542f8aa51ce2fc26beedf Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sun, 7 Dec 2025 04:44:55 -0800 Subject: [PATCH 245/647] Properly load the newbie diffusion model. (#11172) There is still one of the text encoders missing and I didn't actually test it. --- comfy/ldm/lumina/model.py | 35 +++++++++++++++++++++++++++++++++++ comfy/model_base.py | 4 ++++ comfy/model_detection.py | 3 +++ 3 files changed, 42 insertions(+) diff --git a/comfy/ldm/lumina/model.py b/comfy/ldm/lumina/model.py index 6c24fed9b..c47df49ca 100644 --- a/comfy/ldm/lumina/model.py +++ b/comfy/ldm/lumina/model.py @@ -377,6 +377,7 @@ class NextDiT(nn.Module): z_image_modulation=False, time_scale=1.0, pad_tokens_multiple=None, + clip_text_dim=None, image_model=None, device=None, dtype=None, @@ -447,6 +448,31 @@ class NextDiT(nn.Module): ), ) + self.clip_text_pooled_proj = None + + if clip_text_dim is not None: + self.clip_text_dim = clip_text_dim + self.clip_text_pooled_proj = nn.Sequential( + operation_settings.get("operations").RMSNorm(clip_text_dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")), + operation_settings.get("operations").Linear( + clip_text_dim, + clip_text_dim, + bias=True, + device=operation_settings.get("device"), + dtype=operation_settings.get("dtype"), + ), + ) + self.time_text_embed = nn.Sequential( + nn.SiLU(), + operation_settings.get("operations").Linear( + min(dim, 1024) + clip_text_dim, + min(dim, 1024), + bias=True, + device=operation_settings.get("device"), + dtype=operation_settings.get("dtype"), + ), + ) + self.layers = nn.ModuleList( [ JointTransformerBlock( @@ -585,6 +611,15 @@ class NextDiT(nn.Module): cap_feats = self.cap_embedder(cap_feats) # (N, L, D) # todo check if able to batchify w.o. redundant compute + if self.clip_text_pooled_proj is not None: + pooled = kwargs.get("clip_text_pooled", None) + if pooled is not None: + pooled = self.clip_text_pooled_proj(pooled) + else: + pooled = torch.zeros((1, self.clip_text_dim), device=x.device, dtype=x.dtype) + + adaln_input = self.time_text_embed(torch.cat((t, pooled), dim=-1)) + patches = transformer_options.get("patches", {}) x_is_tensor = isinstance(x, torch.Tensor) img, mask, img_size, cap_size, freqs_cis = self.patchify_and_embed(x, cap_feats, cap_mask, t, num_tokens, transformer_options=transformer_options) diff --git a/comfy/model_base.py b/comfy/model_base.py index 0be006cc2..6b8a8454d 100644 --- a/comfy/model_base.py +++ b/comfy/model_base.py @@ -1110,6 +1110,10 @@ class Lumina2(BaseModel): if 'num_tokens' not in out: out['num_tokens'] = comfy.conds.CONDConstant(cross_attn.shape[1]) + clip_text_pooled = kwargs["pooled_output"] # Newbie + if clip_text_pooled is not None: + out['clip_text_pooled'] = comfy.conds.CONDRegular(clip_text_pooled) + return out class WAN21(BaseModel): diff --git a/comfy/model_detection.py b/comfy/model_detection.py index 30b33a486..74c547427 100644 --- a/comfy/model_detection.py +++ b/comfy/model_detection.py @@ -423,6 +423,9 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): dit_config["axes_lens"] = [300, 512, 512] dit_config["rope_theta"] = 10000.0 dit_config["ffn_dim_multiplier"] = 4.0 + ctd_weight = state_dict.get('{}clip_text_pooled_proj.0.weight'.format(key_prefix), None) + if ctd_weight is not None: + dit_config["clip_text_dim"] = ctd_weight.shape[0] elif dit_config["dim"] == 3840: # Z image dit_config["n_heads"] = 30 dit_config["n_kv_heads"] = 30 From ec7f65187d85e22ea23345ce0d919e11768f255e Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Mon, 8 Dec 2025 11:21:41 +0200 Subject: [PATCH 246/647] chore(comfy_api): replace absolute imports with relative (#11145) --- comfy_api/latest/__init__.py | 8 ++++---- comfy_api/latest/_input/video_types.py | 2 +- comfy_api/latest/_input_impl/video_types.py | 4 ++-- comfy_api/latest/_io.py | 2 +- comfy_api/latest/_ui.py | 2 +- comfy_api/latest/_util/video_types.py | 2 +- 6 files changed, 10 insertions(+), 10 deletions(-) diff --git a/comfy_api/latest/__init__.py b/comfy_api/latest/__init__.py index 0fa01d1e7..35e1ac853 100644 --- a/comfy_api/latest/__init__.py +++ b/comfy_api/latest/__init__.py @@ -5,9 +5,9 @@ from typing import Type, TYPE_CHECKING from comfy_api.internal import ComfyAPIBase from comfy_api.internal.singleton import ProxiedSingleton from comfy_api.internal.async_to_sync import create_sync_class -from comfy_api.latest._input import ImageInput, AudioInput, MaskInput, LatentInput, VideoInput -from comfy_api.latest._input_impl import VideoFromFile, VideoFromComponents -from comfy_api.latest._util import VideoCodec, VideoContainer, VideoComponents, MESH, VOXEL +from ._input import ImageInput, AudioInput, MaskInput, LatentInput, VideoInput +from ._input_impl import VideoFromFile, VideoFromComponents +from ._util import VideoCodec, VideoContainer, VideoComponents, MESH, VOXEL from . import _io_public as io from . import _ui_public as ui # from comfy_api.latest._resources import _RESOURCES as resources #noqa: F401 @@ -80,7 +80,7 @@ class ComfyExtension(ABC): async def on_load(self) -> None: """ Called when an extension is loaded. - This should be used to initialize any global resources neeeded by the extension. + This should be used to initialize any global resources needed by the extension. """ @abstractmethod diff --git a/comfy_api/latest/_input/video_types.py b/comfy_api/latest/_input/video_types.py index 87c81d73a..e634a0311 100644 --- a/comfy_api/latest/_input/video_types.py +++ b/comfy_api/latest/_input/video_types.py @@ -4,7 +4,7 @@ from fractions import Fraction from typing import Optional, Union, IO import io import av -from comfy_api.util import VideoContainer, VideoCodec, VideoComponents +from .._util import VideoContainer, VideoCodec, VideoComponents class VideoInput(ABC): """ diff --git a/comfy_api/latest/_input_impl/video_types.py b/comfy_api/latest/_input_impl/video_types.py index a4cd3737d..ea35c6062 100644 --- a/comfy_api/latest/_input_impl/video_types.py +++ b/comfy_api/latest/_input_impl/video_types.py @@ -3,14 +3,14 @@ from av.container import InputContainer from av.subtitles.stream import SubtitleStream from fractions import Fraction from typing import Optional -from comfy_api.latest._input import AudioInput, VideoInput +from .._input import AudioInput, VideoInput import av import io import json import numpy as np import math import torch -from comfy_api.latest._util import VideoContainer, VideoCodec, VideoComponents +from .._util import VideoContainer, VideoCodec, VideoComponents def container_to_output_format(container_format: str | None) -> str | None: diff --git a/comfy_api/latest/_io.py b/comfy_api/latest/_io.py index d7cbe68cf..313a5af20 100644 --- a/comfy_api/latest/_io.py +++ b/comfy_api/latest/_io.py @@ -26,7 +26,7 @@ if TYPE_CHECKING: from comfy_api.input import VideoInput from comfy_api.internal import (_ComfyNodeInternal, _NodeOutputInternal, classproperty, copy_class, first_real_override, is_class, prune_dict, shallow_clone_class) -from comfy_api.latest._resources import Resources, ResourcesLocal +from ._resources import Resources, ResourcesLocal from comfy_execution.graph_utils import ExecutionBlocker from ._util import MESH, VOXEL diff --git a/comfy_api/latest/_ui.py b/comfy_api/latest/_ui.py index 5a75a3aae..2babe209a 100644 --- a/comfy_api/latest/_ui.py +++ b/comfy_api/latest/_ui.py @@ -22,7 +22,7 @@ import folder_paths # used for image preview from comfy.cli_args import args -from comfy_api.latest._io import ComfyNode, FolderType, Image, _UIOutput +from ._io import ComfyNode, FolderType, Image, _UIOutput class SavedResult(dict): diff --git a/comfy_api/latest/_util/video_types.py b/comfy_api/latest/_util/video_types.py index c3e3d8e3a..fd3b5a510 100644 --- a/comfy_api/latest/_util/video_types.py +++ b/comfy_api/latest/_util/video_types.py @@ -3,7 +3,7 @@ from dataclasses import dataclass from enum import Enum from fractions import Fraction from typing import Optional -from comfy_api.latest._input import ImageInput, AudioInput +from .._input import ImageInput, AudioInput class VideoCodec(str, Enum): AUTO = "auto" From 058f084371ef2ed0c456118dfdd3d0bfed17259b Mon Sep 17 00:00:00 2001 From: ComfyUI Wiki Date: Mon, 8 Dec 2025 17:22:51 +0800 Subject: [PATCH 247/647] Update workflow templates to v0.7.51 (#11150) * chore: update workflow templates to v0.7.50 * Update template to 0.7.51 --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index f98848e20..12a7c1089 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,5 +1,5 @@ comfyui-frontend-package==1.33.10 -comfyui-workflow-templates==0.7.25 +comfyui-workflow-templates==0.7.51 comfyui-embedded-docs==0.3.1 torch torchsde From 85c4b4ae262c2de360891dd23c6504da2f5a6014 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Mon, 8 Dec 2025 11:27:02 +0200 Subject: [PATCH 248/647] chore: replace imports of deprecated V1 classes (#11127) --- comfy_api_nodes/apis/veo_api.py | 2 +- comfy_api_nodes/nodes_gemini.py | 19 ++++++++++--------- comfy_api_nodes/nodes_ltxv.py | 17 +++++++---------- comfy_api_nodes/nodes_moonvalley.py | 19 ++++++++----------- comfy_api_nodes/nodes_runway.py | 29 +++++++++++++---------------- comfy_api_nodes/nodes_veo2.py | 12 +++++------- comfy_extras/nodes_video.py | 27 +++++++++++---------------- 7 files changed, 55 insertions(+), 70 deletions(-) diff --git a/comfy_api_nodes/apis/veo_api.py b/comfy_api_nodes/apis/veo_api.py index 8328d1aa4..23ca725b7 100644 --- a/comfy_api_nodes/apis/veo_api.py +++ b/comfy_api_nodes/apis/veo_api.py @@ -85,7 +85,7 @@ class Response1(BaseModel): raiMediaFilteredReasons: Optional[list[str]] = Field( None, description='Reasons why media was filtered by responsible AI policies' ) - videos: Optional[list[Video]] = None + videos: Optional[list[Video]] = Field(None) class VeoGenVidPollResponse(BaseModel): diff --git a/comfy_api_nodes/nodes_gemini.py b/comfy_api_nodes/nodes_gemini.py index 08f7b0f64..0b7422ef7 100644 --- a/comfy_api_nodes/nodes_gemini.py +++ b/comfy_api_nodes/nodes_gemini.py @@ -13,8 +13,7 @@ import torch from typing_extensions import override import folder_paths -from comfy_api.latest import IO, ComfyExtension, Input -from comfy_api.util import VideoCodec, VideoContainer +from comfy_api.latest import IO, ComfyExtension, Input, Types from comfy_api_nodes.apis.gemini_api import ( GeminiContent, GeminiFileData, @@ -68,7 +67,7 @@ class GeminiImageModel(str, Enum): async def create_image_parts( cls: type[IO.ComfyNode], - images: torch.Tensor, + images: Input.Image, image_limit: int = 0, ) -> list[GeminiPart]: image_parts: list[GeminiPart] = [] @@ -154,8 +153,8 @@ def get_text_from_response(response: GeminiGenerateContentResponse) -> str: return "\n".join([part.text for part in parts]) -def get_image_from_response(response: GeminiGenerateContentResponse) -> torch.Tensor: - image_tensors: list[torch.Tensor] = [] +def get_image_from_response(response: GeminiGenerateContentResponse) -> Input.Image: + image_tensors: list[Input.Image] = [] parts = get_parts_by_type(response, "image/png") for part in parts: image_data = base64.b64decode(part.inlineData.data) @@ -293,7 +292,9 @@ class GeminiNode(IO.ComfyNode): def create_video_parts(cls, video_input: Input.Video) -> list[GeminiPart]: """Convert video input to Gemini API compatible parts.""" - base_64_string = video_to_base64_string(video_input, container_format=VideoContainer.MP4, codec=VideoCodec.H264) + base_64_string = video_to_base64_string( + video_input, container_format=Types.VideoContainer.MP4, codec=Types.VideoCodec.H264 + ) return [ GeminiPart( inlineData=GeminiInlineData( @@ -343,7 +344,7 @@ class GeminiNode(IO.ComfyNode): prompt: str, model: str, seed: int, - images: torch.Tensor | None = None, + images: Input.Image | None = None, audio: Input.Audio | None = None, video: Input.Video | None = None, files: list[GeminiPart] | None = None, @@ -542,7 +543,7 @@ class GeminiImage(IO.ComfyNode): prompt: str, model: str, seed: int, - images: torch.Tensor | None = None, + images: Input.Image | None = None, files: list[GeminiPart] | None = None, aspect_ratio: str = "auto", response_modalities: str = "IMAGE+TEXT", @@ -662,7 +663,7 @@ class GeminiImage2(IO.ComfyNode): aspect_ratio: str, resolution: str, response_modalities: str, - images: torch.Tensor | None = None, + images: Input.Image | None = None, files: list[GeminiPart] | None = None, ) -> IO.NodeOutput: validate_string(prompt, strip_whitespace=True, min_length=1) diff --git a/comfy_api_nodes/nodes_ltxv.py b/comfy_api_nodes/nodes_ltxv.py index 0b757a62b..7e61560dc 100644 --- a/comfy_api_nodes/nodes_ltxv.py +++ b/comfy_api_nodes/nodes_ltxv.py @@ -1,12 +1,9 @@ from io import BytesIO -from typing import Optional -import torch from pydantic import BaseModel, Field from typing_extensions import override -from comfy_api.input_impl import VideoFromFile -from comfy_api.latest import IO, ComfyExtension +from comfy_api.latest import IO, ComfyExtension, Input, InputImpl from comfy_api_nodes.util import ( ApiEndpoint, get_number_of_images, @@ -26,9 +23,9 @@ class ExecuteTaskRequest(BaseModel): model: str = Field(...) duration: int = Field(...) resolution: str = Field(...) - fps: Optional[int] = Field(25) - generate_audio: Optional[bool] = Field(True) - image_uri: Optional[str] = Field(None) + fps: int | None = Field(25) + generate_audio: bool | None = Field(True) + image_uri: str | None = Field(None) class TextToVideoNode(IO.ComfyNode): @@ -103,7 +100,7 @@ class TextToVideoNode(IO.ComfyNode): as_binary=True, max_retries=1, ) - return IO.NodeOutput(VideoFromFile(BytesIO(response))) + return IO.NodeOutput(InputImpl.VideoFromFile(BytesIO(response))) class ImageToVideoNode(IO.ComfyNode): @@ -153,7 +150,7 @@ class ImageToVideoNode(IO.ComfyNode): @classmethod async def execute( cls, - image: torch.Tensor, + image: Input.Image, model: str, prompt: str, duration: int, @@ -183,7 +180,7 @@ class ImageToVideoNode(IO.ComfyNode): as_binary=True, max_retries=1, ) - return IO.NodeOutput(VideoFromFile(BytesIO(response))) + return IO.NodeOutput(InputImpl.VideoFromFile(BytesIO(response))) class LtxvApiExtension(ComfyExtension): diff --git a/comfy_api_nodes/nodes_moonvalley.py b/comfy_api_nodes/nodes_moonvalley.py index 7c31d95b3..2771e4790 100644 --- a/comfy_api_nodes/nodes_moonvalley.py +++ b/comfy_api_nodes/nodes_moonvalley.py @@ -1,11 +1,8 @@ import logging -from typing import Optional -import torch from typing_extensions import override -from comfy_api.input import VideoInput -from comfy_api.latest import IO, ComfyExtension +from comfy_api.latest import IO, ComfyExtension, Input from comfy_api_nodes.apis import ( MoonvalleyPromptResponse, MoonvalleyTextToVideoInferenceParams, @@ -61,7 +58,7 @@ def validate_task_creation_response(response) -> None: raise RuntimeError(error_msg) -def validate_video_to_video_input(video: VideoInput) -> VideoInput: +def validate_video_to_video_input(video: Input.Video) -> Input.Video: """ Validates and processes video input for Moonvalley Video-to-Video generation. @@ -82,7 +79,7 @@ def validate_video_to_video_input(video: VideoInput) -> VideoInput: return _validate_and_trim_duration(video) -def _get_video_dimensions(video: VideoInput) -> tuple[int, int]: +def _get_video_dimensions(video: Input.Video) -> tuple[int, int]: """Extracts video dimensions with error handling.""" try: return video.get_dimensions() @@ -106,7 +103,7 @@ def _validate_video_dimensions(width: int, height: int) -> None: raise ValueError(f"Resolution {width}x{height} not supported. Supported: {supported_list}") -def _validate_and_trim_duration(video: VideoInput) -> VideoInput: +def _validate_and_trim_duration(video: Input.Video) -> Input.Video: """Validates video duration and trims to 5 seconds if needed.""" duration = video.get_duration() _validate_minimum_duration(duration) @@ -119,7 +116,7 @@ def _validate_minimum_duration(duration: float) -> None: raise ValueError("Input video must be at least 5 seconds long.") -def _trim_if_too_long(video: VideoInput, duration: float) -> VideoInput: +def _trim_if_too_long(video: Input.Video, duration: float) -> Input.Video: """Trims video to 5 seconds if longer.""" if duration > 5: return trim_video(video, 5) @@ -241,7 +238,7 @@ class MoonvalleyImg2VideoNode(IO.ComfyNode): @classmethod async def execute( cls, - image: torch.Tensor, + image: Input.Image, prompt: str, negative_prompt: str, resolution: str, @@ -362,9 +359,9 @@ class MoonvalleyVideo2VideoNode(IO.ComfyNode): prompt: str, negative_prompt: str, seed: int, - video: Optional[VideoInput] = None, + video: Input.Video | None = None, control_type: str = "Motion Transfer", - motion_intensity: Optional[int] = 100, + motion_intensity: int | None = 100, steps=33, prompt_adherence=4.5, ) -> IO.NodeOutput: diff --git a/comfy_api_nodes/nodes_runway.py b/comfy_api_nodes/nodes_runway.py index 2fdafbbfe..3c55039c9 100644 --- a/comfy_api_nodes/nodes_runway.py +++ b/comfy_api_nodes/nodes_runway.py @@ -11,12 +11,11 @@ User Guides: """ -from typing import Union, Optional -from typing_extensions import override from enum import Enum -import torch +from typing_extensions import override +from comfy_api.latest import IO, ComfyExtension, Input, InputImpl from comfy_api_nodes.apis import ( RunwayImageToVideoRequest, RunwayImageToVideoResponse, @@ -44,8 +43,6 @@ from comfy_api_nodes.util import ( sync_op, poll_op, ) -from comfy_api.input_impl import VideoFromFile -from comfy_api.latest import ComfyExtension, IO PATH_IMAGE_TO_VIDEO = "/proxy/runway/image_to_video" PATH_TEXT_TO_IMAGE = "/proxy/runway/text_to_image" @@ -80,7 +77,7 @@ class RunwayGen3aAspectRatio(str, Enum): field_1280_768 = "1280:768" -def get_video_url_from_task_status(response: TaskStatusResponse) -> Union[str, None]: +def get_video_url_from_task_status(response: TaskStatusResponse) -> str | None: """Returns the video URL from the task status response if it exists.""" if hasattr(response, "output") and len(response.output) > 0: return response.output[0] @@ -89,13 +86,13 @@ def get_video_url_from_task_status(response: TaskStatusResponse) -> Union[str, N def extract_progress_from_task_status( response: TaskStatusResponse, -) -> Union[float, None]: +) -> float | None: if hasattr(response, "progress") and response.progress is not None: return response.progress * 100 return None -def get_image_url_from_task_status(response: TaskStatusResponse) -> Union[str, None]: +def get_image_url_from_task_status(response: TaskStatusResponse) -> str | None: """Returns the image URL from the task status response if it exists.""" if hasattr(response, "output") and len(response.output) > 0: return response.output[0] @@ -103,7 +100,7 @@ def get_image_url_from_task_status(response: TaskStatusResponse) -> Union[str, N async def get_response( - cls: type[IO.ComfyNode], task_id: str, estimated_duration: Optional[int] = None + cls: type[IO.ComfyNode], task_id: str, estimated_duration: int | None = None ) -> TaskStatusResponse: """Poll the task status until it is finished then get the response.""" return await poll_op( @@ -119,8 +116,8 @@ async def get_response( async def generate_video( cls: type[IO.ComfyNode], request: RunwayImageToVideoRequest, - estimated_duration: Optional[int] = None, -) -> VideoFromFile: + estimated_duration: int | None = None, +) -> InputImpl.VideoFromFile: initial_response = await sync_op( cls, endpoint=ApiEndpoint(path=PATH_IMAGE_TO_VIDEO, method="POST"), @@ -193,7 +190,7 @@ class RunwayImageToVideoNodeGen3a(IO.ComfyNode): async def execute( cls, prompt: str, - start_frame: torch.Tensor, + start_frame: Input.Image, duration: str, ratio: str, seed: int, @@ -283,7 +280,7 @@ class RunwayImageToVideoNodeGen4(IO.ComfyNode): async def execute( cls, prompt: str, - start_frame: torch.Tensor, + start_frame: Input.Image, duration: str, ratio: str, seed: int, @@ -381,8 +378,8 @@ class RunwayFirstLastFrameNode(IO.ComfyNode): async def execute( cls, prompt: str, - start_frame: torch.Tensor, - end_frame: torch.Tensor, + start_frame: Input.Image, + end_frame: Input.Image, duration: str, ratio: str, seed: int, @@ -467,7 +464,7 @@ class RunwayTextToImageNode(IO.ComfyNode): cls, prompt: str, ratio: str, - reference_image: Optional[torch.Tensor] = None, + reference_image: Input.Image | None = None, ) -> IO.NodeOutput: validate_string(prompt, min_length=1) diff --git a/comfy_api_nodes/nodes_veo2.py b/comfy_api_nodes/nodes_veo2.py index a54dc13ab..e165b8380 100644 --- a/comfy_api_nodes/nodes_veo2.py +++ b/comfy_api_nodes/nodes_veo2.py @@ -1,11 +1,9 @@ import base64 from io import BytesIO -import torch from typing_extensions import override -from comfy_api.input_impl.video_types import VideoFromFile -from comfy_api.latest import IO, ComfyExtension +from comfy_api.latest import IO, ComfyExtension, Input, InputImpl from comfy_api_nodes.apis.veo_api import ( VeoGenVidPollRequest, VeoGenVidPollResponse, @@ -232,7 +230,7 @@ class VeoVideoGenerationNode(IO.ComfyNode): # Check if video is provided as base64 or URL if hasattr(video, "bytesBase64Encoded") and video.bytesBase64Encoded: - return IO.NodeOutput(VideoFromFile(BytesIO(base64.b64decode(video.bytesBase64Encoded)))) + return IO.NodeOutput(InputImpl.VideoFromFile(BytesIO(base64.b64decode(video.bytesBase64Encoded)))) if hasattr(video, "gcsUri") and video.gcsUri: return IO.NodeOutput(await download_url_to_video_output(video.gcsUri)) @@ -431,8 +429,8 @@ class Veo3FirstLastFrameNode(IO.ComfyNode): aspect_ratio: str, duration: int, seed: int, - first_frame: torch.Tensor, - last_frame: torch.Tensor, + first_frame: Input.Image, + last_frame: Input.Image, model: str, generate_audio: bool, ): @@ -493,7 +491,7 @@ class Veo3FirstLastFrameNode(IO.ComfyNode): if response.videos: video = response.videos[0] if video.bytesBase64Encoded: - return IO.NodeOutput(VideoFromFile(BytesIO(base64.b64decode(video.bytesBase64Encoded)))) + return IO.NodeOutput(InputImpl.VideoFromFile(BytesIO(base64.b64decode(video.bytesBase64Encoded)))) if video.gcsUri: return IO.NodeOutput(await download_url_to_video_output(video.gcsUri)) raise Exception("Video returned but no data or URL was provided") diff --git a/comfy_extras/nodes_video.py b/comfy_extras/nodes_video.py index 6cf6e39bf..c609e03da 100644 --- a/comfy_extras/nodes_video.py +++ b/comfy_extras/nodes_video.py @@ -8,10 +8,7 @@ import json from typing import Optional from typing_extensions import override from fractions import Fraction -from comfy_api.input import AudioInput, ImageInput, VideoInput -from comfy_api.input_impl import VideoFromComponents, VideoFromFile -from comfy_api.util import VideoCodec, VideoComponents, VideoContainer -from comfy_api.latest import ComfyExtension, io, ui +from comfy_api.latest import ComfyExtension, io, ui, Input, InputImpl, Types from comfy.cli_args import args class SaveWEBM(io.ComfyNode): @@ -28,7 +25,6 @@ class SaveWEBM(io.ComfyNode): io.Float.Input("fps", default=24.0, min=0.01, max=1000.0, step=0.01), io.Float.Input("crf", default=32.0, min=0, max=63.0, step=1, tooltip="Higher crf means lower quality with a smaller file size, lower crf means higher quality higher filesize."), ], - outputs=[], hidden=[io.Hidden.prompt, io.Hidden.extra_pnginfo], is_output_node=True, ) @@ -79,16 +75,15 @@ class SaveVideo(io.ComfyNode): inputs=[ io.Video.Input("video", tooltip="The video to save."), io.String.Input("filename_prefix", default="video/ComfyUI", tooltip="The prefix for the file to save. This may include formatting information such as %date:yyyy-MM-dd% or %Empty Latent Image.width% to include values from nodes."), - io.Combo.Input("format", options=VideoContainer.as_input(), default="auto", tooltip="The format to save the video as."), - io.Combo.Input("codec", options=VideoCodec.as_input(), default="auto", tooltip="The codec to use for the video."), + io.Combo.Input("format", options=Types.VideoContainer.as_input(), default="auto", tooltip="The format to save the video as."), + io.Combo.Input("codec", options=Types.VideoCodec.as_input(), default="auto", tooltip="The codec to use for the video."), ], - outputs=[], hidden=[io.Hidden.prompt, io.Hidden.extra_pnginfo], is_output_node=True, ) @classmethod - def execute(cls, video: VideoInput, filename_prefix, format: str, codec) -> io.NodeOutput: + def execute(cls, video: Input.Video, filename_prefix, format: str, codec) -> io.NodeOutput: width, height = video.get_dimensions() full_output_folder, filename, counter, subfolder, filename_prefix = folder_paths.get_save_image_path( filename_prefix, @@ -105,10 +100,10 @@ class SaveVideo(io.ComfyNode): metadata["prompt"] = cls.hidden.prompt if len(metadata) > 0: saved_metadata = metadata - file = f"{filename}_{counter:05}_.{VideoContainer.get_extension(format)}" + file = f"{filename}_{counter:05}_.{Types.VideoContainer.get_extension(format)}" video.save_to( os.path.join(full_output_folder, file), - format=VideoContainer(format), + format=Types.VideoContainer(format), codec=codec, metadata=saved_metadata ) @@ -135,9 +130,9 @@ class CreateVideo(io.ComfyNode): ) @classmethod - def execute(cls, images: ImageInput, fps: float, audio: Optional[AudioInput] = None) -> io.NodeOutput: + def execute(cls, images: Input.Image, fps: float, audio: Optional[Input.Audio] = None) -> io.NodeOutput: return io.NodeOutput( - VideoFromComponents(VideoComponents(images=images, audio=audio, frame_rate=Fraction(fps))) + InputImpl.VideoFromComponents(Types.VideoComponents(images=images, audio=audio, frame_rate=Fraction(fps))) ) class GetVideoComponents(io.ComfyNode): @@ -159,11 +154,11 @@ class GetVideoComponents(io.ComfyNode): ) @classmethod - def execute(cls, video: VideoInput) -> io.NodeOutput: + def execute(cls, video: Input.Video) -> io.NodeOutput: components = video.get_components() - return io.NodeOutput(components.images, components.audio, float(components.frame_rate)) + class LoadVideo(io.ComfyNode): @classmethod def define_schema(cls): @@ -185,7 +180,7 @@ class LoadVideo(io.ComfyNode): @classmethod def execute(cls, file) -> io.NodeOutput: video_path = folder_paths.get_annotated_filepath(file) - return io.NodeOutput(VideoFromFile(video_path)) + return io.NodeOutput(InputImpl.VideoFromFile(video_path)) @classmethod def fingerprint_inputs(s, file): From c3c6313fc7b24a5811efde7cfe10b7cbbea52663 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Mon, 8 Dec 2025 11:28:17 +0200 Subject: [PATCH 249/647] Added "system_prompt" input to Gemini nodes (#11177) --- comfy_api_nodes/apis/gemini_api.py | 10 +----- comfy_api_nodes/nodes_gemini.py | 52 ++++++++++++++++++++++++++++-- 2 files changed, 51 insertions(+), 11 deletions(-) diff --git a/comfy_api_nodes/apis/gemini_api.py b/comfy_api_nodes/apis/gemini_api.py index a380ecc86..f8edc38c9 100644 --- a/comfy_api_nodes/apis/gemini_api.py +++ b/comfy_api_nodes/apis/gemini_api.py @@ -84,15 +84,7 @@ class GeminiSystemInstructionContent(BaseModel): description="A list of ordered parts that make up a single message. " "Different parts may have different IANA MIME types.", ) - role: GeminiRole = Field( - ..., - description="The identity of the entity that creates the message. " - "The following values are supported: " - "user: This indicates that the message is sent by a real person, typically a user-generated message. " - "model: This indicates that the message is generated by the model. " - "The model value is used to insert messages from model into the conversation during multi-turn conversations. " - "For non-multi-turn conversations, this field can be left blank or unset.", - ) + role: GeminiRole | None = Field(..., description="The role field of systemInstruction may be ignored.") class GeminiFunctionDeclaration(BaseModel): diff --git a/comfy_api_nodes/nodes_gemini.py b/comfy_api_nodes/nodes_gemini.py index 0b7422ef7..ad0f4b4d1 100644 --- a/comfy_api_nodes/nodes_gemini.py +++ b/comfy_api_nodes/nodes_gemini.py @@ -26,6 +26,8 @@ from comfy_api_nodes.apis.gemini_api import ( GeminiMimeType, GeminiPart, GeminiRole, + GeminiSystemInstructionContent, + GeminiTextPart, Modality, ) from comfy_api_nodes.util import ( @@ -42,6 +44,14 @@ from comfy_api_nodes.util import ( GEMINI_BASE_ENDPOINT = "/proxy/vertexai/gemini" GEMINI_MAX_INPUT_FILE_SIZE = 20 * 1024 * 1024 # 20 MB +GEMINI_IMAGE_SYS_PROMPT = ( + "You are an expert image-generation engine. You must ALWAYS produce an image.\n" + "Interpret all user input—regardless of " + "format, intent, or abstraction—as literal visual directives for image composition.\n" + "If a prompt is conversational or lacks specific visual details, " + "you must creatively invent a concrete visual scenario that depicts the concept.\n" + "Prioritize generating the visual representation above any text, formatting, or conversational requests." +) class GeminiModel(str, Enum): @@ -276,6 +286,13 @@ class GeminiNode(IO.ComfyNode): tooltip="Optional file(s) to use as context for the model. " "Accepts inputs from the Gemini Generate Content Input Files node.", ), + IO.String.Input( + "system_prompt", + multiline=True, + default="", + optional=True, + tooltip="Foundational instructions that dictate an AI's behavior.", + ), ], outputs=[ IO.String.Output(), @@ -348,6 +365,7 @@ class GeminiNode(IO.ComfyNode): audio: Input.Audio | None = None, video: Input.Video | None = None, files: list[GeminiPart] | None = None, + system_prompt: str = "", ) -> IO.NodeOutput: validate_string(prompt, strip_whitespace=False) @@ -364,7 +382,10 @@ class GeminiNode(IO.ComfyNode): if files is not None: parts.extend(files) - # Create response + gemini_system_prompt = None + if system_prompt: + gemini_system_prompt = GeminiSystemInstructionContent(parts=[GeminiTextPart(text=system_prompt)], role=None) + response = await sync_op( cls, endpoint=ApiEndpoint(path=f"{GEMINI_BASE_ENDPOINT}/{model}", method="POST"), @@ -374,7 +395,8 @@ class GeminiNode(IO.ComfyNode): role=GeminiRole.user, parts=parts, ) - ] + ], + systemInstruction=gemini_system_prompt, ), response_model=GeminiGenerateContentResponse, price_extractor=calculate_tokens_price, @@ -524,6 +546,13 @@ class GeminiImage(IO.ComfyNode): "'IMAGE+TEXT' to return both the generated image and a text response.", optional=True, ), + IO.String.Input( + "system_prompt", + multiline=True, + default=GEMINI_IMAGE_SYS_PROMPT, + optional=True, + tooltip="Foundational instructions that dictate an AI's behavior.", + ), ], outputs=[ IO.Image.Output(), @@ -547,6 +576,7 @@ class GeminiImage(IO.ComfyNode): files: list[GeminiPart] | None = None, aspect_ratio: str = "auto", response_modalities: str = "IMAGE+TEXT", + system_prompt: str = "", ) -> IO.NodeOutput: validate_string(prompt, strip_whitespace=True, min_length=1) parts: list[GeminiPart] = [GeminiPart(text=prompt)] @@ -560,6 +590,10 @@ class GeminiImage(IO.ComfyNode): if files is not None: parts.extend(files) + gemini_system_prompt = None + if system_prompt: + gemini_system_prompt = GeminiSystemInstructionContent(parts=[GeminiTextPart(text=system_prompt)], role=None) + response = await sync_op( cls, endpoint=ApiEndpoint(path=f"{GEMINI_BASE_ENDPOINT}/{model}", method="POST"), @@ -571,6 +605,7 @@ class GeminiImage(IO.ComfyNode): responseModalities=(["IMAGE"] if response_modalities == "IMAGE" else ["TEXT", "IMAGE"]), imageConfig=None if aspect_ratio == "auto" else image_config, ), + systemInstruction=gemini_system_prompt, ), response_model=GeminiGenerateContentResponse, price_extractor=calculate_tokens_price, @@ -641,6 +676,13 @@ class GeminiImage2(IO.ComfyNode): tooltip="Optional file(s) to use as context for the model. " "Accepts inputs from the Gemini Generate Content Input Files node.", ), + IO.String.Input( + "system_prompt", + multiline=True, + default=GEMINI_IMAGE_SYS_PROMPT, + optional=True, + tooltip="Foundational instructions that dictate an AI's behavior.", + ), ], outputs=[ IO.Image.Output(), @@ -665,6 +707,7 @@ class GeminiImage2(IO.ComfyNode): response_modalities: str, images: Input.Image | None = None, files: list[GeminiPart] | None = None, + system_prompt: str = "", ) -> IO.NodeOutput: validate_string(prompt, strip_whitespace=True, min_length=1) @@ -680,6 +723,10 @@ class GeminiImage2(IO.ComfyNode): if aspect_ratio != "auto": image_config.aspectRatio = aspect_ratio + gemini_system_prompt = None + if system_prompt: + gemini_system_prompt = GeminiSystemInstructionContent(parts=[GeminiTextPart(text=system_prompt)], role=None) + response = await sync_op( cls, ApiEndpoint(path=f"{GEMINI_BASE_ENDPOINT}/{model}", method="POST"), @@ -691,6 +738,7 @@ class GeminiImage2(IO.ComfyNode): responseModalities=(["IMAGE"] if response_modalities == "IMAGE" else ["TEXT", "IMAGE"]), imageConfig=image_config, ), + systemInstruction=gemini_system_prompt, ), response_model=GeminiGenerateContentResponse, price_extractor=calculate_tokens_price, From fd271dedfde6e192a1f1a025521070876e89e04a Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Mon, 8 Dec 2025 11:33:46 +0200 Subject: [PATCH 250/647] [API Nodes] add support for seedance-1-0-pro-fast model (#10947) * feat(api-nodes): add support for seedance-1-0-pro-fast model * feat(api-nodes): add support for seedream-4.5 model --- comfy_api_nodes/apis/bytedance_api.py | 144 +++++++++++++++ comfy_api_nodes/nodes_bytedance.py | 255 ++++++-------------------- 2 files changed, 196 insertions(+), 203 deletions(-) create mode 100644 comfy_api_nodes/apis/bytedance_api.py diff --git a/comfy_api_nodes/apis/bytedance_api.py b/comfy_api_nodes/apis/bytedance_api.py new file mode 100644 index 000000000..77cd76f9b --- /dev/null +++ b/comfy_api_nodes/apis/bytedance_api.py @@ -0,0 +1,144 @@ +from typing import Literal + +from pydantic import BaseModel, Field + + +class Text2ImageTaskCreationRequest(BaseModel): + model: str = Field(...) + prompt: str = Field(...) + response_format: str | None = Field("url") + size: str | None = Field(None) + seed: int | None = Field(0, ge=0, le=2147483647) + guidance_scale: float | None = Field(..., ge=1.0, le=10.0) + watermark: bool | None = Field(True) + + +class Image2ImageTaskCreationRequest(BaseModel): + model: str = Field(...) + prompt: str = Field(...) + response_format: str | None = Field("url") + image: str = Field(..., description="Base64 encoded string or image URL") + size: str | None = Field("adaptive") + seed: int | None = Field(..., ge=0, le=2147483647) + guidance_scale: float | None = Field(..., ge=1.0, le=10.0) + watermark: bool | None = Field(True) + + +class Seedream4Options(BaseModel): + max_images: int = Field(15) + + +class Seedream4TaskCreationRequest(BaseModel): + model: str = Field(...) + prompt: str = Field(...) + response_format: str = Field("url") + image: list[str] | None = Field(None, description="Image URLs") + size: str = Field(...) + seed: int = Field(..., ge=0, le=2147483647) + sequential_image_generation: str = Field("disabled") + sequential_image_generation_options: Seedream4Options = Field(Seedream4Options(max_images=15)) + watermark: bool = Field(True) + + +class ImageTaskCreationResponse(BaseModel): + model: str = Field(...) + created: int = Field(..., description="Unix timestamp (in seconds) indicating time when the request was created.") + data: list = Field([], description="Contains information about the generated image(s).") + error: dict = Field({}, description="Contains `code` and `message` fields in case of error.") + + +class TaskTextContent(BaseModel): + type: str = Field("text") + text: str = Field(...) + + +class TaskImageContentUrl(BaseModel): + url: str = Field(...) + + +class TaskImageContent(BaseModel): + type: str = Field("image_url") + image_url: TaskImageContentUrl = Field(...) + role: Literal["first_frame", "last_frame", "reference_image"] | None = Field(None) + + +class Text2VideoTaskCreationRequest(BaseModel): + model: str = Field(...) + content: list[TaskTextContent] = Field(..., min_length=1) + + +class Image2VideoTaskCreationRequest(BaseModel): + model: str = Field(...) + content: list[TaskTextContent | TaskImageContent] = Field(..., min_length=2) + + +class TaskCreationResponse(BaseModel): + id: str = Field(...) + + +class TaskStatusError(BaseModel): + code: str = Field(...) + message: str = Field(...) + + +class TaskStatusResult(BaseModel): + video_url: str = Field(...) + + +class TaskStatusResponse(BaseModel): + id: str = Field(...) + model: str = Field(...) + status: Literal["queued", "running", "cancelled", "succeeded", "failed"] = Field(...) + error: TaskStatusError | None = Field(None) + content: TaskStatusResult | None = Field(None) + + +RECOMMENDED_PRESETS = [ + ("1024x1024 (1:1)", 1024, 1024), + ("864x1152 (3:4)", 864, 1152), + ("1152x864 (4:3)", 1152, 864), + ("1280x720 (16:9)", 1280, 720), + ("720x1280 (9:16)", 720, 1280), + ("832x1248 (2:3)", 832, 1248), + ("1248x832 (3:2)", 1248, 832), + ("1512x648 (21:9)", 1512, 648), + ("2048x2048 (1:1)", 2048, 2048), + ("Custom", None, None), +] + +RECOMMENDED_PRESETS_SEEDREAM_4 = [ + ("2048x2048 (1:1)", 2048, 2048), + ("2304x1728 (4:3)", 2304, 1728), + ("1728x2304 (3:4)", 1728, 2304), + ("2560x1440 (16:9)", 2560, 1440), + ("1440x2560 (9:16)", 1440, 2560), + ("2496x1664 (3:2)", 2496, 1664), + ("1664x2496 (2:3)", 1664, 2496), + ("3024x1296 (21:9)", 3024, 1296), + ("4096x4096 (1:1)", 4096, 4096), + ("Custom", None, None), +] + +# The time in this dictionary are given for 10 seconds duration. +VIDEO_TASKS_EXECUTION_TIME = { + "seedance-1-0-lite-t2v-250428": { + "480p": 40, + "720p": 60, + "1080p": 90, + }, + "seedance-1-0-lite-i2v-250428": { + "480p": 40, + "720p": 60, + "1080p": 90, + }, + "seedance-1-0-pro-250528": { + "480p": 70, + "720p": 85, + "1080p": 115, + }, + "seedance-1-0-pro-fast-251015": { + "480p": 50, + "720p": 65, + "1080p": 100, + }, +} diff --git a/comfy_api_nodes/nodes_bytedance.py b/comfy_api_nodes/nodes_bytedance.py index caced471e..57c0218d0 100644 --- a/comfy_api_nodes/nodes_bytedance.py +++ b/comfy_api_nodes/nodes_bytedance.py @@ -1,13 +1,27 @@ import logging import math -from enum import Enum -from typing import Literal, Optional, Union import torch -from pydantic import BaseModel, Field from typing_extensions import override -from comfy_api.latest import IO, ComfyExtension +from comfy_api.latest import IO, ComfyExtension, Input +from comfy_api_nodes.apis.bytedance_api import ( + RECOMMENDED_PRESETS, + RECOMMENDED_PRESETS_SEEDREAM_4, + VIDEO_TASKS_EXECUTION_TIME, + Image2ImageTaskCreationRequest, + Image2VideoTaskCreationRequest, + ImageTaskCreationResponse, + Seedream4Options, + Seedream4TaskCreationRequest, + TaskCreationResponse, + TaskImageContent, + TaskImageContentUrl, + TaskStatusResponse, + TaskTextContent, + Text2ImageTaskCreationRequest, + Text2VideoTaskCreationRequest, +) from comfy_api_nodes.util import ( ApiEndpoint, download_url_to_image_tensor, @@ -29,162 +43,6 @@ BYTEPLUS_TASK_ENDPOINT = "/proxy/byteplus/api/v3/contents/generations/tasks" BYTEPLUS_TASK_STATUS_ENDPOINT = "/proxy/byteplus/api/v3/contents/generations/tasks" # + /{task_id} -class Text2ImageModelName(str, Enum): - seedream_3 = "seedream-3-0-t2i-250415" - - -class Image2ImageModelName(str, Enum): - seededit_3 = "seededit-3-0-i2i-250628" - - -class Text2VideoModelName(str, Enum): - seedance_1_pro = "seedance-1-0-pro-250528" - seedance_1_lite = "seedance-1-0-lite-t2v-250428" - - -class Image2VideoModelName(str, Enum): - """note(August 31): Pro model only supports FirstFrame: https://docs.byteplus.com/en/docs/ModelArk/1520757""" - - seedance_1_pro = "seedance-1-0-pro-250528" - seedance_1_lite = "seedance-1-0-lite-i2v-250428" - - -class Text2ImageTaskCreationRequest(BaseModel): - model: Text2ImageModelName = Text2ImageModelName.seedream_3 - prompt: str = Field(...) - response_format: Optional[str] = Field("url") - size: Optional[str] = Field(None) - seed: Optional[int] = Field(0, ge=0, le=2147483647) - guidance_scale: Optional[float] = Field(..., ge=1.0, le=10.0) - watermark: Optional[bool] = Field(True) - - -class Image2ImageTaskCreationRequest(BaseModel): - model: Image2ImageModelName = Image2ImageModelName.seededit_3 - prompt: str = Field(...) - response_format: Optional[str] = Field("url") - image: str = Field(..., description="Base64 encoded string or image URL") - size: Optional[str] = Field("adaptive") - seed: Optional[int] = Field(..., ge=0, le=2147483647) - guidance_scale: Optional[float] = Field(..., ge=1.0, le=10.0) - watermark: Optional[bool] = Field(True) - - -class Seedream4Options(BaseModel): - max_images: int = Field(15) - - -class Seedream4TaskCreationRequest(BaseModel): - model: str = Field("seedream-4-0-250828") - prompt: str = Field(...) - response_format: str = Field("url") - image: Optional[list[str]] = Field(None, description="Image URLs") - size: str = Field(...) - seed: int = Field(..., ge=0, le=2147483647) - sequential_image_generation: str = Field("disabled") - sequential_image_generation_options: Seedream4Options = Field(Seedream4Options(max_images=15)) - watermark: bool = Field(True) - - -class ImageTaskCreationResponse(BaseModel): - model: str = Field(...) - created: int = Field(..., description="Unix timestamp (in seconds) indicating time when the request was created.") - data: list = Field([], description="Contains information about the generated image(s).") - error: dict = Field({}, description="Contains `code` and `message` fields in case of error.") - - -class TaskTextContent(BaseModel): - type: str = Field("text") - text: str = Field(...) - - -class TaskImageContentUrl(BaseModel): - url: str = Field(...) - - -class TaskImageContent(BaseModel): - type: str = Field("image_url") - image_url: TaskImageContentUrl = Field(...) - role: Optional[Literal["first_frame", "last_frame", "reference_image"]] = Field(None) - - -class Text2VideoTaskCreationRequest(BaseModel): - model: Text2VideoModelName = Text2VideoModelName.seedance_1_pro - content: list[TaskTextContent] = Field(..., min_length=1) - - -class Image2VideoTaskCreationRequest(BaseModel): - model: Image2VideoModelName = Image2VideoModelName.seedance_1_pro - content: list[Union[TaskTextContent, TaskImageContent]] = Field(..., min_length=2) - - -class TaskCreationResponse(BaseModel): - id: str = Field(...) - - -class TaskStatusError(BaseModel): - code: str = Field(...) - message: str = Field(...) - - -class TaskStatusResult(BaseModel): - video_url: str = Field(...) - - -class TaskStatusResponse(BaseModel): - id: str = Field(...) - model: str = Field(...) - status: Literal["queued", "running", "cancelled", "succeeded", "failed"] = Field(...) - error: Optional[TaskStatusError] = Field(None) - content: Optional[TaskStatusResult] = Field(None) - - -RECOMMENDED_PRESETS = [ - ("1024x1024 (1:1)", 1024, 1024), - ("864x1152 (3:4)", 864, 1152), - ("1152x864 (4:3)", 1152, 864), - ("1280x720 (16:9)", 1280, 720), - ("720x1280 (9:16)", 720, 1280), - ("832x1248 (2:3)", 832, 1248), - ("1248x832 (3:2)", 1248, 832), - ("1512x648 (21:9)", 1512, 648), - ("2048x2048 (1:1)", 2048, 2048), - ("Custom", None, None), -] - -RECOMMENDED_PRESETS_SEEDREAM_4 = [ - ("2048x2048 (1:1)", 2048, 2048), - ("2304x1728 (4:3)", 2304, 1728), - ("1728x2304 (3:4)", 1728, 2304), - ("2560x1440 (16:9)", 2560, 1440), - ("1440x2560 (9:16)", 1440, 2560), - ("2496x1664 (3:2)", 2496, 1664), - ("1664x2496 (2:3)", 1664, 2496), - ("3024x1296 (21:9)", 3024, 1296), - ("4096x4096 (1:1)", 4096, 4096), - ("Custom", None, None), -] - -# The time in this dictionary are given for 10 seconds duration. -VIDEO_TASKS_EXECUTION_TIME = { - "seedance-1-0-lite-t2v-250428": { - "480p": 40, - "720p": 60, - "1080p": 90, - }, - "seedance-1-0-lite-i2v-250428": { - "480p": 40, - "720p": 60, - "1080p": 90, - }, - "seedance-1-0-pro-250528": { - "480p": 70, - "720p": 85, - "1080p": 115, - }, -} - - def get_image_url_from_response(response: ImageTaskCreationResponse) -> str: if response.error: error_msg = f"ByteDance request failed. Code: {response.error['code']}, message: {response.error['message']}" @@ -194,13 +52,6 @@ def get_image_url_from_response(response: ImageTaskCreationResponse) -> str: return response.data[0]["url"] -def get_video_url_from_task_status(response: TaskStatusResponse) -> Union[str, None]: - """Returns the video URL from the task status response if it exists.""" - if hasattr(response, "content") and response.content: - return response.content.video_url - return None - - class ByteDanceImageNode(IO.ComfyNode): @classmethod @@ -211,12 +62,7 @@ class ByteDanceImageNode(IO.ComfyNode): category="api node/image/ByteDance", description="Generate images using ByteDance models via api based on prompt", inputs=[ - IO.Combo.Input( - "model", - options=Text2ImageModelName, - default=Text2ImageModelName.seedream_3, - tooltip="Model name", - ), + IO.Combo.Input("model", options=["seedream-3-0-t2i-250415"]), IO.String.Input( "prompt", multiline=True, @@ -335,12 +181,7 @@ class ByteDanceImageEditNode(IO.ComfyNode): category="api node/image/ByteDance", description="Edit images using ByteDance models via api based on prompt", inputs=[ - IO.Combo.Input( - "model", - options=Image2ImageModelName, - default=Image2ImageModelName.seededit_3, - tooltip="Model name", - ), + IO.Combo.Input("model", options=["seededit-3-0-i2i-250628"]), IO.Image.Input( "image", tooltip="The base image to edit", @@ -394,7 +235,7 @@ class ByteDanceImageEditNode(IO.ComfyNode): async def execute( cls, model: str, - image: torch.Tensor, + image: Input.Image, prompt: str, seed: int, guidance_scale: float, @@ -434,7 +275,7 @@ class ByteDanceSeedreamNode(IO.ComfyNode): inputs=[ IO.Combo.Input( "model", - options=["seedream-4-0-250828"], + options=["seedream-4-5-251128", "seedream-4-0-250828"], tooltip="Model name", ), IO.String.Input( @@ -459,7 +300,7 @@ class ByteDanceSeedreamNode(IO.ComfyNode): default=2048, min=1024, max=4096, - step=64, + step=8, tooltip="Custom width for image. Value is working only if `size_preset` is set to `Custom`", optional=True, ), @@ -468,7 +309,7 @@ class ByteDanceSeedreamNode(IO.ComfyNode): default=2048, min=1024, max=4096, - step=64, + step=8, tooltip="Custom height for image. Value is working only if `size_preset` is set to `Custom`", optional=True, ), @@ -532,7 +373,7 @@ class ByteDanceSeedreamNode(IO.ComfyNode): cls, model: str, prompt: str, - image: torch.Tensor = None, + image: Input.Image | None = None, size_preset: str = RECOMMENDED_PRESETS_SEEDREAM_4[0][0], width: int = 2048, height: int = 2048, @@ -555,6 +396,18 @@ class ByteDanceSeedreamNode(IO.ComfyNode): raise ValueError( f"Custom size out of range: {w}x{h}. " "Both width and height must be between 1024 and 4096 pixels." ) + out_num_pixels = w * h + mp_provided = out_num_pixels / 1_000_000.0 + if "seedream-4-5" in model and out_num_pixels < 3686400: + raise ValueError( + f"Minimum image resolution that Seedream 4.5 can generate is 3.68MP, " + f"but {mp_provided:.2f}MP provided." + ) + if "seedream-4-0" in model and out_num_pixels < 921600: + raise ValueError( + f"Minimum image resolution that the selected model can generate is 0.92MP, " + f"but {mp_provided:.2f}MP provided." + ) n_input_images = get_number_of_images(image) if image is not None else 0 if n_input_images > 10: raise ValueError(f"Maximum of 10 reference images are supported, but {n_input_images} received.") @@ -607,9 +460,8 @@ class ByteDanceTextToVideoNode(IO.ComfyNode): inputs=[ IO.Combo.Input( "model", - options=Text2VideoModelName, - default=Text2VideoModelName.seedance_1_pro, - tooltip="Model name", + options=["seedance-1-0-pro-250528", "seedance-1-0-lite-t2v-250428", "seedance-1-0-pro-fast-251015"], + default="seedance-1-0-pro-fast-251015", ), IO.String.Input( "prompt", @@ -714,9 +566,8 @@ class ByteDanceImageToVideoNode(IO.ComfyNode): inputs=[ IO.Combo.Input( "model", - options=Image2VideoModelName, - default=Image2VideoModelName.seedance_1_pro, - tooltip="Model name", + options=["seedance-1-0-pro-250528", "seedance-1-0-lite-t2v-250428", "seedance-1-0-pro-fast-251015"], + default="seedance-1-0-pro-fast-251015", ), IO.String.Input( "prompt", @@ -787,7 +638,7 @@ class ByteDanceImageToVideoNode(IO.ComfyNode): cls, model: str, prompt: str, - image: torch.Tensor, + image: Input.Image, resolution: str, aspect_ratio: str, duration: int, @@ -833,9 +684,8 @@ class ByteDanceFirstLastFrameNode(IO.ComfyNode): inputs=[ IO.Combo.Input( "model", - options=[model.value for model in Image2VideoModelName], - default=Image2VideoModelName.seedance_1_lite.value, - tooltip="Model name", + options=["seedance-1-0-pro-250528", "seedance-1-0-lite-i2v-250428"], + default="seedance-1-0-lite-i2v-250428", ), IO.String.Input( "prompt", @@ -910,8 +760,8 @@ class ByteDanceFirstLastFrameNode(IO.ComfyNode): cls, model: str, prompt: str, - first_frame: torch.Tensor, - last_frame: torch.Tensor, + first_frame: Input.Image, + last_frame: Input.Image, resolution: str, aspect_ratio: str, duration: int, @@ -968,9 +818,8 @@ class ByteDanceImageReferenceNode(IO.ComfyNode): inputs=[ IO.Combo.Input( "model", - options=[Image2VideoModelName.seedance_1_lite.value], - default=Image2VideoModelName.seedance_1_lite.value, - tooltip="Model name", + options=["seedance-1-0-pro-250528", "seedance-1-0-lite-i2v-250428"], + default="seedance-1-0-lite-i2v-250428", ), IO.String.Input( "prompt", @@ -1034,7 +883,7 @@ class ByteDanceImageReferenceNode(IO.ComfyNode): cls, model: str, prompt: str, - images: torch.Tensor, + images: Input.Image, resolution: str, aspect_ratio: str, duration: int, @@ -1069,8 +918,8 @@ class ByteDanceImageReferenceNode(IO.ComfyNode): async def process_video_task( cls: type[IO.ComfyNode], - payload: Union[Text2VideoTaskCreationRequest, Image2VideoTaskCreationRequest], - estimated_duration: Optional[int], + payload: Text2VideoTaskCreationRequest | Image2VideoTaskCreationRequest, + estimated_duration: int | None, ) -> IO.NodeOutput: initial_response = await sync_op( cls, @@ -1085,7 +934,7 @@ async def process_video_task( estimated_duration=estimated_duration, response_model=TaskStatusResponse, ) - return IO.NodeOutput(await download_url_to_video_output(get_video_url_from_task_status(response))) + return IO.NodeOutput(await download_url_to_video_output(response.content.video_url)) def raise_if_text_params(prompt: str, text_params: list[str]) -> None: From 8e889c535d1fc407bf27dbf8359eef9580f2ed60 Mon Sep 17 00:00:00 2001 From: dxqb <183307934+dxqb@users.noreply.github.com> Date: Mon, 8 Dec 2025 21:17:26 +0100 Subject: [PATCH 251/647] Support "transformer." LoRA prefix for Z-Image (#11135) --- comfy/lora.py | 1 + 1 file changed, 1 insertion(+) diff --git a/comfy/lora.py b/comfy/lora.py index e7202ce97..2ed0acb9d 100644 --- a/comfy/lora.py +++ b/comfy/lora.py @@ -320,6 +320,7 @@ def model_lora_keys_unet(model, key_map={}): to = diffusers_keys[k] key_lora = k[:-len(".weight")] key_map["diffusion_model.{}".format(key_lora)] = to + key_map["transformer.{}".format(key_lora)] = to key_map["lycoris_{}".format(key_lora.replace(".", "_"))] = to if isinstance(model, comfy.model_base.Kandinsky5): From 60ee574748209a17ade1c7524e228be2802d1589 Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Tue, 9 Dec 2025 06:18:06 +1000 Subject: [PATCH 252/647] retune lowVramPatch VRAM accounting (#11173) In the lowvram case, this now does its math in the model dtype in the post de-quantization domain. Account for that. The patching was also put back on the compute stream getting it off-peak so relax the MATH_FACTOR to only x2 so get out of the worst-case assumption of everything peaking at once. --- comfy/model_patcher.py | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index 5b1ccb824..8b5edeb52 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -132,14 +132,14 @@ class LowVramPatch: def __call__(self, weight): return comfy.lora.calculate_weight(self.patches[self.key], weight, self.key, intermediate_dtype=weight.dtype) -#The above patch logic may cast up the weight to fp32, and do math. Go with fp32 x 3 -LOWVRAM_PATCH_ESTIMATE_MATH_FACTOR = 3 +LOWVRAM_PATCH_ESTIMATE_MATH_FACTOR = 2 def low_vram_patch_estimate_vram(model, key): weight, set_func, convert_func = get_key_weight(model, key) if weight is None: return 0 - return weight.numel() * torch.float32.itemsize * LOWVRAM_PATCH_ESTIMATE_MATH_FACTOR + model_dtype = getattr(model, "manual_cast_dtype", torch.float32) + return weight.numel() * model_dtype.itemsize * LOWVRAM_PATCH_ESTIMATE_MATH_FACTOR def get_key_weight(model, key): set_func = None From 935493f6c186de8808508713a465d6bda75e5ce4 Mon Sep 17 00:00:00 2001 From: ComfyUI Wiki Date: Tue, 9 Dec 2025 04:18:53 +0800 Subject: [PATCH 253/647] chore: update workflow templates to v0.7.54 (#11192) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 12a7c1089..4bd4b21c3 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,5 +1,5 @@ comfyui-frontend-package==1.33.10 -comfyui-workflow-templates==0.7.51 +comfyui-workflow-templates==0.7.54 comfyui-embedded-docs==0.3.1 torch torchsde From 3b0368aa34182fc7c97de92d59b609c77138def2 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 8 Dec 2025 14:38:36 -0800 Subject: [PATCH 254/647] Fix regression. (#11194) --- comfy/model_patcher.py | 3 +++ 1 file changed, 3 insertions(+) diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index 8b5edeb52..a7d24ac13 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -139,6 +139,9 @@ def low_vram_patch_estimate_vram(model, key): if weight is None: return 0 model_dtype = getattr(model, "manual_cast_dtype", torch.float32) + if model_dtype is None: + model_dtype = weight.dtype + return weight.numel() * model_dtype.itemsize * LOWVRAM_PATCH_ESTIMATE_MATH_FACTOR def get_key_weight(model, key): From d50f342c90802830c1178ad9d7f2783dc2821af1 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 8 Dec 2025 20:20:04 -0800 Subject: [PATCH 255/647] Fix potential issue. (#11201) --- comfy/model_patcher.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index a7d24ac13..2e8ce2613 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -923,7 +923,7 @@ class ModelPatcher: patch_counter += 1 cast_weight = True - if cast_weight: + if cast_weight and hasattr(m, "comfy_cast_weights"): m.prev_comfy_cast_weights = m.comfy_cast_weights m.comfy_cast_weights = True m.comfy_patched_weights = False From e136b6dbb0b08341388f5bf9a00b1fca29992eb3 Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Tue, 9 Dec 2025 14:21:31 +1000 Subject: [PATCH 256/647] dequantization offload accounting (fixes Flux2 OOMs - incl TEs) (#11171) * make setattr safe for non existent attributes Handle the case where the attribute doesnt exist by returning a static sentinel (distinct from None). If the sentinel is passed in as the set value, del the attr. * Account for dequantization and type-casts in offload costs When measuring the cost of offload, identify weights that need a type change or dequantization and add the size of the conversion result to the offload cost. This is mutually exclusive with lowvram patches which already has a large conservative estimate and wont overlap the dequant cost so\ dont double count. * Set the compute type on CLIP MPs So that the loader can know the size of weights for dequant accounting. --- comfy/model_patcher.py | 19 +++++++++++++------ comfy/sd.py | 2 ++ comfy/utils.py | 9 +++++++-- 3 files changed, 22 insertions(+), 8 deletions(-) diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index 2e8ce2613..a486c2723 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -35,6 +35,7 @@ import comfy.model_management import comfy.patcher_extension import comfy.utils from comfy.comfy_types import UnetWrapperFunction +from comfy.quant_ops import QuantizedTensor from comfy.patcher_extension import CallbacksMP, PatcherInjection, WrappersMP @@ -665,12 +666,18 @@ class ModelPatcher: module_mem = comfy.model_management.module_size(m) module_offload_mem = module_mem if hasattr(m, "comfy_cast_weights"): - weight_key = "{}.weight".format(n) - bias_key = "{}.bias".format(n) - if weight_key in self.patches: - module_offload_mem += low_vram_patch_estimate_vram(self.model, weight_key) - if bias_key in self.patches: - module_offload_mem += low_vram_patch_estimate_vram(self.model, bias_key) + def check_module_offload_mem(key): + if key in self.patches: + return low_vram_patch_estimate_vram(self.model, key) + model_dtype = getattr(self.model, "manual_cast_dtype", None) + weight, _, _ = get_key_weight(self.model, key) + if model_dtype is None or weight is None: + return 0 + if (weight.dtype != model_dtype or isinstance(weight, QuantizedTensor)): + return weight.numel() * model_dtype.itemsize + return 0 + module_offload_mem += check_module_offload_mem("{}.weight".format(n)) + module_offload_mem += check_module_offload_mem("{}.bias".format(n)) loading.append((module_offload_mem, module_mem, n, m, params)) return loading diff --git a/comfy/sd.py b/comfy/sd.py index 754b1703d..a16f2d14f 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -127,6 +127,8 @@ class CLIP: self.tokenizer = tokenizer(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data) self.patcher = comfy.model_patcher.ModelPatcher(self.cond_stage_model, load_device=load_device, offload_device=offload_device) + #Match torch.float32 hardcode upcast in TE implemention + self.patcher.set_model_compute_dtype(torch.float32) self.patcher.hook_mode = comfy.hooks.EnumHookMode.MinVram self.patcher.is_clip = True self.apply_hooks_to_conds = None diff --git a/comfy/utils.py b/comfy/utils.py index 89846bc95..9dc0d76ac 100644 --- a/comfy/utils.py +++ b/comfy/utils.py @@ -803,12 +803,17 @@ def safetensors_header(safetensors_path, max_size=100*1024*1024): return None return f.read(length_of_header) +ATTR_UNSET={} + def set_attr(obj, attr, value): attrs = attr.split(".") for name in attrs[:-1]: obj = getattr(obj, name) - prev = getattr(obj, attrs[-1]) - setattr(obj, attrs[-1], value) + prev = getattr(obj, attrs[-1], ATTR_UNSET) + if value is ATTR_UNSET: + delattr(obj, attrs[-1]) + else: + setattr(obj, attrs[-1], value) return prev def set_attr_param(obj, attr, value): From cabc4d351ff620ece87f18019d98131ebcbdf1aa Mon Sep 17 00:00:00 2001 From: Christian Byrne Date: Mon, 8 Dec 2025 20:22:02 -0800 Subject: [PATCH 257/647] bump comfyui-frontend-package to 1.33.13 (patch) (#11200) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 4bd4b21c3..11a7ac245 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,4 +1,4 @@ -comfyui-frontend-package==1.33.10 +comfyui-frontend-package==1.33.13 comfyui-workflow-templates==0.7.54 comfyui-embedded-docs==0.3.1 torch From b9fb542703085c58f082b4a822329fb6670e8016 Mon Sep 17 00:00:00 2001 From: Lodestone Date: Tue, 9 Dec 2025 11:33:29 +0700 Subject: [PATCH 258/647] add chroma-radiance-x0 mode (#11197) --- comfy/ldm/chroma_radiance/model.py | 20 ++++++++++++++++++-- comfy/model_detection.py | 2 ++ 2 files changed, 20 insertions(+), 2 deletions(-) diff --git a/comfy/ldm/chroma_radiance/model.py b/comfy/ldm/chroma_radiance/model.py index e643b4414..70d173889 100644 --- a/comfy/ldm/chroma_radiance/model.py +++ b/comfy/ldm/chroma_radiance/model.py @@ -37,7 +37,7 @@ class ChromaRadianceParams(ChromaParams): nerf_final_head_type: str # None means use the same dtype as the model. nerf_embedder_dtype: Optional[torch.dtype] - + use_x0: bool class ChromaRadiance(Chroma): """ @@ -159,6 +159,9 @@ class ChromaRadiance(Chroma): self.skip_dit = [] self.lite = False + if params.use_x0: + self.register_buffer("__x0__", torch.tensor([])) + @property def _nerf_final_layer(self) -> nn.Module: if self.params.nerf_final_head_type == "linear": @@ -276,6 +279,12 @@ class ChromaRadiance(Chroma): params_dict |= overrides return params.__class__(**params_dict) + def _apply_x0_residual(self, predicted, noisy, timesteps): + + # non zero during training to prevent 0 div + eps = 0.0 + return (noisy - predicted) / (timesteps.view(-1,1,1,1) + eps) + def _forward( self, x: Tensor, @@ -316,4 +325,11 @@ class ChromaRadiance(Chroma): transformer_options, attn_mask=kwargs.get("attention_mask", None), ) - return self.forward_nerf(img, img_out, params)[:, :, :h, :w] + + out = self.forward_nerf(img, img_out, params)[:, :, :h, :w] + + # If x0 variant → v-pred, just return this instead + if hasattr(self, "__x0__"): + out = self._apply_x0_residual(out, img, timestep) + return out + diff --git a/comfy/model_detection.py b/comfy/model_detection.py index 74c547427..19e6aa954 100644 --- a/comfy/model_detection.py +++ b/comfy/model_detection.py @@ -257,6 +257,8 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): dit_config["nerf_tile_size"] = 512 dit_config["nerf_final_head_type"] = "conv" if f"{key_prefix}nerf_final_layer_conv.norm.scale" in state_dict_keys else "linear" dit_config["nerf_embedder_dtype"] = torch.float32 + if "__x0__" in state_dict_keys: # x0 pred + dit_config["use_x0"] = True else: dit_config["guidance_embed"] = "{}guidance_in.in_layer.weight".format(key_prefix) in state_dict_keys dit_config["yak_mlp"] = '{}double_blocks.0.img_mlp.gate_proj.weight'.format(key_prefix) in state_dict_keys From 9d252f3b70c0e89cbb581e28bb1862593c4e5ceb Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Tue, 9 Dec 2025 15:55:13 +1000 Subject: [PATCH 259/647] ops: delete dead code (#11204) This became dead code in https://github.com/comfyanonymous/ComfyUI/pull/11069 --- comfy/ops.py | 8 -------- 1 file changed, 8 deletions(-) diff --git a/comfy/ops.py b/comfy/ops.py index 35237c9f7..6f34d50fc 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -22,7 +22,6 @@ import comfy.model_management from comfy.cli_args import args, PerformanceFeature import comfy.float import comfy.rmsnorm -import contextlib import json def run_every_op(): @@ -94,13 +93,6 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, of else: offload_stream = None - if offload_stream is not None: - wf_context = offload_stream - if hasattr(wf_context, "as_context"): - wf_context = wf_context.as_context(offload_stream) - else: - wf_context = contextlib.nullcontext() - non_blocking = comfy.model_management.device_supports_non_blocking(device) weight_has_function = len(s.weight_function) > 0 From e2a800e7ef225260c078ce484c75bb40161d9d94 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Jukka=20Sepp=C3=A4nen?= <40791699+kijai@users.noreply.github.com> Date: Tue, 9 Dec 2025 23:59:16 +0200 Subject: [PATCH 260/647] Fix for HunyuanVideo1.5 meanflow distil (#11212) --- comfy/ldm/hunyuan_video/model.py | 3 ++- comfy/model_detection.py | 2 ++ 2 files changed, 4 insertions(+), 1 deletion(-) diff --git a/comfy/ldm/hunyuan_video/model.py b/comfy/ldm/hunyuan_video/model.py index 2749c53f5..55ab550f8 100644 --- a/comfy/ldm/hunyuan_video/model.py +++ b/comfy/ldm/hunyuan_video/model.py @@ -43,6 +43,7 @@ class HunyuanVideoParams: meanflow: bool use_cond_type_embedding: bool vision_in_dim: int + meanflow_sum: bool class SelfAttentionRef(nn.Module): @@ -317,7 +318,7 @@ class HunyuanVideo(nn.Module): timesteps_r = transformer_options['sample_sigmas'][w[0] + 1] timesteps_r = timesteps_r.unsqueeze(0).to(device=timesteps.device, dtype=timesteps.dtype) vec_r = self.time_r_in(timestep_embedding(timesteps_r, 256, time_factor=1000.0).to(img.dtype)) - vec = (vec + vec_r) / 2 + vec = (vec + vec_r) if self.params.meanflow_sum else (vec + vec_r) / 2 if ref_latent is not None: ref_latent_ids = self.img_ids(ref_latent) diff --git a/comfy/model_detection.py b/comfy/model_detection.py index 19e6aa954..1f5d34bdd 100644 --- a/comfy/model_detection.py +++ b/comfy/model_detection.py @@ -180,8 +180,10 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): dit_config["use_cond_type_embedding"] = False if '{}vision_in.proj.0.weight'.format(key_prefix) in state_dict_keys: dit_config["vision_in_dim"] = state_dict['{}vision_in.proj.0.weight'.format(key_prefix)].shape[0] + dit_config["meanflow_sum"] = True else: dit_config["vision_in_dim"] = None + dit_config["meanflow_sum"] = False return dit_config if '{}double_blocks.0.img_attn.norm.key_norm.scale'.format(key_prefix) in state_dict_keys and ('{}img_in.weight'.format(key_prefix) in state_dict_keys or f"{key_prefix}distilled_guidance_layer.norms.0.scale" in state_dict_keys): #Flux, Chroma or Chroma Radiance (has no img_in.weight) From 791e30ff5037fa5e7aa4e1396099ea8d6bfb020b Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 9 Dec 2025 14:03:21 -0800 Subject: [PATCH 261/647] Fix nan issue when quantizing fp16 tensor. (#11213) --- comfy/quant_ops.py | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/comfy/quant_ops.py b/comfy/quant_ops.py index 571d3f760..cd96541d7 100644 --- a/comfy/quant_ops.py +++ b/comfy/quant_ops.py @@ -399,7 +399,10 @@ class TensorCoreFP8Layout(QuantizedLayout): orig_dtype = tensor.dtype if isinstance(scale, str) and scale == "recalculate": - scale = torch.amax(tensor.abs()) / torch.finfo(dtype).max + scale = torch.amax(tensor.abs()).to(dtype=torch.float32) / torch.finfo(dtype).max + if tensor.dtype not in [torch.float32, torch.bfloat16]: # Prevent scale from being too small + tensor_info = torch.finfo(tensor.dtype) + scale = (1.0 / torch.clamp((1.0 / scale), min=tensor_info.min, max=tensor_info.max)) if scale is not None: if not isinstance(scale, torch.Tensor): From fc657f471a29d07696ca16b566000e8e555d67d1 Mon Sep 17 00:00:00 2001 From: comfyanonymous Date: Tue, 9 Dec 2025 18:22:09 -0500 Subject: [PATCH 262/647] ComfyUI version v0.4.0 From now on ComfyUI will do version numbers a bit differently, every stable off the master branch will increment the minor version. Anytime a fix needs to be backported onto a stable version the patch version will be incremented. Example: We release v0.6.0 off the master branch then a day later a bug is discovered and we decide to backport the fix onto the v0.6.0 stable, this will be done in a separate branch in the main repository and this new stable will be tagged v0.6.1 --- comfyui_version.py | 2 +- pyproject.toml | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/comfyui_version.py b/comfyui_version.py index 4b039356e..2f083edaf 100644 --- a/comfyui_version.py +++ b/comfyui_version.py @@ -1,3 +1,3 @@ # This file is automatically generated by the build process when version is # updated in pyproject.toml. -__version__ = "0.3.76" +__version__ = "0.4.0" diff --git a/pyproject.toml b/pyproject.toml index 02b94a0ce..e4d3d616a 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -1,6 +1,6 @@ [project] name = "ComfyUI" -version = "0.3.76" +version = "0.4.0" readme = "README.md" license = { file = "LICENSE" } requires-python = ">=3.9" From f668c2e3c99df40561b416cf62b0fd9eec96007a Mon Sep 17 00:00:00 2001 From: Benjamin Lu Date: Tue, 9 Dec 2025 19:27:07 -0800 Subject: [PATCH 263/647] bump comfyui-frontend-package to 1.34.8 (#11220) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 11a7ac245..9e9b25328 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,4 +1,4 @@ -comfyui-frontend-package==1.33.13 +comfyui-frontend-package==1.34.8 comfyui-workflow-templates==0.7.54 comfyui-embedded-docs==0.3.1 torch From 36357bbcc3c515e37a742457a2b2ab4b7ccc17a8 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Wed, 10 Dec 2025 21:55:09 +0200 Subject: [PATCH 264/647] process the NodeV1 dict results correctly (#11237) --- comfy_api/latest/_io.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy_api/latest/_io.py b/comfy_api/latest/_io.py index 313a5af20..79217c813 100644 --- a/comfy_api/latest/_io.py +++ b/comfy_api/latest/_io.py @@ -1815,7 +1815,7 @@ class NodeOutput(_NodeOutputInternal): ui = data["ui"] if "expand" in data: expand = data["expand"] - return cls(args=args, ui=ui, expand=expand) + return cls(*args, ui=ui, expand=expand) def __getitem__(self, index) -> Any: return self.args[index] From 17c92a9f2843d7b9b727531066be2378b350a6ae Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 10 Dec 2025 16:59:48 -0800 Subject: [PATCH 265/647] Tweak Z Image memory estimation. (#11254) --- comfy/supported_models.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy/supported_models.py b/comfy/supported_models.py index 383c82c3e..dd0f09f32 100644 --- a/comfy/supported_models.py +++ b/comfy/supported_models.py @@ -1026,7 +1026,7 @@ class ZImage(Lumina2): "shift": 3.0, } - memory_usage_factor = 1.7 + memory_usage_factor = 2.0 supported_inference_dtypes = [torch.bfloat16, torch.float16, torch.float32] From 57ddb7fd13d817e7259c2c992a852832b6b0f07a Mon Sep 17 00:00:00 2001 From: Johnpaul Chiwetelu <49923152+Myestery@users.noreply.github.com> Date: Thu, 11 Dec 2025 03:49:49 +0100 Subject: [PATCH 266/647] Fix: filter hidden files from /internal/files endpoint (#11191) --- api_server/routes/internal/internal_routes.py | 7 ++++++- 1 file changed, 6 insertions(+), 1 deletion(-) diff --git a/api_server/routes/internal/internal_routes.py b/api_server/routes/internal/internal_routes.py index 613b0f7c7..b224306da 100644 --- a/api_server/routes/internal/internal_routes.py +++ b/api_server/routes/internal/internal_routes.py @@ -58,8 +58,13 @@ class InternalRoutes: return web.json_response({"error": "Invalid directory type"}, status=400) directory = get_directory_by_type(directory_type) + + def is_visible_file(entry: os.DirEntry) -> bool: + """Filter out hidden files (e.g., .DS_Store on macOS).""" + return entry.is_file() and not entry.name.startswith('.') + sorted_files = sorted( - (entry for entry in os.scandir(directory) if entry.is_file()), + (entry for entry in os.scandir(directory) if is_visible_file(entry)), key=lambda entry: -entry.stat().st_mtime ) return web.json_response([entry.name for entry in sorted_files], status=200) From e711aaf1a75120195c56ebd1f1ce829c6b7b84db Mon Sep 17 00:00:00 2001 From: Farshore <168402472+jiangchengchengark@users.noreply.github.com> Date: Thu, 11 Dec 2025 11:02:26 +0800 Subject: [PATCH 267/647] =?UTF-8?q?Lower=20VAE=20loading=20requirements?= =?UTF-8?q?=EF=BC=9ACreate=20a=20new=20branch=20for=20GPU=20memory=20calcu?= =?UTF-8?q?lations=20in=20qwen-image=20vae=20(#11199)?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- comfy/sd.py | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) diff --git a/comfy/sd.py b/comfy/sd.py index a16f2d14f..1cad98aef 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -549,8 +549,10 @@ class VAE: ddconfig = {"dim": dim, "z_dim": self.latent_channels, "dim_mult": [1, 2, 4, 4], "num_res_blocks": 2, "attn_scales": [], "temperal_downsample": [False, True, True], "dropout": 0.0} self.first_stage_model = comfy.ldm.wan.vae.WanVAE(**ddconfig) self.working_dtypes = [torch.bfloat16, torch.float16, torch.float32] - self.memory_used_encode = lambda shape, dtype: 6000 * shape[3] * shape[4] * model_management.dtype_size(dtype) - self.memory_used_decode = lambda shape, dtype: 7000 * shape[3] * shape[4] * (8 * 8) * model_management.dtype_size(dtype) + self.memory_used_encode = lambda shape, dtype: (1500 if shape[2]<=4 else 6000) * shape[3] * shape[4] * model_management.dtype_size(dtype) + self.memory_used_decode = lambda shape, dtype: (2200 if shape[2]<=4 else 7000) * shape[3] * shape[4] * (8*8) * model_management.dtype_size(dtype) + + # Hunyuan 3d v2 2.0 & 2.1 elif "geo_decoder.cross_attn_decoder.ln_1.bias" in sd: From 93948e3fc598c14082f744fe82fae056b64ff481 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Thu, 11 Dec 2025 08:11:12 +0200 Subject: [PATCH 268/647] feat(api-nodes): enable Kling Omni O1 node (#11229) --- comfy_api_nodes/nodes_kling.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy_api_nodes/nodes_kling.py b/comfy_api_nodes/nodes_kling.py index 6c840dc47..a2cc87d84 100644 --- a/comfy_api_nodes/nodes_kling.py +++ b/comfy_api_nodes/nodes_kling.py @@ -2056,7 +2056,7 @@ class KlingExtension(ComfyExtension): OmniProImageToVideoNode, OmniProVideoToVideoNode, OmniProEditVideoNode, - # OmniProImageNode, # need support from backend + OmniProImageNode, ] From f8321eb57b29a4b34cecd27d5d6365adf5e6e601 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 10 Dec 2025 22:30:31 -0800 Subject: [PATCH 269/647] Adjust memory usage factor. (#11257) --- comfy/supported_models.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy/supported_models.py b/comfy/supported_models.py index dd0f09f32..ef8c75c09 100644 --- a/comfy/supported_models.py +++ b/comfy/supported_models.py @@ -541,7 +541,7 @@ class SD3(supported_models_base.BASE): unet_extra_config = {} latent_format = latent_formats.SD3 - memory_usage_factor = 1.2 + memory_usage_factor = 1.6 text_encoder_key_prefix = ["text_encoders."] From fdebe182966d1dd9bee3138264937137bd2302d8 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Thu, 11 Dec 2025 14:09:35 -0800 Subject: [PATCH 270/647] Fix regular chroma radiance (#11276) --- comfy/model_detection.py | 2 ++ 1 file changed, 2 insertions(+) diff --git a/comfy/model_detection.py b/comfy/model_detection.py index 1f5d34bdd..94b54b7c2 100644 --- a/comfy/model_detection.py +++ b/comfy/model_detection.py @@ -261,6 +261,8 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): dit_config["nerf_embedder_dtype"] = torch.float32 if "__x0__" in state_dict_keys: # x0 pred dit_config["use_x0"] = True + else: + dit_config["use_x0"] = False else: dit_config["guidance_embed"] = "{}guidance_in.in_layer.weight".format(key_prefix) in state_dict_keys dit_config["yak_mlp"] = '{}double_blocks.0.img_mlp.gate_proj.weight'.format(key_prefix) in state_dict_keys From ae65433a602470eea271df47af0eb871d146a002 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Thu, 11 Dec 2025 14:15:00 -0800 Subject: [PATCH 271/647] This only works on radiance. (#11277) --- comfy/model_detection.py | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/comfy/model_detection.py b/comfy/model_detection.py index 94b54b7c2..dd6a703f6 100644 --- a/comfy/model_detection.py +++ b/comfy/model_detection.py @@ -259,10 +259,10 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): dit_config["nerf_tile_size"] = 512 dit_config["nerf_final_head_type"] = "conv" if f"{key_prefix}nerf_final_layer_conv.norm.scale" in state_dict_keys else "linear" dit_config["nerf_embedder_dtype"] = torch.float32 - if "__x0__" in state_dict_keys: # x0 pred - dit_config["use_x0"] = True - else: - dit_config["use_x0"] = False + if "__x0__" in state_dict_keys: # x0 pred + dit_config["use_x0"] = True + else: + dit_config["use_x0"] = False else: dit_config["guidance_embed"] = "{}guidance_in.in_layer.weight".format(key_prefix) in state_dict_keys dit_config["yak_mlp"] = '{}double_blocks.0.img_mlp.gate_proj.weight'.format(key_prefix) in state_dict_keys From eeb020b9b77e1f3c0c2806bc1e38c7ba9576439e Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Thu, 11 Dec 2025 14:33:09 -0800 Subject: [PATCH 272/647] Better chroma radiance and other models vram estimation. (#11278) --- comfy/supported_models.py | 12 ++++++------ 1 file changed, 6 insertions(+), 6 deletions(-) diff --git a/comfy/supported_models.py b/comfy/supported_models.py index ef8c75c09..834dfcffc 100644 --- a/comfy/supported_models.py +++ b/comfy/supported_models.py @@ -965,7 +965,7 @@ class CosmosT2IPredict2(supported_models_base.BASE): def __init__(self, unet_config): super().__init__(unet_config) - self.memory_usage_factor = (unet_config.get("model_channels", 2048) / 2048) * 0.9 + self.memory_usage_factor = (unet_config.get("model_channels", 2048) / 2048) * 0.95 def get_model(self, state_dict, prefix="", device=None): out = model_base.CosmosPredict2(self, device=device) @@ -1289,7 +1289,7 @@ class ChromaRadiance(Chroma): latent_format = comfy.latent_formats.ChromaRadiance # Pixel-space model, no spatial compression for model input. - memory_usage_factor = 0.038 + memory_usage_factor = 0.044 def get_model(self, state_dict, prefix="", device=None): return model_base.ChromaRadiance(self, device=device) @@ -1332,7 +1332,7 @@ class Omnigen2(supported_models_base.BASE): "shift": 2.6, } - memory_usage_factor = 1.65 #TODO + memory_usage_factor = 1.95 #TODO unet_extra_config = {} latent_format = latent_formats.Flux @@ -1397,7 +1397,7 @@ class HunyuanImage21(HunyuanVideo): latent_format = latent_formats.HunyuanImage21 - memory_usage_factor = 7.7 + memory_usage_factor = 8.7 supported_inference_dtypes = [torch.bfloat16, torch.float32] @@ -1488,7 +1488,7 @@ class Kandinsky5(supported_models_base.BASE): unet_extra_config = {} latent_format = latent_formats.HunyuanVideo - memory_usage_factor = 1.1 #TODO + memory_usage_factor = 1.25 #TODO supported_inference_dtypes = [torch.bfloat16, torch.float32] @@ -1517,7 +1517,7 @@ class Kandinsky5Image(Kandinsky5): } latent_format = latent_formats.Flux - memory_usage_factor = 1.1 #TODO + memory_usage_factor = 1.25 #TODO def get_model(self, state_dict, prefix="", device=None): out = model_base.Kandinsky5Image(self, device=device) From 338d9ae3bbf24a9a06996cdf1c2f228acc65fd96 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Thu, 11 Dec 2025 15:56:33 -0800 Subject: [PATCH 273/647] Make portable updater work with repos in unmerged state. (#11281) --- .ci/update_windows/update.py | 10 ++++++++++ 1 file changed, 10 insertions(+) diff --git a/.ci/update_windows/update.py b/.ci/update_windows/update.py index 59ece5130..fe646a6ed 100755 --- a/.ci/update_windows/update.py +++ b/.ci/update_windows/update.py @@ -53,6 +53,16 @@ try: repo.stash(ident) except KeyError: print("nothing to stash") # noqa: T201 +except: + print("Could not stash, cleaning index and trying again.") # noqa: T201 + repo.state_cleanup() + repo.index.read_tree(repo.head.peel().tree) + repo.index.write() + try: + repo.stash(ident) + except KeyError: + print("nothing to stash.") # noqa: T201 + backup_branch_name = 'backup_branch_{}'.format(datetime.today().strftime('%Y-%m-%d_%H_%M_%S')) print("creating backup branch: {}".format(backup_branch_name)) # noqa: T201 try: From 982876d59a659adb085be5e236aacc4f2c54c19c Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Jukka=20Sepp=C3=A4nen?= <40791699+kijai@users.noreply.github.com> Date: Fri, 12 Dec 2025 05:29:34 +0200 Subject: [PATCH 274/647] WanMove support (#11247) --- comfy_api/latest/_io.py | 8 + comfy_extras/nodes_wanmove.py | 535 ++++++++++++++++++++++++++++++++++ nodes.py | 1 + 3 files changed, 544 insertions(+) create mode 100644 comfy_extras/nodes_wanmove.py diff --git a/comfy_api/latest/_io.py b/comfy_api/latest/_io.py index 79217c813..2b634d172 100644 --- a/comfy_api/latest/_io.py +++ b/comfy_api/latest/_io.py @@ -774,6 +774,13 @@ class AudioEncoder(ComfyTypeIO): class AudioEncoderOutput(ComfyTypeIO): Type = Any +@comfytype(io_type="TRACKS") +class Tracks(ComfyTypeIO): + class TrackDict(TypedDict): + track_path: torch.Tensor + track_visibility: torch.Tensor + Type = TrackDict + @comfytype(io_type="COMFY_MULTITYPED_V3") class MultiType: Type = Any @@ -1894,6 +1901,7 @@ __all__ = [ "SEGS", "AnyType", "MultiType", + "Tracks", # Dynamic Types "MatchType", # "DynamicCombo", diff --git a/comfy_extras/nodes_wanmove.py b/comfy_extras/nodes_wanmove.py new file mode 100644 index 000000000..5f39afa46 --- /dev/null +++ b/comfy_extras/nodes_wanmove.py @@ -0,0 +1,535 @@ +import nodes +import node_helpers +import torch +import torchvision.transforms.functional as TF +import comfy.model_management +import comfy.utils +import numpy as np +from typing_extensions import override +from comfy_api.latest import ComfyExtension, io +from comfy_extras.nodes_wan import parse_json_tracks + +# https://github.com/ali-vilab/Wan-Move/blob/main/wan/modules/trajectory.py +from PIL import Image, ImageDraw + +SKIP_ZERO = False + +def get_pos_emb( + pos_k: torch.Tensor, # A 1D tensor containing positions for which to generate embeddings. + pos_emb_dim: int, + theta_func: callable = lambda i, d: torch.pow(10000, torch.mul(2, torch.div(i.to(torch.float32), d))), #Function to compute thetas based on position and embedding dimensions. + device: torch.device = torch.device("cpu"), + dtype: torch.dtype = torch.float32, +) -> torch.Tensor: # The position embeddings (batch_size, pos_emb_dim) + + assert pos_emb_dim % 2 == 0, "The dimension of position embeddings must be even." + pos_k = pos_k.to(device, dtype) + if SKIP_ZERO: + pos_k = pos_k + 1 + batch_size = pos_k.size(0) + + denominator = torch.arange(0, pos_emb_dim // 2, device=device, dtype=dtype) + # Expand denominator to match the shape needed for broadcasting + denominator_expanded = denominator.view(1, -1).expand(batch_size, -1) + + thetas = theta_func(denominator_expanded, pos_emb_dim) + + # Ensure pos_k is in the correct shape for broadcasting + pos_k_expanded = pos_k.view(-1, 1).to(dtype) + sin_thetas = torch.sin(torch.div(pos_k_expanded, thetas)) + cos_thetas = torch.cos(torch.div(pos_k_expanded, thetas)) + + # Concatenate sine and cosine embeddings along the last dimension + pos_emb = torch.cat([sin_thetas, cos_thetas], dim=-1) + + return pos_emb + +def create_pos_embeddings( + pred_tracks: torch.Tensor, # the predicted tracks, [T, N, 2] + pred_visibility: torch.Tensor, # the predicted visibility [T, N] + downsample_ratios: list[int], # the ratios for downsampling time, height, and width + height: int, # the height of the feature map + width: int, # the width of the feature map + track_num: int = -1, # the number of tracks to use + t_down_strategy: str = "sample", # the strategy for downsampling time dimension +): + assert t_down_strategy in ["sample", "average"], "Invalid strategy for downsampling time dimension." + + t, n, _ = pred_tracks.shape + t_down, h_down, w_down = downsample_ratios + track_pos = - torch.ones(n, (t-1) // t_down + 1, 2, dtype=torch.long) + + if track_num == -1: + track_num = n + + tracks_idx = torch.randperm(n)[:track_num] + tracks = pred_tracks[:, tracks_idx] + visibility = pred_visibility[:, tracks_idx] + + for t_idx in range(0, t, t_down): + if t_down_strategy == "sample" or t_idx == 0: + cur_tracks = tracks[t_idx] # [N, 2] + cur_visibility = visibility[t_idx] # [N] + else: + cur_tracks = tracks[t_idx:t_idx+t_down].mean(dim=0) + cur_visibility = torch.any(visibility[t_idx:t_idx+t_down], dim=0) + + for i in range(track_num): + if not cur_visibility[i] or cur_tracks[i][0] < 0 or cur_tracks[i][1] < 0 or cur_tracks[i][0] >= width or cur_tracks[i][1] >= height: + continue + x, y = cur_tracks[i] + x, y = int(x // w_down), int(y // h_down) + track_pos[i, t_idx // t_down, 0], track_pos[i, t_idx // t_down, 1] = y, x + + return track_pos # the position embeddings, [N, T', 2], 2 = height, width + +def replace_feature( + vae_feature: torch.Tensor, # [B, C', T', H', W'] + track_pos: torch.Tensor, # [B, N, T', 2] + strength: float = 1.0 +) -> torch.Tensor: + b, _, t, h, w = vae_feature.shape + assert b == track_pos.shape[0], "Batch size mismatch." + n = track_pos.shape[1] + + # Shuffle the trajectory order + track_pos = track_pos[:, torch.randperm(n), :, :] + + # Extract coordinates at time steps ≥ 1 and generate a valid mask + current_pos = track_pos[:, :, 1:, :] # [B, N, T-1, 2] + mask = (current_pos[..., 0] >= 0) & (current_pos[..., 1] >= 0) # [B, N, T-1] + + # Get all valid indices + valid_indices = mask.nonzero(as_tuple=False) # [num_valid, 3] + num_valid = valid_indices.shape[0] + + if num_valid == 0: + return vae_feature + + # Decompose valid indices into each dimension + batch_idx = valid_indices[:, 0] + track_idx = valid_indices[:, 1] + t_rel = valid_indices[:, 2] + t_target = t_rel + 1 # Convert to original time step indices + + # Extract target position coordinates + h_target = current_pos[batch_idx, track_idx, t_rel, 0].long() # Ensure integer indices + w_target = current_pos[batch_idx, track_idx, t_rel, 1].long() + + # Extract source position coordinates (t=0) + h_source = track_pos[batch_idx, track_idx, 0, 0].long() + w_source = track_pos[batch_idx, track_idx, 0, 1].long() + + # Get source features and assign to target positions + src_features = vae_feature[batch_idx, :, 0, h_source, w_source] + dst_features = vae_feature[batch_idx, :, t_target, h_target, w_target] + + vae_feature[batch_idx, :, t_target, h_target, w_target] = dst_features + (src_features - dst_features) * strength + + + return vae_feature + +# Visualize functions + +def _draw_gradient_polyline_on_overlay(overlay, line_width, points, start_color, opacity=1.0): + draw = ImageDraw.Draw(overlay, 'RGBA') + points = points[::-1] + + # Compute total length + total_length = 0 + segment_lengths = [] + for i in range(len(points) - 1): + dx = points[i + 1][0] - points[i][0] + dy = points[i + 1][1] - points[i][1] + length = (dx * dx + dy * dy) ** 0.5 + segment_lengths.append(length) + total_length += length + + if total_length == 0: + return + + accumulated_length = 0 + + # Draw the gradient polyline + for idx, (start_point, end_point) in enumerate(zip(points[:-1], points[1:])): + segment_length = segment_lengths[idx] + steps = max(int(segment_length), 1) + + for i in range(steps): + current_length = accumulated_length + (i / steps) * segment_length + ratio = current_length / total_length + + alpha = int(255 * (1 - ratio) * opacity) + color = (*start_color, alpha) + + x = int(start_point[0] + (end_point[0] - start_point[0]) * i / steps) + y = int(start_point[1] + (end_point[1] - start_point[1]) * i / steps) + + dynamic_line_width = max(int(line_width * (1 - ratio)), 1) + draw.line([(x, y), (x + 1, y)], fill=color, width=dynamic_line_width) + + accumulated_length += segment_length + + +def add_weighted(rgb, track): + rgb = np.array(rgb) # [H, W, C] "RGB" + track = np.array(track) # [H, W, C] "RGBA" + + alpha = track[:, :, 3] / 255.0 + alpha = np.stack([alpha] * 3, axis=-1) + blend_img = track[:, :, :3] * alpha + rgb * (1 - alpha) + + return Image.fromarray(blend_img.astype(np.uint8)) + +def draw_tracks_on_video(video, tracks, visibility=None, track_frame=24, circle_size=12, opacity=0.5, line_width=16): + color_map = [(102, 153, 255), (0, 255, 255), (255, 255, 0), (255, 102, 204), (0, 255, 0)] + + video = video.byte().cpu().numpy() # (81, 480, 832, 3) + tracks = tracks[0].long().detach().cpu().numpy() + if visibility is not None: + visibility = visibility[0].detach().cpu().numpy() + + num_frames, height, width = video.shape[:3] + num_tracks = tracks.shape[1] + alpha_opacity = int(255 * opacity) + + output_frames = [] + for t in range(num_frames): + frame_rgb = video[t].astype(np.float32) + + # Create a single RGBA overlay for all tracks in this frame + overlay = Image.new("RGBA", (width, height), (0, 0, 0, 0)) + draw_overlay = ImageDraw.Draw(overlay) + + polyline_data = [] + + # Draw all circles on a single overlay + for n in range(num_tracks): + if visibility is not None and visibility[t, n] == 0: + continue + + track_coord = tracks[t, n] + color = color_map[n % len(color_map)] + circle_color = color + (alpha_opacity,) + + draw_overlay.ellipse((track_coord[0] - circle_size, track_coord[1] - circle_size, track_coord[0] + circle_size, track_coord[1] + circle_size), + fill=circle_color + ) + + # Store polyline data for batch processing + tracks_coord = tracks[max(t - track_frame, 0):t + 1, n] + if len(tracks_coord) > 1: + polyline_data.append((tracks_coord, color)) + + # Blend circles overlay once + overlay_np = np.array(overlay) + alpha = overlay_np[:, :, 3:4] / 255.0 + frame_rgb = overlay_np[:, :, :3] * alpha + frame_rgb * (1 - alpha) + + # Draw all polylines on a single overlay + if polyline_data: + polyline_overlay = Image.new("RGBA", (width, height), (0, 0, 0, 0)) + for tracks_coord, color in polyline_data: + _draw_gradient_polyline_on_overlay(polyline_overlay, line_width, tracks_coord, color, opacity) + + # Blend polylines overlay once + polyline_np = np.array(polyline_overlay) + alpha = polyline_np[:, :, 3:4] / 255.0 + frame_rgb = polyline_np[:, :, :3] * alpha + frame_rgb * (1 - alpha) + + output_frames.append(Image.fromarray(frame_rgb.astype(np.uint8))) + + return output_frames + + +class WanMoveVisualizeTracks(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="WanMoveVisualizeTracks", + category="conditioning/video_models", + inputs=[ + io.Image.Input("images"), + io.Tracks.Input("tracks", optional=True), + io.Int.Input("line_resolution", default=24, min=1, max=1024), + io.Int.Input("circle_size", default=12, min=1, max=128), + io.Float.Input("opacity", default=0.75, min=0.0, max=1.0, step=0.01), + io.Int.Input("line_width", default=16, min=1, max=128), + ], + outputs=[ + io.Image.Output(), + ], + ) + + @classmethod + def execute(cls, images, line_resolution, circle_size, opacity, line_width, tracks=None) -> io.NodeOutput: + if tracks is None: + return io.NodeOutput(images) + + track_path = tracks["track_path"].unsqueeze(0) + track_visibility = tracks["track_visibility"].unsqueeze(0) + images_in = images * 255.0 + if images_in.shape[0] != track_path.shape[1]: + repeat_count = track_path.shape[1] // images.shape[0] + images_in = images_in.repeat(repeat_count, 1, 1, 1) + track_video = draw_tracks_on_video(images_in, track_path, track_visibility, track_frame=line_resolution, circle_size=circle_size, opacity=opacity, line_width=line_width) + track_video = torch.stack([TF.to_tensor(frame) for frame in track_video], dim=0).movedim(1, -1).float() + + return io.NodeOutput(track_video.to(comfy.model_management.intermediate_device())) + + +class WanMoveTracksFromCoords(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="WanMoveTracksFromCoords", + category="conditioning/video_models", + inputs=[ + io.String.Input("track_coords", force_input=True, default="[]", optional=True), + io.Mask.Input("track_mask", optional=True), + ], + outputs=[ + io.Tracks.Output(), + io.Int.Output(display_name="track_length"), + ], + ) + + @classmethod + def execute(cls, track_coords, track_mask=None) -> io.NodeOutput: + device=comfy.model_management.intermediate_device() + + tracks_data = parse_json_tracks(track_coords) + track_length = len(tracks_data[0]) + + track_list = [ + [[track[frame]['x'], track[frame]['y']] for track in tracks_data] + for frame in range(len(tracks_data[0])) + ] + tracks = torch.tensor(track_list, dtype=torch.float32, device=device) # [frames, num_tracks, 2] + + num_tracks = tracks.shape[-2] + if track_mask is None: + track_visibility = torch.ones((track_length, num_tracks), dtype=torch.bool, device=device) + else: + track_visibility = (track_mask > 0).any(dim=(1, 2)).unsqueeze(-1) + + out_track_info = {} + out_track_info["track_path"] = tracks + out_track_info["track_visibility"] = track_visibility + return io.NodeOutput(out_track_info, track_length) + + +class GenerateTracks(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="GenerateTracks", + category="conditioning/video_models", + inputs=[ + io.Int.Input("width", default=832, min=16, max=4096, step=16), + io.Int.Input("height", default=480, min=16, max=4096, step=16), + io.Float.Input("start_x", default=0.0, min=0.0, max=1.0, step=0.01, tooltip="Normalized X coordinate (0-1) for start position."), + io.Float.Input("start_y", default=0.0, min=0.0, max=1.0, step=0.01, tooltip="Normalized Y coordinate (0-1) for start position."), + io.Float.Input("end_x", default=1.0, min=0.0, max=1.0, step=0.01, tooltip="Normalized X coordinate (0-1) for end position."), + io.Float.Input("end_y", default=1.0, min=0.0, max=1.0, step=0.01, tooltip="Normalized Y coordinate (0-1) for end position."), + io.Int.Input("num_frames", default=81, min=1, max=1024), + io.Int.Input("num_tracks", default=5, min=1, max=100), + io.Float.Input("track_spread", default=0.025, min=0.0, max=1.0, step=0.001, tooltip="Normalized distance between tracks. Tracks are spread perpendicular to the motion direction."), + io.Boolean.Input("bezier", default=False, tooltip="Enable Bezier curve path using the mid point as control point."), + io.Float.Input("mid_x", default=0.5, min=0.0, max=1.0, step=0.01, tooltip="Normalized X control point for Bezier curve. Only used when 'bezier' is enabled."), + io.Float.Input("mid_y", default=0.5, min=0.0, max=1.0, step=0.01, tooltip="Normalized Y control point for Bezier curve. Only used when 'bezier' is enabled."), + io.Combo.Input( + "interpolation", + options=["linear", "ease_in", "ease_out", "ease_in_out", "constant"], + tooltip="Controls the timing/speed of movement along the path.", + ), + io.Mask.Input("track_mask", optional=True, tooltip="Optional mask to indicate visible frames."), + ], + outputs=[ + io.Tracks.Output(), + io.Int.Output(display_name="track_length"), + ], + ) + + @classmethod + def execute(cls, width, height, start_x, start_y, mid_x, mid_y, end_x, end_y, num_frames, num_tracks, + track_spread, bezier=False, interpolation="linear", track_mask=None) -> io.NodeOutput: + device = comfy.model_management.intermediate_device() + track_length = num_frames + + # normalized coordinates to pixel coordinates + start_x_px = start_x * width + start_y_px = start_y * height + mid_x_px = mid_x * width + mid_y_px = mid_y * height + end_x_px = end_x * width + end_y_px = end_y * height + + track_spread_px = track_spread * (width + height) / 2 # Use average of width/height for spread to keep it proportional + + t = torch.linspace(0, 1, num_frames, device=device) + if interpolation == "constant": # All points stay at start position + interp_values = torch.zeros_like(t) + elif interpolation == "linear": + interp_values = t + elif interpolation == "ease_in": + interp_values = t ** 2 + elif interpolation == "ease_out": + interp_values = 1 - (1 - t) ** 2 + elif interpolation == "ease_in_out": + interp_values = t * t * (3 - 2 * t) + + if bezier: # apply interpolation to t for timing control along the bezier path + t_interp = interp_values + one_minus_t = 1 - t_interp + x_positions = one_minus_t ** 2 * start_x_px + 2 * one_minus_t * t_interp * mid_x_px + t_interp ** 2 * end_x_px + y_positions = one_minus_t ** 2 * start_y_px + 2 * one_minus_t * t_interp * mid_y_px + t_interp ** 2 * end_y_px + tangent_x = 2 * one_minus_t * (mid_x_px - start_x_px) + 2 * t_interp * (end_x_px - mid_x_px) + tangent_y = 2 * one_minus_t * (mid_y_px - start_y_px) + 2 * t_interp * (end_y_px - mid_y_px) + else: # calculate base x and y positions for each frame (center track) + x_positions = start_x_px + (end_x_px - start_x_px) * interp_values + y_positions = start_y_px + (end_y_px - start_y_px) * interp_values + # For non-bezier, tangent is constant (direction from start to end) + tangent_x = torch.full_like(t, end_x_px - start_x_px) + tangent_y = torch.full_like(t, end_y_px - start_y_px) + + track_list = [] + for frame_idx in range(num_frames): + # Calculate perpendicular direction at this frame + tx = tangent_x[frame_idx].item() + ty = tangent_y[frame_idx].item() + length = (tx ** 2 + ty ** 2) ** 0.5 + + if length > 0: # Perpendicular unit vector (rotate 90 degrees) + perp_x = -ty / length + perp_y = tx / length + else: # If tangent is zero, spread horizontally + perp_x = 1.0 + perp_y = 0.0 + + frame_tracks = [] + for track_idx in range(num_tracks): # center tracks around the main path offset ranges from -(num_tracks-1)/2 to +(num_tracks-1)/2 + offset = (track_idx - (num_tracks - 1) / 2) * track_spread_px + track_x = x_positions[frame_idx].item() + perp_x * offset + track_y = y_positions[frame_idx].item() + perp_y * offset + frame_tracks.append([track_x, track_y]) + track_list.append(frame_tracks) + + tracks = torch.tensor(track_list, dtype=torch.float32, device=device) # [frames, num_tracks, 2] + + if track_mask is None: + track_visibility = torch.ones((track_length, num_tracks), dtype=torch.bool, device=device) + else: + track_visibility = (track_mask > 0).any(dim=(1, 2)).unsqueeze(-1) + + out_track_info = {} + out_track_info["track_path"] = tracks + out_track_info["track_visibility"] = track_visibility + return io.NodeOutput(out_track_info, track_length) + + +class WanMoveConcatTrack(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="WanMoveConcatTrack", + category="conditioning/video_models", + inputs=[ + io.Tracks.Input("tracks_1"), + io.Tracks.Input("tracks_2", optional=True), + ], + outputs=[ + io.Tracks.Output(), + ], + ) + + @classmethod + def execute(cls, tracks_1=None, tracks_2=None) -> io.NodeOutput: + if tracks_2 is None: + return io.NodeOutput(tracks_1) + + tracks_out = torch.cat([tracks_1["track_path"], tracks_2["track_path"]], dim=1) # Concatenate along the track dimension + mask_out = torch.cat([tracks_1["track_visibility"], tracks_2["track_visibility"]], dim=-1) + + out_track_info = {} + out_track_info["track_path"] = tracks_out + out_track_info["track_visibility"] = mask_out + return io.NodeOutput(out_track_info) + + +class WanMoveTrackToVideo(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="WanMoveTrackToVideo", + category="conditioning/video_models", + inputs=[ + io.Conditioning.Input("positive"), + io.Conditioning.Input("negative"), + io.Vae.Input("vae"), + io.Tracks.Input("tracks", optional=True), + io.Float.Input("strength", default=1.0, min=0.0, max=100.0, step=0.01, tooltip="Strength of the track conditioning."), + io.Int.Input("width", default=832, min=16, max=nodes.MAX_RESOLUTION, step=16), + io.Int.Input("height", default=480, min=16, max=nodes.MAX_RESOLUTION, step=16), + io.Int.Input("length", default=81, min=1, max=nodes.MAX_RESOLUTION, step=4), + io.Int.Input("batch_size", default=1, min=1, max=4096), + io.Image.Input("start_image"), + io.ClipVisionOutput.Input("clip_vision_output", optional=True), + ], + outputs=[ + io.Conditioning.Output(display_name="positive"), + io.Conditioning.Output(display_name="negative"), + io.Latent.Output(display_name="latent"), + ], + ) + + @classmethod + def execute(cls, positive, negative, vae, width, height, length, batch_size, strength, tracks=None, start_image=None, clip_vision_output=None) -> io.NodeOutput: + device=comfy.model_management.intermediate_device() + latent = torch.zeros([batch_size, 16, ((length - 1) // 4) + 1, height // 8, width // 8], device=device) + if start_image is not None: + start_image = comfy.utils.common_upscale(start_image[:length].movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1) + image = torch.ones((length, height, width, start_image.shape[-1]), device=start_image.device, dtype=start_image.dtype) * 0.5 + image[:start_image.shape[0]] = start_image + + concat_latent_image = vae.encode(image[:, :, :, :3]) + mask = torch.ones((1, 1, latent.shape[2], concat_latent_image.shape[-2], concat_latent_image.shape[-1]), device=start_image.device, dtype=start_image.dtype) + mask[:, :, :((start_image.shape[0] - 1) // 4) + 1] = 0.0 + + if tracks is not None and strength > 0.0: + tracks_path = tracks["track_path"][:length] # [T, N, 2] + num_tracks = tracks_path.shape[-2] + + track_visibility = tracks.get("track_visibility", torch.ones((length, num_tracks), dtype=torch.bool, device=device)) + + track_pos = create_pos_embeddings(tracks_path, track_visibility, [4, 8, 8], height, width, track_num=num_tracks) + track_pos = comfy.utils.resize_to_batch_size(track_pos.unsqueeze(0), batch_size) + concat_latent_image_pos = replace_feature(concat_latent_image, track_pos, strength) + else: + concat_latent_image_pos = concat_latent_image + + positive = node_helpers.conditioning_set_values(positive, {"concat_latent_image": concat_latent_image_pos, "concat_mask": mask}) + negative = node_helpers.conditioning_set_values(negative, {"concat_latent_image": concat_latent_image, "concat_mask": mask}) + + if clip_vision_output is not None: + positive = node_helpers.conditioning_set_values(positive, {"clip_vision_output": clip_vision_output}) + negative = node_helpers.conditioning_set_values(negative, {"clip_vision_output": clip_vision_output}) + + out_latent = {} + out_latent["samples"] = latent + return io.NodeOutput(positive, negative, out_latent) + + +class WanMoveExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[io.ComfyNode]]: + return [ + WanMoveTrackToVideo, + WanMoveTracksFromCoords, + WanMoveConcatTrack, + WanMoveVisualizeTracks, + GenerateTracks, + ] + +async def comfy_entrypoint() -> WanMoveExtension: + return WanMoveExtension() diff --git a/nodes.py b/nodes.py index 8d28a725d..8678f510a 100644 --- a/nodes.py +++ b/nodes.py @@ -2358,6 +2358,7 @@ async def init_builtin_extra_nodes(): "nodes_logic.py", "nodes_nop.py", "nodes_kandinsky5.py", + "nodes_wanmove.py", ] import_failed = [] From 5495589db38409353a85b06df7d10f8de2f9c78d Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Thu, 11 Dec 2025 20:32:27 -0800 Subject: [PATCH 275/647] Respect the dtype the op was initialized in for non quant mixed op. (#11282) --- comfy/ops.py | 11 ++++++++--- 1 file changed, 8 insertions(+), 3 deletions(-) diff --git a/comfy/ops.py b/comfy/ops.py index 6f34d50fc..6ae6e791a 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -497,8 +497,10 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec ) -> None: super().__init__() - self.factory_kwargs = {"device": device, "dtype": MixedPrecisionOps._compute_dtype} - # self.factory_kwargs = {"device": device, "dtype": dtype} + if dtype is None: + dtype = MixedPrecisionOps._compute_dtype + + self.factory_kwargs = {"device": device, "dtype": dtype} self.in_features = in_features self.out_features = out_features @@ -530,7 +532,10 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec layer_conf = json.loads(layer_conf.numpy().tobytes()) if layer_conf is None: - self.weight = torch.nn.Parameter(weight.to(device=device, dtype=MixedPrecisionOps._compute_dtype), requires_grad=False) + dtype = self.factory_kwargs["dtype"] + self.weight = torch.nn.Parameter(weight.to(device=device, dtype=dtype), requires_grad=False) + if dtype != MixedPrecisionOps._compute_dtype: + self.comfy_cast_weights = True else: self.quant_format = layer_conf.get("format", None) if not self._full_precision_mm: From 908fd7d7496f6de88722263e1e00fcd3d22e584f Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Fri, 12 Dec 2025 10:18:31 +0200 Subject: [PATCH 276/647] feat(api-nodes): new TextToVideoWithAudio and ImageToVideoWithAudio nodes (#11267) --- comfy_api_nodes/apis/kling_api.py | 28 ++++- comfy_api_nodes/nodes_kling.py | 169 ++++++++++++++++++++++++++---- 2 files changed, 174 insertions(+), 23 deletions(-) diff --git a/comfy_api_nodes/apis/kling_api.py b/comfy_api_nodes/apis/kling_api.py index d8949f8ac..80a758466 100644 --- a/comfy_api_nodes/apis/kling_api.py +++ b/comfy_api_nodes/apis/kling_api.py @@ -51,25 +51,25 @@ class TaskStatusImageResult(BaseModel): url: str = Field(..., description="URL for generated image") -class OmniTaskStatusResults(BaseModel): +class TaskStatusResults(BaseModel): videos: list[TaskStatusVideoResult] | None = Field(None) images: list[TaskStatusImageResult] | None = Field(None) -class OmniTaskStatusResponseData(BaseModel): +class TaskStatusResponseData(BaseModel): created_at: int | None = Field(None, description="Task creation time") updated_at: int | None = Field(None, description="Task update time") task_status: str | None = None task_status_msg: str | None = Field(None, description="Additional failure reason. Only for polling endpoint.") task_id: str | None = Field(None, description="Task ID") - task_result: OmniTaskStatusResults | None = Field(None) + task_result: TaskStatusResults | None = Field(None) -class OmniTaskStatusResponse(BaseModel): +class TaskStatusResponse(BaseModel): code: int | None = Field(None, description="Error code") message: str | None = Field(None, description="Error message") request_id: str | None = Field(None, description="Request ID") - data: OmniTaskStatusResponseData | None = Field(None) + data: TaskStatusResponseData | None = Field(None) class OmniImageParamImage(BaseModel): @@ -84,3 +84,21 @@ class OmniProImageRequest(BaseModel): mode: str = Field("pro") n: int | None = Field(1, le=9) image_list: list[OmniImageParamImage] | None = Field(..., max_length=10) + + +class TextToVideoWithAudioRequest(BaseModel): + model_name: str = Field(..., description="kling-v2-6") + aspect_ratio: str = Field(..., description="'16:9', '9:16' or '1:1'") + duration: str = Field(..., description="'5' or '10'") + prompt: str = Field(...) + mode: str = Field("pro") + sound: str = Field(..., description="'on' or 'off'") + + +class ImageToVideoWithAudioRequest(BaseModel): + model_name: str = Field(..., description="kling-v2-6") + image: str = Field(...) + duration: str = Field(..., description="'5' or '10'") + prompt: str = Field(...) + mode: str = Field("pro") + sound: str = Field(..., description="'on' or 'off'") diff --git a/comfy_api_nodes/nodes_kling.py b/comfy_api_nodes/nodes_kling.py index a2cc87d84..e545fe490 100644 --- a/comfy_api_nodes/nodes_kling.py +++ b/comfy_api_nodes/nodes_kling.py @@ -50,6 +50,7 @@ from comfy_api_nodes.apis import ( KlingSingleImageEffectModelName, ) from comfy_api_nodes.apis.kling_api import ( + ImageToVideoWithAudioRequest, OmniImageParamImage, OmniParamImage, OmniParamVideo, @@ -57,7 +58,8 @@ from comfy_api_nodes.apis.kling_api import ( OmniProImageRequest, OmniProReferences2VideoRequest, OmniProText2VideoRequest, - OmniTaskStatusResponse, + TaskStatusResponse, + TextToVideoWithAudioRequest, ) from comfy_api_nodes.util import ( ApiEndpoint, @@ -242,7 +244,7 @@ def normalize_omni_prompt_references(prompt: str) -> str: return re.sub(r"(?\d*)(?!\w)", _video_repl, prompt) -async def finish_omni_video_task(cls: type[IO.ComfyNode], response: OmniTaskStatusResponse) -> IO.NodeOutput: +async def finish_omni_video_task(cls: type[IO.ComfyNode], response: TaskStatusResponse) -> IO.NodeOutput: if response.code: raise RuntimeError( f"Kling request failed. Code: {response.code}, Message: {response.message}, Data: {response.data}" @@ -250,7 +252,7 @@ async def finish_omni_video_task(cls: type[IO.ComfyNode], response: OmniTaskStat final_response = await poll_op( cls, ApiEndpoint(path=f"/proxy/kling/v1/videos/omni-video/{response.data.task_id}"), - response_model=OmniTaskStatusResponse, + response_model=TaskStatusResponse, status_extractor=lambda r: (r.data.task_status if r.data else None), max_poll_attempts=160, ) @@ -483,12 +485,12 @@ async def execute_image2video( task_id = task_creation_response.data.task_id final_response = await poll_op( - cls, - ApiEndpoint(path=f"{PATH_IMAGE_TO_VIDEO}/{task_id}"), - response_model=KlingImage2VideoResponse, - estimated_duration=AVERAGE_DURATION_I2V, - status_extractor=lambda r: (r.data.task_status.value if r.data and r.data.task_status else None), - ) + cls, + ApiEndpoint(path=f"{PATH_IMAGE_TO_VIDEO}/{task_id}"), + response_model=KlingImage2VideoResponse, + estimated_duration=AVERAGE_DURATION_I2V, + status_extractor=lambda r: (r.data.task_status.value if r.data and r.data.task_status else None), + ) validate_video_result_response(final_response) video = get_video_from_response(final_response) @@ -834,7 +836,7 @@ class OmniProTextToVideoNode(IO.ComfyNode): response = await sync_op( cls, ApiEndpoint(path="/proxy/kling/v1/videos/omni-video", method="POST"), - response_model=OmniTaskStatusResponse, + response_model=TaskStatusResponse, data=OmniProText2VideoRequest( model_name=model_name, prompt=prompt, @@ -929,7 +931,7 @@ class OmniProFirstLastFrameNode(IO.ComfyNode): response = await sync_op( cls, ApiEndpoint(path="/proxy/kling/v1/videos/omni-video", method="POST"), - response_model=OmniTaskStatusResponse, + response_model=TaskStatusResponse, data=OmniProFirstLastFrameRequest( model_name=model_name, prompt=prompt, @@ -997,7 +999,7 @@ class OmniProImageToVideoNode(IO.ComfyNode): response = await sync_op( cls, ApiEndpoint(path="/proxy/kling/v1/videos/omni-video", method="POST"), - response_model=OmniTaskStatusResponse, + response_model=TaskStatusResponse, data=OmniProReferences2VideoRequest( model_name=model_name, prompt=prompt, @@ -1081,7 +1083,7 @@ class OmniProVideoToVideoNode(IO.ComfyNode): response = await sync_op( cls, ApiEndpoint(path="/proxy/kling/v1/videos/omni-video", method="POST"), - response_model=OmniTaskStatusResponse, + response_model=TaskStatusResponse, data=OmniProReferences2VideoRequest( model_name=model_name, prompt=prompt, @@ -1162,7 +1164,7 @@ class OmniProEditVideoNode(IO.ComfyNode): response = await sync_op( cls, ApiEndpoint(path="/proxy/kling/v1/videos/omni-video", method="POST"), - response_model=OmniTaskStatusResponse, + response_model=TaskStatusResponse, data=OmniProReferences2VideoRequest( model_name=model_name, prompt=prompt, @@ -1237,7 +1239,7 @@ class OmniProImageNode(IO.ComfyNode): response = await sync_op( cls, ApiEndpoint(path="/proxy/kling/v1/images/omni-image", method="POST"), - response_model=OmniTaskStatusResponse, + response_model=TaskStatusResponse, data=OmniProImageRequest( model_name=model_name, prompt=prompt, @@ -1253,7 +1255,7 @@ class OmniProImageNode(IO.ComfyNode): final_response = await poll_op( cls, ApiEndpoint(path=f"/proxy/kling/v1/images/omni-image/{response.data.task_id}"), - response_model=OmniTaskStatusResponse, + response_model=TaskStatusResponse, status_extractor=lambda r: (r.data.task_status if r.data else None), ) return IO.NodeOutput(await download_url_to_image_tensor(final_response.data.task_result.images[0].url)) @@ -1328,9 +1330,8 @@ class KlingImage2VideoNode(IO.ComfyNode): def define_schema(cls) -> IO.Schema: return IO.Schema( node_id="KlingImage2VideoNode", - display_name="Kling Image to Video", + display_name="Kling Image(First Frame) to Video", category="api node/video/Kling", - description="Kling Image to Video Node", inputs=[ IO.Image.Input("start_frame", tooltip="The reference image used to generate the video."), IO.String.Input("prompt", multiline=True, tooltip="Positive text prompt"), @@ -2034,6 +2035,136 @@ class KlingImageGenerationNode(IO.ComfyNode): return IO.NodeOutput(await image_result_to_node_output(images)) +class TextToVideoWithAudio(IO.ComfyNode): + + @classmethod + def define_schema(cls) -> IO.Schema: + return IO.Schema( + node_id="KlingTextToVideoWithAudio", + display_name="Kling Text to Video with Audio", + category="api node/video/Kling", + inputs=[ + IO.Combo.Input("model_name", options=["kling-v2-6"]), + IO.String.Input("prompt", multiline=True, tooltip="Positive text prompt."), + IO.Combo.Input("mode", options=["pro"]), + IO.Combo.Input("aspect_ratio", options=["16:9", "9:16", "1:1"]), + IO.Combo.Input("duration", options=[5, 10]), + IO.Boolean.Input("generate_audio", default=True), + ], + outputs=[ + IO.Video.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + ) + + @classmethod + async def execute( + cls, + model_name: str, + prompt: str, + mode: str, + aspect_ratio: str, + duration: int, + generate_audio: bool, + ) -> IO.NodeOutput: + validate_string(prompt, min_length=1, max_length=2500) + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/kling/v1/videos/text2video", method="POST"), + response_model=TaskStatusResponse, + data=TextToVideoWithAudioRequest( + model_name=model_name, + prompt=prompt, + mode=mode, + aspect_ratio=aspect_ratio, + duration=str(duration), + sound="on" if generate_audio else "off", + ), + ) + if response.code: + raise RuntimeError( + f"Kling request failed. Code: {response.code}, Message: {response.message}, Data: {response.data}" + ) + final_response = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/kling/v1/videos/text2video/{response.data.task_id}"), + response_model=TaskStatusResponse, + status_extractor=lambda r: (r.data.task_status if r.data else None), + ) + return IO.NodeOutput(await download_url_to_video_output(final_response.data.task_result.videos[0].url)) + + +class ImageToVideoWithAudio(IO.ComfyNode): + + @classmethod + def define_schema(cls) -> IO.Schema: + return IO.Schema( + node_id="KlingImageToVideoWithAudio", + display_name="Kling Image(First Frame) to Video with Audio", + category="api node/video/Kling", + inputs=[ + IO.Combo.Input("model_name", options=["kling-v2-6"]), + IO.Image.Input("start_frame"), + IO.String.Input("prompt", multiline=True, tooltip="Positive text prompt."), + IO.Combo.Input("mode", options=["pro"]), + IO.Combo.Input("duration", options=[5, 10]), + IO.Boolean.Input("generate_audio", default=True), + ], + outputs=[ + IO.Video.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + ) + + @classmethod + async def execute( + cls, + model_name: str, + start_frame: Input.Image, + prompt: str, + mode: str, + duration: int, + generate_audio: bool, + ) -> IO.NodeOutput: + validate_string(prompt, min_length=1, max_length=2500) + validate_image_dimensions(start_frame, min_width=300, min_height=300) + validate_image_aspect_ratio(start_frame, (1, 2.5), (2.5, 1)) + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/kling/v1/videos/image2video", method="POST"), + response_model=TaskStatusResponse, + data=ImageToVideoWithAudioRequest( + model_name=model_name, + image=(await upload_images_to_comfyapi(cls, start_frame))[0], + prompt=prompt, + mode=mode, + duration=str(duration), + sound="on" if generate_audio else "off", + ), + ) + if response.code: + raise RuntimeError( + f"Kling request failed. Code: {response.code}, Message: {response.message}, Data: {response.data}" + ) + final_response = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/kling/v1/videos/image2video/{response.data.task_id}"), + response_model=TaskStatusResponse, + status_extractor=lambda r: (r.data.task_status if r.data else None), + ) + return IO.NodeOutput(await download_url_to_video_output(final_response.data.task_result.videos[0].url)) + + class KlingExtension(ComfyExtension): @override async def get_node_list(self) -> list[type[IO.ComfyNode]]: @@ -2057,6 +2188,8 @@ class KlingExtension(ComfyExtension): OmniProVideoToVideoNode, OmniProEditVideoNode, OmniProImageNode, + TextToVideoWithAudio, + ImageToVideoWithAudio, ] From c5a47a16924e1be96241553a1448b298e57e50a1 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Fri, 12 Dec 2025 08:49:35 -0800 Subject: [PATCH 277/647] Fix bias dtype issue in mixed ops. (#11293) --- comfy/ops.py | 14 ++++++++++---- 1 file changed, 10 insertions(+), 4 deletions(-) diff --git a/comfy/ops.py b/comfy/ops.py index 6ae6e791a..0384c8717 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -504,10 +504,7 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec self.in_features = in_features self.out_features = out_features - if bias: - self.bias = torch.nn.Parameter(torch.empty(out_features, **self.factory_kwargs)) - else: - self.register_parameter("bias", None) + self._has_bias = bias self.tensor_class = None self._full_precision_mm = MixedPrecisionOps._full_precision_mm @@ -536,6 +533,10 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec self.weight = torch.nn.Parameter(weight.to(device=device, dtype=dtype), requires_grad=False) if dtype != MixedPrecisionOps._compute_dtype: self.comfy_cast_weights = True + if self._has_bias: + self.bias = torch.nn.Parameter(torch.empty(self.out_features, device=device, dtype=dtype)) + else: + self.register_parameter("bias", None) else: self.quant_format = layer_conf.get("format", None) if not self._full_precision_mm: @@ -565,6 +566,11 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec requires_grad=False ) + if self._has_bias: + self.bias = torch.nn.Parameter(torch.empty(self.out_features, device=device, dtype=MixedPrecisionOps._compute_dtype)) + else: + self.register_parameter("bias", None) + for param_name in qconfig["parameters"]: param_key = f"{prefix}{param_name}" _v = state_dict.pop(param_key, None) From da2bfb5b0af26c7a1c44ec951dbd0fffe413c793 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Fri, 12 Dec 2025 22:39:11 -0800 Subject: [PATCH 278/647] Basic implementation of z image fun control union 2.0 (#11304) The inpaint part is currently missing and will be implemented later. I think they messed up this model pretty bad. They added some control_noise_refiner blocks but don't actually use them. There is a typo in their code so instead of doing control_noise_refiner -> control_layers it runs the whole control_layers twice. Unfortunately they trained with this typo so the model works but is kind of slow and would probably perform a lot better if they corrected their code and trained it again. --- comfy/ldm/lumina/controlnet.py | 95 +++++++++++++++++++++++-------- comfy/ldm/lumina/model.py | 16 +++++- comfy/model_patcher.py | 3 + comfy_extras/nodes_model_patch.py | 72 +++++++++++++++++------ 4 files changed, 142 insertions(+), 44 deletions(-) diff --git a/comfy/ldm/lumina/controlnet.py b/comfy/ldm/lumina/controlnet.py index fd7ce3b5c..8e2de7977 100644 --- a/comfy/ldm/lumina/controlnet.py +++ b/comfy/ldm/lumina/controlnet.py @@ -41,6 +41,11 @@ class ZImage_Control(torch.nn.Module): ffn_dim_multiplier: float = (8.0 / 3.0), norm_eps: float = 1e-5, qk_norm: bool = True, + n_control_layers=6, + control_in_dim=16, + additional_in_dim=0, + broken=False, + refiner_control=False, dtype=None, device=None, operations=None, @@ -49,10 +54,11 @@ class ZImage_Control(torch.nn.Module): super().__init__() operation_settings = {"operations": operations, "device": device, "dtype": dtype} - self.additional_in_dim = 0 - self.control_in_dim = 16 + self.broken = broken + self.additional_in_dim = additional_in_dim + self.control_in_dim = control_in_dim n_refiner_layers = 2 - self.n_control_layers = 6 + self.n_control_layers = n_control_layers self.control_layers = nn.ModuleList( [ ZImageControlTransformerBlock( @@ -74,28 +80,49 @@ class ZImage_Control(torch.nn.Module): all_x_embedder = {} patch_size = 2 f_patch_size = 1 - x_embedder = operations.Linear(f_patch_size * patch_size * patch_size * self.control_in_dim, dim, bias=True, device=device, dtype=dtype) + x_embedder = operations.Linear(f_patch_size * patch_size * patch_size * (self.control_in_dim + self.additional_in_dim), dim, bias=True, device=device, dtype=dtype) all_x_embedder[f"{patch_size}-{f_patch_size}"] = x_embedder + self.refiner_control = refiner_control + self.control_all_x_embedder = nn.ModuleDict(all_x_embedder) - self.control_noise_refiner = nn.ModuleList( - [ - JointTransformerBlock( - layer_id, - dim, - n_heads, - n_kv_heads, - multiple_of, - ffn_dim_multiplier, - norm_eps, - qk_norm, - modulation=True, - z_image_modulation=True, - operation_settings=operation_settings, - ) - for layer_id in range(n_refiner_layers) - ] - ) + if self.refiner_control: + self.control_noise_refiner = nn.ModuleList( + [ + ZImageControlTransformerBlock( + layer_id, + dim, + n_heads, + n_kv_heads, + multiple_of, + ffn_dim_multiplier, + norm_eps, + qk_norm, + block_id=layer_id, + operation_settings=operation_settings, + ) + for layer_id in range(n_refiner_layers) + ] + ) + else: + self.control_noise_refiner = nn.ModuleList( + [ + JointTransformerBlock( + layer_id, + dim, + n_heads, + n_kv_heads, + multiple_of, + ffn_dim_multiplier, + norm_eps, + qk_norm, + modulation=True, + z_image_modulation=True, + operation_settings=operation_settings, + ) + for layer_id in range(n_refiner_layers) + ] + ) def forward(self, cap_feats, control_context, x_freqs_cis, adaln_input): patch_size = 2 @@ -105,9 +132,29 @@ class ZImage_Control(torch.nn.Module): control_context = self.control_all_x_embedder[f"{patch_size}-{f_patch_size}"](control_context.view(B, C, H // pH, pH, W // pW, pW).permute(0, 2, 4, 3, 5, 1).flatten(3).flatten(1, 2)) x_attn_mask = None - for layer in self.control_noise_refiner: - control_context = layer(control_context, x_attn_mask, x_freqs_cis[:control_context.shape[0], :control_context.shape[1]], adaln_input) + if not self.refiner_control: + for layer in self.control_noise_refiner: + control_context = layer(control_context, x_attn_mask, x_freqs_cis[:control_context.shape[0], :control_context.shape[1]], adaln_input) + return control_context + def forward_noise_refiner_block(self, layer_id, control_context, x, x_attn_mask, x_freqs_cis, adaln_input): + if self.refiner_control: + if self.broken: + if layer_id == 0: + return self.control_layers[layer_id](control_context, x, x_mask=x_attn_mask, freqs_cis=x_freqs_cis[:control_context.shape[0], :control_context.shape[1]], adaln_input=adaln_input) + if layer_id > 0: + out = None + for i in range(1, len(self.control_layers)): + o, control_context = self.control_layers[i](control_context, x, x_mask=x_attn_mask, freqs_cis=x_freqs_cis[:control_context.shape[0], :control_context.shape[1]], adaln_input=adaln_input) + if out is None: + out = o + + return (out, control_context) + else: + return self.control_noise_refiner[layer_id](control_context, x, x_mask=x_attn_mask, freqs_cis=x_freqs_cis[:control_context.shape[0], :control_context.shape[1]], adaln_input=adaln_input) + else: + return (None, control_context) + def forward_control_block(self, layer_id, control_context, x, x_attn_mask, x_freqs_cis, adaln_input): return self.control_layers[layer_id](control_context, x, x_mask=x_attn_mask, freqs_cis=x_freqs_cis[:control_context.shape[0], :control_context.shape[1]], adaln_input=adaln_input) diff --git a/comfy/ldm/lumina/model.py b/comfy/ldm/lumina/model.py index c47df49ca..96cb37fa6 100644 --- a/comfy/ldm/lumina/model.py +++ b/comfy/ldm/lumina/model.py @@ -536,6 +536,7 @@ class NextDiT(nn.Module): bsz = len(x) pH = pW = self.patch_size device = x[0].device + orig_x = x if self.pad_tokens_multiple is not None: pad_extra = (-cap_feats.shape[1]) % self.pad_tokens_multiple @@ -572,13 +573,21 @@ class NextDiT(nn.Module): freqs_cis = self.rope_embedder(torch.cat((cap_pos_ids, x_pos_ids), dim=1)).movedim(1, 2) + patches = transformer_options.get("patches", {}) + # refine context for layer in self.context_refiner: cap_feats = layer(cap_feats, cap_mask, freqs_cis[:, :cap_pos_ids.shape[1]], transformer_options=transformer_options) padded_img_mask = None - for layer in self.noise_refiner: + x_input = x + for i, layer in enumerate(self.noise_refiner): x = layer(x, padded_img_mask, freqs_cis[:, cap_pos_ids.shape[1]:], t, transformer_options=transformer_options) + if "noise_refiner" in patches: + for p in patches["noise_refiner"]: + out = p({"img": x, "img_input": x_input, "txt": cap_feats, "pe": freqs_cis[:, cap_pos_ids.shape[1]:], "vec": t, "x": orig_x, "block_index": i, "transformer_options": transformer_options, "block_type": "noise_refiner"}) + if "img" in out: + x = out["img"] padded_full_embed = torch.cat((cap_feats, x), dim=1) mask = None @@ -622,14 +631,15 @@ class NextDiT(nn.Module): patches = transformer_options.get("patches", {}) x_is_tensor = isinstance(x, torch.Tensor) - img, mask, img_size, cap_size, freqs_cis = self.patchify_and_embed(x, cap_feats, cap_mask, t, num_tokens, transformer_options=transformer_options) + img, mask, img_size, cap_size, freqs_cis = self.patchify_and_embed(x, cap_feats, cap_mask, adaln_input, num_tokens, transformer_options=transformer_options) freqs_cis = freqs_cis.to(img.device) + img_input = img for i, layer in enumerate(self.layers): img = layer(img, mask, freqs_cis, adaln_input, transformer_options=transformer_options) if "double_block" in patches: for p in patches["double_block"]: - out = p({"img": img[:, cap_size[0]:], "txt": img[:, :cap_size[0]], "pe": freqs_cis[:, cap_size[0]:], "vec": adaln_input, "x": x, "block_index": i, "transformer_options": transformer_options}) + out = p({"img": img[:, cap_size[0]:], "img_input": img_input[:, cap_size[0]:], "txt": img[:, :cap_size[0]], "pe": freqs_cis[:, cap_size[0]:], "vec": adaln_input, "x": x, "block_index": i, "transformer_options": transformer_options}) if "img" in out: img[:, cap_size[0]:] = out["img"] if "txt" in out: diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index a486c2723..93d26c690 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -454,6 +454,9 @@ class ModelPatcher: def set_model_post_input_patch(self, patch): self.set_model_patch(patch, "post_input") + def set_model_noise_refiner_patch(self, patch): + self.set_model_patch(patch, "noise_refiner") + def set_model_rope_options(self, scale_x, shift_x, scale_y, shift_y, scale_t, shift_t, **kwargs): rope_options = self.model_options["transformer_options"].get("rope_options", {}) rope_options["scale_x"] = scale_x diff --git a/comfy_extras/nodes_model_patch.py b/comfy_extras/nodes_model_patch.py index c61810dbf..ec0e790dc 100644 --- a/comfy_extras/nodes_model_patch.py +++ b/comfy_extras/nodes_model_patch.py @@ -243,7 +243,13 @@ class ModelPatchLoader: model = SigLIPMultiFeatProjModel(device=comfy.model_management.unet_offload_device(), dtype=dtype, operations=comfy.ops.manual_cast) elif 'control_all_x_embedder.2-1.weight' in sd: # alipai z image fun controlnet sd = z_image_convert(sd) - model = comfy.ldm.lumina.controlnet.ZImage_Control(device=comfy.model_management.unet_offload_device(), dtype=dtype, operations=comfy.ops.manual_cast) + config = {} + if 'control_layers.14.adaLN_modulation.0.weight' in sd: + config['n_control_layers'] = 15 + config['additional_in_dim'] = 17 + config['refiner_control'] = True + config['broken'] = True + model = comfy.ldm.lumina.controlnet.ZImage_Control(device=comfy.model_management.unet_offload_device(), dtype=dtype, operations=comfy.ops.manual_cast, **config) model.load_state_dict(sd) model = comfy.model_patcher.ModelPatcher(model, load_device=comfy.model_management.get_torch_device(), offload_device=comfy.model_management.unet_offload_device()) @@ -297,56 +303,86 @@ class DiffSynthCnetPatch: return [self.model_patch] class ZImageControlPatch: - def __init__(self, model_patch, vae, image, strength): + def __init__(self, model_patch, vae, image, strength, inpaint_image=None, mask=None): self.model_patch = model_patch self.vae = vae self.image = image + self.inpaint_image = inpaint_image + self.mask = mask self.strength = strength self.encoded_image = self.encode_latent_cond(image) self.encoded_image_size = (image.shape[1], image.shape[2]) self.temp_data = None - def encode_latent_cond(self, image): - latent_image = comfy.latent_formats.Flux().process_in(self.vae.encode(image)) - return latent_image + def encode_latent_cond(self, control_image, inpaint_image=None): + latent_image = comfy.latent_formats.Flux().process_in(self.vae.encode(control_image)) + if self.model_patch.model.additional_in_dim > 0: + if self.mask is None: + mask_ = torch.zeros_like(latent_image)[:, :1] + else: + mask_ = comfy.utils.common_upscale(self.mask.mean(dim=1, keepdim=True), latent_image.shape[-1], latent_image.shape[-2], "bilinear", "none") + if inpaint_image is None: + inpaint_image = torch.ones_like(control_image) * 0.5 + + inpaint_image_latent = comfy.latent_formats.Flux().process_in(self.vae.encode(inpaint_image)) + + return torch.cat([latent_image, mask_, inpaint_image_latent], dim=1) + else: + return latent_image def __call__(self, kwargs): x = kwargs.get("x") img = kwargs.get("img") + img_input = kwargs.get("img_input") txt = kwargs.get("txt") pe = kwargs.get("pe") vec = kwargs.get("vec") block_index = kwargs.get("block_index") + block_type = kwargs.get("block_type", "") spacial_compression = self.vae.spacial_compression_encode() if self.encoded_image is None or self.encoded_image_size != (x.shape[-2] * spacial_compression, x.shape[-1] * spacial_compression): image_scaled = comfy.utils.common_upscale(self.image.movedim(-1, 1), x.shape[-1] * spacial_compression, x.shape[-2] * spacial_compression, "area", "center") + inpaint_scaled = None + if self.inpaint_image is not None: + inpaint_scaled = comfy.utils.common_upscale(self.inpaint_image.movedim(-1, 1), x.shape[-1] * spacial_compression, x.shape[-2] * spacial_compression, "area", "center").movedim(1, -1) loaded_models = comfy.model_management.loaded_models(only_currently_used=True) - self.encoded_image = self.encode_latent_cond(image_scaled.movedim(1, -1)) + self.encoded_image = self.encode_latent_cond(image_scaled.movedim(1, -1), inpaint_scaled) self.encoded_image_size = (image_scaled.shape[-2], image_scaled.shape[-1]) comfy.model_management.load_models_gpu(loaded_models) - cnet_index = (block_index // 5) - cnet_index_float = (block_index / 5) + cnet_blocks = self.model_patch.model.n_control_layers + div = round(30 / cnet_blocks) + + cnet_index = (block_index // div) + cnet_index_float = (block_index / div) kwargs.pop("img") # we do ops in place kwargs.pop("txt") - cnet_blocks = self.model_patch.model.n_control_layers if cnet_index_float > (cnet_blocks - 1): self.temp_data = None return kwargs if self.temp_data is None or self.temp_data[0] > cnet_index: - self.temp_data = (-1, (None, self.model_patch.model(txt, self.encoded_image.to(img.dtype), pe, vec))) + if block_type == "noise_refiner": + self.temp_data = (-3, (None, self.model_patch.model(txt, self.encoded_image.to(img.dtype), pe, vec))) + else: + self.temp_data = (-1, (None, self.model_patch.model(txt, self.encoded_image.to(img.dtype), pe, vec))) - while self.temp_data[0] < cnet_index and (self.temp_data[0] + 1) < cnet_blocks: + if block_type == "noise_refiner": next_layer = self.temp_data[0] + 1 - self.temp_data = (next_layer, self.model_patch.model.forward_control_block(next_layer, self.temp_data[1][1], img[:, :self.temp_data[1][1].shape[1]], None, pe, vec)) + self.temp_data = (next_layer, self.model_patch.model.forward_noise_refiner_block(block_index, self.temp_data[1][1], img_input[:, :self.temp_data[1][1].shape[1]], None, pe, vec)) + if self.temp_data[1][0] is not None: + img[:, :self.temp_data[1][0].shape[1]] += (self.temp_data[1][0] * self.strength) + else: + while self.temp_data[0] < cnet_index and (self.temp_data[0] + 1) < cnet_blocks: + next_layer = self.temp_data[0] + 1 + self.temp_data = (next_layer, self.model_patch.model.forward_control_block(next_layer, self.temp_data[1][1], img_input[:, :self.temp_data[1][1].shape[1]], None, pe, vec)) - if cnet_index_float == self.temp_data[0]: - img[:, :self.temp_data[1][0].shape[1]] += (self.temp_data[1][0] * self.strength) - if cnet_blocks == self.temp_data[0] + 1: - self.temp_data = None + if cnet_index_float == self.temp_data[0]: + img[:, :self.temp_data[1][0].shape[1]] += (self.temp_data[1][0] * self.strength) + if cnet_blocks == self.temp_data[0] + 1: + self.temp_data = None return kwargs @@ -386,7 +422,9 @@ class QwenImageDiffsynthControlnet: mask = 1.0 - mask if isinstance(model_patch.model, comfy.ldm.lumina.controlnet.ZImage_Control): - model_patched.set_model_double_block_patch(ZImageControlPatch(model_patch, vae, image, strength)) + patch = ZImageControlPatch(model_patch, vae, image, strength, mask=mask) + model_patched.set_model_noise_refiner_patch(patch) + model_patched.set_model_double_block_patch(patch) else: model_patched.set_model_double_block_patch(DiffSynthCnetPatch(model_patch, vae, image, strength, mask)) return (model_patched,) From 971cefe7d4ca15c949d5d901a663cb66562a4f10 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sat, 13 Dec 2025 15:45:23 -0800 Subject: [PATCH 279/647] Fix pytorch warnings. (#11314) --- comfy/ops.py | 2 +- comfy/utils.py | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/comfy/ops.py b/comfy/ops.py index 0384c8717..16889bb82 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -592,7 +592,7 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec quant_conf = {"format": self.quant_format} if self._full_precision_mm: quant_conf["full_precision_matrix_mult"] = True - sd["{}comfy_quant".format(prefix)] = torch.frombuffer(json.dumps(quant_conf).encode('utf-8'), dtype=torch.uint8) + sd["{}comfy_quant".format(prefix)] = torch.tensor(list(json.dumps(quant_conf).encode('utf-8')), dtype=torch.uint8) return sd def _forward(self, input, weight, bias): diff --git a/comfy/utils.py b/comfy/utils.py index 9dc0d76ac..3866cda2e 100644 --- a/comfy/utils.py +++ b/comfy/utils.py @@ -1262,6 +1262,6 @@ def convert_old_quants(state_dict, model_prefix="", metadata={}): if quant_metadata is not None: layers = quant_metadata["layers"] for k, v in layers.items(): - state_dict["{}.comfy_quant".format(k)] = torch.frombuffer(json.dumps(v).encode('utf-8'), dtype=torch.uint8) + state_dict["{}.comfy_quant".format(k)] = torch.tensor(list(json.dumps(v).encode('utf-8')), dtype=torch.uint8) return state_dict, metadata From 6592bffc609da4738b111dbffca1f473972f3574 Mon Sep 17 00:00:00 2001 From: chaObserv <154517000+chaObserv@users.noreply.github.com> Date: Sun, 14 Dec 2025 13:03:29 +0800 Subject: [PATCH 280/647] seeds_2: add phi_2 variant and sampler node (#11309) * Add phi_2 solver type to seeds_2 * Add sampler node of seeds_2 --- comfy/k_diffusion/sampling.py | 15 ++++++++++++--- comfy_extras/nodes_custom_sampler.py | 26 ++++++++++++++++++++++++++ 2 files changed, 38 insertions(+), 3 deletions(-) diff --git a/comfy/k_diffusion/sampling.py b/comfy/k_diffusion/sampling.py index 0e2cda291..753c66afa 100644 --- a/comfy/k_diffusion/sampling.py +++ b/comfy/k_diffusion/sampling.py @@ -1557,10 +1557,13 @@ def sample_er_sde(model, x, sigmas, extra_args=None, callback=None, disable=None @torch.no_grad() -def sample_seeds_2(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, r=0.5): +def sample_seeds_2(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, r=0.5, solver_type="phi_1"): """SEEDS-2 - Stochastic Explicit Exponential Derivative-free Solvers (VP Data Prediction) stage 2. arXiv: https://arxiv.org/abs/2305.14267 (NeurIPS 2023) """ + if solver_type not in {"phi_1", "phi_2"}: + raise ValueError("solver_type must be 'phi_1' or 'phi_2'") + extra_args = {} if extra_args is None else extra_args seed = extra_args.get("seed", None) noise_sampler = default_noise_sampler(x, seed=seed) if noise_sampler is None else noise_sampler @@ -1600,8 +1603,14 @@ def sample_seeds_2(model, x, sigmas, extra_args=None, callback=None, disable=Non denoised_2 = model(x_2, sigma_s_1 * s_in, **extra_args) # Step 2 - denoised_d = torch.lerp(denoised, denoised_2, fac) - x = sigmas[i + 1] / sigmas[i] * (-h * eta).exp() * x - alpha_t * ei_h_phi_1(-h_eta) * denoised_d + if solver_type == "phi_1": + denoised_d = torch.lerp(denoised, denoised_2, fac) + x = sigmas[i + 1] / sigmas[i] * (-h * eta).exp() * x - alpha_t * ei_h_phi_1(-h_eta) * denoised_d + elif solver_type == "phi_2": + b2 = ei_h_phi_2(-h_eta) / r + b1 = ei_h_phi_1(-h_eta) - b2 + x = sigmas[i + 1] / sigmas[i] * (-h * eta).exp() * x - alpha_t * (b1 * denoised + b2 * denoised_2) + if inject_noise: segment_factor = (r - 1) * h * eta sde_noise = sde_noise * segment_factor.exp() diff --git a/comfy_extras/nodes_custom_sampler.py b/comfy_extras/nodes_custom_sampler.py index fbb080886..71ea4e9ec 100644 --- a/comfy_extras/nodes_custom_sampler.py +++ b/comfy_extras/nodes_custom_sampler.py @@ -659,6 +659,31 @@ class SamplerSASolver(io.ComfyNode): get_sampler = execute +class SamplerSEEDS2(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="SamplerSEEDS2", + category="sampling/custom_sampling/samplers", + inputs=[ + io.Combo.Input("solver_type", options=["phi_1", "phi_2"]), + io.Float.Input("eta", default=1.0, min=0.0, max=100.0, step=0.01, round=False, tooltip="Stochastic strength"), + io.Float.Input("s_noise", default=1.0, min=0.0, max=100.0, step=0.01, round=False, tooltip="SDE noise multiplier"), + io.Float.Input("r", default=0.5, min=0.01, max=1.0, step=0.01, round=False, tooltip="Relative step size for the intermediate stage (c2 node)"), + ], + outputs=[io.Sampler.Output()] + ) + + @classmethod + def execute(cls, solver_type, eta, s_noise, r) -> io.NodeOutput: + sampler_name = "seeds_2" + sampler = comfy.samplers.ksampler( + sampler_name, + {"eta": eta, "s_noise": s_noise, "r": r, "solver_type": solver_type}, + ) + return io.NodeOutput(sampler) + + class Noise_EmptyNoise: def __init__(self): self.seed = 0 @@ -996,6 +1021,7 @@ class CustomSamplersExtension(ComfyExtension): SamplerDPMAdaptative, SamplerER_SDE, SamplerSASolver, + SamplerSEEDS2, SplitSigmas, SplitSigmasDenoise, FlipSigmas, From 5ac3b26a7dedb9b13c681abe8733c54f13353273 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sun, 14 Dec 2025 01:02:50 -0800 Subject: [PATCH 281/647] Update warning for old pytorch version. (#11319) Versions below 2.4 are no longer supported. We will not break support on purpose but will not fix it if we do. --- comfy/utils.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy/utils.py b/comfy/utils.py index 3866cda2e..8d4e2b445 100644 --- a/comfy/utils.py +++ b/comfy/utils.py @@ -53,7 +53,7 @@ if hasattr(torch.serialization, "add_safe_globals"): # TODO: this was added in ALWAYS_SAFE_LOAD = True logging.info("Checkpoint files will always be loaded safely.") else: - logging.info("Warning, you are using an old pytorch version and some ckpt/pt files might be loaded unsafely. Upgrading to 2.4 or above is recommended.") + logging.warning("Warning, you are using an old pytorch version and some ckpt/pt files might be loaded unsafely. Upgrading to 2.4 or above is recommended as older versions of pytorch are no longer supported.") def load_torch_file(ckpt, safe_load=False, device=None, return_metadata=False): if device is None: From a5e85017d8574cb99024d320f7a53a77a9e6aa5a Mon Sep 17 00:00:00 2001 From: "Dr.Lt.Data" <128333288+ltdrdata@users.noreply.github.com> Date: Tue, 16 Dec 2025 04:24:01 +0900 Subject: [PATCH 282/647] bump manager requirments to the 4.0.3b5 (#11324) --- manager_requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/manager_requirements.txt b/manager_requirements.txt index b95cefb74..5ef0d3a1d 100644 --- a/manager_requirements.txt +++ b/manager_requirements.txt @@ -1 +1 @@ -comfyui_manager==4.0.3b4 +comfyui_manager==4.0.3b5 From 51347f9fb8a8e60d3add049c6f241822c84c8a87 Mon Sep 17 00:00:00 2001 From: ComfyUI Wiki Date: Tue, 16 Dec 2025 05:28:55 +0800 Subject: [PATCH 283/647] chore: update workflow templates to v0.7.59 (#11337) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 9e9b25328..117260515 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,5 +1,5 @@ comfyui-frontend-package==1.34.8 -comfyui-workflow-templates==0.7.54 +comfyui-workflow-templates==0.7.59 comfyui-embedded-docs==0.3.1 torch torchsde From 5cb1e0c9a0439f1f95a0b372474bd4845e38009c Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 15 Dec 2025 13:49:29 -0800 Subject: [PATCH 284/647] Disable guards on transformer_options when torch.compile (#11317) --- comfy_extras/nodes_torch_compile.py | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/comfy_extras/nodes_torch_compile.py b/comfy_extras/nodes_torch_compile.py index adbeece2f..c43e8ad63 100644 --- a/comfy_extras/nodes_torch_compile.py +++ b/comfy_extras/nodes_torch_compile.py @@ -2,6 +2,8 @@ from typing_extensions import override from comfy_api.latest import ComfyExtension, io from comfy_api.torch_helpers import set_torch_compile_wrapper +def skip_torch_compile_dict(guard_entries): + return [("transformer_options" not in entry.name) for entry in guard_entries] class TorchCompileModel(io.ComfyNode): @classmethod @@ -23,7 +25,7 @@ class TorchCompileModel(io.ComfyNode): @classmethod def execute(cls, model, backend) -> io.NodeOutput: m = model.clone() - set_torch_compile_wrapper(model=m, backend=backend) + set_torch_compile_wrapper(model=m, backend=backend, options={"guard_filter_fn": skip_torch_compile_dict}) return io.NodeOutput(m) From af91eb6c9931d0a2c99cf8a6d4974a6abf9a09fa Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Tue, 16 Dec 2025 01:30:24 +0200 Subject: [PATCH 285/647] api-nodes: drop Kling v1 model (#11307) --- comfy_api_nodes/nodes_kling.py | 12 +++--------- 1 file changed, 3 insertions(+), 9 deletions(-) diff --git a/comfy_api_nodes/nodes_kling.py b/comfy_api_nodes/nodes_kling.py index e545fe490..1a6364fa0 100644 --- a/comfy_api_nodes/nodes_kling.py +++ b/comfy_api_nodes/nodes_kling.py @@ -105,10 +105,6 @@ AVERAGE_DURATION_VIDEO_EXTEND = 320 MODE_TEXT2VIDEO = { - "standard mode / 5s duration / kling-v1": ("std", "5", "kling-v1"), - "standard mode / 10s duration / kling-v1": ("std", "10", "kling-v1"), - "pro mode / 5s duration / kling-v1": ("pro", "5", "kling-v1"), - "pro mode / 10s duration / kling-v1": ("pro", "10", "kling-v1"), "standard mode / 5s duration / kling-v1-6": ("std", "5", "kling-v1-6"), "standard mode / 10s duration / kling-v1-6": ("std", "10", "kling-v1-6"), "pro mode / 5s duration / kling-v2-master": ("pro", "5", "kling-v2-master"), @@ -129,8 +125,6 @@ See: [Kling API Docs Capability Map](https://app.klingai.com/global/dev/document MODE_START_END_FRAME = { - "standard mode / 5s duration / kling-v1": ("std", "5", "kling-v1"), - "pro mode / 5s duration / kling-v1": ("pro", "5", "kling-v1"), "pro mode / 5s duration / kling-v1-5": ("pro", "5", "kling-v1-5"), "pro mode / 10s duration / kling-v1-5": ("pro", "10", "kling-v1-5"), "pro mode / 5s duration / kling-v1-6": ("pro", "5", "kling-v1-6"), @@ -754,7 +748,7 @@ class KlingTextToVideoNode(IO.ComfyNode): IO.Combo.Input( "mode", options=modes, - default=modes[4], + default=modes[8], tooltip="The configuration to use for the video generation following the format: mode / duration / model_name.", ), ], @@ -1489,7 +1483,7 @@ class KlingStartEndFrameNode(IO.ComfyNode): IO.Combo.Input( "mode", options=modes, - default=modes[8], + default=modes[6], tooltip="The configuration to use for the video generation following the format: mode / duration / model_name.", ), ], @@ -1952,7 +1946,7 @@ class KlingImageGenerationNode(IO.ComfyNode): IO.Combo.Input( "model_name", options=[i.value for i in KlingImageGenModelName], - default="kling-v1", + default="kling-v2", ), IO.Combo.Input( "aspect_ratio", From 33c7f1179d4a961e4ca1dd78188c5134e0ee8e8c Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Tue, 16 Dec 2025 01:32:29 +0200 Subject: [PATCH 286/647] drop Pika API nodes (#11306) --- comfy_api_nodes/apis/pika_api.py | 100 ------ comfy_api_nodes/nodes_pika.py | 575 ------------------------------- nodes.py | 1 - 3 files changed, 676 deletions(-) delete mode 100644 comfy_api_nodes/apis/pika_api.py delete mode 100644 comfy_api_nodes/nodes_pika.py diff --git a/comfy_api_nodes/apis/pika_api.py b/comfy_api_nodes/apis/pika_api.py deleted file mode 100644 index 232558cd7..000000000 --- a/comfy_api_nodes/apis/pika_api.py +++ /dev/null @@ -1,100 +0,0 @@ -from typing import Optional -from enum import Enum -from pydantic import BaseModel, Field - - -class Pikaffect(str, Enum): - Cake_ify = "Cake-ify" - Crumble = "Crumble" - Crush = "Crush" - Decapitate = "Decapitate" - Deflate = "Deflate" - Dissolve = "Dissolve" - Explode = "Explode" - Eye_pop = "Eye-pop" - Inflate = "Inflate" - Levitate = "Levitate" - Melt = "Melt" - Peel = "Peel" - Poke = "Poke" - Squish = "Squish" - Ta_da = "Ta-da" - Tear = "Tear" - - -class PikaBodyGenerate22C2vGenerate22PikascenesPost(BaseModel): - aspectRatio: Optional[float] = Field(None, description='Aspect ratio (width / height)') - duration: Optional[int] = Field(5) - ingredientsMode: str = Field(...) - negativePrompt: Optional[str] = Field(None) - promptText: Optional[str] = Field(None) - resolution: Optional[str] = Field('1080p') - seed: Optional[int] = Field(None) - - -class PikaGenerateResponse(BaseModel): - video_id: str = Field(...) - - -class PikaBodyGenerate22I2vGenerate22I2vPost(BaseModel): - duration: Optional[int] = 5 - negativePrompt: Optional[str] = Field(None) - promptText: Optional[str] = Field(None) - resolution: Optional[str] = '1080p' - seed: Optional[int] = Field(None) - - -class PikaBodyGenerate22KeyframeGenerate22PikaframesPost(BaseModel): - duration: Optional[int] = Field(None, ge=5, le=10) - negativePrompt: Optional[str] = Field(None) - promptText: str = Field(...) - resolution: Optional[str] = '1080p' - seed: Optional[int] = Field(None) - - -class PikaBodyGenerate22T2vGenerate22T2vPost(BaseModel): - aspectRatio: Optional[float] = Field( - 1.7777777777777777, - description='Aspect ratio (width / height)', - ge=0.4, - le=2.5, - ) - duration: Optional[int] = 5 - negativePrompt: Optional[str] = Field(None) - promptText: str = Field(...) - resolution: Optional[str] = '1080p' - seed: Optional[int] = Field(None) - - -class PikaBodyGeneratePikadditionsGeneratePikadditionsPost(BaseModel): - negativePrompt: Optional[str] = Field(None) - promptText: Optional[str] = Field(None) - seed: Optional[int] = Field(None) - - -class PikaBodyGeneratePikaffectsGeneratePikaffectsPost(BaseModel): - negativePrompt: Optional[str] = Field(None) - pikaffect: Optional[str] = None - promptText: Optional[str] = Field(None) - seed: Optional[int] = Field(None) - - -class PikaBodyGeneratePikaswapsGeneratePikaswapsPost(BaseModel): - negativePrompt: Optional[str] = Field(None) - promptText: Optional[str] = Field(None) - seed: Optional[int] = Field(None) - modifyRegionRoi: Optional[str] = Field(None) - - -class PikaStatusEnum(str, Enum): - queued = "queued" - started = "started" - finished = "finished" - failed = "failed" - - -class PikaVideoResponse(BaseModel): - id: str = Field(...) - progress: Optional[int] = Field(None) - status: PikaStatusEnum - url: Optional[str] = Field(None) diff --git a/comfy_api_nodes/nodes_pika.py b/comfy_api_nodes/nodes_pika.py deleted file mode 100644 index acd88c391..000000000 --- a/comfy_api_nodes/nodes_pika.py +++ /dev/null @@ -1,575 +0,0 @@ -""" -Pika x ComfyUI API Nodes - -Pika API docs: https://pika-827374fb.mintlify.app/api-reference -""" -from __future__ import annotations - -from io import BytesIO -import logging -from typing import Optional - -import torch - -from typing_extensions import override -from comfy_api.latest import ComfyExtension, IO -from comfy_api.input_impl.video_types import VideoCodec, VideoContainer, VideoInput -from comfy_api_nodes.apis import pika_api as pika_defs -from comfy_api_nodes.util import ( - validate_string, - download_url_to_video_output, - tensor_to_bytesio, - ApiEndpoint, - sync_op, - poll_op, -) - - -PATH_PIKADDITIONS = "/proxy/pika/generate/pikadditions" -PATH_PIKASWAPS = "/proxy/pika/generate/pikaswaps" -PATH_PIKAFFECTS = "/proxy/pika/generate/pikaffects" - -PIKA_API_VERSION = "2.2" -PATH_TEXT_TO_VIDEO = f"/proxy/pika/generate/{PIKA_API_VERSION}/t2v" -PATH_IMAGE_TO_VIDEO = f"/proxy/pika/generate/{PIKA_API_VERSION}/i2v" -PATH_PIKAFRAMES = f"/proxy/pika/generate/{PIKA_API_VERSION}/pikaframes" -PATH_PIKASCENES = f"/proxy/pika/generate/{PIKA_API_VERSION}/pikascenes" - -PATH_VIDEO_GET = "/proxy/pika/videos" - - -async def execute_task( - task_id: str, - cls: type[IO.ComfyNode], -) -> IO.NodeOutput: - final_response: pika_defs.PikaVideoResponse = await poll_op( - cls, - ApiEndpoint(path=f"{PATH_VIDEO_GET}/{task_id}"), - response_model=pika_defs.PikaVideoResponse, - status_extractor=lambda response: (response.status.value if response.status else None), - progress_extractor=lambda response: (response.progress if hasattr(response, "progress") else None), - estimated_duration=60, - max_poll_attempts=240, - ) - if not final_response.url: - error_msg = f"Pika task {task_id} succeeded but no video data found in response:\n{final_response}" - logging.error(error_msg) - raise Exception(error_msg) - video_url = final_response.url - logging.info("Pika task %s succeeded. Video URL: %s", task_id, video_url) - return IO.NodeOutput(await download_url_to_video_output(video_url)) - - -def get_base_inputs_types() -> list[IO.Input]: - """Get the base required inputs types common to all Pika nodes.""" - return [ - IO.String.Input("prompt_text", multiline=True), - IO.String.Input("negative_prompt", multiline=True), - IO.Int.Input("seed", min=0, max=0xFFFFFFFF, control_after_generate=True), - IO.Combo.Input("resolution", options=["1080p", "720p"], default="1080p"), - IO.Combo.Input("duration", options=[5, 10], default=5), - ] - - -class PikaImageToVideo(IO.ComfyNode): - """Pika 2.2 Image to Video Node.""" - - @classmethod - def define_schema(cls) -> IO.Schema: - return IO.Schema( - node_id="PikaImageToVideoNode2_2", - display_name="Pika Image to Video", - description="Sends an image and prompt to the Pika API v2.2 to generate a video.", - category="api node/video/Pika", - inputs=[ - IO.Image.Input("image", tooltip="The image to convert to video"), - *get_base_inputs_types(), - ], - outputs=[IO.Video.Output()], - hidden=[ - IO.Hidden.auth_token_comfy_org, - IO.Hidden.api_key_comfy_org, - IO.Hidden.unique_id, - ], - is_api_node=True, - is_deprecated=True, - ) - - @classmethod - async def execute( - cls, - image: torch.Tensor, - prompt_text: str, - negative_prompt: str, - seed: int, - resolution: str, - duration: int, - ) -> IO.NodeOutput: - image_bytes_io = tensor_to_bytesio(image) - pika_files = {"image": ("image.png", image_bytes_io, "image/png")} - pika_request_data = pika_defs.PikaBodyGenerate22I2vGenerate22I2vPost( - promptText=prompt_text, - negativePrompt=negative_prompt, - seed=seed, - resolution=resolution, - duration=duration, - ) - initial_operation = await sync_op( - cls, - ApiEndpoint(path=PATH_IMAGE_TO_VIDEO, method="POST"), - response_model=pika_defs.PikaGenerateResponse, - data=pika_request_data, - files=pika_files, - content_type="multipart/form-data", - ) - return await execute_task(initial_operation.video_id, cls) - - -class PikaTextToVideoNode(IO.ComfyNode): - """Pika Text2Video v2.2 Node.""" - - @classmethod - def define_schema(cls) -> IO.Schema: - return IO.Schema( - node_id="PikaTextToVideoNode2_2", - display_name="Pika Text to Video", - description="Sends a text prompt to the Pika API v2.2 to generate a video.", - category="api node/video/Pika", - inputs=[ - *get_base_inputs_types(), - IO.Float.Input( - "aspect_ratio", - step=0.001, - min=0.4, - max=2.5, - default=1.7777777777777777, - tooltip="Aspect ratio (width / height)", - ) - ], - outputs=[IO.Video.Output()], - hidden=[ - IO.Hidden.auth_token_comfy_org, - IO.Hidden.api_key_comfy_org, - IO.Hidden.unique_id, - ], - is_api_node=True, - is_deprecated=True, - ) - - @classmethod - async def execute( - cls, - prompt_text: str, - negative_prompt: str, - seed: int, - resolution: str, - duration: int, - aspect_ratio: float, - ) -> IO.NodeOutput: - initial_operation = await sync_op( - cls, - ApiEndpoint(path=PATH_TEXT_TO_VIDEO, method="POST"), - response_model=pika_defs.PikaGenerateResponse, - data=pika_defs.PikaBodyGenerate22T2vGenerate22T2vPost( - promptText=prompt_text, - negativePrompt=negative_prompt, - seed=seed, - resolution=resolution, - duration=duration, - aspectRatio=aspect_ratio, - ), - content_type="application/x-www-form-urlencoded", - ) - return await execute_task(initial_operation.video_id, cls) - - -class PikaScenes(IO.ComfyNode): - """PikaScenes v2.2 Node.""" - - @classmethod - def define_schema(cls) -> IO.Schema: - return IO.Schema( - node_id="PikaScenesV2_2", - display_name="Pika Scenes (Video Image Composition)", - description="Combine your images to create a video with the objects in them. Upload multiple images as ingredients and generate a high-quality video that incorporates all of them.", - category="api node/video/Pika", - inputs=[ - *get_base_inputs_types(), - IO.Combo.Input( - "ingredients_mode", - options=["creative", "precise"], - default="creative", - ), - IO.Float.Input( - "aspect_ratio", - step=0.001, - min=0.4, - max=2.5, - default=1.7777777777777777, - tooltip="Aspect ratio (width / height)", - ), - IO.Image.Input( - "image_ingredient_1", - optional=True, - tooltip="Image that will be used as ingredient to create a video.", - ), - IO.Image.Input( - "image_ingredient_2", - optional=True, - tooltip="Image that will be used as ingredient to create a video.", - ), - IO.Image.Input( - "image_ingredient_3", - optional=True, - tooltip="Image that will be used as ingredient to create a video.", - ), - IO.Image.Input( - "image_ingredient_4", - optional=True, - tooltip="Image that will be used as ingredient to create a video.", - ), - IO.Image.Input( - "image_ingredient_5", - optional=True, - tooltip="Image that will be used as ingredient to create a video.", - ), - ], - outputs=[IO.Video.Output()], - hidden=[ - IO.Hidden.auth_token_comfy_org, - IO.Hidden.api_key_comfy_org, - IO.Hidden.unique_id, - ], - is_api_node=True, - is_deprecated=True, - ) - - @classmethod - async def execute( - cls, - prompt_text: str, - negative_prompt: str, - seed: int, - resolution: str, - duration: int, - ingredients_mode: str, - aspect_ratio: float, - image_ingredient_1: Optional[torch.Tensor] = None, - image_ingredient_2: Optional[torch.Tensor] = None, - image_ingredient_3: Optional[torch.Tensor] = None, - image_ingredient_4: Optional[torch.Tensor] = None, - image_ingredient_5: Optional[torch.Tensor] = None, - ) -> IO.NodeOutput: - all_image_bytes_io = [] - for image in [ - image_ingredient_1, - image_ingredient_2, - image_ingredient_3, - image_ingredient_4, - image_ingredient_5, - ]: - if image is not None: - all_image_bytes_io.append(tensor_to_bytesio(image)) - - pika_files = [ - ("images", (f"image_{i}.png", image_bytes_io, "image/png")) - for i, image_bytes_io in enumerate(all_image_bytes_io) - ] - - pika_request_data = pika_defs.PikaBodyGenerate22C2vGenerate22PikascenesPost( - ingredientsMode=ingredients_mode, - promptText=prompt_text, - negativePrompt=negative_prompt, - seed=seed, - resolution=resolution, - duration=duration, - aspectRatio=aspect_ratio, - ) - initial_operation = await sync_op( - cls, - ApiEndpoint(path=PATH_PIKASCENES, method="POST"), - response_model=pika_defs.PikaGenerateResponse, - data=pika_request_data, - files=pika_files, - content_type="multipart/form-data", - ) - - return await execute_task(initial_operation.video_id, cls) - - -class PikAdditionsNode(IO.ComfyNode): - """Pika Pikadditions Node. Add an image into a video.""" - - @classmethod - def define_schema(cls) -> IO.Schema: - return IO.Schema( - node_id="Pikadditions", - display_name="Pikadditions (Video Object Insertion)", - description="Add any object or image into your video. Upload a video and specify what you'd like to add to create a seamlessly integrated result.", - category="api node/video/Pika", - inputs=[ - IO.Video.Input("video", tooltip="The video to add an image to."), - IO.Image.Input("image", tooltip="The image to add to the video."), - IO.String.Input("prompt_text", multiline=True), - IO.String.Input("negative_prompt", multiline=True), - IO.Int.Input( - "seed", - min=0, - max=0xFFFFFFFF, - control_after_generate=True, - ), - ], - outputs=[IO.Video.Output()], - hidden=[ - IO.Hidden.auth_token_comfy_org, - IO.Hidden.api_key_comfy_org, - IO.Hidden.unique_id, - ], - is_api_node=True, - is_deprecated=True, - ) - - @classmethod - async def execute( - cls, - video: VideoInput, - image: torch.Tensor, - prompt_text: str, - negative_prompt: str, - seed: int, - ) -> IO.NodeOutput: - video_bytes_io = BytesIO() - video.save_to(video_bytes_io, format=VideoContainer.MP4, codec=VideoCodec.H264) - video_bytes_io.seek(0) - - image_bytes_io = tensor_to_bytesio(image) - pika_files = { - "video": ("video.mp4", video_bytes_io, "video/mp4"), - "image": ("image.png", image_bytes_io, "image/png"), - } - pika_request_data = pika_defs.PikaBodyGeneratePikadditionsGeneratePikadditionsPost( - promptText=prompt_text, - negativePrompt=negative_prompt, - seed=seed, - ) - initial_operation = await sync_op( - cls, - ApiEndpoint(path=PATH_PIKADDITIONS, method="POST"), - response_model=pika_defs.PikaGenerateResponse, - data=pika_request_data, - files=pika_files, - content_type="multipart/form-data", - ) - - return await execute_task(initial_operation.video_id, cls) - - -class PikaSwapsNode(IO.ComfyNode): - """Pika Pikaswaps Node.""" - - @classmethod - def define_schema(cls) -> IO.Schema: - return IO.Schema( - node_id="Pikaswaps", - display_name="Pika Swaps (Video Object Replacement)", - description="Swap out any object or region of your video with a new image or object. Define areas to replace either with a mask or coordinates.", - category="api node/video/Pika", - inputs=[ - IO.Video.Input("video", tooltip="The video to swap an object in."), - IO.Image.Input( - "image", - tooltip="The image used to replace the masked object in the video.", - optional=True, - ), - IO.Mask.Input( - "mask", - tooltip="Use the mask to define areas in the video to replace.", - optional=True, - ), - IO.String.Input("prompt_text", multiline=True, optional=True), - IO.String.Input("negative_prompt", multiline=True, optional=True), - IO.Int.Input("seed", min=0, max=0xFFFFFFFF, control_after_generate=True, optional=True), - IO.String.Input( - "region_to_modify", - multiline=True, - optional=True, - tooltip="Plaintext description of the object / region to modify.", - ), - ], - outputs=[IO.Video.Output()], - hidden=[ - IO.Hidden.auth_token_comfy_org, - IO.Hidden.api_key_comfy_org, - IO.Hidden.unique_id, - ], - is_api_node=True, - is_deprecated=True, - ) - - @classmethod - async def execute( - cls, - video: VideoInput, - image: Optional[torch.Tensor] = None, - mask: Optional[torch.Tensor] = None, - prompt_text: str = "", - negative_prompt: str = "", - seed: int = 0, - region_to_modify: str = "", - ) -> IO.NodeOutput: - video_bytes_io = BytesIO() - video.save_to(video_bytes_io, format=VideoContainer.MP4, codec=VideoCodec.H264) - video_bytes_io.seek(0) - pika_files = { - "video": ("video.mp4", video_bytes_io, "video/mp4"), - } - if mask is not None: - pika_files["modifyRegionMask"] = ("mask.png", tensor_to_bytesio(mask), "image/png") - if image is not None: - pika_files["image"] = ("image.png", tensor_to_bytesio(image), "image/png") - - pika_request_data = pika_defs.PikaBodyGeneratePikaswapsGeneratePikaswapsPost( - promptText=prompt_text, - negativePrompt=negative_prompt, - seed=seed, - modifyRegionRoi=region_to_modify if region_to_modify else None, - ) - initial_operation = await sync_op( - cls, - ApiEndpoint(path=PATH_PIKASWAPS, method="POST"), - response_model=pika_defs.PikaGenerateResponse, - data=pika_request_data, - files=pika_files, - content_type="multipart/form-data", - ) - return await execute_task(initial_operation.video_id, cls) - - -class PikaffectsNode(IO.ComfyNode): - """Pika Pikaffects Node.""" - - @classmethod - def define_schema(cls) -> IO.Schema: - return IO.Schema( - node_id="Pikaffects", - display_name="Pikaffects (Video Effects)", - description="Generate a video with a specific Pikaffect. Supported Pikaffects: Cake-ify, Crumble, Crush, Decapitate, Deflate, Dissolve, Explode, Eye-pop, Inflate, Levitate, Melt, Peel, Poke, Squish, Ta-da, Tear", - category="api node/video/Pika", - inputs=[ - IO.Image.Input("image", tooltip="The reference image to apply the Pikaffect to."), - IO.Combo.Input( - "pikaffect", options=pika_defs.Pikaffect, default="Cake-ify" - ), - IO.String.Input("prompt_text", multiline=True), - IO.String.Input("negative_prompt", multiline=True), - IO.Int.Input("seed", min=0, max=0xFFFFFFFF, control_after_generate=True), - ], - outputs=[IO.Video.Output()], - hidden=[ - IO.Hidden.auth_token_comfy_org, - IO.Hidden.api_key_comfy_org, - IO.Hidden.unique_id, - ], - is_api_node=True, - is_deprecated=True, - ) - - @classmethod - async def execute( - cls, - image: torch.Tensor, - pikaffect: str, - prompt_text: str, - negative_prompt: str, - seed: int, - ) -> IO.NodeOutput: - initial_operation = await sync_op( - cls, - ApiEndpoint(path=PATH_PIKAFFECTS, method="POST"), - response_model=pika_defs.PikaGenerateResponse, - data=pika_defs.PikaBodyGeneratePikaffectsGeneratePikaffectsPost( - pikaffect=pikaffect, - promptText=prompt_text, - negativePrompt=negative_prompt, - seed=seed, - ), - files={"image": ("image.png", tensor_to_bytesio(image), "image/png")}, - content_type="multipart/form-data", - ) - return await execute_task(initial_operation.video_id, cls) - - -class PikaStartEndFrameNode(IO.ComfyNode): - """PikaFrames v2.2 Node.""" - - @classmethod - def define_schema(cls) -> IO.Schema: - return IO.Schema( - node_id="PikaStartEndFrameNode2_2", - display_name="Pika Start and End Frame to Video", - description="Generate a video by combining your first and last frame. Upload two images to define the start and end points, and let the AI create a smooth transition between them.", - category="api node/video/Pika", - inputs=[ - IO.Image.Input("image_start", tooltip="The first image to combine."), - IO.Image.Input("image_end", tooltip="The last image to combine."), - *get_base_inputs_types(), - ], - outputs=[IO.Video.Output()], - hidden=[ - IO.Hidden.auth_token_comfy_org, - IO.Hidden.api_key_comfy_org, - IO.Hidden.unique_id, - ], - is_api_node=True, - is_deprecated=True, - ) - - @classmethod - async def execute( - cls, - image_start: torch.Tensor, - image_end: torch.Tensor, - prompt_text: str, - negative_prompt: str, - seed: int, - resolution: str, - duration: int, - ) -> IO.NodeOutput: - validate_string(prompt_text, field_name="prompt_text", min_length=1) - pika_files = [ - ("keyFrames", ("image_start.png", tensor_to_bytesio(image_start), "image/png")), - ("keyFrames", ("image_end.png", tensor_to_bytesio(image_end), "image/png")), - ] - initial_operation = await sync_op( - cls, - ApiEndpoint(path=PATH_PIKAFRAMES, method="POST"), - response_model=pika_defs.PikaGenerateResponse, - data=pika_defs.PikaBodyGenerate22KeyframeGenerate22PikaframesPost( - promptText=prompt_text, - negativePrompt=negative_prompt, - seed=seed, - resolution=resolution, - duration=duration, - ), - files=pika_files, - content_type="multipart/form-data", - ) - return await execute_task(initial_operation.video_id, cls) - - -class PikaApiNodesExtension(ComfyExtension): - @override - async def get_node_list(self) -> list[type[IO.ComfyNode]]: - return [ - PikaImageToVideo, - PikaTextToVideoNode, - PikaScenes, - PikAdditionsNode, - PikaSwapsNode, - PikaffectsNode, - PikaStartEndFrameNode, - ] - - -async def comfy_entrypoint() -> PikaApiNodesExtension: - return PikaApiNodesExtension() diff --git a/nodes.py b/nodes.py index 8678f510a..3fa543294 100644 --- a/nodes.py +++ b/nodes.py @@ -2384,7 +2384,6 @@ async def init_builtin_api_nodes(): "nodes_recraft.py", "nodes_pixverse.py", "nodes_stability.py", - "nodes_pika.py", "nodes_runway.py", "nodes_sora.py", "nodes_topaz.py", From dbd330454ada04609c69fda2ae7c002d7ea05f67 Mon Sep 17 00:00:00 2001 From: "Dr.Lt.Data" <128333288+ltdrdata@users.noreply.github.com> Date: Tue, 16 Dec 2025 08:57:39 +0900 Subject: [PATCH 287/647] feat(preview): add per-queue live preview method override (#11261) - Add set_preview_method() to override live preview method per queue item - Read extra_data.preview_method from /prompt request - Support values: taesd, latent2rgb, none, auto, default - "default" or unset uses server's CLI --preview-method setting - Add 44 tests (37 unit + 7 E2E) --- comfy/cli_args.py | 7 + execution.py | 3 + latent_preview.py | 10 + .../preview_method_override_test.py | 352 +++++++++++++++++ tests/execution/test_preview_method.py | 358 ++++++++++++++++++ 5 files changed, 730 insertions(+) create mode 100644 tests-unit/execution_test/preview_method_override_test.py create mode 100644 tests/execution/test_preview_method.py diff --git a/comfy/cli_args.py b/comfy/cli_args.py index 209fc185b..dae9a895d 100644 --- a/comfy/cli_args.py +++ b/comfy/cli_args.py @@ -97,6 +97,13 @@ class LatentPreviewMethod(enum.Enum): Latent2RGB = "latent2rgb" TAESD = "taesd" + @classmethod + def from_string(cls, value: str): + for member in cls: + if member.value == value: + return member + return None + parser.add_argument("--preview-method", type=LatentPreviewMethod, default=LatentPreviewMethod.NoPreviews, help="Default preview method for sampler nodes.", action=EnumAction) parser.add_argument("--preview-size", type=int, default=512, help="Sets the maximum preview size for sampler nodes.") diff --git a/execution.py b/execution.py index c2186ac98..0c239efd7 100644 --- a/execution.py +++ b/execution.py @@ -13,6 +13,7 @@ import asyncio import torch import comfy.model_management +from latent_preview import set_preview_method import nodes from comfy_execution.caching import ( BasicCache, @@ -669,6 +670,8 @@ class PromptExecutor: asyncio.run(self.execute_async(prompt, prompt_id, extra_data, execute_outputs)) async def execute_async(self, prompt, prompt_id, extra_data={}, execute_outputs=[]): + set_preview_method(extra_data.get("preview_method")) + nodes.interrupt_processing(False) if "client_id" in extra_data: diff --git a/latent_preview.py b/latent_preview.py index 66bded4b9..d52e3f7a1 100644 --- a/latent_preview.py +++ b/latent_preview.py @@ -8,6 +8,8 @@ import folder_paths import comfy.utils import logging +default_preview_method = args.preview_method + MAX_PREVIEW_RESOLUTION = args.preview_size VIDEO_TAES = ["taehv", "lighttaew2_2", "lighttaew2_1", "lighttaehy1_5"] @@ -125,3 +127,11 @@ def prepare_callback(model, steps, x0_output_dict=None): pbar.update_absolute(step + 1, total_steps, preview_bytes) return callback +def set_preview_method(override: str = None): + if override and override != "default": + method = LatentPreviewMethod.from_string(override) + if method is not None: + args.preview_method = method + return + args.preview_method = default_preview_method + diff --git a/tests-unit/execution_test/preview_method_override_test.py b/tests-unit/execution_test/preview_method_override_test.py new file mode 100644 index 000000000..79432d610 --- /dev/null +++ b/tests-unit/execution_test/preview_method_override_test.py @@ -0,0 +1,352 @@ +""" +Unit tests for Queue-specific Preview Method Override feature. + +Tests the preview method override functionality: +- LatentPreviewMethod.from_string() method +- set_preview_method() function in latent_preview.py +- default_preview_method variable +- Integration with args.preview_method +""" +import pytest +from comfy.cli_args import args, LatentPreviewMethod +from latent_preview import set_preview_method, default_preview_method + + +class TestLatentPreviewMethodFromString: + """Test LatentPreviewMethod.from_string() classmethod.""" + + @pytest.mark.parametrize("value,expected", [ + ("auto", LatentPreviewMethod.Auto), + ("latent2rgb", LatentPreviewMethod.Latent2RGB), + ("taesd", LatentPreviewMethod.TAESD), + ("none", LatentPreviewMethod.NoPreviews), + ]) + def test_valid_values_return_enum(self, value, expected): + """Valid string values should return corresponding enum.""" + assert LatentPreviewMethod.from_string(value) == expected + + @pytest.mark.parametrize("invalid", [ + "invalid", + "TAESD", # Case sensitive + "AUTO", # Case sensitive + "Latent2RGB", # Case sensitive + "latent", + "", + "default", # default is special, not a method + ]) + def test_invalid_values_return_none(self, invalid): + """Invalid string values should return None.""" + assert LatentPreviewMethod.from_string(invalid) is None + + +class TestLatentPreviewMethodEnumValues: + """Test LatentPreviewMethod enum has expected values.""" + + def test_enum_values(self): + """Verify enum values match expected strings.""" + assert LatentPreviewMethod.NoPreviews.value == "none" + assert LatentPreviewMethod.Auto.value == "auto" + assert LatentPreviewMethod.Latent2RGB.value == "latent2rgb" + assert LatentPreviewMethod.TAESD.value == "taesd" + + def test_enum_count(self): + """Verify exactly 4 preview methods exist.""" + assert len(LatentPreviewMethod) == 4 + + +class TestSetPreviewMethod: + """Test set_preview_method() function from latent_preview.py.""" + + def setup_method(self): + """Store original value before each test.""" + self.original = args.preview_method + + def teardown_method(self): + """Restore original value after each test.""" + args.preview_method = self.original + + def test_override_with_taesd(self): + """'taesd' should set args.preview_method to TAESD.""" + set_preview_method("taesd") + assert args.preview_method == LatentPreviewMethod.TAESD + + def test_override_with_latent2rgb(self): + """'latent2rgb' should set args.preview_method to Latent2RGB.""" + set_preview_method("latent2rgb") + assert args.preview_method == LatentPreviewMethod.Latent2RGB + + def test_override_with_auto(self): + """'auto' should set args.preview_method to Auto.""" + set_preview_method("auto") + assert args.preview_method == LatentPreviewMethod.Auto + + def test_override_with_none_value(self): + """'none' should set args.preview_method to NoPreviews.""" + set_preview_method("none") + assert args.preview_method == LatentPreviewMethod.NoPreviews + + def test_default_restores_original(self): + """'default' should restore to default_preview_method.""" + # First override to something else + set_preview_method("taesd") + assert args.preview_method == LatentPreviewMethod.TAESD + + # Then use 'default' to restore + set_preview_method("default") + assert args.preview_method == default_preview_method + + def test_none_param_restores_original(self): + """None parameter should restore to default_preview_method.""" + # First override to something else + set_preview_method("taesd") + assert args.preview_method == LatentPreviewMethod.TAESD + + # Then use None to restore + set_preview_method(None) + assert args.preview_method == default_preview_method + + def test_empty_string_restores_original(self): + """Empty string should restore to default_preview_method.""" + set_preview_method("taesd") + set_preview_method("") + assert args.preview_method == default_preview_method + + def test_invalid_value_restores_original(self): + """Invalid value should restore to default_preview_method.""" + set_preview_method("taesd") + set_preview_method("invalid_method") + assert args.preview_method == default_preview_method + + def test_case_sensitive_invalid_restores(self): + """Case-mismatched values should restore to default.""" + set_preview_method("taesd") + set_preview_method("TAESD") # Wrong case + assert args.preview_method == default_preview_method + + +class TestDefaultPreviewMethod: + """Test default_preview_method module variable.""" + + def test_default_is_not_none(self): + """default_preview_method should not be None.""" + assert default_preview_method is not None + + def test_default_is_enum_member(self): + """default_preview_method should be a LatentPreviewMethod enum.""" + assert isinstance(default_preview_method, LatentPreviewMethod) + + def test_default_matches_args_initial(self): + """default_preview_method should match CLI default or user setting.""" + # This tests that default_preview_method was captured at module load + # After set_preview_method(None), args should equal default + original = args.preview_method + set_preview_method("taesd") + set_preview_method(None) + assert args.preview_method == default_preview_method + args.preview_method = original + + +class TestArgsPreviewMethodModification: + """Test args.preview_method can be modified correctly.""" + + def setup_method(self): + """Store original value before each test.""" + self.original = args.preview_method + + def teardown_method(self): + """Restore original value after each test.""" + args.preview_method = self.original + + def test_args_accepts_all_enum_values(self): + """args.preview_method should accept all LatentPreviewMethod values.""" + for method in LatentPreviewMethod: + args.preview_method = method + assert args.preview_method == method + + def test_args_modification_and_restoration(self): + """args.preview_method should be modifiable and restorable.""" + original = args.preview_method + + args.preview_method = LatentPreviewMethod.TAESD + assert args.preview_method == LatentPreviewMethod.TAESD + + args.preview_method = original + assert args.preview_method == original + + +class TestExecutionFlow: + """Test the execution flow pattern used in execution.py.""" + + def setup_method(self): + """Store original value before each test.""" + self.original = args.preview_method + + def teardown_method(self): + """Restore original value after each test.""" + args.preview_method = self.original + + def test_sequential_executions_with_different_methods(self): + """Simulate multiple queue executions with different preview methods.""" + # Execution 1: taesd + set_preview_method("taesd") + assert args.preview_method == LatentPreviewMethod.TAESD + + # Execution 2: none + set_preview_method("none") + assert args.preview_method == LatentPreviewMethod.NoPreviews + + # Execution 3: default (restore) + set_preview_method("default") + assert args.preview_method == default_preview_method + + # Execution 4: auto + set_preview_method("auto") + assert args.preview_method == LatentPreviewMethod.Auto + + # Execution 5: no override (None) + set_preview_method(None) + assert args.preview_method == default_preview_method + + def test_override_then_default_pattern(self): + """Test the pattern: override -> execute -> next call restores.""" + # First execution with override + set_preview_method("latent2rgb") + assert args.preview_method == LatentPreviewMethod.Latent2RGB + + # Second execution without override restores default + set_preview_method(None) + assert args.preview_method == default_preview_method + + def test_extra_data_simulation(self): + """Simulate extra_data.get('preview_method') patterns.""" + # Simulate: extra_data = {"preview_method": "taesd"} + extra_data = {"preview_method": "taesd"} + set_preview_method(extra_data.get("preview_method")) + assert args.preview_method == LatentPreviewMethod.TAESD + + # Simulate: extra_data = {} + extra_data = {} + set_preview_method(extra_data.get("preview_method")) + assert args.preview_method == default_preview_method + + # Simulate: extra_data = {"preview_method": "default"} + extra_data = {"preview_method": "default"} + set_preview_method(extra_data.get("preview_method")) + assert args.preview_method == default_preview_method + + +class TestRealWorldScenarios: + """Tests using real-world prompt data patterns.""" + + def setup_method(self): + """Store original value before each test.""" + self.original = args.preview_method + + def teardown_method(self): + """Restore original value after each test.""" + args.preview_method = self.original + + def test_captured_prompt_without_preview_method(self): + """ + Test with captured prompt that has no preview_method. + Based on: tests-unit/execution_test/fixtures/default_prompt.json + """ + # Real captured extra_data structure (preview_method absent) + extra_data = { + "extra_pnginfo": {"workflow": {}}, + "client_id": "271314f0dabd48e5aaa488ed7a4ceb0d", + "create_time": 1765416558179 + } + + set_preview_method(extra_data.get("preview_method")) + assert args.preview_method == default_preview_method + + def test_captured_prompt_with_preview_method_taesd(self): + """Test captured prompt with preview_method: taesd.""" + extra_data = { + "extra_pnginfo": {"workflow": {}}, + "client_id": "271314f0dabd48e5aaa488ed7a4ceb0d", + "preview_method": "taesd" + } + + set_preview_method(extra_data.get("preview_method")) + assert args.preview_method == LatentPreviewMethod.TAESD + + def test_captured_prompt_with_preview_method_none(self): + """Test captured prompt with preview_method: none (disable preview).""" + extra_data = { + "extra_pnginfo": {"workflow": {}}, + "client_id": "test-client", + "preview_method": "none" + } + + set_preview_method(extra_data.get("preview_method")) + assert args.preview_method == LatentPreviewMethod.NoPreviews + + def test_captured_prompt_with_preview_method_latent2rgb(self): + """Test captured prompt with preview_method: latent2rgb.""" + extra_data = { + "extra_pnginfo": {"workflow": {}}, + "client_id": "test-client", + "preview_method": "latent2rgb" + } + + set_preview_method(extra_data.get("preview_method")) + assert args.preview_method == LatentPreviewMethod.Latent2RGB + + def test_captured_prompt_with_preview_method_auto(self): + """Test captured prompt with preview_method: auto.""" + extra_data = { + "extra_pnginfo": {"workflow": {}}, + "client_id": "test-client", + "preview_method": "auto" + } + + set_preview_method(extra_data.get("preview_method")) + assert args.preview_method == LatentPreviewMethod.Auto + + def test_captured_prompt_with_preview_method_default(self): + """Test captured prompt with preview_method: default (use CLI setting).""" + # First set to something else + set_preview_method("taesd") + assert args.preview_method == LatentPreviewMethod.TAESD + + # Then simulate a prompt with "default" + extra_data = { + "extra_pnginfo": {"workflow": {}}, + "client_id": "test-client", + "preview_method": "default" + } + + set_preview_method(extra_data.get("preview_method")) + assert args.preview_method == default_preview_method + + def test_sequential_queue_with_different_preview_methods(self): + """ + Simulate real queue scenario: multiple prompts with different settings. + This tests the actual usage pattern in ComfyUI. + """ + # Queue 1: User wants TAESD preview + extra_data_1 = {"client_id": "client-1", "preview_method": "taesd"} + set_preview_method(extra_data_1.get("preview_method")) + assert args.preview_method == LatentPreviewMethod.TAESD + + # Queue 2: User wants no preview (faster execution) + extra_data_2 = {"client_id": "client-2", "preview_method": "none"} + set_preview_method(extra_data_2.get("preview_method")) + assert args.preview_method == LatentPreviewMethod.NoPreviews + + # Queue 3: User doesn't specify (use server default) + extra_data_3 = {"client_id": "client-3"} + set_preview_method(extra_data_3.get("preview_method")) + assert args.preview_method == default_preview_method + + # Queue 4: User explicitly wants default + extra_data_4 = {"client_id": "client-4", "preview_method": "default"} + set_preview_method(extra_data_4.get("preview_method")) + assert args.preview_method == default_preview_method + + # Queue 5: User wants latent2rgb + extra_data_5 = {"client_id": "client-5", "preview_method": "latent2rgb"} + set_preview_method(extra_data_5.get("preview_method")) + assert args.preview_method == LatentPreviewMethod.Latent2RGB diff --git a/tests/execution/test_preview_method.py b/tests/execution/test_preview_method.py new file mode 100644 index 000000000..c3037553b --- /dev/null +++ b/tests/execution/test_preview_method.py @@ -0,0 +1,358 @@ +""" +E2E tests for Queue-specific Preview Method Override feature. + +Tests actual execution with different preview_method values. +Requires a running ComfyUI server with models. + +Usage: + COMFYUI_SERVER=http://localhost:8988 pytest test_preview_method_e2e.py -v -m preview_method + +Note: + These tests execute actual image generation and wait for completion. + Tests verify preview image transmission based on preview_method setting. +""" +import os +import json +import pytest +import uuid +import time +import random +import websocket +import urllib.request +from pathlib import Path + + +# Server configuration +SERVER_URL = os.environ.get("COMFYUI_SERVER", "http://localhost:8988") +SERVER_HOST = SERVER_URL.replace("http://", "").replace("https://", "") + +# Use existing inference graph fixture +GRAPH_FILE = Path(__file__).parent.parent / "inference" / "graphs" / "default_graph_sdxl1_0.json" + + +def is_server_running() -> bool: + """Check if ComfyUI server is running.""" + try: + request = urllib.request.Request(f"{SERVER_URL}/system_stats") + with urllib.request.urlopen(request, timeout=2.0): + return True + except Exception: + return False + + +def prepare_graph_for_test(graph: dict, steps: int = 5) -> dict: + """Prepare graph for testing: randomize seeds and reduce steps.""" + adapted = json.loads(json.dumps(graph)) # Deep copy + for node_id, node in adapted.items(): + inputs = node.get("inputs", {}) + # Handle both "seed" and "noise_seed" (used by KSamplerAdvanced) + if "seed" in inputs: + inputs["seed"] = random.randint(0, 2**32 - 1) + if "noise_seed" in inputs: + inputs["noise_seed"] = random.randint(0, 2**32 - 1) + # Reduce steps for faster testing (default 20 -> 5) + if "steps" in inputs: + inputs["steps"] = steps + return adapted + + +# Alias for backward compatibility +randomize_seed = prepare_graph_for_test + + +class PreviewMethodClient: + """Client for testing preview_method with WebSocket execution tracking.""" + + def __init__(self, server_address: str): + self.server_address = server_address + self.client_id = str(uuid.uuid4()) + self.ws = None + + def connect(self): + """Connect to WebSocket.""" + self.ws = websocket.WebSocket() + self.ws.settimeout(120) # 2 minute timeout for sampling + self.ws.connect(f"ws://{self.server_address}/ws?clientId={self.client_id}") + + def close(self): + """Close WebSocket connection.""" + if self.ws: + self.ws.close() + + def queue_prompt(self, prompt: dict, extra_data: dict = None) -> dict: + """Queue a prompt and return response with prompt_id.""" + data = { + "prompt": prompt, + "client_id": self.client_id, + "extra_data": extra_data or {} + } + req = urllib.request.Request( + f"http://{self.server_address}/prompt", + data=json.dumps(data).encode("utf-8"), + headers={"Content-Type": "application/json"} + ) + return json.loads(urllib.request.urlopen(req).read()) + + def wait_for_execution(self, prompt_id: str, timeout: float = 120.0) -> dict: + """ + Wait for execution to complete via WebSocket. + + Returns: + dict with keys: completed, error, preview_count, execution_time + """ + result = { + "completed": False, + "error": None, + "preview_count": 0, + "execution_time": 0.0 + } + + start_time = time.time() + self.ws.settimeout(timeout) + + try: + while True: + out = self.ws.recv() + elapsed = time.time() - start_time + + if isinstance(out, str): + message = json.loads(out) + msg_type = message.get("type") + data = message.get("data", {}) + + if data.get("prompt_id") != prompt_id: + continue + + if msg_type == "executing": + if data.get("node") is None: + # Execution complete + result["completed"] = True + result["execution_time"] = elapsed + break + + elif msg_type == "execution_error": + result["error"] = data + result["execution_time"] = elapsed + break + + elif msg_type == "progress": + # Progress update during sampling + pass + + elif isinstance(out, bytes): + # Binary data = preview image + result["preview_count"] += 1 + + except websocket.WebSocketTimeoutException: + result["error"] = "Timeout waiting for execution" + result["execution_time"] = time.time() - start_time + + return result + + +def load_graph() -> dict: + """Load the SDXL graph fixture with randomized seed.""" + with open(GRAPH_FILE) as f: + graph = json.load(f) + return randomize_seed(graph) # Avoid caching + + +# Skip all tests if server is not running +pytestmark = [ + pytest.mark.skipif( + not is_server_running(), + reason=f"ComfyUI server not running at {SERVER_URL}" + ), + pytest.mark.preview_method, + pytest.mark.execution, +] + + +@pytest.fixture +def client(): + """Create and connect a test client.""" + c = PreviewMethodClient(SERVER_HOST) + c.connect() + yield c + c.close() + + +@pytest.fixture +def graph(): + """Load the test graph.""" + return load_graph() + + +class TestPreviewMethodExecution: + """Test actual execution with different preview methods.""" + + def test_execution_with_latent2rgb(self, client, graph): + """ + Execute with preview_method=latent2rgb. + Should complete and potentially receive preview images. + """ + extra_data = {"preview_method": "latent2rgb"} + + response = client.queue_prompt(graph, extra_data) + assert "prompt_id" in response + + result = client.wait_for_execution(response["prompt_id"]) + + # Should complete (may error if model missing, but that's separate) + assert result["completed"] or result["error"] is not None + # Execution should take some time (sampling) + if result["completed"]: + assert result["execution_time"] > 0.5, "Execution too fast - likely didn't run" + # latent2rgb should produce previews + print(f"latent2rgb: {result['preview_count']} previews in {result['execution_time']:.2f}s") # noqa: T201 + + def test_execution_with_taesd(self, client, graph): + """ + Execute with preview_method=taesd. + TAESD provides higher quality previews. + """ + extra_data = {"preview_method": "taesd"} + + response = client.queue_prompt(graph, extra_data) + assert "prompt_id" in response + + result = client.wait_for_execution(response["prompt_id"]) + + assert result["completed"] or result["error"] is not None + if result["completed"]: + assert result["execution_time"] > 0.5 + # taesd should also produce previews + print(f"taesd: {result['preview_count']} previews in {result['execution_time']:.2f}s") # noqa: T201 + + def test_execution_with_none_preview(self, client, graph): + """ + Execute with preview_method=none. + No preview images should be generated. + """ + extra_data = {"preview_method": "none"} + + response = client.queue_prompt(graph, extra_data) + assert "prompt_id" in response + + result = client.wait_for_execution(response["prompt_id"]) + + assert result["completed"] or result["error"] is not None + if result["completed"]: + # With "none", should receive no preview images + assert result["preview_count"] == 0, \ + f"Expected no previews with 'none', got {result['preview_count']}" + print(f"none: {result['preview_count']} previews in {result['execution_time']:.2f}s") # noqa: T201 + + def test_execution_with_default(self, client, graph): + """ + Execute with preview_method=default. + Should use server's CLI default setting. + """ + extra_data = {"preview_method": "default"} + + response = client.queue_prompt(graph, extra_data) + assert "prompt_id" in response + + result = client.wait_for_execution(response["prompt_id"]) + + assert result["completed"] or result["error"] is not None + if result["completed"]: + print(f"default: {result['preview_count']} previews in {result['execution_time']:.2f}s") # noqa: T201 + + def test_execution_without_preview_method(self, client, graph): + """ + Execute without preview_method in extra_data. + Should use server's default preview method. + """ + extra_data = {} # No preview_method + + response = client.queue_prompt(graph, extra_data) + assert "prompt_id" in response + + result = client.wait_for_execution(response["prompt_id"]) + + assert result["completed"] or result["error"] is not None + if result["completed"]: + print(f"(no override): {result['preview_count']} previews in {result['execution_time']:.2f}s") # noqa: T201 + + +class TestPreviewMethodComparison: + """Compare preview behavior between different methods.""" + + def test_none_vs_latent2rgb_preview_count(self, client, graph): + """ + Compare preview counts: 'none' should have 0, others should have >0. + This is the key verification that preview_method actually works. + """ + results = {} + + # Run with none (randomize seed to avoid caching) + graph_none = randomize_seed(graph) + extra_data_none = {"preview_method": "none"} + response = client.queue_prompt(graph_none, extra_data_none) + results["none"] = client.wait_for_execution(response["prompt_id"]) + + # Run with latent2rgb (randomize seed again) + graph_rgb = randomize_seed(graph) + extra_data_rgb = {"preview_method": "latent2rgb"} + response = client.queue_prompt(graph_rgb, extra_data_rgb) + results["latent2rgb"] = client.wait_for_execution(response["prompt_id"]) + + # Verify both completed + assert results["none"]["completed"], f"'none' execution failed: {results['none']['error']}" + assert results["latent2rgb"]["completed"], f"'latent2rgb' execution failed: {results['latent2rgb']['error']}" + + # Key assertion: 'none' should have 0 previews + assert results["none"]["preview_count"] == 0, \ + f"'none' should have 0 previews, got {results['none']['preview_count']}" + + # 'latent2rgb' should have at least 1 preview (depends on steps) + assert results["latent2rgb"]["preview_count"] > 0, \ + f"'latent2rgb' should have >0 previews, got {results['latent2rgb']['preview_count']}" + + print("\nPreview count comparison:") # noqa: T201 + print(f" none: {results['none']['preview_count']} previews") # noqa: T201 + print(f" latent2rgb: {results['latent2rgb']['preview_count']} previews") # noqa: T201 + + +class TestPreviewMethodSequential: + """Test sequential execution with different preview methods.""" + + def test_sequential_different_methods(self, client, graph): + """ + Execute multiple prompts sequentially with different preview methods. + Each should complete independently with correct preview behavior. + """ + methods = ["latent2rgb", "none", "default"] + results = [] + + for method in methods: + # Randomize seed for each execution to avoid caching + graph_run = randomize_seed(graph) + extra_data = {"preview_method": method} + response = client.queue_prompt(graph_run, extra_data) + + result = client.wait_for_execution(response["prompt_id"]) + results.append({ + "method": method, + "completed": result["completed"], + "preview_count": result["preview_count"], + "execution_time": result["execution_time"], + "error": result["error"] + }) + + # All should complete or have clear errors + for r in results: + assert r["completed"] or r["error"] is not None, \ + f"Method {r['method']} neither completed nor errored" + + # "none" should have zero previews if completed + none_result = next(r for r in results if r["method"] == "none") + if none_result["completed"]: + assert none_result["preview_count"] == 0, \ + f"'none' should have 0 previews, got {none_result['preview_count']}" + + print("\nSequential execution results:") # noqa: T201 + for r in results: + status = "✓" if r["completed"] else f"✗ ({r['error']})" + print(f" {r['method']}: {status}, {r['preview_count']} previews, {r['execution_time']:.2f}s") # noqa: T201 From 43e0d4e3ccfe8b5eac81bcee6f912f661849aafb Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Tue, 16 Dec 2025 02:01:10 +0200 Subject: [PATCH 288/647] comfy_api: remove usage of "Type","List" and "Dict" types (#11238) --- comfy_api/feature_flags.py | 10 +++++----- comfy_api/internal/api_registry.py | 10 +++++----- comfy_api/internal/async_to_sync.py | 14 ++++++------- comfy_api/internal/singleton.py | 6 +++--- comfy_api/latest/__init__.py | 4 ++-- comfy_api/latest/_input/basic_types.py | 4 ++-- comfy_api/latest/_ui.py | 27 +++++++++++++------------- comfy_api/version_list.py | 3 +-- 8 files changed, 38 insertions(+), 40 deletions(-) diff --git a/comfy_api/feature_flags.py b/comfy_api/feature_flags.py index bfb77eb5f..de167f037 100644 --- a/comfy_api/feature_flags.py +++ b/comfy_api/feature_flags.py @@ -5,12 +5,12 @@ This module handles capability negotiation between frontend and backend, allowing graceful protocol evolution while maintaining backward compatibility. """ -from typing import Any, Dict +from typing import Any from comfy.cli_args import args # Default server capabilities -SERVER_FEATURE_FLAGS: Dict[str, Any] = { +SERVER_FEATURE_FLAGS: dict[str, Any] = { "supports_preview_metadata": True, "max_upload_size": args.max_upload_size * 1024 * 1024, # Convert MB to bytes "extension": {"manager": {"supports_v4": True}}, @@ -18,7 +18,7 @@ SERVER_FEATURE_FLAGS: Dict[str, Any] = { def get_connection_feature( - sockets_metadata: Dict[str, Dict[str, Any]], + sockets_metadata: dict[str, dict[str, Any]], sid: str, feature_name: str, default: Any = False @@ -42,7 +42,7 @@ def get_connection_feature( def supports_feature( - sockets_metadata: Dict[str, Dict[str, Any]], + sockets_metadata: dict[str, dict[str, Any]], sid: str, feature_name: str ) -> bool: @@ -60,7 +60,7 @@ def supports_feature( return get_connection_feature(sockets_metadata, sid, feature_name, False) is True -def get_server_features() -> Dict[str, Any]: +def get_server_features() -> dict[str, Any]: """ Get the server's feature flags. diff --git a/comfy_api/internal/api_registry.py b/comfy_api/internal/api_registry.py index 7e3375cf6..2b1cb016a 100644 --- a/comfy_api/internal/api_registry.py +++ b/comfy_api/internal/api_registry.py @@ -1,4 +1,4 @@ -from typing import Type, List, NamedTuple +from typing import NamedTuple from comfy_api.internal.singleton import ProxiedSingleton from packaging import version as packaging_version @@ -10,7 +10,7 @@ class ComfyAPIBase(ProxiedSingleton): class ComfyAPIWithVersion(NamedTuple): version: str - api_class: Type[ComfyAPIBase] + api_class: type[ComfyAPIBase] def parse_version(version_str: str) -> packaging_version.Version: @@ -23,16 +23,16 @@ def parse_version(version_str: str) -> packaging_version.Version: return packaging_version.parse(version_str) -registered_versions: List[ComfyAPIWithVersion] = [] +registered_versions: list[ComfyAPIWithVersion] = [] -def register_versions(versions: List[ComfyAPIWithVersion]): +def register_versions(versions: list[ComfyAPIWithVersion]): versions.sort(key=lambda x: parse_version(x.version)) global registered_versions registered_versions = versions -def get_all_versions() -> List[ComfyAPIWithVersion]: +def get_all_versions() -> list[ComfyAPIWithVersion]: """ Returns a list of all registered ComfyAPI versions. """ diff --git a/comfy_api/internal/async_to_sync.py b/comfy_api/internal/async_to_sync.py index 257ade82e..c9b0576e1 100644 --- a/comfy_api/internal/async_to_sync.py +++ b/comfy_api/internal/async_to_sync.py @@ -8,7 +8,7 @@ import os import textwrap import threading from enum import Enum -from typing import Optional, Type, get_origin, get_args, get_type_hints +from typing import Optional, get_origin, get_args, get_type_hints class TypeTracker: @@ -193,7 +193,7 @@ class AsyncToSyncConverter: return result_container["result"] @classmethod - def create_sync_class(cls, async_class: Type, thread_pool_size=10) -> Type: + def create_sync_class(cls, async_class: type, thread_pool_size=10) -> type: """ Creates a new class with synchronous versions of all async methods. @@ -563,7 +563,7 @@ class AsyncToSyncConverter: @classmethod def _generate_imports( - cls, async_class: Type, type_tracker: TypeTracker + cls, async_class: type, type_tracker: TypeTracker ) -> list[str]: """Generate import statements for the stub file.""" imports = [] @@ -628,7 +628,7 @@ class AsyncToSyncConverter: return imports @classmethod - def _get_class_attributes(cls, async_class: Type) -> list[tuple[str, Type]]: + def _get_class_attributes(cls, async_class: type) -> list[tuple[str, type]]: """Extract class attributes that are classes themselves.""" class_attributes = [] @@ -654,7 +654,7 @@ class AsyncToSyncConverter: def _generate_inner_class_stub( cls, name: str, - attr: Type, + attr: type, indent: str = " ", type_tracker: Optional[TypeTracker] = None, ) -> list[str]: @@ -782,7 +782,7 @@ class AsyncToSyncConverter: return processed @classmethod - def generate_stub_file(cls, async_class: Type, sync_class: Type) -> None: + def generate_stub_file(cls, async_class: type, sync_class: type) -> None: """ Generate a .pyi stub file for the sync class to help IDEs with type checking. """ @@ -988,7 +988,7 @@ class AsyncToSyncConverter: logging.error(traceback.format_exc()) -def create_sync_class(async_class: Type, thread_pool_size=10) -> Type: +def create_sync_class(async_class: type, thread_pool_size=10) -> type: """ Creates a sync version of an async class diff --git a/comfy_api/internal/singleton.py b/comfy_api/internal/singleton.py index 75f16f98e..d89380262 100644 --- a/comfy_api/internal/singleton.py +++ b/comfy_api/internal/singleton.py @@ -1,4 +1,4 @@ -from typing import Type, TypeVar +from typing import TypeVar class SingletonMetaclass(type): T = TypeVar("T", bound="SingletonMetaclass") @@ -11,13 +11,13 @@ class SingletonMetaclass(type): ) return cls._instances[cls] - def inject_instance(cls: Type[T], instance: T) -> None: + def inject_instance(cls: type[T], instance: T) -> None: assert cls not in SingletonMetaclass._instances, ( "Cannot inject instance after first instantiation" ) SingletonMetaclass._instances[cls] = instance - def get_instance(cls: Type[T], *args, **kwargs) -> T: + def get_instance(cls: type[T], *args, **kwargs) -> T: """ Gets the singleton instance of the class, creating it if it doesn't exist. """ diff --git a/comfy_api/latest/__init__.py b/comfy_api/latest/__init__.py index 35e1ac853..fab63c7df 100644 --- a/comfy_api/latest/__init__.py +++ b/comfy_api/latest/__init__.py @@ -1,7 +1,7 @@ from __future__ import annotations from abc import ABC, abstractmethod -from typing import Type, TYPE_CHECKING +from typing import TYPE_CHECKING from comfy_api.internal import ComfyAPIBase from comfy_api.internal.singleton import ProxiedSingleton from comfy_api.internal.async_to_sync import create_sync_class @@ -113,7 +113,7 @@ ComfyAPI = ComfyAPI_latest if TYPE_CHECKING: import comfy_api.latest.generated.ComfyAPISyncStub # type: ignore - ComfyAPISync: Type[comfy_api.latest.generated.ComfyAPISyncStub.ComfyAPISyncStub] + ComfyAPISync: type[comfy_api.latest.generated.ComfyAPISyncStub.ComfyAPISyncStub] ComfyAPISync = create_sync_class(ComfyAPI_latest) # create new aliases for io and ui diff --git a/comfy_api/latest/_input/basic_types.py b/comfy_api/latest/_input/basic_types.py index 245c6cbb1..d73deabd2 100644 --- a/comfy_api/latest/_input/basic_types.py +++ b/comfy_api/latest/_input/basic_types.py @@ -1,5 +1,5 @@ import torch -from typing import TypedDict, List, Optional +from typing import TypedDict, Optional ImageInput = torch.Tensor """ @@ -39,4 +39,4 @@ class LatentInput(TypedDict): Optional noise mask tensor in the same format as samples. """ - batch_index: Optional[List[int]] + batch_index: Optional[list[int]] diff --git a/comfy_api/latest/_ui.py b/comfy_api/latest/_ui.py index 2babe209a..e238cdf3c 100644 --- a/comfy_api/latest/_ui.py +++ b/comfy_api/latest/_ui.py @@ -5,7 +5,6 @@ import os import random import uuid from io import BytesIO -from typing import Type import av import numpy as np @@ -83,7 +82,7 @@ class ImageSaveHelper: return PILImage.fromarray(np.clip(255.0 * image_tensor.cpu().numpy(), 0, 255).astype(np.uint8)) @staticmethod - def _create_png_metadata(cls: Type[ComfyNode] | None) -> PngInfo | None: + def _create_png_metadata(cls: type[ComfyNode] | None) -> PngInfo | None: """Creates a PngInfo object with prompt and extra_pnginfo.""" if args.disable_metadata or cls is None or not cls.hidden: return None @@ -96,7 +95,7 @@ class ImageSaveHelper: return metadata @staticmethod - def _create_animated_png_metadata(cls: Type[ComfyNode] | None) -> PngInfo | None: + def _create_animated_png_metadata(cls: type[ComfyNode] | None) -> PngInfo | None: """Creates a PngInfo object with prompt and extra_pnginfo for animated PNGs (APNG).""" if args.disable_metadata or cls is None or not cls.hidden: return None @@ -121,7 +120,7 @@ class ImageSaveHelper: return metadata @staticmethod - def _create_webp_metadata(pil_image: PILImage.Image, cls: Type[ComfyNode] | None) -> PILImage.Exif: + def _create_webp_metadata(pil_image: PILImage.Image, cls: type[ComfyNode] | None) -> PILImage.Exif: """Creates EXIF metadata bytes for WebP images.""" exif_data = pil_image.getexif() if args.disable_metadata or cls is None or cls.hidden is None: @@ -137,7 +136,7 @@ class ImageSaveHelper: @staticmethod def save_images( - images, filename_prefix: str, folder_type: FolderType, cls: Type[ComfyNode] | None, compress_level = 4, + images, filename_prefix: str, folder_type: FolderType, cls: type[ComfyNode] | None, compress_level = 4, ) -> list[SavedResult]: """Saves a batch of images as individual PNG files.""" full_output_folder, filename, counter, subfolder, _ = folder_paths.get_save_image_path( @@ -155,7 +154,7 @@ class ImageSaveHelper: return results @staticmethod - def get_save_images_ui(images, filename_prefix: str, cls: Type[ComfyNode] | None, compress_level=4) -> SavedImages: + def get_save_images_ui(images, filename_prefix: str, cls: type[ComfyNode] | None, compress_level=4) -> SavedImages: """Saves a batch of images and returns a UI object for the node output.""" return SavedImages( ImageSaveHelper.save_images( @@ -169,7 +168,7 @@ class ImageSaveHelper: @staticmethod def save_animated_png( - images, filename_prefix: str, folder_type: FolderType, cls: Type[ComfyNode] | None, fps: float, compress_level: int + images, filename_prefix: str, folder_type: FolderType, cls: type[ComfyNode] | None, fps: float, compress_level: int ) -> SavedResult: """Saves a batch of images as a single animated PNG.""" full_output_folder, filename, counter, subfolder, _ = folder_paths.get_save_image_path( @@ -191,7 +190,7 @@ class ImageSaveHelper: @staticmethod def get_save_animated_png_ui( - images, filename_prefix: str, cls: Type[ComfyNode] | None, fps: float, compress_level: int + images, filename_prefix: str, cls: type[ComfyNode] | None, fps: float, compress_level: int ) -> SavedImages: """Saves an animated PNG and returns a UI object for the node output.""" result = ImageSaveHelper.save_animated_png( @@ -209,7 +208,7 @@ class ImageSaveHelper: images, filename_prefix: str, folder_type: FolderType, - cls: Type[ComfyNode] | None, + cls: type[ComfyNode] | None, fps: float, lossless: bool, quality: int, @@ -238,7 +237,7 @@ class ImageSaveHelper: def get_save_animated_webp_ui( images, filename_prefix: str, - cls: Type[ComfyNode] | None, + cls: type[ComfyNode] | None, fps: float, lossless: bool, quality: int, @@ -267,7 +266,7 @@ class AudioSaveHelper: audio: dict, filename_prefix: str, folder_type: FolderType, - cls: Type[ComfyNode] | None, + cls: type[ComfyNode] | None, format: str = "flac", quality: str = "128k", ) -> list[SavedResult]: @@ -372,7 +371,7 @@ class AudioSaveHelper: @staticmethod def get_save_audio_ui( - audio, filename_prefix: str, cls: Type[ComfyNode] | None, format: str = "flac", quality: str = "128k", + audio, filename_prefix: str, cls: type[ComfyNode] | None, format: str = "flac", quality: str = "128k", ) -> SavedAudios: """Save and instantly wrap for UI.""" return SavedAudios( @@ -388,7 +387,7 @@ class AudioSaveHelper: class PreviewImage(_UIOutput): - def __init__(self, image: Image.Type, animated: bool = False, cls: Type[ComfyNode] = None, **kwargs): + def __init__(self, image: Image.Type, animated: bool = False, cls: type[ComfyNode] = None, **kwargs): self.values = ImageSaveHelper.save_images( image, filename_prefix="ComfyUI_temp_" + ''.join(random.choice("abcdefghijklmnopqrstupvxyz") for _ in range(5)), @@ -412,7 +411,7 @@ class PreviewMask(PreviewImage): class PreviewAudio(_UIOutput): - def __init__(self, audio: dict, cls: Type[ComfyNode] = None, **kwargs): + def __init__(self, audio: dict, cls: type[ComfyNode] = None, **kwargs): self.values = AudioSaveHelper.save_audio( audio, filename_prefix="ComfyUI_temp_" + "".join(random.choice("abcdefghijklmnopqrstuvwxyz") for _ in range(5)), diff --git a/comfy_api/version_list.py b/comfy_api/version_list.py index 7cb1871d5..be6e1db66 100644 --- a/comfy_api/version_list.py +++ b/comfy_api/version_list.py @@ -2,9 +2,8 @@ from comfy_api.latest import ComfyAPI_latest from comfy_api.v0_0_2 import ComfyAPIAdapter_v0_0_2 from comfy_api.v0_0_1 import ComfyAPIAdapter_v0_0_1 from comfy_api.internal import ComfyAPIBase -from typing import List, Type -supported_versions: List[Type[ComfyAPIBase]] = [ +supported_versions: list[type[ComfyAPIBase]] = [ ComfyAPI_latest, ComfyAPIAdapter_v0_0_2, ComfyAPIAdapter_v0_0_1, From 77b2f7c228a0db6643bb7f29be4db0bff6799db2 Mon Sep 17 00:00:00 2001 From: drozbay <17261091+drozbay@users.noreply.github.com> Date: Mon, 15 Dec 2025 17:06:32 -0700 Subject: [PATCH 289/647] Add context windows callback for custom cond handling (#11208) Co-authored-by: ozbayb <17261091+ozbayb@users.noreply.github.com> --- comfy/context_windows.py | 13 +++++++++++++ 1 file changed, 13 insertions(+) diff --git a/comfy/context_windows.py b/comfy/context_windows.py index 5c412d1c2..2979b3ca1 100644 --- a/comfy/context_windows.py +++ b/comfy/context_windows.py @@ -87,6 +87,7 @@ class IndexListCallbacks: COMBINE_CONTEXT_WINDOW_RESULTS = "combine_context_window_results" EXECUTE_START = "execute_start" EXECUTE_CLEANUP = "execute_cleanup" + RESIZE_COND_ITEM = "resize_cond_item" def init_callbacks(self): return {} @@ -166,6 +167,18 @@ class IndexListContextHandler(ContextHandlerABC): new_cond_item = cond_item.copy() # when in dictionary, look for tensors and CONDCrossAttn [comfy/conds.py] (has cond attr that is a tensor) for cond_key, cond_value in new_cond_item.items(): + # Allow callbacks to handle custom conditioning items + handled = False + for callback in comfy.patcher_extension.get_all_callbacks( + IndexListCallbacks.RESIZE_COND_ITEM, self.callbacks + ): + result = callback(cond_key, cond_value, window, x_in, device, new_cond_item) + if result is not None: + new_cond_item[cond_key] = result + handled = True + break + if handled: + continue if isinstance(cond_value, torch.Tensor): if (self.dim < cond_value.ndim and cond_value(self.dim) == x_in.size(self.dim)) or \ (cond_value.ndim < self.dim and cond_value.size(0) == x_in.size(self.dim)): From 70541d4e7769c6c40eae6594e677355eacd181fe Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 15 Dec 2025 16:20:34 -0800 Subject: [PATCH 290/647] Support the new qwen edit 2511 reference method. (#11340) index_timestep_zero can be selected in the FluxKontextMultiReferenceLatentMethod now with the display name set to the more generic "Edit Model Reference Method" node. --- comfy/ldm/qwen_image/model.py | 47 +++++++++++++++++++++++++++++------ comfy_extras/nodes_flux.py | 3 ++- 2 files changed, 41 insertions(+), 9 deletions(-) diff --git a/comfy/ldm/qwen_image/model.py b/comfy/ldm/qwen_image/model.py index 8c75670cd..96590088b 100644 --- a/comfy/ldm/qwen_image/model.py +++ b/comfy/ldm/qwen_image/model.py @@ -218,9 +218,24 @@ class QwenImageTransformerBlock(nn.Module): operations=operations, ) - def _modulate(self, x: torch.Tensor, mod_params: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]: + def _apply_gate(self, x, y, gate, timestep_zero_index=None): + if timestep_zero_index is not None: + return y + torch.cat((x[:, :timestep_zero_index] * gate[0], x[:, timestep_zero_index:] * gate[1]), dim=1) + else: + return torch.addcmul(y, gate, x) + + def _modulate(self, x: torch.Tensor, mod_params: torch.Tensor, timestep_zero_index=None) -> Tuple[torch.Tensor, torch.Tensor]: shift, scale, gate = torch.chunk(mod_params, 3, dim=-1) - return torch.addcmul(shift.unsqueeze(1), x, 1 + scale.unsqueeze(1)), gate.unsqueeze(1) + if timestep_zero_index is not None: + actual_batch = shift.size(0) // 2 + shift, shift_0 = shift[:actual_batch], shift[actual_batch:] + scale, scale_0 = scale[:actual_batch], scale[actual_batch:] + gate, gate_0 = gate[:actual_batch], gate[actual_batch:] + reg = torch.addcmul(shift.unsqueeze(1), x[:, :timestep_zero_index], 1 + scale.unsqueeze(1)) + zero = torch.addcmul(shift_0.unsqueeze(1), x[:, timestep_zero_index:], 1 + scale_0.unsqueeze(1)) + return torch.cat((reg, zero), dim=1), (gate.unsqueeze(1), gate_0.unsqueeze(1)) + else: + return torch.addcmul(shift.unsqueeze(1), x, 1 + scale.unsqueeze(1)), gate.unsqueeze(1) def forward( self, @@ -229,14 +244,19 @@ class QwenImageTransformerBlock(nn.Module): encoder_hidden_states_mask: torch.Tensor, temb: torch.Tensor, image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, + timestep_zero_index=None, transformer_options={}, ) -> Tuple[torch.Tensor, torch.Tensor]: img_mod_params = self.img_mod(temb) + + if timestep_zero_index is not None: + temb = temb.chunk(2, dim=0)[0] + txt_mod_params = self.txt_mod(temb) img_mod1, img_mod2 = img_mod_params.chunk(2, dim=-1) txt_mod1, txt_mod2 = txt_mod_params.chunk(2, dim=-1) - img_modulated, img_gate1 = self._modulate(self.img_norm1(hidden_states), img_mod1) + img_modulated, img_gate1 = self._modulate(self.img_norm1(hidden_states), img_mod1, timestep_zero_index) del img_mod1 txt_modulated, txt_gate1 = self._modulate(self.txt_norm1(encoder_hidden_states), txt_mod1) del txt_mod1 @@ -251,15 +271,15 @@ class QwenImageTransformerBlock(nn.Module): del img_modulated del txt_modulated - hidden_states = hidden_states + img_gate1 * img_attn_output + hidden_states = self._apply_gate(img_attn_output, hidden_states, img_gate1, timestep_zero_index) encoder_hidden_states = encoder_hidden_states + txt_gate1 * txt_attn_output del img_attn_output del txt_attn_output del img_gate1 del txt_gate1 - img_modulated2, img_gate2 = self._modulate(self.img_norm2(hidden_states), img_mod2) - hidden_states = torch.addcmul(hidden_states, img_gate2, self.img_mlp(img_modulated2)) + img_modulated2, img_gate2 = self._modulate(self.img_norm2(hidden_states), img_mod2, timestep_zero_index) + hidden_states = self._apply_gate(self.img_mlp(img_modulated2), hidden_states, img_gate2, timestep_zero_index) txt_modulated2, txt_gate2 = self._modulate(self.txt_norm2(encoder_hidden_states), txt_mod2) encoder_hidden_states = torch.addcmul(encoder_hidden_states, txt_gate2, self.txt_mlp(txt_modulated2)) @@ -391,11 +411,14 @@ class QwenImageTransformer2DModel(nn.Module): hidden_states, img_ids, orig_shape = self.process_img(x) num_embeds = hidden_states.shape[1] + timestep_zero_index = None if ref_latents is not None: h = 0 w = 0 index = 0 - index_ref_method = kwargs.get("ref_latents_method", "index") == "index" + ref_method = kwargs.get("ref_latents_method", "index") + index_ref_method = (ref_method == "index") or (ref_method == "index_timestep_zero") + timestep_zero = ref_method == "index_timestep_zero" for ref in ref_latents: if index_ref_method: index += 1 @@ -415,6 +438,10 @@ class QwenImageTransformer2DModel(nn.Module): kontext, kontext_ids, _ = self.process_img(ref, index=index, h_offset=h_offset, w_offset=w_offset) hidden_states = torch.cat([hidden_states, kontext], dim=1) img_ids = torch.cat([img_ids, kontext_ids], dim=1) + if timestep_zero: + if index > 0: + timestep = torch.cat([timestep, timestep * 0], dim=0) + timestep_zero_index = num_embeds txt_start = round(max(((x.shape[-1] + (self.patch_size // 2)) // self.patch_size) // 2, ((x.shape[-2] + (self.patch_size // 2)) // self.patch_size) // 2)) txt_ids = torch.arange(txt_start, txt_start + context.shape[1], device=x.device).reshape(1, -1, 1).repeat(x.shape[0], 1, 3) @@ -446,7 +473,7 @@ class QwenImageTransformer2DModel(nn.Module): if ("double_block", i) in blocks_replace: def block_wrap(args): out = {} - out["txt"], out["img"] = block(hidden_states=args["img"], encoder_hidden_states=args["txt"], encoder_hidden_states_mask=encoder_hidden_states_mask, temb=args["vec"], image_rotary_emb=args["pe"], transformer_options=args["transformer_options"]) + out["txt"], out["img"] = block(hidden_states=args["img"], encoder_hidden_states=args["txt"], encoder_hidden_states_mask=encoder_hidden_states_mask, temb=args["vec"], image_rotary_emb=args["pe"], timestep_zero_index=timestep_zero_index, transformer_options=args["transformer_options"]) return out out = blocks_replace[("double_block", i)]({"img": hidden_states, "txt": encoder_hidden_states, "vec": temb, "pe": image_rotary_emb, "transformer_options": transformer_options}, {"original_block": block_wrap}) hidden_states = out["img"] @@ -458,6 +485,7 @@ class QwenImageTransformer2DModel(nn.Module): encoder_hidden_states_mask=encoder_hidden_states_mask, temb=temb, image_rotary_emb=image_rotary_emb, + timestep_zero_index=timestep_zero_index, transformer_options=transformer_options, ) @@ -474,6 +502,9 @@ class QwenImageTransformer2DModel(nn.Module): if add is not None: hidden_states[:, :add.shape[1]] += add + if timestep_zero_index is not None: + temb = temb.chunk(2, dim=0)[0] + hidden_states = self.norm_out(hidden_states, temb) hidden_states = self.proj_out(hidden_states) diff --git a/comfy_extras/nodes_flux.py b/comfy_extras/nodes_flux.py index d9c4bba81..12c8ed3e6 100644 --- a/comfy_extras/nodes_flux.py +++ b/comfy_extras/nodes_flux.py @@ -154,12 +154,13 @@ class FluxKontextMultiReferenceLatentMethod(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="FluxKontextMultiReferenceLatentMethod", + display_name="Edit Model Reference Method", category="advanced/conditioning/flux", inputs=[ io.Conditioning.Input("conditioning"), io.Combo.Input( "reference_latents_method", - options=["offset", "index", "uxo/uno"], + options=["offset", "index", "uxo/uno", "index_timestep_zero"], ), ], outputs=[ From d02d0e5744f2e06fc40834d3c5bb387de4532007 Mon Sep 17 00:00:00 2001 From: seed93 Date: Tue, 16 Dec 2025 09:38:46 +0800 Subject: [PATCH 291/647] [add] tripo3.0 (#10663) * [add] tripo3.0 * [tripo] change paramter order * change order --------- Co-authored-by: liangd --- comfy_api_nodes/apis/tripo_api.py | 46 ++++++++++++++--- comfy_api_nodes/nodes_tripo.py | 86 ++++++++++++++++++++++++++++++- 2 files changed, 122 insertions(+), 10 deletions(-) diff --git a/comfy_api_nodes/apis/tripo_api.py b/comfy_api_nodes/apis/tripo_api.py index 713260e2a..ffaaa7dc1 100644 --- a/comfy_api_nodes/apis/tripo_api.py +++ b/comfy_api_nodes/apis/tripo_api.py @@ -5,11 +5,17 @@ from typing import Optional, List, Dict, Any, Union from pydantic import BaseModel, Field, RootModel class TripoModelVersion(str, Enum): + v3_0_20250812 = 'v3.0-20250812' v2_5_20250123 = 'v2.5-20250123' v2_0_20240919 = 'v2.0-20240919' v1_4_20240625 = 'v1.4-20240625' +class TripoGeometryQuality(str, Enum): + standard = 'standard' + detailed = 'detailed' + + class TripoTextureQuality(str, Enum): standard = 'standard' detailed = 'detailed' @@ -61,14 +67,20 @@ class TripoSpec(str, Enum): class TripoAnimation(str, Enum): IDLE = "preset:idle" WALK = "preset:walk" + RUN = "preset:run" + DIVE = "preset:dive" CLIMB = "preset:climb" JUMP = "preset:jump" - RUN = "preset:run" SLASH = "preset:slash" SHOOT = "preset:shoot" HURT = "preset:hurt" FALL = "preset:fall" TURN = "preset:turn" + QUADRUPED_WALK = "preset:quadruped:walk" + HEXAPOD_WALK = "preset:hexapod:walk" + OCTOPOD_WALK = "preset:octopod:walk" + SERPENTINE_MARCH = "preset:serpentine:march" + AQUATIC_MARCH = "preset:aquatic:march" class TripoStylizeStyle(str, Enum): LEGO = "lego" @@ -105,6 +117,11 @@ class TripoTaskStatus(str, Enum): BANNED = "banned" EXPIRED = "expired" +class TripoFbxPreset(str, Enum): + BLENDER = "blender" + MIXAMO = "mixamo" + _3DSMAX = "3dsmax" + class TripoFileTokenReference(BaseModel): type: Optional[str] = Field(None, description='The type of the reference') file_token: str @@ -142,6 +159,7 @@ class TripoTextToModelRequest(BaseModel): model_seed: Optional[int] = Field(None, description='The seed for the model') texture_seed: Optional[int] = Field(None, description='The seed for the texture') texture_quality: Optional[TripoTextureQuality] = TripoTextureQuality.standard + geometry_quality: Optional[TripoGeometryQuality] = TripoGeometryQuality.standard style: Optional[TripoStyle] = None auto_size: Optional[bool] = Field(False, description='Whether to auto-size the model') quad: Optional[bool] = Field(False, description='Whether to apply quad to the generated model') @@ -156,6 +174,7 @@ class TripoImageToModelRequest(BaseModel): model_seed: Optional[int] = Field(None, description='The seed for the model') texture_seed: Optional[int] = Field(None, description='The seed for the texture') texture_quality: Optional[TripoTextureQuality] = TripoTextureQuality.standard + geometry_quality: Optional[TripoGeometryQuality] = TripoGeometryQuality.standard texture_alignment: Optional[TripoTextureAlignment] = Field(TripoTextureAlignment.ORIGINAL_IMAGE, description='The texture alignment method') style: Optional[TripoStyle] = Field(None, description='The style to apply to the generated model') auto_size: Optional[bool] = Field(False, description='Whether to auto-size the model') @@ -173,6 +192,7 @@ class TripoMultiviewToModelRequest(BaseModel): model_seed: Optional[int] = Field(None, description='The seed for the model') texture_seed: Optional[int] = Field(None, description='The seed for the texture') texture_quality: Optional[TripoTextureQuality] = TripoTextureQuality.standard + geometry_quality: Optional[TripoGeometryQuality] = TripoGeometryQuality.standard texture_alignment: Optional[TripoTextureAlignment] = TripoTextureAlignment.ORIGINAL_IMAGE auto_size: Optional[bool] = Field(False, description='Whether to auto-size the model') orientation: Optional[TripoOrientation] = Field(TripoOrientation.DEFAULT, description='The orientation for the model') @@ -219,14 +239,24 @@ class TripoConvertModelRequest(BaseModel): type: TripoTaskType = Field(TripoTaskType.CONVERT_MODEL, description='Type of task') format: TripoConvertFormat = Field(..., description='The format to convert to') original_model_task_id: str = Field(..., description='The task ID of the original model') - quad: Optional[bool] = Field(False, description='Whether to apply quad to the model') - force_symmetry: Optional[bool] = Field(False, description='Whether to force symmetry') - face_limit: Optional[int] = Field(10000, description='The number of faces to limit the conversion to') - flatten_bottom: Optional[bool] = Field(False, description='Whether to flatten the bottom of the model') - flatten_bottom_threshold: Optional[float] = Field(0.01, description='The threshold for flattening the bottom') - texture_size: Optional[int] = Field(4096, description='The size of the texture') + quad: Optional[bool] = Field(None, description='Whether to apply quad to the model') + force_symmetry: Optional[bool] = Field(None, description='Whether to force symmetry') + face_limit: Optional[int] = Field(None, description='The number of faces to limit the conversion to') + flatten_bottom: Optional[bool] = Field(None, description='Whether to flatten the bottom of the model') + flatten_bottom_threshold: Optional[float] = Field(None, description='The threshold for flattening the bottom') + texture_size: Optional[int] = Field(None, description='The size of the texture') texture_format: Optional[TripoTextureFormat] = Field(TripoTextureFormat.JPEG, description='The format of the texture') - pivot_to_center_bottom: Optional[bool] = Field(False, description='Whether to pivot to the center bottom') + pivot_to_center_bottom: Optional[bool] = Field(None, description='Whether to pivot to the center bottom') + scale_factor: Optional[float] = Field(None, description='The scale factor for the model') + with_animation: Optional[bool] = Field(None, description='Whether to include animations') + pack_uv: Optional[bool] = Field(None, description='Whether to pack the UVs') + bake: Optional[bool] = Field(None, description='Whether to bake the model') + part_names: Optional[List[str]] = Field(None, description='The names of the parts to include') + fbx_preset: Optional[TripoFbxPreset] = Field(None, description='The preset for the FBX export') + export_vertex_colors: Optional[bool] = Field(None, description='Whether to export the vertex colors') + export_orientation: Optional[TripoOrientation] = Field(None, description='The orientation for the export') + animate_in_place: Optional[bool] = Field(None, description='Whether to animate in place') + class TripoTaskRequest(RootModel): root: Union[ diff --git a/comfy_api_nodes/nodes_tripo.py b/comfy_api_nodes/nodes_tripo.py index 697100ff2..bd3c24fb3 100644 --- a/comfy_api_nodes/nodes_tripo.py +++ b/comfy_api_nodes/nodes_tripo.py @@ -102,8 +102,9 @@ class TripoTextToModelNode(IO.ComfyNode): IO.Int.Input("model_seed", default=42, optional=True), IO.Int.Input("texture_seed", default=42, optional=True), IO.Combo.Input("texture_quality", default="standard", options=["standard", "detailed"], optional=True), - IO.Int.Input("face_limit", default=-1, min=-1, max=500000, optional=True), + IO.Int.Input("face_limit", default=-1, min=-1, max=2000000, optional=True), IO.Boolean.Input("quad", default=False, optional=True), + IO.Combo.Input("geometry_quality", default="standard", options=["standard", "detailed"], optional=True), ], outputs=[ IO.String.Output(display_name="model_file"), @@ -131,6 +132,7 @@ class TripoTextToModelNode(IO.ComfyNode): model_seed: Optional[int] = None, texture_seed: Optional[int] = None, texture_quality: Optional[str] = None, + geometry_quality: Optional[str] = None, face_limit: Optional[int] = None, quad: Optional[bool] = None, ) -> IO.NodeOutput: @@ -154,6 +156,7 @@ class TripoTextToModelNode(IO.ComfyNode): texture_seed=texture_seed, texture_quality=texture_quality, face_limit=face_limit, + geometry_quality=geometry_quality, auto_size=True, quad=quad, ), @@ -194,6 +197,7 @@ class TripoImageToModelNode(IO.ComfyNode): ), IO.Int.Input("face_limit", default=-1, min=-1, max=500000, optional=True), IO.Boolean.Input("quad", default=False, optional=True), + IO.Combo.Input("geometry_quality", default="standard", options=["standard", "detailed"], optional=True), ], outputs=[ IO.String.Output(display_name="model_file"), @@ -220,6 +224,7 @@ class TripoImageToModelNode(IO.ComfyNode): orientation=None, texture_seed: Optional[int] = None, texture_quality: Optional[str] = None, + geometry_quality: Optional[str] = None, texture_alignment: Optional[str] = None, face_limit: Optional[int] = None, quad: Optional[bool] = None, @@ -246,6 +251,7 @@ class TripoImageToModelNode(IO.ComfyNode): pbr=pbr, model_seed=model_seed, orientation=orientation, + geometry_quality=geometry_quality, texture_alignment=texture_alignment, texture_seed=texture_seed, texture_quality=texture_quality, @@ -295,6 +301,7 @@ class TripoMultiviewToModelNode(IO.ComfyNode): ), IO.Int.Input("face_limit", default=-1, min=-1, max=500000, optional=True), IO.Boolean.Input("quad", default=False, optional=True), + IO.Combo.Input("geometry_quality", default="standard", options=["standard", "detailed"], optional=True), ], outputs=[ IO.String.Output(display_name="model_file"), @@ -323,6 +330,7 @@ class TripoMultiviewToModelNode(IO.ComfyNode): model_seed: Optional[int] = None, texture_seed: Optional[int] = None, texture_quality: Optional[str] = None, + geometry_quality: Optional[str] = None, texture_alignment: Optional[str] = None, face_limit: Optional[int] = None, quad: Optional[bool] = None, @@ -359,6 +367,7 @@ class TripoMultiviewToModelNode(IO.ComfyNode): model_seed=model_seed, texture_seed=texture_seed, texture_quality=texture_quality, + geometry_quality=geometry_quality, texture_alignment=texture_alignment, face_limit=face_limit, quad=quad, @@ -508,6 +517,8 @@ class TripoRetargetNode(IO.ComfyNode): options=[ "preset:idle", "preset:walk", + "preset:run", + "preset:dive", "preset:climb", "preset:jump", "preset:slash", @@ -515,6 +526,11 @@ class TripoRetargetNode(IO.ComfyNode): "preset:hurt", "preset:fall", "preset:turn", + "preset:quadruped:walk", + "preset:hexapod:walk", + "preset:octopod:walk", + "preset:serpentine:march", + "preset:aquatic:march" ], ), ], @@ -563,7 +579,7 @@ class TripoConversionNode(IO.ComfyNode): "face_limit", default=-1, min=-1, - max=500000, + max=2000000, optional=True, ), IO.Int.Input( @@ -579,6 +595,40 @@ class TripoConversionNode(IO.ComfyNode): default="JPEG", optional=True, ), + IO.Boolean.Input("force_symmetry", default=False, optional=True), + IO.Boolean.Input("flatten_bottom", default=False, optional=True), + IO.Float.Input( + "flatten_bottom_threshold", + default=0.0, + min=0.0, + max=1.0, + optional=True, + ), + IO.Boolean.Input("pivot_to_center_bottom", default=False, optional=True), + IO.Float.Input( + "scale_factor", + default=1.0, + min=0.0, + optional=True, + ), + IO.Boolean.Input("with_animation", default=False, optional=True), + IO.Boolean.Input("pack_uv", default=False, optional=True), + IO.Boolean.Input("bake", default=False, optional=True), + IO.String.Input("part_names", default="", optional=True), # comma-separated list + IO.Combo.Input( + "fbx_preset", + options=["blender", "mixamo", "3dsmax"], + default="blender", + optional=True, + ), + IO.Boolean.Input("export_vertex_colors", default=False, optional=True), + IO.Combo.Input( + "export_orientation", + options=["align_image", "default"], + default="default", + optional=True, + ), + IO.Boolean.Input("animate_in_place", default=False, optional=True), ], outputs=[], hidden=[ @@ -604,12 +654,31 @@ class TripoConversionNode(IO.ComfyNode): original_model_task_id, format: str, quad: bool, + force_symmetry: bool, face_limit: int, + flatten_bottom: bool, + flatten_bottom_threshold: float, texture_size: int, texture_format: str, + pivot_to_center_bottom: bool, + scale_factor: float, + with_animation: bool, + pack_uv: bool, + bake: bool, + part_names: str, + fbx_preset: str, + export_vertex_colors: bool, + export_orientation: str, + animate_in_place: bool, ) -> IO.NodeOutput: if not original_model_task_id: raise RuntimeError("original_model_task_id is required") + + # Parse part_names from comma-separated string to list + part_names_list = None + if part_names and part_names.strip(): + part_names_list = [name.strip() for name in part_names.split(',') if name.strip()] + response = await sync_op( cls, endpoint=ApiEndpoint(path="/proxy/tripo/v2/openapi/task", method="POST"), @@ -618,9 +687,22 @@ class TripoConversionNode(IO.ComfyNode): original_model_task_id=original_model_task_id, format=format, quad=quad if quad else None, + force_symmetry=force_symmetry if force_symmetry else None, face_limit=face_limit if face_limit != -1 else None, + flatten_bottom=flatten_bottom if flatten_bottom else None, + flatten_bottom_threshold=flatten_bottom_threshold if flatten_bottom_threshold != 0.0 else None, texture_size=texture_size if texture_size != 4096 else None, texture_format=texture_format if texture_format != "JPEG" else None, + pivot_to_center_bottom=pivot_to_center_bottom if pivot_to_center_bottom else None, + scale_factor=scale_factor if scale_factor != 1.0 else None, + with_animation=with_animation if with_animation else None, + pack_uv=pack_uv if pack_uv else None, + bake=bake if bake else None, + part_names=part_names_list, + fbx_preset=fbx_preset if fbx_preset != "blender" else None, + export_vertex_colors=export_vertex_colors if export_vertex_colors else None, + export_orientation=export_orientation if export_orientation != "default" else None, + animate_in_place=animate_in_place if animate_in_place else None, ), ) return await poll_until_finished(cls, response, average_duration=30) From 41bcf0619db87d443d468c9ddad4454bdbc1b084 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 15 Dec 2025 17:51:06 -0800 Subject: [PATCH 292/647] Add code to detect if a z image fun controlnet is broken or not. (#11341) --- comfy_extras/nodes_model_patch.py | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/comfy_extras/nodes_model_patch.py b/comfy_extras/nodes_model_patch.py index ec0e790dc..fdd5d0d3f 100644 --- a/comfy_extras/nodes_model_patch.py +++ b/comfy_extras/nodes_model_patch.py @@ -248,7 +248,10 @@ class ModelPatchLoader: config['n_control_layers'] = 15 config['additional_in_dim'] = 17 config['refiner_control'] = True - config['broken'] = True + ref_weight = sd.get("control_noise_refiner.0.after_proj.weight", None) + if ref_weight is not None: + if torch.count_nonzero(ref_weight) == 0: + config['broken'] = True model = comfy.ldm.lumina.controlnet.ZImage_Control(device=comfy.model_management.unet_offload_device(), dtype=dtype, operations=comfy.ops.manual_cast, **config) model.load_state_dict(sd) From fc4af8606880be0374cf1f1f45bc5730e6d22bf5 Mon Sep 17 00:00:00 2001 From: Haoming <73768377+Haoming02@users.noreply.github.com> Date: Tue, 16 Dec 2025 09:57:28 +0800 Subject: [PATCH 293/647] [BlockInfo] Lumina (#11227) * block info * device * Make tensor int again --------- Co-authored-by: Jedrzej Kosinski --- comfy/ldm/lumina/model.py | 3 +++ 1 file changed, 3 insertions(+) diff --git a/comfy/ldm/lumina/model.py b/comfy/ldm/lumina/model.py index 96cb37fa6..5628e2ba3 100644 --- a/comfy/ldm/lumina/model.py +++ b/comfy/ldm/lumina/model.py @@ -634,8 +634,11 @@ class NextDiT(nn.Module): img, mask, img_size, cap_size, freqs_cis = self.patchify_and_embed(x, cap_feats, cap_mask, adaln_input, num_tokens, transformer_options=transformer_options) freqs_cis = freqs_cis.to(img.device) + transformer_options["total_blocks"] = len(self.layers) + transformer_options["block_type"] = "double" img_input = img for i, layer in enumerate(self.layers): + transformer_options["block_index"] = i img = layer(img, mask, freqs_cis, adaln_input, transformer_options=transformer_options) if "double_block" in patches: for p in patches["double_block"]: From ea2c117bc3c9d3b38d68e651905ed0d6dd682f92 Mon Sep 17 00:00:00 2001 From: Haoming <73768377+Haoming02@users.noreply.github.com> Date: Tue, 16 Dec 2025 09:59:16 +0800 Subject: [PATCH 294/647] [BlockInfo] Wan (#10845) * block info * animate * tensor * device * revert --- comfy/ldm/wan/model.py | 21 ++++++++++++++++++--- comfy/ldm/wan/model_animate.py | 3 +++ 2 files changed, 21 insertions(+), 3 deletions(-) diff --git a/comfy/ldm/wan/model.py b/comfy/ldm/wan/model.py index a9d5e10d9..4216ce831 100644 --- a/comfy/ldm/wan/model.py +++ b/comfy/ldm/wan/model.py @@ -568,7 +568,10 @@ class WanModel(torch.nn.Module): patches_replace = transformer_options.get("patches_replace", {}) blocks_replace = patches_replace.get("dit", {}) + transformer_options["total_blocks"] = len(self.blocks) + transformer_options["block_type"] = "double" for i, block in enumerate(self.blocks): + transformer_options["block_index"] = i if ("double_block", i) in blocks_replace: def block_wrap(args): out = {} @@ -763,7 +766,10 @@ class VaceWanModel(WanModel): patches_replace = transformer_options.get("patches_replace", {}) blocks_replace = patches_replace.get("dit", {}) + transformer_options["total_blocks"] = len(self.blocks) + transformer_options["block_type"] = "double" for i, block in enumerate(self.blocks): + transformer_options["block_index"] = i if ("double_block", i) in blocks_replace: def block_wrap(args): out = {} @@ -862,7 +868,10 @@ class CameraWanModel(WanModel): patches_replace = transformer_options.get("patches_replace", {}) blocks_replace = patches_replace.get("dit", {}) + transformer_options["total_blocks"] = len(self.blocks) + transformer_options["block_type"] = "double" for i, block in enumerate(self.blocks): + transformer_options["block_index"] = i if ("double_block", i) in blocks_replace: def block_wrap(args): out = {} @@ -1326,16 +1335,19 @@ class WanModel_S2V(WanModel): patches_replace = transformer_options.get("patches_replace", {}) blocks_replace = patches_replace.get("dit", {}) + transformer_options["total_blocks"] = len(self.blocks) + transformer_options["block_type"] = "double" for i, block in enumerate(self.blocks): + transformer_options["block_index"] = i if ("double_block", i) in blocks_replace: def block_wrap(args): out = {} - out["img"] = block(args["img"], context=args["txt"], e=args["vec"], freqs=args["pe"]) + out["img"] = block(args["img"], context=args["txt"], e=args["vec"], freqs=args["pe"], transformer_options=args["transformer_options"]) return out - out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "vec": e0, "pe": freqs}, {"original_block": block_wrap}) + out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "vec": e0, "pe": freqs, "transformer_options": transformer_options}, {"original_block": block_wrap}) x = out["img"] else: - x = block(x, e=e0, freqs=freqs, context=context) + x = block(x, e=e0, freqs=freqs, context=context, transformer_options=transformer_options) if audio_emb is not None: x = self.audio_injector(x, i, audio_emb, audio_emb_global, seq_len) # head @@ -1574,7 +1586,10 @@ class HumoWanModel(WanModel): patches_replace = transformer_options.get("patches_replace", {}) blocks_replace = patches_replace.get("dit", {}) + transformer_options["total_blocks"] = len(self.blocks) + transformer_options["block_type"] = "double" for i, block in enumerate(self.blocks): + transformer_options["block_index"] = i if ("double_block", i) in blocks_replace: def block_wrap(args): out = {} diff --git a/comfy/ldm/wan/model_animate.py b/comfy/ldm/wan/model_animate.py index 7c87835d4..84d7adec4 100644 --- a/comfy/ldm/wan/model_animate.py +++ b/comfy/ldm/wan/model_animate.py @@ -523,7 +523,10 @@ class AnimateWanModel(WanModel): patches_replace = transformer_options.get("patches_replace", {}) blocks_replace = patches_replace.get("dit", {}) + transformer_options["total_blocks"] = len(self.blocks) + transformer_options["block_type"] = "double" for i, block in enumerate(self.blocks): + transformer_options["block_index"] = i if ("double_block", i) in blocks_replace: def block_wrap(args): out = {} From 683569de5527379d9a095af88a9e1349fb7e46b5 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 15 Dec 2025 19:33:27 -0800 Subject: [PATCH 295/647] Only enable fp16 on ZImage on newer pytorch. (#11344) --- comfy/supported_models.py | 9 ++++++++- 1 file changed, 8 insertions(+), 1 deletion(-) diff --git a/comfy/supported_models.py b/comfy/supported_models.py index 834dfcffc..1888f35ba 100644 --- a/comfy/supported_models.py +++ b/comfy/supported_models.py @@ -28,6 +28,7 @@ from . import supported_models_base from . import latent_formats from . import diffusers_convert +import comfy.model_management class SD15(supported_models_base.BASE): unet_config = { @@ -1028,7 +1029,13 @@ class ZImage(Lumina2): memory_usage_factor = 2.0 - supported_inference_dtypes = [torch.bfloat16, torch.float16, torch.float32] + supported_inference_dtypes = [torch.bfloat16, torch.float32] + + def __init__(self, unet_config): + super().__init__(unet_config) + if comfy.model_management.extended_fp16_support(): + self.supported_inference_dtypes = self.supported_inference_dtypes.copy() + self.supported_inference_dtypes.insert(1, torch.float16) def clip_target(self, state_dict={}): pref = self.text_encoder_key_prefix[0] From 3d082c32065e0653490b9a4ae45dd33b6c7bffb7 Mon Sep 17 00:00:00 2001 From: Christian Byrne Date: Mon, 15 Dec 2025 20:35:37 -0800 Subject: [PATCH 296/647] bump comfyui-frontend-package to 1.34.9 (patch) (#11342) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 117260515..9b9e61683 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,4 +1,4 @@ -comfyui-frontend-package==1.34.8 +comfyui-frontend-package==1.34.9 comfyui-workflow-templates==0.7.59 comfyui-embedded-docs==0.3.1 torch From 645ee1881e739b3013eeb26dbb335280bfbf443e Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 15 Dec 2025 20:38:12 -0800 Subject: [PATCH 297/647] Inpainting for z image fun control. Use the ZImageFunControlnet node. (#11346) image -> control image ex: pose inpaint_image -> image for inpainting mask -> inpaint mask --- comfy_extras/nodes_model_patch.py | 77 ++++++++++++++++++++++++------- 1 file changed, 61 insertions(+), 16 deletions(-) diff --git a/comfy_extras/nodes_model_patch.py b/comfy_extras/nodes_model_patch.py index fdd5d0d3f..2a0cfcf18 100644 --- a/comfy_extras/nodes_model_patch.py +++ b/comfy_extras/nodes_model_patch.py @@ -313,22 +313,46 @@ class ZImageControlPatch: self.inpaint_image = inpaint_image self.mask = mask self.strength = strength - self.encoded_image = self.encode_latent_cond(image) - self.encoded_image_size = (image.shape[1], image.shape[2]) + self.is_inpaint = self.model_patch.model.additional_in_dim > 0 + + skip_encoding = False + if self.image is not None and self.inpaint_image is not None: + if self.image.shape != self.inpaint_image.shape: + skip_encoding = True + + if skip_encoding: + self.encoded_image = None + else: + self.encoded_image = self.encode_latent_cond(self.image, self.inpaint_image) + if self.image is None: + self.encoded_image_size = (self.inpaint_image.shape[1], self.inpaint_image.shape[2]) + else: + self.encoded_image_size = (self.image.shape[1], self.image.shape[2]) self.temp_data = None - def encode_latent_cond(self, control_image, inpaint_image=None): - latent_image = comfy.latent_formats.Flux().process_in(self.vae.encode(control_image)) - if self.model_patch.model.additional_in_dim > 0: - if self.mask is None: - mask_ = torch.zeros_like(latent_image)[:, :1] - else: - mask_ = comfy.utils.common_upscale(self.mask.mean(dim=1, keepdim=True), latent_image.shape[-1], latent_image.shape[-2], "bilinear", "none") + def encode_latent_cond(self, control_image=None, inpaint_image=None): + latent_image = None + if control_image is not None: + latent_image = comfy.latent_formats.Flux().process_in(self.vae.encode(control_image)) + + if self.is_inpaint: if inpaint_image is None: inpaint_image = torch.ones_like(control_image) * 0.5 + if self.mask is not None: + mask_inpaint = comfy.utils.common_upscale(self.mask.view(self.mask.shape[0], -1, self.mask.shape[-2], self.mask.shape[-1]).mean(dim=1, keepdim=True), inpaint_image.shape[-2], inpaint_image.shape[-3], "bilinear", "center") + inpaint_image = ((inpaint_image - 0.5) * mask_inpaint.movedim(1, -1).round()) + 0.5 + inpaint_image_latent = comfy.latent_formats.Flux().process_in(self.vae.encode(inpaint_image)) + if self.mask is None: + mask_ = torch.zeros_like(inpaint_image_latent)[:, :1] + else: + mask_ = comfy.utils.common_upscale(self.mask.view(self.mask.shape[0], -1, self.mask.shape[-2], self.mask.shape[-1]).mean(dim=1, keepdim=True), inpaint_image_latent.shape[-1], inpaint_image_latent.shape[-2], "nearest", "center") + + if latent_image is None: + latent_image = comfy.latent_formats.Flux().process_in(self.vae.encode(torch.ones_like(inpaint_image) * 0.5)) + return torch.cat([latent_image, mask_, inpaint_image_latent], dim=1) else: return latent_image @@ -344,13 +368,18 @@ class ZImageControlPatch: block_type = kwargs.get("block_type", "") spacial_compression = self.vae.spacial_compression_encode() if self.encoded_image is None or self.encoded_image_size != (x.shape[-2] * spacial_compression, x.shape[-1] * spacial_compression): - image_scaled = comfy.utils.common_upscale(self.image.movedim(-1, 1), x.shape[-1] * spacial_compression, x.shape[-2] * spacial_compression, "area", "center") + image_scaled = None + if self.image is not None: + image_scaled = comfy.utils.common_upscale(self.image.movedim(-1, 1), x.shape[-1] * spacial_compression, x.shape[-2] * spacial_compression, "area", "center").movedim(1, -1) + self.encoded_image_size = (image_scaled.shape[-3], image_scaled.shape[-2]) + inpaint_scaled = None if self.inpaint_image is not None: inpaint_scaled = comfy.utils.common_upscale(self.inpaint_image.movedim(-1, 1), x.shape[-1] * spacial_compression, x.shape[-2] * spacial_compression, "area", "center").movedim(1, -1) + self.encoded_image_size = (inpaint_scaled.shape[-3], inpaint_scaled.shape[-2]) + loaded_models = comfy.model_management.loaded_models(only_currently_used=True) - self.encoded_image = self.encode_latent_cond(image_scaled.movedim(1, -1), inpaint_scaled) - self.encoded_image_size = (image_scaled.shape[-2], image_scaled.shape[-1]) + self.encoded_image = self.encode_latent_cond(image_scaled, inpaint_scaled) comfy.model_management.load_models_gpu(loaded_models) cnet_blocks = self.model_patch.model.n_control_layers @@ -391,7 +420,8 @@ class ZImageControlPatch: def to(self, device_or_dtype): if isinstance(device_or_dtype, torch.device): - self.encoded_image = self.encoded_image.to(device_or_dtype) + if self.encoded_image is not None: + self.encoded_image = self.encoded_image.to(device_or_dtype) self.temp_data = None return self @@ -414,9 +444,12 @@ class QwenImageDiffsynthControlnet: CATEGORY = "advanced/loaders/qwen" - def diffsynth_controlnet(self, model, model_patch, vae, image, strength, mask=None): + def diffsynth_controlnet(self, model, model_patch, vae, image=None, strength=1.0, inpaint_image=None, mask=None): model_patched = model.clone() - image = image[:, :, :, :3] + if image is not None: + image = image[:, :, :, :3] + if inpaint_image is not None: + inpaint_image = inpaint_image[:, :, :, :3] if mask is not None: if mask.ndim == 3: mask = mask.unsqueeze(1) @@ -425,13 +458,24 @@ class QwenImageDiffsynthControlnet: mask = 1.0 - mask if isinstance(model_patch.model, comfy.ldm.lumina.controlnet.ZImage_Control): - patch = ZImageControlPatch(model_patch, vae, image, strength, mask=mask) + patch = ZImageControlPatch(model_patch, vae, image, strength, inpaint_image=inpaint_image, mask=mask) model_patched.set_model_noise_refiner_patch(patch) model_patched.set_model_double_block_patch(patch) else: model_patched.set_model_double_block_patch(DiffSynthCnetPatch(model_patch, vae, image, strength, mask)) return (model_patched,) +class ZImageFunControlnet(QwenImageDiffsynthControlnet): + @classmethod + def INPUT_TYPES(s): + return {"required": { "model": ("MODEL",), + "model_patch": ("MODEL_PATCH",), + "vae": ("VAE",), + "strength": ("FLOAT", {"default": 1.0, "min": -10.0, "max": 10.0, "step": 0.01}), + }, + "optional": {"image": ("IMAGE",), "inpaint_image": ("IMAGE",), "mask": ("MASK",)}} + + CATEGORY = "advanced/loaders/zimage" class UsoStyleProjectorPatch: def __init__(self, model_patch, encoded_image): @@ -479,5 +523,6 @@ class USOStyleReference: NODE_CLASS_MAPPINGS = { "ModelPatchLoader": ModelPatchLoader, "QwenImageDiffsynthControlnet": QwenImageDiffsynthControlnet, + "ZImageFunControlnet": ZImageFunControlnet, "USOStyleReference": USOStyleReference, } From bc606d7d645f9edfcac7cca3558210d3ee391d94 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 15 Dec 2025 22:26:55 -0800 Subject: [PATCH 298/647] Add a way to set the default ref method in the qwen image code. (#11349) --- comfy/ldm/qwen_image/model.py | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/comfy/ldm/qwen_image/model.py b/comfy/ldm/qwen_image/model.py index 96590088b..8481f7711 100644 --- a/comfy/ldm/qwen_image/model.py +++ b/comfy/ldm/qwen_image/model.py @@ -322,6 +322,7 @@ class QwenImageTransformer2DModel(nn.Module): pooled_projection_dim: int = 768, guidance_embeds: bool = False, axes_dims_rope: Tuple[int, int, int] = (16, 56, 56), + default_ref_method="index", image_model=None, final_layer=True, dtype=None, @@ -334,6 +335,7 @@ class QwenImageTransformer2DModel(nn.Module): self.in_channels = in_channels self.out_channels = out_channels or in_channels self.inner_dim = num_attention_heads * attention_head_dim + self.default_ref_method = default_ref_method self.pe_embedder = EmbedND(dim=attention_head_dim, theta=10000, axes_dim=list(axes_dims_rope)) @@ -416,7 +418,7 @@ class QwenImageTransformer2DModel(nn.Module): h = 0 w = 0 index = 0 - ref_method = kwargs.get("ref_latents_method", "index") + ref_method = kwargs.get("ref_latents_method", self.default_ref_method) index_ref_method = (ref_method == "index") or (ref_method == "index_timestep_zero") timestep_zero = ref_method == "index_timestep_zero" for ref in ref_latents: From 9304e47351be8d178a093b30bcaf5d72c3a2baf5 Mon Sep 17 00:00:00 2001 From: Benjamin Lu Date: Mon, 15 Dec 2025 23:24:18 -0800 Subject: [PATCH 299/647] Update workflows for new release process (#11064) * Update release workflows for branch process * Adjust branch order in workflow triggers * Revert changes in test workflows --- .github/workflows/test-ci.yml | 1 + .github/workflows/test-execution.yml | 4 ++-- .github/workflows/test-launch.yml | 4 ++-- .github/workflows/test-unit.yml | 4 ++-- .github/workflows/update-version.yml | 1 + 5 files changed, 8 insertions(+), 6 deletions(-) diff --git a/.github/workflows/test-ci.yml b/.github/workflows/test-ci.yml index 1660ec8e3..adfc5dd32 100644 --- a/.github/workflows/test-ci.yml +++ b/.github/workflows/test-ci.yml @@ -5,6 +5,7 @@ on: push: branches: - master + - release/** paths-ignore: - 'app/**' - 'input/**' diff --git a/.github/workflows/test-execution.yml b/.github/workflows/test-execution.yml index 00ef07ebf..9012633d8 100644 --- a/.github/workflows/test-execution.yml +++ b/.github/workflows/test-execution.yml @@ -2,9 +2,9 @@ name: Execution Tests on: push: - branches: [ main, master ] + branches: [ main, master, release/** ] pull_request: - branches: [ main, master ] + branches: [ main, master, release/** ] jobs: test: diff --git a/.github/workflows/test-launch.yml b/.github/workflows/test-launch.yml index 1735fd83b..fd70aff23 100644 --- a/.github/workflows/test-launch.yml +++ b/.github/workflows/test-launch.yml @@ -2,9 +2,9 @@ name: Test server launches without errors on: push: - branches: [ main, master ] + branches: [ main, master, release/** ] pull_request: - branches: [ main, master ] + branches: [ main, master, release/** ] jobs: test: diff --git a/.github/workflows/test-unit.yml b/.github/workflows/test-unit.yml index 00caf5b8a..d05179cd3 100644 --- a/.github/workflows/test-unit.yml +++ b/.github/workflows/test-unit.yml @@ -2,9 +2,9 @@ name: Unit Tests on: push: - branches: [ main, master ] + branches: [ main, master, release/** ] pull_request: - branches: [ main, master ] + branches: [ main, master, release/** ] jobs: test: diff --git a/.github/workflows/update-version.yml b/.github/workflows/update-version.yml index d9d488974..c2343cc39 100644 --- a/.github/workflows/update-version.yml +++ b/.github/workflows/update-version.yml @@ -6,6 +6,7 @@ on: - "pyproject.toml" branches: - master + - release/** jobs: update-version: From 65e2103b09f66e45438445fb0e99709ae7639869 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Tue, 16 Dec 2025 23:51:48 +0200 Subject: [PATCH 300/647] feat(api-nodes): add Wan2.6 model to video nodes (#11357) --- comfy_api_nodes/nodes_wan.py | 162 ++++++++++++++++++++--------------- 1 file changed, 95 insertions(+), 67 deletions(-) diff --git a/comfy_api_nodes/nodes_wan.py b/comfy_api_nodes/nodes_wan.py index 2aab3c2ff..17b680e13 100644 --- a/comfy_api_nodes/nodes_wan.py +++ b/comfy_api_nodes/nodes_wan.py @@ -1,7 +1,5 @@ import re -from typing import Optional -import torch from pydantic import BaseModel, Field from typing_extensions import override @@ -21,26 +19,26 @@ from comfy_api_nodes.util import ( class Text2ImageInputField(BaseModel): prompt: str = Field(...) - negative_prompt: Optional[str] = Field(None) + negative_prompt: str | None = Field(None) class Image2ImageInputField(BaseModel): prompt: str = Field(...) - negative_prompt: Optional[str] = Field(None) + negative_prompt: str | None = Field(None) images: list[str] = Field(..., min_length=1, max_length=2) class Text2VideoInputField(BaseModel): prompt: str = Field(...) - negative_prompt: Optional[str] = Field(None) - audio_url: Optional[str] = Field(None) + negative_prompt: str | None = Field(None) + audio_url: str | None = Field(None) class Image2VideoInputField(BaseModel): prompt: str = Field(...) - negative_prompt: Optional[str] = Field(None) + negative_prompt: str | None = Field(None) img_url: str = Field(...) - audio_url: Optional[str] = Field(None) + audio_url: str | None = Field(None) class Txt2ImageParametersField(BaseModel): @@ -52,7 +50,7 @@ class Txt2ImageParametersField(BaseModel): class Image2ImageParametersField(BaseModel): - size: Optional[str] = Field(None) + size: str | None = Field(None) n: int = Field(1, description="Number of images to generate.") # we support only value=1 seed: int = Field(..., ge=0, le=2147483647) watermark: bool = Field(True) @@ -61,19 +59,21 @@ class Image2ImageParametersField(BaseModel): class Text2VideoParametersField(BaseModel): size: str = Field(...) seed: int = Field(..., ge=0, le=2147483647) - duration: int = Field(5, ge=5, le=10) + duration: int = Field(5, ge=5, le=15) prompt_extend: bool = Field(True) watermark: bool = Field(True) - audio: bool = Field(False, description="Should be audio generated automatically") + audio: bool = Field(False, description="Whether to generate audio automatically.") + shot_type: str = Field("single") class Image2VideoParametersField(BaseModel): resolution: str = Field(...) seed: int = Field(..., ge=0, le=2147483647) - duration: int = Field(5, ge=5, le=10) + duration: int = Field(5, ge=5, le=15) prompt_extend: bool = Field(True) watermark: bool = Field(True) - audio: bool = Field(False, description="Should be audio generated automatically") + audio: bool = Field(False, description="Whether to generate audio automatically.") + shot_type: str = Field("single") class Text2ImageTaskCreationRequest(BaseModel): @@ -106,39 +106,39 @@ class TaskCreationOutputField(BaseModel): class TaskCreationResponse(BaseModel): - output: Optional[TaskCreationOutputField] = Field(None) + output: TaskCreationOutputField | None = Field(None) request_id: str = Field(...) - code: Optional[str] = Field(None, description="The error code of the failed request.") - message: Optional[str] = Field(None, description="Details of the failed request.") + code: str | None = Field(None, description="Error code for the failed request.") + message: str | None = Field(None, description="Details about the failed request.") class TaskResult(BaseModel): - url: Optional[str] = Field(None) - code: Optional[str] = Field(None) - message: Optional[str] = Field(None) + url: str | None = Field(None) + code: str | None = Field(None) + message: str | None = Field(None) class ImageTaskStatusOutputField(TaskCreationOutputField): task_id: str = Field(...) task_status: str = Field(...) - results: Optional[list[TaskResult]] = Field(None) + results: list[TaskResult] | None = Field(None) class VideoTaskStatusOutputField(TaskCreationOutputField): task_id: str = Field(...) task_status: str = Field(...) - video_url: Optional[str] = Field(None) - code: Optional[str] = Field(None) - message: Optional[str] = Field(None) + video_url: str | None = Field(None) + code: str | None = Field(None) + message: str | None = Field(None) class ImageTaskStatusResponse(BaseModel): - output: Optional[ImageTaskStatusOutputField] = Field(None) + output: ImageTaskStatusOutputField | None = Field(None) request_id: str = Field(...) class VideoTaskStatusResponse(BaseModel): - output: Optional[VideoTaskStatusOutputField] = Field(None) + output: VideoTaskStatusOutputField | None = Field(None) request_id: str = Field(...) @@ -152,7 +152,7 @@ class WanTextToImageApi(IO.ComfyNode): node_id="WanTextToImageApi", display_name="Wan Text to Image", category="api node/image/Wan", - description="Generates image based on text prompt.", + description="Generates an image based on a text prompt.", inputs=[ IO.Combo.Input( "model", @@ -164,13 +164,13 @@ class WanTextToImageApi(IO.ComfyNode): "prompt", multiline=True, default="", - tooltip="Prompt used to describe the elements and visual features, supports English/Chinese.", + tooltip="Prompt describing the elements and visual features. Supports English and Chinese.", ), IO.String.Input( "negative_prompt", multiline=True, default="", - tooltip="Negative text prompt to guide what to avoid.", + tooltip="Negative prompt describing what to avoid.", optional=True, ), IO.Int.Input( @@ -209,7 +209,7 @@ class WanTextToImageApi(IO.ComfyNode): IO.Boolean.Input( "watermark", default=True, - tooltip='Whether to add an "AI generated" watermark to the result.', + tooltip="Whether to add an AI-generated watermark to the result.", optional=True, ), ], @@ -252,7 +252,7 @@ class WanTextToImageApi(IO.ComfyNode): ), ) if not initial_response.output: - raise Exception(f"Unknown error occurred: {initial_response.code} - {initial_response.message}") + raise Exception(f"An unknown error occurred: {initial_response.code} - {initial_response.message}") response = await poll_op( cls, ApiEndpoint(path=f"/proxy/wan/api/v1/tasks/{initial_response.output.task_id}"), @@ -272,7 +272,7 @@ class WanImageToImageApi(IO.ComfyNode): display_name="Wan Image to Image", category="api node/image/Wan", description="Generates an image from one or two input images and a text prompt. " - "The output image is currently fixed at 1.6 MP; its aspect ratio matches the input image(s).", + "The output image is currently fixed at 1.6 MP, and its aspect ratio matches the input image(s).", inputs=[ IO.Combo.Input( "model", @@ -282,19 +282,19 @@ class WanImageToImageApi(IO.ComfyNode): ), IO.Image.Input( "image", - tooltip="Single-image editing or multi-image fusion, maximum 2 images.", + tooltip="Single-image editing or multi-image fusion. Maximum 2 images.", ), IO.String.Input( "prompt", multiline=True, default="", - tooltip="Prompt used to describe the elements and visual features, supports English/Chinese.", + tooltip="Prompt describing the elements and visual features. Supports English and Chinese.", ), IO.String.Input( "negative_prompt", multiline=True, default="", - tooltip="Negative text prompt to guide what to avoid.", + tooltip="Negative prompt describing what to avoid.", optional=True, ), # redo this later as an optional combo of recommended resolutions @@ -328,7 +328,7 @@ class WanImageToImageApi(IO.ComfyNode): IO.Boolean.Input( "watermark", default=True, - tooltip='Whether to add an "AI generated" watermark to the result.', + tooltip="Whether to add an AI-generated watermark to the result.", optional=True, ), ], @@ -347,7 +347,7 @@ class WanImageToImageApi(IO.ComfyNode): async def execute( cls, model: str, - image: torch.Tensor, + image: Input.Image, prompt: str, negative_prompt: str = "", # width: int = 1024, @@ -357,7 +357,7 @@ class WanImageToImageApi(IO.ComfyNode): ): n_images = get_number_of_images(image) if n_images not in (1, 2): - raise ValueError(f"Expected 1 or 2 input images, got {n_images}.") + raise ValueError(f"Expected 1 or 2 input images, but got {n_images}.") images = [] for i in image: images.append("data:image/png;base64," + tensor_to_base64_string(i, total_pixels=4096 * 4096)) @@ -376,7 +376,7 @@ class WanImageToImageApi(IO.ComfyNode): ), ) if not initial_response.output: - raise Exception(f"Unknown error occurred: {initial_response.code} - {initial_response.message}") + raise Exception(f"An unknown error occurred: {initial_response.code} - {initial_response.message}") response = await poll_op( cls, ApiEndpoint(path=f"/proxy/wan/api/v1/tasks/{initial_response.output.task_id}"), @@ -395,25 +395,25 @@ class WanTextToVideoApi(IO.ComfyNode): node_id="WanTextToVideoApi", display_name="Wan Text to Video", category="api node/video/Wan", - description="Generates video based on text prompt.", + description="Generates a video based on a text prompt.", inputs=[ IO.Combo.Input( "model", - options=["wan2.5-t2v-preview"], - default="wan2.5-t2v-preview", + options=["wan2.5-t2v-preview", "wan2.6-t2v"], + default="wan2.6-t2v", tooltip="Model to use.", ), IO.String.Input( "prompt", multiline=True, default="", - tooltip="Prompt used to describe the elements and visual features, supports English/Chinese.", + tooltip="Prompt describing the elements and visual features. Supports English and Chinese.", ), IO.String.Input( "negative_prompt", multiline=True, default="", - tooltip="Negative text prompt to guide what to avoid.", + tooltip="Negative prompt describing what to avoid.", optional=True, ), IO.Combo.Input( @@ -433,23 +433,23 @@ class WanTextToVideoApi(IO.ComfyNode): "1080p: 4:3 (1632x1248)", "1080p: 3:4 (1248x1632)", ], - default="480p: 1:1 (624x624)", + default="720p: 1:1 (960x960)", optional=True, ), IO.Int.Input( "duration", default=5, min=5, - max=10, + max=15, step=5, display_mode=IO.NumberDisplay.number, - tooltip="Available durations: 5 and 10 seconds", + tooltip="A 15-second duration is available only for the Wan 2.6 model.", optional=True, ), IO.Audio.Input( "audio", optional=True, - tooltip="Audio must contain a clear, loud voice, without extraneous noise, background music.", + tooltip="Audio must contain a clear, loud voice, without extraneous noise or background music.", ), IO.Int.Input( "seed", @@ -466,7 +466,7 @@ class WanTextToVideoApi(IO.ComfyNode): "generate_audio", default=False, optional=True, - tooltip="If there is no audio input, generate audio automatically.", + tooltip="If no audio input is provided, generate audio automatically.", ), IO.Boolean.Input( "prompt_extend", @@ -477,7 +477,15 @@ class WanTextToVideoApi(IO.ComfyNode): IO.Boolean.Input( "watermark", default=True, - tooltip='Whether to add an "AI generated" watermark to the result.', + tooltip="Whether to add an AI-generated watermark to the result.", + optional=True, + ), + IO.Combo.Input( + "shot_type", + options=["single", "multi"], + tooltip="Specifies the shot type for the generated video, that is, whether the video is a " + "single continuous shot or multiple shots with cuts. " + "This parameter takes effect only when prompt_extend is True.", optional=True, ), ], @@ -498,14 +506,19 @@ class WanTextToVideoApi(IO.ComfyNode): model: str, prompt: str, negative_prompt: str = "", - size: str = "480p: 1:1 (624x624)", + size: str = "720p: 1:1 (960x960)", duration: int = 5, - audio: Optional[Input.Audio] = None, + audio: Input.Audio | None = None, seed: int = 0, generate_audio: bool = False, prompt_extend: bool = True, watermark: bool = True, + shot_type: str = "single", ): + if "480p" in size and model == "wan2.6-t2v": + raise ValueError("The Wan 2.6 model does not support 480p.") + if duration == 15 and model == "wan2.5-t2v-preview": + raise ValueError("A 15-second duration is supported only by the Wan 2.6 model.") width, height = RES_IN_PARENS.search(size).groups() audio_url = None if audio is not None: @@ -526,11 +539,12 @@ class WanTextToVideoApi(IO.ComfyNode): audio=generate_audio, prompt_extend=prompt_extend, watermark=watermark, + shot_type=shot_type, ), ), ) if not initial_response.output: - raise Exception(f"Unknown error occurred: {initial_response.code} - {initial_response.message}") + raise Exception(f"An unknown error occurred: {initial_response.code} - {initial_response.message}") response = await poll_op( cls, ApiEndpoint(path=f"/proxy/wan/api/v1/tasks/{initial_response.output.task_id}"), @@ -549,12 +563,12 @@ class WanImageToVideoApi(IO.ComfyNode): node_id="WanImageToVideoApi", display_name="Wan Image to Video", category="api node/video/Wan", - description="Generates video based on the first frame and text prompt.", + description="Generates a video from the first frame and a text prompt.", inputs=[ IO.Combo.Input( "model", - options=["wan2.5-i2v-preview"], - default="wan2.5-i2v-preview", + options=["wan2.5-i2v-preview", "wan2.6-i2v"], + default="wan2.6-i2v", tooltip="Model to use.", ), IO.Image.Input( @@ -564,13 +578,13 @@ class WanImageToVideoApi(IO.ComfyNode): "prompt", multiline=True, default="", - tooltip="Prompt used to describe the elements and visual features, supports English/Chinese.", + tooltip="Prompt describing the elements and visual features. Supports English and Chinese.", ), IO.String.Input( "negative_prompt", multiline=True, default="", - tooltip="Negative text prompt to guide what to avoid.", + tooltip="Negative prompt describing what to avoid.", optional=True, ), IO.Combo.Input( @@ -580,23 +594,23 @@ class WanImageToVideoApi(IO.ComfyNode): "720P", "1080P", ], - default="480P", + default="720P", optional=True, ), IO.Int.Input( "duration", default=5, min=5, - max=10, + max=15, step=5, display_mode=IO.NumberDisplay.number, - tooltip="Available durations: 5 and 10 seconds", + tooltip="Duration 15 available only for WAN2.6 model.", optional=True, ), IO.Audio.Input( "audio", optional=True, - tooltip="Audio must contain a clear, loud voice, without extraneous noise, background music.", + tooltip="Audio must contain a clear, loud voice, without extraneous noise or background music.", ), IO.Int.Input( "seed", @@ -613,7 +627,7 @@ class WanImageToVideoApi(IO.ComfyNode): "generate_audio", default=False, optional=True, - tooltip="If there is no audio input, generate audio automatically.", + tooltip="If no audio input is provided, generate audio automatically.", ), IO.Boolean.Input( "prompt_extend", @@ -624,7 +638,15 @@ class WanImageToVideoApi(IO.ComfyNode): IO.Boolean.Input( "watermark", default=True, - tooltip='Whether to add an "AI generated" watermark to the result.', + tooltip="Whether to add an AI-generated watermark to the result.", + optional=True, + ), + IO.Combo.Input( + "shot_type", + options=["single", "multi"], + tooltip="Specifies the shot type for the generated video, that is, whether the video is a " + "single continuous shot or multiple shots with cuts. " + "This parameter takes effect only when prompt_extend is True.", optional=True, ), ], @@ -643,19 +665,24 @@ class WanImageToVideoApi(IO.ComfyNode): async def execute( cls, model: str, - image: torch.Tensor, + image: Input.Image, prompt: str, negative_prompt: str = "", - resolution: str = "480P", + resolution: str = "720P", duration: int = 5, - audio: Optional[Input.Audio] = None, + audio: Input.Audio | None = None, seed: int = 0, generate_audio: bool = False, prompt_extend: bool = True, watermark: bool = True, + shot_type: str = "single", ): if get_number_of_images(image) != 1: raise ValueError("Exactly one input image is required.") + if "480P" in resolution and model == "wan2.6-i2v": + raise ValueError("The Wan 2.6 model does not support 480P.") + if duration == 15 and model == "wan2.5-i2v-preview": + raise ValueError("A 15-second duration is supported only by the Wan 2.6 model.") image_url = "data:image/png;base64," + tensor_to_base64_string(image, total_pixels=2000 * 2000) audio_url = None if audio is not None: @@ -677,11 +704,12 @@ class WanImageToVideoApi(IO.ComfyNode): audio=generate_audio, prompt_extend=prompt_extend, watermark=watermark, + shot_type=shot_type, ), ), ) if not initial_response.output: - raise Exception(f"Unknown error occurred: {initial_response.code} - {initial_response.message}") + raise Exception(f"An unknown error occurred: {initial_response.code} - {initial_response.message}") response = await poll_op( cls, ApiEndpoint(path=f"/proxy/wan/api/v1/tasks/{initial_response.output.task_id}"), From ffdd53b327f7ebd48cf81a1c8b06d846cf354a66 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 16 Dec 2025 14:03:17 -0800 Subject: [PATCH 301/647] Check state dict key to auto enable the index_timestep_zero ref method. (#11362) --- comfy/ldm/qwen_image/model.py | 3 +++ comfy/model_detection.py | 4 +++- 2 files changed, 6 insertions(+), 1 deletion(-) diff --git a/comfy/ldm/qwen_image/model.py b/comfy/ldm/qwen_image/model.py index 8481f7711..902af30ed 100644 --- a/comfy/ldm/qwen_image/model.py +++ b/comfy/ldm/qwen_image/model.py @@ -363,6 +363,9 @@ class QwenImageTransformer2DModel(nn.Module): for _ in range(num_layers) ]) + if self.default_ref_method == "index_timestep_zero": + self.register_buffer("__index_timestep_zero__", torch.tensor([])) + if final_layer: self.norm_out = LastLayer(self.inner_dim, self.inner_dim, dtype=dtype, device=device, operations=operations) self.proj_out = operations.Linear(self.inner_dim, patch_size * patch_size * self.out_channels, bias=True, dtype=dtype, device=device) diff --git a/comfy/model_detection.py b/comfy/model_detection.py index dd6a703f6..7148c77fd 100644 --- a/comfy/model_detection.py +++ b/comfy/model_detection.py @@ -259,7 +259,7 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): dit_config["nerf_tile_size"] = 512 dit_config["nerf_final_head_type"] = "conv" if f"{key_prefix}nerf_final_layer_conv.norm.scale" in state_dict_keys else "linear" dit_config["nerf_embedder_dtype"] = torch.float32 - if "__x0__" in state_dict_keys: # x0 pred + if "{}__x0__".format(key_prefix) in state_dict_keys: # x0 pred dit_config["use_x0"] = True else: dit_config["use_x0"] = False @@ -618,6 +618,8 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): dit_config["image_model"] = "qwen_image" dit_config["in_channels"] = state_dict['{}img_in.weight'.format(key_prefix)].shape[1] dit_config["num_layers"] = count_blocks(state_dict_keys, '{}transformer_blocks.'.format(key_prefix) + '{}.') + if "{}__index_timestep_zero__".format(key_prefix) in state_dict_keys: # 2511 + dit_config["default_ref_method"] = "index_timestep_zero" return dit_config if '{}visual_transformer_blocks.0.cross_attention.key_norm.weight'.format(key_prefix) in state_dict_keys: # Kandinsky 5 From 827bb1512b17e349238e69b2d4f463390a5b0d14 Mon Sep 17 00:00:00 2001 From: chaObserv <154517000+chaObserv@users.noreply.github.com> Date: Wed, 17 Dec 2025 12:35:43 +0800 Subject: [PATCH 302/647] Add exp_heun_2_x0 sampler series (#11360) --- comfy/k_diffusion/sampling.py | 11 +++++++++++ comfy/samplers.py | 2 +- comfy_extras/nodes_custom_sampler.py | 11 ++++++++++- 3 files changed, 22 insertions(+), 2 deletions(-) diff --git a/comfy/k_diffusion/sampling.py b/comfy/k_diffusion/sampling.py index 753c66afa..c004b3b47 100644 --- a/comfy/k_diffusion/sampling.py +++ b/comfy/k_diffusion/sampling.py @@ -1618,6 +1618,17 @@ def sample_seeds_2(model, x, sigmas, extra_args=None, callback=None, disable=Non x = x + sde_noise * sigmas[i + 1] * s_noise return x +@torch.no_grad() +def sample_exp_heun_2_x0(model, x, sigmas, extra_args=None, callback=None, disable=None, solver_type="phi_2"): + """Deterministic exponential Heun second order method in data prediction (x0) and logSNR time.""" + return sample_seeds_2(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=0.0, s_noise=0.0, noise_sampler=None, r=1.0, solver_type=solver_type) + + +@torch.no_grad() +def sample_exp_heun_2_x0_sde(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, solver_type="phi_2"): + """Stochastic exponential Heun second order method in data prediction (x0) and logSNR time.""" + return sample_seeds_2(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=eta, s_noise=s_noise, noise_sampler=noise_sampler, r=1.0, solver_type=solver_type) + @torch.no_grad() def sample_seeds_3(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, r_1=1./3, r_2=2./3): diff --git a/comfy/samplers.py b/comfy/samplers.py index fa4640842..8340d376c 100755 --- a/comfy/samplers.py +++ b/comfy/samplers.py @@ -720,7 +720,7 @@ class Sampler: sigma = float(sigmas[0]) return math.isclose(max_sigma, sigma, rel_tol=1e-05) or sigma > max_sigma -KSAMPLER_NAMES = ["euler", "euler_cfg_pp", "euler_ancestral", "euler_ancestral_cfg_pp", "heun", "heunpp2","dpm_2", "dpm_2_ancestral", +KSAMPLER_NAMES = ["euler", "euler_cfg_pp", "euler_ancestral", "euler_ancestral_cfg_pp", "heun", "heunpp2", "exp_heun_2_x0", "exp_heun_2_x0_sde", "dpm_2", "dpm_2_ancestral", "lms", "dpm_fast", "dpm_adaptive", "dpmpp_2s_ancestral", "dpmpp_2s_ancestral_cfg_pp", "dpmpp_sde", "dpmpp_sde_gpu", "dpmpp_2m", "dpmpp_2m_cfg_pp", "dpmpp_2m_sde", "dpmpp_2m_sde_gpu", "dpmpp_2m_sde_heun", "dpmpp_2m_sde_heun_gpu", "dpmpp_3m_sde", "dpmpp_3m_sde_gpu", "ddpm", "lcm", "ipndm", "ipndm_v", "deis", "res_multistep", "res_multistep_cfg_pp", "res_multistep_ancestral", "res_multistep_ancestral_cfg_pp", diff --git a/comfy_extras/nodes_custom_sampler.py b/comfy_extras/nodes_custom_sampler.py index 71ea4e9ec..7ee4caac1 100644 --- a/comfy_extras/nodes_custom_sampler.py +++ b/comfy_extras/nodes_custom_sampler.py @@ -671,7 +671,16 @@ class SamplerSEEDS2(io.ComfyNode): io.Float.Input("s_noise", default=1.0, min=0.0, max=100.0, step=0.01, round=False, tooltip="SDE noise multiplier"), io.Float.Input("r", default=0.5, min=0.01, max=1.0, step=0.01, round=False, tooltip="Relative step size for the intermediate stage (c2 node)"), ], - outputs=[io.Sampler.Output()] + outputs=[io.Sampler.Output()], + description=( + "This sampler node can represent multiple samplers:\n\n" + "seeds_2\n" + "- default setting\n\n" + "exp_heun_2_x0\n" + "- solver_type=phi_2, r=1.0, eta=0.0\n\n" + "exp_heun_2_x0_sde\n" + "- solver_type=phi_2, r=1.0, eta=1.0, s_noise=1.0" + ) ) @classmethod From 3a5f239cb622d7d8b1706d0b63c469dfef2eaf73 Mon Sep 17 00:00:00 2001 From: comfyanonymous Date: Wed, 17 Dec 2025 03:46:11 -0500 Subject: [PATCH 303/647] ComfyUI v0.5.0 --- comfyui_version.py | 2 +- pyproject.toml | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/comfyui_version.py b/comfyui_version.py index 2f083edaf..5edf270e7 100644 --- a/comfyui_version.py +++ b/comfyui_version.py @@ -1,3 +1,3 @@ # This file is automatically generated by the build process when version is # updated in pyproject.toml. -__version__ = "0.4.0" +__version__ = "0.5.0" diff --git a/pyproject.toml b/pyproject.toml index e4d3d616a..c402f278c 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -1,6 +1,6 @@ [project] name = "ComfyUI" -version = "0.4.0" +version = "0.5.0" readme = "README.md" license = { file = "LICENSE" } requires-python = ">=3.9" From 887143854bb2ae1e0f975e4461f376844a1628c8 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Wed, 17 Dec 2025 19:43:41 +0200 Subject: [PATCH 304/647] feat(api-nodes): add GPT-Image-1.5 (#11368) --- comfy_api_nodes/apis/openai_api.py | 52 +++++++ comfy_api_nodes/nodes_openai.py | 209 +++++++++++++++------------- comfy_api_nodes/util/conversions.py | 2 +- 3 files changed, 168 insertions(+), 95 deletions(-) create mode 100644 comfy_api_nodes/apis/openai_api.py diff --git a/comfy_api_nodes/apis/openai_api.py b/comfy_api_nodes/apis/openai_api.py new file mode 100644 index 000000000..ae5bb2673 --- /dev/null +++ b/comfy_api_nodes/apis/openai_api.py @@ -0,0 +1,52 @@ +from pydantic import BaseModel, Field + + +class Datum2(BaseModel): + b64_json: str | None = Field(None, description="Base64 encoded image data") + revised_prompt: str | None = Field(None, description="Revised prompt") + url: str | None = Field(None, description="URL of the image") + + +class InputTokensDetails(BaseModel): + image_tokens: int | None = None + text_tokens: int | None = None + + +class Usage(BaseModel): + input_tokens: int | None = None + input_tokens_details: InputTokensDetails | None = None + output_tokens: int | None = None + total_tokens: int | None = None + + +class OpenAIImageGenerationResponse(BaseModel): + data: list[Datum2] | None = None + usage: Usage | None = None + + +class OpenAIImageEditRequest(BaseModel): + background: str | None = Field(None, description="Background transparency") + model: str = Field(...) + moderation: str | None = Field(None) + n: int | None = Field(None, description="The number of images to generate") + output_compression: int | None = Field(None, description="Compression level for JPEG or WebP (0-100)") + output_format: str | None = Field(None) + prompt: str = Field(...) + quality: str | None = Field(None, description="Size of the image (e.g., 1024x1024, 1536x1024, auto)") + size: str | None = Field(None, description="Size of the output image") + + +class OpenAIImageGenerationRequest(BaseModel): + background: str | None = Field(None, description="Background transparency") + model: str | None = Field(None) + moderation: str | None = Field(None) + n: int | None = Field( + None, + description="The number of images to generate.", + ) + output_compression: int | None = Field(None, description="Compression level for JPEG or WebP (0-100)") + output_format: str | None = Field(None) + prompt: str = Field(...) + quality: str | None = Field(None, description="The quality of the generated image") + size: str | None = Field(None, description="Size of the image (e.g., 1024x1024, 1536x1024, auto)") + style: str | None = Field(None, description="Style of the image (only for dall-e-3)") diff --git a/comfy_api_nodes/nodes_openai.py b/comfy_api_nodes/nodes_openai.py index c8da5464b..a6205a34f 100644 --- a/comfy_api_nodes/nodes_openai.py +++ b/comfy_api_nodes/nodes_openai.py @@ -1,46 +1,45 @@ -from io import BytesIO +import base64 import os from enum import Enum -from inspect import cleandoc +from io import BytesIO + import numpy as np import torch from PIL import Image -import folder_paths -import base64 -from comfy_api.latest import IO, ComfyExtension from typing_extensions import override - +import folder_paths +from comfy_api.latest import IO, ComfyExtension, Input from comfy_api_nodes.apis import ( - OpenAIImageGenerationRequest, - OpenAIImageEditRequest, - OpenAIImageGenerationResponse, - OpenAICreateResponse, - OpenAIResponse, CreateModelResponseProperties, - Item, - OutputContent, - InputImageContent, Detail, - InputTextContent, - InputMessage, - InputMessageContentList, InputContent, InputFileContent, + InputImageContent, + InputMessage, + InputMessageContentList, + InputTextContent, + Item, + OpenAICreateResponse, + OpenAIResponse, + OutputContent, +) +from comfy_api_nodes.apis.openai_api import ( + OpenAIImageEditRequest, + OpenAIImageGenerationRequest, + OpenAIImageGenerationResponse, ) - from comfy_api_nodes.util import ( - downscale_image_tensor, - download_url_to_bytesio, - validate_string, - tensor_to_base64_string, ApiEndpoint, - sync_op, + download_url_to_bytesio, + downscale_image_tensor, poll_op, + sync_op, + tensor_to_base64_string, text_filepath_to_data_uri, + validate_string, ) - RESPONSES_ENDPOINT = "/proxy/openai/v1/responses" STARTING_POINT_ID_PATTERN = r"" @@ -98,9 +97,6 @@ async def validate_and_cast_response(response, timeout: int = None) -> torch.Ten class OpenAIDalle2(IO.ComfyNode): - """ - Generates images synchronously via OpenAI's DALL·E 2 endpoint. - """ @classmethod def define_schema(cls): @@ -108,7 +104,7 @@ class OpenAIDalle2(IO.ComfyNode): node_id="OpenAIDalle2", display_name="OpenAI DALL·E 2", category="api node/image/OpenAI", - description=cleandoc(cls.__doc__ or ""), + description="Generates images synchronously via OpenAI's DALL·E 2 endpoint.", inputs=[ IO.String.Input( "prompt", @@ -234,9 +230,6 @@ class OpenAIDalle2(IO.ComfyNode): class OpenAIDalle3(IO.ComfyNode): - """ - Generates images synchronously via OpenAI's DALL·E 3 endpoint. - """ @classmethod def define_schema(cls): @@ -244,7 +237,7 @@ class OpenAIDalle3(IO.ComfyNode): node_id="OpenAIDalle3", display_name="OpenAI DALL·E 3", category="api node/image/OpenAI", - description=cleandoc(cls.__doc__ or ""), + description="Generates images synchronously via OpenAI's DALL·E 3 endpoint.", inputs=[ IO.String.Input( "prompt", @@ -326,10 +319,16 @@ class OpenAIDalle3(IO.ComfyNode): return IO.NodeOutput(await validate_and_cast_response(response)) +def calculate_tokens_price_image_1(response: OpenAIImageGenerationResponse) -> float | None: + # https://platform.openai.com/docs/pricing + return ((response.usage.input_tokens * 10.0) + (response.usage.output_tokens * 40.0)) / 1_000_000.0 + + +def calculate_tokens_price_image_1_5(response: OpenAIImageGenerationResponse) -> float | None: + return ((response.usage.input_tokens * 8.0) + (response.usage.output_tokens * 32.0)) / 1_000_000.0 + + class OpenAIGPTImage1(IO.ComfyNode): - """ - Generates images synchronously via OpenAI's GPT Image 1 endpoint. - """ @classmethod def define_schema(cls): @@ -337,13 +336,13 @@ class OpenAIGPTImage1(IO.ComfyNode): node_id="OpenAIGPTImage1", display_name="OpenAI GPT Image 1", category="api node/image/OpenAI", - description=cleandoc(cls.__doc__ or ""), + description="Generates images synchronously via OpenAI's GPT Image 1 endpoint.", inputs=[ IO.String.Input( "prompt", default="", multiline=True, - tooltip="Text prompt for GPT Image 1", + tooltip="Text prompt for GPT Image", ), IO.Int.Input( "seed", @@ -365,8 +364,8 @@ class OpenAIGPTImage1(IO.ComfyNode): ), IO.Combo.Input( "background", - default="opaque", - options=["opaque", "transparent"], + default="auto", + options=["auto", "opaque", "transparent"], tooltip="Return image with or without background", optional=True, ), @@ -397,6 +396,11 @@ class OpenAIGPTImage1(IO.ComfyNode): tooltip="Optional mask for inpainting (white areas will be replaced)", optional=True, ), + IO.Combo.Input( + "model", + options=["gpt-image-1", "gpt-image-1.5"], + optional=True, + ), ], outputs=[ IO.Image.Output(), @@ -412,32 +416,34 @@ class OpenAIGPTImage1(IO.ComfyNode): @classmethod async def execute( cls, - prompt, - seed=0, - quality="low", - background="opaque", - image=None, - mask=None, - n=1, - size="1024x1024", + prompt: str, + seed: int = 0, + quality: str = "low", + background: str = "opaque", + image: Input.Image | None = None, + mask: Input.Image | None = None, + n: int = 1, + size: str = "1024x1024", + model: str = "gpt-image-1", ) -> IO.NodeOutput: validate_string(prompt, strip_whitespace=False) - model = "gpt-image-1" - path = "/proxy/openai/images/generations" - content_type = "application/json" - request_class = OpenAIImageGenerationRequest - files = [] + + if mask is not None and image is None: + raise ValueError("Cannot use a mask without an input image") + + if model == "gpt-image-1": + price_extractor = calculate_tokens_price_image_1 + elif model == "gpt-image-1.5": + price_extractor = calculate_tokens_price_image_1_5 + else: + raise ValueError(f"Unknown model: {model}") if image is not None: - path = "/proxy/openai/images/edits" - request_class = OpenAIImageEditRequest - content_type = "multipart/form-data" - + files = [] batch_size = image.shape[0] - for i in range(batch_size): - single_image = image[i : i + 1] - scaled_image = downscale_image_tensor(single_image).squeeze() + single_image = image[i: i + 1] + scaled_image = downscale_image_tensor(single_image, total_pixels=2048*2048).squeeze() image_np = (scaled_image.numpy() * 255).astype(np.uint8) img = Image.fromarray(image_np) @@ -450,44 +456,59 @@ class OpenAIGPTImage1(IO.ComfyNode): else: files.append(("image[]", (f"image_{i}.png", img_byte_arr, "image/png"))) - if mask is not None: - if image is None: - raise Exception("Cannot use a mask without an input image") - if image.shape[0] != 1: - raise Exception("Cannot use a mask with multiple image") - if mask.shape[1:] != image.shape[1:-1]: - raise Exception("Mask and Image must be the same size") - batch, height, width = mask.shape - rgba_mask = torch.zeros(height, width, 4, device="cpu") - rgba_mask[:, :, 3] = 1 - mask.squeeze().cpu() + if mask is not None: + if image.shape[0] != 1: + raise Exception("Cannot use a mask with multiple image") + if mask.shape[1:] != image.shape[1:-1]: + raise Exception("Mask and Image must be the same size") + _, height, width = mask.shape + rgba_mask = torch.zeros(height, width, 4, device="cpu") + rgba_mask[:, :, 3] = 1 - mask.squeeze().cpu() - scaled_mask = downscale_image_tensor(rgba_mask.unsqueeze(0)).squeeze() + scaled_mask = downscale_image_tensor(rgba_mask.unsqueeze(0), total_pixels=2048*2048).squeeze() - mask_np = (scaled_mask.numpy() * 255).astype(np.uint8) - mask_img = Image.fromarray(mask_np) - mask_img_byte_arr = BytesIO() - mask_img.save(mask_img_byte_arr, format="PNG") - mask_img_byte_arr.seek(0) - files.append(("mask", ("mask.png", mask_img_byte_arr, "image/png"))) - - # Build the operation - response = await sync_op( - cls, - ApiEndpoint(path=path, method="POST"), - response_model=OpenAIImageGenerationResponse, - data=request_class( - model=model, - prompt=prompt, - quality=quality, - background=background, - n=n, - seed=seed, - size=size, - ), - files=files if files else None, - content_type=content_type, - ) + mask_np = (scaled_mask.numpy() * 255).astype(np.uint8) + mask_img = Image.fromarray(mask_np) + mask_img_byte_arr = BytesIO() + mask_img.save(mask_img_byte_arr, format="PNG") + mask_img_byte_arr.seek(0) + files.append(("mask", ("mask.png", mask_img_byte_arr, "image/png"))) + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/openai/images/edits", method="POST"), + response_model=OpenAIImageGenerationResponse, + data=OpenAIImageEditRequest( + model=model, + prompt=prompt, + quality=quality, + background=background, + n=n, + seed=seed, + size=size, + moderation="low", + ), + content_type="multipart/form-data", + files=files, + price_extractor=price_extractor, + ) + else: + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/openai/images/generations", method="POST"), + response_model=OpenAIImageGenerationResponse, + data=OpenAIImageGenerationRequest( + model=model, + prompt=prompt, + quality=quality, + background=background, + n=n, + seed=seed, + size=size, + moderation="low", + ), + price_extractor=price_extractor, + ) return IO.NodeOutput(await validate_and_cast_response(response)) diff --git a/comfy_api_nodes/util/conversions.py b/comfy_api_nodes/util/conversions.py index c57457580..d64239c86 100644 --- a/comfy_api_nodes/util/conversions.py +++ b/comfy_api_nodes/util/conversions.py @@ -129,7 +129,7 @@ def pil_to_bytesio(img: Image.Image, mime_type: str = "image/png") -> BytesIO: return img_byte_arr -def downscale_image_tensor(image, total_pixels=1536 * 1024) -> torch.Tensor: +def downscale_image_tensor(image: torch.Tensor, total_pixels: int = 1536 * 1024) -> torch.Tensor: """Downscale input image tensor to roughly the specified total pixels.""" samples = image.movedim(-1, 1) total = int(total_pixels) From c08f97f34407a1bc6cc8d1447d6c12893399acba Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Wed, 17 Dec 2025 20:24:25 +0200 Subject: [PATCH 305/647] fix regression in V3 nodes processing (#11375) --- comfy_api/latest/_io.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/comfy_api/latest/_io.py b/comfy_api/latest/_io.py index 2b634d172..4b14e5ded 100644 --- a/comfy_api/latest/_io.py +++ b/comfy_api/latest/_io.py @@ -1556,12 +1556,12 @@ class _ComfyNodeBaseInternal(_ComfyNodeInternal): @final @classmethod - def PREPARE_CLASS_CLONE(cls, v3_data: V3Data) -> type[ComfyNode]: + def PREPARE_CLASS_CLONE(cls, v3_data: V3Data | None) -> type[ComfyNode]: """Creates clone of real node class to prevent monkey-patching.""" c_type: type[ComfyNode] = cls if is_class(cls) else type(cls) type_clone: type[ComfyNode] = shallow_clone_class(c_type) # set hidden - type_clone.hidden = HiddenHolder.from_dict(v3_data["hidden_inputs"]) + type_clone.hidden = HiddenHolder.from_dict(v3_data["hidden_inputs"] if v3_data else None) return type_clone @final From 5d9ad0c6bf177095aea5026cd872b1faf873669b Mon Sep 17 00:00:00 2001 From: chaObserv <154517000+chaObserv@users.noreply.github.com> Date: Thu, 18 Dec 2025 02:57:40 +0800 Subject: [PATCH 306/647] Fix the last step with non-zero sigma in sa_solver (#11380) --- comfy/k_diffusion/sampling.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/comfy/k_diffusion/sampling.py b/comfy/k_diffusion/sampling.py index c004b3b47..1ba9edad7 100644 --- a/comfy/k_diffusion/sampling.py +++ b/comfy/k_diffusion/sampling.py @@ -1776,7 +1776,7 @@ def sample_sa_solver(model, x, sigmas, extra_args=None, callback=None, disable=F # Predictor if sigmas[i + 1] == 0: # Denoising step - x = denoised + x_pred = denoised else: tau_t = tau_func(sigmas[i + 1]) curr_lambdas = lambdas[i - predictor_order_used + 1:i + 1] @@ -1797,7 +1797,7 @@ def sample_sa_solver(model, x, sigmas, extra_args=None, callback=None, disable=F if tau_t > 0 and s_noise > 0: noise = noise_sampler(sigmas[i], sigmas[i + 1]) * sigmas[i + 1] * (-2 * tau_t ** 2 * h).expm1().neg().sqrt() * s_noise x_pred = x_pred + noise - return x + return x_pred @torch.no_grad() From 16d85ea13342cebc8349a95236c94bde5ac3cb2a Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 17 Dec 2025 16:43:18 -0800 Subject: [PATCH 307/647] Better handle torch being imported by prestartup nodes. (#11383) --- main.py | 66 ++++++++++++++++++++++++++++----------------------------- 1 file changed, 32 insertions(+), 34 deletions(-) diff --git a/main.py b/main.py index 0d02a087b..0e07a95da 100644 --- a/main.py +++ b/main.py @@ -23,6 +23,38 @@ if __name__ == "__main__": setup_logger(log_level=args.verbose, use_stdout=args.log_stdout) +if os.name == "nt": + os.environ['MIMALLOC_PURGE_DELAY'] = '0' + +if __name__ == "__main__": + os.environ['TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL'] = '1' + if args.default_device is not None: + default_dev = args.default_device + devices = list(range(32)) + devices.remove(default_dev) + devices.insert(0, default_dev) + devices = ','.join(map(str, devices)) + os.environ['CUDA_VISIBLE_DEVICES'] = str(devices) + os.environ['HIP_VISIBLE_DEVICES'] = str(devices) + + if args.cuda_device is not None: + os.environ['CUDA_VISIBLE_DEVICES'] = str(args.cuda_device) + os.environ['HIP_VISIBLE_DEVICES'] = str(args.cuda_device) + os.environ["ASCEND_RT_VISIBLE_DEVICES"] = str(args.cuda_device) + logging.info("Set cuda device to: {}".format(args.cuda_device)) + + if args.oneapi_device_selector is not None: + os.environ['ONEAPI_DEVICE_SELECTOR'] = args.oneapi_device_selector + logging.info("Set oneapi device selector to: {}".format(args.oneapi_device_selector)) + + if args.deterministic: + if 'CUBLAS_WORKSPACE_CONFIG' not in os.environ: + os.environ['CUBLAS_WORKSPACE_CONFIG'] = ":4096:8" + + import cuda_malloc + if "rocm" in cuda_malloc.get_torch_version_noimport(): + os.environ['OCL_SET_SVM_SIZE'] = '262144' # set at the request of AMD + def handle_comfyui_manager_unavailable(): if not args.windows_standalone_build: @@ -137,40 +169,6 @@ import shutil import threading import gc - -if os.name == "nt": - os.environ['MIMALLOC_PURGE_DELAY'] = '0' - -if __name__ == "__main__": - os.environ['TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL'] = '1' - if args.default_device is not None: - default_dev = args.default_device - devices = list(range(32)) - devices.remove(default_dev) - devices.insert(0, default_dev) - devices = ','.join(map(str, devices)) - os.environ['CUDA_VISIBLE_DEVICES'] = str(devices) - os.environ['HIP_VISIBLE_DEVICES'] = str(devices) - - if args.cuda_device is not None: - os.environ['CUDA_VISIBLE_DEVICES'] = str(args.cuda_device) - os.environ['HIP_VISIBLE_DEVICES'] = str(args.cuda_device) - os.environ["ASCEND_RT_VISIBLE_DEVICES"] = str(args.cuda_device) - logging.info("Set cuda device to: {}".format(args.cuda_device)) - - if args.oneapi_device_selector is not None: - os.environ['ONEAPI_DEVICE_SELECTOR'] = args.oneapi_device_selector - logging.info("Set oneapi device selector to: {}".format(args.oneapi_device_selector)) - - if args.deterministic: - if 'CUBLAS_WORKSPACE_CONFIG' not in os.environ: - os.environ['CUBLAS_WORKSPACE_CONFIG'] = ":4096:8" - - import cuda_malloc - if "rocm" in cuda_malloc.get_torch_version_noimport(): - os.environ['OCL_SET_SVM_SIZE'] = '262144' # set at the request of AMD - - if 'torch' in sys.modules: logging.warning("WARNING: Potential Error in code: Torch already imported, torch should never be imported before this point.") From ba6080bbab070934ea6e870c5fc30dbf702eb445 Mon Sep 17 00:00:00 2001 From: comfyanonymous Date: Wed, 17 Dec 2025 21:04:50 -0500 Subject: [PATCH 308/647] ComfyUI v0.5.1 --- comfyui_version.py | 2 +- pyproject.toml | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/comfyui_version.py b/comfyui_version.py index 5edf270e7..b45309198 100644 --- a/comfyui_version.py +++ b/comfyui_version.py @@ -1,3 +1,3 @@ # This file is automatically generated by the build process when version is # updated in pyproject.toml. -__version__ = "0.5.0" +__version__ = "0.5.1" diff --git a/pyproject.toml b/pyproject.toml index c402f278c..3a6960811 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -1,6 +1,6 @@ [project] name = "ComfyUI" -version = "0.5.0" +version = "0.5.1" readme = "README.md" license = { file = "LICENSE" } requires-python = ">=3.9" From 86dbb89fc95f0cb652ae5d6cb923f641a58e295d Mon Sep 17 00:00:00 2001 From: Kohaku-Blueleaf <59680068+KohakuBlueleaf@users.noreply.github.com> Date: Thu, 18 Dec 2025 11:15:27 +0800 Subject: [PATCH 309/647] Resolution bucketing and Trainer implementation refactoring (#11117) --- comfy/sampler_helpers.py | 9 +- comfy_extras/nodes_dataset.py | 96 ++- comfy_extras/nodes_post_processing.py | 11 +- comfy_extras/nodes_train.py | 854 +++++++++++++++++++------- 4 files changed, 738 insertions(+), 232 deletions(-) diff --git a/comfy/sampler_helpers.py b/comfy/sampler_helpers.py index e46971afb..e158e8a84 100644 --- a/comfy/sampler_helpers.py +++ b/comfy/sampler_helpers.py @@ -122,20 +122,21 @@ def estimate_memory(model, noise_shape, conds): minimum_memory_required = model.model.memory_required([noise_shape[0]] + list(noise_shape[1:]), cond_shapes=cond_shapes_min) return memory_required, minimum_memory_required -def prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None): +def prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None, skip_load_model=False): executor = comfy.patcher_extension.WrapperExecutor.new_executor( _prepare_sampling, comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.PREPARE_SAMPLING, model_options, is_model_options=True) ) - return executor.execute(model, noise_shape, conds, model_options=model_options) + return executor.execute(model, noise_shape, conds, model_options=model_options, skip_load_model=skip_load_model) -def _prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None): +def _prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None, skip_load_model=False): real_model: BaseModel = None models, inference_memory = get_additional_models(conds, model.model_dtype()) models += get_additional_models_from_model_options(model_options) models += model.get_nested_additional_models() # TODO: does this require inference_memory update? memory_required, minimum_memory_required = estimate_memory(model, noise_shape, conds) - comfy.model_management.load_models_gpu([model] + models, memory_required=memory_required + inference_memory, minimum_memory_required=minimum_memory_required + inference_memory) + models_list = [model] if not skip_load_model else [] + comfy.model_management.load_models_gpu(models_list + models, memory_required=memory_required + inference_memory, minimum_memory_required=minimum_memory_required + inference_memory) real_model = model.model return real_model, conds, models diff --git a/comfy_extras/nodes_dataset.py b/comfy_extras/nodes_dataset.py index 4789d7d53..513aecf3a 100644 --- a/comfy_extras/nodes_dataset.py +++ b/comfy_extras/nodes_dataset.py @@ -1125,6 +1125,99 @@ class MergeTextListsNode(TextProcessingNode): # ========== Training Dataset Nodes ========== +class ResolutionBucket(io.ComfyNode): + """Bucket latents and conditions by resolution for efficient batch training.""" + + @classmethod + def define_schema(cls): + return io.Schema( + node_id="ResolutionBucket", + display_name="Resolution Bucket", + category="dataset", + is_experimental=True, + is_input_list=True, + inputs=[ + io.Latent.Input( + "latents", + tooltip="List of latent dicts to bucket by resolution.", + ), + io.Conditioning.Input( + "conditioning", + tooltip="List of conditioning lists (must match latents length).", + ), + ], + outputs=[ + io.Latent.Output( + display_name="latents", + is_output_list=True, + tooltip="List of batched latent dicts, one per resolution bucket.", + ), + io.Conditioning.Output( + display_name="conditioning", + is_output_list=True, + tooltip="List of condition lists, one per resolution bucket.", + ), + ], + ) + + @classmethod + def execute(cls, latents, conditioning): + # latents: list[{"samples": tensor}] where tensor is (B, C, H, W), typically B=1 + # conditioning: list[list[cond]] + + # Validate lengths match + if len(latents) != len(conditioning): + raise ValueError( + f"Number of latents ({len(latents)}) does not match number of conditions ({len(conditioning)})." + ) + + # Flatten latents and conditions to individual samples + flat_latents = [] # list of (C, H, W) tensors + flat_conditions = [] # list of condition lists + + for latent_dict, cond in zip(latents, conditioning): + samples = latent_dict["samples"] # (B, C, H, W) + batch_size = samples.shape[0] + + # cond is a list of conditions with length == batch_size + for i in range(batch_size): + flat_latents.append(samples[i]) # (C, H, W) + flat_conditions.append(cond[i]) # single condition + + # Group by resolution (H, W) + buckets = {} # (H, W) -> {"latents": list, "conditions": list} + + for latent, cond in zip(flat_latents, flat_conditions): + # latent shape is (..., H, W) (B, C, H, W) or (B, T, C, H ,W) + h, w = latent.shape[-2], latent.shape[-1] + key = (h, w) + + if key not in buckets: + buckets[key] = {"latents": [], "conditions": []} + + buckets[key]["latents"].append(latent) + buckets[key]["conditions"].append(cond) + + # Convert buckets to output format + output_latents = [] # list[{"samples": tensor}] where tensor is (Bi, ..., H, W) + output_conditions = [] # list[list[cond]] where each inner list has Bi conditions + + for (h, w), bucket_data in buckets.items(): + # Stack latents into batch: list of (..., H, W) -> (Bi, ..., H, W) + stacked_latents = torch.stack(bucket_data["latents"], dim=0) + output_latents.append({"samples": stacked_latents}) + + # Conditions stay as list of condition lists + output_conditions.append(bucket_data["conditions"]) + + logging.info( + f"Resolution bucket ({h}x{w}): {len(bucket_data['latents'])} samples" + ) + + logging.info(f"Created {len(buckets)} resolution buckets from {len(flat_latents)} samples") + return io.NodeOutput(output_latents, output_conditions) + + class MakeTrainingDataset(io.ComfyNode): """Encode images with VAE and texts with CLIP to create a training dataset.""" @@ -1373,7 +1466,7 @@ class LoadTrainingDataset(io.ComfyNode): shard_path = os.path.join(dataset_dir, shard_file) with open(shard_path, "rb") as f: - shard_data = torch.load(f, weights_only=True) + shard_data = torch.load(f) all_latents.extend(shard_data["latents"]) all_conditioning.extend(shard_data["conditioning"]) @@ -1425,6 +1518,7 @@ class DatasetExtension(ComfyExtension): MakeTrainingDataset, SaveTrainingDataset, LoadTrainingDataset, + ResolutionBucket, ] diff --git a/comfy_extras/nodes_post_processing.py b/comfy_extras/nodes_post_processing.py index 34c388a5a..ca2cdeb50 100644 --- a/comfy_extras/nodes_post_processing.py +++ b/comfy_extras/nodes_post_processing.py @@ -221,6 +221,7 @@ class ImageScaleToTotalPixels(io.ComfyNode): io.Image.Input("image"), io.Combo.Input("upscale_method", options=cls.upscale_methods), io.Float.Input("megapixels", default=1.0, min=0.01, max=16.0, step=0.01), + io.Int.Input("resolution_steps", default=1, min=1, max=256), ], outputs=[ io.Image.Output(), @@ -228,15 +229,15 @@ class ImageScaleToTotalPixels(io.ComfyNode): ) @classmethod - def execute(cls, image, upscale_method, megapixels) -> io.NodeOutput: + def execute(cls, image, upscale_method, megapixels, resolution_steps) -> io.NodeOutput: samples = image.movedim(-1,1) - total = int(megapixels * 1024 * 1024) + total = megapixels * 1024 * 1024 scale_by = math.sqrt(total / (samples.shape[3] * samples.shape[2])) - width = round(samples.shape[3] * scale_by) - height = round(samples.shape[2] * scale_by) + width = round(samples.shape[3] * scale_by / resolution_steps) * resolution_steps + height = round(samples.shape[2] * scale_by / resolution_steps) * resolution_steps - s = comfy.utils.common_upscale(samples, width, height, upscale_method, "disabled") + s = comfy.utils.common_upscale(samples, int(width), int(height), upscale_method, "disabled") s = s.movedim(1,-1) return io.NodeOutput(s) diff --git a/comfy_extras/nodes_train.py b/comfy_extras/nodes_train.py index 19b8baaf4..88bc8c8e8 100644 --- a/comfy_extras/nodes_train.py +++ b/comfy_extras/nodes_train.py @@ -10,6 +10,7 @@ from PIL import Image, ImageDraw, ImageFont from typing_extensions import override import comfy.samplers +import comfy.sampler_helpers import comfy.sd import comfy.utils import comfy.model_management @@ -21,6 +22,68 @@ from comfy_api.latest import ComfyExtension, io, ui from comfy.utils import ProgressBar +class TrainGuider(comfy_extras.nodes_custom_sampler.Guider_Basic): + """ + CFGGuider with modifications for training specific logic + """ + def outer_sample( + self, + noise, + latent_image, + sampler, + sigmas, + denoise_mask=None, + callback=None, + disable_pbar=False, + seed=None, + latent_shapes=None, + ): + self.inner_model, self.conds, self.loaded_models = ( + comfy.sampler_helpers.prepare_sampling( + self.model_patcher, + noise.shape, + self.conds, + self.model_options, + skip_load_model=True, # skip load model as we manage it in TrainLoraNode.execute() + ) + ) + device = self.model_patcher.load_device + + if denoise_mask is not None: + denoise_mask = comfy.sampler_helpers.prepare_mask( + denoise_mask, noise.shape, device + ) + + noise = noise.to(device) + latent_image = latent_image.to(device) + sigmas = sigmas.to(device) + comfy.samplers.cast_to_load_options( + self.model_options, device=device, dtype=self.model_patcher.model_dtype() + ) + + try: + self.model_patcher.pre_run() + output = self.inner_sample( + noise, + latent_image, + device, + sampler, + sigmas, + denoise_mask, + callback, + disable_pbar, + seed, + latent_shapes=latent_shapes, + ) + finally: + self.model_patcher.cleanup() + + comfy.sampler_helpers.cleanup_models(self.conds, self.loaded_models) + del self.inner_model + del self.loaded_models + return output + + def make_batch_extra_option_dict(d, indicies, full_size=None): new_dict = {} for k, v in d.items(): @@ -65,6 +128,7 @@ class TrainSampler(comfy.samplers.Sampler): seed=0, training_dtype=torch.bfloat16, real_dataset=None, + bucket_latents=None, ): self.loss_fn = loss_fn self.optimizer = optimizer @@ -75,6 +139,28 @@ class TrainSampler(comfy.samplers.Sampler): self.seed = seed self.training_dtype = training_dtype self.real_dataset: list[torch.Tensor] | None = real_dataset + # Bucket mode data + self.bucket_latents: list[torch.Tensor] | None = ( + bucket_latents # list of (Bi, C, Hi, Wi) + ) + # Precompute bucket offsets and weights for sampling + if bucket_latents is not None: + self._init_bucket_data(bucket_latents) + else: + self.bucket_offsets = None + self.bucket_weights = None + self.num_images = None + + def _init_bucket_data(self, bucket_latents): + """Initialize bucket offsets and weights for sampling.""" + self.bucket_offsets = [0] + bucket_sizes = [] + for lat in bucket_latents: + bucket_sizes.append(lat.shape[0]) + self.bucket_offsets.append(self.bucket_offsets[-1] + lat.shape[0]) + self.num_images = self.bucket_offsets[-1] + # Weights for sampling buckets proportional to their size + self.bucket_weights = torch.tensor(bucket_sizes, dtype=torch.float32) def fwd_bwd( self, @@ -115,6 +201,108 @@ class TrainSampler(comfy.samplers.Sampler): bwd_loss.backward() return loss + def _generate_batch_sigmas(self, model_wrap, batch_size, device): + """Generate random sigma values for a batch.""" + batch_sigmas = [ + model_wrap.inner_model.model_sampling.percent_to_sigma( + torch.rand((1,)).item() + ) + for _ in range(batch_size) + ] + return torch.tensor(batch_sigmas).to(device) + + def _train_step_bucket_mode(self, model_wrap, cond, extra_args, noisegen, latent_image, pbar): + """Execute one training step in bucket mode.""" + # Sample bucket (weighted by size), then sample batch from bucket + bucket_idx = torch.multinomial(self.bucket_weights, 1).item() + bucket_latent = self.bucket_latents[bucket_idx] # (Bi, C, Hi, Wi) + bucket_size = bucket_latent.shape[0] + bucket_offset = self.bucket_offsets[bucket_idx] + + # Sample indices from this bucket (use all if bucket_size < batch_size) + actual_batch_size = min(self.batch_size, bucket_size) + relative_indices = torch.randperm(bucket_size)[:actual_batch_size].tolist() + # Convert to absolute indices for fwd_bwd (cond is flattened, use absolute index) + absolute_indices = [bucket_offset + idx for idx in relative_indices] + + batch_latent = bucket_latent[relative_indices].to(latent_image) # (actual_batch_size, C, H, W) + batch_noise = noisegen.generate_noise({"samples": batch_latent}).to( + batch_latent.device + ) + batch_sigmas = self._generate_batch_sigmas(model_wrap, actual_batch_size, batch_latent.device) + + loss = self.fwd_bwd( + model_wrap, + batch_sigmas, + batch_noise, + batch_latent, + cond, # Use flattened cond with absolute indices + absolute_indices, + extra_args, + self.num_images, + bwd=True, + ) + if self.loss_callback: + self.loss_callback(loss.item()) + pbar.set_postfix({"loss": f"{loss.item():.4f}", "bucket": bucket_idx}) + + def _train_step_standard_mode(self, model_wrap, cond, extra_args, noisegen, latent_image, dataset_size, pbar): + """Execute one training step in standard (non-bucket, non-multi-res) mode.""" + indicies = torch.randperm(dataset_size)[: self.batch_size].tolist() + batch_latent = torch.stack([latent_image[i] for i in indicies]) + batch_noise = noisegen.generate_noise({"samples": batch_latent}).to( + batch_latent.device + ) + batch_sigmas = self._generate_batch_sigmas(model_wrap, min(self.batch_size, dataset_size), batch_latent.device) + + loss = self.fwd_bwd( + model_wrap, + batch_sigmas, + batch_noise, + batch_latent, + cond, + indicies, + extra_args, + dataset_size, + bwd=True, + ) + if self.loss_callback: + self.loss_callback(loss.item()) + pbar.set_postfix({"loss": f"{loss.item():.4f}"}) + + def _train_step_multires_mode(self, model_wrap, cond, extra_args, noisegen, latent_image, dataset_size, pbar): + """Execute one training step in multi-resolution mode (real_dataset is set).""" + indicies = torch.randperm(dataset_size)[: self.batch_size].tolist() + total_loss = 0 + for index in indicies: + single_latent = self.real_dataset[index].to(latent_image) + batch_noise = noisegen.generate_noise( + {"samples": single_latent} + ).to(single_latent.device) + batch_sigmas = ( + model_wrap.inner_model.model_sampling.percent_to_sigma( + torch.rand((1,)).item() + ) + ) + batch_sigmas = torch.tensor([batch_sigmas]).to(single_latent.device) + loss = self.fwd_bwd( + model_wrap, + batch_sigmas, + batch_noise, + single_latent, + cond, + [index], + extra_args, + dataset_size, + bwd=False, + ) + total_loss += loss + total_loss = total_loss / self.grad_acc / len(indicies) + total_loss.backward() + if self.loss_callback: + self.loss_callback(total_loss.item()) + pbar.set_postfix({"loss": f"{total_loss.item():.4f}"}) + def sample( self, model_wrap, @@ -142,70 +330,18 @@ class TrainSampler(comfy.samplers.Sampler): noisegen = comfy_extras.nodes_custom_sampler.Noise_RandomNoise( self.seed + i * 1000 ) - indicies = torch.randperm(dataset_size)[: self.batch_size].tolist() - if self.real_dataset is None: - batch_latent = torch.stack([latent_image[i] for i in indicies]) - batch_noise = noisegen.generate_noise({"samples": batch_latent}).to( - batch_latent.device - ) - batch_sigmas = [ - model_wrap.inner_model.model_sampling.percent_to_sigma( - torch.rand((1,)).item() - ) - for _ in range(min(self.batch_size, dataset_size)) - ] - batch_sigmas = torch.tensor(batch_sigmas).to(batch_latent.device) - - loss = self.fwd_bwd( - model_wrap, - batch_sigmas, - batch_noise, - batch_latent, - cond, - indicies, - extra_args, - dataset_size, - bwd=True, - ) - if self.loss_callback: - self.loss_callback(loss.item()) - pbar.set_postfix({"loss": f"{loss.item():.4f}"}) + if self.bucket_latents is not None: + self._train_step_bucket_mode(model_wrap, cond, extra_args, noisegen, latent_image, pbar) + elif self.real_dataset is None: + self._train_step_standard_mode(model_wrap, cond, extra_args, noisegen, latent_image, dataset_size, pbar) else: - total_loss = 0 - for index in indicies: - single_latent = self.real_dataset[index].to(latent_image) - batch_noise = noisegen.generate_noise( - {"samples": single_latent} - ).to(single_latent.device) - batch_sigmas = ( - model_wrap.inner_model.model_sampling.percent_to_sigma( - torch.rand((1,)).item() - ) - ) - batch_sigmas = torch.tensor([batch_sigmas]).to(single_latent.device) - loss = self.fwd_bwd( - model_wrap, - batch_sigmas, - batch_noise, - single_latent, - cond, - [index], - extra_args, - dataset_size, - bwd=False, - ) - total_loss += loss - total_loss = total_loss / self.grad_acc / len(indicies) - total_loss.backward() - if self.loss_callback: - self.loss_callback(total_loss.item()) - pbar.set_postfix({"loss": f"{total_loss.item():.4f}"}) + self._train_step_multires_mode(model_wrap, cond, extra_args, noisegen, latent_image, dataset_size, pbar) if (i + 1) % self.grad_acc == 0: self.optimizer.step() self.optimizer.zero_grad() - ui_pbar.update(1) + ui_pbar.update(1) torch.cuda.empty_cache() return torch.zeros_like(latent_image) @@ -283,6 +419,364 @@ def unpatch(m): del m.org_forward +def _process_latents_bucket_mode(latents): + """Process latents for bucket mode training. + + Args: + latents: list[{"samples": tensor}] where each tensor is (Bi, C, Hi, Wi) + + Returns: + list of latent tensors + """ + bucket_latents = [] + for latent_dict in latents: + bucket_latents.append(latent_dict["samples"]) # (Bi, C, Hi, Wi) + return bucket_latents + + +def _process_latents_standard_mode(latents): + """Process latents for standard (non-bucket) mode training. + + Args: + latents: list of latent dicts or single latent dict + + Returns: + Processed latents (tensor or list of tensors) + """ + if len(latents) == 1: + return latents[0]["samples"] # Single latent dict + + latent_list = [] + for latent in latents: + latent = latent["samples"] + bs = latent.shape[0] + if bs != 1: + for sub_latent in latent: + latent_list.append(sub_latent[None]) + else: + latent_list.append(latent) + return latent_list + + +def _process_conditioning(positive): + """Process conditioning - either single list or list of lists. + + Args: + positive: list of conditioning + + Returns: + Flattened conditioning list + """ + if len(positive) == 1: + return positive[0] # Single conditioning list + + # Multiple conditioning lists - flatten + flat_positive = [] + for cond in positive: + if isinstance(cond, list): + flat_positive.extend(cond) + else: + flat_positive.append(cond) + return flat_positive + + +def _prepare_latents_and_count(latents, dtype, bucket_mode): + """Convert latents to dtype and compute image counts. + + Args: + latents: Latents (tensor, list of tensors, or bucket list) + dtype: Target dtype + bucket_mode: Whether bucket mode is enabled + + Returns: + tuple: (processed_latents, num_images, multi_res) + """ + if bucket_mode: + # In bucket mode, latents is list of tensors (Bi, C, Hi, Wi) + latents = [t.to(dtype) for t in latents] + num_buckets = len(latents) + num_images = sum(t.shape[0] for t in latents) + multi_res = False # Not using multi_res path in bucket mode + + logging.info(f"Bucket mode: {num_buckets} buckets, {num_images} total samples") + for i, lat in enumerate(latents): + logging.info(f" Bucket {i}: shape {lat.shape}") + return latents, num_images, multi_res + + # Non-bucket mode + if isinstance(latents, list): + all_shapes = set() + latents = [t.to(dtype) for t in latents] + for latent in latents: + all_shapes.add(latent.shape) + logging.info(f"Latent shapes: {all_shapes}") + if len(all_shapes) > 1: + multi_res = True + else: + multi_res = False + latents = torch.cat(latents, dim=0) + num_images = len(latents) + elif isinstance(latents, torch.Tensor): + latents = latents.to(dtype) + num_images = latents.shape[0] + multi_res = False + else: + logging.error(f"Invalid latents type: {type(latents)}") + num_images = 0 + multi_res = False + + return latents, num_images, multi_res + + +def _validate_and_expand_conditioning(positive, num_images, bucket_mode): + """Validate conditioning count matches image count, expand if needed. + + Args: + positive: Conditioning list + num_images: Number of images + bucket_mode: Whether bucket mode is enabled + + Returns: + Validated/expanded conditioning list + + Raises: + ValueError: If conditioning count doesn't match image count + """ + if bucket_mode: + return positive # Skip validation in bucket mode + + logging.info(f"Total Images: {num_images}, Total Captions: {len(positive)}") + if len(positive) == 1 and num_images > 1: + return positive * num_images + elif len(positive) != num_images: + raise ValueError( + f"Number of positive conditions ({len(positive)}) does not match number of images ({num_images})." + ) + return positive + + +def _load_existing_lora(existing_lora): + """Load existing LoRA weights if provided. + + Args: + existing_lora: LoRA filename or "[None]" + + Returns: + tuple: (existing_weights dict, existing_steps int) + """ + if existing_lora == "[None]": + return {}, 0 + + lora_path = folder_paths.get_full_path_or_raise("loras", existing_lora) + # Extract steps from filename like "trained_lora_10_steps_20250225_203716" + existing_steps = int(existing_lora.split("_steps_")[0].split("_")[-1]) + existing_weights = {} + if lora_path: + existing_weights = comfy.utils.load_torch_file(lora_path) + return existing_weights, existing_steps + + +def _create_weight_adapter( + module, module_name, existing_weights, algorithm, lora_dtype, rank +): + """Create a weight adapter for a module with weight. + + Args: + module: The module to create adapter for + module_name: Name of the module + existing_weights: Dict of existing LoRA weights + algorithm: Algorithm name for new adapters + lora_dtype: dtype for LoRA weights + rank: Rank for new LoRA adapters + + Returns: + tuple: (train_adapter, lora_params dict) + """ + key = f"{module_name}.weight" + shape = module.weight.shape + lora_params = {} + + if len(shape) >= 2: + alpha = float(existing_weights.get(f"{key}.alpha", 1.0)) + dora_scale = existing_weights.get(f"{key}.dora_scale", None) + + # Try to load existing adapter + existing_adapter = None + for adapter_cls in adapters: + existing_adapter = adapter_cls.load( + module_name, existing_weights, alpha, dora_scale + ) + if existing_adapter is not None: + break + + if existing_adapter is None: + adapter_cls = adapter_maps[algorithm] + + if existing_adapter is not None: + train_adapter = existing_adapter.to_train().to(lora_dtype) + else: + # Use LoRA with alpha=1.0 by default + train_adapter = adapter_cls.create_train( + module.weight, rank=rank, alpha=1.0 + ).to(lora_dtype) + + for name, parameter in train_adapter.named_parameters(): + lora_params[f"{module_name}.{name}"] = parameter + + return train_adapter.train().requires_grad_(True), lora_params + else: + # 1D weight - use BiasDiff + diff = torch.nn.Parameter( + torch.zeros(module.weight.shape, dtype=lora_dtype, requires_grad=True) + ) + diff_module = BiasDiff(diff).train().requires_grad_(True) + lora_params[f"{module_name}.diff"] = diff + return diff_module, lora_params + + +def _create_bias_adapter(module, module_name, lora_dtype): + """Create a bias adapter for a module with bias. + + Args: + module: The module with bias + module_name: Name of the module + lora_dtype: dtype for LoRA weights + + Returns: + tuple: (bias_module, lora_params dict) + """ + bias = torch.nn.Parameter( + torch.zeros(module.bias.shape, dtype=lora_dtype, requires_grad=True) + ) + bias_module = BiasDiff(bias).train().requires_grad_(True) + lora_params = {f"{module_name}.diff_b": bias} + return bias_module, lora_params + + +def _setup_lora_adapters(mp, existing_weights, algorithm, lora_dtype, rank): + """Setup all LoRA adapters on the model. + + Args: + mp: Model patcher + existing_weights: Dict of existing LoRA weights + algorithm: Algorithm name for new adapters + lora_dtype: dtype for LoRA weights + rank: Rank for new LoRA adapters + + Returns: + tuple: (lora_sd dict, all_weight_adapters list) + """ + lora_sd = {} + all_weight_adapters = [] + + for n, m in mp.model.named_modules(): + if hasattr(m, "weight_function"): + if m.weight is not None: + adapter, params = _create_weight_adapter( + m, n, existing_weights, algorithm, lora_dtype, rank + ) + lora_sd.update(params) + key = f"{n}.weight" + mp.add_weight_wrapper(key, adapter) + all_weight_adapters.append(adapter) + + if hasattr(m, "bias") and m.bias is not None: + bias_adapter, bias_params = _create_bias_adapter(m, n, lora_dtype) + lora_sd.update(bias_params) + key = f"{n}.bias" + mp.add_weight_wrapper(key, bias_adapter) + all_weight_adapters.append(bias_adapter) + + return lora_sd, all_weight_adapters + + +def _create_optimizer(optimizer_name, parameters, learning_rate): + """Create optimizer based on name. + + Args: + optimizer_name: Name of optimizer ("Adam", "AdamW", "SGD", "RMSprop") + parameters: Parameters to optimize + learning_rate: Learning rate + + Returns: + Optimizer instance + """ + if optimizer_name == "Adam": + return torch.optim.Adam(parameters, lr=learning_rate) + elif optimizer_name == "AdamW": + return torch.optim.AdamW(parameters, lr=learning_rate) + elif optimizer_name == "SGD": + return torch.optim.SGD(parameters, lr=learning_rate) + elif optimizer_name == "RMSprop": + return torch.optim.RMSprop(parameters, lr=learning_rate) + + +def _create_loss_function(loss_function_name): + """Create loss function based on name. + + Args: + loss_function_name: Name of loss function ("MSE", "L1", "Huber", "SmoothL1") + + Returns: + Loss function instance + """ + if loss_function_name == "MSE": + return torch.nn.MSELoss() + elif loss_function_name == "L1": + return torch.nn.L1Loss() + elif loss_function_name == "Huber": + return torch.nn.HuberLoss() + elif loss_function_name == "SmoothL1": + return torch.nn.SmoothL1Loss() + + +def _run_training_loop( + guider, train_sampler, latents, num_images, seed, bucket_mode, multi_res +): + """Execute the training loop. + + Args: + guider: The guider object + train_sampler: The training sampler + latents: Latent tensors + num_images: Number of images + seed: Random seed + bucket_mode: Whether bucket mode is enabled + multi_res: Whether multi-resolution mode is enabled + """ + sigmas = torch.tensor(range(num_images)) + noise = comfy_extras.nodes_custom_sampler.Noise_RandomNoise(seed) + + if bucket_mode: + # Use first bucket's first latent as dummy for guider + dummy_latent = latents[0][:1].repeat(num_images, 1, 1, 1) + guider.sample( + noise.generate_noise({"samples": dummy_latent}), + dummy_latent, + train_sampler, + sigmas, + seed=noise.seed, + ) + elif multi_res: + # use first latent as dummy latent if multi_res + latents = latents[0].repeat(num_images, 1, 1, 1) + guider.sample( + noise.generate_noise({"samples": latents}), + latents, + train_sampler, + sigmas, + seed=noise.seed, + ) + else: + guider.sample( + noise.generate_noise({"samples": latents}), + latents, + train_sampler, + sigmas, + seed=noise.seed, + ) + + class TrainLoraNode(io.ComfyNode): @classmethod def define_schema(cls): @@ -385,6 +879,11 @@ class TrainLoraNode(io.ComfyNode): default="[None]", tooltip="The existing LoRA to append to. Set to None for new LoRA.", ), + io.Boolean.Input( + "bucket_mode", + default=False, + tooltip="Enable resolution bucket mode. When enabled, expects pre-bucketed latents from ResolutionBucket node.", + ), ], outputs=[ io.Model.Output( @@ -419,6 +918,7 @@ class TrainLoraNode(io.ComfyNode): algorithm, gradient_checkpointing, existing_lora, + bucket_mode, ): # Extract scalars from lists (due to is_input_list=True) model = model[0] @@ -427,215 +927,125 @@ class TrainLoraNode(io.ComfyNode): grad_accumulation_steps = grad_accumulation_steps[0] learning_rate = learning_rate[0] rank = rank[0] - optimizer = optimizer[0] - loss_function = loss_function[0] + optimizer_name = optimizer[0] + loss_function_name = loss_function[0] seed = seed[0] training_dtype = training_dtype[0] lora_dtype = lora_dtype[0] algorithm = algorithm[0] gradient_checkpointing = gradient_checkpointing[0] existing_lora = existing_lora[0] + bucket_mode = bucket_mode[0] - # Handle latents - either single dict or list of dicts - if len(latents) == 1: - latents = latents[0]["samples"] # Single latent dict + # Process latents based on mode + if bucket_mode: + latents = _process_latents_bucket_mode(latents) else: - latent_list = [] - for latent in latents: - latent = latent["samples"] - bs = latent.shape[0] - if bs != 1: - for sub_latent in latent: - latent_list.append(sub_latent[None]) - else: - latent_list.append(latent) - latents = latent_list + latents = _process_latents_standard_mode(latents) - # Handle conditioning - either single list or list of lists - if len(positive) == 1: - positive = positive[0] # Single conditioning list - else: - # Multiple conditioning lists - flatten - flat_positive = [] - for cond in positive: - if isinstance(cond, list): - flat_positive.extend(cond) - else: - flat_positive.append(cond) - positive = flat_positive + # Process conditioning + positive = _process_conditioning(positive) + # Setup model and dtype mp = model.clone() dtype = node_helpers.string_to_torch_dtype(training_dtype) lora_dtype = node_helpers.string_to_torch_dtype(lora_dtype) mp.set_model_compute_dtype(dtype) - # latents here can be list of different size latent or one large batch - if isinstance(latents, list): - all_shapes = set() - latents = [t.to(dtype) for t in latents] - for latent in latents: - all_shapes.add(latent.shape) - logging.info(f"Latent shapes: {all_shapes}") - if len(all_shapes) > 1: - multi_res = True - else: - multi_res = False - latents = torch.cat(latents, dim=0) - num_images = len(latents) - elif isinstance(latents, torch.Tensor): - latents = latents.to(dtype) - num_images = latents.shape[0] - else: - logging.error(f"Invalid latents type: {type(latents)}") + # Prepare latents and compute counts + latents, num_images, multi_res = _prepare_latents_and_count( + latents, dtype, bucket_mode + ) - logging.info(f"Total Images: {num_images}, Total Captions: {len(positive)}") - if len(positive) == 1 and num_images > 1: - positive = positive * num_images - elif len(positive) != num_images: - raise ValueError( - f"Number of positive conditions ({len(positive)}) does not match number of images ({num_images})." - ) + # Validate and expand conditioning + positive = _validate_and_expand_conditioning(positive, num_images, bucket_mode) with torch.inference_mode(False): - lora_sd = {} - generator = torch.Generator() - generator.manual_seed(seed) + # Setup models for training + mp.model.requires_grad_(False) # Load existing LoRA weights if provided - existing_weights = {} - existing_steps = 0 - if existing_lora != "[None]": - lora_path = folder_paths.get_full_path_or_raise("loras", existing_lora) - # Extract steps from filename like "trained_lora_10_steps_20250225_203716" - existing_steps = int(existing_lora.split("_steps_")[0].split("_")[-1]) - if lora_path: - existing_weights = comfy.utils.load_torch_file(lora_path) + existing_weights, existing_steps = _load_existing_lora(existing_lora) - all_weight_adapters = [] - for n, m in mp.model.named_modules(): - if hasattr(m, "weight_function"): - if m.weight is not None: - key = "{}.weight".format(n) - shape = m.weight.shape - if len(shape) >= 2: - alpha = float(existing_weights.get(f"{key}.alpha", 1.0)) - dora_scale = existing_weights.get(f"{key}.dora_scale", None) - for adapter_cls in adapters: - existing_adapter = adapter_cls.load( - n, existing_weights, alpha, dora_scale - ) - if existing_adapter is not None: - break - else: - existing_adapter = None - adapter_cls = adapter_maps[algorithm] + # Setup LoRA adapters + lora_sd, all_weight_adapters = _setup_lora_adapters( + mp, existing_weights, algorithm, lora_dtype, rank + ) - if existing_adapter is not None: - train_adapter = existing_adapter.to_train().to( - lora_dtype - ) - else: - # Use LoRA with alpha=1.0 by default - train_adapter = adapter_cls.create_train( - m.weight, rank=rank, alpha=1.0 - ).to(lora_dtype) - for name, parameter in train_adapter.named_parameters(): - lora_sd[f"{n}.{name}"] = parameter + # Create optimizer and loss function + optimizer = _create_optimizer( + optimizer_name, lora_sd.values(), learning_rate + ) + criterion = _create_loss_function(loss_function_name) - mp.add_weight_wrapper(key, train_adapter) - all_weight_adapters.append(train_adapter) - else: - diff = torch.nn.Parameter( - torch.zeros( - m.weight.shape, dtype=lora_dtype, requires_grad=True - ) - ) - diff_module = BiasDiff(diff) - mp.add_weight_wrapper(key, BiasDiff(diff)) - all_weight_adapters.append(diff_module) - lora_sd["{}.diff".format(n)] = diff - if hasattr(m, "bias") and m.bias is not None: - key = "{}.bias".format(n) - bias = torch.nn.Parameter( - torch.zeros( - m.bias.shape, dtype=lora_dtype, requires_grad=True - ) - ) - bias_module = BiasDiff(bias) - lora_sd["{}.diff_b".format(n)] = bias - mp.add_weight_wrapper(key, BiasDiff(bias)) - all_weight_adapters.append(bias_module) - - if optimizer == "Adam": - optimizer = torch.optim.Adam(lora_sd.values(), lr=learning_rate) - elif optimizer == "AdamW": - optimizer = torch.optim.AdamW(lora_sd.values(), lr=learning_rate) - elif optimizer == "SGD": - optimizer = torch.optim.SGD(lora_sd.values(), lr=learning_rate) - elif optimizer == "RMSprop": - optimizer = torch.optim.RMSprop(lora_sd.values(), lr=learning_rate) - - # Setup loss function based on selection - if loss_function == "MSE": - criterion = torch.nn.MSELoss() - elif loss_function == "L1": - criterion = torch.nn.L1Loss() - elif loss_function == "Huber": - criterion = torch.nn.HuberLoss() - elif loss_function == "SmoothL1": - criterion = torch.nn.SmoothL1Loss() - - # setup models + # Setup gradient checkpointing if gradient_checkpointing: for m in find_all_highest_child_module_with_forward( mp.model.diffusion_model ): patch(m) - mp.model.requires_grad_(False) + + torch.cuda.empty_cache() + # With force_full_load=False we should be able to have offloading + # But for offloading in training we need custom AutoGrad hooks for fwd/bwd comfy.model_management.load_models_gpu( [mp], memory_required=1e20, force_full_load=True ) + torch.cuda.empty_cache() - # Setup sampler and guider like in test script + # Setup loss tracking loss_map = {"loss": []} def loss_callback(loss): loss_map["loss"].append(loss) - train_sampler = TrainSampler( - criterion, - optimizer, - loss_callback=loss_callback, - batch_size=batch_size, - grad_acc=grad_accumulation_steps, - total_steps=steps * grad_accumulation_steps, - seed=seed, - training_dtype=dtype, - real_dataset=latents if multi_res else None, - ) - guider = comfy_extras.nodes_custom_sampler.Guider_Basic(mp) - guider.set_conds(positive) # Set conditioning from input + # Create sampler + if bucket_mode: + train_sampler = TrainSampler( + criterion, + optimizer, + loss_callback=loss_callback, + batch_size=batch_size, + grad_acc=grad_accumulation_steps, + total_steps=steps * grad_accumulation_steps, + seed=seed, + training_dtype=dtype, + bucket_latents=latents, + ) + else: + train_sampler = TrainSampler( + criterion, + optimizer, + loss_callback=loss_callback, + batch_size=batch_size, + grad_acc=grad_accumulation_steps, + total_steps=steps * grad_accumulation_steps, + seed=seed, + training_dtype=dtype, + real_dataset=latents if multi_res else None, + ) - # Training loop + # Setup guider + guider = TrainGuider(mp) + guider.set_conds(positive) + + # Run training loop try: - # Generate dummy sigmas and noise - sigmas = torch.tensor(range(num_images)) - noise = comfy_extras.nodes_custom_sampler.Noise_RandomNoise(seed) - if multi_res: - # use first latent as dummy latent if multi_res - latents = latents[0].repeat((num_images,) + ((1,) * (latents[0].ndim - 1))) - guider.sample( - noise.generate_noise({"samples": latents}), - latents, + _run_training_loop( + guider, train_sampler, - sigmas, - seed=noise.seed, + latents, + num_images, + seed, + bucket_mode, + multi_res, ) finally: for m in mp.model.modules(): unpatch(m) del train_sampler, optimizer + # Finalize adapters for adapter in all_weight_adapters: adapter.requires_grad_(False) @@ -645,7 +1055,7 @@ class TrainLoraNode(io.ComfyNode): return io.NodeOutput(mp, lora_sd, loss_map, steps + existing_steps) -class LoraModelLoader(io.ComfyNode): +class LoraModelLoader(io.ComfyNode):# @classmethod def define_schema(cls): return io.Schema( From bf7dc63bd6acdedca67598856e05080d90eeec90 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 17 Dec 2025 20:29:32 -0800 Subject: [PATCH 310/647] skip_load_model -> force_full_load (#11390) This should be a bit more clear and less prone to potential breakage if the logic of the load models changes a bit. --- comfy/sampler_helpers.py | 9 ++++----- comfy_extras/nodes_train.py | 2 +- 2 files changed, 5 insertions(+), 6 deletions(-) diff --git a/comfy/sampler_helpers.py b/comfy/sampler_helpers.py index e158e8a84..9134e6d71 100644 --- a/comfy/sampler_helpers.py +++ b/comfy/sampler_helpers.py @@ -122,21 +122,20 @@ def estimate_memory(model, noise_shape, conds): minimum_memory_required = model.model.memory_required([noise_shape[0]] + list(noise_shape[1:]), cond_shapes=cond_shapes_min) return memory_required, minimum_memory_required -def prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None, skip_load_model=False): +def prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None, force_full_load=False): executor = comfy.patcher_extension.WrapperExecutor.new_executor( _prepare_sampling, comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.PREPARE_SAMPLING, model_options, is_model_options=True) ) - return executor.execute(model, noise_shape, conds, model_options=model_options, skip_load_model=skip_load_model) + return executor.execute(model, noise_shape, conds, model_options=model_options, force_full_load=force_full_load) -def _prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None, skip_load_model=False): +def _prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None, force_full_load=False): real_model: BaseModel = None models, inference_memory = get_additional_models(conds, model.model_dtype()) models += get_additional_models_from_model_options(model_options) models += model.get_nested_additional_models() # TODO: does this require inference_memory update? memory_required, minimum_memory_required = estimate_memory(model, noise_shape, conds) - models_list = [model] if not skip_load_model else [] - comfy.model_management.load_models_gpu(models_list + models, memory_required=memory_required + inference_memory, minimum_memory_required=minimum_memory_required + inference_memory) + comfy.model_management.load_models_gpu([model] + models, memory_required=memory_required + inference_memory, minimum_memory_required=minimum_memory_required + inference_memory, force_full_load=force_full_load) real_model = model.model return real_model, conds, models diff --git a/comfy_extras/nodes_train.py b/comfy_extras/nodes_train.py index 88bc8c8e8..364804205 100644 --- a/comfy_extras/nodes_train.py +++ b/comfy_extras/nodes_train.py @@ -44,7 +44,7 @@ class TrainGuider(comfy_extras.nodes_custom_sampler.Guider_Basic): noise.shape, self.conds, self.model_options, - skip_load_model=True, # skip load model as we manage it in TrainLoraNode.execute() + force_full_load=True, # mirror behavior in TrainLoraNode.execute() to keep model loaded ) ) device = self.model_patcher.load_device From 1ca89b810e921efce95fb4d254a8c6c93180450b Mon Sep 17 00:00:00 2001 From: ric-yu Date: Wed, 17 Dec 2025 21:44:31 -0800 Subject: [PATCH 311/647] Add unified jobs API with /api/jobs endpoints (#11054) * feat: create a /jobs api to return queue and history jobs * update unused vars * include priority * create jobs helper file * fix ruff * update how we set error message * include execution error in both responses * rename error -> failed, fix output shape * re-use queue and history functions * set workflow id * allow srot by exec duration * fix tests * send priority and remove error msg * use ws messages to get start and end times * revert main.py fully * refactor: move all /jobs business logic to jobs.py * fix failing test * remove some tests * fix non dict nodes * address comments * filter by workflow id and remove null fields * add clearer typing - remove get("..") or .. * refactor query params to top get_job(s) doc, add remove_sensitive_from_queue * add brief comment explaining why we skip animated * comment that format field is for frontend backward compatibility * fix whitespace --------- Co-authored-by: Jedrzej Kosinski Co-authored-by: guill --- comfy_execution/jobs.py | 291 ++++++++++++++++++++++++ server.py | 135 ++++++++++- tests/execution/test_execution.py | 134 +++++++++++ tests/execution/test_jobs.py | 361 ++++++++++++++++++++++++++++++ 4 files changed, 918 insertions(+), 3 deletions(-) create mode 100644 comfy_execution/jobs.py create mode 100644 tests/execution/test_jobs.py diff --git a/comfy_execution/jobs.py b/comfy_execution/jobs.py new file mode 100644 index 000000000..59fb49357 --- /dev/null +++ b/comfy_execution/jobs.py @@ -0,0 +1,291 @@ +""" +Job utilities for the /api/jobs endpoint. +Provides normalization and helper functions for job status tracking. +""" + +from typing import Optional + +from comfy_api.internal import prune_dict + + +class JobStatus: + """Job status constants.""" + PENDING = 'pending' + IN_PROGRESS = 'in_progress' + COMPLETED = 'completed' + FAILED = 'failed' + + ALL = [PENDING, IN_PROGRESS, COMPLETED, FAILED] + + +# Media types that can be previewed in the frontend +PREVIEWABLE_MEDIA_TYPES = frozenset({'images', 'video', 'audio'}) + +# 3D file extensions for preview fallback (no dedicated media_type exists) +THREE_D_EXTENSIONS = frozenset({'.obj', '.fbx', '.gltf', '.glb'}) + + +def _extract_job_metadata(extra_data: dict) -> tuple[Optional[int], Optional[str]]: + """Extract create_time and workflow_id from extra_data. + + Returns: + tuple: (create_time, workflow_id) + """ + create_time = extra_data.get('create_time') + extra_pnginfo = extra_data.get('extra_pnginfo', {}) + workflow_id = extra_pnginfo.get('workflow', {}).get('id') + return create_time, workflow_id + + +def is_previewable(media_type: str, item: dict) -> bool: + """ + Check if an output item is previewable. + Matches frontend logic in ComfyUI_frontend/src/stores/queueStore.ts + Maintains backwards compatibility with existing logic. + + Priority: + 1. media_type is 'images', 'video', or 'audio' + 2. format field starts with 'video/' or 'audio/' + 3. filename has a 3D extension (.obj, .fbx, .gltf, .glb) + """ + if media_type in PREVIEWABLE_MEDIA_TYPES: + return True + + # Check format field (MIME type). + # Maintains backwards compatibility with how custom node outputs are handled in the frontend. + fmt = item.get('format', '') + if fmt and (fmt.startswith('video/') or fmt.startswith('audio/')): + return True + + # Check for 3D files by extension + filename = item.get('filename', '').lower() + if any(filename.endswith(ext) for ext in THREE_D_EXTENSIONS): + return True + + return False + + +def normalize_queue_item(item: tuple, status: str) -> dict: + """Convert queue item tuple to unified job dict. + + Expects item with sensitive data already removed (5 elements). + """ + priority, prompt_id, _, extra_data, _ = item + create_time, workflow_id = _extract_job_metadata(extra_data) + + return prune_dict({ + 'id': prompt_id, + 'status': status, + 'priority': priority, + 'create_time': create_time, + 'outputs_count': 0, + 'workflow_id': workflow_id, + }) + + +def normalize_history_item(prompt_id: str, history_item: dict, include_outputs: bool = False) -> dict: + """Convert history item dict to unified job dict. + + History items have sensitive data already removed (prompt tuple has 5 elements). + """ + prompt_tuple = history_item['prompt'] + priority, _, prompt, extra_data, _ = prompt_tuple + create_time, workflow_id = _extract_job_metadata(extra_data) + + status_info = history_item.get('status', {}) + status_str = status_info.get('status_str') if status_info else None + if status_str == 'success': + status = JobStatus.COMPLETED + elif status_str == 'error': + status = JobStatus.FAILED + else: + status = JobStatus.COMPLETED + + outputs = history_item.get('outputs', {}) + outputs_count, preview_output = get_outputs_summary(outputs) + + execution_error = None + execution_start_time = None + execution_end_time = None + if status_info: + messages = status_info.get('messages', []) + for entry in messages: + if isinstance(entry, (list, tuple)) and len(entry) >= 2: + event_name, event_data = entry[0], entry[1] + if isinstance(event_data, dict): + if event_name == 'execution_start': + execution_start_time = event_data.get('timestamp') + elif event_name in ('execution_success', 'execution_error', 'execution_interrupted'): + execution_end_time = event_data.get('timestamp') + if event_name == 'execution_error': + execution_error = event_data + + job = prune_dict({ + 'id': prompt_id, + 'status': status, + 'priority': priority, + 'create_time': create_time, + 'execution_start_time': execution_start_time, + 'execution_end_time': execution_end_time, + 'execution_error': execution_error, + 'outputs_count': outputs_count, + 'preview_output': preview_output, + 'workflow_id': workflow_id, + }) + + if include_outputs: + job['outputs'] = outputs + job['execution_status'] = status_info + job['workflow'] = { + 'prompt': prompt, + 'extra_data': extra_data, + } + + return job + + +def get_outputs_summary(outputs: dict) -> tuple[int, Optional[dict]]: + """ + Count outputs and find preview in a single pass. + Returns (outputs_count, preview_output). + + Preview priority (matching frontend): + 1. type="output" with previewable media + 2. Any previewable media + """ + count = 0 + preview_output = None + fallback_preview = None + + for node_id, node_outputs in outputs.items(): + if not isinstance(node_outputs, dict): + continue + for media_type, items in node_outputs.items(): + # 'animated' is a boolean flag, not actual output items + if media_type == 'animated' or not isinstance(items, list): + continue + + for item in items: + if not isinstance(item, dict): + continue + count += 1 + + if preview_output is None and is_previewable(media_type, item): + enriched = { + **item, + 'nodeId': node_id, + 'mediaType': media_type + } + if item.get('type') == 'output': + preview_output = enriched + elif fallback_preview is None: + fallback_preview = enriched + + return count, preview_output or fallback_preview + + +def apply_sorting(jobs: list[dict], sort_by: str, sort_order: str) -> list[dict]: + """Sort jobs list by specified field and order.""" + reverse = (sort_order == 'desc') + + if sort_by == 'execution_duration': + def get_sort_key(job): + start = job.get('execution_start_time', 0) + end = job.get('execution_end_time', 0) + return end - start if end and start else 0 + else: + def get_sort_key(job): + return job.get('create_time', 0) + + return sorted(jobs, key=get_sort_key, reverse=reverse) + + +def get_job(prompt_id: str, running: list, queued: list, history: dict) -> Optional[dict]: + """ + Get a single job by prompt_id from history or queue. + + Args: + prompt_id: The prompt ID to look up + running: List of currently running queue items + queued: List of pending queue items + history: Dict of history items keyed by prompt_id + + Returns: + Job dict with full details, or None if not found + """ + if prompt_id in history: + return normalize_history_item(prompt_id, history[prompt_id], include_outputs=True) + + for item in running: + if item[1] == prompt_id: + return normalize_queue_item(item, JobStatus.IN_PROGRESS) + + for item in queued: + if item[1] == prompt_id: + return normalize_queue_item(item, JobStatus.PENDING) + + return None + + +def get_all_jobs( + running: list, + queued: list, + history: dict, + status_filter: Optional[list[str]] = None, + workflow_id: Optional[str] = None, + sort_by: str = "created_at", + sort_order: str = "desc", + limit: Optional[int] = None, + offset: int = 0 +) -> tuple[list[dict], int]: + """ + Get all jobs (running, pending, completed) with filtering and sorting. + + Args: + running: List of currently running queue items + queued: List of pending queue items + history: Dict of history items keyed by prompt_id + status_filter: List of statuses to include (from JobStatus.ALL) + workflow_id: Filter by workflow ID + sort_by: Field to sort by ('created_at', 'execution_duration') + sort_order: 'asc' or 'desc' + limit: Maximum number of items to return + offset: Number of items to skip + + Returns: + tuple: (jobs_list, total_count) + """ + jobs = [] + + if status_filter is None: + status_filter = JobStatus.ALL + + if JobStatus.IN_PROGRESS in status_filter: + for item in running: + jobs.append(normalize_queue_item(item, JobStatus.IN_PROGRESS)) + + if JobStatus.PENDING in status_filter: + for item in queued: + jobs.append(normalize_queue_item(item, JobStatus.PENDING)) + + include_completed = JobStatus.COMPLETED in status_filter + include_failed = JobStatus.FAILED in status_filter + if include_completed or include_failed: + for prompt_id, history_item in history.items(): + is_failed = history_item.get('status', {}).get('status_str') == 'error' + if (is_failed and include_failed) or (not is_failed and include_completed): + jobs.append(normalize_history_item(prompt_id, history_item)) + + if workflow_id: + jobs = [j for j in jobs if j.get('workflow_id') == workflow_id] + + jobs = apply_sorting(jobs, sort_by, sort_order) + + total_count = len(jobs) + + if offset > 0: + jobs = jobs[offset:] + if limit is not None: + jobs = jobs[:limit] + + return (jobs, total_count) diff --git a/server.py b/server.py index ac4f42222..c27f8be7d 100644 --- a/server.py +++ b/server.py @@ -7,6 +7,7 @@ import time import nodes import folder_paths import execution +from comfy_execution.jobs import JobStatus, get_job, get_all_jobs import uuid import urllib import json @@ -47,6 +48,12 @@ from middleware.cache_middleware import cache_control if args.enable_manager: import comfyui_manager + +def _remove_sensitive_from_queue(queue: list) -> list: + """Remove sensitive data (index 5) from queue item tuples.""" + return [item[:5] for item in queue] + + async def send_socket_catch_exception(function, message): try: await function(message) @@ -694,6 +701,129 @@ class PromptServer(): out[node_class] = node_info(node_class) return web.json_response(out) + @routes.get("/api/jobs") + async def get_jobs(request): + """List all jobs with filtering, sorting, and pagination. + + Query parameters: + status: Filter by status (comma-separated): pending, in_progress, completed, failed + workflow_id: Filter by workflow ID + sort_by: Sort field: created_at (default), execution_duration + sort_order: Sort direction: asc, desc (default) + limit: Max items to return (positive integer) + offset: Items to skip (non-negative integer, default 0) + """ + query = request.rel_url.query + + status_param = query.get('status') + workflow_id = query.get('workflow_id') + sort_by = query.get('sort_by', 'created_at').lower() + sort_order = query.get('sort_order', 'desc').lower() + + status_filter = None + if status_param: + status_filter = [s.strip().lower() for s in status_param.split(',') if s.strip()] + invalid_statuses = [s for s in status_filter if s not in JobStatus.ALL] + if invalid_statuses: + return web.json_response( + {"error": f"Invalid status value(s): {', '.join(invalid_statuses)}. Valid values: {', '.join(JobStatus.ALL)}"}, + status=400 + ) + + if sort_by not in {'created_at', 'execution_duration'}: + return web.json_response( + {"error": "sort_by must be 'created_at' or 'execution_duration'"}, + status=400 + ) + + if sort_order not in {'asc', 'desc'}: + return web.json_response( + {"error": "sort_order must be 'asc' or 'desc'"}, + status=400 + ) + + limit = None + + # If limit is provided, validate that it is a positive integer, else continue without a limit + if 'limit' in query: + try: + limit = int(query.get('limit')) + if limit <= 0: + return web.json_response( + {"error": "limit must be a positive integer"}, + status=400 + ) + except (ValueError, TypeError): + return web.json_response( + {"error": "limit must be an integer"}, + status=400 + ) + + offset = 0 + if 'offset' in query: + try: + offset = int(query.get('offset')) + if offset < 0: + offset = 0 + except (ValueError, TypeError): + return web.json_response( + {"error": "offset must be an integer"}, + status=400 + ) + + running, queued = self.prompt_queue.get_current_queue_volatile() + history = self.prompt_queue.get_history() + + running = _remove_sensitive_from_queue(running) + queued = _remove_sensitive_from_queue(queued) + + jobs, total = get_all_jobs( + running, queued, history, + status_filter=status_filter, + workflow_id=workflow_id, + sort_by=sort_by, + sort_order=sort_order, + limit=limit, + offset=offset + ) + + has_more = (offset + len(jobs)) < total + + return web.json_response({ + 'jobs': jobs, + 'pagination': { + 'offset': offset, + 'limit': limit, + 'total': total, + 'has_more': has_more + } + }) + + @routes.get("/api/jobs/{job_id}") + async def get_job_by_id(request): + """Get a single job by ID.""" + job_id = request.match_info.get("job_id", None) + if not job_id: + return web.json_response( + {"error": "job_id is required"}, + status=400 + ) + + running, queued = self.prompt_queue.get_current_queue_volatile() + history = self.prompt_queue.get_history(prompt_id=job_id) + + running = _remove_sensitive_from_queue(running) + queued = _remove_sensitive_from_queue(queued) + + job = get_job(job_id, running, queued, history) + if job is None: + return web.json_response( + {"error": "Job not found"}, + status=404 + ) + + return web.json_response(job) + @routes.get("/history") async def get_history(request): max_items = request.rel_url.query.get("max_items", None) @@ -717,9 +847,8 @@ class PromptServer(): async def get_queue(request): queue_info = {} current_queue = self.prompt_queue.get_current_queue_volatile() - remove_sensitive = lambda queue: [x[:5] for x in queue] - queue_info['queue_running'] = remove_sensitive(current_queue[0]) - queue_info['queue_pending'] = remove_sensitive(current_queue[1]) + queue_info['queue_running'] = _remove_sensitive_from_queue(current_queue[0]) + queue_info['queue_pending'] = _remove_sensitive_from_queue(current_queue[1]) return web.json_response(queue_info) @routes.post("/prompt") diff --git a/tests/execution/test_execution.py b/tests/execution/test_execution.py index ace0d2279..f73ca7e3c 100644 --- a/tests/execution/test_execution.py +++ b/tests/execution/test_execution.py @@ -99,6 +99,37 @@ class ComfyClient: with urllib.request.urlopen(url) as response: return json.loads(response.read()) + def get_jobs(self, status=None, limit=None, offset=None, sort_by=None, sort_order=None): + url = "http://{}/api/jobs".format(self.server_address) + params = {} + if status is not None: + params["status"] = status + if limit is not None: + params["limit"] = limit + if offset is not None: + params["offset"] = offset + if sort_by is not None: + params["sort_by"] = sort_by + if sort_order is not None: + params["sort_order"] = sort_order + + if params: + url_values = urllib.parse.urlencode(params) + url = "{}?{}".format(url, url_values) + + with urllib.request.urlopen(url) as response: + return json.loads(response.read()) + + def get_job(self, job_id): + url = "http://{}/api/jobs/{}".format(self.server_address, job_id) + try: + with urllib.request.urlopen(url) as response: + return json.loads(response.read()) + except urllib.error.HTTPError as e: + if e.code == 404: + return None + raise + def set_test_name(self, name): self.test_name = name @@ -877,3 +908,106 @@ class TestExecution: result = client.get_all_history(max_items=5, offset=len(all_history) - 1) assert len(result) <= 1, "Should return at most 1 item when offset is near end" + + # Jobs API tests + def test_jobs_api_job_structure( + self, client: ComfyClient, builder: GraphBuilder + ): + """Test that job objects have required fields""" + self._create_history_item(client, builder) + + jobs_response = client.get_jobs(status="completed", limit=1) + assert len(jobs_response["jobs"]) > 0, "Should have at least one job" + + job = jobs_response["jobs"][0] + assert "id" in job, "Job should have id" + assert "status" in job, "Job should have status" + assert "create_time" in job, "Job should have create_time" + assert "outputs_count" in job, "Job should have outputs_count" + assert "preview_output" in job, "Job should have preview_output" + + def test_jobs_api_preview_output_structure( + self, client: ComfyClient, builder: GraphBuilder + ): + """Test that preview_output has correct structure""" + self._create_history_item(client, builder) + + jobs_response = client.get_jobs(status="completed", limit=1) + job = jobs_response["jobs"][0] + + if job["preview_output"] is not None: + preview = job["preview_output"] + assert "filename" in preview, "Preview should have filename" + assert "nodeId" in preview, "Preview should have nodeId" + assert "mediaType" in preview, "Preview should have mediaType" + + def test_jobs_api_pagination( + self, client: ComfyClient, builder: GraphBuilder + ): + """Test jobs API pagination""" + for _ in range(5): + self._create_history_item(client, builder) + + first_page = client.get_jobs(limit=2, offset=0) + second_page = client.get_jobs(limit=2, offset=2) + + assert len(first_page["jobs"]) <= 2, "First page should have at most 2 jobs" + assert len(second_page["jobs"]) <= 2, "Second page should have at most 2 jobs" + + first_ids = {j["id"] for j in first_page["jobs"]} + second_ids = {j["id"] for j in second_page["jobs"]} + assert first_ids.isdisjoint(second_ids), "Pages should have different jobs" + + def test_jobs_api_sorting( + self, client: ComfyClient, builder: GraphBuilder + ): + """Test jobs API sorting""" + for _ in range(3): + self._create_history_item(client, builder) + + desc_jobs = client.get_jobs(sort_order="desc") + asc_jobs = client.get_jobs(sort_order="asc") + + if len(desc_jobs["jobs"]) >= 2: + desc_times = [j["create_time"] for j in desc_jobs["jobs"] if j["create_time"]] + asc_times = [j["create_time"] for j in asc_jobs["jobs"] if j["create_time"]] + if len(desc_times) >= 2: + assert desc_times == sorted(desc_times, reverse=True), "Desc should be newest first" + if len(asc_times) >= 2: + assert asc_times == sorted(asc_times), "Asc should be oldest first" + + def test_jobs_api_status_filter( + self, client: ComfyClient, builder: GraphBuilder + ): + """Test jobs API status filtering""" + self._create_history_item(client, builder) + + completed_jobs = client.get_jobs(status="completed") + assert len(completed_jobs["jobs"]) > 0, "Should have completed jobs from history" + + for job in completed_jobs["jobs"]: + assert job["status"] == "completed", "Should only return completed jobs" + + # Pending jobs are transient - just verify filter doesn't error + pending_jobs = client.get_jobs(status="pending") + for job in pending_jobs["jobs"]: + assert job["status"] == "pending", "Should only return pending jobs" + + def test_get_job_by_id( + self, client: ComfyClient, builder: GraphBuilder + ): + """Test getting a single job by ID""" + result = self._create_history_item(client, builder) + prompt_id = result.get_prompt_id() + + job = client.get_job(prompt_id) + assert job is not None, "Should find the job" + assert job["id"] == prompt_id, "Job ID should match" + assert "outputs" in job, "Single job should include outputs" + + def test_get_job_not_found( + self, client: ComfyClient, builder: GraphBuilder + ): + """Test getting a non-existent job returns 404""" + job = client.get_job("nonexistent-job-id") + assert job is None, "Non-existent job should return None" diff --git a/tests/execution/test_jobs.py b/tests/execution/test_jobs.py new file mode 100644 index 000000000..918c8080a --- /dev/null +++ b/tests/execution/test_jobs.py @@ -0,0 +1,361 @@ +"""Unit tests for comfy_execution/jobs.py""" + +from comfy_execution.jobs import ( + JobStatus, + is_previewable, + normalize_queue_item, + normalize_history_item, + get_outputs_summary, + apply_sorting, +) + + +class TestJobStatus: + """Test JobStatus constants.""" + + def test_status_values(self): + """Status constants should have expected string values.""" + assert JobStatus.PENDING == 'pending' + assert JobStatus.IN_PROGRESS == 'in_progress' + assert JobStatus.COMPLETED == 'completed' + assert JobStatus.FAILED == 'failed' + + def test_all_contains_all_statuses(self): + """ALL should contain all status values.""" + assert JobStatus.PENDING in JobStatus.ALL + assert JobStatus.IN_PROGRESS in JobStatus.ALL + assert JobStatus.COMPLETED in JobStatus.ALL + assert JobStatus.FAILED in JobStatus.ALL + assert len(JobStatus.ALL) == 4 + + +class TestIsPreviewable: + """Unit tests for is_previewable()""" + + def test_previewable_media_types(self): + """Images, video, audio media types should be previewable.""" + for media_type in ['images', 'video', 'audio']: + assert is_previewable(media_type, {}) is True + + def test_non_previewable_media_types(self): + """Other media types should not be previewable.""" + for media_type in ['latents', 'text', 'metadata', 'files']: + assert is_previewable(media_type, {}) is False + + def test_3d_extensions_previewable(self): + """3D file extensions should be previewable regardless of media_type.""" + for ext in ['.obj', '.fbx', '.gltf', '.glb']: + item = {'filename': f'model{ext}'} + assert is_previewable('files', item) is True + + def test_3d_extensions_case_insensitive(self): + """3D extension check should be case insensitive.""" + item = {'filename': 'MODEL.GLB'} + assert is_previewable('files', item) is True + + def test_video_format_previewable(self): + """Items with video/ format should be previewable.""" + item = {'format': 'video/mp4'} + assert is_previewable('files', item) is True + + def test_audio_format_previewable(self): + """Items with audio/ format should be previewable.""" + item = {'format': 'audio/wav'} + assert is_previewable('files', item) is True + + def test_other_format_not_previewable(self): + """Items with other format should not be previewable.""" + item = {'format': 'application/json'} + assert is_previewable('files', item) is False + + +class TestGetOutputsSummary: + """Unit tests for get_outputs_summary()""" + + def test_empty_outputs(self): + """Empty outputs should return 0 count and None preview.""" + count, preview = get_outputs_summary({}) + assert count == 0 + assert preview is None + + def test_counts_across_multiple_nodes(self): + """Outputs from multiple nodes should all be counted.""" + outputs = { + 'node1': {'images': [{'filename': 'a.png', 'type': 'output'}]}, + 'node2': {'images': [{'filename': 'b.png', 'type': 'output'}]}, + 'node3': {'images': [ + {'filename': 'c.png', 'type': 'output'}, + {'filename': 'd.png', 'type': 'output'} + ]} + } + count, preview = get_outputs_summary(outputs) + assert count == 4 + + def test_skips_animated_key_and_non_list_values(self): + """The 'animated' key and non-list values should be skipped.""" + outputs = { + 'node1': { + 'images': [{'filename': 'test.png', 'type': 'output'}], + 'animated': [True], # Should skip due to key name + 'metadata': 'string', # Should skip due to non-list + 'count': 42 # Should skip due to non-list + } + } + count, preview = get_outputs_summary(outputs) + assert count == 1 + + def test_preview_prefers_type_output(self): + """Items with type='output' should be preferred for preview.""" + outputs = { + 'node1': { + 'images': [ + {'filename': 'temp.png', 'type': 'temp'}, + {'filename': 'output.png', 'type': 'output'} + ] + } + } + count, preview = get_outputs_summary(outputs) + assert count == 2 + assert preview['filename'] == 'output.png' + + def test_preview_fallback_when_no_output_type(self): + """If no type='output', should use first previewable.""" + outputs = { + 'node1': { + 'images': [ + {'filename': 'temp1.png', 'type': 'temp'}, + {'filename': 'temp2.png', 'type': 'temp'} + ] + } + } + count, preview = get_outputs_summary(outputs) + assert preview['filename'] == 'temp1.png' + + def test_non_previewable_media_types_counted_but_no_preview(self): + """Non-previewable media types should be counted but not used as preview.""" + outputs = { + 'node1': { + 'latents': [ + {'filename': 'latent1.safetensors'}, + {'filename': 'latent2.safetensors'} + ] + } + } + count, preview = get_outputs_summary(outputs) + assert count == 2 + assert preview is None + + def test_previewable_media_types(self): + """Images, video, and audio media types should be previewable.""" + for media_type in ['images', 'video', 'audio']: + outputs = { + 'node1': { + media_type: [{'filename': 'test.file', 'type': 'output'}] + } + } + count, preview = get_outputs_summary(outputs) + assert preview is not None, f"{media_type} should be previewable" + + def test_3d_files_previewable(self): + """3D file extensions should be previewable.""" + for ext in ['.obj', '.fbx', '.gltf', '.glb']: + outputs = { + 'node1': { + 'files': [{'filename': f'model{ext}', 'type': 'output'}] + } + } + count, preview = get_outputs_summary(outputs) + assert preview is not None, f"3D file {ext} should be previewable" + + def test_format_mime_type_previewable(self): + """Files with video/ or audio/ format should be previewable.""" + for fmt in ['video/x-custom', 'audio/x-custom']: + outputs = { + 'node1': { + 'files': [{'filename': 'file.custom', 'format': fmt, 'type': 'output'}] + } + } + count, preview = get_outputs_summary(outputs) + assert preview is not None, f"Format {fmt} should be previewable" + + def test_preview_enriched_with_node_metadata(self): + """Preview should include nodeId, mediaType, and original fields.""" + outputs = { + 'node123': { + 'images': [{'filename': 'test.png', 'type': 'output', 'subfolder': 'outputs'}] + } + } + count, preview = get_outputs_summary(outputs) + assert preview['nodeId'] == 'node123' + assert preview['mediaType'] == 'images' + assert preview['subfolder'] == 'outputs' + + +class TestApplySorting: + """Unit tests for apply_sorting()""" + + def test_sort_by_create_time_desc(self): + """Default sort by create_time descending.""" + jobs = [ + {'id': 'a', 'create_time': 100}, + {'id': 'b', 'create_time': 300}, + {'id': 'c', 'create_time': 200}, + ] + result = apply_sorting(jobs, 'created_at', 'desc') + assert [j['id'] for j in result] == ['b', 'c', 'a'] + + def test_sort_by_create_time_asc(self): + """Sort by create_time ascending.""" + jobs = [ + {'id': 'a', 'create_time': 100}, + {'id': 'b', 'create_time': 300}, + {'id': 'c', 'create_time': 200}, + ] + result = apply_sorting(jobs, 'created_at', 'asc') + assert [j['id'] for j in result] == ['a', 'c', 'b'] + + def test_sort_by_execution_duration(self): + """Sort by execution_duration should order by duration.""" + jobs = [ + {'id': 'a', 'create_time': 100, 'execution_start_time': 100, 'execution_end_time': 5100}, # 5s + {'id': 'b', 'create_time': 300, 'execution_start_time': 300, 'execution_end_time': 1300}, # 1s + {'id': 'c', 'create_time': 200, 'execution_start_time': 200, 'execution_end_time': 3200}, # 3s + ] + result = apply_sorting(jobs, 'execution_duration', 'desc') + assert [j['id'] for j in result] == ['a', 'c', 'b'] + + def test_sort_with_none_values(self): + """Jobs with None values should sort as 0.""" + jobs = [ + {'id': 'a', 'create_time': 100, 'execution_start_time': 100, 'execution_end_time': 5100}, + {'id': 'b', 'create_time': 300, 'execution_start_time': None, 'execution_end_time': None}, + {'id': 'c', 'create_time': 200, 'execution_start_time': 200, 'execution_end_time': 3200}, + ] + result = apply_sorting(jobs, 'execution_duration', 'asc') + assert result[0]['id'] == 'b' # None treated as 0, comes first + + +class TestNormalizeQueueItem: + """Unit tests for normalize_queue_item()""" + + def test_basic_normalization(self): + """Queue item should be normalized to job dict.""" + item = ( + 10, # priority/number + 'prompt-123', # prompt_id + {'nodes': {}}, # prompt + { + 'create_time': 1234567890, + 'extra_pnginfo': {'workflow': {'id': 'workflow-abc'}} + }, # extra_data + ['node1'], # outputs_to_execute + ) + job = normalize_queue_item(item, JobStatus.PENDING) + + assert job['id'] == 'prompt-123' + assert job['status'] == 'pending' + assert job['priority'] == 10 + assert job['create_time'] == 1234567890 + assert 'execution_start_time' not in job + assert 'execution_end_time' not in job + assert 'execution_error' not in job + assert 'preview_output' not in job + assert job['outputs_count'] == 0 + assert job['workflow_id'] == 'workflow-abc' + + +class TestNormalizeHistoryItem: + """Unit tests for normalize_history_item()""" + + def test_completed_job(self): + """Completed history item should have correct status and times from messages.""" + history_item = { + 'prompt': ( + 5, # priority + 'prompt-456', + {'nodes': {}}, + { + 'create_time': 1234567890000, + 'extra_pnginfo': {'workflow': {'id': 'workflow-xyz'}} + }, + ['node1'], + ), + 'status': { + 'status_str': 'success', + 'completed': True, + 'messages': [ + ('execution_start', {'prompt_id': 'prompt-456', 'timestamp': 1234567890500}), + ('execution_success', {'prompt_id': 'prompt-456', 'timestamp': 1234567893000}), + ] + }, + 'outputs': {}, + } + job = normalize_history_item('prompt-456', history_item) + + assert job['id'] == 'prompt-456' + assert job['status'] == 'completed' + assert job['priority'] == 5 + assert job['execution_start_time'] == 1234567890500 + assert job['execution_end_time'] == 1234567893000 + assert job['workflow_id'] == 'workflow-xyz' + + def test_failed_job(self): + """Failed history item should have failed status and error from messages.""" + history_item = { + 'prompt': ( + 5, + 'prompt-789', + {'nodes': {}}, + {'create_time': 1234567890000}, + ['node1'], + ), + 'status': { + 'status_str': 'error', + 'completed': False, + 'messages': [ + ('execution_start', {'prompt_id': 'prompt-789', 'timestamp': 1234567890500}), + ('execution_error', { + 'prompt_id': 'prompt-789', + 'node_id': '5', + 'node_type': 'KSampler', + 'exception_message': 'CUDA out of memory', + 'exception_type': 'RuntimeError', + 'traceback': ['Traceback...', 'RuntimeError: CUDA out of memory'], + 'timestamp': 1234567891000, + }) + ] + }, + 'outputs': {}, + } + + job = normalize_history_item('prompt-789', history_item) + assert job['status'] == 'failed' + assert job['execution_start_time'] == 1234567890500 + assert job['execution_end_time'] == 1234567891000 + assert job['execution_error']['node_id'] == '5' + assert job['execution_error']['node_type'] == 'KSampler' + assert job['execution_error']['exception_message'] == 'CUDA out of memory' + + def test_include_outputs(self): + """When include_outputs=True, should include full output data.""" + history_item = { + 'prompt': ( + 5, + 'prompt-123', + {'nodes': {'1': {}}}, + {'create_time': 1234567890, 'client_id': 'abc'}, + ['node1'], + ), + 'status': {'status_str': 'success', 'completed': True, 'messages': []}, + 'outputs': {'node1': {'images': [{'filename': 'test.png'}]}}, + } + job = normalize_history_item('prompt-123', history_item, include_outputs=True) + + assert 'outputs' in job + assert 'workflow' in job + assert 'execution_status' in job + assert job['outputs'] == {'node1': {'images': [{'filename': 'test.png'}]}} + assert job['workflow'] == { + 'prompt': {'nodes': {'1': {}}}, + 'extra_data': {'create_time': 1234567890, 'client_id': 'abc'}, + } From e8ebbe668e82ab0f3c0842afa79d255329eb76ac Mon Sep 17 00:00:00 2001 From: ComfyUI Wiki Date: Fri, 19 Dec 2025 06:09:29 +0800 Subject: [PATCH 312/647] chore: update workflow templates to v0.7.60 (#11403) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 9b9e61683..54696395f 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,5 +1,5 @@ comfyui-frontend-package==1.34.9 -comfyui-workflow-templates==0.7.59 +comfyui-workflow-templates==0.7.60 comfyui-embedded-docs==0.3.1 torch torchsde From e4fb3a3572c94d8f2ef73ddd18d2a6966ed5a1e5 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Thu, 18 Dec 2025 14:45:33 -0800 Subject: [PATCH 313/647] Support loading Wan/Qwen VAEs with different in/out channels. (#11405) --- comfy/ldm/wan/vae.py | 11 +++++++---- comfy/sd.py | 3 ++- 2 files changed, 9 insertions(+), 5 deletions(-) diff --git a/comfy/ldm/wan/vae.py b/comfy/ldm/wan/vae.py index ccbb25822..08315f1a8 100644 --- a/comfy/ldm/wan/vae.py +++ b/comfy/ldm/wan/vae.py @@ -227,6 +227,7 @@ class Encoder3d(nn.Module): def __init__(self, dim=128, z_dim=4, + input_channels=3, dim_mult=[1, 2, 4, 4], num_res_blocks=2, attn_scales=[], @@ -245,7 +246,7 @@ class Encoder3d(nn.Module): scale = 1.0 # init block - self.conv1 = CausalConv3d(3, dims[0], 3, padding=1) + self.conv1 = CausalConv3d(input_channels, dims[0], 3, padding=1) # downsample blocks downsamples = [] @@ -331,6 +332,7 @@ class Decoder3d(nn.Module): def __init__(self, dim=128, z_dim=4, + output_channels=3, dim_mult=[1, 2, 4, 4], num_res_blocks=2, attn_scales=[], @@ -378,7 +380,7 @@ class Decoder3d(nn.Module): # output blocks self.head = nn.Sequential( RMS_norm(out_dim, images=False), nn.SiLU(), - CausalConv3d(out_dim, 3, 3, padding=1)) + CausalConv3d(out_dim, output_channels, 3, padding=1)) def forward(self, x, feat_cache=None, feat_idx=[0]): ## conv1 @@ -449,6 +451,7 @@ class WanVAE(nn.Module): num_res_blocks=2, attn_scales=[], temperal_downsample=[True, True, False], + image_channels=3, dropout=0.0): super().__init__() self.dim = dim @@ -460,11 +463,11 @@ class WanVAE(nn.Module): self.temperal_upsample = temperal_downsample[::-1] # modules - self.encoder = Encoder3d(dim, z_dim * 2, dim_mult, num_res_blocks, + self.encoder = Encoder3d(dim, z_dim * 2, image_channels, dim_mult, num_res_blocks, attn_scales, self.temperal_downsample, dropout) self.conv1 = CausalConv3d(z_dim * 2, z_dim * 2, 1) self.conv2 = CausalConv3d(z_dim, z_dim, 1) - self.decoder = Decoder3d(dim, z_dim, dim_mult, num_res_blocks, + self.decoder = Decoder3d(dim, z_dim, image_channels, dim_mult, num_res_blocks, attn_scales, self.temperal_upsample, dropout) def encode(self, x): diff --git a/comfy/sd.py b/comfy/sd.py index 1cad98aef..f95c78892 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -546,7 +546,8 @@ class VAE: self.downscale_index_formula = (4, 8, 8) self.latent_dim = 3 self.latent_channels = 16 - ddconfig = {"dim": dim, "z_dim": self.latent_channels, "dim_mult": [1, 2, 4, 4], "num_res_blocks": 2, "attn_scales": [], "temperal_downsample": [False, True, True], "dropout": 0.0} + self.output_channels = sd["encoder.conv1.weight"].shape[1] + ddconfig = {"dim": dim, "z_dim": self.latent_channels, "dim_mult": [1, 2, 4, 4], "num_res_blocks": 2, "attn_scales": [], "temperal_downsample": [False, True, True], "image_channels": self.output_channels, "dropout": 0.0} self.first_stage_model = comfy.ldm.wan.vae.WanVAE(**ddconfig) self.working_dtypes = [torch.bfloat16, torch.float16, torch.float32] self.memory_used_encode = lambda shape, dtype: (1500 if shape[2]<=4 else 6000) * shape[3] * shape[4] * model_management.dtype_size(dtype) From 6a2678ac65ff690e24771a4c64ce96f7a9824fa4 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Thu, 18 Dec 2025 15:22:38 -0800 Subject: [PATCH 314/647] Trim/pad channels in VAE code. (#11406) --- comfy/sd.py | 33 ++++++++++++++++++++++++--------- nodes.py | 4 ++-- 2 files changed, 26 insertions(+), 11 deletions(-) diff --git a/comfy/sd.py b/comfy/sd.py index f95c78892..c2a9728f3 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -321,6 +321,7 @@ class VAE: self.latent_channels = 4 self.latent_dim = 2 self.output_channels = 3 + self.pad_channel_value = None self.process_input = lambda image: image * 2.0 - 1.0 self.process_output = lambda image: torch.clamp((image + 1.0) / 2.0, min=0.0, max=1.0) self.working_dtypes = [torch.bfloat16, torch.float32] @@ -435,6 +436,7 @@ class VAE: self.memory_used_decode = lambda shape, dtype: (1000 * shape[2] * 2048) * model_management.dtype_size(dtype) self.latent_channels = 64 self.output_channels = 2 + self.pad_channel_value = "replicate" self.upscale_ratio = 2048 self.downscale_ratio = 2048 self.latent_dim = 1 @@ -547,6 +549,7 @@ class VAE: self.latent_dim = 3 self.latent_channels = 16 self.output_channels = sd["encoder.conv1.weight"].shape[1] + self.pad_channel_value = 1.0 ddconfig = {"dim": dim, "z_dim": self.latent_channels, "dim_mult": [1, 2, 4, 4], "num_res_blocks": 2, "attn_scales": [], "temperal_downsample": [False, True, True], "image_channels": self.output_channels, "dropout": 0.0} self.first_stage_model = comfy.ldm.wan.vae.WanVAE(**ddconfig) self.working_dtypes = [torch.bfloat16, torch.float16, torch.float32] @@ -583,6 +586,7 @@ class VAE: self.memory_used_decode = lambda shape, dtype: (shape[2] * shape[3] * 87000) * model_management.dtype_size(dtype) self.latent_channels = 8 self.output_channels = 2 + self.pad_channel_value = "replicate" self.upscale_ratio = 4096 self.downscale_ratio = 4096 self.latent_dim = 2 @@ -691,17 +695,28 @@ class VAE: raise RuntimeError("ERROR: VAE is invalid: None\n\nIf the VAE is from a checkpoint loader node your checkpoint does not contain a valid VAE.") def vae_encode_crop_pixels(self, pixels): - if not self.crop_input: - return pixels + if self.crop_input: + downscale_ratio = self.spacial_compression_encode() - downscale_ratio = self.spacial_compression_encode() + dims = pixels.shape[1:-1] + for d in range(len(dims)): + x = (dims[d] // downscale_ratio) * downscale_ratio + x_offset = (dims[d] % downscale_ratio) // 2 + if x != dims[d]: + pixels = pixels.narrow(d + 1, x_offset, x) - dims = pixels.shape[1:-1] - for d in range(len(dims)): - x = (dims[d] // downscale_ratio) * downscale_ratio - x_offset = (dims[d] % downscale_ratio) // 2 - if x != dims[d]: - pixels = pixels.narrow(d + 1, x_offset, x) + if pixels.shape[-1] > self.output_channels: + pixels = pixels[..., :self.output_channels] + elif pixels.shape[-1] < self.output_channels: + if self.pad_channel_value is not None: + if isinstance(self.pad_channel_value, str): + mode = self.pad_channel_value + value = None + else: + mode = "constant" + value = self.pad_channel_value + + pixels = torch.nn.functional.pad(pixels, (0, self.output_channels - pixels.shape[-1]), mode=mode, value=value) return pixels def decode_tiled_(self, samples, tile_x=64, tile_y=64, overlap = 16): diff --git a/nodes.py b/nodes.py index 3fa543294..b13ceb578 100644 --- a/nodes.py +++ b/nodes.py @@ -343,7 +343,7 @@ class VAEEncode: CATEGORY = "latent" def encode(self, vae, pixels): - t = vae.encode(pixels[:,:,:,:3]) + t = vae.encode(pixels) return ({"samples":t}, ) class VAEEncodeTiled: @@ -361,7 +361,7 @@ class VAEEncodeTiled: CATEGORY = "_for_testing" def encode(self, vae, pixels, tile_size, overlap, temporal_size=64, temporal_overlap=8): - t = vae.encode_tiled(pixels[:,:,:,:3], tile_x=tile_size, tile_y=tile_size, overlap=overlap, tile_t=temporal_size, overlap_t=temporal_overlap) + t = vae.encode_tiled(pixels, tile_x=tile_size, tile_y=tile_size, overlap=overlap, tile_t=temporal_size, overlap_t=temporal_overlap) return ({"samples": t}, ) class VAEEncodeForInpaint: From 28eaab608bc34c4e3b1886b1bddbb429453249d8 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Thu, 18 Dec 2025 17:21:14 -0800 Subject: [PATCH 315/647] Diffusion model part of Qwen Image Layered. (#11408) Only thing missing after this is some nodes to make using it easier. --- comfy/ldm/qwen_image/model.py | 63 ++++++++++++++++++++++------------- comfy/model_detection.py | 3 ++ 2 files changed, 42 insertions(+), 24 deletions(-) diff --git a/comfy/ldm/qwen_image/model.py b/comfy/ldm/qwen_image/model.py index 902af30ed..00c597535 100644 --- a/comfy/ldm/qwen_image/model.py +++ b/comfy/ldm/qwen_image/model.py @@ -61,7 +61,7 @@ def apply_rotary_emb(x, freqs_cis): class QwenTimestepProjEmbeddings(nn.Module): - def __init__(self, embedding_dim, pooled_projection_dim, dtype=None, device=None, operations=None): + def __init__(self, embedding_dim, pooled_projection_dim, use_additional_t_cond=False, dtype=None, device=None, operations=None): super().__init__() self.time_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0, scale=1000) self.timestep_embedder = TimestepEmbedding( @@ -72,9 +72,19 @@ class QwenTimestepProjEmbeddings(nn.Module): operations=operations ) - def forward(self, timestep, hidden_states): + self.use_additional_t_cond = use_additional_t_cond + if self.use_additional_t_cond: + self.addition_t_embedding = operations.Embedding(2, embedding_dim, device=device, dtype=dtype) + + def forward(self, timestep, hidden_states, addition_t_cond=None): timesteps_proj = self.time_proj(timestep) timesteps_emb = self.timestep_embedder(timesteps_proj.to(dtype=hidden_states.dtype)) + + if self.use_additional_t_cond: + if addition_t_cond is None: + addition_t_cond = torch.zeros((timesteps_emb.shape[0]), device=timesteps_emb.device, dtype=torch.long) + timesteps_emb += self.addition_t_embedding(addition_t_cond, out_dtype=timesteps_emb.dtype) + return timesteps_emb @@ -320,11 +330,11 @@ class QwenImageTransformer2DModel(nn.Module): num_attention_heads: int = 24, joint_attention_dim: int = 3584, pooled_projection_dim: int = 768, - guidance_embeds: bool = False, axes_dims_rope: Tuple[int, int, int] = (16, 56, 56), default_ref_method="index", image_model=None, final_layer=True, + use_additional_t_cond=False, dtype=None, device=None, operations=None, @@ -342,6 +352,7 @@ class QwenImageTransformer2DModel(nn.Module): self.time_text_embed = QwenTimestepProjEmbeddings( embedding_dim=self.inner_dim, pooled_projection_dim=pooled_projection_dim, + use_additional_t_cond=use_additional_t_cond, dtype=dtype, device=device, operations=operations @@ -375,27 +386,33 @@ class QwenImageTransformer2DModel(nn.Module): patch_size = self.patch_size hidden_states = comfy.ldm.common_dit.pad_to_patch_size(x, (1, self.patch_size, self.patch_size)) orig_shape = hidden_states.shape - hidden_states = hidden_states.view(orig_shape[0], orig_shape[1], orig_shape[-2] // 2, 2, orig_shape[-1] // 2, 2) - hidden_states = hidden_states.permute(0, 2, 4, 1, 3, 5) - hidden_states = hidden_states.reshape(orig_shape[0], (orig_shape[-2] // 2) * (orig_shape[-1] // 2), orig_shape[1] * 4) + hidden_states = hidden_states.view(orig_shape[0], orig_shape[1], orig_shape[-3], orig_shape[-2] // 2, 2, orig_shape[-1] // 2, 2) + hidden_states = hidden_states.permute(0, 2, 3, 5, 1, 4, 6) + hidden_states = hidden_states.reshape(orig_shape[0], orig_shape[-3] * (orig_shape[-2] // 2) * (orig_shape[-1] // 2), orig_shape[1] * 4) + t_len = t h_len = ((h + (patch_size // 2)) // patch_size) w_len = ((w + (patch_size // 2)) // patch_size) h_offset = ((h_offset + (patch_size // 2)) // patch_size) w_offset = ((w_offset + (patch_size // 2)) // patch_size) - img_ids = torch.zeros((h_len, w_len, 3), device=x.device) - img_ids[:, :, 0] = img_ids[:, :, 1] + index - img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(h_offset, h_len - 1 + h_offset, steps=h_len, device=x.device, dtype=x.dtype).unsqueeze(1) - (h_len // 2) - img_ids[:, :, 2] = img_ids[:, :, 2] + torch.linspace(w_offset, w_len - 1 + w_offset, steps=w_len, device=x.device, dtype=x.dtype).unsqueeze(0) - (w_len // 2) - return hidden_states, repeat(img_ids, "h w c -> b (h w) c", b=bs), orig_shape + img_ids = torch.zeros((t_len, h_len, w_len, 3), device=x.device) - def forward(self, x, timestep, context, attention_mask=None, guidance=None, ref_latents=None, transformer_options={}, **kwargs): + if t_len > 1: + img_ids[:, :, :, 0] = img_ids[:, :, :, 0] + torch.linspace(0, t_len - 1, steps=t_len, device=x.device, dtype=x.dtype).unsqueeze(1).unsqueeze(1) + else: + img_ids[:, :, :, 0] = img_ids[:, :, :, 0] + index + + img_ids[:, :, :, 1] = img_ids[:, :, :, 1] + torch.linspace(h_offset, h_len - 1 + h_offset, steps=h_len, device=x.device, dtype=x.dtype).unsqueeze(1).unsqueeze(0) - (h_len // 2) + img_ids[:, :, :, 2] = img_ids[:, :, :, 2] + torch.linspace(w_offset, w_len - 1 + w_offset, steps=w_len, device=x.device, dtype=x.dtype).unsqueeze(0).unsqueeze(0) - (w_len // 2) + return hidden_states, repeat(img_ids, "t h w c -> b (t h w) c", b=bs), orig_shape + + def forward(self, x, timestep, context, attention_mask=None, ref_latents=None, additional_t_cond=None, transformer_options={}, **kwargs): return comfy.patcher_extension.WrapperExecutor.new_class_executor( self._forward, self, comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options) - ).execute(x, timestep, context, attention_mask, guidance, ref_latents, transformer_options, **kwargs) + ).execute(x, timestep, context, attention_mask, ref_latents, additional_t_cond, transformer_options, **kwargs) def _forward( self, @@ -403,8 +420,8 @@ class QwenImageTransformer2DModel(nn.Module): timesteps, context, attention_mask=None, - guidance: torch.Tensor = None, ref_latents=None, + additional_t_cond=None, transformer_options={}, control=None, **kwargs @@ -423,12 +440,17 @@ class QwenImageTransformer2DModel(nn.Module): index = 0 ref_method = kwargs.get("ref_latents_method", self.default_ref_method) index_ref_method = (ref_method == "index") or (ref_method == "index_timestep_zero") + negative_ref_method = ref_method == "negative_index" timestep_zero = ref_method == "index_timestep_zero" for ref in ref_latents: if index_ref_method: index += 1 h_offset = 0 w_offset = 0 + elif negative_ref_method: + index -= 1 + h_offset = 0 + w_offset = 0 else: index = 1 h_offset = 0 @@ -458,14 +480,7 @@ class QwenImageTransformer2DModel(nn.Module): encoder_hidden_states = self.txt_norm(encoder_hidden_states) encoder_hidden_states = self.txt_in(encoder_hidden_states) - if guidance is not None: - guidance = guidance * 1000 - - temb = ( - self.time_text_embed(timestep, hidden_states) - if guidance is None - else self.time_text_embed(timestep, guidance, hidden_states) - ) + temb = self.time_text_embed(timestep, hidden_states, additional_t_cond) patches_replace = transformer_options.get("patches_replace", {}) patches = transformer_options.get("patches", {}) @@ -513,6 +528,6 @@ class QwenImageTransformer2DModel(nn.Module): hidden_states = self.norm_out(hidden_states, temb) hidden_states = self.proj_out(hidden_states) - hidden_states = hidden_states[:, :num_embeds].view(orig_shape[0], orig_shape[-2] // 2, orig_shape[-1] // 2, orig_shape[1], 2, 2) - hidden_states = hidden_states.permute(0, 3, 1, 4, 2, 5) + hidden_states = hidden_states[:, :num_embeds].view(orig_shape[0], orig_shape[-3], orig_shape[-2] // 2, orig_shape[-1] // 2, orig_shape[1], 2, 2) + hidden_states = hidden_states.permute(0, 4, 1, 2, 5, 3, 6) return hidden_states.reshape(orig_shape)[:, :, :, :x.shape[-2], :x.shape[-1]] diff --git a/comfy/model_detection.py b/comfy/model_detection.py index 7148c77fd..84fd409fd 100644 --- a/comfy/model_detection.py +++ b/comfy/model_detection.py @@ -620,6 +620,9 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): dit_config["num_layers"] = count_blocks(state_dict_keys, '{}transformer_blocks.'.format(key_prefix) + '{}.') if "{}__index_timestep_zero__".format(key_prefix) in state_dict_keys: # 2511 dit_config["default_ref_method"] = "index_timestep_zero" + if "{}time_text_embed.addition_t_embedding.weight".format(key_prefix) in state_dict_keys: # Layered + dit_config["use_additional_t_cond"] = True + dit_config["default_ref_method"] = "negative_index" return dit_config if '{}visual_transformer_blocks.0.cross_attention.key_norm.weight'.format(key_prefix) in state_dict_keys: # Kandinsky 5 From 894802b0f9c3a247f5609db89ec3be24eac7fd2b Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Thu, 18 Dec 2025 19:21:40 -0800 Subject: [PATCH 316/647] Add LatentCutToBatch node. (#11411) --- comfy_extras/nodes_latent.py | 43 ++++++++++++++++++++++++++++++++++++ 1 file changed, 43 insertions(+) diff --git a/comfy_extras/nodes_latent.py b/comfy_extras/nodes_latent.py index e439b18ef..2815c5ffc 100644 --- a/comfy_extras/nodes_latent.py +++ b/comfy_extras/nodes_latent.py @@ -5,6 +5,7 @@ import nodes from typing_extensions import override from comfy_api.latest import ComfyExtension, io import logging +import math def reshape_latent_to(target_shape, latent, repeat_batch=True): if latent.shape[1:] != target_shape[1:]: @@ -207,6 +208,47 @@ class LatentCut(io.ComfyNode): samples_out["samples"] = torch.narrow(s1, dim, index, amount) return io.NodeOutput(samples_out) +class LatentCutToBatch(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="LatentCutToBatch", + category="latent/advanced", + inputs=[ + io.Latent.Input("samples"), + io.Combo.Input("dim", options=["t", "x", "y"]), + io.Int.Input("slice_size", default=1, min=1, max=nodes.MAX_RESOLUTION, step=1), + ], + outputs=[ + io.Latent.Output(), + ], + ) + + @classmethod + def execute(cls, samples, dim, slice_size) -> io.NodeOutput: + samples_out = samples.copy() + + s1 = samples["samples"] + + if "x" in dim: + dim = s1.ndim - 1 + elif "y" in dim: + dim = s1.ndim - 2 + elif "t" in dim: + dim = s1.ndim - 3 + + if dim < 2: + return io.NodeOutput(samples) + + s = s1.movedim(dim, 1) + if s.shape[1] < slice_size: + slice_size = s.shape[1] + elif s.shape[1] % slice_size != 0: + s = s[:, :math.floor(s.shape[1] / slice_size) * slice_size] + new_shape = [-1, slice_size] + list(s.shape[2:]) + samples_out["samples"] = s.reshape(new_shape).movedim(1, dim) + return io.NodeOutput(samples_out) + class LatentBatch(io.ComfyNode): @classmethod def define_schema(cls): @@ -435,6 +477,7 @@ class LatentExtension(ComfyExtension): LatentInterpolate, LatentConcat, LatentCut, + LatentCutToBatch, LatentBatch, LatentBatchSeedBehavior, LatentApplyOperation, From 5b4d0664c87dc62a8361fe292b0bdac42681aef8 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Fri, 19 Dec 2025 20:02:49 +0200 Subject: [PATCH 317/647] add Flux2MaxImage API Node (#11420) --- comfy_api_nodes/nodes_bfl.py | 68 ++++++++++++++++-------------------- 1 file changed, 31 insertions(+), 37 deletions(-) diff --git a/comfy_api_nodes/nodes_bfl.py b/comfy_api_nodes/nodes_bfl.py index 8826dea0c..ce077d6b3 100644 --- a/comfy_api_nodes/nodes_bfl.py +++ b/comfy_api_nodes/nodes_bfl.py @@ -1,10 +1,8 @@ -from inspect import cleandoc - import torch from pydantic import BaseModel from typing_extensions import override -from comfy_api.latest import IO, ComfyExtension +from comfy_api.latest import IO, ComfyExtension, Input from comfy_api_nodes.apis.bfl_api import ( BFLFluxExpandImageRequest, BFLFluxFillImageRequest, @@ -28,7 +26,7 @@ from comfy_api_nodes.util import ( ) -def convert_mask_to_image(mask: torch.Tensor): +def convert_mask_to_image(mask: Input.Image): """ Make mask have the expected amount of dims (4) and channels (3) to be recognized as an image. """ @@ -38,9 +36,6 @@ def convert_mask_to_image(mask: torch.Tensor): class FluxProUltraImageNode(IO.ComfyNode): - """ - Generates images using Flux Pro 1.1 Ultra via api based on prompt and resolution. - """ @classmethod def define_schema(cls) -> IO.Schema: @@ -48,7 +43,7 @@ class FluxProUltraImageNode(IO.ComfyNode): node_id="FluxProUltraImageNode", display_name="Flux 1.1 [pro] Ultra Image", category="api node/image/BFL", - description=cleandoc(cls.__doc__ or ""), + description="Generates images using Flux Pro 1.1 Ultra via api based on prompt and resolution.", inputs=[ IO.String.Input( "prompt", @@ -117,7 +112,7 @@ class FluxProUltraImageNode(IO.ComfyNode): prompt_upsampling: bool = False, raw: bool = False, seed: int = 0, - image_prompt: torch.Tensor | None = None, + image_prompt: Input.Image | None = None, image_prompt_strength: float = 0.1, ) -> IO.NodeOutput: if image_prompt is None: @@ -155,9 +150,6 @@ class FluxProUltraImageNode(IO.ComfyNode): class FluxKontextProImageNode(IO.ComfyNode): - """ - Edits images using Flux.1 Kontext [pro] via api based on prompt and aspect ratio. - """ @classmethod def define_schema(cls) -> IO.Schema: @@ -165,7 +157,7 @@ class FluxKontextProImageNode(IO.ComfyNode): node_id=cls.NODE_ID, display_name=cls.DISPLAY_NAME, category="api node/image/BFL", - description=cleandoc(cls.__doc__ or ""), + description="Edits images using Flux.1 Kontext [pro] via api based on prompt and aspect ratio.", inputs=[ IO.String.Input( "prompt", @@ -231,7 +223,7 @@ class FluxKontextProImageNode(IO.ComfyNode): aspect_ratio: str, guidance: float, steps: int, - input_image: torch.Tensor | None = None, + input_image: Input.Image | None = None, seed=0, prompt_upsampling=False, ) -> IO.NodeOutput: @@ -271,20 +263,14 @@ class FluxKontextProImageNode(IO.ComfyNode): class FluxKontextMaxImageNode(FluxKontextProImageNode): - """ - Edits images using Flux.1 Kontext [max] via api based on prompt and aspect ratio. - """ - DESCRIPTION = cleandoc(__doc__ or "") + DESCRIPTION = "Edits images using Flux.1 Kontext [max] via api based on prompt and aspect ratio." BFL_PATH = "/proxy/bfl/flux-kontext-max/generate" NODE_ID = "FluxKontextMaxImageNode" DISPLAY_NAME = "Flux.1 Kontext [max] Image" class FluxProExpandNode(IO.ComfyNode): - """ - Outpaints image based on prompt. - """ @classmethod def define_schema(cls) -> IO.Schema: @@ -292,7 +278,7 @@ class FluxProExpandNode(IO.ComfyNode): node_id="FluxProExpandNode", display_name="Flux.1 Expand Image", category="api node/image/BFL", - description=cleandoc(cls.__doc__ or ""), + description="Outpaints image based on prompt.", inputs=[ IO.Image.Input("image"), IO.String.Input( @@ -371,7 +357,7 @@ class FluxProExpandNode(IO.ComfyNode): @classmethod async def execute( cls, - image: torch.Tensor, + image: Input.Image, prompt: str, prompt_upsampling: bool, top: int, @@ -418,9 +404,6 @@ class FluxProExpandNode(IO.ComfyNode): class FluxProFillNode(IO.ComfyNode): - """ - Inpaints image based on mask and prompt. - """ @classmethod def define_schema(cls) -> IO.Schema: @@ -428,7 +411,7 @@ class FluxProFillNode(IO.ComfyNode): node_id="FluxProFillNode", display_name="Flux.1 Fill Image", category="api node/image/BFL", - description=cleandoc(cls.__doc__ or ""), + description="Inpaints image based on mask and prompt.", inputs=[ IO.Image.Input("image"), IO.Mask.Input("mask"), @@ -480,8 +463,8 @@ class FluxProFillNode(IO.ComfyNode): @classmethod async def execute( cls, - image: torch.Tensor, - mask: torch.Tensor, + image: Input.Image, + mask: Input.Image, prompt: str, prompt_upsampling: bool, steps: int, @@ -525,11 +508,15 @@ class FluxProFillNode(IO.ComfyNode): class Flux2ProImageNode(IO.ComfyNode): + NODE_ID = "Flux2ProImageNode" + DISPLAY_NAME = "Flux.2 [pro] Image" + API_ENDPOINT = "/proxy/bfl/flux-2-pro/generate" + @classmethod def define_schema(cls) -> IO.Schema: return IO.Schema( - node_id="Flux2ProImageNode", - display_name="Flux.2 [pro] Image", + node_id=cls.NODE_ID, + display_name=cls.DISPLAY_NAME, category="api node/image/BFL", description="Generates images synchronously based on prompt and resolution.", inputs=[ @@ -563,12 +550,11 @@ class Flux2ProImageNode(IO.ComfyNode): ), IO.Boolean.Input( "prompt_upsampling", - default=False, + default=True, tooltip="Whether to perform upsampling on the prompt. " - "If active, automatically modifies the prompt for more creative generation, " - "but results are nondeterministic (same seed will not produce exactly the same result).", + "If active, automatically modifies the prompt for more creative generation.", ), - IO.Image.Input("images", optional=True, tooltip="Up to 4 images to be used as references."), + IO.Image.Input("images", optional=True, tooltip="Up to 9 images to be used as references."), ], outputs=[IO.Image.Output()], hidden=[ @@ -587,7 +573,7 @@ class Flux2ProImageNode(IO.ComfyNode): height: int, seed: int, prompt_upsampling: bool, - images: torch.Tensor | None = None, + images: Input.Image | None = None, ) -> IO.NodeOutput: reference_images = {} if images is not None: @@ -598,7 +584,7 @@ class Flux2ProImageNode(IO.ComfyNode): reference_images[key_name] = tensor_to_base64_string(images[image_index], total_pixels=2048 * 2048) initial_response = await sync_op( cls, - ApiEndpoint(path="/proxy/bfl/flux-2-pro/generate", method="POST"), + ApiEndpoint(path=cls.API_ENDPOINT, method="POST"), response_model=BFLFluxProGenerateResponse, data=Flux2ProGenerateRequest( prompt=prompt, @@ -632,6 +618,13 @@ class Flux2ProImageNode(IO.ComfyNode): return IO.NodeOutput(await download_url_to_image_tensor(response.result["sample"])) +class Flux2MaxImageNode(Flux2ProImageNode): + + NODE_ID = "Flux2MaxImageNode" + DISPLAY_NAME = "Flux.2 [max] Image" + API_ENDPOINT = "/proxy/bfl/flux-2-max/generate" + + class BFLExtension(ComfyExtension): @override async def get_node_list(self) -> list[type[IO.ComfyNode]]: @@ -642,6 +635,7 @@ class BFLExtension(ComfyExtension): FluxProExpandNode, FluxProFillNode, Flux2ProImageNode, + Flux2MaxImageNode, ] From 8376ff6831b145eadc3339e1901ffe02386ab86a Mon Sep 17 00:00:00 2001 From: "Dr.Lt.Data" <128333288+ltdrdata@users.noreply.github.com> Date: Sat, 20 Dec 2025 03:41:56 +0900 Subject: [PATCH 318/647] bump comfyui_manager version to the 4.0.3b7 (#11422) --- manager_requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/manager_requirements.txt b/manager_requirements.txt index 5ef0d3a1d..2300f0c70 100644 --- a/manager_requirements.txt +++ b/manager_requirements.txt @@ -1 +1 @@ -comfyui_manager==4.0.3b5 +comfyui_manager==4.0.3b7 From cc4ddba1b68abdc64ef5a701fd0571fcf2faf98d Mon Sep 17 00:00:00 2001 From: BradPepersAMD Date: Fri, 19 Dec 2025 15:01:50 -0700 Subject: [PATCH 319/647] Allow enabling use of MIOpen by setting COMFYUI_ENABLE_MIOPEN=1 as an env var (#11366) --- comfy/model_management.py | 7 +++++-- 1 file changed, 5 insertions(+), 2 deletions(-) diff --git a/comfy/model_management.py b/comfy/model_management.py index 40717b1e4..1889ab0ac 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -26,6 +26,7 @@ import importlib import platform import weakref import gc +import os class VRAMState(Enum): DISABLED = 0 #No vram present: no need to move models to vram @@ -333,13 +334,15 @@ except: SUPPORT_FP8_OPS = args.supports_fp8_compute AMD_RDNA2_AND_OLDER_ARCH = ["gfx1030", "gfx1031", "gfx1010", "gfx1011", "gfx1012", "gfx906", "gfx900", "gfx803"] +AMD_ENABLE_MIOPEN_ENV = 'COMFYUI_ENABLE_MIOPEN' try: if is_amd(): arch = torch.cuda.get_device_properties(get_torch_device()).gcnArchName if not (any((a in arch) for a in AMD_RDNA2_AND_OLDER_ARCH)): - torch.backends.cudnn.enabled = False # Seems to improve things a lot on AMD - logging.info("Set: torch.backends.cudnn.enabled = False for better AMD performance.") + if os.getenv(AMD_ENABLE_MIOPEN_ENV) != '1': + torch.backends.cudnn.enabled = False # Seems to improve things a lot on AMD + logging.info("Set: torch.backends.cudnn.enabled = False for better AMD performance.") try: rocm_version = tuple(map(int, str(torch.version.hip).split(".")[:2])) From 809ce687493db84f6743639adf9b600753b6188e Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Fri, 19 Dec 2025 16:59:25 -0800 Subject: [PATCH 320/647] Support nested tensor denoise masks. (#11431) --- comfy/samplers.py | 21 ++++++++++++++++++--- 1 file changed, 18 insertions(+), 3 deletions(-) diff --git a/comfy/samplers.py b/comfy/samplers.py index 8340d376c..1989ef107 100755 --- a/comfy/samplers.py +++ b/comfy/samplers.py @@ -984,9 +984,6 @@ class CFGGuider: self.inner_model, self.conds, self.loaded_models = comfy.sampler_helpers.prepare_sampling(self.model_patcher, noise.shape, self.conds, self.model_options) device = self.model_patcher.load_device - if denoise_mask is not None: - denoise_mask = comfy.sampler_helpers.prepare_mask(denoise_mask, noise.shape, device) - noise = noise.to(device) latent_image = latent_image.to(device) sigmas = sigmas.to(device) @@ -1013,6 +1010,24 @@ class CFGGuider: else: latent_shapes = [latent_image.shape] + if denoise_mask is not None: + if denoise_mask.is_nested: + denoise_masks = denoise_mask.unbind() + denoise_masks = denoise_masks[:len(latent_shapes)] + else: + denoise_masks = [denoise_mask] + + for i in range(len(denoise_masks), len(latent_shapes)): + denoise_masks.append(torch.ones(latent_shapes[i])) + + for i in range(len(denoise_masks)): + denoise_masks[i] = comfy.sampler_helpers.prepare_mask(denoise_masks[i], latent_shapes[i], self.model_patcher.load_device) + + if len(denoise_masks) > 1: + denoise_mask, _ = comfy.utils.pack_latents(denoise_masks) + else: + denoise_mask = denoise_masks[0] + self.conds = {} for k in self.original_conds: self.conds[k] = list(map(lambda a: a.copy(), self.original_conds[k])) From 514c24d756997c3131c57aa21578a09429096eca Mon Sep 17 00:00:00 2001 From: drozbay <17261091+drozbay@users.noreply.github.com> Date: Fri, 19 Dec 2025 21:22:45 -0700 Subject: [PATCH 321/647] Fix error from logging line (#11423) Co-authored-by: ozbayb <17261091+ozbayb@users.noreply.github.com> --- comfy/context_windows.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy/context_windows.py b/comfy/context_windows.py index 2979b3ca1..1e0f86026 100644 --- a/comfy/context_windows.py +++ b/comfy/context_windows.py @@ -143,7 +143,7 @@ class IndexListContextHandler(ContextHandlerABC): # if multiple conds, split based on primary region if self.split_conds_to_windows and len(cond_in) > 1: region = window.get_region_index(len(cond_in)) - logging.info(f"Splitting conds to windows; using region {region} for window {window[0]}-{window[-1]} with center ratio {window.center_ratio:.3f}") + logging.info(f"Splitting conds to windows; using region {region} for window {window.index_list[0]}-{window.index_list[-1]} with center ratio {window.center_ratio:.3f}") cond_in = [cond_in[region]] # cond object is a list containing a dict - outer list is irrelevant, so just loop through it for actual_cond in cond_in: From 0aa7fa464efc4ecc35a145048c06d325c75fbf2b Mon Sep 17 00:00:00 2001 From: woctordho Date: Sat, 20 Dec 2025 13:16:46 +0800 Subject: [PATCH 322/647] Implement sliding attention in Gemma3 (#11409) --- comfy/text_encoders/llama.py | 16 ++++++++++------ 1 file changed, 10 insertions(+), 6 deletions(-) diff --git a/comfy/text_encoders/llama.py b/comfy/text_encoders/llama.py index 0d07ac8c6..ed29e014d 100644 --- a/comfy/text_encoders/llama.py +++ b/comfy/text_encoders/llama.py @@ -3,7 +3,6 @@ import torch.nn as nn from dataclasses import dataclass from typing import Optional, Any import math -import logging from comfy.ldm.modules.attention import optimized_attention_for_device import comfy.model_management @@ -177,7 +176,7 @@ class Gemma3_4B_Config: num_key_value_heads: int = 4 max_position_embeddings: int = 131072 rms_norm_eps: float = 1e-6 - rope_theta = [10000.0, 1000000.0] + rope_theta = [1000000.0, 10000.0] transformer_type: str = "gemma3" head_dim = 256 rms_norm_add = True @@ -186,8 +185,8 @@ class Gemma3_4B_Config: rope_dims = None q_norm = "gemma3" k_norm = "gemma3" - sliding_attention = [False, False, False, False, False, 1024] - rope_scale = [1.0, 8.0] + sliding_attention = [1024, 1024, 1024, 1024, 1024, False] + rope_scale = [8.0, 1.0] final_norm: bool = True class RMSNorm(nn.Module): @@ -370,7 +369,7 @@ class TransformerBlockGemma2(nn.Module): self.pre_feedforward_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps, add=config.rms_norm_add, device=device, dtype=dtype) self.post_feedforward_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps, add=config.rms_norm_add, device=device, dtype=dtype) - if config.sliding_attention is not None: # TODO: implement. (Not that necessary since models are trained on less than 1024 tokens) + if config.sliding_attention is not None: self.sliding_attention = config.sliding_attention[index % len(config.sliding_attention)] else: self.sliding_attention = False @@ -387,7 +386,12 @@ class TransformerBlockGemma2(nn.Module): if self.transformer_type == 'gemma3': if self.sliding_attention: if x.shape[1] > self.sliding_attention: - logging.warning("Warning: sliding attention not implemented, results may be incorrect") + sliding_mask = torch.full((x.shape[1], x.shape[1]), float("-inf"), device=x.device, dtype=x.dtype) + sliding_mask.tril_(diagonal=-self.sliding_attention) + if attention_mask is not None: + attention_mask = attention_mask + sliding_mask + else: + attention_mask = sliding_mask freqs_cis = freqs_cis[1] else: freqs_cis = freqs_cis[0] From 3ab9748903a8ee51f62ae8d3eeebc1f98847f4bd Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Fri, 19 Dec 2025 21:19:47 -0800 Subject: [PATCH 323/647] Disable prompt weights on newbie te. (#11434) --- comfy/sd1_clip.py | 6 ++++-- comfy/text_encoders/lumina2.py | 2 +- 2 files changed, 5 insertions(+), 3 deletions(-) diff --git a/comfy/sd1_clip.py b/comfy/sd1_clip.py index 962948dae..c512ca5d0 100644 --- a/comfy/sd1_clip.py +++ b/comfy/sd1_clip.py @@ -466,7 +466,7 @@ def load_embed(embedding_name, embedding_directory, embedding_size, embed_key=No return embed_out class SDTokenizer: - def __init__(self, tokenizer_path=None, max_length=77, pad_with_end=True, embedding_directory=None, embedding_size=768, embedding_key='clip_l', tokenizer_class=CLIPTokenizer, has_start_token=True, has_end_token=True, pad_to_max_length=True, min_length=None, pad_token=None, end_token=None, min_padding=None, pad_left=False, tokenizer_data={}, tokenizer_args={}): + def __init__(self, tokenizer_path=None, max_length=77, pad_with_end=True, embedding_directory=None, embedding_size=768, embedding_key='clip_l', tokenizer_class=CLIPTokenizer, has_start_token=True, has_end_token=True, pad_to_max_length=True, min_length=None, pad_token=None, end_token=None, min_padding=None, pad_left=False, disable_weights=False, tokenizer_data={}, tokenizer_args={}): if tokenizer_path is None: tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "sd1_tokenizer") self.tokenizer = tokenizer_class.from_pretrained(tokenizer_path, **tokenizer_args) @@ -513,6 +513,8 @@ class SDTokenizer: self.embedding_size = embedding_size self.embedding_key = embedding_key + self.disable_weights = disable_weights + def _try_get_embedding(self, embedding_name:str): ''' Takes a potential embedding name and tries to retrieve it. @@ -547,7 +549,7 @@ class SDTokenizer: min_padding = tokenizer_options.get("{}_min_padding".format(self.embedding_key), self.min_padding) text = escape_important(text) - if kwargs.get("disable_weights", False): + if kwargs.get("disable_weights", self.disable_weights): parsed_weights = [(text, 1.0)] else: parsed_weights = token_weights(text, 1.0) diff --git a/comfy/text_encoders/lumina2.py b/comfy/text_encoders/lumina2.py index 7a6cfdab2..f82883ba1 100644 --- a/comfy/text_encoders/lumina2.py +++ b/comfy/text_encoders/lumina2.py @@ -14,7 +14,7 @@ class Gemma2BTokenizer(sd1_clip.SDTokenizer): class Gemma3_4BTokenizer(sd1_clip.SDTokenizer): def __init__(self, embedding_directory=None, tokenizer_data={}): tokenizer = tokenizer_data.get("spiece_model", None) - super().__init__(tokenizer, pad_with_end=False, embedding_size=2560, embedding_key='gemma3_4b', tokenizer_class=SPieceTokenizer, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, tokenizer_args={"add_bos": True, "add_eos": False}, tokenizer_data=tokenizer_data) + super().__init__(tokenizer, pad_with_end=False, embedding_size=2560, embedding_key='gemma3_4b', tokenizer_class=SPieceTokenizer, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, tokenizer_args={"add_bos": True, "add_eos": False}, disable_weights=True, tokenizer_data=tokenizer_data) def state_dict(self): return {"spiece_model": self.tokenizer.serialize_model()} From 767ee30f217e72797df6b018417234bf8b3f7b69 Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Fri, 19 Dec 2025 21:22:17 -0800 Subject: [PATCH 324/647] ZImageFunControlNet: Fix mask concatenation in --gpu-only (#11421) This operation trades in latents which in --gpu-only may be out of the GPU The two VAE results will follow the --gpu-only defined behaviour so follow the inpaint image device when calculating the mask in this path. --- comfy_extras/nodes_model_patch.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy_extras/nodes_model_patch.py b/comfy_extras/nodes_model_patch.py index 2a0cfcf18..1355b3c93 100644 --- a/comfy_extras/nodes_model_patch.py +++ b/comfy_extras/nodes_model_patch.py @@ -348,7 +348,7 @@ class ZImageControlPatch: if self.mask is None: mask_ = torch.zeros_like(inpaint_image_latent)[:, :1] else: - mask_ = comfy.utils.common_upscale(self.mask.view(self.mask.shape[0], -1, self.mask.shape[-2], self.mask.shape[-1]).mean(dim=1, keepdim=True), inpaint_image_latent.shape[-1], inpaint_image_latent.shape[-2], "nearest", "center") + mask_ = comfy.utils.common_upscale(self.mask.view(self.mask.shape[0], -1, self.mask.shape[-2], self.mask.shape[-1]).mean(dim=1, keepdim=True).to(device=inpaint_image_latent.device), inpaint_image_latent.shape[-1], inpaint_image_latent.shape[-2], "nearest", "center") if latent_image is None: latent_image = comfy.latent_formats.Flux().process_in(self.vae.encode(torch.ones_like(inpaint_image) * 0.5)) From 31e961736a476851e2579d5d9202ed4177a71720 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Fri, 19 Dec 2025 21:23:51 -0800 Subject: [PATCH 325/647] Fix issue with batches and newbie. (#11435) --- comfy/ldm/lumina/model.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy/ldm/lumina/model.py b/comfy/ldm/lumina/model.py index 5628e2ba3..e80b1c138 100644 --- a/comfy/ldm/lumina/model.py +++ b/comfy/ldm/lumina/model.py @@ -625,7 +625,7 @@ class NextDiT(nn.Module): if pooled is not None: pooled = self.clip_text_pooled_proj(pooled) else: - pooled = torch.zeros((1, self.clip_text_dim), device=x.device, dtype=x.dtype) + pooled = torch.zeros((x.shape[0], self.clip_text_dim), device=x.device, dtype=x.dtype) adaln_input = self.time_text_embed(torch.cat((t, pooled), dim=-1)) From 4c432c11ed6f83466b8ff02569872925753a3c44 Mon Sep 17 00:00:00 2001 From: woctordho Date: Sat, 20 Dec 2025 13:57:22 +0800 Subject: [PATCH 326/647] Implement Jina CLIP v2 and NewBie dual CLIP (#11415) * Implement Jina CLIP v2 * Support quantized Gemma in NewBie dual CLIP --- comfy/model_base.py | 2 +- comfy/model_detection.py | 3 +- comfy/sd.py | 20 +++ comfy/text_encoders/jina_clip_2.py | 219 +++++++++++++++++++++++++++++ comfy/text_encoders/newbie.py | 62 ++++++++ nodes.py | 4 +- 6 files changed, 306 insertions(+), 4 deletions(-) create mode 100644 comfy/text_encoders/jina_clip_2.py create mode 100644 comfy/text_encoders/newbie.py diff --git a/comfy/model_base.py b/comfy/model_base.py index 6b8a8454d..c4f3c0639 100644 --- a/comfy/model_base.py +++ b/comfy/model_base.py @@ -1110,7 +1110,7 @@ class Lumina2(BaseModel): if 'num_tokens' not in out: out['num_tokens'] = comfy.conds.CONDConstant(cross_attn.shape[1]) - clip_text_pooled = kwargs["pooled_output"] # Newbie + clip_text_pooled = kwargs.get("pooled_output", None) # NewBie if clip_text_pooled is not None: out['clip_text_pooled'] = comfy.conds.CONDRegular(clip_text_pooled) diff --git a/comfy/model_detection.py b/comfy/model_detection.py index 84fd409fd..539e296ed 100644 --- a/comfy/model_detection.py +++ b/comfy/model_detection.py @@ -430,8 +430,9 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): dit_config["rope_theta"] = 10000.0 dit_config["ffn_dim_multiplier"] = 4.0 ctd_weight = state_dict.get('{}clip_text_pooled_proj.0.weight'.format(key_prefix), None) - if ctd_weight is not None: + if ctd_weight is not None: # NewBie dit_config["clip_text_dim"] = ctd_weight.shape[0] + # NewBie also sets axes_lens = [1024, 512, 512] but it's not used in ComfyUI elif dit_config["dim"] == 3840: # Z image dit_config["n_heads"] = 30 dit_config["n_kv_heads"] = 30 diff --git a/comfy/sd.py b/comfy/sd.py index c2a9728f3..7de7dd9c6 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -55,6 +55,8 @@ import comfy.text_encoders.hunyuan_image import comfy.text_encoders.z_image import comfy.text_encoders.ovis import comfy.text_encoders.kandinsky5 +import comfy.text_encoders.jina_clip_2 +import comfy.text_encoders.newbie import comfy.model_patcher import comfy.lora @@ -1008,6 +1010,7 @@ class CLIPType(Enum): OVIS = 21 KANDINSKY5 = 22 KANDINSKY5_IMAGE = 23 + NEWBIE = 24 def load_clip(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DIFFUSION, model_options={}): @@ -1038,6 +1041,7 @@ class TEModel(Enum): MISTRAL3_24B_PRUNED_FLUX2 = 15 QWEN3_4B = 16 QWEN3_2B = 17 + JINA_CLIP_2 = 18 def detect_te_model(sd): @@ -1047,6 +1051,8 @@ def detect_te_model(sd): return TEModel.CLIP_H if "text_model.encoder.layers.0.mlp.fc1.weight" in sd: return TEModel.CLIP_L + if "model.encoder.layers.0.mixer.Wqkv.weight" in sd: + return TEModel.JINA_CLIP_2 if "encoder.block.23.layer.1.DenseReluDense.wi_1.weight" in sd: weight = sd["encoder.block.23.layer.1.DenseReluDense.wi_1.weight"] if weight.shape[-1] == 4096: @@ -1207,6 +1213,9 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip elif te_model == TEModel.QWEN3_2B: clip_target.clip = comfy.text_encoders.ovis.te(**llama_detect(clip_data)) clip_target.tokenizer = comfy.text_encoders.ovis.OvisTokenizer + elif te_model == TEModel.JINA_CLIP_2: + clip_target.clip = comfy.text_encoders.jina_clip_2.JinaClip2TextModelWrapper + clip_target.tokenizer = comfy.text_encoders.jina_clip_2.JinaClip2TokenizerWrapper else: # clip_l if clip_type == CLIPType.SD3: @@ -1262,6 +1271,17 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip elif clip_type == CLIPType.KANDINSKY5_IMAGE: clip_target.clip = comfy.text_encoders.kandinsky5.te(**llama_detect(clip_data)) clip_target.tokenizer = comfy.text_encoders.kandinsky5.Kandinsky5TokenizerImage + elif clip_type == CLIPType.NEWBIE: + clip_target.clip = comfy.text_encoders.newbie.te(**llama_detect(clip_data)) + clip_target.tokenizer = comfy.text_encoders.newbie.NewBieTokenizer + if "model.layers.0.self_attn.q_norm.weight" in clip_data[0]: + clip_data_gemma = clip_data[0] + clip_data_jina = clip_data[1] + else: + clip_data_gemma = clip_data[1] + clip_data_jina = clip_data[0] + tokenizer_data["gemma_spiece_model"] = clip_data_gemma.get("spiece_model", None) + tokenizer_data["jina_spiece_model"] = clip_data_jina.get("spiece_model", None) else: clip_target.clip = sdxl_clip.SDXLClipModel clip_target.tokenizer = sdxl_clip.SDXLTokenizer diff --git a/comfy/text_encoders/jina_clip_2.py b/comfy/text_encoders/jina_clip_2.py new file mode 100644 index 000000000..0cffb6d16 --- /dev/null +++ b/comfy/text_encoders/jina_clip_2.py @@ -0,0 +1,219 @@ +# Jina CLIP v2 and Jina Embeddings v3 both use their modified XLM-RoBERTa architecture. Reference implementation: +# Jina CLIP v2 (both text and vision): https://huggingface.co/jinaai/jina-clip-implementation/blob/39e6a55ae971b59bea6e44675d237c99762e7ee2/modeling_clip.py +# Jina XLM-RoBERTa (text only): http://huggingface.co/jinaai/xlm-roberta-flash-implementation/blob/2b6bc3f30750b3a9648fe9b63448c09920efe9be/modeling_xlm_roberta.py + +from dataclasses import dataclass + +import torch +from torch import nn as nn +from torch.nn import functional as F + +import comfy.model_management +import comfy.ops +from comfy import sd1_clip +from .spiece_tokenizer import SPieceTokenizer + +class JinaClip2Tokenizer(sd1_clip.SDTokenizer): + def __init__(self, embedding_directory=None, tokenizer_data={}): + tokenizer = tokenizer_data.get("spiece_model", None) + # The official NewBie uses max_length=8000, but Jina Embeddings v3 actually supports 8192 + super().__init__(tokenizer, pad_with_end=False, embedding_size=1024, embedding_key='jina_clip_2', tokenizer_class=SPieceTokenizer, has_start_token=True, has_end_token=True, pad_to_max_length=False, max_length=8192, min_length=1, pad_token=1, end_token=2, tokenizer_args={"add_bos": True, "add_eos": True}, tokenizer_data=tokenizer_data) + + def state_dict(self): + return {"spiece_model": self.tokenizer.serialize_model()} + +class JinaClip2TokenizerWrapper(sd1_clip.SD1Tokenizer): + def __init__(self, embedding_directory=None, tokenizer_data={}): + super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, tokenizer=JinaClip2Tokenizer, name="jina_clip_2") + +# https://huggingface.co/jinaai/jina-embeddings-v3/blob/343dbf534c76fe845f304fa5c2d1fd87e1e78918/config.json +@dataclass +class XLMRobertaConfig: + vocab_size: int = 250002 + type_vocab_size: int = 1 + hidden_size: int = 1024 + num_hidden_layers: int = 24 + num_attention_heads: int = 16 + rotary_emb_base: float = 20000.0 + intermediate_size: int = 4096 + hidden_act: str = "gelu" + hidden_dropout_prob: float = 0.1 + attention_probs_dropout_prob: float = 0.1 + layer_norm_eps: float = 1e-05 + bos_token_id: int = 0 + eos_token_id: int = 2 + pad_token_id: int = 1 + +class XLMRobertaEmbeddings(nn.Module): + def __init__(self, config, device=None, dtype=None, ops=None): + super().__init__() + embed_dim = config.hidden_size + self.word_embeddings = ops.Embedding(config.vocab_size, embed_dim, padding_idx=config.pad_token_id, device=device, dtype=dtype) + self.token_type_embeddings = ops.Embedding(config.type_vocab_size, embed_dim, device=device, dtype=dtype) + + def forward(self, input_ids=None, embeddings=None): + if input_ids is not None and embeddings is None: + embeddings = self.word_embeddings(input_ids) + + if embeddings is not None: + token_type_ids = torch.zeros(embeddings.shape[1], device=embeddings.device, dtype=torch.int32) + token_type_embeddings = self.token_type_embeddings(token_type_ids) + embeddings = embeddings + token_type_embeddings + return embeddings + +class RotaryEmbedding(nn.Module): + def __init__(self, dim, base, device=None): + super().__init__() + inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, device=device, dtype=torch.float32) / dim)) + self.register_buffer("inv_freq", inv_freq, persistent=False) + self._seq_len_cached = 0 + self._cos_cached = None + self._sin_cached = None + + def _update_cos_sin_cache(self, seqlen, device=None, dtype=None): + if seqlen > self._seq_len_cached or self._cos_cached is None or self._cos_cached.device != device or self._cos_cached.dtype != dtype: + self._seq_len_cached = seqlen + t = torch.arange(seqlen, device=device, dtype=torch.float32) + freqs = torch.outer(t, self.inv_freq.to(device=t.device)) + emb = torch.cat((freqs, freqs), dim=-1) + self._cos_cached = emb.cos().to(dtype) + self._sin_cached = emb.sin().to(dtype) + + def forward(self, q, k): + batch, seqlen, heads, head_dim = q.shape + self._update_cos_sin_cache(seqlen, device=q.device, dtype=q.dtype) + + cos = self._cos_cached[:seqlen].view(1, seqlen, 1, head_dim) + sin = self._sin_cached[:seqlen].view(1, seqlen, 1, head_dim) + + def rotate_half(x): + size = x.shape[-1] // 2 + x1, x2 = x[..., :size], x[..., size:] + return torch.cat((-x2, x1), dim=-1) + + q_embed = (q * cos) + (rotate_half(q) * sin) + k_embed = (k * cos) + (rotate_half(k) * sin) + return q_embed, k_embed + +class MHA(nn.Module): + def __init__(self, config, device=None, dtype=None, ops=None): + super().__init__() + embed_dim = config.hidden_size + self.num_heads = config.num_attention_heads + self.head_dim = embed_dim // config.num_attention_heads + + self.rotary_emb = RotaryEmbedding(self.head_dim, config.rotary_emb_base, device=device) + self.Wqkv = ops.Linear(embed_dim, 3 * embed_dim, device=device, dtype=dtype) + self.out_proj = ops.Linear(embed_dim, embed_dim, device=device, dtype=dtype) + + def forward(self, x, mask=None, optimized_attention=None): + qkv = self.Wqkv(x) + batch_size, seq_len, _ = qkv.shape + qkv = qkv.view(batch_size, seq_len, 3, self.num_heads, self.head_dim) + q, k, v = qkv.unbind(2) + + q, k = self.rotary_emb(q, k) + + # NHD -> HND + q = q.transpose(1, 2) + k = k.transpose(1, 2) + v = v.transpose(1, 2) + + out = optimized_attention(q, k, v, heads=self.num_heads, mask=mask, skip_reshape=True) + return self.out_proj(out) + +class MLP(nn.Module): + def __init__(self, config, device=None, dtype=None, ops=None): + super().__init__() + self.fc1 = ops.Linear(config.hidden_size, config.intermediate_size, device=device, dtype=dtype) + self.activation = F.gelu + self.fc2 = ops.Linear(config.intermediate_size, config.hidden_size, device=device, dtype=dtype) + + def forward(self, x): + x = self.fc1(x) + x = self.activation(x) + x = self.fc2(x) + return x + +class Block(nn.Module): + def __init__(self, config, device=None, dtype=None, ops=None): + super().__init__() + self.mixer = MHA(config, device=device, dtype=dtype, ops=ops) + self.dropout1 = nn.Dropout(config.hidden_dropout_prob) + self.norm1 = ops.LayerNorm(config.hidden_size, eps=config.layer_norm_eps, device=device, dtype=dtype) + self.mlp = MLP(config, device=device, dtype=dtype, ops=ops) + self.dropout2 = nn.Dropout(config.hidden_dropout_prob) + self.norm2 = ops.LayerNorm(config.hidden_size, eps=config.layer_norm_eps, device=device, dtype=dtype) + + def forward(self, hidden_states, mask=None, optimized_attention=None): + mixer_out = self.mixer(hidden_states, mask=mask, optimized_attention=optimized_attention) + hidden_states = self.norm1(self.dropout1(mixer_out) + hidden_states) + mlp_out = self.mlp(hidden_states) + hidden_states = self.norm2(self.dropout2(mlp_out) + hidden_states) + return hidden_states + +class XLMRobertaEncoder(nn.Module): + def __init__(self, config, device=None, dtype=None, ops=None): + super().__init__() + self.layers = nn.ModuleList([Block(config, device=device, dtype=dtype, ops=ops) for _ in range(config.num_hidden_layers)]) + + def forward(self, hidden_states, attention_mask=None): + optimized_attention = comfy.ldm.modules.attention.optimized_attention_for_device(hidden_states.device, mask=attention_mask is not None, small_input=True) + for layer in self.layers: + hidden_states = layer(hidden_states, mask=attention_mask, optimized_attention=optimized_attention) + return hidden_states + +class XLMRobertaModel_(nn.Module): + def __init__(self, config, device=None, dtype=None, ops=None): + super().__init__() + self.embeddings = XLMRobertaEmbeddings(config, device=device, dtype=dtype, ops=ops) + self.emb_ln = ops.LayerNorm(config.hidden_size, eps=config.layer_norm_eps, device=device, dtype=dtype) + self.emb_drop = nn.Dropout(config.hidden_dropout_prob) + self.encoder = XLMRobertaEncoder(config, device=device, dtype=dtype, ops=ops) + + def forward(self, input_ids, attention_mask=None, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None, embeds_info=[]): + x = self.embeddings(input_ids=input_ids, embeddings=embeds) + x = self.emb_ln(x) + x = self.emb_drop(x) + + mask = None + if attention_mask is not None: + mask = 1.0 - attention_mask.to(x.dtype).reshape((attention_mask.shape[0], 1, 1, attention_mask.shape[-1])) + mask = mask.masked_fill(mask.to(torch.bool), -torch.finfo(x.dtype).max) + + sequence_output = self.encoder(x, attention_mask=mask) + + # Mean pool, see https://huggingface.co/jinaai/jina-clip-implementation/blob/39e6a55ae971b59bea6e44675d237c99762e7ee2/hf_model.py + pooled_output = None + if attention_mask is None: + pooled_output = sequence_output.mean(dim=1) + else: + attention_mask = attention_mask.to(sequence_output.dtype) + pooled_output = (sequence_output * attention_mask.unsqueeze(-1)).sum(dim=1) / attention_mask.sum(dim=-1, keepdim=True) + + # Intermediate output is not yet implemented, use None for placeholder + return sequence_output, None, pooled_output + +class XLMRobertaModel(nn.Module): + def __init__(self, config_dict, dtype, device, operations): + super().__init__() + self.config = XLMRobertaConfig(**config_dict) + self.model = XLMRobertaModel_(self.config, device=device, dtype=dtype, ops=operations) + self.num_layers = self.config.num_hidden_layers + + def get_input_embeddings(self): + return self.model.embeddings.word_embeddings + + def set_input_embeddings(self, embeddings): + self.model.embeddings.word_embeddings = embeddings + + def forward(self, *args, **kwargs): + return self.model(*args, **kwargs) + +class JinaClip2TextModel(sd1_clip.SDClipModel): + def __init__(self, device="cpu", dtype=None, model_options={}): + super().__init__(device=device, dtype=dtype, textmodel_json_config={}, model_class=XLMRobertaModel, special_tokens={"start": 0, "end": 2, "pad": 1}, enable_attention_masks=True, return_attention_masks=True, model_options=model_options) + +class JinaClip2TextModelWrapper(sd1_clip.SD1ClipModel): + def __init__(self, device="cpu", dtype=None, model_options={}): + super().__init__(device=device, dtype=dtype, clip_model=JinaClip2TextModel, name="jina_clip_2", model_options=model_options) diff --git a/comfy/text_encoders/newbie.py b/comfy/text_encoders/newbie.py new file mode 100644 index 000000000..31904462b --- /dev/null +++ b/comfy/text_encoders/newbie.py @@ -0,0 +1,62 @@ +import torch + +import comfy.model_management +import comfy.text_encoders.jina_clip_2 +import comfy.text_encoders.lumina2 + +class NewBieTokenizer: + def __init__(self, embedding_directory=None, tokenizer_data={}): + self.gemma = comfy.text_encoders.lumina2.Gemma3_4BTokenizer(embedding_directory=embedding_directory, tokenizer_data={"spiece_model": tokenizer_data["gemma_spiece_model"]}) + self.jina = comfy.text_encoders.jina_clip_2.JinaClip2Tokenizer(embedding_directory=embedding_directory, tokenizer_data={"spiece_model": tokenizer_data["jina_spiece_model"]}) + + def tokenize_with_weights(self, text:str, return_word_ids=False, **kwargs): + out = {} + out["gemma"] = self.gemma.tokenize_with_weights(text, return_word_ids, **kwargs) + out["jina"] = self.jina.tokenize_with_weights(text, return_word_ids, **kwargs) + return out + + def untokenize(self, token_weight_pair): + raise NotImplementedError + + def state_dict(self): + return {} + +class NewBieTEModel(torch.nn.Module): + def __init__(self, dtype_gemma=None, device="cpu", dtype=None, model_options={}): + super().__init__() + dtype_gemma = comfy.model_management.pick_weight_dtype(dtype_gemma, dtype, device) + self.gemma = comfy.text_encoders.lumina2.Gemma3_4BModel(device=device, dtype=dtype_gemma, model_options=model_options) + self.jina = comfy.text_encoders.jina_clip_2.JinaClip2TextModel(device=device, dtype=dtype, model_options=model_options) + self.dtypes = {dtype, dtype_gemma} + + def set_clip_options(self, options): + self.gemma.set_clip_options(options) + self.jina.set_clip_options(options) + + def reset_clip_options(self): + self.gemma.reset_clip_options() + self.jina.reset_clip_options() + + def encode_token_weights(self, token_weight_pairs): + token_weight_pairs_gemma = token_weight_pairs["gemma"] + token_weight_pairs_jina = token_weight_pairs["jina"] + + gemma_out, gemma_pooled, gemma_extra = self.gemma.encode_token_weights(token_weight_pairs_gemma) + jina_out, jina_pooled, jina_extra = self.jina.encode_token_weights(token_weight_pairs_jina) + + return gemma_out, jina_pooled, gemma_extra + + def load_sd(self, sd): + if "model.layers.0.self_attn.q_norm.weight" in sd: + return self.gemma.load_sd(sd) + else: + return self.jina.load_sd(sd) + +def te(dtype_llama=None, llama_quantization_metadata=None): + class NewBieTEModel_(NewBieTEModel): + def __init__(self, device="cpu", dtype=None, model_options={}): + if llama_quantization_metadata is not None: + model_options = model_options.copy() + model_options["quantization_metadata"] = llama_quantization_metadata + super().__init__(dtype_gemma=dtype_llama, device=device, dtype=dtype, model_options=model_options) + return NewBieTEModel_ diff --git a/nodes.py b/nodes.py index b13ceb578..7d83ecb21 100644 --- a/nodes.py +++ b/nodes.py @@ -970,7 +970,7 @@ class DualCLIPLoader: def INPUT_TYPES(s): return {"required": { "clip_name1": (folder_paths.get_filename_list("text_encoders"), ), "clip_name2": (folder_paths.get_filename_list("text_encoders"), ), - "type": (["sdxl", "sd3", "flux", "hunyuan_video", "hidream", "hunyuan_image", "hunyuan_video_15", "kandinsky5", "kandinsky5_image"], ), + "type": (["sdxl", "sd3", "flux", "hunyuan_video", "hidream", "hunyuan_image", "hunyuan_video_15", "kandinsky5", "kandinsky5_image", "newbie"], ), }, "optional": { "device": (["default", "cpu"], {"advanced": True}), @@ -980,7 +980,7 @@ class DualCLIPLoader: CATEGORY = "advanced/loaders" - DESCRIPTION = "[Recipes]\n\nsdxl: clip-l, clip-g\nsd3: clip-l, clip-g / clip-l, t5 / clip-g, t5\nflux: clip-l, t5\nhidream: at least one of t5 or llama, recommended t5 and llama\nhunyuan_image: qwen2.5vl 7b and byt5 small" + DESCRIPTION = "[Recipes]\n\nsdxl: clip-l, clip-g\nsd3: clip-l, clip-g / clip-l, t5 / clip-g, t5\nflux: clip-l, t5\nhidream: at least one of t5 or llama, recommended t5 and llama\nhunyuan_image: qwen2.5vl 7b and byt5 small\nnewbie: gemma-3-4b-it, jina clip v2" def load_clip(self, clip_name1, clip_name2, type, device="default"): clip_type = getattr(comfy.sd.CLIPType, type.upper(), comfy.sd.CLIPType.STABLE_DIFFUSION) From fb478f679a2998c4f2e955bcb895cc4c55f119a4 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Fri, 19 Dec 2025 22:02:43 -0800 Subject: [PATCH 327/647] Only apply gemma quant config to gemma model for newbie. (#11436) --- comfy/text_encoders/lumina2.py | 5 +++++ comfy/text_encoders/newbie.py | 2 +- 2 files changed, 6 insertions(+), 1 deletion(-) diff --git a/comfy/text_encoders/lumina2.py b/comfy/text_encoders/lumina2.py index f82883ba1..b29a7cc87 100644 --- a/comfy/text_encoders/lumina2.py +++ b/comfy/text_encoders/lumina2.py @@ -33,6 +33,11 @@ class Gemma2_2BModel(sd1_clip.SDClipModel): class Gemma3_4BModel(sd1_clip.SDClipModel): def __init__(self, device="cpu", layer="hidden", layer_idx=-2, dtype=None, attention_mask=True, model_options={}): + llama_quantization_metadata = model_options.get("llama_quantization_metadata", None) + if llama_quantization_metadata is not None: + model_options = model_options.copy() + model_options["quantization_metadata"] = llama_quantization_metadata + super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"start": 2, "pad": 0}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Gemma3_4B, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options) class LuminaModel(sd1_clip.SD1ClipModel): diff --git a/comfy/text_encoders/newbie.py b/comfy/text_encoders/newbie.py index 31904462b..db2324576 100644 --- a/comfy/text_encoders/newbie.py +++ b/comfy/text_encoders/newbie.py @@ -57,6 +57,6 @@ def te(dtype_llama=None, llama_quantization_metadata=None): def __init__(self, device="cpu", dtype=None, model_options={}): if llama_quantization_metadata is not None: model_options = model_options.copy() - model_options["quantization_metadata"] = llama_quantization_metadata + model_options["llama_quantization_metadata"] = llama_quantization_metadata super().__init__(dtype_gemma=dtype_llama, device=device, dtype=dtype, model_options=model_options) return NewBieTEModel_ From 0899012ad60db23cbc5990d164fbd22195bafcb2 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Sat, 20 Dec 2025 08:24:37 +0200 Subject: [PATCH 328/647] chore(api-nodes): by default set Watermark generation to False (#11437) --- comfy_api_nodes/apis/bytedance_api.py | 6 +++--- comfy_api_nodes/nodes_bytedance.py | 16 ++++++++-------- comfy_api_nodes/nodes_wan.py | 24 ++++++++++++------------ 3 files changed, 23 insertions(+), 23 deletions(-) diff --git a/comfy_api_nodes/apis/bytedance_api.py b/comfy_api_nodes/apis/bytedance_api.py index 77cd76f9b..b8c2f618b 100644 --- a/comfy_api_nodes/apis/bytedance_api.py +++ b/comfy_api_nodes/apis/bytedance_api.py @@ -10,7 +10,7 @@ class Text2ImageTaskCreationRequest(BaseModel): size: str | None = Field(None) seed: int | None = Field(0, ge=0, le=2147483647) guidance_scale: float | None = Field(..., ge=1.0, le=10.0) - watermark: bool | None = Field(True) + watermark: bool | None = Field(False) class Image2ImageTaskCreationRequest(BaseModel): @@ -21,7 +21,7 @@ class Image2ImageTaskCreationRequest(BaseModel): size: str | None = Field("adaptive") seed: int | None = Field(..., ge=0, le=2147483647) guidance_scale: float | None = Field(..., ge=1.0, le=10.0) - watermark: bool | None = Field(True) + watermark: bool | None = Field(False) class Seedream4Options(BaseModel): @@ -37,7 +37,7 @@ class Seedream4TaskCreationRequest(BaseModel): seed: int = Field(..., ge=0, le=2147483647) sequential_image_generation: str = Field("disabled") sequential_image_generation_options: Seedream4Options = Field(Seedream4Options(max_images=15)) - watermark: bool = Field(True) + watermark: bool = Field(False) class ImageTaskCreationResponse(BaseModel): diff --git a/comfy_api_nodes/nodes_bytedance.py b/comfy_api_nodes/nodes_bytedance.py index 57c0218d0..636cc1265 100644 --- a/comfy_api_nodes/nodes_bytedance.py +++ b/comfy_api_nodes/nodes_bytedance.py @@ -112,7 +112,7 @@ class ByteDanceImageNode(IO.ComfyNode): ), IO.Boolean.Input( "watermark", - default=True, + default=False, tooltip='Whether to add an "AI generated" watermark to the image', optional=True, ), @@ -215,7 +215,7 @@ class ByteDanceImageEditNode(IO.ComfyNode): ), IO.Boolean.Input( "watermark", - default=True, + default=False, tooltip='Whether to add an "AI generated" watermark to the image', optional=True, ), @@ -346,7 +346,7 @@ class ByteDanceSeedreamNode(IO.ComfyNode): ), IO.Boolean.Input( "watermark", - default=True, + default=False, tooltip='Whether to add an "AI generated" watermark to the image.', optional=True, ), @@ -380,7 +380,7 @@ class ByteDanceSeedreamNode(IO.ComfyNode): sequential_image_generation: str = "disabled", max_images: int = 1, seed: int = 0, - watermark: bool = True, + watermark: bool = False, fail_on_partial: bool = True, ) -> IO.NodeOutput: validate_string(prompt, strip_whitespace=True, min_length=1) @@ -507,7 +507,7 @@ class ByteDanceTextToVideoNode(IO.ComfyNode): ), IO.Boolean.Input( "watermark", - default=True, + default=False, tooltip='Whether to add an "AI generated" watermark to the video.', optional=True, ), @@ -617,7 +617,7 @@ class ByteDanceImageToVideoNode(IO.ComfyNode): ), IO.Boolean.Input( "watermark", - default=True, + default=False, tooltip='Whether to add an "AI generated" watermark to the video.', optional=True, ), @@ -739,7 +739,7 @@ class ByteDanceFirstLastFrameNode(IO.ComfyNode): ), IO.Boolean.Input( "watermark", - default=True, + default=False, tooltip='Whether to add an "AI generated" watermark to the video.', optional=True, ), @@ -862,7 +862,7 @@ class ByteDanceImageReferenceNode(IO.ComfyNode): ), IO.Boolean.Input( "watermark", - default=True, + default=False, tooltip='Whether to add an "AI generated" watermark to the video.', optional=True, ), diff --git a/comfy_api_nodes/nodes_wan.py b/comfy_api_nodes/nodes_wan.py index 17b680e13..1675fd863 100644 --- a/comfy_api_nodes/nodes_wan.py +++ b/comfy_api_nodes/nodes_wan.py @@ -46,14 +46,14 @@ class Txt2ImageParametersField(BaseModel): n: int = Field(1, description="Number of images to generate.") # we support only value=1 seed: int = Field(..., ge=0, le=2147483647) prompt_extend: bool = Field(True) - watermark: bool = Field(True) + watermark: bool = Field(False) class Image2ImageParametersField(BaseModel): size: str | None = Field(None) n: int = Field(1, description="Number of images to generate.") # we support only value=1 seed: int = Field(..., ge=0, le=2147483647) - watermark: bool = Field(True) + watermark: bool = Field(False) class Text2VideoParametersField(BaseModel): @@ -61,7 +61,7 @@ class Text2VideoParametersField(BaseModel): seed: int = Field(..., ge=0, le=2147483647) duration: int = Field(5, ge=5, le=15) prompt_extend: bool = Field(True) - watermark: bool = Field(True) + watermark: bool = Field(False) audio: bool = Field(False, description="Whether to generate audio automatically.") shot_type: str = Field("single") @@ -71,7 +71,7 @@ class Image2VideoParametersField(BaseModel): seed: int = Field(..., ge=0, le=2147483647) duration: int = Field(5, ge=5, le=15) prompt_extend: bool = Field(True) - watermark: bool = Field(True) + watermark: bool = Field(False) audio: bool = Field(False, description="Whether to generate audio automatically.") shot_type: str = Field("single") @@ -208,7 +208,7 @@ class WanTextToImageApi(IO.ComfyNode): ), IO.Boolean.Input( "watermark", - default=True, + default=False, tooltip="Whether to add an AI-generated watermark to the result.", optional=True, ), @@ -234,7 +234,7 @@ class WanTextToImageApi(IO.ComfyNode): height: int = 1024, seed: int = 0, prompt_extend: bool = True, - watermark: bool = True, + watermark: bool = False, ): initial_response = await sync_op( cls, @@ -327,7 +327,7 @@ class WanImageToImageApi(IO.ComfyNode): ), IO.Boolean.Input( "watermark", - default=True, + default=False, tooltip="Whether to add an AI-generated watermark to the result.", optional=True, ), @@ -353,7 +353,7 @@ class WanImageToImageApi(IO.ComfyNode): # width: int = 1024, # height: int = 1024, seed: int = 0, - watermark: bool = True, + watermark: bool = False, ): n_images = get_number_of_images(image) if n_images not in (1, 2): @@ -476,7 +476,7 @@ class WanTextToVideoApi(IO.ComfyNode): ), IO.Boolean.Input( "watermark", - default=True, + default=False, tooltip="Whether to add an AI-generated watermark to the result.", optional=True, ), @@ -512,7 +512,7 @@ class WanTextToVideoApi(IO.ComfyNode): seed: int = 0, generate_audio: bool = False, prompt_extend: bool = True, - watermark: bool = True, + watermark: bool = False, shot_type: str = "single", ): if "480p" in size and model == "wan2.6-t2v": @@ -637,7 +637,7 @@ class WanImageToVideoApi(IO.ComfyNode): ), IO.Boolean.Input( "watermark", - default=True, + default=False, tooltip="Whether to add an AI-generated watermark to the result.", optional=True, ), @@ -674,7 +674,7 @@ class WanImageToVideoApi(IO.ComfyNode): seed: int = 0, generate_audio: bool = False, prompt_extend: bool = True, - watermark: bool = True, + watermark: bool = False, shot_type: str = "single", ): if get_number_of_images(image) != 1: From bbb11e26081977474eec72ce36d12ec778b5a9ea Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Sat, 20 Dec 2025 18:48:28 +0200 Subject: [PATCH 329/647] fix(api-nodes): Topaz 4k video upscaling (#11438) --- comfy_api_nodes/nodes_topaz.py | 35 +++++++++++++++++++++++++++------- 1 file changed, 28 insertions(+), 7 deletions(-) diff --git a/comfy_api_nodes/nodes_topaz.py b/comfy_api_nodes/nodes_topaz.py index f522756e5..b04575ad8 100644 --- a/comfy_api_nodes/nodes_topaz.py +++ b/comfy_api_nodes/nodes_topaz.py @@ -23,10 +23,6 @@ UPSCALER_MODELS_MAP = { "Starlight (Astra) Fast": "slf-1", "Starlight (Astra) Creative": "slc-1", } -UPSCALER_VALUES_MAP = { - "FullHD (1080p)": 1920, - "4K (2160p)": 3840, -} class TopazImageEnhance(IO.ComfyNode): @@ -214,7 +210,7 @@ class TopazVideoEnhance(IO.ComfyNode): IO.Video.Input("video"), IO.Boolean.Input("upscaler_enabled", default=True), IO.Combo.Input("upscaler_model", options=list(UPSCALER_MODELS_MAP.keys())), - IO.Combo.Input("upscaler_resolution", options=list(UPSCALER_VALUES_MAP.keys())), + IO.Combo.Input("upscaler_resolution", options=["FullHD (1080p)", "4K (2160p)"]), IO.Combo.Input( "upscaler_creativity", options=["low", "middle", "high"], @@ -306,8 +302,33 @@ class TopazVideoEnhance(IO.ComfyNode): target_frame_rate = src_frame_rate filters = [] if upscaler_enabled: - target_width = UPSCALER_VALUES_MAP[upscaler_resolution] - target_height = UPSCALER_VALUES_MAP[upscaler_resolution] + if "1080p" in upscaler_resolution: + target_pixel_p = 1080 + max_long_side = 1920 + else: + target_pixel_p = 2160 + max_long_side = 3840 + ar = src_width / src_height + if src_width >= src_height: + # Landscape or Square; Attempt to set height to target (e.g., 2160), calculate width + target_height = target_pixel_p + target_width = int(target_height * ar) + # Check if width exceeds standard bounds (for ultra-wide e.g., 21:9 ARs) + if target_width > max_long_side: + target_width = max_long_side + target_height = int(target_width / ar) + else: + # Portrait; Attempt to set width to target (e.g., 2160), calculate height + target_width = target_pixel_p + target_height = int(target_width / ar) + # Check if height exceeds standard bounds + if target_height > max_long_side: + target_height = max_long_side + target_width = int(target_height * ar) + if target_width % 2 != 0: + target_width += 1 + if target_height % 2 != 0: + target_height += 1 filters.append( topaz_api.VideoEnhancementFilter( model=UPSCALER_MODELS_MAP[upscaler_model], From 807538fe6c66bca8c91edbad14414fb4e109cbde Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sat, 20 Dec 2025 17:02:02 -0800 Subject: [PATCH 330/647] Core release process. (#11447) --- README.md | 3 +++ 1 file changed, 3 insertions(+) diff --git a/README.md b/README.md index bae955b1b..b0f62695b 100644 --- a/README.md +++ b/README.md @@ -119,6 +119,9 @@ ComfyUI follows a weekly release cycle targeting Monday but this regularly chang 1. **[ComfyUI Core](https://github.com/comfyanonymous/ComfyUI)** - Releases a new stable version (e.g., v0.7.0) roughly every week. + - Starting from v0.4.0 patch versions will be used for fixes backported onto the current stable release. + - Minor versions will be used for releases off the master branch. + - Patch versions may still be used for releases on the master branch in cases where a backport would not make sense. - Commits outside of the stable release tags may be very unstable and break many custom nodes. - Serves as the foundation for the desktop release From 91bf6b6aa3d5134c1569375a34ff483d3e32e03f Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sun, 21 Dec 2025 16:59:40 -0800 Subject: [PATCH 331/647] Add node to create empty latents for qwen image layered model. (#11460) --- comfy_extras/nodes_qwen.py | 29 ++++++++++++++++++++++++++++- 1 file changed, 28 insertions(+), 1 deletion(-) diff --git a/comfy_extras/nodes_qwen.py b/comfy_extras/nodes_qwen.py index 525239ae5..fde8fac9a 100644 --- a/comfy_extras/nodes_qwen.py +++ b/comfy_extras/nodes_qwen.py @@ -3,7 +3,9 @@ import comfy.utils import math from typing_extensions import override from comfy_api.latest import ComfyExtension, io - +import comfy.model_management +import torch +import nodes class TextEncodeQwenImageEdit(io.ComfyNode): @classmethod @@ -104,12 +106,37 @@ class TextEncodeQwenImageEditPlus(io.ComfyNode): return io.NodeOutput(conditioning) +class EmptyQwenImageLayeredLatentImage(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="EmptyQwenImageLayeredLatentImage", + display_name="Empty Qwen Image Layered Latent", + category="latent/qwen", + inputs=[ + io.Int.Input("width", default=640, min=16, max=nodes.MAX_RESOLUTION, step=16), + io.Int.Input("height", default=640, min=16, max=nodes.MAX_RESOLUTION, step=16), + io.Int.Input("layers", default=3, min=0, max=nodes.MAX_RESOLUTION, step=1), + io.Int.Input("batch_size", default=1, min=1, max=4096), + ], + outputs=[ + io.Latent.Output(), + ], + ) + + @classmethod + def execute(cls, width, height, layers, batch_size=1) -> io.NodeOutput: + latent = torch.zeros([batch_size, 16, layers + 1, height // 8, width // 8], device=comfy.model_management.intermediate_device()) + return io.NodeOutput({"samples": latent}) + + class QwenExtension(ComfyExtension): @override async def get_node_list(self) -> list[type[io.ComfyNode]]: return [ TextEncodeQwenImageEdit, TextEncodeQwenImageEditPlus, + EmptyQwenImageLayeredLatentImage, ] From c176b214cc768d41892add4d4f51c5c5627cbf7b Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Mon, 22 Dec 2025 08:44:49 +0200 Subject: [PATCH 332/647] extend possible duration range for Kling O1 StartEndFrame node (#11451) --- comfy_api_nodes/nodes_kling.py | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-) diff --git a/comfy_api_nodes/nodes_kling.py b/comfy_api_nodes/nodes_kling.py index 1a6364fa0..5294b10d4 100644 --- a/comfy_api_nodes/nodes_kling.py +++ b/comfy_api_nodes/nodes_kling.py @@ -858,7 +858,7 @@ class OmniProFirstLastFrameNode(IO.ComfyNode): tooltip="A text prompt describing the video content. " "This can include both positive and negative descriptions.", ), - IO.Combo.Input("duration", options=["5", "10"]), + IO.Int.Input("duration", default=5, min=3, max=10, display_mode=IO.NumberDisplay.slider), IO.Image.Input("first_frame"), IO.Image.Input( "end_frame", @@ -897,6 +897,10 @@ class OmniProFirstLastFrameNode(IO.ComfyNode): validate_string(prompt, min_length=1, max_length=2500) if end_frame is not None and reference_images is not None: raise ValueError("The 'end_frame' input cannot be used simultaneously with 'reference_images'.") + if duration not in (5, 10) and end_frame is None and reference_images is None: + raise ValueError( + "Duration is only supported for 5 or 10 seconds if there is no end frame or reference images." + ) validate_image_dimensions(first_frame, min_width=300, min_height=300) validate_image_aspect_ratio(first_frame, (1, 2.5), (2.5, 1)) image_list: list[OmniParamImage] = [ From eb0e10aec449eed2bbcda82ae5b56070e61ed86f Mon Sep 17 00:00:00 2001 From: ComfyUI Wiki Date: Tue, 23 Dec 2025 05:02:41 +0800 Subject: [PATCH 333/647] Update workflow templates to v0.7.62 (#11467) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 54696395f..b41dbe1d7 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,5 +1,5 @@ comfyui-frontend-package==1.34.9 -comfyui-workflow-templates==0.7.60 +comfyui-workflow-templates==0.7.62 comfyui-embedded-docs==0.3.1 torch torchsde From 33aa808713f7c36cd9476c53b8b67c745e9bc107 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 22 Dec 2025 13:43:24 -0800 Subject: [PATCH 334/647] Make denoised output on custom sampler nodes work with nested tensors. (#11471) --- comfy_extras/nodes_custom_sampler.py | 12 ++++++++++-- 1 file changed, 10 insertions(+), 2 deletions(-) diff --git a/comfy_extras/nodes_custom_sampler.py b/comfy_extras/nodes_custom_sampler.py index 7ee4caac1..993889d9d 100644 --- a/comfy_extras/nodes_custom_sampler.py +++ b/comfy_extras/nodes_custom_sampler.py @@ -760,8 +760,12 @@ class SamplerCustom(io.ComfyNode): out = latent.copy() out["samples"] = samples if "x0" in x0_output: + x0_out = model.model.process_latent_out(x0_output["x0"].cpu()) + if samples.is_nested: + latent_shapes = [x.shape for x in samples.unbind()] + x0_out = comfy.nested_tensor.NestedTensor(comfy.utils.unpack_latents(x0_out, latent_shapes)) out_denoised = latent.copy() - out_denoised["samples"] = model.model.process_latent_out(x0_output["x0"].cpu()) + out_denoised["samples"] = x0_out else: out_denoised = out return io.NodeOutput(out, out_denoised) @@ -948,8 +952,12 @@ class SamplerCustomAdvanced(io.ComfyNode): out = latent.copy() out["samples"] = samples if "x0" in x0_output: + x0_out = guider.model_patcher.model.process_latent_out(x0_output["x0"].cpu()) + if samples.is_nested: + latent_shapes = [x.shape for x in samples.unbind()] + x0_out = comfy.nested_tensor.NestedTensor(comfy.utils.unpack_latents(x0_out, latent_shapes)) out_denoised = latent.copy() - out_denoised["samples"] = guider.model_patcher.model.process_latent_out(x0_output["x0"].cpu()) + out_denoised["samples"] = x0_out else: out_denoised = out return io.NodeOutput(out, out_denoised) From f4f44bb8073d02597aca61193fec6143292a0b88 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Tue, 23 Dec 2025 22:10:27 +0200 Subject: [PATCH 335/647] api-nodes: use new custom endpoint for Nano Banana (#11311) --- comfy_api_nodes/apis/gemini_api.py | 1 + comfy_api_nodes/nodes_gemini.py | 24 +++++++++++++++--------- 2 files changed, 16 insertions(+), 9 deletions(-) diff --git a/comfy_api_nodes/apis/gemini_api.py b/comfy_api_nodes/apis/gemini_api.py index f8edc38c9..d81337dae 100644 --- a/comfy_api_nodes/apis/gemini_api.py +++ b/comfy_api_nodes/apis/gemini_api.py @@ -133,6 +133,7 @@ class GeminiImageGenerateContentRequest(BaseModel): systemInstruction: GeminiSystemInstructionContent | None = Field(None) tools: list[GeminiTool] | None = Field(None) videoMetadata: GeminiVideoMetadata | None = Field(None) + uploadImagesToStorage: bool = Field(True) class GeminiGenerateContentRequest(BaseModel): diff --git a/comfy_api_nodes/nodes_gemini.py b/comfy_api_nodes/nodes_gemini.py index ad0f4b4d1..e8ed7e797 100644 --- a/comfy_api_nodes/nodes_gemini.py +++ b/comfy_api_nodes/nodes_gemini.py @@ -34,6 +34,7 @@ from comfy_api_nodes.util import ( ApiEndpoint, audio_to_base64_string, bytesio_to_image_tensor, + download_url_to_image_tensor, get_number_of_images, sync_op, tensor_to_base64_string, @@ -141,9 +142,11 @@ def get_parts_by_type(response: GeminiGenerateContentResponse, part_type: Litera ) parts = [] for part in response.candidates[0].content.parts: - if part_type == "text" and hasattr(part, "text") and part.text: + if part_type == "text" and part.text: parts.append(part) - elif hasattr(part, "inlineData") and part.inlineData and part.inlineData.mimeType == part_type: + elif part.inlineData and part.inlineData.mimeType == part_type: + parts.append(part) + elif part.fileData and part.fileData.mimeType == part_type: parts.append(part) # Skip parts that don't match the requested type return parts @@ -163,12 +166,15 @@ def get_text_from_response(response: GeminiGenerateContentResponse) -> str: return "\n".join([part.text for part in parts]) -def get_image_from_response(response: GeminiGenerateContentResponse) -> Input.Image: +async def get_image_from_response(response: GeminiGenerateContentResponse) -> Input.Image: image_tensors: list[Input.Image] = [] parts = get_parts_by_type(response, "image/png") for part in parts: - image_data = base64.b64decode(part.inlineData.data) - returned_image = bytesio_to_image_tensor(BytesIO(image_data)) + if part.inlineData: + image_data = base64.b64decode(part.inlineData.data) + returned_image = bytesio_to_image_tensor(BytesIO(image_data)) + else: + returned_image = await download_url_to_image_tensor(part.fileData.fileUri) image_tensors.append(returned_image) if len(image_tensors) == 0: return torch.zeros((1, 1024, 1024, 4)) @@ -596,7 +602,7 @@ class GeminiImage(IO.ComfyNode): response = await sync_op( cls, - endpoint=ApiEndpoint(path=f"{GEMINI_BASE_ENDPOINT}/{model}", method="POST"), + ApiEndpoint(path=f"/proxy/vertexai/gemini/{model}", method="POST"), data=GeminiImageGenerateContentRequest( contents=[ GeminiContent(role=GeminiRole.user, parts=parts), @@ -610,7 +616,7 @@ class GeminiImage(IO.ComfyNode): response_model=GeminiGenerateContentResponse, price_extractor=calculate_tokens_price, ) - return IO.NodeOutput(get_image_from_response(response), get_text_from_response(response)) + return IO.NodeOutput(await get_image_from_response(response), get_text_from_response(response)) class GeminiImage2(IO.ComfyNode): @@ -729,7 +735,7 @@ class GeminiImage2(IO.ComfyNode): response = await sync_op( cls, - ApiEndpoint(path=f"{GEMINI_BASE_ENDPOINT}/{model}", method="POST"), + ApiEndpoint(path=f"/proxy/vertexai/gemini/{model}", method="POST"), data=GeminiImageGenerateContentRequest( contents=[ GeminiContent(role=GeminiRole.user, parts=parts), @@ -743,7 +749,7 @@ class GeminiImage2(IO.ComfyNode): response_model=GeminiGenerateContentResponse, price_extractor=calculate_tokens_price, ) - return IO.NodeOutput(get_image_from_response(response), get_text_from_response(response)) + return IO.NodeOutput(await get_image_from_response(response), get_text_from_response(response)) class GeminiExtension(ComfyExtension): From 22ff1bbfcb532a294b200a90270b772a339d334e Mon Sep 17 00:00:00 2001 From: ComfyUI Wiki Date: Wed, 24 Dec 2025 09:48:45 +0800 Subject: [PATCH 336/647] chore: update workflow templates to v0.7.63 (#11482) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index b41dbe1d7..59ac599c1 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,5 +1,5 @@ comfyui-frontend-package==1.34.9 -comfyui-workflow-templates==0.7.62 +comfyui-workflow-templates==0.7.63 comfyui-embedded-docs==0.3.1 torch torchsde From e4c61d75555036fa28b6bb34e5fd67b007c9f391 Mon Sep 17 00:00:00 2001 From: comfyanonymous Date: Tue, 23 Dec 2025 20:50:02 -0500 Subject: [PATCH 337/647] ComfyUI v0.6.0 --- comfyui_version.py | 2 +- pyproject.toml | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/comfyui_version.py b/comfyui_version.py index b45309198..1f28e2407 100644 --- a/comfyui_version.py +++ b/comfyui_version.py @@ -1,3 +1,3 @@ # This file is automatically generated by the build process when version is # updated in pyproject.toml. -__version__ = "0.5.1" +__version__ = "0.6.0" diff --git a/pyproject.toml b/pyproject.toml index 3a6960811..35a268bd1 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -1,6 +1,6 @@ [project] name = "ComfyUI" -version = "0.5.1" +version = "0.6.0" readme = "README.md" license = { file = "LICENSE" } requires-python = ">=3.9" From 650e716dda0a966a083f0efe299f3e83336f920e Mon Sep 17 00:00:00 2001 From: Comfy Org PR Bot Date: Wed, 24 Dec 2025 14:29:41 +0900 Subject: [PATCH 338/647] Bump comfyui-frontend-package to 1.35.9 (#11470) Co-authored-by: github-actions[bot] --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 59ac599c1..84b1882aa 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,4 +1,4 @@ -comfyui-frontend-package==1.34.9 +comfyui-frontend-package==1.35.9 comfyui-workflow-templates==0.7.63 comfyui-embedded-docs==0.3.1 torch From 4f067b07fb33cc1b61d91aec73ca968ba7d9c29a Mon Sep 17 00:00:00 2001 From: ComfyUI Wiki Date: Thu, 25 Dec 2025 07:54:21 +0800 Subject: [PATCH 339/647] chore: update workflow templates to v0.7.64 (#11496) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 84b1882aa..8b670b813 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,5 +1,5 @@ comfyui-frontend-package==1.35.9 -comfyui-workflow-templates==0.7.63 +comfyui-workflow-templates==0.7.64 comfyui-embedded-docs==0.3.1 torch torchsde From 532e2850794c7b497174a0a42ac0cb1fe5b62499 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 24 Dec 2025 16:09:37 -0800 Subject: [PATCH 340/647] Add a ManualSigmas node. (#11499) Can be used to manually set the sigmas for a model. This node accepts a list of integer and floating point numbers separated with any non numeric character. --- comfy_extras/nodes_custom_sampler.py | 21 +++++++++++++++++++++ 1 file changed, 21 insertions(+) diff --git a/comfy_extras/nodes_custom_sampler.py b/comfy_extras/nodes_custom_sampler.py index 993889d9d..f19adf4b9 100644 --- a/comfy_extras/nodes_custom_sampler.py +++ b/comfy_extras/nodes_custom_sampler.py @@ -9,6 +9,7 @@ import comfy.utils import node_helpers from typing_extensions import override from comfy_api.latest import ComfyExtension, io +import re class BasicScheduler(io.ComfyNode): @@ -1013,6 +1014,25 @@ class AddNoise(io.ComfyNode): add_noise = execute +class ManualSigmas(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="ManualSigmas", + category="_for_testing/custom_sampling", + is_experimental=True, + inputs=[ + io.String.Input("sigmas", default="1, 0.5", multiline=False) + ], + outputs=[io.Sigmas.Output()] + ) + + @classmethod + def execute(cls, sigmas) -> io.NodeOutput: + sigmas = re.findall(r"[-+]?(?:\d*\.*\d+)", sigmas) + sigmas = [float(i) for i in sigmas] + sigmas = torch.FloatTensor(sigmas) + return io.NodeOutput(sigmas) class CustomSamplersExtension(ComfyExtension): @override @@ -1052,6 +1072,7 @@ class CustomSamplersExtension(ComfyExtension): DisableNoise, AddNoise, SamplerCustomAdvanced, + ManualSigmas, ] From d9a76cf66e3fc6b0047692a07bc1d24f20e16e20 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Thu, 25 Dec 2025 20:46:51 -0800 Subject: [PATCH 341/647] Specify in readme that we only support pytorch 2.4 and up. (#11512) --- README.md | 2 ++ 1 file changed, 2 insertions(+) diff --git a/README.md b/README.md index b0f62695b..6d09758c0 100644 --- a/README.md +++ b/README.md @@ -212,6 +212,8 @@ Python 3.14 works but you may encounter issues with the torch compile node. The Python 3.13 is very well supported. If you have trouble with some custom node dependencies on 3.13 you can try 3.12 +torch 2.4 and above is supported but some features might only work on newer versions. We generally recommend using the latest major version of pytorch unless it is less than 2 weeks old. + ### Instructions: Git clone this repo. From 16fb6849d296259fd2bf106a6f894650d9a12072 Mon Sep 17 00:00:00 2001 From: "Dr.Lt.Data" <128333288+ltdrdata@users.noreply.github.com> Date: Sat, 27 Dec 2025 08:55:59 +0900 Subject: [PATCH 342/647] bump comfyui_manager version to the 4.0.4 (#11521) --- manager_requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/manager_requirements.txt b/manager_requirements.txt index 2300f0c70..6585b0c19 100644 --- a/manager_requirements.txt +++ b/manager_requirements.txt @@ -1 +1 @@ -comfyui_manager==4.0.3b7 +comfyui_manager==4.0.4 From 1e4e342f54386ea4179b273c24b37bd8cbde8f37 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Fri, 26 Dec 2025 19:03:01 -0800 Subject: [PATCH 343/647] Fix noise with ancestral samplers when inferencing on cpu. (#11528) --- comfy/k_diffusion/sampling.py | 3 +++ 1 file changed, 3 insertions(+) diff --git a/comfy/k_diffusion/sampling.py b/comfy/k_diffusion/sampling.py index 1ba9edad7..0949dee44 100644 --- a/comfy/k_diffusion/sampling.py +++ b/comfy/k_diffusion/sampling.py @@ -74,6 +74,9 @@ def get_ancestral_step(sigma_from, sigma_to, eta=1.): def default_noise_sampler(x, seed=None): if seed is not None: + if x.device == torch.device("cpu"): + seed += 1 + generator = torch.Generator(device=x.device) generator.manual_seed(seed) else: From 865568b7fc5fd2a5f626b22a40c363b0a5f0b399 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Sat, 27 Dec 2025 05:16:21 +0200 Subject: [PATCH 344/647] feat(api-nodes): add Kling Motion Control node (#11493) --- comfy_api_nodes/apis/kling_api.py | 9 ++++ comfy_api_nodes/nodes_kling.py | 87 +++++++++++++++++++++++++++++++ 2 files changed, 96 insertions(+) diff --git a/comfy_api_nodes/apis/kling_api.py b/comfy_api_nodes/apis/kling_api.py index 80a758466..bf54ede3e 100644 --- a/comfy_api_nodes/apis/kling_api.py +++ b/comfy_api_nodes/apis/kling_api.py @@ -102,3 +102,12 @@ class ImageToVideoWithAudioRequest(BaseModel): prompt: str = Field(...) mode: str = Field("pro") sound: str = Field(..., description="'on' or 'off'") + + +class MotionControlRequest(BaseModel): + prompt: str = Field(...) + image_url: str = Field(...) + video_url: str = Field(...) + keep_original_sound: str = Field(...) + character_orientation: str = Field(...) + mode: str = Field(..., description="'pro' or 'std'") diff --git a/comfy_api_nodes/nodes_kling.py b/comfy_api_nodes/nodes_kling.py index 5294b10d4..58259e029 100644 --- a/comfy_api_nodes/nodes_kling.py +++ b/comfy_api_nodes/nodes_kling.py @@ -51,6 +51,7 @@ from comfy_api_nodes.apis import ( ) from comfy_api_nodes.apis.kling_api import ( ImageToVideoWithAudioRequest, + MotionControlRequest, OmniImageParamImage, OmniParamImage, OmniParamVideo, @@ -2163,6 +2164,91 @@ class ImageToVideoWithAudio(IO.ComfyNode): return IO.NodeOutput(await download_url_to_video_output(final_response.data.task_result.videos[0].url)) +class MotionControl(IO.ComfyNode): + + @classmethod + def define_schema(cls) -> IO.Schema: + return IO.Schema( + node_id="KlingMotionControl", + display_name="Kling Motion Control", + category="api node/video/Kling", + inputs=[ + IO.String.Input("prompt", multiline=True), + IO.Image.Input("reference_image"), + IO.Video.Input( + "reference_video", + tooltip="Motion reference video used to drive movement/expression.\n" + "Duration limits depend on character_orientation:\n" + " - image: 3–10s (max 10s)\n" + " - video: 3–30s (max 30s)", + ), + IO.Boolean.Input("keep_original_sound", default=True), + IO.Combo.Input( + "character_orientation", + options=["video", "image"], + tooltip="Controls where the character's facing/orientation comes from.\n" + "video: movements, expressions, camera moves, and orientation " + "follow the motion reference video (other details via prompt).\n" + "image: movements and expressions still follow the motion reference video, " + "but the character orientation matches the reference image (camera/other details via prompt).", + ), + IO.Combo.Input("mode", options=["pro", "std"]), + ], + outputs=[ + IO.Video.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + ) + + @classmethod + async def execute( + cls, + prompt: str, + reference_image: Input.Image, + reference_video: Input.Video, + keep_original_sound: bool, + character_orientation: str, + mode: str, + ) -> IO.NodeOutput: + validate_string(prompt, max_length=2500) + validate_image_dimensions(reference_image, min_width=340, min_height=340) + validate_image_aspect_ratio(reference_image, (1, 2.5), (2.5, 1)) + if character_orientation == "image": + validate_video_duration(reference_video, min_duration=3, max_duration=10) + else: + validate_video_duration(reference_video, min_duration=3, max_duration=30) + validate_video_dimensions(reference_video, min_width=340, min_height=340, max_width=3850, max_height=3850) + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/kling/v1/videos/motion-control", method="POST"), + response_model=TaskStatusResponse, + data=MotionControlRequest( + prompt=prompt, + image_url=(await upload_images_to_comfyapi(cls, reference_image))[0], + video_url=await upload_video_to_comfyapi(cls, reference_video), + keep_original_sound="yes" if keep_original_sound else "no", + character_orientation=character_orientation, + mode=mode, + ), + ) + if response.code: + raise RuntimeError( + f"Kling request failed. Code: {response.code}, Message: {response.message}, Data: {response.data}" + ) + final_response = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/kling/v1/videos/motion-control/{response.data.task_id}"), + response_model=TaskStatusResponse, + status_extractor=lambda r: (r.data.task_status if r.data else None), + ) + return IO.NodeOutput(await download_url_to_video_output(final_response.data.task_result.videos[0].url)) + + class KlingExtension(ComfyExtension): @override async def get_node_list(self) -> list[type[IO.ComfyNode]]: @@ -2188,6 +2274,7 @@ class KlingExtension(ComfyExtension): OmniProImageNode, TextToVideoWithAudio, ImageToVideoWithAudio, + MotionControl, ] From eff4ea0b625e851d641b8f6532ff7afe2df16b9d Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Sat, 27 Dec 2025 05:39:02 +0200 Subject: [PATCH 345/647] [V3] converted nodes_images.py to V3 schema (#11206) * converted nodes_images.py to V3 schema * fix test --- comfy_api/latest/_io.py | 5 +- comfy_api/latest/_util/__init__.py | 2 + comfy_api/latest/_util/image_types.py | 18 + comfy_extras/nodes_images.py | 683 +++++++++--------- .../comfy_extras_test/image_stitch_test.py | 2 +- 5 files changed, 351 insertions(+), 359 deletions(-) create mode 100644 comfy_api/latest/_util/image_types.py diff --git a/comfy_api/latest/_io.py b/comfy_api/latest/_io.py index 4b14e5ded..ba0b95498 100644 --- a/comfy_api/latest/_io.py +++ b/comfy_api/latest/_io.py @@ -28,9 +28,8 @@ from comfy_api.internal import (_ComfyNodeInternal, _NodeOutputInternal, classpr prune_dict, shallow_clone_class) from ._resources import Resources, ResourcesLocal from comfy_execution.graph_utils import ExecutionBlocker -from ._util import MESH, VOXEL +from ._util import MESH, VOXEL, SVG as _SVG -# from comfy_extras.nodes_images import SVG as SVG_ # NOTE: needs to be moved before can be imported due to circular reference class FolderType(str, Enum): input = "input" @@ -656,7 +655,7 @@ class Video(ComfyTypeIO): @comfytype(io_type="SVG") class SVG(ComfyTypeIO): - Type = Any # TODO: SVG class is defined in comfy_extras/nodes_images.py, causing circular reference; should be moved to somewhere else before referenced directly in v3 + Type = _SVG @comfytype(io_type="LORA_MODEL") class LoraModel(ComfyTypeIO): diff --git a/comfy_api/latest/_util/__init__.py b/comfy_api/latest/_util/__init__.py index fc5431dda..6313eb01b 100644 --- a/comfy_api/latest/_util/__init__.py +++ b/comfy_api/latest/_util/__init__.py @@ -1,5 +1,6 @@ from .video_types import VideoContainer, VideoCodec, VideoComponents from .geometry_types import VOXEL, MESH +from .image_types import SVG __all__ = [ # Utility Types @@ -8,4 +9,5 @@ __all__ = [ "VideoComponents", "VOXEL", "MESH", + "SVG", ] diff --git a/comfy_api/latest/_util/image_types.py b/comfy_api/latest/_util/image_types.py new file mode 100644 index 000000000..f031ed426 --- /dev/null +++ b/comfy_api/latest/_util/image_types.py @@ -0,0 +1,18 @@ +from io import BytesIO + + +class SVG: + """Stores SVG representations via a list of BytesIO objects.""" + + def __init__(self, data: list[BytesIO]): + self.data = data + + def combine(self, other: 'SVG') -> 'SVG': + return SVG(self.data + other.data) + + @staticmethod + def combine_all(svgs: list['SVG']) -> 'SVG': + all_svgs_list: list[BytesIO] = [] + for svg_item in svgs: + all_svgs_list.extend(svg_item.data) + return SVG(all_svgs_list) diff --git a/comfy_extras/nodes_images.py b/comfy_extras/nodes_images.py index 392aea32c..ce21caade 100644 --- a/comfy_extras/nodes_images.py +++ b/comfy_extras/nodes_images.py @@ -2,280 +2,231 @@ from __future__ import annotations import nodes import folder_paths -from comfy.cli_args import args -from PIL import Image -from PIL.PngImagePlugin import PngInfo - -import numpy as np import json import os import re -from io import BytesIO -from inspect import cleandoc import torch import comfy.utils -from comfy.comfy_types import FileLocator, IO from server import PromptServer +from comfy_api.latest import ComfyExtension, IO, UI +from typing_extensions import override + +SVG = IO.SVG.Type # TODO: temporary solution for backward compatibility, will be removed later. MAX_RESOLUTION = nodes.MAX_RESOLUTION -class ImageCrop: +class ImageCrop(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": { "image": ("IMAGE",), - "width": ("INT", {"default": 512, "min": 1, "max": MAX_RESOLUTION, "step": 1}), - "height": ("INT", {"default": 512, "min": 1, "max": MAX_RESOLUTION, "step": 1}), - "x": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 1}), - "y": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION, "step": 1}), - }} - RETURN_TYPES = ("IMAGE",) - FUNCTION = "crop" + def define_schema(cls): + return IO.Schema( + node_id="ImageCrop", + display_name="Image Crop", + category="image/transform", + inputs=[ + IO.Image.Input("image"), + IO.Int.Input("width", default=512, min=1, max=nodes.MAX_RESOLUTION, step=1), + IO.Int.Input("height", default=512, min=1, max=nodes.MAX_RESOLUTION, step=1), + IO.Int.Input("x", default=0, min=0, max=nodes.MAX_RESOLUTION, step=1), + IO.Int.Input("y", default=0, min=0, max=nodes.MAX_RESOLUTION, step=1), + ], + outputs=[IO.Image.Output()], + ) - CATEGORY = "image/transform" - - def crop(self, image, width, height, x, y): + @classmethod + def execute(cls, image, width, height, x, y) -> IO.NodeOutput: x = min(x, image.shape[2] - 1) y = min(y, image.shape[1] - 1) to_x = width + x to_y = height + y img = image[:,y:to_y, x:to_x, :] - return (img,) + return IO.NodeOutput(img) -class RepeatImageBatch: + crop = execute # TODO: remove + + +class RepeatImageBatch(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": { "image": ("IMAGE",), - "amount": ("INT", {"default": 1, "min": 1, "max": 4096}), - }} - RETURN_TYPES = ("IMAGE",) - FUNCTION = "repeat" + def define_schema(cls): + return IO.Schema( + node_id="RepeatImageBatch", + category="image/batch", + inputs=[ + IO.Image.Input("image"), + IO.Int.Input("amount", default=1, min=1, max=4096), + ], + outputs=[IO.Image.Output()], + ) - CATEGORY = "image/batch" - - def repeat(self, image, amount): + @classmethod + def execute(cls, image, amount) -> IO.NodeOutput: s = image.repeat((amount, 1,1,1)) - return (s,) + return IO.NodeOutput(s) -class ImageFromBatch: + repeat = execute # TODO: remove + + +class ImageFromBatch(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": { "image": ("IMAGE",), - "batch_index": ("INT", {"default": 0, "min": 0, "max": 4095}), - "length": ("INT", {"default": 1, "min": 1, "max": 4096}), - }} - RETURN_TYPES = ("IMAGE",) - FUNCTION = "frombatch" + def define_schema(cls): + return IO.Schema( + node_id="ImageFromBatch", + category="image/batch", + inputs=[ + IO.Image.Input("image"), + IO.Int.Input("batch_index", default=0, min=0, max=4095), + IO.Int.Input("length", default=1, min=1, max=4096), + ], + outputs=[IO.Image.Output()], + ) - CATEGORY = "image/batch" - - def frombatch(self, image, batch_index, length): + @classmethod + def execute(cls, image, batch_index, length) -> IO.NodeOutput: s_in = image batch_index = min(s_in.shape[0] - 1, batch_index) length = min(s_in.shape[0] - batch_index, length) s = s_in[batch_index:batch_index + length].clone() - return (s,) + return IO.NodeOutput(s) + + frombatch = execute # TODO: remove -class ImageAddNoise: +class ImageAddNoise(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": { "image": ("IMAGE",), - "seed": ("INT", {"default": 0, "min": 0, "max": 0xffffffffffffffff, "control_after_generate": True, "tooltip": "The random seed used for creating the noise."}), - "strength": ("FLOAT", {"default": 0.5, "min": 0.0, "max": 1.0, "step": 0.01}), - }} - RETURN_TYPES = ("IMAGE",) - FUNCTION = "repeat" + def define_schema(cls): + return IO.Schema( + node_id="ImageAddNoise", + category="image", + inputs=[ + IO.Image.Input("image"), + IO.Int.Input( + "seed", + default=0, + min=0, + max=0xFFFFFFFFFFFFFFFF, + control_after_generate=True, + tooltip="The random seed used for creating the noise.", + ), + IO.Float.Input("strength", default=0.5, min=0.0, max=1.0, step=0.01), + ], + outputs=[IO.Image.Output()], + ) - CATEGORY = "image" - - def repeat(self, image, seed, strength): + @classmethod + def execute(cls, image, seed, strength) -> IO.NodeOutput: generator = torch.manual_seed(seed) s = torch.clip((image + strength * torch.randn(image.size(), generator=generator, device="cpu").to(image)), min=0.0, max=1.0) - return (s,) + return IO.NodeOutput(s) -class SaveAnimatedWEBP: - def __init__(self): - self.output_dir = folder_paths.get_output_directory() - self.type = "output" - self.prefix_append = "" + repeat = execute # TODO: remove - methods = {"default": 4, "fastest": 0, "slowest": 6} - @classmethod - def INPUT_TYPES(s): - return {"required": - {"images": ("IMAGE", ), - "filename_prefix": ("STRING", {"default": "ComfyUI"}), - "fps": ("FLOAT", {"default": 6.0, "min": 0.01, "max": 1000.0, "step": 0.01}), - "lossless": ("BOOLEAN", {"default": True}), - "quality": ("INT", {"default": 80, "min": 0, "max": 100}), - "method": (list(s.methods.keys()),), - # "num_frames": ("INT", {"default": 0, "min": 0, "max": 8192}), - }, - "hidden": {"prompt": "PROMPT", "extra_pnginfo": "EXTRA_PNGINFO"}, - } - RETURN_TYPES = () - FUNCTION = "save_images" - - OUTPUT_NODE = True - - CATEGORY = "image/animation" - - def save_images(self, images, fps, filename_prefix, lossless, quality, method, num_frames=0, prompt=None, extra_pnginfo=None): - method = self.methods.get(method) - filename_prefix += self.prefix_append - full_output_folder, filename, counter, subfolder, filename_prefix = folder_paths.get_save_image_path(filename_prefix, self.output_dir, images[0].shape[1], images[0].shape[0]) - results: list[FileLocator] = [] - pil_images = [] - for image in images: - i = 255. * image.cpu().numpy() - img = Image.fromarray(np.clip(i, 0, 255).astype(np.uint8)) - pil_images.append(img) - - metadata = pil_images[0].getexif() - if not args.disable_metadata: - if prompt is not None: - metadata[0x0110] = "prompt:{}".format(json.dumps(prompt)) - if extra_pnginfo is not None: - inital_exif = 0x010f - for x in extra_pnginfo: - metadata[inital_exif] = "{}:{}".format(x, json.dumps(extra_pnginfo[x])) - inital_exif -= 1 - - if num_frames == 0: - num_frames = len(pil_images) - - c = len(pil_images) - for i in range(0, c, num_frames): - file = f"{filename}_{counter:05}_.webp" - pil_images[i].save(os.path.join(full_output_folder, file), save_all=True, duration=int(1000.0/fps), append_images=pil_images[i + 1:i + num_frames], exif=metadata, lossless=lossless, quality=quality, method=method) - results.append({ - "filename": file, - "subfolder": subfolder, - "type": self.type - }) - counter += 1 - - animated = num_frames != 1 - return { "ui": { "images": results, "animated": (animated,) } } - -class SaveAnimatedPNG: - def __init__(self): - self.output_dir = folder_paths.get_output_directory() - self.type = "output" - self.prefix_append = "" +class SaveAnimatedWEBP(IO.ComfyNode): + COMPRESS_METHODS = {"default": 4, "fastest": 0, "slowest": 6} @classmethod - def INPUT_TYPES(s): - return {"required": - {"images": ("IMAGE", ), - "filename_prefix": ("STRING", {"default": "ComfyUI"}), - "fps": ("FLOAT", {"default": 6.0, "min": 0.01, "max": 1000.0, "step": 0.01}), - "compress_level": ("INT", {"default": 4, "min": 0, "max": 9}) - }, - "hidden": {"prompt": "PROMPT", "extra_pnginfo": "EXTRA_PNGINFO"}, - } + def define_schema(cls): + return IO.Schema( + node_id="SaveAnimatedWEBP", + category="image/animation", + inputs=[ + IO.Image.Input("images"), + IO.String.Input("filename_prefix", default="ComfyUI"), + IO.Float.Input("fps", default=6.0, min=0.01, max=1000.0, step=0.01), + IO.Boolean.Input("lossless", default=True), + IO.Int.Input("quality", default=80, min=0, max=100), + IO.Combo.Input("method", options=list(cls.COMPRESS_METHODS.keys())), + # "num_frames": ("INT", {"default": 0, "min": 0, "max": 8192}), + ], + hidden=[IO.Hidden.prompt, IO.Hidden.extra_pnginfo], + is_output_node=True, + ) - RETURN_TYPES = () - FUNCTION = "save_images" + @classmethod + def execute(cls, images, fps, filename_prefix, lossless, quality, method, num_frames=0) -> IO.NodeOutput: + return IO.NodeOutput( + ui=UI.ImageSaveHelper.get_save_animated_webp_ui( + images=images, + filename_prefix=filename_prefix, + cls=cls, + fps=fps, + lossless=lossless, + quality=quality, + method=cls.COMPRESS_METHODS.get(method) + ) + ) - OUTPUT_NODE = True - - CATEGORY = "image/animation" - - def save_images(self, images, fps, compress_level, filename_prefix="ComfyUI", prompt=None, extra_pnginfo=None): - filename_prefix += self.prefix_append - full_output_folder, filename, counter, subfolder, filename_prefix = folder_paths.get_save_image_path(filename_prefix, self.output_dir, images[0].shape[1], images[0].shape[0]) - results = list() - pil_images = [] - for image in images: - i = 255. * image.cpu().numpy() - img = Image.fromarray(np.clip(i, 0, 255).astype(np.uint8)) - pil_images.append(img) - - metadata = None - if not args.disable_metadata: - metadata = PngInfo() - if prompt is not None: - metadata.add(b"comf", "prompt".encode("latin-1", "strict") + b"\0" + json.dumps(prompt).encode("latin-1", "strict"), after_idat=True) - if extra_pnginfo is not None: - for x in extra_pnginfo: - metadata.add(b"comf", x.encode("latin-1", "strict") + b"\0" + json.dumps(extra_pnginfo[x]).encode("latin-1", "strict"), after_idat=True) - - file = f"{filename}_{counter:05}_.png" - pil_images[0].save(os.path.join(full_output_folder, file), pnginfo=metadata, compress_level=compress_level, save_all=True, duration=int(1000.0/fps), append_images=pil_images[1:]) - results.append({ - "filename": file, - "subfolder": subfolder, - "type": self.type - }) - - return { "ui": { "images": results, "animated": (True,)} } - -class SVG: - """ - Stores SVG representations via a list of BytesIO objects. - """ - def __init__(self, data: list[BytesIO]): - self.data = data - - def combine(self, other: 'SVG') -> 'SVG': - return SVG(self.data + other.data) - - @staticmethod - def combine_all(svgs: list['SVG']) -> 'SVG': - all_svgs_list: list[BytesIO] = [] - for svg_item in svgs: - all_svgs_list.extend(svg_item.data) - return SVG(all_svgs_list) + save_images = execute # TODO: remove -class ImageStitch: +class SaveAnimatedPNG(IO.ComfyNode): + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="SaveAnimatedPNG", + category="image/animation", + inputs=[ + IO.Image.Input("images"), + IO.String.Input("filename_prefix", default="ComfyUI"), + IO.Float.Input("fps", default=6.0, min=0.01, max=1000.0, step=0.01), + IO.Int.Input("compress_level", default=4, min=0, max=9), + ], + hidden=[IO.Hidden.prompt, IO.Hidden.extra_pnginfo], + is_output_node=True, + ) + + @classmethod + def execute(cls, images, fps, compress_level, filename_prefix="ComfyUI") -> IO.NodeOutput: + return IO.NodeOutput( + ui=UI.ImageSaveHelper.get_save_animated_png_ui( + images=images, + filename_prefix=filename_prefix, + cls=cls, + fps=fps, + compress_level=compress_level, + ) + ) + + save_images = execute # TODO: remove + + +class ImageStitch(IO.ComfyNode): """Upstreamed from https://github.com/kijai/ComfyUI-KJNodes""" @classmethod - def INPUT_TYPES(s): - return { - "required": { - "image1": ("IMAGE",), - "direction": (["right", "down", "left", "up"], {"default": "right"}), - "match_image_size": ("BOOLEAN", {"default": True}), - "spacing_width": ( - "INT", - {"default": 0, "min": 0, "max": 1024, "step": 2}, - ), - "spacing_color": ( - ["white", "black", "red", "green", "blue"], - {"default": "white"}, - ), - }, - "optional": { - "image2": ("IMAGE",), - }, - } + def define_schema(cls): + return IO.Schema( + node_id="ImageStitch", + display_name="Image Stitch", + description="Stitches image2 to image1 in the specified direction.\n" + "If image2 is not provided, returns image1 unchanged.\n" + "Optional spacing can be added between images.", + category="image/transform", + inputs=[ + IO.Image.Input("image1"), + IO.Combo.Input("direction", options=["right", "down", "left", "up"], default="right"), + IO.Boolean.Input("match_image_size", default=True), + IO.Int.Input("spacing_width", default=0, min=0, max=1024, step=2), + IO.Combo.Input("spacing_color", options=["white", "black", "red", "green", "blue"], default="white"), + IO.Image.Input("image2", optional=True), + ], + outputs=[IO.Image.Output()], + ) - RETURN_TYPES = ("IMAGE",) - FUNCTION = "stitch" - CATEGORY = "image/transform" - DESCRIPTION = """ -Stitches image2 to image1 in the specified direction. -If image2 is not provided, returns image1 unchanged. -Optional spacing can be added between images. -""" - - def stitch( - self, + @classmethod + def execute( + cls, image1, direction, match_image_size, spacing_width, spacing_color, image2=None, - ): + ) -> IO.NodeOutput: if image2 is None: - return (image1,) + return IO.NodeOutput(image1) # Handle batch size differences if image1.shape[0] != image2.shape[0]: @@ -412,36 +363,30 @@ Optional spacing can be added between images. images.insert(1, spacing) concat_dim = 2 if direction in ["left", "right"] else 1 - return (torch.cat(images, dim=concat_dim),) + return IO.NodeOutput(torch.cat(images, dim=concat_dim)) + + stitch = execute # TODO: remove + + +class ResizeAndPadImage(IO.ComfyNode): -class ResizeAndPadImage: @classmethod - def INPUT_TYPES(cls): - return { - "required": { - "image": ("IMAGE",), - "target_width": ("INT", { - "default": 512, - "min": 1, - "max": MAX_RESOLUTION, - "step": 1 - }), - "target_height": ("INT", { - "default": 512, - "min": 1, - "max": MAX_RESOLUTION, - "step": 1 - }), - "padding_color": (["white", "black"],), - "interpolation": (["area", "bicubic", "nearest-exact", "bilinear", "lanczos"],), - } - } + def define_schema(cls): + return IO.Schema( + node_id="ResizeAndPadImage", + category="image/transform", + inputs=[ + IO.Image.Input("image"), + IO.Int.Input("target_width", default=512, min=1, max=nodes.MAX_RESOLUTION, step=1), + IO.Int.Input("target_height", default=512, min=1, max=nodes.MAX_RESOLUTION, step=1), + IO.Combo.Input("padding_color", options=["white", "black"]), + IO.Combo.Input("interpolation", options=["area", "bicubic", "nearest-exact", "bilinear", "lanczos"]), + ], + outputs=[IO.Image.Output()], + ) - RETURN_TYPES = ("IMAGE",) - FUNCTION = "resize_and_pad" - CATEGORY = "image/transform" - - def resize_and_pad(self, image, target_width, target_height, padding_color, interpolation): + @classmethod + def execute(cls, image, target_width, target_height, padding_color, interpolation) -> IO.NodeOutput: batch_size, orig_height, orig_width, channels = image.shape scale_w = target_width / orig_width @@ -469,52 +414,47 @@ class ResizeAndPadImage: padded[:, :, y_offset:y_offset + new_height, x_offset:x_offset + new_width] = resized output = padded.permute(0, 2, 3, 1) - return (output,) + return IO.NodeOutput(output) -class SaveSVGNode: - """ - Save SVG files on disk. - """ + resize_and_pad = execute # TODO: remove - def __init__(self): - self.output_dir = folder_paths.get_output_directory() - self.type = "output" - self.prefix_append = "" - RETURN_TYPES = () - DESCRIPTION = cleandoc(__doc__ or "") # Handle potential None value - FUNCTION = "save_svg" - CATEGORY = "image/save" # Changed - OUTPUT_NODE = True +class SaveSVGNode(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return { - "required": { - "svg": ("SVG",), # Changed - "filename_prefix": ("STRING", {"default": "svg/ComfyUI", "tooltip": "The prefix for the file to save. This may include formatting information such as %date:yyyy-MM-dd% or %Empty Latent Image.width% to include values from nodes."}) - }, - "hidden": { - "prompt": "PROMPT", - "extra_pnginfo": "EXTRA_PNGINFO" - } - } + def define_schema(cls): + return IO.Schema( + node_id="SaveSVGNode", + description="Save SVG files on disk.", + category="image/save", + inputs=[ + IO.SVG.Input("svg"), + IO.String.Input( + "filename_prefix", + default="svg/ComfyUI", + tooltip="The prefix for the file to save. This may include formatting information such as %date:yyyy-MM-dd% or %Empty Latent Image.width% to include values from nodes.", + ), + ], + hidden=[IO.Hidden.prompt, IO.Hidden.extra_pnginfo], + is_output_node=True, + ) - def save_svg(self, svg: SVG, filename_prefix="svg/ComfyUI", prompt=None, extra_pnginfo=None): - filename_prefix += self.prefix_append - full_output_folder, filename, counter, subfolder, filename_prefix = folder_paths.get_save_image_path(filename_prefix, self.output_dir) - results = list() + @classmethod + def execute(cls, svg: IO.SVG.Type, filename_prefix="svg/ComfyUI") -> IO.NodeOutput: + full_output_folder, filename, counter, subfolder, filename_prefix = folder_paths.get_save_image_path(filename_prefix, folder_paths.get_output_directory()) + results: list[UI.SavedResult] = [] # Prepare metadata JSON metadata_dict = {} - if prompt is not None: - metadata_dict["prompt"] = prompt - if extra_pnginfo is not None: - metadata_dict.update(extra_pnginfo) + if cls.hidden.prompt is not None: + metadata_dict["prompt"] = cls.hidden.prompt + if cls.hidden.extra_pnginfo is not None: + metadata_dict.update(cls.hidden.extra_pnginfo) # Convert metadata to JSON string metadata_json = json.dumps(metadata_dict, indent=2) if metadata_dict else None + for batch_number, svg_bytes in enumerate(svg.data): filename_with_batch_num = filename.replace("%batch_num%", str(batch_number)) file = f"{filename_with_batch_num}_{counter:05}_.svg" @@ -544,57 +484,64 @@ class SaveSVGNode: with open(os.path.join(full_output_folder, file), 'wb') as svg_file: svg_file.write(svg_content.encode('utf-8')) - results.append({ - "filename": file, - "subfolder": subfolder, - "type": self.type - }) + results.append(UI.SavedResult(filename=file, subfolder=subfolder, type=IO.FolderType.output)) counter += 1 - return { "ui": { "images": results } } + return IO.NodeOutput(ui={"images": results}) -class GetImageSize: + save_svg = execute # TODO: remove + + +class GetImageSize(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return { - "required": { - "image": (IO.IMAGE,), - }, - "hidden": { - "unique_id": "UNIQUE_ID", - } - } + def define_schema(cls): + return IO.Schema( + node_id="GetImageSize", + display_name="Get Image Size", + description="Returns width and height of the image, and passes it through unchanged.", + category="image", + inputs=[ + IO.Image.Input("image"), + ], + outputs=[ + IO.Int.Output(display_name="width"), + IO.Int.Output(display_name="height"), + IO.Int.Output(display_name="batch_size"), + ], + hidden=[IO.Hidden.unique_id], + ) - RETURN_TYPES = (IO.INT, IO.INT, IO.INT) - RETURN_NAMES = ("width", "height", "batch_size") - FUNCTION = "get_size" - - CATEGORY = "image" - DESCRIPTION = """Returns width and height of the image, and passes it through unchanged.""" - - def get_size(self, image, unique_id=None) -> tuple[int, int]: + @classmethod + def execute(cls, image) -> IO.NodeOutput: height = image.shape[1] width = image.shape[2] batch_size = image.shape[0] # Send progress text to display size on the node - if unique_id: - PromptServer.instance.send_progress_text(f"width: {width}, height: {height}\n batch size: {batch_size}", unique_id) + if cls.hidden.unique_id: + PromptServer.instance.send_progress_text(f"width: {width}, height: {height}\n batch size: {batch_size}", cls.hidden.unique_id) - return width, height, batch_size + return IO.NodeOutput(width, height, batch_size) + + get_size = execute # TODO: remove + + +class ImageRotate(IO.ComfyNode): -class ImageRotate: @classmethod - def INPUT_TYPES(s): - return {"required": { "image": (IO.IMAGE,), - "rotation": (["none", "90 degrees", "180 degrees", "270 degrees"],), - }} - RETURN_TYPES = (IO.IMAGE,) - FUNCTION = "rotate" + def define_schema(cls): + return IO.Schema( + node_id="ImageRotate", + category="image/transform", + inputs=[ + IO.Image.Input("image"), + IO.Combo.Input("rotation", options=["none", "90 degrees", "180 degrees", "270 degrees"]), + ], + outputs=[IO.Image.Output()], + ) - CATEGORY = "image/transform" - - def rotate(self, image, rotation): + @classmethod + def execute(cls, image, rotation) -> IO.NodeOutput: rotate_by = 0 if rotation.startswith("90"): rotate_by = 1 @@ -604,41 +551,57 @@ class ImageRotate: rotate_by = 3 image = torch.rot90(image, k=rotate_by, dims=[2, 1]) - return (image,) + return IO.NodeOutput(image) + + rotate = execute # TODO: remove + + +class ImageFlip(IO.ComfyNode): -class ImageFlip: @classmethod - def INPUT_TYPES(s): - return {"required": { "image": (IO.IMAGE,), - "flip_method": (["x-axis: vertically", "y-axis: horizontally"],), - }} - RETURN_TYPES = (IO.IMAGE,) - FUNCTION = "flip" + def define_schema(cls): + return IO.Schema( + node_id="ImageFlip", + category="image/transform", + inputs=[ + IO.Image.Input("image"), + IO.Combo.Input("flip_method", options=["x-axis: vertically", "y-axis: horizontally"]), + ], + outputs=[IO.Image.Output()], + ) - CATEGORY = "image/transform" - - def flip(self, image, flip_method): + @classmethod + def execute(cls, image, flip_method) -> IO.NodeOutput: if flip_method.startswith("x"): image = torch.flip(image, dims=[1]) elif flip_method.startswith("y"): image = torch.flip(image, dims=[2]) - return (image,) + return IO.NodeOutput(image) -class ImageScaleToMaxDimension: - upscale_methods = ["area", "lanczos", "bilinear", "nearest-exact", "bilinear", "bicubic"] + flip = execute # TODO: remove + + +class ImageScaleToMaxDimension(IO.ComfyNode): @classmethod - def INPUT_TYPES(s): - return {"required": {"image": ("IMAGE",), - "upscale_method": (s.upscale_methods,), - "largest_size": ("INT", {"default": 512, "min": 0, "max": MAX_RESOLUTION, "step": 1})}} - RETURN_TYPES = ("IMAGE",) - FUNCTION = "upscale" + def define_schema(cls): + return IO.Schema( + node_id="ImageScaleToMaxDimension", + category="image/upscaling", + inputs=[ + IO.Image.Input("image"), + IO.Combo.Input( + "upscale_method", + options=["area", "lanczos", "bilinear", "nearest-exact", "bilinear", "bicubic"], + ), + IO.Int.Input("largest_size", default=512, min=0, max=MAX_RESOLUTION, step=1), + ], + outputs=[IO.Image.Output()], + ) - CATEGORY = "image/upscaling" - - def upscale(self, image, upscale_method, largest_size): + @classmethod + def execute(cls, image, upscale_method, largest_size) -> IO.NodeOutput: height = image.shape[1] width = image.shape[2] @@ -655,20 +618,30 @@ class ImageScaleToMaxDimension: samples = image.movedim(-1, 1) s = comfy.utils.common_upscale(samples, width, height, upscale_method, "disabled") s = s.movedim(1, -1) - return (s,) + return IO.NodeOutput(s) -NODE_CLASS_MAPPINGS = { - "ImageCrop": ImageCrop, - "RepeatImageBatch": RepeatImageBatch, - "ImageFromBatch": ImageFromBatch, - "ImageAddNoise": ImageAddNoise, - "SaveAnimatedWEBP": SaveAnimatedWEBP, - "SaveAnimatedPNG": SaveAnimatedPNG, - "SaveSVGNode": SaveSVGNode, - "ImageStitch": ImageStitch, - "ResizeAndPadImage": ResizeAndPadImage, - "GetImageSize": GetImageSize, - "ImageRotate": ImageRotate, - "ImageFlip": ImageFlip, - "ImageScaleToMaxDimension": ImageScaleToMaxDimension, -} + upscale = execute # TODO: remove + + +class ImagesExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[IO.ComfyNode]]: + return [ + ImageCrop, + RepeatImageBatch, + ImageFromBatch, + ImageAddNoise, + SaveAnimatedWEBP, + SaveAnimatedPNG, + SaveSVGNode, + ImageStitch, + ResizeAndPadImage, + GetImageSize, + ImageRotate, + ImageFlip, + ImageScaleToMaxDimension, + ] + + +async def comfy_entrypoint() -> ImagesExtension: + return ImagesExtension() diff --git a/tests-unit/comfy_extras_test/image_stitch_test.py b/tests-unit/comfy_extras_test/image_stitch_test.py index b5a0f022c..5c6a15ac4 100644 --- a/tests-unit/comfy_extras_test/image_stitch_test.py +++ b/tests-unit/comfy_extras_test/image_stitch_test.py @@ -25,7 +25,7 @@ class TestImageStitch: result = node.stitch(image1, "right", True, 0, "white", image2=None) - assert len(result) == 1 + assert len(result.result) == 1 assert torch.equal(result[0], image1) def test_basic_horizontal_stitch_right(self): From 0d2e4bdd44f61b198588c5db99bebfd5cdec286b Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Sat, 27 Dec 2025 05:55:30 +0200 Subject: [PATCH 346/647] fix(api-nodes-gemini): always force enhance_prompt to be True (#11503) --- comfy_api_nodes/nodes_veo2.py | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/comfy_api_nodes/nodes_veo2.py b/comfy_api_nodes/nodes_veo2.py index e165b8380..13a6bfd91 100644 --- a/comfy_api_nodes/nodes_veo2.py +++ b/comfy_api_nodes/nodes_veo2.py @@ -168,6 +168,8 @@ class VeoVideoGenerationNode(IO.ComfyNode): # Only add generateAudio for Veo 3 models if model.find("veo-2.0") == -1: parameters["generateAudio"] = generate_audio + # force "enhance_prompt" to True for Veo3 models + parameters["enhancePrompt"] = True initial_response = await sync_op( cls, @@ -291,7 +293,7 @@ class Veo3VideoGenerationNode(VeoVideoGenerationNode): IO.Boolean.Input( "enhance_prompt", default=True, - tooltip="Whether to enhance the prompt with AI assistance", + tooltip="This parameter is deprecated and ignored.", optional=True, ), IO.Combo.Input( From 36deef2c57eacb5d847bd709c5f3068190630612 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Sat, 27 Dec 2025 05:56:52 +0200 Subject: [PATCH 347/647] chore(api-nodes): switch to credits instead of $ (#11489) --- comfy_api_nodes/util/client.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/comfy_api_nodes/util/client.py b/comfy_api_nodes/util/client.py index bf37cba5f..f372ec7b5 100644 --- a/comfy_api_nodes/util/client.py +++ b/comfy_api_nodes/util/client.py @@ -430,9 +430,9 @@ def _display_text( if status: display_lines.append(f"Status: {status.capitalize() if isinstance(status, str) else status}") if price is not None: - p = f"{float(price):,.4f}".rstrip("0").rstrip(".") + p = f"{float(price) * 211:,.1f}".rstrip("0").rstrip(".") if p != "0": - display_lines.append(f"Price: ${p}") + display_lines.append(f"Price: {p} credits") if text is not None: display_lines.append(text) if display_lines: From 2943093a5310fc96aa010a3c68c04f7c16f58a9e Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sat, 27 Dec 2025 15:54:15 -0800 Subject: [PATCH 348/647] Enable async offload by default for AMD. (#11534) --- comfy/model_management.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/comfy/model_management.py b/comfy/model_management.py index 1889ab0ac..e5554e225 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -1019,8 +1019,8 @@ NUM_STREAMS = 0 if args.async_offload is not None: NUM_STREAMS = args.async_offload else: - # Enable by default on Nvidia - if is_nvidia(): + # Enable by default on Nvidia and AMD + if is_nvidia() or is_amd(): NUM_STREAMS = 2 if args.disable_async_offload: From 8fd07170f1b0a7eaaf5a62020cd1926dd3b5092c Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sun, 28 Dec 2025 19:07:25 -0800 Subject: [PATCH 349/647] Comment out unused norm_final in lumina/z image model. (#11545) --- comfy/ldm/lumina/model.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/comfy/ldm/lumina/model.py b/comfy/ldm/lumina/model.py index e80b1c138..afbab2ac7 100644 --- a/comfy/ldm/lumina/model.py +++ b/comfy/ldm/lumina/model.py @@ -491,7 +491,8 @@ class NextDiT(nn.Module): for layer_id in range(n_layers) ] ) - self.norm_final = operation_settings.get("operations").RMSNorm(dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) + # This norm final is in the lumina 2.0 code but isn't actually used for anything. + # self.norm_final = operation_settings.get("operations").RMSNorm(dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) self.final_layer = FinalLayer(dim, patch_size, self.out_channels, z_image_modulation=z_image_modulation, operation_settings=operation_settings) if self.pad_tokens_multiple is not None: From 9ca7e143afe6f09734c9aefcc85f491c5c0dc6e0 Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Mon, 29 Dec 2025 15:19:34 -0800 Subject: [PATCH 350/647] mm: discard async errors from pinning failures (#10738) Pretty much every error cudaHostRegister can throw also queues the same error on the async GPU queue. This was fixed for repinning error case, but there is the bad mmap and just enomem cases that are harder to detect. Do some dummy GPU work to clean the error state. --- comfy/model_management.py | 14 ++++++++++++++ 1 file changed, 14 insertions(+) diff --git a/comfy/model_management.py b/comfy/model_management.py index e5554e225..9fcb699bc 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -1126,6 +1126,16 @@ if not args.disable_pinned_memory: PINNING_ALLOWED_TYPES = set(["Parameter", "QuantizedTensor"]) +def discard_cuda_async_error(): + try: + a = torch.tensor([1], dtype=torch.uint8, device=get_torch_device()) + b = torch.tensor([1], dtype=torch.uint8, device=get_torch_device()) + _ = a + b + torch.cuda.synchronize() + except torch.AcceleratorError: + #Dump it! We already know about it from the synchronous return + pass + def pin_memory(tensor): global TOTAL_PINNED_MEMORY if MAX_PINNED_MEMORY <= 0: @@ -1158,6 +1168,8 @@ def pin_memory(tensor): PINNED_MEMORY[ptr] = size TOTAL_PINNED_MEMORY += size return True + else: + discard_cuda_async_error() return False @@ -1186,6 +1198,8 @@ def unpin_memory(tensor): if len(PINNED_MEMORY) == 0: TOTAL_PINNED_MEMORY = 0 return True + else: + discard_cuda_async_error() return False From 0e6221cc79a3f3cbf0e15a8321bfe75fcffbe667 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 29 Dec 2025 15:26:42 -0800 Subject: [PATCH 351/647] Add some warnings for pin and unpin errors. (#11561) --- comfy/model_management.py | 2 ++ 1 file changed, 2 insertions(+) diff --git a/comfy/model_management.py b/comfy/model_management.py index 9fcb699bc..87baedd73 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -1169,6 +1169,7 @@ def pin_memory(tensor): TOTAL_PINNED_MEMORY += size return True else: + logging.warning("Pin error.") discard_cuda_async_error() return False @@ -1199,6 +1200,7 @@ def unpin_memory(tensor): TOTAL_PINNED_MEMORY = 0 return True else: + logging.warning("Unpin error.") discard_cuda_async_error() return False From d7111e426a48127a97922227b03d31391eb4eba2 Mon Sep 17 00:00:00 2001 From: Tavi Halperin Date: Tue, 30 Dec 2025 03:07:29 +0200 Subject: [PATCH 352/647] ResizeByLongerSide: support video (#11555) (cherry picked from commit 98c6840aa4e5fd5407ba9ab113d209011e474bf6) --- comfy_extras/nodes_dataset.py | 23 +++++++++++++---------- 1 file changed, 13 insertions(+), 10 deletions(-) diff --git a/comfy_extras/nodes_dataset.py b/comfy_extras/nodes_dataset.py index 513aecf3a..5ef851bd0 100644 --- a/comfy_extras/nodes_dataset.py +++ b/comfy_extras/nodes_dataset.py @@ -667,16 +667,19 @@ class ResizeImagesByLongerEdgeNode(ImageProcessingNode): @classmethod def _process(cls, image, longer_edge): - img = tensor_to_pil(image) - w, h = img.size - if w > h: - new_w = longer_edge - new_h = int(h * (longer_edge / w)) - else: - new_h = longer_edge - new_w = int(w * (longer_edge / h)) - img = img.resize((new_w, new_h), Image.Resampling.LANCZOS) - return pil_to_tensor(img) + resized_images = [] + for image_i in image: + img = tensor_to_pil(image_i) + w, h = img.size + if w > h: + new_w = longer_edge + new_h = int(h * (longer_edge / w)) + else: + new_h = longer_edge + new_w = int(w * (longer_edge / h)) + img = img.resize((new_w, new_h), Image.Resampling.LANCZOS) + resized_images.append(pil_to_tensor(img)) + return torch.cat(resized_images, dim=0) class CenterCropImagesNode(ImageProcessingNode): From 25a1bfab4e19b541c2bd6f253a3b83886fb660a1 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Tue, 30 Dec 2025 18:33:34 +0200 Subject: [PATCH 353/647] chore(api-nodes-bytedance): mark "seededit" as deprecated, adjust display name of Seedream (#11490) --- comfy_api_nodes/nodes_bytedance.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/comfy_api_nodes/nodes_bytedance.py b/comfy_api_nodes/nodes_bytedance.py index 636cc1265..d4a2cfae6 100644 --- a/comfy_api_nodes/nodes_bytedance.py +++ b/comfy_api_nodes/nodes_bytedance.py @@ -229,6 +229,7 @@ class ByteDanceImageEditNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + is_deprecated=True, ) @classmethod @@ -269,7 +270,7 @@ class ByteDanceSeedreamNode(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="ByteDanceSeedreamNode", - display_name="ByteDance Seedream 4", + display_name="ByteDance Seedream 4.5", category="api node/image/ByteDance", description="Unified text-to-image generation and precise single-sentence editing at up to 4K resolution.", inputs=[ From 178bdc5e14ec0a55e401c509719e33773cc9b565 Mon Sep 17 00:00:00 2001 From: drozbay <17261091+drozbay@users.noreply.github.com> Date: Tue, 30 Dec 2025 15:40:42 -0700 Subject: [PATCH 354/647] Add handling for vace_context in context windows (#11386) Co-authored-by: ozbayb <17261091+ozbayb@users.noreply.github.com> --- comfy/context_windows.py | 6 ++++++ 1 file changed, 6 insertions(+) diff --git a/comfy/context_windows.py b/comfy/context_windows.py index 1e0f86026..2f82d51da 100644 --- a/comfy/context_windows.py +++ b/comfy/context_windows.py @@ -188,6 +188,12 @@ class IndexListContextHandler(ContextHandlerABC): audio_cond = cond_value.cond if audio_cond.ndim > 1 and audio_cond.size(1) == x_in.size(self.dim): new_cond_item[cond_key] = cond_value._copy_with(window.get_tensor(audio_cond, device, dim=1)) + # Handle vace_context (temporal dim is 3) + elif cond_key == "vace_context" and hasattr(cond_value, "cond") and isinstance(cond_value.cond, torch.Tensor): + vace_cond = cond_value.cond + if vace_cond.ndim >= 4 and vace_cond.size(3) == x_in.size(self.dim): + sliced_vace = window.get_tensor(vace_cond, device, dim=3, retain_index_list=self.cond_retain_index_list) + new_cond_item[cond_key] = cond_value._copy_with(sliced_vace) # if has cond that is a Tensor, check if needs to be subset elif hasattr(cond_value, "cond") and isinstance(cond_value.cond, torch.Tensor): if (self.dim < cond_value.cond.ndim and cond_value.cond.size(self.dim) == x_in.size(self.dim)) or \ From f59f71cf34067d46713f6243312f7f0b360d061f Mon Sep 17 00:00:00 2001 From: comfyanonymous Date: Tue, 30 Dec 2025 22:41:22 -0500 Subject: [PATCH 355/647] ComfyUI version v0.7.0 --- comfyui_version.py | 2 +- pyproject.toml | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/comfyui_version.py b/comfyui_version.py index 1f28e2407..1ed60fe5c 100644 --- a/comfyui_version.py +++ b/comfyui_version.py @@ -1,3 +1,3 @@ # This file is automatically generated by the build process when version is # updated in pyproject.toml. -__version__ = "0.6.0" +__version__ = "0.7.0" diff --git a/pyproject.toml b/pyproject.toml index 35a268bd1..bc1467941 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -1,6 +1,6 @@ [project] name = "ComfyUI" -version = "0.6.0" +version = "0.7.0" readme = "README.md" license = { file = "LICENSE" } requires-python = ">=3.9" From 0357ed7ec4a1bbfe3832874ad6cfc1ca3db1bc0d Mon Sep 17 00:00:00 2001 From: mengqin Date: Tue, 30 Dec 2025 17:53:52 -1000 Subject: [PATCH 356/647] Add support for sage attention 3 in comfyui, enable via new cli arg (#11026) * Add support for sage attention 3 in comfyui, enable via new cli arg --use-sage-attiention3 * Fix some bugs found in PR review. The N dimension at which Sage Attention 3 takes effect is reduced to 1024 (although the improvement is not significant at this scale). * Remove the Sage Attention3 switch, but retain the attention function registration. * Fix a ruff check issue in attention.py --- comfy/ldm/modules/attention.py | 96 ++++++++++++++++++++++++++++++++++ 1 file changed, 96 insertions(+) diff --git a/comfy/ldm/modules/attention.py b/comfy/ldm/modules/attention.py index a8800ded0..ccf690945 100644 --- a/comfy/ldm/modules/attention.py +++ b/comfy/ldm/modules/attention.py @@ -30,6 +30,13 @@ except ImportError as e: raise e exit(-1) +SAGE_ATTENTION3_IS_AVAILABLE = False +try: + from sageattn3 import sageattn3_blackwell + SAGE_ATTENTION3_IS_AVAILABLE = True +except ImportError: + pass + FLASH_ATTENTION_IS_AVAILABLE = False try: from flash_attn import flash_attn_func @@ -563,6 +570,93 @@ def attention_sage(q, k, v, heads, mask=None, attn_precision=None, skip_reshape= out = out.reshape(b, -1, heads * dim_head) return out +@wrap_attn +def attention3_sage(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False, **kwargs): + exception_fallback = False + if (q.device.type != "cuda" or + q.dtype not in (torch.float16, torch.bfloat16) or + mask is not None): + return attention_pytorch( + q, k, v, heads, + mask=mask, + attn_precision=attn_precision, + skip_reshape=skip_reshape, + skip_output_reshape=skip_output_reshape, + **kwargs + ) + + if skip_reshape: + B, H, L, D = q.shape + if H != heads: + return attention_pytorch( + q, k, v, heads, + mask=mask, + attn_precision=attn_precision, + skip_reshape=True, + skip_output_reshape=skip_output_reshape, + **kwargs + ) + q_s, k_s, v_s = q, k, v + N = q.shape[2] + dim_head = D + else: + B, N, inner_dim = q.shape + if inner_dim % heads != 0: + return attention_pytorch( + q, k, v, heads, + mask=mask, + attn_precision=attn_precision, + skip_reshape=False, + skip_output_reshape=skip_output_reshape, + **kwargs + ) + dim_head = inner_dim // heads + + if dim_head >= 256 or N <= 1024: + return attention_pytorch( + q, k, v, heads, + mask=mask, + attn_precision=attn_precision, + skip_reshape=skip_reshape, + skip_output_reshape=skip_output_reshape, + **kwargs + ) + + if not skip_reshape: + q_s, k_s, v_s = map( + lambda t: t.view(B, -1, heads, dim_head).permute(0, 2, 1, 3).contiguous(), + (q, k, v), + ) + B, H, L, D = q_s.shape + + try: + out = sageattn3_blackwell(q_s, k_s, v_s, is_causal=False) + except Exception as e: + exception_fallback = True + logging.error("Error running SageAttention3: %s, falling back to pytorch attention.", e) + + if exception_fallback: + if not skip_reshape: + del q_s, k_s, v_s + return attention_pytorch( + q, k, v, heads, + mask=mask, + attn_precision=attn_precision, + skip_reshape=False, + skip_output_reshape=skip_output_reshape, + **kwargs + ) + + if skip_reshape: + if not skip_output_reshape: + out = out.permute(0, 2, 1, 3).reshape(B, L, H * D) + else: + if skip_output_reshape: + pass + else: + out = out.permute(0, 2, 1, 3).reshape(B, L, H * D) + + return out try: @torch.library.custom_op("flash_attention::flash_attn", mutates_args=()) @@ -650,6 +744,8 @@ optimized_attention_masked = optimized_attention # register core-supported attention functions if SAGE_ATTENTION_IS_AVAILABLE: register_attention_function("sage", attention_sage) +if SAGE_ATTENTION3_IS_AVAILABLE: + register_attention_function("sage3", attention3_sage) if FLASH_ATTENTION_IS_AVAILABLE: register_attention_function("flash", attention_flash) if model_management.xformers_enabled(): From 0be8a76c933026011098d41e61cc6e544739e427 Mon Sep 17 00:00:00 2001 From: Jedrzej Kosinski Date: Tue, 30 Dec 2025 20:09:55 -0800 Subject: [PATCH 357/647] V3 Improvements + DynamicCombo + Autogrow exposed in public API (#11345) * Support Combo outputs in a more sane way * Remove test validate_inputs function on test node * Make curr_prefix be a list of strings instead of string for easier parsing as keys get added to dynamic types * Start to account for id prefixes from frontend, need to fix bug with nested dynamics * Ensure inputs/outputs/hidden are lists in schema finalize function, remove no longer needed 'is not None' checks * Add raw_link and extra_dict to all relevant Inputs * Make nested DynamicCombos work properly with prefixed keys on latest frontend; breaks old Autogrow, but is pretty much ready for upcoming Autogrow keys * Replace ... usage with a MISSING sentinel for clarity in nodes_logic.py * Added CustomCombo node in backend to reflect frontend node * Prepare Autogrow's expand_schema_for_dynamic to work with upcoming frontend changes * Prepare for look up table for dynamic input stuff * More progress towards dynamic input lookup function stuff * Finished converting _expand_schema_for_dynamic to be done via lookup instead of OOP to guarantee working with process isolation, did refactoring to remove old implementation + cleaning INPUT_TYPES definition including v3 hidden definition * Change order of functions * Removed some unneeded functions after dynamic refactor * Make MatchType's output default displayname "MATCHTYPE" * Fix DynamicSlot get_all * Removed redundant code - dynamic stuff no longer happens in OOP way * Natively support AnyType (*) without __ne__ hacks * Remove stray code that made it in * Remove expand_schema_for_dynamic left over on DynamicInput class * get_dynamic() on DynamicInput/Output was not doing anything anymore, so removed it * Make validate_inputs validate combo input correctly * Temporarily comment out conversion to 'new' (9 month old) COMBO format in get_input_info * Remove refrences to resources feature scrapped from V3 * Expose DynamicCombo in public API * satisfy ruff after some code got commented out * Make missing input error prettier for dynamic types * Created a Switch2 node as a side-by-side test, will likely go with Switch2 as the initial switch node * Figured out Switch situation * Pass in v3_data in IsChangedCache.get function's fingerprint_inputs, add a from_v3_data helper method to HiddenHolder * Switch order of Switch and Soft Switch nodes in file * Temp test node for MatchType * Fix missing v3_data for v1 nodes in validation * For now, remove chacking duplicate id's for dynamic types * Add Resize Image/Mask node that thanks to MatchType+DynamicCombo is 16-nodes-in-1 * Made DynamicCombo references in DCTestNode use public interface * Add an AnyTypeTestNode * Make lazy status for specific inputs on DynamicInputs work by having the values of the dictionary for check_lazy_status be a tuple, where the second element is the key of the input that can be returned * Comment out test logic nodes * Make primitive float's step make more sense * Add (and leave commented out) some potential logic nodes * Change default crop option to "center" on Resize Image/Mask node * Changed copy.copy(d) to d.copy() * Autogrow is available in stable frontend, so exposing it in public API * Use outputs id as display_name if no display_name present, remove v3 outputs id restriction that made them have to have unique IDs from the inputs * Enable Custom Combo node as stable frontend now supports it * Make id properly act like display_name on outputs * Add Batch Images/Masks/Latents node * Comment out Batch Images/Masks/Latents node for now, as Autogrow has a bug with MatchType where top connection is disconnected upon refresh * Removed code for a couple test nodes in nodes_logic.py * Add Batch Images, Batch Masks, and Batch Latents nodes with Autogrow, deprecate old Batch Images + LatentBatch nodes --- comfy_api/latest/__init__.py | 1 - comfy_api/latest/_io.py | 370 ++++++++++++++------------ comfy_api/latest/_resources.py | 72 ----- comfy_execution/graph.py | 5 + comfy_execution/validation.py | 14 + comfy_extras/nodes_latent.py | 1 + comfy_extras/nodes_logic.py | 149 +++++++++-- comfy_extras/nodes_post_processing.py | 356 +++++++++++++++++++++++++ comfy_extras/nodes_primitive.py | 2 +- execution.py | 45 ++-- nodes.py | 1 + 11 files changed, 742 insertions(+), 274 deletions(-) delete mode 100644 comfy_api/latest/_resources.py diff --git a/comfy_api/latest/__init__.py b/comfy_api/latest/__init__.py index fab63c7df..b0fa14ff6 100644 --- a/comfy_api/latest/__init__.py +++ b/comfy_api/latest/__init__.py @@ -10,7 +10,6 @@ from ._input_impl import VideoFromFile, VideoFromComponents from ._util import VideoCodec, VideoContainer, VideoComponents, MESH, VOXEL from . import _io_public as io from . import _ui_public as ui -# from comfy_api.latest._resources import _RESOURCES as resources #noqa: F401 from comfy_execution.utils import get_executing_context from comfy_execution.progress import get_progress_state, PreviewImageTuple from PIL import Image diff --git a/comfy_api/latest/_io.py b/comfy_api/latest/_io.py index ba0b95498..764fa8b2b 100644 --- a/comfy_api/latest/_io.py +++ b/comfy_api/latest/_io.py @@ -26,7 +26,6 @@ if TYPE_CHECKING: from comfy_api.input import VideoInput from comfy_api.internal import (_ComfyNodeInternal, _NodeOutputInternal, classproperty, copy_class, first_real_override, is_class, prune_dict, shallow_clone_class) -from ._resources import Resources, ResourcesLocal from comfy_execution.graph_utils import ExecutionBlocker from ._util import MESH, VOXEL, SVG as _SVG @@ -76,16 +75,6 @@ class NumberDisplay(str, Enum): slider = "slider" -class _StringIOType(str): - def __ne__(self, value: object) -> bool: - if self == "*" or value == "*": - return False - if not isinstance(value, str): - return True - a = frozenset(self.split(",")) - b = frozenset(value.split(",")) - return not (b.issubset(a) or a.issubset(b)) - class _ComfyType(ABC): Type = Any io_type: str = None @@ -125,8 +114,7 @@ def comfytype(io_type: str, **kwargs): new_cls.__module__ = cls.__module__ new_cls.__doc__ = cls.__doc__ # assign ComfyType attributes, if needed - # NOTE: use __ne__ trick for io_type (see node_typing.IO.__ne__ for details) - new_cls.io_type = _StringIOType(io_type) + new_cls.io_type = io_type if hasattr(new_cls, "Input") and new_cls.Input is not None: new_cls.Input.Parent = new_cls if hasattr(new_cls, "Output") and new_cls.Output is not None: @@ -165,7 +153,7 @@ class Input(_IO_V3): ''' Base class for a V3 Input. ''' - def __init__(self, id: str, display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, extra_dict=None): + def __init__(self, id: str, display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, extra_dict=None, raw_link: bool=None): super().__init__() self.id = id self.display_name = display_name @@ -173,6 +161,7 @@ class Input(_IO_V3): self.tooltip = tooltip self.lazy = lazy self.extra_dict = extra_dict if extra_dict is not None else {} + self.rawLink = raw_link def as_dict(self): return prune_dict({ @@ -180,10 +169,11 @@ class Input(_IO_V3): "optional": self.optional, "tooltip": self.tooltip, "lazy": self.lazy, + "rawLink": self.rawLink, }) | prune_dict(self.extra_dict) def get_io_type(self): - return _StringIOType(self.io_type) + return self.io_type def get_all(self) -> list[Input]: return [self] @@ -194,8 +184,8 @@ class WidgetInput(Input): ''' def __init__(self, id: str, display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, default: Any=None, - socketless: bool=None, widget_type: str=None, force_input: bool=None, extra_dict=None): - super().__init__(id, display_name, optional, tooltip, lazy, extra_dict) + socketless: bool=None, widget_type: str=None, force_input: bool=None, extra_dict=None, raw_link: bool=None): + super().__init__(id, display_name, optional, tooltip, lazy, extra_dict, raw_link) self.default = default self.socketless = socketless self.widget_type = widget_type @@ -217,13 +207,14 @@ class Output(_IO_V3): def __init__(self, id: str=None, display_name: str=None, tooltip: str=None, is_output_list=False): self.id = id - self.display_name = display_name + self.display_name = display_name if display_name else id self.tooltip = tooltip self.is_output_list = is_output_list def as_dict(self): + display_name = self.display_name if self.display_name else self.id return prune_dict({ - "display_name": self.display_name, + "display_name": display_name, "tooltip": self.tooltip, "is_output_list": self.is_output_list, }) @@ -251,8 +242,8 @@ class Boolean(ComfyTypeIO): '''Boolean input.''' def __init__(self, id: str, display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, default: bool=None, label_on: str=None, label_off: str=None, - socketless: bool=None, force_input: bool=None): - super().__init__(id, display_name, optional, tooltip, lazy, default, socketless, None, force_input) + socketless: bool=None, force_input: bool=None, extra_dict=None, raw_link: bool=None): + super().__init__(id, display_name, optional, tooltip, lazy, default, socketless, None, force_input, extra_dict, raw_link) self.label_on = label_on self.label_off = label_off self.default: bool @@ -271,8 +262,8 @@ class Int(ComfyTypeIO): '''Integer input.''' def __init__(self, id: str, display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, default: int=None, min: int=None, max: int=None, step: int=None, control_after_generate: bool=None, - display_mode: NumberDisplay=None, socketless: bool=None, force_input: bool=None): - super().__init__(id, display_name, optional, tooltip, lazy, default, socketless, None, force_input) + display_mode: NumberDisplay=None, socketless: bool=None, force_input: bool=None, extra_dict=None, raw_link: bool=None): + super().__init__(id, display_name, optional, tooltip, lazy, default, socketless, None, force_input, extra_dict, raw_link) self.min = min self.max = max self.step = step @@ -297,8 +288,8 @@ class Float(ComfyTypeIO): '''Float input.''' def __init__(self, id: str, display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, default: float=None, min: float=None, max: float=None, step: float=None, round: float=None, - display_mode: NumberDisplay=None, socketless: bool=None, force_input: bool=None): - super().__init__(id, display_name, optional, tooltip, lazy, default, socketless, None, force_input) + display_mode: NumberDisplay=None, socketless: bool=None, force_input: bool=None, extra_dict=None, raw_link: bool=None): + super().__init__(id, display_name, optional, tooltip, lazy, default, socketless, None, force_input, extra_dict, raw_link) self.min = min self.max = max self.step = step @@ -323,8 +314,8 @@ class String(ComfyTypeIO): '''String input.''' def __init__(self, id: str, display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, multiline=False, placeholder: str=None, default: str=None, dynamic_prompts: bool=None, - socketless: bool=None, force_input: bool=None): - super().__init__(id, display_name, optional, tooltip, lazy, default, socketless, None, force_input) + socketless: bool=None, force_input: bool=None, extra_dict=None, raw_link: bool=None): + super().__init__(id, display_name, optional, tooltip, lazy, default, socketless, None, force_input, extra_dict, raw_link) self.multiline = multiline self.placeholder = placeholder self.dynamic_prompts = dynamic_prompts @@ -357,12 +348,14 @@ class Combo(ComfyTypeIO): image_folder: FolderType=None, remote: RemoteOptions=None, socketless: bool=None, + extra_dict=None, + raw_link: bool=None, ): if isinstance(options, type) and issubclass(options, Enum): options = [v.value for v in options] if isinstance(default, Enum): default = default.value - super().__init__(id, display_name, optional, tooltip, lazy, default, socketless) + super().__init__(id, display_name, optional, tooltip, lazy, default, socketless, None, None, extra_dict, raw_link) self.multiselect = False self.options = options self.control_after_generate = control_after_generate @@ -386,10 +379,6 @@ class Combo(ComfyTypeIO): super().__init__(id, display_name, tooltip, is_output_list) self.options = options if options is not None else [] - @property - def io_type(self): - return self.options - @comfytype(io_type="COMBO") class MultiCombo(ComfyTypeI): '''Multiselect Combo input (dropdown for selecting potentially more than one value).''' @@ -398,8 +387,8 @@ class MultiCombo(ComfyTypeI): class Input(Combo.Input): def __init__(self, id: str, options: list[str], display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, default: list[str]=None, placeholder: str=None, chip: bool=None, control_after_generate: bool=None, - socketless: bool=None): - super().__init__(id, options, display_name, optional, tooltip, lazy, default, control_after_generate, socketless=socketless) + socketless: bool=None, extra_dict=None, raw_link: bool=None): + super().__init__(id, options, display_name, optional, tooltip, lazy, default, control_after_generate, socketless=socketless, extra_dict=extra_dict, raw_link=raw_link) self.multiselect = True self.placeholder = placeholder self.chip = chip @@ -432,9 +421,9 @@ class Webcam(ComfyTypeIO): Type = str def __init__( self, id: str, display_name: str=None, optional=False, - tooltip: str=None, lazy: bool=None, default: str=None, socketless: bool=None + tooltip: str=None, lazy: bool=None, default: str=None, socketless: bool=None, extra_dict=None, raw_link: bool=None ): - super().__init__(id, display_name, optional, tooltip, lazy, default, socketless) + super().__init__(id, display_name, optional, tooltip, lazy, default, socketless, None, None, extra_dict, raw_link) @comfytype(io_type="MASK") @@ -787,7 +776,7 @@ class MultiType: ''' Input that permits more than one input type; if `id` is an instance of `ComfyType.Input`, then that input will be used to create a widget (if applicable) with overridden values. ''' - def __init__(self, id: str | Input, types: list[type[_ComfyType] | _ComfyType], display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, extra_dict=None): + def __init__(self, id: str | Input, types: list[type[_ComfyType] | _ComfyType], display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, extra_dict=None, raw_link: bool=None): # if id is an Input, then use that Input with overridden values self.input_override = None if isinstance(id, Input): @@ -800,7 +789,7 @@ class MultiType: # if is a widget input, make sure widget_type is set appropriately if isinstance(self.input_override, WidgetInput): self.input_override.widget_type = self.input_override.get_io_type() - super().__init__(id, display_name, optional, tooltip, lazy, extra_dict) + super().__init__(id, display_name, optional, tooltip, lazy, extra_dict, raw_link) self._io_types = types @property @@ -854,8 +843,8 @@ class MatchType(ComfyTypeIO): class Input(Input): def __init__(self, id: str, template: MatchType.Template, - display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, extra_dict=None): - super().__init__(id, display_name, optional, tooltip, lazy, extra_dict) + display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, extra_dict=None, raw_link: bool=None): + super().__init__(id, display_name, optional, tooltip, lazy, extra_dict, raw_link) self.template = template def as_dict(self): @@ -866,6 +855,8 @@ class MatchType(ComfyTypeIO): class Output(Output): def __init__(self, template: MatchType.Template, id: str=None, display_name: str=None, tooltip: str=None, is_output_list=False): + if not id and not display_name: + display_name = "MATCHTYPE" super().__init__(id, display_name, tooltip, is_output_list) self.template = template @@ -878,24 +869,30 @@ class DynamicInput(Input, ABC): ''' Abstract class for dynamic input registration. ''' - def get_dynamic(self) -> list[Input]: - return [] - - def expand_schema_for_dynamic(self, d: dict[str, Any], live_inputs: dict[str, Any], curr_prefix=''): - pass + pass class DynamicOutput(Output, ABC): ''' Abstract class for dynamic output registration. ''' - def __init__(self, id: str=None, display_name: str=None, tooltip: str=None, - is_output_list=False): - super().__init__(id, display_name, tooltip, is_output_list) + pass - def get_dynamic(self) -> list[Output]: - return [] +def handle_prefix(prefix_list: list[str] | None, id: str | None = None) -> list[str]: + if prefix_list is None: + prefix_list = [] + if id is not None: + prefix_list = prefix_list + [id] + return prefix_list + +def finalize_prefix(prefix_list: list[str] | None, id: str | None = None) -> str: + assert not (prefix_list is None and id is None) + if prefix_list is None: + return id + elif id is not None: + prefix_list = prefix_list + [id] + return ".".join(prefix_list) @comfytype(io_type="COMFY_AUTOGROW_V3") class Autogrow(ComfyTypeI): @@ -932,14 +929,6 @@ class Autogrow(ComfyTypeI): def validate(self): self.input.validate() - def expand_schema_for_dynamic(self, d: dict[str, Any], live_inputs: dict[str, Any], curr_prefix=''): - real_inputs = [] - for name, input in self.cached_inputs.items(): - if name in live_inputs: - real_inputs.append(input) - add_to_input_dict_v1(d, real_inputs, live_inputs, curr_prefix) - add_dynamic_id_mapping(d, real_inputs, curr_prefix) - class TemplatePrefix(_AutogrowTemplate): def __init__(self, input: Input, prefix: str, min: int=1, max: int=10): super().__init__(input) @@ -984,22 +973,45 @@ class Autogrow(ComfyTypeI): "template": self.template.as_dict(), }) - def get_dynamic(self) -> list[Input]: - return self.template.get_all() - def get_all(self) -> list[Input]: return [self] + self.template.get_all() def validate(self): self.template.validate() - def expand_schema_for_dynamic(self, d: dict[str, Any], live_inputs: dict[str, Any], curr_prefix=''): - curr_prefix = f"{curr_prefix}{self.id}." - # need to remove self from expected inputs dictionary; replaced by template inputs in frontend - for inner_dict in d.values(): - if self.id in inner_dict: - del inner_dict[self.id] - self.template.expand_schema_for_dynamic(d, live_inputs, curr_prefix) + @staticmethod + def _expand_schema_for_dynamic(out_dict: dict[str, Any], live_inputs: dict[str, Any], value: tuple[str, dict[str, Any]], input_type: str, curr_prefix: list[str] | None): + # NOTE: purposely do not include self in out_dict; instead use only the template inputs + # need to figure out names based on template type + is_names = ("names" in value[1]["template"]) + is_prefix = ("prefix" in value[1]["template"]) + input = value[1]["template"]["input"] + if is_names: + min = value[1]["template"]["min"] + names = value[1]["template"]["names"] + max = len(names) + elif is_prefix: + prefix = value[1]["template"]["prefix"] + min = value[1]["template"]["min"] + max = value[1]["template"]["max"] + names = [f"{prefix}{i}" for i in range(max)] + # need to create a new input based on the contents of input + template_input = None + for _, dict_input in input.items(): + # for now, get just the first value from dict_input + template_input = list(dict_input.values())[0] + new_dict = {} + for i, name in enumerate(names): + expected_id = finalize_prefix(curr_prefix, name) + if expected_id in live_inputs: + # required + if i < min: + type_dict = new_dict.setdefault("required", {}) + # optional + else: + type_dict = new_dict.setdefault("optional", {}) + type_dict[name] = template_input + parse_class_inputs(out_dict, live_inputs, new_dict, curr_prefix) @comfytype(io_type="COMFY_DYNAMICCOMBO_V3") class DynamicCombo(ComfyTypeI): @@ -1022,23 +1034,6 @@ class DynamicCombo(ComfyTypeI): super().__init__(id, display_name, optional, tooltip, lazy, extra_dict) self.options = options - def expand_schema_for_dynamic(self, d: dict[str, Any], live_inputs: dict[str, Any], curr_prefix=''): - # check if dynamic input's id is in live_inputs - if self.id in live_inputs: - curr_prefix = f"{curr_prefix}{self.id}." - key = live_inputs[self.id] - selected_option = None - for option in self.options: - if option.key == key: - selected_option = option - break - if selected_option is not None: - add_to_input_dict_v1(d, selected_option.inputs, live_inputs, curr_prefix) - add_dynamic_id_mapping(d, selected_option.inputs, curr_prefix, self) - - def get_dynamic(self) -> list[Input]: - return [input for option in self.options for input in option.inputs] - def get_all(self) -> list[Input]: return [self] + [input for option in self.options for input in option.inputs] @@ -1053,6 +1048,24 @@ class DynamicCombo(ComfyTypeI): for input in option.inputs: input.validate() + @staticmethod + def _expand_schema_for_dynamic(out_dict: dict[str, Any], live_inputs: dict[str, Any], value: tuple[str, dict[str, Any]], input_type: str, curr_prefix: list[str] | None): + finalized_id = finalize_prefix(curr_prefix) + if finalized_id in live_inputs: + key = live_inputs[finalized_id] + selected_option = None + # get options from dict + options: list[dict[str, str | dict[str, Any]]] = value[1]["options"] + for option in options: + if option["key"] == key: + selected_option = option + break + if selected_option is not None: + parse_class_inputs(out_dict, live_inputs, selected_option["inputs"], curr_prefix) + # add self to inputs + out_dict[input_type][finalized_id] = value + out_dict["dynamic_paths"][finalized_id] = finalize_prefix(curr_prefix, curr_prefix[-1]) + @comfytype(io_type="COMFY_DYNAMICSLOT_V3") class DynamicSlot(ComfyTypeI): Type = dict[str, Any] @@ -1075,17 +1088,8 @@ class DynamicSlot(ComfyTypeI): self.force_input = True self.slot.force_input = True - def expand_schema_for_dynamic(self, d: dict[str, Any], live_inputs: dict[str, Any], curr_prefix=''): - if self.id in live_inputs: - curr_prefix = f"{curr_prefix}{self.id}." - add_to_input_dict_v1(d, self.inputs, live_inputs, curr_prefix) - add_dynamic_id_mapping(d, [self.slot] + self.inputs, curr_prefix) - - def get_dynamic(self) -> list[Input]: - return [self.slot] + self.inputs - def get_all(self) -> list[Input]: - return [self] + [self.slot] + self.inputs + return [self.slot] + self.inputs def as_dict(self): return super().as_dict() | prune_dict({ @@ -1099,17 +1103,41 @@ class DynamicSlot(ComfyTypeI): for input in self.inputs: input.validate() -def add_dynamic_id_mapping(d: dict[str, Any], inputs: list[Input], curr_prefix: str, self: DynamicInput=None): - dynamic = d.setdefault("dynamic_paths", {}) - if self is not None: - dynamic[self.id] = f"{curr_prefix}{self.id}" - for i in inputs: - if not isinstance(i, DynamicInput): - dynamic[f"{i.id}"] = f"{curr_prefix}{i.id}" + @staticmethod + def _expand_schema_for_dynamic(out_dict: dict[str, Any], live_inputs: dict[str, Any], value: tuple[str, dict[str, Any]], input_type: str, curr_prefix: list[str] | None): + finalized_id = finalize_prefix(curr_prefix) + if finalized_id in live_inputs: + inputs = value[1]["inputs"] + parse_class_inputs(out_dict, live_inputs, inputs, curr_prefix) + # add self to inputs + out_dict[input_type][finalized_id] = value + out_dict["dynamic_paths"][finalized_id] = finalize_prefix(curr_prefix, curr_prefix[-1]) + +DYNAMIC_INPUT_LOOKUP: dict[str, Callable[[dict[str, Any], dict[str, Any], tuple[str, dict[str, Any]], str, list[str] | None], None]] = {} +def register_dynamic_input_func(io_type: str, func: Callable[[dict[str, Any], dict[str, Any], tuple[str, dict[str, Any]], str, list[str] | None], None]): + DYNAMIC_INPUT_LOOKUP[io_type] = func + +def get_dynamic_input_func(io_type: str) -> Callable[[dict[str, Any], dict[str, Any], tuple[str, dict[str, Any]], str, list[str] | None], None]: + return DYNAMIC_INPUT_LOOKUP[io_type] + +def setup_dynamic_input_funcs(): + # Autogrow.Input + register_dynamic_input_func(Autogrow.io_type, Autogrow._expand_schema_for_dynamic) + # DynamicCombo.Input + register_dynamic_input_func(DynamicCombo.io_type, DynamicCombo._expand_schema_for_dynamic) + # DynamicSlot.Input + register_dynamic_input_func(DynamicSlot.io_type, DynamicSlot._expand_schema_for_dynamic) + +if len(DYNAMIC_INPUT_LOOKUP) == 0: + setup_dynamic_input_funcs() class V3Data(TypedDict): hidden_inputs: dict[str, Any] + 'Dictionary where the keys are the hidden input ids and the values are the values of the hidden inputs.' dynamic_paths: dict[str, Any] + 'Dictionary where the keys are the input ids and the values dictate how to turn the inputs into a nested dictionary.' + create_dynamic_tuple: bool + 'When True, the value of the dynamic input will be in the format (value, path_key).' class HiddenHolder: def __init__(self, unique_id: str, prompt: Any, @@ -1145,6 +1173,10 @@ class HiddenHolder: api_key_comfy_org=d.get(Hidden.api_key_comfy_org, None), ) + @classmethod + def from_v3_data(cls, v3_data: V3Data | None) -> HiddenHolder: + return cls.from_dict(v3_data["hidden_inputs"] if v3_data else None) + class Hidden(str, Enum): ''' Enumerator for requesting hidden variables in nodes. @@ -1250,61 +1282,56 @@ class Schema: - verify ids on inputs and outputs are unique - both internally and in relation to each other ''' nested_inputs: list[Input] = [] - if self.inputs is not None: - for input in self.inputs: + for input in self.inputs: + if not isinstance(input, DynamicInput): nested_inputs.extend(input.get_all()) - input_ids = [i.id for i in nested_inputs] if nested_inputs is not None else [] - output_ids = [o.id for o in self.outputs] if self.outputs is not None else [] + input_ids = [i.id for i in nested_inputs] + output_ids = [o.id for o in self.outputs] input_set = set(input_ids) output_set = set(output_ids) - issues = [] + issues: list[str] = [] # verify ids are unique per list if len(input_set) != len(input_ids): issues.append(f"Input ids must be unique, but {[item for item, count in Counter(input_ids).items() if count > 1]} are not.") if len(output_set) != len(output_ids): issues.append(f"Output ids must be unique, but {[item for item, count in Counter(output_ids).items() if count > 1]} are not.") - # verify ids are unique between lists - intersection = input_set & output_set - if len(intersection) > 0: - issues.append(f"Ids must be unique between inputs and outputs, but {intersection} are not.") if len(issues) > 0: raise ValueError("\n".join(issues)) # validate inputs and outputs - if self.inputs is not None: - for input in self.inputs: - input.validate() - if self.outputs is not None: - for output in self.outputs: - output.validate() + for input in self.inputs: + input.validate() + for output in self.outputs: + output.validate() def finalize(self): """Add hidden based on selected schema options, and give outputs without ids default ids.""" + # ensure inputs, outputs, and hidden are lists + if self.inputs is None: + self.inputs = [] + if self.outputs is None: + self.outputs = [] + if self.hidden is None: + self.hidden = [] # if is an api_node, will need key-related hidden if self.is_api_node: - if self.hidden is None: - self.hidden = [] if Hidden.auth_token_comfy_org not in self.hidden: self.hidden.append(Hidden.auth_token_comfy_org) if Hidden.api_key_comfy_org not in self.hidden: self.hidden.append(Hidden.api_key_comfy_org) # if is an output_node, will need prompt and extra_pnginfo if self.is_output_node: - if self.hidden is None: - self.hidden = [] if Hidden.prompt not in self.hidden: self.hidden.append(Hidden.prompt) if Hidden.extra_pnginfo not in self.hidden: self.hidden.append(Hidden.extra_pnginfo) # give outputs without ids default ids - if self.outputs is not None: - for i, output in enumerate(self.outputs): - if output.id is None: - output.id = f"_{i}_{output.io_type}_" + for i, output in enumerate(self.outputs): + if output.id is None: + output.id = f"_{i}_{output.io_type}_" - def get_v1_info(self, cls, live_inputs: dict[str, Any]=None) -> NodeInfoV1: - # NOTE: live_inputs will not be used anymore very soon and this will be done another way + def get_v1_info(self, cls) -> NodeInfoV1: # get V1 inputs - input = create_input_dict_v1(self.inputs, live_inputs) + input = create_input_dict_v1(self.inputs) if self.hidden: for hidden in self.hidden: input.setdefault("hidden", {})[hidden.name] = (hidden.value,) @@ -1384,33 +1411,54 @@ class Schema: ) return info +def get_finalized_class_inputs(d: dict[str, Any], live_inputs: dict[str, Any], include_hidden=False) -> tuple[dict[str, Any], V3Data]: + out_dict = { + "required": {}, + "optional": {}, + "dynamic_paths": {}, + } + d = d.copy() + # ignore hidden for parsing + hidden = d.pop("hidden", None) + parse_class_inputs(out_dict, live_inputs, d) + if hidden is not None and include_hidden: + out_dict["hidden"] = hidden + v3_data = {} + dynamic_paths = out_dict.pop("dynamic_paths", None) + if dynamic_paths is not None: + v3_data["dynamic_paths"] = dynamic_paths + return out_dict, hidden, v3_data -def create_input_dict_v1(inputs: list[Input], live_inputs: dict[str, Any]=None) -> dict: +def parse_class_inputs(out_dict: dict[str, Any], live_inputs: dict[str, Any], curr_dict: dict[str, Any], curr_prefix: list[str] | None=None) -> None: + for input_type, inner_d in curr_dict.items(): + for id, value in inner_d.items(): + io_type = value[0] + if io_type in DYNAMIC_INPUT_LOOKUP: + # dynamic inputs need to be handled with lookup functions + dynamic_input_func = get_dynamic_input_func(io_type) + new_prefix = handle_prefix(curr_prefix, id) + dynamic_input_func(out_dict, live_inputs, value, input_type, new_prefix) + else: + # non-dynamic inputs get directly transferred + finalized_id = finalize_prefix(curr_prefix, id) + out_dict[input_type][finalized_id] = value + if curr_prefix: + out_dict["dynamic_paths"][finalized_id] = finalized_id + +def create_input_dict_v1(inputs: list[Input]) -> dict: input = { "required": {} } - add_to_input_dict_v1(input, inputs, live_inputs) + for i in inputs: + add_to_dict_v1(i, input) return input -def add_to_input_dict_v1(d: dict[str, Any], inputs: list[Input], live_inputs: dict[str, Any]=None, curr_prefix=''): - for i in inputs: - if isinstance(i, DynamicInput): - add_to_dict_v1(i, d) - if live_inputs is not None: - i.expand_schema_for_dynamic(d, live_inputs, curr_prefix) - else: - add_to_dict_v1(i, d) - -def add_to_dict_v1(i: Input, d: dict, dynamic_dict: dict=None): +def add_to_dict_v1(i: Input, d: dict): key = "optional" if i.optional else "required" as_dict = i.as_dict() # for v1, we don't want to include the optional key as_dict.pop("optional", None) - if dynamic_dict is None: - value = (i.get_io_type(), as_dict) - else: - value = (i.get_io_type(), as_dict, dynamic_dict) - d.setdefault(key, {})[i.id] = value + d.setdefault(key, {})[i.id] = (i.get_io_type(), as_dict) def add_to_dict_v3(io: Input | Output, d: dict): d[io.id] = (io.get_io_type(), io.as_dict()) @@ -1422,6 +1470,8 @@ def build_nested_inputs(values: dict[str, Any], v3_data: V3Data): values = values.copy() result = {} + create_tuple = v3_data.get("create_dynamic_tuple", False) + for key, path in paths.items(): parts = path.split(".") current = result @@ -1430,7 +1480,10 @@ def build_nested_inputs(values: dict[str, Any], v3_data: V3Data): is_last = (i == len(parts) - 1) if is_last: - current[p] = values.pop(key, None) + value = values.pop(key, None) + if create_tuple: + value = (value, key) + current[p] = value else: current = current.setdefault(p, {}) @@ -1445,7 +1498,6 @@ class _ComfyNodeBaseInternal(_ComfyNodeInternal): SCHEMA = None # filled in during execution - resources: Resources = None hidden: HiddenHolder = None @classmethod @@ -1492,7 +1544,6 @@ class _ComfyNodeBaseInternal(_ComfyNodeInternal): return [name for name in kwargs if kwargs[name] is None] def __init__(self): - self.local_resources: ResourcesLocal = None self.__class__.VALIDATE_CLASS() @classmethod @@ -1560,7 +1611,7 @@ class _ComfyNodeBaseInternal(_ComfyNodeInternal): c_type: type[ComfyNode] = cls if is_class(cls) else type(cls) type_clone: type[ComfyNode] = shallow_clone_class(c_type) # set hidden - type_clone.hidden = HiddenHolder.from_dict(v3_data["hidden_inputs"] if v3_data else None) + type_clone.hidden = HiddenHolder.from_v3_data(v3_data) return type_clone @final @@ -1677,19 +1728,10 @@ class _ComfyNodeBaseInternal(_ComfyNodeInternal): @final @classmethod - def INPUT_TYPES(cls, include_hidden=True, return_schema=False, live_inputs=None) -> dict[str, dict] | tuple[dict[str, dict], Schema, V3Data]: + def INPUT_TYPES(cls) -> dict[str, dict]: schema = cls.FINALIZE_SCHEMA() - info = schema.get_v1_info(cls, live_inputs) - input = info.input - if not include_hidden: - input.pop("hidden", None) - if return_schema: - v3_data: V3Data = {} - dynamic = input.pop("dynamic_paths", None) - if dynamic is not None: - v3_data["dynamic_paths"] = dynamic - return input, schema, v3_data - return input + info = schema.get_v1_info(cls) + return info.input @final @classmethod @@ -1808,7 +1850,7 @@ class NodeOutput(_NodeOutputInternal): return self.args if len(self.args) > 0 else None @classmethod - def from_dict(cls, data: dict[str, Any]) -> "NodeOutput": + def from_dict(cls, data: dict[str, Any]) -> NodeOutput: args = () ui = None expand = None @@ -1903,8 +1945,8 @@ __all__ = [ "Tracks", # Dynamic Types "MatchType", - # "DynamicCombo", - # "Autogrow", + "DynamicCombo", + "Autogrow", # Other classes "HiddenHolder", "Hidden", diff --git a/comfy_api/latest/_resources.py b/comfy_api/latest/_resources.py deleted file mode 100644 index a6bdda972..000000000 --- a/comfy_api/latest/_resources.py +++ /dev/null @@ -1,72 +0,0 @@ -from __future__ import annotations -import comfy.utils -import folder_paths -import logging -from abc import ABC, abstractmethod -from typing import Any -import torch - -class ResourceKey(ABC): - Type = Any - def __init__(self): - ... - -class TorchDictFolderFilename(ResourceKey): - '''Key for requesting a torch file via file_name from a folder category.''' - Type = dict[str, torch.Tensor] - def __init__(self, folder_name: str, file_name: str): - self.folder_name = folder_name - self.file_name = file_name - - def __hash__(self): - return hash((self.folder_name, self.file_name)) - - def __eq__(self, other: object) -> bool: - if not isinstance(other, TorchDictFolderFilename): - return False - return self.folder_name == other.folder_name and self.file_name == other.file_name - - def __str__(self): - return f"{self.folder_name} -> {self.file_name}" - -class Resources(ABC): - def __init__(self): - ... - - @abstractmethod - def get(self, key: ResourceKey, default: Any=...) -> Any: - pass - -class ResourcesLocal(Resources): - def __init__(self): - super().__init__() - self.local_resources: dict[ResourceKey, Any] = {} - - def get(self, key: ResourceKey, default: Any=...) -> Any: - cached = self.local_resources.get(key, None) - if cached is not None: - logging.info(f"Using cached resource '{key}'") - return cached - logging.info(f"Loading resource '{key}'") - to_return = None - if isinstance(key, TorchDictFolderFilename): - if default is ...: - to_return = comfy.utils.load_torch_file(folder_paths.get_full_path_or_raise(key.folder_name, key.file_name), safe_load=True) - else: - full_path = folder_paths.get_full_path(key.folder_name, key.file_name) - if full_path is not None: - to_return = comfy.utils.load_torch_file(full_path, safe_load=True) - - if to_return is not None: - self.local_resources[key] = to_return - return to_return - if default is not ...: - return default - raise Exception(f"Unsupported resource key type: {type(key)}") - - -class _RESOURCES: - ResourceKey = ResourceKey - TorchDictFolderFilename = TorchDictFolderFilename - Resources = Resources - ResourcesLocal = ResourcesLocal diff --git a/comfy_execution/graph.py b/comfy_execution/graph.py index 0d811e354..8fc5846b7 100644 --- a/comfy_execution/graph.py +++ b/comfy_execution/graph.py @@ -97,6 +97,11 @@ def get_input_info( extra_info = input_info[1] else: extra_info = {} + # if input_type is a list, it is a Combo defined in outdated format; convert it. + # NOTE: uncomment this when we are confident old format going away won't cause too much trouble. + # if isinstance(input_type, list): + # extra_info["options"] = input_type + # input_type = IO.Combo.io_type return input_type, input_category, extra_info class TopologicalSort: diff --git a/comfy_execution/validation.py b/comfy_execution/validation.py index 24c0b4ed7..e73624bd1 100644 --- a/comfy_execution/validation.py +++ b/comfy_execution/validation.py @@ -21,14 +21,24 @@ def validate_node_input( """ # If the types are exactly the same, we can return immediately # Use pre-union behaviour: inverse of `__ne__` + # NOTE: this lets legacy '*' Any types work that override the __ne__ method of the str class. if not received_type != input_type: return True + # If one of the types is '*', we can return True immediately; this is the 'Any' type. + if received_type == IO.AnyType.io_type or input_type == IO.AnyType.io_type: + return True + # If the received type or input_type is a MatchType, we can return True immediately; # validation for this is handled by the frontend if received_type == IO.MatchType.io_type or input_type == IO.MatchType.io_type: return True + # This accounts for some custom nodes that output lists of options as the type; + # if we ever want to break them on purpose, this can be removed + if isinstance(received_type, list) and input_type == IO.Combo.io_type: + return True + # Not equal, and not strings if not isinstance(received_type, str) or not isinstance(input_type, str): return False @@ -37,6 +47,10 @@ def validate_node_input( received_types = set(t.strip() for t in received_type.split(",")) input_types = set(t.strip() for t in input_type.split(",")) + # If any of the types is '*', we can return True immediately; this is the 'Any' type. + if IO.AnyType.io_type in received_types or IO.AnyType.io_type in input_types: + return True + if strict: # In strict mode, all received types must be in the input types return received_types.issubset(input_types) diff --git a/comfy_extras/nodes_latent.py b/comfy_extras/nodes_latent.py index 2815c5ffc..9ba1c4ba8 100644 --- a/comfy_extras/nodes_latent.py +++ b/comfy_extras/nodes_latent.py @@ -255,6 +255,7 @@ class LatentBatch(io.ComfyNode): return io.Schema( node_id="LatentBatch", category="latent/batch", + is_deprecated=True, inputs=[ io.Latent.Input("samples1"), io.Latent.Input("samples2"), diff --git a/comfy_extras/nodes_logic.py b/comfy_extras/nodes_logic.py index 95a6ba788..eb888316a 100644 --- a/comfy_extras/nodes_logic.py +++ b/comfy_extras/nodes_logic.py @@ -1,8 +1,11 @@ +from __future__ import annotations from typing import TypedDict from typing_extensions import override from comfy_api.latest import ComfyExtension, io from comfy_api.latest import _io +# sentinel for missing inputs +MISSING = object() class SwitchNode(io.ComfyNode): @@ -14,6 +17,37 @@ class SwitchNode(io.ComfyNode): display_name="Switch", category="logic", is_experimental=True, + inputs=[ + io.Boolean.Input("switch"), + io.MatchType.Input("on_false", template=template, lazy=True), + io.MatchType.Input("on_true", template=template, lazy=True), + ], + outputs=[ + io.MatchType.Output(template=template, display_name="output"), + ], + ) + + @classmethod + def check_lazy_status(cls, switch, on_false=None, on_true=None): + if switch and on_true is None: + return ["on_true"] + if not switch and on_false is None: + return ["on_false"] + + @classmethod + def execute(cls, switch, on_true, on_false) -> io.NodeOutput: + return io.NodeOutput(on_true if switch else on_false) + + +class SoftSwitchNode(io.ComfyNode): + @classmethod + def define_schema(cls): + template = io.MatchType.Template("switch") + return io.Schema( + node_id="ComfySoftSwitchNode", + display_name="Soft Switch", + category="logic", + is_experimental=True, inputs=[ io.Boolean.Input("switch"), io.MatchType.Input("on_false", template=template, lazy=True, optional=True), @@ -25,14 +59,14 @@ class SwitchNode(io.ComfyNode): ) @classmethod - def check_lazy_status(cls, switch, on_false=..., on_true=...): - # We use ... instead of None, as None is passed for connected-but-unevaluated inputs. + def check_lazy_status(cls, switch, on_false=MISSING, on_true=MISSING): + # We use MISSING instead of None, as None is passed for connected-but-unevaluated inputs. # This trick allows us to ignore the value of the switch and still be able to run execute(). # One of the inputs may be missing, in which case we need to evaluate the other input - if on_false is ...: + if on_false is MISSING: return ["on_true"] - if on_true is ...: + if on_true is MISSING: return ["on_false"] # Normal lazy switch operation if switch and on_true is None: @@ -41,22 +75,50 @@ class SwitchNode(io.ComfyNode): return ["on_false"] @classmethod - def validate_inputs(cls, switch, on_false=..., on_true=...): + def validate_inputs(cls, switch, on_false=MISSING, on_true=MISSING): # This check happens before check_lazy_status(), so we can eliminate the case where # both inputs are missing. - if on_false is ... and on_true is ...: + if on_false is MISSING and on_true is MISSING: return "At least one of on_false or on_true must be connected to Switch node" return True @classmethod - def execute(cls, switch, on_true=..., on_false=...) -> io.NodeOutput: - if on_true is ...: + def execute(cls, switch, on_true=MISSING, on_false=MISSING) -> io.NodeOutput: + if on_true is MISSING: return io.NodeOutput(on_false) - if on_false is ...: + if on_false is MISSING: return io.NodeOutput(on_true) return io.NodeOutput(on_true if switch else on_false) +class CustomComboNode(io.ComfyNode): + """ + Frontend node that allows user to write their own options for a combo. + This is here to make sure the node has a backend-representation to avoid some annoyances. + """ + @classmethod + def define_schema(cls): + return io.Schema( + node_id="CustomCombo", + display_name="Custom Combo", + category="utils", + is_experimental=True, + inputs=[io.Combo.Input("choice", options=[])], + outputs=[io.String.Output()] + ) + + @classmethod + def validate_inputs(cls, choice: io.Combo.Type) -> bool: + # NOTE: DO NOT DO THIS unless you want to skip validation entirely on the node's inputs. + # I am doing that here because the widgets (besides the combo dropdown) on this node are fully frontend defined. + # I need to skip checking that the chosen combo option is in the options list, since those are defined by the user. + return True + + @classmethod + def execute(cls, choice: io.Combo.Type) -> io.NodeOutput: + return io.NodeOutput(choice) + + class DCTestNode(io.ComfyNode): class DCValues(TypedDict): combo: str @@ -72,14 +134,14 @@ class DCTestNode(io.ComfyNode): display_name="DCTest", category="logic", is_output_node=True, - inputs=[_io.DynamicCombo.Input("combo", options=[ - _io.DynamicCombo.Option("option1", [io.String.Input("string")]), - _io.DynamicCombo.Option("option2", [io.Int.Input("integer")]), - _io.DynamicCombo.Option("option3", [io.Image.Input("image")]), - _io.DynamicCombo.Option("option4", [ - _io.DynamicCombo.Input("subcombo", options=[ - _io.DynamicCombo.Option("opt1", [io.Float.Input("float_x"), io.Float.Input("float_y")]), - _io.DynamicCombo.Option("opt2", [io.Mask.Input("mask1", optional=True)]), + inputs=[io.DynamicCombo.Input("combo", options=[ + io.DynamicCombo.Option("option1", [io.String.Input("string")]), + io.DynamicCombo.Option("option2", [io.Int.Input("integer")]), + io.DynamicCombo.Option("option3", [io.Image.Input("image")]), + io.DynamicCombo.Option("option4", [ + io.DynamicCombo.Input("subcombo", options=[ + io.DynamicCombo.Option("opt1", [io.Float.Input("float_x"), io.Float.Input("float_y")]), + io.DynamicCombo.Option("opt2", [io.Mask.Input("mask1", optional=True)]), ]) ])] )], @@ -141,14 +203,65 @@ class AutogrowPrefixTestNode(io.ComfyNode): combined = ",".join([str(x) for x in vals]) return io.NodeOutput(combined) +class ComboOutputTestNode(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="ComboOptionTestNode", + display_name="ComboOptionTest", + category="logic", + inputs=[io.Combo.Input("combo", options=["option1", "option2", "option3"]), + io.Combo.Input("combo2", options=["option4", "option5", "option6"])], + outputs=[io.Combo.Output(), io.Combo.Output()], + ) + + @classmethod + def execute(cls, combo: io.Combo.Type, combo2: io.Combo.Type) -> io.NodeOutput: + return io.NodeOutput(combo, combo2) + +class ConvertStringToComboNode(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="ConvertStringToComboNode", + display_name="Convert String to Combo", + category="logic", + inputs=[io.String.Input("string")], + outputs=[io.Combo.Output()], + ) + + @classmethod + def execute(cls, string: str) -> io.NodeOutput: + return io.NodeOutput(string) + +class InvertBooleanNode(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="InvertBooleanNode", + display_name="Invert Boolean", + category="logic", + inputs=[io.Boolean.Input("boolean")], + outputs=[io.Boolean.Output()], + ) + + @classmethod + def execute(cls, boolean: bool) -> io.NodeOutput: + return io.NodeOutput(not boolean) + class LogicExtension(ComfyExtension): @override async def get_node_list(self) -> list[type[io.ComfyNode]]: return [ - # SwitchNode, + SwitchNode, + CustomComboNode, + # SoftSwitchNode, + # ConvertStringToComboNode, # DCTestNode, # AutogrowNamesTestNode, # AutogrowPrefixTestNode, + # ComboOutputTestNode, + # InvertBooleanNode, ] async def comfy_entrypoint() -> LogicExtension: diff --git a/comfy_extras/nodes_post_processing.py b/comfy_extras/nodes_post_processing.py index ca2cdeb50..01afa13a1 100644 --- a/comfy_extras/nodes_post_processing.py +++ b/comfy_extras/nodes_post_processing.py @@ -4,11 +4,15 @@ import torch import torch.nn.functional as F from PIL import Image import math +from enum import Enum +from typing import TypedDict, Literal import comfy.utils import comfy.model_management +from comfy_extras.nodes_latent import reshape_latent_to import node_helpers from comfy_api.latest import ComfyExtension, io +from nodes import MAX_RESOLUTION class Blend(io.ComfyNode): @classmethod @@ -241,6 +245,353 @@ class ImageScaleToTotalPixels(io.ComfyNode): s = s.movedim(1,-1) return io.NodeOutput(s) +class ResizeType(str, Enum): + SCALE_BY = "scale by multiplier" + SCALE_DIMENSIONS = "scale dimensions" + SCALE_LONGER_DIMENSION = "scale longer dimension" + SCALE_SHORTER_DIMENSION = "scale shorter dimension" + SCALE_WIDTH = "scale width" + SCALE_HEIGHT = "scale height" + SCALE_TOTAL_PIXELS = "scale total pixels" + MATCH_SIZE = "match size" + +def is_image(input: torch.Tensor) -> bool: + # images have 4 dimensions: [batch, height, width, channels] + # masks have 3 dimensions: [batch, height, width] + return len(input.shape) == 4 + +def init_image_mask_input(input: torch.Tensor, is_type_image: bool) -> torch.Tensor: + if is_type_image: + input = input.movedim(-1, 1) + else: + input = input.unsqueeze(1) + return input + +def finalize_image_mask_input(input: torch.Tensor, is_type_image: bool) -> torch.Tensor: + if is_type_image: + input = input.movedim(1, -1) + else: + input = input.squeeze(1) + return input + +def scale_by(input: torch.Tensor, multiplier: float, scale_method: str) -> torch.Tensor: + is_type_image = is_image(input) + input = init_image_mask_input(input, is_type_image) + width = round(input.shape[-1] * multiplier) + height = round(input.shape[-2] * multiplier) + + input = comfy.utils.common_upscale(input, width, height, scale_method, "disabled") + input = finalize_image_mask_input(input, is_type_image) + return input + +def scale_dimensions(input: torch.Tensor, width: int, height: int, scale_method: str, crop: str="disabled") -> torch.Tensor: + if width == 0 and height == 0: + return input + is_type_image = is_image(input) + input = init_image_mask_input(input, is_type_image) + + if width == 0: + width = max(1, round(input.shape[-1] * height / input.shape[-2])) + elif height == 0: + height = max(1, round(input.shape[-2] * width / input.shape[-1])) + + input = comfy.utils.common_upscale(input, width, height, scale_method, crop) + input = finalize_image_mask_input(input, is_type_image) + return input + +def scale_longer_dimension(input: torch.Tensor, longer_size: int, scale_method: str) -> torch.Tensor: + is_type_image = is_image(input) + input = init_image_mask_input(input, is_type_image) + width = input.shape[-1] + height = input.shape[-2] + + if height > width: + width = round((width / height) * longer_size) + height = longer_size + elif width > height: + height = round((height / width) * longer_size) + width = longer_size + else: + height = longer_size + width = longer_size + + input = comfy.utils.common_upscale(input, width, height, scale_method, "disabled") + input = finalize_image_mask_input(input, is_type_image) + return input + +def scale_shorter_dimension(input: torch.Tensor, shorter_size: int, scale_method: str) -> torch.Tensor: + is_type_image = is_image(input) + input = init_image_mask_input(input, is_type_image) + width = input.shape[-1] + height = input.shape[-2] + + if height < width: + width = round((width / height) * shorter_size) + height = shorter_size + elif width > height: + height = round((height / width) * shorter_size) + width = shorter_size + else: + height = shorter_size + width = shorter_size + + input = comfy.utils.common_upscale(input, width, height, scale_method, "disabled") + input = finalize_image_mask_input(input, is_type_image) + return input + +def scale_total_pixels(input: torch.Tensor, megapixels: float, scale_method: str) -> torch.Tensor: + is_type_image = is_image(input) + input = init_image_mask_input(input, is_type_image) + total = int(megapixels * 1024 * 1024) + + scale_by = math.sqrt(total / (input.shape[-1] * input.shape[-2])) + width = round(input.shape[-1] * scale_by) + height = round(input.shape[-2] * scale_by) + + input = comfy.utils.common_upscale(input, width, height, scale_method, "disabled") + input = finalize_image_mask_input(input, is_type_image) + return input + +def scale_match_size(input: torch.Tensor, match: torch.Tensor, scale_method: str, crop: str) -> torch.Tensor: + is_type_image = is_image(input) + input = init_image_mask_input(input, is_type_image) + match = init_image_mask_input(match, is_image(match)) + + width = match.shape[-1] + height = match.shape[-2] + input = comfy.utils.common_upscale(input, width, height, scale_method, crop) + input = finalize_image_mask_input(input, is_type_image) + return input + +class ResizeImageMaskNode(io.ComfyNode): + + scale_methods = ["nearest-exact", "bilinear", "area", "bicubic", "lanczos"] + crop_methods = ["disabled", "center"] + + class ResizeTypedDict(TypedDict): + resize_type: ResizeType + scale_method: Literal["nearest-exact", "bilinear", "area", "bicubic", "lanczos"] + crop: Literal["disabled", "center"] + multiplier: float + width: int + height: int + longer_size: int + shorter_size: int + megapixels: float + + @classmethod + def define_schema(cls): + template = io.MatchType.Template("input_type", [io.Image, io.Mask]) + crop_combo = io.Combo.Input("crop", options=cls.crop_methods, default="center") + return io.Schema( + node_id="ResizeImageMaskNode", + display_name="Resize Image/Mask", + category="transform", + inputs=[ + io.MatchType.Input("input", template=template), + io.DynamicCombo.Input("resize_type", options=[ + io.DynamicCombo.Option(ResizeType.SCALE_BY, [ + io.Float.Input("multiplier", default=1.00, min=0.01, max=8.0, step=0.01), + ]), + io.DynamicCombo.Option(ResizeType.SCALE_DIMENSIONS, [ + io.Int.Input("width", default=512, min=0, max=MAX_RESOLUTION, step=1), + io.Int.Input("height", default=512, min=0, max=MAX_RESOLUTION, step=1), + crop_combo, + ]), + io.DynamicCombo.Option(ResizeType.SCALE_LONGER_DIMENSION, [ + io.Int.Input("longer_size", default=512, min=0, max=MAX_RESOLUTION, step=1), + ]), + io.DynamicCombo.Option(ResizeType.SCALE_SHORTER_DIMENSION, [ + io.Int.Input("shorter_size", default=512, min=0, max=MAX_RESOLUTION, step=1), + ]), + io.DynamicCombo.Option(ResizeType.SCALE_WIDTH, [ + io.Int.Input("width", default=512, min=0, max=MAX_RESOLUTION, step=1), + ]), + io.DynamicCombo.Option(ResizeType.SCALE_HEIGHT, [ + io.Int.Input("height", default=512, min=0, max=MAX_RESOLUTION, step=1), + ]), + io.DynamicCombo.Option(ResizeType.SCALE_TOTAL_PIXELS, [ + io.Float.Input("megapixels", default=1.0, min=0.01, max=16.0, step=0.01), + ]), + io.DynamicCombo.Option(ResizeType.MATCH_SIZE, [ + io.MultiType.Input("match", [io.Image, io.Mask]), + crop_combo, + ]), + ]), + io.Combo.Input("scale_method", options=cls.scale_methods, default="area"), + ], + outputs=[io.MatchType.Output(template=template, display_name="resized")] + ) + + @classmethod + def execute(cls, input: io.Image.Type | io.Mask.Type, scale_method: io.Combo.Type, resize_type: ResizeTypedDict) -> io.NodeOutput: + selected_type = resize_type["resize_type"] + if selected_type == ResizeType.SCALE_BY: + return io.NodeOutput(scale_by(input, resize_type["multiplier"], scale_method)) + elif selected_type == ResizeType.SCALE_DIMENSIONS: + return io.NodeOutput(scale_dimensions(input, resize_type["width"], resize_type["height"], scale_method, resize_type["crop"])) + elif selected_type == ResizeType.SCALE_LONGER_DIMENSION: + return io.NodeOutput(scale_longer_dimension(input, resize_type["longer_size"], scale_method)) + elif selected_type == ResizeType.SCALE_SHORTER_DIMENSION: + return io.NodeOutput(scale_shorter_dimension(input, resize_type["shorter_size"], scale_method)) + elif selected_type == ResizeType.SCALE_WIDTH: + return io.NodeOutput(scale_dimensions(input, resize_type["width"], 0, scale_method)) + elif selected_type == ResizeType.SCALE_HEIGHT: + return io.NodeOutput(scale_dimensions(input, 0, resize_type["height"], scale_method)) + elif selected_type == ResizeType.SCALE_TOTAL_PIXELS: + return io.NodeOutput(scale_total_pixels(input, resize_type["megapixels"], scale_method)) + elif selected_type == ResizeType.MATCH_SIZE: + return io.NodeOutput(scale_match_size(input, resize_type["match"], scale_method, resize_type["crop"])) + raise ValueError(f"Unsupported resize type: {selected_type}") + +def batch_images(images: list[torch.Tensor]) -> torch.Tensor | None: + if len(images) == 0: + return None + # first, get the max channels count + max_channels = max(image.shape[-1] for image in images) + # then, pad all images to have the same channels count + padded_images: list[torch.Tensor] = [] + for image in images: + if image.shape[-1] < max_channels: + padded_images.append(torch.nn.functional.pad(image, (0,1), mode='constant', value=1.0)) + else: + padded_images.append(image) + # resize all images to be the same size as the first image + resized_images: list[torch.Tensor] = [] + first_image_shape = padded_images[0].shape + for image in padded_images: + if image.shape[1:] != first_image_shape[1:]: + resized_images.append(comfy.utils.common_upscale(image.movedim(-1,1), first_image_shape[2], first_image_shape[1], "bilinear", "center").movedim(1,-1)) + else: + resized_images.append(image) + # batch the images in the format [b, h, w, c] + return torch.cat(resized_images, dim=0) + +def batch_masks(masks: list[torch.Tensor]) -> torch.Tensor | None: + if len(masks) == 0: + return None + # resize all masks to be the same size as the first mask + resized_masks: list[torch.Tensor] = [] + first_mask_shape = masks[0].shape + for mask in masks: + if mask.shape[1:] != first_mask_shape[1:]: + mask = init_image_mask_input(mask, is_type_image=False) + mask = comfy.utils.common_upscale(mask, first_mask_shape[2], first_mask_shape[1], "bilinear", "center") + resized_masks.append(finalize_image_mask_input(mask, is_type_image=False)) + else: + resized_masks.append(mask) + # batch the masks in the format [b, h, w] + return torch.cat(resized_masks, dim=0) + +def batch_latents(latents: list[dict[str, torch.Tensor]]) -> dict[str, torch.Tensor] | None: + if len(latents) == 0: + return None + samples_out = latents[0].copy() + samples_out["batch_index"] = [] + first_samples = latents[0]["samples"] + tensors: list[torch.Tensor] = [] + for latent in latents: + # first, deal with latent tensors + tensors.append(reshape_latent_to(first_samples.shape, latent["samples"], repeat_batch=False)) + # next, deal with batch_index + samples_out["batch_index"].extend(latent.get("batch_index", [x for x in range(0, latent["samples"].shape[0])])) + samples_out["samples"] = torch.cat(tensors, dim=0) + return samples_out + +class BatchImagesNode(io.ComfyNode): + @classmethod + def define_schema(cls): + autogrow_template = io.Autogrow.TemplatePrefix(io.Image.Input("image"), prefix="image", min=2, max=50) + return io.Schema( + node_id="BatchImagesNode", + display_name="Batch Images", + category="image", + inputs=[ + io.Autogrow.Input("images", template=autogrow_template) + ], + outputs=[ + io.Image.Output() + ] + ) + + @classmethod + def execute(cls, images: io.Autogrow.Type) -> io.NodeOutput: + return io.NodeOutput(batch_images(list(images.values()))) + +class BatchMasksNode(io.ComfyNode): + @classmethod + def define_schema(cls): + autogrow_template = io.Autogrow.TemplatePrefix(io.Mask.Input("mask"), prefix="mask", min=2, max=50) + return io.Schema( + node_id="BatchMasksNode", + display_name="Batch Masks", + category="mask", + inputs=[ + io.Autogrow.Input("masks", template=autogrow_template) + ], + outputs=[ + io.Mask.Output() + ] + ) + + @classmethod + def execute(cls, masks: io.Autogrow.Type) -> io.NodeOutput: + return io.NodeOutput(batch_masks(list(masks.values()))) + +class BatchLatentsNode(io.ComfyNode): + @classmethod + def define_schema(cls): + autogrow_template = io.Autogrow.TemplatePrefix(io.Latent.Input("latent"), prefix="latent", min=2, max=50) + return io.Schema( + node_id="BatchLatentsNode", + display_name="Batch Latents", + category="latent", + inputs=[ + io.Autogrow.Input("latents", template=autogrow_template) + ], + outputs=[ + io.Latent.Output() + ] + ) + + @classmethod + def execute(cls, latents: io.Autogrow.Type) -> io.NodeOutput: + return io.NodeOutput(batch_latents(list(latents.values()))) + +class BatchImagesMasksLatentsNode(io.ComfyNode): + @classmethod + def define_schema(cls): + matchtype_template = io.MatchType.Template("input", allowed_types=[io.Image, io.Mask, io.Latent]) + autogrow_template = io.Autogrow.TemplatePrefix( + io.MatchType.Input("input", matchtype_template), + prefix="input", min=1, max=50) + return io.Schema( + node_id="BatchImagesMasksLatentsNode", + display_name="Batch Images/Masks/Latents", + category="util", + inputs=[ + io.Autogrow.Input("inputs", template=autogrow_template) + ], + outputs=[ + io.MatchType.Output(id=None, template=matchtype_template) + ] + ) + + @classmethod + def execute(cls, inputs: io.Autogrow.Type) -> io.NodeOutput: + batched = None + values = list(inputs.values()) + # latents + if isinstance(values[0], dict): + batched = batch_latents(values) + # images + elif is_image(values[0]): + batched = batch_images(values) + # masks + else: + batched = batch_masks(values) + return io.NodeOutput(batched) + class PostProcessingExtension(ComfyExtension): @override async def get_node_list(self) -> list[type[io.ComfyNode]]: @@ -250,6 +601,11 @@ class PostProcessingExtension(ComfyExtension): Quantize, Sharpen, ImageScaleToTotalPixels, + ResizeImageMaskNode, + BatchImagesNode, + BatchMasksNode, + BatchLatentsNode, + # BatchImagesMasksLatentsNode, ] async def comfy_entrypoint() -> PostProcessingExtension: diff --git a/comfy_extras/nodes_primitive.py b/comfy_extras/nodes_primitive.py index 5a1aeba80..937321800 100644 --- a/comfy_extras/nodes_primitive.py +++ b/comfy_extras/nodes_primitive.py @@ -66,7 +66,7 @@ class Float(io.ComfyNode): display_name="Float", category="utils/primitive", inputs=[ - io.Float.Input("value", min=-sys.maxsize, max=sys.maxsize), + io.Float.Input("value", min=-sys.maxsize, max=sys.maxsize, step=0.1), ], outputs=[io.Float.Output()], ) diff --git a/execution.py b/execution.py index 0c239efd7..38159b1f4 100644 --- a/execution.py +++ b/execution.py @@ -79,7 +79,7 @@ class IsChangedCache: # Intentionally do not use cached outputs here. We only want constants in IS_CHANGED input_data_all, _, v3_data = get_input_data(node["inputs"], class_def, node_id, None) try: - is_changed = await _async_map_node_over_list(self.prompt_id, node_id, class_def, input_data_all, is_changed_name) + is_changed = await _async_map_node_over_list(self.prompt_id, node_id, class_def, input_data_all, is_changed_name, v3_data=v3_data) is_changed = await resolve_map_node_over_list_results(is_changed) node["is_changed"] = [None if isinstance(x, ExecutionBlocker) else x for x in is_changed] except Exception as e: @@ -148,13 +148,12 @@ SENSITIVE_EXTRA_DATA_KEYS = ("auth_token_comfy_org", "api_key_comfy_org") def get_input_data(inputs, class_def, unique_id, execution_list=None, dynprompt=None, extra_data={}): is_v3 = issubclass(class_def, _ComfyNodeInternal) v3_data: io.V3Data = {} + hidden_inputs_v3 = {} + valid_inputs = class_def.INPUT_TYPES() if is_v3: - valid_inputs, schema, v3_data = class_def.INPUT_TYPES(include_hidden=False, return_schema=True, live_inputs=inputs) - else: - valid_inputs = class_def.INPUT_TYPES() + valid_inputs, hidden, v3_data = _io.get_finalized_class_inputs(valid_inputs, inputs) input_data_all = {} missing_keys = {} - hidden_inputs_v3 = {} for x in inputs: input_data = inputs[x] _, input_category, input_info = get_input_info(class_def, x, valid_inputs) @@ -180,18 +179,18 @@ def get_input_data(inputs, class_def, unique_id, execution_list=None, dynprompt= input_data_all[x] = [input_data] if is_v3: - if schema.hidden: - if io.Hidden.prompt in schema.hidden: + if hidden is not None: + if io.Hidden.prompt.name in hidden: hidden_inputs_v3[io.Hidden.prompt] = dynprompt.get_original_prompt() if dynprompt is not None else {} - if io.Hidden.dynprompt in schema.hidden: + if io.Hidden.dynprompt.name in hidden: hidden_inputs_v3[io.Hidden.dynprompt] = dynprompt - if io.Hidden.extra_pnginfo in schema.hidden: + if io.Hidden.extra_pnginfo.name in hidden: hidden_inputs_v3[io.Hidden.extra_pnginfo] = extra_data.get('extra_pnginfo', None) - if io.Hidden.unique_id in schema.hidden: + if io.Hidden.unique_id.name in hidden: hidden_inputs_v3[io.Hidden.unique_id] = unique_id - if io.Hidden.auth_token_comfy_org in schema.hidden: + if io.Hidden.auth_token_comfy_org.name in hidden: hidden_inputs_v3[io.Hidden.auth_token_comfy_org] = extra_data.get("auth_token_comfy_org", None) - if io.Hidden.api_key_comfy_org in schema.hidden: + if io.Hidden.api_key_comfy_org.name in hidden: hidden_inputs_v3[io.Hidden.api_key_comfy_org] = extra_data.get("api_key_comfy_org", None) else: if "hidden" in valid_inputs: @@ -258,7 +257,7 @@ async def _async_map_node_over_list(prompt_id, unique_id, obj, input_data_all, f pre_execute_cb(index) # V3 if isinstance(obj, _ComfyNodeInternal) or (is_class(obj) and issubclass(obj, _ComfyNodeInternal)): - # if is just a class, then assign no resources or state, just create clone + # if is just a class, then assign no state, just create clone if is_class(obj): type_obj = obj obj.VALIDATE_CLASS() @@ -481,7 +480,10 @@ async def execute(server, dynprompt, caches, current_item, extra_data, executed, else: lazy_status_present = getattr(obj, "check_lazy_status", None) is not None if lazy_status_present: - required_inputs = await _async_map_node_over_list(prompt_id, unique_id, obj, input_data_all, "check_lazy_status", allow_interrupt=True, v3_data=v3_data) + # for check_lazy_status, the returned data should include the original key of the input + v3_data_lazy = v3_data.copy() + v3_data_lazy["create_dynamic_tuple"] = True + required_inputs = await _async_map_node_over_list(prompt_id, unique_id, obj, input_data_all, "check_lazy_status", allow_interrupt=True, v3_data=v3_data_lazy) required_inputs = await resolve_map_node_over_list_results(required_inputs) required_inputs = set(sum([r for r in required_inputs if isinstance(r,list)], [])) required_inputs = [x for x in required_inputs if isinstance(x,str) and ( @@ -756,10 +758,13 @@ async def validate_inputs(prompt_id, prompt, item, validated): errors = [] valid = True + v3_data = None validate_function_inputs = [] validate_has_kwargs = False if issubclass(obj_class, _ComfyNodeInternal): - class_inputs, _, _ = obj_class.INPUT_TYPES(include_hidden=False, return_schema=True, live_inputs=inputs) + obj_class: _io._ComfyNodeBaseInternal + class_inputs = obj_class.INPUT_TYPES() + class_inputs, _, v3_data = _io.get_finalized_class_inputs(class_inputs, inputs) validate_function_name = "validate_inputs" validate_function = first_real_override(obj_class, validate_function_name) else: @@ -779,10 +784,11 @@ async def validate_inputs(prompt_id, prompt, item, validated): assert extra_info is not None if x not in inputs: if input_category == "required": + details = f"{x}" if not v3_data else x.split(".")[-1] error = { "type": "required_input_missing", "message": "Required input is missing", - "details": f"{x}", + "details": details, "extra_info": { "input_name": x } @@ -916,8 +922,11 @@ async def validate_inputs(prompt_id, prompt, item, validated): errors.append(error) continue - if isinstance(input_type, list): - combo_options = input_type + if isinstance(input_type, list) or input_type == io.Combo.io_type: + if input_type == io.Combo.io_type: + combo_options = extra_info.get("options", []) + else: + combo_options = input_type if val not in combo_options: input_config = info list_info = "" diff --git a/nodes.py b/nodes.py index 7d83ecb21..d9e4ebd91 100644 --- a/nodes.py +++ b/nodes.py @@ -1863,6 +1863,7 @@ class ImageBatch: FUNCTION = "batch" CATEGORY = "image" + DEPRECATED = True def batch(self, image1, image2): if image1.shape[-1] != image2.shape[-1]: From 6ca3d5c011bc15737131eb665939ae0a39a74254 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Wed, 31 Dec 2025 06:12:38 +0200 Subject: [PATCH 358/647] fix(api-nodes-vidu): preserve percent-encoding for signed URLs (#11564) --- comfy_api_nodes/util/_helpers.py | 20 ++++++++++++++++++++ comfy_api_nodes/util/download_helpers.py | 3 ++- 2 files changed, 22 insertions(+), 1 deletion(-) diff --git a/comfy_api_nodes/util/_helpers.py b/comfy_api_nodes/util/_helpers.py index 491e6b6a8..648defe3d 100644 --- a/comfy_api_nodes/util/_helpers.py +++ b/comfy_api_nodes/util/_helpers.py @@ -1,16 +1,22 @@ import asyncio import contextlib import os +import re import time from collections.abc import Callable from io import BytesIO +from yarl import URL + from comfy.cli_args import args from comfy.model_management import processing_interrupted from comfy_api.latest import IO from .common_exceptions import ProcessingInterrupted +_HAS_PCT_ESC = re.compile(r"%[0-9A-Fa-f]{2}") # any % followed by 2 hex digits +_HAS_BAD_PCT = re.compile(r"%(?![0-9A-Fa-f]{2})") # any % not followed by 2 hex digits + def is_processing_interrupted() -> bool: """Return True if user/runtime requested interruption.""" @@ -69,3 +75,17 @@ def get_fs_object_size(path_or_object: str | BytesIO) -> int: if isinstance(path_or_object, str): return os.path.getsize(path_or_object) return len(path_or_object.getvalue()) + + +def to_aiohttp_url(url: str) -> URL: + """If `url` appears to be already percent-encoded (contains at least one valid %HH + escape and no malformed '%' sequences) and contains no raw whitespace/control + characters preserve the original encoding byte-for-byte (important for signed/presigned URLs). + Otherwise, return `URL(url)` and allow yarl to normalize/quote as needed.""" + if any(c.isspace() for c in url) or any(ord(c) < 0x20 for c in url): + # Avoid encoded=True if URL contains raw whitespace/control chars + return URL(url) + if _HAS_PCT_ESC.search(url) and not _HAS_BAD_PCT.search(url): + # Preserve encoding only if it appears pre-encoded AND has no invalid % sequences + return URL(url, encoded=True) + return URL(url) diff --git a/comfy_api_nodes/util/download_helpers.py b/comfy_api_nodes/util/download_helpers.py index 3e0d0352d..4668d14a9 100644 --- a/comfy_api_nodes/util/download_helpers.py +++ b/comfy_api_nodes/util/download_helpers.py @@ -19,6 +19,7 @@ from ._helpers import ( get_auth_header, is_processing_interrupted, sleep_with_interrupt, + to_aiohttp_url, ) from .client import _diagnose_connectivity from .common_exceptions import ApiServerError, LocalNetworkError, ProcessingInterrupted @@ -94,7 +95,7 @@ async def download_url_to_bytesio( monitor_task = asyncio.create_task(_monitor()) - req_task = asyncio.create_task(session.get(url, headers=headers)) + req_task = asyncio.create_task(session.get(to_aiohttp_url(url), headers=headers)) done, pending = await asyncio.wait({req_task, monitor_task}, return_when=asyncio.FIRST_COMPLETED) if monitor_task in done and req_task in pending: From 236b9e211d5093b33acbe1918f56a6bfb4a5cf17 Mon Sep 17 00:00:00 2001 From: ComfyUI Wiki Date: Thu, 1 Jan 2026 05:38:39 +0800 Subject: [PATCH 359/647] chore: update workflow templates to v0.7.65 (#11579) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 8b670b813..3a05799eb 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,5 +1,5 @@ comfyui-frontend-package==1.35.9 -comfyui-workflow-templates==0.7.64 +comfyui-workflow-templates==0.7.65 comfyui-embedded-docs==0.3.1 torch torchsde From d622a618749b603531b753cef286a6051dd85565 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 31 Dec 2025 14:38:36 -0800 Subject: [PATCH 360/647] Refactor: move clip_preprocess to comfy.clip_model (#11586) --- comfy/clip_model.py | 19 +++++++++++++++++++ comfy/clip_vision.py | 22 ++-------------------- 2 files changed, 21 insertions(+), 20 deletions(-) diff --git a/comfy/clip_model.py b/comfy/clip_model.py index 7c0cadab5..e88872728 100644 --- a/comfy/clip_model.py +++ b/comfy/clip_model.py @@ -2,6 +2,25 @@ import torch from comfy.ldm.modules.attention import optimized_attention_for_device import comfy.ops +def clip_preprocess(image, size=224, mean=[0.48145466, 0.4578275, 0.40821073], std=[0.26862954, 0.26130258, 0.27577711], crop=True): + image = image[:, :, :, :3] if image.shape[3] > 3 else image + mean = torch.tensor(mean, device=image.device, dtype=image.dtype) + std = torch.tensor(std, device=image.device, dtype=image.dtype) + image = image.movedim(-1, 1) + if not (image.shape[2] == size and image.shape[3] == size): + if crop: + scale = (size / min(image.shape[2], image.shape[3])) + scale_size = (round(scale * image.shape[2]), round(scale * image.shape[3])) + else: + scale_size = (size, size) + + image = torch.nn.functional.interpolate(image, size=scale_size, mode="bicubic", antialias=True) + h = (image.shape[2] - size)//2 + w = (image.shape[3] - size)//2 + image = image[:,:,h:h+size,w:w+size] + image = torch.clip((255. * image), 0, 255).round() / 255.0 + return (image - mean.view([3,1,1])) / std.view([3,1,1]) + class CLIPAttention(torch.nn.Module): def __init__(self, embed_dim, heads, dtype, device, operations): super().__init__() diff --git a/comfy/clip_vision.py b/comfy/clip_vision.py index 447b1ce4a..d5fc53497 100644 --- a/comfy/clip_vision.py +++ b/comfy/clip_vision.py @@ -1,6 +1,5 @@ from .utils import load_torch_file, transformers_convert, state_dict_prefix_replace import os -import torch import json import logging @@ -17,24 +16,7 @@ class Output: def __setitem__(self, key, item): setattr(self, key, item) -def clip_preprocess(image, size=224, mean=[0.48145466, 0.4578275, 0.40821073], std=[0.26862954, 0.26130258, 0.27577711], crop=True): - image = image[:, :, :, :3] if image.shape[3] > 3 else image - mean = torch.tensor(mean, device=image.device, dtype=image.dtype) - std = torch.tensor(std, device=image.device, dtype=image.dtype) - image = image.movedim(-1, 1) - if not (image.shape[2] == size and image.shape[3] == size): - if crop: - scale = (size / min(image.shape[2], image.shape[3])) - scale_size = (round(scale * image.shape[2]), round(scale * image.shape[3])) - else: - scale_size = (size, size) - - image = torch.nn.functional.interpolate(image, size=scale_size, mode="bicubic", antialias=True) - h = (image.shape[2] - size)//2 - w = (image.shape[3] - size)//2 - image = image[:,:,h:h+size,w:w+size] - image = torch.clip((255. * image), 0, 255).round() / 255.0 - return (image - mean.view([3,1,1])) / std.view([3,1,1]) +clip_preprocess = comfy.clip_model.clip_preprocess # Prevent some stuff from breaking, TODO: remove eventually IMAGE_ENCODERS = { "clip_vision_model": comfy.clip_model.CLIPVisionModelProjection, @@ -73,7 +55,7 @@ class ClipVisionModel(): def encode_image(self, image, crop=True): comfy.model_management.load_model_gpu(self.patcher) - pixel_values = clip_preprocess(image.to(self.load_device), size=self.image_size, mean=self.image_mean, std=self.image_std, crop=crop).float() + pixel_values = comfy.clip_model.clip_preprocess(image.to(self.load_device), size=self.image_size, mean=self.image_mean, std=self.image_std, crop=crop).float() out = self.model(pixel_values=pixel_values, intermediate_output='all' if self.return_all_hidden_states else -2) outputs = Output() From 1bdc9a947f578733f81c9ae894a5acd5809c7a66 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 31 Dec 2025 16:29:55 -0800 Subject: [PATCH 361/647] Remove duplicate import of model_management (#11587) --- comfy/text_encoders/llama.py | 1 - 1 file changed, 1 deletion(-) diff --git a/comfy/text_encoders/llama.py b/comfy/text_encoders/llama.py index ed29e014d..faa4e1de8 100644 --- a/comfy/text_encoders/llama.py +++ b/comfy/text_encoders/llama.py @@ -8,7 +8,6 @@ from comfy.ldm.modules.attention import optimized_attention_for_device import comfy.model_management import comfy.ldm.common_dit -import comfy.model_management from . import qwen_vl @dataclass From 65cfcf5b1bb0d0618fef7bee08ee64397be5c434 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Thu, 1 Jan 2026 19:06:14 -0800 Subject: [PATCH 362/647] New Year ruff cleanup. (#11595) --- app/model_manager.py | 4 ++-- comfy/hooks.py | 3 ++- comfy/ldm/chroma_radiance/model.py | 2 +- comfy/ldm/hunyuan_video/upsampler.py | 3 ++- comfy/ldm/modules/diffusionmodules/model.py | 6 ++++-- comfy/ldm/modules/ema.py | 4 ++-- comfy/ldm/util.py | 2 +- comfy/taesd/taehv.py | 6 ++++-- comfy_execution/graph.py | 6 +++--- comfy_extras/nodes_apg.py | 3 ++- comfy_extras/nodes_wan.py | 2 +- nodes.py | 6 ++++-- pyproject.toml | 4 ++++ server.py | 6 +++--- 14 files changed, 35 insertions(+), 22 deletions(-) diff --git a/app/model_manager.py b/app/model_manager.py index ab36bca74..f124d1117 100644 --- a/app/model_manager.py +++ b/app/model_manager.py @@ -44,7 +44,7 @@ class ModelFileManager: @routes.get("/experiment/models/{folder}") async def get_all_models(request): folder = request.match_info.get("folder", None) - if not folder in folder_paths.folder_names_and_paths: + if folder not in folder_paths.folder_names_and_paths: return web.Response(status=404) files = self.get_model_file_list(folder) return web.json_response(files) @@ -55,7 +55,7 @@ class ModelFileManager: path_index = int(request.match_info.get("path_index", None)) filename = request.match_info.get("filename", None) - if not folder_name in folder_paths.folder_names_and_paths: + if folder_name not in folder_paths.folder_names_and_paths: return web.Response(status=404) folders = folder_paths.folder_names_and_paths[folder_name] diff --git a/comfy/hooks.py b/comfy/hooks.py index 9d0731072..1a76c7ba4 100644 --- a/comfy/hooks.py +++ b/comfy/hooks.py @@ -527,7 +527,8 @@ class HookKeyframeGroup: if self._current_keyframe.get_effective_guarantee_steps(max_sigma) > 0: break # if eval_c is outside the percent range, stop looking further - else: break + else: + break # update steps current context is used self._current_used_steps += 1 # update current timestep this was performed on diff --git a/comfy/ldm/chroma_radiance/model.py b/comfy/ldm/chroma_radiance/model.py index 70d173889..4fb56165e 100644 --- a/comfy/ldm/chroma_radiance/model.py +++ b/comfy/ldm/chroma_radiance/model.py @@ -270,7 +270,7 @@ class ChromaRadiance(Chroma): bad_keys = tuple( k for k, v in overrides.items() - if type(v) != type(getattr(params, k)) and (v is not None or k not in nullable_keys) + if not isinstance(v, type(getattr(params, k))) and (v is not None or k not in nullable_keys) ) if bad_keys: e = f"Invalid value(s) in transformer_options chroma_radiance_options: {', '.join(bad_keys)}" diff --git a/comfy/ldm/hunyuan_video/upsampler.py b/comfy/ldm/hunyuan_video/upsampler.py index 85f515f67..d9e76922f 100644 --- a/comfy/ldm/hunyuan_video/upsampler.py +++ b/comfy/ldm/hunyuan_video/upsampler.py @@ -3,7 +3,8 @@ import torch.nn as nn import torch.nn.functional as F from comfy.ldm.modules.diffusionmodules.model import ResnetBlock, VideoConv3d from comfy.ldm.hunyuan_video.vae_refiner import RMS_norm -import model_management, model_patcher +import model_management +import model_patcher class SRResidualCausalBlock3D(nn.Module): def __init__(self, channels: int): diff --git a/comfy/ldm/modules/diffusionmodules/model.py b/comfy/ldm/modules/diffusionmodules/model.py index 681a55db5..1ae3ef034 100644 --- a/comfy/ldm/modules/diffusionmodules/model.py +++ b/comfy/ldm/modules/diffusionmodules/model.py @@ -394,7 +394,8 @@ class Model(nn.Module): attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels, resolution, use_timestep=True, use_linear_attn=False, attn_type="vanilla"): super().__init__() - if use_linear_attn: attn_type = "linear" + if use_linear_attn: + attn_type = "linear" self.ch = ch self.temb_ch = self.ch*4 self.num_resolutions = len(ch_mult) @@ -548,7 +549,8 @@ class Encoder(nn.Module): conv3d=False, time_compress=None, **ignore_kwargs): super().__init__() - if use_linear_attn: attn_type = "linear" + if use_linear_attn: + attn_type = "linear" self.ch = ch self.temb_ch = 0 self.num_resolutions = len(ch_mult) diff --git a/comfy/ldm/modules/ema.py b/comfy/ldm/modules/ema.py index bded25019..96ee6e895 100644 --- a/comfy/ldm/modules/ema.py +++ b/comfy/ldm/modules/ema.py @@ -45,7 +45,7 @@ class LitEma(nn.Module): shadow_params[sname] = shadow_params[sname].type_as(m_param[key]) shadow_params[sname].sub_(one_minus_decay * (shadow_params[sname] - m_param[key])) else: - assert not key in self.m_name2s_name + assert key not in self.m_name2s_name def copy_to(self, model): m_param = dict(model.named_parameters()) @@ -54,7 +54,7 @@ class LitEma(nn.Module): if m_param[key].requires_grad: m_param[key].data.copy_(shadow_params[self.m_name2s_name[key]].data) else: - assert not key in self.m_name2s_name + assert key not in self.m_name2s_name def store(self, parameters): """ diff --git a/comfy/ldm/util.py b/comfy/ldm/util.py index 30b4b4721..304936ff4 100644 --- a/comfy/ldm/util.py +++ b/comfy/ldm/util.py @@ -71,7 +71,7 @@ def count_params(model, verbose=False): def instantiate_from_config(config): - if not "target" in config: + if "target" not in config: if config == '__is_first_stage__': return None elif config == "__is_unconditional__": diff --git a/comfy/taesd/taehv.py b/comfy/taesd/taehv.py index 3dfe1e4d4..0e5f9a378 100644 --- a/comfy/taesd/taehv.py +++ b/comfy/taesd/taehv.py @@ -154,7 +154,8 @@ class TAEHV(nn.Module): self._show_progress_bar = value def encode(self, x, **kwargs): - if self.patch_size > 1: x = F.pixel_unshuffle(x, self.patch_size) + if self.patch_size > 1: + x = F.pixel_unshuffle(x, self.patch_size) x = x.movedim(2, 1) # [B, C, T, H, W] -> [B, T, C, H, W] if x.shape[1] % 4 != 0: # pad at end to multiple of 4 @@ -167,5 +168,6 @@ class TAEHV(nn.Module): def decode(self, x, **kwargs): x = self.process_in(x).movedim(2, 1) # [B, C, T, H, W] -> [B, T, C, H, W] x = apply_model_with_memblocks(self.decoder, x, self.parallel, self.show_progress_bar) - if self.patch_size > 1: x = F.pixel_shuffle(x, self.patch_size) + if self.patch_size > 1: + x = F.pixel_shuffle(x, self.patch_size) return x[:, self.frames_to_trim:].movedim(2, 1) diff --git a/comfy_execution/graph.py b/comfy_execution/graph.py index 8fc5846b7..9d170b16e 100644 --- a/comfy_execution/graph.py +++ b/comfy_execution/graph.py @@ -207,15 +207,15 @@ class ExecutionList(TopologicalSort): return self.output_cache.get(node_id) is not None def cache_link(self, from_node_id, to_node_id): - if not to_node_id in self.execution_cache: + if to_node_id not in self.execution_cache: self.execution_cache[to_node_id] = {} self.execution_cache[to_node_id][from_node_id] = self.output_cache.get(from_node_id) - if not from_node_id in self.execution_cache_listeners: + if from_node_id not in self.execution_cache_listeners: self.execution_cache_listeners[from_node_id] = set() self.execution_cache_listeners[from_node_id].add(to_node_id) def get_cache(self, from_node_id, to_node_id): - if not to_node_id in self.execution_cache: + if to_node_id not in self.execution_cache: return None value = self.execution_cache[to_node_id].get(from_node_id) if value is None: diff --git a/comfy_extras/nodes_apg.py b/comfy_extras/nodes_apg.py index f27ae7da8..b9df2dcc9 100644 --- a/comfy_extras/nodes_apg.py +++ b/comfy_extras/nodes_apg.py @@ -55,7 +55,8 @@ class APG(io.ComfyNode): def pre_cfg_function(args): nonlocal running_avg, prev_sigma - if len(args["conds_out"]) == 1: return args["conds_out"] + if len(args["conds_out"]) == 1: + return args["conds_out"] cond = args["conds_out"][0] uncond = args["conds_out"][1] diff --git a/comfy_extras/nodes_wan.py b/comfy_extras/nodes_wan.py index b0bd471bf..d32aad98e 100644 --- a/comfy_extras/nodes_wan.py +++ b/comfy_extras/nodes_wan.py @@ -817,7 +817,7 @@ def get_sample_indices(original_fps, if required_duration > total_frames / original_fps: raise ValueError("required_duration must be less than video length") - if not fixed_start is None and fixed_start >= 0: + if fixed_start is not None and fixed_start >= 0: start_frame = fixed_start else: max_start = total_frames - required_origin_frames diff --git a/nodes.py b/nodes.py index d9e4ebd91..eae2f0086 100644 --- a/nodes.py +++ b/nodes.py @@ -2242,8 +2242,10 @@ async def init_external_custom_nodes(): for possible_module in possible_modules: module_path = os.path.join(custom_node_path, possible_module) - if os.path.isfile(module_path) and os.path.splitext(module_path)[1] != ".py": continue - if module_path.endswith(".disabled"): continue + if os.path.isfile(module_path) and os.path.splitext(module_path)[1] != ".py": + continue + if module_path.endswith(".disabled"): + continue if args.disable_all_custom_nodes and possible_module not in args.whitelist_custom_nodes: logging.info(f"Skipping {possible_module} due to disable_all_custom_nodes and whitelist_custom_nodes") continue diff --git a/pyproject.toml b/pyproject.toml index bc1467941..60378de1e 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -15,12 +15,16 @@ lint.select = [ "N805", # invalid-first-argument-name-for-method "S307", # suspicious-eval-usage "S102", # exec + "E", "T", # print-usage "W", # The "F" series in Ruff stands for "Pyflakes" rules, which catch various Python syntax errors and undefined names. # See all rules here: https://docs.astral.sh/ruff/rules/#pyflakes-f "F", ] + +lint.ignore = ["E501", "E722", "E731", "E712", "E402", "E741"] + exclude = ["*.ipynb", "**/generated/*.pyi"] [tool.pylint] diff --git a/server.py b/server.py index c27f8be7d..70c8b5e3b 100644 --- a/server.py +++ b/server.py @@ -324,7 +324,7 @@ class PromptServer(): @routes.get("/models/{folder}") async def get_models(request): folder = request.match_info.get("folder", None) - if not folder in folder_paths.folder_names_and_paths: + if folder not in folder_paths.folder_names_and_paths: return web.Response(status=404) files = folder_paths.get_filename_list(folder) return web.json_response(files) @@ -579,7 +579,7 @@ class PromptServer(): folder_name = request.match_info.get("folder_name", None) if folder_name is None: return web.Response(status=404) - if not "filename" in request.rel_url.query: + if "filename" not in request.rel_url.query: return web.Response(status=404) filename = request.rel_url.query["filename"] @@ -593,7 +593,7 @@ class PromptServer(): if out is None: return web.Response(status=404) dt = json.loads(out) - if not "__metadata__" in dt: + if "__metadata__" not in dt: return web.Response(status=404) return web.json_response(dt["__metadata__"]) From 9e5f677746463228e35ac6a08f308d758ed620d5 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Fri, 2 Jan 2026 10:35:34 +0200 Subject: [PATCH 363/647] Ignore all frames except the first one for MPO format. (#11569) --- nodes.py | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-) diff --git a/nodes.py b/nodes.py index eae2f0086..662907ae6 100644 --- a/nodes.py +++ b/nodes.py @@ -1663,8 +1663,6 @@ class LoadImage: output_masks = [] w, h = None, None - excluded_formats = ['MPO'] - for i in ImageSequence.Iterator(img): i = node_helpers.pillow(ImageOps.exif_transpose, i) @@ -1692,7 +1690,10 @@ class LoadImage: output_images.append(image) output_masks.append(mask.unsqueeze(0)) - if len(output_images) > 1 and img.format not in excluded_formats: + if img.format == "MPO": + break # ignore all frames except the first one for MPO format + + if len(output_images) > 1: output_image = torch.cat(output_images, dim=0) output_mask = torch.cat(output_masks, dim=0) else: From 303b1735f8785c0d1f947af965567850ca413f61 Mon Sep 17 00:00:00 2001 From: throttlekitty Date: Fri, 2 Jan 2026 01:37:37 -0700 Subject: [PATCH 364/647] Give Mahiro CFG a more appropriate display name (#11580) --- comfy_extras/nodes_mahiro.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy_extras/nodes_mahiro.py b/comfy_extras/nodes_mahiro.py index 07b3353f4..6459ca8c1 100644 --- a/comfy_extras/nodes_mahiro.py +++ b/comfy_extras/nodes_mahiro.py @@ -10,7 +10,7 @@ class Mahiro(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="Mahiro", - display_name="Mahiro is so cute that she deserves a better guidance function!! (。・ω・。)", + display_name="Mahiro CFG", category="_for_testing", description="Modify the guidance to scale more on the 'direction' of the positive prompt rather than the difference between the negative prompt.", inputs=[ From f2fda021ab179ba31d9175698b82474a5dd14359 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Fri, 2 Jan 2026 13:18:43 +0200 Subject: [PATCH 365/647] Tripo3D: pass face_limit parameter only when it differs from default (#11601) --- comfy_api_nodes/nodes_tripo.py | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/comfy_api_nodes/nodes_tripo.py b/comfy_api_nodes/nodes_tripo.py index bd3c24fb3..e72f8e96a 100644 --- a/comfy_api_nodes/nodes_tripo.py +++ b/comfy_api_nodes/nodes_tripo.py @@ -155,7 +155,7 @@ class TripoTextToModelNode(IO.ComfyNode): model_seed=model_seed, texture_seed=texture_seed, texture_quality=texture_quality, - face_limit=face_limit, + face_limit=face_limit if face_limit != -1 else None, geometry_quality=geometry_quality, auto_size=True, quad=quad, @@ -255,7 +255,7 @@ class TripoImageToModelNode(IO.ComfyNode): texture_alignment=texture_alignment, texture_seed=texture_seed, texture_quality=texture_quality, - face_limit=face_limit, + face_limit=face_limit if face_limit != -1 else None, auto_size=True, quad=quad, ), @@ -369,7 +369,7 @@ class TripoMultiviewToModelNode(IO.ComfyNode): texture_quality=texture_quality, geometry_quality=geometry_quality, texture_alignment=texture_alignment, - face_limit=face_limit, + face_limit=face_limit if face_limit != -1 else None, quad=quad, ), ) From 9a552df898ec57f066784cc1f7c475644099b3c1 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Fri, 2 Jan 2026 17:28:10 -0800 Subject: [PATCH 366/647] Remove leftover scaled_fp8 key. (#11603) --- comfy/utils.py | 2 ++ 1 file changed, 2 insertions(+) diff --git a/comfy/utils.py b/comfy/utils.py index 8d4e2b445..e4162d7ac 100644 --- a/comfy/utils.py +++ b/comfy/utils.py @@ -1230,6 +1230,8 @@ def convert_old_quants(state_dict, model_prefix="", metadata={}): out_sd = {} layers = {} for k in list(state_dict.keys()): + if k == scaled_fp8_key: + continue if not k.startswith(model_prefix): out_sd[k] = state_dict[k] continue From 53e762a3af9502ebe61a60eb2d39d783fe8d012b Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sat, 3 Jan 2026 19:28:38 -0800 Subject: [PATCH 367/647] Print memory summary on OOM to help with debugging. (#11613) --- comfy/model_management.py | 4 ++++ execution.py | 1 + 2 files changed, 5 insertions(+) diff --git a/comfy/model_management.py b/comfy/model_management.py index 87baedd73..2501cecb7 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -1542,6 +1542,10 @@ def soft_empty_cache(force=False): def unload_all_models(): free_memory(1e30, get_torch_device()) +def debug_memory_summary(): + if is_amd() or is_nvidia(): + return torch.cuda.memory.memory_summary() + return "" #TODO: might be cleaner to put this somewhere else import threading diff --git a/execution.py b/execution.py index 38159b1f4..648f204ec 100644 --- a/execution.py +++ b/execution.py @@ -601,6 +601,7 @@ async def execute(server, dynprompt, caches, current_item, extra_data, executed, if isinstance(ex, comfy.model_management.OOM_EXCEPTION): tips = "This error means you ran out of memory on your GPU.\n\nTIPS: If the workflow worked before you might have accidentally set the batch_size to a large number." + logging.info("Memory summary: {}".format(comfy.model_management.debug_memory_summary())) logging.error("Got an OOM, unloading all loaded models.") comfy.model_management.unload_all_models() From acbf08cd60fade74b2e9e5009fa0dcad9538356b Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Sun, 4 Jan 2026 09:05:02 +0200 Subject: [PATCH 368/647] feat(api-nodes): add support for 720p resolution for Kling Omni nodes (#11604) --- comfy_api_nodes/nodes_kling.py | 15 +++++++++++++++ 1 file changed, 15 insertions(+) diff --git a/comfy_api_nodes/nodes_kling.py b/comfy_api_nodes/nodes_kling.py index 58259e029..9c707a339 100644 --- a/comfy_api_nodes/nodes_kling.py +++ b/comfy_api_nodes/nodes_kling.py @@ -807,6 +807,7 @@ class OmniProTextToVideoNode(IO.ComfyNode): ), IO.Combo.Input("aspect_ratio", options=["16:9", "9:16", "1:1"]), IO.Combo.Input("duration", options=[5, 10]), + IO.Combo.Input("resolution", options=["1080p", "720p"], optional=True), ], outputs=[ IO.Video.Output(), @@ -826,6 +827,7 @@ class OmniProTextToVideoNode(IO.ComfyNode): prompt: str, aspect_ratio: str, duration: int, + resolution: str = "1080p", ) -> IO.NodeOutput: validate_string(prompt, min_length=1, max_length=2500) response = await sync_op( @@ -837,6 +839,7 @@ class OmniProTextToVideoNode(IO.ComfyNode): prompt=prompt, aspect_ratio=aspect_ratio, duration=str(duration), + mode="pro" if resolution == "1080p" else "std", ), ) return await finish_omni_video_task(cls, response) @@ -872,6 +875,7 @@ class OmniProFirstLastFrameNode(IO.ComfyNode): optional=True, tooltip="Up to 6 additional reference images.", ), + IO.Combo.Input("resolution", options=["1080p", "720p"], optional=True), ], outputs=[ IO.Video.Output(), @@ -893,6 +897,7 @@ class OmniProFirstLastFrameNode(IO.ComfyNode): first_frame: Input.Image, end_frame: Input.Image | None = None, reference_images: Input.Image | None = None, + resolution: str = "1080p", ) -> IO.NodeOutput: prompt = normalize_omni_prompt_references(prompt) validate_string(prompt, min_length=1, max_length=2500) @@ -936,6 +941,7 @@ class OmniProFirstLastFrameNode(IO.ComfyNode): prompt=prompt, duration=str(duration), image_list=image_list, + mode="pro" if resolution == "1080p" else "std", ), ) return await finish_omni_video_task(cls, response) @@ -964,6 +970,7 @@ class OmniProImageToVideoNode(IO.ComfyNode): "reference_images", tooltip="Up to 7 reference images.", ), + IO.Combo.Input("resolution", options=["1080p", "720p"], optional=True), ], outputs=[ IO.Video.Output(), @@ -984,6 +991,7 @@ class OmniProImageToVideoNode(IO.ComfyNode): aspect_ratio: str, duration: int, reference_images: Input.Image, + resolution: str = "1080p", ) -> IO.NodeOutput: prompt = normalize_omni_prompt_references(prompt) validate_string(prompt, min_length=1, max_length=2500) @@ -1005,6 +1013,7 @@ class OmniProImageToVideoNode(IO.ComfyNode): aspect_ratio=aspect_ratio, duration=str(duration), image_list=image_list, + mode="pro" if resolution == "1080p" else "std", ), ) return await finish_omni_video_task(cls, response) @@ -1036,6 +1045,7 @@ class OmniProVideoToVideoNode(IO.ComfyNode): tooltip="Up to 4 additional reference images.", optional=True, ), + IO.Combo.Input("resolution", options=["1080p", "720p"], optional=True), ], outputs=[ IO.Video.Output(), @@ -1058,6 +1068,7 @@ class OmniProVideoToVideoNode(IO.ComfyNode): reference_video: Input.Video, keep_original_sound: bool, reference_images: Input.Image | None = None, + resolution: str = "1080p", ) -> IO.NodeOutput: prompt = normalize_omni_prompt_references(prompt) validate_string(prompt, min_length=1, max_length=2500) @@ -1090,6 +1101,7 @@ class OmniProVideoToVideoNode(IO.ComfyNode): duration=str(duration), image_list=image_list if image_list else None, video_list=video_list, + mode="pro" if resolution == "1080p" else "std", ), ) return await finish_omni_video_task(cls, response) @@ -1119,6 +1131,7 @@ class OmniProEditVideoNode(IO.ComfyNode): tooltip="Up to 4 additional reference images.", optional=True, ), + IO.Combo.Input("resolution", options=["1080p", "720p"], optional=True), ], outputs=[ IO.Video.Output(), @@ -1139,6 +1152,7 @@ class OmniProEditVideoNode(IO.ComfyNode): video: Input.Video, keep_original_sound: bool, reference_images: Input.Image | None = None, + resolution: str = "1080p", ) -> IO.NodeOutput: prompt = normalize_omni_prompt_references(prompt) validate_string(prompt, min_length=1, max_length=2500) @@ -1171,6 +1185,7 @@ class OmniProEditVideoNode(IO.ComfyNode): duration=None, image_list=image_list if image_list else None, video_list=video_list, + mode="pro" if resolution == "1080p" else "std", ), ) return await finish_omni_video_task(cls, response) From 38d049382533c6662d815b08ca3395e96cca9f57 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sun, 4 Jan 2026 16:13:50 -0800 Subject: [PATCH 369/647] Fix case where upscale model wouldn't be moved to cpu. (#11633) --- comfy_extras/nodes_upscale_model.py | 24 +++++++++++++----------- 1 file changed, 13 insertions(+), 11 deletions(-) diff --git a/comfy_extras/nodes_upscale_model.py b/comfy_extras/nodes_upscale_model.py index 4d62b87be..ed587851c 100644 --- a/comfy_extras/nodes_upscale_model.py +++ b/comfy_extras/nodes_upscale_model.py @@ -78,18 +78,20 @@ class ImageUpscaleWithModel(io.ComfyNode): overlap = 32 oom = True - while oom: - try: - steps = in_img.shape[0] * comfy.utils.get_tiled_scale_steps(in_img.shape[3], in_img.shape[2], tile_x=tile, tile_y=tile, overlap=overlap) - pbar = comfy.utils.ProgressBar(steps) - s = comfy.utils.tiled_scale(in_img, lambda a: upscale_model(a), tile_x=tile, tile_y=tile, overlap=overlap, upscale_amount=upscale_model.scale, pbar=pbar) - oom = False - except model_management.OOM_EXCEPTION as e: - tile //= 2 - if tile < 128: - raise e + try: + while oom: + try: + steps = in_img.shape[0] * comfy.utils.get_tiled_scale_steps(in_img.shape[3], in_img.shape[2], tile_x=tile, tile_y=tile, overlap=overlap) + pbar = comfy.utils.ProgressBar(steps) + s = comfy.utils.tiled_scale(in_img, lambda a: upscale_model(a), tile_x=tile, tile_y=tile, overlap=overlap, upscale_amount=upscale_model.scale, pbar=pbar) + oom = False + except model_management.OOM_EXCEPTION as e: + tile //= 2 + if tile < 128: + raise e + finally: + upscale_model.to("cpu") - upscale_model.to("cpu") s = torch.clamp(s.movedim(-3,-1), min=0, max=1.0) return io.NodeOutput(s) From f2b002372b71cf0671a4cf1fa539e1c386d727e4 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sun, 4 Jan 2026 22:58:59 -0800 Subject: [PATCH 370/647] Support the LTXV 2 model. (#11632) --- comfy/latent_formats.py | 3 + comfy/ldm/lightricks/av_model.py | 837 ++++++++++++++++ comfy/ldm/lightricks/embeddings_connector.py | 305 ++++++ comfy/ldm/lightricks/latent_upsampler.py | 292 ++++++ comfy/ldm/lightricks/model.py | 715 +++++++++++--- comfy/ldm/lightricks/symmetric_patchifier.py | 87 +- comfy/ldm/lightricks/vae/audio_vae.py | 286 ++++++ .../vae/causal_audio_autoencoder.py | 909 ++++++++++++++++++ comfy/ldm/lightricks/vocoders/vocoder.py | 213 ++++ comfy/model_base.py | 57 +- comfy/model_detection.py | 2 +- comfy/sd.py | 9 +- comfy/supported_models.py | 17 +- comfy/text_encoders/llama.py | 79 ++ comfy/text_encoders/lt.py | 111 +++ comfy/utils.py | 2 +- comfy_extras/nodes_audio.py | 2 +- comfy_extras/nodes_hunyuan.py | 15 +- comfy_extras/nodes_lt.py | 188 +++- comfy_extras/nodes_lt_audio.py | 183 ++++ comfy_extras/nodes_lt_upsampler.py | 75 ++ nodes.py | 10 +- pyproject.toml | 2 +- 23 files changed, 4214 insertions(+), 185 deletions(-) create mode 100644 comfy/ldm/lightricks/av_model.py create mode 100644 comfy/ldm/lightricks/embeddings_connector.py create mode 100644 comfy/ldm/lightricks/latent_upsampler.py create mode 100644 comfy/ldm/lightricks/vae/audio_vae.py create mode 100644 comfy/ldm/lightricks/vae/causal_audio_autoencoder.py create mode 100644 comfy/ldm/lightricks/vocoders/vocoder.py create mode 100644 comfy_extras/nodes_lt_audio.py create mode 100644 comfy_extras/nodes_lt_upsampler.py diff --git a/comfy/latent_formats.py b/comfy/latent_formats.py index f1ca0151e..9bbe30b53 100644 --- a/comfy/latent_formats.py +++ b/comfy/latent_formats.py @@ -407,6 +407,9 @@ class LTXV(LatentFormat): self.latent_rgb_factors_bias = [-0.0571, -0.1657, -0.2512] +class LTXAV(LTXV): + pass + class HunyuanVideo(LatentFormat): latent_channels = 16 latent_dimensions = 3 diff --git a/comfy/ldm/lightricks/av_model.py b/comfy/ldm/lightricks/av_model.py new file mode 100644 index 000000000..759535501 --- /dev/null +++ b/comfy/ldm/lightricks/av_model.py @@ -0,0 +1,837 @@ +from typing import Tuple +import torch +import torch.nn as nn +from comfy.ldm.lightricks.model import ( + CrossAttention, + FeedForward, + AdaLayerNormSingle, + PixArtAlphaTextProjection, + LTXVModel, +) +from comfy.ldm.lightricks.symmetric_patchifier import AudioPatchifier +import comfy.ldm.common_dit + +class BasicAVTransformerBlock(nn.Module): + def __init__( + self, + v_dim, + a_dim, + v_heads, + a_heads, + vd_head, + ad_head, + v_context_dim=None, + a_context_dim=None, + attn_precision=None, + dtype=None, + device=None, + operations=None, + ): + super().__init__() + + self.attn_precision = attn_precision + + self.attn1 = CrossAttention( + query_dim=v_dim, + heads=v_heads, + dim_head=vd_head, + context_dim=None, + attn_precision=self.attn_precision, + dtype=dtype, + device=device, + operations=operations, + ) + self.audio_attn1 = CrossAttention( + query_dim=a_dim, + heads=a_heads, + dim_head=ad_head, + context_dim=None, + attn_precision=self.attn_precision, + dtype=dtype, + device=device, + operations=operations, + ) + + self.attn2 = CrossAttention( + query_dim=v_dim, + context_dim=v_context_dim, + heads=v_heads, + dim_head=vd_head, + attn_precision=self.attn_precision, + dtype=dtype, + device=device, + operations=operations, + ) + self.audio_attn2 = CrossAttention( + query_dim=a_dim, + context_dim=a_context_dim, + heads=a_heads, + dim_head=ad_head, + attn_precision=self.attn_precision, + dtype=dtype, + device=device, + operations=operations, + ) + + # Q: Video, K,V: Audio + self.audio_to_video_attn = CrossAttention( + query_dim=v_dim, + context_dim=a_dim, + heads=a_heads, + dim_head=ad_head, + attn_precision=self.attn_precision, + dtype=dtype, + device=device, + operations=operations, + ) + + # Q: Audio, K,V: Video + self.video_to_audio_attn = CrossAttention( + query_dim=a_dim, + context_dim=v_dim, + heads=a_heads, + dim_head=ad_head, + attn_precision=self.attn_precision, + dtype=dtype, + device=device, + operations=operations, + ) + + self.ff = FeedForward( + v_dim, dim_out=v_dim, glu=True, dtype=dtype, device=device, operations=operations + ) + self.audio_ff = FeedForward( + a_dim, dim_out=a_dim, glu=True, dtype=dtype, device=device, operations=operations + ) + + self.scale_shift_table = nn.Parameter(torch.empty(6, v_dim, device=device, dtype=dtype)) + self.audio_scale_shift_table = nn.Parameter( + torch.empty(6, a_dim, device=device, dtype=dtype) + ) + + self.scale_shift_table_a2v_ca_audio = nn.Parameter( + torch.empty(5, a_dim, device=device, dtype=dtype) + ) + self.scale_shift_table_a2v_ca_video = nn.Parameter( + torch.empty(5, v_dim, device=device, dtype=dtype) + ) + + def get_ada_values( + self, scale_shift_table: torch.Tensor, batch_size: int, timestep: torch.Tensor, indices: slice = slice(None, None) + ): + num_ada_params = scale_shift_table.shape[0] + + ada_values = ( + scale_shift_table[indices].unsqueeze(0).unsqueeze(0).to(device=timestep.device, dtype=timestep.dtype) + + timestep.reshape(batch_size, timestep.shape[1], num_ada_params, -1)[:, :, indices, :] + ).unbind(dim=2) + return ada_values + + def get_av_ca_ada_values( + self, + scale_shift_table: torch.Tensor, + batch_size: int, + scale_shift_timestep: torch.Tensor, + gate_timestep: torch.Tensor, + num_scale_shift_values: int = 4, + ): + scale_shift_ada_values = self.get_ada_values( + scale_shift_table[:num_scale_shift_values, :], + batch_size, + scale_shift_timestep, + ) + gate_ada_values = self.get_ada_values( + scale_shift_table[num_scale_shift_values:, :], + batch_size, + gate_timestep, + ) + + scale_shift_chunks = [t.squeeze(2) for t in scale_shift_ada_values] + gate_ada_values = [t.squeeze(2) for t in gate_ada_values] + + return (*scale_shift_chunks, *gate_ada_values) + + def forward( + self, + x: Tuple[torch.Tensor, torch.Tensor], + v_context=None, + a_context=None, + attention_mask=None, + v_timestep=None, + a_timestep=None, + v_pe=None, + a_pe=None, + v_cross_pe=None, + a_cross_pe=None, + v_cross_scale_shift_timestep=None, + a_cross_scale_shift_timestep=None, + v_cross_gate_timestep=None, + a_cross_gate_timestep=None, + transformer_options=None, + ) -> Tuple[torch.Tensor, torch.Tensor]: + run_vx = transformer_options.get("run_vx", True) + run_ax = transformer_options.get("run_ax", True) + + vx, ax = x + run_ax = run_ax and ax.numel() > 0 + run_a2v = run_vx and transformer_options.get("a2v_cross_attn", True) and ax.numel() > 0 + run_v2a = run_ax and transformer_options.get("v2a_cross_attn", True) + + if run_vx: + vshift_msa, vscale_msa, vgate_msa = ( + self.get_ada_values(self.scale_shift_table, vx.shape[0], v_timestep, slice(0, 3)) + ) + + norm_vx = comfy.ldm.common_dit.rms_norm(vx) * (1 + vscale_msa) + vshift_msa + vx += self.attn1(norm_vx, pe=v_pe, transformer_options=transformer_options) * vgate_msa + vx += self.attn2( + comfy.ldm.common_dit.rms_norm(vx), + context=v_context, + mask=attention_mask, + transformer_options=transformer_options, + ) + + del vshift_msa, vscale_msa, vgate_msa + + if run_ax: + ashift_msa, ascale_msa, agate_msa = ( + self.get_ada_values(self.audio_scale_shift_table, ax.shape[0], a_timestep, slice(0, 3)) + ) + + norm_ax = comfy.ldm.common_dit.rms_norm(ax) * (1 + ascale_msa) + ashift_msa + ax += ( + self.audio_attn1(norm_ax, pe=a_pe, transformer_options=transformer_options) + * agate_msa + ) + ax += self.audio_attn2( + comfy.ldm.common_dit.rms_norm(ax), + context=a_context, + mask=attention_mask, + transformer_options=transformer_options, + ) + + del ashift_msa, ascale_msa, agate_msa + + # Audio - Video cross attention. + if run_a2v or run_v2a: + # norm3 + vx_norm3 = comfy.ldm.common_dit.rms_norm(vx) + ax_norm3 = comfy.ldm.common_dit.rms_norm(ax) + + ( + scale_ca_audio_hidden_states_a2v, + shift_ca_audio_hidden_states_a2v, + scale_ca_audio_hidden_states_v2a, + shift_ca_audio_hidden_states_v2a, + gate_out_v2a, + ) = self.get_av_ca_ada_values( + self.scale_shift_table_a2v_ca_audio, + ax.shape[0], + a_cross_scale_shift_timestep, + a_cross_gate_timestep, + ) + + ( + scale_ca_video_hidden_states_a2v, + shift_ca_video_hidden_states_a2v, + scale_ca_video_hidden_states_v2a, + shift_ca_video_hidden_states_v2a, + gate_out_a2v, + ) = self.get_av_ca_ada_values( + self.scale_shift_table_a2v_ca_video, + vx.shape[0], + v_cross_scale_shift_timestep, + v_cross_gate_timestep, + ) + + if run_a2v: + vx_scaled = ( + vx_norm3 * (1 + scale_ca_video_hidden_states_a2v) + + shift_ca_video_hidden_states_a2v + ) + ax_scaled = ( + ax_norm3 * (1 + scale_ca_audio_hidden_states_a2v) + + shift_ca_audio_hidden_states_a2v + ) + vx += ( + self.audio_to_video_attn( + vx_scaled, + context=ax_scaled, + pe=v_cross_pe, + k_pe=a_cross_pe, + transformer_options=transformer_options, + ) + * gate_out_a2v + ) + + del gate_out_a2v + del scale_ca_video_hidden_states_a2v,\ + shift_ca_video_hidden_states_a2v,\ + scale_ca_audio_hidden_states_a2v,\ + shift_ca_audio_hidden_states_a2v,\ + + if run_v2a: + ax_scaled = ( + ax_norm3 * (1 + scale_ca_audio_hidden_states_v2a) + + shift_ca_audio_hidden_states_v2a + ) + vx_scaled = ( + vx_norm3 * (1 + scale_ca_video_hidden_states_v2a) + + shift_ca_video_hidden_states_v2a + ) + ax += ( + self.video_to_audio_attn( + ax_scaled, + context=vx_scaled, + pe=a_cross_pe, + k_pe=v_cross_pe, + transformer_options=transformer_options, + ) + * gate_out_v2a + ) + + del gate_out_v2a + del scale_ca_video_hidden_states_v2a,\ + shift_ca_video_hidden_states_v2a,\ + scale_ca_audio_hidden_states_v2a,\ + shift_ca_audio_hidden_states_v2a + + if run_vx: + vshift_mlp, vscale_mlp, vgate_mlp = ( + self.get_ada_values(self.scale_shift_table, vx.shape[0], v_timestep, slice(3, None)) + ) + + vx_scaled = comfy.ldm.common_dit.rms_norm(vx) * (1 + vscale_mlp) + vshift_mlp + vx += self.ff(vx_scaled) * vgate_mlp + del vshift_mlp, vscale_mlp, vgate_mlp + + if run_ax: + ashift_mlp, ascale_mlp, agate_mlp = ( + self.get_ada_values(self.audio_scale_shift_table, ax.shape[0], a_timestep, slice(3, None)) + ) + + ax_scaled = comfy.ldm.common_dit.rms_norm(ax) * (1 + ascale_mlp) + ashift_mlp + ax += self.audio_ff(ax_scaled) * agate_mlp + + del ashift_mlp, ascale_mlp, agate_mlp + + + return vx, ax + + +class LTXAVModel(LTXVModel): + """LTXAV model for audio-video generation.""" + + def __init__( + self, + in_channels=128, + audio_in_channels=128, + cross_attention_dim=4096, + audio_cross_attention_dim=2048, + attention_head_dim=128, + audio_attention_head_dim=64, + num_attention_heads=32, + audio_num_attention_heads=32, + caption_channels=3840, + num_layers=48, + positional_embedding_theta=10000.0, + positional_embedding_max_pos=[20, 2048, 2048], + audio_positional_embedding_max_pos=[20], + causal_temporal_positioning=False, + vae_scale_factors=(8, 32, 32), + use_middle_indices_grid=False, + timestep_scale_multiplier=1000.0, + av_ca_timestep_scale_multiplier=1.0, + dtype=None, + device=None, + operations=None, + **kwargs, + ): + # Store audio-specific parameters + self.audio_in_channels = audio_in_channels + self.audio_cross_attention_dim = audio_cross_attention_dim + self.audio_attention_head_dim = audio_attention_head_dim + self.audio_num_attention_heads = audio_num_attention_heads + self.audio_positional_embedding_max_pos = audio_positional_embedding_max_pos + + # Calculate audio dimensions + self.audio_inner_dim = audio_num_attention_heads * audio_attention_head_dim + self.audio_out_channels = audio_in_channels + + # Audio-specific constants + self.num_audio_channels = 8 + self.audio_frequency_bins = 16 + + self.av_ca_timestep_scale_multiplier = av_ca_timestep_scale_multiplier + + super().__init__( + in_channels=in_channels, + cross_attention_dim=cross_attention_dim, + attention_head_dim=attention_head_dim, + num_attention_heads=num_attention_heads, + caption_channels=caption_channels, + num_layers=num_layers, + positional_embedding_theta=positional_embedding_theta, + positional_embedding_max_pos=positional_embedding_max_pos, + causal_temporal_positioning=causal_temporal_positioning, + vae_scale_factors=vae_scale_factors, + use_middle_indices_grid=use_middle_indices_grid, + timestep_scale_multiplier=timestep_scale_multiplier, + dtype=dtype, + device=device, + operations=operations, + **kwargs, + ) + + def _init_model_components(self, device, dtype, **kwargs): + """Initialize LTXAV-specific components.""" + # Audio-specific projections + self.audio_patchify_proj = self.operations.Linear( + self.audio_in_channels, self.audio_inner_dim, bias=True, dtype=dtype, device=device + ) + + # Audio-specific AdaLN + self.audio_adaln_single = AdaLayerNormSingle( + self.audio_inner_dim, + use_additional_conditions=False, + dtype=dtype, + device=device, + operations=self.operations, + ) + + num_scale_shift_values = 4 + self.av_ca_video_scale_shift_adaln_single = AdaLayerNormSingle( + self.inner_dim, + use_additional_conditions=False, + embedding_coefficient=num_scale_shift_values, + dtype=dtype, + device=device, + operations=self.operations, + ) + self.av_ca_a2v_gate_adaln_single = AdaLayerNormSingle( + self.inner_dim, + use_additional_conditions=False, + embedding_coefficient=1, + dtype=dtype, + device=device, + operations=self.operations, + ) + self.av_ca_audio_scale_shift_adaln_single = AdaLayerNormSingle( + self.audio_inner_dim, + use_additional_conditions=False, + embedding_coefficient=num_scale_shift_values, + dtype=dtype, + device=device, + operations=self.operations, + ) + self.av_ca_v2a_gate_adaln_single = AdaLayerNormSingle( + self.audio_inner_dim, + use_additional_conditions=False, + embedding_coefficient=1, + dtype=dtype, + device=device, + operations=self.operations, + ) + + # Audio caption projection + self.audio_caption_projection = PixArtAlphaTextProjection( + in_features=self.caption_channels, + hidden_size=self.audio_inner_dim, + dtype=dtype, + device=device, + operations=self.operations, + ) + + def _init_transformer_blocks(self, device, dtype, **kwargs): + """Initialize transformer blocks for LTXAV.""" + self.transformer_blocks = nn.ModuleList( + [ + BasicAVTransformerBlock( + v_dim=self.inner_dim, + a_dim=self.audio_inner_dim, + v_heads=self.num_attention_heads, + a_heads=self.audio_num_attention_heads, + vd_head=self.attention_head_dim, + ad_head=self.audio_attention_head_dim, + v_context_dim=self.cross_attention_dim, + a_context_dim=self.audio_cross_attention_dim, + dtype=dtype, + device=device, + operations=self.operations, + ) + for _ in range(self.num_layers) + ] + ) + + def _init_output_components(self, device, dtype): + """Initialize output components for LTXAV.""" + # Video output components + super()._init_output_components(device, dtype) + # Audio output components + self.audio_scale_shift_table = nn.Parameter( + torch.empty(2, self.audio_inner_dim, dtype=dtype, device=device) + ) + self.audio_norm_out = self.operations.LayerNorm( + self.audio_inner_dim, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device + ) + self.audio_proj_out = self.operations.Linear( + self.audio_inner_dim, self.audio_out_channels, dtype=dtype, device=device + ) + self.a_patchifier = AudioPatchifier(1, start_end=True) + + def separate_audio_and_video_latents(self, x, audio_length): + """Separate audio and video latents from combined input.""" + # vx = x[:, : self.in_channels] + # ax = x[:, self.in_channels :] + # + # ax = ax.reshape(ax.shape[0], -1) + # ax = ax[:, : audio_length * self.num_audio_channels * self.audio_frequency_bins] + # + # ax = ax.reshape( + # ax.shape[0], self.num_audio_channels, audio_length, self.audio_frequency_bins + # ) + + vx = x[0] + ax = x[1] if len(x) > 1 else torch.zeros( + (vx.shape[0], self.num_audio_channels, 0, self.audio_frequency_bins), + device=vx.device, dtype=vx.dtype + ) + return vx, ax + + def recombine_audio_and_video_latents(self, vx, ax, target_shape=None): + if ax.numel() == 0: + return vx + else: + return [vx, ax] + """Recombine audio and video latents for output.""" + # if ax.device != vx.device or ax.dtype != vx.dtype: + # logging.warning("Audio and video latents are on different devices or dtypes.") + # ax = ax.to(device=vx.device, dtype=vx.dtype) + # logging.warning(f"Audio audio latent moved to device: {ax.device}, dtype: {ax.dtype}") + # + # ax = ax.reshape(ax.shape[0], -1) + # # pad to f x h x w of the video latents + # divisor = vx.shape[-1] * vx.shape[-2] * vx.shape[-3] + # if target_shape is None: + # repetitions = math.ceil(ax.shape[-1] / divisor) + # else: + # repetitions = target_shape[1] - vx.shape[1] + # padded_len = repetitions * divisor + # ax = F.pad(ax, (0, padded_len - ax.shape[-1])) + # ax = ax.reshape(ax.shape[0], -1, vx.shape[-3], vx.shape[-2], vx.shape[-1]) + # return torch.cat([vx, ax], dim=1) + + def _process_input(self, x, keyframe_idxs, denoise_mask, **kwargs): + """Process input for LTXAV - separate audio and video, then patchify.""" + audio_length = kwargs.get("audio_length", 0) + # Separate audio and video latents + vx, ax = self.separate_audio_and_video_latents(x, audio_length) + [vx, v_pixel_coords, additional_args] = super()._process_input( + vx, keyframe_idxs, denoise_mask, **kwargs + ) + + ax, a_latent_coords = self.a_patchifier.patchify(ax) + ax = self.audio_patchify_proj(ax) + + # additional_args.update({"av_orig_shape": list(x.shape)}) + return [vx, ax], [v_pixel_coords, a_latent_coords], additional_args + + def _prepare_timestep(self, timestep, batch_size, hidden_dtype, **kwargs): + """Prepare timestep embeddings.""" + # TODO: some code reuse is needed here. + grid_mask = kwargs.get("grid_mask", None) + if grid_mask is not None: + timestep = timestep[:, grid_mask] + + timestep = timestep * self.timestep_scale_multiplier + v_timestep, v_embedded_timestep = self.adaln_single( + timestep.flatten(), + {"resolution": None, "aspect_ratio": None}, + batch_size=batch_size, + hidden_dtype=hidden_dtype, + ) + + # Second dimension is 1 or number of tokens (if timestep_per_token) + v_timestep = v_timestep.view(batch_size, -1, v_timestep.shape[-1]) + v_embedded_timestep = v_embedded_timestep.view( + batch_size, -1, v_embedded_timestep.shape[-1] + ) + + # Prepare audio timestep + a_timestep = kwargs.get("a_timestep") + if a_timestep is not None: + a_timestep = a_timestep * self.timestep_scale_multiplier + av_ca_factor = self.av_ca_timestep_scale_multiplier / self.timestep_scale_multiplier + + av_ca_audio_scale_shift_timestep, _ = self.av_ca_audio_scale_shift_adaln_single( + a_timestep.flatten(), + {"resolution": None, "aspect_ratio": None}, + batch_size=batch_size, + hidden_dtype=hidden_dtype, + ) + av_ca_video_scale_shift_timestep, _ = self.av_ca_video_scale_shift_adaln_single( + timestep.flatten(), + {"resolution": None, "aspect_ratio": None}, + batch_size=batch_size, + hidden_dtype=hidden_dtype, + ) + av_ca_a2v_gate_noise_timestep, _ = self.av_ca_a2v_gate_adaln_single( + timestep.flatten() * av_ca_factor, + {"resolution": None, "aspect_ratio": None}, + batch_size=batch_size, + hidden_dtype=hidden_dtype, + ) + av_ca_v2a_gate_noise_timestep, _ = self.av_ca_v2a_gate_adaln_single( + a_timestep.flatten() * av_ca_factor, + {"resolution": None, "aspect_ratio": None}, + batch_size=batch_size, + hidden_dtype=hidden_dtype, + ) + + a_timestep, a_embedded_timestep = self.audio_adaln_single( + a_timestep.flatten(), + {"resolution": None, "aspect_ratio": None}, + batch_size=batch_size, + hidden_dtype=hidden_dtype, + ) + a_timestep = a_timestep.view(batch_size, -1, a_timestep.shape[-1]) + a_embedded_timestep = a_embedded_timestep.view( + batch_size, -1, a_embedded_timestep.shape[-1] + ) + cross_av_timestep_ss = [ + av_ca_audio_scale_shift_timestep, + av_ca_video_scale_shift_timestep, + av_ca_a2v_gate_noise_timestep, + av_ca_v2a_gate_noise_timestep, + ] + cross_av_timestep_ss = list( + [t.view(batch_size, -1, t.shape[-1]) for t in cross_av_timestep_ss] + ) + else: + a_timestep = timestep + a_embedded_timestep = kwargs.get("embedded_timestep") + cross_av_timestep_ss = [] + + return [v_timestep, a_timestep, cross_av_timestep_ss], [ + v_embedded_timestep, + a_embedded_timestep, + ] + + def _prepare_context(self, context, batch_size, x, attention_mask=None): + vx = x[0] + ax = x[1] + v_context, a_context = torch.split( + context, int(context.shape[-1] / 2), len(context.shape) - 1 + ) + + v_context, attention_mask = super()._prepare_context( + v_context, batch_size, vx, attention_mask + ) + if self.audio_caption_projection is not None: + a_context = self.audio_caption_projection(a_context) + a_context = a_context.view(batch_size, -1, ax.shape[-1]) + + return [v_context, a_context], attention_mask + + def _prepare_positional_embeddings(self, pixel_coords, frame_rate, x_dtype): + v_pixel_coords = pixel_coords[0] + v_pe = super()._prepare_positional_embeddings(v_pixel_coords, frame_rate, x_dtype) + + a_latent_coords = pixel_coords[1] + a_pe = self._precompute_freqs_cis( + a_latent_coords, + dim=self.audio_inner_dim, + out_dtype=x_dtype, + max_pos=self.audio_positional_embedding_max_pos, + use_middle_indices_grid=self.use_middle_indices_grid, + num_attention_heads=self.audio_num_attention_heads, + ) + + # calculate positional embeddings for the middle of the token duration, to use in av cross attention layers. + max_pos = max( + self.positional_embedding_max_pos[0], self.audio_positional_embedding_max_pos[0] + ) + v_pixel_coords = v_pixel_coords.to(torch.float32) + v_pixel_coords[:, 0] = v_pixel_coords[:, 0] * (1.0 / frame_rate) + av_cross_video_freq_cis = self._precompute_freqs_cis( + v_pixel_coords[:, 0:1, :], + dim=self.audio_cross_attention_dim, + out_dtype=x_dtype, + max_pos=[max_pos], + use_middle_indices_grid=True, + num_attention_heads=self.audio_num_attention_heads, + ) + av_cross_audio_freq_cis = self._precompute_freqs_cis( + a_latent_coords[:, 0:1, :], + dim=self.audio_cross_attention_dim, + out_dtype=x_dtype, + max_pos=[max_pos], + use_middle_indices_grid=True, + num_attention_heads=self.audio_num_attention_heads, + ) + + return [(v_pe, av_cross_video_freq_cis), (a_pe, av_cross_audio_freq_cis)] + + def _process_transformer_blocks( + self, x, context, attention_mask, timestep, pe, transformer_options={}, **kwargs + ): + vx = x[0] + ax = x[1] + v_context = context[0] + a_context = context[1] + v_timestep = timestep[0] + a_timestep = timestep[1] + v_pe, av_cross_video_freq_cis = pe[0] + a_pe, av_cross_audio_freq_cis = pe[1] + + ( + av_ca_audio_scale_shift_timestep, + av_ca_video_scale_shift_timestep, + av_ca_a2v_gate_noise_timestep, + av_ca_v2a_gate_noise_timestep, + ) = timestep[2] + + """Process transformer blocks for LTXAV.""" + patches_replace = transformer_options.get("patches_replace", {}) + blocks_replace = patches_replace.get("dit", {}) + + # Process transformer blocks + for i, block in enumerate(self.transformer_blocks): + if ("double_block", i) in blocks_replace: + + def block_wrap(args): + out = {} + out["img"] = block( + args["img"], + v_context=args["v_context"], + a_context=args["a_context"], + attention_mask=args["attention_mask"], + v_timestep=args["v_timestep"], + a_timestep=args["a_timestep"], + v_pe=args["v_pe"], + a_pe=args["a_pe"], + v_cross_pe=args["v_cross_pe"], + a_cross_pe=args["a_cross_pe"], + v_cross_scale_shift_timestep=args["v_cross_scale_shift_timestep"], + a_cross_scale_shift_timestep=args["a_cross_scale_shift_timestep"], + v_cross_gate_timestep=args["v_cross_gate_timestep"], + a_cross_gate_timestep=args["a_cross_gate_timestep"], + transformer_options=args["transformer_options"], + ) + return out + + out = blocks_replace[("double_block", i)]( + { + "img": (vx, ax), + "v_context": v_context, + "a_context": a_context, + "attention_mask": attention_mask, + "v_timestep": v_timestep, + "a_timestep": a_timestep, + "v_pe": v_pe, + "a_pe": a_pe, + "v_cross_pe": av_cross_video_freq_cis, + "a_cross_pe": av_cross_audio_freq_cis, + "v_cross_scale_shift_timestep": av_ca_video_scale_shift_timestep, + "a_cross_scale_shift_timestep": av_ca_audio_scale_shift_timestep, + "v_cross_gate_timestep": av_ca_a2v_gate_noise_timestep, + "a_cross_gate_timestep": av_ca_v2a_gate_noise_timestep, + "transformer_options": transformer_options, + }, + {"original_block": block_wrap}, + ) + vx, ax = out["img"] + else: + vx, ax = block( + (vx, ax), + v_context=v_context, + a_context=a_context, + attention_mask=attention_mask, + v_timestep=v_timestep, + a_timestep=a_timestep, + v_pe=v_pe, + a_pe=a_pe, + v_cross_pe=av_cross_video_freq_cis, + a_cross_pe=av_cross_audio_freq_cis, + v_cross_scale_shift_timestep=av_ca_video_scale_shift_timestep, + a_cross_scale_shift_timestep=av_ca_audio_scale_shift_timestep, + v_cross_gate_timestep=av_ca_a2v_gate_noise_timestep, + a_cross_gate_timestep=av_ca_v2a_gate_noise_timestep, + transformer_options=transformer_options, + ) + + return [vx, ax] + + def _process_output(self, x, embedded_timestep, keyframe_idxs, **kwargs): + vx = x[0] + ax = x[1] + v_embedded_timestep = embedded_timestep[0] + a_embedded_timestep = embedded_timestep[1] + vx = super()._process_output(vx, v_embedded_timestep, keyframe_idxs, **kwargs) + + # Process audio output + a_scale_shift_values = ( + self.audio_scale_shift_table[None, None].to(device=a_embedded_timestep.device, dtype=a_embedded_timestep.dtype) + + a_embedded_timestep[:, :, None] + ) + a_shift, a_scale = a_scale_shift_values[:, :, 0], a_scale_shift_values[:, :, 1] + + ax = self.audio_norm_out(ax) + ax = ax * (1 + a_scale) + a_shift + ax = self.audio_proj_out(ax) + + # Unpatchify audio + ax = self.a_patchifier.unpatchify( + ax, channels=self.num_audio_channels, freq=self.audio_frequency_bins + ) + + # Recombine audio and video + original_shape = kwargs.get("av_orig_shape") + return self.recombine_audio_and_video_latents(vx, ax, original_shape) + + def forward( + self, + x, + timestep, + context, + attention_mask=None, + frame_rate=25, + transformer_options={}, + keyframe_idxs=None, + **kwargs, + ): + """ + Forward pass for LTXAV model. + + Args: + x: Combined audio-video input tensor + timestep: Tuple of (video_timestep, audio_timestep) or single timestep + context: Context tensor (e.g., text embeddings) + attention_mask: Attention mask tensor + frame_rate: Frame rate for temporal processing + transformer_options: Additional options for transformer blocks + keyframe_idxs: Keyframe indices for temporal processing + **kwargs: Additional keyword arguments including audio_length + + Returns: + Combined audio-video output tensor + """ + # Handle timestep format + if isinstance(timestep, (tuple, list)) and len(timestep) == 2: + v_timestep, a_timestep = timestep + kwargs["a_timestep"] = a_timestep + timestep = v_timestep + else: + kwargs["a_timestep"] = timestep + + # Call parent forward method + return super().forward( + x, + timestep, + context, + attention_mask, + frame_rate, + transformer_options, + keyframe_idxs, + **kwargs, + ) diff --git a/comfy/ldm/lightricks/embeddings_connector.py b/comfy/ldm/lightricks/embeddings_connector.py new file mode 100644 index 000000000..f7a43f3c3 --- /dev/null +++ b/comfy/ldm/lightricks/embeddings_connector.py @@ -0,0 +1,305 @@ +import math +from typing import Optional + +import comfy.ldm.common_dit +import torch +from comfy.ldm.lightricks.model import ( + CrossAttention, + FeedForward, + generate_freq_grid_np, + interleaved_freqs_cis, + split_freqs_cis, +) +from torch import nn + + +class BasicTransformerBlock1D(nn.Module): + r""" + A basic Transformer block. + + Parameters: + + dim (`int`): The number of channels in the input and output. + num_attention_heads (`int`): The number of heads to use for multi-head attention. + attention_head_dim (`int`): The number of channels in each head. + dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. + activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward. + attention_bias (: + obj: `bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter. + upcast_attention (`bool`, *optional*): + Whether to upcast the attention computation to float32. This is useful for mixed precision training. + norm_elementwise_affine (`bool`, *optional*, defaults to `True`): + Whether to use learnable elementwise affine parameters for normalization. + standardization_norm (`str`, *optional*, defaults to `"layer_norm"`): The type of pre-normalization to use. Can be `"layer_norm"` or `"rms_norm"`. + norm_eps (`float`, *optional*, defaults to 1e-5): Epsilon value for normalization layers. + qk_norm (`str`, *optional*, defaults to None): + Set to 'layer_norm' or `rms_norm` to perform query and key normalization. + final_dropout (`bool` *optional*, defaults to False): + Whether to apply a final dropout after the last feed-forward layer. + ff_inner_dim (`int`, *optional*): Dimension of the inner feed-forward layer. If not provided, defaults to `dim * 4`. + ff_bias (`bool`, *optional*, defaults to `True`): Whether to use bias in the feed-forward layer. + attention_out_bias (`bool`, *optional*, defaults to `True`): Whether to use bias in the attention output layer. + use_rope (`bool`, *optional*, defaults to `False`): Whether to use Rotary Position Embeddings (RoPE). + ffn_dim_mult (`int`, *optional*, defaults to 4): Multiplier for the inner dimension of the feed-forward layer. + """ + + def __init__( + self, + dim, + n_heads, + d_head, + context_dim=None, + attn_precision=None, + dtype=None, + device=None, + operations=None, + ): + super().__init__() + + # Define 3 blocks. Each block has its own normalization layer. + # 1. Self-Attn + self.attn1 = CrossAttention( + query_dim=dim, + heads=n_heads, + dim_head=d_head, + context_dim=None, + dtype=dtype, + device=device, + operations=operations, + ) + + # 3. Feed-forward + self.ff = FeedForward( + dim, + dim_out=dim, + glu=True, + dtype=dtype, + device=device, + operations=operations, + ) + + def forward(self, hidden_states, attention_mask=None, pe=None) -> torch.FloatTensor: + + # Notice that normalization is always applied before the real computation in the following blocks. + + # 1. Normalization Before Self-Attention + norm_hidden_states = comfy.ldm.common_dit.rms_norm(hidden_states) + + norm_hidden_states = norm_hidden_states.squeeze(1) + + # 2. Self-Attention + attn_output = self.attn1(norm_hidden_states, mask=attention_mask, pe=pe) + + hidden_states = attn_output + hidden_states + if hidden_states.ndim == 4: + hidden_states = hidden_states.squeeze(1) + + # 3. Normalization before Feed-Forward + norm_hidden_states = comfy.ldm.common_dit.rms_norm(hidden_states) + + # 4. Feed-forward + ff_output = self.ff(norm_hidden_states) + + hidden_states = ff_output + hidden_states + if hidden_states.ndim == 4: + hidden_states = hidden_states.squeeze(1) + + return hidden_states + + +class Embeddings1DConnector(nn.Module): + _supports_gradient_checkpointing = True + + def __init__( + self, + in_channels=128, + cross_attention_dim=2048, + attention_head_dim=128, + num_attention_heads=30, + num_layers=2, + positional_embedding_theta=10000.0, + positional_embedding_max_pos=[4096], + causal_temporal_positioning=False, + num_learnable_registers: Optional[int] = 128, + dtype=None, + device=None, + operations=None, + split_rope=False, + double_precision_rope=False, + **kwargs, + ): + super().__init__() + self.dtype = dtype + self.out_channels = in_channels + self.num_attention_heads = num_attention_heads + self.inner_dim = num_attention_heads * attention_head_dim + self.causal_temporal_positioning = causal_temporal_positioning + self.positional_embedding_theta = positional_embedding_theta + self.positional_embedding_max_pos = positional_embedding_max_pos + self.split_rope = split_rope + self.double_precision_rope = double_precision_rope + self.transformer_1d_blocks = nn.ModuleList( + [ + BasicTransformerBlock1D( + self.inner_dim, + num_attention_heads, + attention_head_dim, + context_dim=cross_attention_dim, + dtype=dtype, + device=device, + operations=operations, + ) + for _ in range(num_layers) + ] + ) + + inner_dim = num_attention_heads * attention_head_dim + self.num_learnable_registers = num_learnable_registers + if self.num_learnable_registers: + self.learnable_registers = nn.Parameter( + torch.rand( + self.num_learnable_registers, inner_dim, dtype=dtype, device=device + ) + * 2.0 + - 1.0 + ) + + def get_fractional_positions(self, indices_grid): + fractional_positions = torch.stack( + [ + indices_grid[:, i] / self.positional_embedding_max_pos[i] + for i in range(1) + ], + dim=-1, + ) + return fractional_positions + + def precompute_freqs(self, indices_grid, spacing): + source_dtype = indices_grid.dtype + dtype = ( + torch.float32 + if source_dtype in (torch.bfloat16, torch.float16) + else source_dtype + ) + + fractional_positions = self.get_fractional_positions(indices_grid) + indices = ( + generate_freq_grid_np( + self.positional_embedding_theta, + indices_grid.shape[1], + self.inner_dim, + ) + if self.double_precision_rope + else self.generate_freq_grid(spacing, dtype, fractional_positions.device) + ).to(device=fractional_positions.device) + + if spacing == "exp_2": + freqs = ( + (indices * fractional_positions.unsqueeze(-1)) + .transpose(-1, -2) + .flatten(2) + ) + else: + freqs = ( + (indices * (fractional_positions.unsqueeze(-1) * 2 - 1)) + .transpose(-1, -2) + .flatten(2) + ) + return freqs + + def generate_freq_grid(self, spacing, dtype, device): + dim = self.inner_dim + theta = self.positional_embedding_theta + n_pos_dims = 1 + n_elem = 2 * n_pos_dims # 2 for cos and sin e.g. x 3 = 6 + start = 1 + end = theta + + if spacing == "exp": + indices = theta ** (torch.arange(0, dim, n_elem, device="cpu", dtype=torch.float32) / (dim - n_elem)) + indices = indices.to(dtype=dtype, device=device) + elif spacing == "exp_2": + indices = 1.0 / theta ** (torch.arange(0, dim, n_elem, device=device) / dim) + indices = indices.to(dtype=dtype) + elif spacing == "linear": + indices = torch.linspace( + start, end, dim // n_elem, device=device, dtype=dtype + ) + elif spacing == "sqrt": + indices = torch.linspace( + start**2, end**2, dim // n_elem, device=device, dtype=dtype + ).sqrt() + + indices = indices * math.pi / 2 + + return indices + + def precompute_freqs_cis(self, indices_grid, spacing="exp"): + dim = self.inner_dim + n_elem = 2 # 2 because of cos and sin + freqs = self.precompute_freqs(indices_grid, spacing) + if self.split_rope: + expected_freqs = dim // 2 + current_freqs = freqs.shape[-1] + pad_size = expected_freqs - current_freqs + cos_freq, sin_freq = split_freqs_cis( + freqs, pad_size, self.num_attention_heads + ) + else: + cos_freq, sin_freq = interleaved_freqs_cis(freqs, dim % n_elem) + return cos_freq.to(self.dtype), sin_freq.to(self.dtype), self.split_rope + + def forward( + self, + hidden_states: torch.Tensor, + attention_mask: Optional[torch.Tensor] = None, + ): + """ + The [`Transformer2DModel`] forward method. + + Args: + hidden_states (`torch.LongTensor` of shape `(batch size, num latent pixels)` if discrete, `torch.FloatTensor` of shape `(batch size, channel, height, width)` if continuous): + Input `hidden_states`. + indices_grid (`torch.LongTensor` of shape `(batch size, 3, num latent pixels)`): + attention_mask ( `torch.Tensor`, *optional*): + An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask + is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large + negative values to the attention scores corresponding to "discard" tokens. + Returns: + If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a + `tuple` where the first element is the sample tensor. + """ + # 1. Input + + if self.num_learnable_registers: + num_registers_duplications = math.ceil( + max(1024, hidden_states.shape[1]) / self.num_learnable_registers + ) + learnable_registers = torch.tile( + self.learnable_registers, (num_registers_duplications, 1) + ) + + hidden_states = torch.cat((hidden_states, learnable_registers[hidden_states.shape[1]:].unsqueeze(0).repeat(hidden_states.shape[0], 1, 1)), dim=1) + + if attention_mask is not None: + attention_mask = torch.zeros([1, 1, 1, hidden_states.shape[1]], dtype=attention_mask.dtype, device=attention_mask.device) + + indices_grid = torch.arange( + hidden_states.shape[1], dtype=torch.float32, device=hidden_states.device + ) + indices_grid = indices_grid[None, None, :] + freqs_cis = self.precompute_freqs_cis(indices_grid) + + # 2. Blocks + for block_idx, block in enumerate(self.transformer_1d_blocks): + hidden_states = block( + hidden_states, attention_mask=attention_mask, pe=freqs_cis + ) + + # 3. Output + # if self.output_scale is not None: + # hidden_states = hidden_states / self.output_scale + + hidden_states = comfy.ldm.common_dit.rms_norm(hidden_states) + + return hidden_states, attention_mask diff --git a/comfy/ldm/lightricks/latent_upsampler.py b/comfy/ldm/lightricks/latent_upsampler.py new file mode 100644 index 000000000..78ed7653f --- /dev/null +++ b/comfy/ldm/lightricks/latent_upsampler.py @@ -0,0 +1,292 @@ +from typing import Optional, Tuple +import torch +import torch.nn as nn +import torch.nn.functional as F +from einops import rearrange + + +def _rational_for_scale(scale: float) -> Tuple[int, int]: + mapping = {0.75: (3, 4), 1.5: (3, 2), 2.0: (2, 1), 4.0: (4, 1)} + if float(scale) not in mapping: + raise ValueError( + f"Unsupported spatial_scale {scale}. Choose from {list(mapping.keys())}" + ) + return mapping[float(scale)] + + +class PixelShuffleND(nn.Module): + def __init__(self, dims, upscale_factors=(2, 2, 2)): + super().__init__() + assert dims in [1, 2, 3], "dims must be 1, 2, or 3" + self.dims = dims + self.upscale_factors = upscale_factors + + def forward(self, x): + if self.dims == 3: + return rearrange( + x, + "b (c p1 p2 p3) d h w -> b c (d p1) (h p2) (w p3)", + p1=self.upscale_factors[0], + p2=self.upscale_factors[1], + p3=self.upscale_factors[2], + ) + elif self.dims == 2: + return rearrange( + x, + "b (c p1 p2) h w -> b c (h p1) (w p2)", + p1=self.upscale_factors[0], + p2=self.upscale_factors[1], + ) + elif self.dims == 1: + return rearrange( + x, + "b (c p1) f h w -> b c (f p1) h w", + p1=self.upscale_factors[0], + ) + + +class BlurDownsample(nn.Module): + """ + Anti-aliased spatial downsampling by integer stride using a fixed separable binomial kernel. + Applies only on H,W. Works for dims=2 or dims=3 (per-frame). + """ + + def __init__(self, dims: int, stride: int): + super().__init__() + assert dims in (2, 3) + assert stride >= 1 and isinstance(stride, int) + self.dims = dims + self.stride = stride + + # 5x5 separable binomial kernel [1,4,6,4,1] (outer product), normalized + k = torch.tensor([1.0, 4.0, 6.0, 4.0, 1.0]) + k2d = k[:, None] @ k[None, :] + k2d = (k2d / k2d.sum()).float() # shape (5,5) + self.register_buffer("kernel", k2d[None, None, :, :]) # (1,1,5,5) + + def forward(self, x: torch.Tensor) -> torch.Tensor: + if self.stride == 1: + return x + + def _apply_2d(x2d: torch.Tensor) -> torch.Tensor: + # x2d: (B, C, H, W) + B, C, H, W = x2d.shape + weight = self.kernel.expand(C, 1, 5, 5) # depthwise + x2d = F.conv2d( + x2d, weight=weight, bias=None, stride=self.stride, padding=2, groups=C + ) + return x2d + + if self.dims == 2: + return _apply_2d(x) + else: + # dims == 3: apply per-frame on H,W + b, c, f, h, w = x.shape + x = rearrange(x, "b c f h w -> (b f) c h w") + x = _apply_2d(x) + h2, w2 = x.shape[-2:] + x = rearrange(x, "(b f) c h w -> b c f h w", b=b, f=f, h=h2, w=w2) + return x + + +class SpatialRationalResampler(nn.Module): + """ + Fully-learned rational spatial scaling: up by 'num' via PixelShuffle, then anti-aliased + downsample by 'den' using fixed blur + stride. Operates on H,W only. + + For dims==3, work per-frame for spatial scaling (temporal axis untouched). + """ + + def __init__(self, mid_channels: int, scale: float): + super().__init__() + self.scale = float(scale) + self.num, self.den = _rational_for_scale(self.scale) + self.conv = nn.Conv2d( + mid_channels, (self.num**2) * mid_channels, kernel_size=3, padding=1 + ) + self.pixel_shuffle = PixelShuffleND(2, upscale_factors=(self.num, self.num)) + self.blur_down = BlurDownsample(dims=2, stride=self.den) + + def forward(self, x: torch.Tensor) -> torch.Tensor: + b, c, f, h, w = x.shape + x = rearrange(x, "b c f h w -> (b f) c h w") + x = self.conv(x) + x = self.pixel_shuffle(x) + x = self.blur_down(x) + x = rearrange(x, "(b f) c h w -> b c f h w", b=b, f=f) + return x + + +class ResBlock(nn.Module): + def __init__( + self, channels: int, mid_channels: Optional[int] = None, dims: int = 3 + ): + super().__init__() + if mid_channels is None: + mid_channels = channels + + Conv = nn.Conv2d if dims == 2 else nn.Conv3d + + self.conv1 = Conv(channels, mid_channels, kernel_size=3, padding=1) + self.norm1 = nn.GroupNorm(32, mid_channels) + self.conv2 = Conv(mid_channels, channels, kernel_size=3, padding=1) + self.norm2 = nn.GroupNorm(32, channels) + self.activation = nn.SiLU() + + def forward(self, x: torch.Tensor) -> torch.Tensor: + residual = x + x = self.conv1(x) + x = self.norm1(x) + x = self.activation(x) + x = self.conv2(x) + x = self.norm2(x) + x = self.activation(x + residual) + return x + + +class LatentUpsampler(nn.Module): + """ + Model to spatially upsample VAE latents. + + Args: + in_channels (`int`): Number of channels in the input latent + mid_channels (`int`): Number of channels in the middle layers + num_blocks_per_stage (`int`): Number of ResBlocks to use in each stage (pre/post upsampling) + dims (`int`): Number of dimensions for convolutions (2 or 3) + spatial_upsample (`bool`): Whether to spatially upsample the latent + temporal_upsample (`bool`): Whether to temporally upsample the latent + """ + + def __init__( + self, + in_channels: int = 128, + mid_channels: int = 512, + num_blocks_per_stage: int = 4, + dims: int = 3, + spatial_upsample: bool = True, + temporal_upsample: bool = False, + spatial_scale: float = 2.0, + rational_resampler: bool = False, + ): + super().__init__() + + self.in_channels = in_channels + self.mid_channels = mid_channels + self.num_blocks_per_stage = num_blocks_per_stage + self.dims = dims + self.spatial_upsample = spatial_upsample + self.temporal_upsample = temporal_upsample + self.spatial_scale = float(spatial_scale) + self.rational_resampler = rational_resampler + + Conv = nn.Conv2d if dims == 2 else nn.Conv3d + + self.initial_conv = Conv(in_channels, mid_channels, kernel_size=3, padding=1) + self.initial_norm = nn.GroupNorm(32, mid_channels) + self.initial_activation = nn.SiLU() + + self.res_blocks = nn.ModuleList( + [ResBlock(mid_channels, dims=dims) for _ in range(num_blocks_per_stage)] + ) + + if spatial_upsample and temporal_upsample: + self.upsampler = nn.Sequential( + nn.Conv3d(mid_channels, 8 * mid_channels, kernel_size=3, padding=1), + PixelShuffleND(3), + ) + elif spatial_upsample: + if rational_resampler: + self.upsampler = SpatialRationalResampler( + mid_channels=mid_channels, scale=self.spatial_scale + ) + else: + self.upsampler = nn.Sequential( + nn.Conv2d(mid_channels, 4 * mid_channels, kernel_size=3, padding=1), + PixelShuffleND(2), + ) + elif temporal_upsample: + self.upsampler = nn.Sequential( + nn.Conv3d(mid_channels, 2 * mid_channels, kernel_size=3, padding=1), + PixelShuffleND(1), + ) + else: + raise ValueError( + "Either spatial_upsample or temporal_upsample must be True" + ) + + self.post_upsample_res_blocks = nn.ModuleList( + [ResBlock(mid_channels, dims=dims) for _ in range(num_blocks_per_stage)] + ) + + self.final_conv = Conv(mid_channels, in_channels, kernel_size=3, padding=1) + + def forward(self, latent: torch.Tensor) -> torch.Tensor: + b, c, f, h, w = latent.shape + + if self.dims == 2: + x = rearrange(latent, "b c f h w -> (b f) c h w") + x = self.initial_conv(x) + x = self.initial_norm(x) + x = self.initial_activation(x) + + for block in self.res_blocks: + x = block(x) + + x = self.upsampler(x) + + for block in self.post_upsample_res_blocks: + x = block(x) + + x = self.final_conv(x) + x = rearrange(x, "(b f) c h w -> b c f h w", b=b, f=f) + else: + x = self.initial_conv(latent) + x = self.initial_norm(x) + x = self.initial_activation(x) + + for block in self.res_blocks: + x = block(x) + + if self.temporal_upsample: + x = self.upsampler(x) + x = x[:, :, 1:, :, :] + else: + if isinstance(self.upsampler, SpatialRationalResampler): + x = self.upsampler(x) + else: + x = rearrange(x, "b c f h w -> (b f) c h w") + x = self.upsampler(x) + x = rearrange(x, "(b f) c h w -> b c f h w", b=b, f=f) + + for block in self.post_upsample_res_blocks: + x = block(x) + + x = self.final_conv(x) + + return x + + @classmethod + def from_config(cls, config): + return cls( + in_channels=config.get("in_channels", 4), + mid_channels=config.get("mid_channels", 128), + num_blocks_per_stage=config.get("num_blocks_per_stage", 4), + dims=config.get("dims", 2), + spatial_upsample=config.get("spatial_upsample", True), + temporal_upsample=config.get("temporal_upsample", False), + spatial_scale=config.get("spatial_scale", 2.0), + rational_resampler=config.get("rational_resampler", False), + ) + + def config(self): + return { + "_class_name": "LatentUpsampler", + "in_channels": self.in_channels, + "mid_channels": self.mid_channels, + "num_blocks_per_stage": self.num_blocks_per_stage, + "dims": self.dims, + "spatial_upsample": self.spatial_upsample, + "temporal_upsample": self.temporal_upsample, + "spatial_scale": self.spatial_scale, + "rational_resampler": self.rational_resampler, + } diff --git a/comfy/ldm/lightricks/model.py b/comfy/ldm/lightricks/model.py index 593f7940f..d61e19d6e 100644 --- a/comfy/ldm/lightricks/model.py +++ b/comfy/ldm/lightricks/model.py @@ -1,13 +1,47 @@ +from abc import ABC, abstractmethod +from enum import Enum +import functools +import math +from typing import Dict, Optional, Tuple + +from einops import rearrange +import numpy as np import torch from torch import nn import comfy.patcher_extension import comfy.ldm.modules.attention import comfy.ldm.common_dit -import math -from typing import Dict, Optional, Tuple from .symmetric_patchifier import SymmetricPatchifier, latent_to_pixel_coords -from comfy.ldm.flux.math import apply_rope1 + +def _log_base(x, base): + return np.log(x) / np.log(base) + +class LTXRopeType(str, Enum): + INTERLEAVED = "interleaved" + SPLIT = "split" + + KEY = "rope_type" + + @classmethod + def from_dict(cls, kwargs, default=None): + if default is None: + default = cls.INTERLEAVED + return cls(kwargs.get(cls.KEY, default)) + + +class LTXFrequenciesPrecision(str, Enum): + FLOAT32 = "float32" + FLOAT64 = "float64" + + KEY = "frequencies_precision" + + @classmethod + def from_dict(cls, kwargs, default=None): + if default is None: + default = cls.FLOAT32 + return cls(kwargs.get(cls.KEY, default)) + def get_timestep_embedding( timesteps: torch.Tensor, @@ -39,9 +73,7 @@ def get_timestep_embedding( assert len(timesteps.shape) == 1, "Timesteps should be a 1d-array" half_dim = embedding_dim // 2 - exponent = -math.log(max_period) * torch.arange( - start=0, end=half_dim, dtype=torch.float32, device=timesteps.device - ) + exponent = -math.log(max_period) * torch.arange(start=0, end=half_dim, dtype=torch.float32, device=timesteps.device) exponent = exponent / (half_dim - downscale_freq_shift) emb = torch.exp(exponent) @@ -73,7 +105,9 @@ class TimestepEmbedding(nn.Module): post_act_fn: Optional[str] = None, cond_proj_dim=None, sample_proj_bias=True, - dtype=None, device=None, operations=None, + dtype=None, + device=None, + operations=None, ): super().__init__() @@ -90,7 +124,9 @@ class TimestepEmbedding(nn.Module): time_embed_dim_out = out_dim else: time_embed_dim_out = time_embed_dim - self.linear_2 = operations.Linear(time_embed_dim, time_embed_dim_out, sample_proj_bias, dtype=dtype, device=device) + self.linear_2 = operations.Linear( + time_embed_dim, time_embed_dim_out, sample_proj_bias, dtype=dtype, device=device + ) if post_act_fn is None: self.post_act = None @@ -139,12 +175,22 @@ class PixArtAlphaCombinedTimestepSizeEmbeddings(nn.Module): https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L164C9-L168C29 """ - def __init__(self, embedding_dim, size_emb_dim, use_additional_conditions: bool = False, dtype=None, device=None, operations=None): + def __init__( + self, + embedding_dim, + size_emb_dim, + use_additional_conditions: bool = False, + dtype=None, + device=None, + operations=None, + ): super().__init__() self.outdim = size_emb_dim self.time_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0) - self.timestep_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=embedding_dim, dtype=dtype, device=device, operations=operations) + self.timestep_embedder = TimestepEmbedding( + in_channels=256, time_embed_dim=embedding_dim, dtype=dtype, device=device, operations=operations + ) def forward(self, timestep, resolution, aspect_ratio, batch_size, hidden_dtype): timesteps_proj = self.time_proj(timestep) @@ -163,15 +209,22 @@ class AdaLayerNormSingle(nn.Module): use_additional_conditions (`bool`): To use additional conditions for normalization or not. """ - def __init__(self, embedding_dim: int, use_additional_conditions: bool = False, dtype=None, device=None, operations=None): + def __init__( + self, embedding_dim: int, embedding_coefficient: int = 6, use_additional_conditions: bool = False, dtype=None, device=None, operations=None + ): super().__init__() self.emb = PixArtAlphaCombinedTimestepSizeEmbeddings( - embedding_dim, size_emb_dim=embedding_dim // 3, use_additional_conditions=use_additional_conditions, dtype=dtype, device=device, operations=operations + embedding_dim, + size_emb_dim=embedding_dim // 3, + use_additional_conditions=use_additional_conditions, + dtype=dtype, + device=device, + operations=operations, ) self.silu = nn.SiLU() - self.linear = operations.Linear(embedding_dim, 6 * embedding_dim, bias=True, dtype=dtype, device=device) + self.linear = operations.Linear(embedding_dim, embedding_coefficient * embedding_dim, bias=True, dtype=dtype, device=device) def forward( self, @@ -185,6 +238,7 @@ class AdaLayerNormSingle(nn.Module): embedded_timestep = self.emb(timestep, **added_cond_kwargs, batch_size=batch_size, hidden_dtype=hidden_dtype) return self.linear(self.silu(embedded_timestep)), embedded_timestep + class PixArtAlphaTextProjection(nn.Module): """ Projects caption embeddings. Also handles dropout for classifier-free guidance. @@ -192,18 +246,24 @@ class PixArtAlphaTextProjection(nn.Module): Adapted from https://github.com/PixArt-alpha/PixArt-alpha/blob/master/diffusion/model/nets/PixArt_blocks.py """ - def __init__(self, in_features, hidden_size, out_features=None, act_fn="gelu_tanh", dtype=None, device=None, operations=None): + def __init__( + self, in_features, hidden_size, out_features=None, act_fn="gelu_tanh", dtype=None, device=None, operations=None + ): super().__init__() if out_features is None: out_features = hidden_size - self.linear_1 = operations.Linear(in_features=in_features, out_features=hidden_size, bias=True, dtype=dtype, device=device) + self.linear_1 = operations.Linear( + in_features=in_features, out_features=hidden_size, bias=True, dtype=dtype, device=device + ) if act_fn == "gelu_tanh": self.act_1 = nn.GELU(approximate="tanh") elif act_fn == "silu": self.act_1 = nn.SiLU() else: raise ValueError(f"Unknown activation function: {act_fn}") - self.linear_2 = operations.Linear(in_features=hidden_size, out_features=out_features, bias=True, dtype=dtype, device=device) + self.linear_2 = operations.Linear( + in_features=hidden_size, out_features=out_features, bias=True, dtype=dtype, device=device + ) def forward(self, caption): hidden_states = self.linear_1(caption) @@ -222,23 +282,68 @@ class GELU_approx(nn.Module): class FeedForward(nn.Module): - def __init__(self, dim, dim_out, mult=4, glu=False, dropout=0., dtype=None, device=None, operations=None): + def __init__(self, dim, dim_out, mult=4, glu=False, dropout=0.0, dtype=None, device=None, operations=None): super().__init__() inner_dim = int(dim * mult) project_in = GELU_approx(dim, inner_dim, dtype=dtype, device=device, operations=operations) self.net = nn.Sequential( - project_in, - nn.Dropout(dropout), - operations.Linear(inner_dim, dim_out, dtype=dtype, device=device) + project_in, nn.Dropout(dropout), operations.Linear(inner_dim, dim_out, dtype=dtype, device=device) ) def forward(self, x): return self.net(x) +def apply_rotary_emb(input_tensor, freqs_cis): + cos_freqs, sin_freqs = freqs_cis[0], freqs_cis[1] + split_pe = freqs_cis[2] if len(freqs_cis) > 2 else False + return ( + apply_split_rotary_emb(input_tensor, cos_freqs, sin_freqs) + if split_pe else + apply_interleaved_rotary_emb(input_tensor, cos_freqs, sin_freqs) + ) + +def apply_interleaved_rotary_emb(input_tensor, cos_freqs, sin_freqs): # TODO: remove duplicate funcs and pick the best/fastest one + t_dup = rearrange(input_tensor, "... (d r) -> ... d r", r=2) + t1, t2 = t_dup.unbind(dim=-1) + t_dup = torch.stack((-t2, t1), dim=-1) + input_tensor_rot = rearrange(t_dup, "... d r -> ... (d r)") + + out = input_tensor * cos_freqs + input_tensor_rot * sin_freqs + + return out + +def apply_split_rotary_emb(input_tensor, cos, sin): + needs_reshape = False + if input_tensor.ndim != 4 and cos.ndim == 4: + B, H, T, _ = cos.shape + input_tensor = input_tensor.reshape(B, T, H, -1).swapaxes(1, 2) + needs_reshape = True + split_input = rearrange(input_tensor, "... (d r) -> ... d r", d=2) + first_half_input = split_input[..., :1, :] + second_half_input = split_input[..., 1:, :] + output = split_input * cos.unsqueeze(-2) + first_half_output = output[..., :1, :] + second_half_output = output[..., 1:, :] + first_half_output.addcmul_(-sin.unsqueeze(-2), second_half_input) + second_half_output.addcmul_(sin.unsqueeze(-2), first_half_input) + output = rearrange(output, "... d r -> ... (d r)") + return output.swapaxes(1, 2).reshape(B, T, -1) if needs_reshape else output + class CrossAttention(nn.Module): - def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0., attn_precision=None, dtype=None, device=None, operations=None): + def __init__( + self, + query_dim, + context_dim=None, + heads=8, + dim_head=64, + dropout=0.0, + attn_precision=None, + dtype=None, + device=None, + operations=None, + ): super().__init__() inner_dim = dim_head * heads context_dim = query_dim if context_dim is None else context_dim @@ -254,9 +359,11 @@ class CrossAttention(nn.Module): self.to_k = operations.Linear(context_dim, inner_dim, bias=True, dtype=dtype, device=device) self.to_v = operations.Linear(context_dim, inner_dim, bias=True, dtype=dtype, device=device) - self.to_out = nn.Sequential(operations.Linear(inner_dim, query_dim, dtype=dtype, device=device), nn.Dropout(dropout)) + self.to_out = nn.Sequential( + operations.Linear(inner_dim, query_dim, dtype=dtype, device=device), nn.Dropout(dropout) + ) - def forward(self, x, context=None, mask=None, pe=None, transformer_options={}): + def forward(self, x, context=None, mask=None, pe=None, k_pe=None, transformer_options={}): q = self.to_q(x) context = x if context is None else context k = self.to_k(context) @@ -266,8 +373,8 @@ class CrossAttention(nn.Module): k = self.k_norm(k) if pe is not None: - q = apply_rope1(q.unsqueeze(1), pe).squeeze(1) - k = apply_rope1(k.unsqueeze(1), pe).squeeze(1) + q = apply_rotary_emb(q, pe) + k = apply_rotary_emb(k, pe if k_pe is None else k_pe) if mask is None: out = comfy.ldm.modules.attention.optimized_attention(q, k, v, self.heads, attn_precision=self.attn_precision, transformer_options=transformer_options) @@ -277,14 +384,34 @@ class CrossAttention(nn.Module): class BasicTransformerBlock(nn.Module): - def __init__(self, dim, n_heads, d_head, context_dim=None, attn_precision=None, dtype=None, device=None, operations=None): + def __init__( + self, dim, n_heads, d_head, context_dim=None, attn_precision=None, dtype=None, device=None, operations=None + ): super().__init__() self.attn_precision = attn_precision - self.attn1 = CrossAttention(query_dim=dim, heads=n_heads, dim_head=d_head, context_dim=None, attn_precision=self.attn_precision, dtype=dtype, device=device, operations=operations) + self.attn1 = CrossAttention( + query_dim=dim, + heads=n_heads, + dim_head=d_head, + context_dim=None, + attn_precision=self.attn_precision, + dtype=dtype, + device=device, + operations=operations, + ) self.ff = FeedForward(dim, dim_out=dim, glu=True, dtype=dtype, device=device, operations=operations) - self.attn2 = CrossAttention(query_dim=dim, context_dim=context_dim, heads=n_heads, dim_head=d_head, attn_precision=self.attn_precision, dtype=dtype, device=device, operations=operations) + self.attn2 = CrossAttention( + query_dim=dim, + context_dim=context_dim, + heads=n_heads, + dim_head=d_head, + attn_precision=self.attn_precision, + dtype=dtype, + device=device, + operations=operations, + ) self.scale_shift_table = nn.Parameter(torch.empty(6, dim, device=device, dtype=dtype)) @@ -306,116 +433,446 @@ class BasicTransformerBlock(nn.Module): return x def get_fractional_positions(indices_grid, max_pos): + n_pos_dims = indices_grid.shape[1] + assert n_pos_dims == len(max_pos), f'Number of position dimensions ({n_pos_dims}) must match max_pos length ({len(max_pos)})' fractional_positions = torch.stack( - [ - indices_grid[:, i] / max_pos[i] - for i in range(3) - ], - dim=-1, + [indices_grid[:, i] / max_pos[i] for i in range(n_pos_dims)], + axis=-1, ) return fractional_positions -def precompute_freqs_cis(indices_grid, dim, out_dtype, theta=10000.0, max_pos=[20, 2048, 2048]): - dtype = torch.float32 - device = indices_grid.device +@functools.lru_cache(maxsize=5) +def generate_freq_grid_np(positional_embedding_theta, positional_embedding_max_pos_count, inner_dim, _ = None): + theta = positional_embedding_theta + start = 1 + end = theta + + n_elem = 2 * positional_embedding_max_pos_count + pow_indices = np.power( + theta, + np.linspace( + _log_base(start, theta), + _log_base(end, theta), + inner_dim // n_elem, + dtype=np.float64, + ), + ) + return torch.tensor(pow_indices * math.pi / 2, dtype=torch.float32) + +def generate_freq_grid_pytorch(positional_embedding_theta, positional_embedding_max_pos_count, inner_dim, device): + theta = positional_embedding_theta + start = 1 + end = theta + n_elem = 2 * positional_embedding_max_pos_count + + indices = theta ** ( + torch.linspace( + math.log(start, theta), + math.log(end, theta), + inner_dim // n_elem, + device=device, + dtype=torch.float32, + ) + ) + indices = indices.to(dtype=torch.float32) + + indices = indices * math.pi / 2 + + return indices + +def generate_freqs(indices, indices_grid, max_pos, use_middle_indices_grid): + if use_middle_indices_grid: + assert(len(indices_grid.shape) == 4 and indices_grid.shape[-1] ==2) + indices_grid_start, indices_grid_end = indices_grid[..., 0], indices_grid[..., 1] + indices_grid = (indices_grid_start + indices_grid_end) / 2.0 + elif len(indices_grid.shape) == 4: + indices_grid = indices_grid[..., 0] # Get fractional positions and compute frequency indices fractional_positions = get_fractional_positions(indices_grid, max_pos) - indices = theta ** torch.linspace(0, 1, dim // 6, device=device, dtype=dtype) * math.pi / 2 + indices = indices.to(device=fractional_positions.device) - # Compute frequencies and apply cos/sin - freqs = (indices * (fractional_positions.unsqueeze(-1) * 2 - 1)).transpose(-1, -2).flatten(2) - cos_vals = freqs.cos().repeat_interleave(2, dim=-1) - sin_vals = freqs.sin().repeat_interleave(2, dim=-1) + freqs = ( + (indices * (fractional_positions.unsqueeze(-1) * 2 - 1)) + .transpose(-1, -2) + .flatten(2) + ) + return freqs - # Pad if dim is not divisible by 6 - if dim % 6 != 0: - padding_size = dim % 6 - cos_vals = torch.cat([torch.ones_like(cos_vals[:, :, :padding_size]), cos_vals], dim=-1) - sin_vals = torch.cat([torch.zeros_like(sin_vals[:, :, :padding_size]), sin_vals], dim=-1) +def interleaved_freqs_cis(freqs, pad_size): + cos_freq = freqs.cos().repeat_interleave(2, dim=-1) + sin_freq = freqs.sin().repeat_interleave(2, dim=-1) + if pad_size != 0: + cos_padding = torch.ones_like(cos_freq[:, :, : pad_size]) + sin_padding = torch.zeros_like(cos_freq[:, :, : pad_size]) + cos_freq = torch.cat([cos_padding, cos_freq], dim=-1) + sin_freq = torch.cat([sin_padding, sin_freq], dim=-1) + return cos_freq, sin_freq - # Reshape and extract one value per pair (since repeat_interleave duplicates each value) - cos_vals = cos_vals.reshape(*cos_vals.shape[:2], -1, 2)[..., 0].to(out_dtype) # [B, N, dim//2] - sin_vals = sin_vals.reshape(*sin_vals.shape[:2], -1, 2)[..., 0].to(out_dtype) # [B, N, dim//2] +def split_freqs_cis(freqs, pad_size, num_attention_heads): + cos_freq = freqs.cos() + sin_freq = freqs.sin() - # Build rotation matrix [[cos, -sin], [sin, cos]] and add heads dimension - freqs_cis = torch.stack([ - torch.stack([cos_vals, -sin_vals], dim=-1), - torch.stack([sin_vals, cos_vals], dim=-1) - ], dim=-2).unsqueeze(1) # [B, 1, N, dim//2, 2, 2] + if pad_size != 0: + cos_padding = torch.ones_like(cos_freq[:, :, :pad_size]) + sin_padding = torch.zeros_like(sin_freq[:, :, :pad_size]) - return freqs_cis + cos_freq = torch.concatenate([cos_padding, cos_freq], axis=-1) + sin_freq = torch.concatenate([sin_padding, sin_freq], axis=-1) + # Reshape freqs to be compatible with multi-head attention + B , T, half_HD = cos_freq.shape -class LTXVModel(torch.nn.Module): - def __init__(self, - in_channels=128, - cross_attention_dim=2048, - attention_head_dim=64, - num_attention_heads=32, + cos_freq = cos_freq.reshape(B, T, num_attention_heads, half_HD // num_attention_heads) + sin_freq = sin_freq.reshape(B, T, num_attention_heads, half_HD // num_attention_heads) - caption_channels=4096, - num_layers=28, + cos_freq = torch.swapaxes(cos_freq, 1, 2) # (B,H,T,D//2) + sin_freq = torch.swapaxes(sin_freq, 1, 2) # (B,H,T,D//2) + return cos_freq, sin_freq +class LTXBaseModel(torch.nn.Module, ABC): + """ + Abstract base class for LTX models (Lightricks Transformer models). - positional_embedding_theta=10000.0, - positional_embedding_max_pos=[20, 2048, 2048], - causal_temporal_positioning=False, - vae_scale_factors=(8, 32, 32), - dtype=None, device=None, operations=None, **kwargs): + This class defines the common interface and shared functionality for all LTX models, + including LTXV (video) and LTXAV (audio-video) variants. + """ + + def __init__( + self, + in_channels: int, + cross_attention_dim: int, + attention_head_dim: int, + num_attention_heads: int, + caption_channels: int, + num_layers: int, + positional_embedding_theta: float = 10000.0, + positional_embedding_max_pos: list = [20, 2048, 2048], + causal_temporal_positioning: bool = False, + vae_scale_factors: tuple = (8, 32, 32), + use_middle_indices_grid=False, + timestep_scale_multiplier = 1000.0, + dtype=None, + device=None, + operations=None, + **kwargs, + ): super().__init__() self.generator = None self.vae_scale_factors = vae_scale_factors + self.use_middle_indices_grid = use_middle_indices_grid self.dtype = dtype - self.out_channels = in_channels - self.inner_dim = num_attention_heads * attention_head_dim + self.in_channels = in_channels + self.cross_attention_dim = cross_attention_dim + self.attention_head_dim = attention_head_dim + self.num_attention_heads = num_attention_heads + self.caption_channels = caption_channels + self.num_layers = num_layers + self.positional_embedding_theta = positional_embedding_theta + self.positional_embedding_max_pos = positional_embedding_max_pos + self.split_positional_embedding = LTXRopeType.from_dict(kwargs) + self.freq_grid_generator = ( + generate_freq_grid_np if LTXFrequenciesPrecision.from_dict(kwargs) == LTXFrequenciesPrecision.FLOAT64 + else generate_freq_grid_pytorch + ) self.causal_temporal_positioning = causal_temporal_positioning + self.operations = operations + self.timestep_scale_multiplier = timestep_scale_multiplier - self.patchify_proj = operations.Linear(in_channels, self.inner_dim, bias=True, dtype=dtype, device=device) + # Common dimensions + self.inner_dim = num_attention_heads * attention_head_dim + self.out_channels = in_channels + + # Initialize common components + self._init_common_components(device, dtype) + + # Initialize model-specific components + self._init_model_components(device, dtype, **kwargs) + + # Initialize transformer blocks + self._init_transformer_blocks(device, dtype, **kwargs) + + # Initialize output components + self._init_output_components(device, dtype) + + def _init_common_components(self, device, dtype): + """Initialize components common to all LTX models + - patchify_proj: Linear projection for patchifying input + - adaln_single: AdaLN layer for timestep embedding + - caption_projection: Linear projection for caption embedding + """ + self.patchify_proj = self.operations.Linear( + self.in_channels, self.inner_dim, bias=True, dtype=dtype, device=device + ) self.adaln_single = AdaLayerNormSingle( - self.inner_dim, use_additional_conditions=False, dtype=dtype, device=device, operations=operations + self.inner_dim, use_additional_conditions=False, dtype=dtype, device=device, operations=self.operations ) - # self.adaln_single.linear = operations.Linear(self.inner_dim, 4 * self.inner_dim, bias=True, dtype=dtype, device=device) - self.caption_projection = PixArtAlphaTextProjection( - in_features=caption_channels, hidden_size=self.inner_dim, dtype=dtype, device=device, operations=operations + in_features=self.caption_channels, + hidden_size=self.inner_dim, + dtype=dtype, + device=device, + operations=self.operations, ) + @abstractmethod + def _init_model_components(self, device, dtype, **kwargs): + """Initialize model-specific components. Must be implemented by subclasses.""" + pass + + @abstractmethod + def _init_transformer_blocks(self, device, dtype, **kwargs): + """Initialize transformer blocks. Must be implemented by subclasses.""" + pass + + @abstractmethod + def _init_output_components(self, device, dtype): + """Initialize output components. Must be implemented by subclasses.""" + pass + + @abstractmethod + def _process_input(self, x, keyframe_idxs, denoise_mask, **kwargs): + """Process input data. Must be implemented by subclasses.""" + pass + + @abstractmethod + def _process_transformer_blocks(self, x, context, attention_mask, timestep, pe, **kwargs): + """Process transformer blocks. Must be implemented by subclasses.""" + pass + + @abstractmethod + def _process_output(self, x, embedded_timestep, keyframe_idxs, **kwargs): + """Process output data. Must be implemented by subclasses.""" + pass + + def _prepare_timestep(self, timestep, batch_size, hidden_dtype, **kwargs): + """Prepare timestep embeddings.""" + grid_mask = kwargs.get("grid_mask", None) + if grid_mask is not None: + timestep = timestep[:, grid_mask] + + timestep = timestep * self.timestep_scale_multiplier + timestep, embedded_timestep = self.adaln_single( + timestep.flatten(), + {"resolution": None, "aspect_ratio": None}, + batch_size=batch_size, + hidden_dtype=hidden_dtype, + ) + + # Second dimension is 1 or number of tokens (if timestep_per_token) + timestep = timestep.view(batch_size, -1, timestep.shape[-1]) + embedded_timestep = embedded_timestep.view(batch_size, -1, embedded_timestep.shape[-1]) + + return timestep, embedded_timestep + + def _prepare_context(self, context, batch_size, x, attention_mask=None): + """Prepare context for transformer blocks.""" + if self.caption_projection is not None: + context = self.caption_projection(context) + context = context.view(batch_size, -1, x.shape[-1]) + + return context, attention_mask + + def _precompute_freqs_cis( + self, + indices_grid, + dim, + out_dtype, + theta=10000.0, + max_pos=[20, 2048, 2048], + use_middle_indices_grid=False, + num_attention_heads=32, + ): + split_mode = self.split_positional_embedding == LTXRopeType.SPLIT + indices = self.freq_grid_generator(theta, indices_grid.shape[1], dim, indices_grid.device) + freqs = generate_freqs(indices, indices_grid, max_pos, use_middle_indices_grid) + + if split_mode: + expected_freqs = dim // 2 + current_freqs = freqs.shape[-1] + pad_size = expected_freqs - current_freqs + cos_freq, sin_freq = split_freqs_cis(freqs, pad_size, num_attention_heads) + else: + # 2 because of cos and sin by 3 for (t, x, y), 1 for temporal only + n_elem = 2 * indices_grid.shape[1] + cos_freq, sin_freq = interleaved_freqs_cis(freqs, dim % n_elem) + return cos_freq.to(out_dtype), sin_freq.to(out_dtype), split_mode + + def _prepare_positional_embeddings(self, pixel_coords, frame_rate, x_dtype): + """Prepare positional embeddings.""" + fractional_coords = pixel_coords.to(torch.float32) + fractional_coords[:, 0] = fractional_coords[:, 0] * (1.0 / frame_rate) + pe = self._precompute_freqs_cis( + fractional_coords, + dim=self.inner_dim, + out_dtype=x_dtype, + max_pos=self.positional_embedding_max_pos, + use_middle_indices_grid=self.use_middle_indices_grid, + num_attention_heads=self.num_attention_heads, + ) + return pe + + def _prepare_attention_mask(self, attention_mask, x_dtype): + """Prepare attention mask.""" + if attention_mask is not None and not torch.is_floating_point(attention_mask): + attention_mask = (attention_mask - 1).to(x_dtype).reshape( + (attention_mask.shape[0], 1, -1, attention_mask.shape[-1]) + ) * torch.finfo(x_dtype).max + return attention_mask + + def forward( + self, x, timestep, context, attention_mask, frame_rate=25, transformer_options={}, keyframe_idxs=None, denoise_mask=None, **kwargs + ): + """ + Forward pass for LTX models. + + Args: + x: Input tensor + timestep: Timestep tensor + context: Context tensor (e.g., text embeddings) + attention_mask: Attention mask tensor + frame_rate: Frame rate for temporal processing + transformer_options: Additional options for transformer blocks + keyframe_idxs: Keyframe indices for temporal processing + **kwargs: Additional keyword arguments + + Returns: + Processed output tensor + """ + return comfy.patcher_extension.WrapperExecutor.new_class_executor( + self._forward, + self, + comfy.patcher_extension.get_all_wrappers( + comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options + ), + ).execute(x, timestep, context, attention_mask, frame_rate, transformer_options, keyframe_idxs, denoise_mask=denoise_mask, **kwargs) + + def _forward( + self, x, timestep, context, attention_mask, frame_rate=25, transformer_options={}, keyframe_idxs=None, denoise_mask=None, **kwargs + ): + """ + Internal forward pass for LTX models. + + Args: + x: Input tensor + timestep: Timestep tensor + context: Context tensor (e.g., text embeddings) + attention_mask: Attention mask tensor + frame_rate: Frame rate for temporal processing + transformer_options: Additional options for transformer blocks + keyframe_idxs: Keyframe indices for temporal processing + **kwargs: Additional keyword arguments + + Returns: + Processed output tensor + """ + if isinstance(x, list): + input_dtype = x[0].dtype + batch_size = x[0].shape[0] + else: + input_dtype = x.dtype + batch_size = x.shape[0] + # Process input + merged_args = {**transformer_options, **kwargs} + x, pixel_coords, additional_args = self._process_input(x, keyframe_idxs, denoise_mask, **merged_args) + merged_args.update(additional_args) + + # Prepare timestep and context + timestep, embedded_timestep = self._prepare_timestep(timestep, batch_size, input_dtype, **merged_args) + context, attention_mask = self._prepare_context(context, batch_size, x, attention_mask) + + # Prepare attention mask and positional embeddings + attention_mask = self._prepare_attention_mask(attention_mask, input_dtype) + pe = self._prepare_positional_embeddings(pixel_coords, frame_rate, input_dtype) + + # Process transformer blocks + x = self._process_transformer_blocks( + x, context, attention_mask, timestep, pe, transformer_options=transformer_options, **merged_args + ) + + # Process output + x = self._process_output(x, embedded_timestep, keyframe_idxs, **merged_args) + return x + + +class LTXVModel(LTXBaseModel): + """LTXV model for video generation.""" + + def __init__( + self, + in_channels=128, + cross_attention_dim=2048, + attention_head_dim=64, + num_attention_heads=32, + caption_channels=4096, + num_layers=28, + positional_embedding_theta=10000.0, + positional_embedding_max_pos=[20, 2048, 2048], + causal_temporal_positioning=False, + vae_scale_factors=(8, 32, 32), + use_middle_indices_grid=False, + timestep_scale_multiplier = 1000.0, + dtype=None, + device=None, + operations=None, + **kwargs, + ): + super().__init__( + in_channels=in_channels, + cross_attention_dim=cross_attention_dim, + attention_head_dim=attention_head_dim, + num_attention_heads=num_attention_heads, + caption_channels=caption_channels, + num_layers=num_layers, + positional_embedding_theta=positional_embedding_theta, + positional_embedding_max_pos=positional_embedding_max_pos, + causal_temporal_positioning=causal_temporal_positioning, + vae_scale_factors=vae_scale_factors, + use_middle_indices_grid=use_middle_indices_grid, + timestep_scale_multiplier=timestep_scale_multiplier, + dtype=dtype, + device=device, + operations=operations, + **kwargs, + ) + + def _init_model_components(self, device, dtype, **kwargs): + """Initialize LTXV-specific components.""" + # No additional components needed for LTXV beyond base class + pass + + def _init_transformer_blocks(self, device, dtype, **kwargs): + """Initialize transformer blocks for LTXV.""" self.transformer_blocks = nn.ModuleList( [ BasicTransformerBlock( self.inner_dim, - num_attention_heads, - attention_head_dim, - context_dim=cross_attention_dim, - # attn_precision=attn_precision, - dtype=dtype, device=device, operations=operations + self.num_attention_heads, + self.attention_head_dim, + context_dim=self.cross_attention_dim, + dtype=dtype, + device=device, + operations=self.operations, ) - for d in range(num_layers) + for _ in range(self.num_layers) ] ) + def _init_output_components(self, device, dtype): + """Initialize output components for LTXV.""" self.scale_shift_table = nn.Parameter(torch.empty(2, self.inner_dim, dtype=dtype, device=device)) - self.norm_out = operations.LayerNorm(self.inner_dim, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device) - self.proj_out = operations.Linear(self.inner_dim, self.out_channels, dtype=dtype, device=device) - - self.patchifier = SymmetricPatchifier(1) - - def forward(self, x, timestep, context, attention_mask, frame_rate=25, transformer_options={}, keyframe_idxs=None, **kwargs): - return comfy.patcher_extension.WrapperExecutor.new_class_executor( - self._forward, - self, - comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options) - ).execute(x, timestep, context, attention_mask, frame_rate, transformer_options, keyframe_idxs, **kwargs) - - def _forward(self, x, timestep, context, attention_mask, frame_rate=25, transformer_options={}, keyframe_idxs=None, **kwargs): - patches_replace = transformer_options.get("patches_replace", {}) - - orig_shape = list(x.shape) + self.norm_out = self.operations.LayerNorm( + self.inner_dim, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device + ) + self.proj_out = self.operations.Linear(self.inner_dim, self.out_channels, dtype=dtype, device=device) + self.patchifier = SymmetricPatchifier(1, start_end=True) + def _process_input(self, x, keyframe_idxs, denoise_mask, **kwargs): + """Process input for LTXV.""" + additional_args = {"orig_shape": list(x.shape)} x, latent_coords = self.patchifier.patchify(x) pixel_coords = latent_to_pixel_coords( latent_coords=latent_coords, @@ -423,44 +880,30 @@ class LTXVModel(torch.nn.Module): causal_fix=self.causal_temporal_positioning, ) + grid_mask = None if keyframe_idxs is not None: - pixel_coords[:, :, -keyframe_idxs.shape[2]:] = keyframe_idxs + additional_args.update({ "orig_patchified_shape": list(x.shape)}) + denoise_mask = self.patchifier.patchify(denoise_mask)[0] + grid_mask = ~torch.any(denoise_mask < 0, dim=-1)[0] + additional_args.update({"grid_mask": grid_mask}) + x = x[:, grid_mask, :] + pixel_coords = pixel_coords[:, :, grid_mask, ...] - fractional_coords = pixel_coords.to(torch.float32) - fractional_coords[:, 0] = fractional_coords[:, 0] * (1.0 / frame_rate) + kf_grid_mask = grid_mask[-keyframe_idxs.shape[2]:] + keyframe_idxs = keyframe_idxs[..., kf_grid_mask, :] + pixel_coords[:, :, -keyframe_idxs.shape[2]:, :] = keyframe_idxs x = self.patchify_proj(x) - timestep = timestep * 1000.0 - - if attention_mask is not None and not torch.is_floating_point(attention_mask): - attention_mask = (attention_mask - 1).to(x.dtype).reshape((attention_mask.shape[0], 1, -1, attention_mask.shape[-1])) * torch.finfo(x.dtype).max - - pe = precompute_freqs_cis(fractional_coords, dim=self.inner_dim, out_dtype=x.dtype) - - batch_size = x.shape[0] - timestep, embedded_timestep = self.adaln_single( - timestep.flatten(), - {"resolution": None, "aspect_ratio": None}, - batch_size=batch_size, - hidden_dtype=x.dtype, - ) - # Second dimension is 1 or number of tokens (if timestep_per_token) - timestep = timestep.view(batch_size, -1, timestep.shape[-1]) - embedded_timestep = embedded_timestep.view( - batch_size, -1, embedded_timestep.shape[-1] - ) - - # 2. Blocks - if self.caption_projection is not None: - batch_size = x.shape[0] - context = self.caption_projection(context) - context = context.view( - batch_size, -1, x.shape[-1] - ) + return x, pixel_coords, additional_args + def _process_transformer_blocks(self, x, context, attention_mask, timestep, pe, transformer_options={}, **kwargs): + """Process transformer blocks for LTXV.""" + patches_replace = transformer_options.get("patches_replace", {}) blocks_replace = patches_replace.get("dit", {}) + for i, block in enumerate(self.transformer_blocks): if ("double_block", i) in blocks_replace: + def block_wrap(args): out = {} out["img"] = block(args["img"], context=args["txt"], attention_mask=args["attention_mask"], timestep=args["vec"], pe=args["pe"], transformer_options=args["transformer_options"]) @@ -478,16 +921,28 @@ class LTXVModel(torch.nn.Module): transformer_options=transformer_options, ) - # 3. Output + return x + + def _process_output(self, x, embedded_timestep, keyframe_idxs, **kwargs): + """Process output for LTXV.""" + # Apply scale-shift modulation scale_shift_values = ( self.scale_shift_table[None, None].to(device=x.device, dtype=x.dtype) + embedded_timestep[:, :, None] ) shift, scale = scale_shift_values[:, :, 0], scale_shift_values[:, :, 1] + x = self.norm_out(x) - # Modulation - x = torch.addcmul(x, x, scale).add_(shift) + x = x * (1 + scale) + shift x = self.proj_out(x) + if keyframe_idxs is not None: + grid_mask = kwargs["grid_mask"] + orig_patchified_shape = kwargs["orig_patchified_shape"] + full_x = torch.zeros(orig_patchified_shape, dtype=x.dtype, device=x.device) + full_x[:, grid_mask, :] = x + x = full_x + # Unpatchify to restore original dimensions + orig_shape = kwargs["orig_shape"] x = self.patchifier.unpatchify( latents=x, output_height=orig_shape[3], diff --git a/comfy/ldm/lightricks/symmetric_patchifier.py b/comfy/ldm/lightricks/symmetric_patchifier.py index 4b9972b9f..8f9a41186 100644 --- a/comfy/ldm/lightricks/symmetric_patchifier.py +++ b/comfy/ldm/lightricks/symmetric_patchifier.py @@ -21,20 +21,23 @@ def latent_to_pixel_coords( Returns: Tensor: A tensor of pixel coordinates corresponding to the input latent coordinates. """ + shape = [1] * latent_coords.ndim + shape[1] = -1 pixel_coords = ( latent_coords - * torch.tensor(scale_factors, device=latent_coords.device)[None, :, None] + * torch.tensor(scale_factors, device=latent_coords.device).view(*shape) ) if causal_fix: # Fix temporal scale for first frame to 1 due to causality - pixel_coords[:, 0] = (pixel_coords[:, 0] + 1 - scale_factors[0]).clamp(min=0) + pixel_coords[:, 0, ...] = (pixel_coords[:, 0, ...] + 1 - scale_factors[0]).clamp(min=0) return pixel_coords class Patchifier(ABC): - def __init__(self, patch_size: int): + def __init__(self, patch_size: int, start_end: bool=False): super().__init__() self._patch_size = (1, patch_size, patch_size) + self.start_end = start_end @abstractmethod def patchify( @@ -71,11 +74,23 @@ class Patchifier(ABC): torch.arange(0, latent_width, self._patch_size[2], device=device), indexing="ij", ) - latent_sample_coords = torch.stack(latent_sample_coords, dim=0) - latent_coords = latent_sample_coords.unsqueeze(0).repeat(batch_size, 1, 1, 1, 1) - latent_coords = rearrange( - latent_coords, "b c f h w -> b c (f h w)", b=batch_size + latent_sample_coords_start = torch.stack(latent_sample_coords, dim=0) + delta = torch.tensor(self._patch_size, device=latent_sample_coords_start.device, dtype=latent_sample_coords_start.dtype)[:, None, None, None] + latent_sample_coords_end = latent_sample_coords_start + delta + + latent_sample_coords_start = latent_sample_coords_start.unsqueeze(0).repeat(batch_size, 1, 1, 1, 1) + latent_sample_coords_start = rearrange( + latent_sample_coords_start, "b c f h w -> b c (f h w)", b=batch_size ) + if self.start_end: + latent_sample_coords_end = latent_sample_coords_end.unsqueeze(0).repeat(batch_size, 1, 1, 1, 1) + latent_sample_coords_end = rearrange( + latent_sample_coords_end, "b c f h w -> b c (f h w)", b=batch_size + ) + + latent_coords = torch.stack((latent_sample_coords_start, latent_sample_coords_end), dim=-1) + else: + latent_coords = latent_sample_coords_start return latent_coords @@ -115,3 +130,61 @@ class SymmetricPatchifier(Patchifier): q=self._patch_size[2], ) return latents + + +class AudioPatchifier(Patchifier): + def __init__(self, patch_size: int, + sample_rate=16000, + hop_length=160, + audio_latent_downsample_factor=4, + is_causal=True, + start_end=False, + shift = 0 + ): + super().__init__(patch_size, start_end=start_end) + self.hop_length = hop_length + self.sample_rate = sample_rate + self.audio_latent_downsample_factor = audio_latent_downsample_factor + self.is_causal = is_causal + self.shift = shift + + def copy_with_shift(self, shift): + return AudioPatchifier( + self.patch_size, self.sample_rate, self.hop_length, self.audio_latent_downsample_factor, + self.is_causal, self.start_end, shift + ) + + def _get_audio_latent_time_in_sec(self, start_latent, end_latent: int, dtype: torch.dtype, device=torch.device): + audio_latent_frame = torch.arange(start_latent, end_latent, dtype=dtype, device=device) + audio_mel_frame = audio_latent_frame * self.audio_latent_downsample_factor + if self.is_causal: + audio_mel_frame = (audio_mel_frame + 1 - self.audio_latent_downsample_factor).clip(min=0) + return audio_mel_frame * self.hop_length / self.sample_rate + + + def patchify(self, audio_latents: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]: + # audio_latents: (batch, channels, time, freq) + b, _, t, _ = audio_latents.shape + audio_latents = rearrange( + audio_latents, + "b c t f -> b t (c f)", + ) + + audio_latents_start_timings = self._get_audio_latent_time_in_sec(self.shift, t + self.shift, torch.float32, audio_latents.device) + audio_latents_start_timings = audio_latents_start_timings.unsqueeze(0).expand(b, -1).unsqueeze(1) + + if self.start_end: + audio_latents_end_timings = self._get_audio_latent_time_in_sec(self.shift + 1, t + self.shift + 1, torch.float32, audio_latents.device) + audio_latents_end_timings = audio_latents_end_timings.unsqueeze(0).expand(b, -1).unsqueeze(1) + + audio_latents_timings = torch.stack([audio_latents_start_timings, audio_latents_end_timings], dim=-1) + else: + audio_latents_timings = audio_latents_start_timings + return audio_latents, audio_latents_timings + + def unpatchify(self, audio_latents: torch.Tensor, channels: int, freq: int) -> torch.Tensor: + # audio_latents: (batch, time, freq * channels) + audio_latents = rearrange( + audio_latents, "b t (c f) -> b c t f", c=channels, f=freq + ) + return audio_latents diff --git a/comfy/ldm/lightricks/vae/audio_vae.py b/comfy/ldm/lightricks/vae/audio_vae.py new file mode 100644 index 000000000..a9111d3bd --- /dev/null +++ b/comfy/ldm/lightricks/vae/audio_vae.py @@ -0,0 +1,286 @@ +import json +from dataclasses import dataclass +import math +import torch +import torchaudio + +import comfy.model_management +import comfy.model_patcher +import comfy.utils as utils +from comfy.ldm.mmaudio.vae.distributions import DiagonalGaussianDistribution +from comfy.ldm.lightricks.symmetric_patchifier import AudioPatchifier +from comfy.ldm.lightricks.vae.causal_audio_autoencoder import ( + CausalityAxis, + CausalAudioAutoencoder, +) +from comfy.ldm.lightricks.vocoders.vocoder import Vocoder + +LATENT_DOWNSAMPLE_FACTOR = 4 + + +@dataclass(frozen=True) +class AudioVAEComponentConfig: + """Container for model component configuration extracted from metadata.""" + + autoencoder: dict + vocoder: dict + + @classmethod + def from_metadata(cls, metadata: dict) -> "AudioVAEComponentConfig": + assert metadata is not None and "config" in metadata, "Metadata is required for audio VAE" + + raw_config = metadata["config"] + if isinstance(raw_config, str): + parsed_config = json.loads(raw_config) + else: + parsed_config = raw_config + + audio_config = parsed_config.get("audio_vae") + vocoder_config = parsed_config.get("vocoder") + + assert audio_config is not None, "Audio VAE config is required for audio VAE" + assert vocoder_config is not None, "Vocoder config is required for audio VAE" + + return cls(autoencoder=audio_config, vocoder=vocoder_config) + + +class ModelDeviceManager: + """Manages device placement and GPU residency for the composed model.""" + + def __init__(self, module: torch.nn.Module): + load_device = comfy.model_management.get_torch_device() + offload_device = comfy.model_management.vae_offload_device() + self.patcher = comfy.model_patcher.ModelPatcher(module, load_device, offload_device) + + def ensure_model_loaded(self) -> None: + comfy.model_management.free_memory( + self.patcher.model_size(), + self.patcher.load_device, + ) + comfy.model_management.load_model_gpu(self.patcher) + + def move_to_load_device(self, tensor: torch.Tensor) -> torch.Tensor: + return tensor.to(self.patcher.load_device) + + @property + def load_device(self): + return self.patcher.load_device + + +class AudioLatentNormalizer: + """Applies per-channel statistics in patch space and restores original layout.""" + + def __init__(self, patchfier: AudioPatchifier, statistics_processor: torch.nn.Module): + self.patchifier = patchfier + self.statistics = statistics_processor + + def normalize(self, latents: torch.Tensor) -> torch.Tensor: + channels = latents.shape[1] + freq = latents.shape[3] + patched, _ = self.patchifier.patchify(latents) + normalized = self.statistics.normalize(patched) + return self.patchifier.unpatchify(normalized, channels=channels, freq=freq) + + def denormalize(self, latents: torch.Tensor) -> torch.Tensor: + channels = latents.shape[1] + freq = latents.shape[3] + patched, _ = self.patchifier.patchify(latents) + denormalized = self.statistics.un_normalize(patched) + return self.patchifier.unpatchify(denormalized, channels=channels, freq=freq) + + +class AudioPreprocessor: + """Prepares raw waveforms for the autoencoder by matching training conditions.""" + + def __init__(self, target_sample_rate: int, mel_bins: int, mel_hop_length: int, n_fft: int): + self.target_sample_rate = target_sample_rate + self.mel_bins = mel_bins + self.mel_hop_length = mel_hop_length + self.n_fft = n_fft + + def resample(self, waveform: torch.Tensor, source_rate: int) -> torch.Tensor: + if source_rate == self.target_sample_rate: + return waveform + return torchaudio.functional.resample(waveform, source_rate, self.target_sample_rate) + + @staticmethod + def normalize_amplitude( + waveform: torch.Tensor, max_amplitude: float = 0.5, eps: float = 1e-5 + ) -> torch.Tensor: + waveform = waveform - waveform.mean(dim=2, keepdim=True) + peak = torch.max(torch.abs(waveform)) + eps + scale = peak.clamp(max=max_amplitude) / peak + return waveform * scale + + def waveform_to_mel( + self, waveform: torch.Tensor, waveform_sample_rate: int, device + ) -> torch.Tensor: + waveform = self.resample(waveform, waveform_sample_rate) + waveform = self.normalize_amplitude(waveform) + + mel_transform = torchaudio.transforms.MelSpectrogram( + sample_rate=self.target_sample_rate, + n_fft=self.n_fft, + win_length=self.n_fft, + hop_length=self.mel_hop_length, + f_min=0.0, + f_max=self.target_sample_rate / 2.0, + n_mels=self.mel_bins, + window_fn=torch.hann_window, + center=True, + pad_mode="reflect", + power=1.0, + mel_scale="slaney", + norm="slaney", + ).to(device) + + mel = mel_transform(waveform) + mel = torch.log(torch.clamp(mel, min=1e-5)) + return mel.permute(0, 1, 3, 2).contiguous() + + +class AudioVAE(torch.nn.Module): + """High-level Audio VAE wrapper exposing encode and decode entry points.""" + + def __init__(self, state_dict: dict, metadata: dict): + super().__init__() + + component_config = AudioVAEComponentConfig.from_metadata(metadata) + + vae_sd = utils.state_dict_prefix_replace(state_dict, {"audio_vae.": ""}, filter_keys=True) + vocoder_sd = utils.state_dict_prefix_replace(state_dict, {"vocoder.": ""}, filter_keys=True) + + self.autoencoder = CausalAudioAutoencoder(config=component_config.autoencoder) + self.vocoder = Vocoder(config=component_config.vocoder) + + self.autoencoder.load_state_dict(vae_sd, strict=False) + self.vocoder.load_state_dict(vocoder_sd, strict=False) + + autoencoder_config = self.autoencoder.get_config() + self.normalizer = AudioLatentNormalizer( + AudioPatchifier( + patch_size=1, + audio_latent_downsample_factor=LATENT_DOWNSAMPLE_FACTOR, + sample_rate=autoencoder_config["sampling_rate"], + hop_length=autoencoder_config["mel_hop_length"], + is_causal=autoencoder_config["is_causal"], + ), + self.autoencoder.per_channel_statistics, + ) + + self.preprocessor = AudioPreprocessor( + target_sample_rate=autoencoder_config["sampling_rate"], + mel_bins=autoencoder_config["mel_bins"], + mel_hop_length=autoencoder_config["mel_hop_length"], + n_fft=autoencoder_config["n_fft"], + ) + + self.device_manager = ModelDeviceManager(self) + + def encode(self, audio: dict) -> torch.Tensor: + """Encode a waveform dictionary into normalized latent tensors.""" + + waveform = audio["waveform"] + waveform_sample_rate = audio["sample_rate"] + input_device = waveform.device + # Ensure that Audio VAE is loaded on the correct device. + self.device_manager.ensure_model_loaded() + + waveform = self.device_manager.move_to_load_device(waveform) + expected_channels = self.autoencoder.encoder.in_channels + if waveform.shape[1] != expected_channels: + raise ValueError( + f"Input audio must have {expected_channels} channels, got {waveform.shape[1]}" + ) + + mel_spec = self.preprocessor.waveform_to_mel( + waveform, waveform_sample_rate, device=self.device_manager.load_device + ) + + latents = self.autoencoder.encode(mel_spec) + posterior = DiagonalGaussianDistribution(latents) + latent_mode = posterior.mode() + + normalized = self.normalizer.normalize(latent_mode) + return normalized.to(input_device) + + def decode(self, latents: torch.Tensor) -> torch.Tensor: + """Decode normalized latent tensors into an audio waveform.""" + original_shape = latents.shape + + # Ensure that Audio VAE is loaded on the correct device. + self.device_manager.ensure_model_loaded() + + latents = self.device_manager.move_to_load_device(latents) + latents = self.normalizer.denormalize(latents) + + target_shape = self.target_shape_from_latents(original_shape) + mel_spec = self.autoencoder.decode(latents, target_shape=target_shape) + + waveform = self.run_vocoder(mel_spec) + return self.device_manager.move_to_load_device(waveform) + + def target_shape_from_latents(self, latents_shape): + batch, _, time, _ = latents_shape + target_length = time * LATENT_DOWNSAMPLE_FACTOR + if self.autoencoder.causality_axis != CausalityAxis.NONE: + target_length -= LATENT_DOWNSAMPLE_FACTOR - 1 + return ( + batch, + self.autoencoder.decoder.out_ch, + target_length, + self.autoencoder.mel_bins, + ) + + def num_of_latents_from_frames(self, frames_number: int, frame_rate: int) -> int: + return math.ceil((float(frames_number) / frame_rate) * self.latents_per_second) + + def run_vocoder(self, mel_spec: torch.Tensor) -> torch.Tensor: + audio_channels = self.autoencoder.decoder.out_ch + vocoder_input = mel_spec.transpose(2, 3) + + if audio_channels == 1: + vocoder_input = vocoder_input.squeeze(1) + elif audio_channels != 2: + raise ValueError(f"Unsupported audio_channels: {audio_channels}") + + return self.vocoder(vocoder_input) + + @property + def sample_rate(self) -> int: + return int(self.autoencoder.sampling_rate) + + @property + def mel_hop_length(self) -> int: + return int(self.autoencoder.mel_hop_length) + + @property + def mel_bins(self) -> int: + return int(self.autoencoder.mel_bins) + + @property + def latent_channels(self) -> int: + return int(self.autoencoder.decoder.z_channels) + + @property + def latent_frequency_bins(self) -> int: + return int(self.mel_bins // LATENT_DOWNSAMPLE_FACTOR) + + @property + def latents_per_second(self) -> float: + return self.sample_rate / self.mel_hop_length / LATENT_DOWNSAMPLE_FACTOR + + @property + def output_sample_rate(self) -> int: + output_rate = getattr(self.vocoder, "output_sample_rate", None) + if output_rate is not None: + return int(output_rate) + upsample_factor = getattr(self.vocoder, "upsample_factor", None) + if upsample_factor is None: + raise AttributeError( + "Vocoder is missing upsample_factor; cannot infer output sample rate" + ) + return int(self.sample_rate * upsample_factor / self.mel_hop_length) + + def memory_required(self, input_shape): + return self.device_manager.patcher.model_size() diff --git a/comfy/ldm/lightricks/vae/causal_audio_autoencoder.py b/comfy/ldm/lightricks/vae/causal_audio_autoencoder.py new file mode 100644 index 000000000..f12b9bb53 --- /dev/null +++ b/comfy/ldm/lightricks/vae/causal_audio_autoencoder.py @@ -0,0 +1,909 @@ +from __future__ import annotations +import torch +from torch import nn +from torch.nn import functional as F +from typing import Optional +from enum import Enum +from .pixel_norm import PixelNorm +import comfy.ops +import logging + +ops = comfy.ops.disable_weight_init + + +class StringConvertibleEnum(Enum): + """ + Base enum class that provides string-to-enum conversion functionality. + + This mixin adds a str_to_enum() class method that handles conversion from + strings, None, or existing enum instances with case-insensitive matching. + """ + + @classmethod + def str_to_enum(cls, value): + """ + Convert a string, enum instance, or None to the appropriate enum member. + + Args: + value: Can be an enum instance of this class, a string, or None + + Returns: + Enum member of this class + + Raises: + ValueError: If the value cannot be converted to a valid enum member + """ + # Already an enum instance of this class + if isinstance(value, cls): + return value + + # None maps to NONE member if it exists + if value is None: + if hasattr(cls, "NONE"): + return cls.NONE + raise ValueError(f"{cls.__name__} does not have a NONE member to map None to") + + # String conversion (case-insensitive) + if isinstance(value, str): + value_lower = value.lower() + + # Try to match against enum values + for member in cls: + # Handle members with None values + if member.value is None: + if value_lower == "none": + return member + # Handle members with string values + elif isinstance(member.value, str) and member.value.lower() == value_lower: + return member + + # Build helpful error message with valid values + valid_values = [] + for member in cls: + if member.value is None: + valid_values.append("none") + elif isinstance(member.value, str): + valid_values.append(member.value) + + raise ValueError(f"Invalid {cls.__name__} string: '{value}'. " f"Valid values are: {valid_values}") + + raise ValueError( + f"Cannot convert type {type(value).__name__} to {cls.__name__} enum. " + f"Expected string, None, or {cls.__name__} instance." + ) + + +class AttentionType(StringConvertibleEnum): + """Enum for specifying the attention mechanism type.""" + + VANILLA = "vanilla" + LINEAR = "linear" + NONE = "none" + + +class CausalityAxis(StringConvertibleEnum): + """Enum for specifying the causality axis in causal convolutions.""" + + NONE = None + WIDTH = "width" + HEIGHT = "height" + WIDTH_COMPATIBILITY = "width-compatibility" + + +def Normalize(in_channels, *, num_groups=32, normtype="group"): + if normtype == "group": + return torch.nn.GroupNorm(num_groups=num_groups, num_channels=in_channels, eps=1e-6, affine=True) + elif normtype == "pixel": + return PixelNorm(dim=1, eps=1e-6) + else: + raise ValueError(f"Invalid normalization type: {normtype}") + + +class CausalConv2d(nn.Module): + """ + A causal 2D convolution. + + This layer ensures that the output at time `t` only depends on inputs + at time `t` and earlier. It achieves this by applying asymmetric padding + to the time dimension (width) before the convolution. + """ + + def __init__( + self, + in_channels, + out_channels, + kernel_size, + stride=1, + dilation=1, + groups=1, + bias=True, + causality_axis: CausalityAxis = CausalityAxis.HEIGHT, + ): + super().__init__() + + self.causality_axis = causality_axis + + # Ensure kernel_size and dilation are tuples + kernel_size = nn.modules.utils._pair(kernel_size) + dilation = nn.modules.utils._pair(dilation) + + # Calculate padding dimensions + pad_h = (kernel_size[0] - 1) * dilation[0] + pad_w = (kernel_size[1] - 1) * dilation[1] + + # The padding tuple for F.pad is (pad_left, pad_right, pad_top, pad_bottom) + match self.causality_axis: + case CausalityAxis.NONE: + self.padding = (pad_w // 2, pad_w - pad_w // 2, pad_h // 2, pad_h - pad_h // 2) + case CausalityAxis.WIDTH | CausalityAxis.WIDTH_COMPATIBILITY: + self.padding = (pad_w, 0, pad_h // 2, pad_h - pad_h // 2) + case CausalityAxis.HEIGHT: + self.padding = (pad_w // 2, pad_w - pad_w // 2, pad_h, 0) + case _: + raise ValueError(f"Invalid causality_axis: {causality_axis}") + + # The internal convolution layer uses no padding, as we handle it manually + self.conv = ops.Conv2d( + in_channels, + out_channels, + kernel_size, + stride=stride, + padding=0, + dilation=dilation, + groups=groups, + bias=bias, + ) + + def forward(self, x): + # Apply causal padding before convolution + x = F.pad(x, self.padding) + return self.conv(x) + + +def make_conv2d( + in_channels, + out_channels, + kernel_size, + stride=1, + padding=None, + dilation=1, + groups=1, + bias=True, + causality_axis: Optional[CausalityAxis] = None, +): + """ + Create a 2D convolution layer that can be either causal or non-causal. + + Args: + in_channels: Number of input channels + out_channels: Number of output channels + kernel_size: Size of the convolution kernel + stride: Convolution stride + padding: Padding (if None, will be calculated based on causal flag) + dilation: Dilation rate + groups: Number of groups for grouped convolution + bias: Whether to use bias + causality_axis: Dimension along which to apply causality. + + Returns: + Either a regular Conv2d or CausalConv2d layer + """ + if causality_axis is not None: + # For causal convolution, padding is handled internally by CausalConv2d + return CausalConv2d(in_channels, out_channels, kernel_size, stride, dilation, groups, bias, causality_axis) + else: + # For non-causal convolution, use symmetric padding if not specified + if padding is None: + if isinstance(kernel_size, int): + padding = kernel_size // 2 + else: + padding = tuple(k // 2 for k in kernel_size) + return ops.Conv2d( + in_channels, + out_channels, + kernel_size, + stride, + padding, + dilation, + groups, + bias, + ) + + +class Upsample(nn.Module): + def __init__(self, in_channels, with_conv, causality_axis: CausalityAxis = CausalityAxis.HEIGHT): + super().__init__() + self.with_conv = with_conv + self.causality_axis = causality_axis + if self.with_conv: + self.conv = make_conv2d(in_channels, in_channels, kernel_size=3, stride=1, causality_axis=causality_axis) + + def forward(self, x): + x = torch.nn.functional.interpolate(x, scale_factor=2.0, mode="nearest") + if self.with_conv: + x = self.conv(x) + # Drop FIRST element in the causal axis to undo encoder's padding, while keeping the length 1 + 2 * n. + # For example, if the input is [0, 1, 2], after interpolation, the output is [0, 0, 1, 1, 2, 2]. + # The causal convolution will pad the first element as [-, -, 0, 0, 1, 1, 2, 2], + # So the output elements rely on the following windows: + # 0: [-,-,0] + # 1: [-,0,0] + # 2: [0,0,1] + # 3: [0,1,1] + # 4: [1,1,2] + # 5: [1,2,2] + # Notice that the first and second elements in the output rely only on the first element in the input, + # while all other elements rely on two elements in the input. + # So we can drop the first element to undo the padding (rather than the last element). + # This is a no-op for non-causal convolutions. + match self.causality_axis: + case CausalityAxis.NONE: + pass # x remains unchanged + case CausalityAxis.HEIGHT: + x = x[:, :, 1:, :] + case CausalityAxis.WIDTH: + x = x[:, :, :, 1:] + case CausalityAxis.WIDTH_COMPATIBILITY: + pass # x remains unchanged + case _: + raise ValueError(f"Invalid causality_axis: {self.causality_axis}") + + return x + + +class Downsample(nn.Module): + """ + A downsampling layer that can use either a strided convolution + or average pooling. Supports standard and causal padding for the + convolutional mode. + """ + + def __init__(self, in_channels, with_conv, causality_axis: CausalityAxis = CausalityAxis.WIDTH): + super().__init__() + self.with_conv = with_conv + self.causality_axis = causality_axis + + if self.causality_axis != CausalityAxis.NONE and not self.with_conv: + raise ValueError("causality is only supported when `with_conv=True`.") + + if self.with_conv: + # Do time downsampling here + # no asymmetric padding in torch conv, must do it ourselves + self.conv = ops.Conv2d(in_channels, in_channels, kernel_size=3, stride=2, padding=0) + + def forward(self, x): + if self.with_conv: + # (pad_left, pad_right, pad_top, pad_bottom) + match self.causality_axis: + case CausalityAxis.NONE: + pad = (0, 1, 0, 1) + case CausalityAxis.WIDTH: + pad = (2, 0, 0, 1) + case CausalityAxis.HEIGHT: + pad = (0, 1, 2, 0) + case CausalityAxis.WIDTH_COMPATIBILITY: + pad = (1, 0, 0, 1) + case _: + raise ValueError(f"Invalid causality_axis: {self.causality_axis}") + + x = torch.nn.functional.pad(x, pad, mode="constant", value=0) + x = self.conv(x) + else: + # This branch is only taken if with_conv=False, which implies causality_axis is NONE. + x = torch.nn.functional.avg_pool2d(x, kernel_size=2, stride=2) + + return x + + +class ResnetBlock(nn.Module): + def __init__( + self, + *, + in_channels, + out_channels=None, + conv_shortcut=False, + dropout, + temb_channels=512, + norm_type="group", + causality_axis: CausalityAxis = CausalityAxis.HEIGHT, + ): + super().__init__() + self.causality_axis = causality_axis + + if self.causality_axis != CausalityAxis.NONE and norm_type == "group": + raise ValueError("Causal ResnetBlock with GroupNorm is not supported.") + self.in_channels = in_channels + out_channels = in_channels if out_channels is None else out_channels + self.out_channels = out_channels + self.use_conv_shortcut = conv_shortcut + + self.norm1 = Normalize(in_channels, normtype=norm_type) + self.non_linearity = nn.SiLU() + self.conv1 = make_conv2d(in_channels, out_channels, kernel_size=3, stride=1, causality_axis=causality_axis) + if temb_channels > 0: + self.temb_proj = ops.Linear(temb_channels, out_channels) + self.norm2 = Normalize(out_channels, normtype=norm_type) + self.dropout = torch.nn.Dropout(dropout) + self.conv2 = make_conv2d(out_channels, out_channels, kernel_size=3, stride=1, causality_axis=causality_axis) + if self.in_channels != self.out_channels: + if self.use_conv_shortcut: + self.conv_shortcut = make_conv2d( + in_channels, out_channels, kernel_size=3, stride=1, causality_axis=causality_axis + ) + else: + self.nin_shortcut = make_conv2d( + in_channels, out_channels, kernel_size=1, stride=1, causality_axis=causality_axis + ) + + def forward(self, x, temb): + h = x + h = self.norm1(h) + h = self.non_linearity(h) + h = self.conv1(h) + + if temb is not None: + h = h + self.temb_proj(self.non_linearity(temb))[:, :, None, None] + + h = self.norm2(h) + h = self.non_linearity(h) + h = self.dropout(h) + h = self.conv2(h) + + if self.in_channels != self.out_channels: + if self.use_conv_shortcut: + x = self.conv_shortcut(x) + else: + x = self.nin_shortcut(x) + + return x + h + + +class AttnBlock(nn.Module): + def __init__(self, in_channels, norm_type="group"): + super().__init__() + self.in_channels = in_channels + + self.norm = Normalize(in_channels, normtype=norm_type) + self.q = ops.Conv2d(in_channels, in_channels, kernel_size=1, stride=1, padding=0) + self.k = ops.Conv2d(in_channels, in_channels, kernel_size=1, stride=1, padding=0) + self.v = ops.Conv2d(in_channels, in_channels, kernel_size=1, stride=1, padding=0) + self.proj_out = ops.Conv2d(in_channels, in_channels, kernel_size=1, stride=1, padding=0) + + def forward(self, x): + h_ = x + h_ = self.norm(h_) + q = self.q(h_) + k = self.k(h_) + v = self.v(h_) + + # compute attention + b, c, h, w = q.shape + q = q.reshape(b, c, h * w).contiguous() + q = q.permute(0, 2, 1).contiguous() # b,hw,c + k = k.reshape(b, c, h * w).contiguous() # b,c,hw + w_ = torch.bmm(q, k).contiguous() # b,hw,hw w[b,i,j]=sum_c q[b,i,c]k[b,c,j] + w_ = w_ * (int(c) ** (-0.5)) + w_ = torch.nn.functional.softmax(w_, dim=2) + + # attend to values + v = v.reshape(b, c, h * w).contiguous() + w_ = w_.permute(0, 2, 1).contiguous() # b,hw,hw (first hw of k, second of q) + h_ = torch.bmm(v, w_).contiguous() # b, c,hw (hw of q) h_[b,c,j] = sum_i v[b,c,i] w_[b,i,j] + h_ = h_.reshape(b, c, h, w).contiguous() + + h_ = self.proj_out(h_) + + return x + h_ + + +def make_attn(in_channels, attn_type="vanilla", norm_type="group"): + # Convert string to enum if needed + attn_type = AttentionType.str_to_enum(attn_type) + + if attn_type != AttentionType.NONE: + logging.info(f"making attention of type '{attn_type.value}' with {in_channels} in_channels") + else: + logging.info(f"making identity attention with {in_channels} in_channels") + + match attn_type: + case AttentionType.VANILLA: + return AttnBlock(in_channels, norm_type=norm_type) + case AttentionType.NONE: + return nn.Identity(in_channels) + case AttentionType.LINEAR: + raise NotImplementedError(f"Attention type {attn_type.value} is not supported yet.") + case _: + raise ValueError(f"Unknown attention type: {attn_type}") + + +class Encoder(nn.Module): + def __init__( + self, + *, + ch, + out_ch, + ch_mult=(1, 2, 4, 8), + num_res_blocks, + attn_resolutions, + dropout=0.0, + resamp_with_conv=True, + in_channels, + resolution, + z_channels, + double_z=True, + attn_type="vanilla", + mid_block_add_attention=True, + norm_type="group", + causality_axis=CausalityAxis.WIDTH.value, + **ignore_kwargs, + ): + super().__init__() + self.ch = ch + self.temb_ch = 0 + self.num_resolutions = len(ch_mult) + self.num_res_blocks = num_res_blocks + self.resolution = resolution + self.in_channels = in_channels + self.z_channels = z_channels + self.double_z = double_z + self.norm_type = norm_type + # Convert string to enum if needed (for config loading) + causality_axis = CausalityAxis.str_to_enum(causality_axis) + self.attn_type = AttentionType.str_to_enum(attn_type) + + # downsampling + self.conv_in = make_conv2d( + in_channels, + self.ch, + kernel_size=3, + stride=1, + causality_axis=causality_axis, + ) + + self.non_linearity = nn.SiLU() + + curr_res = resolution + in_ch_mult = (1,) + tuple(ch_mult) + self.in_ch_mult = in_ch_mult + self.down = nn.ModuleList() + + for i_level in range(self.num_resolutions): + block = nn.ModuleList() + attn = nn.ModuleList() + block_in = ch * in_ch_mult[i_level] + block_out = ch * ch_mult[i_level] + + for _ in range(self.num_res_blocks): + block.append( + ResnetBlock( + in_channels=block_in, + out_channels=block_out, + temb_channels=self.temb_ch, + dropout=dropout, + norm_type=self.norm_type, + causality_axis=causality_axis, + ) + ) + block_in = block_out + if curr_res in attn_resolutions: + attn.append(make_attn(block_in, attn_type=self.attn_type, norm_type=self.norm_type)) + + down = nn.Module() + down.block = block + down.attn = attn + if i_level != self.num_resolutions - 1: + down.downsample = Downsample(block_in, resamp_with_conv, causality_axis=causality_axis) + curr_res = curr_res // 2 + self.down.append(down) + + # middle + self.mid = nn.Module() + self.mid.block_1 = ResnetBlock( + in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout, + norm_type=self.norm_type, + causality_axis=causality_axis, + ) + if mid_block_add_attention: + self.mid.attn_1 = make_attn(block_in, attn_type=self.attn_type, norm_type=self.norm_type) + else: + self.mid.attn_1 = nn.Identity() + self.mid.block_2 = ResnetBlock( + in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout, + norm_type=self.norm_type, + causality_axis=causality_axis, + ) + + # end + self.norm_out = Normalize(block_in, normtype=self.norm_type) + self.conv_out = make_conv2d( + block_in, + 2 * z_channels if double_z else z_channels, + kernel_size=3, + stride=1, + causality_axis=causality_axis, + ) + + def forward(self, x): + """ + Forward pass through the encoder. + + Args: + x: Input tensor of shape [batch, channels, time, n_mels] + + Returns: + Encoded latent representation + """ + feature_maps = [self.conv_in(x)] + + # Process each resolution level (from high to low resolution) + for resolution_level in range(self.num_resolutions): + # Apply residual blocks at current resolution level + for block_idx in range(self.num_res_blocks): + # Apply ResNet block with optional timestep embedding + current_features = self.down[resolution_level].block[block_idx](feature_maps[-1], temb=None) + + # Apply attention if configured for this resolution level + if len(self.down[resolution_level].attn) > 0: + current_features = self.down[resolution_level].attn[block_idx](current_features) + + # Store processed features + feature_maps.append(current_features) + + # Downsample spatial dimensions (except at the final resolution level) + if resolution_level != self.num_resolutions - 1: + downsampled_features = self.down[resolution_level].downsample(feature_maps[-1]) + feature_maps.append(downsampled_features) + + # === MIDDLE PROCESSING PHASE === + # Take the lowest resolution features for middle processing + bottleneck_features = feature_maps[-1] + + # Apply first middle ResNet block + bottleneck_features = self.mid.block_1(bottleneck_features, temb=None) + + # Apply middle attention block + bottleneck_features = self.mid.attn_1(bottleneck_features) + + # Apply second middle ResNet block + bottleneck_features = self.mid.block_2(bottleneck_features, temb=None) + + # === OUTPUT PHASE === + # Normalize the bottleneck features + output_features = self.norm_out(bottleneck_features) + + # Apply non-linearity (SiLU activation) + output_features = self.non_linearity(output_features) + + # Final convolution to produce latent representation + # [batch, channels, time, n_mels] -> [batch, 2 * z_channels if double_z else z_channels, time, n_mels] + return self.conv_out(output_features) + + +class Decoder(nn.Module): + def __init__( + self, + *, + ch, + out_ch, + ch_mult=(1, 2, 4, 8), + num_res_blocks, + attn_resolutions, + dropout=0.0, + resamp_with_conv=True, + in_channels, + resolution, + z_channels, + give_pre_end=False, + tanh_out=False, + attn_type="vanilla", + mid_block_add_attention=True, + norm_type="group", + causality_axis=CausalityAxis.WIDTH.value, + **ignorekwargs, + ): + super().__init__() + self.ch = ch + self.temb_ch = 0 + self.num_resolutions = len(ch_mult) + self.num_res_blocks = num_res_blocks + self.resolution = resolution + self.in_channels = in_channels + self.out_ch = out_ch + self.give_pre_end = give_pre_end + self.tanh_out = tanh_out + self.norm_type = norm_type + self.z_channels = z_channels + # Convert string to enum if needed (for config loading) + causality_axis = CausalityAxis.str_to_enum(causality_axis) + self.attn_type = AttentionType.str_to_enum(attn_type) + + # compute block_in and curr_res at lowest res + block_in = ch * ch_mult[self.num_resolutions - 1] + curr_res = resolution // 2 ** (self.num_resolutions - 1) + self.z_shape = (1, z_channels, curr_res, curr_res) + + # z to block_in + self.conv_in = make_conv2d(z_channels, block_in, kernel_size=3, stride=1, causality_axis=causality_axis) + + self.non_linearity = nn.SiLU() + + # middle + self.mid = nn.Module() + self.mid.block_1 = ResnetBlock( + in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout, + norm_type=self.norm_type, + causality_axis=causality_axis, + ) + if mid_block_add_attention: + self.mid.attn_1 = make_attn(block_in, attn_type=self.attn_type, norm_type=self.norm_type) + else: + self.mid.attn_1 = nn.Identity() + self.mid.block_2 = ResnetBlock( + in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout, + norm_type=self.norm_type, + causality_axis=causality_axis, + ) + + # upsampling + self.up = nn.ModuleList() + for i_level in reversed(range(self.num_resolutions)): + block = nn.ModuleList() + attn = nn.ModuleList() + block_out = ch * ch_mult[i_level] + for _ in range(self.num_res_blocks + 1): + block.append( + ResnetBlock( + in_channels=block_in, + out_channels=block_out, + temb_channels=self.temb_ch, + dropout=dropout, + norm_type=self.norm_type, + causality_axis=causality_axis, + ) + ) + block_in = block_out + if curr_res in attn_resolutions: + attn.append(make_attn(block_in, attn_type=self.attn_type, norm_type=self.norm_type)) + up = nn.Module() + up.block = block + up.attn = attn + if i_level != 0: + up.upsample = Upsample(block_in, resamp_with_conv, causality_axis=causality_axis) + curr_res = curr_res * 2 + self.up.insert(0, up) # prepend to get consistent order + + # end + self.norm_out = Normalize(block_in, normtype=self.norm_type) + self.conv_out = make_conv2d(block_in, out_ch, kernel_size=3, stride=1, causality_axis=causality_axis) + + def _adjust_output_shape(self, decoded_output, target_shape): + """ + Adjust output shape to match target dimensions for variable-length audio. + + This function handles the common case where decoded audio spectrograms need to be + resized to match a specific target shape. + + Args: + decoded_output: Tensor of shape (batch, channels, time, frequency) + target_shape: Target shape tuple (batch, channels, time, frequency) + + Returns: + Tensor adjusted to match target_shape exactly + """ + # Current output shape: (batch, channels, time, frequency) + _, _, current_time, current_freq = decoded_output.shape + _, target_channels, target_time, target_freq = target_shape + + # Step 1: Crop first to avoid exceeding target dimensions + decoded_output = decoded_output[ + :, :target_channels, : min(current_time, target_time), : min(current_freq, target_freq) + ] + + # Step 2: Calculate padding needed for time and frequency dimensions + time_padding_needed = target_time - decoded_output.shape[2] + freq_padding_needed = target_freq - decoded_output.shape[3] + + # Step 3: Apply padding if needed + if time_padding_needed > 0 or freq_padding_needed > 0: + # PyTorch padding format: (pad_left, pad_right, pad_top, pad_bottom) + # For audio: pad_left/right = frequency, pad_top/bottom = time + padding = ( + 0, + max(freq_padding_needed, 0), # frequency padding (left, right) + 0, + max(time_padding_needed, 0), # time padding (top, bottom) + ) + decoded_output = F.pad(decoded_output, padding) + + # Step 4: Final safety crop to ensure exact target shape + decoded_output = decoded_output[:, :target_channels, :target_time, :target_freq] + + return decoded_output + + def get_config(self): + return { + "ch": self.ch, + "out_ch": self.out_ch, + "ch_mult": self.ch_mult, + "num_res_blocks": self.num_res_blocks, + "in_channels": self.in_channels, + "resolution": self.resolution, + "z_channels": self.z_channels, + } + + def forward(self, latent_features, target_shape=None): + """ + Decode latent features back to audio spectrograms. + + Args: + latent_features: Encoded latent representation of shape (batch, channels, height, width) + target_shape: Optional target output shape (batch, channels, time, frequency) + If provided, output will be cropped/padded to match this shape + + Returns: + Reconstructed audio spectrogram of shape (batch, channels, time, frequency) + """ + assert target_shape is not None, "Target shape is required for CausalAudioAutoencoder Decoder" + + # Transform latent features to decoder's internal feature dimension + hidden_features = self.conv_in(latent_features) + + # Middle processing + hidden_features = self.mid.block_1(hidden_features, temb=None) + hidden_features = self.mid.attn_1(hidden_features) + hidden_features = self.mid.block_2(hidden_features, temb=None) + + # Upsampling + # Progressively increase spatial resolution from lowest to highest + for resolution_level in reversed(range(self.num_resolutions)): + # Apply residual blocks at current resolution level + for block_index in range(self.num_res_blocks + 1): + hidden_features = self.up[resolution_level].block[block_index](hidden_features, temb=None) + + if len(self.up[resolution_level].attn) > 0: + hidden_features = self.up[resolution_level].attn[block_index](hidden_features) + + if resolution_level != 0: + hidden_features = self.up[resolution_level].upsample(hidden_features) + + # Output + if self.give_pre_end: + # Return intermediate features before final processing (for debugging/analysis) + decoded_output = hidden_features + else: + # Standard output path: normalize, activate, and convert to output channels + # Final normalization layer + hidden_features = self.norm_out(hidden_features) + + # Apply SiLU (Swish) activation function + hidden_features = self.non_linearity(hidden_features) + + # Final convolution to map to output channels (typically 2 for stereo audio) + decoded_output = self.conv_out(hidden_features) + + # Optional tanh activation to bound output values to [-1, 1] range + if self.tanh_out: + decoded_output = torch.tanh(decoded_output) + + # Adjust shape for audio data + if target_shape is not None: + decoded_output = self._adjust_output_shape(decoded_output, target_shape) + + return decoded_output + + +class processor(nn.Module): + def __init__(self): + super().__init__() + self.register_buffer("std-of-means", torch.empty(128)) + self.register_buffer("mean-of-means", torch.empty(128)) + + def un_normalize(self, x): + return (x * self.get_buffer("std-of-means").to(x)) + self.get_buffer("mean-of-means").to(x) + + def normalize(self, x): + return (x - self.get_buffer("mean-of-means").to(x)) / self.get_buffer("std-of-means").to(x) + + +class CausalAudioAutoencoder(nn.Module): + def __init__(self, config=None): + super().__init__() + + if config is None: + config = self._guess_config() + + # Extract encoder and decoder configs from the new format + model_config = config.get("model", {}).get("params", {}) + variables_config = config.get("variables", {}) + + self.sampling_rate = variables_config.get( + "sampling_rate", + model_config.get("sampling_rate", config.get("sampling_rate", 16000)), + ) + encoder_config = model_config.get("encoder", model_config.get("ddconfig", {})) + decoder_config = model_config.get("decoder", encoder_config) + + # Load mel spectrogram parameters + self.mel_bins = encoder_config.get("mel_bins", 64) + self.mel_hop_length = model_config.get("preprocessing", {}).get("stft", {}).get("hop_length", 160) + self.n_fft = model_config.get("preprocessing", {}).get("stft", {}).get("filter_length", 1024) + + # Store causality configuration at VAE level (not just in encoder internals) + causality_axis_value = encoder_config.get("causality_axis", CausalityAxis.WIDTH.value) + self.causality_axis = CausalityAxis.str_to_enum(causality_axis_value) + self.is_causal = self.causality_axis == CausalityAxis.HEIGHT + + self.encoder = Encoder(**encoder_config) + self.decoder = Decoder(**decoder_config) + + self.per_channel_statistics = processor() + + def _guess_config(self): + encoder_config = { + # Required parameters - based on ltx-video-av-1679000 model metadata + "ch": 128, + "out_ch": 8, + "ch_mult": [1, 2, 4], # Based on metadata: [1, 2, 4] not [1, 2, 4, 8] + "num_res_blocks": 2, + "attn_resolutions": [], # Based on metadata: empty list, no attention + "dropout": 0.0, + "resamp_with_conv": True, + "in_channels": 2, # stereo + "resolution": 256, + "z_channels": 8, + "double_z": True, + "attn_type": "vanilla", + "mid_block_add_attention": False, # Based on metadata: false + "norm_type": "pixel", + "causality_axis": "height", # Based on metadata + "mel_bins": 64, # Based on metadata: mel_bins = 64 + } + + decoder_config = { + # Inherits encoder config, can override specific params + **encoder_config, + "out_ch": 2, # Stereo audio output (2 channels) + "give_pre_end": False, + "tanh_out": False, + } + + config = { + "_class_name": "CausalAudioAutoencoder", + "sampling_rate": 16000, + "model": { + "params": { + "encoder": encoder_config, + "decoder": decoder_config, + } + }, + } + + return config + + def get_config(self): + return { + "sampling_rate": self.sampling_rate, + "mel_bins": self.mel_bins, + "mel_hop_length": self.mel_hop_length, + "n_fft": self.n_fft, + "causality_axis": self.causality_axis.value, + "is_causal": self.is_causal, + } + + def encode(self, x): + return self.encoder(x) + + def decode(self, x, target_shape=None): + return self.decoder(x, target_shape=target_shape) diff --git a/comfy/ldm/lightricks/vocoders/vocoder.py b/comfy/ldm/lightricks/vocoders/vocoder.py new file mode 100644 index 000000000..b1f15f2c5 --- /dev/null +++ b/comfy/ldm/lightricks/vocoders/vocoder.py @@ -0,0 +1,213 @@ +import torch +import torch.nn.functional as F +import torch.nn as nn +import comfy.ops +import numpy as np + +ops = comfy.ops.disable_weight_init + +LRELU_SLOPE = 0.1 + +def get_padding(kernel_size, dilation=1): + return int((kernel_size * dilation - dilation) / 2) + + +class ResBlock1(torch.nn.Module): + def __init__(self, channels, kernel_size=3, dilation=(1, 3, 5)): + super(ResBlock1, self).__init__() + self.convs1 = nn.ModuleList( + [ + ops.Conv1d( + channels, + channels, + kernel_size, + 1, + dilation=dilation[0], + padding=get_padding(kernel_size, dilation[0]), + ), + ops.Conv1d( + channels, + channels, + kernel_size, + 1, + dilation=dilation[1], + padding=get_padding(kernel_size, dilation[1]), + ), + ops.Conv1d( + channels, + channels, + kernel_size, + 1, + dilation=dilation[2], + padding=get_padding(kernel_size, dilation[2]), + ), + ] + ) + + self.convs2 = nn.ModuleList( + [ + ops.Conv1d( + channels, + channels, + kernel_size, + 1, + dilation=1, + padding=get_padding(kernel_size, 1), + ), + ops.Conv1d( + channels, + channels, + kernel_size, + 1, + dilation=1, + padding=get_padding(kernel_size, 1), + ), + ops.Conv1d( + channels, + channels, + kernel_size, + 1, + dilation=1, + padding=get_padding(kernel_size, 1), + ), + ] + ) + + def forward(self, x): + for c1, c2 in zip(self.convs1, self.convs2): + xt = F.leaky_relu(x, LRELU_SLOPE) + xt = c1(xt) + xt = F.leaky_relu(xt, LRELU_SLOPE) + xt = c2(xt) + x = xt + x + return x + + +class ResBlock2(torch.nn.Module): + def __init__(self, channels, kernel_size=3, dilation=(1, 3)): + super(ResBlock2, self).__init__() + self.convs = nn.ModuleList( + [ + ops.Conv1d( + channels, + channels, + kernel_size, + 1, + dilation=dilation[0], + padding=get_padding(kernel_size, dilation[0]), + ), + ops.Conv1d( + channels, + channels, + kernel_size, + 1, + dilation=dilation[1], + padding=get_padding(kernel_size, dilation[1]), + ), + ] + ) + + def forward(self, x): + for c in self.convs: + xt = F.leaky_relu(x, LRELU_SLOPE) + xt = c(xt) + x = xt + x + return x + + +class Vocoder(torch.nn.Module): + """ + Vocoder model for synthesizing audio from spectrograms, based on: https://github.com/jik876/hifi-gan. + + """ + + def __init__(self, config=None): + super(Vocoder, self).__init__() + + if config is None: + config = self.get_default_config() + + resblock_kernel_sizes = config.get("resblock_kernel_sizes", [3, 7, 11]) + upsample_rates = config.get("upsample_rates", [6, 5, 2, 2, 2]) + upsample_kernel_sizes = config.get("upsample_kernel_sizes", [16, 15, 8, 4, 4]) + resblock_dilation_sizes = config.get("resblock_dilation_sizes", [[1, 3, 5], [1, 3, 5], [1, 3, 5]]) + upsample_initial_channel = config.get("upsample_initial_channel", 1024) + stereo = config.get("stereo", True) + resblock = config.get("resblock", "1") + + self.output_sample_rate = config.get("output_sample_rate") + self.num_kernels = len(resblock_kernel_sizes) + self.num_upsamples = len(upsample_rates) + in_channels = 128 if stereo else 64 + self.conv_pre = ops.Conv1d(in_channels, upsample_initial_channel, 7, 1, padding=3) + resblock_class = ResBlock1 if resblock == "1" else ResBlock2 + + self.ups = nn.ModuleList() + for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)): + self.ups.append( + ops.ConvTranspose1d( + upsample_initial_channel // (2**i), + upsample_initial_channel // (2 ** (i + 1)), + k, + u, + padding=(k - u) // 2, + ) + ) + + self.resblocks = nn.ModuleList() + for i in range(len(self.ups)): + ch = upsample_initial_channel // (2 ** (i + 1)) + for _, (k, d) in enumerate(zip(resblock_kernel_sizes, resblock_dilation_sizes)): + self.resblocks.append(resblock_class(ch, k, d)) + + out_channels = 2 if stereo else 1 + self.conv_post = ops.Conv1d(ch, out_channels, 7, 1, padding=3) + + self.upsample_factor = np.prod([self.ups[i].stride[0] for i in range(len(self.ups))]) + + def get_default_config(self): + """Generate default configuration for the vocoder.""" + + config = { + "resblock_kernel_sizes": [3, 7, 11], + "upsample_rates": [6, 5, 2, 2, 2], + "upsample_kernel_sizes": [16, 15, 8, 4, 4], + "resblock_dilation_sizes": [[1, 3, 5], [1, 3, 5], [1, 3, 5]], + "upsample_initial_channel": 1024, + "stereo": True, + "resblock": "1", + } + + return config + + def forward(self, x): + """ + Forward pass of the vocoder. + + Args: + x: Input spectrogram tensor. Can be: + - 3D: (batch_size, channels, time_steps) for mono + - 4D: (batch_size, 2, channels, time_steps) for stereo + + Returns: + Audio tensor of shape (batch_size, out_channels, audio_length) + """ + if x.dim() == 4: # stereo + assert x.shape[1] == 2, "Input must have 2 channels for stereo" + x = torch.cat((x[:, 0, :, :], x[:, 1, :, :]), dim=1) + x = self.conv_pre(x) + for i in range(self.num_upsamples): + x = F.leaky_relu(x, LRELU_SLOPE) + x = self.ups[i](x) + xs = None + for j in range(self.num_kernels): + if xs is None: + xs = self.resblocks[i * self.num_kernels + j](x) + else: + xs += self.resblocks[i * self.num_kernels + j](x) + x = xs / self.num_kernels + x = F.leaky_relu(x) + x = self.conv_post(x) + x = torch.tanh(x) + + return x diff --git a/comfy/model_base.py b/comfy/model_base.py index c4f3c0639..49efd700b 100644 --- a/comfy/model_base.py +++ b/comfy/model_base.py @@ -20,6 +20,7 @@ import comfy.ldm.hunyuan3dv2_1 import comfy.ldm.hunyuan3dv2_1.hunyuandit import torch import logging +import comfy.ldm.lightricks.av_model from comfy.ldm.modules.diffusionmodules.openaimodel import UNetModel, Timestep from comfy.ldm.cascade.stage_c import StageC from comfy.ldm.cascade.stage_b import StageB @@ -946,7 +947,7 @@ class GenmoMochi(BaseModel): class LTXV(BaseModel): def __init__(self, model_config, model_type=ModelType.FLUX, device=None): - super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.lightricks.model.LTXVModel) #TODO + super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.lightricks.model.LTXVModel) def extra_conds(self, **kwargs): out = super().extra_conds(**kwargs) @@ -977,6 +978,60 @@ class LTXV(BaseModel): def scale_latent_inpaint(self, sigma, noise, latent_image, **kwargs): return latent_image +class LTXAV(BaseModel): + def __init__(self, model_config, model_type=ModelType.FLUX, device=None): + super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.lightricks.av_model.LTXAVModel) #TODO + + def extra_conds(self, **kwargs): + out = super().extra_conds(**kwargs) + attention_mask = kwargs.get("attention_mask", None) + if attention_mask is not None: + out['attention_mask'] = comfy.conds.CONDRegular(attention_mask) + cross_attn = kwargs.get("cross_attn", None) + if cross_attn is not None: + out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn) + + out['frame_rate'] = comfy.conds.CONDConstant(kwargs.get("frame_rate", 25)) + + denoise_mask = kwargs.get("concat_mask", kwargs.get("denoise_mask", None)) + + audio_denoise_mask = None + if denoise_mask is not None and "latent_shapes" in kwargs: + denoise_mask = utils.unpack_latents(denoise_mask, kwargs["latent_shapes"]) + if len(denoise_mask) > 1: + audio_denoise_mask = denoise_mask[1] + denoise_mask = denoise_mask[0] + + if denoise_mask is not None: + out["denoise_mask"] = comfy.conds.CONDRegular(denoise_mask) + + if audio_denoise_mask is not None: + out["audio_denoise_mask"] = comfy.conds.CONDRegular(audio_denoise_mask) + + keyframe_idxs = kwargs.get("keyframe_idxs", None) + if keyframe_idxs is not None: + out['keyframe_idxs'] = comfy.conds.CONDRegular(keyframe_idxs) + + latent_shapes = kwargs.get("latent_shapes", None) + if latent_shapes is not None: + out['latent_shapes'] = comfy.conds.CONDConstant(latent_shapes) + + return out + + def process_timestep(self, timestep, x, denoise_mask=None, audio_denoise_mask=None, **kwargs): + v_timestep = timestep + a_timestep = timestep + + if denoise_mask is not None: + v_timestep = self.diffusion_model.patchifier.patchify(((denoise_mask) * timestep.view([timestep.shape[0]] + [1] * (denoise_mask.ndim - 1)))[:, :1])[0] + if audio_denoise_mask is not None: + a_timestep = self.diffusion_model.a_patchifier.patchify(((audio_denoise_mask) * timestep.view([timestep.shape[0]] + [1] * (audio_denoise_mask.ndim - 1)))[:, :1, :, :1])[0] + + return v_timestep, a_timestep + + def scale_latent_inpaint(self, sigma, noise, latent_image, **kwargs): + return latent_image + class HunyuanVideo(BaseModel): def __init__(self, model_config, model_type=ModelType.FLOW, device=None): super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.hunyuan_video.model.HunyuanVideo) diff --git a/comfy/model_detection.py b/comfy/model_detection.py index 539e296ed..0853b3aec 100644 --- a/comfy/model_detection.py +++ b/comfy/model_detection.py @@ -305,7 +305,7 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): if '{}adaln_single.emb.timestep_embedder.linear_1.bias'.format(key_prefix) in state_dict_keys: #Lightricks ltxv dit_config = {} - dit_config["image_model"] = "ltxv" + dit_config["image_model"] = "ltxav" if f'{key_prefix}audio_adaln_single.linear.weight' in state_dict_keys else "ltxv" dit_config["num_layers"] = count_blocks(state_dict_keys, '{}transformer_blocks.'.format(key_prefix) + '{}.') shape = state_dict['{}transformer_blocks.0.attn2.to_k.weight'.format(key_prefix)].shape dit_config["attention_head_dim"] = shape[0] // 32 diff --git a/comfy/sd.py b/comfy/sd.py index 7de7dd9c6..32157e18b 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -1041,7 +1041,8 @@ class TEModel(Enum): MISTRAL3_24B_PRUNED_FLUX2 = 15 QWEN3_4B = 16 QWEN3_2B = 17 - JINA_CLIP_2 = 18 + GEMMA_3_12B = 18 + JINA_CLIP_2 = 19 def detect_te_model(sd): @@ -1067,6 +1068,8 @@ def detect_te_model(sd): return TEModel.BYT5_SMALL_GLYPH return TEModel.T5_BASE if 'model.layers.0.post_feedforward_layernorm.weight' in sd: + if 'model.layers.47.self_attn.q_norm.weight' in sd: + return TEModel.GEMMA_3_12B if 'model.layers.0.self_attn.q_norm.weight' in sd: return TEModel.GEMMA_3_4B return TEModel.GEMMA_2_2B @@ -1271,6 +1274,10 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip elif clip_type == CLIPType.KANDINSKY5_IMAGE: clip_target.clip = comfy.text_encoders.kandinsky5.te(**llama_detect(clip_data)) clip_target.tokenizer = comfy.text_encoders.kandinsky5.Kandinsky5TokenizerImage + elif clip_type == CLIPType.LTXV: + clip_target.clip = comfy.text_encoders.lt.ltxav_te(**llama_detect(clip_data)) + clip_target.tokenizer = comfy.text_encoders.lt.LTXAVGemmaTokenizer + tokenizer_data["spiece_model"] = clip_data[0].get("spiece_model", None) elif clip_type == CLIPType.NEWBIE: clip_target.clip = comfy.text_encoders.newbie.te(**llama_detect(clip_data)) clip_target.tokenizer = comfy.text_encoders.newbie.NewBieTokenizer diff --git a/comfy/supported_models.py b/comfy/supported_models.py index 1888f35ba..ee9a79001 100644 --- a/comfy/supported_models.py +++ b/comfy/supported_models.py @@ -836,6 +836,21 @@ class LTXV(supported_models_base.BASE): t5_detect = comfy.text_encoders.sd3_clip.t5_xxl_detect(state_dict, "{}t5xxl.transformer.".format(pref)) return supported_models_base.ClipTarget(comfy.text_encoders.lt.LTXVT5Tokenizer, comfy.text_encoders.lt.ltxv_te(**t5_detect)) +class LTXAV(LTXV): + unet_config = { + "image_model": "ltxav", + } + + latent_format = latent_formats.LTXAV + + def __init__(self, unet_config): + super().__init__(unet_config) + self.memory_usage_factor = 0.055 # TODO + + def get_model(self, state_dict, prefix="", device=None): + out = model_base.LTXAV(self, device=device) + return out + class HunyuanVideo(supported_models_base.BASE): unet_config = { "image_model": "hunyuan_video", @@ -1536,6 +1551,6 @@ class Kandinsky5Image(Kandinsky5): return supported_models_base.ClipTarget(comfy.text_encoders.kandinsky5.Kandinsky5TokenizerImage, comfy.text_encoders.kandinsky5.te(**hunyuan_detect)) -models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanVideo15_SR_Distilled, HunyuanVideo15, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, ZImage, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, Omnigen2, QwenImage, Flux2, Kandinsky5Image, Kandinsky5] +models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, LTXAV, HunyuanVideo15_SR_Distilled, HunyuanVideo15, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, ZImage, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, Omnigen2, QwenImage, Flux2, Kandinsky5Image, Kandinsky5] models += [SVD_img2vid] diff --git a/comfy/text_encoders/llama.py b/comfy/text_encoders/llama.py index faa4e1de8..76731576b 100644 --- a/comfy/text_encoders/llama.py +++ b/comfy/text_encoders/llama.py @@ -7,6 +7,7 @@ import math from comfy.ldm.modules.attention import optimized_attention_for_device import comfy.model_management import comfy.ldm.common_dit +import comfy.clip_model from . import qwen_vl @@ -188,6 +189,31 @@ class Gemma3_4B_Config: rope_scale = [8.0, 1.0] final_norm: bool = True +@dataclass +class Gemma3_12B_Config: + vocab_size: int = 262208 + hidden_size: int = 3840 + intermediate_size: int = 15360 + num_hidden_layers: int = 48 + num_attention_heads: int = 16 + num_key_value_heads: int = 8 + max_position_embeddings: int = 131072 + rms_norm_eps: float = 1e-6 + rope_theta = [1000000.0, 10000.0] + transformer_type: str = "gemma3" + head_dim = 256 + rms_norm_add = True + mlp_activation = "gelu_pytorch_tanh" + qkv_bias = False + rope_dims = None + q_norm = "gemma3" + k_norm = "gemma3" + sliding_attention = [1024, 1024, 1024, 1024, 1024, False] + rope_scale = [8.0, 1.0] + final_norm: bool = True + vision_config = {"num_channels": 3, "hidden_act": "gelu_pytorch_tanh", "hidden_size": 1152, "image_size": 896, "intermediate_size": 4304, "model_type": "siglip_vision_model", "num_attention_heads": 16, "num_hidden_layers": 27, "patch_size": 14} + mm_tokens_per_image = 256 + class RMSNorm(nn.Module): def __init__(self, dim: int, eps: float = 1e-5, add=False, device=None, dtype=None): super().__init__() @@ -520,6 +546,41 @@ class Llama2_(nn.Module): return x, intermediate + +class Gemma3MultiModalProjector(torch.nn.Module): + def __init__(self, config, dtype, device, operations): + super().__init__() + + self.mm_input_projection_weight = nn.Parameter( + torch.empty(config.vision_config["hidden_size"], config.hidden_size, device=device, dtype=dtype) + ) + + self.mm_soft_emb_norm = RMSNorm(config.vision_config["hidden_size"], eps=config.rms_norm_eps, add=config.rms_norm_add, device=device, dtype=dtype) + + self.patches_per_image = int(config.vision_config["image_size"] // config.vision_config["patch_size"]) + self.tokens_per_side = int(config.mm_tokens_per_image**0.5) + self.kernel_size = self.patches_per_image // self.tokens_per_side + self.avg_pool = nn.AvgPool2d(kernel_size=self.kernel_size, stride=self.kernel_size) + + def forward(self, vision_outputs: torch.Tensor): + batch_size, _, seq_length = vision_outputs.shape + + reshaped_vision_outputs = vision_outputs.transpose(1, 2) + reshaped_vision_outputs = reshaped_vision_outputs.reshape( + batch_size, seq_length, self.patches_per_image, self.patches_per_image + ) + reshaped_vision_outputs = reshaped_vision_outputs.contiguous() + + pooled_vision_outputs = self.avg_pool(reshaped_vision_outputs) + pooled_vision_outputs = pooled_vision_outputs.flatten(2) + pooled_vision_outputs = pooled_vision_outputs.transpose(1, 2) + + normed_vision_outputs = self.mm_soft_emb_norm(pooled_vision_outputs) + + projected_vision_outputs = torch.matmul(normed_vision_outputs, comfy.model_management.cast_to_device(self.mm_input_projection_weight, device=normed_vision_outputs.device, dtype=normed_vision_outputs.dtype)) + return projected_vision_outputs.type_as(vision_outputs) + + class BaseLlama: def get_input_embeddings(self): return self.model.embed_tokens @@ -636,3 +697,21 @@ class Gemma3_4B(BaseLlama, torch.nn.Module): self.model = Llama2_(config, device=device, dtype=dtype, ops=operations) self.dtype = dtype + +class Gemma3_12B(BaseLlama, torch.nn.Module): + def __init__(self, config_dict, dtype, device, operations): + super().__init__() + config = Gemma3_12B_Config(**config_dict) + self.num_layers = config.num_hidden_layers + + self.model = Llama2_(config, device=device, dtype=dtype, ops=operations) + self.multi_modal_projector = Gemma3MultiModalProjector(config, dtype, device, operations) + self.vision_model = comfy.clip_model.CLIPVision(config.vision_config, dtype, device, operations) + self.dtype = dtype + self.image_size = config.vision_config["image_size"] + + def preprocess_embed(self, embed, device): + if embed["type"] == "image": + image = comfy.clip_model.clip_preprocess(embed["data"], size=self.image_size, mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5], crop=True) + return self.multi_modal_projector(self.vision_model(image.to(device, dtype=torch.float32))[0]), None + return None, None diff --git a/comfy/text_encoders/lt.py b/comfy/text_encoders/lt.py index 48ea67e67..2c2d453e8 100644 --- a/comfy/text_encoders/lt.py +++ b/comfy/text_encoders/lt.py @@ -1,7 +1,11 @@ from comfy import sd1_clip import os from transformers import T5TokenizerFast +from .spiece_tokenizer import SPieceTokenizer import comfy.text_encoders.genmo +from comfy.ldm.lightricks.embeddings_connector import Embeddings1DConnector +import torch +import comfy.utils class T5XXLTokenizer(sd1_clip.SDTokenizer): def __init__(self, embedding_directory=None, tokenizer_data={}): @@ -16,3 +20,110 @@ class LTXVT5Tokenizer(sd1_clip.SD1Tokenizer): def ltxv_te(*args, **kwargs): return comfy.text_encoders.genmo.mochi_te(*args, **kwargs) + + +class Gemma3_12BTokenizer(sd1_clip.SDTokenizer): + def __init__(self, embedding_directory=None, tokenizer_data={}): + tokenizer = tokenizer_data.get("spiece_model", None) + super().__init__(tokenizer, pad_with_end=False, embedding_size=3840, embedding_key='gemma3_12b', tokenizer_class=SPieceTokenizer, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, tokenizer_args={"add_bos": True, "add_eos": False}, tokenizer_data=tokenizer_data) + + def state_dict(self): + return {"spiece_model": self.tokenizer.serialize_model()} + +class LTXAVGemmaTokenizer(sd1_clip.SD1Tokenizer): + def __init__(self, embedding_directory=None, tokenizer_data={}): + super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, name="gemma3_12b", tokenizer=Gemma3_12BTokenizer) + +class Gemma3_12BModel(sd1_clip.SDClipModel): + def __init__(self, device="cpu", layer="all", layer_idx=None, dtype=None, attention_mask=True, model_options={}): + llama_scaled_fp8 = model_options.get("gemma_scaled_fp8", None) + if llama_scaled_fp8 is not None: + model_options = model_options.copy() + model_options["scaled_fp8"] = llama_scaled_fp8 + + super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"start": 2, "pad": 0}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Gemma3_12B, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options) + + def tokenize_with_weights(self, text, return_word_ids=False, llama_template="{}", image_embeds=None, **kwargs): + text = llama_template.format(text) + text_tokens = super().tokenize_with_weights(text, return_word_ids) + embed_count = 0 + for k in text_tokens: + tt = text_tokens[k] + for r in tt: + for i in range(len(r)): + if r[i][0] == 262144: + if image_embeds is not None and embed_count < image_embeds.shape[0]: + r[i] = ({"type": "embedding", "data": image_embeds[embed_count], "original_type": "image"},) + r[i][1:] + embed_count += 1 + return text_tokens + +class LTXAVTEModel(torch.nn.Module): + def __init__(self, dtype_llama=None, device="cpu", dtype=None, model_options={}): + super().__init__() + self.dtypes = set() + self.dtypes.add(dtype) + + self.gemma3_12b = Gemma3_12BModel(device=device, dtype=dtype_llama, model_options=model_options, layer="all", layer_idx=None) + self.dtypes.add(dtype_llama) + + operations = self.gemma3_12b.operations # TODO + self.text_embedding_projection = operations.Linear(3840 * 49, 3840, bias=False, dtype=dtype, device=device) + + self.audio_embeddings_connector = Embeddings1DConnector( + split_rope=True, + double_precision_rope=True, + dtype=dtype, + device=device, + operations=operations, + ) + + self.video_embeddings_connector = Embeddings1DConnector( + split_rope=True, + double_precision_rope=True, + dtype=dtype, + device=device, + operations=operations, + ) + + def set_clip_options(self, options): + self.gemma3_12b.set_clip_options(options) + + def reset_clip_options(self): + self.gemma3_12b.reset_clip_options() + + def encode_token_weights(self, token_weight_pairs): + token_weight_pairs = token_weight_pairs["gemma3_12b"] + + out, pooled, extra = self.gemma3_12b.encode_token_weights(token_weight_pairs) + out_device = out.device + out = out.movedim(1, -1).to(self.text_embedding_projection.weight.device) + out = 8.0 * (out - out.mean(dim=(1, 2), keepdim=True)) / (out.amax(dim=(1, 2), keepdim=True) - out.amin(dim=(1, 2), keepdim=True) + 1e-6) + out = out.reshape((out.shape[0], out.shape[1], -1)) + out = self.text_embedding_projection(out) + out_vid = self.video_embeddings_connector(out)[0] + out_audio = self.audio_embeddings_connector(out)[0] + out = torch.concat((out_vid, out_audio), dim=-1) + + return out.to(out_device), pooled + + def load_sd(self, sd): + if "model.layers.47.self_attn.q_norm.weight" in sd: + return self.gemma3_12b.load_sd(sd) + else: + sdo = comfy.utils.state_dict_prefix_replace(sd, {"text_embedding_projection.aggregate_embed.weight": "text_embedding_projection.weight", "model.diffusion_model.video_embeddings_connector.": "video_embeddings_connector.", "model.diffusion_model.audio_embeddings_connector.": "audio_embeddings_connector."}, filter_keys=True) + if len(sdo) == 0: + sdo = sd + + return self.load_state_dict(sdo, strict=False) + + +def ltxav_te(dtype_llama=None, llama_scaled_fp8=None): + class LTXAVTEModel_(LTXAVTEModel): + def __init__(self, device="cpu", dtype=None, model_options={}): + if llama_scaled_fp8 is not None and "llama_scaled_fp8" not in model_options: + model_options = model_options.copy() + model_options["llama_scaled_fp8"] = llama_scaled_fp8 + if dtype_llama is not None: + dtype = dtype_llama + super().__init__(dtype_llama=dtype_llama, device=device, dtype=dtype, model_options=model_options) + return LTXAVTEModel_ diff --git a/comfy/utils.py b/comfy/utils.py index e4162d7ac..ffa98c9b1 100644 --- a/comfy/utils.py +++ b/comfy/utils.py @@ -1198,7 +1198,7 @@ def unpack_latents(combined_latent, latent_shapes): combined_latent = combined_latent[:, :, cut:] output_tensors.append(tens.reshape([tens.shape[0]] + list(shape)[1:])) else: - output_tensors = combined_latent + output_tensors = [combined_latent] return output_tensors def detect_layer_quantization(state_dict, prefix): diff --git a/comfy_extras/nodes_audio.py b/comfy_extras/nodes_audio.py index c7916443c..94ad5e8a8 100644 --- a/comfy_extras/nodes_audio.py +++ b/comfy_extras/nodes_audio.py @@ -112,7 +112,7 @@ class VAEDecodeAudio(IO.ComfyNode): std = torch.std(audio, dim=[1,2], keepdim=True) * 5.0 std[std < 1.0] = 1.0 audio /= std - return IO.NodeOutput({"waveform": audio, "sample_rate": 44100}) + return IO.NodeOutput({"waveform": audio, "sample_rate": 44100 if "sample_rate" not in samples else samples["sample_rate"]}) decode = execute # TODO: remove diff --git a/comfy_extras/nodes_hunyuan.py b/comfy_extras/nodes_hunyuan.py index 32be182f1..ceff657d3 100644 --- a/comfy_extras/nodes_hunyuan.py +++ b/comfy_extras/nodes_hunyuan.py @@ -5,7 +5,9 @@ import comfy.model_management from typing_extensions import override from comfy_api.latest import ComfyExtension, io from comfy.ldm.hunyuan_video.upsampler import HunyuanVideo15SRModel +from comfy.ldm.lightricks.latent_upsampler import LatentUpsampler import folder_paths +import json class CLIPTextEncodeHunyuanDiT(io.ComfyNode): @classmethod @@ -186,7 +188,7 @@ class LatentUpscaleModelLoader(io.ComfyNode): @classmethod def execute(cls, model_name) -> io.NodeOutput: model_path = folder_paths.get_full_path_or_raise("latent_upscale_models", model_name) - sd = comfy.utils.load_torch_file(model_path, safe_load=True) + sd, metadata = comfy.utils.load_torch_file(model_path, safe_load=True, return_metadata=True) if "blocks.0.block.0.conv.weight" in sd: config = { @@ -197,6 +199,8 @@ class LatentUpscaleModelLoader(io.ComfyNode): "global_residual": False, } model_type = "720p" + model = HunyuanVideo15SRModel(model_type, config) + model.load_sd(sd) elif "up.0.block.0.conv1.conv.weight" in sd: sd = {key.replace("nin_shortcut", "nin_shortcut.conv", 1): value for key, value in sd.items()} config = { @@ -205,9 +209,12 @@ class LatentUpscaleModelLoader(io.ComfyNode): "block_out_channels": tuple(sd[f"up.{i}.block.0.conv1.conv.weight"].shape[0] for i in range(len([k for k in sd.keys() if k.startswith("up.") and k.endswith(".block.0.conv1.conv.weight")]))), } model_type = "1080p" - - model = HunyuanVideo15SRModel(model_type, config) - model.load_sd(sd) + model = HunyuanVideo15SRModel(model_type, config) + model.load_sd(sd) + elif "post_upsample_res_blocks.0.conv2.bias" in sd: + config = json.loads(metadata["config"]) + model = LatentUpsampler.from_config(config).to(dtype=comfy.model_management.vae_dtype(allowed_dtypes=[torch.bfloat16, torch.float32])) + model.load_state_dict(sd) return io.NodeOutput(model) diff --git a/comfy_extras/nodes_lt.py b/comfy_extras/nodes_lt.py index 50da5f4eb..b91a22309 100644 --- a/comfy_extras/nodes_lt.py +++ b/comfy_extras/nodes_lt.py @@ -81,6 +81,59 @@ class LTXVImgToVideo(io.ComfyNode): generate = execute # TODO: remove +class LTXVImgToVideoInplace(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="LTXVImgToVideoInplace", + category="conditioning/video_models", + inputs=[ + io.Vae.Input("vae"), + io.Image.Input("image"), + io.Latent.Input("latent"), + io.Float.Input("strength", default=1.0, min=0.0, max=1.0), + io.Boolean.Input("bypass", default=False, tooltip="Bypass the conditioning.") + ], + outputs=[ + io.Latent.Output(display_name="latent"), + ], + ) + + @classmethod + def execute(cls, vae, image, latent, strength, bypass=False) -> io.NodeOutput: + if bypass: + return (latent,) + + samples = latent["samples"] + _, height_scale_factor, width_scale_factor = ( + vae.downscale_index_formula + ) + + batch, _, latent_frames, latent_height, latent_width = samples.shape + width = latent_width * width_scale_factor + height = latent_height * height_scale_factor + + if image.shape[1] != height or image.shape[2] != width: + pixels = comfy.utils.common_upscale(image.movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1) + else: + pixels = image + encode_pixels = pixels[:, :, :, :3] + t = vae.encode(encode_pixels) + + samples[:, :, :t.shape[2]] = t + + conditioning_latent_frames_mask = torch.ones( + (batch, 1, latent_frames, 1, 1), + dtype=torch.float32, + device=samples.device, + ) + conditioning_latent_frames_mask[:, :, :t.shape[2]] = 1.0 - strength + + return io.NodeOutput({"samples": samples, "noise_mask": conditioning_latent_frames_mask}) + + generate = execute # TODO: remove + + def conditioning_get_any_value(conditioning, key, default=None): for t in conditioning: if key in t[1]: @@ -106,12 +159,12 @@ def get_keyframe_idxs(cond): keyframe_idxs = conditioning_get_any_value(cond, "keyframe_idxs", None) if keyframe_idxs is None: return None, 0 - num_keyframes = torch.unique(keyframe_idxs[:, 0]).shape[0] + # keyframe_idxs contains start/end positions (last dimension), checking for unqiue values only for start + num_keyframes = torch.unique(keyframe_idxs[:, 0, :, 0]).shape[0] return keyframe_idxs, num_keyframes class LTXVAddGuide(io.ComfyNode): - NUM_PREFIX_FRAMES = 2 - PATCHIFIER = SymmetricPatchifier(1) + PATCHIFIER = SymmetricPatchifier(1, start_end=True) @classmethod def define_schema(cls): @@ -182,26 +235,35 @@ class LTXVAddGuide(io.ComfyNode): return node_helpers.conditioning_set_values(cond, {"keyframe_idxs": keyframe_idxs}) @classmethod - def append_keyframe(cls, positive, negative, frame_idx, latent_image, noise_mask, guiding_latent, strength, scale_factors): - _, latent_idx = cls.get_latent_index( - cond=positive, - latent_length=latent_image.shape[2], - guide_length=guiding_latent.shape[2], - frame_idx=frame_idx, - scale_factors=scale_factors, - ) - noise_mask[:, :, latent_idx:latent_idx + guiding_latent.shape[2]] = 1.0 + def append_keyframe(cls, positive, negative, frame_idx, latent_image, noise_mask, guiding_latent, strength, scale_factors, guide_mask=None, in_channels=128): + if latent_image.shape[1] != in_channels or guiding_latent.shape[1] != in_channels: + raise ValueError("Adding guide to a combined AV latent is not supported.") positive = cls.add_keyframe_index(positive, frame_idx, guiding_latent, scale_factors) negative = cls.add_keyframe_index(negative, frame_idx, guiding_latent, scale_factors) - mask = torch.full( - (noise_mask.shape[0], 1, guiding_latent.shape[2], noise_mask.shape[3], noise_mask.shape[4]), - 1.0 - strength, - dtype=noise_mask.dtype, - device=noise_mask.device, - ) + if guide_mask is not None: + target_h = max(noise_mask.shape[3], guide_mask.shape[3]) + target_w = max(noise_mask.shape[4], guide_mask.shape[4]) + if noise_mask.shape[3] == 1 or noise_mask.shape[4] == 1: + noise_mask = noise_mask.expand(-1, -1, -1, target_h, target_w) + + if guide_mask.shape[3] == 1 or guide_mask.shape[4] == 1: + guide_mask = guide_mask.expand(-1, -1, -1, target_h, target_w) + mask = guide_mask - strength + else: + mask = torch.full( + (noise_mask.shape[0], 1, guiding_latent.shape[2], noise_mask.shape[3], noise_mask.shape[4]), + 1.0 - strength, + dtype=noise_mask.dtype, + device=noise_mask.device, + ) + # This solves audio video combined latent case where latent_image has audio latent concatenated + # in channel dimension with video latent. The solution is to pad guiding latent accordingly. + if latent_image.shape[1] > guiding_latent.shape[1]: + pad_len = latent_image.shape[1] - guiding_latent.shape[1] + guiding_latent = torch.nn.functional.pad(guiding_latent, pad=(0, 0, 0, 0, 0, 0, 0, pad_len), value=0) latent_image = torch.cat([latent_image, guiding_latent], dim=2) noise_mask = torch.cat([noise_mask, mask], dim=2) return positive, negative, latent_image, noise_mask @@ -238,33 +300,17 @@ class LTXVAddGuide(io.ComfyNode): frame_idx, latent_idx = cls.get_latent_index(positive, latent_length, len(image), frame_idx, scale_factors) assert latent_idx + t.shape[2] <= latent_length, "Conditioning frames exceed the length of the latent sequence." - num_prefix_frames = min(cls.NUM_PREFIX_FRAMES, t.shape[2]) - positive, negative, latent_image, noise_mask = cls.append_keyframe( positive, negative, frame_idx, latent_image, noise_mask, - t[:, :, :num_prefix_frames], + t, strength, scale_factors, ) - latent_idx += num_prefix_frames - - t = t[:, :, num_prefix_frames:] - if t.shape[2] == 0: - return io.NodeOutput(positive, negative, {"samples": latent_image, "noise_mask": noise_mask}) - - latent_image, noise_mask = cls.replace_latent_frames( - latent_image, - noise_mask, - t, - latent_idx, - strength, - ) - return io.NodeOutput(positive, negative, {"samples": latent_image, "noise_mask": noise_mask}) generate = execute # TODO: remove @@ -507,18 +553,90 @@ class LTXVPreprocess(io.ComfyNode): preprocess = execute # TODO: remove + +import comfy.nested_tensor +class LTXVConcatAVLatent(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="LTXVConcatAVLatent", + category="latent/video/ltxv", + inputs=[ + io.Latent.Input("video_latent"), + io.Latent.Input("audio_latent"), + ], + outputs=[ + io.Latent.Output(display_name="latent"), + ], + ) + + @classmethod + def execute(cls, video_latent, audio_latent) -> io.NodeOutput: + output = {} + output.update(video_latent) + output.update(audio_latent) + video_noise_mask = video_latent.get("noise_mask", None) + audio_noise_mask = audio_latent.get("noise_mask", None) + + if video_noise_mask is not None or audio_noise_mask is not None: + if video_noise_mask is None: + video_noise_mask = torch.ones_like(video_latent["samples"]) + if audio_noise_mask is None: + audio_noise_mask = torch.ones_like(audio_latent["samples"]) + output["noise_mask"] = comfy.nested_tensor.NestedTensor((video_noise_mask, audio_noise_mask)) + + output["samples"] = comfy.nested_tensor.NestedTensor((video_latent["samples"], audio_latent["samples"])) + + return io.NodeOutput(output) + + +class LTXVSeparateAVLatent(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="LTXVSeparateAVLatent", + category="latent/video/ltxv", + description="LTXV Separate AV Latent", + inputs=[ + io.Latent.Input("av_latent"), + ], + outputs=[ + io.Latent.Output(display_name="video_latent"), + io.Latent.Output(display_name="audio_latent"), + ], + ) + + @classmethod + def execute(cls, av_latent) -> io.NodeOutput: + latents = av_latent["samples"].unbind() + video_latent = av_latent.copy() + video_latent["samples"] = latents[0] + audio_latent = av_latent.copy() + audio_latent["samples"] = latents[1] + if "noise_mask" in av_latent: + masks = av_latent["noise_mask"] + if masks is not None: + masks = masks.unbind() + video_latent["noise_mask"] = masks[0] + audio_latent["noise_mask"] = masks[1] + return io.NodeOutput(video_latent, audio_latent) + + class LtxvExtension(ComfyExtension): @override async def get_node_list(self) -> list[type[io.ComfyNode]]: return [ EmptyLTXVLatentVideo, LTXVImgToVideo, + LTXVImgToVideoInplace, ModelSamplingLTXV, LTXVConditioning, LTXVScheduler, LTXVAddGuide, LTXVPreprocess, LTXVCropGuides, + LTXVConcatAVLatent, + LTXVSeparateAVLatent, ] diff --git a/comfy_extras/nodes_lt_audio.py b/comfy_extras/nodes_lt_audio.py new file mode 100644 index 000000000..b0b7000ef --- /dev/null +++ b/comfy_extras/nodes_lt_audio.py @@ -0,0 +1,183 @@ +import folder_paths +import comfy.utils +import comfy.model_management +import torch + +from comfy.ldm.lightricks.vae.audio_vae import AudioVAE +from comfy_api.latest import ComfyExtension, io + + +class LTXVAudioVAELoader(io.ComfyNode): + @classmethod + def define_schema(cls) -> io.Schema: + return io.Schema( + node_id="LTXVAudioVAELoader", + display_name="LTXV Audio VAE Loader", + category="audio", + inputs=[ + io.Combo.Input( + "ckpt_name", + options=folder_paths.get_filename_list("checkpoints"), + tooltip="Audio VAE checkpoint to load.", + ) + ], + outputs=[io.Vae.Output(display_name="Audio VAE")], + ) + + @classmethod + def execute(cls, ckpt_name: str) -> io.NodeOutput: + ckpt_path = folder_paths.get_full_path_or_raise("checkpoints", ckpt_name) + sd, metadata = comfy.utils.load_torch_file(ckpt_path, return_metadata=True) + return io.NodeOutput(AudioVAE(sd, metadata)) + + +class LTXVAudioVAEEncode(io.ComfyNode): + @classmethod + def define_schema(cls) -> io.Schema: + return io.Schema( + node_id="LTXVAudioVAEEncode", + display_name="LTXV Audio VAE Encode", + category="audio", + inputs=[ + io.Audio.Input("audio", tooltip="The audio to be encoded."), + io.Vae.Input( + id="audio_vae", + display_name="Audio VAE", + tooltip="The Audio VAE model to use for encoding.", + ), + ], + outputs=[io.Latent.Output(display_name="Audio Latent")], + ) + + @classmethod + def execute(cls, audio, audio_vae: AudioVAE) -> io.NodeOutput: + audio_latents = audio_vae.encode(audio) + return io.NodeOutput( + { + "samples": audio_latents, + "sample_rate": int(audio_vae.sample_rate), + "type": "audio", + } + ) + + +class LTXVAudioVAEDecode(io.ComfyNode): + @classmethod + def define_schema(cls) -> io.Schema: + return io.Schema( + node_id="LTXVAudioVAEDecode", + display_name="LTXV Audio VAE Decode", + category="audio", + inputs=[ + io.Latent.Input("samples", tooltip="The latent to be decoded."), + io.Vae.Input( + id="audio_vae", + display_name="Audio VAE", + tooltip="The Audio VAE model used for decoding the latent.", + ), + ], + outputs=[io.Audio.Output(display_name="Audio")], + ) + + @classmethod + def execute(cls, samples, audio_vae: AudioVAE) -> io.NodeOutput: + audio_latent = samples["samples"] + if audio_latent.is_nested: + audio_latent = audio_latent.unbind()[-1] + audio = audio_vae.decode(audio_latent).to(audio_latent.device) + output_audio_sample_rate = audio_vae.output_sample_rate + return io.NodeOutput( + { + "waveform": audio, + "sample_rate": int(output_audio_sample_rate), + } + ) + + +class LTXVEmptyLatentAudio(io.ComfyNode): + @classmethod + def define_schema(cls) -> io.Schema: + return io.Schema( + node_id="LTXVEmptyLatentAudio", + display_name="LTXV Empty Latent Audio", + category="latent/audio", + inputs=[ + io.Int.Input( + "frames_number", + default=97, + min=1, + max=1000, + step=1, + display_mode=io.NumberDisplay.number, + tooltip="Number of frames.", + ), + io.Int.Input( + "frame_rate", + default=25, + min=1, + max=1000, + step=1, + display_mode=io.NumberDisplay.number, + tooltip="Number of frames per second.", + ), + io.Int.Input( + "batch_size", + default=1, + min=1, + max=4096, + display_mode=io.NumberDisplay.number, + tooltip="The number of latent audio samples in the batch.", + ), + io.Vae.Input( + id="audio_vae", + display_name="Audio VAE", + tooltip="The Audio VAE model to get configuration from.", + ), + ], + outputs=[io.Latent.Output(display_name="Latent")], + ) + + @classmethod + def execute( + cls, + frames_number: int, + frame_rate: int, + batch_size: int, + audio_vae: AudioVAE, + ) -> io.NodeOutput: + """Generate empty audio latents matching the reference pipeline structure.""" + + assert audio_vae is not None, "Audio VAE model is required" + + z_channels = audio_vae.latent_channels + audio_freq = audio_vae.latent_frequency_bins + sampling_rate = int(audio_vae.sample_rate) + + num_audio_latents = audio_vae.num_of_latents_from_frames(frames_number, frame_rate) + + audio_latents = torch.zeros( + (batch_size, z_channels, num_audio_latents, audio_freq), + device=comfy.model_management.intermediate_device(), + ) + + return io.NodeOutput( + { + "samples": audio_latents, + "sample_rate": sampling_rate, + "type": "audio", + } + ) + + +class LTXVAudioExtension(ComfyExtension): + async def get_node_list(self) -> list[type[io.ComfyNode]]: + return [ + LTXVAudioVAELoader, + LTXVAudioVAEEncode, + LTXVAudioVAEDecode, + LTXVEmptyLatentAudio, + ] + + +async def comfy_entrypoint() -> ComfyExtension: + return LTXVAudioExtension() diff --git a/comfy_extras/nodes_lt_upsampler.py b/comfy_extras/nodes_lt_upsampler.py new file mode 100644 index 000000000..f99ba13fb --- /dev/null +++ b/comfy_extras/nodes_lt_upsampler.py @@ -0,0 +1,75 @@ +from comfy import model_management +import math + +class LTXVLatentUpsampler: + """ + Upsamples a video latent by a factor of 2. + """ + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "samples": ("LATENT",), + "upscale_model": ("LATENT_UPSCALE_MODEL",), + "vae": ("VAE",), + } + } + + RETURN_TYPES = ("LATENT",) + FUNCTION = "upsample_latent" + CATEGORY = "latent/video" + EXPERIMENTAL = True + + def upsample_latent( + self, + samples: dict, + upscale_model, + vae, + ) -> tuple: + """ + Upsample the input latent using the provided model. + + Args: + samples (dict): Input latent samples + upscale_model (LatentUpsampler): Loaded upscale model + vae: VAE model for normalization + auto_tiling (bool): Whether to automatically tile the input for processing + + Returns: + tuple: Tuple containing the upsampled latent + """ + device = model_management.get_torch_device() + memory_required = model_management.module_size(upscale_model) + + model_dtype = next(upscale_model.parameters()).dtype + latents = samples["samples"] + input_dtype = latents.dtype + + memory_required += math.prod(latents.shape) * 3000.0 # TODO: more accurate + model_management.free_memory(memory_required, device) + + try: + upscale_model.to(device) # TODO: use the comfy model management system. + + latents = latents.to(dtype=model_dtype, device=device) + + """Upsample latents without tiling.""" + latents = vae.first_stage_model.per_channel_statistics.un_normalize(latents) + upsampled_latents = upscale_model(latents) + finally: + upscale_model.cpu() + + upsampled_latents = vae.first_stage_model.per_channel_statistics.normalize( + upsampled_latents + ) + upsampled_latents = upsampled_latents.to(dtype=input_dtype, device=model_management.intermediate_device()) + return_dict = samples.copy() + return_dict["samples"] = upsampled_latents + return_dict.pop("noise_mask", None) + return (return_dict,) + + +NODE_CLASS_MAPPINGS = { + "LTXVLatentUpsampler": LTXVLatentUpsampler, +} diff --git a/nodes.py b/nodes.py index 662907ae6..56b74ebe3 100644 --- a/nodes.py +++ b/nodes.py @@ -295,7 +295,11 @@ class VAEDecode: DESCRIPTION = "Decodes latent images back into pixel space images." def decode(self, vae, samples): - images = vae.decode(samples["samples"]) + latent = samples["samples"] + if latent.is_nested: + latent = latent.unbind()[0] + + images = vae.decode(latent) if len(images.shape) == 5: #Combine batches images = images.reshape(-1, images.shape[-3], images.shape[-2], images.shape[-1]) return (images, ) @@ -970,7 +974,7 @@ class DualCLIPLoader: def INPUT_TYPES(s): return {"required": { "clip_name1": (folder_paths.get_filename_list("text_encoders"), ), "clip_name2": (folder_paths.get_filename_list("text_encoders"), ), - "type": (["sdxl", "sd3", "flux", "hunyuan_video", "hidream", "hunyuan_image", "hunyuan_video_15", "kandinsky5", "kandinsky5_image", "newbie"], ), + "type": (["sdxl", "sd3", "flux", "hunyuan_video", "hidream", "hunyuan_image", "hunyuan_video_15", "kandinsky5", "kandinsky5_image", "ltxv", "newbie"], ), }, "optional": { "device": (["default", "cpu"], {"advanced": True}), @@ -2331,6 +2335,8 @@ async def init_builtin_extra_nodes(): "nodes_mochi.py", "nodes_slg.py", "nodes_mahiro.py", + "nodes_lt_upsampler.py", + "nodes_lt_audio.py", "nodes_lt.py", "nodes_hooks.py", "nodes_load_3d.py", diff --git a/pyproject.toml b/pyproject.toml index 60378de1e..a7d159be9 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -3,7 +3,7 @@ name = "ComfyUI" version = "0.7.0" readme = "README.md" license = { file = "LICENSE" } -requires-python = ">=3.9" +requires-python = ">=3.10" [project.urls] homepage = "https://www.comfy.org/" From d1b9822f741843c64b2cbd8e1bcdd49794b182ce Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sun, 4 Jan 2026 23:27:31 -0800 Subject: [PATCH 371/647] Add LTXAVTextEncoderLoader node. (#11634) --- comfy_extras/nodes_lt_audio.py | 33 +++++++++++++++++++++++++++++++++ 1 file changed, 33 insertions(+) diff --git a/comfy_extras/nodes_lt_audio.py b/comfy_extras/nodes_lt_audio.py index b0b7000ef..2d3d103b4 100644 --- a/comfy_extras/nodes_lt_audio.py +++ b/comfy_extras/nodes_lt_audio.py @@ -169,6 +169,38 @@ class LTXVEmptyLatentAudio(io.ComfyNode): ) +class LTXAVTextEncoderLoader(io.ComfyNode): + @classmethod + def define_schema(cls) -> io.Schema: + return io.Schema( + node_id="LTXAVTextEncoderLoader", + display_name="LTXV Audio Text Encoder Loader", + category="advanced/loaders", + description="[Recipes]\n\nltxav: gemma 3 12B", + inputs=[ + io.Combo.Input( + "text_encoder", + options=folder_paths.get_filename_list("text_encoders"), + ), + io.Combo.Input( + "ckpt_name", + options=folder_paths.get_filename_list("checkpoints"), + ) + ], + outputs=[io.Clip.Output(display_name="Audio VAE")], + ) + + @classmethod + def execute(cls, text_encoder, ckpt_name, device="default"): + clip_type = comfy.sd.CLIPType.LTXV + + clip_path1 = folder_paths.get_full_path_or_raise("text_encoders", text_encoder) + clip_path2 = folder_paths.get_full_path_or_raise("checkpoints", ckpt_name) + + clip = comfy.sd.load_clip(ckpt_paths=[clip_path1, clip_path2], embedding_directory=folder_paths.get_folder_paths("embeddings"), clip_type=clip_type) + return io.NodeOutput(clip) + + class LTXVAudioExtension(ComfyExtension): async def get_node_list(self) -> list[type[io.ComfyNode]]: return [ @@ -176,6 +208,7 @@ class LTXVAudioExtension(ComfyExtension): LTXVAudioVAEEncode, LTXVAudioVAEDecode, LTXVEmptyLatentAudio, + LTXAVTextEncoderLoader, ] From d157c3299d6f9e1b57981bdb4931f1d7129e4e8d Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 5 Jan 2026 00:48:31 -0800 Subject: [PATCH 372/647] Refactor module_size function. (#11637) --- comfy/model_management.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy/model_management.py b/comfy/model_management.py index 2501cecb7..7f5a8aee9 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -456,7 +456,7 @@ def module_size(module): sd = module.state_dict() for k in sd: t = sd[k] - module_mem += t.nelement() * t.element_size() + module_mem += t.nbytes return module_mem class LoadedModel: From 4f3f9e72a9d0c15d00c0c362b8e90f1db5af6cfb Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 5 Jan 2026 02:41:23 -0800 Subject: [PATCH 373/647] Fix name. (#11638) --- comfy_extras/nodes_lt_audio.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy_extras/nodes_lt_audio.py b/comfy_extras/nodes_lt_audio.py index 2d3d103b4..26b0160d2 100644 --- a/comfy_extras/nodes_lt_audio.py +++ b/comfy_extras/nodes_lt_audio.py @@ -187,7 +187,7 @@ class LTXAVTextEncoderLoader(io.ComfyNode): options=folder_paths.get_filename_list("checkpoints"), ) ], - outputs=[io.Clip.Output(display_name="Audio VAE")], + outputs=[io.Clip.Output()], ) @classmethod From 6da00dd899e3ee6f2a0a8163b080a9f373395025 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 5 Jan 2026 18:48:58 -0800 Subject: [PATCH 374/647] Initial ops changes to use comfy_kitchen: Initial nvfp4 checkpoint support. (#11635) --------- Co-authored-by: Jedrzej Kosinski --- .github/workflows/test-build.yml | 2 +- .github/workflows/test-launch.yml | 4 +- comfy/model_management.py | 4 +- comfy/ops.py | 164 +++-- comfy/quant_ops.py | 641 +++--------------- requirements.txt | 1 + .../comfy_quant/test_mixed_precision.py | 12 +- tests-unit/comfy_quant/test_quant_registry.py | 190 ------ 8 files changed, 223 insertions(+), 795 deletions(-) delete mode 100644 tests-unit/comfy_quant/test_quant_registry.py diff --git a/.github/workflows/test-build.yml b/.github/workflows/test-build.yml index 419873ad8..9160242e9 100644 --- a/.github/workflows/test-build.yml +++ b/.github/workflows/test-build.yml @@ -18,7 +18,7 @@ jobs: strategy: fail-fast: false matrix: - python-version: ["3.9", "3.10", "3.11", "3.12", "3.13"] + python-version: ["3.10", "3.11", "3.12", "3.13", "3.14"] steps: - uses: actions/checkout@v4 - name: Set up Python ${{ matrix.python-version }} diff --git a/.github/workflows/test-launch.yml b/.github/workflows/test-launch.yml index fd70aff23..ef0d3f123 100644 --- a/.github/workflows/test-launch.yml +++ b/.github/workflows/test-launch.yml @@ -32,7 +32,9 @@ jobs: working-directory: ComfyUI - name: Check for unhandled exceptions in server log run: | - if grep -qE "Exception|Error" console_output.log; then + grep -v "Found comfy_kitchen backend triton: {'available': False, 'disabled': True, 'unavailable_reason': \"ImportError: No module named 'triton'\", 'capabilities': \[\]}" console_output.log | grep -v "Found comfy_kitchen backend triton: {'available': False, 'disabled': False, 'unavailable_reason': \"ImportError: No module named 'triton'\", 'capabilities': \[\]}" > console_output_filtered.log + cat console_output_filtered.log + if grep -qE "Exception|Error" console_output_filtered.log; then echo "Unhandled exception/error found in server log." exit 1 fi diff --git a/comfy/model_management.py b/comfy/model_management.py index 7f5a8aee9..22f4de044 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -1156,7 +1156,7 @@ def pin_memory(tensor): if not tensor.is_contiguous(): return False - size = tensor.numel() * tensor.element_size() + size = tensor.nbytes if (TOTAL_PINNED_MEMORY + size) > MAX_PINNED_MEMORY: return False @@ -1183,7 +1183,7 @@ def unpin_memory(tensor): return False ptr = tensor.data_ptr() - size = tensor.numel() * tensor.element_size() + size = tensor.nbytes size_stored = PINNED_MEMORY.get(ptr, None) if size_stored is None: diff --git a/comfy/ops.py b/comfy/ops.py index 16889bb82..f5e1e9230 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -79,7 +79,7 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, of if input is not None: if dtype is None: if isinstance(input, QuantizedTensor): - dtype = input._layout_params["orig_dtype"] + dtype = input.params.orig_dtype else: dtype = input.dtype if bias_dtype is None: @@ -412,26 +412,34 @@ def fp8_linear(self, input): return None input_dtype = input.dtype + input_shape = input.shape + tensor_3d = input.ndim == 3 - if input.ndim == 3 or input.ndim == 2: - w, bias, offload_stream = cast_bias_weight(self, input, dtype=dtype, bias_dtype=input_dtype, offloadable=True) - scale_weight = torch.ones((), device=input.device, dtype=torch.float32) + if tensor_3d: + input = input.reshape(-1, input_shape[2]) - scale_input = torch.ones((), device=input.device, dtype=torch.float32) - input = torch.clamp(input, min=-448, max=448, out=input) - layout_params_weight = {'scale': scale_input, 'orig_dtype': input_dtype} - quantized_input = QuantizedTensor(input.to(dtype).contiguous(), "TensorCoreFP8Layout", layout_params_weight) + if input.ndim != 2: + return None + w, bias, offload_stream = cast_bias_weight(self, input, dtype=dtype, bias_dtype=input_dtype, offloadable=True) + scale_weight = torch.ones((), device=input.device, dtype=torch.float32) - # Wrap weight in QuantizedTensor - this enables unified dispatch - # Call F.linear - __torch_dispatch__ routes to fp8_linear handler in quant_ops.py! - layout_params_weight = {'scale': scale_weight, 'orig_dtype': input_dtype} - quantized_weight = QuantizedTensor(w, "TensorCoreFP8Layout", layout_params_weight) - o = torch.nn.functional.linear(quantized_input, quantized_weight, bias) + scale_input = torch.ones((), device=input.device, dtype=torch.float32) + input = torch.clamp(input, min=-448, max=448, out=input) + input_fp8 = input.to(dtype).contiguous() + layout_params_input = TensorCoreFP8Layout.Params(scale=scale_input, orig_dtype=input_dtype, orig_shape=tuple(input_fp8.shape)) + quantized_input = QuantizedTensor(input_fp8, TensorCoreFP8Layout, layout_params_input) - uncast_bias_weight(self, w, bias, offload_stream) - return o + # Wrap weight in QuantizedTensor - this enables unified dispatch + # Call F.linear - __torch_dispatch__ routes to fp8_linear handler in quant_ops.py! + layout_params_weight = TensorCoreFP8Layout.Params(scale=scale_weight, orig_dtype=input_dtype, orig_shape=tuple(w.shape)) + quantized_weight = QuantizedTensor(w, TensorCoreFP8Layout, layout_params_weight) + o = torch.nn.functional.linear(quantized_input, quantized_weight, bias) - return None + uncast_bias_weight(self, w, bias, offload_stream) + if tensor_3d: + o = o.reshape((input_shape[0], input_shape[1], w.shape[0])) + + return o class fp8_ops(manual_cast): class Linear(manual_cast.Linear): @@ -477,7 +485,12 @@ if CUBLAS_IS_AVAILABLE: # ============================================================================== # Mixed Precision Operations # ============================================================================== -from .quant_ops import QuantizedTensor, QUANT_ALGOS +from .quant_ops import ( + QuantizedTensor, + QUANT_ALGOS, + TensorCoreFP8Layout, + get_layout_class, +) def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_precision_mm=False): @@ -497,14 +510,15 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec ) -> None: super().__init__() - if dtype is None: - dtype = MixedPrecisionOps._compute_dtype - - self.factory_kwargs = {"device": device, "dtype": dtype} + self.factory_kwargs = {"device": device, "dtype": MixedPrecisionOps._compute_dtype} + # self.factory_kwargs = {"device": device, "dtype": dtype} self.in_features = in_features self.out_features = out_features - self._has_bias = bias + if bias: + self.bias = torch.nn.Parameter(torch.empty(out_features, **self.factory_kwargs)) + else: + self.register_parameter("bias", None) self.tensor_class = None self._full_precision_mm = MixedPrecisionOps._full_precision_mm @@ -512,6 +526,16 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec def reset_parameters(self): return None + def _load_scale_param(self, state_dict, prefix, param_name, device, manually_loaded_keys, dtype=None): + key = f"{prefix}{param_name}" + value = state_dict.pop(key, None) + if value is not None: + value = value.to(device=device) + if dtype is not None: + value = value.view(dtype=dtype) + manually_loaded_keys.append(key) + return value + def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs): @@ -529,14 +553,7 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec layer_conf = json.loads(layer_conf.numpy().tobytes()) if layer_conf is None: - dtype = self.factory_kwargs["dtype"] - self.weight = torch.nn.Parameter(weight.to(device=device, dtype=dtype), requires_grad=False) - if dtype != MixedPrecisionOps._compute_dtype: - self.comfy_cast_weights = True - if self._has_bias: - self.bias = torch.nn.Parameter(torch.empty(self.out_features, device=device, dtype=dtype)) - else: - self.register_parameter("bias", None) + self.weight = torch.nn.Parameter(weight.to(device=device, dtype=MixedPrecisionOps._compute_dtype), requires_grad=False) else: self.quant_format = layer_conf.get("format", None) if not self._full_precision_mm: @@ -547,31 +564,46 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec qconfig = QUANT_ALGOS[self.quant_format] self.layout_type = qconfig["comfy_tensor_layout"] + layout_cls = get_layout_class(self.layout_type) - weight_scale_key = f"{prefix}weight_scale" - scale = state_dict.pop(weight_scale_key, None) - if scale is not None: - scale = scale.to(device) - layout_params = { - 'scale': scale, - 'orig_dtype': MixedPrecisionOps._compute_dtype, - 'block_size': qconfig.get("group_size", None), - } + # Load format-specific parameters + if self.quant_format in ["float8_e4m3fn", "float8_e5m2"]: + # FP8: single tensor scale + scale = self._load_scale_param(state_dict, prefix, "weight_scale", device, manually_loaded_keys) - if scale is not None: - manually_loaded_keys.append(weight_scale_key) + params = layout_cls.Params( + scale=scale, + orig_dtype=MixedPrecisionOps._compute_dtype, + orig_shape=(self.out_features, self.in_features), + ) + + elif self.quant_format == "nvfp4": + # NVFP4: tensor_scale (weight_scale_2) + block_scale (weight_scale) + tensor_scale = self._load_scale_param(state_dict, prefix, "weight_scale_2", device, manually_loaded_keys) + block_scale = self._load_scale_param(state_dict, prefix, "weight_scale", device, manually_loaded_keys, + dtype=torch.float8_e4m3fn) + + if tensor_scale is None or block_scale is None: + raise ValueError(f"Missing NVFP4 scales for layer {layer_name}") + + params = layout_cls.Params( + scale=tensor_scale, + block_scale=block_scale, + orig_dtype=MixedPrecisionOps._compute_dtype, + orig_shape=(self.out_features, self.in_features), + ) + else: + raise ValueError(f"Unsupported quantization format: {self.quant_format}") self.weight = torch.nn.Parameter( - QuantizedTensor(weight.to(device=device, dtype=qconfig.get("storage_t", None)), self.layout_type, layout_params), + QuantizedTensor(weight.to(device=device, dtype=qconfig["storage_t"]), self.layout_type, params), requires_grad=False ) - if self._has_bias: - self.bias = torch.nn.Parameter(torch.empty(self.out_features, device=device, dtype=MixedPrecisionOps._compute_dtype)) - else: - self.register_parameter("bias", None) - for param_name in qconfig["parameters"]: + if param_name in {"weight_scale", "weight_scale_2"}: + continue # Already handled above + param_key = f"{prefix}{param_name}" _v = state_dict.pop(param_key, None) if _v is None: @@ -588,7 +620,15 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec def state_dict(self, *args, destination=None, prefix="", **kwargs): sd = super().state_dict(*args, destination=destination, prefix=prefix, **kwargs) if isinstance(self.weight, QuantizedTensor): - sd["{}weight_scale".format(prefix)] = self.weight._layout_params['scale'] + layout_cls = self.weight._layout_cls + + # Check if it's any FP8 variant (E4M3 or E5M2) + if layout_cls in ("TensorCoreFP8E4M3Layout", "TensorCoreFP8E5M2Layout", "TensorCoreFP8Layout"): + sd["{}weight_scale".format(prefix)] = self.weight._params.scale + elif layout_cls == "TensorCoreNVFP4Layout": + sd["{}weight_scale_2".format(prefix)] = self.weight._params.scale + sd["{}weight_scale".format(prefix)] = self.weight._params.block_scale + quant_conf = {"format": self.quant_format} if self._full_precision_mm: quant_conf["full_precision_matrix_mult"] = True @@ -607,12 +647,33 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec def forward(self, input, *args, **kwargs): run_every_op() + input_shape = input.shape + tensor_3d = input.ndim == 3 + if self._full_precision_mm or self.comfy_cast_weights or len(self.weight_function) > 0 or len(self.bias_function) > 0: return self.forward_comfy_cast_weights(input, *args, **kwargs) + if (getattr(self, 'layout_type', None) is not None and not isinstance(input, QuantizedTensor)): - input = QuantizedTensor.from_float(input, self.layout_type, scale=getattr(self, 'input_scale', None), dtype=self.weight.dtype) - return self._forward(input, self.weight, self.bias) + + # Reshape 3D tensors to 2D for quantization (needed for NVFP4 and others) + if tensor_3d: + input = input.reshape(-1, input_shape[2]) + + if input.ndim != 2: + # Fall back to comfy_cast_weights for non-2D tensors + return self.forward_comfy_cast_weights(input.reshape(input_shape), *args, **kwargs) + + # dtype is now implicit in the layout class + input = QuantizedTensor.from_float(input, self.layout_type, scale=getattr(self, 'input_scale', None)) + + output = self._forward(input, self.weight, self.bias) + + # Reshape output back to 3D if input was 3D + if tensor_3d: + output = output.reshape((input_shape[0], input_shape[1], self.weight.shape[0])) + + return output def convert_weight(self, weight, inplace=False, **kwargs): if isinstance(weight, QuantizedTensor): @@ -622,7 +683,8 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec def set_weight(self, weight, inplace_update=False, seed=None, return_weight=False, **kwargs): if getattr(self, 'layout_type', None) is not None: - weight = QuantizedTensor.from_float(weight, self.layout_type, scale="recalculate", dtype=self.weight.dtype, stochastic_rounding=seed, inplace_ops=True) + # dtype is now implicit in the layout class + weight = QuantizedTensor.from_float(weight, self.layout_type, scale="recalculate", stochastic_rounding=seed, inplace_ops=True) else: weight = weight.to(self.weight.dtype) if return_weight: diff --git a/comfy/quant_ops.py b/comfy/quant_ops.py index cd96541d7..cd737726f 100644 --- a/comfy/quant_ops.py +++ b/comfy/quant_ops.py @@ -1,580 +1,133 @@ import torch import logging -from typing import Tuple, Dict + +try: + import comfy_kitchen as ck + from comfy_kitchen.tensor import ( + QuantizedTensor, + QuantizedLayout, + TensorCoreFP8Layout as _CKFp8Layout, + TensorCoreNVFP4Layout, # Direct import, no wrapper needed + register_layout_op, + register_layout_class, + get_layout_class, + ) + _CK_AVAILABLE = True + ck.registry.disable("triton") + for k, v in ck.list_backends().items(): + logging.info(f"Found comfy_kitchen backend {k}: {v}") +except ImportError as e: + logging.error(f"Failed to import comfy_kitchen, Error: {e}, fp8 and fp4 support will not be available.") + _CK_AVAILABLE = False + + class QuantizedTensor: + pass + + class _CKFp8Layout: + pass + + class TensorCoreNVFP4Layout: + pass + + def register_layout_class(name, cls): + pass + + def get_layout_class(name): + return None + import comfy.float -_LAYOUT_REGISTRY = {} -_GENERIC_UTILS = {} - - -def register_layout_op(torch_op, layout_type): - """ - Decorator to register a layout-specific operation handler. - Args: - torch_op: PyTorch operation (e.g., torch.ops.aten.linear.default) - layout_type: Layout class (e.g., TensorCoreFP8Layout) - Example: - @register_layout_op(torch.ops.aten.linear.default, TensorCoreFP8Layout) - def fp8_linear(func, args, kwargs): - # FP8-specific linear implementation - ... - """ - def decorator(handler_func): - if torch_op not in _LAYOUT_REGISTRY: - _LAYOUT_REGISTRY[torch_op] = {} - _LAYOUT_REGISTRY[torch_op][layout_type] = handler_func - return handler_func - return decorator - - -def register_generic_util(torch_op): - """ - Decorator to register a generic utility that works for all layouts. - Args: - torch_op: PyTorch operation (e.g., torch.ops.aten.detach.default) - - Example: - @register_generic_util(torch.ops.aten.detach.default) - def generic_detach(func, args, kwargs): - # Works for any layout - ... - """ - def decorator(handler_func): - _GENERIC_UTILS[torch_op] = handler_func - return handler_func - return decorator - - -def _get_layout_from_args(args): - for arg in args: - if isinstance(arg, QuantizedTensor): - return arg._layout_type - elif isinstance(arg, (list, tuple)): - for item in arg: - if isinstance(item, QuantizedTensor): - return item._layout_type - return None - - -def _move_layout_params_to_device(params, device): - new_params = {} - for k, v in params.items(): - if isinstance(v, torch.Tensor): - new_params[k] = v.to(device=device) - else: - new_params[k] = v - return new_params - - -def _copy_layout_params(params): - new_params = {} - for k, v in params.items(): - if isinstance(v, torch.Tensor): - new_params[k] = v.clone() - else: - new_params[k] = v - return new_params - -def _copy_layout_params_inplace(src, dst, non_blocking=False): - for k, v in src.items(): - if isinstance(v, torch.Tensor): - dst[k].copy_(v, non_blocking=non_blocking) - else: - dst[k] = v - -class QuantizedLayout: - """ - Base class for quantization layouts. - - A layout encapsulates the format-specific logic for quantization/dequantization - and provides a uniform interface for extracting raw tensors needed for computation. - - New quantization formats should subclass this and implement the required methods. - """ - @classmethod - def quantize(cls, tensor, **kwargs) -> Tuple[torch.Tensor, Dict]: - raise NotImplementedError(f"{cls.__name__} must implement quantize()") - - @staticmethod - def dequantize(qdata, **layout_params) -> torch.Tensor: - raise NotImplementedError("TensorLayout must implement dequantize()") - - @classmethod - def get_plain_tensors(cls, qtensor) -> torch.Tensor: - raise NotImplementedError(f"{cls.__name__} must implement get_plain_tensors()") - - -class QuantizedTensor(torch.Tensor): - """ - Universal quantized tensor that works with any layout. - - This tensor subclass uses a pluggable layout system to support multiple - quantization formats (FP8, INT4, INT8, etc.) without code duplication. - - The layout_type determines format-specific behavior, while common operations - (detach, clone, to) are handled generically. - - Attributes: - _qdata: The quantized tensor data - _layout_type: Layout class (e.g., TensorCoreFP8Layout) - _layout_params: Dict with layout-specific params (scale, zero_point, etc.) - """ - - @staticmethod - def __new__(cls, qdata, layout_type, layout_params): - """ - Create a quantized tensor. - - Args: - qdata: The quantized data tensor - layout_type: Layout class (subclass of QuantizedLayout) - layout_params: Dict with layout-specific parameters - """ - return torch.Tensor._make_wrapper_subclass(cls, qdata.shape, device=qdata.device, dtype=qdata.dtype, requires_grad=False) - - def __init__(self, qdata, layout_type, layout_params): - self._qdata = qdata - self._layout_type = layout_type - self._layout_params = layout_params - - def __repr__(self): - layout_name = self._layout_type - param_str = ", ".join(f"{k}={v}" for k, v in list(self._layout_params.items())[:2]) - return f"QuantizedTensor(shape={self.shape}, layout={layout_name}, {param_str})" - - @property - def layout_type(self): - return self._layout_type - - def __tensor_flatten__(self): - """ - Tensor flattening protocol for proper device movement. - """ - inner_tensors = ["_qdata"] - ctx = { - "layout_type": self._layout_type, - } - - tensor_params = {} - non_tensor_params = {} - for k, v in self._layout_params.items(): - if isinstance(v, torch.Tensor): - tensor_params[k] = v - else: - non_tensor_params[k] = v - - ctx["tensor_param_keys"] = list(tensor_params.keys()) - ctx["non_tensor_params"] = non_tensor_params - - for k, v in tensor_params.items(): - attr_name = f"_layout_param_{k}" - object.__setattr__(self, attr_name, v) - inner_tensors.append(attr_name) - - return inner_tensors, ctx - - @staticmethod - def __tensor_unflatten__(inner_tensors, ctx, outer_size, outer_stride): - """ - Tensor unflattening protocol for proper device movement. - Reconstructs the QuantizedTensor after device movement. - """ - layout_type = ctx["layout_type"] - layout_params = dict(ctx["non_tensor_params"]) - - for key in ctx["tensor_param_keys"]: - attr_name = f"_layout_param_{key}" - layout_params[key] = inner_tensors[attr_name] - - return QuantizedTensor(inner_tensors["_qdata"], layout_type, layout_params) - - @classmethod - def from_float(cls, tensor, layout_type, **quantize_kwargs) -> 'QuantizedTensor': - qdata, layout_params = LAYOUTS[layout_type].quantize(tensor, **quantize_kwargs) - return cls(qdata, layout_type, layout_params) - - def dequantize(self) -> torch.Tensor: - return LAYOUTS[self._layout_type].dequantize(self._qdata, **self._layout_params) - - @classmethod - def __torch_dispatch__(cls, func, types, args=(), kwargs=None): - kwargs = kwargs or {} - - # Step 1: Check generic utilities first (detach, clone, to, etc.) - if func in _GENERIC_UTILS: - return _GENERIC_UTILS[func](func, args, kwargs) - - # Step 2: Check layout-specific handlers (linear, matmul, etc.) - layout_type = _get_layout_from_args(args) - if layout_type and func in _LAYOUT_REGISTRY: - handler = _LAYOUT_REGISTRY[func].get(layout_type) - if handler: - return handler(func, args, kwargs) - - # Step 3: Fallback to dequantization - if isinstance(args[0] if args else None, QuantizedTensor): - logging.info(f"QuantizedTensor: Unhandled operation {func}, falling back to dequantization. kwargs={kwargs}") - return cls._dequant_and_fallback(func, args, kwargs) - - @classmethod - def _dequant_and_fallback(cls, func, args, kwargs): - def dequant_arg(arg): - if isinstance(arg, QuantizedTensor): - return arg.dequantize() - elif isinstance(arg, (list, tuple)): - return type(arg)(dequant_arg(a) for a in arg) - return arg - - new_args = dequant_arg(args) - new_kwargs = dequant_arg(kwargs) - return func(*new_args, **new_kwargs) - - def data_ptr(self): - return self._qdata.data_ptr() - - def is_pinned(self): - return self._qdata.is_pinned() - - def is_contiguous(self, *arg, **kwargs): - return self._qdata.is_contiguous(*arg, **kwargs) - - def storage(self): - return self._qdata.storage() - # ============================================================================== -# Generic Utilities (Layout-Agnostic Operations) +# FP8 Layouts with Comfy-Specific Extensions # ============================================================================== -def _create_transformed_qtensor(qt, transform_fn): - new_data = transform_fn(qt._qdata) - new_params = _copy_layout_params(qt._layout_params) - return QuantizedTensor(new_data, qt._layout_type, new_params) +class _TensorCoreFP8LayoutBase(_CKFp8Layout): + FP8_DTYPE = None # Must be overridden in subclass - -def _handle_device_transfer(qt, target_device, target_dtype=None, target_layout=None, op_name="to"): - if target_layout is not None and target_layout != torch.strided: - logging.warning( - f"QuantizedTensor: layout change requested to {target_layout}, " - f"but not supported. Ignoring layout." - ) - - # Handle device transfer - current_device = qt._qdata.device - if target_device is not None: - # Normalize device for comparison - if isinstance(target_device, str): - target_device = torch.device(target_device) - if isinstance(current_device, str): - current_device = torch.device(current_device) - - if target_device != current_device: - logging.debug(f"QuantizedTensor.{op_name}: Moving from {current_device} to {target_device}") - new_q_data = qt._qdata.to(device=target_device) - new_params = _move_layout_params_to_device(qt._layout_params, target_device) - if target_dtype is not None: - new_params["orig_dtype"] = target_dtype - new_qt = QuantizedTensor(new_q_data, qt._layout_type, new_params) - logging.debug(f"QuantizedTensor.{op_name}: Created new tensor on {target_device}") - return new_qt - - logging.debug(f"QuantizedTensor.{op_name}: No device change needed, returning original") - return qt - - -@register_generic_util(torch.ops.aten.detach.default) -def generic_detach(func, args, kwargs): - """Detach operation - creates a detached copy of the quantized tensor.""" - qt = args[0] - if isinstance(qt, QuantizedTensor): - return _create_transformed_qtensor(qt, lambda x: x.detach()) - return func(*args, **kwargs) - - -@register_generic_util(torch.ops.aten.clone.default) -def generic_clone(func, args, kwargs): - """Clone operation - creates a deep copy of the quantized tensor.""" - qt = args[0] - if isinstance(qt, QuantizedTensor): - return _create_transformed_qtensor(qt, lambda x: x.clone()) - return func(*args, **kwargs) - - -@register_generic_util(torch.ops.aten._to_copy.default) -def generic_to_copy(func, args, kwargs): - """Device/dtype transfer operation - handles .to(device) calls.""" - qt = args[0] - if isinstance(qt, QuantizedTensor): - return _handle_device_transfer( - qt, - target_device=kwargs.get('device', None), - target_dtype=kwargs.get('dtype', None), - op_name="_to_copy" - ) - return func(*args, **kwargs) - - -@register_generic_util(torch.ops.aten.to.dtype_layout) -def generic_to_dtype_layout(func, args, kwargs): - """Handle .to(device) calls using the dtype_layout variant.""" - qt = args[0] - if isinstance(qt, QuantizedTensor): - return _handle_device_transfer( - qt, - target_device=kwargs.get('device', None), - target_dtype=kwargs.get('dtype', None), - target_layout=kwargs.get('layout', None), - op_name="to" - ) - return func(*args, **kwargs) - - -@register_generic_util(torch.ops.aten.copy_.default) -def generic_copy_(func, args, kwargs): - qt_dest = args[0] - src = args[1] - non_blocking = args[2] if len(args) > 2 else False - if isinstance(qt_dest, QuantizedTensor): - if isinstance(src, QuantizedTensor): - # Copy from another quantized tensor - qt_dest._qdata.copy_(src._qdata, non_blocking=non_blocking) - qt_dest._layout_type = src._layout_type - orig_dtype = qt_dest._layout_params["orig_dtype"] - _copy_layout_params_inplace(src._layout_params, qt_dest._layout_params, non_blocking=non_blocking) - qt_dest._layout_params["orig_dtype"] = orig_dtype - else: - # Copy from regular tensor - just copy raw data - qt_dest._qdata.copy_(src) - return qt_dest - return func(*args, **kwargs) - - -@register_generic_util(torch.ops.aten.to.dtype) -def generic_to_dtype(func, args, kwargs): - """Handle .to(dtype) calls - dtype conversion only.""" - src = args[0] - if isinstance(src, QuantizedTensor): - # For dtype-only conversion, just change the orig_dtype, no real cast is needed - target_dtype = args[1] if len(args) > 1 else kwargs.get('dtype') - src._layout_params["orig_dtype"] = target_dtype - return src - return func(*args, **kwargs) - - -@register_generic_util(torch.ops.aten._has_compatible_shallow_copy_type.default) -def generic_has_compatible_shallow_copy_type(func, args, kwargs): - return True - - -@register_generic_util(torch.ops.aten.empty_like.default) -def generic_empty_like(func, args, kwargs): - """Empty_like operation - creates an empty tensor with the same quantized structure.""" - qt = args[0] - if isinstance(qt, QuantizedTensor): - # Create empty tensor with same shape and dtype as the quantized data - hp_dtype = kwargs.pop('dtype', qt._layout_params["orig_dtype"]) - new_qdata = torch.empty_like(qt._qdata, **kwargs) - - # Handle device transfer for layout params - target_device = kwargs.get('device', new_qdata.device) - new_params = _move_layout_params_to_device(qt._layout_params, target_device) - - # Update orig_dtype if dtype is specified - new_params['orig_dtype'] = hp_dtype - - return QuantizedTensor(new_qdata, qt._layout_type, new_params) - return func(*args, **kwargs) - -# ============================================================================== -# FP8 Layout + Operation Handlers -# ============================================================================== -class TensorCoreFP8Layout(QuantizedLayout): - """ - Storage format: - - qdata: FP8 tensor (torch.float8_e4m3fn or torch.float8_e5m2) - - scale: Scalar tensor (float32) for dequantization - - orig_dtype: Original dtype before quantization (for casting back) - """ @classmethod - def quantize(cls, tensor, scale=None, dtype=torch.float8_e4m3fn, stochastic_rounding=0, inplace_ops=False): + def quantize(cls, tensor, scale=None, stochastic_rounding=0, inplace_ops=False): + if cls.FP8_DTYPE is None: + raise NotImplementedError(f"{cls.__name__} must define FP8_DTYPE") + orig_dtype = tensor.dtype + orig_shape = tuple(tensor.shape) if isinstance(scale, str) and scale == "recalculate": - scale = torch.amax(tensor.abs()).to(dtype=torch.float32) / torch.finfo(dtype).max + scale = torch.amax(tensor.abs()).to(dtype=torch.float32) / torch.finfo(cls.FP8_DTYPE).max if tensor.dtype not in [torch.float32, torch.bfloat16]: # Prevent scale from being too small tensor_info = torch.finfo(tensor.dtype) scale = (1.0 / torch.clamp((1.0 / scale), min=tensor_info.min, max=tensor_info.max)) - if scale is not None: - if not isinstance(scale, torch.Tensor): - scale = torch.tensor(scale) - scale = scale.to(device=tensor.device, dtype=torch.float32) + if scale is None: + scale = torch.ones((), device=tensor.device, dtype=torch.float32) + if not isinstance(scale, torch.Tensor): + scale = torch.tensor(scale, device=tensor.device, dtype=torch.float32) + if stochastic_rounding > 0: if inplace_ops: tensor *= (1.0 / scale).to(tensor.dtype) else: tensor = tensor * (1.0 / scale).to(tensor.dtype) + qdata = comfy.float.stochastic_rounding(tensor, dtype=cls.FP8_DTYPE, seed=stochastic_rounding) else: - scale = torch.ones((), device=tensor.device, dtype=torch.float32) + qdata = ck.quantize_per_tensor_fp8(tensor, scale, cls.FP8_DTYPE) - if stochastic_rounding > 0: - tensor = comfy.float.stochastic_rounding(tensor, dtype=dtype, seed=stochastic_rounding) - else: - lp_amax = torch.finfo(dtype).max - torch.clamp(tensor, min=-lp_amax, max=lp_amax, out=tensor) - tensor = tensor.to(dtype, memory_format=torch.contiguous_format) + params = cls.Params(scale=scale.float(), orig_dtype=orig_dtype, orig_shape=orig_shape) + return qdata, params - layout_params = { - 'scale': scale, - 'orig_dtype': orig_dtype - } - return tensor, layout_params - @staticmethod - def dequantize(qdata, scale, orig_dtype, **kwargs): - plain_tensor = torch.ops.aten._to_copy.default(qdata, dtype=orig_dtype) - plain_tensor.mul_(scale) - return plain_tensor +class TensorCoreFP8E4M3Layout(_TensorCoreFP8LayoutBase): + FP8_DTYPE = torch.float8_e4m3fn - @classmethod - def get_plain_tensors(cls, qtensor): - return qtensor._qdata, qtensor._layout_params['scale'] + +class TensorCoreFP8E5M2Layout(_TensorCoreFP8LayoutBase): + FP8_DTYPE = torch.float8_e5m2 + + +# Backward compatibility alias - default to E4M3 +TensorCoreFP8Layout = TensorCoreFP8E4M3Layout + + +# ============================================================================== +# Registry +# ============================================================================== + +register_layout_class("TensorCoreFP8Layout", TensorCoreFP8Layout) +register_layout_class("TensorCoreFP8E4M3Layout", TensorCoreFP8E4M3Layout) +register_layout_class("TensorCoreFP8E5M2Layout", TensorCoreFP8E5M2Layout) +register_layout_class("TensorCoreNVFP4Layout", TensorCoreNVFP4Layout) QUANT_ALGOS = { "float8_e4m3fn": { "storage_t": torch.float8_e4m3fn, "parameters": {"weight_scale", "input_scale"}, - "comfy_tensor_layout": "TensorCoreFP8Layout", + "comfy_tensor_layout": "TensorCoreFP8E4M3Layout", + }, + "float8_e5m2": { + "storage_t": torch.float8_e5m2, + "parameters": {"weight_scale", "input_scale"}, + "comfy_tensor_layout": "TensorCoreFP8E5M2Layout", + }, + "nvfp4": { + "storage_t": torch.uint8, + "parameters": {"weight_scale", "weight_scale_2", "input_scale"}, + "comfy_tensor_layout": "TensorCoreNVFP4Layout", + "group_size": 16, }, } -LAYOUTS = { - "TensorCoreFP8Layout": TensorCoreFP8Layout, -} +# ============================================================================== +# Re-exports for backward compatibility +# ============================================================================== -@register_layout_op(torch.ops.aten.linear.default, "TensorCoreFP8Layout") -def fp8_linear(func, args, kwargs): - input_tensor = args[0] - weight = args[1] - bias = args[2] if len(args) > 2 else None - - if isinstance(input_tensor, QuantizedTensor) and isinstance(weight, QuantizedTensor): - plain_input, scale_a = TensorCoreFP8Layout.get_plain_tensors(input_tensor) - plain_weight, scale_b = TensorCoreFP8Layout.get_plain_tensors(weight) - - out_dtype = kwargs.get("out_dtype") - if out_dtype is None: - out_dtype = input_tensor._layout_params['orig_dtype'] - - weight_t = plain_weight.t() - - tensor_2d = False - if len(plain_input.shape) == 2: - tensor_2d = True - plain_input = plain_input.unsqueeze(1) - - input_shape = plain_input.shape - if len(input_shape) != 3: - return None - - try: - output = torch._scaled_mm( - plain_input.reshape(-1, input_shape[2]).contiguous(), - weight_t, - bias=bias, - scale_a=scale_a, - scale_b=scale_b, - out_dtype=out_dtype, - ) - - if isinstance(output, tuple): # TODO: remove when we drop support for torch 2.4 - output = output[0] - - if not tensor_2d: - output = output.reshape((-1, input_shape[1], weight.shape[0])) - - if output.dtype in [torch.float8_e4m3fn, torch.float8_e5m2]: - output_scale = scale_a * scale_b - output_params = { - 'scale': output_scale, - 'orig_dtype': input_tensor._layout_params['orig_dtype'] - } - return QuantizedTensor(output, "TensorCoreFP8Layout", output_params) - else: - return output - - except Exception as e: - raise RuntimeError(f"FP8 _scaled_mm failed, falling back to dequantization: {e}") - - # Case 2: DQ Fallback - if isinstance(weight, QuantizedTensor): - weight = weight.dequantize() - if isinstance(input_tensor, QuantizedTensor): - input_tensor = input_tensor.dequantize() - - return torch.nn.functional.linear(input_tensor, weight, bias) - -def fp8_mm_(input_tensor, weight, bias=None, out_dtype=None): - if out_dtype is None: - out_dtype = input_tensor._layout_params['orig_dtype'] - - plain_input, scale_a = TensorCoreFP8Layout.get_plain_tensors(input_tensor) - plain_weight, scale_b = TensorCoreFP8Layout.get_plain_tensors(weight) - - output = torch._scaled_mm( - plain_input.contiguous(), - plain_weight, - bias=bias, - scale_a=scale_a, - scale_b=scale_b, - out_dtype=out_dtype, - ) - - if isinstance(output, tuple): # TODO: remove when we drop support for torch 2.4 - output = output[0] - return output - -@register_layout_op(torch.ops.aten.addmm.default, "TensorCoreFP8Layout") -def fp8_addmm(func, args, kwargs): - input_tensor = args[1] - weight = args[2] - bias = args[0] - - if isinstance(input_tensor, QuantizedTensor) and isinstance(weight, QuantizedTensor): - return fp8_mm_(input_tensor, weight, bias=bias, out_dtype=kwargs.get("out_dtype", None)) - - a = list(args) - if isinstance(args[0], QuantizedTensor): - a[0] = args[0].dequantize() - if isinstance(args[1], QuantizedTensor): - a[1] = args[1].dequantize() - if isinstance(args[2], QuantizedTensor): - a[2] = args[2].dequantize() - - return func(*a, **kwargs) - -@register_layout_op(torch.ops.aten.mm.default, "TensorCoreFP8Layout") -def fp8_mm(func, args, kwargs): - input_tensor = args[0] - weight = args[1] - - if isinstance(input_tensor, QuantizedTensor) and isinstance(weight, QuantizedTensor): - return fp8_mm_(input_tensor, weight, bias=None, out_dtype=kwargs.get("out_dtype", None)) - - a = list(args) - if isinstance(args[0], QuantizedTensor): - a[0] = args[0].dequantize() - if isinstance(args[1], QuantizedTensor): - a[1] = args[1].dequantize() - return func(*a, **kwargs) - -@register_layout_op(torch.ops.aten.view.default, "TensorCoreFP8Layout") -@register_layout_op(torch.ops.aten.t.default, "TensorCoreFP8Layout") -def fp8_func(func, args, kwargs): - input_tensor = args[0] - if isinstance(input_tensor, QuantizedTensor): - plain_input, scale_a = TensorCoreFP8Layout.get_plain_tensors(input_tensor) - ar = list(args) - ar[0] = plain_input - return QuantizedTensor(func(*ar, **kwargs), "TensorCoreFP8Layout", input_tensor._layout_params) - return func(*args, **kwargs) +__all__ = [ + "QuantizedTensor", + "QuantizedLayout", + "TensorCoreFP8Layout", + "TensorCoreFP8E4M3Layout", + "TensorCoreFP8E5M2Layout", + "TensorCoreNVFP4Layout", + "QUANT_ALGOS", + "register_layout_op", +] diff --git a/requirements.txt b/requirements.txt index 3a05799eb..0ee152032 100644 --- a/requirements.txt +++ b/requirements.txt @@ -21,6 +21,7 @@ psutil alembic SQLAlchemy av>=14.2.0 +comfy-kitchen>=0.2.0 #non essential dependencies: kornia>=0.7.1 diff --git a/tests-unit/comfy_quant/test_mixed_precision.py b/tests-unit/comfy_quant/test_mixed_precision.py index 3a54941e6..7b2eac940 100644 --- a/tests-unit/comfy_quant/test_mixed_precision.py +++ b/tests-unit/comfy_quant/test_mixed_precision.py @@ -103,18 +103,18 @@ class TestMixedPrecisionOps(unittest.TestCase): # Verify weights are wrapped in QuantizedTensor self.assertIsInstance(model.layer1.weight, QuantizedTensor) - self.assertEqual(model.layer1.weight._layout_type, "TensorCoreFP8Layout") + self.assertEqual(model.layer1.weight._layout_cls, "TensorCoreFP8E4M3Layout") # Layer 2 should NOT be quantized self.assertNotIsInstance(model.layer2.weight, QuantizedTensor) # Layer 3 should be quantized self.assertIsInstance(model.layer3.weight, QuantizedTensor) - self.assertEqual(model.layer3.weight._layout_type, "TensorCoreFP8Layout") + self.assertEqual(model.layer3.weight._layout_cls, "TensorCoreFP8E4M3Layout") # Verify scales were loaded - self.assertEqual(model.layer1.weight._layout_params['scale'].item(), 2.0) - self.assertEqual(model.layer3.weight._layout_params['scale'].item(), 1.5) + self.assertEqual(model.layer1.weight._params.scale.item(), 2.0) + self.assertEqual(model.layer3.weight._params.scale.item(), 1.5) # Forward pass input_tensor = torch.randn(5, 10, dtype=torch.bfloat16) @@ -154,8 +154,8 @@ class TestMixedPrecisionOps(unittest.TestCase): # Verify layer1.weight is a QuantizedTensor with scale preserved self.assertIsInstance(state_dict2["layer1.weight"], QuantizedTensor) - self.assertEqual(state_dict2["layer1.weight"]._layout_params['scale'].item(), 3.0) - self.assertEqual(state_dict2["layer1.weight"]._layout_type, "TensorCoreFP8Layout") + self.assertEqual(state_dict2["layer1.weight"]._params.scale.item(), 3.0) + self.assertEqual(state_dict2["layer1.weight"]._layout_cls, "TensorCoreFP8E4M3Layout") # Verify non-quantized layers are standard tensors self.assertNotIsInstance(state_dict2["layer2.weight"], QuantizedTensor) diff --git a/tests-unit/comfy_quant/test_quant_registry.py b/tests-unit/comfy_quant/test_quant_registry.py deleted file mode 100644 index 9cb54ede8..000000000 --- a/tests-unit/comfy_quant/test_quant_registry.py +++ /dev/null @@ -1,190 +0,0 @@ -import unittest -import torch -import sys -import os - -# Add comfy to path -sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", "..")) - -def has_gpu(): - return torch.cuda.is_available() - -from comfy.cli_args import args -if not has_gpu(): - args.cpu = True - -from comfy.quant_ops import QuantizedTensor, TensorCoreFP8Layout - - -class TestQuantizedTensor(unittest.TestCase): - """Test the QuantizedTensor subclass with FP8 layout""" - - def test_creation(self): - """Test creating a QuantizedTensor with TensorCoreFP8Layout""" - fp8_data = torch.randn(256, 128, dtype=torch.float32).to(torch.float8_e4m3fn) - scale = torch.tensor(2.0) - layout_params = {'scale': scale, 'orig_dtype': torch.bfloat16} - - qt = QuantizedTensor(fp8_data, "TensorCoreFP8Layout", layout_params) - - self.assertIsInstance(qt, QuantizedTensor) - self.assertEqual(qt.shape, (256, 128)) - self.assertEqual(qt.dtype, torch.float8_e4m3fn) - self.assertEqual(qt._layout_params['scale'], scale) - self.assertEqual(qt._layout_params['orig_dtype'], torch.bfloat16) - self.assertEqual(qt._layout_type, "TensorCoreFP8Layout") - - def test_dequantize(self): - """Test explicit dequantization""" - - fp8_data = torch.ones(10, 20, dtype=torch.float32).to(torch.float8_e4m3fn) - scale = torch.tensor(3.0) - layout_params = {'scale': scale, 'orig_dtype': torch.float32} - - qt = QuantizedTensor(fp8_data, "TensorCoreFP8Layout", layout_params) - dequantized = qt.dequantize() - - self.assertEqual(dequantized.dtype, torch.float32) - self.assertTrue(torch.allclose(dequantized, torch.ones(10, 20) * 3.0, rtol=0.1)) - - def test_from_float(self): - """Test creating QuantizedTensor from float tensor""" - float_tensor = torch.randn(64, 32, dtype=torch.float32) - scale = torch.tensor(1.5) - - qt = QuantizedTensor.from_float( - float_tensor, - "TensorCoreFP8Layout", - scale=scale, - dtype=torch.float8_e4m3fn - ) - - self.assertIsInstance(qt, QuantizedTensor) - self.assertEqual(qt.dtype, torch.float8_e4m3fn) - self.assertEqual(qt.shape, (64, 32)) - - # Verify dequantization gives approximately original values - dequantized = qt.dequantize() - mean_rel_error = ((dequantized - float_tensor).abs() / (float_tensor.abs() + 1e-6)).mean() - self.assertLess(mean_rel_error, 0.1) - - -class TestGenericUtilities(unittest.TestCase): - """Test generic utility operations""" - - def test_detach(self): - """Test detach operation on quantized tensor""" - fp8_data = torch.randn(10, 20, dtype=torch.float32).to(torch.float8_e4m3fn) - scale = torch.tensor(1.5) - layout_params = {'scale': scale, 'orig_dtype': torch.float32} - qt = QuantizedTensor(fp8_data, "TensorCoreFP8Layout", layout_params) - - # Detach should return a new QuantizedTensor - qt_detached = qt.detach() - - self.assertIsInstance(qt_detached, QuantizedTensor) - self.assertEqual(qt_detached.shape, qt.shape) - self.assertEqual(qt_detached._layout_type, "TensorCoreFP8Layout") - - def test_clone(self): - """Test clone operation on quantized tensor""" - fp8_data = torch.randn(10, 20, dtype=torch.float32).to(torch.float8_e4m3fn) - scale = torch.tensor(1.5) - layout_params = {'scale': scale, 'orig_dtype': torch.float32} - qt = QuantizedTensor(fp8_data, "TensorCoreFP8Layout", layout_params) - - # Clone should return a new QuantizedTensor - qt_cloned = qt.clone() - - self.assertIsInstance(qt_cloned, QuantizedTensor) - self.assertEqual(qt_cloned.shape, qt.shape) - self.assertEqual(qt_cloned._layout_type, "TensorCoreFP8Layout") - - # Verify it's a deep copy - self.assertIsNot(qt_cloned._qdata, qt._qdata) - - @unittest.skipUnless(has_gpu(), "GPU not available") - def test_to_device(self): - """Test device transfer""" - fp8_data = torch.randn(10, 20, dtype=torch.float32).to(torch.float8_e4m3fn) - scale = torch.tensor(1.5) - layout_params = {'scale': scale, 'orig_dtype': torch.float32} - qt = QuantizedTensor(fp8_data, "TensorCoreFP8Layout", layout_params) - - # Moving to same device should work (CPU to CPU) - qt_cpu = qt.to('cpu') - - self.assertIsInstance(qt_cpu, QuantizedTensor) - self.assertEqual(qt_cpu.device.type, 'cpu') - self.assertEqual(qt_cpu._layout_params['scale'].device.type, 'cpu') - - -class TestTensorCoreFP8Layout(unittest.TestCase): - """Test the TensorCoreFP8Layout implementation""" - - def test_quantize(self): - """Test quantization method""" - float_tensor = torch.randn(32, 64, dtype=torch.float32) - scale = torch.tensor(1.5) - - qdata, layout_params = TensorCoreFP8Layout.quantize( - float_tensor, - scale=scale, - dtype=torch.float8_e4m3fn - ) - - self.assertEqual(qdata.dtype, torch.float8_e4m3fn) - self.assertEqual(qdata.shape, float_tensor.shape) - self.assertIn('scale', layout_params) - self.assertIn('orig_dtype', layout_params) - self.assertEqual(layout_params['orig_dtype'], torch.float32) - - def test_dequantize(self): - """Test dequantization method""" - float_tensor = torch.ones(10, 20, dtype=torch.float32) * 3.0 - scale = torch.tensor(1.0) - - qdata, layout_params = TensorCoreFP8Layout.quantize( - float_tensor, - scale=scale, - dtype=torch.float8_e4m3fn - ) - - dequantized = TensorCoreFP8Layout.dequantize(qdata, **layout_params) - - # Should approximately match original - self.assertTrue(torch.allclose(dequantized, float_tensor, rtol=0.1, atol=0.1)) - - -class TestFallbackMechanism(unittest.TestCase): - """Test fallback for unsupported operations""" - - def test_unsupported_op_dequantizes(self): - """Test that unsupported operations fall back to dequantization""" - # Set seed for reproducibility - torch.manual_seed(42) - - # Create quantized tensor - a_fp32 = torch.randn(10, 20, dtype=torch.float32) - scale = torch.tensor(1.0) - a_q = QuantizedTensor.from_float( - a_fp32, - "TensorCoreFP8Layout", - scale=scale, - dtype=torch.float8_e4m3fn - ) - - # Call an operation that doesn't have a registered handler - # For example, torch.abs - result = torch.abs(a_q) - - # Should work via fallback (dequantize → abs → return) - self.assertNotIsInstance(result, QuantizedTensor) - expected = torch.abs(a_fp32) - # FP8 introduces quantization error, so use loose tolerance - mean_error = (result - expected).abs().mean() - self.assertLess(mean_error, 0.05, f"Mean error {mean_error:.4f} is too large") - - -if __name__ == "__main__": - unittest.main() From 6ef85c49151cf8c4d6bf5e7ccfc566b8d0681cbd Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 5 Jan 2026 19:50:35 -0800 Subject: [PATCH 375/647] Use rope functions from comfy kitchen. (#11647) --- comfy/ldm/flux/math.py | 23 +++++++++++++++-------- 1 file changed, 15 insertions(+), 8 deletions(-) diff --git a/comfy/ldm/flux/math.py b/comfy/ldm/flux/math.py index 6a22df8bc..f9597de5b 100644 --- a/comfy/ldm/flux/math.py +++ b/comfy/ldm/flux/math.py @@ -4,6 +4,7 @@ from torch import Tensor from comfy.ldm.modules.attention import optimized_attention import comfy.model_management +import logging def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor, mask=None, transformer_options={}) -> Tensor: @@ -13,7 +14,6 @@ def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor, mask=None, transforme x = optimized_attention(q, k, v, heads, skip_reshape=True, mask=mask, transformer_options=transformer_options) return x - def rope(pos: Tensor, dim: int, theta: int) -> Tensor: assert dim % 2 == 0 if comfy.model_management.is_device_mps(pos.device) or comfy.model_management.is_intel_xpu() or comfy.model_management.is_directml_enabled(): @@ -28,13 +28,20 @@ def rope(pos: Tensor, dim: int, theta: int) -> Tensor: out = rearrange(out, "b n d (i j) -> b n d i j", i=2, j=2) return out.to(dtype=torch.float32, device=pos.device) -def apply_rope1(x: Tensor, freqs_cis: Tensor): - x_ = x.to(dtype=freqs_cis.dtype).reshape(*x.shape[:-1], -1, 1, 2) - x_out = freqs_cis[..., 0] * x_[..., 0] - x_out.addcmul_(freqs_cis[..., 1], x_[..., 1]) +try: + import comfy.quant_ops + apply_rope = comfy.quant_ops.ck.apply_rope + apply_rope1 = comfy.quant_ops.ck.apply_rope1 +except: + logging.warning("No comfy kitchen, using old apply_rope functions.") + def apply_rope1(x: Tensor, freqs_cis: Tensor): + x_ = x.to(dtype=freqs_cis.dtype).reshape(*x.shape[:-1], -1, 1, 2) - return x_out.reshape(*x.shape).type_as(x) + x_out = freqs_cis[..., 0] * x_[..., 0] + x_out.addcmul_(freqs_cis[..., 1], x_[..., 1]) -def apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor): - return apply_rope1(xq, freqs_cis), apply_rope1(xk, freqs_cis) + return x_out.reshape(*x.shape).type_as(x) + + def apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor): + return apply_rope1(xq, freqs_cis), apply_rope1(xk, freqs_cis) From 161800241117fae7af90e0c938d0cf8cb2f2ddb1 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 5 Jan 2026 20:07:39 -0800 Subject: [PATCH 376/647] Revert "Use rope functions from comfy kitchen. (#11647)" (#11648) This reverts commit 6ef85c49151cf8c4d6bf5e7ccfc566b8d0681cbd. --- comfy/ldm/flux/math.py | 23 ++++++++--------------- 1 file changed, 8 insertions(+), 15 deletions(-) diff --git a/comfy/ldm/flux/math.py b/comfy/ldm/flux/math.py index f9597de5b..6a22df8bc 100644 --- a/comfy/ldm/flux/math.py +++ b/comfy/ldm/flux/math.py @@ -4,7 +4,6 @@ from torch import Tensor from comfy.ldm.modules.attention import optimized_attention import comfy.model_management -import logging def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor, mask=None, transformer_options={}) -> Tensor: @@ -14,6 +13,7 @@ def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor, mask=None, transforme x = optimized_attention(q, k, v, heads, skip_reshape=True, mask=mask, transformer_options=transformer_options) return x + def rope(pos: Tensor, dim: int, theta: int) -> Tensor: assert dim % 2 == 0 if comfy.model_management.is_device_mps(pos.device) or comfy.model_management.is_intel_xpu() or comfy.model_management.is_directml_enabled(): @@ -28,20 +28,13 @@ def rope(pos: Tensor, dim: int, theta: int) -> Tensor: out = rearrange(out, "b n d (i j) -> b n d i j", i=2, j=2) return out.to(dtype=torch.float32, device=pos.device) +def apply_rope1(x: Tensor, freqs_cis: Tensor): + x_ = x.to(dtype=freqs_cis.dtype).reshape(*x.shape[:-1], -1, 1, 2) -try: - import comfy.quant_ops - apply_rope = comfy.quant_ops.ck.apply_rope - apply_rope1 = comfy.quant_ops.ck.apply_rope1 -except: - logging.warning("No comfy kitchen, using old apply_rope functions.") - def apply_rope1(x: Tensor, freqs_cis: Tensor): - x_ = x.to(dtype=freqs_cis.dtype).reshape(*x.shape[:-1], -1, 1, 2) + x_out = freqs_cis[..., 0] * x_[..., 0] + x_out.addcmul_(freqs_cis[..., 1], x_[..., 1]) - x_out = freqs_cis[..., 0] * x_[..., 0] - x_out.addcmul_(freqs_cis[..., 1], x_[..., 1]) + return x_out.reshape(*x.shape).type_as(x) - return x_out.reshape(*x.shape).type_as(x) - - def apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor): - return apply_rope1(xq, freqs_cis), apply_rope1(xk, freqs_cis) +def apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor): + return apply_rope1(xq, freqs_cis), apply_rope1(xk, freqs_cis) From e14f3b661069971163ddc56036b0f486933b9162 Mon Sep 17 00:00:00 2001 From: ComfyUI Wiki Date: Tue, 6 Jan 2026 14:37:11 +0800 Subject: [PATCH 377/647] chore: update workflow templates to v0.7.66 (#11652) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 0ee152032..9c9c0e29e 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,5 +1,5 @@ comfyui-frontend-package==1.35.9 -comfyui-workflow-templates==0.7.65 +comfyui-workflow-templates==0.7.66 comfyui-embedded-docs==0.3.1 torch torchsde From 96e0d0924e027248733bc6e0b8102dcdc8acde33 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 6 Jan 2026 11:43:24 -0800 Subject: [PATCH 378/647] Add helpful message to portable. (#11671) --- .../advanced/run_nvidia_gpu_disable_api_nodes.bat | 2 +- .ci/windows_nvidia_base_files/run_nvidia_gpu.bat | 2 +- .../run_nvidia_gpu_fast_fp16_accumulation.bat | 2 +- 3 files changed, 3 insertions(+), 3 deletions(-) diff --git a/.ci/windows_nvidia_base_files/advanced/run_nvidia_gpu_disable_api_nodes.bat b/.ci/windows_nvidia_base_files/advanced/run_nvidia_gpu_disable_api_nodes.bat index ed00583b6..4501ef9a1 100644 --- a/.ci/windows_nvidia_base_files/advanced/run_nvidia_gpu_disable_api_nodes.bat +++ b/.ci/windows_nvidia_base_files/advanced/run_nvidia_gpu_disable_api_nodes.bat @@ -1,3 +1,3 @@ ..\python_embeded\python.exe -s ..\ComfyUI\main.py --windows-standalone-build --disable-api-nodes -echo If you see this and ComfyUI did not start try updating your Nvidia Drivers to the latest. +echo If you see this and ComfyUI did not start try updating your Nvidia Drivers to the latest. If you get a c10.dll error you need to install vc redist that you can find: https://aka.ms/vc14/vc_redist.x64.exe pause diff --git a/.ci/windows_nvidia_base_files/run_nvidia_gpu.bat b/.ci/windows_nvidia_base_files/run_nvidia_gpu.bat index 4898a424f..6487ac7ce 100755 --- a/.ci/windows_nvidia_base_files/run_nvidia_gpu.bat +++ b/.ci/windows_nvidia_base_files/run_nvidia_gpu.bat @@ -1,3 +1,3 @@ .\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build -echo If you see this and ComfyUI did not start try updating your Nvidia Drivers to the latest. +echo If you see this and ComfyUI did not start try updating your Nvidia Drivers to the latest. If you get a c10.dll error you need to install vc redist that you can find: https://aka.ms/vc14/vc_redist.x64.exe pause diff --git a/.ci/windows_nvidia_base_files/run_nvidia_gpu_fast_fp16_accumulation.bat b/.ci/windows_nvidia_base_files/run_nvidia_gpu_fast_fp16_accumulation.bat index 32611e4af..01c5bb33b 100644 --- a/.ci/windows_nvidia_base_files/run_nvidia_gpu_fast_fp16_accumulation.bat +++ b/.ci/windows_nvidia_base_files/run_nvidia_gpu_fast_fp16_accumulation.bat @@ -1,3 +1,3 @@ .\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --fast fp16_accumulation -echo If you see this and ComfyUI did not start try updating your Nvidia Drivers to the latest. +echo If you see this and ComfyUI did not start try updating your Nvidia Drivers to the latest. If you get a c10.dll error you need to install vc redist that you can find: https://aka.ms/vc14/vc_redist.x64.exe pause From 6ffc159bdd56d1ad73e954081def6a7f163e7a7f Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 6 Jan 2026 12:53:43 -0800 Subject: [PATCH 379/647] Update comfy-kitchen version to 0.2.1 (#11672) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 9c9c0e29e..22cb50e2d 100644 --- a/requirements.txt +++ b/requirements.txt @@ -21,7 +21,7 @@ psutil alembic SQLAlchemy av>=14.2.0 -comfy-kitchen>=0.2.0 +comfy-kitchen>=0.2.1 #non essential dependencies: kornia>=0.7.1 From c3c3e93c5bb3034175c17ef8beeb8fe8626c66ab Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 6 Jan 2026 13:57:50 -0800 Subject: [PATCH 380/647] Use rope functions from comfy kitchen. (#11674) --- comfy/ldm/flux/math.py | 23 +++++++++++++++-------- requirements.txt | 2 +- 2 files changed, 16 insertions(+), 9 deletions(-) diff --git a/comfy/ldm/flux/math.py b/comfy/ldm/flux/math.py index 6a22df8bc..f9597de5b 100644 --- a/comfy/ldm/flux/math.py +++ b/comfy/ldm/flux/math.py @@ -4,6 +4,7 @@ from torch import Tensor from comfy.ldm.modules.attention import optimized_attention import comfy.model_management +import logging def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor, mask=None, transformer_options={}) -> Tensor: @@ -13,7 +14,6 @@ def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor, mask=None, transforme x = optimized_attention(q, k, v, heads, skip_reshape=True, mask=mask, transformer_options=transformer_options) return x - def rope(pos: Tensor, dim: int, theta: int) -> Tensor: assert dim % 2 == 0 if comfy.model_management.is_device_mps(pos.device) or comfy.model_management.is_intel_xpu() or comfy.model_management.is_directml_enabled(): @@ -28,13 +28,20 @@ def rope(pos: Tensor, dim: int, theta: int) -> Tensor: out = rearrange(out, "b n d (i j) -> b n d i j", i=2, j=2) return out.to(dtype=torch.float32, device=pos.device) -def apply_rope1(x: Tensor, freqs_cis: Tensor): - x_ = x.to(dtype=freqs_cis.dtype).reshape(*x.shape[:-1], -1, 1, 2) - x_out = freqs_cis[..., 0] * x_[..., 0] - x_out.addcmul_(freqs_cis[..., 1], x_[..., 1]) +try: + import comfy.quant_ops + apply_rope = comfy.quant_ops.ck.apply_rope + apply_rope1 = comfy.quant_ops.ck.apply_rope1 +except: + logging.warning("No comfy kitchen, using old apply_rope functions.") + def apply_rope1(x: Tensor, freqs_cis: Tensor): + x_ = x.to(dtype=freqs_cis.dtype).reshape(*x.shape[:-1], -1, 1, 2) - return x_out.reshape(*x.shape).type_as(x) + x_out = freqs_cis[..., 0] * x_[..., 0] + x_out.addcmul_(freqs_cis[..., 1], x_[..., 1]) -def apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor): - return apply_rope1(xq, freqs_cis), apply_rope1(xk, freqs_cis) + return x_out.reshape(*x.shape).type_as(x) + + def apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor): + return apply_rope1(xq, freqs_cis), apply_rope1(xk, freqs_cis) diff --git a/requirements.txt b/requirements.txt index 22cb50e2d..7798cb179 100644 --- a/requirements.txt +++ b/requirements.txt @@ -21,7 +21,7 @@ psutil alembic SQLAlchemy av>=14.2.0 -comfy-kitchen>=0.2.1 +comfy-kitchen>=0.2.2 #non essential dependencies: kornia>=0.7.1 From c3566c0d765200068d26d0888f035504a50012f2 Mon Sep 17 00:00:00 2001 From: ComfyUI Wiki Date: Wed, 7 Jan 2026 06:28:29 +0800 Subject: [PATCH 381/647] chore: update workflow templates to v0.7.67 (#11667) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 7798cb179..caad0026a 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,5 +1,5 @@ comfyui-frontend-package==1.35.9 -comfyui-workflow-templates==0.7.66 +comfyui-workflow-templates==0.7.67 comfyui-embedded-docs==0.3.1 torch torchsde From 023cf13721cac256c323e2226319b766d07b1f36 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 6 Jan 2026 14:33:03 -0800 Subject: [PATCH 382/647] Fix lowvram issue with ltxv2 text encoder. (#11675) --- comfy/ldm/lightricks/embeddings_connector.py | 2 +- comfy/text_encoders/lt.py | 4 +++- 2 files changed, 4 insertions(+), 2 deletions(-) diff --git a/comfy/ldm/lightricks/embeddings_connector.py b/comfy/ldm/lightricks/embeddings_connector.py index f7a43f3c3..06f5ada89 100644 --- a/comfy/ldm/lightricks/embeddings_connector.py +++ b/comfy/ldm/lightricks/embeddings_connector.py @@ -276,7 +276,7 @@ class Embeddings1DConnector(nn.Module): max(1024, hidden_states.shape[1]) / self.num_learnable_registers ) learnable_registers = torch.tile( - self.learnable_registers, (num_registers_duplications, 1) + self.learnable_registers.to(hidden_states), (num_registers_duplications, 1) ) hidden_states = torch.cat((hidden_states, learnable_registers[hidden_states.shape[1]:].unsqueeze(0).repeat(hidden_states.shape[0], 1, 1)), dim=1) diff --git a/comfy/text_encoders/lt.py b/comfy/text_encoders/lt.py index 2c2d453e8..e5964e42b 100644 --- a/comfy/text_encoders/lt.py +++ b/comfy/text_encoders/lt.py @@ -86,17 +86,19 @@ class LTXAVTEModel(torch.nn.Module): ) def set_clip_options(self, options): + self.execution_device = options.get("execution_device", self.execution_device) self.gemma3_12b.set_clip_options(options) def reset_clip_options(self): self.gemma3_12b.reset_clip_options() + self.execution_device = None def encode_token_weights(self, token_weight_pairs): token_weight_pairs = token_weight_pairs["gemma3_12b"] out, pooled, extra = self.gemma3_12b.encode_token_weights(token_weight_pairs) out_device = out.device - out = out.movedim(1, -1).to(self.text_embedding_projection.weight.device) + out = out.movedim(1, -1).to(self.execution_device) out = 8.0 * (out - out.mean(dim=(1, 2), keepdim=True)) / (out.amax(dim=(1, 2), keepdim=True) - out.amin(dim=(1, 2), keepdim=True) + 1e-6) out = out.reshape((out.shape[0], out.shape[1], -1)) out = self.text_embedding_projection(out) From 6e9ee55cdd9e0eca6b5144063575b983f3311762 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 6 Jan 2026 14:41:27 -0800 Subject: [PATCH 383/647] Disable ltxav previews. (#11676) --- comfy/latent_formats.py | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/comfy/latent_formats.py b/comfy/latent_formats.py index 9bbe30b53..cb4f52ce1 100644 --- a/comfy/latent_formats.py +++ b/comfy/latent_formats.py @@ -408,7 +408,9 @@ class LTXV(LatentFormat): self.latent_rgb_factors_bias = [-0.0571, -0.1657, -0.2512] class LTXAV(LTXV): - pass + def __init__(self): + self.latent_rgb_factors = None + self.latent_rgb_factors_bias = None class HunyuanVideo(LatentFormat): latent_channels = 16 From 2c03884f5fb7fa213161dfe1e9a09a8e8c4b6062 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 6 Jan 2026 15:07:26 -0800 Subject: [PATCH 384/647] Skip fp4 matrix mult on devices that don't support it. (#11677) --- comfy/model_management.py | 10 ++++++++++ comfy/ops.py | 21 +++++++++++++++++---- 2 files changed, 27 insertions(+), 4 deletions(-) diff --git a/comfy/model_management.py b/comfy/model_management.py index 22f4de044..928282092 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -1504,6 +1504,16 @@ def supports_fp8_compute(device=None): return True +def supports_nvfp4_compute(device=None): + if not is_nvidia(): + return False + + props = torch.cuda.get_device_properties(device) + if props.major < 10: + return False + + return True + def extended_fp16_support(): # TODO: check why some models work with fp16 on newer torch versions but not on older if torch_version_numeric < (2, 7): diff --git a/comfy/ops.py b/comfy/ops.py index f5e1e9230..8f9fdce36 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -493,11 +493,12 @@ from .quant_ops import ( ) -def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_precision_mm=False): +def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_precision_mm=False, disabled=[]): class MixedPrecisionOps(manual_cast): _quant_config = quant_config _compute_dtype = compute_dtype _full_precision_mm = full_precision_mm + _disabled = disabled class Linear(torch.nn.Module, CastWeightBiasOp): def __init__( @@ -522,6 +523,7 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec self.tensor_class = None self._full_precision_mm = MixedPrecisionOps._full_precision_mm + self._full_precision_mm_config = False def reset_parameters(self): return None @@ -556,8 +558,12 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec self.weight = torch.nn.Parameter(weight.to(device=device, dtype=MixedPrecisionOps._compute_dtype), requires_grad=False) else: self.quant_format = layer_conf.get("format", None) + self._full_precision_mm_config = layer_conf.get("full_precision_matrix_mult", False) if not self._full_precision_mm: - self._full_precision_mm = layer_conf.get("full_precision_matrix_mult", False) + self._full_precision_mm = self._full_precision_mm_config + + if self.quant_format in MixedPrecisionOps._disabled: + self._full_precision_mm = True if self.quant_format is None: raise ValueError(f"Unknown quantization format for layer {layer_name}") @@ -630,7 +636,7 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec sd["{}weight_scale".format(prefix)] = self.weight._params.block_scale quant_conf = {"format": self.quant_format} - if self._full_precision_mm: + if self._full_precision_mm_config: quant_conf["full_precision_matrix_mult"] = True sd["{}comfy_quant".format(prefix)] = torch.tensor(list(json.dumps(quant_conf).encode('utf-8')), dtype=torch.uint8) return sd @@ -711,10 +717,17 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec def pick_operations(weight_dtype, compute_dtype, load_device=None, disable_fast_fp8=False, fp8_optimizations=False, model_config=None): fp8_compute = comfy.model_management.supports_fp8_compute(load_device) # TODO: if we support more ops this needs to be more granular + nvfp4_compute = comfy.model_management.supports_nvfp4_compute(load_device) if model_config and hasattr(model_config, 'quant_config') and model_config.quant_config: logging.info("Using mixed precision operations") - return mixed_precision_ops(model_config.quant_config, compute_dtype, full_precision_mm=not fp8_compute) + disabled = set() + if not nvfp4_compute: + disabled.add("nvfp4") + if not fp8_compute: + disabled.add("float8_e4m3fn") + disabled.add("float8_e5m2") + return mixed_precision_ops(model_config.quant_config, compute_dtype, disabled=disabled) if ( fp8_compute and From edee33f55ea27a1931475d3ea788fd6e9a81677b Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 6 Jan 2026 19:13:43 -0800 Subject: [PATCH 385/647] Disable comfy kitchen cuda if pytorch cuda less than 13 (#11681) --- comfy/quant_ops.py | 7 +++++++ 1 file changed, 7 insertions(+) diff --git a/comfy/quant_ops.py b/comfy/quant_ops.py index cd737726f..5a17bc6f5 100644 --- a/comfy/quant_ops.py +++ b/comfy/quant_ops.py @@ -13,6 +13,13 @@ try: get_layout_class, ) _CK_AVAILABLE = True + if torch.version.cuda is None: + ck.registry.disable("cuda") + else: + cuda_version = tuple(map(int, str(torch.version.cuda).split('.'))) + if cuda_version < (13,): + ck.registry.disable("cuda") + ck.registry.disable("triton") for k, v in ck.list_backends().items(): logging.info(f"Found comfy_kitchen backend {k}: {v}") From c5cfb34c07048350f472a9a4f1ccbf75a56ed38f Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 6 Jan 2026 20:51:45 -0800 Subject: [PATCH 386/647] Update comfy-kitchen version to 0.2.3 (#11685) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index caad0026a..bc8346bcf 100644 --- a/requirements.txt +++ b/requirements.txt @@ -21,7 +21,7 @@ psutil alembic SQLAlchemy av>=14.2.0 -comfy-kitchen>=0.2.2 +comfy-kitchen>=0.2.3 #non essential dependencies: kornia>=0.7.1 From ce0000c4f2a7dba12324585dddb784b43e3cd3d0 Mon Sep 17 00:00:00 2001 From: Yoland Yan <4950057+yoland68@users.noreply.github.com> Date: Tue, 6 Jan 2026 21:57:31 -0800 Subject: [PATCH 387/647] Force sequential execution in CI test jobs (#11687) Added max-parallel setting to enforce sequential execution in test jobs. --- .github/workflows/test-ci.yml | 2 ++ 1 file changed, 2 insertions(+) diff --git a/.github/workflows/test-ci.yml b/.github/workflows/test-ci.yml index adfc5dd32..63df2dc3a 100644 --- a/.github/workflows/test-ci.yml +++ b/.github/workflows/test-ci.yml @@ -20,6 +20,7 @@ jobs: test-stable: strategy: fail-fast: false + max-parallel: 1 # This forces sequential execution matrix: # os: [macos, linux, windows] # os: [macos, linux] @@ -74,6 +75,7 @@ jobs: test-unix-nightly: strategy: fail-fast: false + max-parallel: 1 # This forces sequential execution matrix: # os: [macos, linux] os: [linux] From 79e94544bd7ec0cc7a4e5e6167907e7d781c4b76 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Wed, 7 Jan 2026 08:04:50 +0200 Subject: [PATCH 388/647] feat(api-nodes): add WAN2.6 ReferenceToVideo (#11644) --- comfy_api_nodes/nodes_wan.py | 160 +++++++++++++++++++++++++ comfy_api_nodes/util/upload_helpers.py | 2 +- 2 files changed, 161 insertions(+), 1 deletion(-) diff --git a/comfy_api_nodes/nodes_wan.py b/comfy_api_nodes/nodes_wan.py index 1675fd863..3e04786a9 100644 --- a/comfy_api_nodes/nodes_wan.py +++ b/comfy_api_nodes/nodes_wan.py @@ -13,7 +13,9 @@ from comfy_api_nodes.util import ( poll_op, sync_op, tensor_to_base64_string, + upload_video_to_comfyapi, validate_audio_duration, + validate_video_duration, ) @@ -41,6 +43,12 @@ class Image2VideoInputField(BaseModel): audio_url: str | None = Field(None) +class Reference2VideoInputField(BaseModel): + prompt: str = Field(...) + negative_prompt: str | None = Field(None) + reference_video_urls: list[str] = Field(...) + + class Txt2ImageParametersField(BaseModel): size: str = Field(...) n: int = Field(1, description="Number of images to generate.") # we support only value=1 @@ -76,6 +84,14 @@ class Image2VideoParametersField(BaseModel): shot_type: str = Field("single") +class Reference2VideoParametersField(BaseModel): + size: str = Field(...) + duration: int = Field(5, ge=5, le=15) + shot_type: str = Field("single") + seed: int = Field(..., ge=0, le=2147483647) + watermark: bool = Field(False) + + class Text2ImageTaskCreationRequest(BaseModel): model: str = Field(...) input: Text2ImageInputField = Field(...) @@ -100,6 +116,12 @@ class Image2VideoTaskCreationRequest(BaseModel): parameters: Image2VideoParametersField = Field(...) +class Reference2VideoTaskCreationRequest(BaseModel): + model: str = Field(...) + input: Reference2VideoInputField = Field(...) + parameters: Reference2VideoParametersField = Field(...) + + class TaskCreationOutputField(BaseModel): task_id: str = Field(...) task_status: str = Field(...) @@ -721,6 +743,143 @@ class WanImageToVideoApi(IO.ComfyNode): return IO.NodeOutput(await download_url_to_video_output(response.output.video_url)) +class WanReferenceVideoApi(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="WanReferenceVideoApi", + display_name="Wan Reference to Video", + category="api node/video/Wan", + description="Use the character and voice from input videos, combined with a prompt, " + "to generate a new video that maintains character consistency.", + inputs=[ + IO.Combo.Input("model", options=["wan2.6-r2v"]), + IO.String.Input( + "prompt", + multiline=True, + default="", + tooltip="Prompt describing the elements and visual features. Supports English and Chinese. " + "Use identifiers such as `character1` and `character2` to refer to the reference characters.", + ), + IO.String.Input( + "negative_prompt", + multiline=True, + default="", + tooltip="Negative prompt describing what to avoid.", + ), + IO.Autogrow.Input( + "reference_videos", + template=IO.Autogrow.TemplateNames( + IO.Video.Input("reference_video"), + names=["character1", "character2", "character3"], + min=1, + ), + ), + IO.Combo.Input( + "size", + options=[ + "720p: 1:1 (960x960)", + "720p: 16:9 (1280x720)", + "720p: 9:16 (720x1280)", + "720p: 4:3 (1088x832)", + "720p: 3:4 (832x1088)", + "1080p: 1:1 (1440x1440)", + "1080p: 16:9 (1920x1080)", + "1080p: 9:16 (1080x1920)", + "1080p: 4:3 (1632x1248)", + "1080p: 3:4 (1248x1632)", + ], + ), + IO.Int.Input( + "duration", + default=5, + min=5, + max=10, + step=5, + display_mode=IO.NumberDisplay.slider, + ), + IO.Int.Input( + "seed", + default=0, + min=0, + max=2147483647, + step=1, + display_mode=IO.NumberDisplay.number, + control_after_generate=True, + ), + IO.Combo.Input( + "shot_type", + options=["single", "multi"], + tooltip="Specifies the shot type for the generated video, that is, whether the video is a " + "single continuous shot or multiple shots with cuts.", + ), + IO.Boolean.Input( + "watermark", + default=False, + tooltip="Whether to add an AI-generated watermark to the result.", + ), + ], + outputs=[ + IO.Video.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + ) + + @classmethod + async def execute( + cls, + model: str, + prompt: str, + negative_prompt: str, + reference_videos: IO.Autogrow.Type, + size: str, + duration: int, + seed: int, + shot_type: str, + watermark: bool, + ): + reference_video_urls = [] + for i in reference_videos: + validate_video_duration(reference_videos[i], min_duration=2, max_duration=30) + for i in reference_videos: + reference_video_urls.append(await upload_video_to_comfyapi(cls, reference_videos[i])) + width, height = RES_IN_PARENS.search(size).groups() + initial_response = await sync_op( + cls, + ApiEndpoint(path="/proxy/wan/api/v1/services/aigc/video-generation/video-synthesis", method="POST"), + response_model=TaskCreationResponse, + data=Reference2VideoTaskCreationRequest( + model=model, + input=Reference2VideoInputField( + prompt=prompt, negative_prompt=negative_prompt, reference_video_urls=reference_video_urls + ), + parameters=Reference2VideoParametersField( + size=f"{width}*{height}", + duration=duration, + shot_type=shot_type, + watermark=watermark, + seed=seed, + ), + ), + ) + if not initial_response.output: + raise Exception(f"An unknown error occurred: {initial_response.code} - {initial_response.message}") + response = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/wan/api/v1/tasks/{initial_response.output.task_id}"), + response_model=VideoTaskStatusResponse, + status_extractor=lambda x: x.output.task_status, + poll_interval=6, + max_poll_attempts=280, + ) + return IO.NodeOutput(await download_url_to_video_output(response.output.video_url)) + + class WanApiExtension(ComfyExtension): @override async def get_node_list(self) -> list[type[IO.ComfyNode]]: @@ -729,6 +888,7 @@ class WanApiExtension(ComfyExtension): WanImageToImageApi, WanTextToVideoApi, WanImageToVideoApi, + WanReferenceVideoApi, ] diff --git a/comfy_api_nodes/util/upload_helpers.py b/comfy_api_nodes/util/upload_helpers.py index b8d33f4d1..f1ed7fe9c 100644 --- a/comfy_api_nodes/util/upload_helpers.py +++ b/comfy_api_nodes/util/upload_helpers.py @@ -119,7 +119,7 @@ async def upload_video_to_comfyapi( raise ValueError(f"Could not verify video duration from source: {e}") from e upload_mime_type = f"video/{container.value.lower()}" - filename = f"uploaded_video.{container.value.lower()}" + filename = f"{uuid.uuid4()}.{container.value.lower()}" # Convert VideoInput to BytesIO using specified container/codec video_bytes_io = BytesIO() From b7d7cc1d496afe3c82279eec74c4d47399aab8ea Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 6 Jan 2026 22:39:06 -0800 Subject: [PATCH 389/647] Fix fp8 fast issue. (#11688) --- comfy/ops.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/comfy/ops.py b/comfy/ops.py index 8f9fdce36..cd536e22d 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -427,12 +427,12 @@ def fp8_linear(self, input): input = torch.clamp(input, min=-448, max=448, out=input) input_fp8 = input.to(dtype).contiguous() layout_params_input = TensorCoreFP8Layout.Params(scale=scale_input, orig_dtype=input_dtype, orig_shape=tuple(input_fp8.shape)) - quantized_input = QuantizedTensor(input_fp8, TensorCoreFP8Layout, layout_params_input) + quantized_input = QuantizedTensor(input_fp8, "TensorCoreFP8Layout", layout_params_input) # Wrap weight in QuantizedTensor - this enables unified dispatch # Call F.linear - __torch_dispatch__ routes to fp8_linear handler in quant_ops.py! layout_params_weight = TensorCoreFP8Layout.Params(scale=scale_weight, orig_dtype=input_dtype, orig_shape=tuple(w.shape)) - quantized_weight = QuantizedTensor(w, TensorCoreFP8Layout, layout_params_weight) + quantized_weight = QuantizedTensor(w, "TensorCoreFP8Layout", layout_params_weight) o = torch.nn.functional.linear(quantized_input, quantized_weight, bias) uncast_bias_weight(self, w, bias, offload_stream) From fc0cb10bcbee6e73ed3caf34c27f7bde4559a07f Mon Sep 17 00:00:00 2001 From: comfyanonymous Date: Wed, 7 Jan 2026 04:07:31 -0500 Subject: [PATCH 390/647] ComfyUI v0.8.0 --- comfyui_version.py | 2 +- pyproject.toml | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/comfyui_version.py b/comfyui_version.py index 1ed60fe5c..750673f08 100644 --- a/comfyui_version.py +++ b/comfyui_version.py @@ -1,3 +1,3 @@ # This file is automatically generated by the build process when version is # updated in pyproject.toml. -__version__ = "0.7.0" +__version__ = "0.8.0" diff --git a/pyproject.toml b/pyproject.toml index a7d159be9..951c2c978 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -1,6 +1,6 @@ [project] name = "ComfyUI" -version = "0.7.0" +version = "0.8.0" readme = "README.md" license = { file = "LICENSE" } requires-python = ">=3.10" From c0c9720d77774ed2c87981da87189fe1c14a57fa Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 7 Jan 2026 01:48:28 -0800 Subject: [PATCH 391/647] Fix stable release workflow not pulling latest comfy kitchen. (#11695) --- .github/workflows/stable-release.yml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/.github/workflows/stable-release.yml b/.github/workflows/stable-release.yml index 28484a9d1..f501b7b31 100644 --- a/.github/workflows/stable-release.yml +++ b/.github/workflows/stable-release.yml @@ -117,7 +117,7 @@ jobs: ./python.exe get-pip.py ./python.exe -s -m pip install ../${{ inputs.cache_tag }}_python_deps/* - grep comfyui ../ComfyUI/requirements.txt > ./requirements_comfyui.txt + grep comfy ../ComfyUI/requirements.txt > ./requirements_comfyui.txt ./python.exe -s -m pip install -r requirements_comfyui.txt rm requirements_comfyui.txt From 3cd7b32f1b7e7e90395cefe7d9f9b1f89276d8ce Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 7 Jan 2026 02:15:14 -0800 Subject: [PATCH 392/647] Support gemma 12B with quant weights. (#11696) --- comfy/text_encoders/lt.py | 12 ++++++------ 1 file changed, 6 insertions(+), 6 deletions(-) diff --git a/comfy/text_encoders/lt.py b/comfy/text_encoders/lt.py index e5964e42b..130ebaeae 100644 --- a/comfy/text_encoders/lt.py +++ b/comfy/text_encoders/lt.py @@ -36,10 +36,10 @@ class LTXAVGemmaTokenizer(sd1_clip.SD1Tokenizer): class Gemma3_12BModel(sd1_clip.SDClipModel): def __init__(self, device="cpu", layer="all", layer_idx=None, dtype=None, attention_mask=True, model_options={}): - llama_scaled_fp8 = model_options.get("gemma_scaled_fp8", None) - if llama_scaled_fp8 is not None: + llama_quantization_metadata = model_options.get("llama_quantization_metadata", None) + if llama_quantization_metadata is not None: model_options = model_options.copy() - model_options["scaled_fp8"] = llama_scaled_fp8 + model_options["quantization_metadata"] = llama_quantization_metadata super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"start": 2, "pad": 0}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Gemma3_12B, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options) @@ -119,12 +119,12 @@ class LTXAVTEModel(torch.nn.Module): return self.load_state_dict(sdo, strict=False) -def ltxav_te(dtype_llama=None, llama_scaled_fp8=None): +def ltxav_te(dtype_llama=None, llama_quantization_metadata=None): class LTXAVTEModel_(LTXAVTEModel): def __init__(self, device="cpu", dtype=None, model_options={}): - if llama_scaled_fp8 is not None and "llama_scaled_fp8" not in model_options: + if llama_quantization_metadata is not None: model_options = model_options.copy() - model_options["llama_scaled_fp8"] = llama_scaled_fp8 + model_options["llama_quantization_metadata"] = llama_quantization_metadata if dtype_llama is not None: dtype = dtype_llama super().__init__(dtype_llama=dtype_llama, device=device, dtype=dtype, model_options=model_options) From 48e5ea1dfd23a9cb5d118d7af661b026d66743bc Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Wed, 7 Jan 2026 15:39:20 -0800 Subject: [PATCH 393/647] model_patcher: Remove confusing load stat (#11710) If the loader passes 1e32 as the usable memory size, it means force the full load. This happens with CPU loads and a few other misc cases. Removing the confusing number and just leave the other details. --- comfy/model_patcher.py | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-) diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index 93d26c690..4528814ad 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -790,11 +790,12 @@ class ModelPatcher: for param in params: self.pin_weight_to_device("{}.{}".format(n, param)) + usable_stat = "{:.2f} MB usable,".format(lowvram_model_memory / (1024 * 1024)) if lowvram_model_memory < 1e32 else "" if lowvram_counter > 0: - logging.info("loaded partially; {:.2f} MB usable, {:.2f} MB loaded, {:.2f} MB offloaded, {:.2f} MB buffer reserved, lowvram patches: {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (1024 * 1024), lowvram_mem_counter / (1024 * 1024), offload_buffer / (1024 * 1024), patch_counter)) + logging.info("loaded partially; {} {:.2f} MB loaded, {:.2f} MB offloaded, {:.2f} MB buffer reserved, lowvram patches: {}".format(usable_stat, mem_counter / (1024 * 1024), lowvram_mem_counter / (1024 * 1024), offload_buffer / (1024 * 1024), patch_counter)) self.model.model_lowvram = True else: - logging.info("loaded completely; {:.2f} MB usable, {:.2f} MB loaded, full load: {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (1024 * 1024), full_load)) + logging.info("loaded completely; {} {:.2f} MB loaded, full load: {}".format(usable_stat, mem_counter / (1024 * 1024), full_load)) self.model.model_lowvram = False if full_load: self.model.to(device_to) From 1c705f7bfb0fb59f6213dfb85ec5d5dc2ce4300e Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Jukka=20Sepp=C3=A4nen?= <40791699+kijai@users.noreply.github.com> Date: Thu, 8 Jan 2026 01:39:59 +0200 Subject: [PATCH 394/647] Add device selection for LTXAVTextEncoderLoader (#11700) --- comfy_extras/nodes_lt_audio.py | 10 +++++++++- 1 file changed, 9 insertions(+), 1 deletion(-) diff --git a/comfy_extras/nodes_lt_audio.py b/comfy_extras/nodes_lt_audio.py index 26b0160d2..1966fd1bf 100644 --- a/comfy_extras/nodes_lt_audio.py +++ b/comfy_extras/nodes_lt_audio.py @@ -185,6 +185,10 @@ class LTXAVTextEncoderLoader(io.ComfyNode): io.Combo.Input( "ckpt_name", options=folder_paths.get_filename_list("checkpoints"), + ), + io.Combo.Input( + "device", + options=["default", "cpu"], ) ], outputs=[io.Clip.Output()], @@ -197,7 +201,11 @@ class LTXAVTextEncoderLoader(io.ComfyNode): clip_path1 = folder_paths.get_full_path_or_raise("text_encoders", text_encoder) clip_path2 = folder_paths.get_full_path_or_raise("checkpoints", ckpt_name) - clip = comfy.sd.load_clip(ckpt_paths=[clip_path1, clip_path2], embedding_directory=folder_paths.get_folder_paths("embeddings"), clip_type=clip_type) + model_options = {} + if device == "cpu": + model_options["load_device"] = model_options["offload_device"] = torch.device("cpu") + + clip = comfy.sd.load_clip(ckpt_paths=[clip_path1, clip_path2], embedding_directory=folder_paths.get_folder_paths("embeddings"), clip_type=clip_type, model_options=model_options) return io.NodeOutput(clip) From 34751fe9f9ade0c715768202c19211dc0c72e760 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 7 Jan 2026 16:12:15 -0800 Subject: [PATCH 395/647] Lower ltxv text encoder vram use. (#11713) --- comfy/text_encoders/lt.py | 3 +++ 1 file changed, 3 insertions(+) diff --git a/comfy/text_encoders/lt.py b/comfy/text_encoders/lt.py index 130ebaeae..dc0694e0e 100644 --- a/comfy/text_encoders/lt.py +++ b/comfy/text_encoders/lt.py @@ -98,10 +98,13 @@ class LTXAVTEModel(torch.nn.Module): out, pooled, extra = self.gemma3_12b.encode_token_weights(token_weight_pairs) out_device = out.device + if comfy.model_management.should_use_bf16(self.execution_device): + out = out.to(device=self.execution_device, dtype=torch.bfloat16) out = out.movedim(1, -1).to(self.execution_device) out = 8.0 * (out - out.mean(dim=(1, 2), keepdim=True)) / (out.amax(dim=(1, 2), keepdim=True) - out.amin(dim=(1, 2), keepdim=True) + 1e-6) out = out.reshape((out.shape[0], out.shape[1], -1)) out = self.text_embedding_projection(out) + out = out.float() out_vid = self.video_embeddings_connector(out)[0] out_audio = self.audio_embeddings_connector(out)[0] out = torch.concat((out_vid, out_audio), dim=-1) From 007b87e7ac29e55ce0ad2c436f5ae68f3a078080 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 7 Jan 2026 16:48:47 -0800 Subject: [PATCH 396/647] Bump required comfy-kitchen version. (#11714) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index bc8346bcf..13e95afa0 100644 --- a/requirements.txt +++ b/requirements.txt @@ -21,7 +21,7 @@ psutil alembic SQLAlchemy av>=14.2.0 -comfy-kitchen>=0.2.3 +comfy-kitchen>=0.2.5 #non essential dependencies: kornia>=0.7.1 From 3cd19e99c10a25cf6e6b51b82e3c16c501733b8c Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 7 Jan 2026 17:04:56 -0800 Subject: [PATCH 397/647] Increase ltxav mem estimation by a bit. (#11715) --- comfy/supported_models.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy/supported_models.py b/comfy/supported_models.py index ee9a79001..d44c0bc37 100644 --- a/comfy/supported_models.py +++ b/comfy/supported_models.py @@ -845,7 +845,7 @@ class LTXAV(LTXV): def __init__(self, unet_config): super().__init__(unet_config) - self.memory_usage_factor = 0.055 # TODO + self.memory_usage_factor = 0.061 # TODO def get_model(self, state_dict, prefix="", device=None): out = model_base.LTXAV(self, device=device) From 25bc1b5b57d61930d6ab60d8cf7e9241d26e4fe9 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 7 Jan 2026 17:11:22 -0800 Subject: [PATCH 398/647] Add memory estimation function to ltxav text encoder. (#11716) --- comfy/sd.py | 11 +++++++---- comfy/text_encoders/lt.py | 8 ++++++++ 2 files changed, 15 insertions(+), 4 deletions(-) diff --git a/comfy/sd.py b/comfy/sd.py index 32157e18b..efde3839c 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -218,7 +218,7 @@ class CLIP: if unprojected: self.cond_stage_model.set_clip_options({"projected_pooled": False}) - self.load_model() + self.load_model(tokens) self.cond_stage_model.set_clip_options({"execution_device": self.patcher.load_device}) all_hooks.reset() self.patcher.patch_hooks(None) @@ -266,7 +266,7 @@ class CLIP: if return_pooled == "unprojected": self.cond_stage_model.set_clip_options({"projected_pooled": False}) - self.load_model() + self.load_model(tokens) self.cond_stage_model.set_clip_options({"execution_device": self.patcher.load_device}) o = self.cond_stage_model.encode_token_weights(tokens) cond, pooled = o[:2] @@ -299,8 +299,11 @@ class CLIP: sd_clip[k] = sd_tokenizer[k] return sd_clip - def load_model(self): - model_management.load_model_gpu(self.patcher) + def load_model(self, tokens={}): + memory_used = 0 + if hasattr(self.cond_stage_model, "memory_estimation_function"): + memory_used = self.cond_stage_model.memory_estimation_function(tokens, device=self.patcher.load_device) + model_management.load_models_gpu([self.patcher], memory_required=memory_used) return self.patcher def get_key_patches(self): diff --git a/comfy/text_encoders/lt.py b/comfy/text_encoders/lt.py index dc0694e0e..776e25e97 100644 --- a/comfy/text_encoders/lt.py +++ b/comfy/text_encoders/lt.py @@ -121,6 +121,14 @@ class LTXAVTEModel(torch.nn.Module): return self.load_state_dict(sdo, strict=False) + def memory_estimation_function(self, token_weight_pairs, device=None): + constant = 6.0 + if comfy.model_management.should_use_bf16(device): + constant /= 2.0 + + token_weight_pairs = token_weight_pairs.get("gemma3_12b", []) + num_tokens = sum(map(lambda a: len(a), token_weight_pairs)) + return num_tokens * constant * 1024 * 1024 def ltxav_te(dtype_llama=None, llama_quantization_metadata=None): class LTXAVTEModel_(LTXAVTEModel): From b6c79a648a013f477f514f61580d1a06220b15eb Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Wed, 7 Jan 2026 18:01:16 -0800 Subject: [PATCH 399/647] ops: Fix offloading with FP8MM performance (#11697) This logic was checking comfy_cast_weights, and going straight to to the forward_comfy_cast_weights implementation without attempting to downscale input to fp8 in the event comfy_cast_weights is set. The main reason comfy_cast_weights would be set would be for async offload, which is not a good reason to nix FP8MM. So instead, and together the underlying exclusions for FP8MM which are: * having a weight_function (usually LowVramPatch) * force_cast_weights (compute dtype override) * the weight is not Quantized * the input is already quantized * the model or layer has MM explictily disabled. If you get past all of those exclusions, quantize the input tensor. Then hand the new input, quantized or not off to forward_comfy_cast_weights to handle it. If the weight is offloaded but input is quantized you will get an offloaded MM8. --- comfy/model_patcher.py | 1 + comfy/ops.py | 30 +++++++++++++++--------------- 2 files changed, 16 insertions(+), 15 deletions(-) diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index 4528814ad..f6b80a40f 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -718,6 +718,7 @@ class ModelPatcher: continue cast_weight = self.force_cast_weights + m.comfy_force_cast_weights = self.force_cast_weights if lowvram_weight: if hasattr(m, "comfy_cast_weights"): m.weight_function = [] diff --git a/comfy/ops.py b/comfy/ops.py index cd536e22d..8156c42ff 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -654,29 +654,29 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec run_every_op() input_shape = input.shape - tensor_3d = input.ndim == 3 - - if self._full_precision_mm or self.comfy_cast_weights or len(self.weight_function) > 0 or len(self.bias_function) > 0: - return self.forward_comfy_cast_weights(input, *args, **kwargs) + reshaped_3d = False if (getattr(self, 'layout_type', None) is not None and - not isinstance(input, QuantizedTensor)): + not isinstance(input, QuantizedTensor) and not self._full_precision_mm and + not getattr(self, 'comfy_force_cast_weights', False) and + len(self.weight_function) == 0 and len(self.bias_function) == 0): # Reshape 3D tensors to 2D for quantization (needed for NVFP4 and others) - if tensor_3d: - input = input.reshape(-1, input_shape[2]) + input_reshaped = input.reshape(-1, input_shape[2]) if input.ndim == 3 else input - if input.ndim != 2: - # Fall back to comfy_cast_weights for non-2D tensors - return self.forward_comfy_cast_weights(input.reshape(input_shape), *args, **kwargs) + # Fall back to non-quantized for non-2D tensors + if input_reshaped.ndim == 2: + reshaped_3d = input.ndim == 3 + # dtype is now implicit in the layout class + scale = getattr(self, 'input_scale', None) + if scale is not None: + scale = comfy.model_management.cast_to_device(scale, input.device, None) + input = QuantizedTensor.from_float(input_reshaped, self.layout_type, scale=scale) - # dtype is now implicit in the layout class - input = QuantizedTensor.from_float(input, self.layout_type, scale=getattr(self, 'input_scale', None)) - - output = self._forward(input, self.weight, self.bias) + output = self.forward_comfy_cast_weights(input) # Reshape output back to 3D if input was 3D - if tensor_3d: + if reshaped_3d: output = output.reshape((input_shape[0], input_shape[1], self.weight.shape[0])) return output From 21e842508733809354a7b04944b2995ed1169370 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 7 Jan 2026 18:07:26 -0800 Subject: [PATCH 400/647] Add warning for old pytorch. (#11718) --- comfy/quant_ops.py | 1 + 1 file changed, 1 insertion(+) diff --git a/comfy/quant_ops.py b/comfy/quant_ops.py index 5a17bc6f5..8324be42a 100644 --- a/comfy/quant_ops.py +++ b/comfy/quant_ops.py @@ -19,6 +19,7 @@ try: cuda_version = tuple(map(int, str(torch.version.cuda).split('.'))) if cuda_version < (13,): ck.registry.disable("cuda") + logging.warning("WARNING: You need pytorch with cu130 or higher to use optimized CUDA operations.") ck.registry.disable("triton") for k, v in ck.list_backends().items(): From fcd9a236b091bd4e77b177134ddfcf7d7dbd71fd Mon Sep 17 00:00:00 2001 From: ComfyUI Wiki Date: Thu, 8 Jan 2026 10:22:23 +0800 Subject: [PATCH 401/647] Update template to 0.7.69 (#11719) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 13e95afa0..49567ad61 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,5 +1,5 @@ comfyui-frontend-package==1.35.9 -comfyui-workflow-templates==0.7.67 +comfyui-workflow-templates==0.7.69 comfyui-embedded-docs==0.3.1 torch torchsde From ac12f77bed7bbbaf20289533bf7c0bff275e4a41 Mon Sep 17 00:00:00 2001 From: comfyanonymous Date: Wed, 7 Jan 2026 22:10:08 -0500 Subject: [PATCH 402/647] ComfyUI version v0.8.1 --- comfyui_version.py | 2 +- pyproject.toml | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/comfyui_version.py b/comfyui_version.py index 750673f08..4eb6070fe 100644 --- a/comfyui_version.py +++ b/comfyui_version.py @@ -1,3 +1,3 @@ # This file is automatically generated by the build process when version is # updated in pyproject.toml. -__version__ = "0.8.0" +__version__ = "0.8.1" diff --git a/pyproject.toml b/pyproject.toml index 951c2c978..0037abd6c 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -1,6 +1,6 @@ [project] name = "ComfyUI" -version = "0.8.0" +version = "0.8.1" readme = "README.md" license = { file = "LICENSE" } requires-python = ">=3.10" From 50d6e1caf401bf72dca1e9df7e194e722e1bd98b Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 7 Jan 2026 20:07:05 -0800 Subject: [PATCH 403/647] Tweak ltxv vae mem estimation. (#11722) --- comfy/sd.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/comfy/sd.py b/comfy/sd.py index efde3839c..5a7221620 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -479,8 +479,8 @@ class VAE: self.first_stage_model = comfy.ldm.lightricks.vae.causal_video_autoencoder.VideoVAE(version=version, config=vae_config) self.latent_channels = 128 self.latent_dim = 3 - self.memory_used_decode = lambda shape, dtype: (900 * shape[2] * shape[3] * shape[4] * (8 * 8 * 8)) * model_management.dtype_size(dtype) - self.memory_used_encode = lambda shape, dtype: (70 * max(shape[2], 7) * shape[3] * shape[4]) * model_management.dtype_size(dtype) + self.memory_used_decode = lambda shape, dtype: (1200 * shape[2] * shape[3] * shape[4] * (8 * 8 * 8)) * model_management.dtype_size(dtype) + self.memory_used_encode = lambda shape, dtype: (80 * max(shape[2], 7) * shape[3] * shape[4]) * model_management.dtype_size(dtype) self.upscale_ratio = (lambda a: max(0, a * 8 - 7), 32, 32) self.upscale_index_formula = (8, 32, 32) self.downscale_ratio = (lambda a: max(0, math.floor((a + 7) / 8)), 32, 32) From 2e9d51680a90bca9cc375ba7767f7bf3ed27d563 Mon Sep 17 00:00:00 2001 From: comfyanonymous Date: Wed, 7 Jan 2026 23:50:02 -0500 Subject: [PATCH 404/647] ComfyUI version v0.8.2 --- comfyui_version.py | 2 +- pyproject.toml | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/comfyui_version.py b/comfyui_version.py index 4eb6070fe..df82ed4fc 100644 --- a/comfyui_version.py +++ b/comfyui_version.py @@ -1,3 +1,3 @@ # This file is automatically generated by the build process when version is # updated in pyproject.toml. -__version__ = "0.8.1" +__version__ = "0.8.2" diff --git a/pyproject.toml b/pyproject.toml index 0037abd6c..49f1a03fd 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -1,6 +1,6 @@ [project] name = "ComfyUI" -version = "0.8.1" +version = "0.8.2" readme = "README.md" license = { file = "LICENSE" } requires-python = ">=3.10" From a60b7b86c54ea1498e9c5a5c3d6018c0714654d9 Mon Sep 17 00:00:00 2001 From: Yoland Yan <4950057+yoland68@users.noreply.github.com> Date: Wed, 7 Jan 2026 21:41:57 -0800 Subject: [PATCH 405/647] Revert "Force sequential execution in CI test jobs (#11687)" (#11725) This reverts commit ce0000c4f2a7dba12324585dddb784b43e3cd3d0. --- .github/workflows/test-ci.yml | 2 -- 1 file changed, 2 deletions(-) diff --git a/.github/workflows/test-ci.yml b/.github/workflows/test-ci.yml index 63df2dc3a..adfc5dd32 100644 --- a/.github/workflows/test-ci.yml +++ b/.github/workflows/test-ci.yml @@ -20,7 +20,6 @@ jobs: test-stable: strategy: fail-fast: false - max-parallel: 1 # This forces sequential execution matrix: # os: [macos, linux, windows] # os: [macos, linux] @@ -75,7 +74,6 @@ jobs: test-unix-nightly: strategy: fail-fast: false - max-parallel: 1 # This forces sequential execution matrix: # os: [macos, linux] os: [linux] From 5943fbf457d78becbb924a74780e0efc68505a17 Mon Sep 17 00:00:00 2001 From: "Dr.Lt.Data" <128333288+ltdrdata@users.noreply.github.com> Date: Fri, 9 Jan 2026 01:15:42 +0900 Subject: [PATCH 406/647] bump comfyui_manager version to the 4.0.5 (#11732) --- manager_requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/manager_requirements.txt b/manager_requirements.txt index 6585b0c19..bea6d4927 100644 --- a/manager_requirements.txt +++ b/manager_requirements.txt @@ -1 +1 @@ -comfyui_manager==4.0.4 +comfyui_manager==4.0.5 From 0f11869d55c7a459371b8114b1345e55a0274723 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Thu, 8 Jan 2026 14:16:58 -0800 Subject: [PATCH 407/647] Better detection if AMD torch compiled with efficient attention. (#11745) --- comfy/model_management.py | 15 +++++++++++++-- 1 file changed, 13 insertions(+), 2 deletions(-) diff --git a/comfy/model_management.py b/comfy/model_management.py index 928282092..e5de4a5b5 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -22,7 +22,6 @@ from enum import Enum from comfy.cli_args import args, PerformanceFeature import torch import sys -import importlib import platform import weakref import gc @@ -349,10 +348,22 @@ try: except: rocm_version = (6, -1) + def aotriton_supported(gpu_arch): + path = torch.__path__[0] + path = os.path.join(os.path.join(path, "lib"), "aotriton.images") + gfx = set(map(lambda a: a[4:], filter(lambda a: a.startswith("amd-gfx"), os.listdir(path)))) + if gpu_arch in gfx: + return True + if "{}x".format(gpu_arch[:-1]) in gfx: + return True + if "{}xx".format(gpu_arch[:-2]) in gfx: + return True + return False + logging.info("AMD arch: {}".format(arch)) logging.info("ROCm version: {}".format(rocm_version)) if args.use_split_cross_attention == False and args.use_quad_cross_attention == False: - if importlib.util.find_spec('triton') is not None: # AMD efficient attention implementation depends on triton. TODO: better way of detecting if it's compiled in or not. + if aotriton_supported(arch): # AMD efficient attention implementation depends on aotriton. if torch_version_numeric >= (2, 7): # works on 2.6 but doesn't actually seem to improve much if any((a in arch) for a in ["gfx90a", "gfx942", "gfx1100", "gfx1101", "gfx1151"]): # TODO: more arches, TODO: gfx950 ENABLE_PYTORCH_ATTENTION = True From 1a206564487d672561d83ce3eb007517bf018995 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Thu, 8 Jan 2026 14:23:59 -0800 Subject: [PATCH 408/647] Fix import issue. (#11746) --- comfy/ldm/hunyuan_video/upsampler.py | 14 +++++++------- 1 file changed, 7 insertions(+), 7 deletions(-) diff --git a/comfy/ldm/hunyuan_video/upsampler.py b/comfy/ldm/hunyuan_video/upsampler.py index d9e76922f..51b6d1da8 100644 --- a/comfy/ldm/hunyuan_video/upsampler.py +++ b/comfy/ldm/hunyuan_video/upsampler.py @@ -3,8 +3,8 @@ import torch.nn as nn import torch.nn.functional as F from comfy.ldm.modules.diffusionmodules.model import ResnetBlock, VideoConv3d from comfy.ldm.hunyuan_video.vae_refiner import RMS_norm -import model_management -import model_patcher +import comfy.model_management +import comfy.model_patcher class SRResidualCausalBlock3D(nn.Module): def __init__(self, channels: int): @@ -103,13 +103,13 @@ UPSAMPLERS = { class HunyuanVideo15SRModel(): def __init__(self, model_type, config): - self.load_device = model_management.vae_device() - offload_device = model_management.vae_offload_device() - self.dtype = model_management.vae_dtype(self.load_device) + self.load_device = comfy.model_management.vae_device() + offload_device = comfy.model_management.vae_offload_device() + self.dtype = comfy.model_management.vae_dtype(self.load_device) self.model_class = UPSAMPLERS.get(model_type) self.model = self.model_class(**config).eval() - self.patcher = model_patcher.ModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device) + self.patcher = comfy.model_patcher.ModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device) def load_sd(self, sd): return self.model.load_state_dict(sd, strict=True) @@ -118,5 +118,5 @@ class HunyuanVideo15SRModel(): return self.model.state_dict() def resample_latent(self, latent): - model_management.load_model_gpu(self.patcher) + comfy.model_management.load_model_gpu(self.patcher) return self.model(latent.to(self.load_device)) From 027042db6811c875562296f0a6b797c89d59e426 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Jukka=20Sepp=C3=A4nen?= <40791699+kijai@users.noreply.github.com> Date: Fri, 9 Jan 2026 05:14:06 +0200 Subject: [PATCH 409/647] Add node: JoinAudioChannels (#11728) --- comfy_extras/nodes_audio.py | 53 +++++++++++++++++++++++++++++++++++++ 1 file changed, 53 insertions(+) diff --git a/comfy_extras/nodes_audio.py b/comfy_extras/nodes_audio.py index 94ad5e8a8..15b3aa401 100644 --- a/comfy_extras/nodes_audio.py +++ b/comfy_extras/nodes_audio.py @@ -399,6 +399,58 @@ class SplitAudioChannels(IO.ComfyNode): separate = execute # TODO: remove +class JoinAudioChannels(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="JoinAudioChannels", + display_name="Join Audio Channels", + description="Joins left and right mono audio channels into a stereo audio.", + category="audio", + inputs=[ + IO.Audio.Input("audio_left"), + IO.Audio.Input("audio_right"), + ], + outputs=[ + IO.Audio.Output(display_name="audio"), + ], + ) + + @classmethod + def execute(cls, audio_left, audio_right) -> IO.NodeOutput: + waveform_left = audio_left["waveform"] + sample_rate_left = audio_left["sample_rate"] + waveform_right = audio_right["waveform"] + sample_rate_right = audio_right["sample_rate"] + + if waveform_left.shape[1] != 1 or waveform_right.shape[1] != 1: + raise ValueError("AudioJoin: Both input audios must be mono.") + + # Handle different sample rates by resampling to the higher rate + waveform_left, waveform_right, output_sample_rate = match_audio_sample_rates( + waveform_left, sample_rate_left, waveform_right, sample_rate_right + ) + + # Handle different lengths by trimming to the shorter length + length_left = waveform_left.shape[-1] + length_right = waveform_right.shape[-1] + + if length_left != length_right: + min_length = min(length_left, length_right) + if length_left > min_length: + logging.info(f"JoinAudioChannels: Trimming left channel from {length_left} to {min_length} samples.") + waveform_left = waveform_left[..., :min_length] + if length_right > min_length: + logging.info(f"JoinAudioChannels: Trimming right channel from {length_right} to {min_length} samples.") + waveform_right = waveform_right[..., :min_length] + + # Join the channels into stereo + left_channel = waveform_left[..., 0:1, :] + right_channel = waveform_right[..., 0:1, :] + stereo_waveform = torch.cat([left_channel, right_channel], dim=1) + + return IO.NodeOutput({"waveform": stereo_waveform, "sample_rate": output_sample_rate}) + def match_audio_sample_rates(waveform_1, sample_rate_1, waveform_2, sample_rate_2): if sample_rate_1 != sample_rate_2: @@ -616,6 +668,7 @@ class AudioExtension(ComfyExtension): RecordAudio, TrimAudioDuration, SplitAudioChannels, + JoinAudioChannels, AudioConcat, AudioMerge, AudioAdjustVolume, From b48d6a83d4f7012a1b6f6f41e66b0ac3f3253b8a Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Thu, 8 Jan 2026 19:15:50 -0800 Subject: [PATCH 410/647] Fix csp error in frontend when forcing offline. (#11749) --- server.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/server.py b/server.py index 70c8b5e3b..4db3347cb 100644 --- a/server.py +++ b/server.py @@ -184,7 +184,7 @@ def create_block_external_middleware(): else: response = await handler(request) - response.headers['Content-Security-Policy'] = "default-src 'self'; script-src 'self' 'unsafe-inline' 'unsafe-eval' blob:; style-src 'self' 'unsafe-inline'; img-src 'self' data: blob:; font-src 'self'; connect-src 'self'; frame-src 'self'; object-src 'self';" + response.headers['Content-Security-Policy'] = "default-src 'self'; script-src 'self' 'unsafe-inline' 'unsafe-eval' blob:; style-src 'self' 'unsafe-inline'; img-src 'self' data: blob:; font-src 'self'; connect-src 'self' data:; frame-src 'self'; object-src 'self';" return response return block_external_middleware From 114fc73685129bf4e8ddced432247fe67dc6fbff Mon Sep 17 00:00:00 2001 From: Comfy Org PR Bot Date: Fri, 9 Jan 2026 12:16:15 +0900 Subject: [PATCH 411/647] Bump comfyui-frontend-package to 1.36.13 (#11645) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 49567ad61..7686a5f8a 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,4 +1,4 @@ -comfyui-frontend-package==1.35.9 +comfyui-frontend-package==1.36.13 comfyui-workflow-templates==0.7.69 comfyui-embedded-docs==0.3.1 torch From 1dc3da631423b776669a6a9128bb1aeaf5592c55 Mon Sep 17 00:00:00 2001 From: Jedrzej Kosinski Date: Thu, 8 Jan 2026 19:21:51 -0800 Subject: [PATCH 412/647] Add most basic Asset support for models (#11315) * Brought over minimal elements from PR 10045 to reproduce seed_assets and register_assets_system without adding anything to the DB or server routes yet, for now making everything sync (can introduce async once everything is cleaned up and brought over) * Added db script to insert assets stuff, cleaned up some code; assets (models) now get added/rescanned * Added support for 5 http endpoints for assets * Replaced Optional with | None in schemas_in.py and schemas_out.py * Remove two routes that will not be relevant yet in this PR: HEAD /api/assets/hash/ and PUT /api/assets//preview * Remove some functions the two deleted endpoints were using * Don't show assets scan message upon calling /object_info endpoint * removed unsued import to satisfy ruff * Simplified hashing function tpye hint and _hash_file_obj * Satisfied ruff --- alembic_db/versions/0001_assets.py | 174 +++++++++++++++++++ app/assets/api/routes.py | 102 +++++++++++ app/assets/api/schemas_in.py | 94 ++++++++++ app/assets/api/schemas_out.py | 60 +++++++ app/assets/database/bulk_ops.py | 188 ++++++++++++++++++++ app/assets/database/models.py | 233 +++++++++++++++++++++++++ app/assets/database/queries.py | 267 +++++++++++++++++++++++++++++ app/assets/database/tags.py | 62 +++++++ app/assets/hashing.py | 75 ++++++++ app/assets/helpers.py | 216 +++++++++++++++++++++++ app/assets/manager.py | 123 +++++++++++++ app/assets/scanner.py | 229 +++++++++++++++++++++++++ app/database/models.py | 25 ++- comfy/cli_args.py | 1 + main.py | 3 + server.py | 4 + 16 files changed, 1847 insertions(+), 9 deletions(-) create mode 100644 alembic_db/versions/0001_assets.py create mode 100644 app/assets/api/routes.py create mode 100644 app/assets/api/schemas_in.py create mode 100644 app/assets/api/schemas_out.py create mode 100644 app/assets/database/bulk_ops.py create mode 100644 app/assets/database/models.py create mode 100644 app/assets/database/queries.py create mode 100644 app/assets/database/tags.py create mode 100644 app/assets/hashing.py create mode 100644 app/assets/helpers.py create mode 100644 app/assets/manager.py create mode 100644 app/assets/scanner.py diff --git a/alembic_db/versions/0001_assets.py b/alembic_db/versions/0001_assets.py new file mode 100644 index 000000000..1e10b94dc --- /dev/null +++ b/alembic_db/versions/0001_assets.py @@ -0,0 +1,174 @@ +""" +Initial assets schema +Revision ID: 0001_assets +Revises: None +Create Date: 2025-12-10 00:00:00 +""" + +from alembic import op +import sqlalchemy as sa + +revision = "0001_assets" +down_revision = None +branch_labels = None +depends_on = None + + +def upgrade() -> None: + # ASSETS: content identity + op.create_table( + "assets", + sa.Column("id", sa.String(length=36), primary_key=True), + sa.Column("hash", sa.String(length=256), nullable=True), + sa.Column("size_bytes", sa.BigInteger(), nullable=False, server_default="0"), + sa.Column("mime_type", sa.String(length=255), nullable=True), + sa.Column("created_at", sa.DateTime(timezone=False), nullable=False), + sa.CheckConstraint("size_bytes >= 0", name="ck_assets_size_nonneg"), + ) + op.create_index("uq_assets_hash", "assets", ["hash"], unique=True) + op.create_index("ix_assets_mime_type", "assets", ["mime_type"]) + + # ASSETS_INFO: user-visible references + op.create_table( + "assets_info", + sa.Column("id", sa.String(length=36), primary_key=True), + sa.Column("owner_id", sa.String(length=128), nullable=False, server_default=""), + sa.Column("name", sa.String(length=512), nullable=False), + sa.Column("asset_id", sa.String(length=36), sa.ForeignKey("assets.id", ondelete="RESTRICT"), nullable=False), + sa.Column("preview_id", sa.String(length=36), sa.ForeignKey("assets.id", ondelete="SET NULL"), nullable=True), + sa.Column("user_metadata", sa.JSON(), nullable=True), + sa.Column("created_at", sa.DateTime(timezone=False), nullable=False), + sa.Column("updated_at", sa.DateTime(timezone=False), nullable=False), + sa.Column("last_access_time", sa.DateTime(timezone=False), nullable=False), + sa.UniqueConstraint("asset_id", "owner_id", "name", name="uq_assets_info_asset_owner_name"), + ) + op.create_index("ix_assets_info_owner_id", "assets_info", ["owner_id"]) + op.create_index("ix_assets_info_asset_id", "assets_info", ["asset_id"]) + op.create_index("ix_assets_info_name", "assets_info", ["name"]) + op.create_index("ix_assets_info_created_at", "assets_info", ["created_at"]) + op.create_index("ix_assets_info_last_access_time", "assets_info", ["last_access_time"]) + op.create_index("ix_assets_info_owner_name", "assets_info", ["owner_id", "name"]) + + # TAGS: normalized tag vocabulary + op.create_table( + "tags", + sa.Column("name", sa.String(length=512), primary_key=True), + sa.Column("tag_type", sa.String(length=32), nullable=False, server_default="user"), + sa.CheckConstraint("name = lower(name)", name="ck_tags_lowercase"), + ) + op.create_index("ix_tags_tag_type", "tags", ["tag_type"]) + + # ASSET_INFO_TAGS: many-to-many for tags on AssetInfo + op.create_table( + "asset_info_tags", + sa.Column("asset_info_id", sa.String(length=36), sa.ForeignKey("assets_info.id", ondelete="CASCADE"), nullable=False), + sa.Column("tag_name", sa.String(length=512), sa.ForeignKey("tags.name", ondelete="RESTRICT"), nullable=False), + sa.Column("origin", sa.String(length=32), nullable=False, server_default="manual"), + sa.Column("added_at", sa.DateTime(timezone=False), nullable=False), + sa.PrimaryKeyConstraint("asset_info_id", "tag_name", name="pk_asset_info_tags"), + ) + op.create_index("ix_asset_info_tags_tag_name", "asset_info_tags", ["tag_name"]) + op.create_index("ix_asset_info_tags_asset_info_id", "asset_info_tags", ["asset_info_id"]) + + # ASSET_CACHE_STATE: N:1 local cache rows per Asset + op.create_table( + "asset_cache_state", + sa.Column("id", sa.Integer(), primary_key=True, autoincrement=True), + sa.Column("asset_id", sa.String(length=36), sa.ForeignKey("assets.id", ondelete="CASCADE"), nullable=False), + sa.Column("file_path", sa.Text(), nullable=False), # absolute local path to cached file + sa.Column("mtime_ns", sa.BigInteger(), nullable=True), + sa.Column("needs_verify", sa.Boolean(), nullable=False, server_default=sa.text("false")), + sa.CheckConstraint("(mtime_ns IS NULL) OR (mtime_ns >= 0)", name="ck_acs_mtime_nonneg"), + sa.UniqueConstraint("file_path", name="uq_asset_cache_state_file_path"), + ) + op.create_index("ix_asset_cache_state_file_path", "asset_cache_state", ["file_path"]) + op.create_index("ix_asset_cache_state_asset_id", "asset_cache_state", ["asset_id"]) + + # ASSET_INFO_META: typed KV projection of user_metadata for filtering/sorting + op.create_table( + "asset_info_meta", + sa.Column("asset_info_id", sa.String(length=36), sa.ForeignKey("assets_info.id", ondelete="CASCADE"), nullable=False), + sa.Column("key", sa.String(length=256), nullable=False), + sa.Column("ordinal", sa.Integer(), nullable=False, server_default="0"), + sa.Column("val_str", sa.String(length=2048), nullable=True), + sa.Column("val_num", sa.Numeric(38, 10), nullable=True), + sa.Column("val_bool", sa.Boolean(), nullable=True), + sa.Column("val_json", sa.JSON(), nullable=True), + sa.PrimaryKeyConstraint("asset_info_id", "key", "ordinal", name="pk_asset_info_meta"), + ) + op.create_index("ix_asset_info_meta_key", "asset_info_meta", ["key"]) + op.create_index("ix_asset_info_meta_key_val_str", "asset_info_meta", ["key", "val_str"]) + op.create_index("ix_asset_info_meta_key_val_num", "asset_info_meta", ["key", "val_num"]) + op.create_index("ix_asset_info_meta_key_val_bool", "asset_info_meta", ["key", "val_bool"]) + + # Tags vocabulary + tags_table = sa.table( + "tags", + sa.column("name", sa.String(length=512)), + sa.column("tag_type", sa.String()), + ) + op.bulk_insert( + tags_table, + [ + {"name": "models", "tag_type": "system"}, + {"name": "input", "tag_type": "system"}, + {"name": "output", "tag_type": "system"}, + + {"name": "configs", "tag_type": "system"}, + {"name": "checkpoints", "tag_type": "system"}, + {"name": "loras", "tag_type": "system"}, + {"name": "vae", "tag_type": "system"}, + {"name": "text_encoders", "tag_type": "system"}, + {"name": "diffusion_models", "tag_type": "system"}, + {"name": "clip_vision", "tag_type": "system"}, + {"name": "style_models", "tag_type": "system"}, + {"name": "embeddings", "tag_type": "system"}, + {"name": "diffusers", "tag_type": "system"}, + {"name": "vae_approx", "tag_type": "system"}, + {"name": "controlnet", "tag_type": "system"}, + {"name": "gligen", "tag_type": "system"}, + {"name": "upscale_models", "tag_type": "system"}, + {"name": "hypernetworks", "tag_type": "system"}, + {"name": "photomaker", "tag_type": "system"}, + {"name": "classifiers", "tag_type": "system"}, + + {"name": "encoder", "tag_type": "system"}, + {"name": "decoder", "tag_type": "system"}, + + {"name": "missing", "tag_type": "system"}, + {"name": "rescan", "tag_type": "system"}, + ], + ) + + +def downgrade() -> None: + op.drop_index("ix_asset_info_meta_key_val_bool", table_name="asset_info_meta") + op.drop_index("ix_asset_info_meta_key_val_num", table_name="asset_info_meta") + op.drop_index("ix_asset_info_meta_key_val_str", table_name="asset_info_meta") + op.drop_index("ix_asset_info_meta_key", table_name="asset_info_meta") + op.drop_table("asset_info_meta") + + op.drop_index("ix_asset_cache_state_asset_id", table_name="asset_cache_state") + op.drop_index("ix_asset_cache_state_file_path", table_name="asset_cache_state") + op.drop_constraint("uq_asset_cache_state_file_path", table_name="asset_cache_state") + op.drop_table("asset_cache_state") + + op.drop_index("ix_asset_info_tags_asset_info_id", table_name="asset_info_tags") + op.drop_index("ix_asset_info_tags_tag_name", table_name="asset_info_tags") + op.drop_table("asset_info_tags") + + op.drop_index("ix_tags_tag_type", table_name="tags") + op.drop_table("tags") + + op.drop_constraint("uq_assets_info_asset_owner_name", table_name="assets_info") + op.drop_index("ix_assets_info_owner_name", table_name="assets_info") + op.drop_index("ix_assets_info_last_access_time", table_name="assets_info") + op.drop_index("ix_assets_info_created_at", table_name="assets_info") + op.drop_index("ix_assets_info_name", table_name="assets_info") + op.drop_index("ix_assets_info_asset_id", table_name="assets_info") + op.drop_index("ix_assets_info_owner_id", table_name="assets_info") + op.drop_table("assets_info") + + op.drop_index("uq_assets_hash", table_name="assets") + op.drop_index("ix_assets_mime_type", table_name="assets") + op.drop_table("assets") diff --git a/app/assets/api/routes.py b/app/assets/api/routes.py new file mode 100644 index 000000000..30e87a898 --- /dev/null +++ b/app/assets/api/routes.py @@ -0,0 +1,102 @@ +import logging +import uuid +from aiohttp import web + +from pydantic import ValidationError + +import app.assets.manager as manager +from app import user_manager +from app.assets.api import schemas_in +from app.assets.helpers import get_query_dict + +ROUTES = web.RouteTableDef() +USER_MANAGER: user_manager.UserManager | None = None + +# UUID regex (canonical hyphenated form, case-insensitive) +UUID_RE = r"[0-9a-fA-F]{8}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{12}" + +def register_assets_system(app: web.Application, user_manager_instance: user_manager.UserManager) -> None: + global USER_MANAGER + USER_MANAGER = user_manager_instance + app.add_routes(ROUTES) + +def _error_response(status: int, code: str, message: str, details: dict | None = None) -> web.Response: + return web.json_response({"error": {"code": code, "message": message, "details": details or {}}}, status=status) + + +def _validation_error_response(code: str, ve: ValidationError) -> web.Response: + return _error_response(400, code, "Validation failed.", {"errors": ve.json()}) + + +@ROUTES.get("/api/assets") +async def list_assets(request: web.Request) -> web.Response: + """ + GET request to list assets. + """ + query_dict = get_query_dict(request) + try: + q = schemas_in.ListAssetsQuery.model_validate(query_dict) + except ValidationError as ve: + return _validation_error_response("INVALID_QUERY", ve) + + payload = manager.list_assets( + include_tags=q.include_tags, + exclude_tags=q.exclude_tags, + name_contains=q.name_contains, + metadata_filter=q.metadata_filter, + limit=q.limit, + offset=q.offset, + sort=q.sort, + order=q.order, + owner_id=USER_MANAGER.get_request_user_id(request), + ) + return web.json_response(payload.model_dump(mode="json")) + + +@ROUTES.get(f"/api/assets/{{id:{UUID_RE}}}") +async def get_asset(request: web.Request) -> web.Response: + """ + GET request to get an asset's info as JSON. + """ + asset_info_id = str(uuid.UUID(request.match_info["id"])) + try: + result = manager.get_asset( + asset_info_id=asset_info_id, + owner_id=USER_MANAGER.get_request_user_id(request), + ) + except ValueError as e: + return _error_response(404, "ASSET_NOT_FOUND", str(e), {"id": asset_info_id}) + except Exception: + logging.exception( + "get_asset failed for asset_info_id=%s, owner_id=%s", + asset_info_id, + USER_MANAGER.get_request_user_id(request), + ) + return _error_response(500, "INTERNAL", "Unexpected server error.") + return web.json_response(result.model_dump(mode="json"), status=200) + + +@ROUTES.get("/api/tags") +async def get_tags(request: web.Request) -> web.Response: + """ + GET request to list all tags based on query parameters. + """ + query_map = dict(request.rel_url.query) + + try: + query = schemas_in.TagsListQuery.model_validate(query_map) + except ValidationError as e: + return web.json_response( + {"error": {"code": "INVALID_QUERY", "message": "Invalid query parameters", "details": e.errors()}}, + status=400, + ) + + result = manager.list_tags( + prefix=query.prefix, + limit=query.limit, + offset=query.offset, + order=query.order, + include_zero=query.include_zero, + owner_id=USER_MANAGER.get_request_user_id(request), + ) + return web.json_response(result.model_dump(mode="json")) diff --git a/app/assets/api/schemas_in.py b/app/assets/api/schemas_in.py new file mode 100644 index 000000000..200b41aef --- /dev/null +++ b/app/assets/api/schemas_in.py @@ -0,0 +1,94 @@ +import json +import uuid +from typing import Any, Literal + +from pydantic import ( + BaseModel, + ConfigDict, + Field, + conint, + field_validator, +) + + +class ListAssetsQuery(BaseModel): + include_tags: list[str] = Field(default_factory=list) + exclude_tags: list[str] = Field(default_factory=list) + name_contains: str | None = None + + # Accept either a JSON string (query param) or a dict + metadata_filter: dict[str, Any] | None = None + + limit: conint(ge=1, le=500) = 20 + offset: conint(ge=0) = 0 + + sort: Literal["name", "created_at", "updated_at", "size", "last_access_time"] = "created_at" + order: Literal["asc", "desc"] = "desc" + + @field_validator("include_tags", "exclude_tags", mode="before") + @classmethod + def _split_csv_tags(cls, v): + # Accept "a,b,c" or ["a","b"] (we are liberal in what we accept) + if v is None: + return [] + if isinstance(v, str): + return [t.strip() for t in v.split(",") if t.strip()] + if isinstance(v, list): + out: list[str] = [] + for item in v: + if isinstance(item, str): + out.extend([t.strip() for t in item.split(",") if t.strip()]) + return out + return v + + @field_validator("metadata_filter", mode="before") + @classmethod + def _parse_metadata_json(cls, v): + if v is None or isinstance(v, dict): + return v + if isinstance(v, str) and v.strip(): + try: + parsed = json.loads(v) + except Exception as e: + raise ValueError(f"metadata_filter must be JSON: {e}") from e + if not isinstance(parsed, dict): + raise ValueError("metadata_filter must be a JSON object") + return parsed + return None + + +class TagsListQuery(BaseModel): + model_config = ConfigDict(extra="ignore", str_strip_whitespace=True) + + prefix: str | None = Field(None, min_length=1, max_length=256) + limit: int = Field(100, ge=1, le=1000) + offset: int = Field(0, ge=0, le=10_000_000) + order: Literal["count_desc", "name_asc"] = "count_desc" + include_zero: bool = True + + @field_validator("prefix") + @classmethod + def normalize_prefix(cls, v: str | None) -> str | None: + if v is None: + return v + v = v.strip() + return v.lower() or None + + +class SetPreviewBody(BaseModel): + """Set or clear the preview for an AssetInfo. Provide an Asset.id or null.""" + preview_id: str | None = None + + @field_validator("preview_id", mode="before") + @classmethod + def _norm_uuid(cls, v): + if v is None: + return None + s = str(v).strip() + if not s: + return None + try: + uuid.UUID(s) + except Exception: + raise ValueError("preview_id must be a UUID") + return s diff --git a/app/assets/api/schemas_out.py b/app/assets/api/schemas_out.py new file mode 100644 index 000000000..9f8184f20 --- /dev/null +++ b/app/assets/api/schemas_out.py @@ -0,0 +1,60 @@ +from datetime import datetime +from typing import Any + +from pydantic import BaseModel, ConfigDict, Field, field_serializer + + +class AssetSummary(BaseModel): + id: str + name: str + asset_hash: str | None = None + size: int | None = None + mime_type: str | None = None + tags: list[str] = Field(default_factory=list) + preview_url: str | None = None + created_at: datetime | None = None + updated_at: datetime | None = None + last_access_time: datetime | None = None + + model_config = ConfigDict(from_attributes=True) + + @field_serializer("created_at", "updated_at", "last_access_time") + def _ser_dt(self, v: datetime | None, _info): + return v.isoformat() if v else None + + +class AssetsList(BaseModel): + assets: list[AssetSummary] + total: int + has_more: bool + + +class AssetDetail(BaseModel): + id: str + name: str + asset_hash: str | None = None + size: int | None = None + mime_type: str | None = None + tags: list[str] = Field(default_factory=list) + user_metadata: dict[str, Any] = Field(default_factory=dict) + preview_id: str | None = None + created_at: datetime | None = None + last_access_time: datetime | None = None + + model_config = ConfigDict(from_attributes=True) + + @field_serializer("created_at", "last_access_time") + def _ser_dt(self, v: datetime | None, _info): + return v.isoformat() if v else None + + +class TagUsage(BaseModel): + name: str + count: int + type: str + + +class TagsList(BaseModel): + tags: list[TagUsage] = Field(default_factory=list) + total: int + has_more: bool diff --git a/app/assets/database/bulk_ops.py b/app/assets/database/bulk_ops.py new file mode 100644 index 000000000..9352cd65d --- /dev/null +++ b/app/assets/database/bulk_ops.py @@ -0,0 +1,188 @@ +import os +import uuid +import sqlalchemy +from typing import Iterable +from sqlalchemy.orm import Session +from sqlalchemy.dialects import sqlite + +from app.assets.helpers import utcnow +from app.assets.database.models import Asset, AssetCacheState, AssetInfo, AssetInfoTag, AssetInfoMeta + +MAX_BIND_PARAMS = 800 + +def _chunk_rows(rows: list[dict], cols_per_row: int, max_bind_params: int) -> Iterable[list[dict]]: + if not rows: + return [] + rows_per_stmt = max(1, max_bind_params // max(1, cols_per_row)) + for i in range(0, len(rows), rows_per_stmt): + yield rows[i:i + rows_per_stmt] + +def _iter_chunks(seq, n: int): + for i in range(0, len(seq), n): + yield seq[i:i + n] + +def _rows_per_stmt(cols: int) -> int: + return max(1, MAX_BIND_PARAMS // max(1, cols)) + + +def seed_from_paths_batch( + session: Session, + *, + specs: list[dict], + owner_id: str = "", +) -> dict: + """Each spec is a dict with keys: + - abs_path: str + - size_bytes: int + - mtime_ns: int + - info_name: str + - tags: list[str] + - fname: Optional[str] + """ + if not specs: + return {"inserted_infos": 0, "won_states": 0, "lost_states": 0} + + now = utcnow() + asset_rows: list[dict] = [] + state_rows: list[dict] = [] + path_to_asset: dict[str, str] = {} + asset_to_info: dict[str, dict] = {} # asset_id -> prepared info row + path_list: list[str] = [] + + for sp in specs: + ap = os.path.abspath(sp["abs_path"]) + aid = str(uuid.uuid4()) + iid = str(uuid.uuid4()) + path_list.append(ap) + path_to_asset[ap] = aid + + asset_rows.append( + { + "id": aid, + "hash": None, + "size_bytes": sp["size_bytes"], + "mime_type": None, + "created_at": now, + } + ) + state_rows.append( + { + "asset_id": aid, + "file_path": ap, + "mtime_ns": sp["mtime_ns"], + } + ) + asset_to_info[aid] = { + "id": iid, + "owner_id": owner_id, + "name": sp["info_name"], + "asset_id": aid, + "preview_id": None, + "user_metadata": {"filename": sp["fname"]} if sp["fname"] else None, + "created_at": now, + "updated_at": now, + "last_access_time": now, + "_tags": sp["tags"], + "_filename": sp["fname"], + } + + # insert all seed Assets (hash=NULL) + ins_asset = sqlite.insert(Asset) + for chunk in _iter_chunks(asset_rows, _rows_per_stmt(5)): + session.execute(ins_asset, chunk) + + # try to claim AssetCacheState (file_path) + winners_by_path: set[str] = set() + ins_state = ( + sqlite.insert(AssetCacheState) + .on_conflict_do_nothing(index_elements=[AssetCacheState.file_path]) + .returning(AssetCacheState.file_path) + ) + for chunk in _iter_chunks(state_rows, _rows_per_stmt(3)): + winners_by_path.update((session.execute(ins_state, chunk)).scalars().all()) + + all_paths_set = set(path_list) + losers_by_path = all_paths_set - winners_by_path + lost_assets = [path_to_asset[p] for p in losers_by_path] + if lost_assets: # losers get their Asset removed + for id_chunk in _iter_chunks(lost_assets, MAX_BIND_PARAMS): + session.execute(sqlalchemy.delete(Asset).where(Asset.id.in_(id_chunk))) + + if not winners_by_path: + return {"inserted_infos": 0, "won_states": 0, "lost_states": len(losers_by_path)} + + # insert AssetInfo only for winners + winner_info_rows = [asset_to_info[path_to_asset[p]] for p in winners_by_path] + ins_info = ( + sqlite.insert(AssetInfo) + .on_conflict_do_nothing(index_elements=[AssetInfo.asset_id, AssetInfo.owner_id, AssetInfo.name]) + .returning(AssetInfo.id) + ) + + inserted_info_ids: set[str] = set() + for chunk in _iter_chunks(winner_info_rows, _rows_per_stmt(9)): + inserted_info_ids.update((session.execute(ins_info, chunk)).scalars().all()) + + # build and insert tag + meta rows for the AssetInfo + tag_rows: list[dict] = [] + meta_rows: list[dict] = [] + if inserted_info_ids: + for row in winner_info_rows: + iid = row["id"] + if iid not in inserted_info_ids: + continue + for t in row["_tags"]: + tag_rows.append({ + "asset_info_id": iid, + "tag_name": t, + "origin": "automatic", + "added_at": now, + }) + if row["_filename"]: + meta_rows.append( + { + "asset_info_id": iid, + "key": "filename", + "ordinal": 0, + "val_str": row["_filename"], + "val_num": None, + "val_bool": None, + "val_json": None, + } + ) + + bulk_insert_tags_and_meta(session, tag_rows=tag_rows, meta_rows=meta_rows, max_bind_params=MAX_BIND_PARAMS) + return { + "inserted_infos": len(inserted_info_ids), + "won_states": len(winners_by_path), + "lost_states": len(losers_by_path), + } + + +def bulk_insert_tags_and_meta( + session: Session, + *, + tag_rows: list[dict], + meta_rows: list[dict], + max_bind_params: int, +) -> None: + """Batch insert into asset_info_tags and asset_info_meta with ON CONFLICT DO NOTHING. + - tag_rows keys: asset_info_id, tag_name, origin, added_at + - meta_rows keys: asset_info_id, key, ordinal, val_str, val_num, val_bool, val_json + """ + if tag_rows: + ins_links = ( + sqlite.insert(AssetInfoTag) + .on_conflict_do_nothing(index_elements=[AssetInfoTag.asset_info_id, AssetInfoTag.tag_name]) + ) + for chunk in _chunk_rows(tag_rows, cols_per_row=4, max_bind_params=max_bind_params): + session.execute(ins_links, chunk) + if meta_rows: + ins_meta = ( + sqlite.insert(AssetInfoMeta) + .on_conflict_do_nothing( + index_elements=[AssetInfoMeta.asset_info_id, AssetInfoMeta.key, AssetInfoMeta.ordinal] + ) + ) + for chunk in _chunk_rows(meta_rows, cols_per_row=7, max_bind_params=max_bind_params): + session.execute(ins_meta, chunk) diff --git a/app/assets/database/models.py b/app/assets/database/models.py new file mode 100644 index 000000000..3cd28f68b --- /dev/null +++ b/app/assets/database/models.py @@ -0,0 +1,233 @@ +from __future__ import annotations + +import uuid +from datetime import datetime + +from typing import Any +from sqlalchemy import ( + JSON, + BigInteger, + Boolean, + CheckConstraint, + DateTime, + ForeignKey, + Index, + Integer, + Numeric, + String, + Text, + UniqueConstraint, +) +from sqlalchemy.orm import Mapped, foreign, mapped_column, relationship + +from app.assets.helpers import utcnow +from app.database.models import to_dict, Base + + +class Asset(Base): + __tablename__ = "assets" + + id: Mapped[str] = mapped_column(String(36), primary_key=True, default=lambda: str(uuid.uuid4())) + hash: Mapped[str | None] = mapped_column(String(256), nullable=True) + size_bytes: Mapped[int] = mapped_column(BigInteger, nullable=False, default=0) + mime_type: Mapped[str | None] = mapped_column(String(255)) + created_at: Mapped[datetime] = mapped_column( + DateTime(timezone=False), nullable=False, default=utcnow + ) + + infos: Mapped[list[AssetInfo]] = relationship( + "AssetInfo", + back_populates="asset", + primaryjoin=lambda: Asset.id == foreign(AssetInfo.asset_id), + foreign_keys=lambda: [AssetInfo.asset_id], + cascade="all,delete-orphan", + passive_deletes=True, + ) + + preview_of: Mapped[list[AssetInfo]] = relationship( + "AssetInfo", + back_populates="preview_asset", + primaryjoin=lambda: Asset.id == foreign(AssetInfo.preview_id), + foreign_keys=lambda: [AssetInfo.preview_id], + viewonly=True, + ) + + cache_states: Mapped[list[AssetCacheState]] = relationship( + back_populates="asset", + cascade="all, delete-orphan", + passive_deletes=True, + ) + + __table_args__ = ( + Index("uq_assets_hash", "hash", unique=True), + Index("ix_assets_mime_type", "mime_type"), + CheckConstraint("size_bytes >= 0", name="ck_assets_size_nonneg"), + ) + + def to_dict(self, include_none: bool = False) -> dict[str, Any]: + return to_dict(self, include_none=include_none) + + def __repr__(self) -> str: + return f"" + + +class AssetCacheState(Base): + __tablename__ = "asset_cache_state" + + id: Mapped[int] = mapped_column(Integer, primary_key=True, autoincrement=True) + asset_id: Mapped[str] = mapped_column(String(36), ForeignKey("assets.id", ondelete="CASCADE"), nullable=False) + file_path: Mapped[str] = mapped_column(Text, nullable=False) + mtime_ns: Mapped[int | None] = mapped_column(BigInteger, nullable=True) + needs_verify: Mapped[bool] = mapped_column(Boolean, nullable=False, default=False) + + asset: Mapped[Asset] = relationship(back_populates="cache_states") + + __table_args__ = ( + Index("ix_asset_cache_state_file_path", "file_path"), + Index("ix_asset_cache_state_asset_id", "asset_id"), + CheckConstraint("(mtime_ns IS NULL) OR (mtime_ns >= 0)", name="ck_acs_mtime_nonneg"), + UniqueConstraint("file_path", name="uq_asset_cache_state_file_path"), + ) + + def to_dict(self, include_none: bool = False) -> dict[str, Any]: + return to_dict(self, include_none=include_none) + + def __repr__(self) -> str: + return f"" + + +class AssetInfo(Base): + __tablename__ = "assets_info" + + id: Mapped[str] = mapped_column(String(36), primary_key=True, default=lambda: str(uuid.uuid4())) + owner_id: Mapped[str] = mapped_column(String(128), nullable=False, default="") + name: Mapped[str] = mapped_column(String(512), nullable=False) + asset_id: Mapped[str] = mapped_column(String(36), ForeignKey("assets.id", ondelete="RESTRICT"), nullable=False) + preview_id: Mapped[str | None] = mapped_column(String(36), ForeignKey("assets.id", ondelete="SET NULL")) + user_metadata: Mapped[dict[str, Any] | None] = mapped_column(JSON(none_as_null=True)) + created_at: Mapped[datetime] = mapped_column(DateTime(timezone=False), nullable=False, default=utcnow) + updated_at: Mapped[datetime] = mapped_column(DateTime(timezone=False), nullable=False, default=utcnow) + last_access_time: Mapped[datetime] = mapped_column(DateTime(timezone=False), nullable=False, default=utcnow) + + asset: Mapped[Asset] = relationship( + "Asset", + back_populates="infos", + foreign_keys=[asset_id], + lazy="selectin", + ) + preview_asset: Mapped[Asset | None] = relationship( + "Asset", + back_populates="preview_of", + foreign_keys=[preview_id], + ) + + metadata_entries: Mapped[list[AssetInfoMeta]] = relationship( + back_populates="asset_info", + cascade="all,delete-orphan", + passive_deletes=True, + ) + + tag_links: Mapped[list[AssetInfoTag]] = relationship( + back_populates="asset_info", + cascade="all,delete-orphan", + passive_deletes=True, + overlaps="tags,asset_infos", + ) + + tags: Mapped[list[Tag]] = relationship( + secondary="asset_info_tags", + back_populates="asset_infos", + lazy="selectin", + viewonly=True, + overlaps="tag_links,asset_info_links,asset_infos,tag", + ) + + __table_args__ = ( + UniqueConstraint("asset_id", "owner_id", "name", name="uq_assets_info_asset_owner_name"), + Index("ix_assets_info_owner_name", "owner_id", "name"), + Index("ix_assets_info_owner_id", "owner_id"), + Index("ix_assets_info_asset_id", "asset_id"), + Index("ix_assets_info_name", "name"), + Index("ix_assets_info_created_at", "created_at"), + Index("ix_assets_info_last_access_time", "last_access_time"), + ) + + def to_dict(self, include_none: bool = False) -> dict[str, Any]: + data = to_dict(self, include_none=include_none) + data["tags"] = [t.name for t in self.tags] + return data + + def __repr__(self) -> str: + return f"" + + +class AssetInfoMeta(Base): + __tablename__ = "asset_info_meta" + + asset_info_id: Mapped[str] = mapped_column( + String(36), ForeignKey("assets_info.id", ondelete="CASCADE"), primary_key=True + ) + key: Mapped[str] = mapped_column(String(256), primary_key=True) + ordinal: Mapped[int] = mapped_column(Integer, primary_key=True, default=0) + + val_str: Mapped[str | None] = mapped_column(String(2048), nullable=True) + val_num: Mapped[float | None] = mapped_column(Numeric(38, 10), nullable=True) + val_bool: Mapped[bool | None] = mapped_column(Boolean, nullable=True) + val_json: Mapped[Any | None] = mapped_column(JSON(none_as_null=True), nullable=True) + + asset_info: Mapped[AssetInfo] = relationship(back_populates="metadata_entries") + + __table_args__ = ( + Index("ix_asset_info_meta_key", "key"), + Index("ix_asset_info_meta_key_val_str", "key", "val_str"), + Index("ix_asset_info_meta_key_val_num", "key", "val_num"), + Index("ix_asset_info_meta_key_val_bool", "key", "val_bool"), + ) + + +class AssetInfoTag(Base): + __tablename__ = "asset_info_tags" + + asset_info_id: Mapped[str] = mapped_column( + String(36), ForeignKey("assets_info.id", ondelete="CASCADE"), primary_key=True + ) + tag_name: Mapped[str] = mapped_column( + String(512), ForeignKey("tags.name", ondelete="RESTRICT"), primary_key=True + ) + origin: Mapped[str] = mapped_column(String(32), nullable=False, default="manual") + added_at: Mapped[datetime] = mapped_column( + DateTime(timezone=False), nullable=False, default=utcnow + ) + + asset_info: Mapped[AssetInfo] = relationship(back_populates="tag_links") + tag: Mapped[Tag] = relationship(back_populates="asset_info_links") + + __table_args__ = ( + Index("ix_asset_info_tags_tag_name", "tag_name"), + Index("ix_asset_info_tags_asset_info_id", "asset_info_id"), + ) + + +class Tag(Base): + __tablename__ = "tags" + + name: Mapped[str] = mapped_column(String(512), primary_key=True) + tag_type: Mapped[str] = mapped_column(String(32), nullable=False, default="user") + + asset_info_links: Mapped[list[AssetInfoTag]] = relationship( + back_populates="tag", + overlaps="asset_infos,tags", + ) + asset_infos: Mapped[list[AssetInfo]] = relationship( + secondary="asset_info_tags", + back_populates="tags", + viewonly=True, + overlaps="asset_info_links,tag_links,tags,asset_info", + ) + + __table_args__ = ( + Index("ix_tags_tag_type", "tag_type"), + ) + + def __repr__(self) -> str: + return f"" diff --git a/app/assets/database/queries.py b/app/assets/database/queries.py new file mode 100644 index 000000000..0824c0c2f --- /dev/null +++ b/app/assets/database/queries.py @@ -0,0 +1,267 @@ +import sqlalchemy as sa +from collections import defaultdict +from sqlalchemy import select, exists, func +from sqlalchemy.orm import Session, contains_eager, noload +from app.assets.database.models import Asset, AssetInfo, AssetInfoMeta, AssetInfoTag, Tag +from app.assets.helpers import escape_like_prefix, normalize_tags +from typing import Sequence + + +def visible_owner_clause(owner_id: str) -> sa.sql.ClauseElement: + """Build owner visibility predicate for reads. Owner-less rows are visible to everyone.""" + owner_id = (owner_id or "").strip() + if owner_id == "": + return AssetInfo.owner_id == "" + return AssetInfo.owner_id.in_(["", owner_id]) + + +def apply_tag_filters( + stmt: sa.sql.Select, + include_tags: Sequence[str] | None = None, + exclude_tags: Sequence[str] | None = None, +) -> sa.sql.Select: + """include_tags: every tag must be present; exclude_tags: none may be present.""" + include_tags = normalize_tags(include_tags) + exclude_tags = normalize_tags(exclude_tags) + + if include_tags: + for tag_name in include_tags: + stmt = stmt.where( + exists().where( + (AssetInfoTag.asset_info_id == AssetInfo.id) + & (AssetInfoTag.tag_name == tag_name) + ) + ) + + if exclude_tags: + stmt = stmt.where( + ~exists().where( + (AssetInfoTag.asset_info_id == AssetInfo.id) + & (AssetInfoTag.tag_name.in_(exclude_tags)) + ) + ) + return stmt + +def apply_metadata_filter( + stmt: sa.sql.Select, + metadata_filter: dict | None = None, +) -> sa.sql.Select: + """Apply filters using asset_info_meta projection table.""" + if not metadata_filter: + return stmt + + def _exists_for_pred(key: str, *preds) -> sa.sql.ClauseElement: + return sa.exists().where( + AssetInfoMeta.asset_info_id == AssetInfo.id, + AssetInfoMeta.key == key, + *preds, + ) + + def _exists_clause_for_value(key: str, value) -> sa.sql.ClauseElement: + if value is None: + no_row_for_key = sa.not_( + sa.exists().where( + AssetInfoMeta.asset_info_id == AssetInfo.id, + AssetInfoMeta.key == key, + ) + ) + null_row = _exists_for_pred( + key, + AssetInfoMeta.val_json.is_(None), + AssetInfoMeta.val_str.is_(None), + AssetInfoMeta.val_num.is_(None), + AssetInfoMeta.val_bool.is_(None), + ) + return sa.or_(no_row_for_key, null_row) + + if isinstance(value, bool): + return _exists_for_pred(key, AssetInfoMeta.val_bool == bool(value)) + if isinstance(value, (int, float)): + from decimal import Decimal + num = value if isinstance(value, Decimal) else Decimal(str(value)) + return _exists_for_pred(key, AssetInfoMeta.val_num == num) + if isinstance(value, str): + return _exists_for_pred(key, AssetInfoMeta.val_str == value) + return _exists_for_pred(key, AssetInfoMeta.val_json == value) + + for k, v in metadata_filter.items(): + if isinstance(v, list): + ors = [_exists_clause_for_value(k, elem) for elem in v] + if ors: + stmt = stmt.where(sa.or_(*ors)) + else: + stmt = stmt.where(_exists_clause_for_value(k, v)) + return stmt + + +def asset_exists_by_hash(session: Session, asset_hash: str) -> bool: + """ + Check if an asset with a given hash exists in database. + """ + row = ( + session.execute( + select(sa.literal(True)).select_from(Asset).where(Asset.hash == asset_hash).limit(1) + ) + ).first() + return row is not None + +def get_asset_info_by_id(session: Session, asset_info_id: str) -> AssetInfo | None: + return session.get(AssetInfo, asset_info_id) + +def list_asset_infos_page( + session: Session, + owner_id: str = "", + include_tags: Sequence[str] | None = None, + exclude_tags: Sequence[str] | None = None, + name_contains: str | None = None, + metadata_filter: dict | None = None, + limit: int = 20, + offset: int = 0, + sort: str = "created_at", + order: str = "desc", +) -> tuple[list[AssetInfo], dict[str, list[str]], int]: + base = ( + select(AssetInfo) + .join(Asset, Asset.id == AssetInfo.asset_id) + .options(contains_eager(AssetInfo.asset), noload(AssetInfo.tags)) + .where(visible_owner_clause(owner_id)) + ) + + if name_contains: + escaped, esc = escape_like_prefix(name_contains) + base = base.where(AssetInfo.name.ilike(f"%{escaped}%", escape=esc)) + + base = apply_tag_filters(base, include_tags, exclude_tags) + base = apply_metadata_filter(base, metadata_filter) + + sort = (sort or "created_at").lower() + order = (order or "desc").lower() + sort_map = { + "name": AssetInfo.name, + "created_at": AssetInfo.created_at, + "updated_at": AssetInfo.updated_at, + "last_access_time": AssetInfo.last_access_time, + "size": Asset.size_bytes, + } + sort_col = sort_map.get(sort, AssetInfo.created_at) + sort_exp = sort_col.desc() if order == "desc" else sort_col.asc() + + base = base.order_by(sort_exp).limit(limit).offset(offset) + + count_stmt = ( + select(sa.func.count()) + .select_from(AssetInfo) + .join(Asset, Asset.id == AssetInfo.asset_id) + .where(visible_owner_clause(owner_id)) + ) + if name_contains: + escaped, esc = escape_like_prefix(name_contains) + count_stmt = count_stmt.where(AssetInfo.name.ilike(f"%{escaped}%", escape=esc)) + count_stmt = apply_tag_filters(count_stmt, include_tags, exclude_tags) + count_stmt = apply_metadata_filter(count_stmt, metadata_filter) + + total = int((session.execute(count_stmt)).scalar_one() or 0) + + infos = (session.execute(base)).unique().scalars().all() + + id_list: list[str] = [i.id for i in infos] + tag_map: dict[str, list[str]] = defaultdict(list) + if id_list: + rows = session.execute( + select(AssetInfoTag.asset_info_id, Tag.name) + .join(Tag, Tag.name == AssetInfoTag.tag_name) + .where(AssetInfoTag.asset_info_id.in_(id_list)) + ) + for aid, tag_name in rows.all(): + tag_map[aid].append(tag_name) + + return infos, tag_map, total + +def fetch_asset_info_asset_and_tags( + session: Session, + asset_info_id: str, + owner_id: str = "", +) -> tuple[AssetInfo, Asset, list[str]] | None: + stmt = ( + select(AssetInfo, Asset, Tag.name) + .join(Asset, Asset.id == AssetInfo.asset_id) + .join(AssetInfoTag, AssetInfoTag.asset_info_id == AssetInfo.id, isouter=True) + .join(Tag, Tag.name == AssetInfoTag.tag_name, isouter=True) + .where( + AssetInfo.id == asset_info_id, + visible_owner_clause(owner_id), + ) + .options(noload(AssetInfo.tags)) + .order_by(Tag.name.asc()) + ) + + rows = (session.execute(stmt)).all() + if not rows: + return None + + first_info, first_asset, _ = rows[0] + tags: list[str] = [] + seen: set[str] = set() + for _info, _asset, tag_name in rows: + if tag_name and tag_name not in seen: + seen.add(tag_name) + tags.append(tag_name) + return first_info, first_asset, tags + +def list_tags_with_usage( + session: Session, + prefix: str | None = None, + limit: int = 100, + offset: int = 0, + include_zero: bool = True, + order: str = "count_desc", + owner_id: str = "", +) -> tuple[list[tuple[str, str, int]], int]: + counts_sq = ( + select( + AssetInfoTag.tag_name.label("tag_name"), + func.count(AssetInfoTag.asset_info_id).label("cnt"), + ) + .select_from(AssetInfoTag) + .join(AssetInfo, AssetInfo.id == AssetInfoTag.asset_info_id) + .where(visible_owner_clause(owner_id)) + .group_by(AssetInfoTag.tag_name) + .subquery() + ) + + q = ( + select( + Tag.name, + Tag.tag_type, + func.coalesce(counts_sq.c.cnt, 0).label("count"), + ) + .select_from(Tag) + .join(counts_sq, counts_sq.c.tag_name == Tag.name, isouter=True) + ) + + if prefix: + escaped, esc = escape_like_prefix(prefix.strip().lower()) + q = q.where(Tag.name.like(escaped + "%", escape=esc)) + + if not include_zero: + q = q.where(func.coalesce(counts_sq.c.cnt, 0) > 0) + + if order == "name_asc": + q = q.order_by(Tag.name.asc()) + else: + q = q.order_by(func.coalesce(counts_sq.c.cnt, 0).desc(), Tag.name.asc()) + + total_q = select(func.count()).select_from(Tag) + if prefix: + escaped, esc = escape_like_prefix(prefix.strip().lower()) + total_q = total_q.where(Tag.name.like(escaped + "%", escape=esc)) + if not include_zero: + total_q = total_q.where( + Tag.name.in_(select(AssetInfoTag.tag_name).group_by(AssetInfoTag.tag_name)) + ) + + rows = (session.execute(q.limit(limit).offset(offset))).all() + total = (session.execute(total_q)).scalar_one() + + rows_norm = [(name, ttype, int(count or 0)) for (name, ttype, count) in rows] + return rows_norm, int(total or 0) diff --git a/app/assets/database/tags.py b/app/assets/database/tags.py new file mode 100644 index 000000000..3ab6497c2 --- /dev/null +++ b/app/assets/database/tags.py @@ -0,0 +1,62 @@ +from typing import Iterable + +import sqlalchemy +from sqlalchemy.orm import Session +from sqlalchemy.dialects import sqlite + +from app.assets.helpers import normalize_tags, utcnow +from app.assets.database.models import Tag, AssetInfoTag, AssetInfo + + +def ensure_tags_exist(session: Session, names: Iterable[str], tag_type: str = "user") -> None: + wanted = normalize_tags(list(names)) + if not wanted: + return + rows = [{"name": n, "tag_type": tag_type} for n in list(dict.fromkeys(wanted))] + ins = ( + sqlite.insert(Tag) + .values(rows) + .on_conflict_do_nothing(index_elements=[Tag.name]) + ) + return session.execute(ins) + +def add_missing_tag_for_asset_id( + session: Session, + *, + asset_id: str, + origin: str = "automatic", +) -> None: + select_rows = ( + sqlalchemy.select( + AssetInfo.id.label("asset_info_id"), + sqlalchemy.literal("missing").label("tag_name"), + sqlalchemy.literal(origin).label("origin"), + sqlalchemy.literal(utcnow()).label("added_at"), + ) + .where(AssetInfo.asset_id == asset_id) + .where( + sqlalchemy.not_( + sqlalchemy.exists().where((AssetInfoTag.asset_info_id == AssetInfo.id) & (AssetInfoTag.tag_name == "missing")) + ) + ) + ) + session.execute( + sqlite.insert(AssetInfoTag) + .from_select( + ["asset_info_id", "tag_name", "origin", "added_at"], + select_rows, + ) + .on_conflict_do_nothing(index_elements=[AssetInfoTag.asset_info_id, AssetInfoTag.tag_name]) + ) + +def remove_missing_tag_for_asset_id( + session: Session, + *, + asset_id: str, +) -> None: + session.execute( + sqlalchemy.delete(AssetInfoTag).where( + AssetInfoTag.asset_info_id.in_(sqlalchemy.select(AssetInfo.id).where(AssetInfo.asset_id == asset_id)), + AssetInfoTag.tag_name == "missing", + ) + ) diff --git a/app/assets/hashing.py b/app/assets/hashing.py new file mode 100644 index 000000000..4b72084b9 --- /dev/null +++ b/app/assets/hashing.py @@ -0,0 +1,75 @@ +from blake3 import blake3 +from typing import IO +import os +import asyncio + + +DEFAULT_CHUNK = 8 * 1024 *1024 # 8MB + +# NOTE: this allows hashing different representations of a file-like object +def blake3_hash( + fp: str | IO[bytes], + chunk_size: int = DEFAULT_CHUNK, +) -> str: + """ + Returns a BLAKE3 hex digest for ``fp``, which may be: + - a filename (str/bytes) or PathLike + - an open binary file object + If ``fp`` is a file object, it must be opened in **binary** mode and support + ``read``, ``seek``, and ``tell``. The function will seek to the start before + reading and will attempt to restore the original position afterward. + """ + # duck typing to check if input is a file-like object + if hasattr(fp, "read"): + return _hash_file_obj(fp, chunk_size) + + with open(os.fspath(fp), "rb") as f: + return _hash_file_obj(f, chunk_size) + + +async def blake3_hash_async( + fp: str | IO[bytes], + chunk_size: int = DEFAULT_CHUNK, +) -> str: + """Async wrapper for ``blake3_hash_sync``. + Uses a worker thread so the event loop remains responsive. + """ + # If it is a path, open inside the worker thread to keep I/O off the loop. + if hasattr(fp, "read"): + return await asyncio.to_thread(blake3_hash, fp, chunk_size) + + def _worker() -> str: + with open(os.fspath(fp), "rb") as f: + return _hash_file_obj(f, chunk_size) + + return await asyncio.to_thread(_worker) + + +def _hash_file_obj(file_obj: IO, chunk_size: int = DEFAULT_CHUNK) -> str: + """ + Hash an already-open binary file object by streaming in chunks. + - Seeks to the beginning before reading (if supported). + - Restores the original position afterward (if tell/seek are supported). + """ + if chunk_size <= 0: + chunk_size = DEFAULT_CHUNK + + # in case file object is already open and not at the beginning, track so can be restored after hashing + orig_pos = file_obj.tell() + + try: + # seek to the beginning before reading + if orig_pos != 0: + file_obj.seek(0) + + h = blake3() + while True: + chunk = file_obj.read(chunk_size) + if not chunk: + break + h.update(chunk) + return h.hexdigest() + finally: + # restore original position in file object, if needed + if orig_pos != 0: + file_obj.seek(orig_pos) diff --git a/app/assets/helpers.py b/app/assets/helpers.py new file mode 100644 index 000000000..6755d0e56 --- /dev/null +++ b/app/assets/helpers.py @@ -0,0 +1,216 @@ +import contextlib +import os +from aiohttp import web +from datetime import datetime, timezone +from pathlib import Path +from typing import Literal, Any + +import folder_paths + + +RootType = Literal["models", "input", "output"] +ALLOWED_ROOTS: tuple[RootType, ...] = ("models", "input", "output") + +def get_query_dict(request: web.Request) -> dict[str, Any]: + """ + Gets a dictionary of query parameters from the request. + + 'request.query' is a MultiMapping[str], needs to be converted to a dictionary to be validated by Pydantic. + """ + query_dict = { + key: request.query.getall(key) if len(request.query.getall(key)) > 1 else request.query.get(key) + for key in request.query.keys() + } + return query_dict + +def list_tree(base_dir: str) -> list[str]: + out: list[str] = [] + base_abs = os.path.abspath(base_dir) + if not os.path.isdir(base_abs): + return out + for dirpath, _subdirs, filenames in os.walk(base_abs, topdown=True, followlinks=False): + for name in filenames: + out.append(os.path.abspath(os.path.join(dirpath, name))) + return out + +def prefixes_for_root(root: RootType) -> list[str]: + if root == "models": + bases: list[str] = [] + for _bucket, paths in get_comfy_models_folders(): + bases.extend(paths) + return [os.path.abspath(p) for p in bases] + if root == "input": + return [os.path.abspath(folder_paths.get_input_directory())] + if root == "output": + return [os.path.abspath(folder_paths.get_output_directory())] + return [] + +def escape_like_prefix(s: str, escape: str = "!") -> tuple[str, str]: + """Escapes %, _ and the escape char itself in a LIKE prefix. + Returns (escaped_prefix, escape_char). Caller should append '%' and pass escape=escape_char to .like(). + """ + s = s.replace(escape, escape + escape) # escape the escape char first + s = s.replace("%", escape + "%").replace("_", escape + "_") # escape LIKE wildcards + return s, escape + +def fast_asset_file_check( + *, + mtime_db: int | None, + size_db: int | None, + stat_result: os.stat_result, +) -> bool: + if mtime_db is None: + return False + actual_mtime_ns = getattr(stat_result, "st_mtime_ns", int(stat_result.st_mtime * 1_000_000_000)) + if int(mtime_db) != int(actual_mtime_ns): + return False + sz = int(size_db or 0) + if sz > 0: + return int(stat_result.st_size) == sz + return True + +def utcnow() -> datetime: + """Naive UTC timestamp (no tzinfo). We always treat DB datetimes as UTC.""" + return datetime.now(timezone.utc).replace(tzinfo=None) + +def get_comfy_models_folders() -> list[tuple[str, list[str]]]: + """Build a list of (folder_name, base_paths[]) categories that are configured for model locations. + + We trust `folder_paths.folder_names_and_paths` and include a category if + *any* of its base paths lies under the Comfy `models_dir`. + """ + targets: list[tuple[str, list[str]]] = [] + models_root = os.path.abspath(folder_paths.models_dir) + for name, (paths, _exts) in folder_paths.folder_names_and_paths.items(): + if any(os.path.abspath(p).startswith(models_root + os.sep) for p in paths): + targets.append((name, paths)) + return targets + +def compute_relative_filename(file_path: str) -> str | None: + """ + Return the model's path relative to the last well-known folder (the model category), + using forward slashes, eg: + /.../models/checkpoints/flux/123/flux.safetensors -> "flux/123/flux.safetensors" + /.../models/text_encoders/clip_g.safetensors -> "clip_g.safetensors" + + For non-model paths, returns None. + NOTE: this is a temporary helper, used only for initializing metadata["filename"] field. + """ + try: + root_category, rel_path = get_relative_to_root_category_path_of_asset(file_path) + except ValueError: + return None + + p = Path(rel_path) + parts = [seg for seg in p.parts if seg not in (".", "..", p.anchor)] + if not parts: + return None + + if root_category == "models": + # parts[0] is the category ("checkpoints", "vae", etc) – drop it + inside = parts[1:] if len(parts) > 1 else [parts[0]] + return "/".join(inside) + return "/".join(parts) # input/output: keep all parts + + +def get_relative_to_root_category_path_of_asset(file_path: str) -> tuple[Literal["input", "output", "models"], str]: + """Given an absolute or relative file path, determine which root category the path belongs to: + - 'input' if the file resides under `folder_paths.get_input_directory()` + - 'output' if the file resides under `folder_paths.get_output_directory()` + - 'models' if the file resides under any base path of categories returned by `get_comfy_models_folders()` + + Returns: + (root_category, relative_path_inside_that_root) + For 'models', the relative path is prefixed with the category name: + e.g. ('models', 'vae/test/sub/ae.safetensors') + + Raises: + ValueError: if the path does not belong to input, output, or configured model bases. + """ + fp_abs = os.path.abspath(file_path) + + def _is_within(child: str, parent: str) -> bool: + try: + return os.path.commonpath([child, parent]) == parent + except Exception: + return False + + def _rel(child: str, parent: str) -> str: + return os.path.relpath(os.path.join(os.sep, os.path.relpath(child, parent)), os.sep) + + # 1) input + input_base = os.path.abspath(folder_paths.get_input_directory()) + if _is_within(fp_abs, input_base): + return "input", _rel(fp_abs, input_base) + + # 2) output + output_base = os.path.abspath(folder_paths.get_output_directory()) + if _is_within(fp_abs, output_base): + return "output", _rel(fp_abs, output_base) + + # 3) models (check deepest matching base to avoid ambiguity) + best: tuple[int, str, str] | None = None # (base_len, bucket, rel_inside_bucket) + for bucket, bases in get_comfy_models_folders(): + for b in bases: + base_abs = os.path.abspath(b) + if not _is_within(fp_abs, base_abs): + continue + cand = (len(base_abs), bucket, _rel(fp_abs, base_abs)) + if best is None or cand[0] > best[0]: + best = cand + + if best is not None: + _, bucket, rel_inside = best + combined = os.path.join(bucket, rel_inside) + return "models", os.path.relpath(os.path.join(os.sep, combined), os.sep) + + raise ValueError(f"Path is not within input, output, or configured model bases: {file_path}") + +def get_name_and_tags_from_asset_path(file_path: str) -> tuple[str, list[str]]: + """Return a tuple (name, tags) derived from a filesystem path. + + Semantics: + - Root category is determined by `get_relative_to_root_category_path_of_asset`. + - The returned `name` is the base filename with extension from the relative path. + - The returned `tags` are: + [root_category] + parent folders of the relative path (in order) + For 'models', this means: + file '/.../ModelsDir/vae/test_tag/ae.safetensors' + -> root_category='models', some_path='vae/test_tag/ae.safetensors' + -> name='ae.safetensors', tags=['models', 'vae', 'test_tag'] + + Raises: + ValueError: if the path does not belong to input, output, or configured model bases. + """ + root_category, some_path = get_relative_to_root_category_path_of_asset(file_path) + p = Path(some_path) + parent_parts = [part for part in p.parent.parts if part not in (".", "..", p.anchor)] + return p.name, list(dict.fromkeys(normalize_tags([root_category, *parent_parts]))) + +def normalize_tags(tags: list[str] | None) -> list[str]: + """ + Normalize a list of tags by: + - Stripping whitespace and converting to lowercase. + - Removing duplicates. + """ + return [t.strip().lower() for t in (tags or []) if (t or "").strip()] + +def collect_models_files() -> list[str]: + out: list[str] = [] + for folder_name, bases in get_comfy_models_folders(): + rel_files = folder_paths.get_filename_list(folder_name) or [] + for rel_path in rel_files: + abs_path = folder_paths.get_full_path(folder_name, rel_path) + if not abs_path: + continue + abs_path = os.path.abspath(abs_path) + allowed = False + for b in bases: + base_abs = os.path.abspath(b) + with contextlib.suppress(Exception): + if os.path.commonpath([abs_path, base_abs]) == base_abs: + allowed = True + break + if allowed: + out.append(abs_path) + return out diff --git a/app/assets/manager.py b/app/assets/manager.py new file mode 100644 index 000000000..6425e7aa2 --- /dev/null +++ b/app/assets/manager.py @@ -0,0 +1,123 @@ +from typing import Sequence + +from app.database.db import create_session +from app.assets.api import schemas_out +from app.assets.database.queries import ( + asset_exists_by_hash, + fetch_asset_info_asset_and_tags, + list_asset_infos_page, + list_tags_with_usage, +) + + +def _safe_sort_field(requested: str | None) -> str: + if not requested: + return "created_at" + v = requested.lower() + if v in {"name", "created_at", "updated_at", "size", "last_access_time"}: + return v + return "created_at" + + +def asset_exists(asset_hash: str) -> bool: + with create_session() as session: + return asset_exists_by_hash(session, asset_hash=asset_hash) + +def list_assets( + include_tags: Sequence[str] | None = None, + exclude_tags: Sequence[str] | None = None, + name_contains: str | None = None, + metadata_filter: dict | None = None, + limit: int = 20, + offset: int = 0, + sort: str = "created_at", + order: str = "desc", + owner_id: str = "", +) -> schemas_out.AssetsList: + sort = _safe_sort_field(sort) + order = "desc" if (order or "desc").lower() not in {"asc", "desc"} else order.lower() + + with create_session() as session: + infos, tag_map, total = list_asset_infos_page( + session, + owner_id=owner_id, + include_tags=include_tags, + exclude_tags=exclude_tags, + name_contains=name_contains, + metadata_filter=metadata_filter, + limit=limit, + offset=offset, + sort=sort, + order=order, + ) + + summaries: list[schemas_out.AssetSummary] = [] + for info in infos: + asset = info.asset + tags = tag_map.get(info.id, []) + summaries.append( + schemas_out.AssetSummary( + id=info.id, + name=info.name, + asset_hash=asset.hash if asset else None, + size=int(asset.size_bytes) if asset else None, + mime_type=asset.mime_type if asset else None, + tags=tags, + preview_url=f"/api/assets/{info.id}/content", + created_at=info.created_at, + updated_at=info.updated_at, + last_access_time=info.last_access_time, + ) + ) + + return schemas_out.AssetsList( + assets=summaries, + total=total, + has_more=(offset + len(summaries)) < total, + ) + +def get_asset(asset_info_id: str, owner_id: str = "") -> schemas_out.AssetDetail: + with create_session() as session: + res = fetch_asset_info_asset_and_tags(session, asset_info_id=asset_info_id, owner_id=owner_id) + if not res: + raise ValueError(f"AssetInfo {asset_info_id} not found") + info, asset, tag_names = res + preview_id = info.preview_id + + return schemas_out.AssetDetail( + id=info.id, + name=info.name, + asset_hash=asset.hash if asset else None, + size=int(asset.size_bytes) if asset and asset.size_bytes is not None else None, + mime_type=asset.mime_type if asset else None, + tags=tag_names, + user_metadata=info.user_metadata or {}, + preview_id=preview_id, + created_at=info.created_at, + last_access_time=info.last_access_time, + ) + +def list_tags( + prefix: str | None = None, + limit: int = 100, + offset: int = 0, + order: str = "count_desc", + include_zero: bool = True, + owner_id: str = "", +) -> schemas_out.TagsList: + limit = max(1, min(1000, limit)) + offset = max(0, offset) + + with create_session() as session: + rows, total = list_tags_with_usage( + session, + prefix=prefix, + limit=limit, + offset=offset, + include_zero=include_zero, + order=order, + owner_id=owner_id, + ) + + tags = [schemas_out.TagUsage(name=name, count=count, type=tag_type) for (name, tag_type, count) in rows] + return schemas_out.TagsList(tags=tags, total=total, has_more=(offset + len(tags)) < total) diff --git a/app/assets/scanner.py b/app/assets/scanner.py new file mode 100644 index 000000000..a16e41d94 --- /dev/null +++ b/app/assets/scanner.py @@ -0,0 +1,229 @@ +import contextlib +import time +import logging +import os +import sqlalchemy + +import folder_paths +from app.database.db import create_session, dependencies_available +from app.assets.helpers import ( + collect_models_files, compute_relative_filename, fast_asset_file_check, get_name_and_tags_from_asset_path, + list_tree,prefixes_for_root, escape_like_prefix, + RootType +) +from app.assets.database.tags import add_missing_tag_for_asset_id, ensure_tags_exist, remove_missing_tag_for_asset_id +from app.assets.database.bulk_ops import seed_from_paths_batch +from app.assets.database.models import Asset, AssetCacheState, AssetInfo + + +def seed_assets(roots: tuple[RootType, ...], enable_logging: bool = False) -> None: + """ + Scan the given roots and seed the assets into the database. + """ + if not dependencies_available(): + if enable_logging: + logging.warning("Database dependencies not available, skipping assets scan") + return + t_start = time.perf_counter() + created = 0 + skipped_existing = 0 + paths: list[str] = [] + try: + existing_paths: set[str] = set() + for r in roots: + try: + survivors: set[str] = _fast_db_consistency_pass(r, collect_existing_paths=True, update_missing_tags=True) + if survivors: + existing_paths.update(survivors) + except Exception as e: + logging.exception("fast DB scan failed for %s: %s", r, e) + + if "models" in roots: + paths.extend(collect_models_files()) + if "input" in roots: + paths.extend(list_tree(folder_paths.get_input_directory())) + if "output" in roots: + paths.extend(list_tree(folder_paths.get_output_directory())) + + specs: list[dict] = [] + tag_pool: set[str] = set() + for p in paths: + abs_p = os.path.abspath(p) + if abs_p in existing_paths: + skipped_existing += 1 + continue + try: + stat_p = os.stat(abs_p, follow_symlinks=False) + except OSError: + continue + # skip empty files + if not stat_p.st_size: + continue + name, tags = get_name_and_tags_from_asset_path(abs_p) + specs.append( + { + "abs_path": abs_p, + "size_bytes": stat_p.st_size, + "mtime_ns": getattr(stat_p, "st_mtime_ns", int(stat_p.st_mtime * 1_000_000_000)), + "info_name": name, + "tags": tags, + "fname": compute_relative_filename(abs_p), + } + ) + for t in tags: + tag_pool.add(t) + # if no file specs, nothing to do + if not specs: + return + with create_session() as sess: + if tag_pool: + ensure_tags_exist(sess, tag_pool, tag_type="user") + + result = seed_from_paths_batch(sess, specs=specs, owner_id="") + created += result["inserted_infos"] + sess.commit() + finally: + if enable_logging: + logging.info( + "Assets scan(roots=%s) completed in %.3fs (created=%d, skipped_existing=%d, total_seen=%d)", + roots, + time.perf_counter() - t_start, + created, + skipped_existing, + len(paths), + ) + + +def _fast_db_consistency_pass( + root: RootType, + *, + collect_existing_paths: bool = False, + update_missing_tags: bool = False, +) -> set[str] | None: + """Fast DB+FS pass for a root: + - Toggle needs_verify per state using fast check + - For hashed assets with at least one fast-ok state in this root: delete stale missing states + - For seed assets with all states missing: delete Asset and its AssetInfos + - Optionally add/remove 'missing' tags based on fast-ok in this root + - Optionally return surviving absolute paths + """ + prefixes = prefixes_for_root(root) + if not prefixes: + return set() if collect_existing_paths else None + + conds = [] + for p in prefixes: + base = os.path.abspath(p) + if not base.endswith(os.sep): + base += os.sep + escaped, esc = escape_like_prefix(base) + conds.append(AssetCacheState.file_path.like(escaped + "%", escape=esc)) + + with create_session() as sess: + rows = ( + sess.execute( + sqlalchemy.select( + AssetCacheState.id, + AssetCacheState.file_path, + AssetCacheState.mtime_ns, + AssetCacheState.needs_verify, + AssetCacheState.asset_id, + Asset.hash, + Asset.size_bytes, + ) + .join(Asset, Asset.id == AssetCacheState.asset_id) + .where(sqlalchemy.or_(*conds)) + .order_by(AssetCacheState.asset_id.asc(), AssetCacheState.id.asc()) + ) + ).all() + + by_asset: dict[str, dict] = {} + for sid, fp, mtime_db, needs_verify, aid, a_hash, a_size in rows: + acc = by_asset.get(aid) + if acc is None: + acc = {"hash": a_hash, "size_db": int(a_size or 0), "states": []} + by_asset[aid] = acc + + fast_ok = False + try: + exists = True + fast_ok = fast_asset_file_check( + mtime_db=mtime_db, + size_db=acc["size_db"], + stat_result=os.stat(fp, follow_symlinks=True), + ) + except FileNotFoundError: + exists = False + except OSError: + exists = False + + acc["states"].append({ + "sid": sid, + "fp": fp, + "exists": exists, + "fast_ok": fast_ok, + "needs_verify": bool(needs_verify), + }) + + to_set_verify: list[int] = [] + to_clear_verify: list[int] = [] + stale_state_ids: list[int] = [] + survivors: set[str] = set() + + for aid, acc in by_asset.items(): + a_hash = acc["hash"] + states = acc["states"] + any_fast_ok = any(s["fast_ok"] for s in states) + all_missing = all(not s["exists"] for s in states) + + for s in states: + if not s["exists"]: + continue + if s["fast_ok"] and s["needs_verify"]: + to_clear_verify.append(s["sid"]) + if not s["fast_ok"] and not s["needs_verify"]: + to_set_verify.append(s["sid"]) + + if a_hash is None: + if states and all_missing: # remove seed Asset completely, if no valid AssetCache exists + sess.execute(sqlalchemy.delete(AssetInfo).where(AssetInfo.asset_id == aid)) + asset = sess.get(Asset, aid) + if asset: + sess.delete(asset) + else: + for s in states: + if s["exists"]: + survivors.add(os.path.abspath(s["fp"])) + continue + + if any_fast_ok: # if Asset has at least one valid AssetCache record, remove any invalid AssetCache records + for s in states: + if not s["exists"]: + stale_state_ids.append(s["sid"]) + if update_missing_tags: + with contextlib.suppress(Exception): + remove_missing_tag_for_asset_id(sess, asset_id=aid) + elif update_missing_tags: + with contextlib.suppress(Exception): + add_missing_tag_for_asset_id(sess, asset_id=aid, origin="automatic") + + for s in states: + if s["exists"]: + survivors.add(os.path.abspath(s["fp"])) + + if stale_state_ids: + sess.execute(sqlalchemy.delete(AssetCacheState).where(AssetCacheState.id.in_(stale_state_ids))) + if to_set_verify: + sess.execute( + sqlalchemy.update(AssetCacheState) + .where(AssetCacheState.id.in_(to_set_verify)) + .values(needs_verify=True) + ) + if to_clear_verify: + sess.execute( + sqlalchemy.update(AssetCacheState) + .where(AssetCacheState.id.in_(to_clear_verify)) + .values(needs_verify=False) + ) + sess.commit() + return survivors if collect_existing_paths else None diff --git a/app/database/models.py b/app/database/models.py index 6facfb8f2..e7572677a 100644 --- a/app/database/models.py +++ b/app/database/models.py @@ -1,14 +1,21 @@ -from sqlalchemy.orm import declarative_base +from typing import Any +from datetime import datetime +from sqlalchemy.orm import DeclarativeBase -Base = declarative_base() +class Base(DeclarativeBase): + pass - -def to_dict(obj): +def to_dict(obj: Any, include_none: bool = False) -> dict[str, Any]: fields = obj.__table__.columns.keys() - return { - field: (val.to_dict() if hasattr(val, "to_dict") else val) - for field in fields - if (val := getattr(obj, field)) - } + out: dict[str, Any] = {} + for field in fields: + val = getattr(obj, field) + if val is None and not include_none: + continue + if isinstance(val, datetime): + out[field] = val.isoformat() + else: + out[field] = val + return out # TODO: Define models here diff --git a/comfy/cli_args.py b/comfy/cli_args.py index dae9a895d..1716c3de7 100644 --- a/comfy/cli_args.py +++ b/comfy/cli_args.py @@ -231,6 +231,7 @@ database_default_path = os.path.abspath( os.path.join(os.path.dirname(__file__), "..", "user", "comfyui.db") ) parser.add_argument("--database-url", type=str, default=f"sqlite:///{database_default_path}", help="Specify the database URL, e.g. for an in-memory database you can use 'sqlite:///:memory:'.") +parser.add_argument("--disable-assets-autoscan", action="store_true", help="Disable asset scanning on startup for database synchronization.") if comfy.options.args_parsing: args = parser.parse_args() diff --git a/main.py b/main.py index 0e07a95da..37b06c1fa 100644 --- a/main.py +++ b/main.py @@ -7,6 +7,7 @@ import folder_paths import time from comfy.cli_args import args from app.logger import setup_logger +from app.assets.scanner import seed_assets import itertools import utils.extra_config import logging @@ -324,6 +325,8 @@ def setup_database(): from app.database.db import init_db, dependencies_available if dependencies_available(): init_db() + if not args.disable_assets_autoscan: + seed_assets(["models"], enable_logging=True) except Exception as e: logging.error(f"Failed to initialize database. Please ensure you have installed the latest requirements. If the error persists, please report this as in future the database will be required: {e}") diff --git a/server.py b/server.py index 4db3347cb..da2baefd4 100644 --- a/server.py +++ b/server.py @@ -33,6 +33,8 @@ import node_helpers from comfyui_version import __version__ from app.frontend_management import FrontendManager, parse_version from comfy_api.internal import _ComfyNodeInternal +from app.assets.scanner import seed_assets +from app.assets.api.routes import register_assets_system from app.user_manager import UserManager from app.model_manager import ModelFileManager @@ -235,6 +237,7 @@ class PromptServer(): else args.front_end_root ) logging.info(f"[Prompt Server] web root: {self.web_root}") + register_assets_system(self.app, self.user_manager) routes = web.RouteTableDef() self.routes = routes self.last_node_id = None @@ -683,6 +686,7 @@ class PromptServer(): @routes.get("/object_info") async def get_object_info(request): + seed_assets(["models"]) with folder_paths.cache_helper: out = {} for x in nodes.NODE_CLASS_MAPPINGS: From 6207f86c18d2cf2d70ab059987b62d4b38466e77 Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Thu, 8 Jan 2026 20:34:48 -0800 Subject: [PATCH 413/647] Fix VAEEncodeForInpaint to support WAN VAE tuple downscale_ratio (#11572) Use vae.spacial_compression_encode() instead of directly accessing downscale_ratio to handle both standard VAEs (int) and WAN VAEs (tuple). Addresses reviewer feedback on PR #11259. Co-authored-by: ChrisFab16 --- nodes.py | 9 +++++---- 1 file changed, 5 insertions(+), 4 deletions(-) diff --git a/nodes.py b/nodes.py index 56b74ebe3..1aa391f4a 100644 --- a/nodes.py +++ b/nodes.py @@ -378,14 +378,15 @@ class VAEEncodeForInpaint: CATEGORY = "latent/inpaint" def encode(self, vae, pixels, mask, grow_mask_by=6): - x = (pixels.shape[1] // vae.downscale_ratio) * vae.downscale_ratio - y = (pixels.shape[2] // vae.downscale_ratio) * vae.downscale_ratio + downscale_ratio = vae.spacial_compression_encode() + x = (pixels.shape[1] // downscale_ratio) * downscale_ratio + y = (pixels.shape[2] // downscale_ratio) * downscale_ratio mask = torch.nn.functional.interpolate(mask.reshape((-1, 1, mask.shape[-2], mask.shape[-1])), size=(pixels.shape[1], pixels.shape[2]), mode="bilinear") pixels = pixels.clone() if pixels.shape[1] != x or pixels.shape[2] != y: - x_offset = (pixels.shape[1] % vae.downscale_ratio) // 2 - y_offset = (pixels.shape[2] % vae.downscale_ratio) // 2 + x_offset = (pixels.shape[1] % downscale_ratio) // 2 + y_offset = (pixels.shape[2] % downscale_ratio) // 2 pixels = pixels[:,x_offset:x + x_offset, y_offset:y + y_offset,:] mask = mask[:,:,x_offset:x + x_offset, y_offset:y + y_offset] From 4609fcd26081156eef921bd9f43726f670ee6f51 Mon Sep 17 00:00:00 2001 From: Terry Jia Date: Fri, 9 Jan 2026 00:31:19 -0500 Subject: [PATCH 414/647] add node - image compare (#11343) --- comfy_api/latest/_io.py | 13 +++++++ comfy_extras/nodes_image_compare.py | 53 +++++++++++++++++++++++++++++ nodes.py | 1 + 3 files changed, 67 insertions(+) create mode 100644 comfy_extras/nodes_image_compare.py diff --git a/comfy_api/latest/_io.py b/comfy_api/latest/_io.py index 764fa8b2b..50143ff53 100644 --- a/comfy_api/latest/_io.py +++ b/comfy_api/latest/_io.py @@ -1113,6 +1113,18 @@ class DynamicSlot(ComfyTypeI): out_dict[input_type][finalized_id] = value out_dict["dynamic_paths"][finalized_id] = finalize_prefix(curr_prefix, curr_prefix[-1]) +@comfytype(io_type="IMAGECOMPARE") +class ImageCompare(ComfyTypeI): + Type = dict + + class Input(WidgetInput): + def __init__(self, id: str, display_name: str=None, optional=False, tooltip: str=None, + socketless: bool=True): + super().__init__(id, display_name, optional, tooltip, None, None, socketless) + + def as_dict(self): + return super().as_dict() + DYNAMIC_INPUT_LOOKUP: dict[str, Callable[[dict[str, Any], dict[str, Any], tuple[str, dict[str, Any]], str, list[str] | None], None]] = {} def register_dynamic_input_func(io_type: str, func: Callable[[dict[str, Any], dict[str, Any], tuple[str, dict[str, Any]], str, list[str] | None], None]): DYNAMIC_INPUT_LOOKUP[io_type] = func @@ -1958,4 +1970,5 @@ __all__ = [ "add_to_dict_v1", "add_to_dict_v3", "V3Data", + "ImageCompare", ] diff --git a/comfy_extras/nodes_image_compare.py b/comfy_extras/nodes_image_compare.py new file mode 100644 index 000000000..8e9f809e6 --- /dev/null +++ b/comfy_extras/nodes_image_compare.py @@ -0,0 +1,53 @@ +import nodes + +from typing_extensions import override +from comfy_api.latest import IO, ComfyExtension + + +class ImageCompare(IO.ComfyNode): + """Compares two images with a slider interface.""" + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="ImageCompare", + display_name="Image Compare", + description="Compares two images side by side with a slider.", + category="image", + is_experimental=True, + is_output_node=True, + inputs=[ + IO.Image.Input("image_a", optional=True), + IO.Image.Input("image_b", optional=True), + IO.ImageCompare.Input("compare_view"), + ], + outputs=[], + ) + + @classmethod + def execute(cls, image_a=None, image_b=None, compare_view=None) -> IO.NodeOutput: + result = {"a_images": [], "b_images": []} + + preview_node = nodes.PreviewImage() + + if image_a is not None and len(image_a) > 0: + saved = preview_node.save_images(image_a, "comfy.compare.a") + result["a_images"] = saved["ui"]["images"] + + if image_b is not None and len(image_b) > 0: + saved = preview_node.save_images(image_b, "comfy.compare.b") + result["b_images"] = saved["ui"]["images"] + + return IO.NodeOutput(ui=result) + + +class ImageCompareExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[IO.ComfyNode]]: + return [ + ImageCompare, + ] + + +async def comfy_entrypoint() -> ImageCompareExtension: + return ImageCompareExtension() diff --git a/nodes.py b/nodes.py index 1aa391f4a..5a9d42d4a 100644 --- a/nodes.py +++ b/nodes.py @@ -2370,6 +2370,7 @@ async def init_builtin_extra_nodes(): "nodes_nop.py", "nodes_kandinsky5.py", "nodes_wanmove.py", + "nodes_image_compare.py", ] import_failed = [] From 04c49a29b493f3f9037b83cec45f6369b5c4816b Mon Sep 17 00:00:00 2001 From: ric-yu Date: Thu, 8 Jan 2026 21:57:36 -0800 Subject: [PATCH 415/647] feat: add cancelled filter to /jobs (#11680) --- comfy_execution/jobs.py | 31 +++++++++++++++++------------ tests/execution/test_jobs.py | 38 +++++++++++++++++++++++++++++++++++- 2 files changed, 55 insertions(+), 14 deletions(-) diff --git a/comfy_execution/jobs.py b/comfy_execution/jobs.py index 59fb49357..97fd803b8 100644 --- a/comfy_execution/jobs.py +++ b/comfy_execution/jobs.py @@ -14,8 +14,9 @@ class JobStatus: IN_PROGRESS = 'in_progress' COMPLETED = 'completed' FAILED = 'failed' + CANCELLED = 'cancelled' - ALL = [PENDING, IN_PROGRESS, COMPLETED, FAILED] + ALL = [PENDING, IN_PROGRESS, COMPLETED, FAILED, CANCELLED] # Media types that can be previewed in the frontend @@ -94,12 +95,6 @@ def normalize_history_item(prompt_id: str, history_item: dict, include_outputs: status_info = history_item.get('status', {}) status_str = status_info.get('status_str') if status_info else None - if status_str == 'success': - status = JobStatus.COMPLETED - elif status_str == 'error': - status = JobStatus.FAILED - else: - status = JobStatus.COMPLETED outputs = history_item.get('outputs', {}) outputs_count, preview_output = get_outputs_summary(outputs) @@ -107,6 +102,7 @@ def normalize_history_item(prompt_id: str, history_item: dict, include_outputs: execution_error = None execution_start_time = None execution_end_time = None + was_interrupted = False if status_info: messages = status_info.get('messages', []) for entry in messages: @@ -119,6 +115,15 @@ def normalize_history_item(prompt_id: str, history_item: dict, include_outputs: execution_end_time = event_data.get('timestamp') if event_name == 'execution_error': execution_error = event_data + elif event_name == 'execution_interrupted': + was_interrupted = True + + if status_str == 'success': + status = JobStatus.COMPLETED + elif status_str == 'error': + status = JobStatus.CANCELLED if was_interrupted else JobStatus.FAILED + else: + status = JobStatus.COMPLETED job = prune_dict({ 'id': prompt_id, @@ -268,13 +273,13 @@ def get_all_jobs( for item in queued: jobs.append(normalize_queue_item(item, JobStatus.PENDING)) - include_completed = JobStatus.COMPLETED in status_filter - include_failed = JobStatus.FAILED in status_filter - if include_completed or include_failed: + history_statuses = {JobStatus.COMPLETED, JobStatus.FAILED, JobStatus.CANCELLED} + requested_history_statuses = history_statuses & set(status_filter) + if requested_history_statuses: for prompt_id, history_item in history.items(): - is_failed = history_item.get('status', {}).get('status_str') == 'error' - if (is_failed and include_failed) or (not is_failed and include_completed): - jobs.append(normalize_history_item(prompt_id, history_item)) + job = normalize_history_item(prompt_id, history_item) + if job.get('status') in requested_history_statuses: + jobs.append(job) if workflow_id: jobs = [j for j in jobs if j.get('workflow_id') == workflow_id] diff --git a/tests/execution/test_jobs.py b/tests/execution/test_jobs.py index 918c8080a..4d2f9ed36 100644 --- a/tests/execution/test_jobs.py +++ b/tests/execution/test_jobs.py @@ -19,6 +19,7 @@ class TestJobStatus: assert JobStatus.IN_PROGRESS == 'in_progress' assert JobStatus.COMPLETED == 'completed' assert JobStatus.FAILED == 'failed' + assert JobStatus.CANCELLED == 'cancelled' def test_all_contains_all_statuses(self): """ALL should contain all status values.""" @@ -26,7 +27,8 @@ class TestJobStatus: assert JobStatus.IN_PROGRESS in JobStatus.ALL assert JobStatus.COMPLETED in JobStatus.ALL assert JobStatus.FAILED in JobStatus.ALL - assert len(JobStatus.ALL) == 4 + assert JobStatus.CANCELLED in JobStatus.ALL + assert len(JobStatus.ALL) == 5 class TestIsPreviewable: @@ -336,6 +338,40 @@ class TestNormalizeHistoryItem: assert job['execution_error']['node_type'] == 'KSampler' assert job['execution_error']['exception_message'] == 'CUDA out of memory' + def test_cancelled_job(self): + """Cancelled/interrupted history item should have cancelled status.""" + history_item = { + 'prompt': ( + 5, + 'prompt-cancelled', + {'nodes': {}}, + {'create_time': 1234567890000}, + ['node1'], + ), + 'status': { + 'status_str': 'error', + 'completed': False, + 'messages': [ + ('execution_start', {'prompt_id': 'prompt-cancelled', 'timestamp': 1234567890500}), + ('execution_interrupted', { + 'prompt_id': 'prompt-cancelled', + 'node_id': '5', + 'node_type': 'KSampler', + 'executed': ['1', '2', '3'], + 'timestamp': 1234567891000, + }) + ] + }, + 'outputs': {}, + } + + job = normalize_history_item('prompt-cancelled', history_item) + assert job['status'] == 'cancelled' + assert job['execution_start_time'] == 1234567890500 + assert job['execution_end_time'] == 1234567891000 + # Cancelled jobs should not have execution_error set + assert 'execution_error' not in job + def test_include_outputs(self): """When include_outputs=True, should include full output data.""" history_item = { From ec0a832acb25fbe53bd4fc25d286a9ee442a3bcf Mon Sep 17 00:00:00 2001 From: Jedrzej Kosinski Date: Thu, 8 Jan 2026 22:49:12 -0800 Subject: [PATCH 416/647] Add workaround for hacky nodepack(s) that edit folder_names_and_paths to have values with tuples of more than 2. Other things could potentially break with those nodepack(s), so I will hunt for the guilty nodepack(s) now. (#11755) --- app/assets/helpers.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/app/assets/helpers.py b/app/assets/helpers.py index 6755d0e56..08b465b5a 100644 --- a/app/assets/helpers.py +++ b/app/assets/helpers.py @@ -81,7 +81,8 @@ def get_comfy_models_folders() -> list[tuple[str, list[str]]]: """ targets: list[tuple[str, list[str]]] = [] models_root = os.path.abspath(folder_paths.models_dir) - for name, (paths, _exts) in folder_paths.folder_names_and_paths.items(): + for name, values in folder_paths.folder_names_and_paths.items(): + paths, _exts = values[0], values[1] # NOTE: this prevents nodepacks that hackily edit folder_... from breaking ComfyUI if any(os.path.abspath(p).startswith(models_root + os.sep) for p in paths): targets.append((name, paths)) return targets From bd0e6825e84606e0706bbb5645e9ea1f4ad8154d Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Fri, 9 Jan 2026 11:21:06 -0800 Subject: [PATCH 417/647] Be less strict when loading mixed ops weights. (#11769) --- comfy/ops.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/comfy/ops.py b/comfy/ops.py index 8156c42ff..1cf22f0cc 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -546,7 +546,8 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec weight_key = f"{prefix}weight" weight = state_dict.pop(weight_key, None) if weight is None: - raise ValueError(f"Missing weight for layer {layer_name}") + logging.warning(f"Missing weight for layer {layer_name}") + return manually_loaded_keys = [weight_key] From 4484b93d615059012d3a5ce91d1dbbb0cbaa2d76 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Fri, 9 Jan 2026 22:25:56 +0200 Subject: [PATCH 418/647] fix(api-nodes): do not downscale the input image for Topaz Enhance (#11768) --- comfy_api_nodes/nodes_topaz.py | 7 ++++--- comfy_api_nodes/util/upload_helpers.py | 3 ++- 2 files changed, 6 insertions(+), 4 deletions(-) diff --git a/comfy_api_nodes/nodes_topaz.py b/comfy_api_nodes/nodes_topaz.py index b04575ad8..9dc5f45bc 100644 --- a/comfy_api_nodes/nodes_topaz.py +++ b/comfy_api_nodes/nodes_topaz.py @@ -2,7 +2,6 @@ import builtins from io import BytesIO import aiohttp -import torch from typing_extensions import override from comfy_api.latest import IO, ComfyExtension, Input @@ -138,7 +137,7 @@ class TopazImageEnhance(IO.ComfyNode): async def execute( cls, model: str, - image: torch.Tensor, + image: Input.Image, prompt: str = "", subject_detection: str = "All", face_enhancement: bool = True, @@ -153,7 +152,9 @@ class TopazImageEnhance(IO.ComfyNode): ) -> IO.NodeOutput: if get_number_of_images(image) != 1: raise ValueError("Only one input image is supported.") - download_url = await upload_images_to_comfyapi(cls, image, max_images=1, mime_type="image/png") + download_url = await upload_images_to_comfyapi( + cls, image, max_images=1, mime_type="image/png", total_pixels=4096*4096 + ) initial_response = await sync_op( cls, ApiEndpoint(path="/proxy/topaz/image/v1/enhance-gen/async", method="POST"), diff --git a/comfy_api_nodes/util/upload_helpers.py b/comfy_api_nodes/util/upload_helpers.py index f1ed7fe9c..2535a0884 100644 --- a/comfy_api_nodes/util/upload_helpers.py +++ b/comfy_api_nodes/util/upload_helpers.py @@ -49,6 +49,7 @@ async def upload_images_to_comfyapi( mime_type: str | None = None, wait_label: str | None = "Uploading", show_batch_index: bool = True, + total_pixels: int = 2048 * 2048, ) -> list[str]: """ Uploads images to ComfyUI API and returns download URLs. @@ -63,7 +64,7 @@ async def upload_images_to_comfyapi( for idx in range(num_to_upload): tensor = image[idx] if is_batch else image - img_io = tensor_to_bytesio(tensor, mime_type=mime_type) + img_io = tensor_to_bytesio(tensor, total_pixels=total_pixels, mime_type=mime_type) effective_label = wait_label if wait_label and show_batch_index and num_to_upload > 1: From 393d2880ddc6e57c0fa3b878bb50fa2901bd57e6 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Fri, 9 Jan 2026 22:59:38 +0200 Subject: [PATCH 419/647] feat(api-nodes): added nodes for Vidu2 (#11760) --- comfy_api_nodes/apis/vidu.py | 41 +++ comfy_api_nodes/nodes_vidu.py | 588 +++++++++++++++++++++++++--------- 2 files changed, 482 insertions(+), 147 deletions(-) create mode 100644 comfy_api_nodes/apis/vidu.py diff --git a/comfy_api_nodes/apis/vidu.py b/comfy_api_nodes/apis/vidu.py new file mode 100644 index 000000000..a9bb6f7ce --- /dev/null +++ b/comfy_api_nodes/apis/vidu.py @@ -0,0 +1,41 @@ +from pydantic import BaseModel, Field + + +class SubjectReference(BaseModel): + id: str = Field(...) + images: list[str] = Field(...) + + +class TaskCreationRequest(BaseModel): + model: str = Field(...) + prompt: str = Field(..., max_length=2000) + duration: int = Field(...) + seed: int = Field(..., ge=0, le=2147483647) + aspect_ratio: str | None = Field(None) + resolution: str | None = Field(None) + movement_amplitude: str | None = Field(None) + images: list[str] | None = Field(None, description="Base64 encoded string or image URL") + subjects: list[SubjectReference] | None = Field(None) + bgm: bool | None = Field(None) + audio: bool | None = Field(None) + + +class TaskCreationResponse(BaseModel): + task_id: str = Field(...) + state: str = Field(...) + created_at: str = Field(...) + code: int | None = Field(None, description="Error code") + + +class TaskResult(BaseModel): + id: str = Field(..., description="Creation id") + url: str = Field(..., description="The URL of the generated results, valid for one hour") + cover_url: str = Field(..., description="The cover URL of the generated results, valid for one hour") + + +class TaskStatusResponse(BaseModel): + state: str = Field(...) + err_code: str | None = Field(None) + progress: float | None = Field(None) + credits: int | None = Field(None) + creations: list[TaskResult] = Field(..., description="Generated results") diff --git a/comfy_api_nodes/nodes_vidu.py b/comfy_api_nodes/nodes_vidu.py index 7a679f0d9..9d94ae7ad 100644 --- a/comfy_api_nodes/nodes_vidu.py +++ b/comfy_api_nodes/nodes_vidu.py @@ -1,12 +1,13 @@ -import logging -from enum import Enum -from typing import Literal, Optional, TypeVar - -import torch -from pydantic import BaseModel, Field from typing_extensions import override -from comfy_api.latest import IO, ComfyExtension +from comfy_api.latest import IO, ComfyExtension, Input +from comfy_api_nodes.apis.vidu import ( + SubjectReference, + TaskCreationRequest, + TaskCreationResponse, + TaskResult, + TaskStatusResponse, +) from comfy_api_nodes.util import ( ApiEndpoint, download_url_to_video_output, @@ -17,6 +18,7 @@ from comfy_api_nodes.util import ( validate_image_aspect_ratio, validate_image_dimensions, validate_images_aspect_ratio_closeness, + validate_string, ) VIDU_TEXT_TO_VIDEO = "/proxy/vidu/text2video" @@ -25,98 +27,33 @@ VIDU_REFERENCE_VIDEO = "/proxy/vidu/reference2video" VIDU_START_END_VIDEO = "/proxy/vidu/start-end2video" VIDU_GET_GENERATION_STATUS = "/proxy/vidu/tasks/%s/creations" -R = TypeVar("R") - - -class VideoModelName(str, Enum): - vidu_q1 = "viduq1" - - -class AspectRatio(str, Enum): - r_16_9 = "16:9" - r_9_16 = "9:16" - r_1_1 = "1:1" - - -class Resolution(str, Enum): - r_1080p = "1080p" - - -class MovementAmplitude(str, Enum): - auto = "auto" - small = "small" - medium = "medium" - large = "large" - - -class TaskCreationRequest(BaseModel): - model: VideoModelName = VideoModelName.vidu_q1 - prompt: Optional[str] = Field(None, max_length=1500) - duration: Optional[Literal[5]] = 5 - seed: Optional[int] = Field(0, ge=0, le=2147483647) - aspect_ratio: Optional[AspectRatio] = AspectRatio.r_16_9 - resolution: Optional[Resolution] = Resolution.r_1080p - movement_amplitude: Optional[MovementAmplitude] = MovementAmplitude.auto - images: Optional[list[str]] = Field(None, description="Base64 encoded string or image URL") - - -class TaskCreationResponse(BaseModel): - task_id: str = Field(...) - state: str = Field(...) - created_at: str = Field(...) - code: Optional[int] = Field(None, description="Error code") - - -class TaskResult(BaseModel): - id: str = Field(..., description="Creation id") - url: str = Field(..., description="The URL of the generated results, valid for one hour") - cover_url: str = Field(..., description="The cover URL of the generated results, valid for one hour") - - -class TaskStatusResponse(BaseModel): - state: str = Field(...) - err_code: Optional[str] = Field(None) - creations: list[TaskResult] = Field(..., description="Generated results") - - -def get_video_url_from_response(response) -> Optional[str]: - if response.creations: - return response.creations[0].url - return None - - -def get_video_from_response(response) -> TaskResult: - if not response.creations: - error_msg = f"Vidu request does not contain results. State: {response.state}, Error Code: {response.err_code}" - logging.info(error_msg) - raise RuntimeError(error_msg) - logging.info("Vidu task %s succeeded. Video URL: %s", response.creations[0].id, response.creations[0].url) - return response.creations[0] - async def execute_task( cls: type[IO.ComfyNode], vidu_endpoint: str, payload: TaskCreationRequest, - estimated_duration: int, -) -> R: - response = await sync_op( +) -> list[TaskResult]: + task_creation_response = await sync_op( cls, endpoint=ApiEndpoint(path=vidu_endpoint, method="POST"), response_model=TaskCreationResponse, data=payload, ) - if response.state == "failed": - error_msg = f"Vidu request failed. Code: {response.code}" - logging.error(error_msg) - raise RuntimeError(error_msg) - return await poll_op( + if task_creation_response.state == "failed": + raise RuntimeError(f"Vidu request failed. Code: {task_creation_response.code}") + response = await poll_op( cls, - ApiEndpoint(path=VIDU_GET_GENERATION_STATUS % response.task_id), + ApiEndpoint(path=VIDU_GET_GENERATION_STATUS % task_creation_response.task_id), response_model=TaskStatusResponse, status_extractor=lambda r: r.state, - estimated_duration=estimated_duration, + progress_extractor=lambda r: r.progress, + max_poll_attempts=320, ) + if not response.creations: + raise RuntimeError( + f"Vidu request does not contain results. State: {response.state}, Error Code: {response.err_code}" + ) + return response.creations class ViduTextToVideoNode(IO.ComfyNode): @@ -127,14 +64,9 @@ class ViduTextToVideoNode(IO.ComfyNode): node_id="ViduTextToVideoNode", display_name="Vidu Text To Video Generation", category="api node/video/Vidu", - description="Generate video from text prompt", + description="Generate video from a text prompt", inputs=[ - IO.Combo.Input( - "model", - options=VideoModelName, - default=VideoModelName.vidu_q1, - tooltip="Model name", - ), + IO.Combo.Input("model", options=["viduq1"], tooltip="Model name"), IO.String.Input( "prompt", multiline=True, @@ -163,22 +95,19 @@ class ViduTextToVideoNode(IO.ComfyNode): ), IO.Combo.Input( "aspect_ratio", - options=AspectRatio, - default=AspectRatio.r_16_9, + options=["16:9", "9:16", "1:1"], tooltip="The aspect ratio of the output video", optional=True, ), IO.Combo.Input( "resolution", - options=Resolution, - default=Resolution.r_1080p, + options=["1080p"], tooltip="Supported values may vary by model & duration", optional=True, ), IO.Combo.Input( "movement_amplitude", - options=MovementAmplitude, - default=MovementAmplitude.auto, + options=["auto", "small", "medium", "large"], tooltip="The movement amplitude of objects in the frame", optional=True, ), @@ -208,7 +137,7 @@ class ViduTextToVideoNode(IO.ComfyNode): if not prompt: raise ValueError("The prompt field is required and cannot be empty.") payload = TaskCreationRequest( - model_name=model, + model=model, prompt=prompt, duration=duration, seed=seed, @@ -216,8 +145,8 @@ class ViduTextToVideoNode(IO.ComfyNode): resolution=resolution, movement_amplitude=movement_amplitude, ) - results = await execute_task(cls, VIDU_TEXT_TO_VIDEO, payload, 320) - return IO.NodeOutput(await download_url_to_video_output(get_video_from_response(results).url)) + results = await execute_task(cls, VIDU_TEXT_TO_VIDEO, payload) + return IO.NodeOutput(await download_url_to_video_output(results[0].url)) class ViduImageToVideoNode(IO.ComfyNode): @@ -230,12 +159,7 @@ class ViduImageToVideoNode(IO.ComfyNode): category="api node/video/Vidu", description="Generate video from image and optional prompt", inputs=[ - IO.Combo.Input( - "model", - options=VideoModelName, - default=VideoModelName.vidu_q1, - tooltip="Model name", - ), + IO.Combo.Input("model", options=["viduq1"], tooltip="Model name"), IO.Image.Input( "image", tooltip="An image to be used as the start frame of the generated video", @@ -270,15 +194,13 @@ class ViduImageToVideoNode(IO.ComfyNode): ), IO.Combo.Input( "resolution", - options=Resolution, - default=Resolution.r_1080p, + options=["1080p"], tooltip="Supported values may vary by model & duration", optional=True, ), IO.Combo.Input( "movement_amplitude", - options=MovementAmplitude, - default=MovementAmplitude.auto.value, + options=["auto", "small", "medium", "large"], tooltip="The movement amplitude of objects in the frame", optional=True, ), @@ -298,7 +220,7 @@ class ViduImageToVideoNode(IO.ComfyNode): async def execute( cls, model: str, - image: torch.Tensor, + image: Input.Image, prompt: str, duration: int, seed: int, @@ -309,7 +231,7 @@ class ViduImageToVideoNode(IO.ComfyNode): raise ValueError("Only one input image is allowed.") validate_image_aspect_ratio(image, (1, 4), (4, 1)) payload = TaskCreationRequest( - model_name=model, + model=model, prompt=prompt, duration=duration, seed=seed, @@ -322,8 +244,8 @@ class ViduImageToVideoNode(IO.ComfyNode): max_images=1, mime_type="image/png", ) - results = await execute_task(cls, VIDU_IMAGE_TO_VIDEO, payload, 120) - return IO.NodeOutput(await download_url_to_video_output(get_video_from_response(results).url)) + results = await execute_task(cls, VIDU_IMAGE_TO_VIDEO, payload) + return IO.NodeOutput(await download_url_to_video_output(results[0].url)) class ViduReferenceVideoNode(IO.ComfyNode): @@ -334,14 +256,9 @@ class ViduReferenceVideoNode(IO.ComfyNode): node_id="ViduReferenceVideoNode", display_name="Vidu Reference To Video Generation", category="api node/video/Vidu", - description="Generate video from multiple images and prompt", + description="Generate video from multiple images and a prompt", inputs=[ - IO.Combo.Input( - "model", - options=VideoModelName, - default=VideoModelName.vidu_q1, - tooltip="Model name", - ), + IO.Combo.Input("model", options=["viduq1"], tooltip="Model name"), IO.Image.Input( "images", tooltip="Images to use as references to generate a video with consistent subjects (max 7 images).", @@ -374,22 +291,19 @@ class ViduReferenceVideoNode(IO.ComfyNode): ), IO.Combo.Input( "aspect_ratio", - options=AspectRatio, - default=AspectRatio.r_16_9, + options=["16:9", "9:16", "1:1"], tooltip="The aspect ratio of the output video", optional=True, ), IO.Combo.Input( "resolution", - options=[model.value for model in Resolution], - default=Resolution.r_1080p.value, + options=["1080p"], tooltip="Supported values may vary by model & duration", optional=True, ), IO.Combo.Input( "movement_amplitude", - options=[model.value for model in MovementAmplitude], - default=MovementAmplitude.auto.value, + options=["auto", "small", "medium", "large"], tooltip="The movement amplitude of objects in the frame", optional=True, ), @@ -409,7 +323,7 @@ class ViduReferenceVideoNode(IO.ComfyNode): async def execute( cls, model: str, - images: torch.Tensor, + images: Input.Image, prompt: str, duration: int, seed: int, @@ -426,7 +340,7 @@ class ViduReferenceVideoNode(IO.ComfyNode): validate_image_aspect_ratio(image, (1, 4), (4, 1)) validate_image_dimensions(image, min_width=128, min_height=128) payload = TaskCreationRequest( - model_name=model, + model=model, prompt=prompt, duration=duration, seed=seed, @@ -440,8 +354,8 @@ class ViduReferenceVideoNode(IO.ComfyNode): max_images=7, mime_type="image/png", ) - results = await execute_task(cls, VIDU_REFERENCE_VIDEO, payload, 120) - return IO.NodeOutput(await download_url_to_video_output(get_video_from_response(results).url)) + results = await execute_task(cls, VIDU_REFERENCE_VIDEO, payload) + return IO.NodeOutput(await download_url_to_video_output(results[0].url)) class ViduStartEndToVideoNode(IO.ComfyNode): @@ -454,12 +368,7 @@ class ViduStartEndToVideoNode(IO.ComfyNode): category="api node/video/Vidu", description="Generate a video from start and end frames and a prompt", inputs=[ - IO.Combo.Input( - "model", - options=[model.value for model in VideoModelName], - default=VideoModelName.vidu_q1.value, - tooltip="Model name", - ), + IO.Combo.Input("model", options=["viduq1"], tooltip="Model name"), IO.Image.Input( "first_frame", tooltip="Start frame", @@ -497,15 +406,13 @@ class ViduStartEndToVideoNode(IO.ComfyNode): ), IO.Combo.Input( "resolution", - options=[model.value for model in Resolution], - default=Resolution.r_1080p.value, + options=["1080p"], tooltip="Supported values may vary by model & duration", optional=True, ), IO.Combo.Input( "movement_amplitude", - options=[model.value for model in MovementAmplitude], - default=MovementAmplitude.auto.value, + options=["auto", "small", "medium", "large"], tooltip="The movement amplitude of objects in the frame", optional=True, ), @@ -525,8 +432,8 @@ class ViduStartEndToVideoNode(IO.ComfyNode): async def execute( cls, model: str, - first_frame: torch.Tensor, - end_frame: torch.Tensor, + first_frame: Input.Image, + end_frame: Input.Image, prompt: str, duration: int, seed: int, @@ -535,7 +442,7 @@ class ViduStartEndToVideoNode(IO.ComfyNode): ) -> IO.NodeOutput: validate_images_aspect_ratio_closeness(first_frame, end_frame, min_rel=0.8, max_rel=1.25, strict=False) payload = TaskCreationRequest( - model_name=model, + model=model, prompt=prompt, duration=duration, seed=seed, @@ -546,8 +453,391 @@ class ViduStartEndToVideoNode(IO.ComfyNode): (await upload_images_to_comfyapi(cls, frame, max_images=1, mime_type="image/png"))[0] for frame in (first_frame, end_frame) ] - results = await execute_task(cls, VIDU_START_END_VIDEO, payload, 96) - return IO.NodeOutput(await download_url_to_video_output(get_video_from_response(results).url)) + results = await execute_task(cls, VIDU_START_END_VIDEO, payload) + return IO.NodeOutput(await download_url_to_video_output(results[0].url)) + + +class Vidu2TextToVideoNode(IO.ComfyNode): + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="Vidu2TextToVideoNode", + display_name="Vidu2 Text-to-Video Generation", + category="api node/video/Vidu", + description="Generate video from a text prompt", + inputs=[ + IO.Combo.Input("model", options=["viduq2"]), + IO.String.Input( + "prompt", + multiline=True, + tooltip="A textual description for video generation, with a maximum length of 2000 characters.", + ), + IO.Int.Input( + "duration", + default=5, + min=1, + max=10, + step=1, + display_mode=IO.NumberDisplay.slider, + ), + IO.Int.Input( + "seed", + default=1, + min=0, + max=2147483647, + step=1, + display_mode=IO.NumberDisplay.number, + control_after_generate=True, + ), + IO.Combo.Input("aspect_ratio", options=["16:9", "9:16", "3:4", "4:3", "1:1"]), + IO.Combo.Input("resolution", options=["720p", "1080p"]), + IO.Boolean.Input( + "background_music", + default=False, + tooltip="Whether to add background music to the generated video.", + ), + ], + outputs=[ + IO.Video.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + ) + + @classmethod + async def execute( + cls, + model: str, + prompt: str, + duration: int, + seed: int, + aspect_ratio: str, + resolution: str, + background_music: bool, + ) -> IO.NodeOutput: + validate_string(prompt, min_length=1, max_length=2000) + results = await execute_task( + cls, + VIDU_TEXT_TO_VIDEO, + TaskCreationRequest( + model=model, + prompt=prompt, + duration=duration, + seed=seed, + aspect_ratio=aspect_ratio, + resolution=resolution, + bgm=background_music, + ), + ) + return IO.NodeOutput(await download_url_to_video_output(results[0].url)) + + +class Vidu2ImageToVideoNode(IO.ComfyNode): + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="Vidu2ImageToVideoNode", + display_name="Vidu2 Image-to-Video Generation", + category="api node/video/Vidu", + description="Generate a video from an image and an optional prompt.", + inputs=[ + IO.Combo.Input("model", options=["viduq2-pro-fast", "viduq2-pro", "viduq2-turbo"]), + IO.Image.Input( + "image", + tooltip="An image to be used as the start frame of the generated video.", + ), + IO.String.Input( + "prompt", + multiline=True, + default="", + tooltip="An optional text prompt for video generation (max 2000 characters).", + ), + IO.Int.Input( + "duration", + default=5, + min=1, + max=10, + step=1, + display_mode=IO.NumberDisplay.slider, + ), + IO.Int.Input( + "seed", + default=1, + min=0, + max=2147483647, + step=1, + display_mode=IO.NumberDisplay.number, + control_after_generate=True, + ), + IO.Combo.Input( + "resolution", + options=["720p", "1080p"], + ), + IO.Combo.Input( + "movement_amplitude", + options=["auto", "small", "medium", "large"], + tooltip="The movement amplitude of objects in the frame.", + ), + ], + outputs=[ + IO.Video.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + ) + + @classmethod + async def execute( + cls, + model: str, + image: Input.Image, + prompt: str, + duration: int, + seed: int, + resolution: str, + movement_amplitude: str, + ) -> IO.NodeOutput: + if get_number_of_images(image) > 1: + raise ValueError("Only one input image is allowed.") + validate_image_aspect_ratio(image, (1, 4), (4, 1)) + validate_string(prompt, max_length=2000) + results = await execute_task( + cls, + VIDU_IMAGE_TO_VIDEO, + TaskCreationRequest( + model=model, + prompt=prompt, + duration=duration, + seed=seed, + resolution=resolution, + movement_amplitude=movement_amplitude, + images=await upload_images_to_comfyapi( + cls, + image, + max_images=1, + mime_type="image/png", + ), + ), + ) + return IO.NodeOutput(await download_url_to_video_output(results[0].url)) + + +class Vidu2ReferenceVideoNode(IO.ComfyNode): + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="Vidu2ReferenceVideoNode", + display_name="Vidu2 Reference-to-Video Generation", + category="api node/video/Vidu", + description="Generate a video from multiple reference images and a prompt.", + inputs=[ + IO.Combo.Input("model", options=["viduq2"]), + IO.Autogrow.Input( + "subjects", + template=IO.Autogrow.TemplateNames( + IO.Image.Input("reference_images"), + names=["subject1", "subject2", "subject3"], + min=1, + ), + tooltip="For each subject, provide up to 3 reference images (7 images total across all subjects). " + "Reference them in prompts via @subject{subject_id}.", + ), + IO.String.Input( + "prompt", + multiline=True, + tooltip="When enabled, the video will include generated speech and background music " + "based on the prompt.", + ), + IO.Boolean.Input( + "audio", + default=False, + tooltip="When enabled video will contain generated speech and background music based on the prompt.", + ), + IO.Int.Input( + "duration", + default=5, + min=1, + max=10, + step=1, + display_mode=IO.NumberDisplay.slider, + ), + IO.Int.Input( + "seed", + default=1, + min=0, + max=2147483647, + step=1, + display_mode=IO.NumberDisplay.number, + control_after_generate=True, + ), + IO.Combo.Input("aspect_ratio", options=["16:9", "9:16", "4:3", "3:4", "1:1"]), + IO.Combo.Input("resolution", options=["720p"]), + IO.Combo.Input( + "movement_amplitude", + options=["auto", "small", "medium", "large"], + tooltip="The movement amplitude of objects in the frame.", + ), + ], + outputs=[ + IO.Video.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + ) + + @classmethod + async def execute( + cls, + model: str, + subjects: IO.Autogrow.Type, + prompt: str, + audio: bool, + duration: int, + seed: int, + aspect_ratio: str, + resolution: str, + movement_amplitude: str, + ) -> IO.NodeOutput: + validate_string(prompt, min_length=1, max_length=2000) + total_images = 0 + for i in subjects: + if get_number_of_images(subjects[i]) > 3: + raise ValueError("Maximum number of images per subject is 3.") + for im in subjects[i]: + total_images += 1 + validate_image_aspect_ratio(im, (1, 4), (4, 1)) + validate_image_dimensions(im, min_width=128, min_height=128) + if total_images > 7: + raise ValueError("Too many reference images; the maximum allowed is 7.") + subjects_param: list[SubjectReference] = [] + for i in subjects: + subjects_param.append( + SubjectReference( + id=i, + images=await upload_images_to_comfyapi( + cls, + subjects[i], + max_images=3, + mime_type="image/png", + wait_label=f"Uploading reference images for {i}", + ), + ), + ) + payload = TaskCreationRequest( + model=model, + prompt=prompt, + audio=audio, + duration=duration, + seed=seed, + aspect_ratio=aspect_ratio, + resolution=resolution, + movement_amplitude=movement_amplitude, + subjects=subjects_param, + ) + results = await execute_task(cls, VIDU_REFERENCE_VIDEO, payload) + return IO.NodeOutput(await download_url_to_video_output(results[0].url)) + + +class Vidu2StartEndToVideoNode(IO.ComfyNode): + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="Vidu2StartEndToVideoNode", + display_name="Vidu2 Start/End Frame-to-Video Generation", + category="api node/video/Vidu", + description="Generate a video from a start frame, an end frame, and a prompt.", + inputs=[ + IO.Combo.Input("model", options=["viduq2-pro-fast", "viduq2-pro", "viduq2-turbo"]), + IO.Image.Input("first_frame"), + IO.Image.Input("end_frame"), + IO.String.Input( + "prompt", + multiline=True, + tooltip="Prompt description (max 2000 characters).", + ), + IO.Int.Input( + "duration", + default=5, + min=2, + max=8, + step=1, + display_mode=IO.NumberDisplay.slider, + ), + IO.Int.Input( + "seed", + default=1, + min=0, + max=2147483647, + step=1, + display_mode=IO.NumberDisplay.number, + control_after_generate=True, + ), + IO.Combo.Input("resolution", options=["720p", "1080p"]), + IO.Combo.Input( + "movement_amplitude", + options=["auto", "small", "medium", "large"], + tooltip="The movement amplitude of objects in the frame.", + ), + ], + outputs=[ + IO.Video.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + ) + + @classmethod + async def execute( + cls, + model: str, + first_frame: Input.Image, + end_frame: Input.Image, + prompt: str, + duration: int, + seed: int, + resolution: str, + movement_amplitude: str, + ) -> IO.NodeOutput: + validate_string(prompt, max_length=2000) + if get_number_of_images(first_frame) > 1: + raise ValueError("Only one input image is allowed for `first_frame`.") + if get_number_of_images(end_frame) > 1: + raise ValueError("Only one input image is allowed for `end_frame`.") + validate_images_aspect_ratio_closeness(first_frame, end_frame, min_rel=0.8, max_rel=1.25, strict=False) + payload = TaskCreationRequest( + model=model, + prompt=prompt, + duration=duration, + seed=seed, + resolution=resolution, + movement_amplitude=movement_amplitude, + images=[ + (await upload_images_to_comfyapi(cls, frame, max_images=1, mime_type="image/png"))[0] + for frame in (first_frame, end_frame) + ], + ) + results = await execute_task(cls, VIDU_START_END_VIDEO, payload) + return IO.NodeOutput(await download_url_to_video_output(results[0].url)) class ViduExtension(ComfyExtension): @@ -558,6 +848,10 @@ class ViduExtension(ComfyExtension): ViduImageToVideoNode, ViduReferenceVideoNode, ViduStartEndToVideoNode, + Vidu2TextToVideoNode, + Vidu2ImageToVideoNode, + Vidu2ReferenceVideoNode, + Vidu2StartEndToVideoNode, ] From 153bc524bf9db76d723289f6420f418f63966972 Mon Sep 17 00:00:00 2001 From: ComfyUI Wiki Date: Sat, 10 Jan 2026 14:29:30 +0800 Subject: [PATCH 420/647] chore: update embedded docs to v0.4.0 (#11776) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 7686a5f8a..6c1cd86d2 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,6 +1,6 @@ comfyui-frontend-package==1.36.13 comfyui-workflow-templates==0.7.69 -comfyui-embedded-docs==0.3.1 +comfyui-embedded-docs==0.4.0 torch torchsde torchvision From dc202a2e51bf7a6cd00e606b2d2941bc223f2ad2 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Fri, 9 Jan 2026 23:03:57 -0800 Subject: [PATCH 421/647] Properly save mixed ops. (#11772) --- comfy/ops.py | 26 ++++++++++++------- .../comfy_quant/test_mixed_precision.py | 6 ++--- 2 files changed, 20 insertions(+), 12 deletions(-) diff --git a/comfy/ops.py b/comfy/ops.py index 1cf22f0cc..9c0b54ff4 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -625,21 +625,29 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec missing_keys.remove(key) def state_dict(self, *args, destination=None, prefix="", **kwargs): - sd = super().state_dict(*args, destination=destination, prefix=prefix, **kwargs) - if isinstance(self.weight, QuantizedTensor): - layout_cls = self.weight._layout_cls + if destination is not None: + sd = destination + else: + sd = {} - # Check if it's any FP8 variant (E4M3 or E5M2) - if layout_cls in ("TensorCoreFP8E4M3Layout", "TensorCoreFP8E5M2Layout", "TensorCoreFP8Layout"): - sd["{}weight_scale".format(prefix)] = self.weight._params.scale - elif layout_cls == "TensorCoreNVFP4Layout": - sd["{}weight_scale_2".format(prefix)] = self.weight._params.scale - sd["{}weight_scale".format(prefix)] = self.weight._params.block_scale + if self.bias is not None: + sd["{}bias".format(prefix)] = self.bias + + if isinstance(self.weight, QuantizedTensor): + sd_out = self.weight.state_dict("{}weight".format(prefix)) + for k in sd_out: + sd[k] = sd_out[k] quant_conf = {"format": self.quant_format} if self._full_precision_mm_config: quant_conf["full_precision_matrix_mult"] = True sd["{}comfy_quant".format(prefix)] = torch.tensor(list(json.dumps(quant_conf).encode('utf-8')), dtype=torch.uint8) + + input_scale = getattr(self, 'input_scale', None) + if input_scale is not None: + sd["{}input_scale".format(prefix)] = input_scale + else: + sd["{}weight".format(prefix)] = self.weight return sd def _forward(self, input, weight, bias): diff --git a/tests-unit/comfy_quant/test_mixed_precision.py b/tests-unit/comfy_quant/test_mixed_precision.py index 7b2eac940..7c740491d 100644 --- a/tests-unit/comfy_quant/test_mixed_precision.py +++ b/tests-unit/comfy_quant/test_mixed_precision.py @@ -153,9 +153,9 @@ class TestMixedPrecisionOps(unittest.TestCase): state_dict2 = model.state_dict() # Verify layer1.weight is a QuantizedTensor with scale preserved - self.assertIsInstance(state_dict2["layer1.weight"], QuantizedTensor) - self.assertEqual(state_dict2["layer1.weight"]._params.scale.item(), 3.0) - self.assertEqual(state_dict2["layer1.weight"]._layout_cls, "TensorCoreFP8E4M3Layout") + self.assertTrue(torch.equal(state_dict2["layer1.weight"].view(torch.uint8), fp8_weight.view(torch.uint8))) + self.assertEqual(state_dict2["layer1.weight_scale"].item(), 3.0) + self.assertEqual(model.layer1.weight._layout_cls, "TensorCoreFP8E4M3Layout") # Verify non-quantized layers are standard tensors self.assertNotIsInstance(state_dict2["layer2.weight"], QuantizedTensor) From 6e4b1f9d00306fe14d7ca5adf2c7468d631b23d5 Mon Sep 17 00:00:00 2001 From: DELUXA Date: Sat, 10 Jan 2026 23:51:05 +0200 Subject: [PATCH 422/647] pythorch_attn_by_def_on_gfx1200 (#11793) --- comfy/model_management.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy/model_management.py b/comfy/model_management.py index e5de4a5b5..9d39be7b2 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -368,7 +368,7 @@ try: if any((a in arch) for a in ["gfx90a", "gfx942", "gfx1100", "gfx1101", "gfx1151"]): # TODO: more arches, TODO: gfx950 ENABLE_PYTORCH_ATTENTION = True if rocm_version >= (7, 0): - if any((a in arch) for a in ["gfx1201"]): + if any((a in arch) for a in ["gfx1200", "gfx1201"]): ENABLE_PYTORCH_ATTENTION = True if torch_version_numeric >= (2, 7) and rocm_version >= (6, 4): if any((a in arch) for a in ["gfx1200", "gfx1201", "gfx950"]): # TODO: more arches, "gfx942" gives error on pytorch nightly 2.10 1013 rocm7.0 From cd912963f17c9ae00ec12e1869293edb78720831 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sat, 10 Jan 2026 14:31:31 -0800 Subject: [PATCH 423/647] Fix issue with t5 text encoder in fp4. (#11794) --- comfy/model_detection.py | 2 ++ comfy/sd.py | 4 ++-- 2 files changed, 4 insertions(+), 2 deletions(-) diff --git a/comfy/model_detection.py b/comfy/model_detection.py index 0853b3aec..aff5a50b9 100644 --- a/comfy/model_detection.py +++ b/comfy/model_detection.py @@ -237,6 +237,8 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): else: dit_config["vec_in_dim"] = None + dit_config["num_heads"] = dit_config["hidden_size"] // sum(dit_config["axes_dim"]) + dit_config["depth"] = count_blocks(state_dict_keys, '{}double_blocks.'.format(key_prefix) + '{}.') dit_config["depth_single_blocks"] = count_blocks(state_dict_keys, '{}single_blocks.'.format(key_prefix) + '{}.') if '{}distilled_guidance_layer.0.norms.0.scale'.format(key_prefix) in state_dict_keys or '{}distilled_guidance_layer.norms.0.scale'.format(key_prefix) in state_dict_keys: #Chroma diff --git a/comfy/sd.py b/comfy/sd.py index 5a7221620..b689c0dfc 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -1059,9 +1059,9 @@ def detect_te_model(sd): return TEModel.JINA_CLIP_2 if "encoder.block.23.layer.1.DenseReluDense.wi_1.weight" in sd: weight = sd["encoder.block.23.layer.1.DenseReluDense.wi_1.weight"] - if weight.shape[-1] == 4096: + if weight.shape[0] == 10240: return TEModel.T5_XXL - elif weight.shape[-1] == 2048: + elif weight.shape[0] == 5120: return TEModel.T5_XL if 'encoder.block.23.layer.1.DenseReluDense.wi.weight' in sd: return TEModel.T5_XXL_OLD From 2f642d5d9b48ad7cad13bbdd5f8adcf506f565a7 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sat, 10 Jan 2026 14:40:42 -0800 Subject: [PATCH 424/647] Fix chroma fp8 te being treated as fp16. (#11795) --- comfy/text_encoders/cosmos.py | 2 +- comfy/text_encoders/genmo.py | 2 +- comfy/text_encoders/pixart_t5.py | 2 +- 3 files changed, 3 insertions(+), 3 deletions(-) diff --git a/comfy/text_encoders/cosmos.py b/comfy/text_encoders/cosmos.py index 448381fa9..f4b40ac68 100644 --- a/comfy/text_encoders/cosmos.py +++ b/comfy/text_encoders/cosmos.py @@ -36,7 +36,7 @@ def te(dtype_t5=None, t5_quantization_metadata=None): if t5_quantization_metadata is not None: model_options = model_options.copy() model_options["t5xxl_quantization_metadata"] = t5_quantization_metadata - if dtype is None: + if dtype_t5 is not None: dtype = dtype_t5 super().__init__(device=device, dtype=dtype, model_options=model_options) return CosmosTEModel_ diff --git a/comfy/text_encoders/genmo.py b/comfy/text_encoders/genmo.py index 5daea8135..2d7a3fbce 100644 --- a/comfy/text_encoders/genmo.py +++ b/comfy/text_encoders/genmo.py @@ -32,7 +32,7 @@ def mochi_te(dtype_t5=None, t5_quantization_metadata=None): if t5_quantization_metadata is not None: model_options = model_options.copy() model_options["t5xxl_quantization_metadata"] = t5_quantization_metadata - if dtype is None: + if dtype_t5 is not None: dtype = dtype_t5 super().__init__(device=device, dtype=dtype, model_options=model_options) return MochiTEModel_ diff --git a/comfy/text_encoders/pixart_t5.py b/comfy/text_encoders/pixart_t5.py index e5e5f18be..51c6e50c7 100644 --- a/comfy/text_encoders/pixart_t5.py +++ b/comfy/text_encoders/pixart_t5.py @@ -36,7 +36,7 @@ def pixart_te(dtype_t5=None, t5_quantization_metadata=None): if t5_quantization_metadata is not None: model_options = model_options.copy() model_options["t5xxl_quantization_metadata"] = t5_quantization_metadata - if dtype is None: + if dtype_t5 is not None: dtype = dtype_t5 super().__init__(device=device, dtype=dtype, model_options=model_options) return PixArtTEModel_ From 5cd1113236b0fb032a51bf9d63ba196a2510b0d4 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Sun, 11 Jan 2026 13:07:11 +0200 Subject: [PATCH 425/647] fix(api-nodes): use a unique name for uploading audio files (#11778) --- comfy_api_nodes/nodes_kling.py | 2 +- comfy_api_nodes/util/conversions.py | 4 ++-- comfy_api_nodes/util/upload_helpers.py | 3 +-- 3 files changed, 4 insertions(+), 5 deletions(-) diff --git a/comfy_api_nodes/nodes_kling.py b/comfy_api_nodes/nodes_kling.py index 9c707a339..01d9c34f5 100644 --- a/comfy_api_nodes/nodes_kling.py +++ b/comfy_api_nodes/nodes_kling.py @@ -567,7 +567,7 @@ async def execute_lipsync( # Upload the audio file to Comfy API and get download URL if audio: audio_url = await upload_audio_to_comfyapi( - cls, audio, container_format="mp3", codec_name="libmp3lame", mime_type="audio/mpeg", filename="output.mp3" + cls, audio, container_format="mp3", codec_name="libmp3lame", mime_type="audio/mpeg" ) logging.info("Uploaded audio to Comfy API. URL: %s", audio_url) else: diff --git a/comfy_api_nodes/util/conversions.py b/comfy_api_nodes/util/conversions.py index d64239c86..99c302a2a 100644 --- a/comfy_api_nodes/util/conversions.py +++ b/comfy_api_nodes/util/conversions.py @@ -55,7 +55,7 @@ def image_tensor_pair_to_batch(image1: torch.Tensor, image2: torch.Tensor) -> to def tensor_to_bytesio( image: torch.Tensor, - name: str | None = None, + *, total_pixels: int = 2048 * 2048, mime_type: str = "image/png", ) -> BytesIO: @@ -75,7 +75,7 @@ def tensor_to_bytesio( pil_image = tensor_to_pil(image, total_pixels=total_pixels) img_binary = pil_to_bytesio(pil_image, mime_type=mime_type) - img_binary.name = f"{name if name else uuid.uuid4()}.{mimetype_to_extension(mime_type)}" + img_binary.name = f"{uuid.uuid4()}.{mimetype_to_extension(mime_type)}" return img_binary diff --git a/comfy_api_nodes/util/upload_helpers.py b/comfy_api_nodes/util/upload_helpers.py index 2535a0884..cea0d1203 100644 --- a/comfy_api_nodes/util/upload_helpers.py +++ b/comfy_api_nodes/util/upload_helpers.py @@ -82,7 +82,6 @@ async def upload_audio_to_comfyapi( container_format: str = "mp4", codec_name: str = "aac", mime_type: str = "audio/mp4", - filename: str = "uploaded_audio.mp4", ) -> str: """ Uploads a single audio input to ComfyUI API and returns its download URL. @@ -92,7 +91,7 @@ async def upload_audio_to_comfyapi( waveform: torch.Tensor = audio["waveform"] audio_data_np = audio_tensor_to_contiguous_ndarray(waveform) audio_bytes_io = audio_ndarray_to_bytesio(audio_data_np, sample_rate, container_format, codec_name) - return await upload_file_to_comfyapi(cls, audio_bytes_io, filename, mime_type) + return await upload_file_to_comfyapi(cls, audio_bytes_io, f"{uuid.uuid4()}.{container_format}", mime_type) async def upload_video_to_comfyapi( From c6238047ee1ffd87eade7c3ab5a8e53c11d4ce39 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sun, 11 Jan 2026 18:11:53 -0800 Subject: [PATCH 426/647] Put more details about portable in readme. (#11816) --- README.md | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/README.md b/README.md index 6d09758c0..e25f3cda7 100644 --- a/README.md +++ b/README.md @@ -183,7 +183,7 @@ Simply download, extract with [7-Zip](https://7-zip.org) or with the windows exp If you have trouble extracting it, right click the file -> properties -> unblock -Update your Nvidia drivers if it doesn't start. +The portable above currently comes with python 3.13 and pytorch cuda 13.0. Update your Nvidia drivers if it doesn't start. #### Alternative Downloads: @@ -212,7 +212,7 @@ Python 3.14 works but you may encounter issues with the torch compile node. The Python 3.13 is very well supported. If you have trouble with some custom node dependencies on 3.13 you can try 3.12 -torch 2.4 and above is supported but some features might only work on newer versions. We generally recommend using the latest major version of pytorch unless it is less than 2 weeks old. +torch 2.4 and above is supported but some features might only work on newer versions. We generally recommend using the latest major version of pytorch with the latest cuda version unless it is less than 2 weeks old. ### Instructions: From a3b5d4996abcd906c7c99f15b69fde051afcb4be Mon Sep 17 00:00:00 2001 From: kelseyee <971704395@qq.com> Date: Tue, 13 Jan 2026 04:38:46 +0800 Subject: [PATCH 427/647] Support ModelScope-Trainer DiffSynth lora for Z Image. (#11805) --- comfy/lora.py | 1 + 1 file changed, 1 insertion(+) diff --git a/comfy/lora.py b/comfy/lora.py index 2ed0acb9d..e8246bd66 100644 --- a/comfy/lora.py +++ b/comfy/lora.py @@ -322,6 +322,7 @@ def model_lora_keys_unet(model, key_map={}): key_map["diffusion_model.{}".format(key_lora)] = to key_map["transformer.{}".format(key_lora)] = to key_map["lycoris_{}".format(key_lora.replace(".", "_"))] = to + key_map[key_lora] = to if isinstance(model, comfy.model_base.Kandinsky5): for k in sdk: From c881a1d6897d8fee84559a8e3e49b9116efdb959 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 12 Jan 2026 14:05:54 -0800 Subject: [PATCH 428/647] Support the siglip 2 naflex model as a clip vision model. (#11831) Not useful yet. --- comfy/clip_model.py | 62 +++++++++++++++++++++- comfy/clip_vision.py | 25 ++++++--- comfy/clip_vision_siglip2_base_naflex.json | 14 +++++ 3 files changed, 91 insertions(+), 10 deletions(-) create mode 100644 comfy/clip_vision_siglip2_base_naflex.json diff --git a/comfy/clip_model.py b/comfy/clip_model.py index e88872728..d7d3f994c 100644 --- a/comfy/clip_model.py +++ b/comfy/clip_model.py @@ -1,6 +1,7 @@ import torch from comfy.ldm.modules.attention import optimized_attention_for_device import comfy.ops +import math def clip_preprocess(image, size=224, mean=[0.48145466, 0.4578275, 0.40821073], std=[0.26862954, 0.26130258, 0.27577711], crop=True): image = image[:, :, :, :3] if image.shape[3] > 3 else image @@ -21,6 +22,39 @@ def clip_preprocess(image, size=224, mean=[0.48145466, 0.4578275, 0.40821073], s image = torch.clip((255. * image), 0, 255).round() / 255.0 return (image - mean.view([3,1,1])) / std.view([3,1,1]) +def siglip2_flex_calc_resolution(oh, ow, patch_size, max_num_patches, eps=1e-5): + def scale_dim(size, scale): + scaled = math.ceil(size * scale / patch_size) * patch_size + return max(patch_size, int(scaled)) + + # Binary search for optimal scale + lo, hi = eps / 10, 100.0 + while hi - lo >= eps: + mid = (lo + hi) / 2 + h, w = scale_dim(oh, mid), scale_dim(ow, mid) + if (h // patch_size) * (w // patch_size) <= max_num_patches: + lo = mid + else: + hi = mid + + return scale_dim(oh, lo), scale_dim(ow, lo) + +def siglip2_preprocess(image, size, patch_size, num_patches, mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5], crop=True): + if size > 0: + return clip_preprocess(image, size=size, mean=mean, std=std, crop=crop) + + image = image[:, :, :, :3] if image.shape[3] > 3 else image + mean = torch.tensor(mean, device=image.device, dtype=image.dtype) + std = torch.tensor(std, device=image.device, dtype=image.dtype) + image = image.movedim(-1, 1) + + b, c, h, w = image.shape + h, w = siglip2_flex_calc_resolution(h, w, patch_size, num_patches) + + image = torch.nn.functional.interpolate(image, size=(h, w), mode="bilinear", antialias=True) + image = torch.clip((255. * image), 0, 255).round() / 255.0 + return (image - mean.view([3, 1, 1])) / std.view([3, 1, 1]) + class CLIPAttention(torch.nn.Module): def __init__(self, embed_dim, heads, dtype, device, operations): super().__init__() @@ -175,6 +209,27 @@ class CLIPTextModel(torch.nn.Module): out = self.text_projection(x[2]) return (x[0], x[1], out, x[2]) +def siglip2_pos_embed(embed_weight, embeds, orig_shape): + embed_weight_len = round(embed_weight.shape[0] ** 0.5) + embed_weight = comfy.ops.cast_to_input(embed_weight, embeds).movedim(1, 0).reshape(1, -1, embed_weight_len, embed_weight_len) + embed_weight = torch.nn.functional.interpolate(embed_weight, size=orig_shape, mode="bilinear", align_corners=False, antialias=True) + embed_weight = embed_weight.reshape(-1, embed_weight.shape[-2] * embed_weight.shape[-1]).movedim(0, 1) + return embeds + embed_weight + +class Siglip2Embeddings(torch.nn.Module): + def __init__(self, embed_dim, num_channels=3, patch_size=14, image_size=224, model_type="", num_patches=None, dtype=None, device=None, operations=None): + super().__init__() + self.patch_embedding = operations.Linear(num_channels * patch_size * patch_size, embed_dim, dtype=dtype, device=device) + self.position_embedding = operations.Embedding(num_patches, embed_dim, dtype=dtype, device=device) + self.patch_size = patch_size + + def forward(self, pixel_values): + b, c, h, w = pixel_values.shape + img = pixel_values.movedim(1, -1).reshape(b, h // self.patch_size, self.patch_size, w // self.patch_size, self.patch_size, c) + img = img.permute(0, 1, 3, 2, 4, 5) + img = img.reshape(b, img.shape[1] * img.shape[2], -1) + img = self.patch_embedding(img) + return siglip2_pos_embed(self.position_embedding.weight, img, (h // self.patch_size, w // self.patch_size)) class CLIPVisionEmbeddings(torch.nn.Module): def __init__(self, embed_dim, num_channels=3, patch_size=14, image_size=224, model_type="", dtype=None, device=None, operations=None): @@ -218,8 +273,11 @@ class CLIPVision(torch.nn.Module): intermediate_activation = config_dict["hidden_act"] model_type = config_dict["model_type"] - self.embeddings = CLIPVisionEmbeddings(embed_dim, config_dict["num_channels"], config_dict["patch_size"], config_dict["image_size"], model_type=model_type, dtype=dtype, device=device, operations=operations) - if model_type == "siglip_vision_model": + if model_type in ["siglip2_vision_model"]: + self.embeddings = Siglip2Embeddings(embed_dim, config_dict["num_channels"], config_dict["patch_size"], config_dict["image_size"], model_type=model_type, num_patches=config_dict.get("num_patches", None), dtype=dtype, device=device, operations=operations) + else: + self.embeddings = CLIPVisionEmbeddings(embed_dim, config_dict["num_channels"], config_dict["patch_size"], config_dict["image_size"], model_type=model_type, dtype=dtype, device=device, operations=operations) + if model_type in ["siglip_vision_model", "siglip2_vision_model"]: self.pre_layrnorm = lambda a: a self.output_layernorm = True else: diff --git a/comfy/clip_vision.py b/comfy/clip_vision.py index d5fc53497..66f2a9d9c 100644 --- a/comfy/clip_vision.py +++ b/comfy/clip_vision.py @@ -21,6 +21,7 @@ clip_preprocess = comfy.clip_model.clip_preprocess # Prevent some stuff from br IMAGE_ENCODERS = { "clip_vision_model": comfy.clip_model.CLIPVisionModelProjection, "siglip_vision_model": comfy.clip_model.CLIPVisionModelProjection, + "siglip2_vision_model": comfy.clip_model.CLIPVisionModelProjection, "dinov2": comfy.image_encoders.dino2.Dinov2Model, } @@ -32,9 +33,10 @@ class ClipVisionModel(): self.image_size = config.get("image_size", 224) self.image_mean = config.get("image_mean", [0.48145466, 0.4578275, 0.40821073]) self.image_std = config.get("image_std", [0.26862954, 0.26130258, 0.27577711]) - model_type = config.get("model_type", "clip_vision_model") - model_class = IMAGE_ENCODERS.get(model_type) - if model_type == "siglip_vision_model": + self.model_type = config.get("model_type", "clip_vision_model") + self.config = config.copy() + model_class = IMAGE_ENCODERS.get(self.model_type) + if self.model_type == "siglip_vision_model": self.return_all_hidden_states = True else: self.return_all_hidden_states = False @@ -55,7 +57,10 @@ class ClipVisionModel(): def encode_image(self, image, crop=True): comfy.model_management.load_model_gpu(self.patcher) - pixel_values = comfy.clip_model.clip_preprocess(image.to(self.load_device), size=self.image_size, mean=self.image_mean, std=self.image_std, crop=crop).float() + if self.model_type == "siglip2_vision_model": + pixel_values = comfy.clip_model.siglip2_preprocess(image.to(self.load_device), size=self.image_size, patch_size=self.config.get("patch_size", 16), num_patches=self.config.get("num_patches", 256), mean=self.image_mean, std=self.image_std, crop=crop).float() + else: + pixel_values = comfy.clip_model.clip_preprocess(image.to(self.load_device), size=self.image_size, mean=self.image_mean, std=self.image_std, crop=crop).float() out = self.model(pixel_values=pixel_values, intermediate_output='all' if self.return_all_hidden_states else -2) outputs = Output() @@ -107,10 +112,14 @@ def load_clipvision_from_sd(sd, prefix="", convert_keys=False): elif "vision_model.encoder.layers.22.layer_norm1.weight" in sd: embed_shape = sd["vision_model.embeddings.position_embedding.weight"].shape[0] if sd["vision_model.encoder.layers.0.layer_norm1.weight"].shape[0] == 1152: - if embed_shape == 729: - json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_siglip_384.json") - elif embed_shape == 1024: - json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_siglip_512.json") + patch_embedding_shape = sd["vision_model.embeddings.patch_embedding.weight"].shape + if len(patch_embedding_shape) == 2: + json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_siglip2_base_naflex.json") + else: + if embed_shape == 729: + json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_siglip_384.json") + elif embed_shape == 1024: + json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_siglip_512.json") elif embed_shape == 577: if "multi_modal_projector.linear_1.bias" in sd: json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_config_vitl_336_llava.json") diff --git a/comfy/clip_vision_siglip2_base_naflex.json b/comfy/clip_vision_siglip2_base_naflex.json new file mode 100644 index 000000000..6f6b99bd6 --- /dev/null +++ b/comfy/clip_vision_siglip2_base_naflex.json @@ -0,0 +1,14 @@ +{ + "num_channels": 3, + "hidden_act": "gelu_pytorch_tanh", + "hidden_size": 1152, + "image_size": -1, + "intermediate_size": 4304, + "model_type": "siglip2_vision_model", + "num_attention_heads": 16, + "num_hidden_layers": 27, + "patch_size": 16, + "num_patches": 256, + "image_mean": [0.5, 0.5, 0.5], + "image_std": [0.5, 0.5, 0.5] +} From fd5c0755af18530ff1225f5946c7a647b9694032 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Jukka=20Sepp=C3=A4nen?= <40791699+kijai@users.noreply.github.com> Date: Tue, 13 Jan 2026 00:28:59 +0200 Subject: [PATCH 429/647] Reduce LTX2 VRAM use by more efficient timestep embed handling (#11829) --- comfy/ldm/lightricks/av_model.py | 136 ++++++++++++++++++++++++------- 1 file changed, 106 insertions(+), 30 deletions(-) diff --git a/comfy/ldm/lightricks/av_model.py b/comfy/ldm/lightricks/av_model.py index 759535501..c12ace241 100644 --- a/comfy/ldm/lightricks/av_model.py +++ b/comfy/ldm/lightricks/av_model.py @@ -11,6 +11,69 @@ from comfy.ldm.lightricks.model import ( from comfy.ldm.lightricks.symmetric_patchifier import AudioPatchifier import comfy.ldm.common_dit +class CompressedTimestep: + """Store video timestep embeddings in compressed form using per-frame indexing.""" + __slots__ = ('data', 'batch_size', 'num_frames', 'patches_per_frame', 'feature_dim') + + def __init__(self, tensor: torch.Tensor, patches_per_frame: int): + """ + tensor: [batch_size, num_tokens, feature_dim] tensor where num_tokens = num_frames * patches_per_frame + patches_per_frame: Number of spatial patches per frame (height * width in latent space) + """ + self.batch_size, num_tokens, self.feature_dim = tensor.shape + + # Check if compression is valid (num_tokens must be divisible by patches_per_frame) + if num_tokens % patches_per_frame == 0 and num_tokens >= patches_per_frame: + self.patches_per_frame = patches_per_frame + self.num_frames = num_tokens // patches_per_frame + + # Reshape to [batch, frames, patches_per_frame, feature_dim] and store one value per frame + # All patches in a frame are identical, so we only keep the first one + reshaped = tensor.view(self.batch_size, self.num_frames, patches_per_frame, self.feature_dim) + self.data = reshaped[:, :, 0, :].contiguous() # [batch, frames, feature_dim] + else: + # Not divisible or too small - store directly without compression + self.patches_per_frame = 1 + self.num_frames = num_tokens + self.data = tensor + + def expand(self): + """Expand back to original tensor.""" + if self.patches_per_frame == 1: + return self.data + + # [batch, frames, feature_dim] -> [batch, frames, patches_per_frame, feature_dim] -> [batch, tokens, feature_dim] + expanded = self.data.unsqueeze(2).expand(self.batch_size, self.num_frames, self.patches_per_frame, self.feature_dim) + return expanded.reshape(self.batch_size, -1, self.feature_dim) + + def expand_for_computation(self, scale_shift_table: torch.Tensor, batch_size: int, indices: slice = slice(None, None)): + """Compute ada values on compressed per-frame data, then expand spatially.""" + num_ada_params = scale_shift_table.shape[0] + + # No compression - compute directly + if self.patches_per_frame == 1: + num_tokens = self.data.shape[1] + dim_per_param = self.feature_dim // num_ada_params + reshaped = self.data.reshape(batch_size, num_tokens, num_ada_params, dim_per_param)[:, :, indices, :] + table_values = scale_shift_table[indices].unsqueeze(0).unsqueeze(0).to(device=self.data.device, dtype=self.data.dtype) + ada_values = (table_values + reshaped).unbind(dim=2) + return ada_values + + # Compressed: compute on per-frame data then expand spatially + # Reshape: [batch, frames, feature_dim] -> [batch, frames, num_ada_params, dim_per_param] + frame_reshaped = self.data.reshape(batch_size, self.num_frames, num_ada_params, -1)[:, :, indices, :] + table_values = scale_shift_table[indices].unsqueeze(0).unsqueeze(0).to( + device=self.data.device, dtype=self.data.dtype + ) + frame_ada = (table_values + frame_reshaped).unbind(dim=2) + + # Expand each ada parameter spatially: [batch, frames, dim] -> [batch, frames, patches, dim] -> [batch, tokens, dim] + return tuple( + frame_val.unsqueeze(2).expand(batch_size, self.num_frames, self.patches_per_frame, -1) + .reshape(batch_size, -1, frame_val.shape[-1]) + for frame_val in frame_ada + ) + class BasicAVTransformerBlock(nn.Module): def __init__( self, @@ -119,6 +182,9 @@ class BasicAVTransformerBlock(nn.Module): def get_ada_values( self, scale_shift_table: torch.Tensor, batch_size: int, timestep: torch.Tensor, indices: slice = slice(None, None) ): + if isinstance(timestep, CompressedTimestep): + return timestep.expand_for_computation(scale_shift_table, batch_size, indices) + num_ada_params = scale_shift_table.shape[0] ada_values = ( @@ -146,10 +212,7 @@ class BasicAVTransformerBlock(nn.Module): gate_timestep, ) - scale_shift_chunks = [t.squeeze(2) for t in scale_shift_ada_values] - gate_ada_values = [t.squeeze(2) for t in gate_ada_values] - - return (*scale_shift_chunks, *gate_ada_values) + return (*scale_shift_ada_values, *gate_ada_values) def forward( self, @@ -543,72 +606,80 @@ class LTXAVModel(LTXVModel): if grid_mask is not None: timestep = timestep[:, grid_mask] - timestep = timestep * self.timestep_scale_multiplier + timestep_scaled = timestep * self.timestep_scale_multiplier + v_timestep, v_embedded_timestep = self.adaln_single( - timestep.flatten(), + timestep_scaled.flatten(), {"resolution": None, "aspect_ratio": None}, batch_size=batch_size, hidden_dtype=hidden_dtype, ) - # Second dimension is 1 or number of tokens (if timestep_per_token) - v_timestep = v_timestep.view(batch_size, -1, v_timestep.shape[-1]) - v_embedded_timestep = v_embedded_timestep.view( - batch_size, -1, v_embedded_timestep.shape[-1] - ) + # Calculate patches_per_frame from orig_shape: [batch, channels, frames, height, width] + # Video tokens are arranged as (frames * height * width), so patches_per_frame = height * width + orig_shape = kwargs.get("orig_shape") + v_patches_per_frame = None + if orig_shape is not None and len(orig_shape) == 5: + # orig_shape[3] = height, orig_shape[4] = width (in latent space) + v_patches_per_frame = orig_shape[3] * orig_shape[4] + + # Reshape to [batch_size, num_tokens, dim] and compress for storage + v_timestep = CompressedTimestep(v_timestep.view(batch_size, -1, v_timestep.shape[-1]), v_patches_per_frame) + v_embedded_timestep = CompressedTimestep(v_embedded_timestep.view(batch_size, -1, v_embedded_timestep.shape[-1]), v_patches_per_frame) # Prepare audio timestep a_timestep = kwargs.get("a_timestep") if a_timestep is not None: - a_timestep = a_timestep * self.timestep_scale_multiplier + a_timestep_scaled = a_timestep * self.timestep_scale_multiplier + a_timestep_flat = a_timestep_scaled.flatten() + timestep_flat = timestep_scaled.flatten() av_ca_factor = self.av_ca_timestep_scale_multiplier / self.timestep_scale_multiplier + # Cross-attention timesteps - compress these too av_ca_audio_scale_shift_timestep, _ = self.av_ca_audio_scale_shift_adaln_single( - a_timestep.flatten(), + a_timestep_flat, {"resolution": None, "aspect_ratio": None}, batch_size=batch_size, hidden_dtype=hidden_dtype, ) av_ca_video_scale_shift_timestep, _ = self.av_ca_video_scale_shift_adaln_single( - timestep.flatten(), + timestep_flat, {"resolution": None, "aspect_ratio": None}, batch_size=batch_size, hidden_dtype=hidden_dtype, ) av_ca_a2v_gate_noise_timestep, _ = self.av_ca_a2v_gate_adaln_single( - timestep.flatten() * av_ca_factor, + timestep_flat * av_ca_factor, {"resolution": None, "aspect_ratio": None}, batch_size=batch_size, hidden_dtype=hidden_dtype, ) av_ca_v2a_gate_noise_timestep, _ = self.av_ca_v2a_gate_adaln_single( - a_timestep.flatten() * av_ca_factor, + a_timestep_flat * av_ca_factor, {"resolution": None, "aspect_ratio": None}, batch_size=batch_size, hidden_dtype=hidden_dtype, ) + # Compress cross-attention timesteps (only video side, audio is too small to benefit) + cross_av_timestep_ss = [ + av_ca_audio_scale_shift_timestep.view(batch_size, -1, av_ca_audio_scale_shift_timestep.shape[-1]), + CompressedTimestep(av_ca_video_scale_shift_timestep.view(batch_size, -1, av_ca_video_scale_shift_timestep.shape[-1]), v_patches_per_frame), # video - compressed + CompressedTimestep(av_ca_a2v_gate_noise_timestep.view(batch_size, -1, av_ca_a2v_gate_noise_timestep.shape[-1]), v_patches_per_frame), # video - compressed + av_ca_v2a_gate_noise_timestep.view(batch_size, -1, av_ca_v2a_gate_noise_timestep.shape[-1]), + ] + a_timestep, a_embedded_timestep = self.audio_adaln_single( - a_timestep.flatten(), + a_timestep_flat, {"resolution": None, "aspect_ratio": None}, batch_size=batch_size, hidden_dtype=hidden_dtype, ) + # Audio timesteps a_timestep = a_timestep.view(batch_size, -1, a_timestep.shape[-1]) - a_embedded_timestep = a_embedded_timestep.view( - batch_size, -1, a_embedded_timestep.shape[-1] - ) - cross_av_timestep_ss = [ - av_ca_audio_scale_shift_timestep, - av_ca_video_scale_shift_timestep, - av_ca_a2v_gate_noise_timestep, - av_ca_v2a_gate_noise_timestep, - ] - cross_av_timestep_ss = list( - [t.view(batch_size, -1, t.shape[-1]) for t in cross_av_timestep_ss] - ) + a_embedded_timestep = a_embedded_timestep.view(batch_size, -1, a_embedded_timestep.shape[-1]) else: - a_timestep = timestep + a_timestep = timestep_scaled a_embedded_timestep = kwargs.get("embedded_timestep") cross_av_timestep_ss = [] @@ -767,6 +838,11 @@ class LTXAVModel(LTXVModel): ax = x[1] v_embedded_timestep = embedded_timestep[0] a_embedded_timestep = embedded_timestep[1] + + # Expand compressed video timestep if needed + if isinstance(v_embedded_timestep, CompressedTimestep): + v_embedded_timestep = v_embedded_timestep.expand() + vx = super()._process_output(vx, v_embedded_timestep, keyframe_idxs, **kwargs) # Process audio output From c2b65e2fceea821276c143ad52478552633922bf Mon Sep 17 00:00:00 2001 From: ComfyUI Wiki Date: Tue, 13 Jan 2026 06:29:25 +0800 Subject: [PATCH 430/647] Update workflow templates to v0.8.0 (#11828) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 6c1cd86d2..890070d5d 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,5 +1,5 @@ comfyui-frontend-package==1.36.13 -comfyui-workflow-templates==0.7.69 +comfyui-workflow-templates==0.8.0 comfyui-embedded-docs==0.4.0 torch torchsde From ecaeeb990d7a5c3820b6f2373d04335d051d6b47 Mon Sep 17 00:00:00 2001 From: ComfyUI Wiki Date: Tue, 13 Jan 2026 11:18:01 +0800 Subject: [PATCH 431/647] chore: update workflow templates to v0.8.4 (#11835) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 890070d5d..077c8930a 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,5 +1,5 @@ comfyui-frontend-package==1.36.13 -comfyui-workflow-templates==0.8.0 +comfyui-workflow-templates==0.8.4 comfyui-embedded-docs==0.4.0 torch torchsde From b3c0e4de57bfd27e3dd94bd9723bb4c714668a09 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 12 Jan 2026 19:33:54 -0800 Subject: [PATCH 432/647] Make loras work on nvfp4 models. (#11837) The initial applying is a bit slow but will probably be sped up in the future. --- comfy/float.py | 113 +++++++++++++++++++++++++++++++++++++++++++++ comfy/ops.py | 2 +- comfy/quant_ops.py | 37 ++++++++++++++- requirements.txt | 2 +- 4 files changed, 150 insertions(+), 4 deletions(-) diff --git a/comfy/float.py b/comfy/float.py index 521316fd2..e638b1ff7 100644 --- a/comfy/float.py +++ b/comfy/float.py @@ -65,3 +65,116 @@ def stochastic_rounding(value, dtype, seed=0): return output return value.to(dtype=dtype) + + +# TODO: improve this? +def stochastic_float_to_fp4_e2m1(x, generator): + sign = torch.signbit(x).to(torch.uint8) + x_abs = x.abs() + + exp = torch.floor(torch.log2(x_abs) + 1.0).clamp(0, 3) + x += (torch.rand(x.size(), dtype=x.dtype, layout=x.layout, device=x.device, generator=generator) - 0.5) * (2 ** (exp - 2.0)) * 1.25 + + x_abs = x.abs() + exp = torch.floor(torch.log2(x_abs) + 1.1925).clamp(0, 3) + + mantissa = torch.where( + exp > 0, + (x_abs / (2.0 ** (exp - 1)) - 1.0) * 2.0, + (x_abs * 2.0) + ).round().to(torch.uint8) + + fp4 = (sign << 3) | (exp.to(torch.uint8) << 1) | mantissa + + fp4_flat = fp4.view(-1) + packed = (fp4_flat[0::2] << 4) | fp4_flat[1::2] + return packed.reshape(list(x.shape)[:-1] + [-1]) + + +def to_blocked(input_matrix, flatten: bool = True) -> torch.Tensor: + """ + Rearrange a large matrix by breaking it into blocks and applying the rearrangement pattern. + See: + https://docs.nvidia.com/cuda/cublas/index.html#d-block-scaling-factors-layout + + Args: + input_matrix: Input tensor of shape (H, W) + Returns: + Rearranged tensor of shape (32*ceil_div(H,128), 16*ceil_div(W,4)) + """ + + def ceil_div(a, b): + return (a + b - 1) // b + + rows, cols = input_matrix.shape + n_row_blocks = ceil_div(rows, 128) + n_col_blocks = ceil_div(cols, 4) + + # Calculate the padded shape + padded_rows = n_row_blocks * 128 + padded_cols = n_col_blocks * 4 + + padded = input_matrix + if (rows, cols) != (padded_rows, padded_cols): + padded = torch.zeros( + (padded_rows, padded_cols), + device=input_matrix.device, + dtype=input_matrix.dtype, + ) + padded[:rows, :cols] = input_matrix + + # Rearrange the blocks + blocks = padded.view(n_row_blocks, 128, n_col_blocks, 4).permute(0, 2, 1, 3) + rearranged = blocks.reshape(-1, 4, 32, 4).transpose(1, 2).reshape(-1, 32, 16) + if flatten: + return rearranged.flatten() + + return rearranged.reshape(padded_rows, padded_cols) + + +def stochastic_round_quantize_nvfp4(x, per_tensor_scale, pad_16x, seed=0): + F4_E2M1_MAX = 6.0 + F8_E4M3_MAX = 448.0 + + def roundup(x: int, multiple: int) -> int: + """Round up x to the nearest multiple.""" + return ((x + multiple - 1) // multiple) * multiple + + orig_shape = x.shape + + # Handle padding + if pad_16x: + rows, cols = x.shape + padded_rows = roundup(rows, 16) + padded_cols = roundup(cols, 16) + if padded_rows != rows or padded_cols != cols: + x = torch.nn.functional.pad(x, (0, padded_cols - cols, 0, padded_rows - rows)) + # Note: We update orig_shape because the output tensor logic below assumes x.shape matches + # what we want to produce. If we pad here, we want the padded output. + orig_shape = x.shape + + block_size = 16 + + x = x.reshape(orig_shape[0], -1, block_size) + max_abs = torch.amax(torch.abs(x), dim=-1) + block_scale = max_abs / F4_E2M1_MAX + scaled_block_scales = block_scale / per_tensor_scale.to(block_scale.dtype) + scaled_block_scales_fp8 = torch.clamp(scaled_block_scales, max=F8_E4M3_MAX).to(torch.float8_e4m3fn) + total_scale = per_tensor_scale.to(x.dtype) * scaled_block_scales_fp8.to(x.dtype) + + # Handle zero blocks (from padding): avoid 0/0 NaN + zero_scale_mask = (total_scale == 0) + total_scale_safe = torch.where(zero_scale_mask, torch.ones_like(total_scale), total_scale) + + x = x / total_scale_safe.unsqueeze(-1) + + generator = torch.Generator(device=x.device) + generator.manual_seed(seed) + + x = torch.where(zero_scale_mask.unsqueeze(-1), torch.zeros_like(x), x) + + x = x.view(orig_shape) + data_lp = stochastic_float_to_fp4_e2m1(x, generator=generator) + + blocked_scales = to_blocked(scaled_block_scales_fp8, flatten=False) + return data_lp, blocked_scales diff --git a/comfy/ops.py b/comfy/ops.py index 9c0b54ff4..415c39e92 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -699,7 +699,7 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec def set_weight(self, weight, inplace_update=False, seed=None, return_weight=False, **kwargs): if getattr(self, 'layout_type', None) is not None: # dtype is now implicit in the layout class - weight = QuantizedTensor.from_float(weight, self.layout_type, scale="recalculate", stochastic_rounding=seed, inplace_ops=True) + weight = QuantizedTensor.from_float(weight, self.layout_type, scale="recalculate", stochastic_rounding=seed, inplace_ops=True).to(self.weight.dtype) else: weight = weight.to(self.weight.dtype) if return_weight: diff --git a/comfy/quant_ops.py b/comfy/quant_ops.py index 8324be42a..7a61203c3 100644 --- a/comfy/quant_ops.py +++ b/comfy/quant_ops.py @@ -7,7 +7,7 @@ try: QuantizedTensor, QuantizedLayout, TensorCoreFP8Layout as _CKFp8Layout, - TensorCoreNVFP4Layout, # Direct import, no wrapper needed + TensorCoreNVFP4Layout as _CKNvfp4Layout, register_layout_op, register_layout_class, get_layout_class, @@ -34,7 +34,7 @@ except ImportError as e: class _CKFp8Layout: pass - class TensorCoreNVFP4Layout: + class _CKNvfp4Layout: pass def register_layout_class(name, cls): @@ -84,6 +84,39 @@ class _TensorCoreFP8LayoutBase(_CKFp8Layout): return qdata, params +class TensorCoreNVFP4Layout(_CKNvfp4Layout): + @classmethod + def quantize(cls, tensor, scale=None, stochastic_rounding=0, inplace_ops=False): + if tensor.dim() != 2: + raise ValueError(f"NVFP4 requires 2D tensor, got {tensor.dim()}D") + + orig_dtype = tensor.dtype + orig_shape = tuple(tensor.shape) + + if scale is None or (isinstance(scale, str) and scale == "recalculate"): + scale = torch.amax(tensor.abs()) / (ck.float_utils.F8_E4M3_MAX * ck.float_utils.F4_E2M1_MAX) + + if not isinstance(scale, torch.Tensor): + scale = torch.tensor(scale) + scale = scale.to(device=tensor.device, dtype=torch.float32) + + padded_shape = cls.get_padded_shape(orig_shape) + needs_padding = padded_shape != orig_shape + + if stochastic_rounding > 0: + qdata, block_scale = comfy.float.stochastic_round_quantize_nvfp4(tensor, scale, pad_16x=needs_padding, seed=stochastic_rounding) + else: + qdata, block_scale = ck.quantize_nvfp4(tensor, scale, pad_16x=needs_padding) + + params = cls.Params( + scale=scale, + orig_dtype=orig_dtype, + orig_shape=orig_shape, + block_scale=block_scale, + ) + return qdata, params + + class TensorCoreFP8E4M3Layout(_TensorCoreFP8LayoutBase): FP8_DTYPE = torch.float8_e4m3fn diff --git a/requirements.txt b/requirements.txt index 077c8930a..43737056e 100644 --- a/requirements.txt +++ b/requirements.txt @@ -21,7 +21,7 @@ psutil alembic SQLAlchemy av>=14.2.0 -comfy-kitchen>=0.2.5 +comfy-kitchen>=0.2.6 #non essential dependencies: kornia>=0.7.1 From 117e7a5853cc34b6a012e06bb3efcc79ab314184 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 12 Jan 2026 21:01:52 -0800 Subject: [PATCH 433/647] Refactor to try to lower mem usage. (#11840) --- comfy/float.py | 21 +++++++++++++-------- 1 file changed, 13 insertions(+), 8 deletions(-) diff --git a/comfy/float.py b/comfy/float.py index e638b1ff7..c806af76b 100644 --- a/comfy/float.py +++ b/comfy/float.py @@ -69,26 +69,31 @@ def stochastic_rounding(value, dtype, seed=0): # TODO: improve this? def stochastic_float_to_fp4_e2m1(x, generator): + orig_shape = x.shape sign = torch.signbit(x).to(torch.uint8) - x_abs = x.abs() - exp = torch.floor(torch.log2(x_abs) + 1.0).clamp(0, 3) + exp = torch.floor(torch.log2(x.abs()) + 1.0).clamp(0, 3) x += (torch.rand(x.size(), dtype=x.dtype, layout=x.layout, device=x.device, generator=generator) - 0.5) * (2 ** (exp - 2.0)) * 1.25 - x_abs = x.abs() - exp = torch.floor(torch.log2(x_abs) + 1.1925).clamp(0, 3) + x = x.abs() + exp = torch.floor(torch.log2(x) + 1.1925).clamp(0, 3) mantissa = torch.where( exp > 0, - (x_abs / (2.0 ** (exp - 1)) - 1.0) * 2.0, - (x_abs * 2.0) + (x / (2.0 ** (exp - 1)) - 1.0) * 2.0, + (x * 2.0), + out=x ).round().to(torch.uint8) + del x - fp4 = (sign << 3) | (exp.to(torch.uint8) << 1) | mantissa + exp = exp.to(torch.uint8) + + fp4 = (sign << 3) | (exp << 1) | mantissa + del sign, exp, mantissa fp4_flat = fp4.view(-1) packed = (fp4_flat[0::2] << 4) | fp4_flat[1::2] - return packed.reshape(list(x.shape)[:-1] + [-1]) + return packed.reshape(list(orig_shape)[:-1] + [-1]) def to_blocked(input_matrix, flatten: bool = True) -> torch.Tensor: From acd0e536533cdf038bbaa32730cd12a438cc3a60 Mon Sep 17 00:00:00 2001 From: Jedrzej Kosinski Date: Mon, 12 Jan 2026 21:15:24 -0800 Subject: [PATCH 434/647] Make bulk_ops not use .returning to be compatible with python 3.10 and 3.11 sqlalchemy (#11839) --- app/assets/database/bulk_ops.py | 30 +++++++++++++++++++++++------- 1 file changed, 23 insertions(+), 7 deletions(-) diff --git a/app/assets/database/bulk_ops.py b/app/assets/database/bulk_ops.py index 9352cd65d..c7b75290a 100644 --- a/app/assets/database/bulk_ops.py +++ b/app/assets/database/bulk_ops.py @@ -92,14 +92,23 @@ def seed_from_paths_batch( session.execute(ins_asset, chunk) # try to claim AssetCacheState (file_path) - winners_by_path: set[str] = set() + # Insert with ON CONFLICT DO NOTHING, then query to find which paths were actually inserted ins_state = ( sqlite.insert(AssetCacheState) .on_conflict_do_nothing(index_elements=[AssetCacheState.file_path]) - .returning(AssetCacheState.file_path) ) for chunk in _iter_chunks(state_rows, _rows_per_stmt(3)): - winners_by_path.update((session.execute(ins_state, chunk)).scalars().all()) + session.execute(ins_state, chunk) + + # Query to find which of our paths won (were actually inserted) + winners_by_path: set[str] = set() + for chunk in _iter_chunks(path_list, MAX_BIND_PARAMS): + result = session.execute( + sqlalchemy.select(AssetCacheState.file_path) + .where(AssetCacheState.file_path.in_(chunk)) + .where(AssetCacheState.asset_id.in_([path_to_asset[p] for p in chunk])) + ) + winners_by_path.update(result.scalars().all()) all_paths_set = set(path_list) losers_by_path = all_paths_set - winners_by_path @@ -112,16 +121,23 @@ def seed_from_paths_batch( return {"inserted_infos": 0, "won_states": 0, "lost_states": len(losers_by_path)} # insert AssetInfo only for winners + # Insert with ON CONFLICT DO NOTHING, then query to find which were actually inserted winner_info_rows = [asset_to_info[path_to_asset[p]] for p in winners_by_path] ins_info = ( sqlite.insert(AssetInfo) .on_conflict_do_nothing(index_elements=[AssetInfo.asset_id, AssetInfo.owner_id, AssetInfo.name]) - .returning(AssetInfo.id) ) - - inserted_info_ids: set[str] = set() for chunk in _iter_chunks(winner_info_rows, _rows_per_stmt(9)): - inserted_info_ids.update((session.execute(ins_info, chunk)).scalars().all()) + session.execute(ins_info, chunk) + + # Query to find which info rows were actually inserted (by matching our generated IDs) + all_info_ids = [row["id"] for row in winner_info_rows] + inserted_info_ids: set[str] = set() + for chunk in _iter_chunks(all_info_ids, MAX_BIND_PARAMS): + result = session.execute( + sqlalchemy.select(AssetInfo.id).where(AssetInfo.id.in_(chunk)) + ) + inserted_info_ids.update(result.scalars().all()) # build and insert tag + meta rows for the AssetInfo tag_rows: list[dict] = [] From 8af13b439bddaddb6d3b4f7c50b6391e88a10c66 Mon Sep 17 00:00:00 2001 From: Christian Byrne Date: Mon, 12 Jan 2026 22:22:25 -0800 Subject: [PATCH 435/647] Update requirements.txt (#11841) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 43737056e..8650d28ec 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,4 +1,4 @@ -comfyui-frontend-package==1.36.13 +comfyui-frontend-package==1.36.14 comfyui-workflow-templates==0.8.4 comfyui-embedded-docs==0.4.0 torch From db9e6edfa1604be0b1f738b2f67495b46cee5a8c Mon Sep 17 00:00:00 2001 From: comfyanonymous Date: Tue, 13 Jan 2026 01:23:31 -0500 Subject: [PATCH 436/647] ComfyUI v0.9.0 --- comfyui_version.py | 2 +- pyproject.toml | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/comfyui_version.py b/comfyui_version.py index df82ed4fc..09def2c70 100644 --- a/comfyui_version.py +++ b/comfyui_version.py @@ -1,3 +1,3 @@ # This file is automatically generated by the build process when version is # updated in pyproject.toml. -__version__ = "0.8.2" +__version__ = "0.9.0" diff --git a/pyproject.toml b/pyproject.toml index 49f1a03fd..17aac8c3f 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -1,6 +1,6 @@ [project] name = "ComfyUI" -version = "0.8.2" +version = "0.9.0" readme = "README.md" license = { file = "LICENSE" } requires-python = ">=3.10" From 1dcbd9efaf16c74a3aff2770f46de5b4aaaf927e Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 12 Jan 2026 22:42:07 -0800 Subject: [PATCH 437/647] Bump ltxav mem estimation a bit. (#11842) --- comfy/supported_models.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy/supported_models.py b/comfy/supported_models.py index d44c0bc37..1bf54f13f 100644 --- a/comfy/supported_models.py +++ b/comfy/supported_models.py @@ -845,7 +845,7 @@ class LTXAV(LTXV): def __init__(self, unet_config): super().__init__(unet_config) - self.memory_usage_factor = 0.061 # TODO + self.memory_usage_factor = 0.077 # TODO def get_model(self, state_dict, prefix="", device=None): out = model_base.LTXAV(self, device=device) From 5ac13725331c1dfdf7aab977d4588b0a06a3debd Mon Sep 17 00:00:00 2001 From: comfyanonymous Date: Tue, 13 Jan 2026 01:44:06 -0500 Subject: [PATCH 438/647] ComfyUI v0.9.1 --- comfyui_version.py | 2 +- pyproject.toml | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/comfyui_version.py b/comfyui_version.py index 09def2c70..0c9871e35 100644 --- a/comfyui_version.py +++ b/comfyui_version.py @@ -1,3 +1,3 @@ # This file is automatically generated by the build process when version is # updated in pyproject.toml. -__version__ = "0.9.0" +__version__ = "0.9.1" diff --git a/pyproject.toml b/pyproject.toml index 17aac8c3f..dc52218b4 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -1,6 +1,6 @@ [project] name = "ComfyUI" -version = "0.9.0" +version = "0.9.1" readme = "README.md" license = { file = "LICENSE" } requires-python = ">=3.10" From c543ad81c382c8450d2c8de62644c197c3c7416d Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Tue, 13 Jan 2026 18:30:13 +0200 Subject: [PATCH 439/647] fix(api-nodes-gemini): raise exception when no candidates due to safety block (#11848) --- comfy_api_nodes/nodes_gemini.py | 28 +++++++++++++++++++--------- 1 file changed, 19 insertions(+), 9 deletions(-) diff --git a/comfy_api_nodes/nodes_gemini.py b/comfy_api_nodes/nodes_gemini.py index e8ed7e797..35bbf0d2f 100644 --- a/comfy_api_nodes/nodes_gemini.py +++ b/comfy_api_nodes/nodes_gemini.py @@ -130,7 +130,7 @@ def get_parts_by_type(response: GeminiGenerateContentResponse, part_type: Litera Returns: List of response parts matching the requested type. """ - if response.candidates is None: + if not response.candidates: if response.promptFeedback and response.promptFeedback.blockReason: feedback = response.promptFeedback raise ValueError( @@ -141,14 +141,24 @@ def get_parts_by_type(response: GeminiGenerateContentResponse, part_type: Litera "try changing it to `IMAGE+TEXT` to view the model's reasoning and understand why image generation failed." ) parts = [] - for part in response.candidates[0].content.parts: - if part_type == "text" and part.text: - parts.append(part) - elif part.inlineData and part.inlineData.mimeType == part_type: - parts.append(part) - elif part.fileData and part.fileData.mimeType == part_type: - parts.append(part) - # Skip parts that don't match the requested type + blocked_reasons = [] + for candidate in response.candidates: + if candidate.finishReason and candidate.finishReason.upper() == "IMAGE_PROHIBITED_CONTENT": + blocked_reasons.append(candidate.finishReason) + continue + if candidate.content is None or candidate.content.parts is None: + continue + for part in candidate.content.parts: + if part_type == "text" and part.text: + parts.append(part) + elif part.inlineData and part.inlineData.mimeType == part_type: + parts.append(part) + elif part.fileData and part.fileData.mimeType == part_type: + parts.append(part) + + if not parts and blocked_reasons: + raise ValueError(f"Gemini API blocked the request. Reasons: {blocked_reasons}") + return parts From d9dc02a7d602a1918b9dabfc91890e6689f6f16d Mon Sep 17 00:00:00 2001 From: Acly Date: Tue, 13 Jan 2026 21:03:53 +0100 Subject: [PATCH 440/647] Support "lite" version of alibaba-pai Z-Image Controlnet (#11849) * reduced number of control layers (3) compared to full model --- comfy_extras/nodes_model_patch.py | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/comfy_extras/nodes_model_patch.py b/comfy_extras/nodes_model_patch.py index 1355b3c93..f66d28fc9 100644 --- a/comfy_extras/nodes_model_patch.py +++ b/comfy_extras/nodes_model_patch.py @@ -244,6 +244,10 @@ class ModelPatchLoader: elif 'control_all_x_embedder.2-1.weight' in sd: # alipai z image fun controlnet sd = z_image_convert(sd) config = {} + if 'control_layers.4.adaLN_modulation.0.weight' not in sd: + config['n_control_layers'] = 3 + config['additional_in_dim'] = 17 + config['refiner_control'] = True if 'control_layers.14.adaLN_modulation.0.weight' in sd: config['n_control_layers'] = 15 config['additional_in_dim'] = 17 From e4b4fb34798a4710f670c81ae905ec24d58b6373 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Jukka=20Sepp=C3=A4nen?= <40791699+kijai@users.noreply.github.com> Date: Wed, 14 Jan 2026 00:37:21 +0200 Subject: [PATCH 441/647] Load metadata on VAELoader (#11846) Needed to load the proper LTX2 VAE if separated from checkpoint --- nodes.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/nodes.py b/nodes.py index 5a9d42d4a..90c5f2a6e 100644 --- a/nodes.py +++ b/nodes.py @@ -798,8 +798,8 @@ class VAELoader: vae_path = folder_paths.get_full_path_or_raise("vae_approx", vae_name) else: vae_path = folder_paths.get_full_path_or_raise("vae", vae_name) - sd = comfy.utils.load_torch_file(vae_path) - vae = comfy.sd.VAE(sd=sd) + sd, metadata = comfy.utils.load_torch_file(vae_path, return_metadata=True) + vae = comfy.sd.VAE(sd=sd, metadata=metadata) vae.throw_exception_if_invalid() return (vae,) From 79f6bb5e4fca0c2fbd5d09511a65449ca69332a8 Mon Sep 17 00:00:00 2001 From: ric-yu Date: Tue, 13 Jan 2026 16:14:40 -0800 Subject: [PATCH 442/647] add blueprints dir for built-in blueprints (#11853) --- app/subgraph_manager.py | 80 +++++++++++++++++++++------------- blueprints/put_blueprints_here | 0 2 files changed, 50 insertions(+), 30 deletions(-) create mode 100644 blueprints/put_blueprints_here diff --git a/app/subgraph_manager.py b/app/subgraph_manager.py index dbe404541..6a8f586a4 100644 --- a/app/subgraph_manager.py +++ b/app/subgraph_manager.py @@ -10,6 +10,7 @@ import hashlib class Source: custom_node = "custom_node" + templates = "templates" class SubgraphEntry(TypedDict): source: str @@ -38,6 +39,18 @@ class CustomNodeSubgraphEntryInfo(TypedDict): class SubgraphManager: def __init__(self): self.cached_custom_node_subgraphs: dict[SubgraphEntry] | None = None + self.cached_blueprint_subgraphs: dict[SubgraphEntry] | None = None + + def _create_entry(self, file: str, source: str, node_pack: str) -> tuple[str, SubgraphEntry]: + """Create a subgraph entry from a file path. Expects normalized path (forward slashes).""" + entry_id = hashlib.sha256(f"{source}{file}".encode()).hexdigest() + entry: SubgraphEntry = { + "source": source, + "name": os.path.splitext(os.path.basename(file))[0], + "path": file, + "info": {"node_pack": node_pack}, + } + return entry_id, entry async def load_entry_data(self, entry: SubgraphEntry): with open(entry['path'], 'r') as f: @@ -60,53 +73,60 @@ class SubgraphManager: return entries async def get_custom_node_subgraphs(self, loadedModules, force_reload=False): - # if not forced to reload and cached, return cache + """Load subgraphs from custom nodes.""" if not force_reload and self.cached_custom_node_subgraphs is not None: return self.cached_custom_node_subgraphs - # Load subgraphs from custom nodes - subfolder = "subgraphs" - subgraphs_dict: dict[SubgraphEntry] = {} + subgraphs_dict: dict[SubgraphEntry] = {} for folder in folder_paths.get_folder_paths("custom_nodes"): - pattern = os.path.join(folder, f"*/{subfolder}/*.json") - matched_files = glob.glob(pattern) - for file in matched_files: - # replace backslashes with forward slashes + pattern = os.path.join(folder, "*/subgraphs/*.json") + for file in glob.glob(pattern): file = file.replace('\\', '/') - info: CustomNodeSubgraphEntryInfo = { - "node_pack": "custom_nodes." + file.split('/')[-3] - } - source = Source.custom_node - # hash source + path to make sure id will be as unique as possible, but - # reproducible across backend reloads - id = hashlib.sha256(f"{source}{file}".encode()).hexdigest() - entry: SubgraphEntry = { - "source": Source.custom_node, - "name": os.path.splitext(os.path.basename(file))[0], - "path": file, - "info": info, - } - subgraphs_dict[id] = entry + node_pack = "custom_nodes." + file.split('/')[-3] + entry_id, entry = self._create_entry(file, Source.custom_node, node_pack) + subgraphs_dict[entry_id] = entry + self.cached_custom_node_subgraphs = subgraphs_dict return subgraphs_dict - async def get_custom_node_subgraph(self, id: str, loadedModules): - subgraphs = await self.get_custom_node_subgraphs(loadedModules) - entry: SubgraphEntry = subgraphs.get(id, None) - if entry is not None and entry.get('data', None) is None: + async def get_blueprint_subgraphs(self, force_reload=False): + """Load subgraphs from the blueprints directory.""" + if not force_reload and self.cached_blueprint_subgraphs is not None: + return self.cached_blueprint_subgraphs + + subgraphs_dict: dict[SubgraphEntry] = {} + blueprints_dir = os.path.join(os.path.dirname(os.path.dirname(__file__)), 'blueprints') + + if os.path.exists(blueprints_dir): + for file in glob.glob(os.path.join(blueprints_dir, "*.json")): + file = file.replace('\\', '/') + entry_id, entry = self._create_entry(file, Source.templates, "comfyui") + subgraphs_dict[entry_id] = entry + + self.cached_blueprint_subgraphs = subgraphs_dict + return subgraphs_dict + + async def get_all_subgraphs(self, loadedModules, force_reload=False): + """Get all subgraphs from all sources (custom nodes and blueprints).""" + custom_node_subgraphs = await self.get_custom_node_subgraphs(loadedModules, force_reload) + blueprint_subgraphs = await self.get_blueprint_subgraphs(force_reload) + return {**custom_node_subgraphs, **blueprint_subgraphs} + + async def get_subgraph(self, id: str, loadedModules): + """Get a specific subgraph by ID from any source.""" + entry = (await self.get_all_subgraphs(loadedModules)).get(id) + if entry is not None and entry.get('data') is None: await self.load_entry_data(entry) return entry def add_routes(self, routes, loadedModules): @routes.get("/global_subgraphs") async def get_global_subgraphs(request): - subgraphs_dict = await self.get_custom_node_subgraphs(loadedModules) - # NOTE: we may want to include other sources of global subgraphs such as templates in the future; - # that's the reasoning for the current implementation + subgraphs_dict = await self.get_all_subgraphs(loadedModules) return web.json_response(await self.sanitize_entries(subgraphs_dict, remove_data=True)) @routes.get("/global_subgraphs/{id}") async def get_global_subgraph(request): id = request.match_info.get("id", None) - subgraph = await self.get_custom_node_subgraph(id, loadedModules) + subgraph = await self.get_subgraph(id, loadedModules) return web.json_response(await self.sanitize_entry(subgraph)) diff --git a/blueprints/put_blueprints_here b/blueprints/put_blueprints_here new file mode 100644 index 000000000..e69de29bb From 1419047fdbdf26b2311950c041a86fd998a2acbd Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Wed, 14 Jan 2026 02:18:28 +0200 Subject: [PATCH 443/647] [Api Nodes]: Improve Price Badge Declarations (#11582) * api nodes: price badges moved to nodes code * added price badges for 4 more node-packs * added price badges for 10 more node-packs * added new price badges for Omni STD mode * add support for autogrow groups * use full names for "widgets", "inputs" and "groups" * add strict typing for JSONata rules * add price badge for WanReferenceVideoApi node * add support for DynamicCombo * sync price badges changes (https://github.com/Comfy-Org/ComfyUI_frontend/pull/7900) * sync badges for Vidu2 nodes * fixed incorrect price for RecraftCrispUpscaleNode * fixed incorrect price badges for LTXV nodes * fixed price badge for MinimaxHailuoVideoNode * fixed price badges for PixVerse nodes --- comfy_api/latest/_io.py | 79 +++++++++- comfy_api_nodes/nodes_bfl.py | 44 ++++++ comfy_api_nodes/nodes_bytedance.py | 66 +++++++++ comfy_api_nodes/nodes_gemini.py | 40 ++++++ comfy_api_nodes/nodes_ideogram.py | 42 +++++- comfy_api_nodes/nodes_kling.py | 215 ++++++++++++++++++++++++++++ comfy_api_nodes/nodes_ltxv.py | 18 +++ comfy_api_nodes/nodes_luma.py | 76 ++++++++++ comfy_api_nodes/nodes_minimax.py | 20 +++ comfy_api_nodes/nodes_moonvalley.py | 12 ++ comfy_api_nodes/nodes_openai.py | 127 ++++++++++++++++ comfy_api_nodes/nodes_pixverse.py | 30 ++++ comfy_api_nodes/nodes_recraft.py | 32 +++++ comfy_api_nodes/nodes_rodin.py | 12 ++ comfy_api_nodes/nodes_runway.py | 15 ++ comfy_api_nodes/nodes_sora.py | 18 +++ comfy_api_nodes/nodes_stability.py | 31 ++++ comfy_api_nodes/nodes_tripo.py | 164 +++++++++++++++++++++ comfy_api_nodes/nodes_veo2.py | 42 ++++++ comfy_api_nodes/nodes_vidu.py | 100 +++++++++++++ comfy_api_nodes/nodes_wan.py | 43 ++++++ 21 files changed, 1221 insertions(+), 5 deletions(-) diff --git a/comfy_api/latest/_io.py b/comfy_api/latest/_io.py index 50143ff53..e6a0d1821 100644 --- a/comfy_api/latest/_io.py +++ b/comfy_api/latest/_io.py @@ -1225,6 +1225,7 @@ class NodeInfoV1: deprecated: bool=None experimental: bool=None api_node: bool=None + price_badge: dict | None = None @dataclass class NodeInfoV3: @@ -1234,11 +1235,77 @@ class NodeInfoV3: name: str=None display_name: str=None description: str=None + python_module: Any = None category: str=None output_node: bool=None deprecated: bool=None experimental: bool=None api_node: bool=None + price_badge: dict | None = None + + +@dataclass +class PriceBadgeDepends: + widgets: list[str] = field(default_factory=list) + inputs: list[str] = field(default_factory=list) + input_groups: list[str] = field(default_factory=list) + + def validate(self) -> None: + if not isinstance(self.widgets, list) or any(not isinstance(x, str) for x in self.widgets): + raise ValueError("PriceBadgeDepends.widgets must be a list[str].") + if not isinstance(self.inputs, list) or any(not isinstance(x, str) for x in self.inputs): + raise ValueError("PriceBadgeDepends.inputs must be a list[str].") + if not isinstance(self.input_groups, list) or any(not isinstance(x, str) for x in self.input_groups): + raise ValueError("PriceBadgeDepends.input_groups must be a list[str].") + + def as_dict(self, schema_inputs: list["Input"]) -> dict[str, Any]: + # Build lookup: widget_id -> io_type + input_types: dict[str, str] = {} + for inp in schema_inputs: + all_inputs = inp.get_all() + input_types[inp.id] = inp.get_io_type() # First input is always the parent itself + for nested_inp in all_inputs[1:]: + # For DynamicCombo/DynamicSlot, nested inputs are prefixed with parent ID + # to match frontend naming convention (e.g., "should_texture.enable_pbr") + prefixed_id = f"{inp.id}.{nested_inp.id}" + input_types[prefixed_id] = nested_inp.get_io_type() + + # Enrich widgets with type information, raising error for unknown widgets + widgets_data: list[dict[str, str]] = [] + for w in self.widgets: + if w not in input_types: + raise ValueError( + f"PriceBadge depends_on.widgets references unknown widget '{w}'. " + f"Available widgets: {list(input_types.keys())}" + ) + widgets_data.append({"name": w, "type": input_types[w]}) + + return { + "widgets": widgets_data, + "inputs": self.inputs, + "input_groups": self.input_groups, + } + + +@dataclass +class PriceBadge: + expr: str + depends_on: PriceBadgeDepends = field(default_factory=PriceBadgeDepends) + engine: str = field(default="jsonata") + + def validate(self) -> None: + if self.engine != "jsonata": + raise ValueError(f"Unsupported PriceBadge.engine '{self.engine}'. Only 'jsonata' is supported.") + if not isinstance(self.expr, str) or not self.expr.strip(): + raise ValueError("PriceBadge.expr must be a non-empty string.") + self.depends_on.validate() + + def as_dict(self, schema_inputs: list["Input"]) -> dict[str, Any]: + return { + "engine": self.engine, + "depends_on": self.depends_on.as_dict(schema_inputs), + "expr": self.expr, + } @dataclass @@ -1284,6 +1351,8 @@ class Schema: """Flags a node as experimental, informing users that it may change or not work as expected.""" is_api_node: bool=False """Flags a node as an API node. See: https://docs.comfy.org/tutorials/api-nodes/overview.""" + price_badge: PriceBadge | None = None + """Optional client-evaluated pricing badge declaration for this node.""" not_idempotent: bool=False """Flags a node as not idempotent; when True, the node will run and not reuse the cached outputs when identical inputs are provided on a different node in the graph.""" enable_expand: bool=False @@ -1314,6 +1383,8 @@ class Schema: input.validate() for output in self.outputs: output.validate() + if self.price_badge is not None: + self.price_badge.validate() def finalize(self): """Add hidden based on selected schema options, and give outputs without ids default ids.""" @@ -1387,7 +1458,8 @@ class Schema: deprecated=self.is_deprecated, experimental=self.is_experimental, api_node=self.is_api_node, - python_module=getattr(cls, "RELATIVE_PYTHON_MODULE", "nodes") + python_module=getattr(cls, "RELATIVE_PYTHON_MODULE", "nodes"), + price_badge=self.price_badge.as_dict(self.inputs) if self.price_badge is not None else None, ) return info @@ -1419,7 +1491,8 @@ class Schema: deprecated=self.is_deprecated, experimental=self.is_experimental, api_node=self.is_api_node, - python_module=getattr(cls, "RELATIVE_PYTHON_MODULE", "nodes") + python_module=getattr(cls, "RELATIVE_PYTHON_MODULE", "nodes"), + price_badge=self.price_badge.as_dict(self.inputs) if self.price_badge is not None else None, ) return info @@ -1971,4 +2044,6 @@ __all__ = [ "add_to_dict_v3", "V3Data", "ImageCompare", + "PriceBadgeDepends", + "PriceBadge", ] diff --git a/comfy_api_nodes/nodes_bfl.py b/comfy_api_nodes/nodes_bfl.py index ce077d6b3..76021ef7f 100644 --- a/comfy_api_nodes/nodes_bfl.py +++ b/comfy_api_nodes/nodes_bfl.py @@ -97,6 +97,9 @@ class FluxProUltraImageNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.06}""", + ), ) @classmethod @@ -352,6 +355,9 @@ class FluxProExpandNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.05}""", + ), ) @classmethod @@ -458,6 +464,9 @@ class FluxProFillNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.05}""", + ), ) @classmethod @@ -511,6 +520,21 @@ class Flux2ProImageNode(IO.ComfyNode): NODE_ID = "Flux2ProImageNode" DISPLAY_NAME = "Flux.2 [pro] Image" API_ENDPOINT = "/proxy/bfl/flux-2-pro/generate" + PRICE_BADGE_EXPR = """ + ( + $MP := 1024 * 1024; + $outMP := $max([1, $floor(((widgets.width * widgets.height) + $MP - 1) / $MP)]); + $outputCost := 0.03 + 0.015 * ($outMP - 1); + inputs.images.connected + ? { + "type":"range_usd", + "min_usd": $outputCost + 0.015, + "max_usd": $outputCost + 0.12, + "format": { "approximate": true } + } + : {"type":"usd","usd": $outputCost} + ) + """ @classmethod def define_schema(cls) -> IO.Schema: @@ -563,6 +587,10 @@ class Flux2ProImageNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["width", "height"], inputs=["images"]), + expr=cls.PRICE_BADGE_EXPR, + ), ) @classmethod @@ -623,6 +651,22 @@ class Flux2MaxImageNode(Flux2ProImageNode): NODE_ID = "Flux2MaxImageNode" DISPLAY_NAME = "Flux.2 [max] Image" API_ENDPOINT = "/proxy/bfl/flux-2-max/generate" + PRICE_BADGE_EXPR = """ + ( + $MP := 1024 * 1024; + $outMP := $max([1, $floor(((widgets.width * widgets.height) + $MP - 1) / $MP)]); + $outputCost := 0.07 + 0.03 * ($outMP - 1); + + inputs.images.connected + ? { + "type":"range_usd", + "min_usd": $outputCost + 0.03, + "max_usd": $outputCost + 0.24, + "format": { "approximate": true } + } + : {"type":"usd","usd": $outputCost} + ) + """ class BFLExtension(ComfyExtension): diff --git a/comfy_api_nodes/nodes_bytedance.py b/comfy_api_nodes/nodes_bytedance.py index d4a2cfae6..f09a4a0ed 100644 --- a/comfy_api_nodes/nodes_bytedance.py +++ b/comfy_api_nodes/nodes_bytedance.py @@ -126,6 +126,9 @@ class ByteDanceImageNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.03}""", + ), ) @classmethod @@ -367,6 +370,19 @@ class ByteDanceSeedreamNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["model"]), + expr=""" + ( + $price := $contains(widgets.model, "seedream-4-5-251128") ? 0.04 : 0.03; + { + "type":"usd", + "usd": $price, + "format": { "suffix":" x images/Run", "approximate": true } + } + ) + """, + ), ) @classmethod @@ -522,6 +538,7 @@ class ByteDanceTextToVideoNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=PRICE_BADGE_VIDEO, ) @classmethod @@ -632,6 +649,7 @@ class ByteDanceImageToVideoNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=PRICE_BADGE_VIDEO, ) @classmethod @@ -754,6 +772,7 @@ class ByteDanceFirstLastFrameNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=PRICE_BADGE_VIDEO, ) @classmethod @@ -877,6 +896,7 @@ class ByteDanceImageReferenceNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=PRICE_BADGE_VIDEO, ) @classmethod @@ -946,6 +966,52 @@ def raise_if_text_params(prompt: str, text_params: list[str]) -> None: ) +PRICE_BADGE_VIDEO = IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["model", "duration", "resolution"]), + expr=""" + ( + $priceByModel := { + "seedance-1-0-pro": { + "480p":[0.23,0.24], + "720p":[0.51,0.56], + "1080p":[1.18,1.22] + }, + "seedance-1-0-pro-fast": { + "480p":[0.09,0.1], + "720p":[0.21,0.23], + "1080p":[0.47,0.49] + }, + "seedance-1-0-lite": { + "480p":[0.17,0.18], + "720p":[0.37,0.41], + "1080p":[0.85,0.88] + } + }; + $model := widgets.model; + $modelKey := + $contains($model, "seedance-1-0-pro-fast") ? "seedance-1-0-pro-fast" : + $contains($model, "seedance-1-0-pro") ? "seedance-1-0-pro" : + "seedance-1-0-lite"; + $resolution := widgets.resolution; + $resKey := + $contains($resolution, "1080") ? "1080p" : + $contains($resolution, "720") ? "720p" : + "480p"; + $modelPrices := $lookup($priceByModel, $modelKey); + $baseRange := $lookup($modelPrices, $resKey); + $min10s := $baseRange[0]; + $max10s := $baseRange[1]; + $scale := widgets.duration / 10; + $minCost := $min10s * $scale; + $maxCost := $max10s * $scale; + ($minCost = $maxCost) + ? {"type":"usd","usd": $minCost} + : {"type":"range_usd","min_usd": $minCost, "max_usd": $maxCost} + ) + """, +) + + class ByteDanceExtension(ComfyExtension): @override async def get_node_list(self) -> list[type[IO.ComfyNode]]: diff --git a/comfy_api_nodes/nodes_gemini.py b/comfy_api_nodes/nodes_gemini.py index 35bbf0d2f..a2daea50a 100644 --- a/comfy_api_nodes/nodes_gemini.py +++ b/comfy_api_nodes/nodes_gemini.py @@ -319,6 +319,30 @@ class GeminiNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["model"]), + expr=""" + ( + $m := widgets.model; + $contains($m, "gemini-2.5-flash") ? { + "type": "list_usd", + "usd": [0.0003, 0.0025], + "format": { "approximate": true, "separator": "-", "suffix": " per 1K tokens"} + } + : $contains($m, "gemini-2.5-pro") ? { + "type": "list_usd", + "usd": [0.00125, 0.01], + "format": { "approximate": true, "separator": "-", "suffix": " per 1K tokens" } + } + : $contains($m, "gemini-3-pro-preview") ? { + "type": "list_usd", + "usd": [0.002, 0.012], + "format": { "approximate": true, "separator": "-", "suffix": " per 1K tokens" } + } + : {"type":"text", "text":"Token-based"} + ) + """, + ), ) @classmethod @@ -580,6 +604,9 @@ class GeminiImage(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.039,"format":{"suffix":"/Image (1K)","approximate":true}}""", + ), ) @classmethod @@ -710,6 +737,19 @@ class GeminiImage2(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["resolution"]), + expr=""" + ( + $r := widgets.resolution; + ($contains($r,"1k") or $contains($r,"2k")) + ? {"type":"usd","usd":0.134,"format":{"suffix":"/Image","approximate":true}} + : $contains($r,"4k") + ? {"type":"usd","usd":0.24,"format":{"suffix":"/Image","approximate":true}} + : {"type":"text","text":"Token-based"} + ) + """, + ), ) @classmethod diff --git a/comfy_api_nodes/nodes_ideogram.py b/comfy_api_nodes/nodes_ideogram.py index 48f94e612..827b3523a 100644 --- a/comfy_api_nodes/nodes_ideogram.py +++ b/comfy_api_nodes/nodes_ideogram.py @@ -236,7 +236,6 @@ class IdeogramV1(IO.ComfyNode): display_name="Ideogram V1", category="api node/image/Ideogram", description="Generates images using the Ideogram V1 model.", - is_api_node=True, inputs=[ IO.String.Input( "prompt", @@ -298,6 +297,17 @@ class IdeogramV1(IO.ComfyNode): IO.Hidden.api_key_comfy_org, IO.Hidden.unique_id, ], + is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["num_images", "turbo"]), + expr=""" + ( + $n := widgets.num_images; + $base := (widgets.turbo = true) ? 0.0286 : 0.0858; + {"type":"usd","usd": $round($base * $n, 2)} + ) + """, + ), ) @classmethod @@ -351,7 +361,6 @@ class IdeogramV2(IO.ComfyNode): display_name="Ideogram V2", category="api node/image/Ideogram", description="Generates images using the Ideogram V2 model.", - is_api_node=True, inputs=[ IO.String.Input( "prompt", @@ -436,6 +445,17 @@ class IdeogramV2(IO.ComfyNode): IO.Hidden.api_key_comfy_org, IO.Hidden.unique_id, ], + is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["num_images", "turbo"]), + expr=""" + ( + $n := widgets.num_images; + $base := (widgets.turbo = true) ? 0.0715 : 0.1144; + {"type":"usd","usd": $round($base * $n, 2)} + ) + """, + ), ) @classmethod @@ -506,7 +526,6 @@ class IdeogramV3(IO.ComfyNode): category="api node/image/Ideogram", description="Generates images using the Ideogram V3 model. " "Supports both regular image generation from text prompts and image editing with mask.", - is_api_node=True, inputs=[ IO.String.Input( "prompt", @@ -591,6 +610,23 @@ class IdeogramV3(IO.ComfyNode): IO.Hidden.api_key_comfy_org, IO.Hidden.unique_id, ], + is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["rendering_speed", "num_images"], inputs=["character_image"]), + expr=""" + ( + $n := widgets.num_images; + $speed := widgets.rendering_speed; + $hasChar := inputs.character_image.connected; + $base := + $contains($speed,"quality") ? ($hasChar ? 0.286 : 0.1287) : + $contains($speed,"default") ? ($hasChar ? 0.2145 : 0.0858) : + $contains($speed,"turbo") ? ($hasChar ? 0.143 : 0.0429) : + 0.0858; + {"type":"usd","usd": $round($base * $n, 2)} + ) + """, + ), ) @classmethod diff --git a/comfy_api_nodes/nodes_kling.py b/comfy_api_nodes/nodes_kling.py index 01d9c34f5..05dde88b1 100644 --- a/comfy_api_nodes/nodes_kling.py +++ b/comfy_api_nodes/nodes_kling.py @@ -764,6 +764,33 @@ class KlingTextToVideoNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["mode"]), + expr=""" + ( + $m := widgets.mode; + $contains($m,"v2-5-turbo") + ? ($contains($m,"10") ? {"type":"usd","usd":0.7} : {"type":"usd","usd":0.35}) + : $contains($m,"v2-1-master") + ? ($contains($m,"10s") ? {"type":"usd","usd":2.8} : {"type":"usd","usd":1.4}) + : $contains($m,"v2-master") + ? ($contains($m,"10s") ? {"type":"usd","usd":2.8} : {"type":"usd","usd":1.4}) + : $contains($m,"v1-6") + ? ( + $contains($m,"pro") + ? ($contains($m,"10s") ? {"type":"usd","usd":0.98} : {"type":"usd","usd":0.49}) + : ($contains($m,"10s") ? {"type":"usd","usd":0.56} : {"type":"usd","usd":0.28}) + ) + : $contains($m,"v1") + ? ( + $contains($m,"pro") + ? ($contains($m,"10s") ? {"type":"usd","usd":0.98} : {"type":"usd","usd":0.49}) + : ($contains($m,"10s") ? {"type":"usd","usd":0.28} : {"type":"usd","usd":0.14}) + ) + : {"type":"usd","usd":0.14} + ) + """, + ), ) @classmethod @@ -818,6 +845,16 @@ class OmniProTextToVideoNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["duration", "resolution"]), + expr=""" + ( + $mode := (widgets.resolution = "720p") ? "std" : "pro"; + $rates := {"std": 0.084, "pro": 0.112}; + {"type":"usd","usd": $lookup($rates, $mode) * widgets.duration} + ) + """, + ), ) @classmethod @@ -886,6 +923,16 @@ class OmniProFirstLastFrameNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["duration", "resolution"]), + expr=""" + ( + $mode := (widgets.resolution = "720p") ? "std" : "pro"; + $rates := {"std": 0.084, "pro": 0.112}; + {"type":"usd","usd": $lookup($rates, $mode) * widgets.duration} + ) + """, + ), ) @classmethod @@ -981,6 +1028,16 @@ class OmniProImageToVideoNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["duration", "resolution"]), + expr=""" + ( + $mode := (widgets.resolution = "720p") ? "std" : "pro"; + $rates := {"std": 0.084, "pro": 0.112}; + {"type":"usd","usd": $lookup($rates, $mode) * widgets.duration} + ) + """, + ), ) @classmethod @@ -1056,6 +1113,16 @@ class OmniProVideoToVideoNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["duration", "resolution"]), + expr=""" + ( + $mode := (widgets.resolution = "720p") ? "std" : "pro"; + $rates := {"std": 0.126, "pro": 0.168}; + {"type":"usd","usd": $lookup($rates, $mode) * widgets.duration} + ) + """, + ), ) @classmethod @@ -1142,6 +1209,16 @@ class OmniProEditVideoNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["resolution"]), + expr=""" + ( + $mode := (widgets.resolution = "720p") ? "std" : "pro"; + $rates := {"std": 0.126, "pro": 0.168}; + {"type":"usd","usd": $lookup($rates, $mode), "format":{"suffix":"/second"}} + ) + """, + ), ) @classmethod @@ -1228,6 +1305,9 @@ class OmniProImageNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.028}""", + ), ) @classmethod @@ -1313,6 +1393,9 @@ class KlingCameraControlT2VNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.14}""", + ), ) @classmethod @@ -1375,6 +1458,33 @@ class KlingImage2VideoNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["mode", "model_name", "duration"]), + expr=""" + ( + $mode := widgets.mode; + $model := widgets.model_name; + $dur := widgets.duration; + $contains($model,"v2-5-turbo") + ? ($contains($dur,"10") ? {"type":"usd","usd":0.7} : {"type":"usd","usd":0.35}) + : ($contains($model,"v2-1-master") or $contains($model,"v2-master")) + ? ($contains($dur,"10") ? {"type":"usd","usd":2.8} : {"type":"usd","usd":1.4}) + : ($contains($model,"v2-1") or $contains($model,"v1-6") or $contains($model,"v1-5")) + ? ( + $contains($mode,"pro") + ? ($contains($dur,"10") ? {"type":"usd","usd":0.98} : {"type":"usd","usd":0.49}) + : ($contains($dur,"10") ? {"type":"usd","usd":0.56} : {"type":"usd","usd":0.28}) + ) + : $contains($model,"v1") + ? ( + $contains($mode,"pro") + ? ($contains($dur,"10") ? {"type":"usd","usd":0.98} : {"type":"usd","usd":0.49}) + : ($contains($dur,"10") ? {"type":"usd","usd":0.28} : {"type":"usd","usd":0.14}) + ) + : {"type":"usd","usd":0.14} + ) + """, + ), ) @classmethod @@ -1448,6 +1558,9 @@ class KlingCameraControlI2VNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.49}""", + ), ) @classmethod @@ -1518,6 +1631,33 @@ class KlingStartEndFrameNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["mode"]), + expr=""" + ( + $m := widgets.mode; + $contains($m,"v2-5-turbo") + ? ($contains($m,"10") ? {"type":"usd","usd":0.7} : {"type":"usd","usd":0.35}) + : $contains($m,"v2-1") + ? ($contains($m,"10s") ? {"type":"usd","usd":0.98} : {"type":"usd","usd":0.49}) + : $contains($m,"v2-master") + ? ($contains($m,"10s") ? {"type":"usd","usd":2.8} : {"type":"usd","usd":1.4}) + : $contains($m,"v1-6") + ? ( + $contains($m,"pro") + ? ($contains($m,"10s") ? {"type":"usd","usd":0.98} : {"type":"usd","usd":0.49}) + : ($contains($m,"10s") ? {"type":"usd","usd":0.56} : {"type":"usd","usd":0.28}) + ) + : $contains($m,"v1") + ? ( + $contains($m,"pro") + ? ($contains($m,"10s") ? {"type":"usd","usd":0.98} : {"type":"usd","usd":0.49}) + : ($contains($m,"10s") ? {"type":"usd","usd":0.28} : {"type":"usd","usd":0.14}) + ) + : {"type":"usd","usd":0.14} + ) + """, + ), ) @classmethod @@ -1583,6 +1723,9 @@ class KlingVideoExtendNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.28}""", + ), ) @classmethod @@ -1664,6 +1807,29 @@ class KlingDualCharacterVideoEffectNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["mode", "model_name", "duration"]), + expr=""" + ( + $mode := widgets.mode; + $model := widgets.model_name; + $dur := widgets.duration; + ($contains($model,"v1-6") or $contains($model,"v1-5")) + ? ( + $contains($mode,"pro") + ? ($contains($dur,"10") ? {"type":"usd","usd":0.98} : {"type":"usd","usd":0.49}) + : ($contains($dur,"10") ? {"type":"usd","usd":0.56} : {"type":"usd","usd":0.28}) + ) + : $contains($model,"v1") + ? ( + $contains($mode,"pro") + ? ($contains($dur,"10") ? {"type":"usd","usd":0.98} : {"type":"usd","usd":0.49}) + : ($contains($dur,"10") ? {"type":"usd","usd":0.28} : {"type":"usd","usd":0.14}) + ) + : {"type":"usd","usd":0.14} + ) + """, + ), ) @classmethod @@ -1728,6 +1894,16 @@ class KlingSingleImageVideoEffectNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["effect_scene"]), + expr=""" + ( + ($contains(widgets.effect_scene,"dizzydizzy") or $contains(widgets.effect_scene,"bloombloom")) + ? {"type":"usd","usd":0.49} + : {"type":"usd","usd":0.28} + ) + """, + ), ) @classmethod @@ -1782,6 +1958,9 @@ class KlingLipSyncAudioToVideoNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.1,"format":{"approximate":true}}""", + ), ) @classmethod @@ -1842,6 +2021,9 @@ class KlingLipSyncTextToVideoNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.1,"format":{"approximate":true}}""", + ), ) @classmethod @@ -1892,6 +2074,9 @@ class KlingVirtualTryOnNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.7}""", + ), ) @classmethod @@ -1991,6 +2176,19 @@ class KlingImageGenerationNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["model_name", "n"], inputs=["image"]), + expr=""" + ( + $m := widgets.model_name; + $base := + $contains($m,"kling-v1-5") + ? (inputs.image.connected ? 0.028 : 0.014) + : ($contains($m,"kling-v1") ? 0.0035 : 0.014); + {"type":"usd","usd": $base * widgets.n} + ) + """, + ), ) @classmethod @@ -2074,6 +2272,10 @@ class TextToVideoWithAudio(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["duration", "generate_audio"]), + expr="""{"type":"usd","usd": 0.07 * widgets.duration * (widgets.generate_audio ? 2 : 1)}""", + ), ) @classmethod @@ -2138,6 +2340,10 @@ class ImageToVideoWithAudio(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["duration", "generate_audio"]), + expr="""{"type":"usd","usd": 0.07 * widgets.duration * (widgets.generate_audio ? 2 : 1)}""", + ), ) @classmethod @@ -2218,6 +2424,15 @@ class MotionControl(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["mode"]), + expr=""" + ( + $prices := {"std": 0.07, "pro": 0.112}; + {"type":"usd","usd": $lookup($prices, widgets.mode), "format":{"suffix":"/second"}} + ) + """, + ), ) @classmethod diff --git a/comfy_api_nodes/nodes_ltxv.py b/comfy_api_nodes/nodes_ltxv.py index 7e61560dc..c6424af92 100644 --- a/comfy_api_nodes/nodes_ltxv.py +++ b/comfy_api_nodes/nodes_ltxv.py @@ -28,6 +28,22 @@ class ExecuteTaskRequest(BaseModel): image_uri: str | None = Field(None) +PRICE_BADGE = IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["model", "duration", "resolution"]), + expr=""" + ( + $prices := { + "ltx-2 (pro)": {"1920x1080":0.06,"2560x1440":0.12,"3840x2160":0.24}, + "ltx-2 (fast)": {"1920x1080":0.04,"2560x1440":0.08,"3840x2160":0.16} + }; + $modelPrices := $lookup($prices, $lowercase(widgets.model)); + $pps := $lookup($modelPrices, widgets.resolution); + {"type":"usd","usd": $pps * widgets.duration} + ) + """, +) + + class TextToVideoNode(IO.ComfyNode): @classmethod def define_schema(cls): @@ -69,6 +85,7 @@ class TextToVideoNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=PRICE_BADGE, ) @classmethod @@ -145,6 +162,7 @@ class ImageToVideoNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=PRICE_BADGE, ) @classmethod diff --git a/comfy_api_nodes/nodes_luma.py b/comfy_api_nodes/nodes_luma.py index 894f2b08c..95cb442e5 100644 --- a/comfy_api_nodes/nodes_luma.py +++ b/comfy_api_nodes/nodes_luma.py @@ -189,6 +189,19 @@ class LumaImageGenerationNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["model"]), + expr=""" + ( + $m := widgets.model; + $contains($m,"photon-flash-1") + ? {"type":"usd","usd":0.0027} + : $contains($m,"photon-1") + ? {"type":"usd","usd":0.0104} + : {"type":"usd","usd":0.0246} + ) + """, + ), ) @classmethod @@ -303,6 +316,19 @@ class LumaImageModifyNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["model"]), + expr=""" + ( + $m := widgets.model; + $contains($m,"photon-flash-1") + ? {"type":"usd","usd":0.0027} + : $contains($m,"photon-1") + ? {"type":"usd","usd":0.0104} + : {"type":"usd","usd":0.0246} + ) + """, + ), ) @classmethod @@ -395,6 +421,7 @@ class LumaTextToVideoGenerationNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=PRICE_BADGE_VIDEO, ) @classmethod @@ -505,6 +532,8 @@ class LumaImageToVideoGenerationNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=PRICE_BADGE_VIDEO, + ) @classmethod @@ -568,6 +597,53 @@ class LumaImageToVideoGenerationNode(IO.ComfyNode): return LumaKeyframes(frame0=frame0, frame1=frame1) +PRICE_BADGE_VIDEO = IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["model", "resolution", "duration"]), + expr=""" + ( + $p := { + "ray-flash-2": { + "5s": {"4k":3.13,"1080p":0.79,"720p":0.34,"540p":0.2}, + "9s": {"4k":5.65,"1080p":1.42,"720p":0.61,"540p":0.36} + }, + "ray-2": { + "5s": {"4k":9.11,"1080p":2.27,"720p":1.02,"540p":0.57}, + "9s": {"4k":16.4,"1080p":4.1,"720p":1.83,"540p":1.03} + } + }; + + $m := widgets.model; + $d := widgets.duration; + $r := widgets.resolution; + + $modelKey := + $contains($m,"ray-flash-2") ? "ray-flash-2" : + $contains($m,"ray-2") ? "ray-2" : + $contains($m,"ray-1-6") ? "ray-1-6" : + "other"; + + $durKey := $contains($d,"5s") ? "5s" : $contains($d,"9s") ? "9s" : ""; + $resKey := + $contains($r,"4k") ? "4k" : + $contains($r,"1080p") ? "1080p" : + $contains($r,"720p") ? "720p" : + $contains($r,"540p") ? "540p" : ""; + + $modelPrices := $lookup($p, $modelKey); + $durPrices := $lookup($modelPrices, $durKey); + $v := $lookup($durPrices, $resKey); + + $price := + ($modelKey = "ray-1-6") ? 0.5 : + ($modelKey = "other") ? 0.79 : + ($exists($v) ? $v : 0.79); + + {"type":"usd","usd": $price} + ) + """, +) + + class LumaExtension(ComfyExtension): @override async def get_node_list(self) -> list[type[IO.ComfyNode]]: diff --git a/comfy_api_nodes/nodes_minimax.py b/comfy_api_nodes/nodes_minimax.py index 05cbb700f..43a15d50d 100644 --- a/comfy_api_nodes/nodes_minimax.py +++ b/comfy_api_nodes/nodes_minimax.py @@ -134,6 +134,9 @@ class MinimaxTextToVideoNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.43}""", + ), ) @classmethod @@ -197,6 +200,9 @@ class MinimaxImageToVideoNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.43}""", + ), ) @classmethod @@ -340,6 +346,20 @@ class MinimaxHailuoVideoNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["resolution", "duration"]), + expr=""" + ( + $prices := { + "768p": {"6": 0.28, "10": 0.56}, + "1080p": {"6": 0.49} + }; + $resPrices := $lookup($prices, $lowercase(widgets.resolution)); + $price := $lookup($resPrices, $string(widgets.duration)); + {"type":"usd","usd": $price ? $price : 0.43} + ) + """, + ), ) @classmethod diff --git a/comfy_api_nodes/nodes_moonvalley.py b/comfy_api_nodes/nodes_moonvalley.py index 2771e4790..769b171b7 100644 --- a/comfy_api_nodes/nodes_moonvalley.py +++ b/comfy_api_nodes/nodes_moonvalley.py @@ -233,6 +233,10 @@ class MoonvalleyImg2VideoNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(), + expr="""{"type":"usd","usd": 1.5}""", + ), ) @classmethod @@ -351,6 +355,10 @@ class MoonvalleyVideo2VideoNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(), + expr="""{"type":"usd","usd": 2.25}""", + ), ) @classmethod @@ -471,6 +479,10 @@ class MoonvalleyTxt2VideoNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(), + expr="""{"type":"usd","usd": 1.5}""", + ), ) @classmethod diff --git a/comfy_api_nodes/nodes_openai.py b/comfy_api_nodes/nodes_openai.py index a6205a34f..2f144c5c3 100644 --- a/comfy_api_nodes/nodes_openai.py +++ b/comfy_api_nodes/nodes_openai.py @@ -160,6 +160,23 @@ class OpenAIDalle2(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["size", "n"]), + expr=""" + ( + $size := widgets.size; + $nRaw := widgets.n; + $n := ($nRaw != null and $nRaw != 0) ? $nRaw : 1; + + $base := + $contains($size, "256x256") ? 0.016 : + $contains($size, "512x512") ? 0.018 : + 0.02; + + {"type":"usd","usd": $round($base * $n, 3)} + ) + """, + ), ) @classmethod @@ -287,6 +304,25 @@ class OpenAIDalle3(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["size", "quality"]), + expr=""" + ( + $size := widgets.size; + $q := widgets.quality; + $hd := $contains($q, "hd"); + + $price := + $contains($size, "1024x1024") + ? ($hd ? 0.08 : 0.04) + : (($contains($size, "1792x1024") or $contains($size, "1024x1792")) + ? ($hd ? 0.12 : 0.08) + : 0.04); + + {"type":"usd","usd": $price} + ) + """, + ), ) @classmethod @@ -411,6 +447,28 @@ class OpenAIGPTImage1(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["quality", "n"]), + expr=""" + ( + $ranges := { + "low": [0.011, 0.02], + "medium": [0.046, 0.07], + "high": [0.167, 0.3] + }; + $range := $lookup($ranges, widgets.quality); + $n := widgets.n; + ($n = 1) + ? {"type":"range_usd","min_usd": $range[0], "max_usd": $range[1]} + : { + "type":"range_usd", + "min_usd": $range[0], + "max_usd": $range[1], + "format": { "suffix": " x " & $string($n) & "/Run" } + } + ) + """, + ), ) @classmethod @@ -566,6 +624,75 @@ class OpenAIChatNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["model"]), + expr=""" + ( + $m := widgets.model; + $contains($m, "o4-mini") ? { + "type": "list_usd", + "usd": [0.0011, 0.0044], + "format": { "approximate": true, "separator": "-", "suffix": " per 1K tokens" } + } + : $contains($m, "o1-pro") ? { + "type": "list_usd", + "usd": [0.15, 0.6], + "format": { "approximate": true, "separator": "-", "suffix": " per 1K tokens" } + } + : $contains($m, "o1") ? { + "type": "list_usd", + "usd": [0.015, 0.06], + "format": { "approximate": true, "separator": "-", "suffix": " per 1K tokens" } + } + : $contains($m, "o3-mini") ? { + "type": "list_usd", + "usd": [0.0011, 0.0044], + "format": { "approximate": true, "separator": "-", "suffix": " per 1K tokens" } + } + : $contains($m, "o3") ? { + "type": "list_usd", + "usd": [0.01, 0.04], + "format": { "approximate": true, "separator": "-", "suffix": " per 1K tokens" } + } + : $contains($m, "gpt-4o") ? { + "type": "list_usd", + "usd": [0.0025, 0.01], + "format": { "approximate": true, "separator": "-", "suffix": " per 1K tokens" } + } + : $contains($m, "gpt-4.1-nano") ? { + "type": "list_usd", + "usd": [0.0001, 0.0004], + "format": { "approximate": true, "separator": "-", "suffix": " per 1K tokens" } + } + : $contains($m, "gpt-4.1-mini") ? { + "type": "list_usd", + "usd": [0.0004, 0.0016], + "format": { "approximate": true, "separator": "-", "suffix": " per 1K tokens" } + } + : $contains($m, "gpt-4.1") ? { + "type": "list_usd", + "usd": [0.002, 0.008], + "format": { "approximate": true, "separator": "-", "suffix": " per 1K tokens" } + } + : $contains($m, "gpt-5-nano") ? { + "type": "list_usd", + "usd": [0.00005, 0.0004], + "format": { "approximate": true, "separator": "-", "suffix": " per 1K tokens" } + } + : $contains($m, "gpt-5-mini") ? { + "type": "list_usd", + "usd": [0.00025, 0.002], + "format": { "approximate": true, "separator": "-", "suffix": " per 1K tokens" } + } + : $contains($m, "gpt-5") ? { + "type": "list_usd", + "usd": [0.00125, 0.01], + "format": { "approximate": true, "separator": "-", "suffix": " per 1K tokens" } + } + : {"type": "text", "text": "Token-based"} + ) + """, + ), ) @classmethod diff --git a/comfy_api_nodes/nodes_pixverse.py b/comfy_api_nodes/nodes_pixverse.py index 6e1686af0..86ddb3ab9 100644 --- a/comfy_api_nodes/nodes_pixverse.py +++ b/comfy_api_nodes/nodes_pixverse.py @@ -128,6 +128,7 @@ class PixverseTextToVideoNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=PRICE_BADGE_VIDEO, ) @classmethod @@ -242,6 +243,7 @@ class PixverseImageToVideoNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=PRICE_BADGE_VIDEO, ) @classmethod @@ -355,6 +357,7 @@ class PixverseTransitionVideoNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=PRICE_BADGE_VIDEO, ) @classmethod @@ -416,6 +419,33 @@ class PixverseTransitionVideoNode(IO.ComfyNode): return IO.NodeOutput(await download_url_to_video_output(response_poll.Resp.url)) +PRICE_BADGE_VIDEO = IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["duration_seconds", "quality", "motion_mode"]), + expr=""" + ( + $prices := { + "5": { + "1080p": {"normal": 1.2, "fast": 1.2}, + "720p": {"normal": 0.6, "fast": 1.2}, + "540p": {"normal": 0.45, "fast": 0.9}, + "360p": {"normal": 0.45, "fast": 0.9} + }, + "8": { + "1080p": {"normal": 1.2, "fast": 1.2}, + "720p": {"normal": 1.2, "fast": 1.2}, + "540p": {"normal": 0.9, "fast": 1.2}, + "360p": {"normal": 0.9, "fast": 1.2} + } + }; + $durPrices := $lookup($prices, $string(widgets.duration_seconds)); + $qualityPrices := $lookup($durPrices, widgets.quality); + $price := $lookup($qualityPrices, widgets.motion_mode); + {"type":"usd","usd": $price ? $price : 0.9} + ) + """, +) + + class PixVerseExtension(ComfyExtension): @override async def get_node_list(self) -> list[type[IO.ComfyNode]]: diff --git a/comfy_api_nodes/nodes_recraft.py b/comfy_api_nodes/nodes_recraft.py index e3440b946..05dc151ad 100644 --- a/comfy_api_nodes/nodes_recraft.py +++ b/comfy_api_nodes/nodes_recraft.py @@ -378,6 +378,10 @@ class RecraftTextToImageNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["n"]), + expr="""{"type":"usd","usd": $round(0.04 * widgets.n, 2)}""", + ), ) @classmethod @@ -490,6 +494,10 @@ class RecraftImageToImageNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["n"]), + expr="""{"type":"usd","usd": $round(0.04 * widgets.n, 2)}""", + ), ) @classmethod @@ -591,6 +599,10 @@ class RecraftImageInpaintingNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["n"]), + expr="""{"type":"usd","usd": $round(0.04 * widgets.n, 2)}""", + ), ) @classmethod @@ -692,6 +704,10 @@ class RecraftTextToVectorNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["n"]), + expr="""{"type":"usd","usd": $round(0.08 * widgets.n, 2)}""", + ), ) @classmethod @@ -759,6 +775,10 @@ class RecraftVectorizeImageNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(), + expr="""{"type":"usd","usd": 0.01}""", + ), ) @classmethod @@ -817,6 +837,9 @@ class RecraftReplaceBackgroundNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.04}""", + ), ) @classmethod @@ -883,6 +906,9 @@ class RecraftRemoveBackgroundNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.01}""", + ), ) @classmethod @@ -929,6 +955,9 @@ class RecraftCrispUpscaleNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.004}""", + ), ) @classmethod @@ -972,6 +1001,9 @@ class RecraftCreativeUpscaleNode(RecraftCrispUpscaleNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.25}""", + ), ) diff --git a/comfy_api_nodes/nodes_rodin.py b/comfy_api_nodes/nodes_rodin.py index e60e7a6d6..b4420cb93 100644 --- a/comfy_api_nodes/nodes_rodin.py +++ b/comfy_api_nodes/nodes_rodin.py @@ -241,6 +241,9 @@ class Rodin3D_Regular(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.4}""", + ), ) @classmethod @@ -294,6 +297,9 @@ class Rodin3D_Detail(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.4}""", + ), ) @classmethod @@ -347,6 +353,9 @@ class Rodin3D_Smooth(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.4}""", + ), ) @classmethod @@ -406,6 +415,9 @@ class Rodin3D_Sketch(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.4}""", + ), ) @classmethod diff --git a/comfy_api_nodes/nodes_runway.py b/comfy_api_nodes/nodes_runway.py index 3c55039c9..d19fdb365 100644 --- a/comfy_api_nodes/nodes_runway.py +++ b/comfy_api_nodes/nodes_runway.py @@ -184,6 +184,10 @@ class RunwayImageToVideoNodeGen3a(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["duration"]), + expr="""{"type":"usd","usd": 0.0715 * widgets.duration}""", + ), ) @classmethod @@ -274,6 +278,10 @@ class RunwayImageToVideoNodeGen4(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["duration"]), + expr="""{"type":"usd","usd": 0.0715 * widgets.duration}""", + ), ) @classmethod @@ -372,6 +380,10 @@ class RunwayFirstLastFrameNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["duration"]), + expr="""{"type":"usd","usd": 0.0715 * widgets.duration}""", + ), ) @classmethod @@ -457,6 +469,9 @@ class RunwayTextToImageNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.11}""", + ), ) @classmethod diff --git a/comfy_api_nodes/nodes_sora.py b/comfy_api_nodes/nodes_sora.py index 92b225d40..87e663845 100644 --- a/comfy_api_nodes/nodes_sora.py +++ b/comfy_api_nodes/nodes_sora.py @@ -89,6 +89,24 @@ class OpenAIVideoSora2(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["model", "size", "duration"]), + expr=""" + ( + $m := widgets.model; + $size := widgets.size; + $dur := widgets.duration; + $isPro := $contains($m, "sora-2-pro"); + $isSora2 := $contains($m, "sora-2"); + $isProSize := ($size = "1024x1792" or $size = "1792x1024"); + $perSec := + $isPro ? ($isProSize ? 0.5 : 0.3) : + $isSora2 ? 0.1 : + ($isProSize ? 0.5 : 0.1); + {"type":"usd","usd": $round($perSec * $dur, 2)} + ) + """, + ), ) @classmethod diff --git a/comfy_api_nodes/nodes_stability.py b/comfy_api_nodes/nodes_stability.py index bb7ceed78..5c48c1f1e 100644 --- a/comfy_api_nodes/nodes_stability.py +++ b/comfy_api_nodes/nodes_stability.py @@ -127,6 +127,9 @@ class StabilityStableImageUltraNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.08}""", + ), ) @classmethod @@ -264,6 +267,16 @@ class StabilityStableImageSD_3_5Node(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["model"]), + expr=""" + ( + $contains(widgets.model,"large") + ? {"type":"usd","usd":0.065} + : {"type":"usd","usd":0.035} + ) + """, + ), ) @classmethod @@ -382,6 +395,9 @@ class StabilityUpscaleConservativeNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.25}""", + ), ) @classmethod @@ -486,6 +502,9 @@ class StabilityUpscaleCreativeNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.25}""", + ), ) @classmethod @@ -566,6 +585,9 @@ class StabilityUpscaleFastNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.01}""", + ), ) @classmethod @@ -648,6 +670,9 @@ class StabilityTextToAudio(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.2}""", + ), ) @classmethod @@ -732,6 +757,9 @@ class StabilityAudioToAudio(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.2}""", + ), ) @classmethod @@ -828,6 +856,9 @@ class StabilityAudioInpaint(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.2}""", + ), ) @classmethod diff --git a/comfy_api_nodes/nodes_tripo.py b/comfy_api_nodes/nodes_tripo.py index e72f8e96a..aa790143d 100644 --- a/comfy_api_nodes/nodes_tripo.py +++ b/comfy_api_nodes/nodes_tripo.py @@ -117,6 +117,38 @@ class TripoTextToModelNode(IO.ComfyNode): ], is_api_node=True, is_output_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends( + widgets=[ + "model_version", + "style", + "texture", + "pbr", + "quad", + "texture_quality", + "geometry_quality", + ], + ), + expr=""" + ( + $isV14 := $contains(widgets.model_version,"v1.4"); + $style := widgets.style; + $hasStyle := ($style != "" and $style != "none"); + $withTexture := widgets.texture or widgets.pbr; + $isHdTexture := (widgets.texture_quality = "detailed"); + $isDetailedGeometry := (widgets.geometry_quality = "detailed"); + $baseCredits := + $isV14 ? 20 : ($withTexture ? 20 : 10); + $credits := + $baseCredits + + ($hasStyle ? 5 : 0) + + (widgets.quad ? 5 : 0) + + ($isHdTexture ? 10 : 0) + + ($isDetailedGeometry ? 20 : 0); + {"type":"usd","usd": $round($credits * 0.01, 2)} + ) + """, + ), ) @classmethod @@ -210,6 +242,38 @@ class TripoImageToModelNode(IO.ComfyNode): ], is_api_node=True, is_output_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends( + widgets=[ + "model_version", + "style", + "texture", + "pbr", + "quad", + "texture_quality", + "geometry_quality", + ], + ), + expr=""" + ( + $isV14 := $contains(widgets.model_version,"v1.4"); + $style := widgets.style; + $hasStyle := ($style != "" and $style != "none"); + $withTexture := widgets.texture or widgets.pbr; + $isHdTexture := (widgets.texture_quality = "detailed"); + $isDetailedGeometry := (widgets.geometry_quality = "detailed"); + $baseCredits := + $isV14 ? 30 : ($withTexture ? 30 : 20); + $credits := + $baseCredits + + ($hasStyle ? 5 : 0) + + (widgets.quad ? 5 : 0) + + ($isHdTexture ? 10 : 0) + + ($isDetailedGeometry ? 20 : 0); + {"type":"usd","usd": $round($credits * 0.01, 2)} + ) + """, + ), ) @classmethod @@ -314,6 +378,34 @@ class TripoMultiviewToModelNode(IO.ComfyNode): ], is_api_node=True, is_output_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends( + widgets=[ + "model_version", + "texture", + "pbr", + "quad", + "texture_quality", + "geometry_quality", + ], + ), + expr=""" + ( + $isV14 := $contains(widgets.model_version,"v1.4"); + $withTexture := widgets.texture or widgets.pbr; + $isHdTexture := (widgets.texture_quality = "detailed"); + $isDetailedGeometry := (widgets.geometry_quality = "detailed"); + $baseCredits := + $isV14 ? 30 : ($withTexture ? 30 : 20); + $credits := + $baseCredits + + (widgets.quad ? 5 : 0) + + ($isHdTexture ? 10 : 0) + + ($isDetailedGeometry ? 20 : 0); + {"type":"usd","usd": $round($credits * 0.01, 2)} + ) + """, + ), ) @classmethod @@ -405,6 +497,15 @@ class TripoTextureNode(IO.ComfyNode): ], is_api_node=True, is_output_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["texture_quality"]), + expr=""" + ( + $tq := widgets.texture_quality; + {"type":"usd","usd": ($contains($tq,"detailed") ? 0.2 : 0.1)} + ) + """, + ), ) @classmethod @@ -456,6 +557,9 @@ class TripoRefineNode(IO.ComfyNode): ], is_api_node=True, is_output_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.3}""", + ), ) @classmethod @@ -489,6 +593,9 @@ class TripoRigNode(IO.ComfyNode): ], is_api_node=True, is_output_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.25}""", + ), ) @classmethod @@ -545,6 +652,9 @@ class TripoRetargetNode(IO.ComfyNode): ], is_api_node=True, is_output_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.1}""", + ), ) @classmethod @@ -638,6 +748,60 @@ class TripoConversionNode(IO.ComfyNode): ], is_api_node=True, is_output_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends( + widgets=[ + "quad", + "face_limit", + "texture_size", + "texture_format", + "force_symmetry", + "flatten_bottom", + "flatten_bottom_threshold", + "pivot_to_center_bottom", + "scale_factor", + "with_animation", + "pack_uv", + "bake", + "part_names", + "fbx_preset", + "export_vertex_colors", + "export_orientation", + "animate_in_place", + ], + ), + expr=""" + ( + $face := (widgets.face_limit != null) ? widgets.face_limit : -1; + $texSize := (widgets.texture_size != null) ? widgets.texture_size : 4096; + $flatThresh := (widgets.flatten_bottom_threshold != null) ? widgets.flatten_bottom_threshold : 0; + $scale := (widgets.scale_factor != null) ? widgets.scale_factor : 1; + $texFmt := (widgets.texture_format != "" ? widgets.texture_format : "jpeg"); + $part := widgets.part_names; + $fbx := (widgets.fbx_preset != "" ? widgets.fbx_preset : "blender"); + $orient := (widgets.export_orientation != "" ? widgets.export_orientation : "default"); + $advanced := + widgets.quad or + widgets.force_symmetry or + widgets.flatten_bottom or + widgets.pivot_to_center_bottom or + widgets.with_animation or + widgets.pack_uv or + widgets.bake or + widgets.export_vertex_colors or + widgets.animate_in_place or + ($face != -1) or + ($texSize != 4096) or + ($flatThresh != 0) or + ($scale != 1) or + ($texFmt != "jpeg") or + ($part != "") or + ($fbx != "blender") or + ($orient != "default"); + {"type":"usd","usd": ($advanced ? 0.1 : 0.05)} + ) + """, + ), ) @classmethod diff --git a/comfy_api_nodes/nodes_veo2.py b/comfy_api_nodes/nodes_veo2.py index 13a6bfd91..c14d6ad68 100644 --- a/comfy_api_nodes/nodes_veo2.py +++ b/comfy_api_nodes/nodes_veo2.py @@ -122,6 +122,10 @@ class VeoVideoGenerationNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["duration_seconds"]), + expr="""{"type":"usd","usd": 0.5 * widgets.duration_seconds}""", + ), ) @classmethod @@ -347,6 +351,20 @@ class Veo3VideoGenerationNode(VeoVideoGenerationNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["model", "generate_audio"]), + expr=""" + ( + $m := widgets.model; + $a := widgets.generate_audio; + ($contains($m,"veo-3.0-fast-generate-001") or $contains($m,"veo-3.1-fast-generate")) + ? {"type":"usd","usd": ($a ? 1.2 : 0.8)} + : ($contains($m,"veo-3.0-generate-001") or $contains($m,"veo-3.1-generate")) + ? {"type":"usd","usd": ($a ? 3.2 : 1.6)} + : {"type":"range_usd","min_usd":0.8,"max_usd":3.2} + ) + """, + ), ) @@ -420,6 +438,30 @@ class Veo3FirstLastFrameNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["model", "generate_audio", "duration"]), + expr=""" + ( + $prices := { + "veo-3.1-fast-generate": { "audio": 0.15, "no_audio": 0.10 }, + "veo-3.1-generate": { "audio": 0.40, "no_audio": 0.20 } + }; + $m := widgets.model; + $ga := (widgets.generate_audio = "true"); + $seconds := widgets.duration; + $modelKey := + $contains($m, "veo-3.1-fast-generate") ? "veo-3.1-fast-generate" : + $contains($m, "veo-3.1-generate") ? "veo-3.1-generate" : + ""; + $audioKey := $ga ? "audio" : "no_audio"; + $modelPrices := $lookup($prices, $modelKey); + $pps := $lookup($modelPrices, $audioKey); + ($pps != null) + ? {"type":"usd","usd": $pps * $seconds} + : {"type":"range_usd","min_usd": 0.4, "max_usd": 3.2} + ) + """, + ), ) @classmethod diff --git a/comfy_api_nodes/nodes_vidu.py b/comfy_api_nodes/nodes_vidu.py index 9d94ae7ad..8edb02f39 100644 --- a/comfy_api_nodes/nodes_vidu.py +++ b/comfy_api_nodes/nodes_vidu.py @@ -121,6 +121,9 @@ class ViduTextToVideoNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.4}""", + ), ) @classmethod @@ -214,6 +217,9 @@ class ViduImageToVideoNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.4}""", + ), ) @classmethod @@ -317,6 +323,9 @@ class ViduReferenceVideoNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.4}""", + ), ) @classmethod @@ -426,6 +435,9 @@ class ViduStartEndToVideoNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.4}""", + ), ) @classmethod @@ -507,6 +519,17 @@ class Vidu2TextToVideoNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["duration", "resolution"]), + expr=""" + ( + $is1080 := widgets.resolution = "1080p"; + $base := $is1080 ? 0.1 : 0.075; + $perSec := $is1080 ? 0.05 : 0.025; + {"type":"usd","usd": $base + $perSec * (widgets.duration - 1)} + ) + """, + ), ) @classmethod @@ -594,6 +617,39 @@ class Vidu2ImageToVideoNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["model", "duration", "resolution"]), + expr=""" + ( + $m := widgets.model; + $d := widgets.duration; + $is1080 := widgets.resolution = "1080p"; + $contains($m, "pro-fast") + ? ( + $base := $is1080 ? 0.08 : 0.04; + $perSec := $is1080 ? 0.02 : 0.01; + {"type":"usd","usd": $base + $perSec * ($d - 1)} + ) + : $contains($m, "pro") + ? ( + $base := $is1080 ? 0.275 : 0.075; + $perSec := $is1080 ? 0.075 : 0.05; + {"type":"usd","usd": $base + $perSec * ($d - 1)} + ) + : $contains($m, "turbo") + ? ( + $is1080 + ? {"type":"usd","usd": 0.175 + 0.05 * ($d - 1)} + : ( + $d <= 1 ? {"type":"usd","usd": 0.04} + : $d <= 2 ? {"type":"usd","usd": 0.05} + : {"type":"usd","usd": 0.05 + 0.05 * ($d - 2)} + ) + ) + : {"type":"usd","usd": 0.04} + ) + """, + ), ) @classmethod @@ -698,6 +754,18 @@ class Vidu2ReferenceVideoNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["audio", "duration", "resolution"]), + expr=""" + ( + $is1080 := widgets.resolution = "1080p"; + $base := $is1080 ? 0.375 : 0.125; + $perSec := $is1080 ? 0.05 : 0.025; + $audioCost := widgets.audio = true ? 0.075 : 0; + {"type":"usd","usd": $base + $perSec * (widgets.duration - 1) + $audioCost} + ) + """, + ), ) @classmethod @@ -804,6 +872,38 @@ class Vidu2StartEndToVideoNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["model", "duration", "resolution"]), + expr=""" + ( + $m := widgets.model; + $d := widgets.duration; + $is1080 := widgets.resolution = "1080p"; + $contains($m, "pro-fast") + ? ( + $base := $is1080 ? 0.08 : 0.04; + $perSec := $is1080 ? 0.02 : 0.01; + {"type":"usd","usd": $base + $perSec * ($d - 1)} + ) + : $contains($m, "pro") + ? ( + $base := $is1080 ? 0.275 : 0.075; + $perSec := $is1080 ? 0.075 : 0.05; + {"type":"usd","usd": $base + $perSec * ($d - 1)} + ) + : $contains($m, "turbo") + ? ( + $is1080 + ? {"type":"usd","usd": 0.175 + 0.05 * ($d - 1)} + : ( + $d <= 2 ? {"type":"usd","usd": 0.05} + : {"type":"usd","usd": 0.05 + 0.05 * ($d - 2)} + ) + ) + : {"type":"usd","usd": 0.04} + ) + """, + ), ) @classmethod diff --git a/comfy_api_nodes/nodes_wan.py b/comfy_api_nodes/nodes_wan.py index 3e04786a9..a1355d4f1 100644 --- a/comfy_api_nodes/nodes_wan.py +++ b/comfy_api_nodes/nodes_wan.py @@ -244,6 +244,9 @@ class WanTextToImageApi(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.03}""", + ), ) @classmethod @@ -363,6 +366,9 @@ class WanImageToImageApi(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.03}""", + ), ) @classmethod @@ -520,6 +526,17 @@ class WanTextToVideoApi(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["duration", "size"]), + expr=""" + ( + $ppsTable := { "480p": 0.05, "720p": 0.1, "1080p": 0.15 }; + $resKey := $substringBefore(widgets.size, ":"); + $pps := $lookup($ppsTable, $resKey); + { "type": "usd", "usd": $round($pps * widgets.duration, 2) } + ) + """, + ), ) @classmethod @@ -681,6 +698,16 @@ class WanImageToVideoApi(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["duration", "resolution"]), + expr=""" + ( + $ppsTable := { "480p": 0.05, "720p": 0.1, "1080p": 0.15 }; + $pps := $lookup($ppsTable, widgets.resolution); + { "type": "usd", "usd": $round($pps * widgets.duration, 2) } + ) + """, + ), ) @classmethod @@ -828,6 +855,22 @@ class WanReferenceVideoApi(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["size", "duration"]), + expr=""" + ( + $rate := $contains(widgets.size, "1080p") ? 0.15 : 0.10; + $inputMin := 2 * $rate; + $inputMax := 5 * $rate; + $outputPrice := widgets.duration * $rate; + { + "type": "range_usd", + "min_usd": $inputMin + $outputPrice, + "max_usd": $inputMax + $outputPrice + } + ) + """, + ), ) @classmethod From 15b312de7a74a836fa45b989a7697895b01e0cbf Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 13 Jan 2026 16:23:58 -0800 Subject: [PATCH 444/647] Optimize nvfp4 lora applying. (#11854) --- comfy/float.py | 12 ++---------- 1 file changed, 2 insertions(+), 10 deletions(-) diff --git a/comfy/float.py b/comfy/float.py index c806af76b..1a6070bff 100644 --- a/comfy/float.py +++ b/comfy/float.py @@ -165,20 +165,12 @@ def stochastic_round_quantize_nvfp4(x, per_tensor_scale, pad_16x, seed=0): block_scale = max_abs / F4_E2M1_MAX scaled_block_scales = block_scale / per_tensor_scale.to(block_scale.dtype) scaled_block_scales_fp8 = torch.clamp(scaled_block_scales, max=F8_E4M3_MAX).to(torch.float8_e4m3fn) - total_scale = per_tensor_scale.to(x.dtype) * scaled_block_scales_fp8.to(x.dtype) - - # Handle zero blocks (from padding): avoid 0/0 NaN - zero_scale_mask = (total_scale == 0) - total_scale_safe = torch.where(zero_scale_mask, torch.ones_like(total_scale), total_scale) - - x = x / total_scale_safe.unsqueeze(-1) + x /= (per_tensor_scale.to(x.dtype) * scaled_block_scales_fp8.to(x.dtype)).unsqueeze(-1) generator = torch.Generator(device=x.device) generator.manual_seed(seed) - x = torch.where(zero_scale_mask.unsqueeze(-1), torch.zeros_like(x), x) - - x = x.view(orig_shape) + x = x.view(orig_shape).nan_to_num() data_lp = stochastic_float_to_fp4_e2m1(x, generator=generator) blocked_scales = to_blocked(scaled_block_scales_fp8, flatten=False) From eff2b9d412932aa7d49e6302cdf6e7cf24808b6f Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 13 Jan 2026 16:37:19 -0800 Subject: [PATCH 445/647] Optimize nvfp4 lora applying. (#11856) --- comfy/float.py | 6 +----- 1 file changed, 1 insertion(+), 5 deletions(-) diff --git a/comfy/float.py b/comfy/float.py index 1a6070bff..8c303bea0 100644 --- a/comfy/float.py +++ b/comfy/float.py @@ -161,10 +161,7 @@ def stochastic_round_quantize_nvfp4(x, per_tensor_scale, pad_16x, seed=0): block_size = 16 x = x.reshape(orig_shape[0], -1, block_size) - max_abs = torch.amax(torch.abs(x), dim=-1) - block_scale = max_abs / F4_E2M1_MAX - scaled_block_scales = block_scale / per_tensor_scale.to(block_scale.dtype) - scaled_block_scales_fp8 = torch.clamp(scaled_block_scales, max=F8_E4M3_MAX).to(torch.float8_e4m3fn) + scaled_block_scales_fp8 = torch.clamp(((torch.amax(torch.abs(x), dim=-1)) / F4_E2M1_MAX) / per_tensor_scale.to(x.dtype), max=F8_E4M3_MAX).to(torch.float8_e4m3fn) x /= (per_tensor_scale.to(x.dtype) * scaled_block_scales_fp8.to(x.dtype)).unsqueeze(-1) generator = torch.Generator(device=x.device) @@ -172,6 +169,5 @@ def stochastic_round_quantize_nvfp4(x, per_tensor_scale, pad_16x, seed=0): x = x.view(orig_shape).nan_to_num() data_lp = stochastic_float_to_fp4_e2m1(x, generator=generator) - blocked_scales = to_blocked(scaled_block_scales_fp8, flatten=False) return data_lp, blocked_scales From 469dd9c16ad88765ffe4e7bfa57dd80faafbaddf Mon Sep 17 00:00:00 2001 From: nomadoor <124905471+nomadoor@users.noreply.github.com> Date: Wed, 14 Jan 2026 09:48:10 +0900 Subject: [PATCH 446/647] Adds crop to multiple mode to ResizeImageMaskNode. (#11838) * Add crop-to-multiple resize mode * Make scale-to-multiple shape handling explicit --- comfy_extras/nodes_post_processing.py | 44 +++++++++++++++++++++++++++ 1 file changed, 44 insertions(+) diff --git a/comfy_extras/nodes_post_processing.py b/comfy_extras/nodes_post_processing.py index 01afa13a1..0433bbda2 100644 --- a/comfy_extras/nodes_post_processing.py +++ b/comfy_extras/nodes_post_processing.py @@ -254,6 +254,7 @@ class ResizeType(str, Enum): SCALE_HEIGHT = "scale height" SCALE_TOTAL_PIXELS = "scale total pixels" MATCH_SIZE = "match size" + SCALE_TO_MULTIPLE = "scale to multiple" def is_image(input: torch.Tensor) -> bool: # images have 4 dimensions: [batch, height, width, channels] @@ -363,6 +364,43 @@ def scale_match_size(input: torch.Tensor, match: torch.Tensor, scale_method: str input = finalize_image_mask_input(input, is_type_image) return input +def scale_to_multiple_cover(input: torch.Tensor, multiple: int, scale_method: str) -> torch.Tensor: + if multiple <= 1: + return input + is_type_image = is_image(input) + if is_type_image: + _, height, width, _ = input.shape + else: + _, height, width = input.shape + target_w = (width // multiple) * multiple + target_h = (height // multiple) * multiple + if target_w == 0 or target_h == 0: + return input + if target_w == width and target_h == height: + return input + s_w = target_w / width + s_h = target_h / height + if s_w >= s_h: + scaled_w = target_w + scaled_h = int(math.ceil(height * s_w)) + if scaled_h < target_h: + scaled_h = target_h + else: + scaled_h = target_h + scaled_w = int(math.ceil(width * s_h)) + if scaled_w < target_w: + scaled_w = target_w + input = init_image_mask_input(input, is_type_image) + input = comfy.utils.common_upscale(input, scaled_w, scaled_h, scale_method, "disabled") + input = finalize_image_mask_input(input, is_type_image) + x0 = (scaled_w - target_w) // 2 + y0 = (scaled_h - target_h) // 2 + x1 = x0 + target_w + y1 = y0 + target_h + if is_type_image: + return input[:, y0:y1, x0:x1, :] + return input[:, y0:y1, x0:x1] + class ResizeImageMaskNode(io.ComfyNode): scale_methods = ["nearest-exact", "bilinear", "area", "bicubic", "lanczos"] @@ -378,6 +416,7 @@ class ResizeImageMaskNode(io.ComfyNode): longer_size: int shorter_size: int megapixels: float + multiple: int @classmethod def define_schema(cls): @@ -417,6 +456,9 @@ class ResizeImageMaskNode(io.ComfyNode): io.MultiType.Input("match", [io.Image, io.Mask]), crop_combo, ]), + io.DynamicCombo.Option(ResizeType.SCALE_TO_MULTIPLE, [ + io.Int.Input("multiple", default=8, min=1, max=MAX_RESOLUTION, step=1), + ]), ]), io.Combo.Input("scale_method", options=cls.scale_methods, default="area"), ], @@ -442,6 +484,8 @@ class ResizeImageMaskNode(io.ComfyNode): return io.NodeOutput(scale_total_pixels(input, resize_type["megapixels"], scale_method)) elif selected_type == ResizeType.MATCH_SIZE: return io.NodeOutput(scale_match_size(input, resize_type["match"], scale_method, resize_type["crop"])) + elif selected_type == ResizeType.SCALE_TO_MULTIPLE: + return io.NodeOutput(scale_to_multiple_cover(input, resize_type["multiple"], scale_method)) raise ValueError(f"Unsupported resize type: {selected_type}") def batch_images(images: list[torch.Tensor]) -> torch.Tensor | None: From 7eb959ce934da914523b455f9a6e7e0662690325 Mon Sep 17 00:00:00 2001 From: Christian Byrne Date: Tue, 13 Jan 2026 18:03:16 -0800 Subject: [PATCH 447/647] fix: update ComfyUI repo reference to Comfy-Org/ComfyUI (#11858) --- .github/workflows/test-launch.yml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/.github/workflows/test-launch.yml b/.github/workflows/test-launch.yml index ef0d3f123..581c0474b 100644 --- a/.github/workflows/test-launch.yml +++ b/.github/workflows/test-launch.yml @@ -13,7 +13,7 @@ jobs: - name: Checkout ComfyUI uses: actions/checkout@v4 with: - repository: "comfyanonymous/ComfyUI" + repository: "Comfy-Org/ComfyUI" path: "ComfyUI" - uses: actions/setup-python@v4 with: From c9196f355ef5832daf55c4bbe8c6279dec509331 Mon Sep 17 00:00:00 2001 From: nomadoor <124905471+nomadoor@users.noreply.github.com> Date: Wed, 14 Jan 2026 11:25:09 +0900 Subject: [PATCH 448/647] Fix scale_shorter_dimension portrait check (#11862) --- comfy_extras/nodes_post_processing.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy_extras/nodes_post_processing.py b/comfy_extras/nodes_post_processing.py index 0433bbda2..2e559c35c 100644 --- a/comfy_extras/nodes_post_processing.py +++ b/comfy_extras/nodes_post_processing.py @@ -329,7 +329,7 @@ def scale_shorter_dimension(input: torch.Tensor, shorter_size: int, scale_method if height < width: width = round((width / height) * shorter_size) height = shorter_size - elif width > height: + elif width < height: height = round((height / width) * shorter_size) width = shorter_size else: From ac4d8ea9b32f56410860dccdb30ae50a1029d6fd Mon Sep 17 00:00:00 2001 From: Johnpaul Chiwetelu <49923152+Myestery@users.noreply.github.com> Date: Wed, 14 Jan 2026 04:39:22 +0100 Subject: [PATCH 449/647] feat: add CI container version bump automation (#11692) * feat: add CI container version bump automation Adds a workflow that triggers on releases to create PRs in the comfyui-ci-container repo, updating the ComfyUI version in the Dockerfile. Supports both release events and manual workflow dispatch for testing. * feat: add CI container version bump automation Adds a workflow that triggers on releases to create PRs in the comfyui-ci-container repo, updating the ComfyUI version in the Dockerfile. Supports both release events and manual workflow dispatch for testing. * ci: update CI container repository owner * refactor: rename `update-ci-container.yaml` workflow to `update-ci-container.yml` * Remove post-merge instructions from the CI container update workflow. --- .github/workflows/update-ci-container.yml | 59 +++++++++++++++++++++++ 1 file changed, 59 insertions(+) create mode 100644 .github/workflows/update-ci-container.yml diff --git a/.github/workflows/update-ci-container.yml b/.github/workflows/update-ci-container.yml new file mode 100644 index 000000000..f7972e056 --- /dev/null +++ b/.github/workflows/update-ci-container.yml @@ -0,0 +1,59 @@ +name: "CI: Update CI Container" + +on: + release: + types: [published] + workflow_dispatch: + inputs: + version: + description: 'ComfyUI version (e.g., v0.7.0)' + required: true + type: string + +jobs: + update-ci-container: + runs-on: ubuntu-latest + # Skip pre-releases unless manually triggered + if: github.event_name == 'workflow_dispatch' || !github.event.release.prerelease + steps: + - name: Get version + id: version + run: | + if [ "${{ github.event_name }}" = "release" ]; then + VERSION="${{ github.event.release.tag_name }}" + else + VERSION="${{ inputs.version }}" + fi + echo "version=$VERSION" >> $GITHUB_OUTPUT + + - name: Checkout comfyui-ci-container + uses: actions/checkout@v4 + with: + repository: comfy-org/comfyui-ci-container + token: ${{ secrets.CI_CONTAINER_PAT }} + + - name: Check current version + id: current + run: | + CURRENT=$(grep -oP 'ARG COMFYUI_VERSION=\K.*' Dockerfile || echo "unknown") + echo "current_version=$CURRENT" >> $GITHUB_OUTPUT + + - name: Update Dockerfile + run: | + VERSION="${{ steps.version.outputs.version }}" + sed -i "s/^ARG COMFYUI_VERSION=.*/ARG COMFYUI_VERSION=${VERSION}/" Dockerfile + + - name: Create Pull Request + id: create-pr + uses: peter-evans/create-pull-request@v7 + with: + token: ${{ secrets.CI_CONTAINER_PAT }} + branch: automation/comfyui-${{ steps.version.outputs.version }} + title: "chore: bump ComfyUI to ${{ steps.version.outputs.version }}" + body: | + Updates ComfyUI version from `${{ steps.current.outputs.current_version }}` to `${{ steps.version.outputs.version }}` + + **Triggered by:** ${{ github.event_name == 'release' && format('[Release {0}]({1})', github.event.release.tag_name, github.event.release.html_url) || 'Manual workflow dispatch' }} + + labels: automation + commit-message: "chore: bump ComfyUI to ${{ steps.version.outputs.version }}" From 712cca36a13db93a9fa1fde9b7b5f9a5b961209a Mon Sep 17 00:00:00 2001 From: Silver <65376327+silveroxides@users.noreply.github.com> Date: Wed, 14 Jan 2026 04:41:44 +0100 Subject: [PATCH 450/647] feat: throttle ProgressBar updates to reduce WebSocket flooding (#11504) --- comfy/utils.py | 31 ++++++++++++++++++++++++++++++- 1 file changed, 30 insertions(+), 1 deletion(-) diff --git a/comfy/utils.py b/comfy/utils.py index ffa98c9b1..fac13f128 100644 --- a/comfy/utils.py +++ b/comfy/utils.py @@ -30,6 +30,7 @@ from torch.nn.functional import interpolate from einops import rearrange from comfy.cli_args import args import json +import time MMAP_TORCH_FILES = args.mmap_torch_files DISABLE_MMAP = args.disable_mmap @@ -1097,6 +1098,10 @@ def set_progress_bar_global_hook(function): global PROGRESS_BAR_HOOK PROGRESS_BAR_HOOK = function +# Throttle settings for progress bar updates to reduce WebSocket flooding +PROGRESS_THROTTLE_MIN_INTERVAL = 0.1 # 100ms minimum between updates +PROGRESS_THROTTLE_MIN_PERCENT = 0.5 # 0.5% minimum progress change + class ProgressBar: def __init__(self, total, node_id=None): global PROGRESS_BAR_HOOK @@ -1104,6 +1109,8 @@ class ProgressBar: self.current = 0 self.hook = PROGRESS_BAR_HOOK self.node_id = node_id + self._last_update_time = 0.0 + self._last_sent_value = -1 def update_absolute(self, value, total=None, preview=None): if total is not None: @@ -1112,7 +1119,29 @@ class ProgressBar: value = self.total self.current = value if self.hook is not None: - self.hook(self.current, self.total, preview, node_id=self.node_id) + current_time = time.perf_counter() + is_first = (self._last_sent_value < 0) + is_final = (value >= self.total) + has_preview = (preview is not None) + + # Always send immediately for previews, first update, or final update + if has_preview or is_first or is_final: + self.hook(self.current, self.total, preview, node_id=self.node_id) + self._last_update_time = current_time + self._last_sent_value = value + return + + # Apply throttling for regular progress updates + if self.total > 0: + percent_changed = ((value - max(0, self._last_sent_value)) / self.total) * 100 + else: + percent_changed = 100 + time_elapsed = current_time - self._last_update_time + + if time_elapsed >= PROGRESS_THROTTLE_MIN_INTERVAL and percent_changed >= PROGRESS_THROTTLE_MIN_PERCENT: + self.hook(self.current, self.total, preview, node_id=self.node_id) + self._last_update_time = current_time + self._last_sent_value = value def update(self, value): self.update_absolute(self.current + value) From 6165c38cb58c40b15ade879b80051b6c9148587f Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 13 Jan 2026 21:49:38 -0800 Subject: [PATCH 451/647] Optimize nvfp4 lora applying. (#11866) This changes results a bit but it also speeds up things a lot. --- comfy/float.py | 56 ++++++++++++++++++++++++++++++++------- comfy/quant_ops.py | 2 +- comfy/supported_models.py | 2 +- 3 files changed, 49 insertions(+), 11 deletions(-) diff --git a/comfy/float.py b/comfy/float.py index 8c303bea0..88c47cd80 100644 --- a/comfy/float.py +++ b/comfy/float.py @@ -137,10 +137,44 @@ def to_blocked(input_matrix, flatten: bool = True) -> torch.Tensor: return rearranged.reshape(padded_rows, padded_cols) -def stochastic_round_quantize_nvfp4(x, per_tensor_scale, pad_16x, seed=0): +def stochastic_round_quantize_nvfp4_block(x, per_tensor_scale, generator): F4_E2M1_MAX = 6.0 F8_E4M3_MAX = 448.0 + orig_shape = x.shape + + block_size = 16 + + x = x.reshape(orig_shape[0], -1, block_size) + scaled_block_scales_fp8 = torch.clamp(((torch.amax(torch.abs(x), dim=-1)) / F4_E2M1_MAX) / per_tensor_scale.to(x.dtype), max=F8_E4M3_MAX).to(torch.float8_e4m3fn) + x = x / (per_tensor_scale.to(x.dtype) * scaled_block_scales_fp8.to(x.dtype)).unsqueeze(-1) + + x = x.view(orig_shape).nan_to_num() + data_lp = stochastic_float_to_fp4_e2m1(x, generator=generator) + return data_lp, scaled_block_scales_fp8 + + +def stochastic_round_quantize_nvfp4(x, per_tensor_scale, pad_16x, seed=0): + def roundup(x: int, multiple: int) -> int: + """Round up x to the nearest multiple.""" + return ((x + multiple - 1) // multiple) * multiple + + generator = torch.Generator(device=x.device) + generator.manual_seed(seed) + + # Handle padding + if pad_16x: + rows, cols = x.shape + padded_rows = roundup(rows, 16) + padded_cols = roundup(cols, 16) + if padded_rows != rows or padded_cols != cols: + x = torch.nn.functional.pad(x, (0, padded_cols - cols, 0, padded_rows - rows)) + + x, blocked_scaled = stochastic_round_quantize_nvfp4_block(x, per_tensor_scale, generator) + return x, to_blocked(blocked_scaled, flatten=False) + + +def stochastic_round_quantize_nvfp4_by_block(x, per_tensor_scale, pad_16x, seed=0, block_size=4096 * 4096): def roundup(x: int, multiple: int) -> int: """Round up x to the nearest multiple.""" return ((x + multiple - 1) // multiple) * multiple @@ -158,16 +192,20 @@ def stochastic_round_quantize_nvfp4(x, per_tensor_scale, pad_16x, seed=0): # what we want to produce. If we pad here, we want the padded output. orig_shape = x.shape - block_size = 16 + orig_shape = list(orig_shape) - x = x.reshape(orig_shape[0], -1, block_size) - scaled_block_scales_fp8 = torch.clamp(((torch.amax(torch.abs(x), dim=-1)) / F4_E2M1_MAX) / per_tensor_scale.to(x.dtype), max=F8_E4M3_MAX).to(torch.float8_e4m3fn) - x /= (per_tensor_scale.to(x.dtype) * scaled_block_scales_fp8.to(x.dtype)).unsqueeze(-1) + output_fp4 = torch.empty(orig_shape[:-1] + [orig_shape[-1] // 2], dtype=torch.uint8, device=x.device) + output_block = torch.empty(orig_shape[:-1] + [orig_shape[-1] // 16], dtype=torch.float8_e4m3fn, device=x.device) generator = torch.Generator(device=x.device) generator.manual_seed(seed) - x = x.view(orig_shape).nan_to_num() - data_lp = stochastic_float_to_fp4_e2m1(x, generator=generator) - blocked_scales = to_blocked(scaled_block_scales_fp8, flatten=False) - return data_lp, blocked_scales + num_slices = max(1, (x.numel() / block_size)) + slice_size = max(1, (round(x.shape[0] / num_slices))) + + for i in range(0, x.shape[0], slice_size): + fp4, block = stochastic_round_quantize_nvfp4_block(x[i: i + slice_size], per_tensor_scale, generator=generator) + output_fp4[i:i + slice_size].copy_(fp4) + output_block[i:i + slice_size].copy_(block) + + return output_fp4, to_blocked(output_block, flatten=False) diff --git a/comfy/quant_ops.py b/comfy/quant_ops.py index 7a61203c3..15a4f457b 100644 --- a/comfy/quant_ops.py +++ b/comfy/quant_ops.py @@ -104,7 +104,7 @@ class TensorCoreNVFP4Layout(_CKNvfp4Layout): needs_padding = padded_shape != orig_shape if stochastic_rounding > 0: - qdata, block_scale = comfy.float.stochastic_round_quantize_nvfp4(tensor, scale, pad_16x=needs_padding, seed=stochastic_rounding) + qdata, block_scale = comfy.float.stochastic_round_quantize_nvfp4_by_block(tensor, scale, pad_16x=needs_padding, seed=stochastic_rounding) else: qdata, block_scale = ck.quantize_nvfp4(tensor, scale, pad_16x=needs_padding) diff --git a/comfy/supported_models.py b/comfy/supported_models.py index 1bf54f13f..2c4c6b8fc 100644 --- a/comfy/supported_models.py +++ b/comfy/supported_models.py @@ -1042,7 +1042,7 @@ class ZImage(Lumina2): "shift": 3.0, } - memory_usage_factor = 2.0 + memory_usage_factor = 2.8 supported_inference_dtypes = [torch.bfloat16, torch.float32] From d1504404662dfce6e401422701c2a7e24057b1b5 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 14 Jan 2026 10:54:50 -0800 Subject: [PATCH 452/647] Fix VAELoader (#11880) --- nodes.py | 1 + 1 file changed, 1 insertion(+) diff --git a/nodes.py b/nodes.py index 90c5f2a6e..aa8572446 100644 --- a/nodes.py +++ b/nodes.py @@ -788,6 +788,7 @@ class VAELoader: #TODO: scale factor? def load_vae(self, vae_name): + metadata = None if vae_name == "pixel_space": sd = {} sd["pixel_space_vae"] = torch.tensor(1.0) From 07f2462eae7fa2daa34971dd1b15fd525686e958 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Wed, 14 Jan 2026 21:25:38 +0200 Subject: [PATCH 453/647] feat(api-nodes): add Meshy 3D nodes (#11843) * feat(api-nodes): add Meshy 3D nodes * rebased, added JSONata price badges --- comfy_api_nodes/apis/meshy.py | 160 +++++ comfy_api_nodes/nodes_meshy.py | 790 +++++++++++++++++++++++++ comfy_api_nodes/util/upload_helpers.py | 23 +- nodes.py | 1 + 4 files changed, 969 insertions(+), 5 deletions(-) create mode 100644 comfy_api_nodes/apis/meshy.py create mode 100644 comfy_api_nodes/nodes_meshy.py diff --git a/comfy_api_nodes/apis/meshy.py b/comfy_api_nodes/apis/meshy.py new file mode 100644 index 000000000..be46d0d58 --- /dev/null +++ b/comfy_api_nodes/apis/meshy.py @@ -0,0 +1,160 @@ +from typing import TypedDict + +from pydantic import BaseModel, Field + +from comfy_api.latest import Input + + +class InputShouldRemesh(TypedDict): + should_remesh: str + topology: str + target_polycount: int + + +class InputShouldTexture(TypedDict): + should_texture: str + enable_pbr: bool + texture_prompt: str + texture_image: Input.Image | None + + +class MeshyTaskResponse(BaseModel): + result: str = Field(...) + + +class MeshyTextToModelRequest(BaseModel): + mode: str = Field("preview") + prompt: str = Field(..., max_length=600) + art_style: str = Field(..., description="'realistic' or 'sculpture'") + ai_model: str = Field(...) + topology: str | None = Field(..., description="'quad' or 'triangle'") + target_polycount: int | None = Field(..., ge=100, le=300000) + should_remesh: bool = Field( + True, + description="False returns the original mesh, ignoring topology and polycount.", + ) + symmetry_mode: str = Field(..., description="'auto', 'off' or 'on'") + pose_mode: str = Field(...) + seed: int = Field(...) + moderation: bool = Field(False) + + +class MeshyRefineTask(BaseModel): + mode: str = Field("refine") + preview_task_id: str = Field(...) + enable_pbr: bool | None = Field(...) + texture_prompt: str | None = Field(...) + texture_image_url: str | None = Field(...) + ai_model: str = Field(...) + moderation: bool = Field(False) + + +class MeshyImageToModelRequest(BaseModel): + image_url: str = Field(...) + ai_model: str = Field(...) + topology: str | None = Field(..., description="'quad' or 'triangle'") + target_polycount: int | None = Field(..., ge=100, le=300000) + symmetry_mode: str = Field(..., description="'auto', 'off' or 'on'") + should_remesh: bool = Field( + True, + description="False returns the original mesh, ignoring topology and polycount.", + ) + should_texture: bool = Field(...) + enable_pbr: bool | None = Field(...) + pose_mode: str = Field(...) + texture_prompt: str | None = Field(None, max_length=600) + texture_image_url: str | None = Field(None) + seed: int = Field(...) + moderation: bool = Field(False) + + +class MeshyMultiImageToModelRequest(BaseModel): + image_urls: list[str] = Field(...) + ai_model: str = Field(...) + topology: str | None = Field(..., description="'quad' or 'triangle'") + target_polycount: int | None = Field(..., ge=100, le=300000) + symmetry_mode: str = Field(..., description="'auto', 'off' or 'on'") + should_remesh: bool = Field( + True, + description="False returns the original mesh, ignoring topology and polycount.", + ) + should_texture: bool = Field(...) + enable_pbr: bool | None = Field(...) + pose_mode: str = Field(...) + texture_prompt: str | None = Field(None, max_length=600) + texture_image_url: str | None = Field(None) + seed: int = Field(...) + moderation: bool = Field(False) + + +class MeshyRiggingRequest(BaseModel): + input_task_id: str = Field(...) + height_meters: float = Field(...) + texture_image_url: str | None = Field(...) + + +class MeshyAnimationRequest(BaseModel): + rig_task_id: str = Field(...) + action_id: int = Field(...) + + +class MeshyTextureRequest(BaseModel): + input_task_id: str = Field(...) + ai_model: str = Field(...) + enable_original_uv: bool = Field(...) + enable_pbr: bool = Field(...) + text_style_prompt: str | None = Field(...) + image_style_url: str | None = Field(...) + + +class MeshyModelsUrls(BaseModel): + glb: str = Field("") + + +class MeshyRiggedModelsUrls(BaseModel): + rigged_character_glb_url: str = Field("") + + +class MeshyAnimatedModelsUrls(BaseModel): + animation_glb_url: str = Field("") + + +class MeshyResultTextureUrls(BaseModel): + base_color: str = Field(...) + metallic: str | None = Field(None) + normal: str | None = Field(None) + roughness: str | None = Field(None) + + +class MeshyTaskError(BaseModel): + message: str | None = Field(None) + + +class MeshyModelResult(BaseModel): + id: str = Field(...) + type: str = Field(...) + model_urls: MeshyModelsUrls = Field(MeshyModelsUrls()) + thumbnail_url: str = Field(...) + video_url: str | None = Field(None) + status: str = Field(...) + progress: int = Field(0) + texture_urls: list[MeshyResultTextureUrls] | None = Field([]) + task_error: MeshyTaskError | None = Field(None) + + +class MeshyRiggedResult(BaseModel): + id: str = Field(...) + type: str = Field(...) + status: str = Field(...) + progress: int = Field(0) + result: MeshyRiggedModelsUrls = Field(MeshyRiggedModelsUrls()) + task_error: MeshyTaskError | None = Field(None) + + +class MeshyAnimationResult(BaseModel): + id: str = Field(...) + type: str = Field(...) + status: str = Field(...) + progress: int = Field(0) + result: MeshyAnimatedModelsUrls = Field(MeshyAnimatedModelsUrls()) + task_error: MeshyTaskError | None = Field(None) diff --git a/comfy_api_nodes/nodes_meshy.py b/comfy_api_nodes/nodes_meshy.py new file mode 100644 index 000000000..740607983 --- /dev/null +++ b/comfy_api_nodes/nodes_meshy.py @@ -0,0 +1,790 @@ +import os + +from typing_extensions import override + +from comfy_api.latest import IO, ComfyExtension, Input +from comfy_api_nodes.apis.meshy import ( + InputShouldRemesh, + InputShouldTexture, + MeshyAnimationRequest, + MeshyAnimationResult, + MeshyImageToModelRequest, + MeshyModelResult, + MeshyMultiImageToModelRequest, + MeshyRefineTask, + MeshyRiggedResult, + MeshyRiggingRequest, + MeshyTaskResponse, + MeshyTextToModelRequest, + MeshyTextureRequest, +) +from comfy_api_nodes.util import ( + ApiEndpoint, + download_url_to_bytesio, + poll_op, + sync_op, + upload_images_to_comfyapi, + validate_string, +) +from folder_paths import get_output_directory + + +class MeshyTextToModelNode(IO.ComfyNode): + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="MeshyTextToModelNode", + display_name="Meshy: Text to Model", + category="api node/3d/Meshy", + inputs=[ + IO.Combo.Input("model", options=["latest"]), + IO.String.Input("prompt", multiline=True, default=""), + IO.Combo.Input("style", options=["realistic", "sculpture"]), + IO.DynamicCombo.Input( + "should_remesh", + options=[ + IO.DynamicCombo.Option( + "true", + [ + IO.Combo.Input("topology", options=["triangle", "quad"]), + IO.Int.Input( + "target_polycount", + default=300000, + min=100, + max=300000, + display_mode=IO.NumberDisplay.number, + ), + ], + ), + IO.DynamicCombo.Option("false", []), + ], + tooltip="When set to false, returns an unprocessed triangular mesh.", + ), + IO.Combo.Input("symmetry_mode", options=["auto", "on", "off"]), + IO.Combo.Input( + "pose_mode", + options=["", "A-pose", "T-pose"], + tooltip="Specify the pose mode for the generated model.", + ), + IO.Int.Input( + "seed", + default=0, + min=0, + max=2147483647, + display_mode=IO.NumberDisplay.number, + control_after_generate=True, + tooltip="Seed controls whether the node should re-run; " + "results are non-deterministic regardless of seed.", + ), + ], + outputs=[ + IO.String.Output(display_name="model_file"), + IO.Custom("MESHY_TASK_ID").Output(display_name="meshy_task_id"), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + is_output_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.8}""", + ), + ) + + @classmethod + async def execute( + cls, + model: str, + prompt: str, + style: str, + should_remesh: InputShouldRemesh, + symmetry_mode: str, + pose_mode: str, + seed: int, + ) -> IO.NodeOutput: + validate_string(prompt, field_name="prompt", min_length=1, max_length=600) + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/meshy/openapi/v2/text-to-3d", method="POST"), + response_model=MeshyTaskResponse, + data=MeshyTextToModelRequest( + prompt=prompt, + art_style=style, + ai_model=model, + topology=should_remesh.get("topology", None), + target_polycount=should_remesh.get("target_polycount", None), + should_remesh=should_remesh["should_remesh"] == "true", + symmetry_mode=symmetry_mode, + pose_mode=pose_mode.lower(), + seed=seed, + ), + ) + result = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/meshy/openapi/v2/text-to-3d/{response.result}"), + response_model=MeshyModelResult, + status_extractor=lambda r: r.status, + progress_extractor=lambda r: r.progress, + ) + model_file = f"meshy_model_{response.result}.glb" + await download_url_to_bytesio(result.model_urls.glb, os.path.join(get_output_directory(), model_file)) + return IO.NodeOutput(model_file, response.result) + + +class MeshyRefineNode(IO.ComfyNode): + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="MeshyRefineNode", + display_name="Meshy: Refine Draft Model", + category="api node/3d/Meshy", + description="Refine a previously created draft model.", + inputs=[ + IO.Combo.Input("model", options=["latest"]), + IO.Custom("MESHY_TASK_ID").Input("meshy_task_id"), + IO.Boolean.Input( + "enable_pbr", + default=False, + tooltip="Generate PBR Maps (metallic, roughness, normal) in addition to the base color. " + "Note: this should be set to false when using Sculpture style, " + "as Sculpture style generates its own set of PBR maps.", + ), + IO.String.Input( + "texture_prompt", + default="", + multiline=True, + tooltip="Provide a text prompt to guide the texturing process. " + "Maximum 600 characters. Cannot be used at the same time as 'texture_image'.", + ), + IO.Image.Input( + "texture_image", + tooltip="Only one of 'texture_image' or 'texture_prompt' may be used at the same time.", + optional=True, + ), + ], + outputs=[ + IO.String.Output(display_name="model_file"), + IO.Custom("MESHY_TASK_ID").Output(display_name="meshy_task_id"), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + is_output_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.4}""", + ), + ) + + @classmethod + async def execute( + cls, + model: str, + meshy_task_id: str, + enable_pbr: bool, + texture_prompt: str, + texture_image: Input.Image | None = None, + ) -> IO.NodeOutput: + if texture_prompt and texture_image is not None: + raise ValueError("texture_prompt and texture_image cannot be used at the same time") + texture_image_url = None + if texture_prompt: + validate_string(texture_prompt, field_name="texture_prompt", max_length=600) + if texture_image is not None: + texture_image_url = (await upload_images_to_comfyapi(cls, texture_image, wait_label="Uploading texture"))[0] + response = await sync_op( + cls, + endpoint=ApiEndpoint(path="/proxy/meshy/openapi/v2/text-to-3d", method="POST"), + response_model=MeshyTaskResponse, + data=MeshyRefineTask( + preview_task_id=meshy_task_id, + enable_pbr=enable_pbr, + texture_prompt=texture_prompt if texture_prompt else None, + texture_image_url=texture_image_url, + ai_model=model, + ), + ) + result = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/meshy/openapi/v2/text-to-3d/{response.result}"), + response_model=MeshyModelResult, + status_extractor=lambda r: r.status, + progress_extractor=lambda r: r.progress, + ) + model_file = f"meshy_model_{response.result}.glb" + await download_url_to_bytesio(result.model_urls.glb, os.path.join(get_output_directory(), model_file)) + return IO.NodeOutput(model_file, response.result) + + +class MeshyImageToModelNode(IO.ComfyNode): + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="MeshyImageToModelNode", + display_name="Meshy: Image to Model", + category="api node/3d/Meshy", + inputs=[ + IO.Combo.Input("model", options=["latest"]), + IO.Image.Input("image"), + IO.DynamicCombo.Input( + "should_remesh", + options=[ + IO.DynamicCombo.Option( + "true", + [ + IO.Combo.Input("topology", options=["triangle", "quad"]), + IO.Int.Input( + "target_polycount", + default=300000, + min=100, + max=300000, + display_mode=IO.NumberDisplay.number, + ), + ], + ), + IO.DynamicCombo.Option("false", []), + ], + tooltip="When set to false, returns an unprocessed triangular mesh.", + ), + IO.Combo.Input("symmetry_mode", options=["auto", "on", "off"]), + IO.DynamicCombo.Input( + "should_texture", + options=[ + IO.DynamicCombo.Option( + "true", + [ + IO.Boolean.Input( + "enable_pbr", + default=False, + tooltip="Generate PBR Maps (metallic, roughness, normal) " + "in addition to the base color.", + ), + IO.String.Input( + "texture_prompt", + default="", + multiline=True, + tooltip="Provide a text prompt to guide the texturing process. " + "Maximum 600 characters. Cannot be used at the same time as 'texture_image'.", + ), + IO.Image.Input( + "texture_image", + tooltip="Only one of 'texture_image' or 'texture_prompt' " + "may be used at the same time.", + optional=True, + ), + ], + ), + IO.DynamicCombo.Option("false", []), + ], + tooltip="Determines whether textures are generated. " + "Setting it to false skips the texture phase and returns a mesh without textures.", + ), + IO.Combo.Input( + "pose_mode", + options=["", "A-pose", "T-pose"], + tooltip="Specify the pose mode for the generated model.", + ), + IO.Int.Input( + "seed", + default=0, + min=0, + max=2147483647, + display_mode=IO.NumberDisplay.number, + control_after_generate=True, + tooltip="Seed controls whether the node should re-run; " + "results are non-deterministic regardless of seed.", + ), + ], + outputs=[ + IO.String.Output(display_name="model_file"), + IO.Custom("MESHY_TASK_ID").Output(display_name="meshy_task_id"), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + is_output_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["should_texture"]), + expr=""" + ( + $prices := {"true": 1.2, "false": 0.8}; + {"type":"usd","usd": $lookup($prices, widgets.should_texture)} + ) + """, + ), + ) + + @classmethod + async def execute( + cls, + model: str, + image: Input.Image, + should_remesh: InputShouldRemesh, + symmetry_mode: str, + should_texture: InputShouldTexture, + pose_mode: str, + seed: int, + ) -> IO.NodeOutput: + texture = should_texture["should_texture"] == "true" + texture_image_url = texture_prompt = None + if texture: + if should_texture["texture_prompt"] and should_texture["texture_image"] is not None: + raise ValueError("texture_prompt and texture_image cannot be used at the same time") + if should_texture["texture_prompt"]: + validate_string(should_texture["texture_prompt"], field_name="texture_prompt", max_length=600) + texture_prompt = should_texture["texture_prompt"] + if should_texture["texture_image"] is not None: + texture_image_url = ( + await upload_images_to_comfyapi( + cls, should_texture["texture_image"], wait_label="Uploading texture" + ) + )[0] + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/meshy/openapi/v1/image-to-3d", method="POST"), + response_model=MeshyTaskResponse, + data=MeshyImageToModelRequest( + image_url=(await upload_images_to_comfyapi(cls, image, wait_label="Uploading base image"))[0], + ai_model=model, + topology=should_remesh.get("topology", None), + target_polycount=should_remesh.get("target_polycount", None), + symmetry_mode=symmetry_mode, + should_remesh=should_remesh["should_remesh"] == "true", + should_texture=texture, + enable_pbr=should_texture.get("enable_pbr", None), + pose_mode=pose_mode.lower(), + texture_prompt=texture_prompt, + texture_image_url=texture_image_url, + seed=seed, + ), + ) + result = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/meshy/openapi/v1/image-to-3d/{response.result}"), + response_model=MeshyModelResult, + status_extractor=lambda r: r.status, + progress_extractor=lambda r: r.progress, + ) + model_file = f"meshy_model_{response.result}.glb" + await download_url_to_bytesio(result.model_urls.glb, os.path.join(get_output_directory(), model_file)) + return IO.NodeOutput(model_file, response.result) + + +class MeshyMultiImageToModelNode(IO.ComfyNode): + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="MeshyMultiImageToModelNode", + display_name="Meshy: Multi-Image to Model", + category="api node/3d/Meshy", + inputs=[ + IO.Combo.Input("model", options=["latest"]), + IO.Autogrow.Input( + "images", + template=IO.Autogrow.TemplatePrefix(IO.Image.Input("image"), prefix="image", min=2, max=4), + ), + IO.DynamicCombo.Input( + "should_remesh", + options=[ + IO.DynamicCombo.Option( + "true", + [ + IO.Combo.Input("topology", options=["triangle", "quad"]), + IO.Int.Input( + "target_polycount", + default=300000, + min=100, + max=300000, + display_mode=IO.NumberDisplay.number, + ), + ], + ), + IO.DynamicCombo.Option("false", []), + ], + tooltip="When set to false, returns an unprocessed triangular mesh.", + ), + IO.Combo.Input("symmetry_mode", options=["auto", "on", "off"]), + IO.DynamicCombo.Input( + "should_texture", + options=[ + IO.DynamicCombo.Option( + "true", + [ + IO.Boolean.Input( + "enable_pbr", + default=False, + tooltip="Generate PBR Maps (metallic, roughness, normal) " + "in addition to the base color.", + ), + IO.String.Input( + "texture_prompt", + default="", + multiline=True, + tooltip="Provide a text prompt to guide the texturing process. " + "Maximum 600 characters. Cannot be used at the same time as 'texture_image'.", + ), + IO.Image.Input( + "texture_image", + tooltip="Only one of 'texture_image' or 'texture_prompt' " + "may be used at the same time.", + optional=True, + ), + ], + ), + IO.DynamicCombo.Option("false", []), + ], + tooltip="Determines whether textures are generated. " + "Setting it to false skips the texture phase and returns a mesh without textures.", + ), + IO.Combo.Input( + "pose_mode", + options=["", "A-pose", "T-pose"], + tooltip="Specify the pose mode for the generated model.", + ), + IO.Int.Input( + "seed", + default=0, + min=0, + max=2147483647, + display_mode=IO.NumberDisplay.number, + control_after_generate=True, + tooltip="Seed controls whether the node should re-run; " + "results are non-deterministic regardless of seed.", + ), + ], + outputs=[ + IO.String.Output(display_name="model_file"), + IO.Custom("MESHY_TASK_ID").Output(display_name="meshy_task_id"), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + is_output_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["should_texture"]), + expr=""" + ( + $prices := {"true": 0.6, "false": 0.2}; + {"type":"usd","usd": $lookup($prices, widgets.should_texture)} + ) + """, + ), + ) + + @classmethod + async def execute( + cls, + model: str, + images: IO.Autogrow.Type, + should_remesh: InputShouldRemesh, + symmetry_mode: str, + should_texture: InputShouldTexture, + pose_mode: str, + seed: int, + ) -> IO.NodeOutput: + texture = should_texture["should_texture"] == "true" + texture_image_url = texture_prompt = None + if texture: + if should_texture["texture_prompt"] and should_texture["texture_image"] is not None: + raise ValueError("texture_prompt and texture_image cannot be used at the same time") + if should_texture["texture_prompt"]: + validate_string(should_texture["texture_prompt"], field_name="texture_prompt", max_length=600) + texture_prompt = should_texture["texture_prompt"] + if should_texture["texture_image"] is not None: + texture_image_url = ( + await upload_images_to_comfyapi( + cls, should_texture["texture_image"], wait_label="Uploading texture" + ) + )[0] + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/meshy/openapi/v1/multi-image-to-3d", method="POST"), + response_model=MeshyTaskResponse, + data=MeshyMultiImageToModelRequest( + image_urls=await upload_images_to_comfyapi( + cls, list(images.values()), wait_label="Uploading base images" + ), + ai_model=model, + topology=should_remesh.get("topology", None), + target_polycount=should_remesh.get("target_polycount", None), + symmetry_mode=symmetry_mode, + should_remesh=should_remesh["should_remesh"] == "true", + should_texture=texture, + enable_pbr=should_texture.get("enable_pbr", None), + pose_mode=pose_mode.lower(), + texture_prompt=texture_prompt, + texture_image_url=texture_image_url, + seed=seed, + ), + ) + result = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/meshy/openapi/v1/multi-image-to-3d/{response.result}"), + response_model=MeshyModelResult, + status_extractor=lambda r: r.status, + progress_extractor=lambda r: r.progress, + ) + model_file = f"meshy_model_{response.result}.glb" + await download_url_to_bytesio(result.model_urls.glb, os.path.join(get_output_directory(), model_file)) + return IO.NodeOutput(model_file, response.result) + + +class MeshyRigModelNode(IO.ComfyNode): + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="MeshyRigModelNode", + display_name="Meshy: Rig Model", + category="api node/3d/Meshy", + description="Provides a rigged character in standard formats. " + "Auto-rigging is currently not suitable for untextured meshes, non-humanoid assets, " + "or humanoid assets with unclear limb and body structure.", + inputs=[ + IO.Custom("MESHY_TASK_ID").Input("meshy_task_id"), + IO.Float.Input( + "height_meters", + min=0.1, + max=15.0, + default=1.7, + tooltip="The approximate height of the character model in meters. " + "This aids in scaling and rigging accuracy.", + ), + IO.Image.Input( + "texture_image", + tooltip="The model's UV-unwrapped base color texture image.", + optional=True, + ), + ], + outputs=[ + IO.String.Output(display_name="model_file"), + IO.Custom("MESHY_RIGGED_TASK_ID").Output(display_name="rig_task_id"), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + is_output_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.2}""", + ), + ) + + @classmethod + async def execute( + cls, + meshy_task_id: str, + height_meters: float, + texture_image: Input.Image | None = None, + ) -> IO.NodeOutput: + texture_image_url = None + if texture_image is not None: + texture_image_url = (await upload_images_to_comfyapi(cls, texture_image, wait_label="Uploading texture"))[0] + response = await sync_op( + cls, + endpoint=ApiEndpoint(path="/proxy/meshy/openapi/v1/rigging", method="POST"), + response_model=MeshyTaskResponse, + data=MeshyRiggingRequest( + input_task_id=meshy_task_id, + height_meters=height_meters, + texture_image_url=texture_image_url, + ), + ) + result = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/meshy/openapi/v1/rigging/{response.result}"), + response_model=MeshyRiggedResult, + status_extractor=lambda r: r.status, + progress_extractor=lambda r: r.progress, + ) + model_file = f"meshy_model_{response.result}.glb" + await download_url_to_bytesio( + result.result.rigged_character_glb_url, os.path.join(get_output_directory(), model_file) + ) + return IO.NodeOutput(model_file, response.result) + + +class MeshyAnimateModelNode(IO.ComfyNode): + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="MeshyAnimateModelNode", + display_name="Meshy: Animate Model", + category="api node/3d/Meshy", + description="Apply a specific animation action to a previously rigged character.", + inputs=[ + IO.Custom("MESHY_RIGGED_TASK_ID").Input("rig_task_id"), + IO.Int.Input( + "action_id", + default=0, + min=0, + max=696, + tooltip="Visit https://docs.meshy.ai/en/api/animation-library for a list of available values.", + ), + ], + outputs=[ + IO.String.Output(display_name="model_file"), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + is_output_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.12}""", + ), + ) + + @classmethod + async def execute( + cls, + rig_task_id: str, + action_id: int, + ) -> IO.NodeOutput: + response = await sync_op( + cls, + endpoint=ApiEndpoint(path="/proxy/meshy/openapi/v1/animations", method="POST"), + response_model=MeshyTaskResponse, + data=MeshyAnimationRequest( + rig_task_id=rig_task_id, + action_id=action_id, + ), + ) + result = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/meshy/openapi/v1/animations/{response.result}"), + response_model=MeshyAnimationResult, + status_extractor=lambda r: r.status, + progress_extractor=lambda r: r.progress, + ) + model_file = f"meshy_model_{response.result}.glb" + await download_url_to_bytesio(result.result.animation_glb_url, os.path.join(get_output_directory(), model_file)) + return IO.NodeOutput(model_file, response.result) + + +class MeshyTextureNode(IO.ComfyNode): + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="MeshyTextureNode", + display_name="Meshy: Texture Model", + category="api node/3d/Meshy", + inputs=[ + IO.Combo.Input("model", options=["latest"]), + IO.Custom("MESHY_TASK_ID").Input("meshy_task_id"), + IO.Boolean.Input( + "enable_original_uv", + default=True, + tooltip="Use the original UV of the model instead of generating new UVs. " + "When enabled, Meshy preserves existing textures from the uploaded model. " + "If the model has no original UV, the quality of the output might not be as good.", + ), + IO.Boolean.Input("pbr", default=False), + IO.String.Input( + "text_style_prompt", + default="", + multiline=True, + tooltip="Describe your desired texture style of the object using text. Maximum 600 characters." + "Maximum 600 characters. Cannot be used at the same time as 'image_style'.", + ), + IO.Image.Input( + "image_style", + optional=True, + tooltip="A 2d image to guide the texturing process. " + "Can not be used at the same time with 'text_style_prompt'.", + ), + ], + outputs=[ + IO.String.Output(display_name="model_file"), + IO.Custom("MODEL_TASK_ID").Output(display_name="meshy_task_id"), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + is_output_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.4}""", + ), + ) + + @classmethod + async def execute( + cls, + model: str, + meshy_task_id: str, + enable_original_uv: bool, + pbr: bool, + text_style_prompt: str, + image_style: Input.Image | None = None, + ) -> IO.NodeOutput: + if text_style_prompt and image_style is not None: + raise ValueError("text_style_prompt and image_style cannot be used at the same time") + if not text_style_prompt and image_style is None: + raise ValueError("Either text_style_prompt or image_style is required") + image_style_url = None + if image_style is not None: + image_style_url = (await upload_images_to_comfyapi(cls, image_style, wait_label="Uploading style"))[0] + response = await sync_op( + cls, + endpoint=ApiEndpoint(path="/proxy/meshy/openapi/v1/retexture", method="POST"), + response_model=MeshyTaskResponse, + data=MeshyTextureRequest( + input_task_id=meshy_task_id, + ai_model=model, + enable_original_uv=enable_original_uv, + enable_pbr=pbr, + text_style_prompt=text_style_prompt if text_style_prompt else None, + image_style_url=image_style_url, + ), + ) + result = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/meshy/openapi/v1/retexture/{response.result}"), + response_model=MeshyModelResult, + status_extractor=lambda r: r.status, + progress_extractor=lambda r: r.progress, + ) + model_file = f"meshy_model_{response.result}.glb" + await download_url_to_bytesio(result.model_urls.glb, os.path.join(get_output_directory(), model_file)) + return IO.NodeOutput(model_file, response.result) + + +class MeshyExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[IO.ComfyNode]]: + return [ + MeshyTextToModelNode, + MeshyRefineNode, + MeshyImageToModelNode, + MeshyMultiImageToModelNode, + MeshyRigModelNode, + MeshyAnimateModelNode, + MeshyTextureNode, + ] + + +async def comfy_entrypoint() -> MeshyExtension: + return MeshyExtension() diff --git a/comfy_api_nodes/util/upload_helpers.py b/comfy_api_nodes/util/upload_helpers.py index cea0d1203..2794be35c 100644 --- a/comfy_api_nodes/util/upload_helpers.py +++ b/comfy_api_nodes/util/upload_helpers.py @@ -43,7 +43,7 @@ class UploadResponse(BaseModel): async def upload_images_to_comfyapi( cls: type[IO.ComfyNode], - image: torch.Tensor, + image: torch.Tensor | list[torch.Tensor], *, max_images: int = 8, mime_type: str | None = None, @@ -55,15 +55,28 @@ async def upload_images_to_comfyapi( Uploads images to ComfyUI API and returns download URLs. To upload multiple images, stack them in the batch dimension first. """ + tensors: list[torch.Tensor] = [] + if isinstance(image, list): + for img in image: + is_batch = len(img.shape) > 3 + if is_batch: + tensors.extend(img[i] for i in range(img.shape[0])) + else: + tensors.append(img) + else: + is_batch = len(image.shape) > 3 + if is_batch: + tensors.extend(image[i] for i in range(image.shape[0])) + else: + tensors.append(image) + # if batched, try to upload each file if max_images is greater than 0 download_urls: list[str] = [] - is_batch = len(image.shape) > 3 - batch_len = image.shape[0] if is_batch else 1 - num_to_upload = min(batch_len, max_images) + num_to_upload = min(len(tensors), max_images) batch_start_ts = time.monotonic() for idx in range(num_to_upload): - tensor = image[idx] if is_batch else image + tensor = tensors[idx] img_io = tensor_to_bytesio(tensor, total_pixels=total_pixels, mime_type=mime_type) effective_label = wait_label diff --git a/nodes.py b/nodes.py index aa8572446..f19d5fd1c 100644 --- a/nodes.py +++ b/nodes.py @@ -2401,6 +2401,7 @@ async def init_builtin_api_nodes(): "nodes_sora.py", "nodes_topaz.py", "nodes_tripo.py", + "nodes_meshy.py", "nodes_moonvalley.py", "nodes_rodin.py", "nodes_gemini.py", From 80441eb15e807aa280fb462cbb43d14191344ba4 Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Wed, 14 Jan 2026 14:53:16 -0800 Subject: [PATCH 454/647] utils: fix lanczos grayscale upscaling (#11873) --- comfy/utils.py | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/comfy/utils.py b/comfy/utils.py index fac13f128..2e33a4258 100644 --- a/comfy/utils.py +++ b/comfy/utils.py @@ -929,7 +929,9 @@ def bislerp(samples, width, height): return result.to(orig_dtype) def lanczos(samples, width, height): - images = [Image.fromarray(np.clip(255. * image.movedim(0, -1).cpu().numpy(), 0, 255).astype(np.uint8)) for image in samples] + #the below API is strict and expects grayscale to be squeezed + samples = samples.squeeze(1) if samples.shape[1] == 1 else samples.movedim(1, -1) + images = [Image.fromarray(np.clip(255. * image.cpu().numpy(), 0, 255).astype(np.uint8)) for image in samples] images = [image.resize((width, height), resample=Image.Resampling.LANCZOS) for image in images] images = [torch.from_numpy(np.array(image).astype(np.float32) / 255.0).movedim(-1, 0) for image in images] result = torch.stack(images) From be518db5a7daa6010fb1c312c0832b9833a71d10 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Jukka=20Sepp=C3=A4nen?= <40791699+kijai@users.noreply.github.com> Date: Thu, 15 Jan 2026 00:54:04 +0200 Subject: [PATCH 455/647] Remove extraneous clip missing warnings when loading LTX2 embeddings_connector weights (#11874) --- comfy/text_encoders/lt.py | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-) diff --git a/comfy/text_encoders/lt.py b/comfy/text_encoders/lt.py index 776e25e97..c33c77db7 100644 --- a/comfy/text_encoders/lt.py +++ b/comfy/text_encoders/lt.py @@ -118,8 +118,9 @@ class LTXAVTEModel(torch.nn.Module): sdo = comfy.utils.state_dict_prefix_replace(sd, {"text_embedding_projection.aggregate_embed.weight": "text_embedding_projection.weight", "model.diffusion_model.video_embeddings_connector.": "video_embeddings_connector.", "model.diffusion_model.audio_embeddings_connector.": "audio_embeddings_connector."}, filter_keys=True) if len(sdo) == 0: sdo = sd - - return self.load_state_dict(sdo, strict=False) + missing, unexpected = self.load_state_dict(sdo, strict=False) + missing = [k for k in missing if not k.startswith("gemma3_12b.")] # filter out keys that belong to the main gemma model + return (missing, unexpected) def memory_estimation_function(self, token_weight_pairs, device=None): constant = 6.0 From 3b832231bb81024d80bbe31b7d7e51e07b633beb Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Thu, 15 Jan 2026 07:33:15 -0800 Subject: [PATCH 456/647] Flux2 Klein support. (#11890) --- comfy/sd.py | 15 +++++++-- comfy/text_encoders/flux.py | 59 +++++++++++++++++++++++++++++++++++- comfy/text_encoders/llama.py | 31 +++++++++++++++++++ 3 files changed, 102 insertions(+), 3 deletions(-) diff --git a/comfy/sd.py b/comfy/sd.py index b689c0dfc..77700dfd3 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -1014,6 +1014,7 @@ class CLIPType(Enum): KANDINSKY5 = 22 KANDINSKY5_IMAGE = 23 NEWBIE = 24 + FLUX2 = 25 def load_clip(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DIFFUSION, model_options={}): @@ -1046,6 +1047,7 @@ class TEModel(Enum): QWEN3_2B = 17 GEMMA_3_12B = 18 JINA_CLIP_2 = 19 + QWEN3_8B = 20 def detect_te_model(sd): @@ -1089,6 +1091,8 @@ def detect_te_model(sd): return TEModel.QWEN3_4B elif weight.shape[0] == 2048: return TEModel.QWEN3_2B + elif weight.shape[0] == 4096: + return TEModel.QWEN3_8B if weight.shape[0] == 5120: if "model.layers.39.post_attention_layernorm.weight" in sd: return TEModel.MISTRAL3_24B @@ -1214,11 +1218,18 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip clip_target.tokenizer = comfy.text_encoders.flux.Flux2Tokenizer tokenizer_data["tekken_model"] = clip_data[0].get("tekken_model", None) elif te_model == TEModel.QWEN3_4B: - clip_target.clip = comfy.text_encoders.z_image.te(**llama_detect(clip_data)) - clip_target.tokenizer = comfy.text_encoders.z_image.ZImageTokenizer + if clip_type == CLIPType.FLUX or clip_type == CLIPType.FLUX2: + clip_target.clip = comfy.text_encoders.flux.klein_te(**llama_detect(clip_data), model_type="qwen3_4b") + clip_target.tokenizer = comfy.text_encoders.flux.KleinTokenizer + else: + clip_target.clip = comfy.text_encoders.z_image.te(**llama_detect(clip_data)) + clip_target.tokenizer = comfy.text_encoders.z_image.ZImageTokenizer elif te_model == TEModel.QWEN3_2B: clip_target.clip = comfy.text_encoders.ovis.te(**llama_detect(clip_data)) clip_target.tokenizer = comfy.text_encoders.ovis.OvisTokenizer + elif te_model == TEModel.QWEN3_8B: + clip_target.clip = comfy.text_encoders.flux.klein_te(**llama_detect(clip_data), model_type="qwen3_8b") + clip_target.tokenizer = comfy.text_encoders.flux.KleinTokenizer8B elif te_model == TEModel.JINA_CLIP_2: clip_target.clip = comfy.text_encoders.jina_clip_2.JinaClip2TextModelWrapper clip_target.tokenizer = comfy.text_encoders.jina_clip_2.JinaClip2TokenizerWrapper diff --git a/comfy/text_encoders/flux.py b/comfy/text_encoders/flux.py index 21d93d757..4075afca4 100644 --- a/comfy/text_encoders/flux.py +++ b/comfy/text_encoders/flux.py @@ -3,7 +3,7 @@ import comfy.text_encoders.t5 import comfy.text_encoders.sd3_clip import comfy.text_encoders.llama import comfy.model_management -from transformers import T5TokenizerFast, LlamaTokenizerFast +from transformers import T5TokenizerFast, LlamaTokenizerFast, Qwen2Tokenizer import torch import os import json @@ -172,3 +172,60 @@ def flux2_te(dtype_llama=None, llama_quantization_metadata=None, pruned=False): model_options["num_layers"] = 30 super().__init__(device=device, dtype=dtype, model_options=model_options) return Flux2TEModel_ + +class Qwen3Tokenizer(sd1_clip.SDTokenizer): + def __init__(self, embedding_directory=None, tokenizer_data={}): + tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "qwen25_tokenizer") + super().__init__(tokenizer_path, pad_with_end=False, embedding_size=2560, embedding_key='qwen3_4b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=512, pad_token=151643, tokenizer_data=tokenizer_data) + +class Qwen3Tokenizer8B(sd1_clip.SDTokenizer): + def __init__(self, embedding_directory=None, tokenizer_data={}): + tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "qwen25_tokenizer") + super().__init__(tokenizer_path, pad_with_end=False, embedding_size=4096, embedding_key='qwen3_8b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=512, pad_token=151643, tokenizer_data=tokenizer_data) + +class KleinTokenizer(sd1_clip.SD1Tokenizer): + def __init__(self, embedding_directory=None, tokenizer_data={}, name="qwen3_4b"): + if name == "qwen3_4b": + tokenizer = Qwen3Tokenizer + elif name == "qwen3_8b": + tokenizer = Qwen3Tokenizer8B + + super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, name=name, tokenizer=tokenizer) + self.llama_template = "<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n\n\n\n\n" + + def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None, **kwargs): + if llama_template is None: + llama_text = self.llama_template.format(text) + else: + llama_text = llama_template.format(text) + + tokens = super().tokenize_with_weights(llama_text, return_word_ids=return_word_ids, disable_weights=True, **kwargs) + return tokens + +class KleinTokenizer8B(KleinTokenizer): + def __init__(self, embedding_directory=None, tokenizer_data={}, name="qwen3_8b"): + super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, name=name) + +class Qwen3_4BModel(sd1_clip.SDClipModel): + def __init__(self, device="cpu", layer=[9, 18, 27], layer_idx=None, dtype=None, attention_mask=True, model_options={}): + super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"pad": 151643}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Qwen3_4B, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options) + +class Qwen3_8BModel(sd1_clip.SDClipModel): + def __init__(self, device="cpu", layer=[9, 18, 27], layer_idx=None, dtype=None, attention_mask=True, model_options={}): + super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"pad": 151643}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Qwen3_8B, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options) + +def klein_te(dtype_llama=None, llama_quantization_metadata=None, model_type="qwen3_4b"): + if model_type == "qwen3_4b": + model = Qwen3_4BModel + elif model_type == "qwen3_8b": + model = Qwen3_8BModel + + class Flux2TEModel_(Flux2TEModel): + def __init__(self, device="cpu", dtype=None, model_options={}): + if llama_quantization_metadata is not None: + model_options = model_options.copy() + model_options["quantization_metadata"] = llama_quantization_metadata + if dtype_llama is not None: + dtype = dtype_llama + super().__init__(device=device, dtype=dtype, name=model_type, model_options=model_options, clip_model=model) + return Flux2TEModel_ diff --git a/comfy/text_encoders/llama.py b/comfy/text_encoders/llama.py index 76731576b..331a30f61 100644 --- a/comfy/text_encoders/llama.py +++ b/comfy/text_encoders/llama.py @@ -99,6 +99,28 @@ class Qwen3_4BConfig: rope_scale = None final_norm: bool = True +@dataclass +class Qwen3_8BConfig: + vocab_size: int = 151936 + hidden_size: int = 4096 + intermediate_size: int = 12288 + num_hidden_layers: int = 36 + num_attention_heads: int = 32 + num_key_value_heads: int = 8 + max_position_embeddings: int = 40960 + rms_norm_eps: float = 1e-6 + rope_theta: float = 1000000.0 + transformer_type: str = "llama" + head_dim = 128 + rms_norm_add = False + mlp_activation = "silu" + qkv_bias = False + rope_dims = None + q_norm = "gemma3" + k_norm = "gemma3" + rope_scale = None + final_norm: bool = True + @dataclass class Ovis25_2BConfig: vocab_size: int = 151936 @@ -628,6 +650,15 @@ class Qwen3_4B(BaseLlama, torch.nn.Module): self.model = Llama2_(config, device=device, dtype=dtype, ops=operations) self.dtype = dtype +class Qwen3_8B(BaseLlama, torch.nn.Module): + def __init__(self, config_dict, dtype, device, operations): + super().__init__() + config = Qwen3_8BConfig(**config_dict) + self.num_layers = config.num_hidden_layers + + self.model = Llama2_(config, device=device, dtype=dtype, ops=operations) + self.dtype = dtype + class Ovis25_2B(BaseLlama, torch.nn.Module): def __init__(self, config_dict, dtype, device, operations): super().__init__() From 8f40b43e0204d5b9780f3e9618e140e929e80594 Mon Sep 17 00:00:00 2001 From: comfyanonymous Date: Thu, 15 Jan 2026 10:57:35 -0500 Subject: [PATCH 457/647] ComfyUI v0.9.2 --- comfyui_version.py | 2 +- pyproject.toml | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/comfyui_version.py b/comfyui_version.py index 0c9871e35..dbb57b4e5 100644 --- a/comfyui_version.py +++ b/comfyui_version.py @@ -1,3 +1,3 @@ # This file is automatically generated by the build process when version is # updated in pyproject.toml. -__version__ = "0.9.1" +__version__ = "0.9.2" diff --git a/pyproject.toml b/pyproject.toml index dc52218b4..9ea73da05 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -1,6 +1,6 @@ [project] name = "ComfyUI" -version = "0.9.1" +version = "0.9.2" readme = "README.md" license = { file = "LICENSE" } requires-python = ">=3.10" From 12918a5f789d11c7d3c9d9f732891337740fe96f Mon Sep 17 00:00:00 2001 From: ComfyUI Wiki Date: Fri, 16 Jan 2026 03:08:21 +0800 Subject: [PATCH 458/647] chore: update workflow templates to v0.8.7 (#11896) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 8650d28ec..624cd067b 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,5 +1,5 @@ comfyui-frontend-package==1.36.14 -comfyui-workflow-templates==0.8.4 +comfyui-workflow-templates==0.8.7 comfyui-embedded-docs==0.4.0 torch torchsde From 6125b3a5e7215bf01874e402525552a7f5657a41 Mon Sep 17 00:00:00 2001 From: ComfyUI Wiki Date: Fri, 16 Jan 2026 05:12:13 +0800 Subject: [PATCH 459/647] Update workflow templates to v0.8.10 (#11899) * chore: update workflow templates to v0.8.9 * Update requirements.txt --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 624cd067b..996701550 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,5 +1,5 @@ comfyui-frontend-package==1.36.14 -comfyui-workflow-templates==0.8.7 +comfyui-workflow-templates==0.8.10 comfyui-embedded-docs==0.4.0 torch torchsde From 4c816d5c698dafaa31f8fc2c08ab1d81f9bc3239 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Thu, 15 Jan 2026 17:06:40 -0800 Subject: [PATCH 460/647] Adjust memory usage factor calculation for flux2 klein. (#11900) --- comfy/supported_models.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy/supported_models.py b/comfy/supported_models.py index 2c4c6b8fc..c8a7f6efb 100644 --- a/comfy/supported_models.py +++ b/comfy/supported_models.py @@ -763,7 +763,7 @@ class Flux2(Flux): def __init__(self, unet_config): super().__init__(unet_config) - self.memory_usage_factor = self.memory_usage_factor * (2.0 * 2.0) * 2.36 + self.memory_usage_factor = self.memory_usage_factor * (2.0 * 2.0) * (unet_config['hidden_size'] / 2604) def get_model(self, state_dict, prefix="", device=None): out = model_base.Flux2(self, device=device) From 732b707397922dbbec5ed04ecca3c773c878c64e Mon Sep 17 00:00:00 2001 From: Jedrzej Kosinski Date: Thu, 15 Jan 2026 20:15:15 -0800 Subject: [PATCH 461/647] Added try-except around seed_assets call in get_object_info with a logging statement (#11901) --- server.py | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/server.py b/server.py index da2baefd4..04a577488 100644 --- a/server.py +++ b/server.py @@ -686,7 +686,10 @@ class PromptServer(): @routes.get("/object_info") async def get_object_info(request): - seed_assets(["models"]) + try: + seed_assets(["models"]) + except Exception as e: + logging.error(f"Failed to seed assets: {e}") with folder_paths.cache_helper: out = {} for x in nodes.NODE_CLASS_MAPPINGS: From 9125613b53fc6af219d5a3db1d5b202ccc3f41b3 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Fri, 16 Jan 2026 08:09:07 +0200 Subject: [PATCH 462/647] feat(api-nodes): extend ByteDance nodes with seedance-1-5-pro model (#11871) --- comfy_api_nodes/apis/bytedance_api.py | 7 ++ comfy_api_nodes/nodes_bytedance.py | 104 +++++++++++++++++++++++--- 2 files changed, 101 insertions(+), 10 deletions(-) diff --git a/comfy_api_nodes/apis/bytedance_api.py b/comfy_api_nodes/apis/bytedance_api.py index b8c2f618b..400648cca 100644 --- a/comfy_api_nodes/apis/bytedance_api.py +++ b/comfy_api_nodes/apis/bytedance_api.py @@ -65,11 +65,13 @@ class TaskImageContent(BaseModel): class Text2VideoTaskCreationRequest(BaseModel): model: str = Field(...) content: list[TaskTextContent] = Field(..., min_length=1) + generate_audio: bool | None = Field(...) class Image2VideoTaskCreationRequest(BaseModel): model: str = Field(...) content: list[TaskTextContent | TaskImageContent] = Field(..., min_length=2) + generate_audio: bool | None = Field(...) class TaskCreationResponse(BaseModel): @@ -141,4 +143,9 @@ VIDEO_TASKS_EXECUTION_TIME = { "720p": 65, "1080p": 100, }, + "seedance-1-5-pro-251215": { + "480p": 80, + "720p": 100, + "1080p": 150, + }, } diff --git a/comfy_api_nodes/nodes_bytedance.py b/comfy_api_nodes/nodes_bytedance.py index f09a4a0ed..9cb1ca004 100644 --- a/comfy_api_nodes/nodes_bytedance.py +++ b/comfy_api_nodes/nodes_bytedance.py @@ -477,7 +477,12 @@ class ByteDanceTextToVideoNode(IO.ComfyNode): inputs=[ IO.Combo.Input( "model", - options=["seedance-1-0-pro-250528", "seedance-1-0-lite-t2v-250428", "seedance-1-0-pro-fast-251015"], + options=[ + "seedance-1-5-pro-251215", + "seedance-1-0-pro-250528", + "seedance-1-0-lite-t2v-250428", + "seedance-1-0-pro-fast-251015", + ], default="seedance-1-0-pro-fast-251015", ), IO.String.Input( @@ -528,6 +533,12 @@ class ByteDanceTextToVideoNode(IO.ComfyNode): tooltip='Whether to add an "AI generated" watermark to the video.', optional=True, ), + IO.Boolean.Input( + "generate_audio", + default=False, + tooltip="This parameter is ignored for any model except seedance-1-5-pro.", + optional=True, + ), ], outputs=[ IO.Video.Output(), @@ -552,7 +563,10 @@ class ByteDanceTextToVideoNode(IO.ComfyNode): seed: int, camera_fixed: bool, watermark: bool, + generate_audio: bool = False, ) -> IO.NodeOutput: + if model == "seedance-1-5-pro-251215" and duration < 4: + raise ValueError("Minimum supported duration for Seedance 1.5 Pro is 4 seconds.") validate_string(prompt, strip_whitespace=True, min_length=1) raise_if_text_params(prompt, ["resolution", "ratio", "duration", "seed", "camerafixed", "watermark"]) @@ -567,7 +581,11 @@ class ByteDanceTextToVideoNode(IO.ComfyNode): ) return await process_video_task( cls, - payload=Text2VideoTaskCreationRequest(model=model, content=[TaskTextContent(text=prompt)]), + payload=Text2VideoTaskCreationRequest( + model=model, + content=[TaskTextContent(text=prompt)], + generate_audio=generate_audio if model == "seedance-1-5-pro-251215" else None, + ), estimated_duration=max(1, math.ceil(VIDEO_TASKS_EXECUTION_TIME[model][resolution] * (duration / 10.0))), ) @@ -584,7 +602,12 @@ class ByteDanceImageToVideoNode(IO.ComfyNode): inputs=[ IO.Combo.Input( "model", - options=["seedance-1-0-pro-250528", "seedance-1-0-lite-t2v-250428", "seedance-1-0-pro-fast-251015"], + options=[ + "seedance-1-5-pro-251215", + "seedance-1-0-pro-250528", + "seedance-1-0-lite-i2v-250428", + "seedance-1-0-pro-fast-251015", + ], default="seedance-1-0-pro-fast-251015", ), IO.String.Input( @@ -639,6 +662,12 @@ class ByteDanceImageToVideoNode(IO.ComfyNode): tooltip='Whether to add an "AI generated" watermark to the video.', optional=True, ), + IO.Boolean.Input( + "generate_audio", + default=False, + tooltip="This parameter is ignored for any model except seedance-1-5-pro.", + optional=True, + ), ], outputs=[ IO.Video.Output(), @@ -664,7 +693,10 @@ class ByteDanceImageToVideoNode(IO.ComfyNode): seed: int, camera_fixed: bool, watermark: bool, + generate_audio: bool = False, ) -> IO.NodeOutput: + if model == "seedance-1-5-pro-251215" and duration < 4: + raise ValueError("Minimum supported duration for Seedance 1.5 Pro is 4 seconds.") validate_string(prompt, strip_whitespace=True, min_length=1) raise_if_text_params(prompt, ["resolution", "ratio", "duration", "seed", "camerafixed", "watermark"]) validate_image_dimensions(image, min_width=300, min_height=300, max_width=6000, max_height=6000) @@ -686,6 +718,7 @@ class ByteDanceImageToVideoNode(IO.ComfyNode): payload=Image2VideoTaskCreationRequest( model=model, content=[TaskTextContent(text=prompt), TaskImageContent(image_url=TaskImageContentUrl(url=image_url))], + generate_audio=generate_audio if model == "seedance-1-5-pro-251215" else None, ), estimated_duration=max(1, math.ceil(VIDEO_TASKS_EXECUTION_TIME[model][resolution] * (duration / 10.0))), ) @@ -703,7 +736,7 @@ class ByteDanceFirstLastFrameNode(IO.ComfyNode): inputs=[ IO.Combo.Input( "model", - options=["seedance-1-0-pro-250528", "seedance-1-0-lite-i2v-250428"], + options=["seedance-1-5-pro-251215", "seedance-1-0-pro-250528", "seedance-1-0-lite-i2v-250428"], default="seedance-1-0-lite-i2v-250428", ), IO.String.Input( @@ -762,6 +795,12 @@ class ByteDanceFirstLastFrameNode(IO.ComfyNode): tooltip='Whether to add an "AI generated" watermark to the video.', optional=True, ), + IO.Boolean.Input( + "generate_audio", + default=False, + tooltip="This parameter is ignored for any model except seedance-1-5-pro.", + optional=True, + ), ], outputs=[ IO.Video.Output(), @@ -788,7 +827,10 @@ class ByteDanceFirstLastFrameNode(IO.ComfyNode): seed: int, camera_fixed: bool, watermark: bool, + generate_audio: bool = False, ) -> IO.NodeOutput: + if model == "seedance-1-5-pro-251215" and duration < 4: + raise ValueError("Minimum supported duration for Seedance 1.5 Pro is 4 seconds.") validate_string(prompt, strip_whitespace=True, min_length=1) raise_if_text_params(prompt, ["resolution", "ratio", "duration", "seed", "camerafixed", "watermark"]) for i in (first_frame, last_frame): @@ -821,6 +863,7 @@ class ByteDanceFirstLastFrameNode(IO.ComfyNode): TaskImageContent(image_url=TaskImageContentUrl(url=str(download_urls[0])), role="first_frame"), TaskImageContent(image_url=TaskImageContentUrl(url=str(download_urls[1])), role="last_frame"), ], + generate_audio=generate_audio if model == "seedance-1-5-pro-251215" else None, ), estimated_duration=max(1, math.ceil(VIDEO_TASKS_EXECUTION_TIME[model][resolution] * (duration / 10.0))), ) @@ -896,7 +939,41 @@ class ByteDanceImageReferenceNode(IO.ComfyNode): IO.Hidden.unique_id, ], is_api_node=True, - price_badge=PRICE_BADGE_VIDEO, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["model", "duration", "resolution"]), + expr=""" + ( + $priceByModel := { + "seedance-1-0-pro": { + "480p":[0.23,0.24], + "720p":[0.51,0.56] + }, + "seedance-1-0-lite": { + "480p":[0.17,0.18], + "720p":[0.37,0.41] + } + }; + $model := widgets.model; + $modelKey := + $contains($model, "seedance-1-0-pro") ? "seedance-1-0-pro" : + "seedance-1-0-lite"; + $resolution := widgets.resolution; + $resKey := + $contains($resolution, "720") ? "720p" : + "480p"; + $modelPrices := $lookup($priceByModel, $modelKey); + $baseRange := $lookup($modelPrices, $resKey); + $min10s := $baseRange[0]; + $max10s := $baseRange[1]; + $scale := widgets.duration / 10; + $minCost := $min10s * $scale; + $maxCost := $max10s * $scale; + ($minCost = $maxCost) + ? {"type":"usd","usd": $minCost} + : {"type":"range_usd","min_usd": $minCost, "max_usd": $maxCost} + ) + """, + ), ) @classmethod @@ -967,10 +1044,15 @@ def raise_if_text_params(prompt: str, text_params: list[str]) -> None: PRICE_BADGE_VIDEO = IO.PriceBadge( - depends_on=IO.PriceBadgeDepends(widgets=["model", "duration", "resolution"]), + depends_on=IO.PriceBadgeDepends(widgets=["model", "duration", "resolution", "generate_audio"]), expr=""" ( $priceByModel := { + "seedance-1-5-pro": { + "480p":[0.12,0.12], + "720p":[0.26,0.26], + "1080p":[0.58,0.59] + }, "seedance-1-0-pro": { "480p":[0.23,0.24], "720p":[0.51,0.56], @@ -989,6 +1071,7 @@ PRICE_BADGE_VIDEO = IO.PriceBadge( }; $model := widgets.model; $modelKey := + $contains($model, "seedance-1-5-pro") ? "seedance-1-5-pro" : $contains($model, "seedance-1-0-pro-fast") ? "seedance-1-0-pro-fast" : $contains($model, "seedance-1-0-pro") ? "seedance-1-0-pro" : "seedance-1-0-lite"; @@ -1002,11 +1085,12 @@ PRICE_BADGE_VIDEO = IO.PriceBadge( $min10s := $baseRange[0]; $max10s := $baseRange[1]; $scale := widgets.duration / 10; - $minCost := $min10s * $scale; - $maxCost := $max10s * $scale; + $audioMultiplier := ($modelKey = "seedance-1-5-pro" and widgets.generate_audio) ? 2 : 1; + $minCost := $min10s * $scale * $audioMultiplier; + $maxCost := $max10s * $scale * $audioMultiplier; ($minCost = $maxCost) - ? {"type":"usd","usd": $minCost} - : {"type":"range_usd","min_usd": $minCost, "max_usd": $maxCost} + ? {"type":"usd","usd": $minCost, "format": { "approximate": true }} + : {"type":"range_usd","min_usd": $minCost, "max_usd": $maxCost, "format": { "approximate": true }} ) """, ) From 0c6b36c6ac1c34515cdf28f777a63074cd6d563d Mon Sep 17 00:00:00 2001 From: ComfyUI Wiki Date: Sat, 17 Jan 2026 06:22:50 +0800 Subject: [PATCH 463/647] chore: update workflow templates to v0.8.11 (#11918) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 996701550..3876274f9 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,5 +1,5 @@ comfyui-frontend-package==1.36.14 -comfyui-workflow-templates==0.8.10 +comfyui-workflow-templates==0.8.11 comfyui-embedded-docs==0.4.0 torch torchsde From 7ac999bf3069b06648a749212f59237080a75591 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Fri, 16 Jan 2026 20:02:28 -0800 Subject: [PATCH 464/647] Add image sizes to clip vision outputs. (#11923) --- comfy/clip_vision.py | 1 + 1 file changed, 1 insertion(+) diff --git a/comfy/clip_vision.py b/comfy/clip_vision.py index 66f2a9d9c..b28bf636c 100644 --- a/comfy/clip_vision.py +++ b/comfy/clip_vision.py @@ -66,6 +66,7 @@ class ClipVisionModel(): outputs = Output() outputs["last_hidden_state"] = out[0].to(comfy.model_management.intermediate_device()) outputs["image_embeds"] = out[2].to(comfy.model_management.intermediate_device()) + outputs["image_sizes"] = [pixel_values.shape[1:]] * pixel_values.shape[0] if self.return_all_hidden_states: all_hs = out[1].to(comfy.model_management.intermediate_device()) outputs["penultimate_hidden_states"] = all_hs[:, -2] From 00c775950aec5c563f532c8db08dae5e6adc24eb Mon Sep 17 00:00:00 2001 From: Alex Butler Date: Sun, 18 Jan 2026 01:18:04 +0000 Subject: [PATCH 465/647] Update readme rdna3 nightly url (#11937) --- README.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/README.md b/README.md index e25f3cda7..123cc9472 100644 --- a/README.md +++ b/README.md @@ -240,7 +240,7 @@ These have less hardware support than the builds above but they work on windows. RDNA 3 (RX 7000 series): -```pip install --pre torch torchvision torchaudio --index-url https://rocm.nightlies.amd.com/v2/gfx110X-dgpu/``` +```pip install --pre torch torchvision torchaudio --index-url https://rocm.nightlies.amd.com/v2/gfx110X-all/``` RDNA 3.5 (Strix halo/Ryzen AI Max+ 365): From 0fd10ffa09588e0fc7f576ab7d0c93e97ad5fbb0 Mon Sep 17 00:00:00 2001 From: Theephop <144770658+TheephopWS@users.noreply.github.com> Date: Sun, 18 Jan 2026 09:18:24 +0800 Subject: [PATCH 466/647] fix: use .cpu() for waveform conversion in AudioFrame creation (#11787) --- comfy_api/latest/_input_impl/video_types.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy_api/latest/_input_impl/video_types.py b/comfy_api/latest/_input_impl/video_types.py index ea35c6062..1405d0b81 100644 --- a/comfy_api/latest/_input_impl/video_types.py +++ b/comfy_api/latest/_input_impl/video_types.py @@ -374,7 +374,7 @@ class VideoFromComponents(VideoInput): if audio_stream and self.__components.audio: waveform = self.__components.audio['waveform'] waveform = waveform[:, :, :math.ceil((audio_sample_rate / frame_rate) * self.__components.images.shape[0])] - frame = av.AudioFrame.from_ndarray(waveform.movedim(2, 1).reshape(1, -1).float().numpy(), format='flt', layout='mono' if waveform.shape[1] == 1 else 'stereo') + frame = av.AudioFrame.from_ndarray(waveform.movedim(2, 1).reshape(1, -1).float().cpu().numpy(), format='flt', layout='mono' if waveform.shape[1] == 1 else 'stereo') frame.sample_rate = audio_sample_rate frame.pts = 0 output.mux(audio_stream.encode(frame)) From 190c4416cce3b3b97b628935e001d796d565bfc9 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sat, 17 Jan 2026 18:20:35 -0800 Subject: [PATCH 467/647] Bump comfy-kitchen dependency to version 0.2.7 (#11941) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 3876274f9..622256973 100644 --- a/requirements.txt +++ b/requirements.txt @@ -21,7 +21,7 @@ psutil alembic SQLAlchemy av>=14.2.0 -comfy-kitchen>=0.2.6 +comfy-kitchen>=0.2.7 #non essential dependencies: kornia>=0.7.1 From ac26065e6125871e2a742db6960f183fa037a75d Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Sun, 18 Jan 2026 04:52:45 +0200 Subject: [PATCH 468/647] chore(api-nodes): remove non-used; extract model to separate files (#11927) * chore(api-nodes): remove non-used; extract model to separate files * chore(api-nodes): remove non-needed prefix in filenames --- comfy_api_nodes/README.md | 65 ---- comfy_api_nodes/apis/{bfl_api.py => bfl.py} | 0 .../apis/{bytedance_api.py => bytedance.py} | 0 .../apis/{gemini_api.py => gemini.py} | 0 comfy_api_nodes/apis/ideogram.py | 292 +++++++++++++++++ .../apis/{kling_api.py => kling.py} | 0 comfy_api_nodes/apis/{luma_api.py => luma.py} | 0 .../apis/{minimax_api.py => minimax.py} | 0 comfy_api_nodes/apis/moonvalley.py | 152 +++++++++ comfy_api_nodes/apis/openai.py | 170 ++++++++++ comfy_api_nodes/apis/openai_api.py | 52 --- .../apis/{pixverse_api.py => pixverse.py} | 0 .../apis/{recraft_api.py => recraft.py} | 0 .../apis/{rodin_api.py => rodin.py} | 0 comfy_api_nodes/apis/runway.py | 127 ++++++++ .../apis/{stability_api.py => stability.py} | 0 .../apis/{topaz_api.py => topaz.py} | 4 +- .../apis/{tripo_api.py => tripo.py} | 0 comfy_api_nodes/apis/{veo_api.py => veo.py} | 0 comfy_api_nodes/mapper_utils.py | 116 ------- comfy_api_nodes/nodes_bfl.py | 2 +- comfy_api_nodes/nodes_bytedance.py | 2 +- comfy_api_nodes/nodes_gemini.py | 2 +- comfy_api_nodes/nodes_ideogram.py | 2 +- comfy_api_nodes/nodes_kling.py | 2 +- comfy_api_nodes/nodes_luma.py | 2 +- comfy_api_nodes/nodes_minimax.py | 2 +- comfy_api_nodes/nodes_moonvalley.py | 2 +- comfy_api_nodes/nodes_openai.py | 86 ++--- comfy_api_nodes/nodes_pixverse.py | 2 +- comfy_api_nodes/nodes_recraft.py | 2 +- comfy_api_nodes/nodes_rodin.py | 2 +- comfy_api_nodes/nodes_runway.py | 2 +- comfy_api_nodes/nodes_stability.py | 2 +- comfy_api_nodes/nodes_topaz.py | 55 ++-- comfy_api_nodes/nodes_tripo.py | 2 +- comfy_api_nodes/nodes_veo2.py | 2 +- comfy_api_nodes/redocly-dev.yaml | 10 - comfy_api_nodes/redocly.yaml | 10 - .../comfy_api_nodes_test/mapper_utils_test.py | 297 ------------------ 40 files changed, 825 insertions(+), 641 deletions(-) delete mode 100644 comfy_api_nodes/README.md rename comfy_api_nodes/apis/{bfl_api.py => bfl.py} (100%) rename comfy_api_nodes/apis/{bytedance_api.py => bytedance.py} (100%) rename comfy_api_nodes/apis/{gemini_api.py => gemini.py} (100%) create mode 100644 comfy_api_nodes/apis/ideogram.py rename comfy_api_nodes/apis/{kling_api.py => kling.py} (100%) rename comfy_api_nodes/apis/{luma_api.py => luma.py} (100%) rename comfy_api_nodes/apis/{minimax_api.py => minimax.py} (100%) create mode 100644 comfy_api_nodes/apis/moonvalley.py create mode 100644 comfy_api_nodes/apis/openai.py delete mode 100644 comfy_api_nodes/apis/openai_api.py rename comfy_api_nodes/apis/{pixverse_api.py => pixverse.py} (100%) rename comfy_api_nodes/apis/{recraft_api.py => recraft.py} (100%) rename comfy_api_nodes/apis/{rodin_api.py => rodin.py} (100%) create mode 100644 comfy_api_nodes/apis/runway.py rename comfy_api_nodes/apis/{stability_api.py => stability.py} (100%) rename comfy_api_nodes/apis/{topaz_api.py => topaz.py} (97%) rename comfy_api_nodes/apis/{tripo_api.py => tripo.py} (100%) rename comfy_api_nodes/apis/{veo_api.py => veo.py} (100%) delete mode 100644 comfy_api_nodes/mapper_utils.py delete mode 100644 comfy_api_nodes/redocly-dev.yaml delete mode 100644 comfy_api_nodes/redocly.yaml delete mode 100644 tests-unit/comfy_api_nodes_test/mapper_utils_test.py diff --git a/comfy_api_nodes/README.md b/comfy_api_nodes/README.md deleted file mode 100644 index f56d6c860..000000000 --- a/comfy_api_nodes/README.md +++ /dev/null @@ -1,65 +0,0 @@ -# ComfyUI API Nodes - -## Introduction - -Below are a collection of nodes that work by calling external APIs. More information available in our [docs](https://docs.comfy.org/tutorials/api-nodes/overview). - -## Development - -While developing, you should be testing against the Staging environment. To test against staging: - -**Install ComfyUI_frontend** - -Follow the instructions [here](https://github.com/Comfy-Org/ComfyUI_frontend) to start the frontend server. By default, it will connect to Staging authentication. - -> **Hint:** If you use --front-end-version argument for ComfyUI, it will use production authentication. - -```bash -python run main.py --comfy-api-base https://stagingapi.comfy.org -``` - -To authenticate to staging, please login and then ask one of Comfy Org team to whitelist you for access to staging. - -API stubs are generated through automatic codegen tools from OpenAPI definitions. Since the Comfy Org OpenAPI definition contains many things from the Comfy Registry as well, we use redocly/cli to filter out only the paths relevant for API nodes. - -### Redocly Instructions - -**Tip** -When developing locally, use the `redocly-dev.yaml` file to generate pydantic models. This lets you use stubs for APIs that are not marked `Released` yet. - -Before your API node PR merges, make sure to add the `Released` tag to the `openapi.yaml` file and test in staging. - -```bash -# Download the OpenAPI file from staging server. -curl -o openapi.yaml https://stagingapi.comfy.org/openapi - -# Filter out unneeded API definitions. -npm install -g @redocly/cli -redocly bundle openapi.yaml --output filtered-openapi.yaml --config comfy_api_nodes/redocly-dev.yaml --remove-unused-components - -# Generate the pydantic datamodels for validation. -datamodel-codegen --use-subclass-enum --field-constraints --strict-types bytes --input filtered-openapi.yaml --output comfy_api_nodes/apis/__init__.py --output-model-type pydantic_v2.BaseModel - -``` - - -# Merging to Master - -Before merging to comfyanonymous/ComfyUI master, follow these steps: - -1. Add the "Released" tag to the ComfyUI OpenAPI yaml file for each endpoint you are using in the nodes. -1. Make sure the ComfyUI API is deployed to prod with your changes. -1. Run the code generation again with `redocly.yaml` and the production OpenAPI yaml file. - -```bash -# Download the OpenAPI file from prod server. -curl -o openapi.yaml https://api.comfy.org/openapi - -# Filter out unneeded API definitions. -npm install -g @redocly/cli -redocly bundle openapi.yaml --output filtered-openapi.yaml --config comfy_api_nodes/redocly.yaml --remove-unused-components - -# Generate the pydantic datamodels for validation. -datamodel-codegen --use-subclass-enum --field-constraints --strict-types bytes --input filtered-openapi.yaml --output comfy_api_nodes/apis/__init__.py --output-model-type pydantic_v2.BaseModel - -``` diff --git a/comfy_api_nodes/apis/bfl_api.py b/comfy_api_nodes/apis/bfl.py similarity index 100% rename from comfy_api_nodes/apis/bfl_api.py rename to comfy_api_nodes/apis/bfl.py diff --git a/comfy_api_nodes/apis/bytedance_api.py b/comfy_api_nodes/apis/bytedance.py similarity index 100% rename from comfy_api_nodes/apis/bytedance_api.py rename to comfy_api_nodes/apis/bytedance.py diff --git a/comfy_api_nodes/apis/gemini_api.py b/comfy_api_nodes/apis/gemini.py similarity index 100% rename from comfy_api_nodes/apis/gemini_api.py rename to comfy_api_nodes/apis/gemini.py diff --git a/comfy_api_nodes/apis/ideogram.py b/comfy_api_nodes/apis/ideogram.py new file mode 100644 index 000000000..737e18e3b --- /dev/null +++ b/comfy_api_nodes/apis/ideogram.py @@ -0,0 +1,292 @@ +from enum import Enum +from typing import Optional, List, Dict, Any, Union +from datetime import datetime + +from pydantic import BaseModel, Field, RootModel, StrictBytes + + +class IdeogramColorPalette1(BaseModel): + name: str = Field(..., description='Name of the preset color palette') + + +class Member(BaseModel): + color: Optional[str] = Field( + None, description='Hexadecimal color code', pattern='^#[0-9A-Fa-f]{6}$' + ) + weight: Optional[float] = Field( + None, description='Optional weight for the color (0-1)', ge=0.0, le=1.0 + ) + + +class IdeogramColorPalette2(BaseModel): + members: List[Member] = Field( + ..., description='Array of color definitions with optional weights' + ) + + +class IdeogramColorPalette( + RootModel[Union[IdeogramColorPalette1, IdeogramColorPalette2]] +): + root: Union[IdeogramColorPalette1, IdeogramColorPalette2] = Field( + ..., + description='A color palette specification that can either use a preset name or explicit color definitions with weights', + ) + + +class ImageRequest(BaseModel): + aspect_ratio: Optional[str] = Field( + None, + description="Optional. The aspect ratio (e.g., 'ASPECT_16_9', 'ASPECT_1_1'). Cannot be used with resolution. Defaults to 'ASPECT_1_1' if unspecified.", + ) + color_palette: Optional[Dict[str, Any]] = Field( + None, description='Optional. Color palette object. Only for V_2, V_2_TURBO.' + ) + magic_prompt_option: Optional[str] = Field( + None, description="Optional. MagicPrompt usage ('AUTO', 'ON', 'OFF')." + ) + model: str = Field(..., description="The model used (e.g., 'V_2', 'V_2A_TURBO')") + negative_prompt: Optional[str] = Field( + None, + description='Optional. Description of what to exclude. Only for V_1, V_1_TURBO, V_2, V_2_TURBO.', + ) + num_images: Optional[int] = Field( + 1, + description='Optional. Number of images to generate (1-8). Defaults to 1.', + ge=1, + le=8, + ) + prompt: str = Field( + ..., description='Required. The prompt to use to generate the image.' + ) + resolution: Optional[str] = Field( + None, + description="Optional. Resolution (e.g., 'RESOLUTION_1024_1024'). Only for model V_2. Cannot be used with aspect_ratio.", + ) + seed: Optional[int] = Field( + None, + description='Optional. A number between 0 and 2147483647.', + ge=0, + le=2147483647, + ) + style_type: Optional[str] = Field( + None, + description="Optional. Style type ('AUTO', 'GENERAL', 'REALISTIC', 'DESIGN', 'RENDER_3D', 'ANIME'). Only for models V_2 and above.", + ) + + +class IdeogramGenerateRequest(BaseModel): + image_request: ImageRequest = Field( + ..., description='The image generation request parameters.' + ) + + +class Datum(BaseModel): + is_image_safe: Optional[bool] = Field( + None, description='Indicates whether the image is considered safe.' + ) + prompt: Optional[str] = Field( + None, description='The prompt used to generate this image.' + ) + resolution: Optional[str] = Field( + None, description="The resolution of the generated image (e.g., '1024x1024')." + ) + seed: Optional[int] = Field( + None, description='The seed value used for this generation.' + ) + style_type: Optional[str] = Field( + None, + description="The style type used for generation (e.g., 'REALISTIC', 'ANIME').", + ) + url: Optional[str] = Field(None, description='URL to the generated image.') + + +class IdeogramGenerateResponse(BaseModel): + created: Optional[datetime] = Field( + None, description='Timestamp when the generation was created.' + ) + data: Optional[List[Datum]] = Field( + None, description='Array of generated image information.' + ) + + +class StyleCode(RootModel[str]): + root: str = Field(..., pattern='^[0-9A-Fa-f]{8}$') + + +class Datum1(BaseModel): + is_image_safe: Optional[bool] = None + prompt: Optional[str] = None + resolution: Optional[str] = None + seed: Optional[int] = None + style_type: Optional[str] = None + url: Optional[str] = None + + +class IdeogramV3IdeogramResponse(BaseModel): + created: Optional[datetime] = None + data: Optional[List[Datum1]] = None + + +class RenderingSpeed1(str, Enum): + TURBO = 'TURBO' + DEFAULT = 'DEFAULT' + QUALITY = 'QUALITY' + + +class IdeogramV3ReframeRequest(BaseModel): + color_palette: Optional[Dict[str, Any]] = None + image: Optional[StrictBytes] = None + num_images: Optional[int] = Field(None, ge=1, le=8) + rendering_speed: Optional[RenderingSpeed1] = None + resolution: str + seed: Optional[int] = Field(None, ge=0, le=2147483647) + style_codes: Optional[List[str]] = None + style_reference_images: Optional[List[StrictBytes]] = None + + +class MagicPrompt(str, Enum): + AUTO = 'AUTO' + ON = 'ON' + OFF = 'OFF' + + +class StyleType(str, Enum): + AUTO = 'AUTO' + GENERAL = 'GENERAL' + REALISTIC = 'REALISTIC' + DESIGN = 'DESIGN' + + +class IdeogramV3RemixRequest(BaseModel): + aspect_ratio: Optional[str] = None + color_palette: Optional[Dict[str, Any]] = None + image: Optional[StrictBytes] = None + image_weight: Optional[int] = Field(50, ge=1, le=100) + magic_prompt: Optional[MagicPrompt] = None + negative_prompt: Optional[str] = None + num_images: Optional[int] = Field(None, ge=1, le=8) + prompt: str + rendering_speed: Optional[RenderingSpeed1] = None + resolution: Optional[str] = None + seed: Optional[int] = Field(None, ge=0, le=2147483647) + style_codes: Optional[List[str]] = None + style_reference_images: Optional[List[StrictBytes]] = None + style_type: Optional[StyleType] = None + + +class IdeogramV3ReplaceBackgroundRequest(BaseModel): + color_palette: Optional[Dict[str, Any]] = None + image: Optional[StrictBytes] = None + magic_prompt: Optional[MagicPrompt] = None + num_images: Optional[int] = Field(None, ge=1, le=8) + prompt: str + rendering_speed: Optional[RenderingSpeed1] = None + seed: Optional[int] = Field(None, ge=0, le=2147483647) + style_codes: Optional[List[str]] = None + style_reference_images: Optional[List[StrictBytes]] = None + + +class ColorPalette(BaseModel): + name: str = Field(..., description='Name of the color palette', examples=['PASTEL']) + + +class MagicPrompt2(str, Enum): + ON = 'ON' + OFF = 'OFF' + + +class StyleType1(str, Enum): + AUTO = 'AUTO' + GENERAL = 'GENERAL' + REALISTIC = 'REALISTIC' + DESIGN = 'DESIGN' + FICTION = 'FICTION' + + +class RenderingSpeed(str, Enum): + DEFAULT = 'DEFAULT' + TURBO = 'TURBO' + QUALITY = 'QUALITY' + + +class IdeogramV3EditRequest(BaseModel): + color_palette: Optional[IdeogramColorPalette] = None + image: Optional[StrictBytes] = Field( + None, + description='The image being edited (max size 10MB); only JPEG, WebP and PNG formats are supported at this time.', + ) + magic_prompt: Optional[str] = Field( + None, + description='Determine if MagicPrompt should be used in generating the request or not.', + ) + mask: Optional[StrictBytes] = Field( + None, + description='A black and white image of the same size as the image being edited (max size 10MB). Black regions in the mask should match up with the regions of the image that you would like to edit; only JPEG, WebP and PNG formats are supported at this time.', + ) + num_images: Optional[int] = Field( + None, description='The number of images to generate.' + ) + prompt: str = Field( + ..., description='The prompt used to describe the edited result.' + ) + rendering_speed: RenderingSpeed + seed: Optional[int] = Field( + None, description='Random seed. Set for reproducible generation.' + ) + style_codes: Optional[List[StyleCode]] = Field( + None, + description='A list of 8 character hexadecimal codes representing the style of the image. Cannot be used in conjunction with style_reference_images or style_type.', + ) + style_reference_images: Optional[List[StrictBytes]] = Field( + None, + description='A set of images to use as style references (maximum total size 10MB across all style references). The images should be in JPEG, PNG or WebP format.', + ) + character_reference_images: Optional[List[str]] = Field( + None, + description='Generations with character reference are subject to the character reference pricing. A set of images to use as character references (maximum total size 10MB across all character references), currently only supports 1 character reference image. The images should be in JPEG, PNG or WebP format.' + ) + character_reference_images_mask: Optional[List[str]] = Field( + None, + description='Optional masks for character reference images. When provided, must match the number of character_reference_images. Each mask should be a grayscale image of the same dimensions as the corresponding character reference image. The images should be in JPEG, PNG or WebP format.' + ) + + +class IdeogramV3Request(BaseModel): + aspect_ratio: Optional[str] = Field( + None, description='Aspect ratio in format WxH', examples=['1x3'] + ) + color_palette: Optional[ColorPalette] = None + magic_prompt: Optional[MagicPrompt2] = Field( + None, description='Whether to enable magic prompt enhancement' + ) + negative_prompt: Optional[str] = Field( + None, description='Text prompt specifying what to avoid in the generation' + ) + num_images: Optional[int] = Field( + None, description='Number of images to generate', ge=1 + ) + prompt: str = Field(..., description='The text prompt for image generation') + rendering_speed: RenderingSpeed + resolution: Optional[str] = Field( + None, description='Image resolution in format WxH', examples=['1280x800'] + ) + seed: Optional[int] = Field( + None, description='Seed value for reproducible generation' + ) + style_codes: Optional[List[StyleCode]] = Field( + None, description='Array of style codes in hexadecimal format' + ) + style_reference_images: Optional[List[str]] = Field( + None, description='Array of reference image URLs or identifiers' + ) + style_type: Optional[StyleType1] = Field( + None, description='The type of style to apply' + ) + character_reference_images: Optional[List[str]] = Field( + None, + description='Generations with character reference are subject to the character reference pricing. A set of images to use as character references (maximum total size 10MB across all character references), currently only supports 1 character reference image. The images should be in JPEG, PNG or WebP format.' + ) + character_reference_images_mask: Optional[List[str]] = Field( + None, + description='Optional masks for character reference images. When provided, must match the number of character_reference_images. Each mask should be a grayscale image of the same dimensions as the corresponding character reference image. The images should be in JPEG, PNG or WebP format.' + ) diff --git a/comfy_api_nodes/apis/kling_api.py b/comfy_api_nodes/apis/kling.py similarity index 100% rename from comfy_api_nodes/apis/kling_api.py rename to comfy_api_nodes/apis/kling.py diff --git a/comfy_api_nodes/apis/luma_api.py b/comfy_api_nodes/apis/luma.py similarity index 100% rename from comfy_api_nodes/apis/luma_api.py rename to comfy_api_nodes/apis/luma.py diff --git a/comfy_api_nodes/apis/minimax_api.py b/comfy_api_nodes/apis/minimax.py similarity index 100% rename from comfy_api_nodes/apis/minimax_api.py rename to comfy_api_nodes/apis/minimax.py diff --git a/comfy_api_nodes/apis/moonvalley.py b/comfy_api_nodes/apis/moonvalley.py new file mode 100644 index 000000000..7ec7a4ade --- /dev/null +++ b/comfy_api_nodes/apis/moonvalley.py @@ -0,0 +1,152 @@ +from enum import Enum +from typing import Optional, Dict, Any + +from pydantic import BaseModel, Field, StrictBytes + + +class MoonvalleyPromptResponse(BaseModel): + error: Optional[Dict[str, Any]] = None + frame_conditioning: Optional[Dict[str, Any]] = None + id: Optional[str] = None + inference_params: Optional[Dict[str, Any]] = None + meta: Optional[Dict[str, Any]] = None + model_params: Optional[Dict[str, Any]] = None + output_url: Optional[str] = None + prompt_text: Optional[str] = None + status: Optional[str] = None + + +class MoonvalleyTextToVideoInferenceParams(BaseModel): + add_quality_guidance: Optional[bool] = Field( + True, description='Whether to add quality guidance' + ) + caching_coefficient: Optional[float] = Field( + 0.3, description='Caching coefficient for optimization' + ) + caching_cooldown: Optional[int] = Field( + 3, description='Number of caching cooldown steps' + ) + caching_warmup: Optional[int] = Field( + 3, description='Number of caching warmup steps' + ) + clip_value: Optional[float] = Field( + 3, description='CLIP value for generation control' + ) + conditioning_frame_index: Optional[int] = Field( + 0, description='Index of the conditioning frame' + ) + cooldown_steps: Optional[int] = Field( + 75, description='Number of cooldown steps (calculated based on num_frames)' + ) + fps: Optional[int] = Field( + 24, description='Frames per second of the generated video' + ) + guidance_scale: Optional[float] = Field( + 10, description='Guidance scale for generation control' + ) + height: Optional[int] = Field( + 1080, description='Height of the generated video in pixels' + ) + negative_prompt: Optional[str] = Field(None, description='Negative prompt text') + num_frames: Optional[int] = Field(64, description='Number of frames to generate') + seed: Optional[int] = Field( + None, description='Random seed for generation (default: random)' + ) + shift_value: Optional[float] = Field( + 3, description='Shift value for generation control' + ) + steps: Optional[int] = Field(80, description='Number of denoising steps') + use_guidance_schedule: Optional[bool] = Field( + True, description='Whether to use guidance scheduling' + ) + use_negative_prompts: Optional[bool] = Field( + False, description='Whether to use negative prompts' + ) + use_timestep_transform: Optional[bool] = Field( + True, description='Whether to use timestep transformation' + ) + warmup_steps: Optional[int] = Field( + 0, description='Number of warmup steps (calculated based on num_frames)' + ) + width: Optional[int] = Field( + 1920, description='Width of the generated video in pixels' + ) + + +class MoonvalleyTextToVideoRequest(BaseModel): + image_url: Optional[str] = None + inference_params: Optional[MoonvalleyTextToVideoInferenceParams] = None + prompt_text: Optional[str] = None + webhook_url: Optional[str] = None + + +class MoonvalleyUploadFileRequest(BaseModel): + file: Optional[StrictBytes] = None + + +class MoonvalleyUploadFileResponse(BaseModel): + access_url: Optional[str] = None + + +class MoonvalleyVideoToVideoInferenceParams(BaseModel): + add_quality_guidance: Optional[bool] = Field( + True, description='Whether to add quality guidance' + ) + caching_coefficient: Optional[float] = Field( + 0.3, description='Caching coefficient for optimization' + ) + caching_cooldown: Optional[int] = Field( + 3, description='Number of caching cooldown steps' + ) + caching_warmup: Optional[int] = Field( + 3, description='Number of caching warmup steps' + ) + clip_value: Optional[float] = Field( + 3, description='CLIP value for generation control' + ) + conditioning_frame_index: Optional[int] = Field( + 0, description='Index of the conditioning frame' + ) + cooldown_steps: Optional[int] = Field( + 36, description='Number of cooldown steps (calculated based on num_frames)' + ) + guidance_scale: Optional[float] = Field( + 15, description='Guidance scale for generation control' + ) + negative_prompt: Optional[str] = Field(None, description='Negative prompt text') + seed: Optional[int] = Field( + None, description='Random seed for generation (default: random)' + ) + shift_value: Optional[float] = Field( + 3, description='Shift value for generation control' + ) + steps: Optional[int] = Field(80, description='Number of denoising steps') + use_guidance_schedule: Optional[bool] = Field( + True, description='Whether to use guidance scheduling' + ) + use_negative_prompts: Optional[bool] = Field( + False, description='Whether to use negative prompts' + ) + use_timestep_transform: Optional[bool] = Field( + True, description='Whether to use timestep transformation' + ) + warmup_steps: Optional[int] = Field( + 24, description='Number of warmup steps (calculated based on num_frames)' + ) + + +class ControlType(str, Enum): + motion_control = 'motion_control' + pose_control = 'pose_control' + + +class MoonvalleyVideoToVideoRequest(BaseModel): + control_type: ControlType = Field( + ..., description='Supported types for video control' + ) + inference_params: Optional[MoonvalleyVideoToVideoInferenceParams] = None + prompt_text: str = Field(..., description='Describes the video to generate') + video_url: str = Field(..., description='Url to control video') + webhook_url: Optional[str] = Field( + None, description='Optional webhook URL for notifications' + ) diff --git a/comfy_api_nodes/apis/openai.py b/comfy_api_nodes/apis/openai.py new file mode 100644 index 000000000..b85ef252b --- /dev/null +++ b/comfy_api_nodes/apis/openai.py @@ -0,0 +1,170 @@ +from pydantic import BaseModel, Field + + +class Datum2(BaseModel): + b64_json: str | None = Field(None, description="Base64 encoded image data") + revised_prompt: str | None = Field(None, description="Revised prompt") + url: str | None = Field(None, description="URL of the image") + + +class InputTokensDetails(BaseModel): + image_tokens: int | None = Field(None) + text_tokens: int | None = Field(None) + + +class Usage(BaseModel): + input_tokens: int | None = Field(None) + input_tokens_details: InputTokensDetails | None = Field(None) + output_tokens: int | None = Field(None) + total_tokens: int | None = Field(None) + + +class OpenAIImageGenerationResponse(BaseModel): + data: list[Datum2] | None = Field(None) + usage: Usage | None = Field(None) + + +class OpenAIImageEditRequest(BaseModel): + background: str | None = Field(None, description="Background transparency") + model: str = Field(...) + moderation: str | None = Field(None) + n: int | None = Field(None, description="The number of images to generate") + output_compression: int | None = Field(None, description="Compression level for JPEG or WebP (0-100)") + output_format: str | None = Field(None) + prompt: str = Field(...) + quality: str | None = Field(None, description="Size of the image (e.g., 1024x1024, 1536x1024, auto)") + size: str | None = Field(None, description="Size of the output image") + + +class OpenAIImageGenerationRequest(BaseModel): + background: str | None = Field(None, description="Background transparency") + model: str | None = Field(None) + moderation: str | None = Field(None) + n: int | None = Field( + None, + description="The number of images to generate.", + ) + output_compression: int | None = Field(None, description="Compression level for JPEG or WebP (0-100)") + output_format: str | None = Field(None) + prompt: str = Field(...) + quality: str | None = Field(None, description="The quality of the generated image") + size: str | None = Field(None, description="Size of the image (e.g., 1024x1024, 1536x1024, auto)") + style: str | None = Field(None, description="Style of the image (only for dall-e-3)") + + +class ModelResponseProperties(BaseModel): + instructions: str | None = Field(None) + max_output_tokens: int | None = Field(None) + model: str | None = Field(None) + temperature: float | None = Field(1, description="Controls randomness in the response", ge=0.0, le=2.0) + top_p: float | None = Field( + 1, + description="Controls diversity of the response via nucleus sampling", + ge=0.0, + le=1.0, + ) + truncation: str | None = Field("disabled", description="Allowed values: 'auto' or 'disabled'") + + +class ResponseProperties(BaseModel): + instructions: str | None = Field(None) + max_output_tokens: int | None = Field(None) + model: str | None = Field(None) + previous_response_id: str | None = Field(None) + truncation: str | None = Field("disabled", description="Allowed values: 'auto' or 'disabled'") + + +class ResponseError(BaseModel): + code: str = Field(...) + message: str = Field(...) + + +class OutputTokensDetails(BaseModel): + reasoning_tokens: int = Field(..., description="The number of reasoning tokens.") + + +class CachedTokensDetails(BaseModel): + cached_tokens: int = Field( + ..., + description="The number of tokens that were retrieved from the cache.", + ) + + +class ResponseUsage(BaseModel): + input_tokens: int = Field(..., description="The number of input tokens.") + input_tokens_details: CachedTokensDetails = Field(...) + output_tokens: int = Field(..., description="The number of output tokens.") + output_tokens_details: OutputTokensDetails = Field(...) + total_tokens: int = Field(..., description="The total number of tokens used.") + + +class InputTextContent(BaseModel): + text: str = Field(..., description="The text input to the model.") + type: str = Field("input_text") + + +class OutputContent(BaseModel): + type: str = Field(..., description="The type of output content") + text: str | None = Field(None, description="The text content") + data: str | None = Field(None, description="Base64-encoded audio data") + transcript: str | None = Field(None, description="Transcript of the audio") + + +class OutputMessage(BaseModel): + type: str = Field(..., description="The type of output item") + content: list[OutputContent] | None = Field(None, description="The content of the message") + role: str | None = Field(None, description="The role of the message") + + +class OpenAIResponse(ModelResponseProperties, ResponseProperties): + created_at: float | None = Field( + None, + description="Unix timestamp (in seconds) of when this Response was created.", + ) + error: ResponseError | None = Field(None) + id: str | None = Field(None, description="Unique identifier for this Response.") + object: str | None = Field(None, description="The object type of this resource - always set to `response`.") + output: list[OutputMessage] | None = Field(None) + parallel_tool_calls: bool | None = Field(True) + status: str | None = Field( + None, + description="One of `completed`, `failed`, `in_progress`, or `incomplete`.", + ) + usage: ResponseUsage | None = Field(None) + + +class InputImageContent(BaseModel): + detail: str = Field(..., description="One of `high`, `low`, or `auto`. Defaults to `auto`.") + file_id: str | None = Field(None) + image_url: str | None = Field(None) + type: str = Field(..., description="The type of the input item. Always `input_image`.") + + +class InputFileContent(BaseModel): + file_data: str | None = Field(None) + file_id: str | None = Field(None) + filename: str | None = Field(None, description="The name of the file to be sent to the model.") + type: str = Field(..., description="The type of the input item. Always `input_file`.") + + +class InputMessage(BaseModel): + content: list[InputTextContent | InputImageContent | InputFileContent] = Field( + ..., + description="A list of one or many input items to the model, containing different content types.", + ) + role: str | None = Field(None) + type: str | None = Field(None) + + +class OpenAICreateResponse(ModelResponseProperties, ResponseProperties): + include: str | None = Field(None) + input: list[InputMessage] = Field(...) + parallel_tool_calls: bool | None = Field( + True, description="Whether to allow the model to run tool calls in parallel." + ) + store: bool | None = Field( + True, + description="Whether to store the generated model response for later retrieval via API.", + ) + stream: bool | None = Field(False) + usage: ResponseUsage | None = Field(None) diff --git a/comfy_api_nodes/apis/openai_api.py b/comfy_api_nodes/apis/openai_api.py deleted file mode 100644 index ae5bb2673..000000000 --- a/comfy_api_nodes/apis/openai_api.py +++ /dev/null @@ -1,52 +0,0 @@ -from pydantic import BaseModel, Field - - -class Datum2(BaseModel): - b64_json: str | None = Field(None, description="Base64 encoded image data") - revised_prompt: str | None = Field(None, description="Revised prompt") - url: str | None = Field(None, description="URL of the image") - - -class InputTokensDetails(BaseModel): - image_tokens: int | None = None - text_tokens: int | None = None - - -class Usage(BaseModel): - input_tokens: int | None = None - input_tokens_details: InputTokensDetails | None = None - output_tokens: int | None = None - total_tokens: int | None = None - - -class OpenAIImageGenerationResponse(BaseModel): - data: list[Datum2] | None = None - usage: Usage | None = None - - -class OpenAIImageEditRequest(BaseModel): - background: str | None = Field(None, description="Background transparency") - model: str = Field(...) - moderation: str | None = Field(None) - n: int | None = Field(None, description="The number of images to generate") - output_compression: int | None = Field(None, description="Compression level for JPEG or WebP (0-100)") - output_format: str | None = Field(None) - prompt: str = Field(...) - quality: str | None = Field(None, description="Size of the image (e.g., 1024x1024, 1536x1024, auto)") - size: str | None = Field(None, description="Size of the output image") - - -class OpenAIImageGenerationRequest(BaseModel): - background: str | None = Field(None, description="Background transparency") - model: str | None = Field(None) - moderation: str | None = Field(None) - n: int | None = Field( - None, - description="The number of images to generate.", - ) - output_compression: int | None = Field(None, description="Compression level for JPEG or WebP (0-100)") - output_format: str | None = Field(None) - prompt: str = Field(...) - quality: str | None = Field(None, description="The quality of the generated image") - size: str | None = Field(None, description="Size of the image (e.g., 1024x1024, 1536x1024, auto)") - style: str | None = Field(None, description="Style of the image (only for dall-e-3)") diff --git a/comfy_api_nodes/apis/pixverse_api.py b/comfy_api_nodes/apis/pixverse.py similarity index 100% rename from comfy_api_nodes/apis/pixverse_api.py rename to comfy_api_nodes/apis/pixverse.py diff --git a/comfy_api_nodes/apis/recraft_api.py b/comfy_api_nodes/apis/recraft.py similarity index 100% rename from comfy_api_nodes/apis/recraft_api.py rename to comfy_api_nodes/apis/recraft.py diff --git a/comfy_api_nodes/apis/rodin_api.py b/comfy_api_nodes/apis/rodin.py similarity index 100% rename from comfy_api_nodes/apis/rodin_api.py rename to comfy_api_nodes/apis/rodin.py diff --git a/comfy_api_nodes/apis/runway.py b/comfy_api_nodes/apis/runway.py new file mode 100644 index 000000000..df6f2b845 --- /dev/null +++ b/comfy_api_nodes/apis/runway.py @@ -0,0 +1,127 @@ +from enum import Enum +from typing import Optional, List, Union +from datetime import datetime + +from pydantic import BaseModel, Field, RootModel + + +class RunwayAspectRatioEnum(str, Enum): + field_1280_720 = '1280:720' + field_720_1280 = '720:1280' + field_1104_832 = '1104:832' + field_832_1104 = '832:1104' + field_960_960 = '960:960' + field_1584_672 = '1584:672' + field_1280_768 = '1280:768' + field_768_1280 = '768:1280' + + +class Position(str, Enum): + first = 'first' + last = 'last' + + +class RunwayPromptImageDetailedObject(BaseModel): + position: Position = Field( + ..., + description="The position of the image in the output video. 'last' is currently supported for gen3a_turbo only.", + ) + uri: str = Field( + ..., description='A HTTPS URL or data URI containing an encoded image.' + ) + + +class RunwayPromptImageObject( + RootModel[Union[str, List[RunwayPromptImageDetailedObject]]] +): + root: Union[str, List[RunwayPromptImageDetailedObject]] = Field( + ..., + description='Image(s) to use for the video generation. Can be a single URI or an array of image objects with positions.', + ) + + +class RunwayModelEnum(str, Enum): + gen4_turbo = 'gen4_turbo' + gen3a_turbo = 'gen3a_turbo' + + +class RunwayDurationEnum(int, Enum): + integer_5 = 5 + integer_10 = 10 + + +class RunwayImageToVideoRequest(BaseModel): + duration: RunwayDurationEnum + model: RunwayModelEnum + promptImage: RunwayPromptImageObject + promptText: Optional[str] = Field( + None, description='Text prompt for the generation', max_length=1000 + ) + ratio: RunwayAspectRatioEnum + seed: int = Field( + ..., description='Random seed for generation', ge=0, le=4294967295 + ) + + +class RunwayImageToVideoResponse(BaseModel): + id: Optional[str] = Field(None, description='Task ID') + + +class RunwayTaskStatusEnum(str, Enum): + SUCCEEDED = 'SUCCEEDED' + RUNNING = 'RUNNING' + FAILED = 'FAILED' + PENDING = 'PENDING' + CANCELLED = 'CANCELLED' + THROTTLED = 'THROTTLED' + + +class RunwayTaskStatusResponse(BaseModel): + createdAt: datetime = Field(..., description='Task creation timestamp') + id: str = Field(..., description='Task ID') + output: Optional[List[str]] = Field(None, description='Array of output video URLs') + progress: Optional[float] = Field( + None, + description='Float value between 0 and 1 representing the progress of the task. Only available if status is RUNNING.', + ge=0.0, + le=1.0, + ) + status: RunwayTaskStatusEnum + + +class Model4(str, Enum): + gen4_image = 'gen4_image' + + +class ReferenceImage(BaseModel): + uri: Optional[str] = Field( + None, description='A HTTPS URL or data URI containing an encoded image' + ) + + +class RunwayTextToImageAspectRatioEnum(str, Enum): + field_1920_1080 = '1920:1080' + field_1080_1920 = '1080:1920' + field_1024_1024 = '1024:1024' + field_1360_768 = '1360:768' + field_1080_1080 = '1080:1080' + field_1168_880 = '1168:880' + field_1440_1080 = '1440:1080' + field_1080_1440 = '1080:1440' + field_1808_768 = '1808:768' + field_2112_912 = '2112:912' + + +class RunwayTextToImageRequest(BaseModel): + model: Model4 = Field(..., description='Model to use for generation') + promptText: str = Field( + ..., description='Text prompt for the image generation', max_length=1000 + ) + ratio: RunwayTextToImageAspectRatioEnum + referenceImages: Optional[List[ReferenceImage]] = Field( + None, description='Array of reference images to guide the generation' + ) + + +class RunwayTextToImageResponse(BaseModel): + id: Optional[str] = Field(None, description='Task ID') diff --git a/comfy_api_nodes/apis/stability_api.py b/comfy_api_nodes/apis/stability.py similarity index 100% rename from comfy_api_nodes/apis/stability_api.py rename to comfy_api_nodes/apis/stability.py diff --git a/comfy_api_nodes/apis/topaz_api.py b/comfy_api_nodes/apis/topaz.py similarity index 97% rename from comfy_api_nodes/apis/topaz_api.py rename to comfy_api_nodes/apis/topaz.py index 4d9e62e72..a9e6235a7 100644 --- a/comfy_api_nodes/apis/topaz_api.py +++ b/comfy_api_nodes/apis/topaz.py @@ -41,7 +41,7 @@ class Resolution(BaseModel): height: int = Field(...) -class CreateCreateVideoRequestSource(BaseModel): +class CreateVideoRequestSource(BaseModel): container: str = Field(...) size: int = Field(..., description="Size of the video file in bytes") duration: int = Field(..., description="Duration of the video file in seconds") @@ -89,7 +89,7 @@ class Overrides(BaseModel): class CreateVideoRequest(BaseModel): - source: CreateCreateVideoRequestSource = Field(...) + source: CreateVideoRequestSource = Field(...) filters: list[Union[VideoFrameInterpolationFilter, VideoEnhancementFilter]] = Field(...) output: OutputInformationVideo = Field(...) overrides: Overrides = Field(Overrides(isPaidDiffusion=True)) diff --git a/comfy_api_nodes/apis/tripo_api.py b/comfy_api_nodes/apis/tripo.py similarity index 100% rename from comfy_api_nodes/apis/tripo_api.py rename to comfy_api_nodes/apis/tripo.py diff --git a/comfy_api_nodes/apis/veo_api.py b/comfy_api_nodes/apis/veo.py similarity index 100% rename from comfy_api_nodes/apis/veo_api.py rename to comfy_api_nodes/apis/veo.py diff --git a/comfy_api_nodes/mapper_utils.py b/comfy_api_nodes/mapper_utils.py deleted file mode 100644 index 6fab8f4bb..000000000 --- a/comfy_api_nodes/mapper_utils.py +++ /dev/null @@ -1,116 +0,0 @@ -from enum import Enum - -from pydantic.fields import FieldInfo -from pydantic import BaseModel -from pydantic_core import PydanticUndefined - -from comfy.comfy_types.node_typing import IO, InputTypeOptions - -NodeInput = tuple[IO, InputTypeOptions] - - -def _create_base_config(field_info: FieldInfo) -> InputTypeOptions: - config = {} - if hasattr(field_info, "default") and field_info.default is not PydanticUndefined: - config["default"] = field_info.default - if hasattr(field_info, "description") and field_info.description is not None: - config["tooltip"] = field_info.description - return config - - -def _get_number_constraints_config(field_info: FieldInfo) -> dict: - config = {} - if hasattr(field_info, "metadata"): - metadata = field_info.metadata - for constraint in metadata: - if hasattr(constraint, "ge"): - config["min"] = constraint.ge - if hasattr(constraint, "le"): - config["max"] = constraint.le - if hasattr(constraint, "multiple_of"): - config["step"] = constraint.multiple_of - return config - - -def _model_field_to_image_input(field_info: FieldInfo, **kwargs) -> NodeInput: - return IO.IMAGE, { - **_create_base_config(field_info), - **kwargs, - } - - -def _model_field_to_string_input(field_info: FieldInfo, **kwargs) -> NodeInput: - return IO.STRING, { - **_create_base_config(field_info), - **kwargs, - } - - -def _model_field_to_float_input(field_info: FieldInfo, **kwargs) -> NodeInput: - return IO.FLOAT, { - **_create_base_config(field_info), - **_get_number_constraints_config(field_info), - **kwargs, - } - - -def _model_field_to_int_input(field_info: FieldInfo, **kwargs) -> NodeInput: - return IO.INT, { - **_create_base_config(field_info), - **_get_number_constraints_config(field_info), - **kwargs, - } - - -def _model_field_to_combo_input( - field_info: FieldInfo, enum_type: type[Enum] = None, **kwargs -) -> NodeInput: - combo_config = {} - if enum_type is not None: - combo_config["options"] = [option.value for option in enum_type] - combo_config = { - **combo_config, - **_create_base_config(field_info), - **kwargs, - } - return IO.COMBO, combo_config - - -def model_field_to_node_input( - input_type: IO, base_model: type[BaseModel], field_name: str, **kwargs -) -> NodeInput: - """ - Maps a field from a Pydantic model to a Comfy node input. - - Args: - input_type: The type of the input. - base_model: The Pydantic model to map the field from. - field_name: The name of the field to map. - **kwargs: Additional key/values to include in the input options. - - Note: - For combo inputs, pass an `Enum` to the `enum_type` keyword argument to populate the options automatically. - - Example: - >>> model_field_to_node_input(IO.STRING, MyModel, "my_field", multiline=True) - >>> model_field_to_node_input(IO.COMBO, MyModel, "my_field", enum_type=MyEnum) - >>> model_field_to_node_input(IO.FLOAT, MyModel, "my_field", slider=True) - """ - field_info: FieldInfo = base_model.model_fields[field_name] - result: NodeInput - - if input_type == IO.IMAGE: - result = _model_field_to_image_input(field_info, **kwargs) - elif input_type == IO.STRING: - result = _model_field_to_string_input(field_info, **kwargs) - elif input_type == IO.FLOAT: - result = _model_field_to_float_input(field_info, **kwargs) - elif input_type == IO.INT: - result = _model_field_to_int_input(field_info, **kwargs) - elif input_type == IO.COMBO: - result = _model_field_to_combo_input(field_info, **kwargs) - else: - message = f"Invalid input type: {input_type}" - raise ValueError(message) - - return result diff --git a/comfy_api_nodes/nodes_bfl.py b/comfy_api_nodes/nodes_bfl.py index 76021ef7f..61c3b4503 100644 --- a/comfy_api_nodes/nodes_bfl.py +++ b/comfy_api_nodes/nodes_bfl.py @@ -3,7 +3,7 @@ from pydantic import BaseModel from typing_extensions import override from comfy_api.latest import IO, ComfyExtension, Input -from comfy_api_nodes.apis.bfl_api import ( +from comfy_api_nodes.apis.bfl import ( BFLFluxExpandImageRequest, BFLFluxFillImageRequest, BFLFluxKontextProGenerateRequest, diff --git a/comfy_api_nodes/nodes_bytedance.py b/comfy_api_nodes/nodes_bytedance.py index 9cb1ca004..486801150 100644 --- a/comfy_api_nodes/nodes_bytedance.py +++ b/comfy_api_nodes/nodes_bytedance.py @@ -5,7 +5,7 @@ import torch from typing_extensions import override from comfy_api.latest import IO, ComfyExtension, Input -from comfy_api_nodes.apis.bytedance_api import ( +from comfy_api_nodes.apis.bytedance import ( RECOMMENDED_PRESETS, RECOMMENDED_PRESETS_SEEDREAM_4, VIDEO_TASKS_EXECUTION_TIME, diff --git a/comfy_api_nodes/nodes_gemini.py b/comfy_api_nodes/nodes_gemini.py index a2daea50a..3b31caa7b 100644 --- a/comfy_api_nodes/nodes_gemini.py +++ b/comfy_api_nodes/nodes_gemini.py @@ -14,7 +14,7 @@ from typing_extensions import override import folder_paths from comfy_api.latest import IO, ComfyExtension, Input, Types -from comfy_api_nodes.apis.gemini_api import ( +from comfy_api_nodes.apis.gemini import ( GeminiContent, GeminiFileData, GeminiGenerateContentRequest, diff --git a/comfy_api_nodes/nodes_ideogram.py b/comfy_api_nodes/nodes_ideogram.py index 827b3523a..feaf7a858 100644 --- a/comfy_api_nodes/nodes_ideogram.py +++ b/comfy_api_nodes/nodes_ideogram.py @@ -4,7 +4,7 @@ from comfy_api.latest import IO, ComfyExtension from PIL import Image import numpy as np import torch -from comfy_api_nodes.apis import ( +from comfy_api_nodes.apis.ideogram import ( IdeogramGenerateRequest, IdeogramGenerateResponse, ImageRequest, diff --git a/comfy_api_nodes/nodes_kling.py b/comfy_api_nodes/nodes_kling.py index 05dde88b1..3ec71530b 100644 --- a/comfy_api_nodes/nodes_kling.py +++ b/comfy_api_nodes/nodes_kling.py @@ -49,7 +49,7 @@ from comfy_api_nodes.apis import ( KlingCharacterEffectModelName, KlingSingleImageEffectModelName, ) -from comfy_api_nodes.apis.kling_api import ( +from comfy_api_nodes.apis.kling import ( ImageToVideoWithAudioRequest, MotionControlRequest, OmniImageParamImage, diff --git a/comfy_api_nodes/nodes_luma.py b/comfy_api_nodes/nodes_luma.py index 95cb442e5..9ed6cd299 100644 --- a/comfy_api_nodes/nodes_luma.py +++ b/comfy_api_nodes/nodes_luma.py @@ -4,7 +4,7 @@ import torch from typing_extensions import override from comfy_api.latest import IO, ComfyExtension -from comfy_api_nodes.apis.luma_api import ( +from comfy_api_nodes.apis.luma import ( LumaAspectRatio, LumaCharacterRef, LumaConceptChain, diff --git a/comfy_api_nodes/nodes_minimax.py b/comfy_api_nodes/nodes_minimax.py index 43a15d50d..b5d0b461f 100644 --- a/comfy_api_nodes/nodes_minimax.py +++ b/comfy_api_nodes/nodes_minimax.py @@ -4,7 +4,7 @@ import torch from typing_extensions import override from comfy_api.latest import IO, ComfyExtension -from comfy_api_nodes.apis.minimax_api import ( +from comfy_api_nodes.apis.minimax import ( MinimaxFileRetrieveResponse, MiniMaxModel, MinimaxTaskResultResponse, diff --git a/comfy_api_nodes/nodes_moonvalley.py b/comfy_api_nodes/nodes_moonvalley.py index 769b171b7..08315fa2b 100644 --- a/comfy_api_nodes/nodes_moonvalley.py +++ b/comfy_api_nodes/nodes_moonvalley.py @@ -3,7 +3,7 @@ import logging from typing_extensions import override from comfy_api.latest import IO, ComfyExtension, Input -from comfy_api_nodes.apis import ( +from comfy_api_nodes.apis.moonvalley import ( MoonvalleyPromptResponse, MoonvalleyTextToVideoInferenceParams, MoonvalleyTextToVideoRequest, diff --git a/comfy_api_nodes/nodes_openai.py b/comfy_api_nodes/nodes_openai.py index 2f144c5c3..a12acc06b 100644 --- a/comfy_api_nodes/nodes_openai.py +++ b/comfy_api_nodes/nodes_openai.py @@ -10,24 +10,18 @@ from typing_extensions import override import folder_paths from comfy_api.latest import IO, ComfyExtension, Input -from comfy_api_nodes.apis import ( - CreateModelResponseProperties, - Detail, - InputContent, +from comfy_api_nodes.apis.openai import ( InputFileContent, InputImageContent, InputMessage, - InputMessageContentList, InputTextContent, - Item, + ModelResponseProperties, OpenAICreateResponse, - OpenAIResponse, - OutputContent, -) -from comfy_api_nodes.apis.openai_api import ( OpenAIImageEditRequest, OpenAIImageGenerationRequest, OpenAIImageGenerationResponse, + OpenAIResponse, + OutputContent, ) from comfy_api_nodes.util import ( ApiEndpoint, @@ -266,7 +260,7 @@ class OpenAIDalle3(IO.ComfyNode): "seed", default=0, min=0, - max=2 ** 31 - 1, + max=2**31 - 1, step=1, display_mode=IO.NumberDisplay.number, control_after_generate=True, @@ -384,7 +378,7 @@ class OpenAIGPTImage1(IO.ComfyNode): "seed", default=0, min=0, - max=2 ** 31 - 1, + max=2**31 - 1, step=1, display_mode=IO.NumberDisplay.number, control_after_generate=True, @@ -500,8 +494,8 @@ class OpenAIGPTImage1(IO.ComfyNode): files = [] batch_size = image.shape[0] for i in range(batch_size): - single_image = image[i: i + 1] - scaled_image = downscale_image_tensor(single_image, total_pixels=2048*2048).squeeze() + single_image = image[i : i + 1] + scaled_image = downscale_image_tensor(single_image, total_pixels=2048 * 2048).squeeze() image_np = (scaled_image.numpy() * 255).astype(np.uint8) img = Image.fromarray(image_np) @@ -523,7 +517,7 @@ class OpenAIGPTImage1(IO.ComfyNode): rgba_mask = torch.zeros(height, width, 4, device="cpu") rgba_mask[:, :, 3] = 1 - mask.squeeze().cpu() - scaled_mask = downscale_image_tensor(rgba_mask.unsqueeze(0), total_pixels=2048*2048).squeeze() + scaled_mask = downscale_image_tensor(rgba_mask.unsqueeze(0), total_pixels=2048 * 2048).squeeze() mask_np = (scaled_mask.numpy() * 255).astype(np.uint8) mask_img = Image.fromarray(mask_np) @@ -696,29 +690,23 @@ class OpenAIChatNode(IO.ComfyNode): ) @classmethod - def get_message_content_from_response( - cls, response: OpenAIResponse - ) -> list[OutputContent]: + def get_message_content_from_response(cls, response: OpenAIResponse) -> list[OutputContent]: """Extract message content from the API response.""" for output in response.output: - if output.root.type == "message": - return output.root.content + if output.type == "message": + return output.content raise TypeError("No output message found in response") @classmethod - def get_text_from_message_content( - cls, message_content: list[OutputContent] - ) -> str: + def get_text_from_message_content(cls, message_content: list[OutputContent]) -> str: """Extract text content from message content.""" for content_item in message_content: - if content_item.root.type == "output_text": - return str(content_item.root.text) + if content_item.type == "output_text": + return str(content_item.text) return "No text output found in response" @classmethod - def tensor_to_input_image_content( - cls, image: torch.Tensor, detail_level: Detail = "auto" - ) -> InputImageContent: + def tensor_to_input_image_content(cls, image: torch.Tensor, detail_level: str = "auto") -> InputImageContent: """Convert a tensor to an input image content object.""" return InputImageContent( detail=detail_level, @@ -732,9 +720,9 @@ class OpenAIChatNode(IO.ComfyNode): prompt: str, image: torch.Tensor | None = None, files: list[InputFileContent] | None = None, - ) -> InputMessageContentList: + ) -> list[InputTextContent | InputImageContent | InputFileContent]: """Create a list of input message contents from prompt and optional image.""" - content_list: list[InputContent | InputTextContent | InputImageContent | InputFileContent] = [ + content_list: list[InputTextContent | InputImageContent | InputFileContent] = [ InputTextContent(text=prompt, type="input_text"), ] if image is not None: @@ -746,13 +734,9 @@ class OpenAIChatNode(IO.ComfyNode): type="input_image", ) ) - if files is not None: content_list.extend(files) - - return InputMessageContentList( - root=content_list, - ) + return content_list @classmethod async def execute( @@ -762,7 +746,7 @@ class OpenAIChatNode(IO.ComfyNode): model: SupportedOpenAIModel = SupportedOpenAIModel.gpt_5.value, images: torch.Tensor | None = None, files: list[InputFileContent] | None = None, - advanced_options: CreateModelResponseProperties | None = None, + advanced_options: ModelResponseProperties | None = None, ) -> IO.NodeOutput: validate_string(prompt, strip_whitespace=False) @@ -773,36 +757,28 @@ class OpenAIChatNode(IO.ComfyNode): response_model=OpenAIResponse, data=OpenAICreateResponse( input=[ - Item( - root=InputMessage( - content=cls.create_input_message_contents( - prompt, images, files - ), - role="user", - ) + InputMessage( + content=cls.create_input_message_contents(prompt, images, files), + role="user", ), ], store=True, stream=False, model=model, previous_response_id=None, - **( - advanced_options.model_dump(exclude_none=True) - if advanced_options - else {} - ), + **(advanced_options.model_dump(exclude_none=True) if advanced_options else {}), ), ) response_id = create_response.id # Get result output result_response = await poll_op( - cls, - ApiEndpoint(path=f"{RESPONSES_ENDPOINT}/{response_id}"), - response_model=OpenAIResponse, - status_extractor=lambda response: response.status, - completed_statuses=["incomplete", "completed"] - ) + cls, + ApiEndpoint(path=f"{RESPONSES_ENDPOINT}/{response_id}"), + response_model=OpenAIResponse, + status_extractor=lambda response: response.status, + completed_statuses=["incomplete", "completed"], + ) return IO.NodeOutput(cls.get_text_from_message_content(cls.get_message_content_from_response(result_response))) @@ -923,7 +899,7 @@ class OpenAIChatConfig(IO.ComfyNode): remove depending on model choice. """ return IO.NodeOutput( - CreateModelResponseProperties( + ModelResponseProperties( instructions=instructions, truncation=truncation, max_output_tokens=max_output_tokens, diff --git a/comfy_api_nodes/nodes_pixverse.py b/comfy_api_nodes/nodes_pixverse.py index 86ddb3ab9..e17a24ae7 100644 --- a/comfy_api_nodes/nodes_pixverse.py +++ b/comfy_api_nodes/nodes_pixverse.py @@ -1,7 +1,7 @@ import torch from typing_extensions import override from comfy_api.latest import IO, ComfyExtension -from comfy_api_nodes.apis.pixverse_api import ( +from comfy_api_nodes.apis.pixverse import ( PixverseTextVideoRequest, PixverseImageVideoRequest, PixverseTransitionVideoRequest, diff --git a/comfy_api_nodes/nodes_recraft.py b/comfy_api_nodes/nodes_recraft.py index 05dc151ad..c01bcaece 100644 --- a/comfy_api_nodes/nodes_recraft.py +++ b/comfy_api_nodes/nodes_recraft.py @@ -8,7 +8,7 @@ from typing_extensions import override from comfy.utils import ProgressBar from comfy_api.latest import IO, ComfyExtension -from comfy_api_nodes.apis.recraft_api import ( +from comfy_api_nodes.apis.recraft import ( RecraftColor, RecraftColorChain, RecraftControls, diff --git a/comfy_api_nodes/nodes_rodin.py b/comfy_api_nodes/nodes_rodin.py index b4420cb93..3ffdc8b90 100644 --- a/comfy_api_nodes/nodes_rodin.py +++ b/comfy_api_nodes/nodes_rodin.py @@ -14,7 +14,7 @@ from typing import Optional from io import BytesIO from typing_extensions import override from PIL import Image -from comfy_api_nodes.apis.rodin_api import ( +from comfy_api_nodes.apis.rodin import ( Rodin3DGenerateRequest, Rodin3DGenerateResponse, Rodin3DCheckStatusRequest, diff --git a/comfy_api_nodes/nodes_runway.py b/comfy_api_nodes/nodes_runway.py index d19fdb365..573170ba2 100644 --- a/comfy_api_nodes/nodes_runway.py +++ b/comfy_api_nodes/nodes_runway.py @@ -16,7 +16,7 @@ from enum import Enum from typing_extensions import override from comfy_api.latest import IO, ComfyExtension, Input, InputImpl -from comfy_api_nodes.apis import ( +from comfy_api_nodes.apis.runway import ( RunwayImageToVideoRequest, RunwayImageToVideoResponse, RunwayTaskStatusResponse as TaskStatusResponse, diff --git a/comfy_api_nodes/nodes_stability.py b/comfy_api_nodes/nodes_stability.py index 5c48c1f1e..5665109cf 100644 --- a/comfy_api_nodes/nodes_stability.py +++ b/comfy_api_nodes/nodes_stability.py @@ -3,7 +3,7 @@ from typing import Optional from typing_extensions import override from comfy_api.latest import ComfyExtension, Input, IO -from comfy_api_nodes.apis.stability_api import ( +from comfy_api_nodes.apis.stability import ( StabilityUpscaleConservativeRequest, StabilityUpscaleCreativeRequest, StabilityAsyncResponse, diff --git a/comfy_api_nodes/nodes_topaz.py b/comfy_api_nodes/nodes_topaz.py index 9dc5f45bc..c052e7656 100644 --- a/comfy_api_nodes/nodes_topaz.py +++ b/comfy_api_nodes/nodes_topaz.py @@ -5,7 +5,24 @@ import aiohttp from typing_extensions import override from comfy_api.latest import IO, ComfyExtension, Input -from comfy_api_nodes.apis import topaz_api +from comfy_api_nodes.apis.topaz import ( + CreateVideoRequest, + CreateVideoRequestSource, + CreateVideoResponse, + ImageAsyncTaskResponse, + ImageDownloadResponse, + ImageEnhanceRequest, + ImageStatusResponse, + OutputInformationVideo, + Resolution, + VideoAcceptResponse, + VideoCompleteUploadRequest, + VideoCompleteUploadRequestPart, + VideoCompleteUploadResponse, + VideoEnhancementFilter, + VideoFrameInterpolationFilter, + VideoStatusResponse, +) from comfy_api_nodes.util import ( ApiEndpoint, download_url_to_image_tensor, @@ -153,13 +170,13 @@ class TopazImageEnhance(IO.ComfyNode): if get_number_of_images(image) != 1: raise ValueError("Only one input image is supported.") download_url = await upload_images_to_comfyapi( - cls, image, max_images=1, mime_type="image/png", total_pixels=4096*4096 + cls, image, max_images=1, mime_type="image/png", total_pixels=4096 * 4096 ) initial_response = await sync_op( cls, ApiEndpoint(path="/proxy/topaz/image/v1/enhance-gen/async", method="POST"), - response_model=topaz_api.ImageAsyncTaskResponse, - data=topaz_api.ImageEnhanceRequest( + response_model=ImageAsyncTaskResponse, + data=ImageEnhanceRequest( model=model, prompt=prompt, subject_detection=subject_detection, @@ -181,7 +198,7 @@ class TopazImageEnhance(IO.ComfyNode): await poll_op( cls, poll_endpoint=ApiEndpoint(path=f"/proxy/topaz/image/v1/status/{initial_response.process_id}"), - response_model=topaz_api.ImageStatusResponse, + response_model=ImageStatusResponse, status_extractor=lambda x: x.status, progress_extractor=lambda x: getattr(x, "progress", 0), price_extractor=lambda x: x.credits * 0.08, @@ -193,7 +210,7 @@ class TopazImageEnhance(IO.ComfyNode): results = await sync_op( cls, ApiEndpoint(path=f"/proxy/topaz/image/v1/download/{initial_response.process_id}"), - response_model=topaz_api.ImageDownloadResponse, + response_model=ImageDownloadResponse, monitor_progress=False, ) return IO.NodeOutput(await download_url_to_image_tensor(results.download_url)) @@ -331,7 +348,7 @@ class TopazVideoEnhance(IO.ComfyNode): if target_height % 2 != 0: target_height += 1 filters.append( - topaz_api.VideoEnhancementFilter( + VideoEnhancementFilter( model=UPSCALER_MODELS_MAP[upscaler_model], creativity=(upscaler_creativity if UPSCALER_MODELS_MAP[upscaler_model] == "slc-1" else None), isOptimizedMode=(True if UPSCALER_MODELS_MAP[upscaler_model] == "slc-1" else None), @@ -340,7 +357,7 @@ class TopazVideoEnhance(IO.ComfyNode): if interpolation_enabled: target_frame_rate = interpolation_frame_rate filters.append( - topaz_api.VideoFrameInterpolationFilter( + VideoFrameInterpolationFilter( model=interpolation_model, slowmo=interpolation_slowmo, fps=interpolation_frame_rate, @@ -351,19 +368,19 @@ class TopazVideoEnhance(IO.ComfyNode): initial_res = await sync_op( cls, ApiEndpoint(path="/proxy/topaz/video/", method="POST"), - response_model=topaz_api.CreateVideoResponse, - data=topaz_api.CreateVideoRequest( - source=topaz_api.CreateCreateVideoRequestSource( + response_model=CreateVideoResponse, + data=CreateVideoRequest( + source=CreateVideoRequestSource( container="mp4", size=get_fs_object_size(src_video_stream), duration=int(duration_sec), frameCount=video.get_frame_count(), frameRate=src_frame_rate, - resolution=topaz_api.Resolution(width=src_width, height=src_height), + resolution=Resolution(width=src_width, height=src_height), ), filters=filters, - output=topaz_api.OutputInformationVideo( - resolution=topaz_api.Resolution(width=target_width, height=target_height), + output=OutputInformationVideo( + resolution=Resolution(width=target_width, height=target_height), frameRate=target_frame_rate, audioCodec="AAC", audioTransfer="Copy", @@ -379,7 +396,7 @@ class TopazVideoEnhance(IO.ComfyNode): path=f"/proxy/topaz/video/{initial_res.requestId}/accept", method="PATCH", ), - response_model=topaz_api.VideoAcceptResponse, + response_model=VideoAcceptResponse, wait_label="Preparing upload", final_label_on_success="Upload started", ) @@ -402,10 +419,10 @@ class TopazVideoEnhance(IO.ComfyNode): path=f"/proxy/topaz/video/{initial_res.requestId}/complete-upload", method="PATCH", ), - response_model=topaz_api.VideoCompleteUploadResponse, - data=topaz_api.VideoCompleteUploadRequest( + response_model=VideoCompleteUploadResponse, + data=VideoCompleteUploadRequest( uploadResults=[ - topaz_api.VideoCompleteUploadRequestPart( + VideoCompleteUploadRequestPart( partNum=1, eTag=upload_etag, ), @@ -417,7 +434,7 @@ class TopazVideoEnhance(IO.ComfyNode): final_response = await poll_op( cls, ApiEndpoint(path=f"/proxy/topaz/video/{initial_res.requestId}/status"), - response_model=topaz_api.VideoStatusResponse, + response_model=VideoStatusResponse, status_extractor=lambda x: x.status, progress_extractor=lambda x: getattr(x, "progress", 0), price_extractor=lambda x: (x.estimates.cost[0] * 0.08 if x.estimates and x.estimates.cost[0] else None), diff --git a/comfy_api_nodes/nodes_tripo.py b/comfy_api_nodes/nodes_tripo.py index aa790143d..5abf27b4d 100644 --- a/comfy_api_nodes/nodes_tripo.py +++ b/comfy_api_nodes/nodes_tripo.py @@ -5,7 +5,7 @@ import torch from typing_extensions import override from comfy_api.latest import IO, ComfyExtension -from comfy_api_nodes.apis.tripo_api import ( +from comfy_api_nodes.apis.tripo import ( TripoAnimateRetargetRequest, TripoAnimateRigRequest, TripoConvertModelRequest, diff --git a/comfy_api_nodes/nodes_veo2.py b/comfy_api_nodes/nodes_veo2.py index c14d6ad68..2a202fc3b 100644 --- a/comfy_api_nodes/nodes_veo2.py +++ b/comfy_api_nodes/nodes_veo2.py @@ -4,7 +4,7 @@ from io import BytesIO from typing_extensions import override from comfy_api.latest import IO, ComfyExtension, Input, InputImpl -from comfy_api_nodes.apis.veo_api import ( +from comfy_api_nodes.apis.veo import ( VeoGenVidPollRequest, VeoGenVidPollResponse, VeoGenVidRequest, diff --git a/comfy_api_nodes/redocly-dev.yaml b/comfy_api_nodes/redocly-dev.yaml deleted file mode 100644 index d9e3cab70..000000000 --- a/comfy_api_nodes/redocly-dev.yaml +++ /dev/null @@ -1,10 +0,0 @@ -# This file is used to filter the Comfy Org OpenAPI spec for schemas related to API Nodes. -# This is used for development purposes to generate stubs for unreleased API endpoints. -apis: - filter: - root: openapi.yaml - decorators: - filter-in: - property: tags - value: ['API Nodes'] - matchStrategy: all diff --git a/comfy_api_nodes/redocly.yaml b/comfy_api_nodes/redocly.yaml deleted file mode 100644 index d102345b1..000000000 --- a/comfy_api_nodes/redocly.yaml +++ /dev/null @@ -1,10 +0,0 @@ -# This file is used to filter the Comfy Org OpenAPI spec for schemas related to API Nodes. - -apis: - filter: - root: openapi.yaml - decorators: - filter-in: - property: tags - value: ['API Nodes', 'Released'] - matchStrategy: all diff --git a/tests-unit/comfy_api_nodes_test/mapper_utils_test.py b/tests-unit/comfy_api_nodes_test/mapper_utils_test.py deleted file mode 100644 index 69488f691..000000000 --- a/tests-unit/comfy_api_nodes_test/mapper_utils_test.py +++ /dev/null @@ -1,297 +0,0 @@ -from typing import Optional -from enum import Enum - -from pydantic import BaseModel, Field - -from comfy.comfy_types.node_typing import IO -from comfy_api_nodes.mapper_utils import model_field_to_node_input - - -def test_model_field_to_float_input(): - """Tests mapping a float field with constraints.""" - - class ModelWithFloatField(BaseModel): - cfg_scale: Optional[float] = Field( - default=0.5, - description="Flexibility in video generation", - ge=0.0, - le=1.0, - multiple_of=0.001, - ) - - expected_output = ( - IO.FLOAT, - { - "default": 0.5, - "tooltip": "Flexibility in video generation", - "min": 0.0, - "max": 1.0, - "step": 0.001, - }, - ) - - actual_output = model_field_to_node_input( - IO.FLOAT, ModelWithFloatField, "cfg_scale" - ) - - assert actual_output[0] == expected_output[0] - assert actual_output[1] == expected_output[1] - - -def test_model_field_to_float_input_no_constraints(): - """Tests mapping a float field with no constraints.""" - - class ModelWithFloatField(BaseModel): - cfg_scale: Optional[float] = Field(default=0.5) - - expected_output = ( - IO.FLOAT, - { - "default": 0.5, - }, - ) - - actual_output = model_field_to_node_input( - IO.FLOAT, ModelWithFloatField, "cfg_scale" - ) - - assert actual_output[0] == expected_output[0] - assert actual_output[1] == expected_output[1] - - -def test_model_field_to_int_input(): - """Tests mapping an int field with constraints.""" - - class ModelWithIntField(BaseModel): - num_frames: Optional[int] = Field( - default=10, - description="Number of frames to generate", - ge=1, - le=100, - multiple_of=1, - ) - - expected_output = ( - IO.INT, - { - "default": 10, - "tooltip": "Number of frames to generate", - "min": 1, - "max": 100, - "step": 1, - }, - ) - - actual_output = model_field_to_node_input(IO.INT, ModelWithIntField, "num_frames") - - assert actual_output[0] == expected_output[0] - assert actual_output[1] == expected_output[1] - - -def test_model_field_to_string_input(): - """Tests mapping a string field.""" - - class ModelWithStringField(BaseModel): - prompt: Optional[str] = Field( - default="A beautiful sunset over a calm ocean", - description="A prompt for the video generation", - ) - - expected_output = ( - IO.STRING, - { - "default": "A beautiful sunset over a calm ocean", - "tooltip": "A prompt for the video generation", - }, - ) - - actual_output = model_field_to_node_input(IO.STRING, ModelWithStringField, "prompt") - - assert actual_output[0] == expected_output[0] - assert actual_output[1] == expected_output[1] - - -def test_model_field_to_string_input_multiline(): - """Tests mapping a string field.""" - - class ModelWithStringField(BaseModel): - prompt: Optional[str] = Field( - default="A beautiful sunset over a calm ocean", - description="A prompt for the video generation", - ) - - expected_output = ( - IO.STRING, - { - "default": "A beautiful sunset over a calm ocean", - "tooltip": "A prompt for the video generation", - "multiline": True, - }, - ) - - actual_output = model_field_to_node_input( - IO.STRING, ModelWithStringField, "prompt", multiline=True - ) - - assert actual_output[0] == expected_output[0] - assert actual_output[1] == expected_output[1] - - -def test_model_field_to_combo_input(): - """Tests mapping a combo field.""" - - class MockEnum(str, Enum): - option_1 = "option 1" - option_2 = "option 2" - option_3 = "option 3" - - class ModelWithComboField(BaseModel): - model_name: Optional[MockEnum] = Field("option 1", description="Model Name") - - expected_output = ( - IO.COMBO, - { - "options": ["option 1", "option 2", "option 3"], - "default": "option 1", - "tooltip": "Model Name", - }, - ) - - actual_output = model_field_to_node_input( - IO.COMBO, ModelWithComboField, "model_name", enum_type=MockEnum - ) - - assert actual_output[0] == expected_output[0] - assert actual_output[1] == expected_output[1] - - -def test_model_field_to_combo_input_no_options(): - """Tests mapping a combo field with no options.""" - - class ModelWithComboField(BaseModel): - model_name: Optional[str] = Field(description="Model Name") - - expected_output = ( - IO.COMBO, - { - "tooltip": "Model Name", - }, - ) - - actual_output = model_field_to_node_input( - IO.COMBO, ModelWithComboField, "model_name" - ) - - assert actual_output[0] == expected_output[0] - assert actual_output[1] == expected_output[1] - - -def test_model_field_to_image_input(): - """Tests mapping an image field.""" - - class ModelWithImageField(BaseModel): - image: Optional[str] = Field( - default=None, - description="An image for the video generation", - ) - - expected_output = ( - IO.IMAGE, - { - "default": None, - "tooltip": "An image for the video generation", - }, - ) - - actual_output = model_field_to_node_input(IO.IMAGE, ModelWithImageField, "image") - - assert actual_output[0] == expected_output[0] - assert actual_output[1] == expected_output[1] - - -def test_model_field_to_node_input_no_description(): - """Tests mapping a field with no description.""" - - class ModelWithNoDescriptionField(BaseModel): - field: Optional[str] = Field(default="default value") - - expected_output = ( - IO.STRING, - { - "default": "default value", - }, - ) - - actual_output = model_field_to_node_input( - IO.STRING, ModelWithNoDescriptionField, "field" - ) - - assert actual_output[0] == expected_output[0] - assert actual_output[1] == expected_output[1] - - -def test_model_field_to_node_input_no_default(): - """Tests mapping a field with no default.""" - - class ModelWithNoDefaultField(BaseModel): - field: Optional[str] = Field(description="A field with no default") - - expected_output = ( - IO.STRING, - { - "tooltip": "A field with no default", - }, - ) - - actual_output = model_field_to_node_input( - IO.STRING, ModelWithNoDefaultField, "field" - ) - - assert actual_output[0] == expected_output[0] - assert actual_output[1] == expected_output[1] - - -def test_model_field_to_node_input_no_metadata(): - """Tests mapping a field with no metadata or properties defined on the schema.""" - - class ModelWithNoMetadataField(BaseModel): - field: Optional[str] = Field() - - expected_output = ( - IO.STRING, - {}, - ) - - actual_output = model_field_to_node_input( - IO.STRING, ModelWithNoMetadataField, "field" - ) - - assert actual_output[0] == expected_output[0] - assert actual_output[1] == expected_output[1] - - -def test_model_field_to_node_input_default_is_none(): - """ - Tests mapping a field with a default of `None`. - I.e., the default field should be included as the schema explicitly sets it to `None`. - """ - - class ModelWithNoneDefaultField(BaseModel): - field: Optional[str] = Field( - default=None, description="A field with a default of None" - ) - - expected_output = ( - IO.STRING, - { - "default": None, - "tooltip": "A field with a default of None", - }, - ) - - actual_output = model_field_to_node_input( - IO.STRING, ModelWithNoneDefaultField, "field" - ) - - assert actual_output[0] == expected_output[0] - assert actual_output[1] == expected_output[1] From f7ca41ff6226eecbf6c9ee475c1de714cb8f04e9 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Sun, 18 Jan 2026 04:57:57 +0200 Subject: [PATCH 469/647] chore(api-nodes): remove check for pyav>=14.2 in code (it was added to requirements.txt long ago) (#11934) --- comfy_api_nodes/canary.py | 10 ---------- nodes.py | 3 --- 2 files changed, 13 deletions(-) delete mode 100644 comfy_api_nodes/canary.py diff --git a/comfy_api_nodes/canary.py b/comfy_api_nodes/canary.py deleted file mode 100644 index 4df7590b6..000000000 --- a/comfy_api_nodes/canary.py +++ /dev/null @@ -1,10 +0,0 @@ -import av - -ver = av.__version__.split(".") -if int(ver[0]) < 14: - raise Exception("INSTALL NEW VERSION OF PYAV TO USE API NODES.") - -if int(ver[0]) == 14 and int(ver[1]) < 2: - raise Exception("INSTALL NEW VERSION OF PYAV TO USE API NODES.") - -NODE_CLASS_MAPPINGS = {} diff --git a/nodes.py b/nodes.py index f19d5fd1c..8b5279b36 100644 --- a/nodes.py +++ b/nodes.py @@ -2409,9 +2409,6 @@ async def init_builtin_api_nodes(): "nodes_wan.py", ] - if not await load_custom_node(os.path.join(api_nodes_dir, "canary.py"), module_parent="comfy_api_nodes"): - return api_nodes_files - import_failed = [] for node_file in api_nodes_files: if not await load_custom_node(os.path.join(api_nodes_dir, node_file), module_parent="comfy_api_nodes"): From a498556d0dcde3d7a7c19e1f5c733c8c2a2ffb10 Mon Sep 17 00:00:00 2001 From: Christian Byrne Date: Sat, 17 Jan 2026 19:06:03 -0800 Subject: [PATCH 470/647] feat: add advanced parameter to Input classes for advanced widgets support (#11939) Add 'advanced' boolean parameter to Input and WidgetInput base classes and propagate to all typed Input subclasses (Boolean, Int, Float, String, Combo, MultiCombo, Webcam, MultiType, MatchType, ImageCompare). When set to True, the frontend will hide these inputs by default in a collapsible 'Advanced Inputs' section in the right side panel, reducing visual clutter for power-user options. This enables nodes to expose advanced configuration options (like encoding parameters, quality settings, etc.) without overwhelming typical users. Frontend support: ComfyUI_frontend PR #7812 --- comfy_api/latest/_io.py | 47 ++++++++++++++++++++++------------------- 1 file changed, 25 insertions(+), 22 deletions(-) diff --git a/comfy_api/latest/_io.py b/comfy_api/latest/_io.py index e6a0d1821..c30d92aaa 100644 --- a/comfy_api/latest/_io.py +++ b/comfy_api/latest/_io.py @@ -153,7 +153,7 @@ class Input(_IO_V3): ''' Base class for a V3 Input. ''' - def __init__(self, id: str, display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, extra_dict=None, raw_link: bool=None): + def __init__(self, id: str, display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, extra_dict=None, raw_link: bool=None, advanced: bool=None): super().__init__() self.id = id self.display_name = display_name @@ -162,6 +162,7 @@ class Input(_IO_V3): self.lazy = lazy self.extra_dict = extra_dict if extra_dict is not None else {} self.rawLink = raw_link + self.advanced = advanced def as_dict(self): return prune_dict({ @@ -170,6 +171,7 @@ class Input(_IO_V3): "tooltip": self.tooltip, "lazy": self.lazy, "rawLink": self.rawLink, + "advanced": self.advanced, }) | prune_dict(self.extra_dict) def get_io_type(self): @@ -184,8 +186,8 @@ class WidgetInput(Input): ''' def __init__(self, id: str, display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, default: Any=None, - socketless: bool=None, widget_type: str=None, force_input: bool=None, extra_dict=None, raw_link: bool=None): - super().__init__(id, display_name, optional, tooltip, lazy, extra_dict, raw_link) + socketless: bool=None, widget_type: str=None, force_input: bool=None, extra_dict=None, raw_link: bool=None, advanced: bool=None): + super().__init__(id, display_name, optional, tooltip, lazy, extra_dict, raw_link, advanced) self.default = default self.socketless = socketless self.widget_type = widget_type @@ -242,8 +244,8 @@ class Boolean(ComfyTypeIO): '''Boolean input.''' def __init__(self, id: str, display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, default: bool=None, label_on: str=None, label_off: str=None, - socketless: bool=None, force_input: bool=None, extra_dict=None, raw_link: bool=None): - super().__init__(id, display_name, optional, tooltip, lazy, default, socketless, None, force_input, extra_dict, raw_link) + socketless: bool=None, force_input: bool=None, extra_dict=None, raw_link: bool=None, advanced: bool=None): + super().__init__(id, display_name, optional, tooltip, lazy, default, socketless, None, force_input, extra_dict, raw_link, advanced) self.label_on = label_on self.label_off = label_off self.default: bool @@ -262,8 +264,8 @@ class Int(ComfyTypeIO): '''Integer input.''' def __init__(self, id: str, display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, default: int=None, min: int=None, max: int=None, step: int=None, control_after_generate: bool=None, - display_mode: NumberDisplay=None, socketless: bool=None, force_input: bool=None, extra_dict=None, raw_link: bool=None): - super().__init__(id, display_name, optional, tooltip, lazy, default, socketless, None, force_input, extra_dict, raw_link) + display_mode: NumberDisplay=None, socketless: bool=None, force_input: bool=None, extra_dict=None, raw_link: bool=None, advanced: bool=None): + super().__init__(id, display_name, optional, tooltip, lazy, default, socketless, None, force_input, extra_dict, raw_link, advanced) self.min = min self.max = max self.step = step @@ -288,8 +290,8 @@ class Float(ComfyTypeIO): '''Float input.''' def __init__(self, id: str, display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, default: float=None, min: float=None, max: float=None, step: float=None, round: float=None, - display_mode: NumberDisplay=None, socketless: bool=None, force_input: bool=None, extra_dict=None, raw_link: bool=None): - super().__init__(id, display_name, optional, tooltip, lazy, default, socketless, None, force_input, extra_dict, raw_link) + display_mode: NumberDisplay=None, socketless: bool=None, force_input: bool=None, extra_dict=None, raw_link: bool=None, advanced: bool=None): + super().__init__(id, display_name, optional, tooltip, lazy, default, socketless, None, force_input, extra_dict, raw_link, advanced) self.min = min self.max = max self.step = step @@ -314,8 +316,8 @@ class String(ComfyTypeIO): '''String input.''' def __init__(self, id: str, display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, multiline=False, placeholder: str=None, default: str=None, dynamic_prompts: bool=None, - socketless: bool=None, force_input: bool=None, extra_dict=None, raw_link: bool=None): - super().__init__(id, display_name, optional, tooltip, lazy, default, socketless, None, force_input, extra_dict, raw_link) + socketless: bool=None, force_input: bool=None, extra_dict=None, raw_link: bool=None, advanced: bool=None): + super().__init__(id, display_name, optional, tooltip, lazy, default, socketless, None, force_input, extra_dict, raw_link, advanced) self.multiline = multiline self.placeholder = placeholder self.dynamic_prompts = dynamic_prompts @@ -350,12 +352,13 @@ class Combo(ComfyTypeIO): socketless: bool=None, extra_dict=None, raw_link: bool=None, + advanced: bool=None, ): if isinstance(options, type) and issubclass(options, Enum): options = [v.value for v in options] if isinstance(default, Enum): default = default.value - super().__init__(id, display_name, optional, tooltip, lazy, default, socketless, None, None, extra_dict, raw_link) + super().__init__(id, display_name, optional, tooltip, lazy, default, socketless, None, None, extra_dict, raw_link, advanced) self.multiselect = False self.options = options self.control_after_generate = control_after_generate @@ -387,8 +390,8 @@ class MultiCombo(ComfyTypeI): class Input(Combo.Input): def __init__(self, id: str, options: list[str], display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, default: list[str]=None, placeholder: str=None, chip: bool=None, control_after_generate: bool=None, - socketless: bool=None, extra_dict=None, raw_link: bool=None): - super().__init__(id, options, display_name, optional, tooltip, lazy, default, control_after_generate, socketless=socketless, extra_dict=extra_dict, raw_link=raw_link) + socketless: bool=None, extra_dict=None, raw_link: bool=None, advanced: bool=None): + super().__init__(id, options, display_name, optional, tooltip, lazy, default, control_after_generate, socketless=socketless, extra_dict=extra_dict, raw_link=raw_link, advanced=advanced) self.multiselect = True self.placeholder = placeholder self.chip = chip @@ -421,9 +424,9 @@ class Webcam(ComfyTypeIO): Type = str def __init__( self, id: str, display_name: str=None, optional=False, - tooltip: str=None, lazy: bool=None, default: str=None, socketless: bool=None, extra_dict=None, raw_link: bool=None + tooltip: str=None, lazy: bool=None, default: str=None, socketless: bool=None, extra_dict=None, raw_link: bool=None, advanced: bool=None ): - super().__init__(id, display_name, optional, tooltip, lazy, default, socketless, None, None, extra_dict, raw_link) + super().__init__(id, display_name, optional, tooltip, lazy, default, socketless, None, None, extra_dict, raw_link, advanced) @comfytype(io_type="MASK") @@ -776,7 +779,7 @@ class MultiType: ''' Input that permits more than one input type; if `id` is an instance of `ComfyType.Input`, then that input will be used to create a widget (if applicable) with overridden values. ''' - def __init__(self, id: str | Input, types: list[type[_ComfyType] | _ComfyType], display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, extra_dict=None, raw_link: bool=None): + def __init__(self, id: str | Input, types: list[type[_ComfyType] | _ComfyType], display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, extra_dict=None, raw_link: bool=None, advanced: bool=None): # if id is an Input, then use that Input with overridden values self.input_override = None if isinstance(id, Input): @@ -789,7 +792,7 @@ class MultiType: # if is a widget input, make sure widget_type is set appropriately if isinstance(self.input_override, WidgetInput): self.input_override.widget_type = self.input_override.get_io_type() - super().__init__(id, display_name, optional, tooltip, lazy, extra_dict, raw_link) + super().__init__(id, display_name, optional, tooltip, lazy, extra_dict, raw_link, advanced) self._io_types = types @property @@ -843,8 +846,8 @@ class MatchType(ComfyTypeIO): class Input(Input): def __init__(self, id: str, template: MatchType.Template, - display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, extra_dict=None, raw_link: bool=None): - super().__init__(id, display_name, optional, tooltip, lazy, extra_dict, raw_link) + display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, extra_dict=None, raw_link: bool=None, advanced: bool=None): + super().__init__(id, display_name, optional, tooltip, lazy, extra_dict, raw_link, advanced) self.template = template def as_dict(self): @@ -1119,8 +1122,8 @@ class ImageCompare(ComfyTypeI): class Input(WidgetInput): def __init__(self, id: str, display_name: str=None, optional=False, tooltip: str=None, - socketless: bool=True): - super().__init__(id, display_name, optional, tooltip, None, None, socketless) + socketless: bool=True, advanced: bool=None): + super().__init__(id, display_name, optional, tooltip, None, None, socketless, None, None, None, None, advanced) def as_dict(self): return super().as_dict() From 034fac70549dd9c35b155b80a3ff627ad07b1015 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Sun, 18 Jan 2026 08:40:39 +0200 Subject: [PATCH 471/647] chore(api-nodes): auto-discover all nodes_*.py files to avoid merge conflicts when adding new API nodes (#11943) --- nodes.py | 30 ++++-------------------------- 1 file changed, 4 insertions(+), 26 deletions(-) diff --git a/nodes.py b/nodes.py index 8b5279b36..cba8eacc2 100644 --- a/nodes.py +++ b/nodes.py @@ -5,6 +5,7 @@ import torch import os import sys import json +import glob import hashlib import inspect import traceback @@ -2384,35 +2385,12 @@ async def init_builtin_extra_nodes(): async def init_builtin_api_nodes(): api_nodes_dir = os.path.join(os.path.dirname(os.path.realpath(__file__)), "comfy_api_nodes") - api_nodes_files = [ - "nodes_ideogram.py", - "nodes_openai.py", - "nodes_minimax.py", - "nodes_veo2.py", - "nodes_kling.py", - "nodes_bfl.py", - "nodes_bytedance.py", - "nodes_ltxv.py", - "nodes_luma.py", - "nodes_recraft.py", - "nodes_pixverse.py", - "nodes_stability.py", - "nodes_runway.py", - "nodes_sora.py", - "nodes_topaz.py", - "nodes_tripo.py", - "nodes_meshy.py", - "nodes_moonvalley.py", - "nodes_rodin.py", - "nodes_gemini.py", - "nodes_vidu.py", - "nodes_wan.py", - ] + api_nodes_files = sorted(glob.glob(os.path.join(api_nodes_dir, "nodes_*.py"))) import_failed = [] for node_file in api_nodes_files: - if not await load_custom_node(os.path.join(api_nodes_dir, node_file), module_parent="comfy_api_nodes"): - import_failed.append(node_file) + if not await load_custom_node(node_file, module_parent="comfy_api_nodes"): + import_failed.append(os.path.basename(node_file)) return import_failed From 1a72bf20469dee31ad156f819c14f0172cbad222 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sun, 18 Jan 2026 19:53:43 -0800 Subject: [PATCH 472/647] Readme update. (#11957) --- README.md | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/README.md b/README.md index 123cc9472..c56e05d07 100644 --- a/README.md +++ b/README.md @@ -108,7 +108,7 @@ See what ComfyUI can do with the [example workflows](https://comfyanonymous.gith - [LCM models and Loras](https://comfyanonymous.github.io/ComfyUI_examples/lcm/) - Latent previews with [TAESD](#how-to-show-high-quality-previews) - Works fully offline: core will never download anything unless you want to. -- Optional API nodes to use paid models from external providers through the online [Comfy API](https://docs.comfy.org/tutorials/api-nodes/overview). +- Optional API nodes to use paid models from external providers through the online [Comfy API](https://docs.comfy.org/tutorials/api-nodes/overview) disable with: `--disable-api-nodes` - [Config file](extra_model_paths.yaml.example) to set the search paths for models. Workflow examples can be found on the [Examples page](https://comfyanonymous.github.io/ComfyUI_examples/) @@ -212,7 +212,7 @@ Python 3.14 works but you may encounter issues with the torch compile node. The Python 3.13 is very well supported. If you have trouble with some custom node dependencies on 3.13 you can try 3.12 -torch 2.4 and above is supported but some features might only work on newer versions. We generally recommend using the latest major version of pytorch with the latest cuda version unless it is less than 2 weeks old. +torch 2.4 and above is supported but some features and optimizations might only work on newer versions. We generally recommend using the latest major version of pytorch with the latest cuda version unless it is less than 2 weeks old. ### Instructions: @@ -229,7 +229,7 @@ AMD users can install rocm and pytorch with pip if you don't have it already ins ```pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm6.4``` -This is the command to install the nightly with ROCm 7.0 which might have some performance improvements: +This is the command to install the nightly with ROCm 7.1 which might have some performance improvements: ```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm7.1``` From 866a4619db2db56c77a86e5fc9968a2454928627 Mon Sep 17 00:00:00 2001 From: ComfyUI Wiki Date: Tue, 20 Jan 2026 06:21:35 +0800 Subject: [PATCH 473/647] chore: update workflow templates to v0.8.14 (#11974) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 622256973..312c7c137 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,5 +1,5 @@ comfyui-frontend-package==1.36.14 -comfyui-workflow-templates==0.8.11 +comfyui-workflow-templates==0.8.14 comfyui-embedded-docs==0.4.0 torch torchsde From b931b37e30bb19b6e13ad8623e193ccdaf671a23 Mon Sep 17 00:00:00 2001 From: Jedrzej Kosinski Date: Mon, 19 Jan 2026 16:47:14 -0800 Subject: [PATCH 474/647] feat(api-nodes): add Bria Edit node (#11978) Co-authored-by: Alexander Piskun --- comfy_api_nodes/apis/bria.py | 61 +++++++++ comfy_api_nodes/nodes_bria.py | 198 ++++++++++++++++++++++++++++ comfy_api_nodes/util/__init__.py | 2 + comfy_api_nodes/util/conversions.py | 6 + 4 files changed, 267 insertions(+) create mode 100644 comfy_api_nodes/apis/bria.py create mode 100644 comfy_api_nodes/nodes_bria.py diff --git a/comfy_api_nodes/apis/bria.py b/comfy_api_nodes/apis/bria.py new file mode 100644 index 000000000..9119cacc6 --- /dev/null +++ b/comfy_api_nodes/apis/bria.py @@ -0,0 +1,61 @@ +from typing import TypedDict + +from pydantic import BaseModel, Field + + +class InputModerationSettings(TypedDict): + prompt_content_moderation: bool + visual_input_moderation: bool + visual_output_moderation: bool + + +class BriaEditImageRequest(BaseModel): + instruction: str | None = Field(...) + structured_instruction: str | None = Field( + ..., + description="Use this instead of instruction for precise, programmatic control.", + ) + images: list[str] = Field( + ..., + description="Required. Publicly available URL or Base64-encoded. Must contain exactly one item.", + ) + mask: str | None = Field( + None, + description="Mask image (black and white). Black areas will be preserved, white areas will be edited. " + "If omitted, the edit applies to the entire image. " + "The input image and the the input mask must be of the same size.", + ) + negative_prompt: str | None = Field(None) + guidance_scale: float = Field(...) + model_version: str = Field(...) + steps_num: int = Field(...) + seed: int = Field(...) + ip_signal: bool = Field( + False, + description="If true, returns a warning for potential IP content in the instruction.", + ) + prompt_content_moderation: bool = Field( + False, description="If true, returns 422 on instruction moderation failure." + ) + visual_input_content_moderation: bool = Field( + False, description="If true, returns 422 on images or mask moderation failure." + ) + visual_output_content_moderation: bool = Field( + False, description="If true, returns 422 on visual output moderation failure." + ) + + +class BriaStatusResponse(BaseModel): + request_id: str = Field(...) + status_url: str = Field(...) + warning: str | None = Field(None) + + +class BriaResult(BaseModel): + structured_prompt: str = Field(...) + image_url: str = Field(...) + + +class BriaResponse(BaseModel): + status: str = Field(...) + result: BriaResult | None = Field(None) diff --git a/comfy_api_nodes/nodes_bria.py b/comfy_api_nodes/nodes_bria.py new file mode 100644 index 000000000..72a3055a7 --- /dev/null +++ b/comfy_api_nodes/nodes_bria.py @@ -0,0 +1,198 @@ +from typing_extensions import override + +from comfy_api.latest import IO, ComfyExtension, Input +from comfy_api_nodes.apis.bria import ( + BriaEditImageRequest, + BriaResponse, + BriaStatusResponse, + InputModerationSettings, +) +from comfy_api_nodes.util import ( + ApiEndpoint, + convert_mask_to_image, + download_url_to_image_tensor, + get_number_of_images, + poll_op, + sync_op, + upload_images_to_comfyapi, +) + + +class BriaImageEditNode(IO.ComfyNode): + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="BriaImageEditNode", + display_name="Bria Image Edit", + category="api node/image/Bria", + description="Edit images using Bria latest model", + inputs=[ + IO.Combo.Input("model", options=["FIBO"]), + IO.Image.Input("image"), + IO.String.Input( + "prompt", + multiline=True, + default="", + tooltip="Instruction to edit image", + ), + IO.String.Input("negative_prompt", multiline=True, default=""), + IO.String.Input( + "structured_prompt", + multiline=True, + default="", + tooltip="A string containing the structured edit prompt in JSON format. " + "Use this instead of usual prompt for precise, programmatic control.", + ), + IO.Int.Input( + "seed", + default=1, + min=1, + max=2147483647, + step=1, + display_mode=IO.NumberDisplay.number, + control_after_generate=True, + ), + IO.Float.Input( + "guidance_scale", + default=3, + min=3, + max=5, + step=0.01, + display_mode=IO.NumberDisplay.number, + tooltip="Higher value makes the image follow the prompt more closely.", + ), + IO.Int.Input( + "steps", + default=50, + min=20, + max=50, + step=1, + display_mode=IO.NumberDisplay.number, + ), + IO.DynamicCombo.Input( + "moderation", + options=[ + IO.DynamicCombo.Option( + "true", + [ + IO.Boolean.Input( + "prompt_content_moderation", default=False + ), + IO.Boolean.Input( + "visual_input_moderation", default=False + ), + IO.Boolean.Input( + "visual_output_moderation", default=True + ), + ], + ), + IO.DynamicCombo.Option("false", []), + ], + tooltip="Moderation settings", + ), + IO.Mask.Input( + "mask", + tooltip="If omitted, the edit applies to the entire image.", + optional=True, + ), + ], + outputs=[ + IO.Image.Output(), + IO.String.Output(display_name="structured_prompt"), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.04}""", + ), + ) + + @classmethod + async def execute( + cls, + model: str, + image: Input.Image, + prompt: str, + negative_prompt: str, + structured_prompt: str, + seed: int, + guidance_scale: float, + steps: int, + moderation: InputModerationSettings, + mask: Input.Image | None = None, + ) -> IO.NodeOutput: + if not prompt and not structured_prompt: + raise ValueError( + "One of prompt or structured_prompt is required to be non-empty." + ) + if get_number_of_images(image) != 1: + raise ValueError("Exactly one input image is required.") + mask_url = None + if mask is not None: + mask_url = ( + await upload_images_to_comfyapi( + cls, + convert_mask_to_image(mask), + max_images=1, + mime_type="image/png", + wait_label="Uploading mask", + ) + )[0] + response = await sync_op( + cls, + ApiEndpoint(path="proxy/bria/v2/image/edit", method="POST"), + data=BriaEditImageRequest( + instruction=prompt if prompt else None, + structured_instruction=structured_prompt if structured_prompt else None, + images=await upload_images_to_comfyapi( + cls, + image, + max_images=1, + mime_type="image/png", + wait_label="Uploading image", + ), + mask=mask_url, + negative_prompt=negative_prompt if negative_prompt else None, + guidance_scale=guidance_scale, + seed=seed, + model_version=model, + steps_num=steps, + prompt_content_moderation=moderation.get( + "prompt_content_moderation", False + ), + visual_input_content_moderation=moderation.get( + "visual_input_moderation", False + ), + visual_output_content_moderation=moderation.get( + "visual_output_moderation", False + ), + ), + response_model=BriaStatusResponse, + ) + response = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/bria/v2/status/{response.request_id}"), + status_extractor=lambda r: r.status, + response_model=BriaResponse, + ) + return IO.NodeOutput( + await download_url_to_image_tensor(response.result.image_url), + response.result.structured_prompt, + ) + + +class BriaExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[IO.ComfyNode]]: + return [ + BriaImageEditNode, + ] + + +async def comfy_entrypoint() -> BriaExtension: + return BriaExtension() diff --git a/comfy_api_nodes/util/__init__.py b/comfy_api_nodes/util/__init__.py index 4cc22abfb..364976000 100644 --- a/comfy_api_nodes/util/__init__.py +++ b/comfy_api_nodes/util/__init__.py @@ -11,6 +11,7 @@ from .conversions import ( audio_input_to_mp3, audio_to_base64_string, bytesio_to_image_tensor, + convert_mask_to_image, downscale_image_tensor, image_tensor_pair_to_batch, pil_to_bytesio, @@ -72,6 +73,7 @@ __all__ = [ "audio_input_to_mp3", "audio_to_base64_string", "bytesio_to_image_tensor", + "convert_mask_to_image", "downscale_image_tensor", "image_tensor_pair_to_batch", "pil_to_bytesio", diff --git a/comfy_api_nodes/util/conversions.py b/comfy_api_nodes/util/conversions.py index 99c302a2a..546741b7b 100644 --- a/comfy_api_nodes/util/conversions.py +++ b/comfy_api_nodes/util/conversions.py @@ -451,6 +451,12 @@ def resize_mask_to_image( return mask +def convert_mask_to_image(mask: Input.Image) -> torch.Tensor: + """Make mask have the expected amount of dims (4) and channels (3) to be recognized as an image.""" + mask = mask.unsqueeze(-1) + return torch.cat([mask] * 3, dim=-1) + + def text_filepath_to_base64_string(filepath: str) -> str: """Converts a text file to a base64 string.""" with open(filepath, "rb") as f: From 7458e20465a0efcf91eafc0c65d1929ab7b2238d Mon Sep 17 00:00:00 2001 From: Jedrzej Kosinski Date: Mon, 19 Jan 2026 16:58:30 -0800 Subject: [PATCH 475/647] Make Autogrow validation work properly (#11977) * In-progress autogrow validation fixes - properly looks at required/optional inputs, now working on the edge case that all inputs are optional and nothing is plugged in (should just be an empty dictionary passed into node) * Allow autogrow to work with all inputs being optional * Revert accidentally pushed changes to nodes_logic.py --- comfy_api/latest/_io.py | 53 ++++++++++++++++++++++++++++++++++------- 1 file changed, 44 insertions(+), 9 deletions(-) diff --git a/comfy_api/latest/_io.py b/comfy_api/latest/_io.py index c30d92aaa..4969d3506 100644 --- a/comfy_api/latest/_io.py +++ b/comfy_api/latest/_io.py @@ -1000,20 +1000,38 @@ class Autogrow(ComfyTypeI): names = [f"{prefix}{i}" for i in range(max)] # need to create a new input based on the contents of input template_input = None - for _, dict_input in input.items(): - # for now, get just the first value from dict_input + template_required = True + for _input_type, dict_input in input.items(): + # for now, get just the first value from dict_input; if not required, min can be ignored + if len(dict_input) == 0: + continue template_input = list(dict_input.values())[0] + template_required = _input_type == "required" + break + if template_input is None: + raise Exception("template_input could not be determined from required or optional; this should never happen.") new_dict = {} + new_dict_added_to = False + # first, add possible inputs into out_dict for i, name in enumerate(names): expected_id = finalize_prefix(curr_prefix, name) + # required + if i < min and template_required: + out_dict["required"][expected_id] = template_input + type_dict = new_dict.setdefault("required", {}) + # optional + else: + out_dict["optional"][expected_id] = template_input + type_dict = new_dict.setdefault("optional", {}) if expected_id in live_inputs: - # required - if i < min: - type_dict = new_dict.setdefault("required", {}) - # optional - else: - type_dict = new_dict.setdefault("optional", {}) + # NOTE: prefix gets added in parse_class_inputs type_dict[name] = template_input + new_dict_added_to = True + # account for the edge case that all inputs are optional and no values are received + if not new_dict_added_to: + finalized_prefix = finalize_prefix(curr_prefix) + out_dict["dynamic_paths"][finalized_prefix] = finalized_prefix + out_dict["dynamic_paths_default_value"][finalized_prefix] = DynamicPathsDefaultValue.EMPTY_DICT parse_class_inputs(out_dict, live_inputs, new_dict, curr_prefix) @comfytype(io_type="COMFY_DYNAMICCOMBO_V3") @@ -1151,6 +1169,8 @@ class V3Data(TypedDict): 'Dictionary where the keys are the hidden input ids and the values are the values of the hidden inputs.' dynamic_paths: dict[str, Any] 'Dictionary where the keys are the input ids and the values dictate how to turn the inputs into a nested dictionary.' + dynamic_paths_default_value: dict[str, Any] + 'Dictionary where the keys are the input ids and the values are a string from DynamicPathsDefaultValue for the inputs if value is None.' create_dynamic_tuple: bool 'When True, the value of the dynamic input will be in the format (value, path_key).' @@ -1504,6 +1524,7 @@ def get_finalized_class_inputs(d: dict[str, Any], live_inputs: dict[str, Any], i "required": {}, "optional": {}, "dynamic_paths": {}, + "dynamic_paths_default_value": {}, } d = d.copy() # ignore hidden for parsing @@ -1513,8 +1534,12 @@ def get_finalized_class_inputs(d: dict[str, Any], live_inputs: dict[str, Any], i out_dict["hidden"] = hidden v3_data = {} dynamic_paths = out_dict.pop("dynamic_paths", None) - if dynamic_paths is not None: + if dynamic_paths is not None and len(dynamic_paths) > 0: v3_data["dynamic_paths"] = dynamic_paths + # this list is used for autogrow, in the case all inputs are optional and no values are passed + dynamic_paths_default_value = out_dict.pop("dynamic_paths_default_value", None) + if dynamic_paths_default_value is not None and len(dynamic_paths_default_value) > 0: + v3_data["dynamic_paths_default_value"] = dynamic_paths_default_value return out_dict, hidden, v3_data def parse_class_inputs(out_dict: dict[str, Any], live_inputs: dict[str, Any], curr_dict: dict[str, Any], curr_prefix: list[str] | None=None) -> None: @@ -1551,11 +1576,16 @@ def add_to_dict_v1(i: Input, d: dict): def add_to_dict_v3(io: Input | Output, d: dict): d[io.id] = (io.get_io_type(), io.as_dict()) +class DynamicPathsDefaultValue: + EMPTY_DICT = "empty_dict" + def build_nested_inputs(values: dict[str, Any], v3_data: V3Data): paths = v3_data.get("dynamic_paths", None) + default_value_dict = v3_data.get("dynamic_paths_default_value", {}) if paths is None: return values values = values.copy() + result = {} create_tuple = v3_data.get("create_dynamic_tuple", False) @@ -1569,6 +1599,11 @@ def build_nested_inputs(values: dict[str, Any], v3_data: V3Data): if is_last: value = values.pop(key, None) + if value is None: + # see if a default value was provided for this key + default_option = default_value_dict.get(key, None) + if default_option == DynamicPathsDefaultValue.EMPTY_DICT: + value = {} if create_tuple: value = (value, key) current[p] = value From e0eacb06883c1f7ddf8af249cd461d7c2ebcbaae Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 19 Jan 2026 19:00:36 -0800 Subject: [PATCH 476/647] Simpler way to implement the #11980 loras. (#11981) --- comfy/utils.py | 2 ++ 1 file changed, 2 insertions(+) diff --git a/comfy/utils.py b/comfy/utils.py index 2e33a4258..5e79fb449 100644 --- a/comfy/utils.py +++ b/comfy/utils.py @@ -639,6 +639,8 @@ def flux_to_diffusers(mmdit_config, output_prefix=""): "proj_out.bias": "linear2.bias", "attn.norm_q.weight": "norm.query_norm.scale", "attn.norm_k.weight": "norm.key_norm.scale", + "attn.to_qkv_mlp_proj.weight": "linear1.weight", # Flux 2 + "attn.to_out.weight": "linear2.weight", # Flux 2 } for k in block_map: From 0da5a0fe58ae940726a61b94698e303fb39d73c1 Mon Sep 17 00:00:00 2001 From: rkfg Date: Tue, 20 Jan 2026 06:12:02 +0300 Subject: [PATCH 477/647] Convert mono audio to fake stereo for LTXV VAE encoding (#11965) --- comfy/ldm/lightricks/vae/audio_vae.py | 9 ++++++--- 1 file changed, 6 insertions(+), 3 deletions(-) diff --git a/comfy/ldm/lightricks/vae/audio_vae.py b/comfy/ldm/lightricks/vae/audio_vae.py index a9111d3bd..29d9e6c29 100644 --- a/comfy/ldm/lightricks/vae/audio_vae.py +++ b/comfy/ldm/lightricks/vae/audio_vae.py @@ -189,9 +189,12 @@ class AudioVAE(torch.nn.Module): waveform = self.device_manager.move_to_load_device(waveform) expected_channels = self.autoencoder.encoder.in_channels if waveform.shape[1] != expected_channels: - raise ValueError( - f"Input audio must have {expected_channels} channels, got {waveform.shape[1]}" - ) + if waveform.shape[1] == 1: + waveform = waveform.expand(-1, expected_channels, *waveform.shape[2:]) + else: + raise ValueError( + f"Input audio must have {expected_channels} channels, got {waveform.shape[1]}" + ) mel_spec = self.preprocessor.waveform_to_mel( waveform, waveform_sample_rate, device=self.device_manager.load_device From 70c91b8248e08492cf16bfebdc83579b801a6ee0 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 19 Jan 2026 19:32:40 -0800 Subject: [PATCH 478/647] Fix #11963 (#11982) --- comfy/text_encoders/ovis.py | 1 + comfy/text_encoders/z_image.py | 1 + 2 files changed, 2 insertions(+) diff --git a/comfy/text_encoders/ovis.py b/comfy/text_encoders/ovis.py index 5754424d2..2cc0867c3 100644 --- a/comfy/text_encoders/ovis.py +++ b/comfy/text_encoders/ovis.py @@ -61,6 +61,7 @@ def te(dtype_llama=None, llama_quantization_metadata=None): if dtype_llama is not None: dtype = dtype_llama if llama_quantization_metadata is not None: + model_options = model_options.copy() model_options["quantization_metadata"] = llama_quantization_metadata super().__init__(device=device, dtype=dtype, model_options=model_options) return OvisTEModel_ diff --git a/comfy/text_encoders/z_image.py b/comfy/text_encoders/z_image.py index 19adde0b7..ad41bfb1e 100644 --- a/comfy/text_encoders/z_image.py +++ b/comfy/text_encoders/z_image.py @@ -40,6 +40,7 @@ def te(dtype_llama=None, llama_quantization_metadata=None): if dtype_llama is not None: dtype = dtype_llama if llama_quantization_metadata is not None: + model_options = model_options.copy() model_options["quantization_metadata"] = llama_quantization_metadata super().__init__(device=device, dtype=dtype, model_options=model_options) return ZImageTEModel_ From 9d273d3ab1fb1d2c8b34de4d54cabe50a5a3e5bc Mon Sep 17 00:00:00 2001 From: comfyanonymous Date: Mon, 19 Jan 2026 22:40:18 -0500 Subject: [PATCH 479/647] ComfyUI v0.10.0 --- comfyui_version.py | 2 +- pyproject.toml | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/comfyui_version.py b/comfyui_version.py index dbb57b4e5..952d413db 100644 --- a/comfyui_version.py +++ b/comfyui_version.py @@ -1,3 +1,3 @@ # This file is automatically generated by the build process when version is # updated in pyproject.toml. -__version__ = "0.9.2" +__version__ = "0.10.0" diff --git a/pyproject.toml b/pyproject.toml index 9ea73da05..120b6c751 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -1,6 +1,6 @@ [project] name = "ComfyUI" -version = "0.9.2" +version = "0.10.0" readme = "README.md" license = { file = "LICENSE" } requires-python = ">=3.10" From 2108167f9f70cfd4874945b31a916680f959a6d7 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 19 Jan 2026 20:17:38 -0800 Subject: [PATCH 480/647] Support zimage omni base model. (#11979) --- comfy/ldm/lumina/model.py | 317 ++++++++++++++++++++++++++++------- comfy/model_base.py | 30 ++++ comfy/model_detection.py | 3 + comfy_extras/nodes_zimage.py | 88 ++++++++++ nodes.py | 1 + 5 files changed, 381 insertions(+), 58 deletions(-) create mode 100644 comfy_extras/nodes_zimage.py diff --git a/comfy/ldm/lumina/model.py b/comfy/ldm/lumina/model.py index afbab2ac7..139f879a1 100644 --- a/comfy/ldm/lumina/model.py +++ b/comfy/ldm/lumina/model.py @@ -13,10 +13,53 @@ from comfy.ldm.modules.attention import optimized_attention_masked from comfy.ldm.flux.layers import EmbedND from comfy.ldm.flux.math import apply_rope import comfy.patcher_extension +import comfy.utils -def modulate(x, scale): - return x * (1 + scale.unsqueeze(1)) +def invert_slices(slices, length): + sorted_slices = sorted(slices) + result = [] + current = 0 + + for start, end in sorted_slices: + if current < start: + result.append((current, start)) + current = max(current, end) + + if current < length: + result.append((current, length)) + + return result + + +def modulate(x, scale, timestep_zero_index=None): + if timestep_zero_index is None: + return x * (1 + scale.unsqueeze(1)) + else: + scale = (1 + scale.unsqueeze(1)) + actual_batch = scale.size(0) // 2 + slices = timestep_zero_index + invert = invert_slices(timestep_zero_index, x.shape[1]) + for s in slices: + x[:, s[0]:s[1]] *= scale[actual_batch:] + for s in invert: + x[:, s[0]:s[1]] *= scale[:actual_batch] + return x + + +def apply_gate(gate, x, timestep_zero_index=None): + if timestep_zero_index is None: + return gate * x + else: + actual_batch = gate.size(0) // 2 + + slices = timestep_zero_index + invert = invert_slices(timestep_zero_index, x.shape[1]) + for s in slices: + x[:, s[0]:s[1]] *= gate[actual_batch:] + for s in invert: + x[:, s[0]:s[1]] *= gate[:actual_batch] + return x ############################################################################# # Core NextDiT Model # @@ -258,6 +301,7 @@ class JointTransformerBlock(nn.Module): x_mask: torch.Tensor, freqs_cis: torch.Tensor, adaln_input: Optional[torch.Tensor]=None, + timestep_zero_index=None, transformer_options={}, ): """ @@ -276,18 +320,18 @@ class JointTransformerBlock(nn.Module): assert adaln_input is not None scale_msa, gate_msa, scale_mlp, gate_mlp = self.adaLN_modulation(adaln_input).chunk(4, dim=1) - x = x + gate_msa.unsqueeze(1).tanh() * self.attention_norm2( + x = x + apply_gate(gate_msa.unsqueeze(1).tanh(), self.attention_norm2( clamp_fp16(self.attention( - modulate(self.attention_norm1(x), scale_msa), + modulate(self.attention_norm1(x), scale_msa, timestep_zero_index=timestep_zero_index), x_mask, freqs_cis, transformer_options=transformer_options, - )) + ))), timestep_zero_index=timestep_zero_index ) - x = x + gate_mlp.unsqueeze(1).tanh() * self.ffn_norm2( + x = x + apply_gate(gate_mlp.unsqueeze(1).tanh(), self.ffn_norm2( clamp_fp16(self.feed_forward( - modulate(self.ffn_norm1(x), scale_mlp), - )) + modulate(self.ffn_norm1(x), scale_mlp, timestep_zero_index=timestep_zero_index), + ))), timestep_zero_index=timestep_zero_index ) else: assert adaln_input is None @@ -345,13 +389,37 @@ class FinalLayer(nn.Module): ), ) - def forward(self, x, c): + def forward(self, x, c, timestep_zero_index=None): scale = self.adaLN_modulation(c) - x = modulate(self.norm_final(x), scale) + x = modulate(self.norm_final(x), scale, timestep_zero_index=timestep_zero_index) x = self.linear(x) return x +def pad_zimage(feats, pad_token, pad_tokens_multiple): + pad_extra = (-feats.shape[1]) % pad_tokens_multiple + return torch.cat((feats, pad_token.to(device=feats.device, dtype=feats.dtype, copy=True).unsqueeze(0).repeat(feats.shape[0], pad_extra, 1)), dim=1), pad_extra + + +def pos_ids_x(start_t, H_tokens, W_tokens, batch_size, device, transformer_options={}): + rope_options = transformer_options.get("rope_options", None) + h_scale = 1.0 + w_scale = 1.0 + h_start = 0 + w_start = 0 + if rope_options is not None: + h_scale = rope_options.get("scale_y", 1.0) + w_scale = rope_options.get("scale_x", 1.0) + + h_start = rope_options.get("shift_y", 0.0) + w_start = rope_options.get("shift_x", 0.0) + x_pos_ids = torch.zeros((batch_size, H_tokens * W_tokens, 3), dtype=torch.float32, device=device) + x_pos_ids[:, :, 0] = start_t + x_pos_ids[:, :, 1] = (torch.arange(H_tokens, dtype=torch.float32, device=device) * h_scale + h_start).view(-1, 1).repeat(1, W_tokens).flatten() + x_pos_ids[:, :, 2] = (torch.arange(W_tokens, dtype=torch.float32, device=device) * w_scale + w_start).view(1, -1).repeat(H_tokens, 1).flatten() + return x_pos_ids + + class NextDiT(nn.Module): """ Diffusion model with a Transformer backbone. @@ -378,6 +446,7 @@ class NextDiT(nn.Module): time_scale=1.0, pad_tokens_multiple=None, clip_text_dim=None, + siglip_feat_dim=None, image_model=None, device=None, dtype=None, @@ -491,6 +560,41 @@ class NextDiT(nn.Module): for layer_id in range(n_layers) ] ) + + if siglip_feat_dim is not None: + self.siglip_embedder = nn.Sequential( + operation_settings.get("operations").RMSNorm(siglip_feat_dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")), + operation_settings.get("operations").Linear( + siglip_feat_dim, + dim, + bias=True, + device=operation_settings.get("device"), + dtype=operation_settings.get("dtype"), + ), + ) + self.siglip_refiner = nn.ModuleList( + [ + JointTransformerBlock( + layer_id, + dim, + n_heads, + n_kv_heads, + multiple_of, + ffn_dim_multiplier, + norm_eps, + qk_norm, + modulation=False, + operation_settings=operation_settings, + ) + for layer_id in range(n_refiner_layers) + ] + ) + self.siglip_pad_token = nn.Parameter(torch.empty((1, dim), device=device, dtype=dtype)) + else: + self.siglip_embedder = None + self.siglip_refiner = None + self.siglip_pad_token = None + # This norm final is in the lumina 2.0 code but isn't actually used for anything. # self.norm_final = operation_settings.get("operations").RMSNorm(dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) self.final_layer = FinalLayer(dim, patch_size, self.out_channels, z_image_modulation=z_image_modulation, operation_settings=operation_settings) @@ -531,70 +635,166 @@ class NextDiT(nn.Module): imgs = torch.stack(imgs, dim=0) return imgs - def patchify_and_embed( - self, x: List[torch.Tensor] | torch.Tensor, cap_feats: torch.Tensor, cap_mask: torch.Tensor, t: torch.Tensor, num_tokens, transformer_options={} - ) -> Tuple[torch.Tensor, torch.Tensor, List[Tuple[int, int]], List[int], torch.Tensor]: - bsz = len(x) - pH = pW = self.patch_size - device = x[0].device - orig_x = x - - if self.pad_tokens_multiple is not None: - pad_extra = (-cap_feats.shape[1]) % self.pad_tokens_multiple - cap_feats = torch.cat((cap_feats, self.cap_pad_token.to(device=cap_feats.device, dtype=cap_feats.dtype, copy=True).unsqueeze(0).repeat(cap_feats.shape[0], pad_extra, 1)), dim=1) + def embed_cap(self, cap_feats=None, offset=0, bsz=1, device=None, dtype=None): + if cap_feats is not None: + cap_feats = self.cap_embedder(cap_feats) + cap_feats_len = cap_feats.shape[1] + if self.pad_tokens_multiple is not None: + cap_feats, _ = pad_zimage(cap_feats, self.cap_pad_token, self.pad_tokens_multiple) + else: + cap_feats_len = 0 + cap_feats = self.cap_pad_token.to(device=device, dtype=dtype, copy=True).unsqueeze(0).repeat(bsz, self.pad_tokens_multiple, 1) cap_pos_ids = torch.zeros(bsz, cap_feats.shape[1], 3, dtype=torch.float32, device=device) - cap_pos_ids[:, :, 0] = torch.arange(cap_feats.shape[1], dtype=torch.float32, device=device) + 1.0 + cap_pos_ids[:, :, 0] = torch.arange(cap_feats.shape[1], dtype=torch.float32, device=device) + 1.0 + offset + embeds = (cap_feats,) + freqs_cis = (self.rope_embedder(cap_pos_ids).movedim(1, 2),) + return embeds, freqs_cis, cap_feats_len + + def embed_all(self, x, cap_feats=None, siglip_feats=None, offset=0, omni=False, transformer_options={}): + bsz = 1 + pH = pW = self.patch_size + device = x.device + embeds, freqs_cis, cap_feats_len = self.embed_cap(cap_feats, offset=offset, bsz=bsz, device=device, dtype=x.dtype) + + if not omni: + cap_feats_len = embeds[0].shape[1] + offset + embeds += (None,) + freqs_cis += (None,) + else: + cap_feats_len += offset + if siglip_feats is not None: + b, h, w, c = siglip_feats.shape + siglip_feats = siglip_feats.permute(0, 3, 1, 2).reshape(b, h * w, c) + siglip_feats = self.siglip_embedder(siglip_feats) + siglip_pos_ids = torch.zeros((bsz, siglip_feats.shape[1], 3), dtype=torch.float32, device=device) + siglip_pos_ids[:, :, 0] = cap_feats_len + 2 + siglip_pos_ids[:, :, 1] = (torch.linspace(0, h * 8 - 1, steps=h, dtype=torch.float32, device=device).floor()).view(-1, 1).repeat(1, w).flatten() + siglip_pos_ids[:, :, 2] = (torch.linspace(0, w * 8 - 1, steps=w, dtype=torch.float32, device=device).floor()).view(1, -1).repeat(h, 1).flatten() + if self.siglip_pad_token is not None: + siglip_feats, pad_extra = pad_zimage(siglip_feats, self.siglip_pad_token, self.pad_tokens_multiple) # TODO: double check + siglip_pos_ids = torch.nn.functional.pad(siglip_pos_ids, (0, 0, 0, pad_extra)) + else: + siglip_feats = self.siglip_pad_token.to(device=device, dtype=x.dtype, copy=True).unsqueeze(0).repeat(bsz, self.pad_tokens_multiple, 1) + siglip_pos_ids = torch.zeros((bsz, siglip_feats.shape[1], 3), dtype=torch.float32, device=device) + + if siglip_feats is None: + embeds += (None,) + freqs_cis += (None,) + else: + embeds += (siglip_feats,) + freqs_cis += (self.rope_embedder(siglip_pos_ids).movedim(1, 2),) B, C, H, W = x.shape x = self.x_embedder(x.view(B, C, H // pH, pH, W // pW, pW).permute(0, 2, 4, 3, 5, 1).flatten(3).flatten(1, 2)) - - rope_options = transformer_options.get("rope_options", None) - h_scale = 1.0 - w_scale = 1.0 - h_start = 0 - w_start = 0 - if rope_options is not None: - h_scale = rope_options.get("scale_y", 1.0) - w_scale = rope_options.get("scale_x", 1.0) - - h_start = rope_options.get("shift_y", 0.0) - w_start = rope_options.get("shift_x", 0.0) - - H_tokens, W_tokens = H // pH, W // pW - x_pos_ids = torch.zeros((bsz, x.shape[1], 3), dtype=torch.float32, device=device) - x_pos_ids[:, :, 0] = cap_feats.shape[1] + 1 - x_pos_ids[:, :, 1] = (torch.arange(H_tokens, dtype=torch.float32, device=device) * h_scale + h_start).view(-1, 1).repeat(1, W_tokens).flatten() - x_pos_ids[:, :, 2] = (torch.arange(W_tokens, dtype=torch.float32, device=device) * w_scale + w_start).view(1, -1).repeat(H_tokens, 1).flatten() - + x_pos_ids = pos_ids_x(cap_feats_len + 1, H // pH, W // pW, bsz, device, transformer_options=transformer_options) if self.pad_tokens_multiple is not None: - pad_extra = (-x.shape[1]) % self.pad_tokens_multiple - x = torch.cat((x, self.x_pad_token.to(device=x.device, dtype=x.dtype, copy=True).unsqueeze(0).repeat(x.shape[0], pad_extra, 1)), dim=1) + x, pad_extra = pad_zimage(x, self.x_pad_token, self.pad_tokens_multiple) x_pos_ids = torch.nn.functional.pad(x_pos_ids, (0, 0, 0, pad_extra)) - freqs_cis = self.rope_embedder(torch.cat((cap_pos_ids, x_pos_ids), dim=1)).movedim(1, 2) + embeds += (x,) + freqs_cis += (self.rope_embedder(x_pos_ids).movedim(1, 2),) + return embeds, freqs_cis, cap_feats_len + len(freqs_cis) - 1 + + + def patchify_and_embed( + self, x: torch.Tensor, cap_feats: torch.Tensor, cap_mask: torch.Tensor, t: torch.Tensor, num_tokens, ref_latents=[], ref_contexts=[], siglip_feats=[], transformer_options={} + ) -> Tuple[torch.Tensor, torch.Tensor, List[Tuple[int, int]], List[int], torch.Tensor]: + bsz = x.shape[0] + cap_mask = None # TODO? + main_siglip = None + orig_x = x + + embeds = ([], [], []) + freqs_cis = ([], [], []) + leftover_cap = [] + + start_t = 0 + omni = len(ref_latents) > 0 + if omni: + for i, ref in enumerate(ref_latents): + if i < len(ref_contexts): + ref_con = ref_contexts[i] + else: + ref_con = None + if i < len(siglip_feats): + sig_feat = siglip_feats[i] + else: + sig_feat = None + + out = self.embed_all(ref, ref_con, sig_feat, offset=start_t, omni=omni, transformer_options=transformer_options) + for i, e in enumerate(out[0]): + embeds[i].append(comfy.utils.repeat_to_batch_size(e, bsz)) + freqs_cis[i].append(out[1][i]) + start_t = out[2] + leftover_cap = ref_contexts[len(ref_latents):] + + H, W = x.shape[-2], x.shape[-1] + img_sizes = [(H, W)] * bsz + out = self.embed_all(x, cap_feats, main_siglip, offset=start_t, omni=omni, transformer_options=transformer_options) + img_len = out[0][-1].shape[1] + cap_len = out[0][0].shape[1] + for i, e in enumerate(out[0]): + if e is not None: + e = comfy.utils.repeat_to_batch_size(e, bsz) + embeds[i].append(e) + freqs_cis[i].append(out[1][i]) + start_t = out[2] + + for cap in leftover_cap: + out = self.embed_cap(cap, offset=start_t, bsz=bsz, device=x.device, dtype=x.dtype) + cap_len += out[0][0].shape[1] + embeds[0].append(comfy.utils.repeat_to_batch_size(out[0][0], bsz)) + freqs_cis[0].append(out[1][0]) + start_t += out[2] patches = transformer_options.get("patches", {}) # refine context + cap_feats = torch.cat(embeds[0], dim=1) + cap_freqs_cis = torch.cat(freqs_cis[0], dim=1) for layer in self.context_refiner: - cap_feats = layer(cap_feats, cap_mask, freqs_cis[:, :cap_pos_ids.shape[1]], transformer_options=transformer_options) + cap_feats = layer(cap_feats, cap_mask, cap_freqs_cis, transformer_options=transformer_options) + + feats = (cap_feats,) + fc = (cap_freqs_cis,) + + if omni: + siglip_mask = None + siglip_feats_combined = torch.cat(embeds[1], dim=1) + siglip_feats_freqs_cis = torch.cat(freqs_cis[1], dim=1) + if self.siglip_refiner is not None: + for layer in self.siglip_refiner: + siglip_feats_combined = layer(siglip_feats_combined, siglip_mask, siglip_feats_freqs_cis, transformer_options=transformer_options) + feats += (siglip_feats_combined,) + fc += (siglip_feats_freqs_cis,) padded_img_mask = None + x = torch.cat(embeds[-1], dim=1) + fc_x = torch.cat(freqs_cis[-1], dim=1) + if omni: + timestep_zero_index = [(x.shape[1] - img_len, x.shape[1])] + else: + timestep_zero_index = None + x_input = x for i, layer in enumerate(self.noise_refiner): - x = layer(x, padded_img_mask, freqs_cis[:, cap_pos_ids.shape[1]:], t, transformer_options=transformer_options) + x = layer(x, padded_img_mask, fc_x, t, timestep_zero_index=timestep_zero_index, transformer_options=transformer_options) if "noise_refiner" in patches: for p in patches["noise_refiner"]: - out = p({"img": x, "img_input": x_input, "txt": cap_feats, "pe": freqs_cis[:, cap_pos_ids.shape[1]:], "vec": t, "x": orig_x, "block_index": i, "transformer_options": transformer_options, "block_type": "noise_refiner"}) + out = p({"img": x, "img_input": x_input, "txt": cap_feats, "pe": fc_x, "vec": t, "x": orig_x, "block_index": i, "transformer_options": transformer_options, "block_type": "noise_refiner"}) if "img" in out: x = out["img"] - padded_full_embed = torch.cat((cap_feats, x), dim=1) + padded_full_embed = torch.cat(feats + (x,), dim=1) + if timestep_zero_index is not None: + ind = padded_full_embed.shape[1] - x.shape[1] + timestep_zero_index = [(ind + x.shape[1] - img_len, ind + x.shape[1])] + timestep_zero_index.append((feats[0].shape[1] - cap_len, feats[0].shape[1])) + mask = None - img_sizes = [(H, W)] * bsz - l_effective_cap_len = [cap_feats.shape[1]] * bsz - return padded_full_embed, mask, img_sizes, l_effective_cap_len, freqs_cis + l_effective_cap_len = [padded_full_embed.shape[1] - img_len] * bsz + return padded_full_embed, mask, img_sizes, l_effective_cap_len, torch.cat(fc + (fc_x,), dim=1), timestep_zero_index def forward(self, x, timesteps, context, num_tokens, attention_mask=None, **kwargs): return comfy.patcher_extension.WrapperExecutor.new_class_executor( @@ -604,7 +804,11 @@ class NextDiT(nn.Module): ).execute(x, timesteps, context, num_tokens, attention_mask, **kwargs) # def forward(self, x, t, cap_feats, cap_mask): - def _forward(self, x, timesteps, context, num_tokens, attention_mask=None, transformer_options={}, **kwargs): + def _forward(self, x, timesteps, context, num_tokens, attention_mask=None, ref_latents=[], ref_contexts=[], siglip_feats=[], transformer_options={}, **kwargs): + omni = len(ref_latents) > 0 + if omni: + timesteps = torch.cat([timesteps * 0, timesteps], dim=0) + t = 1.0 - timesteps cap_feats = context cap_mask = attention_mask @@ -619,8 +823,6 @@ class NextDiT(nn.Module): t = self.t_embedder(t * self.time_scale, dtype=x.dtype) # (N, D) adaln_input = t - cap_feats = self.cap_embedder(cap_feats) # (N, L, D) # todo check if able to batchify w.o. redundant compute - if self.clip_text_pooled_proj is not None: pooled = kwargs.get("clip_text_pooled", None) if pooled is not None: @@ -632,7 +834,7 @@ class NextDiT(nn.Module): patches = transformer_options.get("patches", {}) x_is_tensor = isinstance(x, torch.Tensor) - img, mask, img_size, cap_size, freqs_cis = self.patchify_and_embed(x, cap_feats, cap_mask, adaln_input, num_tokens, transformer_options=transformer_options) + img, mask, img_size, cap_size, freqs_cis, timestep_zero_index = self.patchify_and_embed(x, cap_feats, cap_mask, adaln_input, num_tokens, ref_latents=ref_latents, ref_contexts=ref_contexts, siglip_feats=siglip_feats, transformer_options=transformer_options) freqs_cis = freqs_cis.to(img.device) transformer_options["total_blocks"] = len(self.layers) @@ -640,7 +842,7 @@ class NextDiT(nn.Module): img_input = img for i, layer in enumerate(self.layers): transformer_options["block_index"] = i - img = layer(img, mask, freqs_cis, adaln_input, transformer_options=transformer_options) + img = layer(img, mask, freqs_cis, adaln_input, timestep_zero_index=timestep_zero_index, transformer_options=transformer_options) if "double_block" in patches: for p in patches["double_block"]: out = p({"img": img[:, cap_size[0]:], "img_input": img_input[:, cap_size[0]:], "txt": img[:, :cap_size[0]], "pe": freqs_cis[:, cap_size[0]:], "vec": adaln_input, "x": x, "block_index": i, "transformer_options": transformer_options}) @@ -649,8 +851,7 @@ class NextDiT(nn.Module): if "txt" in out: img[:, :cap_size[0]] = out["txt"] - img = self.final_layer(img, adaln_input) + img = self.final_layer(img, adaln_input, timestep_zero_index=timestep_zero_index) img = self.unpatchify(img, img_size, cap_size, return_tensor=x_is_tensor)[:, :, :h, :w] - return -img diff --git a/comfy/model_base.py b/comfy/model_base.py index 49efd700b..28ba2643e 100644 --- a/comfy/model_base.py +++ b/comfy/model_base.py @@ -1150,6 +1150,7 @@ class CosmosPredict2(BaseModel): class Lumina2(BaseModel): def __init__(self, model_config, model_type=ModelType.FLOW, device=None): super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.lumina.model.NextDiT) + self.memory_usage_factor_conds = ("ref_latents",) def extra_conds(self, **kwargs): out = super().extra_conds(**kwargs) @@ -1169,6 +1170,35 @@ class Lumina2(BaseModel): if clip_text_pooled is not None: out['clip_text_pooled'] = comfy.conds.CONDRegular(clip_text_pooled) + clip_vision_outputs = kwargs.get("clip_vision_outputs", list(map(lambda a: a.get("clip_vision_output"), kwargs.get("unclip_conditioning", [{}])))) # Z Image omni + if clip_vision_outputs is not None and len(clip_vision_outputs) > 0: + sigfeats = [] + for clip_vision_output in clip_vision_outputs: + if clip_vision_output is not None: + image_size = clip_vision_output.image_sizes[0] + shape = clip_vision_output.last_hidden_state.shape + sigfeats.append(clip_vision_output.last_hidden_state.reshape(shape[0], image_size[1] // 16, image_size[2] // 16, shape[-1])) + if len(sigfeats) > 0: + out['siglip_feats'] = comfy.conds.CONDList(sigfeats) + + ref_latents = kwargs.get("reference_latents", None) + if ref_latents is not None: + latents = [] + for lat in ref_latents: + latents.append(self.process_latent_in(lat)) + out['ref_latents'] = comfy.conds.CONDList(latents) + + ref_contexts = kwargs.get("reference_latents_text_embeds", None) + if ref_contexts is not None: + out['ref_contexts'] = comfy.conds.CONDList(ref_contexts) + + return out + + def extra_conds_shapes(self, **kwargs): + out = {} + ref_latents = kwargs.get("reference_latents", None) + if ref_latents is not None: + out['ref_latents'] = list([1, 16, sum(map(lambda a: math.prod(a.size()[2:]), ref_latents))]) return out class WAN21(BaseModel): diff --git a/comfy/model_detection.py b/comfy/model_detection.py index aff5a50b9..42884f797 100644 --- a/comfy/model_detection.py +++ b/comfy/model_detection.py @@ -446,6 +446,9 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): dit_config["time_scale"] = 1000.0 if '{}cap_pad_token'.format(key_prefix) in state_dict_keys: dit_config["pad_tokens_multiple"] = 32 + sig_weight = state_dict.get('{}siglip_embedder.0.weight'.format(key_prefix), None) + if sig_weight is not None: + dit_config["siglip_feat_dim"] = sig_weight.shape[0] return dit_config diff --git a/comfy_extras/nodes_zimage.py b/comfy_extras/nodes_zimage.py new file mode 100644 index 000000000..2ee3c43b1 --- /dev/null +++ b/comfy_extras/nodes_zimage.py @@ -0,0 +1,88 @@ +import node_helpers +from typing_extensions import override +from comfy_api.latest import ComfyExtension, io +import math +import comfy.utils + + +class TextEncodeZImageOmni(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="TextEncodeZImageOmni", + category="advanced/conditioning", + is_experimental=True, + inputs=[ + io.Clip.Input("clip"), + io.ClipVision.Input("image_encoder", optional=True), + io.String.Input("prompt", multiline=True, dynamic_prompts=True), + io.Boolean.Input("auto_resize_images", default=True), + io.Vae.Input("vae", optional=True), + io.Image.Input("image1", optional=True), + io.Image.Input("image2", optional=True), + io.Image.Input("image3", optional=True), + ], + outputs=[ + io.Conditioning.Output(), + ], + ) + + @classmethod + def execute(cls, clip, prompt, image_encoder=None, auto_resize_images=True, vae=None, image1=None, image2=None, image3=None) -> io.NodeOutput: + ref_latents = [] + images = list(filter(lambda a: a is not None, [image1, image2, image3])) + + prompt_list = [] + template = None + if len(images) > 0: + prompt_list = ["<|im_start|>user\n<|vision_start|>"] + prompt_list += ["<|vision_end|><|vision_start|>"] * (len(images) - 1) + prompt_list += ["<|vision_end|><|im_end|>"] + template = "<|vision_end|>{}<|im_end|>\n<|im_start|>assistant\n<|vision_start|>" + + encoded_images = [] + + for i, image in enumerate(images): + if image_encoder is not None: + encoded_images.append(image_encoder.encode_image(image)) + + if vae is not None: + if auto_resize_images: + samples = image.movedim(-1, 1) + total = int(1024 * 1024) + scale_by = math.sqrt(total / (samples.shape[3] * samples.shape[2])) + width = round(samples.shape[3] * scale_by / 8.0) * 8 + height = round(samples.shape[2] * scale_by / 8.0) * 8 + + image = comfy.utils.common_upscale(samples, width, height, "area", "disabled").movedim(1, -1) + ref_latents.append(vae.encode(image)) + + tokens = clip.tokenize(prompt, llama_template=template) + conditioning = clip.encode_from_tokens_scheduled(tokens) + + extra_text_embeds = [] + for p in prompt_list: + tokens = clip.tokenize(p, llama_template="{}") + text_embeds = clip.encode_from_tokens_scheduled(tokens) + extra_text_embeds.append(text_embeds[0][0]) + + if len(ref_latents) > 0: + conditioning = node_helpers.conditioning_set_values(conditioning, {"reference_latents": ref_latents}, append=True) + if len(encoded_images) > 0: + conditioning = node_helpers.conditioning_set_values(conditioning, {"clip_vision_outputs": encoded_images}, append=True) + if len(extra_text_embeds) > 0: + conditioning = node_helpers.conditioning_set_values(conditioning, {"reference_latents_text_embeds": extra_text_embeds}, append=True) + + return io.NodeOutput(conditioning) + + +class ZImageExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[io.ComfyNode]]: + return [ + TextEncodeZImageOmni, + ] + + +async def comfy_entrypoint() -> ZImageExtension: + return ZImageExtension() diff --git a/nodes.py b/nodes.py index cba8eacc2..ea5d6e525 100644 --- a/nodes.py +++ b/nodes.py @@ -2373,6 +2373,7 @@ async def init_builtin_extra_nodes(): "nodes_kandinsky5.py", "nodes_wanmove.py", "nodes_image_compare.py", + "nodes_zimage.py", ] import_failed = [] From 0fc3b6e3a6f1d8fdffca3a51cb4d10a06f4e079d Mon Sep 17 00:00:00 2001 From: ComfyUI Wiki Date: Tue, 20 Jan 2026 12:17:56 +0800 Subject: [PATCH 481/647] chore: update workflow templates to v0.8.15 (#11984) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 312c7c137..35543525d 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,5 +1,5 @@ comfyui-frontend-package==1.36.14 -comfyui-workflow-templates==0.8.14 +comfyui-workflow-templates==0.8.15 comfyui-embedded-docs==0.4.0 torch torchsde From 4edb87aa50190139a38a2ccd6b6ee35ba9df4da1 Mon Sep 17 00:00:00 2001 From: Comfy Org PR Bot Date: Tue, 20 Jan 2026 13:57:50 +0900 Subject: [PATCH 482/647] Bump comfyui-frontend-package to 1.37.11 (#11976) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 35543525d..ec89dccd2 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,4 +1,4 @@ -comfyui-frontend-package==1.36.14 +comfyui-frontend-package==1.37.11 comfyui-workflow-templates==0.8.15 comfyui-embedded-docs==0.4.0 torch From 8ccc0c94fa0d8e43fffe7190e6a36551a53df54a Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 19 Jan 2026 21:32:00 -0800 Subject: [PATCH 483/647] Make omni stuff work on regular z image for easier testing. (#11985) --- comfy/ldm/lumina/model.py | 14 ++++++++------ 1 file changed, 8 insertions(+), 6 deletions(-) diff --git a/comfy/ldm/lumina/model.py b/comfy/ldm/lumina/model.py index 139f879a1..b114d9e31 100644 --- a/comfy/ldm/lumina/model.py +++ b/comfy/ldm/lumina/model.py @@ -657,7 +657,7 @@ class NextDiT(nn.Module): device = x.device embeds, freqs_cis, cap_feats_len = self.embed_cap(cap_feats, offset=offset, bsz=bsz, device=device, dtype=x.dtype) - if not omni: + if (not omni) or self.siglip_embedder is None: cap_feats_len = embeds[0].shape[1] + offset embeds += (None,) freqs_cis += (None,) @@ -675,8 +675,9 @@ class NextDiT(nn.Module): siglip_feats, pad_extra = pad_zimage(siglip_feats, self.siglip_pad_token, self.pad_tokens_multiple) # TODO: double check siglip_pos_ids = torch.nn.functional.pad(siglip_pos_ids, (0, 0, 0, pad_extra)) else: - siglip_feats = self.siglip_pad_token.to(device=device, dtype=x.dtype, copy=True).unsqueeze(0).repeat(bsz, self.pad_tokens_multiple, 1) - siglip_pos_ids = torch.zeros((bsz, siglip_feats.shape[1], 3), dtype=torch.float32, device=device) + if self.siglip_pad_token is not None: + siglip_feats = self.siglip_pad_token.to(device=device, dtype=x.dtype, copy=True).unsqueeze(0).repeat(bsz, self.pad_tokens_multiple, 1) + siglip_pos_ids = torch.zeros((bsz, siglip_feats.shape[1], 3), dtype=torch.float32, device=device) if siglip_feats is None: embeds += (None,) @@ -724,8 +725,9 @@ class NextDiT(nn.Module): out = self.embed_all(ref, ref_con, sig_feat, offset=start_t, omni=omni, transformer_options=transformer_options) for i, e in enumerate(out[0]): - embeds[i].append(comfy.utils.repeat_to_batch_size(e, bsz)) - freqs_cis[i].append(out[1][i]) + if e is not None: + embeds[i].append(comfy.utils.repeat_to_batch_size(e, bsz)) + freqs_cis[i].append(out[1][i]) start_t = out[2] leftover_cap = ref_contexts[len(ref_latents):] @@ -759,7 +761,7 @@ class NextDiT(nn.Module): feats = (cap_feats,) fc = (cap_freqs_cis,) - if omni: + if omni and len(embeds[1]) > 0: siglip_mask = None siglip_feats_combined = torch.cat(embeds[1], dim=1) siglip_feats_freqs_cis = torch.cat(freqs_cis[1], dim=1) From ddc541ffdae0fe626de5a33192001f31c6ab93c6 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Tue, 20 Jan 2026 23:05:40 +0200 Subject: [PATCH 484/647] feat(api-nodes): add WaveSpeed nodes (#11945) --- comfy_api_nodes/apis/wavespeed.py | 35 ++++++ comfy_api_nodes/nodes_wavespeed.py | 178 +++++++++++++++++++++++++++++ 2 files changed, 213 insertions(+) create mode 100644 comfy_api_nodes/apis/wavespeed.py create mode 100644 comfy_api_nodes/nodes_wavespeed.py diff --git a/comfy_api_nodes/apis/wavespeed.py b/comfy_api_nodes/apis/wavespeed.py new file mode 100644 index 000000000..07a7bfa5d --- /dev/null +++ b/comfy_api_nodes/apis/wavespeed.py @@ -0,0 +1,35 @@ +from pydantic import BaseModel, Field + + +class SeedVR2ImageRequest(BaseModel): + image: str = Field(...) + target_resolution: str = Field(...) + output_format: str = Field("png") + enable_sync_mode: bool = Field(False) + + +class FlashVSRRequest(BaseModel): + target_resolution: str = Field(...) + video: str = Field(...) + duration: float = Field(...) + + +class TaskCreatedDataResponse(BaseModel): + id: str = Field(...) + + +class TaskCreatedResponse(BaseModel): + code: int = Field(...) + message: str = Field(...) + data: TaskCreatedDataResponse | None = Field(None) + + +class TaskResultDataResponse(BaseModel): + status: str = Field(...) + outputs: list[str] = Field([]) + + +class TaskResultResponse(BaseModel): + code: int = Field(...) + message: str = Field(...) + data: TaskResultDataResponse | None = Field(None) diff --git a/comfy_api_nodes/nodes_wavespeed.py b/comfy_api_nodes/nodes_wavespeed.py new file mode 100644 index 000000000..c59fafd3b --- /dev/null +++ b/comfy_api_nodes/nodes_wavespeed.py @@ -0,0 +1,178 @@ +from typing_extensions import override + +from comfy_api.latest import IO, ComfyExtension, Input +from comfy_api_nodes.apis.wavespeed import ( + FlashVSRRequest, + TaskCreatedResponse, + TaskResultResponse, + SeedVR2ImageRequest, +) +from comfy_api_nodes.util import ( + ApiEndpoint, + download_url_to_video_output, + poll_op, + sync_op, + upload_video_to_comfyapi, + validate_container_format_is_mp4, + validate_video_duration, + upload_images_to_comfyapi, + get_number_of_images, + download_url_to_image_tensor, +) + + +class WavespeedFlashVSRNode(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="WavespeedFlashVSRNode", + display_name="FlashVSR Video Upscale", + category="api node/video/WaveSpeed", + description="Fast, high-quality video upscaler that " + "boosts resolution and restores clarity for low-resolution or blurry footage.", + inputs=[ + IO.Video.Input("video"), + IO.Combo.Input("target_resolution", options=["720p", "1080p", "2K", "4K"]), + ], + outputs=[ + IO.Video.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["target_resolution"]), + expr=""" + ( + $price_for_1sec := {"720p": 0.012, "1080p": 0.018, "2k": 0.024, "4k": 0.032}; + { + "type":"usd", + "usd": $lookup($price_for_1sec, widgets.target_resolution), + "format":{"suffix": "/second", "approximate": true} + } + ) + """, + ), + ) + + @classmethod + async def execute( + cls, + video: Input.Video, + target_resolution: str, + ) -> IO.NodeOutput: + validate_container_format_is_mp4(video) + validate_video_duration(video, min_duration=5, max_duration=60 * 10) + initial_res = await sync_op( + cls, + ApiEndpoint(path="/proxy/wavespeed/api/v3/wavespeed-ai/flashvsr", method="POST"), + response_model=TaskCreatedResponse, + data=FlashVSRRequest( + target_resolution=target_resolution.lower(), + video=await upload_video_to_comfyapi(cls, video), + duration=video.get_duration(), + ), + ) + if initial_res.code != 200: + raise ValueError(f"Task creation fails with code={initial_res.code} and message={initial_res.message}") + final_response = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/wavespeed/api/v3/predictions/{initial_res.data.id}/result"), + response_model=TaskResultResponse, + status_extractor=lambda x: "failed" if x.data is None else x.data.status, + poll_interval=10.0, + max_poll_attempts=480, + ) + if final_response.code != 200: + raise ValueError( + f"Task processing failed with code={final_response.code} and message={final_response.message}" + ) + return IO.NodeOutput(await download_url_to_video_output(final_response.data.outputs[0])) + + +class WavespeedImageUpscaleNode(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="WavespeedImageUpscaleNode", + display_name="WaveSpeed Image Upscale", + category="api node/image/WaveSpeed", + description="Boost image resolution and quality, upscaling photos to 4K or 8K for sharp, detailed results.", + inputs=[ + IO.Combo.Input("model", options=["SeedVR2", "Ultimate"]), + IO.Image.Input("image"), + IO.Combo.Input("target_resolution", options=["2K", "4K", "8K"]), + ], + outputs=[ + IO.Image.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["model"]), + expr=""" + ( + $prices := {"seedvr2": 0.01, "ultimate": 0.06}; + {"type":"usd", "usd": $lookup($prices, widgets.model)} + ) + """, + ), + ) + + @classmethod + async def execute( + cls, + model: str, + image: Input.Image, + target_resolution: str, + ) -> IO.NodeOutput: + if get_number_of_images(image) != 1: + raise ValueError("Exactly one input image is required.") + if model == "SeedVR2": + model_path = "seedvr2/image" + else: + model_path = "ultimate-image-upscaler" + initial_res = await sync_op( + cls, + ApiEndpoint(path=f"/proxy/wavespeed/api/v3/wavespeed-ai/{model_path}", method="POST"), + response_model=TaskCreatedResponse, + data=SeedVR2ImageRequest( + target_resolution=target_resolution.lower(), + image=(await upload_images_to_comfyapi(cls, image, max_images=1))[0], + ), + ) + if initial_res.code != 200: + raise ValueError(f"Task creation fails with code={initial_res.code} and message={initial_res.message}") + final_response = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/wavespeed/api/v3/predictions/{initial_res.data.id}/result"), + response_model=TaskResultResponse, + status_extractor=lambda x: "failed" if x.data is None else x.data.status, + poll_interval=10.0, + max_poll_attempts=480, + ) + if final_response.code != 200: + raise ValueError( + f"Task processing failed with code={final_response.code} and message={final_response.message}" + ) + return IO.NodeOutput(await download_url_to_image_tensor(final_response.data.outputs[0])) + + +class WavespeedExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[IO.ComfyNode]]: + return [ + WavespeedFlashVSRNode, + WavespeedImageUpscaleNode, + ] + + +async def comfy_entrypoint() -> WavespeedExtension: + return WavespeedExtension() From 965d0ed509ce46a3328c342aee23a234ba6e4f88 Mon Sep 17 00:00:00 2001 From: Ivan Zorin Date: Wed, 21 Jan 2026 01:44:28 +0200 Subject: [PATCH 485/647] fix: remove normalization of audio in LTX Mel spectrogram creation (#11990) For LTX Audio VAE, remove normalization of audio during MEL spectrogram creation. This aligs inference with training and prevents loud audio from being attenuated. --- comfy/ldm/lightricks/vae/audio_vae.py | 10 ---------- 1 file changed, 10 deletions(-) diff --git a/comfy/ldm/lightricks/vae/audio_vae.py b/comfy/ldm/lightricks/vae/audio_vae.py index 29d9e6c29..55a074661 100644 --- a/comfy/ldm/lightricks/vae/audio_vae.py +++ b/comfy/ldm/lightricks/vae/audio_vae.py @@ -103,20 +103,10 @@ class AudioPreprocessor: return waveform return torchaudio.functional.resample(waveform, source_rate, self.target_sample_rate) - @staticmethod - def normalize_amplitude( - waveform: torch.Tensor, max_amplitude: float = 0.5, eps: float = 1e-5 - ) -> torch.Tensor: - waveform = waveform - waveform.mean(dim=2, keepdim=True) - peak = torch.max(torch.abs(waveform)) + eps - scale = peak.clamp(max=max_amplitude) / peak - return waveform * scale - def waveform_to_mel( self, waveform: torch.Tensor, waveform_sample_rate: int, device ) -> torch.Tensor: waveform = self.resample(waveform, waveform_sample_rate) - waveform = self.normalize_amplitude(waveform) mel_transform = torchaudio.transforms.MelSpectrogram( sample_rate=self.target_sample_rate, From c4a14df9a35336dbfff096683c5015ce726c269d Mon Sep 17 00:00:00 2001 From: Mylo <36931363+gitmylo@users.noreply.github.com> Date: Wed, 21 Jan 2026 00:46:11 +0100 Subject: [PATCH 486/647] Dynamically detect chroma radiance patch size (#11991) --- comfy/model_detection.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy/model_detection.py b/comfy/model_detection.py index 42884f797..dad206a2f 100644 --- a/comfy/model_detection.py +++ b/comfy/model_detection.py @@ -253,7 +253,7 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): dit_config["image_model"] = "chroma_radiance" dit_config["in_channels"] = 3 dit_config["out_channels"] = 3 - dit_config["patch_size"] = 16 + dit_config["patch_size"] = state_dict.get('{}img_in_patch.weight'.format(key_prefix)).size(dim=-1) dit_config["nerf_hidden_size"] = 64 dit_config["nerf_mlp_ratio"] = 4 dit_config["nerf_depth"] = 4 From e755268e7b7843695f52b87595afcb09c1e9fd87 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 20 Jan 2026 20:08:31 -0800 Subject: [PATCH 487/647] Config for Qwen 3 0.6B model. (#11998) --- comfy/text_encoders/llama.py | 31 +++++++++++++++++++++++++++++++ 1 file changed, 31 insertions(+) diff --git a/comfy/text_encoders/llama.py b/comfy/text_encoders/llama.py index 331a30f61..3080a3e09 100644 --- a/comfy/text_encoders/llama.py +++ b/comfy/text_encoders/llama.py @@ -77,6 +77,28 @@ class Qwen25_3BConfig: rope_scale = None final_norm: bool = True +@dataclass +class Qwen3_06BConfig: + vocab_size: int = 151936 + hidden_size: int = 1024 + intermediate_size: int = 3072 + num_hidden_layers: int = 28 + num_attention_heads: int = 16 + num_key_value_heads: int = 8 + max_position_embeddings: int = 32768 + rms_norm_eps: float = 1e-6 + rope_theta: float = 1000000.0 + transformer_type: str = "llama" + head_dim = 128 + rms_norm_add = False + mlp_activation = "silu" + qkv_bias = False + rope_dims = None + q_norm = "gemma3" + k_norm = "gemma3" + rope_scale = None + final_norm: bool = True + @dataclass class Qwen3_4BConfig: vocab_size: int = 151936 @@ -641,6 +663,15 @@ class Qwen25_3B(BaseLlama, torch.nn.Module): self.model = Llama2_(config, device=device, dtype=dtype, ops=operations) self.dtype = dtype +class Qwen3_06B(BaseLlama, torch.nn.Module): + def __init__(self, config_dict, dtype, device, operations): + super().__init__() + config = Qwen3_06BConfig(**config_dict) + self.num_layers = config.num_hidden_layers + + self.model = Llama2_(config, device=device, dtype=dtype, ops=operations) + self.dtype = dtype + class Qwen3_4B(BaseLlama, torch.nn.Module): def __init__(self, config_dict, dtype, device, operations): super().__init__() From 0fc15700be9b555f351034942b5bd7243bdf6bcc Mon Sep 17 00:00:00 2001 From: Markury Date: Tue, 20 Jan 2026 23:18:33 -0500 Subject: [PATCH 488/647] Add LyCoris LoKr MLP layer support for Flux2 (#11997) --- comfy/utils.py | 8 ++++++++ 1 file changed, 8 insertions(+) diff --git a/comfy/utils.py b/comfy/utils.py index 5e79fb449..d97d753e6 100644 --- a/comfy/utils.py +++ b/comfy/utils.py @@ -611,6 +611,14 @@ def flux_to_diffusers(mmdit_config, output_prefix=""): "ff_context.net.0.proj.bias": "txt_mlp.0.bias", "ff_context.net.2.weight": "txt_mlp.2.weight", "ff_context.net.2.bias": "txt_mlp.2.bias", + "ff.linear_in.weight": "img_mlp.0.weight", # LyCoris LoKr + "ff.linear_in.bias": "img_mlp.0.bias", + "ff.linear_out.weight": "img_mlp.2.weight", + "ff.linear_out.bias": "img_mlp.2.bias", + "ff_context.linear_in.weight": "txt_mlp.0.weight", + "ff_context.linear_in.bias": "txt_mlp.0.bias", + "ff_context.linear_out.weight": "txt_mlp.2.weight", + "ff_context.linear_out.bias": "txt_mlp.2.bias", "attn.norm_q.weight": "img_attn.norm.query_norm.scale", "attn.norm_k.weight": "img_attn.norm.key_norm.scale", "attn.norm_added_q.weight": "txt_attn.norm.query_norm.scale", From 451af7015435df22e6313ae79f25fe2ef336a96d Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Wed, 21 Jan 2026 14:03:45 +0200 Subject: [PATCH 489/647] fix(api-nodes-Vidu): allow passing up to 7 subjects in Vidu Reference node (#12002) --- comfy_api_nodes/nodes_vidu.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/comfy_api_nodes/nodes_vidu.py b/comfy_api_nodes/nodes_vidu.py index 8edb02f39..b9114c4bb 100644 --- a/comfy_api_nodes/nodes_vidu.py +++ b/comfy_api_nodes/nodes_vidu.py @@ -703,7 +703,7 @@ class Vidu2ReferenceVideoNode(IO.ComfyNode): "subjects", template=IO.Autogrow.TemplateNames( IO.Image.Input("reference_images"), - names=["subject1", "subject2", "subject3"], + names=["subject1", "subject2", "subject3", "subject4", "subject5", "subject6", "subject7"], min=1, ), tooltip="For each subject, provide up to 3 reference images (7 images total across all subjects). " @@ -738,7 +738,7 @@ class Vidu2ReferenceVideoNode(IO.ComfyNode): control_after_generate=True, ), IO.Combo.Input("aspect_ratio", options=["16:9", "9:16", "4:3", "3:4", "1:1"]), - IO.Combo.Input("resolution", options=["720p"]), + IO.Combo.Input("resolution", options=["720p", "1080p"]), IO.Combo.Input( "movement_amplitude", options=["auto", "small", "medium", "large"], From bdeac8897e522b9637a6a427fdc8a50a6abd6b20 Mon Sep 17 00:00:00 2001 From: Christian Byrne Date: Wed, 21 Jan 2026 15:36:02 -0800 Subject: [PATCH 490/647] feat: Add search_aliases field to node schema (#12010) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit * feat: Add search_aliases field to node schema Adds `search_aliases` field to improve node discoverability. Users can define alternative search terms for nodes (e.g., "text concat" → StringConcatenate). Changes: - Add `search_aliases: list[str]` to V3 Schema - Add `SEARCH_ALIASES` support for V1 nodes - Include field in `/object_info` response - Add aliases to high-priority core nodes V1 usage: ```python class MyNode: SEARCH_ALIASES = ["alt name", "synonym"] ``` V3 usage: ```python io.Schema( node_id="MyNode", search_aliases=["alt name", "synonym"], ... ) ``` ## Related PRs - Frontend: Comfy-Org/ComfyUI_frontend#XXXX (draft - merge after this) - Docs: Comfy-Org/docs#XXXX (draft - merge after stable) * Propagate search_aliases through V3 Schema.get_v1_info to NodeInfoV1 --- comfy_api/latest/_io.py | 4 ++++ comfy_extras/nodes_post_processing.py | 1 + comfy_extras/nodes_preview_any.py | 1 + comfy_extras/nodes_string.py | 1 + comfy_extras/nodes_upscale_model.py | 1 + nodes.py | 15 +++++++++++++++ server.py | 2 ++ 7 files changed, 25 insertions(+) diff --git a/comfy_api/latest/_io.py b/comfy_api/latest/_io.py index 4969d3506..a60020ca8 100644 --- a/comfy_api/latest/_io.py +++ b/comfy_api/latest/_io.py @@ -1249,6 +1249,7 @@ class NodeInfoV1: experimental: bool=None api_node: bool=None price_badge: dict | None = None + search_aliases: list[str]=None @dataclass class NodeInfoV3: @@ -1346,6 +1347,8 @@ class Schema: hidden: list[Hidden] = field(default_factory=list) description: str="" """Node description, shown as a tooltip when hovering over the node.""" + search_aliases: list[str] = field(default_factory=list) + """Alternative names for search. Useful for synonyms, abbreviations, or old names after renaming.""" is_input_list: bool = False """A flag indicating if this node implements the additional code necessary to deal with OUTPUT_IS_LIST nodes. @@ -1483,6 +1486,7 @@ class Schema: api_node=self.is_api_node, python_module=getattr(cls, "RELATIVE_PYTHON_MODULE", "nodes"), price_badge=self.price_badge.as_dict(self.inputs) if self.price_badge is not None else None, + search_aliases=self.search_aliases if self.search_aliases else None, ) return info diff --git a/comfy_extras/nodes_post_processing.py b/comfy_extras/nodes_post_processing.py index 2e559c35c..6011275d6 100644 --- a/comfy_extras/nodes_post_processing.py +++ b/comfy_extras/nodes_post_processing.py @@ -550,6 +550,7 @@ class BatchImagesNode(io.ComfyNode): node_id="BatchImagesNode", display_name="Batch Images", category="image", + search_aliases=["batch", "image batch", "batch images", "combine images", "merge images", "stack images"], inputs=[ io.Autogrow.Input("images", template=autogrow_template) ], diff --git a/comfy_extras/nodes_preview_any.py b/comfy_extras/nodes_preview_any.py index 139b07c93..91502ebf2 100644 --- a/comfy_extras/nodes_preview_any.py +++ b/comfy_extras/nodes_preview_any.py @@ -16,6 +16,7 @@ class PreviewAny(): OUTPUT_NODE = True CATEGORY = "utils" + SEARCH_ALIASES = ["preview", "show", "display", "view", "show text", "display text", "preview text", "show output", "inspect", "debug"] def main(self, source=None): value = 'None' diff --git a/comfy_extras/nodes_string.py b/comfy_extras/nodes_string.py index 571d89f62..a2d5f0d94 100644 --- a/comfy_extras/nodes_string.py +++ b/comfy_extras/nodes_string.py @@ -11,6 +11,7 @@ class StringConcatenate(io.ComfyNode): node_id="StringConcatenate", display_name="Concatenate", category="utils/string", + search_aliases=["text concat", "join text", "merge text", "combine strings", "concat", "concatenate", "append text", "combine text", "string"], inputs=[ io.String.Input("string_a", multiline=True), io.String.Input("string_b", multiline=True), diff --git a/comfy_extras/nodes_upscale_model.py b/comfy_extras/nodes_upscale_model.py index ed587851c..97b9e948d 100644 --- a/comfy_extras/nodes_upscale_model.py +++ b/comfy_extras/nodes_upscale_model.py @@ -53,6 +53,7 @@ class ImageUpscaleWithModel(io.ComfyNode): node_id="ImageUpscaleWithModel", display_name="Upscale Image (using Model)", category="image/upscaling", + search_aliases=["upscale", "upscaler", "upsc", "enlarge image", "super resolution", "hires", "superres", "increase resolution"], inputs=[ io.UpscaleModel.Input("upscale_model"), io.Image.Input("image"), diff --git a/nodes.py b/nodes.py index ea5d6e525..67b61dcfe 100644 --- a/nodes.py +++ b/nodes.py @@ -70,6 +70,7 @@ class CLIPTextEncode(ComfyNodeABC): CATEGORY = "conditioning" DESCRIPTION = "Encodes a text prompt using a CLIP model into an embedding that can be used to guide the diffusion model towards generating specific images." + SEARCH_ALIASES = ["text", "prompt", "text prompt", "positive prompt", "negative prompt", "encode text", "text encoder", "encode prompt"] def encode(self, clip, text): if clip is None: @@ -86,6 +87,7 @@ class ConditioningCombine: FUNCTION = "combine" CATEGORY = "conditioning" + SEARCH_ALIASES = ["combine", "merge conditioning", "combine prompts", "merge prompts", "mix prompts", "add prompt"] def combine(self, conditioning_1, conditioning_2): return (conditioning_1 + conditioning_2, ) @@ -294,6 +296,7 @@ class VAEDecode: CATEGORY = "latent" DESCRIPTION = "Decodes latent images back into pixel space images." + SEARCH_ALIASES = ["decode", "decode latent", "latent to image", "render latent"] def decode(self, vae, samples): latent = samples["samples"] @@ -346,6 +349,7 @@ class VAEEncode: FUNCTION = "encode" CATEGORY = "latent" + SEARCH_ALIASES = ["encode", "encode image", "image to latent"] def encode(self, vae, pixels): t = vae.encode(pixels) @@ -581,6 +585,7 @@ class CheckpointLoaderSimple: CATEGORY = "loaders" DESCRIPTION = "Loads a diffusion model checkpoint, diffusion models are used to denoise latents." + SEARCH_ALIASES = ["load model", "checkpoint", "model loader", "load checkpoint", "ckpt", "model"] def load_checkpoint(self, ckpt_name): ckpt_path = folder_paths.get_full_path_or_raise("checkpoints", ckpt_name) @@ -667,6 +672,7 @@ class LoraLoader: CATEGORY = "loaders" DESCRIPTION = "LoRAs are used to modify diffusion and CLIP models, altering the way in which latents are denoised such as applying styles. Multiple LoRA nodes can be linked together." + SEARCH_ALIASES = ["lora", "load lora", "apply lora", "lora loader", "lora model"] def load_lora(self, model, clip, lora_name, strength_model, strength_clip): if strength_model == 0 and strength_clip == 0: @@ -814,6 +820,7 @@ class ControlNetLoader: FUNCTION = "load_controlnet" CATEGORY = "loaders" + SEARCH_ALIASES = ["controlnet", "control net", "cn", "load controlnet", "controlnet loader"] def load_controlnet(self, control_net_name): controlnet_path = folder_paths.get_full_path_or_raise("controlnet", control_net_name) @@ -890,6 +897,7 @@ class ControlNetApplyAdvanced: FUNCTION = "apply_controlnet" CATEGORY = "conditioning/controlnet" + SEARCH_ALIASES = ["controlnet", "apply controlnet", "use controlnet", "control net"] def apply_controlnet(self, positive, negative, control_net, image, strength, start_percent, end_percent, vae=None, extra_concat=[]): if strength == 0: @@ -1200,6 +1208,7 @@ class EmptyLatentImage: CATEGORY = "latent" DESCRIPTION = "Create a new batch of empty latent images to be denoised via sampling." + SEARCH_ALIASES = ["empty", "empty latent", "new latent", "create latent", "blank latent", "blank"] def generate(self, width, height, batch_size=1): latent = torch.zeros([batch_size, 4, height // 8, width // 8], device=self.device) @@ -1540,6 +1549,7 @@ class KSampler: CATEGORY = "sampling" DESCRIPTION = "Uses the provided model, positive and negative conditioning to denoise the latent image." + SEARCH_ALIASES = ["sampler", "sample", "generate", "denoise", "diffuse", "txt2img", "img2img"] def sample(self, model, seed, steps, cfg, sampler_name, scheduler, positive, negative, latent_image, denoise=1.0): return common_ksampler(model, seed, steps, cfg, sampler_name, scheduler, positive, negative, latent_image, denoise=denoise) @@ -1604,6 +1614,7 @@ class SaveImage: CATEGORY = "image" DESCRIPTION = "Saves the input images to your ComfyUI output directory." + SEARCH_ALIASES = ["save", "save image", "export image", "output image", "write image", "download"] def save_images(self, images, filename_prefix="ComfyUI", prompt=None, extra_pnginfo=None): filename_prefix += self.prefix_append @@ -1640,6 +1651,8 @@ class PreviewImage(SaveImage): self.prefix_append = "_temp_" + ''.join(random.choice("abcdefghijklmnopqrstupvxyz") for x in range(5)) self.compress_level = 1 + SEARCH_ALIASES = ["preview", "preview image", "show image", "view image", "display image", "image viewer"] + @classmethod def INPUT_TYPES(s): return {"required": @@ -1658,6 +1671,7 @@ class LoadImage: } CATEGORY = "image" + SEARCH_ALIASES = ["load image", "open image", "import image", "image input", "upload image", "read image", "image loader"] RETURN_TYPES = ("IMAGE", "MASK") FUNCTION = "load_image" @@ -1810,6 +1824,7 @@ class ImageScale: FUNCTION = "upscale" CATEGORY = "image/upscaling" + SEARCH_ALIASES = ["resize", "resize image", "scale image", "image resize", "zoom", "zoom in", "change size"] def upscale(self, image, upscale_method, width, height, crop): if width == 0 and height == 0: diff --git a/server.py b/server.py index 04a577488..1888745b7 100644 --- a/server.py +++ b/server.py @@ -682,6 +682,8 @@ class PromptServer(): if hasattr(obj_class, 'API_NODE'): info['api_node'] = obj_class.API_NODE + + info['search_aliases'] = getattr(obj_class, 'SEARCH_ALIASES', []) return info @routes.get("/object_info") From abe2ec26a61ff670b9c0e71e4821c873368c8728 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 21 Jan 2026 16:44:28 -0800 Subject: [PATCH 491/647] Support the Anima model. (#12012) --- comfy/ldm/anima/model.py | 202 +++++++++++++++++++++++++++++++++++ comfy/model_base.py | 22 ++++ comfy/model_detection.py | 2 + comfy/sd.py | 7 ++ comfy/supported_models.py | 33 +++++- comfy/text_encoders/anima.py | 61 +++++++++++ 6 files changed, 326 insertions(+), 1 deletion(-) create mode 100644 comfy/ldm/anima/model.py create mode 100644 comfy/text_encoders/anima.py diff --git a/comfy/ldm/anima/model.py b/comfy/ldm/anima/model.py new file mode 100644 index 000000000..2e6ed58fa --- /dev/null +++ b/comfy/ldm/anima/model.py @@ -0,0 +1,202 @@ +from comfy.ldm.cosmos.predict2 import MiniTrainDIT +import torch +from torch import nn +import torch.nn.functional as F + + +def rotate_half(x): + x1 = x[..., : x.shape[-1] // 2] + x2 = x[..., x.shape[-1] // 2 :] + return torch.cat((-x2, x1), dim=-1) + + +def apply_rotary_pos_emb(x, cos, sin, unsqueeze_dim=1): + cos = cos.unsqueeze(unsqueeze_dim) + sin = sin.unsqueeze(unsqueeze_dim) + x_embed = (x * cos) + (rotate_half(x) * sin) + return x_embed + + +class RotaryEmbedding(nn.Module): + def __init__(self, head_dim): + super().__init__() + self.rope_theta = 10000 + inv_freq = 1.0 / (self.rope_theta ** (torch.arange(0, head_dim, 2, dtype=torch.int64).to(dtype=torch.float) / head_dim)) + self.register_buffer("inv_freq", inv_freq, persistent=False) + + @torch.no_grad() + def forward(self, x, position_ids): + inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device) + position_ids_expanded = position_ids[:, None, :].float() + + device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu" + with torch.autocast(device_type=device_type, enabled=False): # Force float32 + freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2) + emb = torch.cat((freqs, freqs), dim=-1) + cos = emb.cos() + sin = emb.sin() + + return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype) + + +class Attention(nn.Module): + def __init__(self, query_dim, context_dim, n_heads, head_dim, device=None, dtype=None, operations=None): + super().__init__() + + inner_dim = head_dim * n_heads + self.n_heads = n_heads + self.head_dim = head_dim + self.query_dim = query_dim + self.context_dim = context_dim + + self.q_proj = operations.Linear(query_dim, inner_dim, bias=False, device=device, dtype=dtype) + self.q_norm = operations.RMSNorm(self.head_dim, eps=1e-6, device=device, dtype=dtype) + + self.k_proj = operations.Linear(context_dim, inner_dim, bias=False, device=device, dtype=dtype) + self.k_norm = operations.RMSNorm(self.head_dim, eps=1e-6, device=device, dtype=dtype) + + self.v_proj = operations.Linear(context_dim, inner_dim, bias=False, device=device, dtype=dtype) + + self.o_proj = operations.Linear(inner_dim, query_dim, bias=False, device=device, dtype=dtype) + + def forward(self, x, mask=None, context=None, position_embeddings=None, position_embeddings_context=None): + context = x if context is None else context + input_shape = x.shape[:-1] + q_shape = (*input_shape, self.n_heads, self.head_dim) + context_shape = context.shape[:-1] + kv_shape = (*context_shape, self.n_heads, self.head_dim) + + query_states = self.q_norm(self.q_proj(x).view(q_shape)).transpose(1, 2) + key_states = self.k_norm(self.k_proj(context).view(kv_shape)).transpose(1, 2) + value_states = self.v_proj(context).view(kv_shape).transpose(1, 2) + + if position_embeddings is not None: + assert position_embeddings_context is not None + cos, sin = position_embeddings + query_states = apply_rotary_pos_emb(query_states, cos, sin) + cos, sin = position_embeddings_context + key_states = apply_rotary_pos_emb(key_states, cos, sin) + + attn_output = F.scaled_dot_product_attention(query_states, key_states, value_states, attn_mask=mask) + + attn_output = attn_output.transpose(1, 2).reshape(*input_shape, -1).contiguous() + attn_output = self.o_proj(attn_output) + return attn_output + + def init_weights(self): + torch.nn.init.zeros_(self.o_proj.weight) + + +class TransformerBlock(nn.Module): + def __init__(self, source_dim, model_dim, num_heads=16, mlp_ratio=4.0, use_self_attn=False, layer_norm=False, device=None, dtype=None, operations=None): + super().__init__() + self.use_self_attn = use_self_attn + + if self.use_self_attn: + self.norm_self_attn = operations.LayerNorm(model_dim, device=device, dtype=dtype) if layer_norm else operations.RMSNorm(model_dim, eps=1e-6, device=device, dtype=dtype) + self.self_attn = Attention( + query_dim=model_dim, + context_dim=model_dim, + n_heads=num_heads, + head_dim=model_dim//num_heads, + device=device, + dtype=dtype, + operations=operations, + ) + + self.norm_cross_attn = operations.LayerNorm(model_dim, device=device, dtype=dtype) if layer_norm else operations.RMSNorm(model_dim, eps=1e-6, device=device, dtype=dtype) + self.cross_attn = Attention( + query_dim=model_dim, + context_dim=source_dim, + n_heads=num_heads, + head_dim=model_dim//num_heads, + device=device, + dtype=dtype, + operations=operations, + ) + + self.norm_mlp = operations.LayerNorm(model_dim, device=device, dtype=dtype) if layer_norm else operations.RMSNorm(model_dim, eps=1e-6, device=device, dtype=dtype) + self.mlp = nn.Sequential( + operations.Linear(model_dim, int(model_dim * mlp_ratio), device=device, dtype=dtype), + nn.GELU(), + operations.Linear(int(model_dim * mlp_ratio), model_dim, device=device, dtype=dtype) + ) + + def forward(self, x, context, target_attention_mask=None, source_attention_mask=None, position_embeddings=None, position_embeddings_context=None): + if self.use_self_attn: + normed = self.norm_self_attn(x) + attn_out = self.self_attn(normed, mask=target_attention_mask, position_embeddings=position_embeddings, position_embeddings_context=position_embeddings) + x = x + attn_out + + normed = self.norm_cross_attn(x) + attn_out = self.cross_attn(normed, mask=source_attention_mask, context=context, position_embeddings=position_embeddings, position_embeddings_context=position_embeddings_context) + x = x + attn_out + + x = x + self.mlp(self.norm_mlp(x)) + return x + + def init_weights(self): + torch.nn.init.zeros_(self.mlp[2].weight) + self.cross_attn.init_weights() + + +class LLMAdapter(nn.Module): + def __init__( + self, + source_dim=1024, + target_dim=1024, + model_dim=1024, + num_layers=6, + num_heads=16, + use_self_attn=True, + layer_norm=False, + device=None, + dtype=None, + operations=None, + ): + super().__init__() + + self.embed = operations.Embedding(32128, target_dim, device=device, dtype=dtype) + if model_dim != target_dim: + self.in_proj = operations.Linear(target_dim, model_dim, device=device, dtype=dtype) + else: + self.in_proj = nn.Identity() + self.rotary_emb = RotaryEmbedding(model_dim//num_heads) + self.blocks = nn.ModuleList([ + TransformerBlock(source_dim, model_dim, num_heads=num_heads, use_self_attn=use_self_attn, layer_norm=layer_norm, device=device, dtype=dtype, operations=operations) for _ in range(num_layers) + ]) + self.out_proj = operations.Linear(model_dim, target_dim, device=device, dtype=dtype) + self.norm = operations.RMSNorm(target_dim, eps=1e-6, device=device, dtype=dtype) + + def forward(self, source_hidden_states, target_input_ids, target_attention_mask=None, source_attention_mask=None): + if target_attention_mask is not None: + target_attention_mask = target_attention_mask.to(torch.bool) + if target_attention_mask.ndim == 2: + target_attention_mask = target_attention_mask.unsqueeze(1).unsqueeze(1) + + if source_attention_mask is not None: + source_attention_mask = source_attention_mask.to(torch.bool) + if source_attention_mask.ndim == 2: + source_attention_mask = source_attention_mask.unsqueeze(1).unsqueeze(1) + + x = self.in_proj(self.embed(target_input_ids)) + context = source_hidden_states + position_ids = torch.arange(x.shape[1], device=x.device).unsqueeze(0) + position_ids_context = torch.arange(context.shape[1], device=x.device).unsqueeze(0) + position_embeddings = self.rotary_emb(x, position_ids) + position_embeddings_context = self.rotary_emb(x, position_ids_context) + for block in self.blocks: + x = block(x, context, target_attention_mask=target_attention_mask, source_attention_mask=source_attention_mask, position_embeddings=position_embeddings, position_embeddings_context=position_embeddings_context) + return self.norm(self.out_proj(x)) + + +class Anima(MiniTrainDIT): + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + self.llm_adapter = LLMAdapter(device=kwargs.get("device"), dtype=kwargs.get("dtype"), operations=kwargs.get("operations")) + + def preprocess_text_embeds(self, text_embeds, text_ids): + if text_ids is not None: + return self.llm_adapter(text_embeds, text_ids) + else: + return text_embeds diff --git a/comfy/model_base.py b/comfy/model_base.py index 28ba2643e..1d57562cc 100644 --- a/comfy/model_base.py +++ b/comfy/model_base.py @@ -49,6 +49,7 @@ import comfy.ldm.ace.model import comfy.ldm.omnigen.omnigen2 import comfy.ldm.qwen_image.model import comfy.ldm.kandinsky5.model +import comfy.ldm.anima.model import comfy.model_management import comfy.patcher_extension @@ -1147,6 +1148,27 @@ class CosmosPredict2(BaseModel): sigma = (sigma / (sigma + 1)) return latent_image / (1.0 - sigma) +class Anima(BaseModel): + def __init__(self, model_config, model_type=ModelType.FLOW, device=None): + super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.anima.model.Anima) + + def extra_conds(self, **kwargs): + out = super().extra_conds(**kwargs) + cross_attn = kwargs.get("cross_attn", None) + t5xxl_ids = kwargs.get("t5xxl_ids", None) + t5xxl_weights = kwargs.get("t5xxl_weights", None) + device = kwargs["device"] + if cross_attn is not None: + if t5xxl_ids is not None: + cross_attn = self.diffusion_model.preprocess_text_embeds(cross_attn.to(device=device, dtype=self.get_dtype()), t5xxl_ids.unsqueeze(0).to(device=device)) + if t5xxl_weights is not None: + cross_attn *= t5xxl_weights.unsqueeze(0).unsqueeze(-1).to(cross_attn) + + if cross_attn.shape[1] < 512: + cross_attn = torch.nn.functional.pad(cross_attn, (0, 0, 0, 512 - cross_attn.shape[1])) + out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn) + return out + class Lumina2(BaseModel): def __init__(self, model_config, model_type=ModelType.FLOW, device=None): super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.lumina.model.NextDiT) diff --git a/comfy/model_detection.py b/comfy/model_detection.py index dad206a2f..b29a033cc 100644 --- a/comfy/model_detection.py +++ b/comfy/model_detection.py @@ -550,6 +550,8 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): if '{}blocks.0.mlp.layer1.weight'.format(key_prefix) in state_dict_keys: # Cosmos predict2 dit_config = {} dit_config["image_model"] = "cosmos_predict2" + if "{}llm_adapter.blocks.0.cross_attn.q_proj.weight".format(key_prefix) in state_dict_keys: + dit_config["image_model"] = "anima" dit_config["max_img_h"] = 240 dit_config["max_img_w"] = 240 dit_config["max_frames"] = 128 diff --git a/comfy/sd.py b/comfy/sd.py index 77700dfd3..f7f6a44a0 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -57,6 +57,7 @@ import comfy.text_encoders.ovis import comfy.text_encoders.kandinsky5 import comfy.text_encoders.jina_clip_2 import comfy.text_encoders.newbie +import comfy.text_encoders.anima import comfy.model_patcher import comfy.lora @@ -1048,6 +1049,7 @@ class TEModel(Enum): GEMMA_3_12B = 18 JINA_CLIP_2 = 19 QWEN3_8B = 20 + QWEN3_06B = 21 def detect_te_model(sd): @@ -1093,6 +1095,8 @@ def detect_te_model(sd): return TEModel.QWEN3_2B elif weight.shape[0] == 4096: return TEModel.QWEN3_8B + elif weight.shape[0] == 1024: + return TEModel.QWEN3_06B if weight.shape[0] == 5120: if "model.layers.39.post_attention_layernorm.weight" in sd: return TEModel.MISTRAL3_24B @@ -1233,6 +1237,9 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip elif te_model == TEModel.JINA_CLIP_2: clip_target.clip = comfy.text_encoders.jina_clip_2.JinaClip2TextModelWrapper clip_target.tokenizer = comfy.text_encoders.jina_clip_2.JinaClip2TokenizerWrapper + elif te_model == TEModel.QWEN3_06B: + clip_target.clip = comfy.text_encoders.anima.te(**llama_detect(clip_data)) + clip_target.tokenizer = comfy.text_encoders.anima.AnimaTokenizer else: # clip_l if clip_type == CLIPType.SD3: diff --git a/comfy/supported_models.py b/comfy/supported_models.py index c8a7f6efb..70abebf46 100644 --- a/comfy/supported_models.py +++ b/comfy/supported_models.py @@ -23,6 +23,7 @@ import comfy.text_encoders.qwen_image import comfy.text_encoders.hunyuan_image import comfy.text_encoders.kandinsky5 import comfy.text_encoders.z_image +import comfy.text_encoders.anima from . import supported_models_base from . import latent_formats @@ -992,6 +993,36 @@ class CosmosT2IPredict2(supported_models_base.BASE): t5_detect = comfy.text_encoders.sd3_clip.t5_xxl_detect(state_dict, "{}t5xxl.transformer.".format(pref)) return supported_models_base.ClipTarget(comfy.text_encoders.cosmos.CosmosT5Tokenizer, comfy.text_encoders.cosmos.te(**t5_detect)) +class Anima(supported_models_base.BASE): + unet_config = { + "image_model": "anima", + } + + sampling_settings = { + "multiplier": 1.0, + "shift": 3.0, + } + + unet_extra_config = {} + latent_format = latent_formats.Wan21 + + memory_usage_factor = 1.0 + + supported_inference_dtypes = [torch.bfloat16, torch.float32] + + def __init__(self, unet_config): + super().__init__(unet_config) + self.memory_usage_factor = (unet_config.get("model_channels", 2048) / 2048) * 0.95 + + def get_model(self, state_dict, prefix="", device=None): + out = model_base.Anima(self, device=device) + return out + + def clip_target(self, state_dict={}): + pref = self.text_encoder_key_prefix[0] + detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen3_06b.transformer.".format(pref)) + return supported_models_base.ClipTarget(comfy.text_encoders.anima.AnimaTokenizer, comfy.text_encoders.anima.te(**detect)) + class CosmosI2VPredict2(CosmosT2IPredict2): unet_config = { "image_model": "cosmos_predict2", @@ -1551,6 +1582,6 @@ class Kandinsky5Image(Kandinsky5): return supported_models_base.ClipTarget(comfy.text_encoders.kandinsky5.Kandinsky5TokenizerImage, comfy.text_encoders.kandinsky5.te(**hunyuan_detect)) -models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, LTXAV, HunyuanVideo15_SR_Distilled, HunyuanVideo15, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, ZImage, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, Omnigen2, QwenImage, Flux2, Kandinsky5Image, Kandinsky5] +models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, LTXAV, HunyuanVideo15_SR_Distilled, HunyuanVideo15, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, ZImage, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, Omnigen2, QwenImage, Flux2, Kandinsky5Image, Kandinsky5, Anima] models += [SVD_img2vid] diff --git a/comfy/text_encoders/anima.py b/comfy/text_encoders/anima.py new file mode 100644 index 000000000..41f95bcb6 --- /dev/null +++ b/comfy/text_encoders/anima.py @@ -0,0 +1,61 @@ +from transformers import Qwen2Tokenizer, T5TokenizerFast +import comfy.text_encoders.llama +from comfy import sd1_clip +import os +import torch + + +class Qwen3Tokenizer(sd1_clip.SDTokenizer): + def __init__(self, embedding_directory=None, tokenizer_data={}): + tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "qwen25_tokenizer") + super().__init__(tokenizer_path, pad_with_end=False, embedding_size=1024, embedding_key='qwen3_06b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, pad_token=151643, tokenizer_data=tokenizer_data) + +class T5XXLTokenizer(sd1_clip.SDTokenizer): + def __init__(self, embedding_directory=None, tokenizer_data={}): + tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_tokenizer") + super().__init__(tokenizer_path, embedding_directory=embedding_directory, pad_with_end=False, embedding_size=4096, embedding_key='t5xxl', tokenizer_class=T5TokenizerFast, has_start_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, tokenizer_data=tokenizer_data) + +class AnimaTokenizer: + def __init__(self, embedding_directory=None, tokenizer_data={}): + self.qwen3_06b = Qwen3Tokenizer(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data) + self.t5xxl = T5XXLTokenizer(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data) + + def tokenize_with_weights(self, text:str, return_word_ids=False, **kwargs): + out = {} + qwen_ids = self.qwen3_06b.tokenize_with_weights(text, return_word_ids, **kwargs) + out["qwen3_06b"] = [[(token, 1.0) for token, _ in inner_list] for inner_list in qwen_ids] # Set weights to 1.0 + out["t5xxl"] = self.t5xxl.tokenize_with_weights(text, return_word_ids, **kwargs) + return out + + def untokenize(self, token_weight_pair): + return self.t5xxl.untokenize(token_weight_pair) + + def state_dict(self): + return {} + + +class Qwen3_06BModel(sd1_clip.SDClipModel): + def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, attention_mask=True, model_options={}): + super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"pad": 151643}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Qwen3_06B, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options) + + +class AnimaTEModel(sd1_clip.SD1ClipModel): + def __init__(self, device="cpu", dtype=None, model_options={}): + super().__init__(device=device, dtype=dtype, name="qwen3_06b", clip_model=Qwen3_06BModel, model_options=model_options) + + def encode_token_weights(self, token_weight_pairs): + out = super().encode_token_weights(token_weight_pairs) + out[2]["t5xxl_ids"] = torch.tensor(list(map(lambda a: a[0], token_weight_pairs["t5xxl"][0])), dtype=torch.int) + out[2]["t5xxl_weights"] = torch.tensor(list(map(lambda a: a[1], token_weight_pairs["t5xxl"][0]))) + return out + +def te(dtype_llama=None, llama_quantization_metadata=None): + class AnimaTEModel_(AnimaTEModel): + def __init__(self, device="cpu", dtype=None, model_options={}): + if dtype_llama is not None: + dtype = dtype_llama + if llama_quantization_metadata is not None: + model_options = model_options.copy() + model_options["quantization_metadata"] = llama_quantization_metadata + super().__init__(device=device, dtype=dtype, model_options=model_options) + return AnimaTEModel_ From f09904720dc8b56bc6823ebdaf5de69465448e46 Mon Sep 17 00:00:00 2001 From: Jedrzej Kosinski Date: Wed, 21 Jan 2026 20:01:35 -0800 Subject: [PATCH 492/647] Fix for edge case of EasyCache when conditionings change during a sampling run (like with timestep scheduling) (#12020) --- comfy_extras/nodes_easycache.py | 9 +++++++-- 1 file changed, 7 insertions(+), 2 deletions(-) diff --git a/comfy_extras/nodes_easycache.py b/comfy_extras/nodes_easycache.py index 11b23ffdb..90d730df6 100644 --- a/comfy_extras/nodes_easycache.py +++ b/comfy_extras/nodes_easycache.py @@ -29,8 +29,10 @@ def easycache_forward_wrapper(executor, *args, **kwargs): do_easycache = easycache.should_do_easycache(sigmas) if do_easycache: easycache.check_metadata(x) + # if there isn't a cache diff for current conds, we cannot skip this step + can_apply_cache_diff = easycache.can_apply_cache_diff(uuids) # if first cond marked this step for skipping, skip it and use appropriate cached values - if easycache.skip_current_step: + if easycache.skip_current_step and can_apply_cache_diff: if easycache.verbose: logging.info(f"EasyCache [verbose] - was marked to skip this step by {easycache.first_cond_uuid}. Present uuids: {uuids}") return easycache.apply_cache_diff(x, uuids) @@ -44,7 +46,7 @@ def easycache_forward_wrapper(executor, *args, **kwargs): if easycache.has_output_prev_norm() and easycache.has_relative_transformation_rate(): approx_output_change_rate = (easycache.relative_transformation_rate * input_change) / easycache.output_prev_norm easycache.cumulative_change_rate += approx_output_change_rate - if easycache.cumulative_change_rate < easycache.reuse_threshold: + if easycache.cumulative_change_rate < easycache.reuse_threshold and can_apply_cache_diff: if easycache.verbose: logging.info(f"EasyCache [verbose] - skipping step; cumulative_change_rate: {easycache.cumulative_change_rate}, reuse_threshold: {easycache.reuse_threshold}") # other conds should also skip this step, and instead use their cached values @@ -240,6 +242,9 @@ class EasyCacheHolder: return to_return.clone() return to_return + def can_apply_cache_diff(self, uuids: list[UUID]) -> bool: + return all(uuid in self.uuid_cache_diffs for uuid in uuids) + def apply_cache_diff(self, x: torch.Tensor, uuids: list[UUID]): if self.first_cond_uuid in uuids: self.total_steps_skipped += 1 From 3365ad18a5e0c86b23c6272e5adcedd333fc45cb Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Jukka=20Sepp=C3=A4nen?= <40791699+kijai@users.noreply.github.com> Date: Thu, 22 Jan 2026 06:03:51 +0200 Subject: [PATCH 493/647] Support LTX2 tiny vae (taeltx_2) (#11929) --- comfy/sd.py | 5 ++--- comfy/taesd/taehv.py | 53 ++++++++++++++++++++++++++++---------------- latent_preview.py | 2 +- nodes.py | 2 +- 4 files changed, 38 insertions(+), 24 deletions(-) diff --git a/comfy/sd.py b/comfy/sd.py index f7f6a44a0..ce7e6bcff 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -636,14 +636,13 @@ class VAE: self.upscale_index_formula = (4, 16, 16) self.downscale_ratio = (lambda a: max(0, math.floor((a + 3) / 4)), 16, 16) self.downscale_index_formula = (4, 16, 16) - if self.latent_channels == 48: # Wan 2.2 + if self.latent_channels in [48, 128]: # Wan 2.2 and LTX2 self.first_stage_model = comfy.taesd.taehv.TAEHV(latent_channels=self.latent_channels, latent_format=None) # taehv doesn't need scaling - self.process_input = lambda image: (_ for _ in ()).throw(NotImplementedError("This light tae doesn't support encoding currently")) + self.process_input = self.process_output = lambda image: image self.process_output = lambda image: image self.memory_used_decode = lambda shape, dtype: (1800 * (max(1, (shape[-3] ** 0.7 * 0.1)) * shape[-2] * shape[-1] * 16 * 16) * model_management.dtype_size(dtype)) elif self.latent_channels == 32 and sd["decoder.22.bias"].shape[0] == 12: # lighttae_hv15 self.first_stage_model = comfy.taesd.taehv.TAEHV(latent_channels=self.latent_channels, latent_format=comfy.latent_formats.HunyuanVideo15) - self.process_input = lambda image: (_ for _ in ()).throw(NotImplementedError("This light tae doesn't support encoding currently")) self.memory_used_decode = lambda shape, dtype: (1200 * (max(1, (shape[-3] ** 0.7 * 0.05)) * shape[-2] * shape[-1] * 32 * 32) * model_management.dtype_size(dtype)) else: if sd["decoder.1.weight"].dtype == torch.float16: # taehv currently only available in float16, so assume it's not lighttaew2_1 as otherwise state dicts are identical diff --git a/comfy/taesd/taehv.py b/comfy/taesd/taehv.py index 0e5f9a378..6c06ce19d 100644 --- a/comfy/taesd/taehv.py +++ b/comfy/taesd/taehv.py @@ -112,7 +112,8 @@ def apply_model_with_memblocks(model, x, parallel, show_progress_bar): class TAEHV(nn.Module): - def __init__(self, latent_channels, parallel=False, decoder_time_upscale=(True, True), decoder_space_upscale=(True, True, True), latent_format=None, show_progress_bar=True): + def __init__(self, latent_channels, parallel=False, encoder_time_downscale=(True, True, False), decoder_time_upscale=(False, True, True), decoder_space_upscale=(True, True, True), + latent_format=None, show_progress_bar=False): super().__init__() self.image_channels = 3 self.patch_size = 1 @@ -124,6 +125,9 @@ class TAEHV(nn.Module): self.process_out = latent_format().process_out if latent_format is not None else (lambda x: x) if self.latent_channels in [48, 32]: # Wan 2.2 and HunyuanVideo1.5 self.patch_size = 2 + elif self.latent_channels == 128: # LTX2 + self.patch_size, self.latent_channels, encoder_time_downscale, decoder_time_upscale = 4, 128, (True, True, True), (True, True, True) + if self.latent_channels == 32: # HunyuanVideo1.5 act_func = nn.LeakyReLU(0.2, inplace=True) else: # HunyuanVideo, Wan 2.1 @@ -131,41 +135,52 @@ class TAEHV(nn.Module): self.encoder = nn.Sequential( conv(self.image_channels*self.patch_size**2, 64), act_func, - TPool(64, 2), conv(64, 64, stride=2, bias=False), MemBlock(64, 64, act_func), MemBlock(64, 64, act_func), MemBlock(64, 64, act_func), - TPool(64, 2), conv(64, 64, stride=2, bias=False), MemBlock(64, 64, act_func), MemBlock(64, 64, act_func), MemBlock(64, 64, act_func), - TPool(64, 1), conv(64, 64, stride=2, bias=False), MemBlock(64, 64, act_func), MemBlock(64, 64, act_func), MemBlock(64, 64, act_func), + TPool(64, 2 if encoder_time_downscale[0] else 1), conv(64, 64, stride=2, bias=False), MemBlock(64, 64, act_func), MemBlock(64, 64, act_func), MemBlock(64, 64, act_func), + TPool(64, 2 if encoder_time_downscale[1] else 1), conv(64, 64, stride=2, bias=False), MemBlock(64, 64, act_func), MemBlock(64, 64, act_func), MemBlock(64, 64, act_func), + TPool(64, 2 if encoder_time_downscale[2] else 1), conv(64, 64, stride=2, bias=False), MemBlock(64, 64, act_func), MemBlock(64, 64, act_func), MemBlock(64, 64, act_func), conv(64, self.latent_channels), ) n_f = [256, 128, 64, 64] - self.frames_to_trim = 2**sum(decoder_time_upscale) - 1 + self.decoder = nn.Sequential( Clamp(), conv(self.latent_channels, n_f[0]), act_func, - MemBlock(n_f[0], n_f[0], act_func), MemBlock(n_f[0], n_f[0], act_func), MemBlock(n_f[0], n_f[0], act_func), nn.Upsample(scale_factor=2 if decoder_space_upscale[0] else 1), TGrow(n_f[0], 1), conv(n_f[0], n_f[1], bias=False), - MemBlock(n_f[1], n_f[1], act_func), MemBlock(n_f[1], n_f[1], act_func), MemBlock(n_f[1], n_f[1], act_func), nn.Upsample(scale_factor=2 if decoder_space_upscale[1] else 1), TGrow(n_f[1], 2 if decoder_time_upscale[0] else 1), conv(n_f[1], n_f[2], bias=False), - MemBlock(n_f[2], n_f[2], act_func), MemBlock(n_f[2], n_f[2], act_func), MemBlock(n_f[2], n_f[2], act_func), nn.Upsample(scale_factor=2 if decoder_space_upscale[2] else 1), TGrow(n_f[2], 2 if decoder_time_upscale[1] else 1), conv(n_f[2], n_f[3], bias=False), + MemBlock(n_f[0], n_f[0], act_func), MemBlock(n_f[0], n_f[0], act_func), MemBlock(n_f[0], n_f[0], act_func), nn.Upsample(scale_factor=2 if decoder_space_upscale[0] else 1), TGrow(n_f[0], 2 if decoder_time_upscale[0] else 1), conv(n_f[0], n_f[1], bias=False), + MemBlock(n_f[1], n_f[1], act_func), MemBlock(n_f[1], n_f[1], act_func), MemBlock(n_f[1], n_f[1], act_func), nn.Upsample(scale_factor=2 if decoder_space_upscale[1] else 1), TGrow(n_f[1], 2 if decoder_time_upscale[1] else 1), conv(n_f[1], n_f[2], bias=False), + MemBlock(n_f[2], n_f[2], act_func), MemBlock(n_f[2], n_f[2], act_func), MemBlock(n_f[2], n_f[2], act_func), nn.Upsample(scale_factor=2 if decoder_space_upscale[2] else 1), TGrow(n_f[2], 2 if decoder_time_upscale[2] else 1), conv(n_f[2], n_f[3], bias=False), act_func, conv(n_f[3], self.image_channels*self.patch_size**2), ) - @property - def show_progress_bar(self): - return self._show_progress_bar - @show_progress_bar.setter - def show_progress_bar(self, value): - self._show_progress_bar = value + self.t_downscale = 2**sum(t.stride == 2 for t in self.encoder if isinstance(t, TPool)) + self.t_upscale = 2**sum(t.stride == 2 for t in self.decoder if isinstance(t, TGrow)) + self.frames_to_trim = self.t_upscale - 1 + self._show_progress_bar = show_progress_bar + + @property + def show_progress_bar(self): + return self._show_progress_bar + + @show_progress_bar.setter + def show_progress_bar(self, value): + self._show_progress_bar = value def encode(self, x, **kwargs): - if self.patch_size > 1: - x = F.pixel_unshuffle(x, self.patch_size) x = x.movedim(2, 1) # [B, C, T, H, W] -> [B, T, C, H, W] - if x.shape[1] % 4 != 0: - # pad at end to multiple of 4 - n_pad = 4 - x.shape[1] % 4 + if self.patch_size > 1: + B, T, C, H, W = x.shape + x = x.reshape(B * T, C, H, W) + x = F.pixel_unshuffle(x, self.patch_size) + x = x.reshape(B, T, C * self.patch_size ** 2, H // self.patch_size, W // self.patch_size) + if x.shape[1] % self.t_downscale != 0: + # pad at end to multiple of t_downscale + n_pad = self.t_downscale - x.shape[1] % self.t_downscale padding = x[:, -1:].repeat_interleave(n_pad, dim=1) x = torch.cat([x, padding], 1) x = apply_model_with_memblocks(self.encoder, x, self.parallel, self.show_progress_bar).movedim(2, 1) return self.process_out(x) def decode(self, x, **kwargs): + x = x.unsqueeze(0) if x.ndim == 4 else x # [T, C, H, W] -> [1, T, C, H, W] + x = x.movedim(1, 2) if x.shape[1] != self.latent_channels else x # [B, T, C, H, W] or [B, C, T, H, W] x = self.process_in(x).movedim(2, 1) # [B, C, T, H, W] -> [B, T, C, H, W] x = apply_model_with_memblocks(self.decoder, x, self.parallel, self.show_progress_bar) if self.patch_size > 1: diff --git a/latent_preview.py b/latent_preview.py index d52e3f7a1..a9d777661 100644 --- a/latent_preview.py +++ b/latent_preview.py @@ -11,7 +11,7 @@ import logging default_preview_method = args.preview_method MAX_PREVIEW_RESOLUTION = args.preview_size -VIDEO_TAES = ["taehv", "lighttaew2_2", "lighttaew2_1", "lighttaehy1_5"] +VIDEO_TAES = ["taehv", "lighttaew2_2", "lighttaew2_1", "lighttaehy1_5", "taeltx_2"] def preview_to_image(latent_image, do_scale=True): if do_scale: diff --git a/nodes.py b/nodes.py index 67b61dcfe..8864fda60 100644 --- a/nodes.py +++ b/nodes.py @@ -707,7 +707,7 @@ class LoraLoaderModelOnly(LoraLoader): return (self.load_lora(model, None, lora_name, strength_model, 0)[0],) class VAELoader: - video_taes = ["taehv", "lighttaew2_2", "lighttaew2_1", "lighttaehy1_5"] + video_taes = ["taehv", "lighttaew2_2", "lighttaew2_1", "lighttaehy1_5", "taeltx_2"] image_taes = ["taesd", "taesdxl", "taesd3", "taef1"] @staticmethod def vae_list(s): From 245f6139b65899112d11ff294d36a820f2d69496 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Jukka=20Sepp=C3=A4nen?= <40791699+kijai@users.noreply.github.com> Date: Thu, 22 Jan 2026 06:05:06 +0200 Subject: [PATCH 494/647] More targeted embedding_connector loading for LTX2 text encoder (#11992) Reduces errors --- comfy/text_encoders/lt.py | 15 ++++++++++++--- 1 file changed, 12 insertions(+), 3 deletions(-) diff --git a/comfy/text_encoders/lt.py b/comfy/text_encoders/lt.py index c33c77db7..e49161964 100644 --- a/comfy/text_encoders/lt.py +++ b/comfy/text_encoders/lt.py @@ -118,9 +118,18 @@ class LTXAVTEModel(torch.nn.Module): sdo = comfy.utils.state_dict_prefix_replace(sd, {"text_embedding_projection.aggregate_embed.weight": "text_embedding_projection.weight", "model.diffusion_model.video_embeddings_connector.": "video_embeddings_connector.", "model.diffusion_model.audio_embeddings_connector.": "audio_embeddings_connector."}, filter_keys=True) if len(sdo) == 0: sdo = sd - missing, unexpected = self.load_state_dict(sdo, strict=False) - missing = [k for k in missing if not k.startswith("gemma3_12b.")] # filter out keys that belong to the main gemma model - return (missing, unexpected) + + missing_all = [] + unexpected_all = [] + + for prefix, component in [("text_embedding_projection.", self.text_embedding_projection), ("video_embeddings_connector.", self.video_embeddings_connector), ("audio_embeddings_connector.", self.audio_embeddings_connector)]: + component_sd = {k.replace(prefix, ""): v for k, v in sdo.items() if k.startswith(prefix)} + if component_sd: + missing, unexpected = component.load_state_dict(component_sd, strict=False) + missing_all.extend([f"{prefix}{k}" for k in missing]) + unexpected_all.extend([f"{prefix}{k}" for k in unexpected]) + + return (missing_all, unexpected_all) def memory_estimation_function(self, token_weight_pairs, device=None): constant = 6.0 From 16b9aabd52c3b81b365fbf562bbcc4528111ef6b Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Jukka=20Sepp=C3=A4nen?= <40791699+kijai@users.noreply.github.com> Date: Thu, 22 Jan 2026 06:09:48 +0200 Subject: [PATCH 495/647] Support Multi/InfiniteTalk (#10179) * re-init * Update model_multitalk.py * whitespace... * Update model_multitalk.py * remove print * this is redundant * remove import * Restore preview functionality * Move block_idx to transformer_options * Remove LoopingSamplerCustomAdvanced * Remove looping functionality, keep extension functionality * Update model_multitalk.py * Handle ref_attn_mask with separate patch to avoid having to always return q and k from self_attn * Chunk attention map calculation for multiple speakers to reduce peak VRAM usage * Update model_multitalk.py * Add ModelPatch type back * Fix for latest upstream * Use DynamicCombo for cleaner node Basically just so that single_speaker mode hides mask inputs and 2nd audio input * Update nodes_wan.py --- comfy/ldm/wan/model.py | 17 +- comfy/ldm/wan/model_multitalk.py | 500 ++++++++++++++++++++++++++++++ comfy_api/latest/_io.py | 3 +- comfy_extras/nodes_model_patch.py | 41 +++ comfy_extras/nodes_wan.py | 169 +++++++++- 5 files changed, 727 insertions(+), 3 deletions(-) create mode 100644 comfy/ldm/wan/model_multitalk.py diff --git a/comfy/ldm/wan/model.py b/comfy/ldm/wan/model.py index 4216ce831..ea123acb4 100644 --- a/comfy/ldm/wan/model.py +++ b/comfy/ldm/wan/model.py @@ -62,6 +62,8 @@ class WanSelfAttention(nn.Module): x(Tensor): Shape [B, L, num_heads, C / num_heads] freqs(Tensor): Rope freqs, shape [1024, C / num_heads / 2] """ + patches = transformer_options.get("patches", {}) + b, s, n, d = *x.shape[:2], self.num_heads, self.head_dim def qkv_fn_q(x): @@ -86,6 +88,10 @@ class WanSelfAttention(nn.Module): transformer_options=transformer_options, ) + if "attn1_patch" in patches: + for p in patches["attn1_patch"]: + x = p({"x": x, "q": q, "k": k, "transformer_options": transformer_options}) + x = self.o(x) return x @@ -225,6 +231,8 @@ class WanAttentionBlock(nn.Module): """ # assert e.dtype == torch.float32 + patches = transformer_options.get("patches", {}) + if e.ndim < 4: e = (comfy.model_management.cast_to(self.modulation, dtype=x.dtype, device=x.device) + e).chunk(6, dim=1) else: @@ -242,6 +250,11 @@ class WanAttentionBlock(nn.Module): # cross-attention & ffn x = x + self.cross_attn(self.norm3(x), context, context_img_len=context_img_len, transformer_options=transformer_options) + + if "attn2_patch" in patches: + for p in patches["attn2_patch"]: + x = p({"x": x, "transformer_options": transformer_options}) + y = self.ffn(torch.addcmul(repeat_e(e[3], x), self.norm2(x), 1 + repeat_e(e[4], x))) x = torch.addcmul(x, y, repeat_e(e[5], x)) return x @@ -488,7 +501,7 @@ class WanModel(torch.nn.Module): self.blocks = nn.ModuleList([ wan_attn_block_class(cross_attn_type, dim, ffn_dim, num_heads, window_size, qk_norm, cross_attn_norm, eps, operation_settings=operation_settings) - for _ in range(num_layers) + for i in range(num_layers) ]) # head @@ -541,6 +554,7 @@ class WanModel(torch.nn.Module): # embeddings x = self.patch_embedding(x.float()).to(x.dtype) grid_sizes = x.shape[2:] + transformer_options["grid_sizes"] = grid_sizes x = x.flatten(2).transpose(1, 2) # time embeddings @@ -738,6 +752,7 @@ class VaceWanModel(WanModel): # embeddings x = self.patch_embedding(x.float()).to(x.dtype) grid_sizes = x.shape[2:] + transformer_options["grid_sizes"] = grid_sizes x = x.flatten(2).transpose(1, 2) # time embeddings diff --git a/comfy/ldm/wan/model_multitalk.py b/comfy/ldm/wan/model_multitalk.py new file mode 100644 index 000000000..c9dd98c4d --- /dev/null +++ b/comfy/ldm/wan/model_multitalk.py @@ -0,0 +1,500 @@ +import torch +from einops import rearrange, repeat +import comfy +from comfy.ldm.modules.attention import optimized_attention + + +def calculate_x_ref_attn_map(visual_q, ref_k, ref_target_masks, split_num=8): + scale = 1.0 / visual_q.shape[-1] ** 0.5 + visual_q = visual_q.transpose(1, 2) * scale + + B, H, x_seqlens, K = visual_q.shape + + x_ref_attn_maps = [] + for class_idx, ref_target_mask in enumerate(ref_target_masks): + ref_target_mask = ref_target_mask.view(1, 1, 1, -1) + + x_ref_attnmap = torch.zeros(B, H, x_seqlens, device=visual_q.device, dtype=visual_q.dtype) + chunk_size = min(max(x_seqlens // split_num, 1), x_seqlens) + + for i in range(0, x_seqlens, chunk_size): + end_i = min(i + chunk_size, x_seqlens) + + attn_chunk = visual_q[:, :, i:end_i] @ ref_k.permute(0, 2, 3, 1) # B, H, chunk, ref_seqlens + + # Apply softmax + attn_max = attn_chunk.max(dim=-1, keepdim=True).values + attn_chunk = (attn_chunk - attn_max).exp() + attn_sum = attn_chunk.sum(dim=-1, keepdim=True) + attn_chunk = attn_chunk / (attn_sum + 1e-8) + + # Apply mask and sum + masked_attn = attn_chunk * ref_target_mask + x_ref_attnmap[:, :, i:end_i] = masked_attn.sum(-1) / (ref_target_mask.sum() + 1e-8) + + del attn_chunk, masked_attn + + # Average across heads + x_ref_attnmap = x_ref_attnmap.mean(dim=1) # B, x_seqlens + x_ref_attn_maps.append(x_ref_attnmap) + + del visual_q, ref_k + + return torch.cat(x_ref_attn_maps, dim=0) + +def get_attn_map_with_target(visual_q, ref_k, shape, ref_target_masks=None, split_num=2): + """Args: + query (torch.tensor): B M H K + key (torch.tensor): B M H K + shape (tuple): (N_t, N_h, N_w) + ref_target_masks: [B, N_h * N_w] + """ + + N_t, N_h, N_w = shape + + x_seqlens = N_h * N_w + ref_k = ref_k[:, :x_seqlens] + _, seq_lens, heads, _ = visual_q.shape + class_num, _ = ref_target_masks.shape + x_ref_attn_maps = torch.zeros(class_num, seq_lens).to(visual_q) + + split_chunk = heads // split_num + + for i in range(split_num): + x_ref_attn_maps_perhead = calculate_x_ref_attn_map( + visual_q[:, :, i*split_chunk:(i+1)*split_chunk, :], + ref_k[:, :, i*split_chunk:(i+1)*split_chunk, :], + ref_target_masks + ) + x_ref_attn_maps += x_ref_attn_maps_perhead + + return x_ref_attn_maps / split_num + + +def normalize_and_scale(column, source_range, target_range, epsilon=1e-8): + source_min, source_max = source_range + new_min, new_max = target_range + normalized = (column - source_min) / (source_max - source_min + epsilon) + scaled = normalized * (new_max - new_min) + new_min + return scaled + + +def rotate_half(x): + x = rearrange(x, "... (d r) -> ... d r", r=2) + x1, x2 = x.unbind(dim=-1) + x = torch.stack((-x2, x1), dim=-1) + return rearrange(x, "... d r -> ... (d r)") + + +def get_audio_embeds(encoded_audio, audio_start, audio_end): + audio_embs = [] + human_num = len(encoded_audio) + audio_frames = encoded_audio[0].shape[0] + + indices = (torch.arange(4 + 1) - 2) * 1 + + for human_idx in range(human_num): + if audio_end > audio_frames: # in case of not enough audio for current window, pad with first audio frame as that's most likely silence + pad_len = audio_end - audio_frames + pad_shape = list(encoded_audio[human_idx].shape) + pad_shape[0] = pad_len + pad_tensor = encoded_audio[human_idx][:1].repeat(pad_len, *([1] * (encoded_audio[human_idx].dim() - 1))) + encoded_audio_in = torch.cat([encoded_audio[human_idx], pad_tensor], dim=0) + else: + encoded_audio_in = encoded_audio[human_idx] + center_indices = torch.arange(audio_start, audio_end, 1).unsqueeze(1) + indices.unsqueeze(0) + center_indices = torch.clamp(center_indices, min=0, max=encoded_audio_in.shape[0] - 1) + audio_emb = encoded_audio_in[center_indices].unsqueeze(0) + audio_embs.append(audio_emb) + + return torch.cat(audio_embs, dim=0) + + +def project_audio_features(audio_proj, encoded_audio, audio_start, audio_end): + audio_embs = get_audio_embeds(encoded_audio, audio_start, audio_end) + + first_frame_audio_emb_s = audio_embs[:, :1, ...] + latter_frame_audio_emb = audio_embs[:, 1:, ...] + latter_frame_audio_emb = rearrange(latter_frame_audio_emb, "b (n_t n) w s c -> b n_t n w s c", n=4) + + middle_index = audio_proj.seq_len // 2 + + latter_first_frame_audio_emb = latter_frame_audio_emb[:, :, :1, :middle_index+1, ...] + latter_first_frame_audio_emb = rearrange(latter_first_frame_audio_emb, "b n_t n w s c -> b n_t (n w) s c") + latter_last_frame_audio_emb = latter_frame_audio_emb[:, :, -1:, middle_index:, ...] + latter_last_frame_audio_emb = rearrange(latter_last_frame_audio_emb, "b n_t n w s c -> b n_t (n w) s c") + latter_middle_frame_audio_emb = latter_frame_audio_emb[:, :, 1:-1, middle_index:middle_index+1, ...] + latter_middle_frame_audio_emb = rearrange(latter_middle_frame_audio_emb, "b n_t n w s c -> b n_t (n w) s c") + latter_frame_audio_emb_s = torch.cat([latter_first_frame_audio_emb, latter_middle_frame_audio_emb, latter_last_frame_audio_emb], dim=2) + + audio_emb = audio_proj(first_frame_audio_emb_s, latter_frame_audio_emb_s) + audio_emb = torch.cat(audio_emb.split(1), dim=2) + + return audio_emb + + +class RotaryPositionalEmbedding1D(torch.nn.Module): + def __init__(self, + head_dim, + ): + super().__init__() + self.head_dim = head_dim + self.base = 10000 + + def precompute_freqs_cis_1d(self, pos_indices): + freqs = 1.0 / (self.base ** (torch.arange(0, self.head_dim, 2)[: (self.head_dim // 2)].float() / self.head_dim)) + freqs = freqs.to(pos_indices.device) + freqs = torch.einsum("..., f -> ... f", pos_indices.float(), freqs) + freqs = repeat(freqs, "... n -> ... (n r)", r=2) + return freqs + + def forward(self, x, pos_indices): + freqs_cis = self.precompute_freqs_cis_1d(pos_indices) + + x_ = x.float() + + freqs_cis = freqs_cis.float().to(x.device) + cos, sin = freqs_cis.cos(), freqs_cis.sin() + cos, sin = rearrange(cos, 'n d -> 1 1 n d'), rearrange(sin, 'n d -> 1 1 n d') + x_ = (x_ * cos) + (rotate_half(x_) * sin) + + return x_.type_as(x) + +class SingleStreamAttention(torch.nn.Module): + def __init__( + self, + dim: int, + encoder_hidden_states_dim: int, + num_heads: int, + qkv_bias: bool, + device=None, dtype=None, operations=None + ) -> None: + super().__init__() + self.dim = dim + self.encoder_hidden_states_dim = encoder_hidden_states_dim + self.num_heads = num_heads + self.head_dim = dim // num_heads + + self.q_linear = operations.Linear(dim, dim, bias=qkv_bias, device=device, dtype=dtype) + self.proj = operations.Linear(dim, dim, device=device, dtype=dtype) + self.kv_linear = operations.Linear(encoder_hidden_states_dim, dim * 2, bias=qkv_bias, device=device, dtype=dtype) + + def forward(self, x: torch.Tensor, encoder_hidden_states: torch.Tensor, shape=None) -> torch.Tensor: + N_t, N_h, N_w = shape + + expected_tokens = N_t * N_h * N_w + actual_tokens = x.shape[1] + x_extra = None + + if actual_tokens != expected_tokens: + x_extra = x[:, -N_h * N_w:, :] + x = x[:, :-N_h * N_w, :] + N_t = N_t - 1 + + B = x.shape[0] + S = N_h * N_w + x = x.view(B * N_t, S, self.dim) + + # get q for hidden_state + q = self.q_linear(x).view(B * N_t, S, self.num_heads, self.head_dim) + + # get kv from encoder_hidden_states # shape: (B, N, num_heads, head_dim) + kv = self.kv_linear(encoder_hidden_states) + encoder_k, encoder_v = kv.view(B * N_t, encoder_hidden_states.shape[1], 2, self.num_heads, self.head_dim).unbind(2) + + #print("q.shape", q.shape) #torch.Size([21, 1024, 40, 128]) + x = optimized_attention( + q.transpose(1, 2), + encoder_k.transpose(1, 2), + encoder_v.transpose(1, 2), + heads=self.num_heads, skip_reshape=True, skip_output_reshape=True).transpose(1, 2) + + # linear transform + x = self.proj(x.reshape(B * N_t, S, self.dim)) + x = x.view(B, N_t * S, self.dim) + + if x_extra is not None: + x = torch.cat([x, torch.zeros_like(x_extra)], dim=1) + + return x + +class SingleStreamMultiAttention(SingleStreamAttention): + def __init__( + self, + dim: int, + encoder_hidden_states_dim: int, + num_heads: int, + qkv_bias: bool, + class_range: int = 24, + class_interval: int = 4, + device=None, dtype=None, operations=None + ) -> None: + super().__init__( + dim=dim, + encoder_hidden_states_dim=encoder_hidden_states_dim, + num_heads=num_heads, + qkv_bias=qkv_bias, + device=device, + dtype=dtype, + operations=operations + ) + + # Rotary-embedding layout parameters + self.class_interval = class_interval + self.class_range = class_range + self.max_humans = self.class_range // self.class_interval + + # Constant bucket used for background tokens + self.rope_bak = int(self.class_range // 2) + + self.rope_1d = RotaryPositionalEmbedding1D(self.head_dim) + + def forward( + self, + x: torch.Tensor, + encoder_hidden_states: torch.Tensor, + shape=None, + x_ref_attn_map=None + ) -> torch.Tensor: + encoder_hidden_states = encoder_hidden_states.squeeze(0).to(x.device) + human_num = x_ref_attn_map.shape[0] if x_ref_attn_map is not None else 1 + # Single-speaker fall-through + if human_num <= 1: + return super().forward(x, encoder_hidden_states, shape) + + N_t, N_h, N_w = shape + + x_extra = None + if x.shape[0] * N_t != encoder_hidden_states.shape[0]: + x_extra = x[:, -N_h * N_w:, :] + x = x[:, :-N_h * N_w, :] + N_t = N_t - 1 + x = rearrange(x, "B (N_t S) C -> (B N_t) S C", N_t=N_t) + + # Query projection + B, N, C = x.shape + q = self.q_linear(x) + q = q.view(B, N, self.num_heads, self.head_dim).permute(0, 2, 1, 3) + + # Use `class_range` logic for 2 speakers + rope_h1 = (0, self.class_interval) + rope_h2 = (self.class_range - self.class_interval, self.class_range) + rope_bak = int(self.class_range // 2) + + # Normalize and scale attention maps for each speaker + max_values = x_ref_attn_map.max(1).values[:, None, None] + min_values = x_ref_attn_map.min(1).values[:, None, None] + max_min_values = torch.cat([max_values, min_values], dim=2) + + human1_max_value, human1_min_value = max_min_values[0, :, 0].max(), max_min_values[0, :, 1].min() + human2_max_value, human2_min_value = max_min_values[1, :, 0].max(), max_min_values[1, :, 1].min() + + human1 = normalize_and_scale(x_ref_attn_map[0], (human1_min_value, human1_max_value), rope_h1) + human2 = normalize_and_scale(x_ref_attn_map[1], (human2_min_value, human2_max_value), rope_h2) + back = torch.full((x_ref_attn_map.size(1),), rope_bak, dtype=human1.dtype, device=human1.device) + + # Token-wise speaker dominance + max_indices = x_ref_attn_map.argmax(dim=0) + normalized_map = torch.stack([human1, human2, back], dim=1) + normalized_pos = normalized_map[torch.arange(x_ref_attn_map.size(1)), max_indices] + + # Apply rotary to Q + q = rearrange(q, "(B N_t) H S C -> B H (N_t S) C", N_t=N_t) + q = self.rope_1d(q, normalized_pos) + q = rearrange(q, "B H (N_t S) C -> (B N_t) H S C", N_t=N_t) + + # Keys / Values + _, N_a, _ = encoder_hidden_states.shape + encoder_kv = self.kv_linear(encoder_hidden_states) + encoder_kv = encoder_kv.view(B, N_a, 2, self.num_heads, self.head_dim).permute(2, 0, 3, 1, 4) + encoder_k, encoder_v = encoder_kv.unbind(0) + + # Rotary for keys – assign centre of each speaker bucket to its context tokens + per_frame = torch.zeros(N_a, dtype=encoder_k.dtype, device=encoder_k.device) + per_frame[: per_frame.size(0) // 2] = (rope_h1[0] + rope_h1[1]) / 2 + per_frame[per_frame.size(0) // 2 :] = (rope_h2[0] + rope_h2[1]) / 2 + encoder_pos = torch.cat([per_frame] * N_t, dim=0) + + encoder_k = rearrange(encoder_k, "(B N_t) H S C -> B H (N_t S) C", N_t=N_t) + encoder_k = self.rope_1d(encoder_k, encoder_pos) + encoder_k = rearrange(encoder_k, "B H (N_t S) C -> (B N_t) H S C", N_t=N_t) + + # Final attention + q = rearrange(q, "B H M K -> B M H K") + encoder_k = rearrange(encoder_k, "B H M K -> B M H K") + encoder_v = rearrange(encoder_v, "B H M K -> B M H K") + + x = optimized_attention( + q.transpose(1, 2), + encoder_k.transpose(1, 2), + encoder_v.transpose(1, 2), + heads=self.num_heads, skip_reshape=True, skip_output_reshape=True).transpose(1, 2) + + # Linear projection + x = x.reshape(B, N, C) + x = self.proj(x) + + # Restore original layout + x = rearrange(x, "(B N_t) S C -> B (N_t S) C", N_t=N_t) + if x_extra is not None: + x = torch.cat([x, torch.zeros_like(x_extra)], dim=1) + + return x + + +class MultiTalkAudioProjModel(torch.nn.Module): + def __init__( + self, + seq_len: int = 5, + seq_len_vf: int = 12, + blocks: int = 12, + channels: int = 768, + intermediate_dim: int = 512, + out_dim: int = 768, + context_tokens: int = 32, + device=None, dtype=None, operations=None + ): + super().__init__() + + self.seq_len = seq_len + self.blocks = blocks + self.channels = channels + self.input_dim = seq_len * blocks * channels + self.input_dim_vf = seq_len_vf * blocks * channels + self.intermediate_dim = intermediate_dim + self.context_tokens = context_tokens + self.out_dim = out_dim + + # define multiple linear layers + self.proj1 = operations.Linear(self.input_dim, intermediate_dim, device=device, dtype=dtype) + self.proj1_vf = operations.Linear(self.input_dim_vf, intermediate_dim, device=device, dtype=dtype) + self.proj2 = operations.Linear(intermediate_dim, intermediate_dim, device=device, dtype=dtype) + self.proj3 = operations.Linear(intermediate_dim, context_tokens * out_dim, device=device, dtype=dtype) + self.norm = operations.LayerNorm(out_dim, device=device, dtype=dtype) + + def forward(self, audio_embeds, audio_embeds_vf): + video_length = audio_embeds.shape[1] + audio_embeds_vf.shape[1] + B, _, _, S, C = audio_embeds.shape + + # process audio of first frame + audio_embeds = rearrange(audio_embeds, "bz f w b c -> (bz f) w b c") + batch_size, window_size, blocks, channels = audio_embeds.shape + audio_embeds = audio_embeds.view(batch_size, window_size * blocks * channels) + + # process audio of latter frame + audio_embeds_vf = rearrange(audio_embeds_vf, "bz f w b c -> (bz f) w b c") + batch_size_vf, window_size_vf, blocks_vf, channels_vf = audio_embeds_vf.shape + audio_embeds_vf = audio_embeds_vf.view(batch_size_vf, window_size_vf * blocks_vf * channels_vf) + + # first projection + audio_embeds = torch.relu(self.proj1(audio_embeds)) + audio_embeds_vf = torch.relu(self.proj1_vf(audio_embeds_vf)) + audio_embeds = rearrange(audio_embeds, "(bz f) c -> bz f c", bz=B) + audio_embeds_vf = rearrange(audio_embeds_vf, "(bz f) c -> bz f c", bz=B) + audio_embeds_c = torch.concat([audio_embeds, audio_embeds_vf], dim=1) + batch_size_c, N_t, C_a = audio_embeds_c.shape + audio_embeds_c = audio_embeds_c.view(batch_size_c*N_t, C_a) + + # second projection + audio_embeds_c = torch.relu(self.proj2(audio_embeds_c)) + + context_tokens = self.proj3(audio_embeds_c).reshape(batch_size_c*N_t, self.context_tokens, self.out_dim) + + # normalization and reshape + context_tokens = self.norm(context_tokens) + context_tokens = rearrange(context_tokens, "(bz f) m c -> bz f m c", f=video_length) + + return context_tokens + + +class WanMultiTalkAttentionBlock(torch.nn.Module): + def __init__(self, in_dim=5120, out_dim=768, device=None, dtype=None, operations=None): + super().__init__() + self.audio_cross_attn = SingleStreamMultiAttention(in_dim, out_dim, num_heads=40, qkv_bias=True, device=device, dtype=dtype, operations=operations) + self.norm_x = operations.LayerNorm(in_dim, device=device, dtype=dtype, elementwise_affine=True) + + +class MultiTalkGetAttnMapPatch: + def __init__(self, ref_target_masks=None): + self.ref_target_masks = ref_target_masks + + def __call__(self, kwargs): + transformer_options = kwargs.get("transformer_options", {}) + x = kwargs["x"] + + if self.ref_target_masks is not None: + x_ref_attn_map = get_attn_map_with_target(kwargs["q"], kwargs["k"], transformer_options["grid_sizes"], ref_target_masks=self.ref_target_masks.to(x.device)) + transformer_options["x_ref_attn_map"] = x_ref_attn_map + return x + + +class MultiTalkCrossAttnPatch: + def __init__(self, model_patch, audio_scale=1.0, ref_target_masks=None): + self.model_patch = model_patch + self.audio_scale = audio_scale + self.ref_target_masks = ref_target_masks + + def __call__(self, kwargs): + transformer_options = kwargs.get("transformer_options", {}) + block_idx = transformer_options.get("block_index", None) + x = kwargs["x"] + if block_idx is None: + return torch.zeros_like(x) + + audio_embeds = transformer_options.get("audio_embeds") + x_ref_attn_map = transformer_options.pop("x_ref_attn_map", None) + + norm_x = self.model_patch.model.blocks[block_idx].norm_x(x) + x_audio = self.model_patch.model.blocks[block_idx].audio_cross_attn( + norm_x, audio_embeds.to(x.dtype), + shape=transformer_options["grid_sizes"], + x_ref_attn_map=x_ref_attn_map + ) + x = x + x_audio * self.audio_scale + return x + + def models(self): + return [self.model_patch] + +class MultiTalkApplyModelWrapper: + def __init__(self, init_latents): + self.init_latents = init_latents + + def __call__(self, executor, x, *args, **kwargs): + x[:, :, :self.init_latents.shape[2]] = self.init_latents.to(x) + samples = executor(x, *args, **kwargs) + return samples + + +class InfiniteTalkOuterSampleWrapper: + def __init__(self, motion_frames_latent, model_patch, is_extend=False): + self.motion_frames_latent = motion_frames_latent + self.model_patch = model_patch + self.is_extend = is_extend + + def __call__(self, executor, *args, **kwargs): + model_patcher = executor.class_obj.model_patcher + model_options = executor.class_obj.model_options + process_latent_in = model_patcher.model.process_latent_in + + # for InfiniteTalk, model input first latent(s) need to always be replaced on every step + if self.motion_frames_latent is not None: + wrappers = model_options["transformer_options"]["wrappers"] + w = wrappers.setdefault(comfy.patcher_extension.WrappersMP.APPLY_MODEL, {}) + w["MultiTalk_apply_model"] = [MultiTalkApplyModelWrapper(process_latent_in(self.motion_frames_latent))] + + # run the sampling process + result = executor(*args, **kwargs) + + # insert motion frames before decoding + if self.is_extend: + overlap = self.motion_frames_latent.shape[2] + result = torch.cat([self.motion_frames_latent.to(result), result[:, :, overlap:]], dim=2) + + return result + + def to(self, device_or_dtype): + if isinstance(device_or_dtype, torch.device): + if self.motion_frames_latent is not None: + self.motion_frames_latent = self.motion_frames_latent.to(device_or_dtype) + return self diff --git a/comfy_api/latest/_io.py b/comfy_api/latest/_io.py index a60020ca8..2ec8d6e4b 100644 --- a/comfy_api/latest/_io.py +++ b/comfy_api/latest/_io.py @@ -754,7 +754,7 @@ class AnyType(ComfyTypeIO): Type = Any @comfytype(io_type="MODEL_PATCH") -class MODEL_PATCH(ComfyTypeIO): +class ModelPatch(ComfyTypeIO): Type = Any @comfytype(io_type="AUDIO_ENCODER") @@ -2038,6 +2038,7 @@ __all__ = [ "ControlNet", "Vae", "Model", + "ModelPatch", "ClipVision", "ClipVisionOutput", "AudioEncoder", diff --git a/comfy_extras/nodes_model_patch.py b/comfy_extras/nodes_model_patch.py index f66d28fc9..82c4754a3 100644 --- a/comfy_extras/nodes_model_patch.py +++ b/comfy_extras/nodes_model_patch.py @@ -7,6 +7,7 @@ import comfy.model_management import comfy.ldm.common_dit import comfy.latent_formats import comfy.ldm.lumina.controlnet +from comfy.ldm.wan.model_multitalk import WanMultiTalkAttentionBlock, MultiTalkAudioProjModel class BlockWiseControlBlock(torch.nn.Module): @@ -257,6 +258,14 @@ class ModelPatchLoader: if torch.count_nonzero(ref_weight) == 0: config['broken'] = True model = comfy.ldm.lumina.controlnet.ZImage_Control(device=comfy.model_management.unet_offload_device(), dtype=dtype, operations=comfy.ops.manual_cast, **config) + elif "audio_proj.proj1.weight" in sd: + model = MultiTalkModelPatch( + audio_window=5, context_tokens=32, vae_scale=4, + in_dim=sd["blocks.0.audio_cross_attn.proj.weight"].shape[0], + intermediate_dim=sd["audio_proj.proj1.weight"].shape[0], + out_dim=sd["audio_proj.norm.weight"].shape[0], + device=comfy.model_management.unet_offload_device(), + operations=comfy.ops.manual_cast) model.load_state_dict(sd) model = comfy.model_patcher.ModelPatcher(model, load_device=comfy.model_management.get_torch_device(), offload_device=comfy.model_management.unet_offload_device()) @@ -524,6 +533,38 @@ class USOStyleReference: return (model_patched,) +class MultiTalkModelPatch(torch.nn.Module): + def __init__( + self, + audio_window: int = 5, + intermediate_dim: int = 512, + in_dim: int = 5120, + out_dim: int = 768, + context_tokens: int = 32, + vae_scale: int = 4, + num_layers: int = 40, + + device=None, dtype=None, operations=None + ): + super().__init__() + self.audio_proj = MultiTalkAudioProjModel( + seq_len=audio_window, + seq_len_vf=audio_window+vae_scale-1, + intermediate_dim=intermediate_dim, + out_dim=out_dim, + context_tokens=context_tokens, + device=device, + dtype=dtype, + operations=operations + ) + self.blocks = torch.nn.ModuleList( + [ + WanMultiTalkAttentionBlock(in_dim, out_dim, device=device, dtype=dtype, operations=operations) + for _ in range(num_layers) + ] + ) + + NODE_CLASS_MAPPINGS = { "ModelPatchLoader": ModelPatchLoader, "QwenImageDiffsynthControlnet": QwenImageDiffsynthControlnet, diff --git a/comfy_extras/nodes_wan.py b/comfy_extras/nodes_wan.py index d32aad98e..90deb0077 100644 --- a/comfy_extras/nodes_wan.py +++ b/comfy_extras/nodes_wan.py @@ -8,9 +8,10 @@ import comfy.latent_formats import comfy.clip_vision import json import numpy as np -from typing import Tuple +from typing import Tuple, TypedDict from typing_extensions import override from comfy_api.latest import ComfyExtension, io +import logging class WanImageToVideo(io.ComfyNode): @classmethod @@ -1288,6 +1289,171 @@ class Wan22ImageToVideoLatent(io.ComfyNode): return io.NodeOutput(out_latent) +from comfy.ldm.wan.model_multitalk import InfiniteTalkOuterSampleWrapper, MultiTalkCrossAttnPatch, MultiTalkGetAttnMapPatch, project_audio_features +class WanInfiniteTalkToVideo(io.ComfyNode): + class DCValues(TypedDict): + mode: str + audio_encoder_output_2: io.AudioEncoderOutput.Type + mask: io.Mask.Type + + @classmethod + def define_schema(cls): + return io.Schema( + node_id="WanInfiniteTalkToVideo", + category="conditioning/video_models", + inputs=[ + io.DynamicCombo.Input("mode", options=[ + io.DynamicCombo.Option("single_speaker", []), + io.DynamicCombo.Option("two_speakers", [ + io.AudioEncoderOutput.Input("audio_encoder_output_2", optional=True), + io.Mask.Input("mask_1", optional=True, tooltip="Mask for the first speaker, required if using two audio inputs."), + io.Mask.Input("mask_2", optional=True, tooltip="Mask for the second speaker, required if using two audio inputs."), + ]), + ]), + io.Model.Input("model"), + io.ModelPatch.Input("model_patch"), + io.Conditioning.Input("positive"), + io.Conditioning.Input("negative"), + io.Vae.Input("vae"), + io.Int.Input("width", default=832, min=16, max=nodes.MAX_RESOLUTION, step=16), + io.Int.Input("height", default=480, min=16, max=nodes.MAX_RESOLUTION, step=16), + io.Int.Input("length", default=81, min=1, max=nodes.MAX_RESOLUTION, step=4), + io.ClipVisionOutput.Input("clip_vision_output", optional=True), + io.Image.Input("start_image", optional=True), + io.AudioEncoderOutput.Input("audio_encoder_output_1"), + io.Int.Input("motion_frame_count", default=9, min=1, max=33, step=1, tooltip="Number of previous frames to use as motion context."), + io.Float.Input("audio_scale", default=1.0, min=-10.0, max=10.0, step=0.01), + io.Image.Input("previous_frames", optional=True), + ], + outputs=[ + io.Model.Output(display_name="model"), + io.Conditioning.Output(display_name="positive"), + io.Conditioning.Output(display_name="negative"), + io.Latent.Output(display_name="latent"), + io.Int.Output(display_name="trim_image"), + ], + ) + + @classmethod + def execute(cls, mode: DCValues, model, model_patch, positive, negative, vae, width, height, length, audio_encoder_output_1, motion_frame_count, + start_image=None, previous_frames=None, audio_scale=None, clip_vision_output=None, audio_encoder_output_2=None, mask_1=None, mask_2=None) -> io.NodeOutput: + + if previous_frames is not None and previous_frames.shape[0] < motion_frame_count: + raise ValueError("Not enough previous frames provided.") + + if mode["mode"] == "two_speakers": + audio_encoder_output_2 = mode["audio_encoder_output_2"] + mask_1 = mode["mask_1"] + mask_2 = mode["mask_2"] + + if audio_encoder_output_2 is not None: + if mask_1 is None or mask_2 is None: + raise ValueError("Masks must be provided if two audio encoder outputs are used.") + + ref_masks = None + if mask_1 is not None and mask_2 is not None: + if audio_encoder_output_2 is None: + raise ValueError("Second audio encoder output must be provided if two masks are used.") + ref_masks = torch.cat([mask_1, mask_2]) + + latent = torch.zeros([1, 16, ((length - 1) // 4) + 1, height // 8, width // 8], device=comfy.model_management.intermediate_device()) + if start_image is not None: + start_image = comfy.utils.common_upscale(start_image[:length].movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1) + image = torch.ones((length, height, width, start_image.shape[-1]), device=start_image.device, dtype=start_image.dtype) * 0.5 + image[:start_image.shape[0]] = start_image + + concat_latent_image = vae.encode(image[:, :, :, :3]) + concat_mask = torch.ones((1, 1, latent.shape[2], concat_latent_image.shape[-2], concat_latent_image.shape[-1]), device=start_image.device, dtype=start_image.dtype) + concat_mask[:, :, :((start_image.shape[0] - 1) // 4) + 1] = 0.0 + + positive = node_helpers.conditioning_set_values(positive, {"concat_latent_image": concat_latent_image, "concat_mask": concat_mask}) + negative = node_helpers.conditioning_set_values(negative, {"concat_latent_image": concat_latent_image, "concat_mask": concat_mask}) + + if clip_vision_output is not None: + positive = node_helpers.conditioning_set_values(positive, {"clip_vision_output": clip_vision_output}) + negative = node_helpers.conditioning_set_values(negative, {"clip_vision_output": clip_vision_output}) + + model_patched = model.clone() + + encoded_audio_list = [] + seq_lengths = [] + + for audio_encoder_output in [audio_encoder_output_1, audio_encoder_output_2]: + if audio_encoder_output is None: + continue + all_layers = audio_encoder_output["encoded_audio_all_layers"] + encoded_audio = torch.stack(all_layers, dim=0).squeeze(1)[1:] # shape: [num_layers, T, 512] + encoded_audio = linear_interpolation(encoded_audio, input_fps=50, output_fps=25).movedim(0, 1) # shape: [T, num_layers, 512] + encoded_audio_list.append(encoded_audio) + seq_lengths.append(encoded_audio.shape[0]) + + # Pad / combine depending on multi_audio_type + multi_audio_type = "add" + if len(encoded_audio_list) > 1: + if multi_audio_type == "para": + max_len = max(seq_lengths) + padded = [] + for emb in encoded_audio_list: + if emb.shape[0] < max_len: + pad = torch.zeros(max_len - emb.shape[0], *emb.shape[1:], dtype=emb.dtype) + emb = torch.cat([emb, pad], dim=0) + padded.append(emb) + encoded_audio_list = padded + elif multi_audio_type == "add": + total_len = sum(seq_lengths) + full_list = [] + offset = 0 + for emb, seq_len in zip(encoded_audio_list, seq_lengths): + full = torch.zeros(total_len, *emb.shape[1:], dtype=emb.dtype) + full[offset:offset+seq_len] = emb + full_list.append(full) + offset += seq_len + encoded_audio_list = full_list + + token_ref_target_masks = None + if ref_masks is not None: + token_ref_target_masks = torch.nn.functional.interpolate( + ref_masks.unsqueeze(0), size=(latent.shape[-2] // 2, latent.shape[-1] // 2), mode='nearest')[0] + token_ref_target_masks = (token_ref_target_masks > 0).view(token_ref_target_masks.shape[0], -1) + + # when extending from previous frames + if previous_frames is not None: + motion_frames = comfy.utils.common_upscale(previous_frames[-motion_frame_count:].movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1) + frame_offset = previous_frames.shape[0] - motion_frame_count + + audio_start = frame_offset + audio_end = audio_start + length + logging.info(f"InfiniteTalk: Processing audio frames {audio_start} - {audio_end}") + + motion_frames_latent = vae.encode(motion_frames[:, :, :, :3]) + trim_image = motion_frame_count + else: + audio_start = trim_image = 0 + audio_end = length + motion_frames_latent = concat_latent_image[:, :, :1] + + audio_embed = project_audio_features(model_patch.model.audio_proj, encoded_audio_list, audio_start, audio_end).to(model_patched.model_dtype()) + model_patched.model_options["transformer_options"]["audio_embeds"] = audio_embed + + # add outer sample wrapper + model_patched.add_wrapper_with_key( + comfy.patcher_extension.WrappersMP.OUTER_SAMPLE, + "infinite_talk_outer_sample", + InfiniteTalkOuterSampleWrapper( + motion_frames_latent, + model_patch, + is_extend=previous_frames is not None, + )) + # add cross-attention patch + model_patched.set_model_patch(MultiTalkCrossAttnPatch(model_patch, audio_scale), "attn2_patch") + if token_ref_target_masks is not None: + model_patched.set_model_patch(MultiTalkGetAttnMapPatch(token_ref_target_masks), "attn1_patch") + + out_latent = {} + out_latent["samples"] = latent + return io.NodeOutput(model_patched, positive, negative, out_latent, trim_image) + + class WanExtension(ComfyExtension): @override async def get_node_list(self) -> list[type[io.ComfyNode]]: @@ -1307,6 +1473,7 @@ class WanExtension(ComfyExtension): WanHuMoImageToVideo, WanAnimateToVideo, Wan22ImageToVideoLatent, + WanInfiniteTalkToVideo, ] async def comfy_entrypoint() -> WanExtension: From 72f6be1690868af852a624084a949b785fc056ea Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Thu, 22 Jan 2026 09:42:04 +0200 Subject: [PATCH 496/647] chore(api-nodes): rename BriaImage and OpenAIGImage nodes (#12022) --- comfy_api_nodes/nodes_bria.py | 2 +- comfy_api_nodes/nodes_openai.py | 5 +++-- 2 files changed, 4 insertions(+), 3 deletions(-) diff --git a/comfy_api_nodes/nodes_bria.py b/comfy_api_nodes/nodes_bria.py index 72a3055a7..d3a52bc1b 100644 --- a/comfy_api_nodes/nodes_bria.py +++ b/comfy_api_nodes/nodes_bria.py @@ -24,7 +24,7 @@ class BriaImageEditNode(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="BriaImageEditNode", - display_name="Bria Image Edit", + display_name="Bria FIBO Image Edit", category="api node/image/Bria", description="Edit images using Bria latest model", inputs=[ diff --git a/comfy_api_nodes/nodes_openai.py b/comfy_api_nodes/nodes_openai.py index a12acc06b..f05aaab7b 100644 --- a/comfy_api_nodes/nodes_openai.py +++ b/comfy_api_nodes/nodes_openai.py @@ -364,9 +364,9 @@ class OpenAIGPTImage1(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="OpenAIGPTImage1", - display_name="OpenAI GPT Image 1", + display_name="OpenAI GPT Image 1.5", category="api node/image/OpenAI", - description="Generates images synchronously via OpenAI's GPT Image 1 endpoint.", + description="Generates images synchronously via OpenAI's GPT Image endpoint.", inputs=[ IO.String.Input( "prompt", @@ -429,6 +429,7 @@ class OpenAIGPTImage1(IO.ComfyNode): IO.Combo.Input( "model", options=["gpt-image-1", "gpt-image-1.5"], + default="gpt-image-1.5", optional=True, ), ], From 8490eedadfc0ab00cb131bab681059163c2ebbcd Mon Sep 17 00:00:00 2001 From: Terry Jia Date: Thu, 22 Jan 2026 12:46:56 -0500 Subject: [PATCH 497/647] add ply & 3dgs format in 3d node (#11474) --- comfy_extras/nodes_load_3d.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy_extras/nodes_load_3d.py b/comfy_extras/nodes_load_3d.py index 545588ef8..a16b8c8f3 100644 --- a/comfy_extras/nodes_load_3d.py +++ b/comfy_extras/nodes_load_3d.py @@ -24,7 +24,7 @@ class Load3D(IO.ComfyNode): files = [ normalize_path(str(file_path.relative_to(base_path))) for file_path in input_path.rglob("*") - if file_path.suffix.lower() in {'.gltf', '.glb', '.obj', '.fbx', '.stl'} + if file_path.suffix.lower() in {'.gltf', '.glb', '.obj', '.fbx', '.stl', '.spz', '.splat', '.ply', '.ksplat'} ] return IO.Schema( node_id="Load3D", From 0fd1b787360a70dd37aa14089ccb5fc1820f9e17 Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Thu, 22 Jan 2026 13:54:18 -0800 Subject: [PATCH 498/647] Reduce LTX2 VAE VRAM consumption (#12028) * causal_video_ae: Remove attention ResNet This attention_head_dim argument does not exist on this constructor so this is dead code. Remove as generic attention mid VAE conflicts with temporal roll. * ltx-vae: consoldate causal/non-causal code paths * ltx-vae: add cache rolling adder * ltx-vae: use cached adder for resnet * ltx-vae: Implement rolling VAE Implement a temporal rolling VAE for the LTX2 VAE. Usually when doing temporal rolling VAEs you can just chunk on time relying on causality and cache behind you as you go. The LTX VAE is however non-causal. So go whole hog and implement per layer run ahead and backpressure between the decoder layers using recursive state beween the layers. Operations are ammended with temporal_cache_state{} which they can use to hold any state then need for partial execution. Convolutions cache their inputs behind the up to N-1 frames, and skip connections need to cache the mismatch between convolution input and output that happens due to missing future (non-causal) input. Each call to run_up() processes a layer accross a range on input that may or may not be complete. It goes depth first to process as much as possible to try and digest frames to the final output ASAP. If layers run out of input due to convolution losses, they simply return without action effectively applying back-pressure to the earlier layers. As the earlier layers do more work and caller deeper, the partial states are reconciled and output continues to digest depth first as much as possible. Chunking is done using a size quota rather than a fixed frame length and any layer can initiate chunking, and multiple layers can chunk at different granulatiries. This remove the old limitation of always having to process 1 latent frame to entirety and having to hold 8 full decoded frames as the VRAM peak. --- comfy/ldm/lightricks/vae/causal_conv3d.py | 43 +++-- .../vae/causal_video_autoencoder.py | 176 +++++++++++++----- comfy/ldm/modules/diffusionmodules/model.py | 5 +- 3 files changed, 160 insertions(+), 64 deletions(-) diff --git a/comfy/ldm/lightricks/vae/causal_conv3d.py b/comfy/ldm/lightricks/vae/causal_conv3d.py index 70d612e86..b8341edbc 100644 --- a/comfy/ldm/lightricks/vae/causal_conv3d.py +++ b/comfy/ldm/lightricks/vae/causal_conv3d.py @@ -1,11 +1,11 @@ from typing import Tuple, Union +import threading import torch import torch.nn as nn import comfy.ops ops = comfy.ops.disable_weight_init - class CausalConv3d(nn.Module): def __init__( self, @@ -42,23 +42,34 @@ class CausalConv3d(nn.Module): padding_mode=spatial_padding_mode, groups=groups, ) + self.temporal_cache_state={} def forward(self, x, causal: bool = True): - if causal: - first_frame_pad = x[:, :, :1, :, :].repeat( - (1, 1, self.time_kernel_size - 1, 1, 1) - ) - x = torch.concatenate((first_frame_pad, x), dim=2) - else: - first_frame_pad = x[:, :, :1, :, :].repeat( - (1, 1, (self.time_kernel_size - 1) // 2, 1, 1) - ) - last_frame_pad = x[:, :, -1:, :, :].repeat( - (1, 1, (self.time_kernel_size - 1) // 2, 1, 1) - ) - x = torch.concatenate((first_frame_pad, x, last_frame_pad), dim=2) - x = self.conv(x) - return x + tid = threading.get_ident() + + cached, is_end = self.temporal_cache_state.get(tid, (None, False)) + if cached is None: + padding_length = self.time_kernel_size - 1 + if not causal: + padding_length = padding_length // 2 + if x.shape[2] == 0: + return x + cached = x[:, :, :1, :, :].repeat((1, 1, padding_length, 1, 1)) + pieces = [ cached, x ] + if is_end and not causal: + pieces.append(x[:, :, -1:, :, :].repeat((1, 1, (self.time_kernel_size - 1) // 2, 1, 1))) + + needs_caching = not is_end + if needs_caching and x.shape[2] >= self.time_kernel_size - 1: + needs_caching = False + self.temporal_cache_state[tid] = (x[:, :, -(self.time_kernel_size - 1):, :, :], False) + + x = torch.cat(pieces, dim=2) + + if needs_caching: + self.temporal_cache_state[tid] = (x[:, :, -(self.time_kernel_size - 1):, :, :], False) + + return self.conv(x) if x.shape[2] >= self.time_kernel_size else x[:, :, :0, :, :] @property def weight(self): diff --git a/comfy/ldm/lightricks/vae/causal_video_autoencoder.py b/comfy/ldm/lightricks/vae/causal_video_autoencoder.py index 75ed069ad..cbfdf412d 100644 --- a/comfy/ldm/lightricks/vae/causal_video_autoencoder.py +++ b/comfy/ldm/lightricks/vae/causal_video_autoencoder.py @@ -1,4 +1,5 @@ from __future__ import annotations +import threading import torch from torch import nn from functools import partial @@ -6,12 +7,35 @@ import math from einops import rearrange from typing import List, Optional, Tuple, Union from .conv_nd_factory import make_conv_nd, make_linear_nd +from .causal_conv3d import CausalConv3d from .pixel_norm import PixelNorm from ..model import PixArtAlphaCombinedTimestepSizeEmbeddings import comfy.ops +from comfy.ldm.modules.diffusionmodules.model import torch_cat_if_needed ops = comfy.ops.disable_weight_init +def mark_conv3d_ended(module): + tid = threading.get_ident() + for _, m in module.named_modules(): + if isinstance(m, CausalConv3d): + current = m.temporal_cache_state.get(tid, (None, False)) + m.temporal_cache_state[tid] = (current[0], True) + +def split2(tensor, split_point, dim=2): + return torch.split(tensor, [split_point, tensor.shape[dim] - split_point], dim=dim) + +def add_exchange_cache(dest, cache_in, new_input, dim=2): + if dest is not None: + if cache_in is not None: + cache_to_dest = min(dest.shape[dim], cache_in.shape[dim]) + lead_in_dest, dest = split2(dest, cache_to_dest, dim=dim) + lead_in_source, cache_in = split2(cache_in, cache_to_dest, dim=dim) + lead_in_dest.add_(lead_in_source) + body, new_input = split2(new_input, dest.shape[dim], dim) + dest.add_(body) + return torch_cat_if_needed([cache_in, new_input], dim=dim) + class Encoder(nn.Module): r""" The `Encoder` layer of a variational autoencoder that encodes its input into a latent representation. @@ -205,7 +229,7 @@ class Encoder(nn.Module): self.gradient_checkpointing = False - def forward(self, sample: torch.FloatTensor) -> torch.FloatTensor: + def forward_orig(self, sample: torch.FloatTensor) -> torch.FloatTensor: r"""The forward method of the `Encoder` class.""" sample = patchify(sample, patch_size_hw=self.patch_size, patch_size_t=1) @@ -254,6 +278,22 @@ class Encoder(nn.Module): return sample + def forward(self, *args, **kwargs): + #No encoder support so just flag the end so it doesnt use the cache. + mark_conv3d_ended(self) + try: + return self.forward_orig(*args, **kwargs) + finally: + tid = threading.get_ident() + for _, module in self.named_modules(): + # ComfyUI doesn't thread this kind of stuff today, but just in case + # we key on the thread to make it thread safe. + tid = threading.get_ident() + if hasattr(module, "temporal_cache_state"): + module.temporal_cache_state.pop(tid, None) + + +MAX_CHUNK_SIZE=(128 * 1024 ** 2) class Decoder(nn.Module): r""" @@ -341,18 +381,6 @@ class Decoder(nn.Module): timestep_conditioning=timestep_conditioning, spatial_padding_mode=spatial_padding_mode, ) - elif block_name == "attn_res_x": - block = UNetMidBlock3D( - dims=dims, - in_channels=input_channel, - num_layers=block_params["num_layers"], - resnet_groups=norm_num_groups, - norm_layer=norm_layer, - inject_noise=block_params.get("inject_noise", False), - timestep_conditioning=timestep_conditioning, - attention_head_dim=block_params["attention_head_dim"], - spatial_padding_mode=spatial_padding_mode, - ) elif block_name == "res_x_y": output_channel = output_channel // block_params.get("multiplier", 2) block = ResnetBlock3D( @@ -428,8 +456,9 @@ class Decoder(nn.Module): ) self.last_scale_shift_table = nn.Parameter(torch.empty(2, output_channel)) + # def forward(self, sample: torch.FloatTensor, target_shape) -> torch.FloatTensor: - def forward( + def forward_orig( self, sample: torch.FloatTensor, timestep: Optional[torch.Tensor] = None, @@ -437,6 +466,7 @@ class Decoder(nn.Module): r"""The forward method of the `Decoder` class.""" batch_size = sample.shape[0] + mark_conv3d_ended(self.conv_in) sample = self.conv_in(sample, causal=self.causal) checkpoint_fn = ( @@ -445,24 +475,12 @@ class Decoder(nn.Module): else lambda x: x ) - scaled_timestep = None + timestep_shift_scale = None if self.timestep_conditioning: assert ( timestep is not None ), "should pass timestep with timestep_conditioning=True" scaled_timestep = timestep * self.timestep_scale_multiplier.to(dtype=sample.dtype, device=sample.device) - - for up_block in self.up_blocks: - if self.timestep_conditioning and isinstance(up_block, UNetMidBlock3D): - sample = checkpoint_fn(up_block)( - sample, causal=self.causal, timestep=scaled_timestep - ) - else: - sample = checkpoint_fn(up_block)(sample, causal=self.causal) - - sample = self.conv_norm_out(sample) - - if self.timestep_conditioning: embedded_timestep = self.last_time_embedder( timestep=scaled_timestep.flatten(), resolution=None, @@ -483,16 +501,62 @@ class Decoder(nn.Module): embedded_timestep.shape[-2], embedded_timestep.shape[-1], ) - shift, scale = ada_values.unbind(dim=1) - sample = sample * (1 + scale) + shift + timestep_shift_scale = ada_values.unbind(dim=1) - sample = self.conv_act(sample) - sample = self.conv_out(sample, causal=self.causal) + output = [] + + def run_up(idx, sample, ended): + if idx >= len(self.up_blocks): + sample = self.conv_norm_out(sample) + if timestep_shift_scale is not None: + shift, scale = timestep_shift_scale + sample = sample * (1 + scale) + shift + sample = self.conv_act(sample) + if ended: + mark_conv3d_ended(self.conv_out) + sample = self.conv_out(sample, causal=self.causal) + if sample is not None and sample.shape[2] > 0: + output.append(sample) + return + + up_block = self.up_blocks[idx] + if (ended): + mark_conv3d_ended(up_block) + if self.timestep_conditioning and isinstance(up_block, UNetMidBlock3D): + sample = checkpoint_fn(up_block)( + sample, causal=self.causal, timestep=scaled_timestep + ) + else: + sample = checkpoint_fn(up_block)(sample, causal=self.causal) + + if sample is None or sample.shape[2] == 0: + return + + total_bytes = sample.numel() * sample.element_size() + num_chunks = (total_bytes + MAX_CHUNK_SIZE - 1) // MAX_CHUNK_SIZE + samples = torch.chunk(sample, chunks=num_chunks, dim=2) + + for chunk_idx, sample1 in enumerate(samples): + run_up(idx + 1, sample1, ended and chunk_idx == len(samples) - 1) + + run_up(0, sample, True) + sample = torch.cat(output, dim=2) sample = unpatchify(sample, patch_size_hw=self.patch_size, patch_size_t=1) return sample + def forward(self, *args, **kwargs): + try: + return self.forward_orig(*args, **kwargs) + finally: + for _, module in self.named_modules(): + #ComfyUI doesn't thread this kind of stuff today, but just incase + #we key on the thread to make it thread safe. + tid = threading.get_ident() + if hasattr(module, "temporal_cache_state"): + module.temporal_cache_state.pop(tid, None) + class UNetMidBlock3D(nn.Module): """ @@ -663,8 +727,22 @@ class DepthToSpaceUpsample(nn.Module): ) self.residual = residual self.out_channels_reduction_factor = out_channels_reduction_factor + self.temporal_cache_state = {} def forward(self, x, causal: bool = True, timestep: Optional[torch.Tensor] = None): + tid = threading.get_ident() + cached, drop_first_conv, drop_first_res = self.temporal_cache_state.get(tid, (None, True, True)) + y = self.conv(x, causal=causal) + y = rearrange( + y, + "b (c p1 p2 p3) d h w -> b c (d p1) (h p2) (w p3)", + p1=self.stride[0], + p2=self.stride[1], + p3=self.stride[2], + ) + if self.stride[0] == 2 and y.shape[2] > 0 and drop_first_conv: + y = y[:, :, 1:, :, :] + drop_first_conv = False if self.residual: # Reshape and duplicate the input to match the output shape x_in = rearrange( @@ -676,21 +754,20 @@ class DepthToSpaceUpsample(nn.Module): ) num_repeat = math.prod(self.stride) // self.out_channels_reduction_factor x_in = x_in.repeat(1, num_repeat, 1, 1, 1) - if self.stride[0] == 2: + if self.stride[0] == 2 and x_in.shape[2] > 0 and drop_first_res: x_in = x_in[:, :, 1:, :, :] - x = self.conv(x, causal=causal) - x = rearrange( - x, - "b (c p1 p2 p3) d h w -> b c (d p1) (h p2) (w p3)", - p1=self.stride[0], - p2=self.stride[1], - p3=self.stride[2], - ) - if self.stride[0] == 2: - x = x[:, :, 1:, :, :] - if self.residual: - x = x + x_in - return x + drop_first_res = False + + if y.shape[2] == 0: + y = None + + cached = add_exchange_cache(y, cached, x_in, dim=2) + self.temporal_cache_state[tid] = (cached, drop_first_conv, drop_first_res) + + else: + self.temporal_cache_state[tid] = (None, drop_first_conv, False) + + return y class LayerNorm(nn.Module): def __init__(self, dim, eps, elementwise_affine=True) -> None: @@ -807,6 +884,8 @@ class ResnetBlock3D(nn.Module): torch.randn(4, in_channels) / in_channels**0.5 ) + self.temporal_cache_state={} + def _feed_spatial_noise( self, hidden_states: torch.FloatTensor, per_channel_scale: torch.FloatTensor ) -> torch.FloatTensor: @@ -880,9 +959,12 @@ class ResnetBlock3D(nn.Module): input_tensor = self.conv_shortcut(input_tensor) - output_tensor = input_tensor + hidden_states + tid = threading.get_ident() + cached = self.temporal_cache_state.get(tid, None) + cached = add_exchange_cache(hidden_states, cached, input_tensor, dim=2) + self.temporal_cache_state[tid] = cached - return output_tensor + return hidden_states def patchify(x, patch_size_hw, patch_size_t=1): diff --git a/comfy/ldm/modules/diffusionmodules/model.py b/comfy/ldm/modules/diffusionmodules/model.py index 1ae3ef034..5a22ef030 100644 --- a/comfy/ldm/modules/diffusionmodules/model.py +++ b/comfy/ldm/modules/diffusionmodules/model.py @@ -14,10 +14,13 @@ if model_management.xformers_enabled_vae(): import xformers.ops def torch_cat_if_needed(xl, dim): + xl = [x for x in xl if x is not None and x.shape[dim] > 0] if len(xl) > 1: return torch.cat(xl, dim) - else: + elif len(xl) == 1: return xl[0] + else: + return None def get_timestep_embedding(timesteps, embedding_dim): """ From 09a2e67151c6753a0038f6e01f3c3d93fcc3ec98 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Thu, 22 Jan 2026 15:20:48 -0800 Subject: [PATCH 499/647] Support loading flux 2 klein checkpoints saved with SaveCheckpoint. (#12033) --- comfy/supported_models.py | 20 +++++++++++++++++--- comfy/text_encoders/hunyuan_video.py | 8 +++++--- 2 files changed, 22 insertions(+), 6 deletions(-) diff --git a/comfy/supported_models.py b/comfy/supported_models.py index 70abebf46..45d913fa6 100644 --- a/comfy/supported_models.py +++ b/comfy/supported_models.py @@ -771,10 +771,24 @@ class Flux2(Flux): return out def clip_target(self, state_dict={}): - return None # TODO pref = self.text_encoder_key_prefix[0] - t5_detect = comfy.text_encoders.sd3_clip.t5_xxl_detect(state_dict, "{}t5xxl.transformer.".format(pref)) - return supported_models_base.ClipTarget(comfy.text_encoders.flux.FluxTokenizer, comfy.text_encoders.flux.flux_clip(**t5_detect)) + detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen3_4b.transformer.".format(pref)) + if len(detect) > 0: + detect["model_type"] = "qwen3_4b" + return supported_models_base.ClipTarget(comfy.text_encoders.flux.KleinTokenizer, comfy.text_encoders.flux.klein_te(**detect)) + + detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen3_8b.transformer.".format(pref)) + if len(detect) > 0: + detect["model_type"] = "qwen3_8b" + return supported_models_base.ClipTarget(comfy.text_encoders.flux.KleinTokenizer8B, comfy.text_encoders.flux.klein_te(**detect)) + + detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}mistral3_24b.transformer.".format(pref)) + if len(detect) > 0: + if "{}mistral3_24b.transformer.model.layers.39.post_attention_layernorm.weight".format(pref) not in state_dict: + detect["pruned"] = True + return supported_models_base.ClipTarget(comfy.text_encoders.flux.Flux2Tokenizer, comfy.text_encoders.flux.flux2_te(**detect)) + + return None class GenmoMochi(supported_models_base.BASE): unet_config = { diff --git a/comfy/text_encoders/hunyuan_video.py b/comfy/text_encoders/hunyuan_video.py index a9a6c525e..2ddb4da60 100644 --- a/comfy/text_encoders/hunyuan_video.py +++ b/comfy/text_encoders/hunyuan_video.py @@ -10,9 +10,11 @@ import comfy.utils def llama_detect(state_dict, prefix=""): out = {} - t5_key = "{}model.norm.weight".format(prefix) - if t5_key in state_dict: - out["dtype_llama"] = state_dict[t5_key].dtype + norm_keys = ["{}model.norm.weight".format(prefix), "{}model.layers.0.input_layernorm.weight".format(prefix)] + for norm_key in norm_keys: + if norm_key in state_dict: + out["dtype_llama"] = state_dict[norm_key].dtype + break quant = comfy.utils.detect_layer_quantization(state_dict, prefix) if quant is not None: From d7f3241bf6b11f67ada34c51097fbaad0c01124a Mon Sep 17 00:00:00 2001 From: Omri Marom <110098005+maromri@users.noreply.github.com> Date: Fri, 23 Jan 2026 03:02:31 +0200 Subject: [PATCH 500/647] qwen_image: propagate attention mask. (#11966) --- comfy/ldm/qwen_image/model.py | 11 ++++++++++- comfy/model_base.py | 3 +++ 2 files changed, 13 insertions(+), 1 deletion(-) diff --git a/comfy/ldm/qwen_image/model.py b/comfy/ldm/qwen_image/model.py index 00c597535..6eb744286 100644 --- a/comfy/ldm/qwen_image/model.py +++ b/comfy/ldm/qwen_image/model.py @@ -170,8 +170,14 @@ class Attention(nn.Module): joint_query = apply_rope1(joint_query, image_rotary_emb) joint_key = apply_rope1(joint_key, image_rotary_emb) + if encoder_hidden_states_mask is not None: + attn_mask = torch.zeros((batch_size, 1, seq_txt + seq_img), dtype=hidden_states.dtype, device=hidden_states.device) + attn_mask[:, 0, :seq_txt] = encoder_hidden_states_mask + else: + attn_mask = None + joint_hidden_states = optimized_attention_masked(joint_query, joint_key, joint_value, self.heads, - attention_mask, transformer_options=transformer_options, + attn_mask, transformer_options=transformer_options, skip_reshape=True) txt_attn_output = joint_hidden_states[:, :seq_txt, :] @@ -430,6 +436,9 @@ class QwenImageTransformer2DModel(nn.Module): encoder_hidden_states = context encoder_hidden_states_mask = attention_mask + if encoder_hidden_states_mask is not None and not torch.is_floating_point(encoder_hidden_states_mask): + encoder_hidden_states_mask = (encoder_hidden_states_mask - 1).to(x.dtype) * torch.finfo(x.dtype).max + hidden_states, img_ids, orig_shape = self.process_img(x) num_embeds = hidden_states.shape[1] diff --git a/comfy/model_base.py b/comfy/model_base.py index 1d57562cc..66e52864d 100644 --- a/comfy/model_base.py +++ b/comfy/model_base.py @@ -1578,6 +1578,9 @@ class QwenImage(BaseModel): def extra_conds(self, **kwargs): out = super().extra_conds(**kwargs) + attention_mask = kwargs.get("attention_mask", None) + if attention_mask is not None: + out['attention_mask'] = comfy.conds.CONDRegular(attention_mask) cross_attn = kwargs.get("cross_attn", None) if cross_attn is not None: out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn) From bbb8864778a93eb0fa60c76201383e2b5a63aa38 Mon Sep 17 00:00:00 2001 From: Christian Byrne Date: Thu, 22 Jan 2026 18:36:58 -0800 Subject: [PATCH 501/647] add search aliases to all nodes (#12035) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit * feat: Add search_aliases field to node schema Adds `search_aliases` field to improve node discoverability. Users can define alternative search terms for nodes (e.g., "text concat" → StringConcatenate). Changes: - Add `search_aliases: list[str]` to V3 Schema - Add `SEARCH_ALIASES` support for V1 nodes - Include field in `/object_info` response - Add aliases to high-priority core nodes V1 usage: ```python class MyNode: SEARCH_ALIASES = ["alt name", "synonym"] ``` V3 usage: ```python io.Schema( node_id="MyNode", search_aliases=["alt name", "synonym"], ... ) ``` ## Related PRs - Frontend: Comfy-Org/ComfyUI_frontend#XXXX (draft - merge after this) - Docs: Comfy-Org/docs#XXXX (draft - merge after stable) * Propagate search_aliases through V3 Schema.get_v1_info to NodeInfoV1 * feat: add SEARCH_ALIASES for core nodes (#12016) Add search aliases to 22 core nodes in nodes.py to improve node discoverability: - Checkpoint/model loaders: CheckpointLoader, DiffusersLoader - Conditioning nodes: ConditioningAverage, ConditioningSetArea, ConditioningSetMask, ConditioningZeroOut - Style nodes: StyleModelApply - Image nodes: LoadImageMask, LoadImageOutput, ImageBatch, ImageInvert, ImagePadForOutpaint - Latent nodes: LoadLatent, SaveLatent, LatentBlend, LatentComposite, LatentCrop, LatentFlip, LatentFromBatch, LatentUpscale, LatentUpscaleBy, RepeatLatentBatch * feat: add SEARCH_ALIASES for image, mask, and string nodes (#12017) Add search aliases to nodes in comfy_extras for better discoverability: - nodes_mask.py: mask manipulation nodes - nodes_images.py: image processing nodes - nodes_post_processing.py: post-processing effect nodes - nodes_string.py: string manipulation nodes - nodes_compositing.py: compositing nodes - nodes_morphology.py: morphological operation nodes - nodes_latent.py: latent space nodes Uses search_aliases parameter in io.Schema() for v3 nodes. * feat: add SEARCH_ALIASES for audio and video nodes (#12018) Add search aliases to audio and video nodes for better discoverability: - nodes_audio.py: audio loading, saving, and processing nodes - nodes_video.py: video loading and processing nodes - nodes_wan.py: WAN model nodes Uses search_aliases parameter in io.Schema() for v3 nodes. * feat: add SEARCH_ALIASES for model and misc nodes (#12019) Add search aliases to model-related and miscellaneous nodes: - Model nodes: nodes_model_merging.py, nodes_model_advanced.py, nodes_lora_extract.py - Sampler nodes: nodes_custom_sampler.py, nodes_align_your_steps.py - Control nodes: nodes_controlnet.py, nodes_attention_multiply.py, nodes_hooks.py - Training nodes: nodes_train.py, nodes_dataset.py - Utility nodes: nodes_logic.py, nodes_canny.py, nodes_differential_diffusion.py - Architecture-specific: nodes_sd3.py, nodes_pixart.py, nodes_lumina2.py, nodes_kandinsky5.py, nodes_hidream.py, nodes_fresca.py, nodes_hunyuan3d.py - Media nodes: nodes_load_3d.py, nodes_webcam.py, nodes_preview_any.py, nodes_wanmove.py Uses search_aliases parameter in io.Schema() for v3 nodes, SEARCH_ALIASES class attribute for legacy nodes. --- comfy_extras/nodes_align_your_steps.py | 1 + comfy_extras/nodes_attention_multiply.py | 1 + comfy_extras/nodes_audio.py | 14 +++++++ comfy_extras/nodes_canny.py | 1 + comfy_extras/nodes_compositing.py | 3 ++ comfy_extras/nodes_controlnet.py | 1 + comfy_extras/nodes_custom_sampler.py | 4 ++ comfy_extras/nodes_dataset.py | 6 +-- comfy_extras/nodes_differential_diffusion.py | 1 + comfy_extras/nodes_fresca.py | 1 + comfy_extras/nodes_hidream.py | 1 + comfy_extras/nodes_hooks.py | 5 +++ comfy_extras/nodes_hunyuan3d.py | 1 + comfy_extras/nodes_images.py | 16 +++++--- comfy_extras/nodes_kandinsky5.py | 1 + comfy_extras/nodes_latent.py | 10 +++++ comfy_extras/nodes_load_3d.py | 1 + comfy_extras/nodes_logic.py | 2 + comfy_extras/nodes_lora_extract.py | 1 + comfy_extras/nodes_lumina2.py | 1 + comfy_extras/nodes_mask.py | 12 ++++++ comfy_extras/nodes_model_advanced.py | 1 + comfy_extras/nodes_model_merging.py | 4 ++ comfy_extras/nodes_morphology.py | 3 ++ comfy_extras/nodes_pixart.py | 1 + comfy_extras/nodes_post_processing.py | 5 ++- comfy_extras/nodes_preview_any.py | 2 +- comfy_extras/nodes_sd3.py | 1 + comfy_extras/nodes_string.py | 10 +++++ comfy_extras/nodes_train.py | 2 + comfy_extras/nodes_video.py | 5 +++ comfy_extras/nodes_wan.py | 2 + comfy_extras/nodes_wanmove.py | 1 + comfy_extras/nodes_webcam.py | 1 + nodes.py | 41 ++++++++++++++++++++ 35 files changed, 152 insertions(+), 11 deletions(-) diff --git a/comfy_extras/nodes_align_your_steps.py b/comfy_extras/nodes_align_your_steps.py index edd5dadd4..4fc511d2c 100644 --- a/comfy_extras/nodes_align_your_steps.py +++ b/comfy_extras/nodes_align_your_steps.py @@ -28,6 +28,7 @@ class AlignYourStepsScheduler(io.ComfyNode): def define_schema(cls) -> io.Schema: return io.Schema( node_id="AlignYourStepsScheduler", + search_aliases=["AYS scheduler"], category="sampling/custom_sampling/schedulers", inputs=[ io.Combo.Input("model_type", options=["SD1", "SDXL", "SVD"]), diff --git a/comfy_extras/nodes_attention_multiply.py b/comfy_extras/nodes_attention_multiply.py index c0e494c2a..67c4e2ed0 100644 --- a/comfy_extras/nodes_attention_multiply.py +++ b/comfy_extras/nodes_attention_multiply.py @@ -71,6 +71,7 @@ class CLIPAttentionMultiply(io.ComfyNode): def define_schema(cls) -> io.Schema: return io.Schema( node_id="CLIPAttentionMultiply", + search_aliases=["clip attention scale", "text encoder attention"], category="_for_testing/attention_experiments", inputs=[ io.Clip.Input("clip"), diff --git a/comfy_extras/nodes_audio.py b/comfy_extras/nodes_audio.py index 15b3aa401..271b75fbd 100644 --- a/comfy_extras/nodes_audio.py +++ b/comfy_extras/nodes_audio.py @@ -69,6 +69,7 @@ class VAEEncodeAudio(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="VAEEncodeAudio", + search_aliases=["audio to latent"], display_name="VAE Encode Audio", category="latent/audio", inputs=[ @@ -97,6 +98,7 @@ class VAEDecodeAudio(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="VAEDecodeAudio", + search_aliases=["latent to audio"], display_name="VAE Decode Audio", category="latent/audio", inputs=[ @@ -122,6 +124,7 @@ class SaveAudio(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="SaveAudio", + search_aliases=["export flac"], display_name="Save Audio (FLAC)", category="audio", inputs=[ @@ -146,6 +149,7 @@ class SaveAudioMP3(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="SaveAudioMP3", + search_aliases=["export mp3"], display_name="Save Audio (MP3)", category="audio", inputs=[ @@ -173,6 +177,7 @@ class SaveAudioOpus(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="SaveAudioOpus", + search_aliases=["export opus"], display_name="Save Audio (Opus)", category="audio", inputs=[ @@ -200,6 +205,7 @@ class PreviewAudio(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="PreviewAudio", + search_aliases=["play audio"], display_name="Preview Audio", category="audio", inputs=[ @@ -259,6 +265,7 @@ class LoadAudio(IO.ComfyNode): files = folder_paths.filter_files_content_types(os.listdir(input_dir), ["audio", "video"]) return IO.Schema( node_id="LoadAudio", + search_aliases=["import audio", "open audio", "audio file"], display_name="Load Audio", category="audio", inputs=[ @@ -296,6 +303,7 @@ class RecordAudio(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="RecordAudio", + search_aliases=["microphone input", "audio capture", "voice input"], display_name="Record Audio", category="audio", inputs=[ @@ -320,6 +328,7 @@ class TrimAudioDuration(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="TrimAudioDuration", + search_aliases=["cut audio", "audio clip", "shorten audio"], display_name="Trim Audio Duration", description="Trim audio tensor into chosen time range.", category="audio", @@ -372,6 +381,7 @@ class SplitAudioChannels(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="SplitAudioChannels", + search_aliases=["stereo to mono"], display_name="Split Audio Channels", description="Separates the audio into left and right channels.", category="audio", @@ -472,6 +482,7 @@ class AudioConcat(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="AudioConcat", + search_aliases=["join audio", "combine audio", "append audio"], display_name="Audio Concat", description="Concatenates the audio1 to audio2 in the specified direction.", category="audio", @@ -519,6 +530,7 @@ class AudioMerge(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="AudioMerge", + search_aliases=["mix audio", "overlay audio", "layer audio"], display_name="Audio Merge", description="Combine two audio tracks by overlaying their waveforms.", category="audio", @@ -579,6 +591,7 @@ class AudioAdjustVolume(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="AudioAdjustVolume", + search_aliases=["audio gain", "loudness", "audio level"], display_name="Audio Adjust Volume", category="audio", inputs=[ @@ -614,6 +627,7 @@ class EmptyAudio(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="EmptyAudio", + search_aliases=["blank audio"], display_name="Empty Audio", category="audio", inputs=[ diff --git a/comfy_extras/nodes_canny.py b/comfy_extras/nodes_canny.py index 576f3640a..6e0fadca5 100644 --- a/comfy_extras/nodes_canny.py +++ b/comfy_extras/nodes_canny.py @@ -10,6 +10,7 @@ class Canny(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="Canny", + search_aliases=["edge detection", "outline", "contour detection", "line art"], category="image/preprocessors", inputs=[ io.Image.Input("image"), diff --git a/comfy_extras/nodes_compositing.py b/comfy_extras/nodes_compositing.py index e4e4e1cbc..3bc9fccb3 100644 --- a/comfy_extras/nodes_compositing.py +++ b/comfy_extras/nodes_compositing.py @@ -109,6 +109,7 @@ class PorterDuffImageComposite(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="PorterDuffImageComposite", + search_aliases=["alpha composite", "blend modes", "layer blend", "transparency blend"], display_name="Porter-Duff Image Composite", category="mask/compositing", inputs=[ @@ -165,6 +166,7 @@ class SplitImageWithAlpha(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="SplitImageWithAlpha", + search_aliases=["extract alpha", "separate transparency", "remove alpha"], display_name="Split Image with Alpha", category="mask/compositing", inputs=[ @@ -188,6 +190,7 @@ class JoinImageWithAlpha(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="JoinImageWithAlpha", + search_aliases=["add transparency", "apply alpha", "composite alpha", "RGBA"], display_name="Join Image with Alpha", category="mask/compositing", inputs=[ diff --git a/comfy_extras/nodes_controlnet.py b/comfy_extras/nodes_controlnet.py index e835feed7..0c1d7f0d4 100644 --- a/comfy_extras/nodes_controlnet.py +++ b/comfy_extras/nodes_controlnet.py @@ -38,6 +38,7 @@ class ControlNetInpaintingAliMamaApply(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="ControlNetInpaintingAliMamaApply", + search_aliases=["masked controlnet"], category="conditioning/controlnet", inputs=[ io.Conditioning.Input("positive"), diff --git a/comfy_extras/nodes_custom_sampler.py b/comfy_extras/nodes_custom_sampler.py index f19adf4b9..3eb40e937 100644 --- a/comfy_extras/nodes_custom_sampler.py +++ b/comfy_extras/nodes_custom_sampler.py @@ -297,6 +297,7 @@ class ExtendIntermediateSigmas(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="ExtendIntermediateSigmas", + search_aliases=["interpolate sigmas"], category="sampling/custom_sampling/sigmas", inputs=[ io.Sigmas.Input("sigmas"), @@ -856,6 +857,7 @@ class DualCFGGuider(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="DualCFGGuider", + search_aliases=["dual prompt guidance"], category="sampling/custom_sampling/guiders", inputs=[ io.Model.Input("model"), @@ -883,6 +885,7 @@ class DisableNoise(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="DisableNoise", + search_aliases=["zero noise"], category="sampling/custom_sampling/noise", inputs=[], outputs=[io.Noise.Output()] @@ -1019,6 +1022,7 @@ class ManualSigmas(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="ManualSigmas", + search_aliases=["custom noise schedule", "define sigmas"], category="_for_testing/custom_sampling", is_experimental=True, inputs=[ diff --git a/comfy_extras/nodes_dataset.py b/comfy_extras/nodes_dataset.py index 5ef851bd0..fb9409ac3 100644 --- a/comfy_extras/nodes_dataset.py +++ b/comfy_extras/nodes_dataset.py @@ -1223,11 +1223,11 @@ class ResolutionBucket(io.ComfyNode): class MakeTrainingDataset(io.ComfyNode): """Encode images with VAE and texts with CLIP to create a training dataset.""" - @classmethod def define_schema(cls): return io.Schema( node_id="MakeTrainingDataset", + search_aliases=["encode dataset"], display_name="Make Training Dataset", category="dataset", is_experimental=True, @@ -1309,11 +1309,11 @@ class MakeTrainingDataset(io.ComfyNode): class SaveTrainingDataset(io.ComfyNode): """Save encoded training dataset (latents + conditioning) to disk.""" - @classmethod def define_schema(cls): return io.Schema( node_id="SaveTrainingDataset", + search_aliases=["export training data"], display_name="Save Training Dataset", category="dataset", is_experimental=True, @@ -1410,11 +1410,11 @@ class SaveTrainingDataset(io.ComfyNode): class LoadTrainingDataset(io.ComfyNode): """Load encoded training dataset from disk.""" - @classmethod def define_schema(cls): return io.Schema( node_id="LoadTrainingDataset", + search_aliases=["import dataset", "training data"], display_name="Load Training Dataset", category="dataset", is_experimental=True, diff --git a/comfy_extras/nodes_differential_diffusion.py b/comfy_extras/nodes_differential_diffusion.py index 6dfdf466c..34ffb9a89 100644 --- a/comfy_extras/nodes_differential_diffusion.py +++ b/comfy_extras/nodes_differential_diffusion.py @@ -11,6 +11,7 @@ class DifferentialDiffusion(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="DifferentialDiffusion", + search_aliases=["inpaint gradient", "variable denoise strength"], display_name="Differential Diffusion", category="_for_testing", inputs=[ diff --git a/comfy_extras/nodes_fresca.py b/comfy_extras/nodes_fresca.py index f308eb0c1..3d590af4b 100644 --- a/comfy_extras/nodes_fresca.py +++ b/comfy_extras/nodes_fresca.py @@ -58,6 +58,7 @@ class FreSca(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="FreSca", + search_aliases=["frequency guidance"], display_name="FreSca", category="_for_testing", description="Applies frequency-dependent scaling to the guidance", diff --git a/comfy_extras/nodes_hidream.py b/comfy_extras/nodes_hidream.py index eee683ee1..e345fe51d 100644 --- a/comfy_extras/nodes_hidream.py +++ b/comfy_extras/nodes_hidream.py @@ -38,6 +38,7 @@ class CLIPTextEncodeHiDream(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="CLIPTextEncodeHiDream", + search_aliases=["hidream prompt"], category="advanced/conditioning", inputs=[ io.Clip.Input("clip"), diff --git a/comfy_extras/nodes_hooks.py b/comfy_extras/nodes_hooks.py index 1edc06f3d..58e511ef5 100644 --- a/comfy_extras/nodes_hooks.py +++ b/comfy_extras/nodes_hooks.py @@ -259,6 +259,7 @@ class SetClipHooks: return (clip,) class ConditioningTimestepsRange: + SEARCH_ALIASES = ["prompt scheduling", "timestep segments", "conditioning phases"] NodeId = 'ConditioningTimestepsRange' NodeName = 'Timesteps Range' @classmethod @@ -468,6 +469,7 @@ class SetHookKeyframes: return (hooks,) class CreateHookKeyframe: + SEARCH_ALIASES = ["hook scheduling", "strength animation", "timed hook"] NodeId = 'CreateHookKeyframe' NodeName = 'Create Hook Keyframe' @classmethod @@ -497,6 +499,7 @@ class CreateHookKeyframe: return (prev_hook_kf,) class CreateHookKeyframesInterpolated: + SEARCH_ALIASES = ["ease hook strength", "smooth hook transition", "interpolate keyframes"] NodeId = 'CreateHookKeyframesInterpolated' NodeName = 'Create Hook Keyframes Interp.' @classmethod @@ -544,6 +547,7 @@ class CreateHookKeyframesInterpolated: return (prev_hook_kf,) class CreateHookKeyframesFromFloats: + SEARCH_ALIASES = ["batch keyframes", "strength list to keyframes"] NodeId = 'CreateHookKeyframesFromFloats' NodeName = 'Create Hook Keyframes From Floats' @classmethod @@ -618,6 +622,7 @@ class SetModelHooksOnCond: # Combine Hooks #------------------------------------------ class CombineHooks: + SEARCH_ALIASES = ["merge hooks"] NodeId = 'CombineHooks2' NodeName = 'Combine Hooks [2]' @classmethod diff --git a/comfy_extras/nodes_hunyuan3d.py b/comfy_extras/nodes_hunyuan3d.py index adca14f62..5bb5df48e 100644 --- a/comfy_extras/nodes_hunyuan3d.py +++ b/comfy_extras/nodes_hunyuan3d.py @@ -618,6 +618,7 @@ class SaveGLB(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="SaveGLB", + search_aliases=["export 3d model", "save mesh"], category="3d", is_output_node=True, inputs=[ diff --git a/comfy_extras/nodes_images.py b/comfy_extras/nodes_images.py index ce21caade..cb4fb24a1 100644 --- a/comfy_extras/nodes_images.py +++ b/comfy_extras/nodes_images.py @@ -22,6 +22,7 @@ class ImageCrop(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="ImageCrop", + search_aliases=["trim"], display_name="Image Crop", category="image/transform", inputs=[ @@ -51,6 +52,7 @@ class RepeatImageBatch(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="RepeatImageBatch", + search_aliases=["duplicate image", "clone image"], category="image/batch", inputs=[ IO.Image.Input("image"), @@ -72,6 +74,7 @@ class ImageFromBatch(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="ImageFromBatch", + search_aliases=["select image", "pick from batch", "extract image"], category="image/batch", inputs=[ IO.Image.Input("image"), @@ -97,6 +100,7 @@ class ImageAddNoise(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="ImageAddNoise", + search_aliases=["film grain"], category="image", inputs=[ IO.Image.Input("image"), @@ -194,11 +198,11 @@ class SaveAnimatedPNG(IO.ComfyNode): class ImageStitch(IO.ComfyNode): """Upstreamed from https://github.com/kijai/ComfyUI-KJNodes""" - @classmethod def define_schema(cls): return IO.Schema( node_id="ImageStitch", + search_aliases=["combine images", "join images", "concatenate images", "side by side"], display_name="Image Stitch", description="Stitches image2 to image1 in the specified direction.\n" "If image2 is not provided, returns image1 unchanged.\n" @@ -369,11 +373,11 @@ class ImageStitch(IO.ComfyNode): class ResizeAndPadImage(IO.ComfyNode): - @classmethod def define_schema(cls): return IO.Schema( node_id="ResizeAndPadImage", + search_aliases=["fit to size"], category="image/transform", inputs=[ IO.Image.Input("image"), @@ -420,11 +424,11 @@ class ResizeAndPadImage(IO.ComfyNode): class SaveSVGNode(IO.ComfyNode): - @classmethod def define_schema(cls): return IO.Schema( node_id="SaveSVGNode", + search_aliases=["export vector", "save vector graphics"], description="Save SVG files on disk.", category="image/save", inputs=[ @@ -492,11 +496,11 @@ class SaveSVGNode(IO.ComfyNode): class GetImageSize(IO.ComfyNode): - @classmethod def define_schema(cls): return IO.Schema( node_id="GetImageSize", + search_aliases=["dimensions", "resolution", "image info"], display_name="Get Image Size", description="Returns width and height of the image, and passes it through unchanged.", category="image", @@ -527,11 +531,11 @@ class GetImageSize(IO.ComfyNode): class ImageRotate(IO.ComfyNode): - @classmethod def define_schema(cls): return IO.Schema( node_id="ImageRotate", + search_aliases=["turn", "flip orientation"], category="image/transform", inputs=[ IO.Image.Input("image"), @@ -557,11 +561,11 @@ class ImageRotate(IO.ComfyNode): class ImageFlip(IO.ComfyNode): - @classmethod def define_schema(cls): return IO.Schema( node_id="ImageFlip", + search_aliases=["mirror", "reflect"], category="image/transform", inputs=[ IO.Image.Input("image"), diff --git a/comfy_extras/nodes_kandinsky5.py b/comfy_extras/nodes_kandinsky5.py index 9cb234be1..346c50cde 100644 --- a/comfy_extras/nodes_kandinsky5.py +++ b/comfy_extras/nodes_kandinsky5.py @@ -104,6 +104,7 @@ class CLIPTextEncodeKandinsky5(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="CLIPTextEncodeKandinsky5", + search_aliases=["kandinsky prompt"], category="advanced/conditioning/kandinsky5", inputs=[ io.Clip.Input("clip"), diff --git a/comfy_extras/nodes_latent.py b/comfy_extras/nodes_latent.py index 9ba1c4ba8..6aecf1561 100644 --- a/comfy_extras/nodes_latent.py +++ b/comfy_extras/nodes_latent.py @@ -21,6 +21,7 @@ class LatentAdd(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="LatentAdd", + search_aliases=["combine latents", "sum latents"], category="latent/advanced", inputs=[ io.Latent.Input("samples1"), @@ -47,6 +48,7 @@ class LatentSubtract(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="LatentSubtract", + search_aliases=["difference latent", "remove features"], category="latent/advanced", inputs=[ io.Latent.Input("samples1"), @@ -73,6 +75,7 @@ class LatentMultiply(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="LatentMultiply", + search_aliases=["scale latent", "amplify latent", "latent gain"], category="latent/advanced", inputs=[ io.Latent.Input("samples"), @@ -96,6 +99,7 @@ class LatentInterpolate(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="LatentInterpolate", + search_aliases=["blend latent", "mix latent", "lerp latent", "transition"], category="latent/advanced", inputs=[ io.Latent.Input("samples1"), @@ -134,6 +138,7 @@ class LatentConcat(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="LatentConcat", + search_aliases=["join latents", "stitch latents"], category="latent/advanced", inputs=[ io.Latent.Input("samples1"), @@ -173,6 +178,7 @@ class LatentCut(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="LatentCut", + search_aliases=["crop latent", "slice latent", "extract region"], category="latent/advanced", inputs=[ io.Latent.Input("samples"), @@ -213,6 +219,7 @@ class LatentCutToBatch(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="LatentCutToBatch", + search_aliases=["slice to batch", "split latent", "tile latent"], category="latent/advanced", inputs=[ io.Latent.Input("samples"), @@ -254,6 +261,7 @@ class LatentBatch(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="LatentBatch", + search_aliases=["combine latents", "merge latents", "join latents"], category="latent/batch", is_deprecated=True, inputs=[ @@ -310,6 +318,7 @@ class LatentApplyOperation(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="LatentApplyOperation", + search_aliases=["transform latent"], category="latent/advanced/operations", is_experimental=True, inputs=[ @@ -365,6 +374,7 @@ class LatentOperationTonemapReinhard(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="LatentOperationTonemapReinhard", + search_aliases=["hdr latent"], category="latent/advanced/operations", is_experimental=True, inputs=[ diff --git a/comfy_extras/nodes_load_3d.py b/comfy_extras/nodes_load_3d.py index a16b8c8f3..4b8d950ae 100644 --- a/comfy_extras/nodes_load_3d.py +++ b/comfy_extras/nodes_load_3d.py @@ -75,6 +75,7 @@ class Preview3D(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="Preview3D", + search_aliases=["view mesh", "3d viewer"], display_name="Preview 3D & Animation", category="3d", is_experimental=True, diff --git a/comfy_extras/nodes_logic.py b/comfy_extras/nodes_logic.py index eb888316a..1ed060205 100644 --- a/comfy_extras/nodes_logic.py +++ b/comfy_extras/nodes_logic.py @@ -224,6 +224,7 @@ class ConvertStringToComboNode(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="ConvertStringToComboNode", + search_aliases=["string to dropdown", "text to combo"], display_name="Convert String to Combo", category="logic", inputs=[io.String.Input("string")], @@ -239,6 +240,7 @@ class InvertBooleanNode(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="InvertBooleanNode", + search_aliases=["not", "toggle", "negate", "flip boolean"], display_name="Invert Boolean", category="logic", inputs=[io.Boolean.Input("boolean")], diff --git a/comfy_extras/nodes_lora_extract.py b/comfy_extras/nodes_lora_extract.py index a2375cba7..fb89e03f4 100644 --- a/comfy_extras/nodes_lora_extract.py +++ b/comfy_extras/nodes_lora_extract.py @@ -78,6 +78,7 @@ class LoraSave(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="LoraSave", + search_aliases=["export lora"], display_name="Extract and Save Lora", category="_for_testing", inputs=[ diff --git a/comfy_extras/nodes_lumina2.py b/comfy_extras/nodes_lumina2.py index 89ff2397a..2550475ae 100644 --- a/comfy_extras/nodes_lumina2.py +++ b/comfy_extras/nodes_lumina2.py @@ -79,6 +79,7 @@ class CLIPTextEncodeLumina2(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="CLIPTextEncodeLumina2", + search_aliases=["lumina prompt"], display_name="CLIP Text Encode for Lumina2", category="conditioning", description="Encodes a system prompt and a user prompt using a CLIP model into an embedding " diff --git a/comfy_extras/nodes_mask.py b/comfy_extras/nodes_mask.py index 290e6f55e..98e8fef8f 100644 --- a/comfy_extras/nodes_mask.py +++ b/comfy_extras/nodes_mask.py @@ -50,6 +50,7 @@ class LatentCompositeMasked(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="LatentCompositeMasked", + search_aliases=["overlay latent", "layer latent", "paste latent", "inpaint latent"], category="latent", inputs=[ IO.Latent.Input("destination"), @@ -78,6 +79,7 @@ class ImageCompositeMasked(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="ImageCompositeMasked", + search_aliases=["paste image", "overlay", "layer"], category="image", inputs=[ IO.Image.Input("destination"), @@ -105,6 +107,7 @@ class MaskToImage(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="MaskToImage", + search_aliases=["convert mask"], display_name="Convert Mask to Image", category="mask", inputs=[ @@ -126,6 +129,7 @@ class ImageToMask(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="ImageToMask", + search_aliases=["extract channel", "channel to mask"], display_name="Convert Image to Mask", category="mask", inputs=[ @@ -149,6 +153,7 @@ class ImageColorToMask(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="ImageColorToMask", + search_aliases=["color keying", "chroma key"], category="mask", inputs=[ IO.Image.Input("image"), @@ -194,6 +199,7 @@ class InvertMask(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="InvertMask", + search_aliases=["reverse mask", "flip mask"], category="mask", inputs=[ IO.Mask.Input("mask"), @@ -214,6 +220,7 @@ class CropMask(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="CropMask", + search_aliases=["cut mask", "extract mask region", "mask slice"], category="mask", inputs=[ IO.Mask.Input("mask"), @@ -239,6 +246,7 @@ class MaskComposite(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="MaskComposite", + search_aliases=["combine masks", "blend masks", "layer masks"], category="mask", inputs=[ IO.Mask.Input("destination"), @@ -287,6 +295,7 @@ class FeatherMask(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="FeatherMask", + search_aliases=["soft edge mask", "blur mask edges", "gradient mask edge"], category="mask", inputs=[ IO.Mask.Input("mask"), @@ -333,6 +342,7 @@ class GrowMask(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="GrowMask", + search_aliases=["expand mask", "shrink mask"], display_name="Grow Mask", category="mask", inputs=[ @@ -370,6 +380,7 @@ class ThresholdMask(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="ThresholdMask", + search_aliases=["binary mask"], category="mask", inputs=[ IO.Mask.Input("mask"), @@ -394,6 +405,7 @@ class MaskPreview(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="MaskPreview", + search_aliases=["show mask", "view mask", "inspect mask", "debug mask"], display_name="Preview Mask", category="mask", description="Saves the input images to your ComfyUI output directory.", diff --git a/comfy_extras/nodes_model_advanced.py b/comfy_extras/nodes_model_advanced.py index ae5d2c563..f22b333fc 100644 --- a/comfy_extras/nodes_model_advanced.py +++ b/comfy_extras/nodes_model_advanced.py @@ -299,6 +299,7 @@ class RescaleCFG: return (m, ) class ModelComputeDtype: + SEARCH_ALIASES = ["model precision", "change dtype"] @classmethod def INPUT_TYPES(s): return {"required": { "model": ("MODEL",), diff --git a/comfy_extras/nodes_model_merging.py b/comfy_extras/nodes_model_merging.py index f20beab7d..5384ed531 100644 --- a/comfy_extras/nodes_model_merging.py +++ b/comfy_extras/nodes_model_merging.py @@ -91,6 +91,7 @@ class CLIPMergeSimple: class CLIPSubtract: + SEARCH_ALIASES = ["clip difference", "text encoder subtract"] @classmethod def INPUT_TYPES(s): return {"required": { "clip1": ("CLIP",), @@ -113,6 +114,7 @@ class CLIPSubtract: class CLIPAdd: + SEARCH_ALIASES = ["combine clip"] @classmethod def INPUT_TYPES(s): return {"required": { "clip1": ("CLIP",), @@ -225,6 +227,7 @@ def save_checkpoint(model, clip=None, vae=None, clip_vision=None, filename_prefi comfy.sd.save_checkpoint(output_checkpoint, model, clip, vae, clip_vision, metadata=metadata, extra_keys=extra_keys) class CheckpointSave: + SEARCH_ALIASES = ["save model", "export checkpoint", "merge save"] def __init__(self): self.output_dir = folder_paths.get_output_directory() @@ -337,6 +340,7 @@ class VAESave: return {} class ModelSave: + SEARCH_ALIASES = ["export model", "checkpoint save"] def __init__(self): self.output_dir = folder_paths.get_output_directory() diff --git a/comfy_extras/nodes_morphology.py b/comfy_extras/nodes_morphology.py index 67377e1bc..4ab2fb7e8 100644 --- a/comfy_extras/nodes_morphology.py +++ b/comfy_extras/nodes_morphology.py @@ -12,6 +12,7 @@ class Morphology(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="Morphology", + search_aliases=["erode", "dilate"], display_name="ImageMorphology", category="image/postprocessing", inputs=[ @@ -57,6 +58,7 @@ class ImageRGBToYUV(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="ImageRGBToYUV", + search_aliases=["color space conversion"], category="image/batch", inputs=[ io.Image.Input("image"), @@ -78,6 +80,7 @@ class ImageYUVToRGB(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="ImageYUVToRGB", + search_aliases=["color space conversion"], category="image/batch", inputs=[ io.Image.Input("Y"), diff --git a/comfy_extras/nodes_pixart.py b/comfy_extras/nodes_pixart.py index a23e87b1f..2f1b73e60 100644 --- a/comfy_extras/nodes_pixart.py +++ b/comfy_extras/nodes_pixart.py @@ -7,6 +7,7 @@ class CLIPTextEncodePixArtAlpha(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="CLIPTextEncodePixArtAlpha", + search_aliases=["pixart prompt"], category="advanced/conditioning", description="Encodes text and sets the resolution conditioning for PixArt Alpha. Does not apply to PixArt Sigma.", inputs=[ diff --git a/comfy_extras/nodes_post_processing.py b/comfy_extras/nodes_post_processing.py index 6011275d6..ab002daca 100644 --- a/comfy_extras/nodes_post_processing.py +++ b/comfy_extras/nodes_post_processing.py @@ -402,7 +402,6 @@ def scale_to_multiple_cover(input: torch.Tensor, multiple: int, scale_method: st return input[:, y0:y1, x0:x1] class ResizeImageMaskNode(io.ComfyNode): - scale_methods = ["nearest-exact", "bilinear", "area", "bicubic", "lanczos"] crop_methods = ["disabled", "center"] @@ -424,6 +423,7 @@ class ResizeImageMaskNode(io.ComfyNode): crop_combo = io.Combo.Input("crop", options=cls.crop_methods, default="center") return io.Schema( node_id="ResizeImageMaskNode", + search_aliases=["scale image", "scale mask"], display_name="Resize Image/Mask", category="transform", inputs=[ @@ -569,6 +569,7 @@ class BatchMasksNode(io.ComfyNode): autogrow_template = io.Autogrow.TemplatePrefix(io.Mask.Input("mask"), prefix="mask", min=2, max=50) return io.Schema( node_id="BatchMasksNode", + search_aliases=["combine masks", "stack masks", "merge masks"], display_name="Batch Masks", category="mask", inputs=[ @@ -589,6 +590,7 @@ class BatchLatentsNode(io.ComfyNode): autogrow_template = io.Autogrow.TemplatePrefix(io.Latent.Input("latent"), prefix="latent", min=2, max=50) return io.Schema( node_id="BatchLatentsNode", + search_aliases=["combine latents", "stack latents", "merge latents"], display_name="Batch Latents", category="latent", inputs=[ @@ -612,6 +614,7 @@ class BatchImagesMasksLatentsNode(io.ComfyNode): prefix="input", min=1, max=50) return io.Schema( node_id="BatchImagesMasksLatentsNode", + search_aliases=["combine batch", "merge batch", "stack inputs"], display_name="Batch Images/Masks/Latents", category="util", inputs=[ diff --git a/comfy_extras/nodes_preview_any.py b/comfy_extras/nodes_preview_any.py index 91502ebf2..b0a6f279d 100644 --- a/comfy_extras/nodes_preview_any.py +++ b/comfy_extras/nodes_preview_any.py @@ -16,7 +16,7 @@ class PreviewAny(): OUTPUT_NODE = True CATEGORY = "utils" - SEARCH_ALIASES = ["preview", "show", "display", "view", "show text", "display text", "preview text", "show output", "inspect", "debug"] + SEARCH_ALIASES = ["show output", "inspect", "debug", "print value", "show text"] def main(self, source=None): value = 'None' diff --git a/comfy_extras/nodes_sd3.py b/comfy_extras/nodes_sd3.py index 14782cb2b..02e5e7dd8 100644 --- a/comfy_extras/nodes_sd3.py +++ b/comfy_extras/nodes_sd3.py @@ -65,6 +65,7 @@ class CLIPTextEncodeSD3(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="CLIPTextEncodeSD3", + search_aliases=["sd3 prompt"], category="advanced/conditioning", inputs=[ io.Clip.Input("clip"), diff --git a/comfy_extras/nodes_string.py b/comfy_extras/nodes_string.py index a2d5f0d94..8d3e65cc5 100644 --- a/comfy_extras/nodes_string.py +++ b/comfy_extras/nodes_string.py @@ -32,6 +32,7 @@ class StringSubstring(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="StringSubstring", + search_aliases=["extract text", "text portion"], display_name="Substring", category="utils/string", inputs=[ @@ -54,6 +55,7 @@ class StringLength(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="StringLength", + search_aliases=["character count", "text size"], display_name="Length", category="utils/string", inputs=[ @@ -74,6 +76,7 @@ class CaseConverter(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="CaseConverter", + search_aliases=["text case", "uppercase", "lowercase", "capitalize"], display_name="Case Converter", category="utils/string", inputs=[ @@ -106,6 +109,7 @@ class StringTrim(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="StringTrim", + search_aliases=["clean whitespace", "remove whitespace"], display_name="Trim", category="utils/string", inputs=[ @@ -136,6 +140,7 @@ class StringReplace(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="StringReplace", + search_aliases=["find and replace", "substitute", "swap text"], display_name="Replace", category="utils/string", inputs=[ @@ -158,6 +163,7 @@ class StringContains(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="StringContains", + search_aliases=["text includes", "string includes"], display_name="Contains", category="utils/string", inputs=[ @@ -185,6 +191,7 @@ class StringCompare(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="StringCompare", + search_aliases=["text match", "string equals", "starts with", "ends with"], display_name="Compare", category="utils/string", inputs=[ @@ -220,6 +227,7 @@ class RegexMatch(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="RegexMatch", + search_aliases=["pattern match", "text contains", "string match"], display_name="Regex Match", category="utils/string", inputs=[ @@ -260,6 +268,7 @@ class RegexExtract(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="RegexExtract", + search_aliases=["pattern extract", "text parser", "parse text"], display_name="Regex Extract", category="utils/string", inputs=[ @@ -334,6 +343,7 @@ class RegexReplace(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="RegexReplace", + search_aliases=["pattern replace", "find and replace", "substitution"], display_name="Regex Replace", category="utils/string", description="Find and replace text using regex patterns.", diff --git a/comfy_extras/nodes_train.py b/comfy_extras/nodes_train.py index 364804205..68a73cf13 100644 --- a/comfy_extras/nodes_train.py +++ b/comfy_extras/nodes_train.py @@ -1101,6 +1101,7 @@ class SaveLoRA(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="SaveLoRA", + search_aliases=["export lora"], display_name="Save LoRA Weights", category="loaders", is_experimental=True, @@ -1144,6 +1145,7 @@ class LossGraphNode(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="LossGraphNode", + search_aliases=["training chart", "training visualization", "plot loss"], display_name="Plot Loss Graph", category="training", is_experimental=True, diff --git a/comfy_extras/nodes_video.py b/comfy_extras/nodes_video.py index c609e03da..ccf7b63d3 100644 --- a/comfy_extras/nodes_video.py +++ b/comfy_extras/nodes_video.py @@ -16,6 +16,7 @@ class SaveWEBM(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="SaveWEBM", + search_aliases=["export webm"], category="image/video", is_experimental=True, inputs=[ @@ -69,6 +70,7 @@ class SaveVideo(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="SaveVideo", + search_aliases=["export video"], display_name="Save Video", category="image/video", description="Saves the input images to your ComfyUI output directory.", @@ -116,6 +118,7 @@ class CreateVideo(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="CreateVideo", + search_aliases=["images to video"], display_name="Create Video", category="image/video", description="Create a video from images.", @@ -140,6 +143,7 @@ class GetVideoComponents(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="GetVideoComponents", + search_aliases=["extract frames", "split video", "video to images", "demux"], display_name="Get Video Components", category="image/video", description="Extracts all components from a video: frames, audio, and framerate.", @@ -167,6 +171,7 @@ class LoadVideo(io.ComfyNode): files = folder_paths.filter_files_content_types(files, ["video"]) return io.Schema( node_id="LoadVideo", + search_aliases=["import video", "open video", "video file"], display_name="Load Video", category="image/video", inputs=[ diff --git a/comfy_extras/nodes_wan.py b/comfy_extras/nodes_wan.py index 90deb0077..2ff012134 100644 --- a/comfy_extras/nodes_wan.py +++ b/comfy_extras/nodes_wan.py @@ -287,6 +287,7 @@ class WanVaceToVideo(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="WanVaceToVideo", + search_aliases=["video conditioning", "video control"], category="conditioning/video_models", inputs=[ io.Conditioning.Input("positive"), @@ -705,6 +706,7 @@ class WanTrackToVideo(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="WanTrackToVideo", + search_aliases=["motion tracking", "trajectory video", "point tracking", "keypoint animation"], category="conditioning/video_models", inputs=[ io.Conditioning.Input("positive"), diff --git a/comfy_extras/nodes_wanmove.py b/comfy_extras/nodes_wanmove.py index 5f39afa46..d60baf230 100644 --- a/comfy_extras/nodes_wanmove.py +++ b/comfy_extras/nodes_wanmove.py @@ -324,6 +324,7 @@ class GenerateTracks(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="GenerateTracks", + search_aliases=["motion paths", "camera movement", "trajectory"], category="conditioning/video_models", inputs=[ io.Int.Input("width", default=832, min=16, max=4096, step=16), diff --git a/comfy_extras/nodes_webcam.py b/comfy_extras/nodes_webcam.py index 5bf80b4c6..6349ac017 100644 --- a/comfy_extras/nodes_webcam.py +++ b/comfy_extras/nodes_webcam.py @@ -5,6 +5,7 @@ MAX_RESOLUTION = nodes.MAX_RESOLUTION class WebcamCapture(nodes.LoadImage): + SEARCH_ALIASES = ["camera input", "live capture", "camera feed", "snapshot"] @classmethod def INPUT_TYPES(s): return { diff --git a/nodes.py b/nodes.py index 8864fda60..158106686 100644 --- a/nodes.py +++ b/nodes.py @@ -93,6 +93,8 @@ class ConditioningCombine: return (conditioning_1 + conditioning_2, ) class ConditioningAverage : + SEARCH_ALIASES = ["blend prompts", "interpolate conditioning", "mix prompts", "style fusion", "weighted blend"] + @classmethod def INPUT_TYPES(s): return {"required": {"conditioning_to": ("CONDITIONING", ), "conditioning_from": ("CONDITIONING", ), @@ -159,6 +161,8 @@ class ConditioningConcat: return (out, ) class ConditioningSetArea: + SEARCH_ALIASES = ["regional prompt", "area prompt", "spatial conditioning", "localized prompt"] + @classmethod def INPUT_TYPES(s): return {"required": {"conditioning": ("CONDITIONING", ), @@ -217,6 +221,8 @@ class ConditioningSetAreaStrength: class ConditioningSetMask: + SEARCH_ALIASES = ["masked prompt", "regional inpaint conditioning", "mask conditioning"] + @classmethod def INPUT_TYPES(s): return {"required": {"conditioning": ("CONDITIONING", ), @@ -242,6 +248,8 @@ class ConditioningSetMask: return (c, ) class ConditioningZeroOut: + SEARCH_ALIASES = ["null conditioning", "clear conditioning"] + @classmethod def INPUT_TYPES(s): return {"required": {"conditioning": ("CONDITIONING", )}} @@ -467,6 +475,8 @@ class InpaintModelConditioning: class SaveLatent: + SEARCH_ALIASES = ["export latent"] + def __init__(self): self.output_dir = folder_paths.get_output_directory() @@ -518,6 +528,8 @@ class SaveLatent: class LoadLatent: + SEARCH_ALIASES = ["import latent", "open latent"] + @classmethod def INPUT_TYPES(s): input_dir = folder_paths.get_input_directory() @@ -554,6 +566,8 @@ class LoadLatent: class CheckpointLoader: + SEARCH_ALIASES = ["load model", "model loader"] + @classmethod def INPUT_TYPES(s): return {"required": { "config_name": (folder_paths.get_filename_list("configs"), ), @@ -593,6 +607,8 @@ class CheckpointLoaderSimple: return out[:3] class DiffusersLoader: + SEARCH_ALIASES = ["load diffusers model"] + @classmethod def INPUT_TYPES(cls): paths = [] @@ -1063,6 +1079,8 @@ class StyleModelLoader: class StyleModelApply: + SEARCH_ALIASES = ["style transfer"] + @classmethod def INPUT_TYPES(s): return {"required": {"conditioning": ("CONDITIONING", ), @@ -1216,6 +1234,8 @@ class EmptyLatentImage: class LatentFromBatch: + SEARCH_ALIASES = ["select from batch", "pick latent", "batch subset"] + @classmethod def INPUT_TYPES(s): return {"required": { "samples": ("LATENT",), @@ -1248,6 +1268,8 @@ class LatentFromBatch: return (s,) class RepeatLatentBatch: + SEARCH_ALIASES = ["duplicate latent", "clone latent"] + @classmethod def INPUT_TYPES(s): return {"required": { "samples": ("LATENT",), @@ -1274,6 +1296,8 @@ class RepeatLatentBatch: return (s,) class LatentUpscale: + SEARCH_ALIASES = ["enlarge latent", "resize latent"] + upscale_methods = ["nearest-exact", "bilinear", "area", "bicubic", "bislerp"] crop_methods = ["disabled", "center"] @@ -1308,6 +1332,8 @@ class LatentUpscale: return (s,) class LatentUpscaleBy: + SEARCH_ALIASES = ["enlarge latent", "resize latent", "scale latent"] + upscale_methods = ["nearest-exact", "bilinear", "area", "bicubic", "bislerp"] @classmethod @@ -1351,6 +1377,8 @@ class LatentRotate: return (s,) class LatentFlip: + SEARCH_ALIASES = ["mirror latent"] + @classmethod def INPUT_TYPES(s): return {"required": { "samples": ("LATENT",), @@ -1371,6 +1399,8 @@ class LatentFlip: return (s,) class LatentComposite: + SEARCH_ALIASES = ["overlay latent", "layer latent", "paste latent"] + @classmethod def INPUT_TYPES(s): return {"required": { "samples_to": ("LATENT",), @@ -1413,6 +1443,8 @@ class LatentComposite: return (samples_out,) class LatentBlend: + SEARCH_ALIASES = ["mix latents", "interpolate latents"] + @classmethod def INPUT_TYPES(s): return {"required": { @@ -1454,6 +1486,8 @@ class LatentBlend: raise ValueError(f"Unsupported blend mode: {mode}") class LatentCrop: + SEARCH_ALIASES = ["trim latent", "cut latent"] + @classmethod def INPUT_TYPES(s): return {"required": { "samples": ("LATENT",), @@ -1739,6 +1773,8 @@ class LoadImage: return True class LoadImageMask: + SEARCH_ALIASES = ["import mask", "alpha mask", "channel mask"] + _color_channels = ["alpha", "red", "green", "blue"] @classmethod def INPUT_TYPES(s): @@ -1789,6 +1825,8 @@ class LoadImageMask: class LoadImageOutput(LoadImage): + SEARCH_ALIASES = ["output image", "previous generation"] + @classmethod def INPUT_TYPES(s): return { @@ -1862,6 +1900,7 @@ class ImageScaleBy: return (s,) class ImageInvert: + SEARCH_ALIASES = ["reverse colors"] @classmethod def INPUT_TYPES(s): @@ -1877,6 +1916,7 @@ class ImageInvert: return (s,) class ImageBatch: + SEARCH_ALIASES = ["combine images", "merge images", "stack images"] @classmethod def INPUT_TYPES(s): @@ -1922,6 +1962,7 @@ class EmptyImage: return (torch.cat((r, g, b), dim=-1), ) class ImagePadForOutpaint: + SEARCH_ALIASES = ["extend canvas", "expand image"] @classmethod def INPUT_TYPES(s): From 4e3038114a725d15166a726860a29cbab0dda4e3 Mon Sep 17 00:00:00 2001 From: Christian Byrne Date: Thu, 22 Jan 2026 18:46:55 -0800 Subject: [PATCH 502/647] feat: Improve ResizeImageMaskNode UX with tooltips and search aliases (#12013) - Add search_aliases for discoverability: resize, scale, dimensions, etc. - Add node description for hover tooltip - Add tooltips to all inputs explaining their behavior - Reorder options: most common (scale dimensions) first, most technical (scale to multiple) last Addresses user feedback that 'resize' search returned nothing useful and options like 'match size' and 'scale to multiple' were not self-explanatory. --- comfy_extras/nodes_post_processing.py | 66 +++++++++++++++++---------- 1 file changed, 41 insertions(+), 25 deletions(-) diff --git a/comfy_extras/nodes_post_processing.py b/comfy_extras/nodes_post_processing.py index ab002daca..32ab2f70d 100644 --- a/comfy_extras/nodes_post_processing.py +++ b/comfy_extras/nodes_post_processing.py @@ -420,47 +420,63 @@ class ResizeImageMaskNode(io.ComfyNode): @classmethod def define_schema(cls): template = io.MatchType.Template("input_type", [io.Image, io.Mask]) - crop_combo = io.Combo.Input("crop", options=cls.crop_methods, default="center") + crop_combo = io.Combo.Input( + "crop", + options=cls.crop_methods, + default="center", + tooltip="How to handle aspect ratio mismatch: 'disabled' stretches to fit, 'center' crops to maintain aspect ratio.", + ) return io.Schema( node_id="ResizeImageMaskNode", search_aliases=["scale image", "scale mask"], display_name="Resize Image/Mask", + description="Resize an image or mask using various scaling methods.", category="transform", + search_aliases=["resize", "resize image", "resize mask", "scale", "scale image", "image resize", "change size", "dimensions", "shrink", "enlarge"], inputs=[ io.MatchType.Input("input", template=template), - io.DynamicCombo.Input("resize_type", options=[ - io.DynamicCombo.Option(ResizeType.SCALE_BY, [ - io.Float.Input("multiplier", default=1.00, min=0.01, max=8.0, step=0.01), + io.DynamicCombo.Input( + "resize_type", + tooltip="Select how to resize: by exact dimensions, scale factor, matching another image, etc.", + options=[ + io.DynamicCombo.Option(ResizeType.SCALE_DIMENSIONS, [ + io.Int.Input("width", default=512, min=0, max=MAX_RESOLUTION, step=1, tooltip="Target width in pixels. Set to 0 to auto-calculate from height while preserving aspect ratio."), + io.Int.Input("height", default=512, min=0, max=MAX_RESOLUTION, step=1, tooltip="Target height in pixels. Set to 0 to auto-calculate from width while preserving aspect ratio."), + crop_combo, ]), - io.DynamicCombo.Option(ResizeType.SCALE_DIMENSIONS, [ - io.Int.Input("width", default=512, min=0, max=MAX_RESOLUTION, step=1), - io.Int.Input("height", default=512, min=0, max=MAX_RESOLUTION, step=1), - crop_combo, + io.DynamicCombo.Option(ResizeType.SCALE_BY, [ + io.Float.Input("multiplier", default=1.00, min=0.01, max=8.0, step=0.01, tooltip="Scale factor (e.g., 2.0 doubles size, 0.5 halves size)."), ]), - io.DynamicCombo.Option(ResizeType.SCALE_LONGER_DIMENSION, [ - io.Int.Input("longer_size", default=512, min=0, max=MAX_RESOLUTION, step=1), + io.DynamicCombo.Option(ResizeType.SCALE_LONGER_DIMENSION, [ + io.Int.Input("longer_size", default=512, min=0, max=MAX_RESOLUTION, step=1, tooltip="The longer edge will be resized to this value. Aspect ratio is preserved."), ]), - io.DynamicCombo.Option(ResizeType.SCALE_SHORTER_DIMENSION, [ - io.Int.Input("shorter_size", default=512, min=0, max=MAX_RESOLUTION, step=1), + io.DynamicCombo.Option(ResizeType.SCALE_SHORTER_DIMENSION, [ + io.Int.Input("shorter_size", default=512, min=0, max=MAX_RESOLUTION, step=1, tooltip="The shorter edge will be resized to this value. Aspect ratio is preserved."), ]), - io.DynamicCombo.Option(ResizeType.SCALE_WIDTH, [ - io.Int.Input("width", default=512, min=0, max=MAX_RESOLUTION, step=1), + io.DynamicCombo.Option(ResizeType.SCALE_WIDTH, [ + io.Int.Input("width", default=512, min=0, max=MAX_RESOLUTION, step=1, tooltip="Target width in pixels. Height auto-adjusts to preserve aspect ratio."), ]), - io.DynamicCombo.Option(ResizeType.SCALE_HEIGHT, [ - io.Int.Input("height", default=512, min=0, max=MAX_RESOLUTION, step=1), + io.DynamicCombo.Option(ResizeType.SCALE_HEIGHT, [ + io.Int.Input("height", default=512, min=0, max=MAX_RESOLUTION, step=1, tooltip="Target height in pixels. Width auto-adjusts to preserve aspect ratio."), ]), - io.DynamicCombo.Option(ResizeType.SCALE_TOTAL_PIXELS, [ - io.Float.Input("megapixels", default=1.0, min=0.01, max=16.0, step=0.01), + io.DynamicCombo.Option(ResizeType.SCALE_TOTAL_PIXELS, [ + io.Float.Input("megapixels", default=1.0, min=0.01, max=16.0, step=0.01, tooltip="Target total megapixels (e.g., 1.0 ≈ 1024×1024). Aspect ratio is preserved."), ]), - io.DynamicCombo.Option(ResizeType.MATCH_SIZE, [ - io.MultiType.Input("match", [io.Image, io.Mask]), - crop_combo, + io.DynamicCombo.Option(ResizeType.MATCH_SIZE, [ + io.MultiType.Input("match", [io.Image, io.Mask], tooltip="Resize input to match the dimensions of this reference image or mask."), + crop_combo, ]), - io.DynamicCombo.Option(ResizeType.SCALE_TO_MULTIPLE, [ - io.Int.Input("multiple", default=8, min=1, max=MAX_RESOLUTION, step=1), + io.DynamicCombo.Option(ResizeType.SCALE_TO_MULTIPLE, [ + io.Int.Input("multiple", default=8, min=1, max=MAX_RESOLUTION, step=1, tooltip="Resize so width and height are divisible by this number. Useful for latent alignment (e.g., 8 or 64)."), ]), - ]), - io.Combo.Input("scale_method", options=cls.scale_methods, default="area"), + ], + ), + io.Combo.Input( + "scale_method", + options=cls.scale_methods, + default="area", + tooltip="Interpolation algorithm. 'area' is best for downscaling, 'lanczos' for upscaling, 'nearest-exact' for pixel art.", + ), ], outputs=[io.MatchType.Output(template=template, display_name="resized")] ) From f443b9f2ca3109f7e3ef6c5de3cdd22330fbf34c Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Thu, 22 Jan 2026 20:02:37 -0800 Subject: [PATCH 503/647] =?UTF-8?q?Revert=20"feat:=20Improve=20ResizeImage?= =?UTF-8?q?MaskNode=20UX=20with=20tooltips=20and=20search=20aliases?= =?UTF-8?q?=E2=80=A6"=20(#12038)?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit This reverts commit 4e3038114a725d15166a726860a29cbab0dda4e3. --- comfy_extras/nodes_post_processing.py | 66 ++++++++++----------------- 1 file changed, 25 insertions(+), 41 deletions(-) diff --git a/comfy_extras/nodes_post_processing.py b/comfy_extras/nodes_post_processing.py index 32ab2f70d..ab002daca 100644 --- a/comfy_extras/nodes_post_processing.py +++ b/comfy_extras/nodes_post_processing.py @@ -420,63 +420,47 @@ class ResizeImageMaskNode(io.ComfyNode): @classmethod def define_schema(cls): template = io.MatchType.Template("input_type", [io.Image, io.Mask]) - crop_combo = io.Combo.Input( - "crop", - options=cls.crop_methods, - default="center", - tooltip="How to handle aspect ratio mismatch: 'disabled' stretches to fit, 'center' crops to maintain aspect ratio.", - ) + crop_combo = io.Combo.Input("crop", options=cls.crop_methods, default="center") return io.Schema( node_id="ResizeImageMaskNode", search_aliases=["scale image", "scale mask"], display_name="Resize Image/Mask", - description="Resize an image or mask using various scaling methods.", category="transform", - search_aliases=["resize", "resize image", "resize mask", "scale", "scale image", "image resize", "change size", "dimensions", "shrink", "enlarge"], inputs=[ io.MatchType.Input("input", template=template), - io.DynamicCombo.Input( - "resize_type", - tooltip="Select how to resize: by exact dimensions, scale factor, matching another image, etc.", - options=[ - io.DynamicCombo.Option(ResizeType.SCALE_DIMENSIONS, [ - io.Int.Input("width", default=512, min=0, max=MAX_RESOLUTION, step=1, tooltip="Target width in pixels. Set to 0 to auto-calculate from height while preserving aspect ratio."), - io.Int.Input("height", default=512, min=0, max=MAX_RESOLUTION, step=1, tooltip="Target height in pixels. Set to 0 to auto-calculate from width while preserving aspect ratio."), - crop_combo, + io.DynamicCombo.Input("resize_type", options=[ + io.DynamicCombo.Option(ResizeType.SCALE_BY, [ + io.Float.Input("multiplier", default=1.00, min=0.01, max=8.0, step=0.01), ]), - io.DynamicCombo.Option(ResizeType.SCALE_BY, [ - io.Float.Input("multiplier", default=1.00, min=0.01, max=8.0, step=0.01, tooltip="Scale factor (e.g., 2.0 doubles size, 0.5 halves size)."), + io.DynamicCombo.Option(ResizeType.SCALE_DIMENSIONS, [ + io.Int.Input("width", default=512, min=0, max=MAX_RESOLUTION, step=1), + io.Int.Input("height", default=512, min=0, max=MAX_RESOLUTION, step=1), + crop_combo, ]), - io.DynamicCombo.Option(ResizeType.SCALE_LONGER_DIMENSION, [ - io.Int.Input("longer_size", default=512, min=0, max=MAX_RESOLUTION, step=1, tooltip="The longer edge will be resized to this value. Aspect ratio is preserved."), + io.DynamicCombo.Option(ResizeType.SCALE_LONGER_DIMENSION, [ + io.Int.Input("longer_size", default=512, min=0, max=MAX_RESOLUTION, step=1), ]), - io.DynamicCombo.Option(ResizeType.SCALE_SHORTER_DIMENSION, [ - io.Int.Input("shorter_size", default=512, min=0, max=MAX_RESOLUTION, step=1, tooltip="The shorter edge will be resized to this value. Aspect ratio is preserved."), + io.DynamicCombo.Option(ResizeType.SCALE_SHORTER_DIMENSION, [ + io.Int.Input("shorter_size", default=512, min=0, max=MAX_RESOLUTION, step=1), ]), - io.DynamicCombo.Option(ResizeType.SCALE_WIDTH, [ - io.Int.Input("width", default=512, min=0, max=MAX_RESOLUTION, step=1, tooltip="Target width in pixels. Height auto-adjusts to preserve aspect ratio."), + io.DynamicCombo.Option(ResizeType.SCALE_WIDTH, [ + io.Int.Input("width", default=512, min=0, max=MAX_RESOLUTION, step=1), ]), - io.DynamicCombo.Option(ResizeType.SCALE_HEIGHT, [ - io.Int.Input("height", default=512, min=0, max=MAX_RESOLUTION, step=1, tooltip="Target height in pixels. Width auto-adjusts to preserve aspect ratio."), + io.DynamicCombo.Option(ResizeType.SCALE_HEIGHT, [ + io.Int.Input("height", default=512, min=0, max=MAX_RESOLUTION, step=1), ]), - io.DynamicCombo.Option(ResizeType.SCALE_TOTAL_PIXELS, [ - io.Float.Input("megapixels", default=1.0, min=0.01, max=16.0, step=0.01, tooltip="Target total megapixels (e.g., 1.0 ≈ 1024×1024). Aspect ratio is preserved."), + io.DynamicCombo.Option(ResizeType.SCALE_TOTAL_PIXELS, [ + io.Float.Input("megapixels", default=1.0, min=0.01, max=16.0, step=0.01), ]), - io.DynamicCombo.Option(ResizeType.MATCH_SIZE, [ - io.MultiType.Input("match", [io.Image, io.Mask], tooltip="Resize input to match the dimensions of this reference image or mask."), - crop_combo, + io.DynamicCombo.Option(ResizeType.MATCH_SIZE, [ + io.MultiType.Input("match", [io.Image, io.Mask]), + crop_combo, ]), - io.DynamicCombo.Option(ResizeType.SCALE_TO_MULTIPLE, [ - io.Int.Input("multiple", default=8, min=1, max=MAX_RESOLUTION, step=1, tooltip="Resize so width and height are divisible by this number. Useful for latent alignment (e.g., 8 or 64)."), + io.DynamicCombo.Option(ResizeType.SCALE_TO_MULTIPLE, [ + io.Int.Input("multiple", default=8, min=1, max=MAX_RESOLUTION, step=1), ]), - ], - ), - io.Combo.Input( - "scale_method", - options=cls.scale_methods, - default="area", - tooltip="Interpolation algorithm. 'area' is best for downscaling, 'lanczos' for upscaling, 'nearest-exact' for pixel art.", - ), + ]), + io.Combo.Input("scale_method", options=cls.scale_methods, default="area"), ], outputs=[io.MatchType.Output(template=template, display_name="resized")] ) From 79cdbc81cb552b363430d1e88c98c4b4b4b4cf62 Mon Sep 17 00:00:00 2001 From: Christian Byrne Date: Thu, 22 Jan 2026 22:04:27 -0800 Subject: [PATCH 504/647] feat: Improve ResizeImageMaskNode UX with tooltips and search aliases (#12040) - Add search_aliases for discoverability: resize, scale, dimensions, etc. - Add node description for hover tooltip - Add tooltips to all inputs explaining their behavior - Reorder options: most common (scale dimensions) first, most technical (scale to multiple) last Addresses user feedback that 'resize' search returned nothing useful and options like 'match size' and 'scale to multiple' were not self-explanatory. --- comfy_extras/nodes_post_processing.py | 67 ++++++++++++++++----------- 1 file changed, 41 insertions(+), 26 deletions(-) diff --git a/comfy_extras/nodes_post_processing.py b/comfy_extras/nodes_post_processing.py index ab002daca..a52a90e2c 100644 --- a/comfy_extras/nodes_post_processing.py +++ b/comfy_extras/nodes_post_processing.py @@ -420,47 +420,62 @@ class ResizeImageMaskNode(io.ComfyNode): @classmethod def define_schema(cls): template = io.MatchType.Template("input_type", [io.Image, io.Mask]) - crop_combo = io.Combo.Input("crop", options=cls.crop_methods, default="center") + crop_combo = io.Combo.Input( + "crop", + options=cls.crop_methods, + default="center", + tooltip="How to handle aspect ratio mismatch: 'disabled' stretches to fit, 'center' crops to maintain aspect ratio.", + ) return io.Schema( node_id="ResizeImageMaskNode", - search_aliases=["scale image", "scale mask"], display_name="Resize Image/Mask", + description="Resize an image or mask using various scaling methods.", category="transform", + search_aliases=["resize", "resize image", "resize mask", "scale", "scale image", "scale mask", "image resize", "change size", "dimensions", "shrink", "enlarge"], inputs=[ io.MatchType.Input("input", template=template), - io.DynamicCombo.Input("resize_type", options=[ - io.DynamicCombo.Option(ResizeType.SCALE_BY, [ - io.Float.Input("multiplier", default=1.00, min=0.01, max=8.0, step=0.01), + io.DynamicCombo.Input( + "resize_type", + tooltip="Select how to resize: by exact dimensions, scale factor, matching another image, etc.", + options=[ + io.DynamicCombo.Option(ResizeType.SCALE_DIMENSIONS, [ + io.Int.Input("width", default=512, min=0, max=MAX_RESOLUTION, step=1, tooltip="Target width in pixels. Set to 0 to auto-calculate from height while preserving aspect ratio."), + io.Int.Input("height", default=512, min=0, max=MAX_RESOLUTION, step=1, tooltip="Target height in pixels. Set to 0 to auto-calculate from width while preserving aspect ratio."), + crop_combo, ]), - io.DynamicCombo.Option(ResizeType.SCALE_DIMENSIONS, [ - io.Int.Input("width", default=512, min=0, max=MAX_RESOLUTION, step=1), - io.Int.Input("height", default=512, min=0, max=MAX_RESOLUTION, step=1), - crop_combo, + io.DynamicCombo.Option(ResizeType.SCALE_BY, [ + io.Float.Input("multiplier", default=1.00, min=0.01, max=8.0, step=0.01, tooltip="Scale factor (e.g., 2.0 doubles size, 0.5 halves size)."), ]), - io.DynamicCombo.Option(ResizeType.SCALE_LONGER_DIMENSION, [ - io.Int.Input("longer_size", default=512, min=0, max=MAX_RESOLUTION, step=1), + io.DynamicCombo.Option(ResizeType.SCALE_LONGER_DIMENSION, [ + io.Int.Input("longer_size", default=512, min=0, max=MAX_RESOLUTION, step=1, tooltip="The longer edge will be resized to this value. Aspect ratio is preserved."), ]), - io.DynamicCombo.Option(ResizeType.SCALE_SHORTER_DIMENSION, [ - io.Int.Input("shorter_size", default=512, min=0, max=MAX_RESOLUTION, step=1), + io.DynamicCombo.Option(ResizeType.SCALE_SHORTER_DIMENSION, [ + io.Int.Input("shorter_size", default=512, min=0, max=MAX_RESOLUTION, step=1, tooltip="The shorter edge will be resized to this value. Aspect ratio is preserved."), ]), - io.DynamicCombo.Option(ResizeType.SCALE_WIDTH, [ - io.Int.Input("width", default=512, min=0, max=MAX_RESOLUTION, step=1), + io.DynamicCombo.Option(ResizeType.SCALE_WIDTH, [ + io.Int.Input("width", default=512, min=0, max=MAX_RESOLUTION, step=1, tooltip="Target width in pixels. Height auto-adjusts to preserve aspect ratio."), ]), - io.DynamicCombo.Option(ResizeType.SCALE_HEIGHT, [ - io.Int.Input("height", default=512, min=0, max=MAX_RESOLUTION, step=1), + io.DynamicCombo.Option(ResizeType.SCALE_HEIGHT, [ + io.Int.Input("height", default=512, min=0, max=MAX_RESOLUTION, step=1, tooltip="Target height in pixels. Width auto-adjusts to preserve aspect ratio."), ]), - io.DynamicCombo.Option(ResizeType.SCALE_TOTAL_PIXELS, [ - io.Float.Input("megapixels", default=1.0, min=0.01, max=16.0, step=0.01), + io.DynamicCombo.Option(ResizeType.SCALE_TOTAL_PIXELS, [ + io.Float.Input("megapixels", default=1.0, min=0.01, max=16.0, step=0.01, tooltip="Target total megapixels (e.g., 1.0 ≈ 1024×1024). Aspect ratio is preserved."), ]), - io.DynamicCombo.Option(ResizeType.MATCH_SIZE, [ - io.MultiType.Input("match", [io.Image, io.Mask]), - crop_combo, + io.DynamicCombo.Option(ResizeType.MATCH_SIZE, [ + io.MultiType.Input("match", [io.Image, io.Mask], tooltip="Resize input to match the dimensions of this reference image or mask."), + crop_combo, ]), - io.DynamicCombo.Option(ResizeType.SCALE_TO_MULTIPLE, [ - io.Int.Input("multiple", default=8, min=1, max=MAX_RESOLUTION, step=1), + io.DynamicCombo.Option(ResizeType.SCALE_TO_MULTIPLE, [ + io.Int.Input("multiple", default=8, min=1, max=MAX_RESOLUTION, step=1, tooltip="Resize so width and height are divisible by this number. Useful for latent alignment (e.g., 8 or 64)."), ]), - ]), - io.Combo.Input("scale_method", options=cls.scale_methods, default="area"), + ], + ), + io.Combo.Input( + "scale_method", + options=cls.scale_methods, + default="area", + tooltip="Interpolation algorithm. 'area' is best for downscaling, 'lanczos' for upscaling, 'nearest-exact' for pixel art.", + ), ], outputs=[io.MatchType.Output(template=template, display_name="resized")] ) From 55bd606e92ea0a0ef1cc83a7fa4f6decf0128b12 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Jukka=20Sepp=C3=A4nen?= <40791699+kijai@users.noreply.github.com> Date: Fri, 23 Jan 2026 22:26:38 +0200 Subject: [PATCH 505/647] LTX2: Refactor forward function for better VRAM efficiency and fix spatial inpainting (#12046) * Disable timestep embed compression when inpainting Spatial inpainting not compatible with the compression * Reduce crossattn peak VRAM * LTX2: Refactor forward function for better VRAM efficiency --- comfy/ldm/lightricks/av_model.py | 230 +++++++++++++------------------ 1 file changed, 94 insertions(+), 136 deletions(-) diff --git a/comfy/ldm/lightricks/av_model.py b/comfy/ldm/lightricks/av_model.py index c12ace241..2c6954ecd 100644 --- a/comfy/ldm/lightricks/av_model.py +++ b/comfy/ldm/lightricks/av_model.py @@ -18,12 +18,12 @@ class CompressedTimestep: def __init__(self, tensor: torch.Tensor, patches_per_frame: int): """ tensor: [batch_size, num_tokens, feature_dim] tensor where num_tokens = num_frames * patches_per_frame - patches_per_frame: Number of spatial patches per frame (height * width in latent space) + patches_per_frame: Number of spatial patches per frame (height * width in latent space), or None to disable compression """ self.batch_size, num_tokens, self.feature_dim = tensor.shape # Check if compression is valid (num_tokens must be divisible by patches_per_frame) - if num_tokens % patches_per_frame == 0 and num_tokens >= patches_per_frame: + if patches_per_frame is not None and num_tokens % patches_per_frame == 0 and num_tokens >= patches_per_frame: self.patches_per_frame = patches_per_frame self.num_frames = num_tokens // patches_per_frame @@ -215,22 +215,9 @@ class BasicAVTransformerBlock(nn.Module): return (*scale_shift_ada_values, *gate_ada_values) def forward( - self, - x: Tuple[torch.Tensor, torch.Tensor], - v_context=None, - a_context=None, - attention_mask=None, - v_timestep=None, - a_timestep=None, - v_pe=None, - a_pe=None, - v_cross_pe=None, - a_cross_pe=None, - v_cross_scale_shift_timestep=None, - a_cross_scale_shift_timestep=None, - v_cross_gate_timestep=None, - a_cross_gate_timestep=None, - transformer_options=None, + self, x: Tuple[torch.Tensor, torch.Tensor], v_context=None, a_context=None, attention_mask=None, v_timestep=None, a_timestep=None, + v_pe=None, a_pe=None, v_cross_pe=None, a_cross_pe=None, v_cross_scale_shift_timestep=None, a_cross_scale_shift_timestep=None, + v_cross_gate_timestep=None, a_cross_gate_timestep=None, transformer_options=None, ) -> Tuple[torch.Tensor, torch.Tensor]: run_vx = transformer_options.get("run_vx", True) run_ax = transformer_options.get("run_ax", True) @@ -240,144 +227,102 @@ class BasicAVTransformerBlock(nn.Module): run_a2v = run_vx and transformer_options.get("a2v_cross_attn", True) and ax.numel() > 0 run_v2a = run_ax and transformer_options.get("v2a_cross_attn", True) + # video if run_vx: - vshift_msa, vscale_msa, vgate_msa = ( - self.get_ada_values(self.scale_shift_table, vx.shape[0], v_timestep, slice(0, 3)) - ) - + # video self-attention + vshift_msa, vscale_msa = (self.get_ada_values(self.scale_shift_table, vx.shape[0], v_timestep, slice(0, 2))) norm_vx = comfy.ldm.common_dit.rms_norm(vx) * (1 + vscale_msa) + vshift_msa - vx += self.attn1(norm_vx, pe=v_pe, transformer_options=transformer_options) * vgate_msa - vx += self.attn2( - comfy.ldm.common_dit.rms_norm(vx), - context=v_context, - mask=attention_mask, - transformer_options=transformer_options, - ) - - del vshift_msa, vscale_msa, vgate_msa + del vshift_msa, vscale_msa + attn1_out = self.attn1(norm_vx, pe=v_pe, transformer_options=transformer_options) + del norm_vx + # video cross-attention + vgate_msa = self.get_ada_values(self.scale_shift_table, vx.shape[0], v_timestep, slice(2, 3))[0] + vx.addcmul_(attn1_out, vgate_msa) + del vgate_msa, attn1_out + vx.add_(self.attn2(comfy.ldm.common_dit.rms_norm(vx), context=v_context, mask=attention_mask, transformer_options=transformer_options)) + # audio if run_ax: - ashift_msa, ascale_msa, agate_msa = ( - self.get_ada_values(self.audio_scale_shift_table, ax.shape[0], a_timestep, slice(0, 3)) - ) - + # audio self-attention + ashift_msa, ascale_msa = (self.get_ada_values(self.audio_scale_shift_table, ax.shape[0], a_timestep, slice(0, 2))) norm_ax = comfy.ldm.common_dit.rms_norm(ax) * (1 + ascale_msa) + ashift_msa - ax += ( - self.audio_attn1(norm_ax, pe=a_pe, transformer_options=transformer_options) - * agate_msa - ) - ax += self.audio_attn2( - comfy.ldm.common_dit.rms_norm(ax), - context=a_context, - mask=attention_mask, - transformer_options=transformer_options, - ) + del ashift_msa, ascale_msa + attn1_out = self.audio_attn1(norm_ax, pe=a_pe, transformer_options=transformer_options) + del norm_ax + # audio cross-attention + agate_msa = self.get_ada_values(self.audio_scale_shift_table, ax.shape[0], a_timestep, slice(2, 3))[0] + ax.addcmul_(attn1_out, agate_msa) + del agate_msa, attn1_out + ax.add_(self.audio_attn2(comfy.ldm.common_dit.rms_norm(ax), context=a_context, mask=attention_mask, transformer_options=transformer_options)) - del ashift_msa, ascale_msa, agate_msa - - # Audio - Video cross attention. + # video - audio cross attention. if run_a2v or run_v2a: - # norm3 vx_norm3 = comfy.ldm.common_dit.rms_norm(vx) ax_norm3 = comfy.ldm.common_dit.rms_norm(ax) - ( - scale_ca_audio_hidden_states_a2v, - shift_ca_audio_hidden_states_a2v, - scale_ca_audio_hidden_states_v2a, - shift_ca_audio_hidden_states_v2a, - gate_out_v2a, - ) = self.get_av_ca_ada_values( - self.scale_shift_table_a2v_ca_audio, - ax.shape[0], - a_cross_scale_shift_timestep, - a_cross_gate_timestep, - ) - - ( - scale_ca_video_hidden_states_a2v, - shift_ca_video_hidden_states_a2v, - scale_ca_video_hidden_states_v2a, - shift_ca_video_hidden_states_v2a, - gate_out_a2v, - ) = self.get_av_ca_ada_values( - self.scale_shift_table_a2v_ca_video, - vx.shape[0], - v_cross_scale_shift_timestep, - v_cross_gate_timestep, - ) - + # audio to video cross attention if run_a2v: - vx_scaled = ( - vx_norm3 * (1 + scale_ca_video_hidden_states_a2v) - + shift_ca_video_hidden_states_a2v - ) - ax_scaled = ( - ax_norm3 * (1 + scale_ca_audio_hidden_states_a2v) - + shift_ca_audio_hidden_states_a2v - ) - vx += ( - self.audio_to_video_attn( - vx_scaled, - context=ax_scaled, - pe=v_cross_pe, - k_pe=a_cross_pe, - transformer_options=transformer_options, - ) - * gate_out_a2v - ) + scale_ca_audio_hidden_states_a2v, shift_ca_audio_hidden_states_a2v = self.get_ada_values( + self.scale_shift_table_a2v_ca_audio[:4, :], ax.shape[0], a_cross_scale_shift_timestep)[:2] + scale_ca_video_hidden_states_a2v_v, shift_ca_video_hidden_states_a2v_v = self.get_ada_values( + self.scale_shift_table_a2v_ca_video[:4, :], vx.shape[0], v_cross_scale_shift_timestep)[:2] - del gate_out_a2v - del scale_ca_video_hidden_states_a2v,\ - shift_ca_video_hidden_states_a2v,\ - scale_ca_audio_hidden_states_a2v,\ - shift_ca_audio_hidden_states_a2v,\ + vx_scaled = vx_norm3 * (1 + scale_ca_video_hidden_states_a2v_v) + shift_ca_video_hidden_states_a2v_v + ax_scaled = ax_norm3 * (1 + scale_ca_audio_hidden_states_a2v) + shift_ca_audio_hidden_states_a2v + del scale_ca_video_hidden_states_a2v_v, shift_ca_video_hidden_states_a2v_v, scale_ca_audio_hidden_states_a2v, shift_ca_audio_hidden_states_a2v + a2v_out = self.audio_to_video_attn(vx_scaled, context=ax_scaled, pe=v_cross_pe, k_pe=a_cross_pe, transformer_options=transformer_options) + del vx_scaled, ax_scaled + + gate_out_a2v = self.get_ada_values(self.scale_shift_table_a2v_ca_video[4:, :], vx.shape[0], v_cross_gate_timestep)[0] + vx.addcmul_(a2v_out, gate_out_a2v) + del gate_out_a2v, a2v_out + + # video to audio cross attention if run_v2a: - ax_scaled = ( - ax_norm3 * (1 + scale_ca_audio_hidden_states_v2a) - + shift_ca_audio_hidden_states_v2a - ) - vx_scaled = ( - vx_norm3 * (1 + scale_ca_video_hidden_states_v2a) - + shift_ca_video_hidden_states_v2a - ) - ax += ( - self.video_to_audio_attn( - ax_scaled, - context=vx_scaled, - pe=a_cross_pe, - k_pe=v_cross_pe, - transformer_options=transformer_options, - ) - * gate_out_v2a - ) + scale_ca_audio_hidden_states_v2a, shift_ca_audio_hidden_states_v2a = self.get_ada_values( + self.scale_shift_table_a2v_ca_audio[:4, :], ax.shape[0], a_cross_scale_shift_timestep)[2:4] + scale_ca_video_hidden_states_v2a, shift_ca_video_hidden_states_v2a = self.get_ada_values( + self.scale_shift_table_a2v_ca_video[:4, :], vx.shape[0], v_cross_scale_shift_timestep)[2:4] - del gate_out_v2a - del scale_ca_video_hidden_states_v2a,\ - shift_ca_video_hidden_states_v2a,\ - scale_ca_audio_hidden_states_v2a,\ - shift_ca_audio_hidden_states_v2a + ax_scaled = ax_norm3 * (1 + scale_ca_audio_hidden_states_v2a) + shift_ca_audio_hidden_states_v2a + vx_scaled = vx_norm3 * (1 + scale_ca_video_hidden_states_v2a) + shift_ca_video_hidden_states_v2a + del scale_ca_video_hidden_states_v2a, shift_ca_video_hidden_states_v2a, scale_ca_audio_hidden_states_v2a, shift_ca_audio_hidden_states_v2a + v2a_out = self.video_to_audio_attn(ax_scaled, context=vx_scaled, pe=a_cross_pe, k_pe=v_cross_pe, transformer_options=transformer_options) + del ax_scaled, vx_scaled + + gate_out_v2a = self.get_ada_values(self.scale_shift_table_a2v_ca_audio[4:, :], ax.shape[0], a_cross_gate_timestep)[0] + ax.addcmul_(v2a_out, gate_out_v2a) + del gate_out_v2a, v2a_out + + del vx_norm3, ax_norm3 + + # video feedforward if run_vx: - vshift_mlp, vscale_mlp, vgate_mlp = ( - self.get_ada_values(self.scale_shift_table, vx.shape[0], v_timestep, slice(3, None)) - ) - + vshift_mlp, vscale_mlp = self.get_ada_values(self.scale_shift_table, vx.shape[0], v_timestep, slice(3, 5)) vx_scaled = comfy.ldm.common_dit.rms_norm(vx) * (1 + vscale_mlp) + vshift_mlp - vx += self.ff(vx_scaled) * vgate_mlp - del vshift_mlp, vscale_mlp, vgate_mlp + del vshift_mlp, vscale_mlp + ff_out = self.ff(vx_scaled) + del vx_scaled + + vgate_mlp = self.get_ada_values(self.scale_shift_table, vx.shape[0], v_timestep, slice(5, 6))[0] + vx.addcmul_(ff_out, vgate_mlp) + del vgate_mlp, ff_out + + # audio feedforward if run_ax: - ashift_mlp, ascale_mlp, agate_mlp = ( - self.get_ada_values(self.audio_scale_shift_table, ax.shape[0], a_timestep, slice(3, None)) - ) - + ashift_mlp, ascale_mlp = self.get_ada_values(self.audio_scale_shift_table, ax.shape[0], a_timestep, slice(3, 5)) ax_scaled = comfy.ldm.common_dit.rms_norm(ax) * (1 + ascale_mlp) + ashift_mlp - ax += self.audio_ff(ax_scaled) * agate_mlp + del ashift_mlp, ascale_mlp - del ashift_mlp, ascale_mlp, agate_mlp + ff_out = self.audio_ff(ax_scaled) + del ax_scaled + agate_mlp = self.get_ada_values(self.audio_scale_shift_table, ax.shape[0], a_timestep, slice(5, 6))[0] + ax.addcmul_(ff_out, agate_mlp) + del agate_mlp, ff_out return vx, ax @@ -589,9 +534,20 @@ class LTXAVModel(LTXVModel): audio_length = kwargs.get("audio_length", 0) # Separate audio and video latents vx, ax = self.separate_audio_and_video_latents(x, audio_length) + + has_spatial_mask = False + if denoise_mask is not None: + # check if any frame has spatial variation (inpainting) + for frame_idx in range(denoise_mask.shape[2]): + frame_mask = denoise_mask[0, 0, frame_idx] + if frame_mask.numel() > 0 and frame_mask.min() != frame_mask.max(): + has_spatial_mask = True + break + [vx, v_pixel_coords, additional_args] = super()._process_input( vx, keyframe_idxs, denoise_mask, **kwargs ) + additional_args["has_spatial_mask"] = has_spatial_mask ax, a_latent_coords = self.a_patchifier.patchify(ax) ax = self.audio_patchify_proj(ax) @@ -618,8 +574,9 @@ class LTXAVModel(LTXVModel): # Calculate patches_per_frame from orig_shape: [batch, channels, frames, height, width] # Video tokens are arranged as (frames * height * width), so patches_per_frame = height * width orig_shape = kwargs.get("orig_shape") + has_spatial_mask = kwargs.get("has_spatial_mask", None) v_patches_per_frame = None - if orig_shape is not None and len(orig_shape) == 5: + if not has_spatial_mask and orig_shape is not None and len(orig_shape) == 5: # orig_shape[3] = height, orig_shape[4] = width (in latent space) v_patches_per_frame = orig_shape[3] * orig_shape[4] @@ -662,10 +619,11 @@ class LTXAVModel(LTXVModel): ) # Compress cross-attention timesteps (only video side, audio is too small to benefit) + # v_patches_per_frame is None for spatial masks, set for temporal masks or no mask cross_av_timestep_ss = [ av_ca_audio_scale_shift_timestep.view(batch_size, -1, av_ca_audio_scale_shift_timestep.shape[-1]), - CompressedTimestep(av_ca_video_scale_shift_timestep.view(batch_size, -1, av_ca_video_scale_shift_timestep.shape[-1]), v_patches_per_frame), # video - compressed - CompressedTimestep(av_ca_a2v_gate_noise_timestep.view(batch_size, -1, av_ca_a2v_gate_noise_timestep.shape[-1]), v_patches_per_frame), # video - compressed + CompressedTimestep(av_ca_video_scale_shift_timestep.view(batch_size, -1, av_ca_video_scale_shift_timestep.shape[-1]), v_patches_per_frame), # video - compressed if possible + CompressedTimestep(av_ca_a2v_gate_noise_timestep.view(batch_size, -1, av_ca_a2v_gate_noise_timestep.shape[-1]), v_patches_per_frame), # video - compressed if possible av_ca_v2a_gate_noise_timestep.view(batch_size, -1, av_ca_v2a_gate_noise_timestep.shape[-1]), ] From e89b22993aa2e2b27f4ab1585754cee3a7ca1ff5 Mon Sep 17 00:00:00 2001 From: ComfyUI Wiki Date: Sat, 24 Jan 2026 04:27:49 +0800 Subject: [PATCH 506/647] Support ModelScope-Trainer/DiffSynth LoRA format for Flux.2 Klein models (#12042) --- comfy/lora.py | 1 + 1 file changed, 1 insertion(+) diff --git a/comfy/lora.py b/comfy/lora.py index e8246bd66..7b31d055c 100644 --- a/comfy/lora.py +++ b/comfy/lora.py @@ -260,6 +260,7 @@ def model_lora_keys_unet(model, key_map={}): key_map["transformer.{}".format(k[:-len(".weight")])] = to #simpletrainer and probably regular diffusers flux lora format key_map["lycoris_{}".format(k[:-len(".weight")].replace(".", "_"))] = to #simpletrainer lycoris key_map["lora_transformer_{}".format(k[:-len(".weight")].replace(".", "_"))] = to #onetrainer + key_map[k[:-len(".weight")]] = to #DiffSynth lora format for k in sdk: hidden_size = model.model_config.unet_config.get("hidden_size", 0) if k.endswith(".weight") and ".linear1." in k: From 9cf299a9f9488e4cb9b3f7cef3bc94c185c19f73 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Fri, 23 Jan 2026 16:50:48 -0800 Subject: [PATCH 507/647] Make regular empty latent node work properly on flux 2 variants. (#12050) --- comfy/latent_formats.py | 3 +++ comfy/sample.py | 12 +++++++++--- comfy_extras/nodes_custom_sampler.py | 6 ++++-- comfy_extras/nodes_sd3.py | 2 +- nodes.py | 5 +++-- 5 files changed, 20 insertions(+), 8 deletions(-) diff --git a/comfy/latent_formats.py b/comfy/latent_formats.py index cb4f52ce1..5600825ed 100644 --- a/comfy/latent_formats.py +++ b/comfy/latent_formats.py @@ -8,6 +8,7 @@ class LatentFormat: latent_rgb_factors_bias = None latent_rgb_factors_reshape = None taesd_decoder_name = None + spacial_downscale_ratio = 8 def process_in(self, latent): return latent * self.scale_factor @@ -181,6 +182,7 @@ class Flux(SD3): class Flux2(LatentFormat): latent_channels = 128 + spacial_downscale_ratio = 16 def __init__(self): self.latent_rgb_factors =[ @@ -749,6 +751,7 @@ class ACEAudio(LatentFormat): class ChromaRadiance(LatentFormat): latent_channels = 3 + spacial_downscale_ratio = 1 def __init__(self): self.latent_rgb_factors = [ diff --git a/comfy/sample.py b/comfy/sample.py index 2f8f3a51c..a2a39b527 100644 --- a/comfy/sample.py +++ b/comfy/sample.py @@ -37,12 +37,18 @@ def prepare_noise(latent_image, seed, noise_inds=None): return noises -def fix_empty_latent_channels(model, latent_image): +def fix_empty_latent_channels(model, latent_image, downscale_ratio_spacial=None): if latent_image.is_nested: return latent_image latent_format = model.get_model_object("latent_format") #Resize the empty latent image so it has the right number of channels - if latent_format.latent_channels != latent_image.shape[1] and torch.count_nonzero(latent_image) == 0: - latent_image = comfy.utils.repeat_to_batch_size(latent_image, latent_format.latent_channels, dim=1) + if torch.count_nonzero(latent_image) == 0: + if latent_format.latent_channels != latent_image.shape[1]: + latent_image = comfy.utils.repeat_to_batch_size(latent_image, latent_format.latent_channels, dim=1) + if downscale_ratio_spacial is not None: + if downscale_ratio_spacial != latent_format.spacial_downscale_ratio: + ratio = downscale_ratio_spacial / latent_format.spacial_downscale_ratio + latent_image = comfy.utils.common_upscale(latent_image, round(latent_image.shape[-1] * ratio), round(latent_image.shape[-2] * ratio), "nearest-exact", crop="disabled") + if latent_format.latent_dimensions == 3 and latent_image.ndim == 4: latent_image = latent_image.unsqueeze(2) return latent_image diff --git a/comfy_extras/nodes_custom_sampler.py b/comfy_extras/nodes_custom_sampler.py index 3eb40e937..a4d84ddf7 100644 --- a/comfy_extras/nodes_custom_sampler.py +++ b/comfy_extras/nodes_custom_sampler.py @@ -741,7 +741,7 @@ class SamplerCustom(io.ComfyNode): latent = latent_image latent_image = latent["samples"] latent = latent.copy() - latent_image = comfy.sample.fix_empty_latent_channels(model, latent_image) + latent_image = comfy.sample.fix_empty_latent_channels(model, latent_image, latent.get("downscale_ratio_spacial", None)) latent["samples"] = latent_image if not add_noise: @@ -760,6 +760,7 @@ class SamplerCustom(io.ComfyNode): samples = comfy.sample.sample_custom(model, noise, cfg, sampler, sigmas, positive, negative, latent_image, noise_mask=noise_mask, callback=callback, disable_pbar=disable_pbar, seed=noise_seed) out = latent.copy() + out.pop("downscale_ratio_spacial", None) out["samples"] = samples if "x0" in x0_output: x0_out = model.model.process_latent_out(x0_output["x0"].cpu()) @@ -939,7 +940,7 @@ class SamplerCustomAdvanced(io.ComfyNode): latent = latent_image latent_image = latent["samples"] latent = latent.copy() - latent_image = comfy.sample.fix_empty_latent_channels(guider.model_patcher, latent_image) + latent_image = comfy.sample.fix_empty_latent_channels(guider.model_patcher, latent_image, latent.get("downscale_ratio_spacial", None)) latent["samples"] = latent_image noise_mask = None @@ -954,6 +955,7 @@ class SamplerCustomAdvanced(io.ComfyNode): samples = samples.to(comfy.model_management.intermediate_device()) out = latent.copy() + out.pop("downscale_ratio_spacial", None) out["samples"] = samples if "x0" in x0_output: x0_out = guider.model_patcher.model.process_latent_out(x0_output["x0"].cpu()) diff --git a/comfy_extras/nodes_sd3.py b/comfy_extras/nodes_sd3.py index 02e5e7dd8..736213a47 100644 --- a/comfy_extras/nodes_sd3.py +++ b/comfy_extras/nodes_sd3.py @@ -55,7 +55,7 @@ class EmptySD3LatentImage(io.ComfyNode): @classmethod def execute(cls, width, height, batch_size=1) -> io.NodeOutput: latent = torch.zeros([batch_size, 16, height // 8, width // 8], device=comfy.model_management.intermediate_device()) - return io.NodeOutput({"samples":latent}) + return io.NodeOutput({"samples": latent, "downscale_ratio_spacial": 8}) generate = execute # TODO: remove diff --git a/nodes.py b/nodes.py index 158106686..b75247665 100644 --- a/nodes.py +++ b/nodes.py @@ -1230,7 +1230,7 @@ class EmptyLatentImage: def generate(self, width, height, batch_size=1): latent = torch.zeros([batch_size, 4, height // 8, width // 8], device=self.device) - return ({"samples":latent}, ) + return ({"samples": latent, "downscale_ratio_spacial": 8}, ) class LatentFromBatch: @@ -1538,7 +1538,7 @@ class SetLatentNoiseMask: def common_ksampler(model, seed, steps, cfg, sampler_name, scheduler, positive, negative, latent, denoise=1.0, disable_noise=False, start_step=None, last_step=None, force_full_denoise=False): latent_image = latent["samples"] - latent_image = comfy.sample.fix_empty_latent_channels(model, latent_image) + latent_image = comfy.sample.fix_empty_latent_channels(model, latent_image, latent.get("downscale_ratio_spacial", None)) if disable_noise: noise = torch.zeros(latent_image.size(), dtype=latent_image.dtype, layout=latent_image.layout, device="cpu") @@ -1556,6 +1556,7 @@ def common_ksampler(model, seed, steps, cfg, sampler_name, scheduler, positive, denoise=denoise, disable_noise=disable_noise, start_step=start_step, last_step=last_step, force_full_denoise=force_full_denoise, noise_mask=noise_mask, callback=callback, disable_pbar=disable_pbar, seed=seed) out = latent.copy() + out.pop("downscale_ratio_spacial", None) out["samples"] = samples return (out, ) From 4e6a1b66a93ef91848bc4bbf2a84e0ea98efcfc9 Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Fri, 23 Jan 2026 16:56:14 -0800 Subject: [PATCH 508/647] speed up and reduce VRAM of QWEN VAE and WAN (less so) (#12036) * ops: introduce autopad for conv3d This works around pytorch missing ability to causal pad as part of the kernel and avoids massive weight duplications for padding. * wan-vae: rework causal padding This currently uses F.pad which takes a full deep copy and is liable to be the VRAM peak. Instead, kick spatial padding back to the op and consolidate the temporal padding with the cat for the cache. * wan-vae: implement zero pad fast path The WAN VAE is also QWEN where it is used single-image. These convolutions are however zero padded 3d convolutions, which means the VAE is actually just 2D down the last element of the conv weight in the temporal dimension. Fast path this, to avoid adding zeros that then just evaporate in convoluton math but cost computation. --- comfy/ldm/wan/vae.py | 27 +++++++++++++++++---------- comfy/ops.py | 10 ++++++---- 2 files changed, 23 insertions(+), 14 deletions(-) diff --git a/comfy/ldm/wan/vae.py b/comfy/ldm/wan/vae.py index 08315f1a8..40e767213 100644 --- a/comfy/ldm/wan/vae.py +++ b/comfy/ldm/wan/vae.py @@ -5,7 +5,7 @@ import torch import torch.nn as nn import torch.nn.functional as F from einops import rearrange -from comfy.ldm.modules.diffusionmodules.model import vae_attention +from comfy.ldm.modules.diffusionmodules.model import vae_attention, torch_cat_if_needed import comfy.ops ops = comfy.ops.disable_weight_init @@ -20,22 +20,29 @@ class CausalConv3d(ops.Conv3d): def __init__(self, *args, **kwargs): super().__init__(*args, **kwargs) - self._padding = (self.padding[2], self.padding[2], self.padding[1], - self.padding[1], 2 * self.padding[0], 0) - self.padding = (0, 0, 0) + self._padding = 2 * self.padding[0] + self.padding = (0, self.padding[1], self.padding[2]) def forward(self, x, cache_x=None, cache_list=None, cache_idx=None): if cache_list is not None: cache_x = cache_list[cache_idx] cache_list[cache_idx] = None - padding = list(self._padding) - if cache_x is not None and self._padding[4] > 0: - cache_x = cache_x.to(x.device) - x = torch.cat([cache_x, x], dim=2) - padding[4] -= cache_x.shape[2] + if cache_x is None and x.shape[2] == 1: + #Fast path - the op will pad for use by truncating the weight + #and save math on a pile of zeros. + return super().forward(x, autopad="causal_zero") + + if self._padding > 0: + padding_needed = self._padding + if cache_x is not None: + cache_x = cache_x.to(x.device) + padding_needed = max(0, padding_needed - cache_x.shape[2]) + padding_shape = list(x.shape) + padding_shape[2] = padding_needed + padding = torch.zeros(padding_shape, device=x.device, dtype=x.dtype) + x = torch_cat_if_needed([padding, cache_x, x], dim=2) del cache_x - x = F.pad(x, padding) return super().forward(x) diff --git a/comfy/ops.py b/comfy/ops.py index 415c39e92..e406ba7ed 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -203,7 +203,9 @@ class disable_weight_init: def reset_parameters(self): return None - def _conv_forward(self, input, weight, bias, *args, **kwargs): + def _conv_forward(self, input, weight, bias, autopad=None, *args, **kwargs): + if autopad == "causal_zero": + weight = weight[:, :, -input.shape[2]:, :, :] if NVIDIA_MEMORY_CONV_BUG_WORKAROUND and weight.dtype in (torch.float16, torch.bfloat16): out = torch.cudnn_convolution(input, weight, self.padding, self.stride, self.dilation, self.groups, benchmark=False, deterministic=False, allow_tf32=True) if bias is not None: @@ -212,15 +214,15 @@ class disable_weight_init: else: return super()._conv_forward(input, weight, bias, *args, **kwargs) - def forward_comfy_cast_weights(self, input): + def forward_comfy_cast_weights(self, input, autopad=None): weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True) - x = self._conv_forward(input, weight, bias) + x = self._conv_forward(input, weight, bias, autopad=autopad) uncast_bias_weight(self, weight, bias, offload_stream) return x def forward(self, *args, **kwargs): run_every_op() - if self.comfy_cast_weights or len(self.weight_function) > 0 or len(self.bias_function) > 0: + if self.comfy_cast_weights or len(self.weight_function) > 0 or len(self.bias_function) > 0 or "autopad" in kwargs: return self.forward_comfy_cast_weights(*args, **kwargs) else: return super().forward(*args, **kwargs) From aef4e135889638812fc1ceab6f323d3441b48f5d Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sat, 24 Jan 2026 16:23:20 -0800 Subject: [PATCH 509/647] Make empty latent node work with other models. (#12062) --- comfy/latent_formats.py | 2 ++ 1 file changed, 2 insertions(+) diff --git a/comfy/latent_formats.py b/comfy/latent_formats.py index 5600825ed..38f18a83f 100644 --- a/comfy/latent_formats.py +++ b/comfy/latent_formats.py @@ -594,6 +594,7 @@ class Wan22(Wan21): class HunyuanImage21(LatentFormat): latent_channels = 64 latent_dimensions = 2 + spacial_downscale_ratio = 32 scale_factor = 0.75289 latent_rgb_factors = [ @@ -727,6 +728,7 @@ class HunyuanVideo15(LatentFormat): latent_rgb_factors_bias = [ 0.0456, -0.0202, -0.0644] latent_channels = 32 latent_dimensions = 3 + spacial_downscale_ratio = 16 scale_factor = 1.03682 taesd_decoder_name = "lighttaehy1_5" From bc72d7f8d11a664bc59941affc05a3f515239171 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Sun, 25 Jan 2026 03:10:09 +0200 Subject: [PATCH 510/647] [API Nodes] add TencentHunyuan3D nodes (#12026) * feat(api-nodes): add TencentHunyuan3D nodes * add "(Pro)" to display name --------- Co-authored-by: Jedrzej Kosinski --- comfy_api_nodes/apis/hunyuan3d.py | 66 ++++++ comfy_api_nodes/nodes_hunyuan3d.py | 297 +++++++++++++++++++++++++ comfy_api_nodes/nodes_kling.py | 1 - comfy_api_nodes/nodes_sora.py | 1 - comfy_api_nodes/nodes_topaz.py | 1 - comfy_api_nodes/util/__init__.py | 4 + comfy_api_nodes/util/client.py | 4 +- comfy_api_nodes/util/conversions.py | 15 ++ comfy_api_nodes/util/upload_helpers.py | 22 ++ 9 files changed, 406 insertions(+), 5 deletions(-) create mode 100644 comfy_api_nodes/apis/hunyuan3d.py create mode 100644 comfy_api_nodes/nodes_hunyuan3d.py diff --git a/comfy_api_nodes/apis/hunyuan3d.py b/comfy_api_nodes/apis/hunyuan3d.py new file mode 100644 index 000000000..6421c9bd5 --- /dev/null +++ b/comfy_api_nodes/apis/hunyuan3d.py @@ -0,0 +1,66 @@ +from typing import TypedDict + +from pydantic import BaseModel, Field, model_validator + + +class InputGenerateType(TypedDict): + generate_type: str + polygon_type: str + pbr: bool + + +class Hunyuan3DViewImage(BaseModel): + ViewType: str = Field(..., description="Valid values: back, left, right.") + ViewImageUrl: str = Field(...) + + +class To3DProTaskRequest(BaseModel): + Model: str = Field(...) + Prompt: str | None = Field(None) + ImageUrl: str | None = Field(None) + MultiViewImages: list[Hunyuan3DViewImage] | None = Field(None) + EnablePBR: bool | None = Field(...) + FaceCount: int | None = Field(...) + GenerateType: str | None = Field(...) + PolygonType: str | None = Field(...) + + +class RequestError(BaseModel): + Code: str = Field("") + Message: str = Field("") + + +class To3DProTaskCreateResponse(BaseModel): + JobId: str | None = Field(None) + Error: RequestError | None = Field(None) + + @model_validator(mode="before") + @classmethod + def unwrap_data(cls, values: dict) -> dict: + if "Response" in values and isinstance(values["Response"], dict): + return values["Response"] + return values + + +class ResultFile3D(BaseModel): + Type: str = Field(...) + Url: str = Field(...) + PreviewImageUrl: str = Field("") + + +class To3DProTaskResultResponse(BaseModel): + ErrorCode: str = Field("") + ErrorMessage: str = Field("") + ResultFile3Ds: list[ResultFile3D] = Field([]) + Status: str = Field(...) + + @model_validator(mode="before") + @classmethod + def unwrap_data(cls, values: dict) -> dict: + if "Response" in values and isinstance(values["Response"], dict): + return values["Response"] + return values + + +class To3DProTaskQueryRequest(BaseModel): + JobId: str = Field(...) diff --git a/comfy_api_nodes/nodes_hunyuan3d.py b/comfy_api_nodes/nodes_hunyuan3d.py new file mode 100644 index 000000000..b3a736643 --- /dev/null +++ b/comfy_api_nodes/nodes_hunyuan3d.py @@ -0,0 +1,297 @@ +import os + +from typing_extensions import override + +from comfy_api.latest import IO, ComfyExtension, Input +from comfy_api_nodes.apis.hunyuan3d import ( + Hunyuan3DViewImage, + InputGenerateType, + ResultFile3D, + To3DProTaskCreateResponse, + To3DProTaskQueryRequest, + To3DProTaskRequest, + To3DProTaskResultResponse, +) +from comfy_api_nodes.util import ( + ApiEndpoint, + download_url_to_bytesio, + downscale_image_tensor_by_max_side, + poll_op, + sync_op, + upload_image_to_comfyapi, + validate_image_dimensions, + validate_string, +) +from folder_paths import get_output_directory + + +def get_glb_obj_from_response(response_objs: list[ResultFile3D]) -> ResultFile3D: + for i in response_objs: + if i.Type.lower() == "glb": + return i + raise ValueError("No GLB file found in response. Please report this to the developers.") + + +class TencentTextToModelNode(IO.ComfyNode): + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="TencentTextToModelNode", + display_name="Hunyuan3D: Text to Model (Pro)", + category="api node/3d/Tencent", + inputs=[ + IO.Combo.Input( + "model", + options=["3.0", "3.1"], + tooltip="The LowPoly option is unavailable for the `3.1` model.", + ), + IO.String.Input("prompt", multiline=True, default="", tooltip="Supports up to 1024 characters."), + IO.Int.Input("face_count", default=500000, min=40000, max=1500000), + IO.DynamicCombo.Input( + "generate_type", + options=[ + IO.DynamicCombo.Option("Normal", [IO.Boolean.Input("pbr", default=False)]), + IO.DynamicCombo.Option( + "LowPoly", + [ + IO.Combo.Input("polygon_type", options=["triangle", "quadrilateral"]), + IO.Boolean.Input("pbr", default=False), + ], + ), + IO.DynamicCombo.Option("Geometry", []), + ], + ), + IO.Int.Input( + "seed", + default=0, + min=0, + max=2147483647, + display_mode=IO.NumberDisplay.number, + control_after_generate=True, + tooltip="Seed controls whether the node should re-run; " + "results are non-deterministic regardless of seed.", + ), + ], + outputs=[ + IO.String.Output(display_name="model_file"), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + is_output_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["generate_type", "generate_type.pbr", "face_count"]), + expr=""" + ( + $base := widgets.generate_type = "normal" ? 25 : widgets.generate_type = "lowpoly" ? 30 : 15; + $pbr := $lookup(widgets, "generate_type.pbr") ? 10 : 0; + $face := widgets.face_count != 500000 ? 10 : 0; + {"type":"usd","usd": ($base + $pbr + $face) * 0.02} + ) + """, + ), + ) + + @classmethod + async def execute( + cls, + model: str, + prompt: str, + face_count: int, + generate_type: InputGenerateType, + seed: int, + ) -> IO.NodeOutput: + _ = seed + validate_string(prompt, field_name="prompt", min_length=1, max_length=1024) + if model == "3.1" and generate_type["generate_type"].lower() == "lowpoly": + raise ValueError("The LowPoly option is currently unavailable for the 3.1 model.") + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/tencent/hunyuan/3d-pro", method="POST"), + response_model=To3DProTaskCreateResponse, + data=To3DProTaskRequest( + Model=model, + Prompt=prompt, + FaceCount=face_count, + GenerateType=generate_type["generate_type"], + EnablePBR=generate_type.get("pbr", None), + PolygonType=generate_type.get("polygon_type", None), + ), + ) + if response.Error: + raise ValueError(f"Task creation failed with code {response.Error.Code}: {response.Error.Message}") + result = await poll_op( + cls, + ApiEndpoint(path="/proxy/tencent/hunyuan/3d-pro/query", method="POST"), + data=To3DProTaskQueryRequest(JobId=response.JobId), + response_model=To3DProTaskResultResponse, + status_extractor=lambda r: r.Status, + ) + model_file = f"hunyuan_model_{response.JobId}.glb" + await download_url_to_bytesio( + get_glb_obj_from_response(result.ResultFile3Ds).Url, + os.path.join(get_output_directory(), model_file), + ) + return IO.NodeOutput(model_file) + + +class TencentImageToModelNode(IO.ComfyNode): + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="TencentImageToModelNode", + display_name="Hunyuan3D: Image(s) to Model (Pro)", + category="api node/3d/Tencent", + inputs=[ + IO.Combo.Input( + "model", + options=["3.0", "3.1"], + tooltip="The LowPoly option is unavailable for the `3.1` model.", + ), + IO.Image.Input("image"), + IO.Image.Input("image_left", optional=True), + IO.Image.Input("image_right", optional=True), + IO.Image.Input("image_back", optional=True), + IO.Int.Input("face_count", default=500000, min=40000, max=1500000), + IO.DynamicCombo.Input( + "generate_type", + options=[ + IO.DynamicCombo.Option("Normal", [IO.Boolean.Input("pbr", default=False)]), + IO.DynamicCombo.Option( + "LowPoly", + [ + IO.Combo.Input("polygon_type", options=["triangle", "quadrilateral"]), + IO.Boolean.Input("pbr", default=False), + ], + ), + IO.DynamicCombo.Option("Geometry", []), + ], + ), + IO.Int.Input( + "seed", + default=0, + min=0, + max=2147483647, + display_mode=IO.NumberDisplay.number, + control_after_generate=True, + tooltip="Seed controls whether the node should re-run; " + "results are non-deterministic regardless of seed.", + ), + ], + outputs=[ + IO.String.Output(display_name="model_file"), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + is_output_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends( + widgets=["generate_type", "generate_type.pbr", "face_count"], + inputs=["image_left", "image_right", "image_back"], + ), + expr=""" + ( + $base := widgets.generate_type = "normal" ? 25 : widgets.generate_type = "lowpoly" ? 30 : 15; + $multiview := ( + inputs.image_left.connected or inputs.image_right.connected or inputs.image_back.connected + ) ? 10 : 0; + $pbr := $lookup(widgets, "generate_type.pbr") ? 10 : 0; + $face := widgets.face_count != 500000 ? 10 : 0; + {"type":"usd","usd": ($base + $multiview + $pbr + $face) * 0.02} + ) + """, + ), + ) + + @classmethod + async def execute( + cls, + model: str, + image: Input.Image, + face_count: int, + generate_type: InputGenerateType, + seed: int, + image_left: Input.Image | None = None, + image_right: Input.Image | None = None, + image_back: Input.Image | None = None, + ) -> IO.NodeOutput: + _ = seed + if model == "3.1" and generate_type["generate_type"].lower() == "lowpoly": + raise ValueError("The LowPoly option is currently unavailable for the 3.1 model.") + validate_image_dimensions(image, min_width=128, min_height=128) + multiview_images = [] + for k, v in { + "left": image_left, + "right": image_right, + "back": image_back, + }.items(): + if v is None: + continue + validate_image_dimensions(v, min_width=128, min_height=128) + multiview_images.append( + Hunyuan3DViewImage( + ViewType=k, + ViewImageUrl=await upload_image_to_comfyapi( + cls, + downscale_image_tensor_by_max_side(v, max_side=4900), + mime_type="image/webp", + total_pixels=24_010_000, + ), + ) + ) + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/tencent/hunyuan/3d-pro", method="POST"), + response_model=To3DProTaskCreateResponse, + data=To3DProTaskRequest( + Model=model, + FaceCount=face_count, + GenerateType=generate_type["generate_type"], + ImageUrl=await upload_image_to_comfyapi( + cls, + downscale_image_tensor_by_max_side(image, max_side=4900), + mime_type="image/webp", + total_pixels=24_010_000, + ), + MultiViewImages=multiview_images if multiview_images else None, + EnablePBR=generate_type.get("pbr", None), + PolygonType=generate_type.get("polygon_type", None), + ), + ) + if response.Error: + raise ValueError(f"Task creation failed with code {response.Error.Code}: {response.Error.Message}") + result = await poll_op( + cls, + ApiEndpoint(path="/proxy/tencent/hunyuan/3d-pro/query", method="POST"), + data=To3DProTaskQueryRequest(JobId=response.JobId), + response_model=To3DProTaskResultResponse, + status_extractor=lambda r: r.Status, + ) + model_file = f"hunyuan_model_{response.JobId}.glb" + await download_url_to_bytesio( + get_glb_obj_from_response(result.ResultFile3Ds).Url, + os.path.join(get_output_directory(), model_file), + ) + return IO.NodeOutput(model_file) + + +class TencentHunyuan3DExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[IO.ComfyNode]]: + return [ + TencentTextToModelNode, + TencentImageToModelNode, + ] + + +async def comfy_entrypoint() -> TencentHunyuan3DExtension: + return TencentHunyuan3DExtension() diff --git a/comfy_api_nodes/nodes_kling.py b/comfy_api_nodes/nodes_kling.py index 3ec71530b..739fe1855 100644 --- a/comfy_api_nodes/nodes_kling.py +++ b/comfy_api_nodes/nodes_kling.py @@ -249,7 +249,6 @@ async def finish_omni_video_task(cls: type[IO.ComfyNode], response: TaskStatusRe ApiEndpoint(path=f"/proxy/kling/v1/videos/omni-video/{response.data.task_id}"), response_model=TaskStatusResponse, status_extractor=lambda r: (r.data.task_status if r.data else None), - max_poll_attempts=160, ) return IO.NodeOutput(await download_url_to_video_output(final_response.data.task_result.videos[0].url)) diff --git a/comfy_api_nodes/nodes_sora.py b/comfy_api_nodes/nodes_sora.py index 87e663845..afc18bb25 100644 --- a/comfy_api_nodes/nodes_sora.py +++ b/comfy_api_nodes/nodes_sora.py @@ -149,7 +149,6 @@ class OpenAIVideoSora2(IO.ComfyNode): response_model=Sora2GenerationResponse, status_extractor=lambda x: x.status, poll_interval=8.0, - max_poll_attempts=160, estimated_duration=int(45 * (duration / 4) * model_time_multiplier), ) return IO.NodeOutput( diff --git a/comfy_api_nodes/nodes_topaz.py b/comfy_api_nodes/nodes_topaz.py index c052e7656..8fccde25a 100644 --- a/comfy_api_nodes/nodes_topaz.py +++ b/comfy_api_nodes/nodes_topaz.py @@ -203,7 +203,6 @@ class TopazImageEnhance(IO.ComfyNode): progress_extractor=lambda x: getattr(x, "progress", 0), price_extractor=lambda x: x.credits * 0.08, poll_interval=8.0, - max_poll_attempts=160, estimated_duration=60, ) diff --git a/comfy_api_nodes/util/__init__.py b/comfy_api_nodes/util/__init__.py index 364976000..c3c9ff4bf 100644 --- a/comfy_api_nodes/util/__init__.py +++ b/comfy_api_nodes/util/__init__.py @@ -13,6 +13,7 @@ from .conversions import ( bytesio_to_image_tensor, convert_mask_to_image, downscale_image_tensor, + downscale_image_tensor_by_max_side, image_tensor_pair_to_batch, pil_to_bytesio, resize_mask_to_image, @@ -33,6 +34,7 @@ from .download_helpers import ( from .upload_helpers import ( upload_audio_to_comfyapi, upload_file_to_comfyapi, + upload_image_to_comfyapi, upload_images_to_comfyapi, upload_video_to_comfyapi, ) @@ -61,6 +63,7 @@ __all__ = [ # Upload helpers "upload_audio_to_comfyapi", "upload_file_to_comfyapi", + "upload_image_to_comfyapi", "upload_images_to_comfyapi", "upload_video_to_comfyapi", # Download helpers @@ -75,6 +78,7 @@ __all__ = [ "bytesio_to_image_tensor", "convert_mask_to_image", "downscale_image_tensor", + "downscale_image_tensor_by_max_side", "image_tensor_pair_to_batch", "pil_to_bytesio", "resize_mask_to_image", diff --git a/comfy_api_nodes/util/client.py b/comfy_api_nodes/util/client.py index f372ec7b5..8a1259506 100644 --- a/comfy_api_nodes/util/client.py +++ b/comfy_api_nodes/util/client.py @@ -141,7 +141,7 @@ async def poll_op( queued_statuses: list[str | int] | None = None, data: BaseModel | None = None, poll_interval: float = 5.0, - max_poll_attempts: int = 120, + max_poll_attempts: int = 160, timeout_per_poll: float = 120.0, max_retries_per_poll: int = 3, retry_delay_per_poll: float = 1.0, @@ -238,7 +238,7 @@ async def poll_op_raw( queued_statuses: list[str | int] | None = None, data: dict[str, Any] | BaseModel | None = None, poll_interval: float = 5.0, - max_poll_attempts: int = 120, + max_poll_attempts: int = 160, timeout_per_poll: float = 120.0, max_retries_per_poll: int = 3, retry_delay_per_poll: float = 1.0, diff --git a/comfy_api_nodes/util/conversions.py b/comfy_api_nodes/util/conversions.py index 546741b7b..0e15a0efe 100644 --- a/comfy_api_nodes/util/conversions.py +++ b/comfy_api_nodes/util/conversions.py @@ -144,6 +144,21 @@ def downscale_image_tensor(image: torch.Tensor, total_pixels: int = 1536 * 1024) return s +def downscale_image_tensor_by_max_side(image: torch.Tensor, *, max_side: int) -> torch.Tensor: + """Downscale input image tensor so the largest dimension is at most max_side pixels.""" + samples = image.movedim(-1, 1) + height, width = samples.shape[2], samples.shape[3] + max_dim = max(width, height) + if max_dim <= max_side: + return image + scale_by = max_side / max_dim + new_width = round(width * scale_by) + new_height = round(height * scale_by) + s = common_upscale(samples, new_width, new_height, "lanczos", "disabled") + s = s.movedim(1, -1) + return s + + def tensor_to_data_uri( image_tensor: torch.Tensor, total_pixels: int = 2048 * 2048, diff --git a/comfy_api_nodes/util/upload_helpers.py b/comfy_api_nodes/util/upload_helpers.py index 2794be35c..2190f9639 100644 --- a/comfy_api_nodes/util/upload_helpers.py +++ b/comfy_api_nodes/util/upload_helpers.py @@ -88,6 +88,28 @@ async def upload_images_to_comfyapi( return download_urls +async def upload_image_to_comfyapi( + cls: type[IO.ComfyNode], + image: torch.Tensor, + *, + mime_type: str | None = None, + wait_label: str | None = "Uploading", + total_pixels: int = 2048 * 2048, +) -> str: + """Uploads a single image to ComfyUI API and returns its download URL.""" + return ( + await upload_images_to_comfyapi( + cls, + image, + max_images=1, + mime_type=mime_type, + wait_label=wait_label, + show_batch_index=False, + total_pixels=total_pixels, + ) + )[0] + + async def upload_audio_to_comfyapi( cls: type[IO.ComfyNode], audio: Input.Audio, From ed6002cb60e0709a493d4d8f56793ce0bce12e7e Mon Sep 17 00:00:00 2001 From: pythongosssss <125205205+pythongosssss@users.noreply.github.com> Date: Sat, 24 Jan 2026 17:30:40 -0800 Subject: [PATCH 511/647] add support for kwargs inputs to allow arbitrary inputs from frontend (#12063) used to output selected combo index Co-authored-by: Jedrzej Kosinski --- comfy_api/latest/_io.py | 12 ++++++++++++ comfy_extras/nodes_logic.py | 12 ++++++++---- execution.py | 2 +- 3 files changed, 21 insertions(+), 5 deletions(-) diff --git a/comfy_api/latest/_io.py b/comfy_api/latest/_io.py index 2ec8d6e4b..03c77a531 100644 --- a/comfy_api/latest/_io.py +++ b/comfy_api/latest/_io.py @@ -1383,6 +1383,8 @@ class Schema: """Flags a node as not idempotent; when True, the node will run and not reuse the cached outputs when identical inputs are provided on a different node in the graph.""" enable_expand: bool=False """Flags a node as expandable, allowing NodeOutput to include 'expand' property.""" + accept_all_inputs: bool=False + """When True, all inputs from the prompt will be passed to the node as kwargs, even if not defined in the schema.""" def validate(self): '''Validate the schema: @@ -1853,6 +1855,14 @@ class _ComfyNodeBaseInternal(_ComfyNodeInternal): cls.GET_SCHEMA() return cls._NOT_IDEMPOTENT + _ACCEPT_ALL_INPUTS = None + @final + @classproperty + def ACCEPT_ALL_INPUTS(cls): # noqa + if cls._ACCEPT_ALL_INPUTS is None: + cls.GET_SCHEMA() + return cls._ACCEPT_ALL_INPUTS + @final @classmethod def INPUT_TYPES(cls) -> dict[str, dict]: @@ -1891,6 +1901,8 @@ class _ComfyNodeBaseInternal(_ComfyNodeInternal): cls._INPUT_IS_LIST = schema.is_input_list if cls._NOT_IDEMPOTENT is None: cls._NOT_IDEMPOTENT = schema.not_idempotent + if cls._ACCEPT_ALL_INPUTS is None: + cls._ACCEPT_ALL_INPUTS = schema.accept_all_inputs if cls._RETURN_TYPES is None: output = [] diff --git a/comfy_extras/nodes_logic.py b/comfy_extras/nodes_logic.py index 1ed060205..c066064ac 100644 --- a/comfy_extras/nodes_logic.py +++ b/comfy_extras/nodes_logic.py @@ -104,19 +104,23 @@ class CustomComboNode(io.ComfyNode): category="utils", is_experimental=True, inputs=[io.Combo.Input("choice", options=[])], - outputs=[io.String.Output()] + outputs=[ + io.String.Output(display_name="STRING"), + io.Int.Output(display_name="INDEX"), + ], + accept_all_inputs=True, ) @classmethod - def validate_inputs(cls, choice: io.Combo.Type) -> bool: + def validate_inputs(cls, choice: io.Combo.Type, index: int = 0, **kwargs) -> bool: # NOTE: DO NOT DO THIS unless you want to skip validation entirely on the node's inputs. # I am doing that here because the widgets (besides the combo dropdown) on this node are fully frontend defined. # I need to skip checking that the chosen combo option is in the options list, since those are defined by the user. return True @classmethod - def execute(cls, choice: io.Combo.Type) -> io.NodeOutput: - return io.NodeOutput(choice) + def execute(cls, choice: io.Combo.Type, index: int = 0, **kwargs) -> io.NodeOutput: + return io.NodeOutput(choice, index) class DCTestNode(io.ComfyNode): diff --git a/execution.py b/execution.py index 648f204ec..4b4f63c80 100644 --- a/execution.py +++ b/execution.py @@ -175,7 +175,7 @@ def get_input_data(inputs, class_def, unique_id, execution_list=None, dynprompt= continue obj = cached.outputs[output_index] input_data_all[x] = obj - elif input_category is not None: + elif input_category is not None or (is_v3 and class_def.ACCEPT_ALL_INPUTS): input_data_all[x] = [input_data] if is_v3: From 635406e283e9c0c8964f2fde3ff1ff4a8b31201e Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sat, 24 Jan 2026 19:32:28 -0800 Subject: [PATCH 512/647] Only enable fp16 on z image models that actually support it. (#12065) --- comfy/ldm/lumina/model.py | 1 + comfy/model_detection.py | 4 ++++ comfy/supported_models.py | 2 +- 3 files changed, 6 insertions(+), 1 deletion(-) diff --git a/comfy/ldm/lumina/model.py b/comfy/ldm/lumina/model.py index b114d9e31..77d1abc97 100644 --- a/comfy/ldm/lumina/model.py +++ b/comfy/ldm/lumina/model.py @@ -451,6 +451,7 @@ class NextDiT(nn.Module): device=None, dtype=None, operations=None, + **kwargs, ) -> None: super().__init__() self.dtype = dtype diff --git a/comfy/model_detection.py b/comfy/model_detection.py index b29a033cc..8cea16e50 100644 --- a/comfy/model_detection.py +++ b/comfy/model_detection.py @@ -444,6 +444,10 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): dit_config["ffn_dim_multiplier"] = (8.0 / 3.0) dit_config["z_image_modulation"] = True dit_config["time_scale"] = 1000.0 + try: + dit_config["allow_fp16"] = torch.std(state_dict['{}layers.{}.ffn_norm1.weight'.format(key_prefix, dit_config["n_layers"] - 2)], unbiased=False).item() < 0.42 + except Exception: + pass if '{}cap_pad_token'.format(key_prefix) in state_dict_keys: dit_config["pad_tokens_multiple"] = 32 sig_weight = state_dict.get('{}siglip_embedder.0.weight'.format(key_prefix), None) diff --git a/comfy/supported_models.py b/comfy/supported_models.py index 45d913fa6..d25271d6e 100644 --- a/comfy/supported_models.py +++ b/comfy/supported_models.py @@ -1093,7 +1093,7 @@ class ZImage(Lumina2): def __init__(self, unet_config): super().__init__(unet_config) - if comfy.model_management.extended_fp16_support(): + if comfy.model_management.extended_fp16_support() and unet_config.get("allow_fp16", False): self.supported_inference_dtypes = self.supported_inference_dtypes.copy() self.supported_inference_dtypes.insert(1, torch.float16) From a97c98068f6301b1f87ce89e7bd942ee2db3155d Mon Sep 17 00:00:00 2001 From: Kohaku-Blueleaf <59680068+KohakuBlueleaf@users.noreply.github.com> Date: Sun, 25 Jan 2026 11:56:22 +0800 Subject: [PATCH 513/647] [Weight-adapter/Trainer] Bypass forward mode in Weight adapter system (#11958) * Add API of bypass forward module * bypass implementation * add bypass fwd into nodes list/trainer --- comfy/sd.py | 100 +++++++ comfy/weight_adapter/__init__.py | 8 + comfy/weight_adapter/base.py | 231 +++++++++++++++- comfy/weight_adapter/boft.py | 119 ++++++++- comfy/weight_adapter/bypass.py | 437 +++++++++++++++++++++++++++++++ comfy/weight_adapter/glora.py | 219 +++++++++++++++- comfy/weight_adapter/loha.py | 186 +++++++++++-- comfy/weight_adapter/lokr.py | 311 ++++++++++++++++++++-- comfy/weight_adapter/lora.py | 165 +++++++++++- comfy/weight_adapter/oft.py | 186 ++++++++++++- comfy_extras/nodes_train.py | 111 +++++++- nodes.py | 67 +++++ 12 files changed, 2039 insertions(+), 101 deletions(-) create mode 100644 comfy/weight_adapter/bypass.py diff --git a/comfy/sd.py b/comfy/sd.py index ce7e6bcff..f627f7d55 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -20,6 +20,7 @@ import comfy.ldm.ace.vae.music_dcae_pipeline import comfy.ldm.hunyuan_video.vae import comfy.ldm.mmaudio.vae.autoencoder import comfy.pixel_space_convert +import comfy.weight_adapter import yaml import math import os @@ -101,6 +102,105 @@ def load_lora_for_models(model, clip, lora, strength_model, strength_clip): return (new_modelpatcher, new_clip) +def load_bypass_lora_for_models(model, clip, lora, strength_model, strength_clip): + """ + Load LoRA in bypass mode without modifying base model weights. + + Instead of patching weights, this injects the LoRA computation into the + forward pass: output = base_forward(x) + lora_path(x) + + Non-adapter patches (bias diff, weight diff, etc.) are applied as regular patches. + + This is useful for training and when model weights are offloaded. + """ + key_map = {} + if model is not None: + key_map = comfy.lora.model_lora_keys_unet(model.model, key_map) + if clip is not None: + key_map = comfy.lora.model_lora_keys_clip(clip.cond_stage_model, key_map) + + logging.debug(f"[BypassLoRA] key_map has {len(key_map)} entries") + + lora = comfy.lora_convert.convert_lora(lora) + loaded = comfy.lora.load_lora(lora, key_map) + + logging.debug(f"[BypassLoRA] loaded has {len(loaded)} entries") + + # Separate adapters (for bypass) from other patches (for regular patching) + bypass_patches = {} # WeightAdapterBase instances -> bypass mode + regular_patches = {} # diff, set, bias patches -> regular weight patching + + for key, patch_data in loaded.items(): + if isinstance(patch_data, comfy.weight_adapter.WeightAdapterBase): + bypass_patches[key] = patch_data + else: + regular_patches[key] = patch_data + + logging.debug(f"[BypassLoRA] {len(bypass_patches)} bypass adapters, {len(regular_patches)} regular patches") + + k = set() + k1 = set() + + if model is not None: + new_modelpatcher = model.clone() + + # Apply regular patches (bias diff, weight diff, etc.) via normal patching + if regular_patches: + patched_keys = new_modelpatcher.add_patches(regular_patches, strength_model) + k.update(patched_keys) + + # Apply adapter patches via bypass injection + manager = comfy.weight_adapter.BypassInjectionManager() + model_sd_keys = set(new_modelpatcher.model.state_dict().keys()) + + for key, adapter in bypass_patches.items(): + if key in model_sd_keys: + manager.add_adapter(key, adapter, strength=strength_model) + k.add(key) + else: + logging.warning(f"[BypassLoRA] Adapter key not in model state_dict: {key}") + + injections = manager.create_injections(new_modelpatcher.model) + + if manager.get_hook_count() > 0: + new_modelpatcher.set_injections("bypass_lora", injections) + else: + new_modelpatcher = None + + if clip is not None: + new_clip = clip.clone() + + # Apply regular patches to clip + if regular_patches: + patched_keys = new_clip.add_patches(regular_patches, strength_clip) + k1.update(patched_keys) + + # Apply adapter patches via bypass injection + clip_manager = comfy.weight_adapter.BypassInjectionManager() + clip_sd_keys = set(new_clip.cond_stage_model.state_dict().keys()) + + for key, adapter in bypass_patches.items(): + if key in clip_sd_keys: + clip_manager.add_adapter(key, adapter, strength=strength_clip) + k1.add(key) + + clip_injections = clip_manager.create_injections(new_clip.cond_stage_model) + if clip_manager.get_hook_count() > 0: + new_clip.patcher.set_injections("bypass_lora", clip_injections) + else: + new_clip = None + + for x in loaded: + if (x not in k) and (x not in k1): + patch_data = loaded[x] + patch_type = type(patch_data).__name__ + if isinstance(patch_data, tuple): + patch_type = f"tuple({patch_data[0]})" + logging.warning(f"NOT LOADED: {x} (type={patch_type})") + + return (new_modelpatcher, new_clip) + + class CLIP: def __init__(self, target=None, embedding_directory=None, no_init=False, tokenizer_data={}, parameters=0, state_dict=[], model_options={}): if no_init: diff --git a/comfy/weight_adapter/__init__.py b/comfy/weight_adapter/__init__.py index b40f920e4..b9fa8d5cf 100644 --- a/comfy/weight_adapter/__init__.py +++ b/comfy/weight_adapter/__init__.py @@ -5,6 +5,11 @@ from .lokr import LoKrAdapter from .glora import GLoRAAdapter from .oft import OFTAdapter from .boft import BOFTAdapter +from .bypass import ( + BypassInjectionManager, + BypassForwardHook, + create_bypass_injections_from_patches, +) adapters: list[type[WeightAdapterBase]] = [ @@ -31,4 +36,7 @@ __all__ = [ "WeightAdapterTrainBase", "adapters", "adapter_maps", + "BypassInjectionManager", + "BypassForwardHook", + "create_bypass_injections_from_patches", ] + [a.__name__ for a in adapters] diff --git a/comfy/weight_adapter/base.py b/comfy/weight_adapter/base.py index 43644b106..bce89a0e2 100644 --- a/comfy/weight_adapter/base.py +++ b/comfy/weight_adapter/base.py @@ -1,4 +1,4 @@ -from typing import Optional +from typing import Callable, Optional import torch import torch.nn as nn @@ -7,12 +7,35 @@ import comfy.model_management class WeightAdapterBase: + """ + Base class for weight adapters (LoRA, LoHa, LoKr, OFT, etc.) + + Bypass Mode: + All adapters follow the pattern: bypass(f)(x) = g(f(x) + h(x)) + + - h(x): Additive component (LoRA path). Returns delta to add to base output. + - g(y): Output transformation. Applied after base + h(x). + + For LoRA/LoHa/LoKr: g = identity, h = adapter(x) + For OFT/BOFT: g = transform, h = 0 + """ + name: str loaded_keys: set[str] weights: list[torch.Tensor] + # Attributes set by bypass system + multiplier: float = 1.0 + shape: tuple = None # (out_features, in_features) or (out_ch, in_ch, *kernel) + @classmethod - def load(cls, x: str, lora: dict[str, torch.Tensor], alpha: float, dora_scale: torch.Tensor) -> Optional["WeightAdapterBase"]: + def load( + cls, + x: str, + lora: dict[str, torch.Tensor], + alpha: float, + dora_scale: torch.Tensor, + ) -> Optional["WeightAdapterBase"]: raise NotImplementedError def to_train(self) -> "WeightAdapterTrainBase": @@ -39,18 +62,202 @@ class WeightAdapterBase: ): raise NotImplementedError + # ===== Bypass Mode Methods ===== + # + # IMPORTANT: Bypass mode is designed for quantized models where original weights + # may not be accessible in a usable format. Therefore, h() and bypass_forward() + # do NOT take org_weight as a parameter. All necessary information (out_channels, + # in_channels, conv params, etc.) is provided via attributes set by BypassForwardHook. + + def h(self, x: torch.Tensor, base_out: torch.Tensor) -> torch.Tensor: + """ + Additive bypass component: h(x, base_out) + + Computes the adapter's contribution to be added to base forward output. + For adapters that only transform output (OFT/BOFT), returns zeros. + + Note: + This method does NOT access original model weights. Bypass mode is + designed for quantized models where weights may not be in a usable format. + All shape info comes from module attributes set by BypassForwardHook. + + Args: + x: Input tensor + base_out: Output from base forward f(x), can be used for shape reference + + Returns: + Delta tensor to add to base output. Shape matches base output. + + Reference: LyCORIS LoConModule.bypass_forward_diff + """ + # Default: no additive component (for OFT/BOFT) + # Simply return zeros matching base_out shape + return torch.zeros_like(base_out) + + def g(self, y: torch.Tensor) -> torch.Tensor: + """ + Output transformation: g(y) + + Applied after base forward + h(x). For most adapters this is identity. + OFT/BOFT override this to apply orthogonal transformation. + + Args: + y: Combined output (base + h(x)) + + Returns: + Transformed output + + Reference: LyCORIS OFTModule applies orthogonal transform here + """ + # Default: identity (for LoRA/LoHa/LoKr) + return y + + def bypass_forward( + self, + org_forward: Callable, + x: torch.Tensor, + *args, + **kwargs, + ) -> torch.Tensor: + """ + Full bypass forward: g(f(x) + h(x, f(x))) + + Note: + This method does NOT take org_weight/org_bias parameters. Bypass mode + is designed for quantized models where weights may not be accessible. + The original forward function handles weight access internally. + + Args: + org_forward: Original module forward function + x: Input tensor + *args, **kwargs: Additional arguments for org_forward + + Returns: + Output with adapter applied in bypass mode + + Reference: LyCORIS LoConModule.bypass_forward + """ + # Base forward: f(x) + base_out = org_forward(x, *args, **kwargs) + + # Additive component: h(x, base_out) - base_out provided for shape reference + h_out = self.h(x, base_out) + + # Output transformation: g(base + h) + return self.g(base_out + h_out) + class WeightAdapterTrainBase(nn.Module): - # We follow the scheme of PR #7032 + """ + Base class for trainable weight adapters (LoRA, LoHa, LoKr, OFT, etc.) + + Bypass Mode: + All adapters follow the pattern: bypass(f)(x) = g(f(x) + h(x)) + + - h(x): Additive component (LoRA path). Returns delta to add to base output. + - g(y): Output transformation. Applied after base + h(x). + + For LoRA/LoHa/LoKr: g = identity, h = adapter(x) + For OFT: g = transform, h = 0 + + Note: + Unlike WeightAdapterBase, TrainBase classes have simplified weight formats + with fewer branches (e.g., LoKr only has w1/w2, not w1_a/w1_b decomposition). + + We follow the scheme of PR #7032 + """ + + # Attributes set by bypass system (BypassForwardHook) + # These are set before h()/g()/bypass_forward() are called + multiplier: float = 1.0 + is_conv: bool = False + conv_dim: int = 0 # 0=linear, 1=conv1d, 2=conv2d, 3=conv3d + kw_dict: dict = {} # Conv kwargs: stride, padding, dilation, groups + kernel_size: tuple = () + in_channels: int = None + out_channels: int = None + def __init__(self): super().__init__() def __call__(self, w): """ - w: The original weight tensor to be modified. + Weight modification mode: returns modified weight. + + Args: + w: The original weight tensor to be modified. + + Returns: + Modified weight tensor. """ raise NotImplementedError + # ===== Bypass Mode Methods ===== + + def h(self, x: torch.Tensor, base_out: torch.Tensor) -> torch.Tensor: + """ + Additive bypass component: h(x, base_out) + + Computes the adapter's contribution to be added to base forward output. + For adapters that only transform output (OFT), returns zeros. + + Args: + x: Input tensor + base_out: Output from base forward f(x), can be used for shape reference + + Returns: + Delta tensor to add to base output. Shape matches base output. + + Subclasses should override this method. + """ + raise NotImplementedError( + f"{self.__class__.__name__}.h() not implemented. " + "Subclasses must implement h() for bypass mode." + ) + + def g(self, y: torch.Tensor) -> torch.Tensor: + """ + Output transformation: g(y) + + Applied after base forward + h(x). For most adapters this is identity. + OFT overrides this to apply orthogonal transformation. + + Args: + y: Combined output (base + h(x)) + + Returns: + Transformed output + """ + # Default: identity (for LoRA/LoHa/LoKr) + return y + + def bypass_forward( + self, + org_forward: Callable, + x: torch.Tensor, + *args, + **kwargs, + ) -> torch.Tensor: + """ + Full bypass forward: g(f(x) + h(x, f(x))) + + Args: + org_forward: Original module forward function + x: Input tensor + *args, **kwargs: Additional arguments for org_forward + + Returns: + Output with adapter applied in bypass mode + """ + # Base forward: f(x) + base_out = org_forward(x, *args, **kwargs) + + # Additive component: h(x, base_out) - base_out provided for shape reference + h_out = self.h(x, base_out) + + # Output transformation: g(base + h) + return self.g(base_out + h_out) + def passive_memory_usage(self): raise NotImplementedError("passive_memory_usage is not implemented") @@ -59,8 +266,12 @@ class WeightAdapterTrainBase(nn.Module): return self.passive_memory_usage() -def weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype, function): - dora_scale = comfy.model_management.cast_to_device(dora_scale, weight.device, intermediate_dtype) +def weight_decompose( + dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype, function +): + dora_scale = comfy.model_management.cast_to_device( + dora_scale, weight.device, intermediate_dtype + ) lora_diff *= alpha weight_calc = weight + function(lora_diff).type(weight.dtype) @@ -106,10 +317,14 @@ def pad_tensor_to_shape(tensor: torch.Tensor, new_shape: list[int]) -> torch.Ten the original tensor will be truncated in that dimension. """ if any([new_shape[i] < tensor.shape[i] for i in range(len(new_shape))]): - raise ValueError("The new shape must be larger than the original tensor in all dimensions") + raise ValueError( + "The new shape must be larger than the original tensor in all dimensions" + ) if len(new_shape) != len(tensor.shape): - raise ValueError("The new shape must have the same number of dimensions as the original tensor") + raise ValueError( + "The new shape must have the same number of dimensions as the original tensor" + ) # Create a new tensor filled with zeros padded_tensor = torch.zeros(new_shape, dtype=tensor.dtype, device=tensor.device) diff --git a/comfy/weight_adapter/boft.py b/comfy/weight_adapter/boft.py index b2a2f1bd4..02a8dc130 100644 --- a/comfy/weight_adapter/boft.py +++ b/comfy/weight_adapter/boft.py @@ -62,9 +62,13 @@ class BOFTAdapter(WeightAdapterBase): alpha = v[2] dora_scale = v[3] - blocks = comfy.model_management.cast_to_device(blocks, weight.device, intermediate_dtype) + blocks = comfy.model_management.cast_to_device( + blocks, weight.device, intermediate_dtype + ) if rescale is not None: - rescale = comfy.model_management.cast_to_device(rescale, weight.device, intermediate_dtype) + rescale = comfy.model_management.cast_to_device( + rescale, weight.device, intermediate_dtype + ) boft_m, block_num, boft_b, *_ = blocks.shape @@ -74,7 +78,7 @@ class BOFTAdapter(WeightAdapterBase): # for Q = -Q^T q = blocks - blocks.transpose(-1, -2) normed_q = q - if alpha > 0: # alpha in boft/bboft is for constraint + if alpha > 0: # alpha in boft/bboft is for constraint q_norm = torch.norm(q) + 1e-8 if q_norm > alpha: normed_q = q * alpha / q_norm @@ -83,13 +87,13 @@ class BOFTAdapter(WeightAdapterBase): r = r.to(weight) inp = org = weight - r_b = boft_b//2 + r_b = boft_b // 2 for i in range(boft_m): bi = r[i] g = 2 k = 2**i * r_b if strength != 1: - bi = bi * strength + (1-strength) * I + bi = bi * strength + (1 - strength) * I inp = ( inp.unflatten(0, (-1, g, k)) .transpose(1, 2) @@ -98,18 +102,117 @@ class BOFTAdapter(WeightAdapterBase): ) inp = torch.einsum("b i j, b j ...-> b i ...", bi, inp) inp = ( - inp.flatten(0, 1).unflatten(0, (-1, k, g)).transpose(1, 2).flatten(0, 2) + inp.flatten(0, 1) + .unflatten(0, (-1, k, g)) + .transpose(1, 2) + .flatten(0, 2) ) if rescale is not None: inp = inp * rescale lora_diff = inp - org - lora_diff = comfy.model_management.cast_to_device(lora_diff, weight.device, intermediate_dtype) + lora_diff = comfy.model_management.cast_to_device( + lora_diff, weight.device, intermediate_dtype + ) if dora_scale is not None: - weight = weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype, function) + weight = weight_decompose( + dora_scale, + weight, + lora_diff, + alpha, + strength, + intermediate_dtype, + function, + ) else: weight += function((strength * lora_diff).type(weight.dtype)) except Exception as e: logging.error("ERROR {} {} {}".format(self.name, key, e)) return weight + + def _get_orthogonal_matrices(self, device, dtype): + """Compute the orthogonal rotation matrices R from BOFT blocks.""" + v = self.weights + blocks = v[0].to(device=device, dtype=dtype) + alpha = v[2] + if alpha is None: + alpha = 0 + + boft_m, block_num, boft_b, _ = blocks.shape + I = torch.eye(boft_b, device=device, dtype=dtype) + + # Q = blocks - blocks^T (skew-symmetric) + q = blocks - blocks.transpose(-1, -2) + normed_q = q + + # Apply constraint if alpha > 0 + if alpha > 0: + q_norm = torch.norm(q) + 1e-8 + if q_norm > alpha: + normed_q = q * alpha / q_norm + + # Cayley transform: R = (I + Q)(I - Q)^-1 + r = (I + normed_q) @ (I - normed_q).float().inverse() + return r, boft_m, boft_b + + def g(self, y: torch.Tensor) -> torch.Tensor: + """ + Output transformation for BOFT: applies butterfly orthogonal transform. + + BOFT uses multiple stages of butterfly-structured orthogonal transforms. + + Reference: LyCORIS ButterflyOFTModule._bypass_forward + """ + v = self.weights + rescale = v[1] + + r, boft_m, boft_b = self._get_orthogonal_matrices(y.device, y.dtype) + r_b = boft_b // 2 + + # Apply multiplier + multiplier = getattr(self, "multiplier", 1.0) + I = torch.eye(boft_b, device=y.device, dtype=y.dtype) + + # Use module info from bypass injection to determine conv vs linear + is_conv = getattr(self, "is_conv", y.dim() > 2) + + if is_conv: + # Conv output: (N, C, H, W, ...) -> transpose to (N, H, W, ..., C) + y = y.transpose(1, -1) + + # Apply butterfly transform stages + inp = y + for i in range(boft_m): + bi = r[i] # (block_num, boft_b, boft_b) + g = 2 + k = 2**i * r_b + + # Interpolate with identity based on multiplier + if multiplier != 1: + bi = bi * multiplier + (1 - multiplier) * I + + # Reshape for butterfly: unflatten last dim, transpose, flatten, unflatten + inp = ( + inp.unflatten(-1, (-1, g, k)) + .transpose(-2, -1) + .flatten(-3) + .unflatten(-1, (-1, boft_b)) + ) + # Apply block-diagonal orthogonal transform + inp = torch.einsum("b i j, ... b j -> ... b i", bi, inp) + # Reshape back + inp = ( + inp.flatten(-2).unflatten(-1, (-1, k, g)).transpose(-2, -1).flatten(-3) + ) + + # Apply rescale if present + if rescale is not None: + rescale = rescale.to(device=y.device, dtype=y.dtype) + inp = inp * rescale.transpose(0, -1) + + if is_conv: + # Transpose back: (N, H, W, ..., C) -> (N, C, H, W, ...) + inp = inp.transpose(1, -1) + + return inp diff --git a/comfy/weight_adapter/bypass.py b/comfy/weight_adapter/bypass.py new file mode 100644 index 000000000..d4aaf98ca --- /dev/null +++ b/comfy/weight_adapter/bypass.py @@ -0,0 +1,437 @@ +""" +Bypass mode implementation for weight adapters (LoRA, LoKr, LoHa, etc.) + +Bypass mode applies adapters during forward pass without modifying base weights: + bypass(f)(x) = g(f(x) + h(x)) + +Where: + - f(x): Original layer forward + - h(x): Additive component from adapter (LoRA path) + - g(y): Output transformation (identity for most adapters) + +This is useful for: + - Training with gradient checkpointing + - Avoiding weight modifications when weights are offloaded + - Supporting multiple adapters with different strengths dynamically +""" + +import logging +from typing import Optional, Union + +import torch +import torch.nn as nn + +from .base import WeightAdapterBase, WeightAdapterTrainBase +from comfy.patcher_extension import PatcherInjection + +# Type alias for adapters that support bypass mode +BypassAdapter = Union[WeightAdapterBase, WeightAdapterTrainBase] + + +def get_module_type_info(module: nn.Module) -> dict: + """ + Determine module type and extract conv parameters from module class. + + This is more reliable than checking weight.ndim, especially for quantized layers + where weight shape might be different. + + Returns: + dict with keys: is_conv, conv_dim, stride, padding, dilation, groups + """ + info = { + "is_conv": False, + "conv_dim": 0, + "stride": (1,), + "padding": (0,), + "dilation": (1,), + "groups": 1, + "kernel_size": (1,), + "in_channels": None, + "out_channels": None, + } + + # Determine conv type + if isinstance(module, nn.Conv1d): + info["is_conv"] = True + info["conv_dim"] = 1 + elif isinstance(module, nn.Conv2d): + info["is_conv"] = True + info["conv_dim"] = 2 + elif isinstance(module, nn.Conv3d): + info["is_conv"] = True + info["conv_dim"] = 3 + elif isinstance(module, nn.Linear): + info["is_conv"] = False + info["conv_dim"] = 0 + else: + # Try to infer from class name for custom/quantized layers + class_name = type(module).__name__.lower() + if "conv3d" in class_name: + info["is_conv"] = True + info["conv_dim"] = 3 + elif "conv2d" in class_name: + info["is_conv"] = True + info["conv_dim"] = 2 + elif "conv1d" in class_name: + info["is_conv"] = True + info["conv_dim"] = 1 + elif "conv" in class_name: + info["is_conv"] = True + info["conv_dim"] = 2 + + # Extract conv parameters if it's a conv layer + if info["is_conv"]: + # Try to get stride, padding, dilation, groups, kernel_size from module + info["stride"] = getattr(module, "stride", (1,) * info["conv_dim"]) + info["padding"] = getattr(module, "padding", (0,) * info["conv_dim"]) + info["dilation"] = getattr(module, "dilation", (1,) * info["conv_dim"]) + info["groups"] = getattr(module, "groups", 1) + info["kernel_size"] = getattr(module, "kernel_size", (1,) * info["conv_dim"]) + info["in_channels"] = getattr(module, "in_channels", None) + info["out_channels"] = getattr(module, "out_channels", None) + + # Ensure they're tuples + if isinstance(info["stride"], int): + info["stride"] = (info["stride"],) * info["conv_dim"] + if isinstance(info["padding"], int): + info["padding"] = (info["padding"],) * info["conv_dim"] + if isinstance(info["dilation"], int): + info["dilation"] = (info["dilation"],) * info["conv_dim"] + if isinstance(info["kernel_size"], int): + info["kernel_size"] = (info["kernel_size"],) * info["conv_dim"] + + return info + + +class BypassForwardHook: + """ + Hook that wraps a layer's forward to apply adapter in bypass mode. + + Stores the original forward and replaces it with bypass version. + + Supports both: + - WeightAdapterBase: Inference adapters (uses self.weights tuple) + - WeightAdapterTrainBase: Training adapters (nn.Module with parameters) + """ + + def __init__( + self, + module: nn.Module, + adapter: BypassAdapter, + multiplier: float = 1.0, + ): + self.module = module + self.adapter = adapter + self.multiplier = multiplier + self.original_forward = None + + # Determine layer type and conv params from module class (works for quantized layers) + module_info = get_module_type_info(module) + + # Set multiplier and layer type info on adapter for use in h() + adapter.multiplier = multiplier + adapter.is_conv = module_info["is_conv"] + adapter.conv_dim = module_info["conv_dim"] + adapter.kernel_size = module_info["kernel_size"] + adapter.in_channels = module_info["in_channels"] + adapter.out_channels = module_info["out_channels"] + # Store kw_dict for conv operations (like LyCORIS extra_args) + if module_info["is_conv"]: + adapter.kw_dict = { + "stride": module_info["stride"], + "padding": module_info["padding"], + "dilation": module_info["dilation"], + "groups": module_info["groups"], + } + else: + adapter.kw_dict = {} + + def _bypass_forward(self, x: torch.Tensor, *args, **kwargs) -> torch.Tensor: + """Bypass forward: uses adapter's bypass_forward or default g(f(x) + h(x)) + + Note: + Bypass mode does NOT access original model weights (org_weight). + This is intentional - bypass mode is designed for quantized models + where weights may not be in a usable format. All necessary shape + information is provided via adapter attributes set during inject(). + """ + # Check if adapter has custom bypass_forward (e.g., GLoRA) + adapter_bypass = getattr(self.adapter, "bypass_forward", None) + if adapter_bypass is not None: + # Check if it's overridden (not the base class default) + # Need to check both base classes since adapter could be either type + adapter_type = type(self.adapter) + is_default_bypass = ( + adapter_type.bypass_forward is WeightAdapterBase.bypass_forward + or adapter_type.bypass_forward is WeightAdapterTrainBase.bypass_forward + ) + if not is_default_bypass: + return adapter_bypass(self.original_forward, x, *args, **kwargs) + + # Default bypass: g(f(x) + h(x, f(x))) + base_out = self.original_forward(x, *args, **kwargs) + h_out = self.adapter.h(x, base_out) + return self.adapter.g(base_out + h_out) + + def inject(self): + """Replace module forward with bypass version.""" + if self.original_forward is not None: + logging.debug( + f"[BypassHook] Already injected for {type(self.module).__name__}" + ) + return # Already injected + + # Move adapter weights to module's device to avoid CPU-GPU transfer on every forward + device = None + dtype = None + if hasattr(self.module, "weight") and self.module.weight is not None: + device = self.module.weight.device + dtype = self.module.weight.dtype + elif hasattr(self.module, "W_q"): # Quantized layers might use different attr + device = self.module.W_q.device + dtype = self.module.W_q.dtype + + if device is not None: + self._move_adapter_weights_to_device(device, dtype) + + self.original_forward = self.module.forward + self.module.forward = self._bypass_forward + logging.debug( + f"[BypassHook] Injected bypass forward for {type(self.module).__name__} (adapter={type(self.adapter).__name__})" + ) + + def _move_adapter_weights_to_device(self, device, dtype=None): + """Move adapter weights to specified device to avoid per-forward transfers. + + Handles both: + - WeightAdapterBase: has self.weights tuple of tensors + - WeightAdapterTrainBase: nn.Module with parameters, uses .to() method + """ + adapter = self.adapter + + # Check if adapter is an nn.Module (WeightAdapterTrainBase) + if isinstance(adapter, nn.Module): + # In training mode we don't touch dtype as trainer will handle it + adapter.to(device=device) + logging.debug( + f"[BypassHook] Moved training adapter (nn.Module) to {device}" + ) + return + + # WeightAdapterBase: handle self.weights tuple + if not hasattr(adapter, "weights") or adapter.weights is None: + return + + weights = adapter.weights + if isinstance(weights, (list, tuple)): + new_weights = [] + for w in weights: + if isinstance(w, torch.Tensor): + if dtype is not None: + new_weights.append(w.to(device=device, dtype=dtype)) + else: + new_weights.append(w.to(device=device)) + else: + new_weights.append(w) + adapter.weights = ( + tuple(new_weights) if isinstance(weights, tuple) else new_weights + ) + elif isinstance(weights, torch.Tensor): + if dtype is not None: + adapter.weights = weights.to(device=device, dtype=dtype) + else: + adapter.weights = weights.to(device=device) + + logging.debug(f"[BypassHook] Moved adapter weights to {device}") + + def eject(self): + """Restore original module forward.""" + if self.original_forward is None: + logging.debug(f"[BypassHook] Not injected for {type(self.module).__name__}") + return # Not injected + + self.module.forward = self.original_forward + self.original_forward = None + logging.debug( + f"[BypassHook] Ejected bypass forward for {type(self.module).__name__}" + ) + + +class BypassInjectionManager: + """ + Manages bypass mode injection for a collection of adapters. + + Creates PatcherInjection objects that can be used with ModelPatcher. + + Supports both inference adapters (WeightAdapterBase) and training adapters + (WeightAdapterTrainBase). + + Usage: + manager = BypassInjectionManager() + manager.add_adapter("model.layers.0.self_attn.q_proj", lora_adapter, strength=0.8) + manager.add_adapter("model.layers.0.self_attn.k_proj", lora_adapter, strength=0.8) + + injections = manager.create_injections(model) + model_patcher.set_injections("bypass_lora", injections) + """ + + def __init__(self): + self.adapters: dict[str, tuple[BypassAdapter, float]] = {} + self.hooks: list[BypassForwardHook] = [] + + def add_adapter( + self, + key: str, + adapter: BypassAdapter, + strength: float = 1.0, + ): + """ + Add an adapter for a specific weight key. + + Args: + key: Weight key (e.g., "model.layers.0.self_attn.q_proj.weight") + adapter: The weight adapter (LoRAAdapter, LoKrAdapter, etc.) + strength: Multiplier for adapter effect + """ + # Remove .weight suffix if present for module lookup + module_key = key + if module_key.endswith(".weight"): + module_key = module_key[:-7] + logging.debug( + f"[BypassManager] Stripped .weight suffix: {key} -> {module_key}" + ) + + self.adapters[module_key] = (adapter, strength) + logging.debug( + f"[BypassManager] Added adapter: {module_key} (type={type(adapter).__name__}, strength={strength})" + ) + + def clear_adapters(self): + """Remove all adapters.""" + self.adapters.clear() + + def _get_module_by_key(self, model: nn.Module, key: str) -> Optional[nn.Module]: + """Get a submodule by dot-separated key.""" + parts = key.split(".") + module = model + try: + for i, part in enumerate(parts): + if part.isdigit(): + module = module[int(part)] + else: + module = getattr(module, part) + logging.debug( + f"[BypassManager] Found module for key {key}: {type(module).__name__}" + ) + return module + except (AttributeError, IndexError, KeyError) as e: + logging.error(f"[BypassManager] Failed to find module for key {key}: {e}") + logging.error( + f"[BypassManager] Failed at part index {i}, part={part}, current module type={type(module).__name__}" + ) + return None + + def create_injections(self, model: nn.Module) -> list[PatcherInjection]: + """ + Create PatcherInjection objects for all registered adapters. + + Args: + model: The model to inject into (e.g., model_patcher.model) + + Returns: + List of PatcherInjection objects to use with model_patcher.set_injections() + """ + self.hooks.clear() + + logging.debug( + f"[BypassManager] create_injections called with {len(self.adapters)} adapters" + ) + logging.debug(f"[BypassManager] Model type: {type(model).__name__}") + + for key, (adapter, strength) in self.adapters.items(): + logging.debug(f"[BypassManager] Looking for module: {key}") + module = self._get_module_by_key(model, key) + + if module is None: + logging.warning(f"[BypassManager] Module not found for key {key}") + continue + + if not hasattr(module, "weight"): + logging.warning( + f"[BypassManager] Module {key} has no weight attribute (type={type(module).__name__})" + ) + continue + + logging.debug( + f"[BypassManager] Creating hook for {key} (module type={type(module).__name__}, weight shape={module.weight.shape})" + ) + hook = BypassForwardHook(module, adapter, multiplier=strength) + self.hooks.append(hook) + + logging.debug(f"[BypassManager] Created {len(self.hooks)} hooks") + + # Create single injection that manages all hooks + def inject_all(model_patcher): + logging.debug( + f"[BypassManager] inject_all called, injecting {len(self.hooks)} hooks" + ) + for hook in self.hooks: + hook.inject() + logging.debug( + f"[BypassManager] Injected hook for {type(hook.module).__name__}" + ) + + def eject_all(model_patcher): + logging.debug( + f"[BypassManager] eject_all called, ejecting {len(self.hooks)} hooks" + ) + for hook in self.hooks: + hook.eject() + + return [PatcherInjection(inject=inject_all, eject=eject_all)] + + def get_hook_count(self) -> int: + """Return number of hooks that will be/are injected.""" + return len(self.hooks) + + +def create_bypass_injections_from_patches( + model: nn.Module, + patches: dict, + strength: float = 1.0, +) -> list[PatcherInjection]: + """ + Convenience function to create bypass injections from a patches dict. + + This is useful when you have patches in the format used by model_patcher.add_patches() + and want to apply them in bypass mode instead. + + Args: + model: The model to inject into + patches: Dict mapping weight keys to adapter data + strength: Global strength multiplier + + Returns: + List of PatcherInjection objects + """ + manager = BypassInjectionManager() + + for key, patch_list in patches.items(): + if not patch_list: + continue + + # patches format: list of (strength_patch, patch_data, strength_model, offset, function) + for patch in patch_list: + patch_strength, patch_data, strength_model, offset, function = patch + + # patch_data should be a WeightAdapterBase/WeightAdapterTrainBase or tuple + if isinstance(patch_data, (WeightAdapterBase, WeightAdapterTrainBase)): + adapter = patch_data + else: + # Skip non-adapter patches + continue + + combined_strength = strength * patch_strength + manager.add_adapter(key, adapter, strength=combined_strength) + + return manager.create_injections(model) diff --git a/comfy/weight_adapter/glora.py b/comfy/weight_adapter/glora.py index 939abbba5..d6b97a23b 100644 --- a/comfy/weight_adapter/glora.py +++ b/comfy/weight_adapter/glora.py @@ -1,7 +1,8 @@ import logging -from typing import Optional +from typing import Callable, Optional import torch +import torch.nn.functional as F import comfy.model_management from .base import WeightAdapterBase, weight_decompose @@ -29,7 +30,14 @@ class GLoRAAdapter(WeightAdapterBase): b1_name = "{}.b1.weight".format(x) b2_name = "{}.b2.weight".format(x) if a1_name in lora: - weights = (lora[a1_name], lora[a2_name], lora[b1_name], lora[b2_name], alpha, dora_scale) + weights = ( + lora[a1_name], + lora[a2_name], + lora[b1_name], + lora[b2_name], + alpha, + dora_scale, + ) loaded_keys.add(a1_name) loaded_keys.add(a2_name) loaded_keys.add(b1_name) @@ -58,16 +66,28 @@ class GLoRAAdapter(WeightAdapterBase): old_glora = True if v[3].shape[0] == v[2].shape[1] == v[0].shape[1] == v[1].shape[0]: - if old_glora and v[1].shape[0] == weight.shape[0] and weight.shape[0] == weight.shape[1]: + if ( + old_glora + and v[1].shape[0] == weight.shape[0] + and weight.shape[0] == weight.shape[1] + ): pass else: old_glora = False rank = v[1].shape[0] - a1 = comfy.model_management.cast_to_device(v[0].flatten(start_dim=1), weight.device, intermediate_dtype) - a2 = comfy.model_management.cast_to_device(v[1].flatten(start_dim=1), weight.device, intermediate_dtype) - b1 = comfy.model_management.cast_to_device(v[2].flatten(start_dim=1), weight.device, intermediate_dtype) - b2 = comfy.model_management.cast_to_device(v[3].flatten(start_dim=1), weight.device, intermediate_dtype) + a1 = comfy.model_management.cast_to_device( + v[0].flatten(start_dim=1), weight.device, intermediate_dtype + ) + a2 = comfy.model_management.cast_to_device( + v[1].flatten(start_dim=1), weight.device, intermediate_dtype + ) + b1 = comfy.model_management.cast_to_device( + v[2].flatten(start_dim=1), weight.device, intermediate_dtype + ) + b2 = comfy.model_management.cast_to_device( + v[3].flatten(start_dim=1), weight.device, intermediate_dtype + ) if v[4] is not None: alpha = v[4] / rank @@ -76,18 +96,195 @@ class GLoRAAdapter(WeightAdapterBase): try: if old_glora: - lora_diff = (torch.mm(b2, b1) + torch.mm(torch.mm(weight.flatten(start_dim=1).to(dtype=intermediate_dtype), a2), a1)).reshape(weight.shape) #old lycoris glora + lora_diff = ( + torch.mm(b2, b1) + + torch.mm( + torch.mm( + weight.flatten(start_dim=1).to(dtype=intermediate_dtype), a2 + ), + a1, + ) + ).reshape( + weight.shape + ) # old lycoris glora else: if weight.dim() > 2: - lora_diff = torch.einsum("o i ..., i j -> o j ...", torch.einsum("o i ..., i j -> o j ...", weight.to(dtype=intermediate_dtype), a1), a2).reshape(weight.shape) + lora_diff = torch.einsum( + "o i ..., i j -> o j ...", + torch.einsum( + "o i ..., i j -> o j ...", + weight.to(dtype=intermediate_dtype), + a1, + ), + a2, + ).reshape(weight.shape) else: - lora_diff = torch.mm(torch.mm(weight.to(dtype=intermediate_dtype), a1), a2).reshape(weight.shape) + lora_diff = torch.mm( + torch.mm(weight.to(dtype=intermediate_dtype), a1), a2 + ).reshape(weight.shape) lora_diff += torch.mm(b1, b2).reshape(weight.shape) if dora_scale is not None: - weight = weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype, function) + weight = weight_decompose( + dora_scale, + weight, + lora_diff, + alpha, + strength, + intermediate_dtype, + function, + ) else: weight += function(((strength * alpha) * lora_diff).type(weight.dtype)) except Exception as e: logging.error("ERROR {} {} {}".format(self.name, key, e)) return weight + + def _compute_paths(self, x: torch.Tensor): + """ + Compute A path and B path outputs for GLoRA bypass. + + GLoRA: f(x) = Wx + WAx + Bx + - A path: a1(a2(x)) - modifies input to base forward + - B path: b1(b2(x)) - additive component + + Note: + Does not access original model weights - bypass mode is designed + for quantized models where weights may not be accessible. + + Returns: (a_out, b_out) + """ + v = self.weights + # v = (a1, a2, b1, b2, alpha, dora_scale) + a1 = v[0] + a2 = v[1] + b1 = v[2] + b2 = v[3] + alpha = v[4] + + dtype = x.dtype + + # Cast dtype (weights should already be on correct device from inject()) + a1 = a1.to(dtype=dtype) + a2 = a2.to(dtype=dtype) + b1 = b1.to(dtype=dtype) + b2 = b2.to(dtype=dtype) + + # Determine rank and scale + # Check for old vs new glora format + old_glora = False + if b2.shape[1] == b1.shape[0] == a1.shape[0] == a2.shape[1]: + rank = a1.shape[0] + old_glora = True + + if b2.shape[0] == b1.shape[1] == a1.shape[1] == a2.shape[0]: + if old_glora and a2.shape[0] == x.shape[-1] and x.shape[-1] == x.shape[-1]: + pass + else: + old_glora = False + rank = a2.shape[0] + + if alpha is not None: + scale = alpha / rank + else: + scale = 1.0 + + # Apply multiplier + multiplier = getattr(self, "multiplier", 1.0) + scale = scale * multiplier + + # Use module info from bypass injection, not input tensor shape + is_conv = getattr(self, "is_conv", False) + conv_dim = getattr(self, "conv_dim", 0) + kw_dict = getattr(self, "kw_dict", {}) + + if is_conv: + # Conv case - conv_dim is 1/2/3 for conv1d/2d/3d + conv_fn = (F.conv1d, F.conv2d, F.conv3d)[conv_dim - 1] + + # Get module's stride/padding for spatial dimension handling + module_stride = kw_dict.get("stride", (1,) * conv_dim) + module_padding = kw_dict.get("padding", (0,) * conv_dim) + kernel_size = getattr(self, "kernel_size", (1,) * conv_dim) + in_channels = getattr(self, "in_channels", None) + + # Ensure weights are in conv shape + # a1, a2, b1 are always 1x1 kernels + if a1.ndim == 2: + a1 = a1.view(*a1.shape, *([1] * conv_dim)) + if a2.ndim == 2: + a2 = a2.view(*a2.shape, *([1] * conv_dim)) + if b1.ndim == 2: + b1 = b1.view(*b1.shape, *([1] * conv_dim)) + # b2 has actual kernel_size (like LoRA down) + if b2.ndim == 2: + if in_channels is not None: + b2 = b2.view(b2.shape[0], in_channels, *kernel_size) + else: + b2 = b2.view(*b2.shape, *([1] * conv_dim)) + + # A path: a2(x) -> a1(...) - 1x1 convs, no stride/padding needed, a_out is added to x + a2_out = conv_fn(x, a2) + a_out = conv_fn(a2_out, a1) * scale + + # B path: b2(x) with kernel/stride/padding -> b1(...) 1x1 + b2_out = conv_fn(x, b2, stride=module_stride, padding=module_padding) + b_out = conv_fn(b2_out, b1) * scale + else: + # Linear case + if old_glora: + # Old format: a1 @ a2 @ x, b2 @ b1 + a_out = F.linear(F.linear(x, a2), a1) * scale + b_out = F.linear(F.linear(x, b1), b2) * scale + else: + # New format: x @ a1 @ a2, b1 @ b2 + a_out = F.linear(F.linear(x, a1), a2) * scale + b_out = F.linear(F.linear(x, b2), b1) * scale + + return a_out, b_out + + def bypass_forward( + self, + org_forward: Callable, + x: torch.Tensor, + *args, + **kwargs, + ) -> torch.Tensor: + """ + GLoRA bypass forward: f(x + a(x)) + b(x) + + Unlike standard adapters, GLoRA modifies the input to the base forward + AND adds the B path output. + + Note: + Does not access original model weights - bypass mode is designed + for quantized models where weights may not be accessible. + + Reference: LyCORIS GLoRAModule._bypass_forward + """ + a_out, b_out = self._compute_paths(x) + + # Call base forward with modified input + base_out = org_forward(x + a_out, *args, **kwargs) + + # Add B path + return base_out + b_out + + def h(self, x: torch.Tensor, base_out: torch.Tensor) -> torch.Tensor: + """ + For GLoRA, h() returns the B path output. + + Note: + GLoRA's full bypass requires overriding bypass_forward() since + it also modifies the input to org_forward. This h() is provided for + compatibility but bypass_forward() should be used for correct behavior. + + Does not access original model weights - bypass mode is designed + for quantized models where weights may not be accessible. + + Args: + x: Input tensor + base_out: Output from base forward (unused, for API consistency) + """ + _, b_out = self._compute_paths(x) + return b_out diff --git a/comfy/weight_adapter/loha.py b/comfy/weight_adapter/loha.py index 0abb2d403..8007b7b44 100644 --- a/comfy/weight_adapter/loha.py +++ b/comfy/weight_adapter/loha.py @@ -1,11 +1,22 @@ import logging +from functools import cache from typing import Optional import torch +import torch.nn.functional as F import comfy.model_management from .base import WeightAdapterBase, WeightAdapterTrainBase, weight_decompose +@cache +def _warn_loha_bypass_inefficient(): + """One-time warning about LoHa bypass inefficiency.""" + logging.warning( + "LoHa bypass mode is inefficient: full weight diff is computed each forward pass. " + "Consider using LoRA or LoKr for training with bypass mode." + ) + + class HadaWeight(torch.autograd.Function): @staticmethod def forward(ctx, w1u, w1d, w2u, w2d, scale=torch.tensor(1)): @@ -105,9 +116,19 @@ class LohaDiff(WeightAdapterTrainBase): scale = self.alpha / self.rank if self.use_tucker: - diff_weight = HadaWeightTucker.apply(self.hada_t1, self.hada_w1_a, self.hada_w1_b, self.hada_t2, self.hada_w2_a, self.hada_w2_b, scale) + diff_weight = HadaWeightTucker.apply( + self.hada_t1, + self.hada_w1_a, + self.hada_w1_b, + self.hada_t2, + self.hada_w2_a, + self.hada_w2_b, + scale, + ) else: - diff_weight = HadaWeight.apply(self.hada_w1_a, self.hada_w1_b, self.hada_w2_a, self.hada_w2_b, scale) + diff_weight = HadaWeight.apply( + self.hada_w1_a, self.hada_w1_b, self.hada_w2_a, self.hada_w2_b, scale + ) # Add the scaled difference to the original weight weight = w.to(diff_weight) + diff_weight.reshape(w.shape) @@ -138,9 +159,7 @@ class LoHaAdapter(WeightAdapterBase): mat4 = torch.empty(rank, in_dim, device=weight.device, dtype=torch.float32) torch.nn.init.normal_(mat3, 0.1) torch.nn.init.normal_(mat4, 0.01) - return LohaDiff( - (mat1, mat2, alpha, mat3, mat4, None, None, None) - ) + return LohaDiff((mat1, mat2, alpha, mat3, mat4, None, None, None)) def to_train(self): return LohaDiff(self.weights) @@ -172,7 +191,16 @@ class LoHaAdapter(WeightAdapterBase): loaded_keys.add(hada_t1_name) loaded_keys.add(hada_t2_name) - weights = (lora[hada_w1_a_name], lora[hada_w1_b_name], alpha, lora[hada_w2_a_name], lora[hada_w2_b_name], hada_t1, hada_t2, dora_scale) + weights = ( + lora[hada_w1_a_name], + lora[hada_w1_b_name], + alpha, + lora[hada_w2_a_name], + lora[hada_w2_b_name], + hada_t1, + hada_t2, + dora_scale, + ) loaded_keys.add(hada_w1_a_name) loaded_keys.add(hada_w1_b_name) loaded_keys.add(hada_w2_a_name) @@ -203,30 +231,148 @@ class LoHaAdapter(WeightAdapterBase): w2a = v[3] w2b = v[4] dora_scale = v[7] - if v[5] is not None: #cp decomposition + if v[5] is not None: # cp decomposition t1 = v[5] t2 = v[6] - m1 = torch.einsum('i j k l, j r, i p -> p r k l', - comfy.model_management.cast_to_device(t1, weight.device, intermediate_dtype), - comfy.model_management.cast_to_device(w1b, weight.device, intermediate_dtype), - comfy.model_management.cast_to_device(w1a, weight.device, intermediate_dtype)) + m1 = torch.einsum( + "i j k l, j r, i p -> p r k l", + comfy.model_management.cast_to_device( + t1, weight.device, intermediate_dtype + ), + comfy.model_management.cast_to_device( + w1b, weight.device, intermediate_dtype + ), + comfy.model_management.cast_to_device( + w1a, weight.device, intermediate_dtype + ), + ) - m2 = torch.einsum('i j k l, j r, i p -> p r k l', - comfy.model_management.cast_to_device(t2, weight.device, intermediate_dtype), - comfy.model_management.cast_to_device(w2b, weight.device, intermediate_dtype), - comfy.model_management.cast_to_device(w2a, weight.device, intermediate_dtype)) + m2 = torch.einsum( + "i j k l, j r, i p -> p r k l", + comfy.model_management.cast_to_device( + t2, weight.device, intermediate_dtype + ), + comfy.model_management.cast_to_device( + w2b, weight.device, intermediate_dtype + ), + comfy.model_management.cast_to_device( + w2a, weight.device, intermediate_dtype + ), + ) else: - m1 = torch.mm(comfy.model_management.cast_to_device(w1a, weight.device, intermediate_dtype), - comfy.model_management.cast_to_device(w1b, weight.device, intermediate_dtype)) - m2 = torch.mm(comfy.model_management.cast_to_device(w2a, weight.device, intermediate_dtype), - comfy.model_management.cast_to_device(w2b, weight.device, intermediate_dtype)) + m1 = torch.mm( + comfy.model_management.cast_to_device( + w1a, weight.device, intermediate_dtype + ), + comfy.model_management.cast_to_device( + w1b, weight.device, intermediate_dtype + ), + ) + m2 = torch.mm( + comfy.model_management.cast_to_device( + w2a, weight.device, intermediate_dtype + ), + comfy.model_management.cast_to_device( + w2b, weight.device, intermediate_dtype + ), + ) try: lora_diff = (m1 * m2).reshape(weight.shape) if dora_scale is not None: - weight = weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype, function) + weight = weight_decompose( + dora_scale, + weight, + lora_diff, + alpha, + strength, + intermediate_dtype, + function, + ) else: weight += function(((strength * alpha) * lora_diff).type(weight.dtype)) except Exception as e: logging.error("ERROR {} {} {}".format(self.name, key, e)) return weight + + def h(self, x: torch.Tensor, base_out: torch.Tensor) -> torch.Tensor: + """ + Additive bypass component for LoHa: h(x) = diff_weight @ x + + WARNING: Inefficient - computes full Hadamard product each forward. + + Note: + Does not access original model weights - bypass mode is designed + for quantized models where weights may not be accessible. + + Args: + x: Input tensor + base_out: Output from base forward (unused, for API consistency) + + Reference: LyCORIS functional/loha.py bypass_forward_diff + """ + _warn_loha_bypass_inefficient() + + # FUNC_LIST: [None, None, F.linear, F.conv1d, F.conv2d, F.conv3d] + FUNC_LIST = [None, None, F.linear, F.conv1d, F.conv2d, F.conv3d] + + v = self.weights + # v[0]=w1a, v[1]=w1b, v[2]=alpha, v[3]=w2a, v[4]=w2b, v[5]=t1, v[6]=t2, v[7]=dora + w1a = v[0] + w1b = v[1] + alpha = v[2] + w2a = v[3] + w2b = v[4] + t1 = v[5] + t2 = v[6] + + # Compute scale + rank = w1b.shape[0] + scale = (alpha / rank if alpha is not None else 1.0) * getattr( + self, "multiplier", 1.0 + ) + + # Cast dtype + w1a = w1a.to(dtype=x.dtype) + w1b = w1b.to(dtype=x.dtype) + w2a = w2a.to(dtype=x.dtype) + w2b = w2b.to(dtype=x.dtype) + + # Use module info from bypass injection, not weight dimension + is_conv = getattr(self, "is_conv", False) + conv_dim = getattr(self, "conv_dim", 0) + kw_dict = getattr(self, "kw_dict", {}) + + # Compute diff weight using Hadamard product + if t1 is not None and t2 is not None: + t1 = t1.to(dtype=x.dtype) + t2 = t2.to(dtype=x.dtype) + m1 = torch.einsum("i j k l, j r, i p -> p r k l", t1, w1b, w1a) + m2 = torch.einsum("i j k l, j r, i p -> p r k l", t2, w2b, w2a) + diff_weight = (m1 * m2) * scale + else: + m1 = w1a @ w1b + m2 = w2a @ w2b + diff_weight = (m1 * m2) * scale + + if is_conv: + op = FUNC_LIST[conv_dim + 2] + kernel_size = getattr(self, "kernel_size", (1,) * conv_dim) + in_channels = getattr(self, "in_channels", None) + + # Reshape 2D diff_weight to conv format using kernel_size + # diff_weight: [out_channels, in_channels * prod(kernel_size)] -> [out_channels, in_channels, *kernel_size] + if diff_weight.dim() == 2: + if in_channels is not None: + diff_weight = diff_weight.view( + diff_weight.shape[0], in_channels, *kernel_size + ) + else: + diff_weight = diff_weight.view( + *diff_weight.shape, *([1] * conv_dim) + ) + else: + op = F.linear + kw_dict = {} + + return op(x, diff_weight, **kw_dict) diff --git a/comfy/weight_adapter/lokr.py b/comfy/weight_adapter/lokr.py index 9b2aff2d7..b83750012 100644 --- a/comfy/weight_adapter/lokr.py +++ b/comfy/weight_adapter/lokr.py @@ -2,6 +2,7 @@ import logging from typing import Optional import torch +import torch.nn.functional as F import comfy.model_management from .base import ( WeightAdapterBase, @@ -14,7 +15,17 @@ from .base import ( class LokrDiff(WeightAdapterTrainBase): def __init__(self, weights): super().__init__() - (lokr_w1, lokr_w2, alpha, lokr_w1_a, lokr_w1_b, lokr_w2_a, lokr_w2_b, lokr_t2, dora_scale) = weights + ( + lokr_w1, + lokr_w2, + alpha, + lokr_w1_a, + lokr_w1_b, + lokr_w2_a, + lokr_w2_b, + lokr_t2, + dora_scale, + ) = weights self.use_tucker = False if lokr_w1_a is not None: _, rank_a = lokr_w1_a.shape[0], lokr_w1_a.shape[1] @@ -57,10 +68,10 @@ class LokrDiff(WeightAdapterTrainBase): if self.w2_rebuild: if self.use_tucker: w2 = torch.einsum( - 'i j k l, j r, i p -> p r k l', + "i j k l, j r, i p -> p r k l", self.lokr_t2, self.lokr_w2_b, - self.lokr_w2_a + self.lokr_w2_a, ) else: w2 = self.lokr_w2_a @ self.lokr_w2_b @@ -69,9 +80,89 @@ class LokrDiff(WeightAdapterTrainBase): return self.lokr_w2 def __call__(self, w): - diff = torch.kron(self.w1, self.w2) + w1 = self.w1 + w2 = self.w2 + # Unsqueeze w1 to match w2 dims for proper kron product (like LyCORIS make_kron) + for _ in range(w2.dim() - w1.dim()): + w1 = w1.unsqueeze(-1) + diff = torch.kron(w1, w2) return w + diff.reshape(w.shape).to(w) + def h(self, x: torch.Tensor, base_out: torch.Tensor) -> torch.Tensor: + """ + Additive bypass component for LoKr training: efficient Kronecker product. + + Uses w1/w2 properties which handle both direct and decomposed cases. + For create_train (direct w1/w2), no alpha scaling in properties. + For to_train (decomposed), alpha/rank scaling is in properties. + + Args: + x: Input tensor + base_out: Output from base forward (unused, for API consistency) + """ + # Get w1, w2 from properties (handles rebuild vs direct) + w1 = self.w1 + w2 = self.w2 + + # Multiplier from bypass injection + multiplier = getattr(self, "multiplier", 1.0) + + # Get module info from bypass injection + is_conv = getattr(self, "is_conv", False) + conv_dim = getattr(self, "conv_dim", 0) + kw_dict = getattr(self, "kw_dict", {}) + + # Efficient Kronecker application without materializing full weight + # kron(w1, w2) @ x can be computed as nested operations + # w1: [out_l, in_m], w2: [out_k, in_n, *k_size] + # Full weight would be [out_l*out_k, in_m*in_n, *k_size] + + uq = w1.size(1) # in_m - inner grouping dimension + + if is_conv: + conv_fn = (F.conv1d, F.conv2d, F.conv3d)[conv_dim - 1] + + B, C_in, *spatial = x.shape + # Reshape input for grouped application: [B * uq, C_in // uq, *spatial] + h_in_group = x.reshape(B * uq, -1, *spatial) + + # Ensure w2 has conv dims + if w2.dim() == 2: + w2 = w2.view(*w2.shape, *([1] * conv_dim)) + + # Apply w2 path with stride/padding + hb = conv_fn(h_in_group, w2, **kw_dict) + + # Reshape for cross-group operation + hb = hb.view(B, -1, *hb.shape[1:]) + h_cross = hb.transpose(1, -1) + + # Apply w1 (always 2D, applied as linear on channel dim) + hc = F.linear(h_cross, w1) + hc = hc.transpose(1, -1) + + # Reshape to output + out = hc.reshape(B, -1, *hc.shape[3:]) + else: + # Linear case + # Reshape input: [..., in_m * in_n] -> [..., uq (in_m), in_n] + h_in_group = x.reshape(*x.shape[:-1], uq, -1) + + # Apply w2: [..., uq, in_n] @ [out_k, in_n].T -> [..., uq, out_k] + hb = F.linear(h_in_group, w2) + + # Transpose for w1: [..., uq, out_k] -> [..., out_k, uq] + h_cross = hb.transpose(-1, -2) + + # Apply w1: [..., out_k, uq] @ [out_l, uq].T -> [..., out_k, out_l] + hc = F.linear(h_cross, w1) + + # Transpose back and flatten: [..., out_k, out_l] -> [..., out_l * out_k] + hc = hc.transpose(-1, -2) + out = hc.reshape(*hc.shape[:-2], -1) + + return out * multiplier + def passive_memory_usage(self): return sum(param.numel() * param.element_size() for param in self.parameters()) @@ -86,16 +177,22 @@ class LoKrAdapter(WeightAdapterBase): @classmethod def create_train(cls, weight, rank=1, alpha=1.0): out_dim = weight.shape[0] - in_dim = weight.shape[1:].numel() - out1, out2 = factorization(out_dim, rank) - in1, in2 = factorization(in_dim, rank) - mat1 = torch.empty(out1, in1, device=weight.device, dtype=torch.float32) - mat2 = torch.empty(out2, in2, device=weight.device, dtype=torch.float32) + in_dim = weight.shape[1] # Just in_channels, not flattened with kernel + k_size = weight.shape[2:] if weight.dim() > 2 else () + + out_l, out_k = factorization(out_dim, rank) + in_m, in_n = factorization(in_dim, rank) + + # w1: [out_l, in_m] + mat1 = torch.empty(out_l, in_m, device=weight.device, dtype=torch.float32) + # w2: [out_k, in_n, *k_size] for conv, [out_k, in_n] for linear + mat2 = torch.empty( + out_k, in_n, *k_size, device=weight.device, dtype=torch.float32 + ) + torch.nn.init.kaiming_uniform_(mat2, a=5**0.5) torch.nn.init.constant_(mat1, 0.0) - return LokrDiff( - (mat1, mat2, alpha, None, None, None, None, None, None) - ) + return LokrDiff((mat1, mat2, alpha, None, None, None, None, None, None)) def to_train(self): return LokrDiff(self.weights) @@ -154,8 +251,23 @@ class LoKrAdapter(WeightAdapterBase): lokr_t2 = lora[lokr_t2_name] loaded_keys.add(lokr_t2_name) - if (lokr_w1 is not None) or (lokr_w2 is not None) or (lokr_w1_a is not None) or (lokr_w2_a is not None): - weights = (lokr_w1, lokr_w2, alpha, lokr_w1_a, lokr_w1_b, lokr_w2_a, lokr_w2_b, lokr_t2, dora_scale) + if ( + (lokr_w1 is not None) + or (lokr_w2 is not None) + or (lokr_w1_a is not None) + or (lokr_w2_a is not None) + ): + weights = ( + lokr_w1, + lokr_w2, + alpha, + lokr_w1_a, + lokr_w1_b, + lokr_w2_a, + lokr_w2_b, + lokr_t2, + dora_scale, + ) return cls(loaded_keys, weights) else: return None @@ -184,23 +296,47 @@ class LoKrAdapter(WeightAdapterBase): if w1 is None: dim = w1_b.shape[0] - w1 = torch.mm(comfy.model_management.cast_to_device(w1_a, weight.device, intermediate_dtype), - comfy.model_management.cast_to_device(w1_b, weight.device, intermediate_dtype)) + w1 = torch.mm( + comfy.model_management.cast_to_device( + w1_a, weight.device, intermediate_dtype + ), + comfy.model_management.cast_to_device( + w1_b, weight.device, intermediate_dtype + ), + ) else: - w1 = comfy.model_management.cast_to_device(w1, weight.device, intermediate_dtype) + w1 = comfy.model_management.cast_to_device( + w1, weight.device, intermediate_dtype + ) if w2 is None: dim = w2_b.shape[0] if t2 is None: - w2 = torch.mm(comfy.model_management.cast_to_device(w2_a, weight.device, intermediate_dtype), - comfy.model_management.cast_to_device(w2_b, weight.device, intermediate_dtype)) + w2 = torch.mm( + comfy.model_management.cast_to_device( + w2_a, weight.device, intermediate_dtype + ), + comfy.model_management.cast_to_device( + w2_b, weight.device, intermediate_dtype + ), + ) else: - w2 = torch.einsum('i j k l, j r, i p -> p r k l', - comfy.model_management.cast_to_device(t2, weight.device, intermediate_dtype), - comfy.model_management.cast_to_device(w2_b, weight.device, intermediate_dtype), - comfy.model_management.cast_to_device(w2_a, weight.device, intermediate_dtype)) + w2 = torch.einsum( + "i j k l, j r, i p -> p r k l", + comfy.model_management.cast_to_device( + t2, weight.device, intermediate_dtype + ), + comfy.model_management.cast_to_device( + w2_b, weight.device, intermediate_dtype + ), + comfy.model_management.cast_to_device( + w2_a, weight.device, intermediate_dtype + ), + ) else: - w2 = comfy.model_management.cast_to_device(w2, weight.device, intermediate_dtype) + w2 = comfy.model_management.cast_to_device( + w2, weight.device, intermediate_dtype + ) if len(w2.shape) == 4: w1 = w1.unsqueeze(2).unsqueeze(2) @@ -212,9 +348,134 @@ class LoKrAdapter(WeightAdapterBase): try: lora_diff = torch.kron(w1, w2).reshape(weight.shape) if dora_scale is not None: - weight = weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype, function) + weight = weight_decompose( + dora_scale, + weight, + lora_diff, + alpha, + strength, + intermediate_dtype, + function, + ) else: weight += function(((strength * alpha) * lora_diff).type(weight.dtype)) except Exception as e: logging.error("ERROR {} {} {}".format(self.name, key, e)) return weight + + def h(self, x: torch.Tensor, base_out: torch.Tensor) -> torch.Tensor: + """ + Additive bypass component for LoKr: efficient Kronecker product application. + + Note: + Does not access original model weights - bypass mode is designed + for quantized models where weights may not be accessible. + + Args: + x: Input tensor + base_out: Output from base forward (unused, for API consistency) + + Reference: LyCORIS functional/lokr.py bypass_forward_diff + """ + # FUNC_LIST: [None, None, F.linear, F.conv1d, F.conv2d, F.conv3d] + FUNC_LIST = [None, None, F.linear, F.conv1d, F.conv2d, F.conv3d] + + v = self.weights + # v[0]=w1, v[1]=w2, v[2]=alpha, v[3]=w1_a, v[4]=w1_b, v[5]=w2_a, v[6]=w2_b, v[7]=t2, v[8]=dora + w1 = v[0] + w2 = v[1] + alpha = v[2] + w1_a = v[3] + w1_b = v[4] + w2_a = v[5] + w2_b = v[6] + t2 = v[7] + + use_w1 = w1 is not None + use_w2 = w2 is not None + tucker = t2 is not None + + # Use module info from bypass injection, not weight dimension + is_conv = getattr(self, "is_conv", False) + conv_dim = getattr(self, "conv_dim", 0) + kw_dict = getattr(self, "kw_dict", {}) if is_conv else {} + + if is_conv: + op = FUNC_LIST[conv_dim + 2] + else: + op = F.linear + + # Determine rank and scale + rank = w1_b.size(0) if not use_w1 else w2_b.size(0) if not use_w2 else alpha + scale = (alpha / rank if alpha is not None else 1.0) * getattr( + self, "multiplier", 1.0 + ) + + # Build c (w1) + if use_w1: + c = w1.to(dtype=x.dtype) + else: + c = w1_a.to(dtype=x.dtype) @ w1_b.to(dtype=x.dtype) + uq = c.size(1) + + # Build w2 components + if use_w2: + ba = w2.to(dtype=x.dtype) + else: + a = w2_b.to(dtype=x.dtype) + b = w2_a.to(dtype=x.dtype) + if is_conv: + if tucker: + # Tucker: a, b get 1s appended (kernel is in t2) + if a.dim() == 2: + a = a.view(*a.shape, *([1] * conv_dim)) + if b.dim() == 2: + b = b.view(*b.shape, *([1] * conv_dim)) + else: + # Non-tucker conv: b may need 1s appended + if b.dim() == 2: + b = b.view(*b.shape, *([1] * conv_dim)) + + # Reshape input by uq groups + if is_conv: + B, _, *rest = x.shape + h_in_group = x.reshape(B * uq, -1, *rest) + else: + h_in_group = x.reshape(*x.shape[:-1], uq, -1) + + # Apply w2 path + if use_w2: + hb = op(h_in_group, ba, **kw_dict) + else: + if is_conv: + if tucker: + t = t2.to(dtype=x.dtype) + if t.dim() == 2: + t = t.view(*t.shape, *([1] * conv_dim)) + ha = op(h_in_group, a) + ht = op(ha, t, **kw_dict) + hb = op(ht, b) + else: + ha = op(h_in_group, a, **kw_dict) + hb = op(ha, b) + else: + ha = op(h_in_group, a) + hb = op(ha, b) + + # Reshape and apply c (w1) + if is_conv: + hb = hb.view(B, -1, *hb.shape[1:]) + h_cross_group = hb.transpose(1, -1) + else: + h_cross_group = hb.transpose(-1, -2) + + hc = F.linear(h_cross_group, c) + + if is_conv: + hc = hc.transpose(1, -1) + out = hc.reshape(B, -1, *hc.shape[3:]) + else: + hc = hc.transpose(-1, -2) + out = hc.reshape(*hc.shape[:-2], -1) + + return out * scale diff --git a/comfy/weight_adapter/lora.py b/comfy/weight_adapter/lora.py index 3cc60bb1b..bc4260a8f 100644 --- a/comfy/weight_adapter/lora.py +++ b/comfy/weight_adapter/lora.py @@ -2,6 +2,7 @@ import logging from typing import Optional import torch +import torch.nn.functional as F import comfy.model_management from .base import ( WeightAdapterBase, @@ -20,11 +21,7 @@ class LoraDiff(WeightAdapterTrainBase): rank, in_dim = mat2.shape[0], mat2.shape[1] if mid is not None: convdim = mid.ndim - 2 - layer = ( - torch.nn.Conv1d, - torch.nn.Conv2d, - torch.nn.Conv3d - )[convdim] + layer = (torch.nn.Conv1d, torch.nn.Conv2d, torch.nn.Conv3d)[convdim] else: layer = torch.nn.Linear self.lora_up = layer(rank, out_dim, bias=False) @@ -51,6 +48,78 @@ class LoraDiff(WeightAdapterTrainBase): weight = w + scale * diff.reshape(w.shape) return weight.to(org_dtype) + def h(self, x: torch.Tensor, base_out: torch.Tensor) -> torch.Tensor: + """ + Additive bypass component for LoRA training: h(x) = up(down(x)) * scale + + Simple implementation using the nn.Module weights directly. + No mid/dora/reshape branches (create_train doesn't create them). + + Args: + x: Input tensor + base_out: Output from base forward (unused, for API consistency) + """ + # Compute scale = alpha / rank * multiplier + scale = (self.alpha / self.rank) * getattr(self, "multiplier", 1.0) + + # Get module info from bypass injection + is_conv = getattr(self, "is_conv", False) + conv_dim = getattr(self, "conv_dim", 0) + kw_dict = getattr(self, "kw_dict", {}) + + # Get weights (keep in original dtype for numerical stability) + down_weight = self.lora_down.weight + up_weight = self.lora_up.weight + + if is_conv: + # Conv path: use functional conv + # conv_dim: 1=conv1d, 2=conv2d, 3=conv3d + conv_fn = (F.conv1d, F.conv2d, F.conv3d)[conv_dim - 1] + + # Reshape 2D weights to conv format if needed + # down: [rank, in_features] -> [rank, in_channels, *kernel_size] + # up: [out_features, rank] -> [out_features, rank, 1, 1, ...] + if down_weight.dim() == 2: + kernel_size = getattr(self, "kernel_size", (1,) * conv_dim) + in_channels = getattr(self, "in_channels", None) + if in_channels is not None: + down_weight = down_weight.view( + down_weight.shape[0], in_channels, *kernel_size + ) + else: + # Fallback: assume 1x1 kernel + down_weight = down_weight.view( + *down_weight.shape, *([1] * conv_dim) + ) + if up_weight.dim() == 2: + # up always uses 1x1 kernel + up_weight = up_weight.view(*up_weight.shape, *([1] * conv_dim)) + + # down conv uses stride/padding from module, up is 1x1 + hidden = conv_fn(x, down_weight, **kw_dict) + + # mid layer if exists (tucker decomposition) + if self.lora_mid is not None: + mid_weight = self.lora_mid.weight + if mid_weight.dim() == 2: + mid_weight = mid_weight.view(*mid_weight.shape, *([1] * conv_dim)) + hidden = conv_fn(hidden, mid_weight) + + # up conv is always 1x1 (no stride/padding) + out = conv_fn(hidden, up_weight) + else: + # Linear path: simple matmul chain + hidden = F.linear(x, down_weight) + + # mid layer if exists + if self.lora_mid is not None: + mid_weight = self.lora_mid.weight + hidden = F.linear(hidden, mid_weight) + + out = F.linear(hidden, up_weight) + + return out * scale + def passive_memory_usage(self): return sum(param.numel() * param.element_size() for param in self.parameters()) @@ -70,9 +139,7 @@ class LoRAAdapter(WeightAdapterBase): mat2 = torch.empty(rank, in_dim, device=weight.device, dtype=torch.float32) torch.nn.init.kaiming_uniform_(mat1, a=5**0.5) torch.nn.init.constant_(mat2, 0.0) - return LoraDiff( - (mat1, mat2, alpha, None, None, None) - ) + return LoraDiff((mat1, mat2, alpha, None, None, None)) def to_train(self): return LoraDiff(self.weights) @@ -210,3 +277,85 @@ class LoRAAdapter(WeightAdapterBase): except Exception as e: logging.error("ERROR {} {} {}".format(self.name, key, e)) return weight + + def h(self, x: torch.Tensor, base_out: torch.Tensor) -> torch.Tensor: + """ + Additive bypass component for LoRA: h(x) = up(down(x)) * scale + + Note: + Does not access original model weights - bypass mode is designed + for quantized models where weights may not be accessible. + + Args: + x: Input tensor + base_out: Output from base forward (unused, for API consistency) + + Reference: LyCORIS functional/locon.py bypass_forward_diff + """ + # FUNC_LIST: [None, None, F.linear, F.conv1d, F.conv2d, F.conv3d] + FUNC_LIST = [None, None, F.linear, F.conv1d, F.conv2d, F.conv3d] + + v = self.weights + # v[0]=up, v[1]=down, v[2]=alpha, v[3]=mid, v[4]=dora_scale, v[5]=reshape + up = v[0] + down = v[1] + alpha = v[2] + mid = v[3] + + # Compute scale = alpha / rank + rank = down.shape[0] + if alpha is not None: + scale = alpha / rank + else: + scale = 1.0 + scale = scale * getattr(self, "multiplier", 1.0) + + # Cast dtype + up = up.to(dtype=x.dtype) + down = down.to(dtype=x.dtype) + + # Use module info from bypass injection, not weight dimension + is_conv = getattr(self, "is_conv", False) + conv_dim = getattr(self, "conv_dim", 0) + kw_dict = getattr(self, "kw_dict", {}) + + if is_conv: + op = FUNC_LIST[ + conv_dim + 2 + ] # conv_dim 1->conv1d(3), 2->conv2d(4), 3->conv3d(5) + kernel_size = getattr(self, "kernel_size", (1,) * conv_dim) + in_channels = getattr(self, "in_channels", None) + + # Reshape 2D weights to conv format using kernel_size + # down: [rank, in_channels * prod(kernel_size)] -> [rank, in_channels, *kernel_size] + # up: [out_channels, rank] -> [out_channels, rank, 1, 1, ...] (1x1 kernel) + if down.dim() == 2: + # down.shape[1] = in_channels * prod(kernel_size) + if in_channels is not None: + down = down.view(down.shape[0], in_channels, *kernel_size) + else: + # Fallback: assume 1x1 kernel if in_channels unknown + down = down.view(*down.shape, *([1] * conv_dim)) + if up.dim() == 2: + # up always uses 1x1 kernel + up = up.view(*up.shape, *([1] * conv_dim)) + if mid is not None: + mid = mid.to(dtype=x.dtype) + if mid.dim() == 2: + mid = mid.view(*mid.shape, *([1] * conv_dim)) + else: + op = F.linear + kw_dict = {} # linear doesn't take stride/padding + + # Simple chain: down -> mid (if tucker) -> up + if mid is not None: + if not is_conv: + mid = mid.to(dtype=x.dtype) + hidden = op(x, down) + hidden = op(hidden, mid, **kw_dict) + out = op(hidden, up) + else: + hidden = op(x, down, **kw_dict) + out = op(hidden, up) + + return out * scale diff --git a/comfy/weight_adapter/oft.py b/comfy/weight_adapter/oft.py index c0aab9635..bc83cf8e8 100644 --- a/comfy/weight_adapter/oft.py +++ b/comfy/weight_adapter/oft.py @@ -3,13 +3,18 @@ from typing import Optional import torch import comfy.model_management -from .base import WeightAdapterBase, WeightAdapterTrainBase, weight_decompose, factorization +from .base import ( + WeightAdapterBase, + WeightAdapterTrainBase, + weight_decompose, + factorization, +) class OFTDiff(WeightAdapterTrainBase): def __init__(self, weights): super().__init__() - # Unpack weights tuple from LoHaAdapter + # Unpack weights tuple from OFTAdapter blocks, rescale, alpha, _ = weights # Create trainable parameters @@ -52,6 +57,78 @@ class OFTDiff(WeightAdapterTrainBase): weight = self.rescale * weight return weight.to(org_dtype) + def _get_orthogonal_matrix(self, device, dtype): + """Compute the orthogonal rotation matrix R from OFT blocks.""" + blocks = self.oft_blocks.to(device=device, dtype=dtype) + I = torch.eye(self.block_size, device=device, dtype=dtype) + + # Q = blocks - blocks^T (skew-symmetric) + q = blocks - blocks.transpose(1, 2) + normed_q = q + + # Apply constraint if set + if self.constraint: + q_norm = torch.norm(q) + 1e-8 + if q_norm > self.constraint: + normed_q = q * self.constraint / q_norm + + # Cayley transform: R = (I + Q)(I - Q)^-1 + r = (I + normed_q) @ (I - normed_q).float().inverse() + return r.to(dtype) + + def h(self, x: torch.Tensor, base_out: torch.Tensor) -> torch.Tensor: + """ + OFT has no additive component - returns zeros matching base_out shape. + + OFT only transforms the output via g(), it doesn't add to it. + """ + return torch.zeros_like(base_out) + + def g(self, y: torch.Tensor) -> torch.Tensor: + """ + Output transformation for OFT: applies orthogonal rotation. + + OFT transforms output channels using block-diagonal orthogonal matrices. + """ + r = self._get_orthogonal_matrix(y.device, y.dtype) + + # Apply multiplier to interpolate between identity and full transform + multiplier = getattr(self, "multiplier", 1.0) + I = torch.eye(self.block_size, device=y.device, dtype=y.dtype) + r = r * multiplier + (1 - multiplier) * I + + # Use module info from bypass injection + is_conv = getattr(self, "is_conv", y.dim() > 2) + + if is_conv: + # Conv output: (N, C, H, W, ...) -> transpose to (N, H, W, ..., C) + y = y.transpose(1, -1) + + # y now has channels in last dim + *batch_shape, out_features = y.shape + + # Reshape to apply block-diagonal transform + # (*, out_features) -> (*, block_num, block_size) + y_blocked = y.reshape(*batch_shape, self.block_num, self.block_size) + + # Apply orthogonal transform: R @ y for each block + # r: (block_num, block_size, block_size), y_blocked: (*, block_num, block_size) + out_blocked = torch.einsum("k n m, ... k n -> ... k m", r, y_blocked) + + # Reshape back: (*, block_num, block_size) -> (*, out_features) + out = out_blocked.reshape(*batch_shape, out_features) + + # Apply rescale if present + if self.rescaled: + rescale = self.rescale.to(device=y.device, dtype=y.dtype) + out = out * rescale.view(-1) + + if is_conv: + # Transpose back: (N, H, W, ..., C) -> (N, C, H, W, ...) + out = out.transpose(1, -1) + + return out + def passive_memory_usage(self): """Calculates memory usage of the trainable parameters.""" return sum(param.numel() * param.element_size() for param in self.parameters()) @@ -68,10 +145,10 @@ class OFTAdapter(WeightAdapterBase): def create_train(cls, weight, rank=1, alpha=1.0): out_dim = weight.shape[0] block_size, block_num = factorization(out_dim, rank) - block = torch.zeros(block_num, block_size, block_size, device=weight.device, dtype=torch.float32) - return OFTDiff( - (block, None, alpha, None) + block = torch.zeros( + block_num, block_size, block_size, device=weight.device, dtype=torch.float32 ) + return OFTDiff((block, None, alpha, None)) def to_train(self): return OFTDiff(self.weights) @@ -127,9 +204,13 @@ class OFTAdapter(WeightAdapterBase): alpha = 0 dora_scale = v[3] - blocks = comfy.model_management.cast_to_device(blocks, weight.device, intermediate_dtype) + blocks = comfy.model_management.cast_to_device( + blocks, weight.device, intermediate_dtype + ) if rescale is not None: - rescale = comfy.model_management.cast_to_device(rescale, weight.device, intermediate_dtype) + rescale = comfy.model_management.cast_to_device( + rescale, weight.device, intermediate_dtype + ) block_num, block_size, *_ = blocks.shape @@ -139,23 +220,108 @@ class OFTAdapter(WeightAdapterBase): # for Q = -Q^T q = blocks - blocks.transpose(1, 2) normed_q = q - if alpha > 0: # alpha in oft/boft is for constraint + if alpha > 0: # alpha in oft/boft is for constraint q_norm = torch.norm(q) + 1e-8 if q_norm > alpha: normed_q = q * alpha / q_norm # use float() to prevent unsupported type in .inverse() r = (I + normed_q) @ (I - normed_q).float().inverse() r = r.to(weight) + # Create I in weight's dtype for the einsum + I_w = torch.eye(block_size, device=weight.device, dtype=weight.dtype) _, *shape = weight.shape lora_diff = torch.einsum( "k n m, k n ... -> k m ...", - (r * strength) - strength * I, + (r * strength) - strength * I_w, weight.view(block_num, block_size, *shape), ).view(-1, *shape) if dora_scale is not None: - weight = weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype, function) + weight = weight_decompose( + dora_scale, + weight, + lora_diff, + alpha, + strength, + intermediate_dtype, + function, + ) else: weight += function((strength * lora_diff).type(weight.dtype)) except Exception as e: logging.error("ERROR {} {} {}".format(self.name, key, e)) return weight + + def _get_orthogonal_matrix(self, device, dtype): + """Compute the orthogonal rotation matrix R from OFT blocks.""" + v = self.weights + blocks = v[0].to(device=device, dtype=dtype) + alpha = v[2] + if alpha is None: + alpha = 0 + + block_num, block_size, _ = blocks.shape + I = torch.eye(block_size, device=device, dtype=dtype) + + # Q = blocks - blocks^T (skew-symmetric) + q = blocks - blocks.transpose(1, 2) + normed_q = q + + # Apply constraint if alpha > 0 + if alpha > 0: + q_norm = torch.norm(q) + 1e-8 + if q_norm > alpha: + normed_q = q * alpha / q_norm + + # Cayley transform: R = (I + Q)(I - Q)^-1 + r = (I + normed_q) @ (I - normed_q).float().inverse() + return r, block_num, block_size + + def g(self, y: torch.Tensor) -> torch.Tensor: + """ + Output transformation for OFT: applies orthogonal rotation to output. + + OFT transforms the output channels using block-diagonal orthogonal matrices. + + Reference: LyCORIS DiagOFTModule._bypass_forward + """ + v = self.weights + rescale = v[1] + + r, block_num, block_size = self._get_orthogonal_matrix(y.device, y.dtype) + + # Apply multiplier to interpolate between identity and full transform + multiplier = getattr(self, "multiplier", 1.0) + I = torch.eye(block_size, device=y.device, dtype=y.dtype) + r = r * multiplier + (1 - multiplier) * I + + # Use module info from bypass injection to determine conv vs linear + is_conv = getattr(self, "is_conv", y.dim() > 2) + + if is_conv: + # Conv output: (N, C, H, W, ...) -> transpose to (N, H, W, ..., C) + y = y.transpose(1, -1) + + # y now has channels in last dim + *batch_shape, out_features = y.shape + + # Reshape to apply block-diagonal transform + # (*, out_features) -> (*, block_num, block_size) + y_blocked = y.view(*batch_shape, block_num, block_size) + + # Apply orthogonal transform: R @ y for each block + # r: (block_num, block_size, block_size), y_blocked: (*, block_num, block_size) + out_blocked = torch.einsum("k n m, ... k n -> ... k m", r, y_blocked) + + # Reshape back: (*, block_num, block_size) -> (*, out_features) + out = out_blocked.view(*batch_shape, out_features) + + # Apply rescale if present + if rescale is not None: + rescale = rescale.to(device=y.device, dtype=y.dtype) + out = out * rescale.view(-1) + + if is_conv: + # Transpose back: (N, H, W, ..., C) -> (N, C, H, W, ...) + out = out.transpose(1, -1) + + return out diff --git a/comfy_extras/nodes_train.py b/comfy_extras/nodes_train.py index 68a73cf13..024a89391 100644 --- a/comfy_extras/nodes_train.py +++ b/comfy_extras/nodes_train.py @@ -18,6 +18,7 @@ import comfy_extras.nodes_custom_sampler import folder_paths import node_helpers from comfy.weight_adapter import adapters, adapter_maps +from comfy.weight_adapter.bypass import BypassInjectionManager from comfy_api.latest import ComfyExtension, io, ui from comfy.utils import ProgressBar @@ -339,6 +340,11 @@ class TrainSampler(comfy.samplers.Sampler): self._train_step_multires_mode(model_wrap, cond, extra_args, noisegen, latent_image, dataset_size, pbar) if (i + 1) % self.grad_acc == 0: + for param_groups in self.optimizer.param_groups: + for param in param_groups["params"]: + if param.grad is None: + continue + param.grad.data = param.grad.data.to(param.data.dtype) self.optimizer.step() self.optimizer.zero_grad() ui_pbar.update(1) @@ -498,9 +504,9 @@ def _prepare_latents_and_count(latents, dtype, bucket_mode): num_images = sum(t.shape[0] for t in latents) multi_res = False # Not using multi_res path in bucket mode - logging.info(f"Bucket mode: {num_buckets} buckets, {num_images} total samples") + logging.debug(f"Bucket mode: {num_buckets} buckets, {num_images} total samples") for i, lat in enumerate(latents): - logging.info(f" Bucket {i}: shape {lat.shape}") + logging.debug(f" Bucket {i}: shape {lat.shape}") return latents, num_images, multi_res # Non-bucket mode @@ -509,7 +515,7 @@ def _prepare_latents_and_count(latents, dtype, bucket_mode): latents = [t.to(dtype) for t in latents] for latent in latents: all_shapes.add(latent.shape) - logging.info(f"Latent shapes: {all_shapes}") + logging.debug(f"Latent shapes: {all_shapes}") if len(all_shapes) > 1: multi_res = True else: @@ -545,7 +551,7 @@ def _validate_and_expand_conditioning(positive, num_images, bucket_mode): if bucket_mode: return positive # Skip validation in bucket mode - logging.info(f"Total Images: {num_images}, Total Captions: {len(positive)}") + logging.debug(f"Total Images: {num_images}, Total Captions: {len(positive)}") if len(positive) == 1 and num_images > 1: return positive * num_images elif len(positive) != num_images: @@ -596,6 +602,8 @@ def _create_weight_adapter( shape = module.weight.shape lora_params = {} + logging.debug(f"Creating weight adapter for {key} with shape {shape}") + if len(shape) >= 2: alpha = float(existing_weights.get(f"{key}.alpha", 1.0)) dora_scale = existing_weights.get(f"{key}.dora_scale", None) @@ -690,6 +698,61 @@ def _setup_lora_adapters(mp, existing_weights, algorithm, lora_dtype, rank): return lora_sd, all_weight_adapters +def _setup_lora_adapters_bypass(mp, existing_weights, algorithm, lora_dtype, rank): + """Setup LoRA adapters in bypass mode. + + In bypass mode: + - Weight adapters (lora/lokr/oft) use bypass injection (forward hook) + - Bias/norm adapters (BiasDiff) still use weight wrapper (direct modification) + + This is useful when the base model weights are quantized and cannot be + directly modified. + + Args: + mp: Model patcher + existing_weights: Dict of existing LoRA weights + algorithm: Algorithm name for new adapters + lora_dtype: dtype for LoRA weights + rank: Rank for new LoRA adapters + + Returns: + tuple: (lora_sd dict, all_weight_adapters list, bypass_manager) + """ + lora_sd = {} + all_weight_adapters = [] + bypass_manager = BypassInjectionManager() + + for n, m in mp.model.named_modules(): + if hasattr(m, "weight_function"): + if m.weight is not None: + adapter, params = _create_weight_adapter( + m, n, existing_weights, algorithm, lora_dtype, rank + ) + lora_sd.update(params) + all_weight_adapters.append(adapter) + + key = f"{n}.weight" + # BiasDiff (for 1D weights like norm) uses weight wrapper, not bypass + # Only use bypass for adapters that have h() method (lora/lokr/oft) + if isinstance(adapter, BiasDiff): + mp.add_weight_wrapper(key, adapter) + logging.debug(f"[BypassMode] Added 1D weight adapter (weight wrapper) for {key}") + else: + bypass_manager.add_adapter(key, adapter, strength=1.0) + logging.debug(f"[BypassMode] Added weight adapter (bypass) for {key}") + + if hasattr(m, "bias") and m.bias is not None: + # Bias adapters still use weight wrapper (bias is usually not quantized) + bias_adapter, bias_params = _create_bias_adapter(m, n, lora_dtype) + lora_sd.update(bias_params) + key = f"{n}.bias" + mp.add_weight_wrapper(key, bias_adapter) + all_weight_adapters.append(bias_adapter) + logging.debug(f"[BypassMode] Added bias adapter (weight wrapper) for {key}") + + return lora_sd, all_weight_adapters, bypass_manager + + def _create_optimizer(optimizer_name, parameters, learning_rate): """Create optimizer based on name. @@ -884,11 +947,13 @@ class TrainLoraNode(io.ComfyNode): default=False, tooltip="Enable resolution bucket mode. When enabled, expects pre-bucketed latents from ResolutionBucket node.", ), + io.Boolean.Input( + "bypass_mode", + default=False, + tooltip="Enable bypass mode for training. When enabled, adapters are applied via forward hooks instead of weight modification. Useful for quantized models where weights cannot be directly modified.", + ), ], outputs=[ - io.Model.Output( - display_name="model", tooltip="Model with LoRA applied" - ), io.Custom("LORA_MODEL").Output( display_name="lora", tooltip="LoRA weights" ), @@ -919,6 +984,7 @@ class TrainLoraNode(io.ComfyNode): gradient_checkpointing, existing_lora, bucket_mode, + bypass_mode, ): # Extract scalars from lists (due to is_input_list=True) model = model[0] @@ -936,6 +1002,7 @@ class TrainLoraNode(io.ComfyNode): gradient_checkpointing = gradient_checkpointing[0] existing_lora = existing_lora[0] bucket_mode = bucket_mode[0] + bypass_mode = bypass_mode[0] # Process latents based on mode if bucket_mode: @@ -968,9 +1035,16 @@ class TrainLoraNode(io.ComfyNode): existing_weights, existing_steps = _load_existing_lora(existing_lora) # Setup LoRA adapters - lora_sd, all_weight_adapters = _setup_lora_adapters( - mp, existing_weights, algorithm, lora_dtype, rank - ) + bypass_manager = None + if bypass_mode: + logging.debug("Using bypass mode for training") + lora_sd, all_weight_adapters, bypass_manager = _setup_lora_adapters_bypass( + mp, existing_weights, algorithm, lora_dtype, rank + ) + else: + lora_sd, all_weight_adapters = _setup_lora_adapters( + mp, existing_weights, algorithm, lora_dtype, rank + ) # Create optimizer and loss function optimizer = _create_optimizer( @@ -1029,6 +1103,14 @@ class TrainLoraNode(io.ComfyNode): guider = TrainGuider(mp) guider.set_conds(positive) + # Inject bypass hooks if bypass mode is enabled + bypass_injections = None + if bypass_manager is not None: + bypass_injections = bypass_manager.create_injections(mp.model) + for injection in bypass_injections: + injection.inject(mp) + logging.debug(f"[BypassMode] Injected {bypass_manager.get_hook_count()} bypass hooks") + # Run training loop try: _run_training_loop( @@ -1041,6 +1123,11 @@ class TrainLoraNode(io.ComfyNode): multi_res, ) finally: + # Eject bypass hooks if they were injected + if bypass_injections is not None: + for injection in bypass_injections: + injection.eject(mp) + logging.debug("[BypassMode] Ejected bypass hooks") for m in mp.model.modules(): unpatch(m) del train_sampler, optimizer @@ -1052,7 +1139,9 @@ class TrainLoraNode(io.ComfyNode): for param in lora_sd: lora_sd[param] = lora_sd[param].to(lora_dtype) - return io.NodeOutput(mp, lora_sd, loss_map, steps + existing_steps) + # mp in train node is highly specialized for training + # use it in inference will result in bad behavior so we don't return it + return io.NodeOutput(lora_sd, loss_map, steps + existing_steps) class LoraModelLoader(io.ComfyNode):# diff --git a/nodes.py b/nodes.py index b75247665..8a8df9246 100644 --- a/nodes.py +++ b/nodes.py @@ -722,6 +722,69 @@ class LoraLoaderModelOnly(LoraLoader): def load_lora_model_only(self, model, lora_name, strength_model): return (self.load_lora(model, None, lora_name, strength_model, 0)[0],) +class LoraLoaderBypass: + """ + Apply LoRA in bypass mode without modifying base model weights. + + Bypass mode computes: output = base_forward(x) + lora_path(x) + This is useful for training and when model weights are offloaded. + """ + + def __init__(self): + self.loaded_lora = None + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "model": ("MODEL", {"tooltip": "The diffusion model the LoRA will be applied to."}), + "clip": ("CLIP", {"tooltip": "The CLIP model the LoRA will be applied to."}), + "lora_name": (folder_paths.get_filename_list("loras"), {"tooltip": "The name of the LoRA."}), + "strength_model": ("FLOAT", {"default": 1.0, "min": -100.0, "max": 100.0, "step": 0.01, "tooltip": "How strongly to modify the diffusion model. This value can be negative."}), + "strength_clip": ("FLOAT", {"default": 1.0, "min": -100.0, "max": 100.0, "step": 0.01, "tooltip": "How strongly to modify the CLIP model. This value can be negative."}), + } + } + + RETURN_TYPES = ("MODEL", "CLIP") + OUTPUT_TOOLTIPS = ("The modified diffusion model.", "The modified CLIP model.") + FUNCTION = "load_lora" + + CATEGORY = "loaders" + DESCRIPTION = "Apply LoRA in bypass mode. Unlike regular LoRA, this doesn't modify model weights - instead it injects the LoRA computation during forward pass. Useful for training scenarios." + + def load_lora(self, model, clip, lora_name, strength_model, strength_clip): + if strength_model == 0 and strength_clip == 0: + return (model, clip) + + lora_path = folder_paths.get_full_path_or_raise("loras", lora_name) + lora = None + if self.loaded_lora is not None: + if self.loaded_lora[0] == lora_path: + lora = self.loaded_lora[1] + else: + self.loaded_lora = None + + if lora is None: + lora = comfy.utils.load_torch_file(lora_path, safe_load=True) + self.loaded_lora = (lora_path, lora) + + model_lora, clip_lora = comfy.sd.load_bypass_lora_for_models(model, clip, lora, strength_model, strength_clip) + return (model_lora, clip_lora) + + +class LoraLoaderBypassModelOnly(LoraLoaderBypass): + @classmethod + def INPUT_TYPES(s): + return {"required": { "model": ("MODEL",), + "lora_name": (folder_paths.get_filename_list("loras"), ), + "strength_model": ("FLOAT", {"default": 1.0, "min": -100.0, "max": 100.0, "step": 0.01}), + }} + RETURN_TYPES = ("MODEL",) + FUNCTION = "load_lora_model_only" + + def load_lora_model_only(self, model, lora_name, strength_model): + return (self.load_lora(model, None, lora_name, strength_model, 0)[0],) + class VAELoader: video_taes = ["taehv", "lighttaew2_2", "lighttaew2_1", "lighttaehy1_5", "taeltx_2"] image_taes = ["taesd", "taesdxl", "taesd3", "taef1"] @@ -2067,6 +2130,8 @@ NODE_CLASS_MAPPINGS = { "LatentFlip": LatentFlip, "LatentCrop": LatentCrop, "LoraLoader": LoraLoader, + "LoraLoaderBypass": LoraLoaderBypass, + "LoraLoaderBypassModelOnly": LoraLoaderBypassModelOnly, "CLIPLoader": CLIPLoader, "UNETLoader": UNETLoader, "DualCLIPLoader": DualCLIPLoader, @@ -2106,6 +2171,8 @@ NODE_DISPLAY_NAME_MAPPINGS = { "CheckpointLoaderSimple": "Load Checkpoint", "VAELoader": "Load VAE", "LoraLoader": "Load LoRA", + "LoraLoaderBypass": "Load LoRA (Bypass)", + "LoraLoaderBypassModelOnly": "Load LoRA (Bypass, Model Only)", "CLIPLoader": "Load CLIP", "ControlNetLoader": "Load ControlNet Model", "DiffControlNetLoader": "Load ControlNet Model (diff)", From 26c5bbb8751071cb499b65d48e218b54e856572d Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sat, 24 Jan 2026 20:02:32 -0800 Subject: [PATCH 514/647] Move nodes from previous PR into their own file. (#12066) --- comfy_extras/nodes_lora_debug.py | 79 ++++++++++++++++++++++++++++++++ nodes.py | 68 +-------------------------- 2 files changed, 80 insertions(+), 67 deletions(-) create mode 100644 comfy_extras/nodes_lora_debug.py diff --git a/comfy_extras/nodes_lora_debug.py b/comfy_extras/nodes_lora_debug.py new file mode 100644 index 000000000..937a0fbfb --- /dev/null +++ b/comfy_extras/nodes_lora_debug.py @@ -0,0 +1,79 @@ +import folder_paths +import comfy.utils +import comfy.sd + + +class LoraLoaderBypass: + """ + Apply LoRA in bypass mode without modifying base model weights. + + Bypass mode computes: output = base_forward(x) + lora_path(x) + This is useful for training and when model weights are offloaded. + """ + + def __init__(self): + self.loaded_lora = None + + @classmethod + def INPUT_TYPES(s): + return { + "required": { + "model": ("MODEL", {"tooltip": "The diffusion model the LoRA will be applied to."}), + "clip": ("CLIP", {"tooltip": "The CLIP model the LoRA will be applied to."}), + "lora_name": (folder_paths.get_filename_list("loras"), {"tooltip": "The name of the LoRA."}), + "strength_model": ("FLOAT", {"default": 1.0, "min": -100.0, "max": 100.0, "step": 0.01, "tooltip": "How strongly to modify the diffusion model. This value can be negative."}), + "strength_clip": ("FLOAT", {"default": 1.0, "min": -100.0, "max": 100.0, "step": 0.01, "tooltip": "How strongly to modify the CLIP model. This value can be negative."}), + } + } + + RETURN_TYPES = ("MODEL", "CLIP") + OUTPUT_TOOLTIPS = ("The modified diffusion model.", "The modified CLIP model.") + FUNCTION = "load_lora" + + CATEGORY = "loaders" + DESCRIPTION = "Apply LoRA in bypass mode. Unlike regular LoRA, this doesn't modify model weights - instead it injects the LoRA computation during forward pass. Useful for training scenarios." + EXPERIMENTAL = True + + def load_lora(self, model, clip, lora_name, strength_model, strength_clip): + if strength_model == 0 and strength_clip == 0: + return (model, clip) + + lora_path = folder_paths.get_full_path_or_raise("loras", lora_name) + lora = None + if self.loaded_lora is not None: + if self.loaded_lora[0] == lora_path: + lora = self.loaded_lora[1] + else: + self.loaded_lora = None + + if lora is None: + lora = comfy.utils.load_torch_file(lora_path, safe_load=True) + self.loaded_lora = (lora_path, lora) + + model_lora, clip_lora = comfy.sd.load_bypass_lora_for_models(model, clip, lora, strength_model, strength_clip) + return (model_lora, clip_lora) + + +class LoraLoaderBypassModelOnly(LoraLoaderBypass): + @classmethod + def INPUT_TYPES(s): + return {"required": { "model": ("MODEL",), + "lora_name": (folder_paths.get_filename_list("loras"), ), + "strength_model": ("FLOAT", {"default": 1.0, "min": -100.0, "max": 100.0, "step": 0.01}), + }} + RETURN_TYPES = ("MODEL",) + FUNCTION = "load_lora_model_only" + + def load_lora_model_only(self, model, lora_name, strength_model): + return (self.load_lora(model, None, lora_name, strength_model, 0)[0],) + + +NODE_CLASS_MAPPINGS = { + "LoraLoaderBypass": LoraLoaderBypass, + "LoraLoaderBypassModelOnly": LoraLoaderBypassModelOnly, +} + +NODE_DISPLAY_NAME_MAPPINGS = { + "LoraLoaderBypass": "Load LoRA (Bypass) (For debugging)", + "LoraLoaderBypassModelOnly": "Load LoRA (Bypass, Model Only) (for debugging)", +} diff --git a/nodes.py b/nodes.py index 8a8df9246..2535b4ec6 100644 --- a/nodes.py +++ b/nodes.py @@ -722,69 +722,6 @@ class LoraLoaderModelOnly(LoraLoader): def load_lora_model_only(self, model, lora_name, strength_model): return (self.load_lora(model, None, lora_name, strength_model, 0)[0],) -class LoraLoaderBypass: - """ - Apply LoRA in bypass mode without modifying base model weights. - - Bypass mode computes: output = base_forward(x) + lora_path(x) - This is useful for training and when model weights are offloaded. - """ - - def __init__(self): - self.loaded_lora = None - - @classmethod - def INPUT_TYPES(s): - return { - "required": { - "model": ("MODEL", {"tooltip": "The diffusion model the LoRA will be applied to."}), - "clip": ("CLIP", {"tooltip": "The CLIP model the LoRA will be applied to."}), - "lora_name": (folder_paths.get_filename_list("loras"), {"tooltip": "The name of the LoRA."}), - "strength_model": ("FLOAT", {"default": 1.0, "min": -100.0, "max": 100.0, "step": 0.01, "tooltip": "How strongly to modify the diffusion model. This value can be negative."}), - "strength_clip": ("FLOAT", {"default": 1.0, "min": -100.0, "max": 100.0, "step": 0.01, "tooltip": "How strongly to modify the CLIP model. This value can be negative."}), - } - } - - RETURN_TYPES = ("MODEL", "CLIP") - OUTPUT_TOOLTIPS = ("The modified diffusion model.", "The modified CLIP model.") - FUNCTION = "load_lora" - - CATEGORY = "loaders" - DESCRIPTION = "Apply LoRA in bypass mode. Unlike regular LoRA, this doesn't modify model weights - instead it injects the LoRA computation during forward pass. Useful for training scenarios." - - def load_lora(self, model, clip, lora_name, strength_model, strength_clip): - if strength_model == 0 and strength_clip == 0: - return (model, clip) - - lora_path = folder_paths.get_full_path_or_raise("loras", lora_name) - lora = None - if self.loaded_lora is not None: - if self.loaded_lora[0] == lora_path: - lora = self.loaded_lora[1] - else: - self.loaded_lora = None - - if lora is None: - lora = comfy.utils.load_torch_file(lora_path, safe_load=True) - self.loaded_lora = (lora_path, lora) - - model_lora, clip_lora = comfy.sd.load_bypass_lora_for_models(model, clip, lora, strength_model, strength_clip) - return (model_lora, clip_lora) - - -class LoraLoaderBypassModelOnly(LoraLoaderBypass): - @classmethod - def INPUT_TYPES(s): - return {"required": { "model": ("MODEL",), - "lora_name": (folder_paths.get_filename_list("loras"), ), - "strength_model": ("FLOAT", {"default": 1.0, "min": -100.0, "max": 100.0, "step": 0.01}), - }} - RETURN_TYPES = ("MODEL",) - FUNCTION = "load_lora_model_only" - - def load_lora_model_only(self, model, lora_name, strength_model): - return (self.load_lora(model, None, lora_name, strength_model, 0)[0],) - class VAELoader: video_taes = ["taehv", "lighttaew2_2", "lighttaew2_1", "lighttaehy1_5", "taeltx_2"] image_taes = ["taesd", "taesdxl", "taesd3", "taef1"] @@ -2130,8 +2067,6 @@ NODE_CLASS_MAPPINGS = { "LatentFlip": LatentFlip, "LatentCrop": LatentCrop, "LoraLoader": LoraLoader, - "LoraLoaderBypass": LoraLoaderBypass, - "LoraLoaderBypassModelOnly": LoraLoaderBypassModelOnly, "CLIPLoader": CLIPLoader, "UNETLoader": UNETLoader, "DualCLIPLoader": DualCLIPLoader, @@ -2171,8 +2106,6 @@ NODE_DISPLAY_NAME_MAPPINGS = { "CheckpointLoaderSimple": "Load Checkpoint", "VAELoader": "Load VAE", "LoraLoader": "Load LoRA", - "LoraLoaderBypass": "Load LoRA (Bypass)", - "LoraLoaderBypassModelOnly": "Load LoRA (Bypass, Model Only)", "CLIPLoader": "Load CLIP", "ControlNetLoader": "Load ControlNet Model", "DiffControlNetLoader": "Load ControlNet Model (diff)", @@ -2498,6 +2431,7 @@ async def init_builtin_extra_nodes(): "nodes_wanmove.py", "nodes_image_compare.py", "nodes_zimage.py", + "nodes_lora_debug.py" ] import_failed = [] From 7ee77ff038937bdfdbea5d603ad8d4c487c14fd6 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sun, 25 Jan 2026 18:01:55 -0800 Subject: [PATCH 515/647] Add name to LoraLoaderModelOnly. (#12078) --- nodes.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/nodes.py b/nodes.py index 2535b4ec6..ad474d3cd 100644 --- a/nodes.py +++ b/nodes.py @@ -2105,7 +2105,8 @@ NODE_DISPLAY_NAME_MAPPINGS = { "CheckpointLoader": "Load Checkpoint With Config (DEPRECATED)", "CheckpointLoaderSimple": "Load Checkpoint", "VAELoader": "Load VAE", - "LoraLoader": "Load LoRA", + "LoraLoader": "Load LoRA (Model and CLIP)", + "LoraLoaderModelOnly": "Load LoRA", "CLIPLoader": "Load CLIP", "ControlNetLoader": "Load ControlNet Model", "DiffControlNetLoader": "Load ControlNet Model (diff)", From 2129e7d27854057737808438ec5b9db195bb81bb Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 26 Jan 2026 08:39:00 -0800 Subject: [PATCH 516/647] Fix mistral 3 tokenizer code failing on latest transformers version and other breakage. (#12095) * Fix mistral 3 tokenizer code failing on latest transformers version. * Add requests to the requirements --- comfy/sd1_clip.py | 15 +++++++++++---- comfy/text_encoders/flux.py | 2 +- requirements.txt | 1 + 3 files changed, 13 insertions(+), 5 deletions(-) diff --git a/comfy/sd1_clip.py b/comfy/sd1_clip.py index c512ca5d0..d4f22120b 100644 --- a/comfy/sd1_clip.py +++ b/comfy/sd1_clip.py @@ -466,7 +466,7 @@ def load_embed(embedding_name, embedding_directory, embedding_size, embed_key=No return embed_out class SDTokenizer: - def __init__(self, tokenizer_path=None, max_length=77, pad_with_end=True, embedding_directory=None, embedding_size=768, embedding_key='clip_l', tokenizer_class=CLIPTokenizer, has_start_token=True, has_end_token=True, pad_to_max_length=True, min_length=None, pad_token=None, end_token=None, min_padding=None, pad_left=False, disable_weights=False, tokenizer_data={}, tokenizer_args={}): + def __init__(self, tokenizer_path=None, max_length=77, pad_with_end=True, embedding_directory=None, embedding_size=768, embedding_key='clip_l', tokenizer_class=CLIPTokenizer, has_start_token=True, has_end_token=True, pad_to_max_length=True, min_length=None, pad_token=None, end_token=None, start_token=None, min_padding=None, pad_left=False, disable_weights=False, tokenizer_data={}, tokenizer_args={}): if tokenizer_path is None: tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "sd1_tokenizer") self.tokenizer = tokenizer_class.from_pretrained(tokenizer_path, **tokenizer_args) @@ -479,8 +479,15 @@ class SDTokenizer: empty = self.tokenizer('')["input_ids"] self.tokenizer_adds_end_token = has_end_token if has_start_token: - self.tokens_start = 1 - self.start_token = empty[0] + if len(empty) > 0: + self.tokens_start = 1 + self.start_token = empty[0] + else: + self.tokens_start = 0 + self.start_token = start_token + if start_token is None: + logging.warning("WARNING: There's something wrong with your tokenizers.'") + if end_token is not None: self.end_token = end_token else: @@ -488,7 +495,7 @@ class SDTokenizer: self.end_token = empty[1] else: self.tokens_start = 0 - self.start_token = None + self.start_token = start_token if end_token is not None: self.end_token = end_token else: diff --git a/comfy/text_encoders/flux.py b/comfy/text_encoders/flux.py index 4075afca4..f67a5f805 100644 --- a/comfy/text_encoders/flux.py +++ b/comfy/text_encoders/flux.py @@ -118,7 +118,7 @@ class MistralTokenizerClass: class Mistral3Tokenizer(sd1_clip.SDTokenizer): def __init__(self, embedding_directory=None, tokenizer_data={}): self.tekken_data = tokenizer_data.get("tekken_model", None) - super().__init__("", pad_with_end=False, embedding_size=5120, embedding_key='mistral3_24b', tokenizer_class=MistralTokenizerClass, has_end_token=False, pad_to_max_length=False, pad_token=11, max_length=99999999, min_length=1, pad_left=True, tokenizer_args=load_mistral_tokenizer(self.tekken_data), tokenizer_data=tokenizer_data) + super().__init__("", pad_with_end=False, embedding_size=5120, embedding_key='mistral3_24b', tokenizer_class=MistralTokenizerClass, has_end_token=False, pad_to_max_length=False, pad_token=11, start_token=1, max_length=99999999, min_length=1, pad_left=True, tokenizer_args=load_mistral_tokenizer(self.tekken_data), tokenizer_data=tokenizer_data) def state_dict(self): return {"tekken_model": self.tekken_data} diff --git a/requirements.txt b/requirements.txt index ec89dccd2..8d38c114b 100644 --- a/requirements.txt +++ b/requirements.txt @@ -22,6 +22,7 @@ alembic SQLAlchemy av>=14.2.0 comfy-kitchen>=0.2.7 +requests #non essential dependencies: kornia>=0.7.1 From bfe31d0b9dce106cbb3bc073660d4ab1d7d9e992 Mon Sep 17 00:00:00 2001 From: Tavi Halperin Date: Mon, 26 Jan 2026 22:33:19 +0200 Subject: [PATCH 517/647] IC-LoRA: support small grid (#12074) --- comfy_extras/nodes_lt.py | 21 +++++++++++++++++---- 1 file changed, 17 insertions(+), 4 deletions(-) diff --git a/comfy_extras/nodes_lt.py b/comfy_extras/nodes_lt.py index b91a22309..2aec62f61 100644 --- a/comfy_extras/nodes_lt.py +++ b/comfy_extras/nodes_lt.py @@ -223,11 +223,24 @@ class LTXVAddGuide(io.ComfyNode): return frame_idx, latent_idx @classmethod - def add_keyframe_index(cls, cond, frame_idx, guiding_latent, scale_factors): + def add_keyframe_index(cls, cond, frame_idx, guiding_latent, scale_factors, latent_downscale_factor=1): keyframe_idxs, _ = get_keyframe_idxs(cond) _, latent_coords = cls.PATCHIFIER.patchify(guiding_latent) pixel_coords = latent_to_pixel_coords(latent_coords, scale_factors, causal_fix=frame_idx == 0) # we need the causal fix only if we're placing the new latents at index 0 pixel_coords[:, 0] += frame_idx + + # The following adjusts keyframe end positions for small grid IC-LoRA. + # After dilation, the small grid has the same size and position as the large grid, + # but each token encodes a larger image patch. We adjust the end position (not start) + # so that RoPE represents the correct middle point of each token. + # keyframe_idxs dims: (batch, spatial_dim [t,h,w], token_id, [start, end]) + # We only adjust h,w (not t) in dim 1, and only end (not start) in dim 3. + spatial_end_offset = (latent_downscale_factor - 1) * torch.tensor( + scale_factors[1:], + device=pixel_coords.device, + ).view(1, -1, 1, 1) + pixel_coords[:, 1:, :, 1:] += spatial_end_offset.to(pixel_coords.dtype) + if keyframe_idxs is None: keyframe_idxs = pixel_coords else: @@ -235,12 +248,12 @@ class LTXVAddGuide(io.ComfyNode): return node_helpers.conditioning_set_values(cond, {"keyframe_idxs": keyframe_idxs}) @classmethod - def append_keyframe(cls, positive, negative, frame_idx, latent_image, noise_mask, guiding_latent, strength, scale_factors, guide_mask=None, in_channels=128): + def append_keyframe(cls, positive, negative, frame_idx, latent_image, noise_mask, guiding_latent, strength, scale_factors, guide_mask=None, in_channels=128, latent_downscale_factor=1): if latent_image.shape[1] != in_channels or guiding_latent.shape[1] != in_channels: raise ValueError("Adding guide to a combined AV latent is not supported.") - positive = cls.add_keyframe_index(positive, frame_idx, guiding_latent, scale_factors) - negative = cls.add_keyframe_index(negative, frame_idx, guiding_latent, scale_factors) + positive = cls.add_keyframe_index(positive, frame_idx, guiding_latent, scale_factors, latent_downscale_factor) + negative = cls.add_keyframe_index(negative, frame_idx, guiding_latent, scale_factors, latent_downscale_factor) if guide_mask is not None: target_h = max(noise_mask.shape[3], guide_mask.shape[3]) From cd4985e2f33e6c339a6f176b2caed155309a1c6f Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Mon, 26 Jan 2026 23:58:33 +0200 Subject: [PATCH 518/647] chore(api-nodes): remove ByteDanceImageEditNode node (seededit) (#12069) Co-authored-by: Jedrzej Kosinski --- comfy_api_nodes/apis/bytedance.py | 11 ---- comfy_api_nodes/nodes_bytedance.py | 95 ------------------------------ 2 files changed, 106 deletions(-) diff --git a/comfy_api_nodes/apis/bytedance.py b/comfy_api_nodes/apis/bytedance.py index 400648cca..23cbe2372 100644 --- a/comfy_api_nodes/apis/bytedance.py +++ b/comfy_api_nodes/apis/bytedance.py @@ -13,17 +13,6 @@ class Text2ImageTaskCreationRequest(BaseModel): watermark: bool | None = Field(False) -class Image2ImageTaskCreationRequest(BaseModel): - model: str = Field(...) - prompt: str = Field(...) - response_format: str | None = Field("url") - image: str = Field(..., description="Base64 encoded string or image URL") - size: str | None = Field("adaptive") - seed: int | None = Field(..., ge=0, le=2147483647) - guidance_scale: float | None = Field(..., ge=1.0, le=10.0) - watermark: bool | None = Field(False) - - class Seedream4Options(BaseModel): max_images: int = Field(15) diff --git a/comfy_api_nodes/nodes_bytedance.py b/comfy_api_nodes/nodes_bytedance.py index 486801150..0cb5e3be8 100644 --- a/comfy_api_nodes/nodes_bytedance.py +++ b/comfy_api_nodes/nodes_bytedance.py @@ -9,7 +9,6 @@ from comfy_api_nodes.apis.bytedance import ( RECOMMENDED_PRESETS, RECOMMENDED_PRESETS_SEEDREAM_4, VIDEO_TASKS_EXECUTION_TIME, - Image2ImageTaskCreationRequest, Image2VideoTaskCreationRequest, ImageTaskCreationResponse, Seedream4Options, @@ -174,99 +173,6 @@ class ByteDanceImageNode(IO.ComfyNode): return IO.NodeOutput(await download_url_to_image_tensor(get_image_url_from_response(response))) -class ByteDanceImageEditNode(IO.ComfyNode): - - @classmethod - def define_schema(cls): - return IO.Schema( - node_id="ByteDanceImageEditNode", - display_name="ByteDance Image Edit", - category="api node/image/ByteDance", - description="Edit images using ByteDance models via api based on prompt", - inputs=[ - IO.Combo.Input("model", options=["seededit-3-0-i2i-250628"]), - IO.Image.Input( - "image", - tooltip="The base image to edit", - ), - IO.String.Input( - "prompt", - multiline=True, - default="", - tooltip="Instruction to edit image", - ), - IO.Int.Input( - "seed", - default=0, - min=0, - max=2147483647, - step=1, - display_mode=IO.NumberDisplay.number, - control_after_generate=True, - tooltip="Seed to use for generation", - optional=True, - ), - IO.Float.Input( - "guidance_scale", - default=5.5, - min=1.0, - max=10.0, - step=0.01, - display_mode=IO.NumberDisplay.number, - tooltip="Higher value makes the image follow the prompt more closely", - optional=True, - ), - IO.Boolean.Input( - "watermark", - default=False, - tooltip='Whether to add an "AI generated" watermark to the image', - optional=True, - ), - ], - outputs=[ - IO.Image.Output(), - ], - hidden=[ - IO.Hidden.auth_token_comfy_org, - IO.Hidden.api_key_comfy_org, - IO.Hidden.unique_id, - ], - is_api_node=True, - is_deprecated=True, - ) - - @classmethod - async def execute( - cls, - model: str, - image: Input.Image, - prompt: str, - seed: int, - guidance_scale: float, - watermark: bool, - ) -> IO.NodeOutput: - validate_string(prompt, strip_whitespace=True, min_length=1) - if get_number_of_images(image) != 1: - raise ValueError("Exactly one input image is required.") - validate_image_aspect_ratio(image, (1, 3), (3, 1)) - source_url = (await upload_images_to_comfyapi(cls, image, max_images=1, mime_type="image/png"))[0] - payload = Image2ImageTaskCreationRequest( - model=model, - prompt=prompt, - image=source_url, - seed=seed, - guidance_scale=guidance_scale, - watermark=watermark, - ) - response = await sync_op( - cls, - ApiEndpoint(path=BYTEPLUS_IMAGE_ENDPOINT, method="POST"), - data=payload, - response_model=ImageTaskCreationResponse, - ) - return IO.NodeOutput(await download_url_to_image_tensor(get_image_url_from_response(response))) - - class ByteDanceSeedreamNode(IO.ComfyNode): @classmethod @@ -1101,7 +1007,6 @@ class ByteDanceExtension(ComfyExtension): async def get_node_list(self) -> list[type[IO.ComfyNode]]: return [ ByteDanceImageNode, - ByteDanceImageEditNode, ByteDanceSeedreamNode, ByteDanceTextToVideoNode, ByteDanceImageToVideoNode, From 29011ba87eb2131c7943bf0eaf9ac8c0a6ff3c7f Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Tue, 27 Jan 2026 00:10:09 +0200 Subject: [PATCH 519/647] [API Nodes] add Magnific nodes (#11986) * feat(api-nodes): add Magnific nodes * aggressive downscaling should not be performed * disable upscaler nodes --------- Co-authored-by: Jedrzej Kosinski --- comfy_api_nodes/apis/magnific.py | 122 ++++ comfy_api_nodes/nodes_magnific.py | 889 +++++++++++++++++++++++++ comfy_api_nodes/util/conversions.py | 18 +- comfy_api_nodes/util/upload_helpers.py | 2 +- 4 files changed, 1021 insertions(+), 10 deletions(-) create mode 100644 comfy_api_nodes/apis/magnific.py create mode 100644 comfy_api_nodes/nodes_magnific.py diff --git a/comfy_api_nodes/apis/magnific.py b/comfy_api_nodes/apis/magnific.py new file mode 100644 index 000000000..b9f148def --- /dev/null +++ b/comfy_api_nodes/apis/magnific.py @@ -0,0 +1,122 @@ +from typing import TypedDict + +from pydantic import AliasChoices, BaseModel, Field, model_validator + + +class InputPortraitMode(TypedDict): + portrait_mode: str + portrait_style: str + portrait_beautifier: str + + +class InputAdvancedSettings(TypedDict): + advanced_settings: str + whites: int + blacks: int + brightness: int + contrast: int + saturation: int + engine: str + transfer_light_a: str + transfer_light_b: str + fixed_generation: bool + + +class InputSkinEnhancerMode(TypedDict): + mode: str + skin_detail: int + optimized_for: str + + +class ImageUpscalerCreativeRequest(BaseModel): + image: str = Field(...) + scale_factor: str = Field(...) + optimized_for: str = Field(...) + prompt: str | None = Field(None) + creativity: int = Field(...) + hdr: int = Field(...) + resemblance: int = Field(...) + fractality: int = Field(...) + engine: str = Field(...) + + +class ImageUpscalerPrecisionV2Request(BaseModel): + image: str = Field(...) + sharpen: int = Field(...) + smart_grain: int = Field(...) + ultra_detail: int = Field(...) + flavor: str = Field(...) + scale_factor: int = Field(...) + + +class ImageRelightAdvancedSettingsRequest(BaseModel): + whites: int = Field(...) + blacks: int = Field(...) + brightness: int = Field(...) + contrast: int = Field(...) + saturation: int = Field(...) + engine: str = Field(...) + transfer_light_a: str = Field(...) + transfer_light_b: str = Field(...) + fixed_generation: bool = Field(...) + + +class ImageRelightRequest(BaseModel): + image: str = Field(...) + prompt: str | None = Field(None) + transfer_light_from_reference_image: str | None = Field(None) + light_transfer_strength: int = Field(...) + interpolate_from_original: bool = Field(...) + change_background: bool = Field(...) + style: str = Field(...) + preserve_details: bool = Field(...) + advanced_settings: ImageRelightAdvancedSettingsRequest | None = Field(...) + + +class ImageStyleTransferRequest(BaseModel): + image: str = Field(...) + reference_image: str = Field(...) + prompt: str | None = Field(None) + style_strength: int = Field(...) + structure_strength: int = Field(...) + is_portrait: bool = Field(...) + portrait_style: str | None = Field(...) + portrait_beautifier: str | None = Field(...) + flavor: str = Field(...) + engine: str = Field(...) + fixed_generation: bool = Field(...) + + +class ImageSkinEnhancerCreativeRequest(BaseModel): + image: str = Field(...) + sharpen: int = Field(...) + smart_grain: int = Field(...) + + +class ImageSkinEnhancerFaithfulRequest(BaseModel): + image: str = Field(...) + sharpen: int = Field(...) + smart_grain: int = Field(...) + skin_detail: int = Field(...) + + +class ImageSkinEnhancerFlexibleRequest(BaseModel): + image: str = Field(...) + sharpen: int = Field(...) + smart_grain: int = Field(...) + optimized_for: str = Field(...) + + +class TaskResponse(BaseModel): + """Unified response model that handles both wrapped and unwrapped API responses.""" + + task_id: str = Field(...) + status: str = Field(validation_alias=AliasChoices("status", "task_status")) + generated: list[str] | None = Field(None) + + @model_validator(mode="before") + @classmethod + def unwrap_data(cls, values: dict) -> dict: + if "data" in values and isinstance(values["data"], dict): + return values["data"] + return values diff --git a/comfy_api_nodes/nodes_magnific.py b/comfy_api_nodes/nodes_magnific.py new file mode 100644 index 000000000..013e71cc8 --- /dev/null +++ b/comfy_api_nodes/nodes_magnific.py @@ -0,0 +1,889 @@ +import math + +from typing_extensions import override + +from comfy_api.latest import IO, ComfyExtension, Input +from comfy_api_nodes.apis.magnific import ( + ImageRelightAdvancedSettingsRequest, + ImageRelightRequest, + ImageSkinEnhancerCreativeRequest, + ImageSkinEnhancerFaithfulRequest, + ImageSkinEnhancerFlexibleRequest, + ImageStyleTransferRequest, + ImageUpscalerCreativeRequest, + ImageUpscalerPrecisionV2Request, + InputAdvancedSettings, + InputPortraitMode, + InputSkinEnhancerMode, + TaskResponse, +) +from comfy_api_nodes.util import ( + ApiEndpoint, + download_url_to_image_tensor, + downscale_image_tensor, + get_image_dimensions, + get_number_of_images, + poll_op, + sync_op, + upload_images_to_comfyapi, + validate_image_aspect_ratio, + validate_image_dimensions, +) + + +class MagnificImageUpscalerCreativeNode(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="MagnificImageUpscalerCreativeNode", + display_name="Magnific Image Upscale (Creative)", + category="api node/image/Magnific", + description="Prompt‑guided enhancement, stylization, and 2x/4x/8x/16x upscaling. " + "Maximum output: 25.3 megapixels.", + inputs=[ + IO.Image.Input("image"), + IO.String.Input("prompt", multiline=True, default=""), + IO.Combo.Input("scale_factor", options=["2x", "4x", "8x", "16x"]), + IO.Combo.Input( + "optimized_for", + options=[ + "standard", + "soft_portraits", + "hard_portraits", + "art_n_illustration", + "videogame_assets", + "nature_n_landscapes", + "films_n_photography", + "3d_renders", + "science_fiction_n_horror", + ], + ), + IO.Int.Input("creativity", min=-10, max=10, default=0, display_mode=IO.NumberDisplay.slider), + IO.Int.Input( + "hdr", + min=-10, + max=10, + default=0, + tooltip="The level of definition and detail.", + display_mode=IO.NumberDisplay.slider, + ), + IO.Int.Input( + "resemblance", + min=-10, + max=10, + default=0, + tooltip="The level of resemblance to the original image.", + display_mode=IO.NumberDisplay.slider, + ), + IO.Int.Input( + "fractality", + min=-10, + max=10, + default=0, + tooltip="The strength of the prompt and intricacy per square pixel.", + display_mode=IO.NumberDisplay.slider, + ), + IO.Combo.Input( + "engine", + options=["automatic", "magnific_illusio", "magnific_sharpy", "magnific_sparkle"], + ), + IO.Boolean.Input( + "auto_downscale", + default=False, + tooltip="Automatically downscale input image if output would exceed maximum pixel limit.", + ), + ], + outputs=[ + IO.Image.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["scale_factor"]), + expr=""" + ( + $max := widgets.scale_factor = "2x" ? 1.326 : 1.657; + {"type": "range_usd", "min_usd": 0.11, "max_usd": $max} + ) + """, + ), + ) + + @classmethod + async def execute( + cls, + image: Input.Image, + prompt: str, + scale_factor: str, + optimized_for: str, + creativity: int, + hdr: int, + resemblance: int, + fractality: int, + engine: str, + auto_downscale: bool, + ) -> IO.NodeOutput: + if get_number_of_images(image) != 1: + raise ValueError("Exactly one input image is required.") + validate_image_aspect_ratio(image, (1, 3), (3, 1), strict=False) + validate_image_dimensions(image, min_height=160, min_width=160) + + max_output_pixels = 25_300_000 + height, width = get_image_dimensions(image) + requested_scale = int(scale_factor.rstrip("x")) + output_pixels = height * width * requested_scale * requested_scale + + if output_pixels > max_output_pixels: + if auto_downscale: + # Find optimal scale factor that doesn't require >2x downscale. + # Server upscales in 2x steps, so aggressive downscaling degrades quality. + input_pixels = width * height + scale = 2 + max_input_pixels = max_output_pixels // 4 + for candidate in [16, 8, 4, 2]: + if candidate > requested_scale: + continue + scale_output_pixels = input_pixels * candidate * candidate + if scale_output_pixels <= max_output_pixels: + scale = candidate + max_input_pixels = None + break + downscale_ratio = math.sqrt(scale_output_pixels / max_output_pixels) + if downscale_ratio <= 2.0: + scale = candidate + max_input_pixels = max_output_pixels // (candidate * candidate) + break + + if max_input_pixels is not None: + image = downscale_image_tensor(image, total_pixels=max_input_pixels) + scale_factor = f"{scale}x" + else: + raise ValueError( + f"Output size ({width * requested_scale}x{height * requested_scale} = {output_pixels:,} pixels) " + f"exceeds maximum allowed size of {max_output_pixels:,} pixels. " + f"Use a smaller input image or lower scale factor." + ) + + initial_res = await sync_op( + cls, + ApiEndpoint(path="/proxy/freepik/v1/ai/image-upscaler", method="POST"), + response_model=TaskResponse, + data=ImageUpscalerCreativeRequest( + image=(await upload_images_to_comfyapi(cls, image, max_images=1, total_pixels=None))[0], + scale_factor=scale_factor, + optimized_for=optimized_for, + creativity=creativity, + hdr=hdr, + resemblance=resemblance, + fractality=fractality, + engine=engine, + prompt=prompt if prompt else None, + ), + ) + final_response = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/freepik/v1/ai/image-upscaler/{initial_res.task_id}"), + response_model=TaskResponse, + status_extractor=lambda x: x.status, + poll_interval=10.0, + max_poll_attempts=480, + ) + return IO.NodeOutput(await download_url_to_image_tensor(final_response.generated[0])) + + +class MagnificImageUpscalerPreciseV2Node(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="MagnificImageUpscalerPreciseV2Node", + display_name="Magnific Image Upscale (Precise V2)", + category="api node/image/Magnific", + description="High-fidelity upscaling with fine control over sharpness, grain, and detail. " + "Maximum output: 10060×10060 pixels.", + inputs=[ + IO.Image.Input("image"), + IO.Combo.Input("scale_factor", options=["2x", "4x", "8x", "16x"]), + IO.Combo.Input( + "flavor", + options=["sublime", "photo", "photo_denoiser"], + tooltip="Processing style: " + "sublime for general use, photo for photographs, photo_denoiser for noisy photos.", + ), + IO.Int.Input( + "sharpen", + min=0, + max=100, + default=7, + tooltip="Image sharpness intensity. Higher values increase edge definition and clarity.", + display_mode=IO.NumberDisplay.slider, + ), + IO.Int.Input( + "smart_grain", + min=0, + max=100, + default=7, + tooltip="Intelligent grain/texture enhancement to prevent the image from " + "looking too smooth or artificial.", + display_mode=IO.NumberDisplay.slider, + ), + IO.Int.Input( + "ultra_detail", + min=0, + max=100, + default=30, + tooltip="Controls fine detail, textures, and micro-details added during upscaling.", + display_mode=IO.NumberDisplay.slider, + ), + IO.Boolean.Input( + "auto_downscale", + default=False, + tooltip="Automatically downscale input image if output would exceed maximum resolution.", + ), + ], + outputs=[ + IO.Image.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["scale_factor"]), + expr=""" + ( + $max := widgets.scale_factor = "2x" ? 1.326 : 1.657; + {"type": "range_usd", "min_usd": 0.11, "max_usd": $max} + ) + """, + ), + ) + + @classmethod + async def execute( + cls, + image: Input.Image, + scale_factor: str, + flavor: str, + sharpen: int, + smart_grain: int, + ultra_detail: int, + auto_downscale: bool, + ) -> IO.NodeOutput: + if get_number_of_images(image) != 1: + raise ValueError("Exactly one input image is required.") + validate_image_aspect_ratio(image, (1, 3), (3, 1), strict=False) + validate_image_dimensions(image, min_height=160, min_width=160) + + max_output_dimension = 10060 + height, width = get_image_dimensions(image) + requested_scale = int(scale_factor.strip("x")) + output_width = width * requested_scale + output_height = height * requested_scale + + if output_width > max_output_dimension or output_height > max_output_dimension: + if auto_downscale: + # Find optimal scale factor that doesn't require >2x downscale. + # Server upscales in 2x steps, so aggressive downscaling degrades quality. + max_dim = max(width, height) + scale = 2 + max_input_dim = max_output_dimension // 2 + scale_ratio = max_input_dim / max_dim + max_input_pixels = int(width * height * scale_ratio * scale_ratio) + for candidate in [16, 8, 4, 2]: + if candidate > requested_scale: + continue + output_dim = max_dim * candidate + if output_dim <= max_output_dimension: + scale = candidate + max_input_pixels = None + break + downscale_ratio = output_dim / max_output_dimension + if downscale_ratio <= 2.0: + scale = candidate + max_input_dim = max_output_dimension // candidate + scale_ratio = max_input_dim / max_dim + max_input_pixels = int(width * height * scale_ratio * scale_ratio) + break + + if max_input_pixels is not None: + image = downscale_image_tensor(image, total_pixels=max_input_pixels) + requested_scale = scale + else: + raise ValueError( + f"Output dimensions ({output_width}x{output_height}) exceed maximum allowed " + f"resolution of {max_output_dimension}x{max_output_dimension} pixels. " + f"Use a smaller input image or lower scale factor." + ) + + initial_res = await sync_op( + cls, + ApiEndpoint(path="/proxy/freepik/v1/ai/image-upscaler-precision-v2", method="POST"), + response_model=TaskResponse, + data=ImageUpscalerPrecisionV2Request( + image=(await upload_images_to_comfyapi(cls, image, max_images=1, total_pixels=None))[0], + scale_factor=requested_scale, + flavor=flavor, + sharpen=sharpen, + smart_grain=smart_grain, + ultra_detail=ultra_detail, + ), + ) + final_response = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/freepik/v1/ai/image-upscaler-precision-v2/{initial_res.task_id}"), + response_model=TaskResponse, + status_extractor=lambda x: x.status, + poll_interval=10.0, + max_poll_attempts=480, + ) + return IO.NodeOutput(await download_url_to_image_tensor(final_response.generated[0])) + + +class MagnificImageStyleTransferNode(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="MagnificImageStyleTransferNode", + display_name="Magnific Image Style Transfer", + category="api node/image/Magnific", + description="Transfer the style from a reference image to your input image.", + inputs=[ + IO.Image.Input("image", tooltip="The image to apply style transfer to."), + IO.Image.Input("reference_image", tooltip="The reference image to extract style from."), + IO.String.Input("prompt", multiline=True, default=""), + IO.Int.Input( + "style_strength", + min=0, + max=100, + default=100, + tooltip="Percentage of style strength.", + display_mode=IO.NumberDisplay.slider, + ), + IO.Int.Input( + "structure_strength", + min=0, + max=100, + default=50, + tooltip="Maintains the structure of the original image.", + display_mode=IO.NumberDisplay.slider, + ), + IO.Combo.Input( + "flavor", + options=["faithful", "gen_z", "psychedelia", "detaily", "clear", "donotstyle", "donotstyle_sharp"], + tooltip="Style transfer flavor.", + ), + IO.Combo.Input( + "engine", + options=[ + "balanced", + "definio", + "illusio", + "3d_cartoon", + "colorful_anime", + "caricature", + "real", + "super_real", + "softy", + ], + tooltip="Processing engine selection.", + ), + IO.DynamicCombo.Input( + "portrait_mode", + options=[ + IO.DynamicCombo.Option("disabled", []), + IO.DynamicCombo.Option( + "enabled", + [ + IO.Combo.Input( + "portrait_style", + options=["standard", "pop", "super_pop"], + tooltip="Visual style applied to portrait images.", + ), + IO.Combo.Input( + "portrait_beautifier", + options=["none", "beautify_face", "beautify_face_max"], + tooltip="Facial beautification intensity on portraits.", + ), + ], + ), + ], + tooltip="Enable portrait mode for facial enhancements.", + ), + IO.Boolean.Input( + "fixed_generation", + default=True, + tooltip="When disabled, expect each generation to introduce a degree of randomness, " + "leading to more diverse outcomes.", + ), + ], + outputs=[ + IO.Image.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.11}""", + ), + ) + + @classmethod + async def execute( + cls, + image: Input.Image, + reference_image: Input.Image, + prompt: str, + style_strength: int, + structure_strength: int, + flavor: str, + engine: str, + portrait_mode: InputPortraitMode, + fixed_generation: bool, + ) -> IO.NodeOutput: + if get_number_of_images(image) != 1: + raise ValueError("Exactly one input image is required.") + if get_number_of_images(reference_image) != 1: + raise ValueError("Exactly one reference image is required.") + validate_image_aspect_ratio(image, (1, 3), (3, 1), strict=False) + validate_image_aspect_ratio(reference_image, (1, 3), (3, 1), strict=False) + validate_image_dimensions(image, min_height=160, min_width=160) + validate_image_dimensions(reference_image, min_height=160, min_width=160) + + is_portrait = portrait_mode["portrait_mode"] == "enabled" + portrait_style = portrait_mode.get("portrait_style", "standard") + portrait_beautifier = portrait_mode.get("portrait_beautifier", "none") + + uploaded_urls = await upload_images_to_comfyapi(cls, [image, reference_image], max_images=2) + + initial_res = await sync_op( + cls, + ApiEndpoint(path="/proxy/freepik/v1/ai/image-style-transfer", method="POST"), + response_model=TaskResponse, + data=ImageStyleTransferRequest( + image=uploaded_urls[0], + reference_image=uploaded_urls[1], + prompt=prompt if prompt else None, + style_strength=style_strength, + structure_strength=structure_strength, + is_portrait=is_portrait, + portrait_style=portrait_style if is_portrait else None, + portrait_beautifier=portrait_beautifier if is_portrait and portrait_beautifier != "none" else None, + flavor=flavor, + engine=engine, + fixed_generation=fixed_generation, + ), + ) + final_response = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/freepik/v1/ai/image-style-transfer/{initial_res.task_id}"), + response_model=TaskResponse, + status_extractor=lambda x: x.status, + poll_interval=10.0, + max_poll_attempts=480, + ) + return IO.NodeOutput(await download_url_to_image_tensor(final_response.generated[0])) + + +class MagnificImageRelightNode(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="MagnificImageRelightNode", + display_name="Magnific Image Relight", + category="api node/image/Magnific", + description="Relight an image with lighting adjustments and optional reference-based light transfer.", + inputs=[ + IO.Image.Input("image", tooltip="The image to relight."), + IO.String.Input( + "prompt", + multiline=True, + default="", + tooltip="Descriptive guidance for lighting. Supports emphasis notation (1-1.4).", + ), + IO.Int.Input( + "light_transfer_strength", + min=0, + max=100, + default=100, + tooltip="Intensity of light transfer application.", + display_mode=IO.NumberDisplay.slider, + ), + IO.Combo.Input( + "style", + options=[ + "standard", + "darker_but_realistic", + "clean", + "smooth", + "brighter", + "contrasted_n_hdr", + "just_composition", + ], + tooltip="Stylistic output preference.", + ), + IO.Boolean.Input( + "interpolate_from_original", + default=False, + tooltip="Restricts generation freedom to match original more closely.", + ), + IO.Boolean.Input( + "change_background", + default=True, + tooltip="Modifies background based on prompt/reference.", + ), + IO.Boolean.Input( + "preserve_details", + default=True, + tooltip="Maintains texture and fine details from original.", + ), + IO.DynamicCombo.Input( + "advanced_settings", + options=[ + IO.DynamicCombo.Option("disabled", []), + IO.DynamicCombo.Option( + "enabled", + [ + IO.Int.Input( + "whites", + min=0, + max=100, + default=50, + tooltip="Adjusts the brightest tones in the image.", + display_mode=IO.NumberDisplay.slider, + ), + IO.Int.Input( + "blacks", + min=0, + max=100, + default=50, + tooltip="Adjusts the darkest tones in the image.", + display_mode=IO.NumberDisplay.slider, + ), + IO.Int.Input( + "brightness", + min=0, + max=100, + default=50, + tooltip="Overall brightness adjustment.", + display_mode=IO.NumberDisplay.slider, + ), + IO.Int.Input( + "contrast", + min=0, + max=100, + default=50, + tooltip="Contrast adjustment.", + display_mode=IO.NumberDisplay.slider, + ), + IO.Int.Input( + "saturation", + min=0, + max=100, + default=50, + tooltip="Color saturation adjustment.", + display_mode=IO.NumberDisplay.slider, + ), + IO.Combo.Input( + "engine", + options=[ + "automatic", + "balanced", + "cool", + "real", + "illusio", + "fairy", + "colorful_anime", + "hard_transform", + "softy", + ], + tooltip="Processing engine selection.", + ), + IO.Combo.Input( + "transfer_light_a", + options=["automatic", "low", "medium", "normal", "high", "high_on_faces"], + tooltip="The intensity of light transfer.", + ), + IO.Combo.Input( + "transfer_light_b", + options=[ + "automatic", + "composition", + "straight", + "smooth_in", + "smooth_out", + "smooth_both", + "reverse_both", + "soft_in", + "soft_out", + "soft_mid", + # "strong_mid", # Commented out because requests fail when this is set. + "style_shift", + "strong_shift", + ], + tooltip="Also modifies light transfer intensity. " + "Can be combined with the previous control for varied effects.", + ), + IO.Boolean.Input( + "fixed_generation", + default=True, + tooltip="Ensures consistent output with the same settings.", + ), + ], + ), + ], + tooltip="Fine-tuning options for advanced lighting control.", + ), + IO.Image.Input( + "reference_image", + optional=True, + tooltip="Optional reference image to transfer lighting from.", + ), + ], + outputs=[ + IO.Image.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.11}""", + ), + ) + + @classmethod + async def execute( + cls, + image: Input.Image, + prompt: str, + light_transfer_strength: int, + style: str, + interpolate_from_original: bool, + change_background: bool, + preserve_details: bool, + advanced_settings: InputAdvancedSettings, + reference_image: Input.Image | None = None, + ) -> IO.NodeOutput: + if get_number_of_images(image) != 1: + raise ValueError("Exactly one input image is required.") + if reference_image is not None and get_number_of_images(reference_image) != 1: + raise ValueError("Exactly one reference image is required.") + validate_image_aspect_ratio(image, (1, 3), (3, 1), strict=False) + validate_image_dimensions(image, min_height=160, min_width=160) + if reference_image is not None: + validate_image_aspect_ratio(reference_image, (1, 3), (3, 1), strict=False) + validate_image_dimensions(reference_image, min_height=160, min_width=160) + + image_url = (await upload_images_to_comfyapi(cls, image, max_images=1))[0] + reference_url = None + if reference_image is not None: + reference_url = (await upload_images_to_comfyapi(cls, reference_image, max_images=1))[0] + + adv_settings = None + if advanced_settings["advanced_settings"] == "enabled": + adv_settings = ImageRelightAdvancedSettingsRequest( + whites=advanced_settings["whites"], + blacks=advanced_settings["blacks"], + brightness=advanced_settings["brightness"], + contrast=advanced_settings["contrast"], + saturation=advanced_settings["saturation"], + engine=advanced_settings["engine"], + transfer_light_a=advanced_settings["transfer_light_a"], + transfer_light_b=advanced_settings["transfer_light_b"], + fixed_generation=advanced_settings["fixed_generation"], + ) + + initial_res = await sync_op( + cls, + ApiEndpoint(path="/proxy/freepik/v1/ai/image-relight", method="POST"), + response_model=TaskResponse, + data=ImageRelightRequest( + image=image_url, + prompt=prompt if prompt else None, + transfer_light_from_reference_image=reference_url, + light_transfer_strength=light_transfer_strength, + interpolate_from_original=interpolate_from_original, + change_background=change_background, + style=style, + preserve_details=preserve_details, + advanced_settings=adv_settings, + ), + ) + final_response = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/freepik/v1/ai/image-relight/{initial_res.task_id}"), + response_model=TaskResponse, + status_extractor=lambda x: x.status, + poll_interval=10.0, + max_poll_attempts=480, + ) + return IO.NodeOutput(await download_url_to_image_tensor(final_response.generated[0])) + + +class MagnificImageSkinEnhancerNode(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="MagnificImageSkinEnhancerNode", + display_name="Magnific Image Skin Enhancer", + category="api node/image/Magnific", + description="Skin enhancement for portraits with multiple processing modes.", + inputs=[ + IO.Image.Input("image", tooltip="The portrait image to enhance."), + IO.Int.Input( + "sharpen", + min=0, + max=100, + default=0, + tooltip="Sharpening intensity level.", + display_mode=IO.NumberDisplay.slider, + ), + IO.Int.Input( + "smart_grain", + min=0, + max=100, + default=2, + tooltip="Smart grain intensity level.", + display_mode=IO.NumberDisplay.slider, + ), + IO.DynamicCombo.Input( + "mode", + options=[ + IO.DynamicCombo.Option("creative", []), + IO.DynamicCombo.Option( + "faithful", + [ + IO.Int.Input( + "skin_detail", + min=0, + max=100, + default=80, + tooltip="Skin detail enhancement level.", + display_mode=IO.NumberDisplay.slider, + ), + ], + ), + IO.DynamicCombo.Option( + "flexible", + [ + IO.Combo.Input( + "optimized_for", + options=[ + "enhance_skin", + "improve_lighting", + "enhance_everything", + "transform_to_real", + "no_make_up", + ], + tooltip="Enhancement optimization target.", + ), + ], + ), + ], + tooltip="Processing mode: creative for artistic enhancement, " + "faithful for preserving original appearance, " + "flexible for targeted optimization.", + ), + ], + outputs=[ + IO.Image.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["mode"]), + expr=""" + ( + $rates := {"creative": 0.29, "faithful": 0.37, "flexible": 0.45}; + {"type":"usd","usd": $lookup($rates, widgets.mode)} + ) + """, + ), + ) + + @classmethod + async def execute( + cls, + image: Input.Image, + sharpen: int, + smart_grain: int, + mode: InputSkinEnhancerMode, + ) -> IO.NodeOutput: + if get_number_of_images(image) != 1: + raise ValueError("Exactly one input image is required.") + validate_image_aspect_ratio(image, (1, 3), (3, 1), strict=False) + validate_image_dimensions(image, min_height=160, min_width=160) + + image_url = (await upload_images_to_comfyapi(cls, image, max_images=1, total_pixels=4096 * 4096))[0] + selected_mode = mode["mode"] + + if selected_mode == "creative": + endpoint = "creative" + data = ImageSkinEnhancerCreativeRequest( + image=image_url, + sharpen=sharpen, + smart_grain=smart_grain, + ) + elif selected_mode == "faithful": + endpoint = "faithful" + data = ImageSkinEnhancerFaithfulRequest( + image=image_url, + sharpen=sharpen, + smart_grain=smart_grain, + skin_detail=mode["skin_detail"], + ) + else: # flexible + endpoint = "flexible" + data = ImageSkinEnhancerFlexibleRequest( + image=image_url, + sharpen=sharpen, + smart_grain=smart_grain, + optimized_for=mode["optimized_for"], + ) + + initial_res = await sync_op( + cls, + ApiEndpoint(path=f"/proxy/freepik/v1/ai/skin-enhancer/{endpoint}", method="POST"), + response_model=TaskResponse, + data=data, + ) + final_response = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/freepik/v1/ai/skin-enhancer/{initial_res.task_id}"), + response_model=TaskResponse, + status_extractor=lambda x: x.status, + poll_interval=10.0, + max_poll_attempts=480, + ) + return IO.NodeOutput(await download_url_to_image_tensor(final_response.generated[0])) + + +class MagnificExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[IO.ComfyNode]]: + return [ + # MagnificImageUpscalerCreativeNode, + # MagnificImageUpscalerPreciseV2Node, + MagnificImageStyleTransferNode, + MagnificImageRelightNode, + MagnificImageSkinEnhancerNode, + ] + + +async def comfy_entrypoint() -> MagnificExtension: + return MagnificExtension() diff --git a/comfy_api_nodes/util/conversions.py b/comfy_api_nodes/util/conversions.py index 0e15a0efe..3e37e8a8c 100644 --- a/comfy_api_nodes/util/conversions.py +++ b/comfy_api_nodes/util/conversions.py @@ -56,15 +56,14 @@ def image_tensor_pair_to_batch(image1: torch.Tensor, image2: torch.Tensor) -> to def tensor_to_bytesio( image: torch.Tensor, *, - total_pixels: int = 2048 * 2048, + total_pixels: int | None = 2048 * 2048, mime_type: str = "image/png", ) -> BytesIO: """Converts a torch.Tensor image to a named BytesIO object. Args: image: Input torch.Tensor image. - name: Optional filename for the BytesIO object. - total_pixels: Maximum total pixels for potential downscaling. + total_pixels: Maximum total pixels for downscaling. If None, no downscaling is performed. mime_type: Target image MIME type (e.g., 'image/png', 'image/jpeg', 'image/webp', 'video/mp4'). Returns: @@ -79,13 +78,14 @@ def tensor_to_bytesio( return img_binary -def tensor_to_pil(image: torch.Tensor, total_pixels: int = 2048 * 2048) -> Image.Image: +def tensor_to_pil(image: torch.Tensor, total_pixels: int | None = 2048 * 2048) -> Image.Image: """Converts a single torch.Tensor image [H, W, C] to a PIL Image, optionally downscaling.""" if len(image.shape) > 3: image = image[0] # TODO: remove alpha if not allowed and present input_tensor = image.cpu() - input_tensor = downscale_image_tensor(input_tensor.unsqueeze(0), total_pixels=total_pixels).squeeze() + if total_pixels is not None: + input_tensor = downscale_image_tensor(input_tensor.unsqueeze(0), total_pixels=total_pixels).squeeze() image_np = (input_tensor.numpy() * 255).astype(np.uint8) img = Image.fromarray(image_np) return img @@ -93,14 +93,14 @@ def tensor_to_pil(image: torch.Tensor, total_pixels: int = 2048 * 2048) -> Image def tensor_to_base64_string( image_tensor: torch.Tensor, - total_pixels: int = 2048 * 2048, + total_pixels: int | None = 2048 * 2048, mime_type: str = "image/png", ) -> str: """Convert [B, H, W, C] or [H, W, C] tensor to a base64 string. Args: image_tensor: Input torch.Tensor image. - total_pixels: Maximum total pixels for potential downscaling. + total_pixels: Maximum total pixels for downscaling. If None, no downscaling is performed. mime_type: Target image MIME type (e.g., 'image/png', 'image/jpeg', 'image/webp', 'video/mp4'). Returns: @@ -161,14 +161,14 @@ def downscale_image_tensor_by_max_side(image: torch.Tensor, *, max_side: int) - def tensor_to_data_uri( image_tensor: torch.Tensor, - total_pixels: int = 2048 * 2048, + total_pixels: int | None = 2048 * 2048, mime_type: str = "image/png", ) -> str: """Converts a tensor image to a Data URI string. Args: image_tensor: Input torch.Tensor image. - total_pixels: Maximum total pixels for potential downscaling. + total_pixels: Maximum total pixels for downscaling. If None, no downscaling is performed. mime_type: Target image MIME type (e.g., 'image/png', 'image/jpeg', 'image/webp'). Returns: diff --git a/comfy_api_nodes/util/upload_helpers.py b/comfy_api_nodes/util/upload_helpers.py index 2190f9639..3153f2b98 100644 --- a/comfy_api_nodes/util/upload_helpers.py +++ b/comfy_api_nodes/util/upload_helpers.py @@ -49,7 +49,7 @@ async def upload_images_to_comfyapi( mime_type: str | None = None, wait_label: str | None = "Uploading", show_batch_index: bool = True, - total_pixels: int = 2048 * 2048, + total_pixels: int | None = 2048 * 2048, ) -> list[str]: """ Uploads images to ComfyUI API and returns download URLs. From ad53e78f11aba710166e6658f4ad508ee52beb58 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Jukka=20Sepp=C3=A4nen?= <40791699+kijai@users.noreply.github.com> Date: Tue, 27 Jan 2026 02:25:00 +0200 Subject: [PATCH 520/647] Fix Noise_EmptyNoise when using nested latents (#12089) --- comfy_extras/nodes_custom_sampler.py | 9 ++++++++- 1 file changed, 8 insertions(+), 1 deletion(-) diff --git a/comfy_extras/nodes_custom_sampler.py b/comfy_extras/nodes_custom_sampler.py index a4d84ddf7..8afd13acf 100644 --- a/comfy_extras/nodes_custom_sampler.py +++ b/comfy_extras/nodes_custom_sampler.py @@ -701,7 +701,14 @@ class Noise_EmptyNoise: def generate_noise(self, input_latent): latent_image = input_latent["samples"] - return torch.zeros(latent_image.shape, dtype=latent_image.dtype, layout=latent_image.layout, device="cpu") + if latent_image.is_nested: + tensors = latent_image.unbind() + zeros = [] + for t in tensors: + zeros.append(torch.zeros(t.shape, dtype=t.dtype, layout=t.layout, device="cpu")) + return comfy.nested_tensor.NestedTensor(zeros) + else: + return torch.zeros(latent_image.shape, dtype=latent_image.dtype, layout=latent_image.layout, device="cpu") class Noise_RandomNoise: From 6516ab335da4da56242d0cb4180aed3a572a94e6 Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Mon, 26 Jan 2026 16:40:19 -0800 Subject: [PATCH 521/647] wan-vae: Switch off feature cache for single frame (#12090) The code throughout is None safe to just skip the feature cache saving step if none. Set it none in single frame use so qwen doesn't burn VRAM on the unused cache. --- comfy/ldm/wan/vae.py | 9 ++++++--- 1 file changed, 6 insertions(+), 3 deletions(-) diff --git a/comfy/ldm/wan/vae.py b/comfy/ldm/wan/vae.py index 40e767213..fd125ceed 100644 --- a/comfy/ldm/wan/vae.py +++ b/comfy/ldm/wan/vae.py @@ -479,10 +479,12 @@ class WanVAE(nn.Module): def encode(self, x): conv_idx = [0] - feat_map = [None] * count_conv3d(self.decoder) ## cache t = x.shape[2] iter_ = 1 + (t - 1) // 4 + feat_map = None + if iter_ > 1: + feat_map = [None] * count_conv3d(self.decoder) ## 对encode输入的x，按时间拆分为1、4、4、4.... for i in range(iter_): conv_idx = [0] @@ -502,10 +504,11 @@ class WanVAE(nn.Module): def decode(self, z): conv_idx = [0] - feat_map = [None] * count_conv3d(self.decoder) # z: [b,c,t,h,w] - iter_ = z.shape[2] + feat_map = None + if iter_ > 1: + feat_map = [None] * count_conv3d(self.decoder) x = self.conv2(z) for i in range(iter_): conv_idx = [0] From 2db3b0ff90032564e8a4d45a5d68cbf9f6c41af7 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 26 Jan 2026 16:49:31 -0800 Subject: [PATCH 522/647] Update amd portable for rocm 7.2 (#12101) * Update amd portable for rocm 7.2 * Update Python patch version in release workflow --- .github/workflows/release-stable-all.yml | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/.github/workflows/release-stable-all.yml b/.github/workflows/release-stable-all.yml index d72ece2ce..8f07a7b1c 100644 --- a/.github/workflows/release-stable-all.yml +++ b/.github/workflows/release-stable-all.yml @@ -20,7 +20,7 @@ jobs: git_tag: ${{ inputs.git_tag }} cache_tag: "cu130" python_minor: "13" - python_patch: "9" + python_patch: "11" rel_name: "nvidia" rel_extra_name: "" test_release: true @@ -65,11 +65,11 @@ jobs: contents: "write" packages: "write" pull-requests: "read" - name: "Release AMD ROCm 7.1.1" + name: "Release AMD ROCm 7.2" uses: ./.github/workflows/stable-release.yml with: git_tag: ${{ inputs.git_tag }} - cache_tag: "rocm711" + cache_tag: "rocm72" python_minor: "12" python_patch: "10" rel_name: "amd" From ac91c340f4f58b1503276599c9b32f437eecc62a Mon Sep 17 00:00:00 2001 From: ComfyUI Wiki Date: Tue, 27 Jan 2026 10:39:39 +0800 Subject: [PATCH 523/647] Update workflow templates to v0.8.23 (#12102) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 8d38c114b..31493a44c 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,5 +1,5 @@ comfyui-frontend-package==1.37.11 -comfyui-workflow-templates==0.8.15 +comfyui-workflow-templates==0.8.23 comfyui-embedded-docs==0.4.0 torch torchsde From 5f62440fbbfb2a57ec4b4d4ad370332fed5e6625 Mon Sep 17 00:00:00 2001 From: ComfyUI Wiki Date: Tue, 27 Jan 2026 11:47:33 +0800 Subject: [PATCH 524/647] chore: update workflow templates to v0.8.24 (#12103) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 31493a44c..666a0e35b 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,5 +1,5 @@ comfyui-frontend-package==1.37.11 -comfyui-workflow-templates==0.8.23 +comfyui-workflow-templates==0.8.24 comfyui-embedded-docs==0.4.0 torch torchsde From 09725967cf76304371c390ca1d6483e04061da48 Mon Sep 17 00:00:00 2001 From: comfyanonymous Date: Mon, 26 Jan 2026 23:08:01 -0500 Subject: [PATCH 525/647] ComfyUI version v0.11.0 --- comfyui_version.py | 2 +- pyproject.toml | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/comfyui_version.py b/comfyui_version.py index 952d413db..d56466db2 100644 --- a/comfyui_version.py +++ b/comfyui_version.py @@ -1,3 +1,3 @@ # This file is automatically generated by the build process when version is # updated in pyproject.toml. -__version__ = "0.10.0" +__version__ = "0.11.0" diff --git a/pyproject.toml b/pyproject.toml index 120b6c751..c0e787abd 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -1,6 +1,6 @@ [project] name = "ComfyUI" -version = "0.10.0" +version = "0.11.0" readme = "README.md" license = { file = "LICENSE" } requires-python = ">=3.10" From dcff27fe3fca80b610de5c648eb55c68e6154c57 Mon Sep 17 00:00:00 2001 From: guill Date: Tue, 27 Jan 2026 13:03:29 -0800 Subject: [PATCH 526/647] Add support for dev-only nodes. (#12106) When a node is declared as dev-only, it doesn't show in the default UI unless the dev mode is enabled in the settings. The intention is to allow nodes related to unit testing to be included in ComfyUI distributions without confusing the average user. --- comfy/comfy_types/node_typing.py | 2 ++ comfy_api/latest/_io.py | 16 ++++++++++++++++ server.py | 2 ++ 3 files changed, 20 insertions(+) diff --git a/comfy/comfy_types/node_typing.py b/comfy/comfy_types/node_typing.py index 071b98332..0194b7d70 100644 --- a/comfy/comfy_types/node_typing.py +++ b/comfy/comfy_types/node_typing.py @@ -236,6 +236,8 @@ class ComfyNodeABC(ABC): """Flags a node as experimental, informing users that it may change or not work as expected.""" DEPRECATED: bool """Flags a node as deprecated, indicating to users that they should find alternatives to this node.""" + DEV_ONLY: bool + """Flags a node as dev-only, hiding it from search/menus unless dev mode is enabled.""" API_NODE: Optional[bool] """Flags a node as an API node. See: https://docs.comfy.org/tutorials/api-nodes/overview.""" diff --git a/comfy_api/latest/_io.py b/comfy_api/latest/_io.py index 03c77a531..be759952e 100644 --- a/comfy_api/latest/_io.py +++ b/comfy_api/latest/_io.py @@ -1247,6 +1247,7 @@ class NodeInfoV1: output_node: bool=None deprecated: bool=None experimental: bool=None + dev_only: bool=None api_node: bool=None price_badge: dict | None = None search_aliases: list[str]=None @@ -1264,6 +1265,7 @@ class NodeInfoV3: output_node: bool=None deprecated: bool=None experimental: bool=None + dev_only: bool=None api_node: bool=None price_badge: dict | None = None @@ -1375,6 +1377,8 @@ class Schema: """Flags a node as deprecated, indicating to users that they should find alternatives to this node.""" is_experimental: bool=False """Flags a node as experimental, informing users that it may change or not work as expected.""" + is_dev_only: bool=False + """Flags a node as dev-only, hiding it from search/menus unless dev mode is enabled.""" is_api_node: bool=False """Flags a node as an API node. See: https://docs.comfy.org/tutorials/api-nodes/overview.""" price_badge: PriceBadge | None = None @@ -1485,6 +1489,7 @@ class Schema: output_node=self.is_output_node, deprecated=self.is_deprecated, experimental=self.is_experimental, + dev_only=self.is_dev_only, api_node=self.is_api_node, python_module=getattr(cls, "RELATIVE_PYTHON_MODULE", "nodes"), price_badge=self.price_badge.as_dict(self.inputs) if self.price_badge is not None else None, @@ -1519,6 +1524,7 @@ class Schema: output_node=self.is_output_node, deprecated=self.is_deprecated, experimental=self.is_experimental, + dev_only=self.is_dev_only, api_node=self.is_api_node, python_module=getattr(cls, "RELATIVE_PYTHON_MODULE", "nodes"), price_badge=self.price_badge.as_dict(self.inputs) if self.price_badge is not None else None, @@ -1791,6 +1797,14 @@ class _ComfyNodeBaseInternal(_ComfyNodeInternal): cls.GET_SCHEMA() return cls._DEPRECATED + _DEV_ONLY = None + @final + @classproperty + def DEV_ONLY(cls): # noqa + if cls._DEV_ONLY is None: + cls.GET_SCHEMA() + return cls._DEV_ONLY + _API_NODE = None @final @classproperty @@ -1893,6 +1907,8 @@ class _ComfyNodeBaseInternal(_ComfyNodeInternal): cls._EXPERIMENTAL = schema.is_experimental if cls._DEPRECATED is None: cls._DEPRECATED = schema.is_deprecated + if cls._DEV_ONLY is None: + cls._DEV_ONLY = schema.is_dev_only if cls._API_NODE is None: cls._API_NODE = schema.is_api_node if cls._OUTPUT_NODE is None: diff --git a/server.py b/server.py index 1888745b7..2aee5cc06 100644 --- a/server.py +++ b/server.py @@ -679,6 +679,8 @@ class PromptServer(): info['deprecated'] = True if getattr(obj_class, "EXPERIMENTAL", False): info['experimental'] = True + if getattr(obj_class, "DEV_ONLY", False): + info['dev_only'] = True if hasattr(obj_class, 'API_NODE'): info['api_node'] = obj_class.API_NODE From 4f5bd39b1c81cc9b71fa3ada2e4910dfa62a8692 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 27 Jan 2026 16:58:48 -0800 Subject: [PATCH 527/647] Update Python 3.14 compatibility notes in README (#12127) --- README.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/README.md b/README.md index c56e05d07..96dc2904b 100644 --- a/README.md +++ b/README.md @@ -208,7 +208,7 @@ comfy install ## Manual Install (Windows, Linux) -Python 3.14 works but you may encounter issues with the torch compile node. The free threaded variant is still missing some dependencies. +Python 3.14 works but some custom nodes may have issues. The free threaded variant works but some dependencies will enable the GIL so it's not fully supported. Python 3.13 is very well supported. If you have trouble with some custom node dependencies on 3.13 you can try 3.12 From 6c5f906bf2443c488c409b6998e9da4bb0afe55a Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Wed, 28 Jan 2026 19:46:57 +0200 Subject: [PATCH 528/647] feat(api-nodes): add Grok Imagine nodes (#12136) --- comfy_api_nodes/apis/grok.py | 67 ++++++ comfy_api_nodes/nodes_grok.py | 417 ++++++++++++++++++++++++++++++++++ 2 files changed, 484 insertions(+) create mode 100644 comfy_api_nodes/apis/grok.py create mode 100644 comfy_api_nodes/nodes_grok.py diff --git a/comfy_api_nodes/apis/grok.py b/comfy_api_nodes/apis/grok.py new file mode 100644 index 000000000..8e3c79ab9 --- /dev/null +++ b/comfy_api_nodes/apis/grok.py @@ -0,0 +1,67 @@ +from pydantic import BaseModel, Field + + +class ImageGenerationRequest(BaseModel): + model: str = Field(...) + prompt: str = Field(...) + aspect_ratio: str = Field(...) + n: int = Field(...) + seed: int = Field(...) + response_for: str = Field("url") + + +class InputUrlObject(BaseModel): + url: str = Field(...) + + +class ImageEditRequest(BaseModel): + model: str = Field(...) + image: InputUrlObject = Field(...) + prompt: str = Field(...) + resolution: str = Field(...) + n: int = Field(...) + seed: int = Field(...) + response_for: str = Field("url") + + +class VideoGenerationRequest(BaseModel): + model: str = Field(...) + prompt: str = Field(...) + image: InputUrlObject | None = Field(...) + duration: int = Field(...) + aspect_ratio: str | None = Field(...) + resolution: str = Field(...) + seed: int = Field(...) + + +class VideoEditRequest(BaseModel): + model: str = Field(...) + prompt: str = Field(...) + video: InputUrlObject = Field(...) + seed: int = Field(...) + + +class ImageResponseObject(BaseModel): + url: str | None = Field(None) + b64_json: str | None = Field(None) + revised_prompt: str | None = Field(None) + + +class ImageGenerationResponse(BaseModel): + data: list[ImageResponseObject] = Field(...) + + +class VideoGenerationResponse(BaseModel): + request_id: str = Field(...) + + +class VideoResponseObject(BaseModel): + url: str = Field(...) + upsampled_prompt: str | None = Field(None) + duration: int = Field(...) + + +class VideoStatusResponse(BaseModel): + status: str | None = Field(None) + video: VideoResponseObject | None = Field(None) + model: str | None = Field(None) diff --git a/comfy_api_nodes/nodes_grok.py b/comfy_api_nodes/nodes_grok.py new file mode 100644 index 000000000..da15e97ea --- /dev/null +++ b/comfy_api_nodes/nodes_grok.py @@ -0,0 +1,417 @@ +import torch +from typing_extensions import override + +from comfy_api.latest import IO, ComfyExtension, Input +from comfy_api_nodes.apis.grok import ( + ImageEditRequest, + ImageGenerationRequest, + ImageGenerationResponse, + InputUrlObject, + VideoEditRequest, + VideoGenerationRequest, + VideoGenerationResponse, + VideoStatusResponse, +) +from comfy_api_nodes.util import ( + ApiEndpoint, + download_url_to_image_tensor, + download_url_to_video_output, + get_fs_object_size, + get_number_of_images, + poll_op, + sync_op, + tensor_to_base64_string, + upload_video_to_comfyapi, + validate_string, + validate_video_duration, +) + + +class GrokImageNode(IO.ComfyNode): + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="GrokImageNode", + display_name="Grok Image", + category="api node/image/Grok", + description="Generate images using Grok based on a text prompt", + inputs=[ + IO.Combo.Input("model", options=["grok-imagine-image-beta"]), + IO.String.Input( + "prompt", + multiline=True, + tooltip="The text prompt used to generate the image", + ), + IO.Combo.Input( + "aspect_ratio", + options=[ + "1:1", + "2:3", + "3:2", + "3:4", + "4:3", + "9:16", + "16:9", + "9:19.5", + "19.5:9", + "9:20", + "20:9", + "1:2", + "2:1", + ], + ), + IO.Int.Input( + "number_of_images", + default=1, + min=1, + max=10, + step=1, + tooltip="Number of images to generate", + display_mode=IO.NumberDisplay.number, + ), + IO.Int.Input( + "seed", + default=0, + min=0, + max=2147483647, + step=1, + display_mode=IO.NumberDisplay.number, + control_after_generate=True, + tooltip="Seed to determine if node should re-run; " + "actual results are nondeterministic regardless of seed.", + ), + ], + outputs=[ + IO.Image.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["number_of_images"]), + expr="""{"type":"usd","usd":0.033 * widgets.number_of_images}""", + ), + ) + + @classmethod + async def execute( + cls, + model: str, + prompt: str, + aspect_ratio: str, + number_of_images: int, + seed: int, + ) -> IO.NodeOutput: + validate_string(prompt, strip_whitespace=True, min_length=1) + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/xai/v1/images/generations", method="POST"), + data=ImageGenerationRequest( + model=model, + prompt=prompt, + aspect_ratio=aspect_ratio, + n=number_of_images, + seed=seed, + ), + response_model=ImageGenerationResponse, + ) + if len(response.data) == 1: + return IO.NodeOutput(await download_url_to_image_tensor(response.data[0].url)) + return IO.NodeOutput( + torch.cat( + [await download_url_to_image_tensor(i) for i in [str(d.url) for d in response.data if d.url]], + ) + ) + + +class GrokImageEditNode(IO.ComfyNode): + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="GrokImageEditNode", + display_name="Grok Image Edit", + category="api node/image/Grok", + description="Modify an existing image based on a text prompt", + inputs=[ + IO.Combo.Input("model", options=["grok-imagine-image-beta"]), + IO.Image.Input("image"), + IO.String.Input( + "prompt", + multiline=True, + tooltip="The text prompt used to generate the image", + ), + IO.Combo.Input("resolution", options=["1K"]), + IO.Int.Input( + "number_of_images", + default=1, + min=1, + max=10, + step=1, + tooltip="Number of edited images to generate", + display_mode=IO.NumberDisplay.number, + ), + IO.Int.Input( + "seed", + default=0, + min=0, + max=2147483647, + step=1, + display_mode=IO.NumberDisplay.number, + control_after_generate=True, + tooltip="Seed to determine if node should re-run; " + "actual results are nondeterministic regardless of seed.", + ), + ], + outputs=[ + IO.Image.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["number_of_images"]), + expr="""{"type":"usd","usd":0.002 + 0.033 * widgets.number_of_images}""", + ), + ) + + @classmethod + async def execute( + cls, + model: str, + image: Input.Image, + prompt: str, + resolution: str, + number_of_images: int, + seed: int, + ) -> IO.NodeOutput: + validate_string(prompt, strip_whitespace=True, min_length=1) + if get_number_of_images(image) != 1: + raise ValueError("Only one input image is supported.") + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/xai/v1/images/edits", method="POST"), + data=ImageEditRequest( + model=model, + image=InputUrlObject(url=f"data:image/png;base64,{tensor_to_base64_string(image)}"), + prompt=prompt, + resolution=resolution.lower(), + n=number_of_images, + seed=seed, + ), + response_model=ImageGenerationResponse, + ) + if len(response.data) == 1: + return IO.NodeOutput(await download_url_to_image_tensor(response.data[0].url)) + return IO.NodeOutput( + torch.cat( + [await download_url_to_image_tensor(i) for i in [str(d.url) for d in response.data if d.url]], + ) + ) + + +class GrokVideoNode(IO.ComfyNode): + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="GrokVideoNode", + display_name="Grok Video", + category="api node/video/Grok", + description="Generate video from a prompt or an image", + inputs=[ + IO.Combo.Input("model", options=["grok-imagine-video-beta"]), + IO.String.Input( + "prompt", + multiline=True, + tooltip="Text description of the desired video.", + ), + IO.Combo.Input( + "resolution", + options=["480p", "720p"], + tooltip="The resolution of the output video.", + ), + IO.Combo.Input( + "aspect_ratio", + options=["auto", "16:9", "4:3", "3:2", "1:1", "2:3", "3:4", "9:16"], + tooltip="The aspect ratio of the output video.", + ), + IO.Int.Input( + "duration", + default=6, + min=1, + max=15, + step=1, + tooltip="The duration of the output video in seconds.", + display_mode=IO.NumberDisplay.slider, + ), + IO.Int.Input( + "seed", + default=0, + min=0, + max=2147483647, + step=1, + display_mode=IO.NumberDisplay.number, + control_after_generate=True, + tooltip="Seed to determine if node should re-run; " + "actual results are nondeterministic regardless of seed.", + ), + IO.Image.Input("image", optional=True), + ], + outputs=[ + IO.Video.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["duration"], inputs=["image"]), + expr=""" + ( + $base := 0.181 * widgets.duration; + {"type":"usd","usd": inputs.image.connected ? $base + 0.002 : $base} + ) + """, + ), + ) + + @classmethod + async def execute( + cls, + model: str, + prompt: str, + resolution: str, + aspect_ratio: str, + duration: int, + seed: int, + image: Input.Image | None = None, + ) -> IO.NodeOutput: + image_url = None + if image is not None: + if get_number_of_images(image) != 1: + raise ValueError("Only one input image is supported.") + image_url = InputUrlObject(url=f"data:image/png;base64,{tensor_to_base64_string(image)}") + validate_string(prompt, strip_whitespace=True, min_length=1) + initial_response = await sync_op( + cls, + ApiEndpoint(path="/proxy/xai/v1/videos/generations", method="POST"), + data=VideoGenerationRequest( + model=model, + image=image_url, + prompt=prompt, + resolution=resolution, + duration=duration, + aspect_ratio=None if aspect_ratio == "auto" else aspect_ratio, + seed=seed, + ), + response_model=VideoGenerationResponse, + ) + response = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/xai/v1/videos/{initial_response.request_id}"), + status_extractor=lambda r: r.status if r.status is not None else "complete", + response_model=VideoStatusResponse, + ) + return IO.NodeOutput(await download_url_to_video_output(response.video.url)) + + +class GrokVideoEditNode(IO.ComfyNode): + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="GrokVideoEditNode", + display_name="Grok Video Edit", + category="api node/video/Grok", + description="Edit an existing video based on a text prompt.", + inputs=[ + IO.Combo.Input("model", options=["grok-imagine-video-beta"]), + IO.String.Input( + "prompt", + multiline=True, + tooltip="Text description of the desired video.", + ), + IO.Video.Input("video", tooltip="Maximum supported duration is 8.7 seconds and 50MB file size."), + IO.Int.Input( + "seed", + default=0, + min=0, + max=2147483647, + step=1, + display_mode=IO.NumberDisplay.number, + control_after_generate=True, + tooltip="Seed to determine if node should re-run; " + "actual results are nondeterministic regardless of seed.", + ), + ], + outputs=[ + IO.Video.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd": 0.191, "format": {"suffix": "/sec", "approximate": true}}""", + ), + ) + + @classmethod + async def execute( + cls, + model: str, + prompt: str, + video: Input.Video, + seed: int, + ) -> IO.NodeOutput: + validate_string(prompt, strip_whitespace=True, min_length=1) + validate_video_duration(video, min_duration=1, max_duration=8.7) + video_stream = video.get_stream_source() + video_size = get_fs_object_size(video_stream) + if video_size > 50 * 1024 * 1024: + raise ValueError(f"Video size ({video_size / 1024 / 1024:.1f}MB) exceeds 50MB limit.") + initial_response = await sync_op( + cls, + ApiEndpoint(path="/proxy/xai/v1/videos/edits", method="POST"), + data=VideoEditRequest( + model=model, + video=InputUrlObject(url=await upload_video_to_comfyapi(cls, video)), + prompt=prompt, + seed=seed, + ), + response_model=VideoGenerationResponse, + ) + response = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/xai/v1/videos/{initial_response.request_id}"), + status_extractor=lambda r: r.status if r.status is not None else "complete", + response_model=VideoStatusResponse, + ) + return IO.NodeOutput(await download_url_to_video_output(response.video.url)) + + +class GrokExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[IO.ComfyNode]]: + return [ + GrokImageNode, + GrokImageEditNode, + GrokVideoNode, + GrokVideoEditNode, + ] + + +async def comfy_entrypoint() -> GrokExtension: + return GrokExtension() From d9b856754741a680da34095c6e8afa7792ccede3 Mon Sep 17 00:00:00 2001 From: "Dr.Lt.Data" <128333288+ltdrdata@users.noreply.github.com> Date: Thu, 29 Jan 2026 02:47:37 +0900 Subject: [PATCH 529/647] bump manager version to 4.1b1 (#12140) --- manager_requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/manager_requirements.txt b/manager_requirements.txt index bea6d4927..c420cc48e 100644 --- a/manager_requirements.txt +++ b/manager_requirements.txt @@ -1 +1 @@ -comfyui_manager==4.0.5 +comfyui_manager==4.1b1 From 1711020904edd33bad7556bb70ef3ec15d4f8e5a Mon Sep 17 00:00:00 2001 From: ComfyUI Wiki Date: Thu, 29 Jan 2026 01:48:02 +0800 Subject: [PATCH 530/647] chore: update workflow templates to v0.8.27 (#12141) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 666a0e35b..4ac94cb16 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,5 +1,5 @@ comfyui-frontend-package==1.37.11 -comfyui-workflow-templates==0.8.24 +comfyui-workflow-templates==0.8.27 comfyui-embedded-docs==0.4.0 torch torchsde From c9b633d84f1dd9c2e6fb5dbc733a0687e5a9764b Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 28 Jan 2026 17:52:51 -0800 Subject: [PATCH 531/647] Add missing spacial downscale ratios. (#12146) --- comfy/latent_formats.py | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/comfy/latent_formats.py b/comfy/latent_formats.py index 38f18a83f..4b3a3798c 100644 --- a/comfy/latent_formats.py +++ b/comfy/latent_formats.py @@ -81,6 +81,7 @@ class SD_X4(LatentFormat): class SC_Prior(LatentFormat): latent_channels = 16 + spacial_downscale_ratio = 42 def __init__(self): self.scale_factor = 1.0 self.latent_rgb_factors = [ @@ -103,6 +104,7 @@ class SC_Prior(LatentFormat): ] class SC_B(LatentFormat): + spacial_downscale_ratio = 4 def __init__(self): self.scale_factor = 1.0 / 0.43 self.latent_rgb_factors = [ @@ -274,6 +276,7 @@ class Mochi(LatentFormat): class LTXV(LatentFormat): latent_channels = 128 latent_dimensions = 3 + spacial_downscale_ratio = 32 def __init__(self): self.latent_rgb_factors = [ @@ -517,6 +520,7 @@ class Wan21(LatentFormat): class Wan22(Wan21): latent_channels = 48 latent_dimensions = 3 + spacial_downscale_ratio = 16 latent_rgb_factors = [ [ 0.0119, 0.0103, 0.0046], From b0d9708974f50fce7d2448ac84e9260c87f7ade3 Mon Sep 17 00:00:00 2001 From: comfyanonymous Date: Thu, 29 Jan 2026 00:27:23 -0500 Subject: [PATCH 532/647] ComfyUI v0.11.1 --- comfyui_version.py | 2 +- pyproject.toml | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/comfyui_version.py b/comfyui_version.py index d56466db2..b1ebaa115 100644 --- a/comfyui_version.py +++ b/comfyui_version.py @@ -1,3 +1,3 @@ # This file is automatically generated by the build process when version is # updated in pyproject.toml. -__version__ = "0.11.0" +__version__ = "0.11.1" diff --git a/pyproject.toml b/pyproject.toml index c0e787abd..042f124e4 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -1,6 +1,6 @@ [project] name = "ComfyUI" -version = "0.11.0" +version = "0.11.1" readme = "README.md" license = { file = "LICENSE" } requires-python = ">=3.10" From 3aace5c8dc81d7bf1c3020edcbc4e6d37caa0a29 Mon Sep 17 00:00:00 2001 From: Christian Byrne Date: Thu, 29 Jan 2026 17:10:08 -0800 Subject: [PATCH 533/647] fix: count non-dict items in outputs_count (#12166) Move count increment before isinstance(item, dict) check so that non-dict output items (like text strings from PreviewAny node) are included in outputs_count. This aligns OSS Python with Cloud's Go implementation which uses len(itemsArray) to count ALL items regardless of type. Amp-Thread-ID: https://ampcode.com/threads/T-019c0bb5-14e0-744f-8808-1e57653f3ae3 Co-authored-by: Amp --- comfy_execution/jobs.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/comfy_execution/jobs.py b/comfy_execution/jobs.py index 97fd803b8..bf091a448 100644 --- a/comfy_execution/jobs.py +++ b/comfy_execution/jobs.py @@ -171,9 +171,10 @@ def get_outputs_summary(outputs: dict) -> tuple[int, Optional[dict]]: continue for item in items: + count += 1 + if not isinstance(item, dict): continue - count += 1 if preview_output is None and is_previewable(media_type, item): enriched = { From bbe2c13a7075bcf4de3b6744f96d84d12c334350 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Thu, 29 Jan 2026 20:52:22 -0800 Subject: [PATCH 534/647] Make empty hunyuan latent 1.0 work with the 1.5 model. (#12171) --- comfy_extras/nodes_hunyuan.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/comfy_extras/nodes_hunyuan.py b/comfy_extras/nodes_hunyuan.py index ceff657d3..774da75a3 100644 --- a/comfy_extras/nodes_hunyuan.py +++ b/comfy_extras/nodes_hunyuan.py @@ -56,7 +56,7 @@ class EmptyHunyuanLatentVideo(io.ComfyNode): @classmethod def execute(cls, width, height, length, batch_size=1) -> io.NodeOutput: latent = torch.zeros([batch_size, 16, ((length - 1) // 4) + 1, height // 8, width // 8], device=comfy.model_management.intermediate_device()) - return io.NodeOutput({"samples":latent}) + return io.NodeOutput({"samples": latent, "downscale_ratio_spacial": 8}) generate = execute # TODO: remove @@ -73,7 +73,7 @@ class EmptyHunyuanVideo15Latent(EmptyHunyuanLatentVideo): def execute(cls, width, height, length, batch_size=1) -> io.NodeOutput: # Using scale factor of 16 instead of 8 latent = torch.zeros([batch_size, 32, ((length - 1) // 4) + 1, height // 16, width // 16], device=comfy.model_management.intermediate_device()) - return io.NodeOutput({"samples": latent}) + return io.NodeOutput({"samples": latent, "downscale_ratio_spacial": 16}) class HunyuanVideo15ImageToVideo(io.ComfyNode): From 0a7993729c70933e6fe19e26c2d7f60f6ebaf46a Mon Sep 17 00:00:00 2001 From: Jedrzej Kosinski Date: Fri, 30 Jan 2026 10:21:48 -0800 Subject: [PATCH 535/647] Remove NodeInfoV3-related code; we are almost 100% guaranteed to stick with NodeInfoV1 for the foreseable future (#12147) Co-authored-by: guill --- comfy_api/latest/_io.py | 63 ----------------------------------------- 1 file changed, 63 deletions(-) diff --git a/comfy_api/latest/_io.py b/comfy_api/latest/_io.py index be759952e..f5445554f 100644 --- a/comfy_api/latest/_io.py +++ b/comfy_api/latest/_io.py @@ -1252,23 +1252,6 @@ class NodeInfoV1: price_badge: dict | None = None search_aliases: list[str]=None -@dataclass -class NodeInfoV3: - input: dict=None - output: dict=None - hidden: list[str]=None - name: str=None - display_name: str=None - description: str=None - python_module: Any = None - category: str=None - output_node: bool=None - deprecated: bool=None - experimental: bool=None - dev_only: bool=None - api_node: bool=None - price_badge: dict | None = None - @dataclass class PriceBadgeDepends: @@ -1497,40 +1480,6 @@ class Schema: ) return info - - def get_v3_info(self, cls) -> NodeInfoV3: - input_dict = {} - output_dict = {} - hidden_list = [] - # TODO: make sure dynamic types will be handled correctly - if self.inputs: - for input in self.inputs: - add_to_dict_v3(input, input_dict) - if self.outputs: - for output in self.outputs: - add_to_dict_v3(output, output_dict) - if self.hidden: - for hidden in self.hidden: - hidden_list.append(hidden.value) - - info = NodeInfoV3( - input=input_dict, - output=output_dict, - hidden=hidden_list, - name=self.node_id, - display_name=self.display_name, - description=self.description, - category=self.category, - output_node=self.is_output_node, - deprecated=self.is_deprecated, - experimental=self.is_experimental, - dev_only=self.is_dev_only, - api_node=self.is_api_node, - python_module=getattr(cls, "RELATIVE_PYTHON_MODULE", "nodes"), - price_badge=self.price_badge.as_dict(self.inputs) if self.price_badge is not None else None, - ) - return info - def get_finalized_class_inputs(d: dict[str, Any], live_inputs: dict[str, Any], include_hidden=False) -> tuple[dict[str, Any], V3Data]: out_dict = { "required": {}, @@ -1585,9 +1534,6 @@ def add_to_dict_v1(i: Input, d: dict): as_dict.pop("optional", None) d.setdefault(key, {})[i.id] = (i.get_io_type(), as_dict) -def add_to_dict_v3(io: Input | Output, d: dict): - d[io.id] = (io.get_io_type(), io.as_dict()) - class DynamicPathsDefaultValue: EMPTY_DICT = "empty_dict" @@ -1748,13 +1694,6 @@ class _ComfyNodeBaseInternal(_ComfyNodeInternal): # set hidden type_clone.hidden = HiddenHolder.from_v3_data(v3_data) return type_clone - - @final - @classmethod - def GET_NODE_INFO_V3(cls) -> dict[str, Any]: - schema = cls.GET_SCHEMA() - info = schema.get_v3_info(cls) - return asdict(info) ############################################# # V1 Backwards Compatibility code #-------------------------------------------- @@ -2107,12 +2046,10 @@ __all__ = [ "HiddenHolder", "Hidden", "NodeInfoV1", - "NodeInfoV3", "Schema", "ComfyNode", "NodeOutput", "add_to_dict_v1", - "add_to_dict_v3", "V3Data", "ImageCompare", "PriceBadgeDepends", From 016765378178cf2a565d4e87a99d2dabd4150d22 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Sat, 31 Jan 2026 00:04:43 +0200 Subject: [PATCH 536/647] feat(api-nodes): add RecraftCreateStyleNode node (#12055) Co-authored-by: Jedrzej Kosinski --- comfy_api_nodes/apis/recraft.py | 45 ++++++++++--------- comfy_api_nodes/nodes_recraft.py | 74 +++++++++++++++++++++++++++++++- 2 files changed, 98 insertions(+), 21 deletions(-) diff --git a/comfy_api_nodes/apis/recraft.py b/comfy_api_nodes/apis/recraft.py index c36d95f24..0bd7d23b3 100644 --- a/comfy_api_nodes/apis/recraft.py +++ b/comfy_api_nodes/apis/recraft.py @@ -1,11 +1,8 @@ from __future__ import annotations - - from enum import Enum -from typing import Optional -from pydantic import BaseModel, Field, conint, confloat +from pydantic import BaseModel, Field class RecraftColor: @@ -229,24 +226,24 @@ class RecraftColorObject(BaseModel): class RecraftControlsObject(BaseModel): - colors: Optional[list[RecraftColorObject]] = Field(None, description='An array of preferable colors') - background_color: Optional[RecraftColorObject] = Field(None, description='Use given color as a desired background color') - no_text: Optional[bool] = Field(None, description='Do not embed text layouts') - artistic_level: Optional[conint(ge=0, le=5)] = Field(None, description='Defines artistic tone of your image. At a simple level, the person looks straight at the camera in a static and clean style. Dynamic and eccentric levels introduce movement and creativity. The value should be in range [0..5].') + colors: list[RecraftColorObject] | None = Field(None, description='An array of preferable colors') + background_color: RecraftColorObject | None = Field(None, description='Use given color as a desired background color') + no_text: bool | None = Field(None, description='Do not embed text layouts') + artistic_level: int | None = Field(None, description='Defines artistic tone of your image. At a simple level, the person looks straight at the camera in a static and clean style. Dynamic and eccentric levels introduce movement and creativity. The value should be in range [0..5].') class RecraftImageGenerationRequest(BaseModel): prompt: str = Field(..., description='The text prompt describing the image to generate') - size: Optional[RecraftImageSize] = Field(None, description='The size of the generated image (e.g., "1024x1024")') - n: conint(ge=1, le=6) = Field(..., description='The number of images to generate') - negative_prompt: Optional[str] = Field(None, description='A text description of undesired elements on an image') - model: Optional[RecraftModel] = Field(RecraftModel.recraftv3, description='The model to use for generation (e.g., "recraftv3")') - style: Optional[str] = Field(None, description='The style to apply to the generated image (e.g., "digital_illustration")') - substyle: Optional[str] = Field(None, description='The substyle to apply to the generated image, depending on the style input') - controls: Optional[RecraftControlsObject] = Field(None, description='A set of custom parameters to tweak generation process') - style_id: Optional[str] = Field(None, description='Use a previously uploaded style as a reference; UUID') - strength: Optional[confloat(ge=0.0, le=1.0)] = Field(None, description='Defines the difference with the original image, should lie in [0, 1], where 0 means almost identical, and 1 means miserable similarity') - random_seed: Optional[int] = Field(None, description="Seed for video generation") + size: RecraftImageSize | None = Field(None, description='The size of the generated image (e.g., "1024x1024")') + n: int = Field(..., description='The number of images to generate') + negative_prompt: str | None = Field(None, description='A text description of undesired elements on an image') + model: RecraftModel | None = Field(RecraftModel.recraftv3, description='The model to use for generation (e.g., "recraftv3")') + style: str | None = Field(None, description='The style to apply to the generated image (e.g., "digital_illustration")') + substyle: str | None = Field(None, description='The substyle to apply to the generated image, depending on the style input') + controls: RecraftControlsObject | None = Field(None, description='A set of custom parameters to tweak generation process') + style_id: str | None = Field(None, description='Use a previously uploaded style as a reference; UUID') + strength: float | None = Field(None, description='Defines the difference with the original image, should lie in [0, 1], where 0 means almost identical, and 1 means miserable similarity') + random_seed: int | None = Field(None, description="Seed for video generation") # text_layout @@ -258,5 +255,13 @@ class RecraftReturnedObject(BaseModel): class RecraftImageGenerationResponse(BaseModel): created: int = Field(..., description='Unix timestamp when the generation was created') credits: int = Field(..., description='Number of credits used for the generation') - data: Optional[list[RecraftReturnedObject]] = Field(None, description='Array of generated image information') - image: Optional[RecraftReturnedObject] = Field(None, description='Single generated image') + data: list[RecraftReturnedObject] | None = Field(None, description='Array of generated image information') + image: RecraftReturnedObject | None = Field(None, description='Single generated image') + + +class RecraftCreateStyleRequest(BaseModel): + style: str = Field(..., description="realistic_image, digital_illustration, vector_illustration, or icon") + + +class RecraftCreateStyleResponse(BaseModel): + id: str = Field(..., description="UUID of the created style") diff --git a/comfy_api_nodes/nodes_recraft.py b/comfy_api_nodes/nodes_recraft.py index c01bcaece..3a1f32263 100644 --- a/comfy_api_nodes/nodes_recraft.py +++ b/comfy_api_nodes/nodes_recraft.py @@ -12,6 +12,8 @@ from comfy_api_nodes.apis.recraft import ( RecraftColor, RecraftColorChain, RecraftControls, + RecraftCreateStyleRequest, + RecraftCreateStyleResponse, RecraftImageGenerationRequest, RecraftImageGenerationResponse, RecraftImageSize, @@ -323,6 +325,75 @@ class RecraftStyleInfiniteStyleLibrary(IO.ComfyNode): return IO.NodeOutput(RecraftStyle(style_id=style_id)) +class RecraftCreateStyleNode(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="RecraftCreateStyleNode", + display_name="Recraft Create Style", + category="api node/image/Recraft", + description="Create a custom style from reference images. " + "Upload 1-5 images to use as style references. " + "Total size of all images is limited to 5 MB.", + inputs=[ + IO.Combo.Input( + "style", + options=["realistic_image", "digital_illustration"], + tooltip="The base style of the generated images.", + ), + IO.Autogrow.Input( + "images", + template=IO.Autogrow.TemplatePrefix( + IO.Image.Input("image"), + prefix="image", + min=1, + max=5, + ), + ), + ], + outputs=[ + IO.String.Output(display_name="style_id"), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd": 0.04}""", + ), + ) + + @classmethod + async def execute( + cls, + style: str, + images: IO.Autogrow.Type, + ) -> IO.NodeOutput: + files = [] + total_size = 0 + max_total_size = 5 * 1024 * 1024 # 5 MB limit + for i, img in enumerate(list(images.values())): + file_bytes = tensor_to_bytesio(img, total_pixels=2048 * 2048, mime_type="image/webp").read() + total_size += len(file_bytes) + if total_size > max_total_size: + raise Exception("Total size of all images exceeds 5 MB limit.") + files.append((f"file{i + 1}", file_bytes)) + + response = await sync_op( + cls, + endpoint=ApiEndpoint(path="/proxy/recraft/styles", method="POST"), + response_model=RecraftCreateStyleResponse, + files=files, + data=RecraftCreateStyleRequest(style=style), + content_type="multipart/form-data", + max_retries=1, + ) + + return IO.NodeOutput(response.id) + + class RecraftTextToImageNode(IO.ComfyNode): @classmethod def define_schema(cls): @@ -395,7 +466,7 @@ class RecraftTextToImageNode(IO.ComfyNode): negative_prompt: str = None, recraft_controls: RecraftControls = None, ) -> IO.NodeOutput: - validate_string(prompt, strip_whitespace=False, max_length=1000) + validate_string(prompt, strip_whitespace=False, min_length=1, max_length=1000) default_style = RecraftStyle(RecraftStyleV3.realistic_image) if recraft_style is None: recraft_style = default_style @@ -1024,6 +1095,7 @@ class RecraftExtension(ComfyExtension): RecraftStyleV3DigitalIllustrationNode, RecraftStyleV3LogoRasterNode, RecraftStyleInfiniteStyleLibrary, + RecraftCreateStyleNode, RecraftColorRGBNode, RecraftControlsNode, ] From 8aabe2403eb2a182e936d6b9514751ab075a623b Mon Sep 17 00:00:00 2001 From: pythongosssss <125205205+pythongosssss@users.noreply.github.com> Date: Fri, 30 Jan 2026 15:01:33 -0800 Subject: [PATCH 537/647] Add color type and Color to RGB Int node (#12145) * add color type and color to rgb int node * review fix for allowing output --------- Co-authored-by: Jedrzej Kosinski --- comfy_api/latest/_io.py | 15 +++++++++++++ comfy_extras/nodes_color.py | 42 +++++++++++++++++++++++++++++++++++++ nodes.py | 3 ++- 3 files changed, 59 insertions(+), 1 deletion(-) create mode 100644 comfy_extras/nodes_color.py diff --git a/comfy_api/latest/_io.py b/comfy_api/latest/_io.py index f5445554f..78f77d4b2 100644 --- a/comfy_api/latest/_io.py +++ b/comfy_api/latest/_io.py @@ -1146,6 +1146,20 @@ class ImageCompare(ComfyTypeI): def as_dict(self): return super().as_dict() + +@comfytype(io_type="COLOR") +class Color(ComfyTypeIO): + Type = str + + class Input(WidgetInput): + def __init__(self, id: str, display_name: str=None, optional=False, tooltip: str=None, + socketless: bool=True, advanced: bool=None, default: str="#ffffff"): + super().__init__(id, display_name, optional, tooltip, None, default, socketless, None, None, None, None, advanced) + self.default: str + + def as_dict(self): + return super().as_dict() + DYNAMIC_INPUT_LOOKUP: dict[str, Callable[[dict[str, Any], dict[str, Any], tuple[str, dict[str, Any]], str, list[str] | None], None]] = {} def register_dynamic_input_func(io_type: str, func: Callable[[dict[str, Any], dict[str, Any], tuple[str, dict[str, Any]], str, list[str] | None], None]): DYNAMIC_INPUT_LOOKUP[io_type] = func @@ -2038,6 +2052,7 @@ __all__ = [ "AnyType", "MultiType", "Tracks", + "Color", # Dynamic Types "MatchType", "DynamicCombo", diff --git a/comfy_extras/nodes_color.py b/comfy_extras/nodes_color.py new file mode 100644 index 000000000..80ba121cd --- /dev/null +++ b/comfy_extras/nodes_color.py @@ -0,0 +1,42 @@ +from typing_extensions import override +from comfy_api.latest import ComfyExtension, io + + +class ColorToRGBInt(io.ComfyNode): + @classmethod + def define_schema(cls) -> io.Schema: + return io.Schema( + node_id="ColorToRGBInt", + display_name="Color to RGB Int", + category="utils", + description="Convert a color to a RGB integer value.", + inputs=[ + io.Color.Input("color"), + ], + outputs=[ + io.Int.Output(display_name="rgb_int"), + ], + ) + + @classmethod + def execute( + cls, + color: str, + ) -> io.NodeOutput: + # expect format #RRGGBB + if len(color) != 7 or color[0] != "#": + raise ValueError("Color must be in format #RRGGBB") + r = int(color[1:3], 16) + g = int(color[3:5], 16) + b = int(color[5:7], 16) + return io.NodeOutput(r * 256 * 256 + g * 256 + b) + + +class ColorExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[io.ComfyNode]]: + return [ColorToRGBInt] + + +async def comfy_entrypoint() -> ColorExtension: + return ColorExtension() diff --git a/nodes.py b/nodes.py index ad474d3cd..1cb43d9e2 100644 --- a/nodes.py +++ b/nodes.py @@ -2432,7 +2432,8 @@ async def init_builtin_extra_nodes(): "nodes_wanmove.py", "nodes_image_compare.py", "nodes_zimage.py", - "nodes_lora_debug.py" + "nodes_lora_debug.py", + "nodes_color.py" ] import_failed = [] From 4064062e7d2d5abdca6767b6944e331b70065ee8 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Fri, 30 Jan 2026 17:20:06 -0800 Subject: [PATCH 538/647] Update python patch version in dep workflow. (#12184) --- .github/workflows/windows_release_dependencies.yml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/.github/workflows/windows_release_dependencies.yml b/.github/workflows/windows_release_dependencies.yml index f61ee21a2..93e01ac93 100644 --- a/.github/workflows/windows_release_dependencies.yml +++ b/.github/workflows/windows_release_dependencies.yml @@ -29,7 +29,7 @@ on: description: 'python patch version' required: true type: string - default: "9" + default: "11" # push: # branches: # - master From b8f848bfe38be79b9dba693bc7a214c9aefd597a Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Fri, 30 Jan 2026 21:12:48 -0800 Subject: [PATCH 539/647] Fix model not working with any res. (#12186) --- comfy/ldm/cosmos/predict2.py | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/comfy/ldm/cosmos/predict2.py b/comfy/ldm/cosmos/predict2.py index 07a4fc79f..c270e6333 100644 --- a/comfy/ldm/cosmos/predict2.py +++ b/comfy/ldm/cosmos/predict2.py @@ -13,6 +13,7 @@ from torchvision import transforms import comfy.patcher_extension from comfy.ldm.modules.attention import optimized_attention +import comfy.ldm.common_dit def apply_rotary_pos_emb( t: torch.Tensor, @@ -835,6 +836,8 @@ class MiniTrainDIT(nn.Module): padding_mask: Optional[torch.Tensor] = None, **kwargs, ): + orig_shape = list(x.shape) + x = comfy.ldm.common_dit.pad_to_patch_size(x, (self.patch_temporal, self.patch_spatial, self.patch_spatial)) x_B_C_T_H_W = x timesteps_B_T = timesteps crossattn_emb = context @@ -882,5 +885,5 @@ class MiniTrainDIT(nn.Module): ) x_B_T_H_W_O = self.final_layer(x_B_T_H_W_D, t_embedding_B_T_D, adaln_lora_B_T_3D=adaln_lora_B_T_3D) - x_B_C_Tt_Hp_Wp = self.unpatchify(x_B_T_H_W_O) + x_B_C_Tt_Hp_Wp = self.unpatchify(x_B_T_H_W_O)[:, :, :orig_shape[-3], :orig_shape[-2], :orig_shape[-1]] return x_B_C_Tt_Hp_Wp From 6e469a3f355782067b80fa072448a1b5d1322719 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Sat, 31 Jan 2026 08:44:08 +0200 Subject: [PATCH 540/647] feat(api-nodes): add Q3 models and support for Extend and MultiFrame Vidu endpoints (#12175) Co-authored-by: Jedrzej Kosinski --- comfy_api_nodes/apis/vidu.py | 24 ++ comfy_api_nodes/nodes_vidu.py | 549 +++++++++++++++++++++++++++++++++- 2 files changed, 571 insertions(+), 2 deletions(-) diff --git a/comfy_api_nodes/apis/vidu.py b/comfy_api_nodes/apis/vidu.py index a9bb6f7ce..469adcdbc 100644 --- a/comfy_api_nodes/apis/vidu.py +++ b/comfy_api_nodes/apis/vidu.py @@ -6,6 +6,30 @@ class SubjectReference(BaseModel): images: list[str] = Field(...) +class FrameSetting(BaseModel): + prompt: str = Field(...) + key_image: str = Field(...) + duration: int = Field(...) + + +class TaskMultiFrameCreationRequest(BaseModel): + model: str = Field(...) + seed: int = Field(..., ge=0, le=2147483647) + resolution: str = Field(...) + start_image: str = Field(...) + image_settings: list[FrameSetting] = Field(...) + + +class TaskExtendCreationRequest(BaseModel): + model: str = Field(...) + prompt: str = Field(..., max_length=2000) + duration: int = Field(...) + seed: int = Field(..., ge=0, le=2147483647) + resolution: str = Field(...) + images: list[str] | None = Field(None, description="Base64 encoded string or image URL") + video_url: str = Field(..., description="URL of the video to extend") + + class TaskCreationRequest(BaseModel): model: str = Field(...) prompt: str = Field(..., max_length=2000) diff --git a/comfy_api_nodes/nodes_vidu.py b/comfy_api_nodes/nodes_vidu.py index b9114c4bb..80de14dfe 100644 --- a/comfy_api_nodes/nodes_vidu.py +++ b/comfy_api_nodes/nodes_vidu.py @@ -2,9 +2,12 @@ from typing_extensions import override from comfy_api.latest import IO, ComfyExtension, Input from comfy_api_nodes.apis.vidu import ( + FrameSetting, SubjectReference, TaskCreationRequest, TaskCreationResponse, + TaskExtendCreationRequest, + TaskMultiFrameCreationRequest, TaskResult, TaskStatusResponse, ) @@ -14,11 +17,14 @@ from comfy_api_nodes.util import ( get_number_of_images, poll_op, sync_op, + upload_image_to_comfyapi, upload_images_to_comfyapi, + upload_video_to_comfyapi, validate_image_aspect_ratio, validate_image_dimensions, validate_images_aspect_ratio_closeness, validate_string, + validate_video_duration, ) VIDU_TEXT_TO_VIDEO = "/proxy/vidu/text2video" @@ -31,7 +37,8 @@ VIDU_GET_GENERATION_STATUS = "/proxy/vidu/tasks/%s/creations" async def execute_task( cls: type[IO.ComfyNode], vidu_endpoint: str, - payload: TaskCreationRequest, + payload: TaskCreationRequest | TaskExtendCreationRequest | TaskMultiFrameCreationRequest, + max_poll_attempts: int = 320, ) -> list[TaskResult]: task_creation_response = await sync_op( cls, @@ -47,7 +54,7 @@ async def execute_task( response_model=TaskStatusResponse, status_extractor=lambda r: r.state, progress_extractor=lambda r: r.progress, - max_poll_attempts=320, + max_poll_attempts=max_poll_attempts, ) if not response.creations: raise RuntimeError( @@ -940,6 +947,540 @@ class Vidu2StartEndToVideoNode(IO.ComfyNode): return IO.NodeOutput(await download_url_to_video_output(results[0].url)) +class ViduExtendVideoNode(IO.ComfyNode): + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="ViduExtendVideoNode", + display_name="Vidu Video Extension", + category="api node/video/Vidu", + description="Extend an existing video by generating additional frames.", + inputs=[ + IO.DynamicCombo.Input( + "model", + options=[ + IO.DynamicCombo.Option( + "viduq2-pro", + [ + IO.Int.Input( + "duration", + default=4, + min=1, + max=7, + step=1, + display_mode=IO.NumberDisplay.slider, + tooltip="Duration of the extended video in seconds.", + ), + IO.Combo.Input( + "resolution", + options=["720p", "1080p"], + tooltip="Resolution of the output video.", + ), + ], + ), + IO.DynamicCombo.Option( + "viduq2-turbo", + [ + IO.Int.Input( + "duration", + default=4, + min=1, + max=7, + step=1, + display_mode=IO.NumberDisplay.slider, + tooltip="Duration of the extended video in seconds.", + ), + IO.Combo.Input( + "resolution", + options=["720p", "1080p"], + tooltip="Resolution of the output video.", + ), + ], + ), + ], + tooltip="Model to use for video extension.", + ), + IO.Video.Input( + "video", + tooltip="The source video to extend.", + ), + IO.String.Input( + "prompt", + multiline=True, + default="", + tooltip="An optional text prompt for the extended video (max 2000 characters).", + ), + IO.Int.Input( + "seed", + default=1, + min=0, + max=2147483647, + step=1, + display_mode=IO.NumberDisplay.number, + control_after_generate=True, + ), + IO.Image.Input("end_frame", optional=True), + ], + outputs=[ + IO.Video.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["model", "model.duration", "model.resolution"]), + expr=""" + ( + $m := widgets.model; + $d := $lookup(widgets, "model.duration"); + $res := $lookup(widgets, "model.resolution"); + $contains($m, "pro") + ? ( + $base := $lookup({"720p": 0.15, "1080p": 0.3}, $res); + $perSec := $lookup({"720p": 0.05, "1080p": 0.075}, $res); + {"type":"usd","usd": $base + $perSec * ($d - 1)} + ) + : ( + $base := $lookup({"720p": 0.075, "1080p": 0.2}, $res); + $perSec := $lookup({"720p": 0.025, "1080p": 0.05}, $res); + {"type":"usd","usd": $base + $perSec * ($d - 1)} + ) + ) + """, + ), + ) + + @classmethod + async def execute( + cls, + model: dict, + video: Input.Video, + prompt: str, + seed: int, + end_frame: Input.Image | None = None, + ) -> IO.NodeOutput: + validate_string(prompt, max_length=2000) + validate_video_duration(video, min_duration=4, max_duration=55) + image_url = None + if end_frame is not None: + validate_image_aspect_ratio(end_frame, (1, 4), (4, 1)) + validate_image_dimensions(end_frame, min_width=128, min_height=128) + image_url = await upload_image_to_comfyapi(cls, end_frame, wait_label="Uploading end frame") + results = await execute_task( + cls, + "/proxy/vidu/extend", + TaskExtendCreationRequest( + model=model["model"], + prompt=prompt, + duration=model["duration"], + seed=seed, + resolution=model["resolution"], + video_url=await upload_video_to_comfyapi(cls, video, wait_label="Uploading video"), + images=[image_url] if image_url else None, + ), + max_poll_attempts=480, + ) + return IO.NodeOutput(await download_url_to_video_output(results[0].url)) + + +def _generate_frame_inputs(count: int) -> list: + """Generate input widgets for a given number of frames.""" + inputs = [] + for i in range(1, count + 1): + inputs.extend( + [ + IO.String.Input( + f"prompt{i}", + multiline=True, + default="", + tooltip=f"Text prompt for frame {i} transition.", + ), + IO.Image.Input( + f"end_image{i}", + tooltip=f"End frame image for segment {i}. Aspect ratio must be between 1:4 and 4:1.", + ), + IO.Int.Input( + f"duration{i}", + default=4, + min=2, + max=7, + step=1, + display_mode=IO.NumberDisplay.slider, + tooltip=f"Duration for segment {i} in seconds.", + ), + ] + ) + return inputs + + +class ViduMultiFrameVideoNode(IO.ComfyNode): + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="ViduMultiFrameVideoNode", + display_name="Vidu Multi-Frame Video Generation", + category="api node/video/Vidu", + description="Generate a video with multiple keyframe transitions.", + inputs=[ + IO.Combo.Input("model", options=["viduq2-pro", "viduq2-turbo"]), + IO.Image.Input( + "start_image", + tooltip="The starting frame image. Aspect ratio must be between 1:4 and 4:1.", + ), + IO.Int.Input( + "seed", + default=1, + min=0, + max=2147483647, + step=1, + display_mode=IO.NumberDisplay.number, + control_after_generate=True, + ), + IO.Combo.Input("resolution", options=["720p", "1080p"]), + IO.DynamicCombo.Input( + "frames", + options=[ + IO.DynamicCombo.Option("2", _generate_frame_inputs(2)), + IO.DynamicCombo.Option("3", _generate_frame_inputs(3)), + IO.DynamicCombo.Option("4", _generate_frame_inputs(4)), + IO.DynamicCombo.Option("5", _generate_frame_inputs(5)), + IO.DynamicCombo.Option("6", _generate_frame_inputs(6)), + IO.DynamicCombo.Option("7", _generate_frame_inputs(7)), + IO.DynamicCombo.Option("8", _generate_frame_inputs(8)), + IO.DynamicCombo.Option("9", _generate_frame_inputs(9)), + ], + tooltip="Number of keyframe transitions (2-9).", + ), + ], + outputs=[ + IO.Video.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends( + widgets=[ + "model", + "resolution", + "frames", + "frames.duration1", + "frames.duration2", + "frames.duration3", + "frames.duration4", + "frames.duration5", + "frames.duration6", + "frames.duration7", + "frames.duration8", + "frames.duration9", + ] + ), + expr=""" + ( + $m := widgets.model; + $n := $number(widgets.frames); + $is1080 := widgets.resolution = "1080p"; + $d1 := $lookup(widgets, "frames.duration1"); + $d2 := $lookup(widgets, "frames.duration2"); + $d3 := $n >= 3 ? $lookup(widgets, "frames.duration3") : 0; + $d4 := $n >= 4 ? $lookup(widgets, "frames.duration4") : 0; + $d5 := $n >= 5 ? $lookup(widgets, "frames.duration5") : 0; + $d6 := $n >= 6 ? $lookup(widgets, "frames.duration6") : 0; + $d7 := $n >= 7 ? $lookup(widgets, "frames.duration7") : 0; + $d8 := $n >= 8 ? $lookup(widgets, "frames.duration8") : 0; + $d9 := $n >= 9 ? $lookup(widgets, "frames.duration9") : 0; + $totalDuration := $d1 + $d2 + $d3 + $d4 + $d5 + $d6 + $d7 + $d8 + $d9; + $contains($m, "pro") + ? ( + $base := $is1080 ? 0.3 : 0.15; + $perSec := $is1080 ? 0.075 : 0.05; + {"type":"usd","usd": $n * $base + $perSec * $totalDuration} + ) + : ( + $base := $is1080 ? 0.2 : 0.075; + $perSec := $is1080 ? 0.05 : 0.025; + {"type":"usd","usd": $n * $base + $perSec * $totalDuration} + ) + ) + """, + ), + ) + + @classmethod + async def execute( + cls, + model: str, + start_image: Input.Image, + seed: int, + resolution: str, + frames: dict, + ) -> IO.NodeOutput: + validate_image_aspect_ratio(start_image, (1, 4), (4, 1)) + frame_count = int(frames["frames"]) + image_settings: list[FrameSetting] = [] + for i in range(1, frame_count + 1): + validate_image_aspect_ratio(frames[f"end_image{i}"], (1, 4), (4, 1)) + validate_string(frames[f"prompt{i}"], max_length=2000) + start_image_url = await upload_image_to_comfyapi( + cls, + start_image, + mime_type="image/png", + wait_label="Uploading start image", + ) + for i in range(1, frame_count + 1): + image_settings.append( + FrameSetting( + prompt=frames[f"prompt{i}"], + key_image=await upload_image_to_comfyapi( + cls, + frames[f"end_image{i}"], + mime_type="image/png", + wait_label=f"Uploading end image({i})", + ), + duration=frames[f"duration{i}"], + ) + ) + results = await execute_task( + cls, + "/proxy/vidu/multiframe", + TaskMultiFrameCreationRequest( + model=model, + seed=seed, + resolution=resolution, + start_image=start_image_url, + image_settings=image_settings, + ), + max_poll_attempts=480 * frame_count, + ) + return IO.NodeOutput(await download_url_to_video_output(results[0].url)) + + +class Vidu3TextToVideoNode(IO.ComfyNode): + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="Vidu3TextToVideoNode", + display_name="Vidu Q3 Text-to-Video Generation", + category="api node/video/Vidu", + description="Generate video from a text prompt.", + inputs=[ + IO.DynamicCombo.Input( + "model", + options=[ + IO.DynamicCombo.Option( + "viduq3-pro", + [ + IO.Combo.Input( + "aspect_ratio", + options=["16:9", "9:16", "3:4", "4:3", "1:1"], + tooltip="The aspect ratio of the output video.", + ), + IO.Combo.Input( + "resolution", + options=["720p", "1080p"], + tooltip="Resolution of the output video.", + ), + IO.Int.Input( + "duration", + default=5, + min=1, + max=16, + step=1, + display_mode=IO.NumberDisplay.slider, + tooltip="Duration of the output video in seconds.", + ), + IO.Boolean.Input( + "audio", + default=False, + tooltip="When enabled, outputs video with sound " + "(including dialogue and sound effects).", + ), + ], + ), + ], + tooltip="Model to use for video generation.", + ), + IO.String.Input( + "prompt", + multiline=True, + tooltip="A textual description for video generation, with a maximum length of 2000 characters.", + ), + IO.Int.Input( + "seed", + default=1, + min=0, + max=2147483647, + step=1, + display_mode=IO.NumberDisplay.number, + control_after_generate=True, + ), + ], + outputs=[ + IO.Video.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["model.duration", "model.resolution"]), + expr=""" + ( + $res := $lookup(widgets, "model.resolution"); + $base := $lookup({"720p": 0.075, "1080p": 0.1}, $res); + $perSec := $lookup({"720p": 0.025, "1080p": 0.05}, $res); + {"type":"usd","usd": $base + $perSec * ($lookup(widgets, "model.duration") - 1)} + ) + """, + ), + ) + + @classmethod + async def execute( + cls, + model: dict, + prompt: str, + seed: int, + ) -> IO.NodeOutput: + validate_string(prompt, min_length=1, max_length=2000) + results = await execute_task( + cls, + VIDU_TEXT_TO_VIDEO, + TaskCreationRequest( + model=model["model"], + prompt=prompt, + duration=model["duration"], + seed=seed, + aspect_ratio=model["aspect_ratio"], + resolution=model["resolution"], + audio=model["audio"], + ), + max_poll_attempts=640, + ) + return IO.NodeOutput(await download_url_to_video_output(results[0].url)) + + +class Vidu3ImageToVideoNode(IO.ComfyNode): + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="Vidu3ImageToVideoNode", + display_name="Vidu Q3 Image-to-Video Generation", + category="api node/video/Vidu", + description="Generate a video from an image and an optional prompt.", + inputs=[ + IO.DynamicCombo.Input( + "model", + options=[ + IO.DynamicCombo.Option( + "viduq3-pro", + [ + IO.Combo.Input( + "resolution", + options=["720p", "1080p", "2K"], + tooltip="Resolution of the output video.", + ), + IO.Int.Input( + "duration", + default=5, + min=1, + max=16, + step=1, + display_mode=IO.NumberDisplay.slider, + tooltip="Duration of the output video in seconds.", + ), + IO.Boolean.Input( + "audio", + default=False, + tooltip="When enabled, outputs video with sound " + "(including dialogue and sound effects).", + ), + ], + ), + ], + tooltip="Model to use for video generation.", + ), + IO.Image.Input( + "image", + tooltip="An image to be used as the start frame of the generated video.", + ), + IO.String.Input( + "prompt", + multiline=True, + default="", + tooltip="An optional text prompt for video generation (max 2000 characters).", + ), + IO.Int.Input( + "seed", + default=1, + min=0, + max=2147483647, + step=1, + display_mode=IO.NumberDisplay.number, + control_after_generate=True, + ), + ], + outputs=[ + IO.Video.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["model.duration", "model.resolution"]), + expr=""" + ( + $res := $lookup(widgets, "model.resolution"); + $base := $lookup({"720p": 0.075, "1080p": 0.275, "2k": 0.35}, $res); + $perSec := $lookup({"720p": 0.05, "1080p": 0.075, "2k": 0.075}, $res); + {"type":"usd","usd": $base + $perSec * ($lookup(widgets, "model.duration") - 1)} + ) + """, + ), + ) + + @classmethod + async def execute( + cls, + model: dict, + image: Input.Image, + prompt: str, + seed: int, + ) -> IO.NodeOutput: + validate_image_aspect_ratio(image, (1, 4), (4, 1)) + validate_string(prompt, max_length=2000) + results = await execute_task( + cls, + VIDU_IMAGE_TO_VIDEO, + TaskCreationRequest( + model=model["model"], + prompt=prompt, + duration=model["duration"], + seed=seed, + resolution=model["resolution"], + audio=model["audio"], + images=[await upload_image_to_comfyapi(cls, image)], + ), + max_poll_attempts=720, + ) + return IO.NodeOutput(await download_url_to_video_output(results[0].url)) + + class ViduExtension(ComfyExtension): @override async def get_node_list(self) -> list[type[IO.ComfyNode]]: @@ -952,6 +1493,10 @@ class ViduExtension(ComfyExtension): Vidu2ImageToVideoNode, Vidu2ReferenceVideoNode, Vidu2StartEndToVideoNode, + ViduExtendVideoNode, + ViduMultiFrameVideoNode, + Vidu3TextToVideoNode, + Vidu3ImageToVideoNode, ] From 6ea8c128a3770763f150e97aca4b29bd2aec60ad Mon Sep 17 00:00:00 2001 From: Jedrzej Kosinski Date: Fri, 30 Jan 2026 23:22:05 -0800 Subject: [PATCH 541/647] Assets Part 2 - add more endpoints (#12125) --- app/assets/api/routes.py | 414 +++++++++- app/assets/api/schemas_in.py | 196 ++++- app/assets/api/schemas_out.py | 33 + app/assets/database/queries.py | 719 +++++++++++++++++- app/assets/helpers.py | 97 ++- app/assets/manager.py | 401 +++++++++- app/assets/scanner.py | 36 +- tests-unit/assets_test/conftest.py | 271 +++++++ .../assets_test/test_assets_missing_sync.py | 348 +++++++++ tests-unit/assets_test/test_crud.py | 306 ++++++++ tests-unit/assets_test/test_downloads.py | 166 ++++ tests-unit/assets_test/test_list_filter.py | 342 +++++++++ .../assets_test/test_metadata_filters.py | 395 ++++++++++ .../assets_test/test_prune_orphaned_assets.py | 141 ++++ tests-unit/assets_test/test_tags.py | 225 ++++++ tests-unit/assets_test/test_uploads.py | 281 +++++++ tests-unit/requirements.txt | 1 + 17 files changed, 4347 insertions(+), 25 deletions(-) create mode 100644 tests-unit/assets_test/conftest.py create mode 100644 tests-unit/assets_test/test_assets_missing_sync.py create mode 100644 tests-unit/assets_test/test_crud.py create mode 100644 tests-unit/assets_test/test_downloads.py create mode 100644 tests-unit/assets_test/test_list_filter.py create mode 100644 tests-unit/assets_test/test_metadata_filters.py create mode 100644 tests-unit/assets_test/test_prune_orphaned_assets.py create mode 100644 tests-unit/assets_test/test_tags.py create mode 100644 tests-unit/assets_test/test_uploads.py diff --git a/app/assets/api/routes.py b/app/assets/api/routes.py index 30e87a898..7676e50b4 100644 --- a/app/assets/api/routes.py +++ b/app/assets/api/routes.py @@ -1,5 +1,8 @@ import logging import uuid +import urllib.parse +import os +import contextlib from aiohttp import web from pydantic import ValidationError @@ -8,6 +11,9 @@ import app.assets.manager as manager from app import user_manager from app.assets.api import schemas_in from app.assets.helpers import get_query_dict +from app.assets.scanner import seed_assets + +import folder_paths ROUTES = web.RouteTableDef() USER_MANAGER: user_manager.UserManager | None = None @@ -15,6 +21,9 @@ USER_MANAGER: user_manager.UserManager | None = None # UUID regex (canonical hyphenated form, case-insensitive) UUID_RE = r"[0-9a-fA-F]{8}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{12}" +# Note to any custom node developers reading this code: +# The assets system is not yet fully implemented, do not rely on the code in /app/assets remaining the same. + def register_assets_system(app: web.Application, user_manager_instance: user_manager.UserManager) -> None: global USER_MANAGER USER_MANAGER = user_manager_instance @@ -28,6 +37,18 @@ def _validation_error_response(code: str, ve: ValidationError) -> web.Response: return _error_response(400, code, "Validation failed.", {"errors": ve.json()}) +@ROUTES.head("/api/assets/hash/{hash}") +async def head_asset_by_hash(request: web.Request) -> web.Response: + hash_str = request.match_info.get("hash", "").strip().lower() + if not hash_str or ":" not in hash_str: + return _error_response(400, "INVALID_HASH", "hash must be like 'blake3:'") + algo, digest = hash_str.split(":", 1) + if algo != "blake3" or not digest or any(c for c in digest if c not in "0123456789abcdef"): + return _error_response(400, "INVALID_HASH", "hash must be like 'blake3:'") + exists = manager.asset_exists(asset_hash=hash_str) + return web.Response(status=200 if exists else 404) + + @ROUTES.get("/api/assets") async def list_assets(request: web.Request) -> web.Response: """ @@ -50,7 +71,7 @@ async def list_assets(request: web.Request) -> web.Response: order=q.order, owner_id=USER_MANAGER.get_request_user_id(request), ) - return web.json_response(payload.model_dump(mode="json")) + return web.json_response(payload.model_dump(mode="json", exclude_none=True)) @ROUTES.get(f"/api/assets/{{id:{UUID_RE}}}") @@ -76,6 +97,314 @@ async def get_asset(request: web.Request) -> web.Response: return web.json_response(result.model_dump(mode="json"), status=200) +@ROUTES.get(f"/api/assets/{{id:{UUID_RE}}}/content") +async def download_asset_content(request: web.Request) -> web.Response: + # question: do we need disposition? could we just stick with one of these? + disposition = request.query.get("disposition", "attachment").lower().strip() + if disposition not in {"inline", "attachment"}: + disposition = "attachment" + + try: + abs_path, content_type, filename = manager.resolve_asset_content_for_download( + asset_info_id=str(uuid.UUID(request.match_info["id"])), + owner_id=USER_MANAGER.get_request_user_id(request), + ) + except ValueError as ve: + return _error_response(404, "ASSET_NOT_FOUND", str(ve)) + except NotImplementedError as nie: + return _error_response(501, "BACKEND_UNSUPPORTED", str(nie)) + except FileNotFoundError: + return _error_response(404, "FILE_NOT_FOUND", "Underlying file not found on disk.") + + quoted = (filename or "").replace("\r", "").replace("\n", "").replace('"', "'") + cd = f'{disposition}; filename="{quoted}"; filename*=UTF-8\'\'{urllib.parse.quote(filename)}' + + file_size = os.path.getsize(abs_path) + logging.info( + "download_asset_content: path=%s, size=%d bytes (%.2f MB), content_type=%s, filename=%s", + abs_path, + file_size, + file_size / (1024 * 1024), + content_type, + filename, + ) + + async def file_sender(): + chunk_size = 64 * 1024 + with open(abs_path, "rb") as f: + while True: + chunk = f.read(chunk_size) + if not chunk: + break + yield chunk + + return web.Response( + body=file_sender(), + content_type=content_type, + headers={ + "Content-Disposition": cd, + "Content-Length": str(file_size), + }, + ) + + +@ROUTES.post("/api/assets/from-hash") +async def create_asset_from_hash(request: web.Request) -> web.Response: + try: + payload = await request.json() + body = schemas_in.CreateFromHashBody.model_validate(payload) + except ValidationError as ve: + return _validation_error_response("INVALID_BODY", ve) + except Exception: + return _error_response(400, "INVALID_JSON", "Request body must be valid JSON.") + + result = manager.create_asset_from_hash( + hash_str=body.hash, + name=body.name, + tags=body.tags, + user_metadata=body.user_metadata, + owner_id=USER_MANAGER.get_request_user_id(request), + ) + if result is None: + return _error_response(404, "ASSET_NOT_FOUND", f"Asset content {body.hash} does not exist") + return web.json_response(result.model_dump(mode="json"), status=201) + + +@ROUTES.post("/api/assets") +async def upload_asset(request: web.Request) -> web.Response: + """Multipart/form-data endpoint for Asset uploads.""" + if not (request.content_type or "").lower().startswith("multipart/"): + return _error_response(415, "UNSUPPORTED_MEDIA_TYPE", "Use multipart/form-data for uploads.") + + reader = await request.multipart() + + file_present = False + file_client_name: str | None = None + tags_raw: list[str] = [] + provided_name: str | None = None + user_metadata_raw: str | None = None + provided_hash: str | None = None + provided_hash_exists: bool | None = None + + file_written = 0 + tmp_path: str | None = None + while True: + field = await reader.next() + if field is None: + break + + fname = getattr(field, "name", "") or "" + + if fname == "hash": + try: + s = ((await field.text()) or "").strip().lower() + except Exception: + return _error_response(400, "INVALID_HASH", "hash must be like 'blake3:'") + + if s: + if ":" not in s: + return _error_response(400, "INVALID_HASH", "hash must be like 'blake3:'") + algo, digest = s.split(":", 1) + if algo != "blake3" or not digest or any(c for c in digest if c not in "0123456789abcdef"): + return _error_response(400, "INVALID_HASH", "hash must be like 'blake3:'") + provided_hash = f"{algo}:{digest}" + try: + provided_hash_exists = manager.asset_exists(asset_hash=provided_hash) + except Exception: + provided_hash_exists = None # do not fail the whole request here + + elif fname == "file": + file_present = True + file_client_name = (field.filename or "").strip() + + if provided_hash and provided_hash_exists is True: + # If client supplied a hash that we know exists, drain but do not write to disk + try: + while True: + chunk = await field.read_chunk(8 * 1024 * 1024) + if not chunk: + break + file_written += len(chunk) + except Exception: + return _error_response(500, "UPLOAD_IO_ERROR", "Failed to receive uploaded file.") + continue # Do not create temp file; we will create AssetInfo from the existing content + + # Otherwise, store to temp for hashing/ingest + uploads_root = os.path.join(folder_paths.get_temp_directory(), "uploads") + unique_dir = os.path.join(uploads_root, uuid.uuid4().hex) + os.makedirs(unique_dir, exist_ok=True) + tmp_path = os.path.join(unique_dir, ".upload.part") + + try: + with open(tmp_path, "wb") as f: + while True: + chunk = await field.read_chunk(8 * 1024 * 1024) + if not chunk: + break + f.write(chunk) + file_written += len(chunk) + except Exception: + try: + if os.path.exists(tmp_path or ""): + os.remove(tmp_path) + finally: + return _error_response(500, "UPLOAD_IO_ERROR", "Failed to receive and store uploaded file.") + elif fname == "tags": + tags_raw.append((await field.text()) or "") + elif fname == "name": + provided_name = (await field.text()) or None + elif fname == "user_metadata": + user_metadata_raw = (await field.text()) or None + + # If client did not send file, and we are not doing a from-hash fast path -> error + if not file_present and not (provided_hash and provided_hash_exists): + return _error_response(400, "MISSING_FILE", "Form must include a 'file' part or a known 'hash'.") + + if file_present and file_written == 0 and not (provided_hash and provided_hash_exists): + # Empty upload is only acceptable if we are fast-pathing from existing hash + try: + if tmp_path and os.path.exists(tmp_path): + os.remove(tmp_path) + finally: + return _error_response(400, "EMPTY_UPLOAD", "Uploaded file is empty.") + + try: + spec = schemas_in.UploadAssetSpec.model_validate({ + "tags": tags_raw, + "name": provided_name, + "user_metadata": user_metadata_raw, + "hash": provided_hash, + }) + except ValidationError as ve: + try: + if tmp_path and os.path.exists(tmp_path): + os.remove(tmp_path) + finally: + return _validation_error_response("INVALID_BODY", ve) + + # Validate models category against configured folders (consistent with previous behavior) + if spec.tags and spec.tags[0] == "models": + if len(spec.tags) < 2 or spec.tags[1] not in folder_paths.folder_names_and_paths: + if tmp_path and os.path.exists(tmp_path): + os.remove(tmp_path) + return _error_response( + 400, "INVALID_BODY", f"unknown models category '{spec.tags[1] if len(spec.tags) >= 2 else ''}'" + ) + + owner_id = USER_MANAGER.get_request_user_id(request) + + # Fast path: if a valid provided hash exists, create AssetInfo without writing anything + if spec.hash and provided_hash_exists is True: + try: + result = manager.create_asset_from_hash( + hash_str=spec.hash, + name=spec.name or (spec.hash.split(":", 1)[1]), + tags=spec.tags, + user_metadata=spec.user_metadata or {}, + owner_id=owner_id, + ) + except Exception: + logging.exception("create_asset_from_hash failed for hash=%s, owner_id=%s", spec.hash, owner_id) + return _error_response(500, "INTERNAL", "Unexpected server error.") + + if result is None: + return _error_response(404, "ASSET_NOT_FOUND", f"Asset content {spec.hash} does not exist") + + # Drain temp if we accidentally saved (e.g., hash field came after file) + if tmp_path and os.path.exists(tmp_path): + with contextlib.suppress(Exception): + os.remove(tmp_path) + + status = 200 if (not result.created_new) else 201 + return web.json_response(result.model_dump(mode="json"), status=status) + + # Otherwise, we must have a temp file path to ingest + if not tmp_path or not os.path.exists(tmp_path): + # The only case we reach here without a temp file is: client sent a hash that does not exist and no file + return _error_response(404, "ASSET_NOT_FOUND", "Provided hash not found and no file uploaded.") + + try: + created = manager.upload_asset_from_temp_path( + spec, + temp_path=tmp_path, + client_filename=file_client_name, + owner_id=owner_id, + expected_asset_hash=spec.hash, + ) + status = 201 if created.created_new else 200 + return web.json_response(created.model_dump(mode="json"), status=status) + except ValueError as e: + if tmp_path and os.path.exists(tmp_path): + os.remove(tmp_path) + msg = str(e) + if "HASH_MISMATCH" in msg or msg.strip().upper() == "HASH_MISMATCH": + return _error_response( + 400, + "HASH_MISMATCH", + "Uploaded file hash does not match provided hash.", + ) + return _error_response(400, "BAD_REQUEST", "Invalid inputs.") + except Exception: + if tmp_path and os.path.exists(tmp_path): + os.remove(tmp_path) + logging.exception("upload_asset_from_temp_path failed for tmp_path=%s, owner_id=%s", tmp_path, owner_id) + return _error_response(500, "INTERNAL", "Unexpected server error.") + + +@ROUTES.put(f"/api/assets/{{id:{UUID_RE}}}") +async def update_asset(request: web.Request) -> web.Response: + asset_info_id = str(uuid.UUID(request.match_info["id"])) + try: + body = schemas_in.UpdateAssetBody.model_validate(await request.json()) + except ValidationError as ve: + return _validation_error_response("INVALID_BODY", ve) + except Exception: + return _error_response(400, "INVALID_JSON", "Request body must be valid JSON.") + + try: + result = manager.update_asset( + asset_info_id=asset_info_id, + name=body.name, + user_metadata=body.user_metadata, + owner_id=USER_MANAGER.get_request_user_id(request), + ) + except (ValueError, PermissionError) as ve: + return _error_response(404, "ASSET_NOT_FOUND", str(ve), {"id": asset_info_id}) + except Exception: + logging.exception( + "update_asset failed for asset_info_id=%s, owner_id=%s", + asset_info_id, + USER_MANAGER.get_request_user_id(request), + ) + return _error_response(500, "INTERNAL", "Unexpected server error.") + return web.json_response(result.model_dump(mode="json"), status=200) + + +@ROUTES.delete(f"/api/assets/{{id:{UUID_RE}}}") +async def delete_asset(request: web.Request) -> web.Response: + asset_info_id = str(uuid.UUID(request.match_info["id"])) + delete_content = request.query.get("delete_content") + delete_content = True if delete_content is None else delete_content.lower() not in {"0", "false", "no"} + + try: + deleted = manager.delete_asset_reference( + asset_info_id=asset_info_id, + owner_id=USER_MANAGER.get_request_user_id(request), + delete_content_if_orphan=delete_content, + ) + except Exception: + logging.exception( + "delete_asset_reference failed for asset_info_id=%s, owner_id=%s", + asset_info_id, + USER_MANAGER.get_request_user_id(request), + ) + return _error_response(500, "INTERNAL", "Unexpected server error.") + + if not deleted: + return _error_response(404, "ASSET_NOT_FOUND", f"AssetInfo {asset_info_id} not found.") + return web.Response(status=204) + + @ROUTES.get("/api/tags") async def get_tags(request: web.Request) -> web.Response: """ @@ -100,3 +429,86 @@ async def get_tags(request: web.Request) -> web.Response: owner_id=USER_MANAGER.get_request_user_id(request), ) return web.json_response(result.model_dump(mode="json")) + + +@ROUTES.post(f"/api/assets/{{id:{UUID_RE}}}/tags") +async def add_asset_tags(request: web.Request) -> web.Response: + asset_info_id = str(uuid.UUID(request.match_info["id"])) + try: + payload = await request.json() + data = schemas_in.TagsAdd.model_validate(payload) + except ValidationError as ve: + return _error_response(400, "INVALID_BODY", "Invalid JSON body for tags add.", {"errors": ve.errors()}) + except Exception: + return _error_response(400, "INVALID_JSON", "Request body must be valid JSON.") + + try: + result = manager.add_tags_to_asset( + asset_info_id=asset_info_id, + tags=data.tags, + origin="manual", + owner_id=USER_MANAGER.get_request_user_id(request), + ) + except (ValueError, PermissionError) as ve: + return _error_response(404, "ASSET_NOT_FOUND", str(ve), {"id": asset_info_id}) + except Exception: + logging.exception( + "add_tags_to_asset failed for asset_info_id=%s, owner_id=%s", + asset_info_id, + USER_MANAGER.get_request_user_id(request), + ) + return _error_response(500, "INTERNAL", "Unexpected server error.") + + return web.json_response(result.model_dump(mode="json"), status=200) + + +@ROUTES.delete(f"/api/assets/{{id:{UUID_RE}}}/tags") +async def delete_asset_tags(request: web.Request) -> web.Response: + asset_info_id = str(uuid.UUID(request.match_info["id"])) + try: + payload = await request.json() + data = schemas_in.TagsRemove.model_validate(payload) + except ValidationError as ve: + return _error_response(400, "INVALID_BODY", "Invalid JSON body for tags remove.", {"errors": ve.errors()}) + except Exception: + return _error_response(400, "INVALID_JSON", "Request body must be valid JSON.") + + try: + result = manager.remove_tags_from_asset( + asset_info_id=asset_info_id, + tags=data.tags, + owner_id=USER_MANAGER.get_request_user_id(request), + ) + except ValueError as ve: + return _error_response(404, "ASSET_NOT_FOUND", str(ve), {"id": asset_info_id}) + except Exception: + logging.exception( + "remove_tags_from_asset failed for asset_info_id=%s, owner_id=%s", + asset_info_id, + USER_MANAGER.get_request_user_id(request), + ) + return _error_response(500, "INTERNAL", "Unexpected server error.") + + return web.json_response(result.model_dump(mode="json"), status=200) + + +@ROUTES.post("/api/assets/seed") +async def seed_assets_endpoint(request: web.Request) -> web.Response: + """Trigger asset seeding for specified roots (models, input, output).""" + try: + payload = await request.json() + roots = payload.get("roots", ["models", "input", "output"]) + except Exception: + roots = ["models", "input", "output"] + + valid_roots = [r for r in roots if r in ("models", "input", "output")] + if not valid_roots: + return _error_response(400, "INVALID_BODY", "No valid roots specified") + + try: + seed_assets(tuple(valid_roots)) + except Exception: + logging.exception("seed_assets failed for roots=%s", valid_roots) + return _error_response(500, "INTERNAL", "Seed operation failed") + + return web.json_response({"seeded": valid_roots}, status=200) diff --git a/app/assets/api/schemas_in.py b/app/assets/api/schemas_in.py index 200b41aef..6707ffb0c 100644 --- a/app/assets/api/schemas_in.py +++ b/app/assets/api/schemas_in.py @@ -1,5 +1,4 @@ import json -import uuid from typing import Any, Literal from pydantic import ( @@ -8,9 +7,9 @@ from pydantic import ( Field, conint, field_validator, + model_validator, ) - class ListAssetsQuery(BaseModel): include_tags: list[str] = Field(default_factory=list) exclude_tags: list[str] = Field(default_factory=list) @@ -57,6 +56,57 @@ class ListAssetsQuery(BaseModel): return None +class UpdateAssetBody(BaseModel): + name: str | None = None + user_metadata: dict[str, Any] | None = None + + @model_validator(mode="after") + def _at_least_one(self): + if self.name is None and self.user_metadata is None: + raise ValueError("Provide at least one of: name, user_metadata.") + return self + + +class CreateFromHashBody(BaseModel): + model_config = ConfigDict(extra="ignore", str_strip_whitespace=True) + + hash: str + name: str + tags: list[str] = Field(default_factory=list) + user_metadata: dict[str, Any] = Field(default_factory=dict) + + @field_validator("hash") + @classmethod + def _require_blake3(cls, v): + s = (v or "").strip().lower() + if ":" not in s: + raise ValueError("hash must be 'blake3:'") + algo, digest = s.split(":", 1) + if algo != "blake3": + raise ValueError("only canonical 'blake3:' is accepted here") + if not digest or any(c for c in digest if c not in "0123456789abcdef"): + raise ValueError("hash digest must be lowercase hex") + return s + + @field_validator("tags", mode="before") + @classmethod + def _tags_norm(cls, v): + if v is None: + return [] + if isinstance(v, list): + out = [str(t).strip().lower() for t in v if str(t).strip()] + seen = set() + dedup = [] + for t in out: + if t not in seen: + seen.add(t) + dedup.append(t) + return dedup + if isinstance(v, str): + return [t.strip().lower() for t in v.split(",") if t.strip()] + return [] + + class TagsListQuery(BaseModel): model_config = ConfigDict(extra="ignore", str_strip_whitespace=True) @@ -75,20 +125,140 @@ class TagsListQuery(BaseModel): return v.lower() or None -class SetPreviewBody(BaseModel): - """Set or clear the preview for an AssetInfo. Provide an Asset.id or null.""" - preview_id: str | None = None +class TagsAdd(BaseModel): + model_config = ConfigDict(extra="ignore") + tags: list[str] = Field(..., min_length=1) - @field_validator("preview_id", mode="before") + @field_validator("tags") @classmethod - def _norm_uuid(cls, v): + def normalize_tags(cls, v: list[str]) -> list[str]: + out = [] + for t in v: + if not isinstance(t, str): + raise TypeError("tags must be strings") + tnorm = t.strip().lower() + if tnorm: + out.append(tnorm) + seen = set() + deduplicated = [] + for x in out: + if x not in seen: + seen.add(x) + deduplicated.append(x) + return deduplicated + + +class TagsRemove(TagsAdd): + pass + + +class UploadAssetSpec(BaseModel): + """Upload Asset operation. + - tags: ordered; first is root ('models'|'input'|'output'); + if root == 'models', second must be a valid category from folder_paths.folder_names_and_paths + - name: display name + - user_metadata: arbitrary JSON object (optional) + - hash: optional canonical 'blake3:' provided by the client for validation / fast-path + + Files created via this endpoint are stored on disk using the **content hash** as the filename stem + and the original extension is preserved when available. + """ + model_config = ConfigDict(extra="ignore", str_strip_whitespace=True) + + tags: list[str] = Field(..., min_length=1) + name: str | None = Field(default=None, max_length=512, description="Display Name") + user_metadata: dict[str, Any] = Field(default_factory=dict) + hash: str | None = Field(default=None) + + @field_validator("hash", mode="before") + @classmethod + def _parse_hash(cls, v): if v is None: return None - s = str(v).strip() + s = str(v).strip().lower() if not s: return None - try: - uuid.UUID(s) - except Exception: - raise ValueError("preview_id must be a UUID") - return s + if ":" not in s: + raise ValueError("hash must be 'blake3:'") + algo, digest = s.split(":", 1) + if algo != "blake3": + raise ValueError("only canonical 'blake3:' is accepted here") + if not digest or any(c for c in digest if c not in "0123456789abcdef"): + raise ValueError("hash digest must be lowercase hex") + return f"{algo}:{digest}" + + @field_validator("tags", mode="before") + @classmethod + def _parse_tags(cls, v): + """ + Accepts a list of strings (possibly multiple form fields), + where each string can be: + - JSON array (e.g., '["models","loras","foo"]') + - comma-separated ('models, loras, foo') + - single token ('models') + Returns a normalized, deduplicated, ordered list. + """ + items: list[str] = [] + if v is None: + return [] + if isinstance(v, str): + v = [v] + + if isinstance(v, list): + for item in v: + if item is None: + continue + s = str(item).strip() + if not s: + continue + if s.startswith("["): + try: + arr = json.loads(s) + if isinstance(arr, list): + items.extend(str(x) for x in arr) + continue + except Exception: + pass # fallback to CSV parse below + items.extend([p for p in s.split(",") if p.strip()]) + else: + return [] + + # normalize + dedupe + norm = [] + seen = set() + for t in items: + tnorm = str(t).strip().lower() + if tnorm and tnorm not in seen: + seen.add(tnorm) + norm.append(tnorm) + return norm + + @field_validator("user_metadata", mode="before") + @classmethod + def _parse_metadata_json(cls, v): + if v is None or isinstance(v, dict): + return v or {} + if isinstance(v, str): + s = v.strip() + if not s: + return {} + try: + parsed = json.loads(s) + except Exception as e: + raise ValueError(f"user_metadata must be JSON: {e}") from e + if not isinstance(parsed, dict): + raise ValueError("user_metadata must be a JSON object") + return parsed + return {} + + @model_validator(mode="after") + def _validate_order(self): + if not self.tags: + raise ValueError("tags must be provided and non-empty") + root = self.tags[0] + if root not in {"models", "input", "output"}: + raise ValueError("first tag must be one of: models, input, output") + if root == "models": + if len(self.tags) < 2: + raise ValueError("models uploads require a category tag as the second tag") + return self diff --git a/app/assets/api/schemas_out.py b/app/assets/api/schemas_out.py index 9f8184f20..b6fb3da0c 100644 --- a/app/assets/api/schemas_out.py +++ b/app/assets/api/schemas_out.py @@ -29,6 +29,21 @@ class AssetsList(BaseModel): has_more: bool +class AssetUpdated(BaseModel): + id: str + name: str + asset_hash: str | None = None + tags: list[str] = Field(default_factory=list) + user_metadata: dict[str, Any] = Field(default_factory=dict) + updated_at: datetime | None = None + + model_config = ConfigDict(from_attributes=True) + + @field_serializer("updated_at") + def _ser_updated(self, v: datetime | None, _info): + return v.isoformat() if v else None + + class AssetDetail(BaseModel): id: str name: str @@ -48,6 +63,10 @@ class AssetDetail(BaseModel): return v.isoformat() if v else None +class AssetCreated(AssetDetail): + created_new: bool + + class TagUsage(BaseModel): name: str count: int @@ -58,3 +77,17 @@ class TagsList(BaseModel): tags: list[TagUsage] = Field(default_factory=list) total: int has_more: bool + + +class TagsAdd(BaseModel): + model_config = ConfigDict(str_strip_whitespace=True) + added: list[str] = Field(default_factory=list) + already_present: list[str] = Field(default_factory=list) + total_tags: list[str] = Field(default_factory=list) + + +class TagsRemove(BaseModel): + model_config = ConfigDict(str_strip_whitespace=True) + removed: list[str] = Field(default_factory=list) + not_present: list[str] = Field(default_factory=list) + total_tags: list[str] = Field(default_factory=list) diff --git a/app/assets/database/queries.py b/app/assets/database/queries.py index 0824c0c2f..d6b33ec7b 100644 --- a/app/assets/database/queries.py +++ b/app/assets/database/queries.py @@ -1,9 +1,17 @@ +import os +import logging import sqlalchemy as sa from collections import defaultdict -from sqlalchemy import select, exists, func +from datetime import datetime +from typing import Iterable, Any +from sqlalchemy import select, delete, exists, func +from sqlalchemy.dialects import sqlite +from sqlalchemy.exc import IntegrityError from sqlalchemy.orm import Session, contains_eager, noload -from app.assets.database.models import Asset, AssetInfo, AssetInfoMeta, AssetInfoTag, Tag -from app.assets.helpers import escape_like_prefix, normalize_tags +from app.assets.database.models import Asset, AssetInfo, AssetCacheState, AssetInfoMeta, AssetInfoTag, Tag +from app.assets.helpers import ( + compute_relative_filename, escape_like_prefix, normalize_tags, project_kv, utcnow +) from typing import Sequence @@ -15,6 +23,22 @@ def visible_owner_clause(owner_id: str) -> sa.sql.ClauseElement: return AssetInfo.owner_id.in_(["", owner_id]) +def pick_best_live_path(states: Sequence[AssetCacheState]) -> str: + """ + Return the best on-disk path among cache states: + 1) Prefer a path that exists with needs_verify == False (already verified). + 2) Otherwise, pick the first path that exists. + 3) Otherwise return empty string. + """ + alive = [s for s in states if getattr(s, "file_path", None) and os.path.isfile(s.file_path)] + if not alive: + return "" + for s in alive: + if not getattr(s, "needs_verify", False): + return s.file_path + return alive[0].file_path + + def apply_tag_filters( stmt: sa.sql.Select, include_tags: Sequence[str] | None = None, @@ -42,6 +66,7 @@ def apply_tag_filters( ) return stmt + def apply_metadata_filter( stmt: sa.sql.Select, metadata_filter: dict | None = None, @@ -94,7 +119,11 @@ def apply_metadata_filter( return stmt -def asset_exists_by_hash(session: Session, asset_hash: str) -> bool: +def asset_exists_by_hash( + session: Session, + *, + asset_hash: str, +) -> bool: """ Check if an asset with a given hash exists in database. """ @@ -105,9 +134,39 @@ def asset_exists_by_hash(session: Session, asset_hash: str) -> bool: ).first() return row is not None -def get_asset_info_by_id(session: Session, asset_info_id: str) -> AssetInfo | None: + +def asset_info_exists_for_asset_id( + session: Session, + *, + asset_id: str, +) -> bool: + q = ( + select(sa.literal(True)) + .select_from(AssetInfo) + .where(AssetInfo.asset_id == asset_id) + .limit(1) + ) + return (session.execute(q)).first() is not None + + +def get_asset_by_hash( + session: Session, + *, + asset_hash: str, +) -> Asset | None: + return ( + session.execute(select(Asset).where(Asset.hash == asset_hash).limit(1)) + ).scalars().first() + + +def get_asset_info_by_id( + session: Session, + *, + asset_info_id: str, +) -> AssetInfo | None: return session.get(AssetInfo, asset_info_id) + def list_asset_infos_page( session: Session, owner_id: str = "", @@ -171,12 +230,14 @@ def list_asset_infos_page( select(AssetInfoTag.asset_info_id, Tag.name) .join(Tag, Tag.name == AssetInfoTag.tag_name) .where(AssetInfoTag.asset_info_id.in_(id_list)) + .order_by(AssetInfoTag.added_at) ) for aid, tag_name in rows.all(): tag_map[aid].append(tag_name) return infos, tag_map, total + def fetch_asset_info_asset_and_tags( session: Session, asset_info_id: str, @@ -208,6 +269,494 @@ def fetch_asset_info_asset_and_tags( tags.append(tag_name) return first_info, first_asset, tags + +def fetch_asset_info_and_asset( + session: Session, + *, + asset_info_id: str, + owner_id: str = "", +) -> tuple[AssetInfo, Asset] | None: + stmt = ( + select(AssetInfo, Asset) + .join(Asset, Asset.id == AssetInfo.asset_id) + .where( + AssetInfo.id == asset_info_id, + visible_owner_clause(owner_id), + ) + .limit(1) + .options(noload(AssetInfo.tags)) + ) + row = session.execute(stmt) + pair = row.first() + if not pair: + return None + return pair[0], pair[1] + +def list_cache_states_by_asset_id( + session: Session, *, asset_id: str +) -> Sequence[AssetCacheState]: + return ( + session.execute( + select(AssetCacheState) + .where(AssetCacheState.asset_id == asset_id) + .order_by(AssetCacheState.id.asc()) + ) + ).scalars().all() + + +def touch_asset_info_by_id( + session: Session, + *, + asset_info_id: str, + ts: datetime | None = None, + only_if_newer: bool = True, +) -> None: + ts = ts or utcnow() + stmt = sa.update(AssetInfo).where(AssetInfo.id == asset_info_id) + if only_if_newer: + stmt = stmt.where( + sa.or_(AssetInfo.last_access_time.is_(None), AssetInfo.last_access_time < ts) + ) + session.execute(stmt.values(last_access_time=ts)) + + +def create_asset_info_for_existing_asset( + session: Session, + *, + asset_hash: str, + name: str, + user_metadata: dict | None = None, + tags: Sequence[str] | None = None, + tag_origin: str = "manual", + owner_id: str = "", +) -> AssetInfo: + """Create or return an existing AssetInfo for an Asset identified by asset_hash.""" + now = utcnow() + asset = get_asset_by_hash(session, asset_hash=asset_hash) + if not asset: + raise ValueError(f"Unknown asset hash {asset_hash}") + + info = AssetInfo( + owner_id=owner_id, + name=name, + asset_id=asset.id, + preview_id=None, + created_at=now, + updated_at=now, + last_access_time=now, + ) + try: + with session.begin_nested(): + session.add(info) + session.flush() + except IntegrityError: + existing = ( + session.execute( + select(AssetInfo) + .options(noload(AssetInfo.tags)) + .where( + AssetInfo.asset_id == asset.id, + AssetInfo.name == name, + AssetInfo.owner_id == owner_id, + ) + .limit(1) + ) + ).unique().scalars().first() + if not existing: + raise RuntimeError("AssetInfo upsert failed to find existing row after conflict.") + return existing + + # metadata["filename"] hack + new_meta = dict(user_metadata or {}) + computed_filename = None + try: + p = pick_best_live_path(list_cache_states_by_asset_id(session, asset_id=asset.id)) + if p: + computed_filename = compute_relative_filename(p) + except Exception: + computed_filename = None + if computed_filename: + new_meta["filename"] = computed_filename + if new_meta: + replace_asset_info_metadata_projection( + session, + asset_info_id=info.id, + user_metadata=new_meta, + ) + + if tags is not None: + set_asset_info_tags( + session, + asset_info_id=info.id, + tags=tags, + origin=tag_origin, + ) + return info + + +def set_asset_info_tags( + session: Session, + *, + asset_info_id: str, + tags: Sequence[str], + origin: str = "manual", +) -> dict: + desired = normalize_tags(tags) + + current = set( + tag_name for (tag_name,) in ( + session.execute(select(AssetInfoTag.tag_name).where(AssetInfoTag.asset_info_id == asset_info_id)) + ).all() + ) + + to_add = [t for t in desired if t not in current] + to_remove = [t for t in current if t not in desired] + + if to_add: + ensure_tags_exist(session, to_add, tag_type="user") + session.add_all([ + AssetInfoTag(asset_info_id=asset_info_id, tag_name=t, origin=origin, added_at=utcnow()) + for t in to_add + ]) + session.flush() + + if to_remove: + session.execute( + delete(AssetInfoTag) + .where(AssetInfoTag.asset_info_id == asset_info_id, AssetInfoTag.tag_name.in_(to_remove)) + ) + session.flush() + + return {"added": to_add, "removed": to_remove, "total": desired} + + +def replace_asset_info_metadata_projection( + session: Session, + *, + asset_info_id: str, + user_metadata: dict | None = None, +) -> None: + info = session.get(AssetInfo, asset_info_id) + if not info: + raise ValueError(f"AssetInfo {asset_info_id} not found") + + info.user_metadata = user_metadata or {} + info.updated_at = utcnow() + session.flush() + + session.execute(delete(AssetInfoMeta).where(AssetInfoMeta.asset_info_id == asset_info_id)) + session.flush() + + if not user_metadata: + return + + rows: list[AssetInfoMeta] = [] + for k, v in user_metadata.items(): + for r in project_kv(k, v): + rows.append( + AssetInfoMeta( + asset_info_id=asset_info_id, + key=r["key"], + ordinal=int(r["ordinal"]), + val_str=r.get("val_str"), + val_num=r.get("val_num"), + val_bool=r.get("val_bool"), + val_json=r.get("val_json"), + ) + ) + if rows: + session.add_all(rows) + session.flush() + + +def ingest_fs_asset( + session: Session, + *, + asset_hash: str, + abs_path: str, + size_bytes: int, + mtime_ns: int, + mime_type: str | None = None, + info_name: str | None = None, + owner_id: str = "", + preview_id: str | None = None, + user_metadata: dict | None = None, + tags: Sequence[str] = (), + tag_origin: str = "manual", + require_existing_tags: bool = False, +) -> dict: + """ + Idempotently upsert: + - Asset by content hash (create if missing) + - AssetCacheState(file_path) pointing to asset_id + - Optionally AssetInfo + tag links and metadata projection + Returns flags and ids. + """ + locator = os.path.abspath(abs_path) + now = utcnow() + + if preview_id: + if not session.get(Asset, preview_id): + preview_id = None + + out: dict[str, Any] = { + "asset_created": False, + "asset_updated": False, + "state_created": False, + "state_updated": False, + "asset_info_id": None, + } + + # 1) Asset by hash + asset = ( + session.execute(select(Asset).where(Asset.hash == asset_hash).limit(1)) + ).scalars().first() + if not asset: + vals = { + "hash": asset_hash, + "size_bytes": int(size_bytes), + "mime_type": mime_type, + "created_at": now, + } + res = session.execute( + sqlite.insert(Asset) + .values(**vals) + .on_conflict_do_nothing(index_elements=[Asset.hash]) + ) + if int(res.rowcount or 0) > 0: + out["asset_created"] = True + asset = ( + session.execute( + select(Asset).where(Asset.hash == asset_hash).limit(1) + ) + ).scalars().first() + if not asset: + raise RuntimeError("Asset row not found after upsert.") + else: + changed = False + if asset.size_bytes != int(size_bytes) and int(size_bytes) > 0: + asset.size_bytes = int(size_bytes) + changed = True + if mime_type and asset.mime_type != mime_type: + asset.mime_type = mime_type + changed = True + if changed: + out["asset_updated"] = True + + # 2) AssetCacheState upsert by file_path (unique) + vals = { + "asset_id": asset.id, + "file_path": locator, + "mtime_ns": int(mtime_ns), + } + ins = ( + sqlite.insert(AssetCacheState) + .values(**vals) + .on_conflict_do_nothing(index_elements=[AssetCacheState.file_path]) + ) + + res = session.execute(ins) + if int(res.rowcount or 0) > 0: + out["state_created"] = True + else: + upd = ( + sa.update(AssetCacheState) + .where(AssetCacheState.file_path == locator) + .where( + sa.or_( + AssetCacheState.asset_id != asset.id, + AssetCacheState.mtime_ns.is_(None), + AssetCacheState.mtime_ns != int(mtime_ns), + ) + ) + .values(asset_id=asset.id, mtime_ns=int(mtime_ns)) + ) + res2 = session.execute(upd) + if int(res2.rowcount or 0) > 0: + out["state_updated"] = True + + # 3) Optional AssetInfo + tags + metadata + if info_name: + try: + with session.begin_nested(): + info = AssetInfo( + owner_id=owner_id, + name=info_name, + asset_id=asset.id, + preview_id=preview_id, + created_at=now, + updated_at=now, + last_access_time=now, + ) + session.add(info) + session.flush() + out["asset_info_id"] = info.id + except IntegrityError: + pass + + existing_info = ( + session.execute( + select(AssetInfo) + .where( + AssetInfo.asset_id == asset.id, + AssetInfo.name == info_name, + (AssetInfo.owner_id == owner_id), + ) + .limit(1) + ) + ).unique().scalar_one_or_none() + if not existing_info: + raise RuntimeError("Failed to update or insert AssetInfo.") + + if preview_id and existing_info.preview_id != preview_id: + existing_info.preview_id = preview_id + + existing_info.updated_at = now + if existing_info.last_access_time < now: + existing_info.last_access_time = now + session.flush() + out["asset_info_id"] = existing_info.id + + norm = [t.strip().lower() for t in (tags or []) if (t or "").strip()] + if norm and out["asset_info_id"] is not None: + if not require_existing_tags: + ensure_tags_exist(session, norm, tag_type="user") + + existing_tag_names = set( + name for (name,) in (session.execute(select(Tag.name).where(Tag.name.in_(norm)))).all() + ) + missing = [t for t in norm if t not in existing_tag_names] + if missing and require_existing_tags: + raise ValueError(f"Unknown tags: {missing}") + + existing_links = set( + tag_name + for (tag_name,) in ( + session.execute( + select(AssetInfoTag.tag_name).where(AssetInfoTag.asset_info_id == out["asset_info_id"]) + ) + ).all() + ) + to_add = [t for t in norm if t in existing_tag_names and t not in existing_links] + if to_add: + session.add_all( + [ + AssetInfoTag( + asset_info_id=out["asset_info_id"], + tag_name=t, + origin=tag_origin, + added_at=now, + ) + for t in to_add + ] + ) + session.flush() + + # metadata["filename"] hack + if out["asset_info_id"] is not None: + primary_path = pick_best_live_path(list_cache_states_by_asset_id(session, asset_id=asset.id)) + computed_filename = compute_relative_filename(primary_path) if primary_path else None + + current_meta = existing_info.user_metadata or {} + new_meta = dict(current_meta) + if user_metadata is not None: + for k, v in user_metadata.items(): + new_meta[k] = v + if computed_filename: + new_meta["filename"] = computed_filename + + if new_meta != current_meta: + replace_asset_info_metadata_projection( + session, + asset_info_id=out["asset_info_id"], + user_metadata=new_meta, + ) + + try: + remove_missing_tag_for_asset_id(session, asset_id=asset.id) + except Exception: + logging.exception("Failed to clear 'missing' tag for asset %s", asset.id) + return out + + +def update_asset_info_full( + session: Session, + *, + asset_info_id: str, + name: str | None = None, + tags: Sequence[str] | None = None, + user_metadata: dict | None = None, + tag_origin: str = "manual", + asset_info_row: Any = None, +) -> AssetInfo: + if not asset_info_row: + info = session.get(AssetInfo, asset_info_id) + if not info: + raise ValueError(f"AssetInfo {asset_info_id} not found") + else: + info = asset_info_row + + touched = False + if name is not None and name != info.name: + info.name = name + touched = True + + computed_filename = None + try: + p = pick_best_live_path(list_cache_states_by_asset_id(session, asset_id=info.asset_id)) + if p: + computed_filename = compute_relative_filename(p) + except Exception: + computed_filename = None + + if user_metadata is not None: + new_meta = dict(user_metadata) + if computed_filename: + new_meta["filename"] = computed_filename + replace_asset_info_metadata_projection( + session, asset_info_id=asset_info_id, user_metadata=new_meta + ) + touched = True + else: + if computed_filename: + current_meta = info.user_metadata or {} + if current_meta.get("filename") != computed_filename: + new_meta = dict(current_meta) + new_meta["filename"] = computed_filename + replace_asset_info_metadata_projection( + session, asset_info_id=asset_info_id, user_metadata=new_meta + ) + touched = True + + if tags is not None: + set_asset_info_tags( + session, + asset_info_id=asset_info_id, + tags=tags, + origin=tag_origin, + ) + touched = True + + if touched and user_metadata is None: + info.updated_at = utcnow() + session.flush() + + return info + + +def delete_asset_info_by_id( + session: Session, + *, + asset_info_id: str, + owner_id: str, +) -> bool: + stmt = sa.delete(AssetInfo).where( + AssetInfo.id == asset_info_id, + visible_owner_clause(owner_id), + ) + return int((session.execute(stmt)).rowcount or 0) > 0 + + def list_tags_with_usage( session: Session, prefix: str | None = None, @@ -265,3 +814,163 @@ def list_tags_with_usage( rows_norm = [(name, ttype, int(count or 0)) for (name, ttype, count) in rows] return rows_norm, int(total or 0) + + +def ensure_tags_exist(session: Session, names: Iterable[str], tag_type: str = "user") -> None: + wanted = normalize_tags(list(names)) + if not wanted: + return + rows = [{"name": n, "tag_type": tag_type} for n in list(dict.fromkeys(wanted))] + ins = ( + sqlite.insert(Tag) + .values(rows) + .on_conflict_do_nothing(index_elements=[Tag.name]) + ) + session.execute(ins) + + +def get_asset_tags(session: Session, *, asset_info_id: str) -> list[str]: + return [ + tag_name for (tag_name,) in ( + session.execute( + select(AssetInfoTag.tag_name).where(AssetInfoTag.asset_info_id == asset_info_id) + ) + ).all() + ] + + +def add_tags_to_asset_info( + session: Session, + *, + asset_info_id: str, + tags: Sequence[str], + origin: str = "manual", + create_if_missing: bool = True, + asset_info_row: Any = None, +) -> dict: + if not asset_info_row: + info = session.get(AssetInfo, asset_info_id) + if not info: + raise ValueError(f"AssetInfo {asset_info_id} not found") + + norm = normalize_tags(tags) + if not norm: + total = get_asset_tags(session, asset_info_id=asset_info_id) + return {"added": [], "already_present": [], "total_tags": total} + + if create_if_missing: + ensure_tags_exist(session, norm, tag_type="user") + + current = { + tag_name + for (tag_name,) in ( + session.execute( + sa.select(AssetInfoTag.tag_name).where(AssetInfoTag.asset_info_id == asset_info_id) + ) + ).all() + } + + want = set(norm) + to_add = sorted(want - current) + + if to_add: + with session.begin_nested() as nested: + try: + session.add_all( + [ + AssetInfoTag( + asset_info_id=asset_info_id, + tag_name=t, + origin=origin, + added_at=utcnow(), + ) + for t in to_add + ] + ) + session.flush() + except IntegrityError: + nested.rollback() + + after = set(get_asset_tags(session, asset_info_id=asset_info_id)) + return { + "added": sorted(((after - current) & want)), + "already_present": sorted(want & current), + "total_tags": sorted(after), + } + + +def remove_tags_from_asset_info( + session: Session, + *, + asset_info_id: str, + tags: Sequence[str], +) -> dict: + info = session.get(AssetInfo, asset_info_id) + if not info: + raise ValueError(f"AssetInfo {asset_info_id} not found") + + norm = normalize_tags(tags) + if not norm: + total = get_asset_tags(session, asset_info_id=asset_info_id) + return {"removed": [], "not_present": [], "total_tags": total} + + existing = { + tag_name + for (tag_name,) in ( + session.execute( + sa.select(AssetInfoTag.tag_name).where(AssetInfoTag.asset_info_id == asset_info_id) + ) + ).all() + } + + to_remove = sorted(set(t for t in norm if t in existing)) + not_present = sorted(set(t for t in norm if t not in existing)) + + if to_remove: + session.execute( + delete(AssetInfoTag) + .where( + AssetInfoTag.asset_info_id == asset_info_id, + AssetInfoTag.tag_name.in_(to_remove), + ) + ) + session.flush() + + total = get_asset_tags(session, asset_info_id=asset_info_id) + return {"removed": to_remove, "not_present": not_present, "total_tags": total} + + +def remove_missing_tag_for_asset_id( + session: Session, + *, + asset_id: str, +) -> None: + session.execute( + sa.delete(AssetInfoTag).where( + AssetInfoTag.asset_info_id.in_(sa.select(AssetInfo.id).where(AssetInfo.asset_id == asset_id)), + AssetInfoTag.tag_name == "missing", + ) + ) + + +def set_asset_info_preview( + session: Session, + *, + asset_info_id: str, + preview_asset_id: str | None = None, +) -> None: + """Set or clear preview_id and bump updated_at. Raises on unknown IDs.""" + info = session.get(AssetInfo, asset_info_id) + if not info: + raise ValueError(f"AssetInfo {asset_info_id} not found") + + if preview_asset_id is None: + info.preview_id = None + else: + # validate preview asset exists + if not session.get(Asset, preview_asset_id): + raise ValueError(f"Preview Asset {preview_asset_id} not found") + info.preview_id = preview_asset_id + + info.updated_at = utcnow() + session.flush() diff --git a/app/assets/helpers.py b/app/assets/helpers.py index 08b465b5a..5030b123a 100644 --- a/app/assets/helpers.py +++ b/app/assets/helpers.py @@ -1,5 +1,6 @@ import contextlib import os +from decimal import Decimal from aiohttp import web from datetime import datetime, timezone from pathlib import Path @@ -87,6 +88,40 @@ def get_comfy_models_folders() -> list[tuple[str, list[str]]]: targets.append((name, paths)) return targets +def resolve_destination_from_tags(tags: list[str]) -> tuple[str, list[str]]: + """Validates and maps tags -> (base_dir, subdirs_for_fs)""" + root = tags[0] + if root == "models": + if len(tags) < 2: + raise ValueError("at least two tags required for model asset") + try: + bases = folder_paths.folder_names_and_paths[tags[1]][0] + except KeyError: + raise ValueError(f"unknown model category '{tags[1]}'") + if not bases: + raise ValueError(f"no base path configured for category '{tags[1]}'") + base_dir = os.path.abspath(bases[0]) + raw_subdirs = tags[2:] + else: + base_dir = os.path.abspath( + folder_paths.get_input_directory() if root == "input" else folder_paths.get_output_directory() + ) + raw_subdirs = tags[1:] + for i in raw_subdirs: + if i in (".", ".."): + raise ValueError("invalid path component in tags") + + return base_dir, raw_subdirs if raw_subdirs else [] + +def ensure_within_base(candidate: str, base: str) -> None: + cand_abs = os.path.abspath(candidate) + base_abs = os.path.abspath(base) + try: + if os.path.commonpath([cand_abs, base_abs]) != base_abs: + raise ValueError("destination escapes base directory") + except Exception: + raise ValueError("invalid destination path") + def compute_relative_filename(file_path: str) -> str | None: """ Return the model's path relative to the last well-known folder (the model category), @@ -113,7 +148,6 @@ def compute_relative_filename(file_path: str) -> str | None: return "/".join(inside) return "/".join(parts) # input/output: keep all parts - def get_relative_to_root_category_path_of_asset(file_path: str) -> tuple[Literal["input", "output", "models"], str]: """Given an absolute or relative file path, determine which root category the path belongs to: - 'input' if the file resides under `folder_paths.get_input_directory()` @@ -215,3 +249,64 @@ def collect_models_files() -> list[str]: if allowed: out.append(abs_path) return out + +def is_scalar(v): + if v is None: + return True + if isinstance(v, bool): + return True + if isinstance(v, (int, float, Decimal, str)): + return True + return False + +def project_kv(key: str, value): + """ + Turn a metadata key/value into typed projection rows. + Returns list[dict] with keys: + key, ordinal, and one of val_str / val_num / val_bool / val_json (others None) + """ + rows: list[dict] = [] + + def _null_row(ordinal: int) -> dict: + return { + "key": key, "ordinal": ordinal, + "val_str": None, "val_num": None, "val_bool": None, "val_json": None + } + + if value is None: + rows.append(_null_row(0)) + return rows + + if is_scalar(value): + if isinstance(value, bool): + rows.append({"key": key, "ordinal": 0, "val_bool": bool(value)}) + elif isinstance(value, (int, float, Decimal)): + num = value if isinstance(value, Decimal) else Decimal(str(value)) + rows.append({"key": key, "ordinal": 0, "val_num": num}) + elif isinstance(value, str): + rows.append({"key": key, "ordinal": 0, "val_str": value}) + else: + rows.append({"key": key, "ordinal": 0, "val_json": value}) + return rows + + if isinstance(value, list): + if all(is_scalar(x) for x in value): + for i, x in enumerate(value): + if x is None: + rows.append(_null_row(i)) + elif isinstance(x, bool): + rows.append({"key": key, "ordinal": i, "val_bool": bool(x)}) + elif isinstance(x, (int, float, Decimal)): + num = x if isinstance(x, Decimal) else Decimal(str(x)) + rows.append({"key": key, "ordinal": i, "val_num": num}) + elif isinstance(x, str): + rows.append({"key": key, "ordinal": i, "val_str": x}) + else: + rows.append({"key": key, "ordinal": i, "val_json": x}) + return rows + for i, x in enumerate(value): + rows.append({"key": key, "ordinal": i, "val_json": x}) + return rows + + rows.append({"key": key, "ordinal": 0, "val_json": value}) + return rows diff --git a/app/assets/manager.py b/app/assets/manager.py index 6425e7aa2..a68c8c8ae 100644 --- a/app/assets/manager.py +++ b/app/assets/manager.py @@ -1,13 +1,33 @@ +import os +import mimetypes +import contextlib from typing import Sequence from app.database.db import create_session -from app.assets.api import schemas_out +from app.assets.api import schemas_out, schemas_in from app.assets.database.queries import ( asset_exists_by_hash, + asset_info_exists_for_asset_id, + get_asset_by_hash, + get_asset_info_by_id, fetch_asset_info_asset_and_tags, + fetch_asset_info_and_asset, + create_asset_info_for_existing_asset, + touch_asset_info_by_id, + update_asset_info_full, + delete_asset_info_by_id, + list_cache_states_by_asset_id, list_asset_infos_page, list_tags_with_usage, + get_asset_tags, + add_tags_to_asset_info, + remove_tags_from_asset_info, + pick_best_live_path, + ingest_fs_asset, + set_asset_info_preview, ) +from app.assets.helpers import resolve_destination_from_tags, ensure_within_base +from app.assets.database.models import Asset def _safe_sort_field(requested: str | None) -> str: @@ -19,11 +39,28 @@ def _safe_sort_field(requested: str | None) -> str: return "created_at" -def asset_exists(asset_hash: str) -> bool: +def _get_size_mtime_ns(path: str) -> tuple[int, int]: + st = os.stat(path, follow_symlinks=True) + return st.st_size, getattr(st, "st_mtime_ns", int(st.st_mtime * 1_000_000_000)) + + +def _safe_filename(name: str | None, fallback: str) -> str: + n = os.path.basename((name or "").strip() or fallback) + if n: + return n + return fallback + + +def asset_exists(*, asset_hash: str) -> bool: + """ + Check if an asset with a given hash exists in database. + """ with create_session() as session: return asset_exists_by_hash(session, asset_hash=asset_hash) + def list_assets( + *, include_tags: Sequence[str] | None = None, exclude_tags: Sequence[str] | None = None, name_contains: str | None = None, @@ -63,7 +100,6 @@ def list_assets( size=int(asset.size_bytes) if asset else None, mime_type=asset.mime_type if asset else None, tags=tags, - preview_url=f"/api/assets/{info.id}/content", created_at=info.created_at, updated_at=info.updated_at, last_access_time=info.last_access_time, @@ -76,7 +112,12 @@ def list_assets( has_more=(offset + len(summaries)) < total, ) -def get_asset(asset_info_id: str, owner_id: str = "") -> schemas_out.AssetDetail: + +def get_asset( + *, + asset_info_id: str, + owner_id: str = "", +) -> schemas_out.AssetDetail: with create_session() as session: res = fetch_asset_info_asset_and_tags(session, asset_info_id=asset_info_id, owner_id=owner_id) if not res: @@ -97,6 +138,358 @@ def get_asset(asset_info_id: str, owner_id: str = "") -> schemas_out.AssetDetail last_access_time=info.last_access_time, ) + +def resolve_asset_content_for_download( + *, + asset_info_id: str, + owner_id: str = "", +) -> tuple[str, str, str]: + with create_session() as session: + pair = fetch_asset_info_and_asset(session, asset_info_id=asset_info_id, owner_id=owner_id) + if not pair: + raise ValueError(f"AssetInfo {asset_info_id} not found") + + info, asset = pair + states = list_cache_states_by_asset_id(session, asset_id=asset.id) + abs_path = pick_best_live_path(states) + if not abs_path: + raise FileNotFoundError + + touch_asset_info_by_id(session, asset_info_id=asset_info_id) + session.commit() + + ctype = asset.mime_type or mimetypes.guess_type(info.name or abs_path)[0] or "application/octet-stream" + download_name = info.name or os.path.basename(abs_path) + return abs_path, ctype, download_name + + +def upload_asset_from_temp_path( + spec: schemas_in.UploadAssetSpec, + *, + temp_path: str, + client_filename: str | None = None, + owner_id: str = "", + expected_asset_hash: str | None = None, +) -> schemas_out.AssetCreated: + """ + Create new asset or update existing asset from a temporary file path. + """ + try: + # NOTE: blake3 is not required right now, so this will fail if blake3 is not installed in local environment + import app.assets.hashing as hashing + digest = hashing.blake3_hash(temp_path) + except Exception as e: + raise RuntimeError(f"failed to hash uploaded file: {e}") + asset_hash = "blake3:" + digest + + if expected_asset_hash and asset_hash != expected_asset_hash.strip().lower(): + raise ValueError("HASH_MISMATCH") + + with create_session() as session: + existing = get_asset_by_hash(session, asset_hash=asset_hash) + if existing is not None: + with contextlib.suppress(Exception): + if temp_path and os.path.exists(temp_path): + os.remove(temp_path) + + display_name = _safe_filename(spec.name or (client_filename or ""), fallback=digest) + info = create_asset_info_for_existing_asset( + session, + asset_hash=asset_hash, + name=display_name, + user_metadata=spec.user_metadata or {}, + tags=spec.tags or [], + tag_origin="manual", + owner_id=owner_id, + ) + tag_names = get_asset_tags(session, asset_info_id=info.id) + session.commit() + + return schemas_out.AssetCreated( + id=info.id, + name=info.name, + asset_hash=existing.hash, + size=int(existing.size_bytes) if existing.size_bytes is not None else None, + mime_type=existing.mime_type, + tags=tag_names, + user_metadata=info.user_metadata or {}, + preview_id=info.preview_id, + created_at=info.created_at, + last_access_time=info.last_access_time, + created_new=False, + ) + + base_dir, subdirs = resolve_destination_from_tags(spec.tags) + dest_dir = os.path.join(base_dir, *subdirs) if subdirs else base_dir + os.makedirs(dest_dir, exist_ok=True) + + src_for_ext = (client_filename or spec.name or "").strip() + _ext = os.path.splitext(os.path.basename(src_for_ext))[1] if src_for_ext else "" + ext = _ext if 0 < len(_ext) <= 16 else "" + hashed_basename = f"{digest}{ext}" + dest_abs = os.path.abspath(os.path.join(dest_dir, hashed_basename)) + ensure_within_base(dest_abs, base_dir) + + content_type = ( + mimetypes.guess_type(os.path.basename(src_for_ext), strict=False)[0] + or mimetypes.guess_type(hashed_basename, strict=False)[0] + or "application/octet-stream" + ) + + try: + os.replace(temp_path, dest_abs) + except Exception as e: + raise RuntimeError(f"failed to move uploaded file into place: {e}") + + try: + size_bytes, mtime_ns = _get_size_mtime_ns(dest_abs) + except OSError as e: + raise RuntimeError(f"failed to stat destination file: {e}") + + with create_session() as session: + result = ingest_fs_asset( + session, + asset_hash=asset_hash, + abs_path=dest_abs, + size_bytes=size_bytes, + mtime_ns=mtime_ns, + mime_type=content_type, + info_name=_safe_filename(spec.name or (client_filename or ""), fallback=digest), + owner_id=owner_id, + preview_id=None, + user_metadata=spec.user_metadata or {}, + tags=spec.tags, + tag_origin="manual", + require_existing_tags=False, + ) + info_id = result["asset_info_id"] + if not info_id: + raise RuntimeError("failed to create asset metadata") + + pair = fetch_asset_info_and_asset(session, asset_info_id=info_id, owner_id=owner_id) + if not pair: + raise RuntimeError("inconsistent DB state after ingest") + info, asset = pair + tag_names = get_asset_tags(session, asset_info_id=info.id) + created_result = schemas_out.AssetCreated( + id=info.id, + name=info.name, + asset_hash=asset.hash, + size=int(asset.size_bytes), + mime_type=asset.mime_type, + tags=tag_names, + user_metadata=info.user_metadata or {}, + preview_id=info.preview_id, + created_at=info.created_at, + last_access_time=info.last_access_time, + created_new=result["asset_created"], + ) + session.commit() + + return created_result + + +def update_asset( + *, + asset_info_id: str, + name: str | None = None, + tags: list[str] | None = None, + user_metadata: dict | None = None, + owner_id: str = "", +) -> schemas_out.AssetUpdated: + with create_session() as session: + info_row = get_asset_info_by_id(session, asset_info_id=asset_info_id) + if not info_row: + raise ValueError(f"AssetInfo {asset_info_id} not found") + if info_row.owner_id and info_row.owner_id != owner_id: + raise PermissionError("not owner") + + info = update_asset_info_full( + session, + asset_info_id=asset_info_id, + name=name, + tags=tags, + user_metadata=user_metadata, + tag_origin="manual", + asset_info_row=info_row, + ) + + tag_names = get_asset_tags(session, asset_info_id=asset_info_id) + result = schemas_out.AssetUpdated( + id=info.id, + name=info.name, + asset_hash=info.asset.hash if info.asset else None, + tags=tag_names, + user_metadata=info.user_metadata or {}, + updated_at=info.updated_at, + ) + session.commit() + + return result + + +def set_asset_preview( + *, + asset_info_id: str, + preview_asset_id: str | None = None, + owner_id: str = "", +) -> schemas_out.AssetDetail: + with create_session() as session: + info_row = get_asset_info_by_id(session, asset_info_id=asset_info_id) + if not info_row: + raise ValueError(f"AssetInfo {asset_info_id} not found") + if info_row.owner_id and info_row.owner_id != owner_id: + raise PermissionError("not owner") + + set_asset_info_preview( + session, + asset_info_id=asset_info_id, + preview_asset_id=preview_asset_id, + ) + + res = fetch_asset_info_asset_and_tags(session, asset_info_id=asset_info_id, owner_id=owner_id) + if not res: + raise RuntimeError("State changed during preview update") + info, asset, tags = res + result = schemas_out.AssetDetail( + id=info.id, + name=info.name, + asset_hash=asset.hash if asset else None, + size=int(asset.size_bytes) if asset and asset.size_bytes is not None else None, + mime_type=asset.mime_type if asset else None, + tags=tags, + user_metadata=info.user_metadata or {}, + preview_id=info.preview_id, + created_at=info.created_at, + last_access_time=info.last_access_time, + ) + session.commit() + + return result + + +def delete_asset_reference(*, asset_info_id: str, owner_id: str, delete_content_if_orphan: bool = True) -> bool: + with create_session() as session: + info_row = get_asset_info_by_id(session, asset_info_id=asset_info_id) + asset_id = info_row.asset_id if info_row else None + deleted = delete_asset_info_by_id(session, asset_info_id=asset_info_id, owner_id=owner_id) + if not deleted: + session.commit() + return False + + if not delete_content_if_orphan or not asset_id: + session.commit() + return True + + still_exists = asset_info_exists_for_asset_id(session, asset_id=asset_id) + if still_exists: + session.commit() + return True + + states = list_cache_states_by_asset_id(session, asset_id=asset_id) + file_paths = [s.file_path for s in (states or []) if getattr(s, "file_path", None)] + + asset_row = session.get(Asset, asset_id) + if asset_row is not None: + session.delete(asset_row) + + session.commit() + for p in file_paths: + with contextlib.suppress(Exception): + if p and os.path.isfile(p): + os.remove(p) + return True + + +def create_asset_from_hash( + *, + hash_str: str, + name: str, + tags: list[str] | None = None, + user_metadata: dict | None = None, + owner_id: str = "", +) -> schemas_out.AssetCreated | None: + canonical = hash_str.strip().lower() + with create_session() as session: + asset = get_asset_by_hash(session, asset_hash=canonical) + if not asset: + return None + + info = create_asset_info_for_existing_asset( + session, + asset_hash=canonical, + name=_safe_filename(name, fallback=canonical.split(":", 1)[1]), + user_metadata=user_metadata or {}, + tags=tags or [], + tag_origin="manual", + owner_id=owner_id, + ) + tag_names = get_asset_tags(session, asset_info_id=info.id) + result = schemas_out.AssetCreated( + id=info.id, + name=info.name, + asset_hash=asset.hash, + size=int(asset.size_bytes), + mime_type=asset.mime_type, + tags=tag_names, + user_metadata=info.user_metadata or {}, + preview_id=info.preview_id, + created_at=info.created_at, + last_access_time=info.last_access_time, + created_new=False, + ) + session.commit() + + return result + + +def add_tags_to_asset( + *, + asset_info_id: str, + tags: list[str], + origin: str = "manual", + owner_id: str = "", +) -> schemas_out.TagsAdd: + with create_session() as session: + info_row = get_asset_info_by_id(session, asset_info_id=asset_info_id) + if not info_row: + raise ValueError(f"AssetInfo {asset_info_id} not found") + if info_row.owner_id and info_row.owner_id != owner_id: + raise PermissionError("not owner") + data = add_tags_to_asset_info( + session, + asset_info_id=asset_info_id, + tags=tags, + origin=origin, + create_if_missing=True, + asset_info_row=info_row, + ) + session.commit() + return schemas_out.TagsAdd(**data) + + +def remove_tags_from_asset( + *, + asset_info_id: str, + tags: list[str], + owner_id: str = "", +) -> schemas_out.TagsRemove: + with create_session() as session: + info_row = get_asset_info_by_id(session, asset_info_id=asset_info_id) + if not info_row: + raise ValueError(f"AssetInfo {asset_info_id} not found") + if info_row.owner_id and info_row.owner_id != owner_id: + raise PermissionError("not owner") + + data = remove_tags_from_asset_info( + session, + asset_info_id=asset_info_id, + tags=tags, + ) + session.commit() + return schemas_out.TagsRemove(**data) + + def list_tags( prefix: str | None = None, limit: int = 100, diff --git a/app/assets/scanner.py b/app/assets/scanner.py index a16e41d94..0172a5c2f 100644 --- a/app/assets/scanner.py +++ b/app/assets/scanner.py @@ -27,6 +27,7 @@ def seed_assets(roots: tuple[RootType, ...], enable_logging: bool = False) -> No t_start = time.perf_counter() created = 0 skipped_existing = 0 + orphans_pruned = 0 paths: list[str] = [] try: existing_paths: set[str] = set() @@ -38,6 +39,11 @@ def seed_assets(roots: tuple[RootType, ...], enable_logging: bool = False) -> No except Exception as e: logging.exception("fast DB scan failed for %s: %s", r, e) + try: + orphans_pruned = _prune_orphaned_assets(roots) + except Exception as e: + logging.exception("orphan pruning failed: %s", e) + if "models" in roots: paths.extend(collect_models_files()) if "input" in roots: @@ -85,15 +91,43 @@ def seed_assets(roots: tuple[RootType, ...], enable_logging: bool = False) -> No finally: if enable_logging: logging.info( - "Assets scan(roots=%s) completed in %.3fs (created=%d, skipped_existing=%d, total_seen=%d)", + "Assets scan(roots=%s) completed in %.3fs (created=%d, skipped_existing=%d, orphans_pruned=%d, total_seen=%d)", roots, time.perf_counter() - t_start, created, skipped_existing, + orphans_pruned, len(paths), ) +def _prune_orphaned_assets(roots: tuple[RootType, ...]) -> int: + """Prune cache states outside configured prefixes, then delete orphaned seed assets.""" + all_prefixes = [os.path.abspath(p) for r in roots for p in prefixes_for_root(r)] + if not all_prefixes: + return 0 + + def make_prefix_condition(prefix: str): + base = prefix if prefix.endswith(os.sep) else prefix + os.sep + escaped, esc = escape_like_prefix(base) + return AssetCacheState.file_path.like(escaped + "%", escape=esc) + + matches_valid_prefix = sqlalchemy.or_(*[make_prefix_condition(p) for p in all_prefixes]) + + orphan_subq = ( + sqlalchemy.select(Asset.id) + .outerjoin(AssetCacheState, AssetCacheState.asset_id == Asset.id) + .where(Asset.hash.is_(None), AssetCacheState.id.is_(None)) + ).scalar_subquery() + + with create_session() as sess: + sess.execute(sqlalchemy.delete(AssetCacheState).where(~matches_valid_prefix)) + sess.execute(sqlalchemy.delete(AssetInfo).where(AssetInfo.asset_id.in_(orphan_subq))) + result = sess.execute(sqlalchemy.delete(Asset).where(Asset.id.in_(orphan_subq))) + sess.commit() + return result.rowcount + + def _fast_db_consistency_pass( root: RootType, *, diff --git a/tests-unit/assets_test/conftest.py b/tests-unit/assets_test/conftest.py new file mode 100644 index 000000000..0a57dd7b5 --- /dev/null +++ b/tests-unit/assets_test/conftest.py @@ -0,0 +1,271 @@ +import contextlib +import json +import os +import socket +import subprocess +import sys +import tempfile +import time +from pathlib import Path +from typing import Callable, Iterator, Optional + +import pytest +import requests + + +def pytest_addoption(parser: pytest.Parser) -> None: + """ + Allow overriding the database URL used by the spawned ComfyUI process. + Priority: + 1) --db-url command line option + 2) ASSETS_TEST_DB_URL environment variable (used by CI) + 3) default: None (will use file-backed sqlite in temp dir) + """ + parser.addoption( + "--db-url", + action="store", + default=os.environ.get("ASSETS_TEST_DB_URL"), + help="SQLAlchemy DB URL (e.g. sqlite:///path/to/db.sqlite3)", + ) + + +def _free_port() -> int: + with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s: + s.bind(("127.0.0.1", 0)) + return s.getsockname()[1] + + +def _make_base_dirs(root: Path) -> None: + for sub in ("models", "custom_nodes", "input", "output", "temp", "user"): + (root / sub).mkdir(parents=True, exist_ok=True) + + +def _wait_http_ready(base: str, session: requests.Session, timeout: float = 90.0) -> None: + start = time.time() + last_err = None + while time.time() - start < timeout: + try: + r = session.get(base + "/api/assets", timeout=5) + if r.status_code in (200, 400): + return + except Exception as e: + last_err = e + time.sleep(0.25) + raise RuntimeError(f"ComfyUI HTTP did not become ready: {last_err}") + + +@pytest.fixture(scope="session") +def comfy_tmp_base_dir() -> Path: + env_base = os.environ.get("ASSETS_TEST_BASE_DIR") + created_by_fixture = False + if env_base: + tmp = Path(env_base) + tmp.mkdir(parents=True, exist_ok=True) + else: + tmp = Path(tempfile.mkdtemp(prefix="comfyui-assets-tests-")) + created_by_fixture = True + _make_base_dirs(tmp) + yield tmp + if created_by_fixture: + with contextlib.suppress(Exception): + for p in sorted(tmp.rglob("*"), reverse=True): + if p.is_file() or p.is_symlink(): + p.unlink(missing_ok=True) + for p in sorted(tmp.glob("**/*"), reverse=True): + with contextlib.suppress(Exception): + p.rmdir() + tmp.rmdir() + + +@pytest.fixture(scope="session") +def comfy_url_and_proc(comfy_tmp_base_dir: Path, request: pytest.FixtureRequest): + """ + Boot ComfyUI subprocess with: + - sandbox base dir + - file-backed sqlite DB in temp dir + - autoscan disabled + Returns (base_url, process, port) + """ + port = _free_port() + db_url = request.config.getoption("--db-url") + if not db_url: + # Use a file-backed sqlite database in the temp directory + db_path = comfy_tmp_base_dir / "assets-test.sqlite3" + db_url = f"sqlite:///{db_path}" + + logs_dir = comfy_tmp_base_dir / "logs" + logs_dir.mkdir(exist_ok=True) + out_log = open(logs_dir / "stdout.log", "w", buffering=1) + err_log = open(logs_dir / "stderr.log", "w", buffering=1) + + comfy_root = Path(__file__).resolve().parent.parent.parent + if not (comfy_root / "main.py").is_file(): + raise FileNotFoundError(f"main.py not found under {comfy_root}") + + proc = subprocess.Popen( + args=[ + sys.executable, + "main.py", + f"--base-directory={str(comfy_tmp_base_dir)}", + f"--database-url={db_url}", + "--disable-assets-autoscan", + "--listen", + "127.0.0.1", + "--port", + str(port), + "--cpu", + ], + stdout=out_log, + stderr=err_log, + cwd=str(comfy_root), + env={**os.environ}, + ) + + for _ in range(50): + if proc.poll() is not None: + out_log.flush() + err_log.flush() + raise RuntimeError(f"ComfyUI exited early with code {proc.returncode}") + time.sleep(0.1) + + base_url = f"http://127.0.0.1:{port}" + try: + with requests.Session() as s: + _wait_http_ready(base_url, s, timeout=90.0) + yield base_url, proc, port + except Exception as e: + with contextlib.suppress(Exception): + proc.terminate() + proc.wait(timeout=10) + with contextlib.suppress(Exception): + out_log.flush() + err_log.flush() + raise RuntimeError(f"ComfyUI did not become ready: {e}") + + if proc and proc.poll() is None: + with contextlib.suppress(Exception): + proc.terminate() + proc.wait(timeout=15) + out_log.close() + err_log.close() + + +@pytest.fixture +def http() -> Iterator[requests.Session]: + with requests.Session() as s: + s.timeout = 120 + yield s + + +@pytest.fixture +def api_base(comfy_url_and_proc) -> str: + base_url, _proc, _port = comfy_url_and_proc + return base_url + + +def _post_multipart_asset( + session: requests.Session, + base: str, + *, + name: str, + tags: list[str], + meta: dict, + data: bytes, + extra_fields: Optional[dict] = None, +) -> tuple[int, dict]: + files = {"file": (name, data, "application/octet-stream")} + form_data = { + "tags": json.dumps(tags), + "name": name, + "user_metadata": json.dumps(meta), + } + if extra_fields: + for k, v in extra_fields.items(): + form_data[k] = v + r = session.post(base + "/api/assets", files=files, data=form_data, timeout=120) + return r.status_code, r.json() + + +@pytest.fixture +def make_asset_bytes() -> Callable[[str, int], bytes]: + def _make(name: str, size: int = 8192) -> bytes: + seed = sum(ord(c) for c in name) % 251 + return bytes((i * 31 + seed) % 256 for i in range(size)) + return _make + + +@pytest.fixture +def asset_factory(http: requests.Session, api_base: str): + """ + Returns create(name, tags, meta, data) -> response dict + Tracks created ids and deletes them after the test. + """ + created: list[str] = [] + + def create(name: str, tags: list[str], meta: dict, data: bytes) -> dict: + status, body = _post_multipart_asset(http, api_base, name=name, tags=tags, meta=meta, data=data) + assert status in (200, 201), body + created.append(body["id"]) + return body + + yield create + + for aid in created: + with contextlib.suppress(Exception): + http.delete(f"{api_base}/api/assets/{aid}", timeout=30) + + +@pytest.fixture +def seeded_asset(request: pytest.FixtureRequest, http: requests.Session, api_base: str) -> dict: + """ + Upload one asset with ".safetensors" extension into models/checkpoints/unit-tests/. + Returns response dict with id, asset_hash, tags, etc. + """ + name = "unit_1_example.safetensors" + p = getattr(request, "param", {}) or {} + tags: Optional[list[str]] = p.get("tags") + if tags is None: + tags = ["models", "checkpoints", "unit-tests", "alpha"] + meta = {"purpose": "test", "epoch": 1, "flags": ["x", "y"], "nullable": None} + files = {"file": (name, b"A" * 4096, "application/octet-stream")} + form_data = { + "tags": json.dumps(tags), + "name": name, + "user_metadata": json.dumps(meta), + } + r = http.post(api_base + "/api/assets", files=files, data=form_data, timeout=120) + body = r.json() + assert r.status_code == 201, body + return body + + +@pytest.fixture(autouse=True) +def autoclean_unit_test_assets(http: requests.Session, api_base: str): + """Ensure isolation by removing all AssetInfo rows tagged with 'unit-tests' after each test.""" + yield + + while True: + r = http.get( + api_base + "/api/assets", + params={"include_tags": "unit-tests", "limit": "500", "sort": "name"}, + timeout=30, + ) + if r.status_code != 200: + break + body = r.json() + ids = [a["id"] for a in body.get("assets", [])] + if not ids: + break + for aid in ids: + with contextlib.suppress(Exception): + http.delete(f"{api_base}/api/assets/{aid}", timeout=30) + + +def trigger_sync_seed_assets(session: requests.Session, base_url: str) -> None: + """Force a fast sync/seed pass by calling the seed endpoint.""" + session.post(base_url + "/api/assets/seed", json={"roots": ["models", "input", "output"]}, timeout=30) + time.sleep(0.2) + + +def get_asset_filename(asset_hash: str, extension: str) -> str: + return asset_hash.removeprefix("blake3:") + extension diff --git a/tests-unit/assets_test/test_assets_missing_sync.py b/tests-unit/assets_test/test_assets_missing_sync.py new file mode 100644 index 000000000..78fa7b404 --- /dev/null +++ b/tests-unit/assets_test/test_assets_missing_sync.py @@ -0,0 +1,348 @@ +import os +import uuid +from pathlib import Path + +import pytest +import requests +from conftest import get_asset_filename, trigger_sync_seed_assets + + + + +@pytest.mark.parametrize("root", ["input", "output"]) +def test_seed_asset_removed_when_file_is_deleted( + root: str, + http: requests.Session, + api_base: str, + comfy_tmp_base_dir: Path, +): + """Asset without hash (seed) whose file disappears: + after triggering sync_seed_assets, Asset + AssetInfo disappear. + """ + # Create a file directly under input/unit-tests/ so tags include "unit-tests" + case_dir = comfy_tmp_base_dir / root / "unit-tests" / "syncseed" + case_dir.mkdir(parents=True, exist_ok=True) + name = f"seed_{uuid.uuid4().hex[:8]}.bin" + fp = case_dir / name + fp.write_bytes(b"Z" * 2048) + + # Trigger a seed sync so DB sees this path (seed asset => hash is NULL) + trigger_sync_seed_assets(http, api_base) + + # Verify it is visible via API and carries no hash (seed) + r1 = http.get( + api_base + "/api/assets", + params={"include_tags": "unit-tests,syncseed", "name_contains": name}, + timeout=120, + ) + body1 = r1.json() + assert r1.status_code == 200 + # there should be exactly one with that name + matches = [a for a in body1.get("assets", []) if a.get("name") == name] + assert matches + assert matches[0].get("asset_hash") is None + asset_info_id = matches[0]["id"] + + # Remove the underlying file and sync again + if fp.exists(): + fp.unlink() + + trigger_sync_seed_assets(http, api_base) + + # It should disappear (AssetInfo and seed Asset gone) + r2 = http.get( + api_base + "/api/assets", + params={"include_tags": "unit-tests,syncseed", "name_contains": name}, + timeout=120, + ) + body2 = r2.json() + assert r2.status_code == 200 + matches2 = [a for a in body2.get("assets", []) if a.get("name") == name] + assert not matches2, f"Seed asset {asset_info_id} should be gone after sync" + + +@pytest.mark.skip(reason="Requires computing hashes of files in directories to verify and clear missing tags") +def test_hashed_asset_missing_tag_added_then_removed_after_scan( + http: requests.Session, + api_base: str, + comfy_tmp_base_dir: Path, + asset_factory, + make_asset_bytes, +): + """Hashed asset with a single cache_state: + 1. delete its file -> sync adds 'missing' + 2. restore file -> sync removes 'missing' + """ + name = "missing_tag_test.png" + tags = ["input", "unit-tests", "msync2"] + data = make_asset_bytes(name, 4096) + a = asset_factory(name, tags, {}, data) + + # Compute its on-disk path and remove it + dest = comfy_tmp_base_dir / "input" / "unit-tests" / "msync2" / get_asset_filename(a["asset_hash"], ".png") + assert dest.exists(), f"Expected asset file at {dest}" + dest.unlink() + + # Fast sync should add 'missing' to the AssetInfo + trigger_sync_seed_assets(http, api_base) + + g1 = http.get(f"{api_base}/api/assets/{a['id']}", timeout=120) + d1 = g1.json() + assert g1.status_code == 200, d1 + assert "missing" in set(d1.get("tags", [])), "Expected 'missing' tag after deletion" + + # Restore the file with the exact same content and sync again + dest.parent.mkdir(parents=True, exist_ok=True) + dest.write_bytes(data) + + trigger_sync_seed_assets(http, api_base) + + g2 = http.get(f"{api_base}/api/assets/{a['id']}", timeout=120) + d2 = g2.json() + assert g2.status_code == 200, d2 + assert "missing" not in set(d2.get("tags", [])), "Missing tag should be cleared after verify" + + +def test_hashed_asset_two_asset_infos_both_get_missing( + http: requests.Session, + api_base: str, + comfy_tmp_base_dir: Path, + asset_factory, +): + """Hashed asset with a single cache_state, but two AssetInfo rows: + deleting the single file then syncing should add 'missing' to both infos. + """ + # Upload one hashed asset + name = "two_infos_one_path.png" + base_tags = ["input", "unit-tests", "multiinfo"] + created = asset_factory(name, base_tags, {}, b"A" * 2048) + + # Create second AssetInfo for the same Asset via from-hash + payload = { + "hash": created["asset_hash"], + "name": "two_infos_one_path_copy.png", + "tags": base_tags, # keep it in our unit-tests scope for cleanup + "user_metadata": {"k": "v"}, + } + r2 = http.post(api_base + "/api/assets/from-hash", json=payload, timeout=120) + b2 = r2.json() + assert r2.status_code == 201, b2 + second_id = b2["id"] + + # Remove the single underlying file + p = comfy_tmp_base_dir / "input" / "unit-tests" / "multiinfo" / get_asset_filename(b2["asset_hash"], ".png") + assert p.exists() + p.unlink() + + r0 = http.get(api_base + "/api/tags", params={"limit": "1000", "include_zero": "false"}, timeout=120) + tags0 = r0.json() + assert r0.status_code == 200, tags0 + byname0 = {t["name"]: t for t in tags0.get("tags", [])} + old_missing = int(byname0.get("missing", {}).get("count", 0)) + + # Sync -> both AssetInfos for this asset must receive 'missing' + trigger_sync_seed_assets(http, api_base) + + ga = http.get(f"{api_base}/api/assets/{created['id']}", timeout=120) + da = ga.json() + assert ga.status_code == 200, da + assert "missing" in set(da.get("tags", [])) + + gb = http.get(f"{api_base}/api/assets/{second_id}", timeout=120) + db = gb.json() + assert gb.status_code == 200, db + assert "missing" in set(db.get("tags", [])) + + # Tag usage for 'missing' increased by exactly 2 (two AssetInfos) + r1 = http.get(api_base + "/api/tags", params={"limit": "1000", "include_zero": "false"}, timeout=120) + tags1 = r1.json() + assert r1.status_code == 200, tags1 + byname1 = {t["name"]: t for t in tags1.get("tags", [])} + new_missing = int(byname1.get("missing", {}).get("count", 0)) + assert new_missing == old_missing + 2 + + +@pytest.mark.skip(reason="Requires computing hashes of files in directories to deduplicate into multiple cache states") +def test_hashed_asset_two_cache_states_partial_delete_then_full_delete( + http: requests.Session, + api_base: str, + comfy_tmp_base_dir: Path, + asset_factory, + make_asset_bytes, + run_scan_and_wait, +): + """Hashed asset with two cache_state rows: + 1. delete one file -> sync should NOT add 'missing' + 2. delete second file -> sync should add 'missing' + """ + name = "two_cache_states_partial_delete.png" + tags = ["input", "unit-tests", "dual"] + data = make_asset_bytes(name, 3072) + + created = asset_factory(name, tags, {}, data) + path1 = comfy_tmp_base_dir / "input" / "unit-tests" / "dual" / get_asset_filename(created["asset_hash"], ".png") + assert path1.exists() + + # Create a second on-disk copy under the same root but different subfolder + path2 = comfy_tmp_base_dir / "input" / "unit-tests" / "dual_copy" / name + path2.parent.mkdir(parents=True, exist_ok=True) + path2.write_bytes(data) + + # Fast seed so the second path appears (as a seed initially) + trigger_sync_seed_assets(http, api_base) + + # Deduplication of AssetInfo-s will not happen as first AssetInfo has owner='default' and second has empty owner. + run_scan_and_wait("input") + + # Remove only one file and sync -> asset should still be healthy (no 'missing') + path1.unlink() + trigger_sync_seed_assets(http, api_base) + + g1 = http.get(f"{api_base}/api/assets/{created['id']}", timeout=120) + d1 = g1.json() + assert g1.status_code == 200, d1 + assert "missing" not in set(d1.get("tags", [])), "Should not be missing while one valid path remains" + + # Baseline 'missing' usage count just before last file removal + r0 = http.get(api_base + "/api/tags", params={"limit": "1000", "include_zero": "false"}, timeout=120) + tags0 = r0.json() + assert r0.status_code == 200, tags0 + old_missing = int({t["name"]: t for t in tags0.get("tags", [])}.get("missing", {}).get("count", 0)) + + # Remove the second (last) file and sync -> now we expect 'missing' on this AssetInfo + path2.unlink() + trigger_sync_seed_assets(http, api_base) + + g2 = http.get(f"{api_base}/api/assets/{created['id']}", timeout=120) + d2 = g2.json() + assert g2.status_code == 200, d2 + assert "missing" in set(d2.get("tags", [])), "Missing must be set once no valid paths remain" + + # Tag usage for 'missing' increased by exactly 2 (two AssetInfo for one Asset) + r1 = http.get(api_base + "/api/tags", params={"limit": "1000", "include_zero": "false"}, timeout=120) + tags1 = r1.json() + assert r1.status_code == 200, tags1 + new_missing = int({t["name"]: t for t in tags1.get("tags", [])}.get("missing", {}).get("count", 0)) + assert new_missing == old_missing + 2 + + +@pytest.mark.parametrize("root", ["input", "output"]) +def test_missing_tag_clears_on_fastpass_when_mtime_and_size_match( + root: str, + http: requests.Session, + api_base: str, + comfy_tmp_base_dir: Path, + asset_factory, + make_asset_bytes, +): + """ + Fast pass alone clears 'missing' when size and mtime match exactly: + 1) upload (hashed), record original mtime_ns + 2) delete -> fast pass adds 'missing' + 3) restore same bytes and set mtime back to the original value + 4) run fast pass again -> 'missing' is removed (no slow scan) + """ + scope = f"fastclear-{uuid.uuid4().hex[:6]}" + name = "fastpass_clear.bin" + data = make_asset_bytes(name, 3072) + + a = asset_factory(name, [root, "unit-tests", scope], {}, data) + aid = a["id"] + base = comfy_tmp_base_dir / root / "unit-tests" / scope + p = base / get_asset_filename(a["asset_hash"], ".bin") + st0 = p.stat() + orig_mtime_ns = getattr(st0, "st_mtime_ns", int(st0.st_mtime * 1_000_000_000)) + + # Delete -> fast pass adds 'missing' + p.unlink() + trigger_sync_seed_assets(http, api_base) + g1 = http.get(f"{api_base}/api/assets/{aid}", timeout=120) + d1 = g1.json() + assert g1.status_code == 200, d1 + assert "missing" in set(d1.get("tags", [])) + + # Restore same bytes and revert mtime to the original value + p.parent.mkdir(parents=True, exist_ok=True) + p.write_bytes(data) + # set both atime and mtime in ns to ensure exact match + os.utime(p, ns=(orig_mtime_ns, orig_mtime_ns)) + + # Fast pass should clear 'missing' without a scan + trigger_sync_seed_assets(http, api_base) + g2 = http.get(f"{api_base}/api/assets/{aid}", timeout=120) + d2 = g2.json() + assert g2.status_code == 200, d2 + assert "missing" not in set(d2.get("tags", [])), "Fast pass should clear 'missing' when size+mtime match" + + +@pytest.mark.skip(reason="Requires computing hashes of files in directories to deduplicate into multiple cache states") +@pytest.mark.parametrize("root", ["input", "output"]) +def test_fastpass_removes_stale_state_row_no_missing( + root: str, + http: requests.Session, + api_base: str, + comfy_tmp_base_dir: Path, + asset_factory, + make_asset_bytes, + run_scan_and_wait, +): + """ + Hashed asset with two states: + - delete one file + - run fast pass only + Expect: + - asset stays healthy (no 'missing') + - stale AssetCacheState row for the deleted path is removed. + We verify this behaviorally by recreating the deleted path and running fast pass again: + a new *seed* AssetInfo is created, which proves the old state row was not reused. + """ + scope = f"stale-{uuid.uuid4().hex[:6]}" + name = "two_states.bin" + data = make_asset_bytes(name, 2048) + + # Upload hashed asset at path1 + a = asset_factory(name, [root, "unit-tests", scope], {}, data) + base = comfy_tmp_base_dir / root / "unit-tests" / scope + a1_filename = get_asset_filename(a["asset_hash"], ".bin") + p1 = base / a1_filename + assert p1.exists() + + aid = a["id"] + h = a["asset_hash"] + + # Create second state path2, seed+scan to dedupe into the same Asset + p2 = base / "copy" / name + p2.parent.mkdir(parents=True, exist_ok=True) + p2.write_bytes(data) + trigger_sync_seed_assets(http, api_base) + run_scan_and_wait(root) + + # Delete path1 and run fast pass -> no 'missing' and stale state row should be removed + p1.unlink() + trigger_sync_seed_assets(http, api_base) + g1 = http.get(f"{api_base}/api/assets/{aid}", timeout=120) + d1 = g1.json() + assert g1.status_code == 200, d1 + assert "missing" not in set(d1.get("tags", [])) + + # Recreate path1 and run fast pass again. + # If the stale state row was removed, a NEW seed AssetInfo will appear for this path. + p1.write_bytes(data) + trigger_sync_seed_assets(http, api_base) + + rl = http.get( + api_base + "/api/assets", + params={"include_tags": f"unit-tests,{scope}"}, + timeout=120, + ) + bl = rl.json() + assert rl.status_code == 200, bl + items = bl.get("assets", []) + # one hashed AssetInfo (asset_hash == h) + one seed AssetInfo (asset_hash == null) + hashes = [it.get("asset_hash") for it in items if it.get("name") in (name, a1_filename)] + assert h in hashes + assert any(x is None for x in hashes), "Expected a new seed AssetInfo for the recreated path" + + # Asset identity still healthy + rh = http.head(f"{api_base}/api/assets/hash/{h}", timeout=120) + assert rh.status_code == 200 diff --git a/tests-unit/assets_test/test_crud.py b/tests-unit/assets_test/test_crud.py new file mode 100644 index 000000000..d2b69f475 --- /dev/null +++ b/tests-unit/assets_test/test_crud.py @@ -0,0 +1,306 @@ +import uuid +from concurrent.futures import ThreadPoolExecutor +from pathlib import Path + +import pytest +import requests +from conftest import get_asset_filename, trigger_sync_seed_assets + + +def test_create_from_hash_success( + http: requests.Session, api_base: str, seeded_asset: dict +): + h = seeded_asset["asset_hash"] + payload = { + "hash": h, + "name": "from_hash_ok.safetensors", + "tags": ["models", "checkpoints", "unit-tests", "from-hash"], + "user_metadata": {"k": "v"}, + } + r1 = http.post(f"{api_base}/api/assets/from-hash", json=payload, timeout=120) + b1 = r1.json() + assert r1.status_code == 201, b1 + assert b1["asset_hash"] == h + assert b1["created_new"] is False + aid = b1["id"] + + # Calling again with the same name should return the same AssetInfo id + r2 = http.post(f"{api_base}/api/assets/from-hash", json=payload, timeout=120) + b2 = r2.json() + assert r2.status_code == 201, b2 + assert b2["id"] == aid + + +def test_get_and_delete_asset(http: requests.Session, api_base: str, seeded_asset: dict): + aid = seeded_asset["id"] + + # GET detail + rg = http.get(f"{api_base}/api/assets/{aid}", timeout=120) + detail = rg.json() + assert rg.status_code == 200, detail + assert detail["id"] == aid + assert "user_metadata" in detail + assert "filename" in detail["user_metadata"] + + # DELETE + rd = http.delete(f"{api_base}/api/assets/{aid}", timeout=120) + assert rd.status_code == 204 + + # GET again -> 404 + rg2 = http.get(f"{api_base}/api/assets/{aid}", timeout=120) + body = rg2.json() + assert rg2.status_code == 404 + assert body["error"]["code"] == "ASSET_NOT_FOUND" + + +def test_delete_upon_reference_count( + http: requests.Session, api_base: str, seeded_asset: dict +): + # Create a second reference to the same asset via from-hash + src_hash = seeded_asset["asset_hash"] + payload = { + "hash": src_hash, + "name": "unit_ref_copy.safetensors", + "tags": ["models", "checkpoints", "unit-tests", "del-flow"], + "user_metadata": {"note": "copy"}, + } + r2 = http.post(f"{api_base}/api/assets/from-hash", json=payload, timeout=120) + copy = r2.json() + assert r2.status_code == 201, copy + assert copy["asset_hash"] == src_hash + assert copy["created_new"] is False + + # Delete original reference -> asset identity must remain + aid1 = seeded_asset["id"] + rd1 = http.delete(f"{api_base}/api/assets/{aid1}", timeout=120) + assert rd1.status_code == 204 + + rh1 = http.head(f"{api_base}/api/assets/hash/{src_hash}", timeout=120) + assert rh1.status_code == 200 # identity still present + + # Delete the last reference with default semantics -> identity and cached files removed + aid2 = copy["id"] + rd2 = http.delete(f"{api_base}/api/assets/{aid2}", timeout=120) + assert rd2.status_code == 204 + + rh2 = http.head(f"{api_base}/api/assets/hash/{src_hash}", timeout=120) + assert rh2.status_code == 404 # orphan content removed + + +def test_update_asset_fields(http: requests.Session, api_base: str, seeded_asset: dict): + aid = seeded_asset["id"] + original_tags = seeded_asset["tags"] + + payload = { + "name": "unit_1_renamed.safetensors", + "user_metadata": {"purpose": "updated", "epoch": 2}, + } + ru = http.put(f"{api_base}/api/assets/{aid}", json=payload, timeout=120) + body = ru.json() + assert ru.status_code == 200, body + assert body["name"] == payload["name"] + assert body["tags"] == original_tags # tags unchanged + assert body["user_metadata"]["purpose"] == "updated" + # filename should still be present and normalized by server + assert "filename" in body["user_metadata"] + + +def test_head_asset_by_hash(http: requests.Session, api_base: str, seeded_asset: dict): + h = seeded_asset["asset_hash"] + + # Existing + rh1 = http.head(f"{api_base}/api/assets/hash/{h}", timeout=120) + assert rh1.status_code == 200 + + # Non-existent + rh2 = http.head(f"{api_base}/api/assets/hash/blake3:{'0'*64}", timeout=120) + assert rh2.status_code == 404 + + +def test_head_asset_bad_hash_returns_400_and_no_body(http: requests.Session, api_base: str): + # Invalid format; handler returns a JSON error, but HEAD responses must not carry a payload. + # requests exposes an empty body for HEAD, so validate status and that there is no payload. + rh = http.head(f"{api_base}/api/assets/hash/not_a_hash", timeout=120) + assert rh.status_code == 400 + body = rh.content + assert body == b"" + + +def test_delete_nonexistent_returns_404(http: requests.Session, api_base: str): + bogus = str(uuid.uuid4()) + r = http.delete(f"{api_base}/api/assets/{bogus}", timeout=120) + body = r.json() + assert r.status_code == 404 + assert body["error"]["code"] == "ASSET_NOT_FOUND" + + +def test_create_from_hash_invalids(http: requests.Session, api_base: str): + # Bad hash algorithm + bad = { + "hash": "sha256:" + "0" * 64, + "name": "x.bin", + "tags": ["models", "checkpoints", "unit-tests"], + } + r1 = http.post(f"{api_base}/api/assets/from-hash", json=bad, timeout=120) + b1 = r1.json() + assert r1.status_code == 400 + assert b1["error"]["code"] == "INVALID_BODY" + + # Invalid JSON body + r2 = http.post(f"{api_base}/api/assets/from-hash", data=b"{not json}", timeout=120) + b2 = r2.json() + assert r2.status_code == 400 + assert b2["error"]["code"] == "INVALID_JSON" + + +def test_get_update_download_bad_ids(http: requests.Session, api_base: str): + # All endpoints should be not found, as we UUID regex directly in the route definition. + bad_id = "not-a-uuid" + + r1 = http.get(f"{api_base}/api/assets/{bad_id}", timeout=120) + assert r1.status_code == 404 + + r3 = http.get(f"{api_base}/api/assets/{bad_id}/content", timeout=120) + assert r3.status_code == 404 + + +def test_update_requires_at_least_one_field(http: requests.Session, api_base: str, seeded_asset: dict): + aid = seeded_asset["id"] + r = http.put(f"{api_base}/api/assets/{aid}", json={}, timeout=120) + body = r.json() + assert r.status_code == 400 + assert body["error"]["code"] == "INVALID_BODY" + + +@pytest.mark.parametrize("root", ["input", "output"]) +def test_concurrent_delete_same_asset_info_single_204( + root: str, + http: requests.Session, + api_base: str, + asset_factory, + make_asset_bytes, +): + """ + Many concurrent DELETE for the same AssetInfo should result in: + - exactly one 204 No Content (the one that actually deleted) + - all others 404 Not Found (row already gone) + """ + scope = f"conc-del-{uuid.uuid4().hex[:6]}" + name = "to_delete.bin" + data = make_asset_bytes(name, 1536) + + created = asset_factory(name, [root, "unit-tests", scope], {}, data) + aid = created["id"] + + # Hit the same endpoint N times in parallel. + n_tests = 4 + url = f"{api_base}/api/assets/{aid}?delete_content=false" + + def _do_delete(delete_url): + with requests.Session() as s: + return s.delete(delete_url, timeout=120).status_code + + with ThreadPoolExecutor(max_workers=n_tests) as ex: + statuses = list(ex.map(_do_delete, [url] * n_tests)) + + # Exactly one actual delete, the rest must be 404 + assert statuses.count(204) == 1, f"Expected exactly one 204; got: {statuses}" + assert statuses.count(404) == n_tests - 1, f"Expected {n_tests-1} 404; got: {statuses}" + + # The resource must be gone. + rg = http.get(f"{api_base}/api/assets/{aid}", timeout=120) + assert rg.status_code == 404 + + +@pytest.mark.parametrize("root", ["input", "output"]) +def test_metadata_filename_is_set_for_seed_asset_without_hash( + root: str, + http: requests.Session, + api_base: str, + comfy_tmp_base_dir: Path, +): + """Seed ingest (no hash yet) must compute user_metadata['filename'] immediately.""" + scope = f"seedmeta-{uuid.uuid4().hex[:6]}" + name = "seed_filename.bin" + + base = comfy_tmp_base_dir / root / "unit-tests" / scope / "a" / "b" + base.mkdir(parents=True, exist_ok=True) + fp = base / name + fp.write_bytes(b"Z" * 2048) + + trigger_sync_seed_assets(http, api_base) + + r1 = http.get( + api_base + "/api/assets", + params={"include_tags": f"unit-tests,{scope}", "name_contains": name}, + timeout=120, + ) + body = r1.json() + assert r1.status_code == 200, body + matches = [a for a in body.get("assets", []) if a.get("name") == name] + assert matches, "Seed asset should be visible after sync" + assert matches[0].get("asset_hash") is None # still a seed + aid = matches[0]["id"] + + r2 = http.get(f"{api_base}/api/assets/{aid}", timeout=120) + detail = r2.json() + assert r2.status_code == 200, detail + filename = (detail.get("user_metadata") or {}).get("filename") + expected = str(fp.relative_to(comfy_tmp_base_dir / root)).replace("\\", "/") + assert filename == expected, f"expected filename={expected}, got {filename!r}" + + +@pytest.mark.skip(reason="Requires computing hashes of files in directories to retarget cache states") +@pytest.mark.parametrize("root", ["input", "output"]) +def test_metadata_filename_computed_and_updated_on_retarget( + root: str, + http: requests.Session, + api_base: str, + comfy_tmp_base_dir: Path, + asset_factory, + make_asset_bytes, + run_scan_and_wait, +): + """ + 1) Ingest under {root}/unit-tests//a/b/ -> filename reflects relative path. + 2) Retarget by copying to {root}/unit-tests//x/, remove old file, + run fast pass + scan -> filename updates to new relative path. + """ + scope = f"meta-fn-{uuid.uuid4().hex[:6]}" + name1 = "compute_metadata_filename.png" + name2 = "compute_changed_metadata_filename.png" + data = make_asset_bytes(name1, 2100) + + # Upload into nested path a/b + a = asset_factory(name1, [root, "unit-tests", scope, "a", "b"], {}, data) + aid = a["id"] + + root_base = comfy_tmp_base_dir / root + p1 = (root_base / "unit-tests" / scope / "a" / "b" / get_asset_filename(a["asset_hash"], ".png")) + assert p1.exists() + + # filename at ingest should be the path relative to root + rel1 = str(p1.relative_to(root_base)).replace("\\", "/") + g1 = http.get(f"{api_base}/api/assets/{aid}", timeout=120) + d1 = g1.json() + assert g1.status_code == 200, d1 + fn1 = d1["user_metadata"].get("filename") + assert fn1 == rel1 + + # Retarget: copy to x/, remove old, then sync+scan + p2 = root_base / "unit-tests" / scope / "x" / name2 + p2.parent.mkdir(parents=True, exist_ok=True) + p2.write_bytes(data) + if p1.exists(): + p1.unlink() + + trigger_sync_seed_assets(http, api_base) # seed the new path + run_scan_and_wait(root) # verify/hash and reconcile + + # filename should now point at x/ + rel2 = str(p2.relative_to(root_base)).replace("\\", "/") + g2 = http.get(f"{api_base}/api/assets/{aid}", timeout=120) + d2 = g2.json() + assert g2.status_code == 200, d2 + fn2 = d2["user_metadata"].get("filename") + assert fn2 == rel2 diff --git a/tests-unit/assets_test/test_downloads.py b/tests-unit/assets_test/test_downloads.py new file mode 100644 index 000000000..cdebf9082 --- /dev/null +++ b/tests-unit/assets_test/test_downloads.py @@ -0,0 +1,166 @@ +import time +import uuid +from datetime import datetime +from pathlib import Path +from typing import Optional + +import pytest +import requests +from conftest import get_asset_filename, trigger_sync_seed_assets + + +def test_download_attachment_and_inline(http: requests.Session, api_base: str, seeded_asset: dict): + aid = seeded_asset["id"] + + # default attachment + r1 = http.get(f"{api_base}/api/assets/{aid}/content", timeout=120) + data = r1.content + assert r1.status_code == 200 + cd = r1.headers.get("Content-Disposition", "") + assert "attachment" in cd + assert data and len(data) == 4096 + + # inline requested + r2 = http.get(f"{api_base}/api/assets/{aid}/content?disposition=inline", timeout=120) + r2.content + assert r2.status_code == 200 + cd2 = r2.headers.get("Content-Disposition", "") + assert "inline" in cd2 + + +@pytest.mark.skip(reason="Requires computing hashes of files in directories to deduplicate into multiple cache states") +@pytest.mark.parametrize("root", ["input", "output"]) +def test_download_chooses_existing_state_and_updates_access_time( + root: str, + http: requests.Session, + api_base: str, + comfy_tmp_base_dir: Path, + asset_factory, + make_asset_bytes, + run_scan_and_wait, +): + """ + Hashed asset with two state paths: if the first one disappears, + GET /content still serves from the remaining path and bumps last_access_time. + """ + scope = f"dl-first-{uuid.uuid4().hex[:6]}" + name = "first_existing_state.bin" + data = make_asset_bytes(name, 3072) + + # Upload -> path1 + a = asset_factory(name, [root, "unit-tests", scope], {}, data) + aid = a["id"] + + base = comfy_tmp_base_dir / root / "unit-tests" / scope + path1 = base / get_asset_filename(a["asset_hash"], ".bin") + assert path1.exists() + + # Seed path2 by copying, then scan to dedupe into a second state + path2 = base / "alt" / name + path2.parent.mkdir(parents=True, exist_ok=True) + path2.write_bytes(data) + trigger_sync_seed_assets(http, api_base) + run_scan_and_wait(root) + + # Remove path1 so server must fall back to path2 + path1.unlink() + + # last_access_time before + rg0 = http.get(f"{api_base}/api/assets/{aid}", timeout=120) + d0 = rg0.json() + assert rg0.status_code == 200, d0 + ts0 = d0.get("last_access_time") + + time.sleep(0.05) + r = http.get(f"{api_base}/api/assets/{aid}/content", timeout=120) + blob = r.content + assert r.status_code == 200 + assert blob == data # must serve from the surviving state (same bytes) + + rg1 = http.get(f"{api_base}/api/assets/{aid}", timeout=120) + d1 = rg1.json() + assert rg1.status_code == 200, d1 + ts1 = d1.get("last_access_time") + + def _parse_iso8601(s: Optional[str]) -> Optional[float]: + if not s: + return None + s = s[:-1] if s.endswith("Z") else s + return datetime.fromisoformat(s).timestamp() + + t0 = _parse_iso8601(ts0) + t1 = _parse_iso8601(ts1) + assert t1 is not None + if t0 is not None: + assert t1 > t0 + + +@pytest.mark.parametrize("seeded_asset", [{"tags": ["models", "checkpoints"]}], indirect=True) +def test_download_missing_file_returns_404( + http: requests.Session, api_base: str, comfy_tmp_base_dir: Path, seeded_asset: dict +): + # Remove the underlying file then attempt download. + # We initialize fixture without additional tags to know exactly the asset file path. + try: + aid = seeded_asset["id"] + rg = http.get(f"{api_base}/api/assets/{aid}", timeout=120) + detail = rg.json() + assert rg.status_code == 200 + asset_filename = get_asset_filename(detail["asset_hash"], ".safetensors") + abs_path = comfy_tmp_base_dir / "models" / "checkpoints" / asset_filename + assert abs_path.exists() + abs_path.unlink() + + r2 = http.get(f"{api_base}/api/assets/{aid}/content", timeout=120) + assert r2.status_code == 404 + body = r2.json() + assert body["error"]["code"] == "FILE_NOT_FOUND" + finally: + # We created asset without the "unit-tests" tag(see `autoclean_unit_test_assets`), we need to clear it manually. + dr = http.delete(f"{api_base}/api/assets/{aid}", timeout=120) + dr.content + + +@pytest.mark.skip(reason="Requires computing hashes of files in directories to deduplicate into multiple cache states") +@pytest.mark.parametrize("root", ["input", "output"]) +def test_download_404_if_all_states_missing( + root: str, + http: requests.Session, + api_base: str, + comfy_tmp_base_dir: Path, + asset_factory, + make_asset_bytes, + run_scan_and_wait, +): + """Multi-state asset: after the last remaining on-disk file is removed, download must return 404.""" + scope = f"dl-404-{uuid.uuid4().hex[:6]}" + name = "missing_all_states.bin" + data = make_asset_bytes(name, 2048) + + # Upload -> path1 + a = asset_factory(name, [root, "unit-tests", scope], {}, data) + aid = a["id"] + + base = comfy_tmp_base_dir / root / "unit-tests" / scope + p1 = base / get_asset_filename(a["asset_hash"], ".bin") + assert p1.exists() + + # Seed a second state and dedupe + p2 = base / "copy" / name + p2.parent.mkdir(parents=True, exist_ok=True) + p2.write_bytes(data) + trigger_sync_seed_assets(http, api_base) + run_scan_and_wait(root) + + # Remove first file -> download should still work via the second state + p1.unlink() + ok1 = http.get(f"{api_base}/api/assets/{aid}/content", timeout=120) + b1 = ok1.content + assert ok1.status_code == 200 and b1 == data + + # Remove the last file -> download must 404 + p2.unlink() + r2 = http.get(f"{api_base}/api/assets/{aid}/content", timeout=120) + body = r2.json() + assert r2.status_code == 404 + assert body["error"]["code"] == "FILE_NOT_FOUND" diff --git a/tests-unit/assets_test/test_list_filter.py b/tests-unit/assets_test/test_list_filter.py new file mode 100644 index 000000000..82e109832 --- /dev/null +++ b/tests-unit/assets_test/test_list_filter.py @@ -0,0 +1,342 @@ +import time +import uuid + +import requests + + +def test_list_assets_paging_and_sort(http: requests.Session, api_base: str, asset_factory, make_asset_bytes): + names = ["a1_u.safetensors", "a2_u.safetensors", "a3_u.safetensors"] + for n in names: + asset_factory( + n, + ["models", "checkpoints", "unit-tests", "paging"], + {"epoch": 1}, + make_asset_bytes(n, size=2048), + ) + + # name ascending for stable order + r1 = http.get( + api_base + "/api/assets", + params={"include_tags": "unit-tests,paging", "sort": "name", "order": "asc", "limit": "2", "offset": "0"}, + timeout=120, + ) + b1 = r1.json() + assert r1.status_code == 200 + got1 = [a["name"] for a in b1["assets"]] + assert got1 == sorted(names)[:2] + assert b1["has_more"] is True + + r2 = http.get( + api_base + "/api/assets", + params={"include_tags": "unit-tests,paging", "sort": "name", "order": "asc", "limit": "2", "offset": "2"}, + timeout=120, + ) + b2 = r2.json() + assert r2.status_code == 200 + got2 = [a["name"] for a in b2["assets"]] + assert got2 == sorted(names)[2:] + assert b2["has_more"] is False + + +def test_list_assets_include_exclude_and_name_contains(http: requests.Session, api_base: str, asset_factory): + a = asset_factory("inc_a.safetensors", ["models", "checkpoints", "unit-tests", "alpha"], {}, b"X" * 1024) + b = asset_factory("inc_b.safetensors", ["models", "checkpoints", "unit-tests", "beta"], {}, b"Y" * 1024) + + r = http.get( + api_base + "/api/assets", + params={"include_tags": "unit-tests,alpha", "exclude_tags": "beta", "limit": "50"}, + timeout=120, + ) + body = r.json() + assert r.status_code == 200 + names = [x["name"] for x in body["assets"]] + assert a["name"] in names + assert b["name"] not in names + + r2 = http.get( + api_base + "/api/assets", + params={"include_tags": "unit-tests", "name_contains": "inc_"}, + timeout=120, + ) + body2 = r2.json() + assert r2.status_code == 200 + names2 = [x["name"] for x in body2["assets"]] + assert a["name"] in names2 + assert b["name"] in names2 + + r2 = http.get( + api_base + "/api/assets", + params={"include_tags": "non-existing-tag"}, + timeout=120, + ) + body3 = r2.json() + assert r2.status_code == 200 + assert not body3["assets"] + + +def test_list_assets_sort_by_size_both_orders(http, api_base, asset_factory, make_asset_bytes): + t = ["models", "checkpoints", "unit-tests", "lf-size"] + n1, n2, n3 = "sz1.safetensors", "sz2.safetensors", "sz3.safetensors" + asset_factory(n1, t, {}, make_asset_bytes(n1, 1024)) + asset_factory(n2, t, {}, make_asset_bytes(n2, 2048)) + asset_factory(n3, t, {}, make_asset_bytes(n3, 3072)) + + r1 = http.get( + api_base + "/api/assets", + params={"include_tags": "unit-tests,lf-size", "sort": "size", "order": "asc"}, + timeout=120, + ) + b1 = r1.json() + names = [a["name"] for a in b1["assets"]] + assert names[:3] == [n1, n2, n3] + + r2 = http.get( + api_base + "/api/assets", + params={"include_tags": "unit-tests,lf-size", "sort": "size", "order": "desc"}, + timeout=120, + ) + b2 = r2.json() + names2 = [a["name"] for a in b2["assets"]] + assert names2[:3] == [n3, n2, n1] + + + +def test_list_assets_sort_by_updated_at_desc(http, api_base, asset_factory, make_asset_bytes): + t = ["models", "checkpoints", "unit-tests", "lf-upd"] + a1 = asset_factory("upd_a.safetensors", t, {}, make_asset_bytes("upd_a", 1200)) + a2 = asset_factory("upd_b.safetensors", t, {}, make_asset_bytes("upd_b", 1200)) + + # Rename the second asset to bump updated_at + rp = http.put(f"{api_base}/api/assets/{a2['id']}", json={"name": "upd_b_renamed.safetensors"}, timeout=120) + upd = rp.json() + assert rp.status_code == 200, upd + + r = http.get( + api_base + "/api/assets", + params={"include_tags": "unit-tests,lf-upd", "sort": "updated_at", "order": "desc"}, + timeout=120, + ) + body = r.json() + assert r.status_code == 200 + names = [x["name"] for x in body["assets"]] + assert names[0] == "upd_b_renamed.safetensors" + assert a1["name"] in names + + + +def test_list_assets_sort_by_last_access_time_desc(http, api_base, asset_factory, make_asset_bytes): + t = ["models", "checkpoints", "unit-tests", "lf-access"] + asset_factory("acc_a.safetensors", t, {}, make_asset_bytes("acc_a", 1100)) + time.sleep(0.02) + a2 = asset_factory("acc_b.safetensors", t, {}, make_asset_bytes("acc_b", 1100)) + + # Touch last_access_time of b by downloading its content + time.sleep(0.02) + dl = http.get(f"{api_base}/api/assets/{a2['id']}/content", timeout=120) + assert dl.status_code == 200 + dl.content + + r = http.get( + api_base + "/api/assets", + params={"include_tags": "unit-tests,lf-access", "sort": "last_access_time", "order": "desc"}, + timeout=120, + ) + body = r.json() + assert r.status_code == 200 + names = [x["name"] for x in body["assets"]] + assert names[0] == a2["name"] + + +def test_list_assets_include_tags_variants_and_case(http, api_base, asset_factory, make_asset_bytes): + t = ["models", "checkpoints", "unit-tests", "lf-include"] + a = asset_factory("incvar_alpha.safetensors", [*t, "alpha"], {}, make_asset_bytes("iva")) + asset_factory("incvar_beta.safetensors", [*t, "beta"], {}, make_asset_bytes("ivb")) + + # CSV + case-insensitive + r1 = http.get( + api_base + "/api/assets", + params={"include_tags": "UNIT-TESTS,LF-INCLUDE,alpha"}, + timeout=120, + ) + b1 = r1.json() + assert r1.status_code == 200 + names1 = [x["name"] for x in b1["assets"]] + assert a["name"] in names1 + assert not any("beta" in x for x in names1) + + # Repeated query params for include_tags + params_multi = [ + ("include_tags", "unit-tests"), + ("include_tags", "lf-include"), + ("include_tags", "alpha"), + ] + r2 = http.get(api_base + "/api/assets", params=params_multi, timeout=120) + b2 = r2.json() + assert r2.status_code == 200 + names2 = [x["name"] for x in b2["assets"]] + assert a["name"] in names2 + assert not any("beta" in x for x in names2) + + # Duplicates and spaces in CSV + r3 = http.get( + api_base + "/api/assets", + params={"include_tags": " unit-tests , lf-include , alpha , alpha "}, + timeout=120, + ) + b3 = r3.json() + assert r3.status_code == 200 + names3 = [x["name"] for x in b3["assets"]] + assert a["name"] in names3 + + +def test_list_assets_exclude_tags_dedup_and_case(http, api_base, asset_factory, make_asset_bytes): + t = ["models", "checkpoints", "unit-tests", "lf-exclude"] + a = asset_factory("ex_a_alpha.safetensors", [*t, "alpha"], {}, make_asset_bytes("exa", 900)) + asset_factory("ex_b_beta.safetensors", [*t, "beta"], {}, make_asset_bytes("exb", 900)) + + # Exclude uppercase should work + r1 = http.get( + api_base + "/api/assets", + params={"include_tags": "unit-tests,lf-exclude", "exclude_tags": "BETA"}, + timeout=120, + ) + b1 = r1.json() + assert r1.status_code == 200 + names1 = [x["name"] for x in b1["assets"]] + assert a["name"] in names1 + # Repeated excludes with duplicates + params_multi = [ + ("include_tags", "unit-tests"), + ("include_tags", "lf-exclude"), + ("exclude_tags", "beta"), + ("exclude_tags", "beta"), + ] + r2 = http.get(api_base + "/api/assets", params=params_multi, timeout=120) + b2 = r2.json() + assert r2.status_code == 200 + names2 = [x["name"] for x in b2["assets"]] + assert all("beta" not in x for x in names2) + + +def test_list_assets_name_contains_case_and_specials(http, api_base, asset_factory, make_asset_bytes): + t = ["models", "checkpoints", "unit-tests", "lf-name"] + a1 = asset_factory("CaseMix.SAFE", t, {}, make_asset_bytes("cm", 800)) + a2 = asset_factory("case-other.safetensors", t, {}, make_asset_bytes("co", 800)) + + r1 = http.get( + api_base + "/api/assets", + params={"include_tags": "unit-tests,lf-name", "name_contains": "casemix"}, + timeout=120, + ) + b1 = r1.json() + assert r1.status_code == 200 + names1 = [x["name"] for x in b1["assets"]] + assert a1["name"] in names1 + + r2 = http.get( + api_base + "/api/assets", + params={"include_tags": "unit-tests,lf-name", "name_contains": ".SAFE"}, + timeout=120, + ) + b2 = r2.json() + assert r2.status_code == 200 + names2 = [x["name"] for x in b2["assets"]] + assert a1["name"] in names2 + + r3 = http.get( + api_base + "/api/assets", + params={"include_tags": "unit-tests,lf-name", "name_contains": "case-"}, + timeout=120, + ) + b3 = r3.json() + assert r3.status_code == 200 + names3 = [x["name"] for x in b3["assets"]] + assert a2["name"] in names3 + + +def test_list_assets_offset_beyond_total_and_limit_boundary(http, api_base, asset_factory, make_asset_bytes): + t = ["models", "checkpoints", "unit-tests", "lf-pagelimits"] + asset_factory("pl1.safetensors", t, {}, make_asset_bytes("pl1", 600)) + asset_factory("pl2.safetensors", t, {}, make_asset_bytes("pl2", 600)) + asset_factory("pl3.safetensors", t, {}, make_asset_bytes("pl3", 600)) + + # Offset far beyond total + r1 = http.get( + api_base + "/api/assets", + params={"include_tags": "unit-tests,lf-pagelimits", "limit": "2", "offset": "10"}, + timeout=120, + ) + b1 = r1.json() + assert r1.status_code == 200 + assert not b1["assets"] + assert b1["has_more"] is False + + # Boundary large limit (<=500 is valid) + r2 = http.get( + api_base + "/api/assets", + params={"include_tags": "unit-tests,lf-pagelimits", "limit": "500"}, + timeout=120, + ) + b2 = r2.json() + assert r2.status_code == 200 + assert len(b2["assets"]) == 3 + assert b2["has_more"] is False + + +def test_list_assets_offset_negative_and_limit_nonint_rejected(http, api_base): + r1 = http.get(api_base + "/api/assets", params={"offset": "-1"}, timeout=120) + b1 = r1.json() + assert r1.status_code == 400 + assert b1["error"]["code"] == "INVALID_QUERY" + + r2 = http.get(api_base + "/api/assets", params={"limit": "abc"}, timeout=120) + b2 = r2.json() + assert r2.status_code == 400 + assert b2["error"]["code"] == "INVALID_QUERY" + + +def test_list_assets_invalid_query_rejected(http: requests.Session, api_base: str): + # limit too small + r1 = http.get(api_base + "/api/assets", params={"limit": "0"}, timeout=120) + b1 = r1.json() + assert r1.status_code == 400 + assert b1["error"]["code"] == "INVALID_QUERY" + + # bad metadata JSON + r2 = http.get(api_base + "/api/assets", params={"metadata_filter": "{not json"}, timeout=120) + b2 = r2.json() + assert r2.status_code == 400 + assert b2["error"]["code"] == "INVALID_QUERY" + + +def test_list_assets_name_contains_literal_underscore( + http, + api_base, + asset_factory, + make_asset_bytes, +): + """'name_contains' must treat '_' literally, not as a SQL wildcard. + We create: + - foo_bar.safetensors (should match) + - fooxbar.safetensors (must NOT match if '_' is escaped) + - foobar.safetensors (must NOT match) + """ + scope = f"lf-underscore-{uuid.uuid4().hex[:6]}" + tags = ["models", "checkpoints", "unit-tests", scope] + + a = asset_factory("foo_bar.safetensors", tags, {}, make_asset_bytes("a", 700)) + b = asset_factory("fooxbar.safetensors", tags, {}, make_asset_bytes("b", 700)) + c = asset_factory("foobar.safetensors", tags, {}, make_asset_bytes("c", 700)) + + r = http.get( + api_base + "/api/assets", + params={"include_tags": f"unit-tests,{scope}", "name_contains": "foo_bar"}, + timeout=120, + ) + body = r.json() + assert r.status_code == 200, body + names = [x["name"] for x in body["assets"]] + assert a["name"] in names, f"Expected literal underscore match to include {a['name']}" + assert b["name"] not in names, "Underscore must be escaped — should not match 'fooxbar'" + assert c["name"] not in names, "Underscore must be escaped — should not match 'foobar'" + assert body["total"] == 1 diff --git a/tests-unit/assets_test/test_metadata_filters.py b/tests-unit/assets_test/test_metadata_filters.py new file mode 100644 index 000000000..20285a3b3 --- /dev/null +++ b/tests-unit/assets_test/test_metadata_filters.py @@ -0,0 +1,395 @@ +import json + + +def test_meta_and_across_keys_and_types( + http, api_base: str, asset_factory, make_asset_bytes +): + name = "mf_and_mix.safetensors" + tags = ["models", "checkpoints", "unit-tests", "mf-and"] + meta = {"purpose": "mix", "epoch": 1, "active": True, "score": 1.23} + asset_factory(name, tags, meta, make_asset_bytes(name, 4096)) + + # All keys must match (AND semantics) + f_ok = {"purpose": "mix", "epoch": 1, "active": True, "score": 1.23} + r1 = http.get( + api_base + "/api/assets", + params={ + "include_tags": "unit-tests,mf-and", + "metadata_filter": json.dumps(f_ok), + }, + timeout=120, + ) + b1 = r1.json() + assert r1.status_code == 200 + names = [a["name"] for a in b1["assets"]] + assert name in names + + # One key mismatched -> no result + f_bad = {"purpose": "mix", "epoch": 2, "active": True} + r2 = http.get( + api_base + "/api/assets", + params={ + "include_tags": "unit-tests,mf-and", + "metadata_filter": json.dumps(f_bad), + }, + timeout=120, + ) + b2 = r2.json() + assert r2.status_code == 200 + assert not b2["assets"] + + +def test_meta_type_strictness_int_vs_str_and_bool(http, api_base, asset_factory, make_asset_bytes): + name = "mf_types.safetensors" + tags = ["models", "checkpoints", "unit-tests", "mf-types"] + meta = {"epoch": 1, "active": True} + asset_factory(name, tags, meta, make_asset_bytes(name)) + + # int filter matches numeric + r1 = http.get( + api_base + "/api/assets", + params={ + "include_tags": "unit-tests,mf-types", + "metadata_filter": json.dumps({"epoch": 1}), + }, + timeout=120, + ) + b1 = r1.json() + assert r1.status_code == 200 and any(a["name"] == name for a in b1["assets"]) + + # string "1" must NOT match numeric 1 + r2 = http.get( + api_base + "/api/assets", + params={ + "include_tags": "unit-tests,mf-types", + "metadata_filter": json.dumps({"epoch": "1"}), + }, + timeout=120, + ) + b2 = r2.json() + assert r2.status_code == 200 and not b2["assets"] + + # bool True matches, string "true" must NOT match + r3 = http.get( + api_base + "/api/assets", + params={ + "include_tags": "unit-tests,mf-types", + "metadata_filter": json.dumps({"active": True}), + }, + timeout=120, + ) + b3 = r3.json() + assert r3.status_code == 200 and any(a["name"] == name for a in b3["assets"]) + + r4 = http.get( + api_base + "/api/assets", + params={ + "include_tags": "unit-tests,mf-types", + "metadata_filter": json.dumps({"active": "true"}), + }, + timeout=120, + ) + b4 = r4.json() + assert r4.status_code == 200 and not b4["assets"] + + +def test_meta_any_of_list_of_scalars(http, api_base, asset_factory, make_asset_bytes): + name = "mf_list_scalars.safetensors" + tags = ["models", "checkpoints", "unit-tests", "mf-list"] + meta = {"flags": ["red", "green"]} + asset_factory(name, tags, meta, make_asset_bytes(name, 3000)) + + # Any-of should match because "green" is present + filt_ok = {"flags": ["blue", "green"]} + r1 = http.get( + api_base + "/api/assets", + params={"include_tags": "unit-tests,mf-list", "metadata_filter": json.dumps(filt_ok)}, + timeout=120, + ) + b1 = r1.json() + assert r1.status_code == 200 and any(a["name"] == name for a in b1["assets"]) + + # None of provided flags present -> no match + filt_miss = {"flags": ["blue", "yellow"]} + r2 = http.get( + api_base + "/api/assets", + params={"include_tags": "unit-tests,mf-list", "metadata_filter": json.dumps(filt_miss)}, + timeout=120, + ) + b2 = r2.json() + assert r2.status_code == 200 and not b2["assets"] + + # Duplicates in list should not break matching + filt_dup = {"flags": ["green", "green", "green"]} + r3 = http.get( + api_base + "/api/assets", + params={"include_tags": "unit-tests,mf-list", "metadata_filter": json.dumps(filt_dup)}, + timeout=120, + ) + b3 = r3.json() + assert r3.status_code == 200 and any(a["name"] == name for a in b3["assets"]) + + +def test_meta_none_semantics_missing_or_null_and_any_of_with_none( + http, api_base, asset_factory, make_asset_bytes +): + # a1: key missing; a2: explicit null; a3: concrete value + t = ["models", "checkpoints", "unit-tests", "mf-none"] + a1 = asset_factory("mf_none_missing.safetensors", t, {"x": 1}, make_asset_bytes("a1")) + a2 = asset_factory("mf_none_null.safetensors", t, {"maybe": None}, make_asset_bytes("a2")) + a3 = asset_factory("mf_none_value.safetensors", t, {"maybe": "x"}, make_asset_bytes("a3")) + + # Filter {maybe: None} must match a1 and a2, not a3 + filt = {"maybe": None} + r1 = http.get( + api_base + "/api/assets", + params={"include_tags": "unit-tests,mf-none", "metadata_filter": json.dumps(filt), "sort": "name"}, + timeout=120, + ) + b1 = r1.json() + assert r1.status_code == 200 + got = [a["name"] for a in b1["assets"]] + assert a1["name"] in got and a2["name"] in got and a3["name"] not in got + + # Any-of with None should include missing/null plus value matches + filt_any = {"maybe": [None, "x"]} + r2 = http.get( + api_base + "/api/assets", + params={"include_tags": "unit-tests,mf-none", "metadata_filter": json.dumps(filt_any), "sort": "name"}, + timeout=120, + ) + b2 = r2.json() + assert r2.status_code == 200 + got2 = [a["name"] for a in b2["assets"]] + assert a1["name"] in got2 and a2["name"] in got2 and a3["name"] in got2 + + +def test_meta_nested_json_object_equality(http, api_base, asset_factory, make_asset_bytes): + name = "mf_nested_json.safetensors" + tags = ["models", "checkpoints", "unit-tests", "mf-nested"] + cfg = {"optimizer": "adam", "lr": 0.001, "schedule": {"type": "cosine", "warmup": 100}} + asset_factory(name, tags, {"config": cfg}, make_asset_bytes(name, 2200)) + + # Exact JSON object equality (same structure) + r1 = http.get( + api_base + "/api/assets", + params={ + "include_tags": "unit-tests,mf-nested", + "metadata_filter": json.dumps({"config": cfg}), + }, + timeout=120, + ) + b1 = r1.json() + assert r1.status_code == 200 and any(a["name"] == name for a in b1["assets"]) + + # Different JSON object should not match + r2 = http.get( + api_base + "/api/assets", + params={ + "include_tags": "unit-tests,mf-nested", + "metadata_filter": json.dumps({"config": {"optimizer": "sgd"}}), + }, + timeout=120, + ) + b2 = r2.json() + assert r2.status_code == 200 and not b2["assets"] + + +def test_meta_list_of_objects_any_of(http, api_base, asset_factory, make_asset_bytes): + name = "mf_list_objects.safetensors" + tags = ["models", "checkpoints", "unit-tests", "mf-objlist"] + transforms = [{"type": "crop", "size": 128}, {"type": "flip", "p": 0.5}] + asset_factory(name, tags, {"transforms": transforms}, make_asset_bytes(name, 2048)) + + # Any-of for list of objects should match when one element equals the filter object + r1 = http.get( + api_base + "/api/assets", + params={ + "include_tags": "unit-tests,mf-objlist", + "metadata_filter": json.dumps({"transforms": {"type": "flip", "p": 0.5}}), + }, + timeout=120, + ) + b1 = r1.json() + assert r1.status_code == 200 and any(a["name"] == name for a in b1["assets"]) + + # Non-matching object -> no match + r2 = http.get( + api_base + "/api/assets", + params={ + "include_tags": "unit-tests,mf-objlist", + "metadata_filter": json.dumps({"transforms": {"type": "rotate", "deg": 90}}), + }, + timeout=120, + ) + b2 = r2.json() + assert r2.status_code == 200 and not b2["assets"] + + +def test_meta_with_special_and_unicode_keys(http, api_base, asset_factory, make_asset_bytes): + name = "mf_keys_unicode.safetensors" + tags = ["models", "checkpoints", "unit-tests", "mf-keys"] + meta = { + "weird.key": "v1", + "path/like": 7, + "with:colon": True, + "ключ": "значение", + "emoji": "🐍", + } + asset_factory(name, tags, meta, make_asset_bytes(name, 1500)) + + # Match all the special keys + filt = {"weird.key": "v1", "path/like": 7, "with:colon": True, "emoji": "🐍"} + r1 = http.get( + api_base + "/api/assets", + params={"include_tags": "unit-tests,mf-keys", "metadata_filter": json.dumps(filt)}, + timeout=120, + ) + b1 = r1.json() + assert r1.status_code == 200 and any(a["name"] == name for a in b1["assets"]) + + # Unicode key match + r2 = http.get( + api_base + "/api/assets", + params={"include_tags": "unit-tests,mf-keys", "metadata_filter": json.dumps({"ключ": "значение"})}, + timeout=120, + ) + b2 = r2.json() + assert r2.status_code == 200 and any(a["name"] == name for a in b2["assets"]) + + +def test_meta_with_zero_and_boolean_lists(http, api_base, asset_factory, make_asset_bytes): + t = ["models", "checkpoints", "unit-tests", "mf-zero-bool"] + a0 = asset_factory("mf_zero_count.safetensors", t, {"count": 0}, make_asset_bytes("z", 1025)) + a1 = asset_factory("mf_bool_list.safetensors", t, {"choices": [True, False]}, make_asset_bytes("b", 1026)) + + # count == 0 must match only a0 + r1 = http.get( + api_base + "/api/assets", + params={"include_tags": "unit-tests,mf-zero-bool", "metadata_filter": json.dumps({"count": 0})}, + timeout=120, + ) + b1 = r1.json() + assert r1.status_code == 200 + names1 = [a["name"] for a in b1["assets"]] + assert a0["name"] in names1 and a1["name"] not in names1 + + # Any-of list of booleans: True matches second asset + r2 = http.get( + api_base + "/api/assets", + params={"include_tags": "unit-tests,mf-zero-bool", "metadata_filter": json.dumps({"choices": True})}, + timeout=120, + ) + b2 = r2.json() + assert r2.status_code == 200 and any(a["name"] == a1["name"] for a in b2["assets"]) + + +def test_meta_mixed_list_types_and_strictness(http, api_base, asset_factory, make_asset_bytes): + name = "mf_mixed_list.safetensors" + tags = ["models", "checkpoints", "unit-tests", "mf-mixed"] + meta = {"mix": ["1", 1, True, None]} + asset_factory(name, tags, meta, make_asset_bytes(name, 1999)) + + # Should match because 1 is present + r1 = http.get( + api_base + "/api/assets", + params={"include_tags": "unit-tests,mf-mixed", "metadata_filter": json.dumps({"mix": [2, 1]})}, + timeout=120, + ) + b1 = r1.json() + assert r1.status_code == 200 and any(a["name"] == name for a in b1["assets"]) + + # Should NOT match for False + r2 = http.get( + api_base + "/api/assets", + params={"include_tags": "unit-tests,mf-mixed", "metadata_filter": json.dumps({"mix": False})}, + timeout=120, + ) + b2 = r2.json() + assert r2.status_code == 200 and not b2["assets"] + + +def test_meta_unknown_key_and_none_behavior_with_scope_tags(http, api_base, asset_factory, make_asset_bytes): + # Use a unique scope tag to avoid interference + t = ["models", "checkpoints", "unit-tests", "mf-unknown-scope"] + x = asset_factory("mf_unknown_a.safetensors", t, {"k1": 1}, make_asset_bytes("ua")) + y = asset_factory("mf_unknown_b.safetensors", t, {"k2": 2}, make_asset_bytes("ub")) + + # Filtering by unknown key with None should return both (missing key OR null) + r1 = http.get( + api_base + "/api/assets", + params={"include_tags": "unit-tests,mf-unknown-scope", "metadata_filter": json.dumps({"unknown": None})}, + timeout=120, + ) + b1 = r1.json() + assert r1.status_code == 200 + names = {a["name"] for a in b1["assets"]} + assert x["name"] in names and y["name"] in names + + # Filtering by unknown key with concrete value should return none + r2 = http.get( + api_base + "/api/assets", + params={"include_tags": "unit-tests,mf-unknown-scope", "metadata_filter": json.dumps({"unknown": "x"})}, + timeout=120, + ) + b2 = r2.json() + assert r2.status_code == 200 and not b2["assets"] + + +def test_meta_with_tags_include_exclude_and_name_contains(http, api_base, asset_factory, make_asset_bytes): + # alpha matches epoch=1; beta has epoch=2 + a = asset_factory( + "mf_tag_alpha.safetensors", + ["models", "checkpoints", "unit-tests", "mf-tag", "alpha"], + {"epoch": 1}, + make_asset_bytes("alpha"), + ) + b = asset_factory( + "mf_tag_beta.safetensors", + ["models", "checkpoints", "unit-tests", "mf-tag", "beta"], + {"epoch": 2}, + make_asset_bytes("beta"), + ) + + params = { + "include_tags": "unit-tests,mf-tag,alpha", + "exclude_tags": "beta", + "name_contains": "mf_tag_", + "metadata_filter": json.dumps({"epoch": 1}), + } + r = http.get(api_base + "/api/assets", params=params, timeout=120) + body = r.json() + assert r.status_code == 200 + names = [x["name"] for x in body["assets"]] + assert a["name"] in names + assert b["name"] not in names + + +def test_meta_sort_and_paging_under_filter(http, api_base, asset_factory, make_asset_bytes): + # Three assets in same scope with different sizes and a common filter key + t = ["models", "checkpoints", "unit-tests", "mf-sort"] + n1, n2, n3 = "mf_sort_1.safetensors", "mf_sort_2.safetensors", "mf_sort_3.safetensors" + asset_factory(n1, t, {"group": "g"}, make_asset_bytes(n1, 1024)) + asset_factory(n2, t, {"group": "g"}, make_asset_bytes(n2, 2048)) + asset_factory(n3, t, {"group": "g"}, make_asset_bytes(n3, 3072)) + + # Sort by size ascending with paging + q = { + "include_tags": "unit-tests,mf-sort", + "metadata_filter": json.dumps({"group": "g"}), + "sort": "size", "order": "asc", "limit": "2", + } + r1 = http.get(api_base + "/api/assets", params=q, timeout=120) + b1 = r1.json() + assert r1.status_code == 200 + got1 = [a["name"] for a in b1["assets"]] + assert got1 == [n1, n2] + assert b1["has_more"] is True + + q2 = {**q, "offset": "2"} + r2 = http.get(api_base + "/api/assets", params=q2, timeout=120) + b2 = r2.json() + assert r2.status_code == 200 + got2 = [a["name"] for a in b2["assets"]] + assert got2 == [n3] + assert b2["has_more"] is False diff --git a/tests-unit/assets_test/test_prune_orphaned_assets.py b/tests-unit/assets_test/test_prune_orphaned_assets.py new file mode 100644 index 000000000..f602e5a77 --- /dev/null +++ b/tests-unit/assets_test/test_prune_orphaned_assets.py @@ -0,0 +1,141 @@ +import uuid +from pathlib import Path + +import pytest +import requests +from conftest import get_asset_filename, trigger_sync_seed_assets + + +@pytest.fixture +def create_seed_file(comfy_tmp_base_dir: Path): + """Create a file on disk that will become a seed asset after sync.""" + created: list[Path] = [] + + def _create(root: str, scope: str, name: str | None = None, data: bytes = b"TEST") -> Path: + name = name or f"seed_{uuid.uuid4().hex[:8]}.bin" + path = comfy_tmp_base_dir / root / "unit-tests" / scope / name + path.parent.mkdir(parents=True, exist_ok=True) + path.write_bytes(data) + created.append(path) + return path + + yield _create + + for p in created: + p.unlink(missing_ok=True) + + +@pytest.fixture +def find_asset(http: requests.Session, api_base: str): + """Query API for assets matching scope and optional name.""" + def _find(scope: str, name: str | None = None) -> list[dict]: + params = {"include_tags": f"unit-tests,{scope}"} + if name: + params["name_contains"] = name + r = http.get(f"{api_base}/api/assets", params=params, timeout=120) + assert r.status_code == 200 + assets = r.json().get("assets", []) + if name: + return [a for a in assets if a.get("name") == name] + return assets + + return _find + + +@pytest.mark.parametrize("root", ["input", "output"]) +def test_orphaned_seed_asset_is_pruned( + root: str, + create_seed_file, + find_asset, + http: requests.Session, + api_base: str, +): + """Seed asset with deleted file is removed; with file present, it survives.""" + scope = f"prune-{uuid.uuid4().hex[:6]}" + fp = create_seed_file(root, scope) + name = fp.name + + trigger_sync_seed_assets(http, api_base) + assert find_asset(scope, name), "Seed asset should exist" + + fp.unlink() + trigger_sync_seed_assets(http, api_base) + assert not find_asset(scope, name), "Orphaned seed should be pruned" + + +def test_seed_asset_with_file_survives_prune( + create_seed_file, + find_asset, + http: requests.Session, + api_base: str, +): + """Seed asset with file still on disk is NOT pruned.""" + scope = f"keep-{uuid.uuid4().hex[:6]}" + fp = create_seed_file("input", scope) + + trigger_sync_seed_assets(http, api_base) + trigger_sync_seed_assets(http, api_base) + + assert find_asset(scope, fp.name), "Seed with valid file should survive" + + +def test_hashed_asset_not_pruned_when_file_missing( + http: requests.Session, + api_base: str, + comfy_tmp_base_dir: Path, + asset_factory, + make_asset_bytes, +): + """Hashed assets are never deleted by prune, even without file.""" + scope = f"hashed-{uuid.uuid4().hex[:6]}" + data = make_asset_bytes("test", 2048) + a = asset_factory("test.bin", ["input", "unit-tests", scope], {}, data) + + path = comfy_tmp_base_dir / "input" / "unit-tests" / scope / get_asset_filename(a["asset_hash"], ".bin") + path.unlink() + + trigger_sync_seed_assets(http, api_base) + + r = http.get(f"{api_base}/api/assets/{a['id']}", timeout=120) + assert r.status_code == 200, "Hashed asset should NOT be pruned" + + +def test_prune_across_multiple_roots( + create_seed_file, + find_asset, + http: requests.Session, + api_base: str, +): + """Prune correctly handles assets across input and output roots.""" + scope = f"multi-{uuid.uuid4().hex[:6]}" + input_fp = create_seed_file("input", scope, "input.bin") + create_seed_file("output", scope, "output.bin") + + trigger_sync_seed_assets(http, api_base) + assert len(find_asset(scope)) == 2 + + input_fp.unlink() + trigger_sync_seed_assets(http, api_base) + + remaining = find_asset(scope) + assert len(remaining) == 1 + assert remaining[0]["name"] == "output.bin" + + +@pytest.mark.parametrize("dirname", ["100%_done", "my_folder_name", "has spaces"]) +def test_special_chars_in_path_escaped_correctly( + dirname: str, + create_seed_file, + find_asset, + http: requests.Session, + api_base: str, + comfy_tmp_base_dir: Path, +): + """SQL LIKE wildcards (%, _) and spaces in paths don't cause false matches.""" + scope = f"special-{uuid.uuid4().hex[:6]}/{dirname}" + fp = create_seed_file("input", scope) + + trigger_sync_seed_assets(http, api_base) + trigger_sync_seed_assets(http, api_base) + + assert find_asset(scope.split("/")[0], fp.name), "Asset with special chars should survive" diff --git a/tests-unit/assets_test/test_tags.py b/tests-unit/assets_test/test_tags.py new file mode 100644 index 000000000..6b1047802 --- /dev/null +++ b/tests-unit/assets_test/test_tags.py @@ -0,0 +1,225 @@ +import json +import uuid + +import requests + + +def test_tags_present(http: requests.Session, api_base: str, seeded_asset: dict): + # Include zero-usage tags by default + r1 = http.get(api_base + "/api/tags", params={"limit": "50"}, timeout=120) + body1 = r1.json() + assert r1.status_code == 200 + names = [t["name"] for t in body1["tags"]] + # A few system tags from migration should exist: + assert "models" in names + assert "checkpoints" in names + + # Only used tags before we add anything new from this test cycle + r2 = http.get(api_base + "/api/tags", params={"include_zero": "false"}, timeout=120) + body2 = r2.json() + assert r2.status_code == 200 + # We already seeded one asset via fixture, so used tags must be non-empty + used_names = [t["name"] for t in body2["tags"]] + assert "models" in used_names + assert "checkpoints" in used_names + + # Prefix filter should refine the list + r3 = http.get(api_base + "/api/tags", params={"include_zero": "false", "prefix": "uni"}, timeout=120) + b3 = r3.json() + assert r3.status_code == 200 + names3 = [t["name"] for t in b3["tags"]] + assert "unit-tests" in names3 + assert "models" not in names3 # filtered out by prefix + + # Order by name ascending should be stable + r4 = http.get(api_base + "/api/tags", params={"include_zero": "false", "order": "name_asc"}, timeout=120) + b4 = r4.json() + assert r4.status_code == 200 + names4 = [t["name"] for t in b4["tags"]] + assert names4 == sorted(names4) + + +def test_tags_empty_usage(http: requests.Session, api_base: str, asset_factory, make_asset_bytes): + # Baseline: system tags exist when include_zero (default) is true + r1 = http.get(api_base + "/api/tags", params={"limit": "500"}, timeout=120) + body1 = r1.json() + assert r1.status_code == 200 + names = [t["name"] for t in body1["tags"]] + assert "models" in names and "checkpoints" in names + + # Create a short-lived asset under input with a unique custom tag + scope = f"tags-empty-usage-{uuid.uuid4().hex[:6]}" + custom_tag = f"temp-{uuid.uuid4().hex[:8]}" + name = "tag_seed.bin" + _asset = asset_factory( + name, + ["input", "unit-tests", scope, custom_tag], + {}, + make_asset_bytes(name, 512), + ) + + # While the asset exists, the custom tag must appear when include_zero=false + r2 = http.get( + api_base + "/api/tags", + params={"include_zero": "false", "prefix": custom_tag, "limit": "50"}, + timeout=120, + ) + body2 = r2.json() + assert r2.status_code == 200 + used_names = [t["name"] for t in body2["tags"]] + assert custom_tag in used_names + + # Delete the asset so the tag usage drops to zero + rd = http.delete(f"{api_base}/api/assets/{_asset['id']}", timeout=120) + assert rd.status_code == 204 + + # Now the custom tag must not be returned when include_zero=false + r3 = http.get( + api_base + "/api/tags", + params={"include_zero": "false", "prefix": custom_tag, "limit": "50"}, + timeout=120, + ) + body3 = r3.json() + assert r3.status_code == 200 + names_after = [t["name"] for t in body3["tags"]] + assert custom_tag not in names_after + assert not names_after # filtered view should be empty now + + +def test_add_and_remove_tags(http: requests.Session, api_base: str, seeded_asset: dict): + aid = seeded_asset["id"] + + # Add tags with duplicates and mixed case + payload_add = {"tags": ["NewTag", "unit-tests", "newtag", "BETA"]} + r1 = http.post(f"{api_base}/api/assets/{aid}/tags", json=payload_add, timeout=120) + b1 = r1.json() + assert r1.status_code == 200, b1 + # normalized, deduplicated; 'unit-tests' was already present from the seed + assert set(b1["added"]) == {"newtag", "beta"} + assert set(b1["already_present"]) == {"unit-tests"} + assert "newtag" in b1["total_tags"] and "beta" in b1["total_tags"] + + rg = http.get(f"{api_base}/api/assets/{aid}", timeout=120) + g = rg.json() + assert rg.status_code == 200 + tags_now = set(g["tags"]) + assert {"newtag", "beta"}.issubset(tags_now) + + # Remove a tag and a non-existent tag + payload_del = {"tags": ["newtag", "does-not-exist"]} + r2 = http.delete(f"{api_base}/api/assets/{aid}/tags", json=payload_del, timeout=120) + b2 = r2.json() + assert r2.status_code == 200 + assert set(b2["removed"]) == {"newtag"} + assert set(b2["not_present"]) == {"does-not-exist"} + + # Verify remaining tags after deletion + rg2 = http.get(f"{api_base}/api/assets/{aid}", timeout=120) + g2 = rg2.json() + assert rg2.status_code == 200 + tags_later = set(g2["tags"]) + assert "newtag" not in tags_later + assert "beta" in tags_later # still present + + +def test_tags_list_order_and_prefix(http: requests.Session, api_base: str, seeded_asset: dict): + aid = seeded_asset["id"] + h = seeded_asset["asset_hash"] + + # Add both tags to the seeded asset (usage: orderaaa=1, orderbbb=1) + r_add = http.post(f"{api_base}/api/assets/{aid}/tags", json={"tags": ["orderaaa", "orderbbb"]}, timeout=120) + add_body = r_add.json() + assert r_add.status_code == 200, add_body + + # Create another AssetInfo from the same content but tagged ONLY with 'orderbbb'. + payload = { + "hash": h, + "name": "order_only_bbb.safetensors", + "tags": ["input", "unit-tests", "orderbbb"], + "user_metadata": {}, + } + r_copy = http.post(f"{api_base}/api/assets/from-hash", json=payload, timeout=120) + copy_body = r_copy.json() + assert r_copy.status_code == 201, copy_body + + # 1) Default order (count_desc): 'orderbbb' should come before 'orderaaa' + # because it has higher usage (2 vs 1). + r1 = http.get(api_base + "/api/tags", params={"prefix": "order", "include_zero": "false"}, timeout=120) + b1 = r1.json() + assert r1.status_code == 200, b1 + names1 = [t["name"] for t in b1["tags"]] + counts1 = {t["name"]: t["count"] for t in b1["tags"]} + # Both must be present within the prefix subset + assert "orderaaa" in names1 and "orderbbb" in names1 + # Usage of 'orderbbb' must be >= 'orderaaa'; in our setup it's 2 vs 1 + assert counts1["orderbbb"] >= counts1["orderaaa"] + # And with count_desc, 'orderbbb' appears earlier than 'orderaaa' + assert names1.index("orderbbb") < names1.index("orderaaa") + + # 2) name_asc: lexical order should flip the relative order + r2 = http.get( + api_base + "/api/tags", + params={"prefix": "order", "include_zero": "false", "order": "name_asc"}, + timeout=120, + ) + b2 = r2.json() + assert r2.status_code == 200, b2 + names2 = [t["name"] for t in b2["tags"]] + assert "orderaaa" in names2 and "orderbbb" in names2 + assert names2.index("orderaaa") < names2.index("orderbbb") + + # 3) invalid limit rejected (existing negative case retained) + r3 = http.get(api_base + "/api/tags", params={"limit": "1001"}, timeout=120) + b3 = r3.json() + assert r3.status_code == 400 + assert b3["error"]["code"] == "INVALID_QUERY" + + +def test_tags_endpoints_invalid_bodies(http: requests.Session, api_base: str, seeded_asset: dict): + aid = seeded_asset["id"] + + # Add with empty list + r1 = http.post(f"{api_base}/api/assets/{aid}/tags", json={"tags": []}, timeout=120) + b1 = r1.json() + assert r1.status_code == 400 + assert b1["error"]["code"] == "INVALID_BODY" + + # Remove with wrong type + r2 = http.delete(f"{api_base}/api/assets/{aid}/tags", json={"tags": [123]}, timeout=120) + b2 = r2.json() + assert r2.status_code == 400 + assert b2["error"]["code"] == "INVALID_BODY" + + # metadata_filter provided as JSON array should be rejected (must be object) + r3 = http.get( + api_base + "/api/assets", + params={"metadata_filter": json.dumps([{"x": 1}])}, + timeout=120, + ) + b3 = r3.json() + assert r3.status_code == 400 + assert b3["error"]["code"] == "INVALID_QUERY" + + +def test_tags_prefix_treats_underscore_literal( + http, + api_base, + asset_factory, + make_asset_bytes, +): + """'prefix' for /api/tags must treat '_' literally, not as a wildcard.""" + base = f"pref_{uuid.uuid4().hex[:6]}" + tag_ok = f"{base}_ok" # should match prefix=f"{base}_" + tag_bad = f"{base}xok" # must NOT match if '_' is escaped + scope = f"tags-underscore-{uuid.uuid4().hex[:6]}" + + asset_factory("t1.bin", ["input", "unit-tests", scope, tag_ok], {}, make_asset_bytes("t1", 512)) + asset_factory("t2.bin", ["input", "unit-tests", scope, tag_bad], {}, make_asset_bytes("t2", 512)) + + r = http.get(api_base + "/api/tags", params={"include_zero": "false", "prefix": f"{base}_"}, timeout=120) + body = r.json() + assert r.status_code == 200, body + names = [t["name"] for t in body["tags"]] + assert tag_ok in names, f"Expected {tag_ok} to be returned for prefix '{base}_'" + assert tag_bad not in names, f"'{tag_bad}' must not match — '_' is not a wildcard" + assert body["total"] == 1 diff --git a/tests-unit/assets_test/test_uploads.py b/tests-unit/assets_test/test_uploads.py new file mode 100644 index 000000000..137d7391a --- /dev/null +++ b/tests-unit/assets_test/test_uploads.py @@ -0,0 +1,281 @@ +import json +import uuid +from concurrent.futures import ThreadPoolExecutor + +import requests +import pytest + + +def test_upload_ok_duplicate_reference(http: requests.Session, api_base: str, make_asset_bytes): + name = "dup_a.safetensors" + tags = ["models", "checkpoints", "unit-tests", "alpha"] + meta = {"purpose": "dup"} + data = make_asset_bytes(name) + files = {"file": (name, data, "application/octet-stream")} + form = {"tags": json.dumps(tags), "name": name, "user_metadata": json.dumps(meta)} + r1 = http.post(api_base + "/api/assets", data=form, files=files, timeout=120) + a1 = r1.json() + assert r1.status_code == 201, a1 + assert a1["created_new"] is True + + # Second upload with the same data and name should return created_new == False and the same asset + files = {"file": (name, data, "application/octet-stream")} + form = {"tags": json.dumps(tags), "name": name, "user_metadata": json.dumps(meta)} + r2 = http.post(api_base + "/api/assets", data=form, files=files, timeout=120) + a2 = r2.json() + assert r2.status_code == 200, a2 + assert a2["created_new"] is False + assert a2["asset_hash"] == a1["asset_hash"] + assert a2["id"] == a1["id"] # old reference + + # Third upload with the same data but new name should return created_new == False and the new AssetReference + files = {"file": (name, data, "application/octet-stream")} + form = {"tags": json.dumps(tags), "name": name + "_d", "user_metadata": json.dumps(meta)} + r2 = http.post(api_base + "/api/assets", data=form, files=files, timeout=120) + a3 = r2.json() + assert r2.status_code == 200, a3 + assert a3["created_new"] is False + assert a3["asset_hash"] == a1["asset_hash"] + assert a3["id"] != a1["id"] # old reference + + +def test_upload_fastpath_from_existing_hash_no_file(http: requests.Session, api_base: str): + # Seed a small file first + name = "fastpath_seed.safetensors" + tags = ["models", "checkpoints", "unit-tests"] + meta = {} + files = {"file": (name, b"B" * 1024, "application/octet-stream")} + form = {"tags": json.dumps(tags), "name": name, "user_metadata": json.dumps(meta)} + r1 = http.post(api_base + "/api/assets", data=form, files=files, timeout=120) + b1 = r1.json() + assert r1.status_code == 201, b1 + h = b1["asset_hash"] + + # Now POST /api/assets with only hash and no file + files = [ + ("hash", (None, h)), + ("tags", (None, json.dumps(tags))), + ("name", (None, "fastpath_copy.safetensors")), + ("user_metadata", (None, json.dumps({"purpose": "copy"}))), + ] + r2 = http.post(api_base + "/api/assets", files=files, timeout=120) + b2 = r2.json() + assert r2.status_code == 200, b2 # fast path returns 200 with created_new == False + assert b2["created_new"] is False + assert b2["asset_hash"] == h + + +def test_upload_fastpath_with_known_hash_and_file( + http: requests.Session, api_base: str +): + # Seed + files = {"file": ("seed.safetensors", b"C" * 128, "application/octet-stream")} + form = {"tags": json.dumps(["models", "checkpoints", "unit-tests", "fp"]), "name": "seed.safetensors", "user_metadata": json.dumps({})} + r1 = http.post(api_base + "/api/assets", data=form, files=files, timeout=120) + b1 = r1.json() + assert r1.status_code == 201, b1 + h = b1["asset_hash"] + + # Send both file and hash of existing content -> server must drain file and create from hash (200) + files = {"file": ("ignored.bin", b"ignored" * 10, "application/octet-stream")} + form = {"hash": h, "tags": json.dumps(["models", "checkpoints", "unit-tests", "fp"]), "name": "copy_from_hash.safetensors", "user_metadata": json.dumps({})} + r2 = http.post(api_base + "/api/assets", data=form, files=files, timeout=120) + b2 = r2.json() + assert r2.status_code == 200, b2 + assert b2["created_new"] is False + assert b2["asset_hash"] == h + + +def test_upload_multiple_tags_fields_are_merged(http: requests.Session, api_base: str): + data = [ + ("tags", "models,checkpoints"), + ("tags", json.dumps(["unit-tests", "alpha"])), + ("name", "merge.safetensors"), + ("user_metadata", json.dumps({"u": 1})), + ] + files = {"file": ("merge.safetensors", b"B" * 256, "application/octet-stream")} + r1 = http.post(api_base + "/api/assets", data=data, files=files, timeout=120) + created = r1.json() + assert r1.status_code in (200, 201), created + aid = created["id"] + + # Verify all tags are present on the resource + rg = http.get(f"{api_base}/api/assets/{aid}", timeout=120) + detail = rg.json() + assert rg.status_code == 200, detail + tags = set(detail["tags"]) + assert {"models", "checkpoints", "unit-tests", "alpha"}.issubset(tags) + + +@pytest.mark.parametrize("root", ["input", "output"]) +def test_concurrent_upload_identical_bytes_different_names( + root: str, + http: requests.Session, + api_base: str, + make_asset_bytes, +): + """ + Two concurrent uploads of identical bytes but different names. + Expect a single Asset (same hash), two AssetInfo rows, and exactly one created_new=True. + """ + scope = f"concupload-{uuid.uuid4().hex[:6]}" + name1, name2 = "cu_a.bin", "cu_b.bin" + data = make_asset_bytes("concurrent", 4096) + tags = [root, "unit-tests", scope] + + def _do_upload(args): + url, form_data, files_data = args + with requests.Session() as s: + return s.post(url, data=form_data, files=files_data, timeout=120) + + url = api_base + "/api/assets" + form1 = {"tags": json.dumps(tags), "name": name1, "user_metadata": json.dumps({})} + files1 = {"file": (name1, data, "application/octet-stream")} + form2 = {"tags": json.dumps(tags), "name": name2, "user_metadata": json.dumps({})} + files2 = {"file": (name2, data, "application/octet-stream")} + + with ThreadPoolExecutor(max_workers=2) as executor: + futures = list(executor.map(_do_upload, [(url, form1, files1), (url, form2, files2)])) + r1, r2 = futures + + b1, b2 = r1.json(), r2.json() + assert r1.status_code in (200, 201), b1 + assert r2.status_code in (200, 201), b2 + assert b1["asset_hash"] == b2["asset_hash"] + assert b1["id"] != b2["id"] + + created_flags = sorted([bool(b1.get("created_new")), bool(b2.get("created_new"))]) + assert created_flags == [False, True] + + rl = http.get( + api_base + "/api/assets", + params={"include_tags": f"unit-tests,{scope}", "sort": "name"}, + timeout=120, + ) + bl = rl.json() + assert rl.status_code == 200, bl + names = [a["name"] for a in bl.get("assets", [])] + assert set([name1, name2]).issubset(names) + + +def test_create_from_hash_endpoint_404(http: requests.Session, api_base: str): + payload = { + "hash": "blake3:" + "0" * 64, + "name": "nonexistent.bin", + "tags": ["models", "checkpoints", "unit-tests"], + } + r = http.post(api_base + "/api/assets/from-hash", json=payload, timeout=120) + body = r.json() + assert r.status_code == 404 + assert body["error"]["code"] == "ASSET_NOT_FOUND" + + +def test_upload_zero_byte_rejected(http: requests.Session, api_base: str): + files = {"file": ("empty.safetensors", b"", "application/octet-stream")} + form = {"tags": json.dumps(["models", "checkpoints", "unit-tests", "edge"]), "name": "empty.safetensors", "user_metadata": json.dumps({})} + r = http.post(api_base + "/api/assets", data=form, files=files, timeout=120) + body = r.json() + assert r.status_code == 400 + assert body["error"]["code"] == "EMPTY_UPLOAD" + + +def test_upload_invalid_root_tag_rejected(http: requests.Session, api_base: str): + files = {"file": ("badroot.bin", b"A" * 64, "application/octet-stream")} + form = {"tags": json.dumps(["not-a-root", "whatever"]), "name": "badroot.bin", "user_metadata": json.dumps({})} + r = http.post(api_base + "/api/assets", data=form, files=files, timeout=120) + body = r.json() + assert r.status_code == 400 + assert body["error"]["code"] == "INVALID_BODY" + + +def test_upload_user_metadata_must_be_json(http: requests.Session, api_base: str): + files = {"file": ("badmeta.bin", b"A" * 128, "application/octet-stream")} + form = {"tags": json.dumps(["models", "checkpoints", "unit-tests", "edge"]), "name": "badmeta.bin", "user_metadata": "{not json}"} + r = http.post(api_base + "/api/assets", data=form, files=files, timeout=120) + body = r.json() + assert r.status_code == 400 + assert body["error"]["code"] == "INVALID_BODY" + + +def test_upload_requires_multipart(http: requests.Session, api_base: str): + r = http.post(api_base + "/api/assets", json={"foo": "bar"}, timeout=120) + body = r.json() + assert r.status_code == 415 + assert body["error"]["code"] == "UNSUPPORTED_MEDIA_TYPE" + + +def test_upload_missing_file_and_hash(http: requests.Session, api_base: str): + files = [ + ("tags", (None, json.dumps(["models", "checkpoints", "unit-tests"]))), + ("name", (None, "x.safetensors")), + ] + r = http.post(api_base + "/api/assets", files=files, timeout=120) + body = r.json() + assert r.status_code == 400 + assert body["error"]["code"] == "MISSING_FILE" + + +def test_upload_models_unknown_category(http: requests.Session, api_base: str): + files = {"file": ("m.safetensors", b"A" * 128, "application/octet-stream")} + form = {"tags": json.dumps(["models", "no_such_category", "unit-tests"]), "name": "m.safetensors"} + r = http.post(api_base + "/api/assets", data=form, files=files, timeout=120) + body = r.json() + assert r.status_code == 400 + assert body["error"]["code"] == "INVALID_BODY" + assert body["error"]["message"].startswith("unknown models category") + + +def test_upload_models_requires_category(http: requests.Session, api_base: str): + files = {"file": ("nocat.safetensors", b"A" * 64, "application/octet-stream")} + form = {"tags": json.dumps(["models"]), "name": "nocat.safetensors", "user_metadata": json.dumps({})} + r = http.post(api_base + "/api/assets", data=form, files=files, timeout=120) + body = r.json() + assert r.status_code == 400 + assert body["error"]["code"] == "INVALID_BODY" + + +def test_upload_tags_traversal_guard(http: requests.Session, api_base: str): + files = {"file": ("evil.safetensors", b"A" * 256, "application/octet-stream")} + form = {"tags": json.dumps(["models", "checkpoints", "unit-tests", "..", "zzz"]), "name": "evil.safetensors"} + r = http.post(api_base + "/api/assets", data=form, files=files, timeout=120) + body = r.json() + assert r.status_code == 400 + assert body["error"]["code"] in ("BAD_REQUEST", "INVALID_BODY") + + +@pytest.mark.parametrize("root", ["input", "output"]) +def test_duplicate_upload_same_display_name_does_not_clobber( + root: str, + http: requests.Session, + api_base: str, + asset_factory, + make_asset_bytes, +): + """ + Two uploads use the same tags and the same display name but different bytes. + With hash-based filenames, they must NOT overwrite each other. Both assets + remain accessible and serve their original content. + """ + scope = f"dup-path-{uuid.uuid4().hex[:6]}" + display_name = "same_display.bin" + + d1 = make_asset_bytes(scope + "-v1", 1536) + d2 = make_asset_bytes(scope + "-v2", 2048) + tags = [root, "unit-tests", scope] + + first = asset_factory(display_name, tags, {}, d1) + second = asset_factory(display_name, tags, {}, d2) + + assert first["id"] != second["id"] + assert first["asset_hash"] != second["asset_hash"] # different content + assert first["name"] == second["name"] == display_name + + # Both must be independently retrievable + r1 = http.get(f"{api_base}/api/assets/{first['id']}/content", timeout=120) + b1 = r1.content + assert r1.status_code == 200 + assert b1 == d1 + r2 = http.get(f"{api_base}/api/assets/{second['id']}/content", timeout=120) + b2 = r2.content + assert r2.status_code == 200 + assert b2 == d2 diff --git a/tests-unit/requirements.txt b/tests-unit/requirements.txt index 3a6790ee0..2355b8000 100644 --- a/tests-unit/requirements.txt +++ b/tests-unit/requirements.txt @@ -2,3 +2,4 @@ pytest>=7.8.0 pytest-aiohttp pytest-asyncio websocket-client +blake3 From aa6f7a83bb313f121ee032a8b21aa72ea32cfee1 Mon Sep 17 00:00:00 2001 From: Jedrzej Kosinski Date: Sat, 31 Jan 2026 17:05:11 -0800 Subject: [PATCH 542/647] Send is_input_list on v1 and v3 schema to frontend (#12188) --- comfy_api/latest/_io.py | 2 ++ server.py | 1 + 2 files changed, 3 insertions(+) diff --git a/comfy_api/latest/_io.py b/comfy_api/latest/_io.py index 78f77d4b2..eeea9781a 100644 --- a/comfy_api/latest/_io.py +++ b/comfy_api/latest/_io.py @@ -1248,6 +1248,7 @@ class Hidden(str, Enum): class NodeInfoV1: input: dict=None input_order: dict[str, list[str]]=None + is_input_list: bool=None output: list[str]=None output_is_list: list[bool]=None output_name: list[str]=None @@ -1474,6 +1475,7 @@ class Schema: info = NodeInfoV1( input=input, input_order={key: list(value.keys()) for (key, value) in input.items()}, + is_input_list=self.is_input_list, output=output, output_is_list=output_is_list, output_name=output_name, diff --git a/server.py b/server.py index 2aee5cc06..2300393b2 100644 --- a/server.py +++ b/server.py @@ -656,6 +656,7 @@ class PromptServer(): info = {} info['input'] = obj_class.INPUT_TYPES() info['input_order'] = {key: list(value.keys()) for (key, value) in obj_class.INPUT_TYPES().items()} + info['is_input_list'] = getattr(obj_class, "INPUT_IS_LIST", False) info['output'] = obj_class.RETURN_TYPES info['output_is_list'] = obj_class.OUTPUT_IS_LIST if hasattr(obj_class, 'OUTPUT_IS_LIST') else [False] * len(obj_class.RETURN_TYPES) info['output_name'] = obj_class.RETURN_NAMES if hasattr(obj_class, 'RETURN_NAMES') else info['output'] From 873de5f37a1f4c39c0203b2989e6736b0cb12b98 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sat, 31 Jan 2026 18:11:11 -0800 Subject: [PATCH 543/647] KV cache implementation for using llama models for text generation. (#12195) --- comfy/text_encoders/llama.py | 91 +++++++++++++++++++++++++++++------- 1 file changed, 74 insertions(+), 17 deletions(-) diff --git a/comfy/text_encoders/llama.py b/comfy/text_encoders/llama.py index 3080a3e09..68ac1e804 100644 --- a/comfy/text_encoders/llama.py +++ b/comfy/text_encoders/llama.py @@ -1,7 +1,7 @@ import torch import torch.nn as nn from dataclasses import dataclass -from typing import Optional, Any +from typing import Optional, Any, Tuple import math from comfy.ldm.modules.attention import optimized_attention_for_device @@ -32,6 +32,7 @@ class Llama2Config: k_norm = None rope_scale = None final_norm: bool = True + lm_head: bool = False @dataclass class Mistral3Small24BConfig: @@ -54,6 +55,7 @@ class Mistral3Small24BConfig: k_norm = None rope_scale = None final_norm: bool = True + lm_head: bool = False @dataclass class Qwen25_3BConfig: @@ -76,6 +78,7 @@ class Qwen25_3BConfig: k_norm = None rope_scale = None final_norm: bool = True + lm_head: bool = False @dataclass class Qwen3_06BConfig: @@ -98,6 +101,7 @@ class Qwen3_06BConfig: k_norm = "gemma3" rope_scale = None final_norm: bool = True + lm_head: bool = False @dataclass class Qwen3_4BConfig: @@ -120,6 +124,7 @@ class Qwen3_4BConfig: k_norm = "gemma3" rope_scale = None final_norm: bool = True + lm_head: bool = False @dataclass class Qwen3_8BConfig: @@ -142,6 +147,7 @@ class Qwen3_8BConfig: k_norm = "gemma3" rope_scale = None final_norm: bool = True + lm_head: bool = False @dataclass class Ovis25_2BConfig: @@ -164,6 +170,7 @@ class Ovis25_2BConfig: k_norm = "gemma3" rope_scale = None final_norm: bool = True + lm_head: bool = False @dataclass class Qwen25_7BVLI_Config: @@ -186,6 +193,7 @@ class Qwen25_7BVLI_Config: k_norm = None rope_scale = None final_norm: bool = True + lm_head: bool = False @dataclass class Gemma2_2B_Config: @@ -209,6 +217,7 @@ class Gemma2_2B_Config: sliding_attention = None rope_scale = None final_norm: bool = True + lm_head: bool = False @dataclass class Gemma3_4B_Config: @@ -232,6 +241,7 @@ class Gemma3_4B_Config: sliding_attention = [1024, 1024, 1024, 1024, 1024, False] rope_scale = [8.0, 1.0] final_norm: bool = True + lm_head: bool = False @dataclass class Gemma3_12B_Config: @@ -255,6 +265,7 @@ class Gemma3_12B_Config: sliding_attention = [1024, 1024, 1024, 1024, 1024, False] rope_scale = [8.0, 1.0] final_norm: bool = True + lm_head: bool = False vision_config = {"num_channels": 3, "hidden_act": "gelu_pytorch_tanh", "hidden_size": 1152, "image_size": 896, "intermediate_size": 4304, "model_type": "siglip_vision_model", "num_attention_heads": 16, "num_hidden_layers": 27, "patch_size": 14} mm_tokens_per_image = 256 @@ -356,6 +367,7 @@ class Attention(nn.Module): attention_mask: Optional[torch.Tensor] = None, freqs_cis: Optional[torch.Tensor] = None, optimized_attention=None, + past_key_value: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, ): batch_size, seq_length, _ = hidden_states.shape xq = self.q_proj(hidden_states) @@ -373,11 +385,30 @@ class Attention(nn.Module): xq, xk = apply_rope(xq, xk, freqs_cis=freqs_cis) + present_key_value = None + if past_key_value is not None: + index = 0 + num_tokens = xk.shape[2] + if len(past_key_value) > 0: + past_key, past_value, index = past_key_value + if past_key.shape[2] >= (index + num_tokens): + past_key[:, :, index:index + xk.shape[2]] = xk + past_value[:, :, index:index + xv.shape[2]] = xv + xk = past_key[:, :, :index + xk.shape[2]] + xv = past_value[:, :, :index + xv.shape[2]] + present_key_value = (past_key, past_value, index + num_tokens) + else: + xk = torch.cat((past_key[:, :, :index], xk), dim=2) + xv = torch.cat((past_value[:, :, :index], xv), dim=2) + present_key_value = (xk, xv, index + num_tokens) + else: + present_key_value = (xk, xv, index + num_tokens) + xk = xk.repeat_interleave(self.num_heads // self.num_kv_heads, dim=1) xv = xv.repeat_interleave(self.num_heads // self.num_kv_heads, dim=1) output = optimized_attention(xq, xk, xv, self.num_heads, mask=attention_mask, skip_reshape=True) - return self.o_proj(output) + return self.o_proj(output), present_key_value class MLP(nn.Module): def __init__(self, config: Llama2Config, device=None, dtype=None, ops: Any = None): @@ -408,15 +439,17 @@ class TransformerBlock(nn.Module): attention_mask: Optional[torch.Tensor] = None, freqs_cis: Optional[torch.Tensor] = None, optimized_attention=None, + past_key_value: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, ): # Self Attention residual = x x = self.input_layernorm(x) - x = self.self_attn( + x, present_key_value = self.self_attn( hidden_states=x, attention_mask=attention_mask, freqs_cis=freqs_cis, optimized_attention=optimized_attention, + past_key_value=past_key_value, ) x = residual + x @@ -426,7 +459,7 @@ class TransformerBlock(nn.Module): x = self.mlp(x) x = residual + x - return x + return x, present_key_value class TransformerBlockGemma2(nn.Module): def __init__(self, config: Llama2Config, index, device=None, dtype=None, ops: Any = None): @@ -451,6 +484,7 @@ class TransformerBlockGemma2(nn.Module): attention_mask: Optional[torch.Tensor] = None, freqs_cis: Optional[torch.Tensor] = None, optimized_attention=None, + past_key_value: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, ): if self.transformer_type == 'gemma3': if self.sliding_attention: @@ -468,11 +502,12 @@ class TransformerBlockGemma2(nn.Module): # Self Attention residual = x x = self.input_layernorm(x) - x = self.self_attn( + x, present_key_value = self.self_attn( hidden_states=x, attention_mask=attention_mask, freqs_cis=freqs_cis, optimized_attention=optimized_attention, + past_key_value=past_key_value, ) x = self.post_attention_layernorm(x) @@ -485,7 +520,7 @@ class TransformerBlockGemma2(nn.Module): x = self.post_feedforward_layernorm(x) x = residual + x - return x + return x, present_key_value class Llama2_(nn.Module): def __init__(self, config, device=None, dtype=None, ops=None): @@ -516,9 +551,10 @@ class Llama2_(nn.Module): else: self.norm = None - # self.lm_head = ops.Linear(config.hidden_size, config.vocab_size, bias=False, device=device, dtype=dtype) + if config.lm_head: + self.lm_head = ops.Linear(config.hidden_size, config.vocab_size, bias=False, device=device, dtype=dtype) - def forward(self, x, attention_mask=None, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None, position_ids=None, embeds_info=[]): + def forward(self, x, attention_mask=None, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None, position_ids=None, embeds_info=[], past_key_values=None): if embeds is not None: x = embeds else: @@ -527,8 +563,13 @@ class Llama2_(nn.Module): if self.normalize_in: x *= self.config.hidden_size ** 0.5 + seq_len = x.shape[1] + past_len = 0 + if past_key_values is not None and len(past_key_values) > 0: + past_len = past_key_values[0][2] + if position_ids is None: - position_ids = torch.arange(0, x.shape[1], device=x.device).unsqueeze(0) + position_ids = torch.arange(past_len, past_len + seq_len, device=x.device).unsqueeze(0) freqs_cis = precompute_freqs_cis(self.config.head_dim, position_ids, @@ -539,14 +580,16 @@ class Llama2_(nn.Module): mask = None if attention_mask is not None: - mask = 1.0 - attention_mask.to(x.dtype).reshape((attention_mask.shape[0], 1, -1, attention_mask.shape[-1])).expand(attention_mask.shape[0], 1, attention_mask.shape[-1], attention_mask.shape[-1]) + mask = 1.0 - attention_mask.to(x.dtype).reshape((attention_mask.shape[0], 1, -1, attention_mask.shape[-1])).expand(attention_mask.shape[0], 1, seq_len, attention_mask.shape[-1]) mask = mask.masked_fill(mask.to(torch.bool), float("-inf")) - causal_mask = torch.empty(x.shape[1], x.shape[1], dtype=x.dtype, device=x.device).fill_(float("-inf")).triu_(1) - if mask is not None: - mask += causal_mask - else: - mask = causal_mask + if seq_len > 1: + causal_mask = torch.empty(past_len + seq_len, past_len + seq_len, dtype=x.dtype, device=x.device).fill_(float("-inf")).triu_(1) + if mask is not None: + mask += causal_mask + else: + mask = causal_mask + optimized_attention = optimized_attention_for_device(x.device, mask=mask is not None, small_input=True) intermediate = None @@ -562,16 +605,27 @@ class Llama2_(nn.Module): elif intermediate_output < 0: intermediate_output = len(self.layers) + intermediate_output + next_key_values = [] for i, layer in enumerate(self.layers): if all_intermediate is not None: if only_layers is None or (i in only_layers): all_intermediate.append(x.unsqueeze(1).clone()) - x = layer( + + past_kv = None + if past_key_values is not None: + past_kv = past_key_values[i] if len(past_key_values) > 0 else [] + + x, current_kv = layer( x=x, attention_mask=mask, freqs_cis=freqs_cis, optimized_attention=optimized_attention, + past_key_value=past_kv, ) + + if current_kv is not None: + next_key_values.append(current_kv) + if i == intermediate_output: intermediate = x.clone() @@ -588,7 +642,10 @@ class Llama2_(nn.Module): if intermediate is not None and final_layer_norm_intermediate and self.norm is not None: intermediate = self.norm(intermediate) - return x, intermediate + if len(next_key_values) > 0: + return x, intermediate, next_key_values + else: + return x, intermediate class Gemma3MultiModalProjector(torch.nn.Module): From f8acd9c402f41efc2c748503e4c5d9e0027965ab Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Sat, 31 Jan 2026 22:01:11 -0800 Subject: [PATCH 544/647] Reduce RAM usage, fix VRAM OOMs, and fix Windows shared memory spilling with adaptive model loading (#11845) --- comfy/audio_encoders/audio_encoders.py | 4 +- comfy/cli_args.py | 4 + comfy/clip_vision.py | 4 +- comfy/controlnet.py | 2 +- comfy/k_diffusion/sampling.py | 33 ++- comfy/ldm/hunyuan_video/upsampler.py | 4 +- comfy/memory_management.py | 81 +++++++ comfy/model_base.py | 15 +- comfy/model_management.py | 170 ++++++++++++-- comfy/model_patcher.py | 313 +++++++++++++++++++++++-- comfy/ops.py | 221 +++++++++++++++-- comfy/pinned_memory.py | 30 +++ comfy/samplers.py | 3 +- comfy/sd.py | 59 +++-- comfy/sd1_clip.py | 2 +- comfy/text_encoders/lt.py | 2 +- comfy/utils.py | 80 ++++++- comfy/windows.py | 52 ++++ comfy_extras/nodes_model_patch.py | 6 +- cuda_malloc.py | 12 +- execution.py | 16 +- main.py | 30 ++- requirements.txt | 1 + 23 files changed, 1030 insertions(+), 114 deletions(-) create mode 100644 comfy/memory_management.py create mode 100644 comfy/pinned_memory.py create mode 100644 comfy/windows.py diff --git a/comfy/audio_encoders/audio_encoders.py b/comfy/audio_encoders/audio_encoders.py index 46ef21c95..16998af94 100644 --- a/comfy/audio_encoders/audio_encoders.py +++ b/comfy/audio_encoders/audio_encoders.py @@ -25,11 +25,11 @@ class AudioEncoderModel(): elif model_type == "whisper3": self.model = WhisperLargeV3(**model_config) self.model.eval() - self.patcher = comfy.model_patcher.ModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device) + self.patcher = comfy.model_patcher.CoreModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device) self.model_sample_rate = 16000 def load_sd(self, sd): - return self.model.load_state_dict(sd, strict=False) + return self.model.load_state_dict(sd, strict=False, assign=self.patcher.is_dynamic()) def get_sd(self): return self.model.state_dict() diff --git a/comfy/cli_args.py b/comfy/cli_args.py index 1716c3de7..63daca861 100644 --- a/comfy/cli_args.py +++ b/comfy/cli_args.py @@ -159,6 +159,7 @@ class PerformanceFeature(enum.Enum): Fp8MatrixMultiplication = "fp8_matrix_mult" CublasOps = "cublas_ops" AutoTune = "autotune" + DynamicVRAM = "dynamic_vram" parser.add_argument("--fast", nargs="*", type=PerformanceFeature, help="Enable some untested and potentially quality deteriorating optimizations. This is used to test new features so using it might crash your comfyui. --fast with no arguments enables everything. You can pass a list specific optimizations if you only want to enable specific ones. Current valid optimizations: {}".format(" ".join(map(lambda c: c.value, PerformanceFeature)))) @@ -257,3 +258,6 @@ elif args.fast == []: # '--fast' is provided with a list of performance features, use that list else: args.fast = set(args.fast) + +def enables_dynamic_vram(): + return PerformanceFeature.DynamicVRAM in args.fast and not args.highvram and not args.gpu_only diff --git a/comfy/clip_vision.py b/comfy/clip_vision.py index b28bf636c..1691fca81 100644 --- a/comfy/clip_vision.py +++ b/comfy/clip_vision.py @@ -47,10 +47,10 @@ class ClipVisionModel(): self.model = model_class(config, self.dtype, offload_device, comfy.ops.manual_cast) self.model.eval() - self.patcher = comfy.model_patcher.ModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device) + self.patcher = comfy.model_patcher.CoreModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device) def load_sd(self, sd): - return self.model.load_state_dict(sd, strict=False) + return self.model.load_state_dict(sd, strict=False, assign=self.patcher.is_dynamic()) def get_sd(self): return self.model.state_dict() diff --git a/comfy/controlnet.py b/comfy/controlnet.py index 0b5e30f52..9e1e704e0 100644 --- a/comfy/controlnet.py +++ b/comfy/controlnet.py @@ -203,7 +203,7 @@ class ControlNet(ControlBase): self.control_model = control_model self.load_device = load_device if control_model is not None: - self.control_model_wrapped = comfy.model_patcher.ModelPatcher(self.control_model, load_device=load_device, offload_device=comfy.model_management.unet_offload_device()) + self.control_model_wrapped = comfy.model_patcher.CoreModelPatcher(self.control_model, load_device=load_device, offload_device=comfy.model_management.unet_offload_device()) self.compression_ratio = compression_ratio self.global_average_pooling = global_average_pooling diff --git a/comfy/k_diffusion/sampling.py b/comfy/k_diffusion/sampling.py index 0949dee44..c0c51d51a 100644 --- a/comfy/k_diffusion/sampling.py +++ b/comfy/k_diffusion/sampling.py @@ -1,11 +1,12 @@ import math +import time from functools import partial from scipy import integrate import torch from torch import nn import torchsde -from tqdm.auto import trange, tqdm +from tqdm.auto import trange as trange_, tqdm from . import utils from . import deis @@ -13,6 +14,36 @@ from . import sa_solver import comfy.model_patcher import comfy.model_sampling +import comfy.memory_management + + +def trange(*args, **kwargs): + if comfy.memory_management.aimdo_allocator is None: + return trange_(*args, **kwargs) + + pbar = trange_(*args, **kwargs, smoothing=1.0) + pbar._i = 0 + pbar.set_postfix_str(" Model Initializing ... ") + + _update = pbar.update + + def warmup_update(n=1): + pbar._i += 1 + if pbar._i == 1: + pbar.i1_time = time.time() + pbar.set_postfix_str(" Model Initialization complete! ") + elif pbar._i == 2: + #bring forward the effective start time based the the diff between first and second iteration + #to attempt to remove load overhead from the final step rate estimate. + pbar.start_t = pbar.i1_time - (time.time() - pbar.i1_time) + pbar.set_postfix_str("") + + _update(n) + + pbar.update = warmup_update + return pbar + + def append_zero(x): return torch.cat([x, x.new_zeros([1])]) diff --git a/comfy/ldm/hunyuan_video/upsampler.py b/comfy/ldm/hunyuan_video/upsampler.py index 51b6d1da8..1f68144e2 100644 --- a/comfy/ldm/hunyuan_video/upsampler.py +++ b/comfy/ldm/hunyuan_video/upsampler.py @@ -109,10 +109,10 @@ class HunyuanVideo15SRModel(): self.model_class = UPSAMPLERS.get(model_type) self.model = self.model_class(**config).eval() - self.patcher = comfy.model_patcher.ModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device) + self.patcher = comfy.model_patcher.CoreModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device) def load_sd(self, sd): - return self.model.load_state_dict(sd, strict=True) + return self.model.load_state_dict(sd, strict=True, assign=self.patcher.is_dynamic()) def get_sd(self): return self.model.state_dict() diff --git a/comfy/memory_management.py b/comfy/memory_management.py new file mode 100644 index 000000000..858bd4cc7 --- /dev/null +++ b/comfy/memory_management.py @@ -0,0 +1,81 @@ +import math +import torch +from typing import NamedTuple + +from comfy.quant_ops import QuantizedTensor + +class TensorGeometry(NamedTuple): + shape: any + dtype: torch.dtype + + def element_size(self): + info = torch.finfo(self.dtype) if self.dtype.is_floating_point else torch.iinfo(self.dtype) + return info.bits // 8 + + def numel(self): + return math.prod(self.shape) + +def tensors_to_geometries(tensors, dtype=None): + geometries = [] + for t in tensors: + if t is None or isinstance(t, QuantizedTensor): + geometries.append(t) + continue + tdtype = t.dtype + if hasattr(t, "_model_dtype"): + tdtype = t._model_dtype + if dtype is not None: + tdtype = dtype + geometries.append(TensorGeometry(shape=t.shape, dtype=tdtype)) + return geometries + +def vram_aligned_size(tensor): + if isinstance(tensor, list): + return sum([vram_aligned_size(t) for t in tensor]) + + if isinstance(tensor, QuantizedTensor): + inner_tensors, _ = tensor.__tensor_flatten__() + return vram_aligned_size([ getattr(tensor, attr) for attr in inner_tensors ]) + + if tensor is None: + return 0 + + size = tensor.numel() * tensor.element_size() + aligment_req = 1024 + return (size + aligment_req - 1) // aligment_req * aligment_req + +def interpret_gathered_like(tensors, gathered): + offset = 0 + dest_views = [] + + if gathered.dim() != 1 or gathered.element_size() != 1: + raise ValueError(f"Buffer must be 1D and single-byte (got {gathered.dim()}D {gathered.dtype})") + + for tensor in tensors: + + if tensor is None: + dest_views.append(None) + continue + + if isinstance(tensor, QuantizedTensor): + inner_tensors, qt_ctx = tensor.__tensor_flatten__() + templates = { attr: getattr(tensor, attr) for attr in inner_tensors } + else: + templates = { "data": tensor } + + actuals = {} + for attr, template in templates.items(): + size = template.numel() * template.element_size() + if offset + size > gathered.numel(): + raise ValueError(f"Buffer too small: needs {offset + size} bytes, but only has {gathered.numel()}. ") + actuals[attr] = gathered[offset:offset+size].view(dtype=template.dtype).view(template.shape) + offset += vram_aligned_size(template) + + if isinstance(tensor, QuantizedTensor): + dest_views.append(QuantizedTensor.__tensor_unflatten__(actuals, qt_ctx, 0, 0)) + else: + dest_views.append(actuals["data"]) + + return dest_views + +aimdo_allocator = None diff --git a/comfy/model_base.py b/comfy/model_base.py index 66e52864d..85acdb66a 100644 --- a/comfy/model_base.py +++ b/comfy/model_base.py @@ -149,6 +149,8 @@ class BaseModel(torch.nn.Module): self.model_type = model_type self.model_sampling = model_sampling(model_config, model_type) + comfy.model_management.archive_model_dtypes(self.diffusion_model) + self.adm_channels = unet_config.get("adm_in_channels", None) if self.adm_channels is None: self.adm_channels = 0 @@ -299,7 +301,7 @@ class BaseModel(torch.nn.Module): return out - def load_model_weights(self, sd, unet_prefix=""): + def load_model_weights(self, sd, unet_prefix="", assign=False): to_load = {} keys = list(sd.keys()) for k in keys: @@ -307,7 +309,7 @@ class BaseModel(torch.nn.Module): to_load[k[len(unet_prefix):]] = sd.pop(k) to_load = self.model_config.process_unet_state_dict(to_load) - m, u = self.diffusion_model.load_state_dict(to_load, strict=False) + m, u = self.diffusion_model.load_state_dict(to_load, strict=False, assign=assign) if len(m) > 0: logging.warning("unet missing: {}".format(m)) @@ -322,7 +324,7 @@ class BaseModel(torch.nn.Module): def process_latent_out(self, latent): return self.latent_format.process_out(latent) - def state_dict_for_saving(self, clip_state_dict=None, vae_state_dict=None, clip_vision_state_dict=None): + def state_dict_for_saving(self, unet_state_dict, clip_state_dict=None, vae_state_dict=None, clip_vision_state_dict=None): extra_sds = [] if clip_state_dict is not None: extra_sds.append(self.model_config.process_clip_state_dict_for_saving(clip_state_dict)) @@ -330,10 +332,7 @@ class BaseModel(torch.nn.Module): extra_sds.append(self.model_config.process_vae_state_dict_for_saving(vae_state_dict)) if clip_vision_state_dict is not None: extra_sds.append(self.model_config.process_clip_vision_state_dict_for_saving(clip_vision_state_dict)) - - unet_state_dict = self.diffusion_model.state_dict() unet_state_dict = self.model_config.process_unet_state_dict_for_saving(unet_state_dict) - if self.model_type == ModelType.V_PREDICTION: unet_state_dict["v_pred"] = torch.tensor([]) @@ -776,8 +775,8 @@ class StableAudio1(BaseModel): out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn) return out - def state_dict_for_saving(self, clip_state_dict=None, vae_state_dict=None, clip_vision_state_dict=None): - sd = super().state_dict_for_saving(clip_state_dict=clip_state_dict, vae_state_dict=vae_state_dict, clip_vision_state_dict=clip_vision_state_dict) + def state_dict_for_saving(self, unet_state_dict, clip_state_dict=None, vae_state_dict=None, clip_vision_state_dict=None): + sd = super().state_dict_for_saving(unet_state_dict, clip_state_dict=clip_state_dict, vae_state_dict=vae_state_dict, clip_vision_state_dict=clip_vision_state_dict) d = {"conditioner.conditioners.seconds_start.": self.seconds_start_embedder.state_dict(), "conditioner.conditioners.seconds_total.": self.seconds_total_embedder.state_dict()} for k in d: s = d[k] diff --git a/comfy/model_management.py b/comfy/model_management.py index 9d39be7b2..758e718e8 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -26,6 +26,13 @@ import platform import weakref import gc import os +from contextlib import nullcontext +import comfy.memory_management +import comfy.utils +import comfy.quant_ops + +import comfy_aimdo.torch +import comfy_aimdo.model_vbar class VRAMState(Enum): DISABLED = 0 #No vram present: no need to move models to vram @@ -578,9 +585,15 @@ WINDOWS = any(platform.win32_ver()) EXTRA_RESERVED_VRAM = 400 * 1024 * 1024 if WINDOWS: + import comfy.windows EXTRA_RESERVED_VRAM = 600 * 1024 * 1024 #Windows is higher because of the shared vram issue if total_vram > (15 * 1024): # more extra reserved vram on 16GB+ cards EXTRA_RESERVED_VRAM += 100 * 1024 * 1024 + def get_free_ram(): + return comfy.windows.get_free_ram() +else: + def get_free_ram(): + return psutil.virtual_memory().available if args.reserve_vram is not None: EXTRA_RESERVED_VRAM = args.reserve_vram * 1024 * 1024 * 1024 @@ -592,7 +605,7 @@ def extra_reserved_memory(): def minimum_inference_memory(): return (1024 * 1024 * 1024) * 0.8 + extra_reserved_memory() -def free_memory(memory_required, device, keep_loaded=[]): +def free_memory(memory_required, device, keep_loaded=[], for_dynamic=False, ram_required=0): cleanup_models_gc() unloaded_model = [] can_unload = [] @@ -607,15 +620,23 @@ def free_memory(memory_required, device, keep_loaded=[]): for x in sorted(can_unload): i = x[-1] - memory_to_free = None + memory_to_free = 1e32 + ram_to_free = 1e32 if not DISABLE_SMART_MEMORY: - free_mem = get_free_memory(device) - if free_mem > memory_required: - break - memory_to_free = memory_required - free_mem - logging.debug(f"Unloading {current_loaded_models[i].model.model.__class__.__name__}") - if current_loaded_models[i].model_unload(memory_to_free): + memory_to_free = memory_required - get_free_memory(device) + ram_to_free = ram_required - get_free_ram() + + if current_loaded_models[i].model.is_dynamic() and for_dynamic: + #don't actually unload dynamic models for the sake of other dynamic models + #as that works on-demand. + memory_required -= current_loaded_models[i].model.loaded_size() + memory_to_free = 0 + if memory_to_free > 0 and current_loaded_models[i].model_unload(memory_to_free): + logging.debug(f"Unloading {current_loaded_models[i].model.model.__class__.__name__}") unloaded_model.append(i) + if ram_to_free > 0: + logging.debug(f"RAM Unloading {current_loaded_models[i].model.model.__class__.__name__}") + current_loaded_models[i].model.partially_unload_ram(ram_to_free) for i in sorted(unloaded_model, reverse=True): unloaded_models.append(current_loaded_models.pop(i)) @@ -650,7 +671,10 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimu models_to_load = [] + free_for_dynamic=True for x in models: + if not x.is_dynamic(): + free_for_dynamic = False loaded_model = LoadedModel(x) try: loaded_model_index = current_loaded_models.index(loaded_model) @@ -676,19 +700,25 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimu model_to_unload.model.detach(unpatch_all=False) model_to_unload.model_finalizer.detach() + total_memory_required = {} + total_ram_required = {} for loaded_model in models_to_load: total_memory_required[loaded_model.device] = total_memory_required.get(loaded_model.device, 0) + loaded_model.model_memory_required(loaded_model.device) + #x2, one to make sure the OS can fit the model for loading in disk cache, and for us to do any pinning we + #want to do. + #FIXME: This should subtract off the to_load current pin consumption. + total_ram_required[loaded_model.device] = total_ram_required.get(loaded_model.device, 0) + loaded_model.model_memory() * 2 for device in total_memory_required: if device != torch.device("cpu"): - free_memory(total_memory_required[device] * 1.1 + extra_mem, device) + free_memory(total_memory_required[device] * 1.1 + extra_mem, device, for_dynamic=free_for_dynamic, ram_required=total_ram_required[device]) for device in total_memory_required: if device != torch.device("cpu"): free_mem = get_free_memory(device) if free_mem < minimum_memory_required: - models_l = free_memory(minimum_memory_required, device) + models_l = free_memory(minimum_memory_required, device, for_dynamic=free_for_dynamic) logging.info("{} models unloaded.".format(len(models_l))) for loaded_model in models_to_load: @@ -732,6 +762,9 @@ def loaded_models(only_currently_used=False): def cleanup_models_gc(): do_gc = False + + reset_cast_buffers() + for i in range(len(current_loaded_models)): cur = current_loaded_models[i] if cur.is_dead(): @@ -749,6 +782,11 @@ def cleanup_models_gc(): logging.warning("WARNING, memory leak with model {}. Please make sure it is not being referenced from somewhere.".format(cur.real_model().__class__.__name__)) +def archive_model_dtypes(model): + for name, module in model.named_modules(): + for param_name, param in module.named_parameters(recurse=False): + setattr(module, f"{param_name}_comfy_model_dtype", param.dtype) + def cleanup_models(): to_delete = [] @@ -792,7 +830,7 @@ def unet_inital_load_device(parameters, dtype): mem_dev = get_free_memory(torch_dev) mem_cpu = get_free_memory(cpu_dev) - if mem_dev > mem_cpu and model_size < mem_dev: + if mem_dev > mem_cpu and model_size < mem_dev and comfy.memory_management.aimdo_allocator is None: return torch_dev else: return cpu_dev @@ -1051,6 +1089,53 @@ def current_stream(device): return None stream_counters = {} + +STREAM_CAST_BUFFERS = {} +LARGEST_CASTED_WEIGHT = (None, 0) + +def get_cast_buffer(offload_stream, device, size, ref): + global LARGEST_CASTED_WEIGHT + + if offload_stream is not None: + wf_context = offload_stream + if hasattr(wf_context, "as_context"): + wf_context = wf_context.as_context(offload_stream) + else: + wf_context = nullcontext() + + cast_buffer = STREAM_CAST_BUFFERS.get(offload_stream, None) + if cast_buffer is None or cast_buffer.numel() < size: + if ref is LARGEST_CASTED_WEIGHT[0]: + #If there is one giant weight we do not want both streams to + #allocate a buffer for it. It's up to the caster to get the other + #offload stream in this corner case + return None + if cast_buffer is not None and cast_buffer.numel() > 50 * (1024 ** 2): + #I want my wrongly sized 50MB+ of VRAM back from the caching allocator right now + torch.cuda.synchronize() + del STREAM_CAST_BUFFERS[offload_stream] + del cast_buffer + #FIXME: This doesn't work in Aimdo because mempool cant clear cache + torch.cuda.empty_cache() + with wf_context: + cast_buffer = torch.empty((size), dtype=torch.int8, device=device) + STREAM_CAST_BUFFERS[offload_stream] = cast_buffer + + if size > LARGEST_CASTED_WEIGHT[1]: + LARGEST_CASTED_WEIGHT = (ref, size) + + return cast_buffer + +def reset_cast_buffers(): + global LARGEST_CASTED_WEIGHT + LARGEST_CASTED_WEIGHT = (None, 0) + for offload_stream in STREAM_CAST_BUFFERS: + offload_stream.synchronize() + STREAM_CAST_BUFFERS.clear() + if comfy.memory_management.aimdo_allocator is None: + #Pytorch 2.7 and earlier crashes if you try and empty_cache when mempools exist + torch.cuda.empty_cache() + def get_offload_stream(device): stream_counter = stream_counters.get(device, 0) if NUM_STREAMS == 0: @@ -1093,7 +1178,53 @@ def sync_stream(device, stream): return current_stream(device).wait_stream(stream) -def cast_to(weight, dtype=None, device=None, non_blocking=False, copy=False, stream=None): + +def cast_to_gathered(tensors, r, non_blocking=False, stream=None): + wf_context = nullcontext() + if stream is not None: + wf_context = stream + if hasattr(wf_context, "as_context"): + wf_context = wf_context.as_context(stream) + + dest_views = comfy.memory_management.interpret_gathered_like(tensors, r) + with wf_context: + for tensor in tensors: + dest_view = dest_views.pop(0) + if tensor is None: + continue + dest_view.copy_(tensor, non_blocking=non_blocking) + + +def cast_to(weight, dtype=None, device=None, non_blocking=False, copy=False, stream=None, r=None): + if hasattr(weight, "_v"): + #Unexpected usage patterns. There is no reason these don't work but they + #have no testing and no callers do this. + assert r is None + assert stream is None + + r = torch.empty_like(weight, dtype=weight._model_dtype, device=device) + + signature = comfy_aimdo.model_vbar.vbar_fault(weight._v) + if signature is not None: + raw_tensor = comfy_aimdo.torch.aimdo_to_tensor(weight._v, device) + v_tensor = comfy.memory_management.interpret_gathered_like([r], raw_tensor)[0] + + if comfy_aimdo.model_vbar.vbar_signature_compare(signature, weight._v_signature): + #always take a deep copy even if _v is good, as we have no reasonable point to unpin + #a non comfy weight + r.copy_(v_tensor) + comfy_aimdo.model_vbar.vbar_unpin(weight._v) + return r + + r.copy_(weight, non_blocking=non_blocking) + + if signature is not None: + weight._v_signature = signature + v_tensor.copy_(r) + comfy_aimdo.model_vbar.vbar_unpin(weight._v) + + return r + if device is None or weight.device == device: if not copy: if dtype is None or weight.dtype == dtype: @@ -1112,10 +1243,12 @@ def cast_to(weight, dtype=None, device=None, non_blocking=False, copy=False, str if hasattr(wf_context, "as_context"): wf_context = wf_context.as_context(stream) with wf_context: - r = torch.empty_like(weight, dtype=dtype, device=device) + if r is None: + r = torch.empty_like(weight, dtype=dtype, device=device) r.copy_(weight, non_blocking=non_blocking) else: - r = torch.empty_like(weight, dtype=dtype, device=device) + if r is None: + r = torch.empty_like(weight, dtype=dtype, device=device) r.copy_(weight, non_blocking=non_blocking) return r @@ -1135,7 +1268,7 @@ if not args.disable_pinned_memory: MAX_PINNED_MEMORY = get_total_memory(torch.device("cpu")) * 0.95 logging.info("Enabled pinned memory {}".format(MAX_PINNED_MEMORY // (1024 * 1024))) -PINNING_ALLOWED_TYPES = set(["Parameter", "QuantizedTensor"]) +PINNING_ALLOWED_TYPES = set(["Tensor", "Parameter", "QuantizedTensor"]) def discard_cuda_async_error(): try: @@ -1557,8 +1690,11 @@ def soft_empty_cache(force=False): elif is_mlu(): torch.mlu.empty_cache() elif torch.cuda.is_available(): - torch.cuda.empty_cache() - torch.cuda.ipc_collect() + if comfy.memory_management.aimdo_allocator is None: + #Pytorch 2.7 and earlier crashes if you try and empty_cache when mempools exist + torch.cuda.synchronize() + torch.cuda.empty_cache() + torch.cuda.ipc_collect() def unload_all_models(): free_memory(1e30, get_torch_device()) diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index f6b80a40f..57b53d8c5 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -38,19 +38,7 @@ from comfy.comfy_types import UnetWrapperFunction from comfy.quant_ops import QuantizedTensor from comfy.patcher_extension import CallbacksMP, PatcherInjection, WrappersMP - -def string_to_seed(data): - crc = 0xFFFFFFFF - for byte in data: - if isinstance(byte, str): - byte = ord(byte) - crc ^= byte - for _ in range(8): - if crc & 1: - crc = (crc >> 1) ^ 0xEDB88320 - else: - crc >>= 1 - return crc ^ 0xFFFFFFFF +import comfy_aimdo.model_vbar def set_model_options_patch_replace(model_options, patch, name, block_name, number, transformer_index=None): to = model_options["transformer_options"].copy() @@ -212,6 +200,27 @@ class MemoryCounter: def decrement(self, used: int): self.value -= used +CustomTorchDevice = collections.namedtuple("FakeDevice", ["type", "index"])("comfy-lazy-caster", 0) + +class LazyCastingParam(torch.nn.Parameter): + def __new__(cls, model, key, tensor): + return super().__new__(cls, tensor) + + def __init__(self, model, key, tensor): + self.model = model + self.key = key + + @property + def device(self): + return CustomTorchDevice + + #safetensors will .to() us to the cpu which we catch here to cast on demand. The returned tensor is + #then just a short lived thing in the safetensors serialization logic inside its big for loop over + #all weights getting garbage collected per-weight + def to(self, *args, **kwargs): + return self.model.patch_weight_to_device(self.key, device_to=self.model.load_device, return_weight=True).to("cpu") + + class ModelPatcher: def __init__(self, model, load_device, offload_device, size=0, weight_inplace_update=False): self.size = size @@ -269,6 +278,9 @@ class ModelPatcher: if not hasattr(self.model, 'model_offload_buffer_memory'): self.model.model_offload_buffer_memory = 0 + def is_dynamic(self): + return False + def model_size(self): if self.size > 0: return self.size @@ -284,6 +296,9 @@ class ModelPatcher: def lowvram_patch_counter(self): return self.model.lowvram_patch_counter + def get_free_memory(self, device): + return comfy.model_management.get_free_memory(device) + def clone(self): n = self.__class__(self.model, self.load_device, self.offload_device, self.model_size(), weight_inplace_update=self.weight_inplace_update) n.patches = {} @@ -611,14 +626,14 @@ class ModelPatcher: sd.pop(k) return sd - def patch_weight_to_device(self, key, device_to=None, inplace_update=False): - if key not in self.patches: - return - + def patch_weight_to_device(self, key, device_to=None, inplace_update=False, return_weight=False): weight, set_func, convert_func = get_key_weight(self.model, key) + if key not in self.patches: + return weight + inplace_update = self.weight_inplace_update or inplace_update - if key not in self.backup: + if key not in self.backup and not return_weight: self.backup[key] = collections.namedtuple('Dimension', ['weight', 'inplace_update'])(weight.to(device=self.offload_device, copy=inplace_update), inplace_update) temp_dtype = comfy.model_management.lora_compute_dtype(device_to) @@ -631,13 +646,15 @@ class ModelPatcher: out_weight = comfy.lora.calculate_weight(self.patches[key], temp_weight, key) if set_func is None: - out_weight = comfy.float.stochastic_rounding(out_weight, weight.dtype, seed=string_to_seed(key)) - if inplace_update: + out_weight = comfy.float.stochastic_rounding(out_weight, weight.dtype, seed=comfy.utils.string_to_seed(key)) + if return_weight: + return out_weight + elif inplace_update: comfy.utils.copy_to_param(self.model, key, out_weight) else: comfy.utils.set_attr_param(self.model, key, out_weight) else: - set_func(out_weight, inplace_update=inplace_update, seed=string_to_seed(key)) + return set_func(out_weight, inplace_update=inplace_update, seed=comfy.utils.string_to_seed(key), return_weight=return_weight) def pin_weight_to_device(self, key): weight, set_func, convert_func = get_key_weight(self.model, key) @@ -654,7 +671,7 @@ class ModelPatcher: for key in list(self.pinned): self.unpin_weight(key) - def _load_list(self): + def _load_list(self, prio_comfy_cast_weights=False): loading = [] for n, m in self.model.named_modules(): params = [] @@ -681,7 +698,8 @@ class ModelPatcher: return 0 module_offload_mem += check_module_offload_mem("{}.weight".format(n)) module_offload_mem += check_module_offload_mem("{}.bias".format(n)) - loading.append((module_offload_mem, module_mem, n, m, params)) + prepend = (not hasattr(m, "comfy_cast_weights"),) if prio_comfy_cast_weights else () + loading.append(prepend + (module_offload_mem, module_mem, n, m, params)) return loading def load(self, device_to=None, lowvram_model_memory=0, force_patch_weights=False, full_load=False): @@ -984,6 +1002,9 @@ class ModelPatcher: return self.model.model_loaded_weight_memory - current_used + def partially_unload_ram(self, ram_to_unload): + pass + def detach(self, unpatch_all=True): self.eject_model() self.model_patches_to(self.offload_device) @@ -1317,10 +1338,10 @@ class ModelPatcher: key, original_weights=original_weights) del original_weights[key] if set_func is None: - out_weight = comfy.float.stochastic_rounding(out_weight, weight.dtype, seed=string_to_seed(key)) + out_weight = comfy.float.stochastic_rounding(out_weight, weight.dtype, seed=comfy.utils.string_to_seed(key)) comfy.utils.copy_to_param(self.model, key, out_weight) else: - set_func(out_weight, inplace_update=True, seed=string_to_seed(key)) + set_func(out_weight, inplace_update=True, seed=comfy.utils.string_to_seed(key)) if self.hook_mode == comfy.hooks.EnumHookMode.MaxSpeed: # TODO: disable caching if not enough system RAM to do so target_device = self.offload_device @@ -1355,7 +1376,249 @@ class ModelPatcher: self.unpatch_hooks() self.clear_cached_hook_weights() + def state_dict_for_saving(self, clip_state_dict=None, vae_state_dict=None, clip_vision_state_dict=None): + unet_state_dict = self.model.diffusion_model.state_dict() + for k, v in unet_state_dict.items(): + op_keys = k.rsplit('.', 1) + if (len(op_keys) < 2) or op_keys[1] not in ["weight", "bias"]: + continue + try: + op = comfy.utils.get_attr(self.model.diffusion_model, op_keys[0]) + except: + continue + if not op or not hasattr(op, "comfy_cast_weights") or \ + (hasattr(op, "comfy_patched_weights") and op.comfy_patched_weights == True): + continue + key = "diffusion_model." + k + unet_state_dict[k] = LazyCastingParam(self, key, comfy.utils.get_attr(self.model, key)) + return self.model.state_dict_for_saving(unet_state_dict) + def __del__(self): self.unpin_all_weights() self.detach(unpatch_all=False) +class ModelPatcherDynamic(ModelPatcher): + + def __new__(cls, model=None, load_device=None, offload_device=None, size=0, weight_inplace_update=False): + if load_device is not None and comfy.model_management.is_device_cpu(load_device): + #reroute to default MP for CPUs + return ModelPatcher(model, load_device, offload_device, size, weight_inplace_update) + return super().__new__(cls) + + def __init__(self, model, load_device, offload_device, size=0, weight_inplace_update=False): + super().__init__(model, load_device, offload_device, size, weight_inplace_update) + #this is now way more dynamic and we dont support the same base model for both Dynamic + #and non-dynamic patchers. + if hasattr(self.model, "model_loaded_weight_memory"): + del self.model.model_loaded_weight_memory + if not hasattr(self.model, "dynamic_vbars"): + self.model.dynamic_vbars = {} + assert load_device is not None + + def is_dynamic(self): + return True + + def _vbar_get(self, create=False): + if self.load_device == torch.device("cpu"): + return None + vbar = self.model.dynamic_vbars.get(self.load_device, None) + if create and vbar is None: + # x10. We dont know what model defined type casts we have in the vbar, but virtual address + # space is pretty free. This will cover someone casting an entire model from FP4 to FP32 + # with some left over. + vbar = comfy_aimdo.model_vbar.ModelVBAR(self.model_size() * 10, self.load_device.index) + self.model.dynamic_vbars[self.load_device] = vbar + return vbar + + def loaded_size(self): + vbar = self._vbar_get() + if vbar is None: + return 0 + return vbar.loaded_size() + + def get_free_memory(self, device): + #NOTE: on high condition / batch counts, estimate should have already vacated + #all non-dynamic models so this is safe even if its not 100% true that this + #would all be avaiable for inference use. + return comfy.model_management.get_total_memory(device) - self.model_size() + + #Pinning is deferred to ops time. Assert against this API to avoid pin leaks. + + def pin_weight_to_device(self, key): + raise RuntimeError("pin_weight_to_device invalid for dymamic weight loading") + + def unpin_weight(self, key): + raise RuntimeError("unpin_weight invalid for dymamic weight loading") + + def unpin_all_weights(self): + self.partially_unload_ram(1e32) + + def memory_required(self, input_shape): + #Pad this significantly. We are trying to get away from precise estimates. This + #estimate is only used when using the ModelPatcherDynamic after ModelPatcher. If you + #use all ModelPatcherDynamic this is ignored and its all done dynamically. + return super().memory_required(input_shape=input_shape) * 1.3 + (1024 ** 3) + + + def load(self, device_to=None, lowvram_model_memory=0, force_patch_weights=False, full_load=False, dirty=False): + + #Force patching doesn't make sense in Dynamic loading, as you dont know what does and + #doesn't need to be forced at this stage. The only thing you could do would be patch + #it all on CPU which consumes huge RAM. + assert not force_patch_weights + + #Full load doesn't make sense as we dont actually have any loader capability here and + #now. + assert not full_load + + assert device_to == self.load_device + + num_patches = 0 + allocated_size = 0 + + with self.use_ejected(): + self.unpatch_hooks() + + vbar = self._vbar_get(create=True) + if vbar is not None: + vbar.prioritize() + + #We have way more tools for acceleration on comfy weight offloading, so always + #prioritize the non-comfy weights (note the order reverse). + loading = self._load_list(prio_comfy_cast_weights=True) + loading.sort(reverse=True) + + for x in loading: + _, _, _, n, m, params = x + + def set_dirty(item, dirty): + if dirty or not hasattr(item, "_v_signature"): + item._v_signature = None + + def setup_param(self, m, n, param_key): + nonlocal num_patches + key = "{}.{}".format(n, param_key) + + weight_function = [] + + weight, _, _ = get_key_weight(self.model, key) + if weight is None: + return 0 + if key in self.patches: + setattr(m, param_key + "_lowvram_function", LowVramPatch(key, self.patches)) + num_patches += 1 + else: + setattr(m, param_key + "_lowvram_function", None) + + if key in self.weight_wrapper_patches: + weight_function.extend(self.weight_wrapper_patches[key]) + setattr(m, param_key + "_function", weight_function) + geometry = weight + if not isinstance(weight, QuantizedTensor): + model_dtype = getattr(m, param_key + "_comfy_model_dtype", weight.dtype) + weight._model_dtype = model_dtype + geometry = comfy.memory_management.TensorGeometry(shape=weight.shape, dtype=model_dtype) + return comfy.memory_management.vram_aligned_size(geometry) + + if hasattr(m, "comfy_cast_weights"): + m.comfy_cast_weights = True + m.pin_failed = False + m.seed_key = n + set_dirty(m, dirty) + + v_weight_size = 0 + v_weight_size += setup_param(self, m, n, "weight") + v_weight_size += setup_param(self, m, n, "bias") + + if vbar is not None and not hasattr(m, "_v"): + m._v = vbar.alloc(v_weight_size) + allocated_size += v_weight_size + + else: + for param in params: + key = "{}.{}".format(n, param) + weight, _, _ = get_key_weight(self.model, key) + weight.seed_key = key + set_dirty(weight, dirty) + geometry = weight + model_dtype = getattr(m, param + "_comfy_model_dtype", weight.dtype) + geometry = comfy.memory_management.TensorGeometry(shape=weight.shape, dtype=model_dtype) + weight_size = geometry.numel() * geometry.element_size() + if vbar is not None and not hasattr(weight, "_v"): + weight._v = vbar.alloc(weight_size) + weight._model_dtype = model_dtype + allocated_size += weight_size + + logging.info(f"Model {self.model.__class__.__name__} prepared for dynamic VRAM loading. {allocated_size // (1024 ** 2)}MB Staged. {num_patches} patches attached.") + + self.model.device = device_to + self.model.current_weight_patches_uuid = self.patches_uuid + + for callback in self.get_all_callbacks(CallbacksMP.ON_LOAD): + #These are all super dangerous. Who knows what the custom nodes actually do here... + callback(self, device_to, lowvram_model_memory, force_patch_weights, full_load) + + self.apply_hooks(self.forced_hooks, force_apply=True) + + def partially_unload(self, device_to, memory_to_free=0, force_patch_weights=False): + assert not force_patch_weights #See above + assert self.load_device != torch.device("cpu") + + vbar = self._vbar_get() + return 0 if vbar is None else vbar.free_memory(memory_to_free) + + def partially_unload_ram(self, ram_to_unload): + loading = self._load_list(prio_comfy_cast_weights=True) + for x in loading: + _, _, _, _, m, _ = x + ram_to_unload -= comfy.pinned_memory.unpin_memory(m) + if ram_to_unload <= 0: + return + + def patch_model(self, device_to=None, lowvram_model_memory=0, load_weights=True, force_patch_weights=False): + #This isn't used by the core at all and can only be to load a model out of + #the control of proper model_managment. If you are a custom node author reading + #this, the correct pattern is to call load_models_gpu() to get a proper + #managed load of your model. + assert not load_weights + return super().patch_model(load_weights=load_weights, force_patch_weights=force_patch_weights) + + def unpatch_model(self, device_to=None, unpatch_weights=True): + super().unpatch_model(device_to=None, unpatch_weights=False) + + if unpatch_weights: + self.partially_unload_ram(1e32) + self.partially_unload(None) + + def partially_load(self, device_to, extra_memory=0, force_patch_weights=False): + assert not force_patch_weights #See above + with self.use_ejected(skip_and_inject_on_exit_only=True): + dirty = self.model.current_weight_patches_uuid is not None and (self.model.current_weight_patches_uuid != self.patches_uuid) + + self.unpatch_model(self.offload_device, unpatch_weights=False) + self.patch_model(load_weights=False) + + try: + self.load(device_to, dirty=dirty) + except Exception as e: + self.detach() + raise e + #ModelPatcher::partially_load returns a number on what got loaded but + #nothing in core uses this and we have no data in the Dynamic world. Hit + #the custom node devs with a None rather than a 0 that would mislead any + #logic they might have. + return None + + def patch_cached_hook_weights(self, cached_weights: dict, key: str, memory_counter: MemoryCounter): + assert False #Should be unreachable - we dont ever cache in the new implementation + + def patch_hook_weight_to_device(self, hooks: comfy.hooks.HookGroup, combined_patches: dict, key: str, original_weights: dict, memory_counter: MemoryCounter): + if key not in combined_patches: + return + + raise RuntimeError("Hooks not implemented in ModelPatcherDynamic. Please remove --fast arguments form ComfyUI startup") + + def unpatch_hooks(self, whitelist_keys_set: set[str]=None) -> None: + pass + +CoreModelPatcher = ModelPatcher diff --git a/comfy/ops.py b/comfy/ops.py index e406ba7ed..c3a1825ce 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -19,10 +19,16 @@ import torch import logging import comfy.model_management -from comfy.cli_args import args, PerformanceFeature +from comfy.cli_args import args, PerformanceFeature, enables_dynamic_vram import comfy.float import comfy.rmsnorm import json +import comfy.memory_management +import comfy.pinned_memory +import comfy.utils + +import comfy_aimdo.model_vbar +import comfy_aimdo.torch def run_every_op(): if torch.compiler.is_compiling(): @@ -72,7 +78,115 @@ def cast_to_input(weight, input, non_blocking=False, copy=True): return comfy.model_management.cast_to(weight, input.dtype, input.device, non_blocking=non_blocking, copy=copy) -def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, offloadable=False): +def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compute_dtype): + offload_stream = None + xfer_dest = None + cast_geometry = comfy.memory_management.tensors_to_geometries([ s.weight, s.bias ]) + + signature = comfy_aimdo.model_vbar.vbar_fault(s._v) + if signature is not None: + xfer_dest = comfy_aimdo.torch.aimdo_to_tensor(s._v, device) + resident = comfy_aimdo.model_vbar.vbar_signature_compare(signature, s._v_signature) + + if not resident: + cast_dest = None + + xfer_source = [ s.weight, s.bias ] + + pin = comfy.pinned_memory.get_pin(s) + if pin is not None: + xfer_source = [ pin ] + else: + for data, geometry in zip([ s.weight, s.bias ], cast_geometry): + if data is None: + continue + if data.dtype != geometry.dtype: + cast_dest = xfer_dest + if cast_dest is None: + cast_dest = torch.empty((comfy.memory_management.vram_aligned_size(cast_geometry),), dtype=torch.uint8, device=device) + xfer_dest = None + break + + dest_size = comfy.memory_management.vram_aligned_size(xfer_source) + offload_stream = comfy.model_management.get_offload_stream(device) + if xfer_dest is None and offload_stream is not None: + xfer_dest = comfy.model_management.get_cast_buffer(offload_stream, device, dest_size, s) + if xfer_dest is None: + offload_stream = comfy.model_management.get_offload_stream(device) + xfer_dest = comfy.model_management.get_cast_buffer(offload_stream, device, dest_size, s) + if xfer_dest is None: + xfer_dest = torch.empty((dest_size,), dtype=torch.uint8, device=device) + offload_stream = None + + if signature is None and pin is None: + comfy.pinned_memory.pin_memory(s) + pin = comfy.pinned_memory.get_pin(s) + else: + pin = None + + if pin is not None: + comfy.model_management.cast_to_gathered(xfer_source, pin) + xfer_source = [ pin ] + #send it over + comfy.model_management.cast_to_gathered(xfer_source, xfer_dest, non_blocking=non_blocking, stream=offload_stream) + comfy.model_management.sync_stream(device, offload_stream) + + if cast_dest is not None: + for pre_cast, post_cast in zip(comfy.memory_management.interpret_gathered_like(xfer_source, xfer_dest), + comfy.memory_management.interpret_gathered_like(cast_geometry, cast_dest)): + if post_cast is not None: + post_cast.copy_(pre_cast) + xfer_dest = cast_dest + + params = comfy.memory_management.interpret_gathered_like(cast_geometry, xfer_dest) + weight = params[0] + bias = params[1] + + def post_cast(s, param_key, x, dtype, resident, update_weight): + lowvram_fn = getattr(s, param_key + "_lowvram_function", None) + fns = getattr(s, param_key + "_function", []) + + orig = x + + def to_dequant(tensor, dtype): + tensor = tensor.to(dtype=dtype) + if isinstance(tensor, QuantizedTensor): + tensor = tensor.dequantize() + return tensor + + if orig.dtype != dtype or len(fns) > 0: + x = to_dequant(x, dtype) + if not resident and lowvram_fn is not None: + x = to_dequant(x, dtype if compute_dtype is None else compute_dtype) + #FIXME: this is not accurate, we need to be sensitive to the compute dtype + x = lowvram_fn(x) + if (isinstance(orig, QuantizedTensor) and + (orig.dtype == dtype and len(fns) == 0 or update_weight)): + seed = comfy.utils.string_to_seed(s.seed_key) + y = QuantizedTensor.from_float(x, s.layout_type, scale="recalculate", stochastic_rounding=seed) + if orig.dtype == dtype and len(fns) == 0: + #The layer actually wants our freshly saved QT + x = y + else: + y = x + if update_weight: + orig.copy_(y) + for f in fns: + x = f(x) + return x + + update_weight = signature is not None + + weight = post_cast(s, "weight", weight, dtype, resident, update_weight) + if s.bias is not None: + bias = post_cast(s, "bias", bias, bias_dtype, resident, update_weight) + s._v_signature=signature + + #FIXME: weird offload return protocol + return weight, bias, (offload_stream, device if signature is not None else None, None) + + +def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, offloadable=False, compute_dtype=None): # NOTE: offloadable=False is a a legacy and if you are a custom node author reading this please pass # offloadable=True and call uncast_bias_weight() after your last usage of the weight/bias. This # will add async-offload support to your cast and improve performance. @@ -87,22 +201,38 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, of if device is None: device = input.device + non_blocking = comfy.model_management.device_supports_non_blocking(device) + + if hasattr(s, "_v"): + return cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compute_dtype) + if offloadable and (device != s.weight.device or (s.bias is not None and device != s.bias.device)): offload_stream = comfy.model_management.get_offload_stream(device) else: offload_stream = None - non_blocking = comfy.model_management.device_supports_non_blocking(device) + bias = None + weight = None + + if offload_stream is not None and not args.cuda_malloc: + cast_buffer_size = comfy.memory_management.vram_aligned_size([ s.weight, s.bias ]) + cast_buffer = comfy.model_management.get_cast_buffer(offload_stream, device, cast_buffer_size, s) + #The streams can be uneven in buffer capability and reject us. Retry to get the other stream + if cast_buffer is None: + offload_stream = comfy.model_management.get_offload_stream(device) + cast_buffer = comfy.model_management.get_cast_buffer(offload_stream, device, cast_buffer_size, s) + params = comfy.memory_management.interpret_gathered_like([ s.weight, s.bias ], cast_buffer) + weight = params[0] + bias = params[1] weight_has_function = len(s.weight_function) > 0 bias_has_function = len(s.bias_function) > 0 - weight = comfy.model_management.cast_to(s.weight, None, device, non_blocking=non_blocking, copy=weight_has_function, stream=offload_stream) + weight = comfy.model_management.cast_to(s.weight, None, device, non_blocking=non_blocking, copy=weight_has_function, stream=offload_stream, r=weight) - bias = None if s.bias is not None: - bias = comfy.model_management.cast_to(s.bias, bias_dtype, device, non_blocking=non_blocking, copy=bias_has_function, stream=offload_stream) + bias = comfy.model_management.cast_to(s.bias, None, device, non_blocking=non_blocking, copy=bias_has_function, stream=offload_stream, r=bias) comfy.model_management.sync_stream(device, offload_stream) @@ -110,6 +240,7 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, of weight_a = weight if s.bias is not None: + bias = bias.to(dtype=bias_dtype) for f in s.bias_function: bias = f(bias) @@ -131,14 +262,20 @@ def uncast_bias_weight(s, weight, bias, offload_stream): if offload_stream is None: return os, weight_a, bias_a = offload_stream + device=None + #FIXME: This is not good RTTI + if not isinstance(weight_a, torch.Tensor): + comfy_aimdo.model_vbar.vbar_unpin(s._v) + device = weight_a if os is None: return - if weight_a is not None: - device = weight_a.device - else: - if bias_a is None: - return - device = bias_a.device + if device is None: + if weight_a is not None: + device = weight_a.device + else: + if bias_a is None: + return + device = bias_a.device os.wait_stream(comfy.model_management.current_stream(device)) @@ -149,6 +286,57 @@ class CastWeightBiasOp: class disable_weight_init: class Linear(torch.nn.Linear, CastWeightBiasOp): + + def __init__(self, in_features, out_features, bias=True, device=None, dtype=None): + if not comfy.model_management.WINDOWS or not enables_dynamic_vram(): + super().__init__(in_features, out_features, bias, device, dtype) + return + + # Issue is with `torch.empty` still reserving the full memory for the layer. + # Windows doesn't over-commit memory so without this, We are momentarily commit + # charged for the weight even though we might zero-copy it when we load the + # state dict. If the commit charge exceeds the ceiling we can destabilize the + # system. + torch.nn.Module.__init__(self) + self.in_features = in_features + self.out_features = out_features + self.weight = None + self.bias = None + self.comfy_need_lazy_init_bias=bias + self.weight_comfy_model_dtype = dtype + self.bias_comfy_model_dtype = dtype + + def _load_from_state_dict(self, state_dict, prefix, local_metadata, + strict, missing_keys, unexpected_keys, error_msgs): + + if not comfy.model_management.WINDOWS or not enables_dynamic_vram(): + return super()._load_from_state_dict(state_dict, prefix, local_metadata, strict, + missing_keys, unexpected_keys, error_msgs) + assign_to_params_buffers = local_metadata.get("assign_to_params_buffers", False) + prefix_len = len(prefix) + for k,v in state_dict.items(): + if k[prefix_len:] == "weight": + if not assign_to_params_buffers: + v = v.clone() + self.weight = torch.nn.Parameter(v, requires_grad=False) + elif k[prefix_len:] == "bias" and v is not None: + if not assign_to_params_buffers: + v = v.clone() + self.bias = torch.nn.Parameter(v, requires_grad=False) + else: + unexpected_keys.append(k) + + #Reconcile default construction of the weight if its missing. + if self.weight is None: + v = torch.zeros(self.in_features, self.out_features) + self.weight = torch.nn.Parameter(v, requires_grad=False) + missing_keys.append(prefix+"weight") + if self.bias is None and self.comfy_need_lazy_init_bias: + v = torch.zeros(self.out_features,) + self.bias = torch.nn.Parameter(v, requires_grad=False) + missing_keys.append(prefix+"bias") + + def reset_parameters(self): return None @@ -655,8 +843,8 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec def _forward(self, input, weight, bias): return torch.nn.functional.linear(input, weight, bias) - def forward_comfy_cast_weights(self, input): - weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True) + def forward_comfy_cast_weights(self, input, compute_dtype=None): + weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True, compute_dtype=compute_dtype) x = self._forward(input, weight, bias) uncast_bias_weight(self, weight, bias, offload_stream) return x @@ -666,6 +854,8 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec input_shape = input.shape reshaped_3d = False + #If cast needs to apply lora, it should be done in the compute dtype + compute_dtype = input.dtype if (getattr(self, 'layout_type', None) is not None and not isinstance(input, QuantizedTensor) and not self._full_precision_mm and @@ -684,7 +874,8 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec scale = comfy.model_management.cast_to_device(scale, input.device, None) input = QuantizedTensor.from_float(input_reshaped, self.layout_type, scale=scale) - output = self.forward_comfy_cast_weights(input) + + output = self.forward_comfy_cast_weights(input, compute_dtype) # Reshape output back to 3D if input was 3D if reshaped_3d: diff --git a/comfy/pinned_memory.py b/comfy/pinned_memory.py new file mode 100644 index 000000000..0650e4d1a --- /dev/null +++ b/comfy/pinned_memory.py @@ -0,0 +1,30 @@ +import torch +import comfy.model_management +import comfy.memory_management + +from comfy.cli_args import args + +def get_pin(module): + return getattr(module, "_pin", None) + +def pin_memory(module): + if module.pin_failed or args.disable_pinned_memory or get_pin(module) is not None: + return + #FIXME: This is a RAM cache trigger event + params = comfy.memory_management.tensors_to_geometries([ module.weight, module.bias ]) + size = comfy.memory_management.vram_aligned_size(params) + pin = torch.empty((size,), dtype=torch.uint8) + if comfy.model_management.pin_memory(pin): + module._pin = pin + else: + module.pin_failed = True + return False + return True + +def unpin_memory(module): + if get_pin(module) is None: + return 0 + size = module._pin.numel() * module._pin.element_size() + comfy.model_management.unpin_memory(module._pin) + del module._pin + return size diff --git a/comfy/samplers.py b/comfy/samplers.py index 1989ef107..8b9782956 100755 --- a/comfy/samplers.py +++ b/comfy/samplers.py @@ -9,7 +9,6 @@ if TYPE_CHECKING: import torch from functools import partial import collections -from comfy import model_management import math import logging import comfy.sampler_helpers @@ -260,7 +259,7 @@ def _calc_cond_batch(model: BaseModel, conds: list[list[dict]], x_in: torch.Tens to_batch_temp.reverse() to_batch = to_batch_temp[:1] - free_memory = model_management.get_free_memory(x_in.device) + free_memory = model.current_patcher.get_free_memory(x_in.device) for i in range(1, len(to_batch_temp) + 1): batch_amount = to_batch_temp[:len(to_batch_temp)//i] input_shape = [len(batch_amount) * first_shape[0]] + list(first_shape)[1:] diff --git a/comfy/sd.py b/comfy/sd.py index f627f7d55..fd0ac85e7 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -228,8 +228,10 @@ class CLIP: self.cond_stage_model.to(offload_device) logging.warning("Had to shift TE back.") + model_management.archive_model_dtypes(self.cond_stage_model) + self.tokenizer = tokenizer(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data) - self.patcher = comfy.model_patcher.ModelPatcher(self.cond_stage_model, load_device=load_device, offload_device=offload_device) + self.patcher = comfy.model_patcher.CoreModelPatcher(self.cond_stage_model, load_device=load_device, offload_device=offload_device) #Match torch.float32 hardcode upcast in TE implemention self.patcher.set_model_compute_dtype(torch.float32) self.patcher.hook_mode = comfy.hooks.EnumHookMode.MinVram @@ -389,8 +391,18 @@ class CLIP: def load_sd(self, sd, full_model=False): if full_model: - return self.cond_stage_model.load_state_dict(sd, strict=False) + return self.cond_stage_model.load_state_dict(sd, strict=False, assign=self.patcher.is_dynamic()) else: + can_assign = self.patcher.is_dynamic() + self.cond_stage_model.can_assign_sd = can_assign + + # The CLIP models are a pretty complex web of wrappers and its + # a bit of an API change to plumb this all the way through. + # So spray paint the model with this flag that the loading + # nn.Module can then inspect for itself. + for m in self.cond_stage_model.modules(): + m.can_assign_sd = can_assign + return self.cond_stage_model.load_sd(sd) def get_sd(self): @@ -765,12 +777,7 @@ class VAE: self.first_stage_model = AutoencoderKL(**(config['params'])) self.first_stage_model = self.first_stage_model.eval() - m, u = self.first_stage_model.load_state_dict(sd, strict=False) - if len(m) > 0: - logging.warning("Missing VAE keys {}".format(m)) - - if len(u) > 0: - logging.debug("Leftover VAE keys {}".format(u)) + model_management.archive_model_dtypes(self.first_stage_model) if device is None: device = model_management.vae_device() @@ -782,7 +789,18 @@ class VAE: self.first_stage_model.to(self.vae_dtype) self.output_device = model_management.intermediate_device() - self.patcher = comfy.model_patcher.ModelPatcher(self.first_stage_model, load_device=self.device, offload_device=offload_device) + mp = comfy.model_patcher.CoreModelPatcher + if self.disable_offload: + mp = comfy.model_patcher.ModelPatcher + self.patcher = mp(self.first_stage_model, load_device=self.device, offload_device=offload_device) + + m, u = self.first_stage_model.load_state_dict(sd, strict=False, assign=self.patcher.is_dynamic()) + if len(m) > 0: + logging.warning("Missing VAE keys {}".format(m)) + + if len(u) > 0: + logging.debug("Leftover VAE keys {}".format(u)) + logging.info("VAE load device: {}, offload device: {}, dtype: {}".format(self.device, offload_device, self.vae_dtype)) self.model_size() @@ -897,7 +915,7 @@ class VAE: try: memory_used = self.memory_used_decode(samples_in.shape, self.vae_dtype) model_management.load_models_gpu([self.patcher], memory_required=memory_used, force_full_load=self.disable_offload) - free_memory = model_management.get_free_memory(self.device) + free_memory = self.patcher.get_free_memory(self.device) batch_number = int(free_memory / memory_used) batch_number = max(1, batch_number) @@ -971,7 +989,7 @@ class VAE: try: memory_used = self.memory_used_encode(pixel_samples.shape, self.vae_dtype) model_management.load_models_gpu([self.patcher], memory_required=memory_used, force_full_load=self.disable_offload) - free_memory = model_management.get_free_memory(self.device) + free_memory = self.patcher.get_free_memory(self.device) batch_number = int(free_memory / max(1, memory_used)) batch_number = max(1, batch_number) samples = None @@ -1432,7 +1450,7 @@ def load_gligen(ckpt_path): model = gligen.load_gligen(data) if model_management.should_use_fp16(): model = model.half() - return comfy.model_patcher.ModelPatcher(model, load_device=model_management.get_torch_device(), offload_device=model_management.unet_offload_device()) + return comfy.model_patcher.CoreModelPatcher(model, load_device=model_management.get_torch_device(), offload_device=model_management.unet_offload_device()) def model_detection_error_hint(path, state_dict): filename = os.path.basename(path) @@ -1520,7 +1538,8 @@ def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_c if output_model: inital_load_device = model_management.unet_inital_load_device(parameters, unet_dtype) model = model_config.get_model(sd, diffusion_model_prefix, device=inital_load_device) - model.load_model_weights(sd, diffusion_model_prefix) + model_patcher = comfy.model_patcher.CoreModelPatcher(model, load_device=load_device, offload_device=model_management.unet_offload_device()) + model.load_model_weights(sd, diffusion_model_prefix, assign=model_patcher.is_dynamic()) if output_vae: vae_sd = comfy.utils.state_dict_prefix_replace(sd, {k: "" for k in model_config.vae_key_prefix}, filter_keys=True) @@ -1563,7 +1582,6 @@ def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_c logging.debug("left over keys: {}".format(left_over)) if output_model: - model_patcher = comfy.model_patcher.ModelPatcher(model, load_device=load_device, offload_device=model_management.unet_offload_device()) if inital_load_device != torch.device("cpu"): logging.info("loaded diffusion model directly to GPU") model_management.load_models_gpu([model_patcher], force_full_load=True) @@ -1655,13 +1673,14 @@ def load_diffusion_model_state_dict(sd, model_options={}, metadata=None): model_config.optimizations["fp8"] = True model = model_config.get_model(new_sd, "") - model = model.to(offload_device) - model.load_model_weights(new_sd, "") + model_patcher = comfy.model_patcher.CoreModelPatcher(model, load_device=load_device, offload_device=offload_device) + if not model_management.is_device_cpu(offload_device): + model.to(offload_device) + model.load_model_weights(new_sd, "", assign=model_patcher.is_dynamic()) left_over = sd.keys() if len(left_over) > 0: logging.info("left over keys in diffusion model: {}".format(left_over)) - return comfy.model_patcher.ModelPatcher(model, load_device=load_device, offload_device=offload_device) - + return model_patcher def load_diffusion_model(unet_path, model_options={}): sd, metadata = comfy.utils.load_torch_file(unet_path, return_metadata=True) @@ -1692,9 +1711,9 @@ def save_checkpoint(output_path, model, clip=None, vae=None, clip_vision=None, m if metadata is None: metadata = {} - model_management.load_models_gpu(load_models, force_patch_weights=True) + model_management.load_models_gpu(load_models) clip_vision_sd = clip_vision.get_sd() if clip_vision is not None else None - sd = model.model.state_dict_for_saving(clip_sd, vae_sd, clip_vision_sd) + sd = model.state_dict_for_saving(clip_sd, vae_sd, clip_vision_sd) for k in extra_keys: sd[k] = extra_keys[k] diff --git a/comfy/sd1_clip.py b/comfy/sd1_clip.py index d4f22120b..9ecfc9c55 100644 --- a/comfy/sd1_clip.py +++ b/comfy/sd1_clip.py @@ -297,7 +297,7 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder): return self(tokens) def load_sd(self, sd): - return self.transformer.load_state_dict(sd, strict=False) + return self.transformer.load_state_dict(sd, strict=False, assign=getattr(self, "can_assign_sd", False)) def parse_parentheses(string): result = [] diff --git a/comfy/text_encoders/lt.py b/comfy/text_encoders/lt.py index e49161964..26573fb12 100644 --- a/comfy/text_encoders/lt.py +++ b/comfy/text_encoders/lt.py @@ -125,7 +125,7 @@ class LTXAVTEModel(torch.nn.Module): for prefix, component in [("text_embedding_projection.", self.text_embedding_projection), ("video_embeddings_connector.", self.video_embeddings_connector), ("audio_embeddings_connector.", self.audio_embeddings_connector)]: component_sd = {k.replace(prefix, ""): v for k, v in sdo.items() if k.startswith(prefix)} if component_sd: - missing, unexpected = component.load_state_dict(component_sd, strict=False) + missing, unexpected = component.load_state_dict(component_sd, strict=False, assign=getattr(self, "can_assign_sd", False)) missing_all.extend([f"{prefix}{k}" for k in missing]) unexpected_all.extend([f"{prefix}{k}" for k in unexpected]) diff --git a/comfy/utils.py b/comfy/utils.py index d97d753e6..9e98eb176 100644 --- a/comfy/utils.py +++ b/comfy/utils.py @@ -28,9 +28,10 @@ import logging import itertools from torch.nn.functional import interpolate from einops import rearrange -from comfy.cli_args import args +from comfy.cli_args import args, enables_dynamic_vram import json import time +import mmap MMAP_TORCH_FILES = args.mmap_torch_files DISABLE_MMAP = args.disable_mmap @@ -56,21 +57,67 @@ if hasattr(torch.serialization, "add_safe_globals"): # TODO: this was added in else: logging.warning("Warning, you are using an old pytorch version and some ckpt/pt files might be loaded unsafely. Upgrading to 2.4 or above is recommended as older versions of pytorch are no longer supported.") +# Current as of safetensors 0.7.0 +_TYPES = { + "F64": torch.float64, + "F32": torch.float32, + "F16": torch.float16, + "BF16": torch.bfloat16, + "I64": torch.int64, + "I32": torch.int32, + "I16": torch.int16, + "I8": torch.int8, + "U8": torch.uint8, + "BOOL": torch.bool, + "F8_E4M3": torch.float8_e4m3fn, + "F8_E5M2": torch.float8_e5m2, + "C64": torch.complex64, + + "U64": torch.uint64, + "U32": torch.uint32, + "U16": torch.uint16, +} + +def load_safetensors(ckpt): + f = open(ckpt, "rb") + mapping = mmap.mmap(f.fileno(), 0, access=mmap.ACCESS_READ) + + header_size = struct.unpack(" 0: message = e.args[0] @@ -1308,3 +1355,16 @@ def convert_old_quants(state_dict, model_prefix="", metadata={}): state_dict["{}.comfy_quant".format(k)] = torch.tensor(list(json.dumps(v).encode('utf-8')), dtype=torch.uint8) return state_dict, metadata + +def string_to_seed(data): + crc = 0xFFFFFFFF + for byte in data: + if isinstance(byte, str): + byte = ord(byte) + crc ^= byte + for _ in range(8): + if crc & 1: + crc = (crc >> 1) ^ 0xEDB88320 + else: + crc >>= 1 + return crc ^ 0xFFFFFFFF diff --git a/comfy/windows.py b/comfy/windows.py new file mode 100644 index 000000000..213dc481d --- /dev/null +++ b/comfy/windows.py @@ -0,0 +1,52 @@ +import ctypes +import logging +import psutil +from ctypes import wintypes + +import comfy_aimdo.control + +psapi = ctypes.WinDLL("psapi") +kernel32 = ctypes.WinDLL("kernel32") + +class PERFORMANCE_INFORMATION(ctypes.Structure): + _fields_ = [ + ("cb", wintypes.DWORD), + ("CommitTotal", ctypes.c_size_t), + ("CommitLimit", ctypes.c_size_t), + ("CommitPeak", ctypes.c_size_t), + ("PhysicalTotal", ctypes.c_size_t), + ("PhysicalAvailable", ctypes.c_size_t), + ("SystemCache", ctypes.c_size_t), + ("KernelTotal", ctypes.c_size_t), + ("KernelPaged", ctypes.c_size_t), + ("KernelNonpaged", ctypes.c_size_t), + ("PageSize", ctypes.c_size_t), + ("HandleCount", wintypes.DWORD), + ("ProcessCount", wintypes.DWORD), + ("ThreadCount", wintypes.DWORD), + ] + +def get_free_ram(): + #Windows is way too conservative and chalks recently used uncommitted model RAM + #as "in-use". So, calculate free RAM for the sake of general use as the greater of: + # + #1: What psutil says + #2: Total Memory - (Committed Memory - VRAM in use) + # + #We have to subtract VRAM in use from the comitted memory as WDDM creates a naked + #commit charge for all VRAM used just incase it wants to page it all out. This just + #isn't realistic so "overcommit" on our calculations by just subtracting it off. + + pi = PERFORMANCE_INFORMATION() + pi.cb = ctypes.sizeof(pi) + + if not psapi.GetPerformanceInfo(ctypes.byref(pi), pi.cb): + logging.warning("WARNING: Failed to query windows performance info. RAM usage may be sub optimal") + return psutil.virtual_memory().available + + committed = pi.CommitTotal * pi.PageSize + total = pi.PhysicalTotal * pi.PageSize + + return max(psutil.virtual_memory().available, + total - (committed - comfy_aimdo.control.get_total_vram_usage())) + diff --git a/comfy_extras/nodes_model_patch.py b/comfy_extras/nodes_model_patch.py index 82c4754a3..176e6bc2f 100644 --- a/comfy_extras/nodes_model_patch.py +++ b/comfy_extras/nodes_model_patch.py @@ -267,9 +267,9 @@ class ModelPatchLoader: device=comfy.model_management.unet_offload_device(), operations=comfy.ops.manual_cast) - model.load_state_dict(sd) - model = comfy.model_patcher.ModelPatcher(model, load_device=comfy.model_management.get_torch_device(), offload_device=comfy.model_management.unet_offload_device()) - return (model,) + model_patcher = comfy.model_patcher.CoreModelPatcher(model, load_device=comfy.model_management.get_torch_device(), offload_device=comfy.model_management.unet_offload_device()) + model.load_state_dict(sd, assign=model_patcher.is_dynamic()) + return (model_patcher,) class DiffSynthCnetPatch: diff --git a/cuda_malloc.py b/cuda_malloc.py index ee2bc4b69..b2182df37 100644 --- a/cuda_malloc.py +++ b/cuda_malloc.py @@ -1,8 +1,10 @@ import os import importlib.util -from comfy.cli_args import args, PerformanceFeature +from comfy.cli_args import args, PerformanceFeature, enables_dynamic_vram import subprocess +import comfy_aimdo.control + #Can't use pytorch to get the GPU names because the cuda malloc has to be set before the first import. def get_gpu_names(): if os.name == 'nt': @@ -85,8 +87,14 @@ if not args.cuda_malloc: except: pass +if enables_dynamic_vram() and comfy_aimdo.control.init(): + args.cuda_malloc = False + os.environ['PYTORCH_CUDA_ALLOC_CONF'] = "" -if args.cuda_malloc and not args.disable_cuda_malloc: +if args.disable_cuda_malloc: + args.cuda_malloc = False + +if args.cuda_malloc: env_var = os.environ.get('PYTORCH_CUDA_ALLOC_CONF', None) if env_var is None: env_var = "backend:cudaMallocAsync" diff --git a/execution.py b/execution.py index 4b4f63c80..93fafc4a2 100644 --- a/execution.py +++ b/execution.py @@ -9,9 +9,11 @@ import traceback from enum import Enum from typing import List, Literal, NamedTuple, Optional, Union import asyncio +from contextlib import nullcontext import torch +import comfy.memory_management import comfy.model_management from latent_preview import set_preview_method import nodes @@ -515,7 +517,19 @@ async def execute(server, dynprompt, caches, current_item, extra_data, executed, def pre_execute_cb(call_index): # TODO - How to handle this with async functions without contextvars (which requires Python 3.12)? GraphBuilder.set_default_prefix(unique_id, call_index, 0) - output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data) + + #Do comfy_aimdo mempool chunking here on the per-node level. Multi-model workflows + #will cause all sorts of incompatible memory shapes to fragment the pytorch alloc + #that we just want to cull out each model run. + allocator = comfy.memory_management.aimdo_allocator + with nullcontext() if allocator is None else torch.cuda.use_mem_pool(torch.cuda.MemPool(allocator.allocator())): + try: + output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data) + finally: + if allocator is not None: + comfy.model_management.reset_cast_buffers() + torch.cuda.synchronize() + if has_pending_tasks: pending_async_nodes[unique_id] = output_data unblock = execution_list.add_external_block(unique_id) diff --git a/main.py b/main.py index 37b06c1fa..b8c951375 100644 --- a/main.py +++ b/main.py @@ -5,7 +5,7 @@ import os import importlib.util import folder_paths import time -from comfy.cli_args import args +from comfy.cli_args import args, enables_dynamic_vram from app.logger import setup_logger from app.assets.scanner import seed_assets import itertools @@ -173,6 +173,7 @@ import gc if 'torch' in sys.modules: logging.warning("WARNING: Potential Error in code: Torch already imported, torch should never be imported before this point.") + import comfy.utils import execution @@ -184,6 +185,33 @@ import comfyui_version import app.logger import hook_breaker_ac10a0 +import comfy.memory_management +import comfy.model_patcher + +import comfy_aimdo.control +import comfy_aimdo.torch + +if enables_dynamic_vram(): + if comfy_aimdo.control.init_device(comfy.model_management.get_torch_device().index): + if args.verbose == 'DEBUG': + comfy_aimdo.control.set_log_debug() + elif args.verbose == 'CRITICAL': + comfy_aimdo.control.set_log_critical() + elif args.verbose == 'ERROR': + comfy_aimdo.control.set_log_error() + elif args.verbose == 'WARNING': + comfy_aimdo.control.set_log_warning() + else: #INFO + comfy_aimdo.control.set_log_info() + + comfy.model_patcher.CoreModelPatcher = comfy.model_patcher.ModelPatcherDynamic + comfy.memory_management.aimdo_allocator = comfy_aimdo.torch.get_torch_allocator() + logging.info("DynamicVRAM support detected and enabled") + else: + logging.info("No working comfy-aimdo install detected. DynamicVRAM support disabled. Falling back to legacy ModelPatcher. VRAM estimates may be unreliable especially on Windows") + comfy.memory_management.aimdo_allocator = None + + def cuda_malloc_warning(): device = comfy.model_management.get_torch_device() device_name = comfy.model_management.get_torch_device_name(device) diff --git a/requirements.txt b/requirements.txt index 4ac94cb16..be0bc537e 100644 --- a/requirements.txt +++ b/requirements.txt @@ -22,6 +22,7 @@ alembic SQLAlchemy av>=14.2.0 comfy-kitchen>=0.2.7 +comfy-aimdo>=0.1.6 requests #non essential dependencies: From 32621c6a11f6ce1c0dd46aeb08e35dd64ec804f0 Mon Sep 17 00:00:00 2001 From: Christian Byrne Date: Sat, 31 Jan 2026 22:13:48 -0800 Subject: [PATCH 545/647] fix: improve error message when node type is missing (#12194) - Change error type from 'invalid_prompt' to 'missing_node_type' for frontend detection - Add extra_info with node_id, class_type, and node_title (from _meta.title) - Improve user-facing message: 'Node X not found. The custom node may not be installed.' --- execution.py | 24 ++++++++++++++++++------ 1 file changed, 18 insertions(+), 6 deletions(-) diff --git a/execution.py b/execution.py index 93fafc4a2..3dbab82e6 100644 --- a/execution.py +++ b/execution.py @@ -1014,22 +1014,34 @@ async def validate_prompt(prompt_id, prompt, partial_execution_list: Union[list[ outputs = set() for x in prompt: if 'class_type' not in prompt[x]: + node_data = prompt[x] + node_title = node_data.get('_meta', {}).get('title') error = { - "type": "invalid_prompt", - "message": "Cannot execute because a node is missing the class_type property.", + "type": "missing_node_type", + "message": f"Node '{node_title or f'ID #{x}'}' has no class_type. The workflow may be corrupted or a custom node is missing.", "details": f"Node ID '#{x}'", - "extra_info": {} + "extra_info": { + "node_id": x, + "class_type": None, + "node_title": node_title + } } return (False, error, [], {}) class_type = prompt[x]['class_type'] class_ = nodes.NODE_CLASS_MAPPINGS.get(class_type, None) if class_ is None: + node_data = prompt[x] + node_title = node_data.get('_meta', {}).get('title', class_type) error = { - "type": "invalid_prompt", - "message": f"Cannot execute because node {class_type} does not exist.", + "type": "missing_node_type", + "message": f"Node '{node_title}' not found. The custom node may not be installed.", "details": f"Node ID '#{x}'", - "extra_info": {} + "extra_info": { + "node_id": x, + "class_type": class_type, + "node_title": node_title + } } return (False, error, [], {}) From 667a1b8878187f7a5c8955b272e33c16e4c46bb7 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sat, 31 Jan 2026 22:55:18 -0800 Subject: [PATCH 546/647] Fix some custom nodes breaking. (#12203) --- comfy/model_patcher.py | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index 57b53d8c5..c8e6f088f 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -111,6 +111,10 @@ def move_weight_functions(m, device): memory += f.move_to(device=device) return memory +def string_to_seed(data): + logging.warning("WARNING: string_to_seed has moved from comfy.model_patcher to comfy.utils") + return comfy.utils.string_to_seed(data) + class LowVramPatch: def __init__(self, key, patches, convert_func=None, set_func=None): self.key = key From 361b9a82a3e2445fc22e352df187138b0c8f67fb Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Sun, 1 Feb 2026 08:42:32 -0800 Subject: [PATCH 547/647] fix pinning with model defined dtype (#12208) pinned memory was converted back to pinning the CPU side weight without any changes. Fix the pinner to use the CPU weight and not the model defined geometry. This will either save RAM or stop buffer overruns when the types mismatch. Fix the model defined weight caster to use the [ s.weight, s.bias ] interpretation, as xfer_dest might be the flattened pin now. Fix the detection of needing to cast to not be conditional on !pin. --- comfy/ops.py | 22 +++++++++++----------- comfy/pinned_memory.py | 3 +-- 2 files changed, 12 insertions(+), 13 deletions(-) diff --git a/comfy/ops.py b/comfy/ops.py index c3a1825ce..53c5e4dc3 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -96,16 +96,16 @@ def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compu pin = comfy.pinned_memory.get_pin(s) if pin is not None: xfer_source = [ pin ] - else: - for data, geometry in zip([ s.weight, s.bias ], cast_geometry): - if data is None: - continue - if data.dtype != geometry.dtype: - cast_dest = xfer_dest - if cast_dest is None: - cast_dest = torch.empty((comfy.memory_management.vram_aligned_size(cast_geometry),), dtype=torch.uint8, device=device) - xfer_dest = None - break + + for data, geometry in zip([ s.weight, s.bias ], cast_geometry): + if data is None: + continue + if data.dtype != geometry.dtype: + cast_dest = xfer_dest + if cast_dest is None: + cast_dest = torch.empty((comfy.memory_management.vram_aligned_size(cast_geometry),), dtype=torch.uint8, device=device) + xfer_dest = None + break dest_size = comfy.memory_management.vram_aligned_size(xfer_source) offload_stream = comfy.model_management.get_offload_stream(device) @@ -132,7 +132,7 @@ def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compu comfy.model_management.sync_stream(device, offload_stream) if cast_dest is not None: - for pre_cast, post_cast in zip(comfy.memory_management.interpret_gathered_like(xfer_source, xfer_dest), + for pre_cast, post_cast in zip(comfy.memory_management.interpret_gathered_like([s.weight, s.bias ], xfer_dest), comfy.memory_management.interpret_gathered_like(cast_geometry, cast_dest)): if post_cast is not None: post_cast.copy_(pre_cast) diff --git a/comfy/pinned_memory.py b/comfy/pinned_memory.py index 0650e4d1a..8acc327a7 100644 --- a/comfy/pinned_memory.py +++ b/comfy/pinned_memory.py @@ -11,8 +11,7 @@ def pin_memory(module): if module.pin_failed or args.disable_pinned_memory or get_pin(module) is not None: return #FIXME: This is a RAM cache trigger event - params = comfy.memory_management.tensors_to_geometries([ module.weight, module.bias ]) - size = comfy.memory_management.vram_aligned_size(params) + size = comfy.memory_management.vram_aligned_size([ module.weight, module.bias ]) pin = torch.empty((size,), dtype=torch.uint8) if comfy.model_management.pin_memory(pin): module._pin = pin From 794d05bdb12be95f0e58dbb1728542ac4c7998aa Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Sun, 1 Feb 2026 17:09:21 -0800 Subject: [PATCH 548/647] dynamic_vram: respect argument cast dtypes in non-comfy weights (#12209) This function has a dtype argument that allows the caller to set the dtype in the cast. TIL Some models override this on weight casts, which means its the highest priority. Priority scheme is: argument > model dtype > state dict dtype --- comfy/model_management.py | 27 ++++++++++++++++++--------- 1 file changed, 18 insertions(+), 9 deletions(-) diff --git a/comfy/model_management.py b/comfy/model_management.py index 758e718e8..6b1166b94 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -1202,27 +1202,36 @@ def cast_to(weight, dtype=None, device=None, non_blocking=False, copy=False, str assert r is None assert stream is None - r = torch.empty_like(weight, dtype=weight._model_dtype, device=device) + cast_geometry = comfy.memory_management.tensors_to_geometries([ weight ]) + + if dtype is None: + dtype = weight._model_dtype + + r = torch.empty_like(weight, dtype=dtype, device=device) signature = comfy_aimdo.model_vbar.vbar_fault(weight._v) if signature is not None: raw_tensor = comfy_aimdo.torch.aimdo_to_tensor(weight._v, device) - v_tensor = comfy.memory_management.interpret_gathered_like([r], raw_tensor)[0] - - if comfy_aimdo.model_vbar.vbar_signature_compare(signature, weight._v_signature): + v_tensor = comfy.memory_management.interpret_gathered_like(cast_geometry, raw_tensor)[0] + if not comfy_aimdo.model_vbar.vbar_signature_compare(signature, weight._v_signature): + weight._v_signature = signature + #Send it over + v_tensor.copy_(weight, non_blocking=non_blocking) #always take a deep copy even if _v is good, as we have no reasonable point to unpin #a non comfy weight r.copy_(v_tensor) comfy_aimdo.model_vbar.vbar_unpin(weight._v) return r + if weight.dtype != r.dtype and weight.dtype != weight._model_dtype: + #Offloaded casting could skip this, however it would make the quantizations + #inconsistent between loaded and offloaded weights. So force the double casting + #that would happen in regular flow to make offload deterministic. + cast_buffer = torch.empty_like(weight, dtype=weight._model_dtype, device=device) + cast_buffer.copy_(weight, non_blocking=non_blocking) + weight = cast_buffer r.copy_(weight, non_blocking=non_blocking) - if signature is not None: - weight._v_signature = signature - v_tensor.copy_(r) - comfy_aimdo.model_vbar.vbar_unpin(weight._v) - return r if device is None or weight.device == device: From 2b5da3b72ebb8017661b0d58b548d4aeb53b7e42 Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Sun, 1 Feb 2026 17:09:55 -0800 Subject: [PATCH 549/647] dynamic_vram: silence pytorch buffer warning (#12210) This is log clutter and concerning to users. Its a false alarm. --- comfy/utils.py | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-) diff --git a/comfy/utils.py b/comfy/utils.py index 9e98eb176..c1b536833 100644 --- a/comfy/utils.py +++ b/comfy/utils.py @@ -32,6 +32,7 @@ from comfy.cli_args import args, enables_dynamic_vram import json import time import mmap +import warnings MMAP_TORCH_FILES = args.mmap_torch_files DISABLE_MMAP = args.disable_mmap @@ -85,7 +86,10 @@ def load_safetensors(ckpt): header_size = struct.unpack(" Date: Sun, 1 Feb 2026 17:10:15 -0800 Subject: [PATCH 550/647] requirements: bump comfy-aimdo to 0.1.7 (#12211) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index be0bc537e..3ca417dd8 100644 --- a/requirements.txt +++ b/requirements.txt @@ -22,7 +22,7 @@ alembic SQLAlchemy av>=14.2.0 comfy-kitchen>=0.2.7 -comfy-aimdo>=0.1.6 +comfy-aimdo>=0.1.7 requests #non essential dependencies: From 021ba2071985f9e3b2984f396a238095c4a64832 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sun, 1 Feb 2026 17:12:52 -0800 Subject: [PATCH 551/647] Fix issue with parameters on root model object. (#12216) --- comfy/model_patcher.py | 17 +++++++++++------ 1 file changed, 11 insertions(+), 6 deletions(-) diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index c8e6f088f..b70c031bf 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -161,6 +161,11 @@ def get_key_weight(model, key): return weight, set_func, convert_func +def key_param_name_to_key(key, param): + if len(key) == 0: + return param + return "{}.{}".format(key, param) + class AutoPatcherEjector: def __init__(self, model: 'ModelPatcher', skip_and_inject_on_exit_only=False): self.model = model @@ -795,7 +800,7 @@ class ModelPatcher: continue for param in params: - key = "{}.{}".format(n, param) + key = key_param_name_to_key(n, param) self.unpin_weight(key) self.patch_weight_to_device(key, device_to=device_to) if comfy.model_management.is_device_cuda(device_to): @@ -811,7 +816,7 @@ class ModelPatcher: n = x[1] params = x[3] for param in params: - self.pin_weight_to_device("{}.{}".format(n, param)) + self.pin_weight_to_device(key_param_name_to_key(n, param)) usable_stat = "{:.2f} MB usable,".format(lowvram_model_memory / (1024 * 1024)) if lowvram_model_memory < 1e32 else "" if lowvram_counter > 0: @@ -917,7 +922,7 @@ class ModelPatcher: if hasattr(m, "comfy_patched_weights") and m.comfy_patched_weights == True: move_weight = True for param in params: - key = "{}.{}".format(n, param) + key = key_param_name_to_key(n, param) bk = self.backup.get(key, None) if bk is not None: if not lowvram_possible: @@ -968,7 +973,7 @@ class ModelPatcher: logging.debug("freed {}".format(n)) for param in params: - self.pin_weight_to_device("{}.{}".format(n, param)) + self.pin_weight_to_device(key_param_name_to_key(n, param)) self.model.model_lowvram = True @@ -1501,7 +1506,7 @@ class ModelPatcherDynamic(ModelPatcher): def setup_param(self, m, n, param_key): nonlocal num_patches - key = "{}.{}".format(n, param_key) + key = key_param_name_to_key(n, param_key) weight_function = [] @@ -1540,7 +1545,7 @@ class ModelPatcherDynamic(ModelPatcher): else: for param in params: - key = "{}.{}".format(n, param) + key = key_param_name_to_key(n, param) weight, _, _ = get_key_weight(self.model, key) weight.seed_key = key set_dirty(weight, dirty) From dd86b155210df9b34f479d70dad675aa782a30ef Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 2 Feb 2026 00:51:09 -0800 Subject: [PATCH 552/647] Enable embeddings for some qwen 3 models. (#12218) --- comfy/text_encoders/anima.py | 2 +- comfy/text_encoders/flux.py | 6 +++--- comfy/text_encoders/z_image.py | 2 +- 3 files changed, 5 insertions(+), 5 deletions(-) diff --git a/comfy/text_encoders/anima.py b/comfy/text_encoders/anima.py index 41f95bcb6..b6f58cb25 100644 --- a/comfy/text_encoders/anima.py +++ b/comfy/text_encoders/anima.py @@ -8,7 +8,7 @@ import torch class Qwen3Tokenizer(sd1_clip.SDTokenizer): def __init__(self, embedding_directory=None, tokenizer_data={}): tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "qwen25_tokenizer") - super().__init__(tokenizer_path, pad_with_end=False, embedding_size=1024, embedding_key='qwen3_06b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, pad_token=151643, tokenizer_data=tokenizer_data) + super().__init__(tokenizer_path, pad_with_end=False, embedding_directory=embedding_directory, embedding_size=1024, embedding_key='qwen3_06b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, pad_token=151643, tokenizer_data=tokenizer_data) class T5XXLTokenizer(sd1_clip.SDTokenizer): def __init__(self, embedding_directory=None, tokenizer_data={}): diff --git a/comfy/text_encoders/flux.py b/comfy/text_encoders/flux.py index f67a5f805..1ae398789 100644 --- a/comfy/text_encoders/flux.py +++ b/comfy/text_encoders/flux.py @@ -118,7 +118,7 @@ class MistralTokenizerClass: class Mistral3Tokenizer(sd1_clip.SDTokenizer): def __init__(self, embedding_directory=None, tokenizer_data={}): self.tekken_data = tokenizer_data.get("tekken_model", None) - super().__init__("", pad_with_end=False, embedding_size=5120, embedding_key='mistral3_24b', tokenizer_class=MistralTokenizerClass, has_end_token=False, pad_to_max_length=False, pad_token=11, start_token=1, max_length=99999999, min_length=1, pad_left=True, tokenizer_args=load_mistral_tokenizer(self.tekken_data), tokenizer_data=tokenizer_data) + super().__init__("", pad_with_end=False, embedding_directory=embedding_directory, embedding_size=5120, embedding_key='mistral3_24b', tokenizer_class=MistralTokenizerClass, has_end_token=False, pad_to_max_length=False, pad_token=11, start_token=1, max_length=99999999, min_length=1, pad_left=True, tokenizer_args=load_mistral_tokenizer(self.tekken_data), tokenizer_data=tokenizer_data) def state_dict(self): return {"tekken_model": self.tekken_data} @@ -176,12 +176,12 @@ def flux2_te(dtype_llama=None, llama_quantization_metadata=None, pruned=False): class Qwen3Tokenizer(sd1_clip.SDTokenizer): def __init__(self, embedding_directory=None, tokenizer_data={}): tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "qwen25_tokenizer") - super().__init__(tokenizer_path, pad_with_end=False, embedding_size=2560, embedding_key='qwen3_4b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=512, pad_token=151643, tokenizer_data=tokenizer_data) + super().__init__(tokenizer_path, pad_with_end=False, embedding_directory=embedding_directory, embedding_size=2560, embedding_key='qwen3_4b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=512, pad_token=151643, tokenizer_data=tokenizer_data) class Qwen3Tokenizer8B(sd1_clip.SDTokenizer): def __init__(self, embedding_directory=None, tokenizer_data={}): tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "qwen25_tokenizer") - super().__init__(tokenizer_path, pad_with_end=False, embedding_size=4096, embedding_key='qwen3_8b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=512, pad_token=151643, tokenizer_data=tokenizer_data) + super().__init__(tokenizer_path, pad_with_end=False, embedding_directory=embedding_directory, embedding_size=4096, embedding_key='qwen3_8b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=512, pad_token=151643, tokenizer_data=tokenizer_data) class KleinTokenizer(sd1_clip.SD1Tokenizer): def __init__(self, embedding_directory=None, tokenizer_data={}, name="qwen3_4b"): diff --git a/comfy/text_encoders/z_image.py b/comfy/text_encoders/z_image.py index ad41bfb1e..33b7cf594 100644 --- a/comfy/text_encoders/z_image.py +++ b/comfy/text_encoders/z_image.py @@ -6,7 +6,7 @@ import os class Qwen3Tokenizer(sd1_clip.SDTokenizer): def __init__(self, embedding_directory=None, tokenizer_data={}): tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "qwen25_tokenizer") - super().__init__(tokenizer_path, pad_with_end=False, embedding_size=2560, embedding_key='qwen3_4b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, pad_token=151643, tokenizer_data=tokenizer_data) + super().__init__(tokenizer_path, pad_with_end=False, embedding_directory=embedding_directory, embedding_size=2560, embedding_key='qwen3_4b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, pad_token=151643, tokenizer_data=tokenizer_data) class ZImageTokenizer(sd1_clip.SD1Tokenizer): From 37f711d4a1d429e6b390b01729510525155385e1 Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Mon, 2 Feb 2026 14:34:46 -0800 Subject: [PATCH 553/647] mm: Fix cast buffers with intel offloading (#12229) Intel has offloading support but there were some nvidia calls in the new cast buffer stuff. --- comfy/model_management.py | 16 ++++++++++------ 1 file changed, 10 insertions(+), 6 deletions(-) diff --git a/comfy/model_management.py b/comfy/model_management.py index 6b1166b94..2167f81bf 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -1112,11 +1112,11 @@ def get_cast_buffer(offload_stream, device, size, ref): return None if cast_buffer is not None and cast_buffer.numel() > 50 * (1024 ** 2): #I want my wrongly sized 50MB+ of VRAM back from the caching allocator right now - torch.cuda.synchronize() + synchronize() del STREAM_CAST_BUFFERS[offload_stream] del cast_buffer #FIXME: This doesn't work in Aimdo because mempool cant clear cache - torch.cuda.empty_cache() + soft_empty_cache() with wf_context: cast_buffer = torch.empty((size), dtype=torch.int8, device=device) STREAM_CAST_BUFFERS[offload_stream] = cast_buffer @@ -1132,9 +1132,7 @@ def reset_cast_buffers(): for offload_stream in STREAM_CAST_BUFFERS: offload_stream.synchronize() STREAM_CAST_BUFFERS.clear() - if comfy.memory_management.aimdo_allocator is None: - #Pytorch 2.7 and earlier crashes if you try and empty_cache when mempools exist - torch.cuda.empty_cache() + soft_empty_cache() def get_offload_stream(device): stream_counter = stream_counters.get(device, 0) @@ -1284,7 +1282,7 @@ def discard_cuda_async_error(): a = torch.tensor([1], dtype=torch.uint8, device=get_torch_device()) b = torch.tensor([1], dtype=torch.uint8, device=get_torch_device()) _ = a + b - torch.cuda.synchronize() + synchronize() except torch.AcceleratorError: #Dump it! We already know about it from the synchronous return pass @@ -1688,6 +1686,12 @@ def lora_compute_dtype(device): LORA_COMPUTE_DTYPES[device] = dtype return dtype +def synchronize(): + if is_intel_xpu(): + torch.xpu.synchronize() + elif torch.cuda.is_available(): + torch.cuda.synchronize() + def soft_empty_cache(force=False): global cpu_state if cpu_state == CPUState.MPS: From de9ada6a4147f4626f903cd975a5d2c134af3915 Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Mon, 2 Feb 2026 14:35:20 -0800 Subject: [PATCH 554/647] Dynamic VRAM unloading fix (#12227) * mp: fix full dynamic unloading This was not unloading dynamic models when requesting a full unload via the unpatch() code path. This was ok, i your workflow was all dynamic models but fails with big VRAM leaks if you need to fully unload something for a regular ModelPatcher It also fices the "unload models" button. * mm: load models outside of Aimdo Mempool In dynamic_vram mode, escape the Aimdo mempool and load into the regular mempool. Use a dummy thread to do it. --- comfy/model_management.py | 29 ++++++++++++++++++++++------- comfy/model_patcher.py | 2 +- 2 files changed, 23 insertions(+), 8 deletions(-) diff --git a/comfy/model_management.py b/comfy/model_management.py index 2167f81bf..cd035f017 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -19,7 +19,8 @@ import psutil import logging from enum import Enum -from comfy.cli_args import args, PerformanceFeature +from comfy.cli_args import args, PerformanceFeature, enables_dynamic_vram +import threading import torch import sys import platform @@ -650,7 +651,7 @@ def free_memory(memory_required, device, keep_loaded=[], for_dynamic=False, ram_ soft_empty_cache() return unloaded_models -def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimum_memory_required=None, force_full_load=False): +def load_models_gpu_orig(models, memory_required=0, force_patch_weights=False, minimum_memory_required=None, force_full_load=False): cleanup_models_gc() global vram_state @@ -746,8 +747,25 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimu current_loaded_models.insert(0, loaded_model) return -def load_model_gpu(model): - return load_models_gpu([model]) +def load_models_gpu_thread(models, memory_required, force_patch_weights, minimum_memory_required, force_full_load): + with torch.inference_mode(): + load_models_gpu_orig(models, memory_required, force_patch_weights, minimum_memory_required, force_full_load) + soft_empty_cache() + +def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimum_memory_required=None, force_full_load=False): + #Deliberately load models outside of the Aimdo mempool so they can be retained accross + #nodes. Use a dummy thread to do it as pytorch documents that mempool contexts are + #thread local. So exploit that to escape context + if enables_dynamic_vram(): + t = threading.Thread( + target=load_models_gpu_thread, + args=(models, memory_required, force_patch_weights, minimum_memory_required, force_full_load) + ) + t.start() + t.join() + else: + load_models_gpu_orig(models, memory_required=memory_required, force_patch_weights=force_patch_weights, + minimum_memory_required=minimum_memory_required, force_full_load=force_full_load) def loaded_models(only_currently_used=False): output = [] @@ -1717,9 +1735,6 @@ def debug_memory_summary(): return torch.cuda.memory.memory_summary() return "" -#TODO: might be cleaner to put this somewhere else -import threading - class InterruptProcessingException(Exception): pass diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index b70c031bf..cdf289395 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -1597,7 +1597,7 @@ class ModelPatcherDynamic(ModelPatcher): if unpatch_weights: self.partially_unload_ram(1e32) - self.partially_unload(None) + self.partially_unload(None, 1e32) def partially_load(self, device_to, extra_memory=0, force_patch_weights=False): assert not force_patch_weights #See above From c05a08ae66d8114c30f8d3606b9de3117cb61ef8 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 2 Feb 2026 16:52:07 -0800 Subject: [PATCH 555/647] Add back function. (#12234) --- comfy/model_management.py | 3 +++ 1 file changed, 3 insertions(+) diff --git a/comfy/model_management.py b/comfy/model_management.py index cd035f017..72348258b 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -767,6 +767,9 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimu load_models_gpu_orig(models, memory_required=memory_required, force_patch_weights=force_patch_weights, minimum_memory_required=minimum_memory_required, force_full_load=force_full_load) +def load_model_gpu(model): + return load_models_gpu([model]) + def loaded_models(only_currently_used=False): output = [] for m in current_loaded_models: From ba5bf3f1a801c6d4c24f015215314b6a653f8752 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Tue, 3 Feb 2026 05:17:59 +0200 Subject: [PATCH 556/647] [API Nodes] HitPaw API nodes (#12117) * feat(api-nodes): add HitPaw API nodes * remove face_soft_2x model as not working --------- Co-authored-by: Robin Huang --- comfy_api_nodes/apis/hitpaw.py | 51 ++++ comfy_api_nodes/nodes_hitpaw.py | 342 +++++++++++++++++++++++++ comfy_api_nodes/util/upload_helpers.py | 2 +- 3 files changed, 394 insertions(+), 1 deletion(-) create mode 100644 comfy_api_nodes/apis/hitpaw.py create mode 100644 comfy_api_nodes/nodes_hitpaw.py diff --git a/comfy_api_nodes/apis/hitpaw.py b/comfy_api_nodes/apis/hitpaw.py new file mode 100644 index 000000000..b23c5d9eb --- /dev/null +++ b/comfy_api_nodes/apis/hitpaw.py @@ -0,0 +1,51 @@ +from typing import TypedDict + +from pydantic import BaseModel, Field + + +class InputVideoModel(TypedDict): + model: str + resolution: str + + +class ImageEnhanceTaskCreateRequest(BaseModel): + model_name: str = Field(...) + img_url: str = Field(...) + extension: str = Field(".png") + exif: bool = Field(False) + DPI: int | None = Field(None) + + +class VideoEnhanceTaskCreateRequest(BaseModel): + video_url: str = Field(...) + extension: str = Field(".mp4") + model_name: str | None = Field(...) + resolution: list[int] = Field(..., description="Target resolution [width, height]") + original_resolution: list[int] = Field(..., description="Original video resolution [width, height]") + + +class TaskCreateDataResponse(BaseModel): + job_id: str = Field(...) + consume_coins: int | None = Field(None) + + +class TaskStatusPollRequest(BaseModel): + job_id: str = Field(...) + + +class TaskCreateResponse(BaseModel): + code: int = Field(...) + message: str = Field(...) + data: TaskCreateDataResponse | None = Field(None) + + +class TaskStatusDataResponse(BaseModel): + job_id: str = Field(...) + status: str = Field(...) + res_url: str = Field("") + + +class TaskStatusResponse(BaseModel): + code: int = Field(...) + message: str = Field(...) + data: TaskStatusDataResponse = Field(...) diff --git a/comfy_api_nodes/nodes_hitpaw.py b/comfy_api_nodes/nodes_hitpaw.py new file mode 100644 index 000000000..488080a74 --- /dev/null +++ b/comfy_api_nodes/nodes_hitpaw.py @@ -0,0 +1,342 @@ +import math + +from typing_extensions import override + +from comfy_api.latest import IO, ComfyExtension, Input +from comfy_api_nodes.apis.hitpaw import ( + ImageEnhanceTaskCreateRequest, + InputVideoModel, + TaskCreateDataResponse, + TaskCreateResponse, + TaskStatusPollRequest, + TaskStatusResponse, + VideoEnhanceTaskCreateRequest, +) +from comfy_api_nodes.util import ( + ApiEndpoint, + download_url_to_image_tensor, + download_url_to_video_output, + downscale_image_tensor, + get_image_dimensions, + poll_op, + sync_op, + upload_image_to_comfyapi, + upload_video_to_comfyapi, + validate_video_duration, +) + +VIDEO_MODELS_MODELS_MAP = { + "Portrait Restore Model (1x)": "portrait_restore_1x", + "Portrait Restore Model (2x)": "portrait_restore_2x", + "General Restore Model (1x)": "general_restore_1x", + "General Restore Model (2x)": "general_restore_2x", + "General Restore Model (4x)": "general_restore_4x", + "Ultra HD Model (2x)": "ultrahd_restore_2x", + "Generative Model (1x)": "generative_1x", +} + +# Resolution name to target dimension (shorter side) in pixels +RESOLUTION_TARGET_MAP = { + "720p": 720, + "1080p": 1080, + "2K/QHD": 1440, + "4K/UHD": 2160, + "8K": 4320, +} + +# Square (1:1) resolutions use standard square dimensions +RESOLUTION_SQUARE_MAP = { + "720p": 720, + "1080p": 1080, + "2K/QHD": 1440, + "4K/UHD": 2048, # DCI 4K square + "8K": 4096, # DCI 8K square +} + +# Models with limited resolution support (no 8K) +LIMITED_RESOLUTION_MODELS = {"Generative Model (1x)"} + +# Resolution options for different model types +RESOLUTIONS_LIMITED = ["original", "720p", "1080p", "2K/QHD", "4K/UHD"] +RESOLUTIONS_FULL = ["original", "720p", "1080p", "2K/QHD", "4K/UHD", "8K"] + +# Maximum output resolution in pixels +MAX_PIXELS_GENERATIVE = 32_000_000 +MAX_MP_GENERATIVE = MAX_PIXELS_GENERATIVE // 1_000_000 + + +class HitPawGeneralImageEnhance(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="HitPawGeneralImageEnhance", + display_name="HitPaw General Image Enhance", + category="api node/image/HitPaw", + description="Upscale low-resolution images to super-resolution, eliminate artifacts and noise. " + f"Maximum output: {MAX_MP_GENERATIVE} megapixels.", + inputs=[ + IO.Combo.Input("model", options=["generative_portrait", "generative"]), + IO.Image.Input("image"), + IO.Combo.Input("upscale_factor", options=[1, 2, 4]), + IO.Boolean.Input( + "auto_downscale", + default=False, + tooltip="Automatically downscale input image if output would exceed the limit.", + ), + ], + outputs=[ + IO.Image.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["model"]), + expr=""" + ( + $prices := { + "generative_portrait": {"min": 0.02, "max": 0.06}, + "generative": {"min": 0.05, "max": 0.15} + }; + $price := $lookup($prices, widgets.model); + { + "type": "range_usd", + "min_usd": $price.min, + "max_usd": $price.max + } + ) + """, + ), + ) + + @classmethod + async def execute( + cls, + model: str, + image: Input.Image, + upscale_factor: int, + auto_downscale: bool, + ) -> IO.NodeOutput: + height, width = get_image_dimensions(image) + requested_scale = upscale_factor + output_pixels = height * width * requested_scale * requested_scale + if output_pixels > MAX_PIXELS_GENERATIVE: + if auto_downscale: + input_pixels = width * height + scale = 1 + max_input_pixels = MAX_PIXELS_GENERATIVE + + for candidate in [4, 2, 1]: + if candidate > requested_scale: + continue + scale_output_pixels = input_pixels * candidate * candidate + if scale_output_pixels <= MAX_PIXELS_GENERATIVE: + scale = candidate + max_input_pixels = None + break + # Check if we can downscale input by at most 2x to fit + downscale_ratio = math.sqrt(scale_output_pixels / MAX_PIXELS_GENERATIVE) + if downscale_ratio <= 2.0: + scale = candidate + max_input_pixels = MAX_PIXELS_GENERATIVE // (candidate * candidate) + break + + if max_input_pixels is not None: + image = downscale_image_tensor(image, total_pixels=max_input_pixels) + upscale_factor = scale + else: + output_width = width * requested_scale + output_height = height * requested_scale + raise ValueError( + f"Output size ({output_width}x{output_height} = {output_pixels:,} pixels) " + f"exceeds maximum allowed size of {MAX_PIXELS_GENERATIVE:,} pixels ({MAX_MP_GENERATIVE}MP). " + f"Enable auto_downscale or use a smaller input image or a lower upscale factor." + ) + + initial_res = await sync_op( + cls, + ApiEndpoint(path="/proxy/hitpaw/api/photo-enhancer", method="POST"), + response_model=TaskCreateResponse, + data=ImageEnhanceTaskCreateRequest( + model_name=f"{model}_{upscale_factor}x", + img_url=await upload_image_to_comfyapi(cls, image, total_pixels=None), + ), + wait_label="Creating task", + final_label_on_success="Task created", + ) + if initial_res.code != 200: + raise ValueError(f"Task creation failed with code {initial_res.code}: {initial_res.message}") + request_price = initial_res.data.consume_coins / 1000 + final_response = await poll_op( + cls, + ApiEndpoint(path="/proxy/hitpaw/api/task-status", method="POST"), + data=TaskCreateDataResponse(job_id=initial_res.data.job_id), + response_model=TaskStatusResponse, + status_extractor=lambda x: x.data.status, + price_extractor=lambda x: request_price, + poll_interval=10.0, + max_poll_attempts=480, + ) + return IO.NodeOutput(await download_url_to_image_tensor(final_response.data.res_url)) + + +class HitPawVideoEnhance(IO.ComfyNode): + @classmethod + def define_schema(cls): + model_options = [] + for model_name in VIDEO_MODELS_MODELS_MAP: + if model_name in LIMITED_RESOLUTION_MODELS: + resolutions = RESOLUTIONS_LIMITED + else: + resolutions = RESOLUTIONS_FULL + model_options.append( + IO.DynamicCombo.Option( + model_name, + [IO.Combo.Input("resolution", options=resolutions)], + ) + ) + + return IO.Schema( + node_id="HitPawVideoEnhance", + display_name="HitPaw Video Enhance", + category="api node/video/HitPaw", + description="Upscale low-resolution videos to high resolution, eliminate artifacts and noise. " + "Prices shown are per second of video.", + inputs=[ + IO.DynamicCombo.Input("model", options=model_options), + IO.Video.Input("video"), + ], + outputs=[ + IO.Video.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends(widgets=["model", "model.resolution"]), + expr=""" + ( + $m := $lookup(widgets, "model"); + $res := $lookup(widgets, "model.resolution"); + $standard_model_prices := { + "original": {"min": 0.01, "max": 0.198}, + "720p": {"min": 0.01, "max": 0.06}, + "1080p": {"min": 0.015, "max": 0.09}, + "2k/qhd": {"min": 0.02, "max": 0.117}, + "4k/uhd": {"min": 0.025, "max": 0.152}, + "8k": {"min": 0.033, "max": 0.198} + }; + $ultra_hd_model_prices := { + "original": {"min": 0.015, "max": 0.264}, + "720p": {"min": 0.015, "max": 0.092}, + "1080p": {"min": 0.02, "max": 0.12}, + "2k/qhd": {"min": 0.026, "max": 0.156}, + "4k/uhd": {"min": 0.034, "max": 0.203}, + "8k": {"min": 0.044, "max": 0.264} + }; + $generative_model_prices := { + "original": {"min": 0.015, "max": 0.338}, + "720p": {"min": 0.008, "max": 0.090}, + "1080p": {"min": 0.05, "max": 0.15}, + "2k/qhd": {"min": 0.038, "max": 0.225}, + "4k/uhd": {"min": 0.056, "max": 0.338} + }; + $prices := $contains($m, "ultra hd") ? $ultra_hd_model_prices : + $contains($m, "generative") ? $generative_model_prices : + $standard_model_prices; + $price := $lookup($prices, $res); + { + "type": "range_usd", + "min_usd": $price.min, + "max_usd": $price.max, + "format": {"approximate": true, "suffix": "/second"} + } + ) + """, + ), + ) + + @classmethod + async def execute( + cls, + model: InputVideoModel, + video: Input.Video, + ) -> IO.NodeOutput: + validate_video_duration(video, min_duration=0.5, max_duration=60 * 60) + resolution = model["resolution"] + src_width, src_height = video.get_dimensions() + + if resolution == "original": + output_width = src_width + output_height = src_height + else: + if src_width == src_height: + target_size = RESOLUTION_SQUARE_MAP[resolution] + if target_size < src_width: + raise ValueError( + f"Selected resolution {resolution} ({target_size}x{target_size}) is smaller than " + f"the input video ({src_width}x{src_height}). Please select a higher resolution or 'original'." + ) + output_width = target_size + output_height = target_size + else: + min_dimension = min(src_width, src_height) + target_size = RESOLUTION_TARGET_MAP[resolution] + if target_size < min_dimension: + raise ValueError( + f"Selected resolution {resolution} ({target_size}p) is smaller than " + f"the input video's shorter dimension ({min_dimension}p). " + f"Please select a higher resolution or 'original'." + ) + if src_width > src_height: + output_height = target_size + output_width = int(target_size * (src_width / src_height)) + else: + output_width = target_size + output_height = int(target_size * (src_height / src_width)) + initial_res = await sync_op( + cls, + ApiEndpoint(path="/proxy/hitpaw/api/video-enhancer", method="POST"), + response_model=TaskCreateResponse, + data=VideoEnhanceTaskCreateRequest( + video_url=await upload_video_to_comfyapi(cls, video), + resolution=[output_width, output_height], + original_resolution=[src_width, src_height], + model_name=VIDEO_MODELS_MODELS_MAP[model["model"]], + ), + wait_label="Creating task", + final_label_on_success="Task created", + ) + request_price = initial_res.data.consume_coins / 1000 + if initial_res.code != 200: + raise ValueError(f"Task creation failed with code {initial_res.code}: {initial_res.message}") + final_response = await poll_op( + cls, + ApiEndpoint(path="/proxy/hitpaw/api/task-status", method="POST"), + data=TaskStatusPollRequest(job_id=initial_res.data.job_id), + response_model=TaskStatusResponse, + status_extractor=lambda x: x.data.status, + price_extractor=lambda x: request_price, + poll_interval=10.0, + max_poll_attempts=320, + ) + return IO.NodeOutput(await download_url_to_video_output(final_response.data.res_url)) + + +class HitPawExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[IO.ComfyNode]]: + return [ + HitPawGeneralImageEnhance, + HitPawVideoEnhance, + ] + + +async def comfy_entrypoint() -> HitPawExtension: + return HitPawExtension() diff --git a/comfy_api_nodes/util/upload_helpers.py b/comfy_api_nodes/util/upload_helpers.py index 3153f2b98..83d936ce1 100644 --- a/comfy_api_nodes/util/upload_helpers.py +++ b/comfy_api_nodes/util/upload_helpers.py @@ -94,7 +94,7 @@ async def upload_image_to_comfyapi( *, mime_type: str | None = None, wait_label: str | None = "Uploading", - total_pixels: int = 2048 * 2048, + total_pixels: int | None = 2048 * 2048, ) -> str: """Uploads a single image to ComfyUI API and returns its download URL.""" return ( From 3c1a1a2df82fc6c7aa20a3c8301d1c632e1a1d87 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 2 Feb 2026 21:06:18 -0800 Subject: [PATCH 557/647] Basic support for the ace step 1.5 model. (#12237) --- comfy/latent_formats.py | 4 + comfy/ldm/ace/ace_step15.py | 1093 ++++++++++++++++++++++++++++++++++ comfy/model_base.py | 42 ++ comfy/model_detection.py | 5 + comfy/sd.py | 21 +- comfy/sd1_clip.py | 2 + comfy/supported_models.py | 35 +- comfy/text_encoders/ace15.py | 218 +++++++ comfy/text_encoders/llama.py | 67 +++ comfy_extras/nodes_ace.py | 75 +++ comfy_extras/nodes_audio.py | 10 +- nodes.py | 2 +- 12 files changed, 1566 insertions(+), 8 deletions(-) create mode 100644 comfy/ldm/ace/ace_step15.py create mode 100644 comfy/text_encoders/ace15.py diff --git a/comfy/latent_formats.py b/comfy/latent_formats.py index 4b3a3798c..f59999af6 100644 --- a/comfy/latent_formats.py +++ b/comfy/latent_formats.py @@ -755,6 +755,10 @@ class ACEAudio(LatentFormat): latent_channels = 8 latent_dimensions = 2 +class ACEAudio15(LatentFormat): + latent_channels = 64 + latent_dimensions = 1 + class ChromaRadiance(LatentFormat): latent_channels = 3 spacial_downscale_ratio = 1 diff --git a/comfy/ldm/ace/ace_step15.py b/comfy/ldm/ace/ace_step15.py new file mode 100644 index 000000000..d90549658 --- /dev/null +++ b/comfy/ldm/ace/ace_step15.py @@ -0,0 +1,1093 @@ +import math +import torch +import torch.nn as nn +import torch.nn.functional as F +import itertools +from comfy.ldm.modules.attention import optimized_attention +import comfy.model_management +from comfy.ldm.flux.layers import timestep_embedding + +def get_layer_class(operations, layer_name): + if operations is not None and hasattr(operations, layer_name): + return getattr(operations, layer_name) + return getattr(nn, layer_name) + +class RotaryEmbedding(nn.Module): + def __init__(self, dim, max_position_embeddings=32768, base=1000000.0, dtype=None, device=None, operations=None): + super().__init__() + self.dim = dim + self.base = base + self.max_position_embeddings = max_position_embeddings + + inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2, dtype=torch.float32, device=device) / self.dim)) + self.register_buffer("inv_freq", inv_freq, persistent=False) + self._set_cos_sin_cache(max_position_embeddings, device=device, dtype=torch.get_default_dtype() if dtype is None else dtype) + + def _set_cos_sin_cache(self, seq_len, device, dtype): + self.max_seq_len_cached = seq_len + t = torch.arange(self.max_seq_len_cached, device=device, dtype=torch.float32) + freqs = torch.outer(t, self.inv_freq) + emb = torch.cat((freqs, freqs), dim=-1) + self.register_buffer("cos_cached", emb.cos().to(dtype), persistent=False) + self.register_buffer("sin_cached", emb.sin().to(dtype), persistent=False) + + def forward(self, x, seq_len=None): + if seq_len > self.max_seq_len_cached: + self._set_cos_sin_cache(seq_len, x.device, x.dtype) + return ( + self.cos_cached[:seq_len].to(dtype=x.dtype, device=x.device), + self.sin_cached[:seq_len].to(dtype=x.dtype, device=x.device), + ) + +def rotate_half(x): + x1 = x[..., : x.shape[-1] // 2] + x2 = x[..., x.shape[-1] // 2 :] + return torch.cat((-x2, x1), dim=-1) + +def apply_rotary_pos_emb(q, k, cos, sin): + cos = cos.unsqueeze(0).unsqueeze(0) + sin = sin.unsqueeze(0).unsqueeze(0) + q_embed = (q * cos) + (rotate_half(q) * sin) + k_embed = (k * cos) + (rotate_half(k) * sin) + return q_embed, k_embed + +class MLP(nn.Module): + def __init__(self, hidden_size, intermediate_size, dtype=None, device=None, operations=None): + super().__init__() + Linear = get_layer_class(operations, "Linear") + self.gate_proj = Linear(hidden_size, intermediate_size, bias=False, dtype=dtype, device=device) + self.up_proj = Linear(hidden_size, intermediate_size, bias=False, dtype=dtype, device=device) + self.down_proj = Linear(intermediate_size, hidden_size, bias=False, dtype=dtype, device=device) + self.act_fn = nn.SiLU() + + def forward(self, x): + return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x)) + +class TimestepEmbedding(nn.Module): + def __init__(self, in_channels: int, time_embed_dim: int, scale: float = 1000, dtype=None, device=None, operations=None): + super().__init__() + Linear = get_layer_class(operations, "Linear") + self.linear_1 = Linear(in_channels, time_embed_dim, bias=True, dtype=dtype, device=device) + self.act1 = nn.SiLU() + self.linear_2 = Linear(time_embed_dim, time_embed_dim, bias=True, dtype=dtype, device=device) + self.in_channels = in_channels + self.act2 = nn.SiLU() + self.time_proj = Linear(time_embed_dim, time_embed_dim * 6, dtype=dtype, device=device) + self.scale = scale + + def forward(self, t, dtype=None): + t_freq = timestep_embedding(t, self.in_channels, time_factor=self.scale) + temb = self.linear_1(t_freq.to(dtype=dtype)) + temb = self.act1(temb) + temb = self.linear_2(temb) + timestep_proj = self.time_proj(self.act2(temb)).view(-1, 6, temb.shape[-1]) + return temb, timestep_proj + +class AceStepAttention(nn.Module): + def __init__( + self, + hidden_size, + num_heads, + num_kv_heads, + head_dim, + rms_norm_eps=1e-6, + is_cross_attention=False, + sliding_window=None, + dtype=None, + device=None, + operations=None + ): + super().__init__() + self.hidden_size = hidden_size + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = head_dim + self.is_cross_attention = is_cross_attention + self.sliding_window = sliding_window + + Linear = get_layer_class(operations, "Linear") + + self.q_proj = Linear(hidden_size, num_heads * head_dim, bias=False, dtype=dtype, device=device) + self.k_proj = Linear(hidden_size, num_kv_heads * head_dim, bias=False, dtype=dtype, device=device) + self.v_proj = Linear(hidden_size, num_kv_heads * head_dim, bias=False, dtype=dtype, device=device) + self.o_proj = Linear(num_heads * head_dim, hidden_size, bias=False, dtype=dtype, device=device) + + self.q_norm = operations.RMSNorm(head_dim, eps=rms_norm_eps, dtype=dtype, device=device) + self.k_norm = operations.RMSNorm(head_dim, eps=rms_norm_eps, dtype=dtype, device=device) + + def forward( + self, + hidden_states, + encoder_hidden_states=None, + attention_mask=None, + position_embeddings=None, + ): + bsz, q_len, _ = hidden_states.size() + + query_states = self.q_proj(hidden_states) + query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim) + query_states = self.q_norm(query_states) + query_states = query_states.transpose(1, 2) + + if self.is_cross_attention and encoder_hidden_states is not None: + bsz_enc, kv_len, _ = encoder_hidden_states.size() + key_states = self.k_proj(encoder_hidden_states) + value_states = self.v_proj(encoder_hidden_states) + + key_states = key_states.view(bsz_enc, kv_len, self.num_kv_heads, self.head_dim) + key_states = self.k_norm(key_states) + value_states = value_states.view(bsz_enc, kv_len, self.num_kv_heads, self.head_dim) + + key_states = key_states.transpose(1, 2) + value_states = value_states.transpose(1, 2) + else: + kv_len = q_len + key_states = self.k_proj(hidden_states) + value_states = self.v_proj(hidden_states) + + key_states = key_states.view(bsz, q_len, self.num_kv_heads, self.head_dim) + key_states = self.k_norm(key_states) + value_states = value_states.view(bsz, q_len, self.num_kv_heads, self.head_dim) + + key_states = key_states.transpose(1, 2) + value_states = value_states.transpose(1, 2) + + if position_embeddings is not None: + cos, sin = position_embeddings + query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin) + + n_rep = self.num_heads // self.num_kv_heads + if n_rep > 1: + key_states = key_states.repeat_interleave(n_rep, dim=1) + value_states = value_states.repeat_interleave(n_rep, dim=1) + + attn_bias = None + if self.sliding_window is not None and not self.is_cross_attention: + indices = torch.arange(q_len, device=query_states.device) + diff = indices.unsqueeze(1) - indices.unsqueeze(0) + in_window = torch.abs(diff) <= self.sliding_window + + window_bias = torch.zeros((q_len, kv_len), device=query_states.device, dtype=query_states.dtype) + min_value = torch.finfo(query_states.dtype).min + window_bias.masked_fill_(~in_window, min_value) + + window_bias = window_bias.unsqueeze(0).unsqueeze(0) + + if attn_bias is not None: + if attn_bias.dtype == torch.bool: + base_bias = torch.zeros_like(window_bias) + base_bias.masked_fill_(~attn_bias, min_value) + attn_bias = base_bias + window_bias + else: + attn_bias = attn_bias + window_bias + else: + attn_bias = window_bias + + attn_output = optimized_attention(query_states, key_states, value_states, self.num_heads, attn_bias, skip_reshape=True) + attn_output = self.o_proj(attn_output) + + return attn_output + +class AceStepDiTLayer(nn.Module): + def __init__( + self, + hidden_size, + num_heads, + num_kv_heads, + head_dim, + intermediate_size, + rms_norm_eps=1e-6, + layer_type="full_attention", + sliding_window=128, + dtype=None, + device=None, + operations=None + ): + super().__init__() + + self_attn_window = sliding_window if layer_type == "sliding_attention" else None + + self.self_attn_norm = operations.RMSNorm(hidden_size, eps=rms_norm_eps, dtype=dtype, device=device) + self.self_attn = AceStepAttention( + hidden_size, num_heads, num_kv_heads, head_dim, rms_norm_eps, + is_cross_attention=False, sliding_window=self_attn_window, + dtype=dtype, device=device, operations=operations + ) + + self.cross_attn_norm = operations.RMSNorm(hidden_size, eps=rms_norm_eps, dtype=dtype, device=device) + self.cross_attn = AceStepAttention( + hidden_size, num_heads, num_kv_heads, head_dim, rms_norm_eps, + is_cross_attention=True, dtype=dtype, device=device, operations=operations + ) + + self.mlp_norm = operations.RMSNorm(hidden_size, eps=rms_norm_eps, dtype=dtype, device=device) + self.mlp = MLP(hidden_size, intermediate_size, dtype=dtype, device=device, operations=operations) + + self.scale_shift_table = nn.Parameter(torch.empty(1, 6, hidden_size, dtype=dtype, device=device)) + + def forward( + self, + hidden_states, + temb, + encoder_hidden_states, + position_embeddings, + attention_mask=None, + encoder_attention_mask=None + ): + modulation = comfy.model_management.cast_to(self.scale_shift_table, dtype=temb.dtype, device=temb.device) + temb + shift_msa, scale_msa, gate_msa, c_shift_msa, c_scale_msa, c_gate_msa = modulation.chunk(6, dim=1) + + norm_hidden = self.self_attn_norm(hidden_states) + norm_hidden = norm_hidden * (1 + scale_msa) + shift_msa + + attn_out = self.self_attn( + norm_hidden, + position_embeddings=position_embeddings, + attention_mask=attention_mask + ) + hidden_states = hidden_states + attn_out * gate_msa + + norm_hidden = self.cross_attn_norm(hidden_states) + attn_out = self.cross_attn( + norm_hidden, + encoder_hidden_states=encoder_hidden_states, + attention_mask=encoder_attention_mask + ) + hidden_states = hidden_states + attn_out + + norm_hidden = self.mlp_norm(hidden_states) + norm_hidden = norm_hidden * (1 + c_scale_msa) + c_shift_msa + + mlp_out = self.mlp(norm_hidden) + hidden_states = hidden_states + mlp_out * c_gate_msa + + return hidden_states + +class AceStepEncoderLayer(nn.Module): + def __init__( + self, + hidden_size, + num_heads, + num_kv_heads, + head_dim, + intermediate_size, + rms_norm_eps=1e-6, + dtype=None, + device=None, + operations=None + ): + super().__init__() + self.self_attn = AceStepAttention( + hidden_size, num_heads, num_kv_heads, head_dim, rms_norm_eps, + is_cross_attention=False, dtype=dtype, device=device, operations=operations + ) + self.input_layernorm = operations.RMSNorm(hidden_size, eps=rms_norm_eps, dtype=dtype, device=device) + self.post_attention_layernorm = operations.RMSNorm(hidden_size, eps=rms_norm_eps, dtype=dtype, device=device) + self.mlp = MLP(hidden_size, intermediate_size, dtype=dtype, device=device, operations=operations) + + def forward(self, hidden_states, position_embeddings, attention_mask=None): + residual = hidden_states + hidden_states = self.input_layernorm(hidden_states) + hidden_states = self.self_attn( + hidden_states=hidden_states, + position_embeddings=position_embeddings, + attention_mask=attention_mask + ) + hidden_states = residual + hidden_states + + residual = hidden_states + hidden_states = self.post_attention_layernorm(hidden_states) + hidden_states = self.mlp(hidden_states) + hidden_states = residual + hidden_states + return hidden_states + +class AceStepLyricEncoder(nn.Module): + def __init__( + self, + text_hidden_dim, + hidden_size, + num_layers, + num_heads, + num_kv_heads, + head_dim, + intermediate_size, + rms_norm_eps=1e-6, + dtype=None, + device=None, + operations=None + ): + super().__init__() + Linear = get_layer_class(operations, "Linear") + self.embed_tokens = Linear(text_hidden_dim, hidden_size, dtype=dtype, device=device) + self.norm = operations.RMSNorm(hidden_size, eps=rms_norm_eps, dtype=dtype, device=device) + + self.rotary_emb = RotaryEmbedding( + head_dim, + base=1000000.0, + dtype=dtype, + device=device, + operations=operations + ) + + self.layers = nn.ModuleList([ + AceStepEncoderLayer( + hidden_size, num_heads, num_kv_heads, head_dim, intermediate_size, rms_norm_eps, + dtype=dtype, device=device, operations=operations + ) + for _ in range(num_layers) + ]) + + def forward(self, inputs_embeds, attention_mask=None): + hidden_states = self.embed_tokens(inputs_embeds) + seq_len = hidden_states.shape[1] + cos, sin = self.rotary_emb(hidden_states, seq_len=seq_len) + position_embeddings = (cos, sin) + + for layer in self.layers: + hidden_states = layer( + hidden_states, + position_embeddings=position_embeddings, + attention_mask=attention_mask + ) + + hidden_states = self.norm(hidden_states) + return hidden_states + +class AceStepTimbreEncoder(nn.Module): + def __init__( + self, + timbre_hidden_dim, + hidden_size, + num_layers, + num_heads, + num_kv_heads, + head_dim, + intermediate_size, + rms_norm_eps=1e-6, + dtype=None, + device=None, + operations=None + ): + super().__init__() + Linear = get_layer_class(operations, "Linear") + self.embed_tokens = Linear(timbre_hidden_dim, hidden_size, dtype=dtype, device=device) + self.norm = operations.RMSNorm(hidden_size, eps=rms_norm_eps, dtype=dtype, device=device) + + self.rotary_emb = RotaryEmbedding( + head_dim, + base=1000000.0, + dtype=dtype, + device=device, + operations=operations + ) + + self.layers = nn.ModuleList([ + AceStepEncoderLayer( + hidden_size, num_heads, num_kv_heads, head_dim, intermediate_size, rms_norm_eps, + dtype=dtype, device=device, operations=operations + ) + for _ in range(num_layers) + ]) + self.special_token = nn.Parameter(torch.empty(1, 1, hidden_size, device=device, dtype=dtype)) + + def unpack_timbre_embeddings(self, timbre_embs_packed, refer_audio_order_mask): + N, d = timbre_embs_packed.shape + device = timbre_embs_packed.device + B = N + counts = torch.bincount(refer_audio_order_mask, minlength=B) + max_count = counts.max().item() + + sorted_indices = torch.argsort( + refer_audio_order_mask * N + torch.arange(N, device=device), + stable=True + ) + sorted_batch_ids = refer_audio_order_mask[sorted_indices] + + positions = torch.arange(N, device=device) + batch_starts = torch.cat([torch.tensor([0], device=device), torch.cumsum(counts, dim=0)[:-1]]) + positions_in_sorted = positions - batch_starts[sorted_batch_ids] + + inverse_indices = torch.empty_like(sorted_indices) + inverse_indices[sorted_indices] = torch.arange(N, device=device) + positions_in_batch = positions_in_sorted[inverse_indices] + + indices_2d = refer_audio_order_mask * max_count + positions_in_batch + one_hot = F.one_hot(indices_2d, num_classes=B * max_count).to(timbre_embs_packed.dtype) + + timbre_embs_flat = one_hot.t() @ timbre_embs_packed + timbre_embs_unpack = timbre_embs_flat.view(B, max_count, d) + + mask_flat = (one_hot.sum(dim=0) > 0).long() + new_mask = mask_flat.view(B, max_count) + return timbre_embs_unpack, new_mask + + def forward(self, refer_audio_acoustic_hidden_states_packed, refer_audio_order_mask, attention_mask=None): + hidden_states = self.embed_tokens(refer_audio_acoustic_hidden_states_packed) + if hidden_states.dim() == 2: + hidden_states = hidden_states.unsqueeze(0) + + seq_len = hidden_states.shape[1] + cos, sin = self.rotary_emb(hidden_states, seq_len=seq_len) + + for layer in self.layers: + hidden_states = layer( + hidden_states, + position_embeddings=(cos, sin), + attention_mask=attention_mask + ) + hidden_states = self.norm(hidden_states) + + flat_states = hidden_states[:, 0, :] + unpacked_embs, unpacked_mask = self.unpack_timbre_embeddings(flat_states, refer_audio_order_mask) + return unpacked_embs, unpacked_mask + + +def pack_sequences(hidden1, hidden2, mask1, mask2): + hidden_cat = torch.cat([hidden1, hidden2], dim=1) + B, L, D = hidden_cat.shape + + if mask1 is not None and mask2 is not None: + mask_cat = torch.cat([mask1, mask2], dim=1) + sort_idx = mask_cat.argsort(dim=1, descending=True, stable=True) + gather_idx = sort_idx.unsqueeze(-1).expand(B, L, D) + hidden_sorted = torch.gather(hidden_cat, 1, gather_idx) + lengths = mask_cat.sum(dim=1) + new_mask = (torch.arange(L, device=hidden_cat.device).unsqueeze(0) < lengths.unsqueeze(1)) + else: + new_mask = None + hidden_sorted = hidden_cat + + return hidden_sorted, new_mask + +class AceStepConditionEncoder(nn.Module): + def __init__( + self, + text_hidden_dim, + timbre_hidden_dim, + hidden_size, + num_lyric_layers, + num_timbre_layers, + num_heads, + num_kv_heads, + head_dim, + intermediate_size, + rms_norm_eps=1e-6, + dtype=None, + device=None, + operations=None + ): + super().__init__() + Linear = get_layer_class(operations, "Linear") + self.text_projector = Linear(text_hidden_dim, hidden_size, bias=False, dtype=dtype, device=device) + + self.lyric_encoder = AceStepLyricEncoder( + text_hidden_dim=text_hidden_dim, + hidden_size=hidden_size, + num_layers=num_lyric_layers, + num_heads=num_heads, + num_kv_heads=num_kv_heads, + head_dim=head_dim, + intermediate_size=intermediate_size, + rms_norm_eps=rms_norm_eps, + dtype=dtype, + device=device, + operations=operations + ) + + self.timbre_encoder = AceStepTimbreEncoder( + timbre_hidden_dim=timbre_hidden_dim, + hidden_size=hidden_size, + num_layers=num_timbre_layers, + num_heads=num_heads, + num_kv_heads=num_kv_heads, + head_dim=head_dim, + intermediate_size=intermediate_size, + rms_norm_eps=rms_norm_eps, + dtype=dtype, + device=device, + operations=operations + ) + + def forward( + self, + text_hidden_states=None, + text_attention_mask=None, + lyric_hidden_states=None, + lyric_attention_mask=None, + refer_audio_acoustic_hidden_states_packed=None, + refer_audio_order_mask=None + ): + text_emb = self.text_projector(text_hidden_states) + + lyric_emb = self.lyric_encoder( + inputs_embeds=lyric_hidden_states, + attention_mask=lyric_attention_mask + ) + + timbre_emb, timbre_mask = self.timbre_encoder( + refer_audio_acoustic_hidden_states_packed, + refer_audio_order_mask + ) + + merged_emb, merged_mask = pack_sequences(lyric_emb, timbre_emb, lyric_attention_mask, timbre_mask) + final_emb, final_mask = pack_sequences(merged_emb, text_emb, merged_mask, text_attention_mask) + + return final_emb, final_mask + +# -------------------------------------------------------------------------------- +# Main Diffusion Model (DiT) +# -------------------------------------------------------------------------------- + +class AceStepDiTModel(nn.Module): + def __init__( + self, + in_channels, + hidden_size, + num_layers, + num_heads, + num_kv_heads, + head_dim, + intermediate_size, + patch_size, + audio_acoustic_hidden_dim, + layer_types=None, + sliding_window=128, + rms_norm_eps=1e-6, + dtype=None, + device=None, + operations=None + ): + super().__init__() + self.patch_size = patch_size + self.rotary_emb = RotaryEmbedding( + head_dim, + base=1000000.0, + dtype=dtype, + device=device, + operations=operations + ) + + Conv1d = get_layer_class(operations, "Conv1d") + ConvTranspose1d = get_layer_class(operations, "ConvTranspose1d") + Linear = get_layer_class(operations, "Linear") + + self.proj_in = nn.Sequential( + nn.Identity(), + Conv1d( + in_channels, hidden_size, kernel_size=patch_size, stride=patch_size, + dtype=dtype, device=device)) + + self.time_embed = TimestepEmbedding(256, hidden_size, dtype=dtype, device=device, operations=operations) + self.time_embed_r = TimestepEmbedding(256, hidden_size, dtype=dtype, device=device, operations=operations) + self.condition_embedder = Linear(hidden_size, hidden_size, dtype=dtype, device=device) + + if layer_types is None: + layer_types = ["full_attention"] * num_layers + + if len(layer_types) < num_layers: + layer_types = list(itertools.islice(itertools.cycle(layer_types), num_layers)) + + self.layers = nn.ModuleList([ + AceStepDiTLayer( + hidden_size, num_heads, num_kv_heads, head_dim, intermediate_size, rms_norm_eps, + layer_type=layer_types[i], + sliding_window=sliding_window, + dtype=dtype, device=device, operations=operations + ) for i in range(num_layers) + ]) + + self.norm_out = operations.RMSNorm(hidden_size, eps=rms_norm_eps, dtype=dtype, device=device) + self.proj_out = nn.Sequential( + nn.Identity(), + ConvTranspose1d(hidden_size, audio_acoustic_hidden_dim, kernel_size=patch_size, stride=patch_size, dtype=dtype, device=device) + ) + + self.scale_shift_table = nn.Parameter(torch.empty(1, 2, hidden_size, dtype=dtype, device=device)) + + def forward( + self, + hidden_states, + timestep, + timestep_r, + attention_mask, + encoder_hidden_states, + encoder_attention_mask, + context_latents + ): + temb_t, proj_t = self.time_embed(timestep, dtype=hidden_states.dtype) + temb_r, proj_r = self.time_embed_r(timestep - timestep_r, dtype=hidden_states.dtype) + temb = temb_t + temb_r + timestep_proj = proj_t + proj_r + + x = torch.cat([context_latents, hidden_states], dim=-1) + original_seq_len = x.shape[1] + + pad_length = 0 + if x.shape[1] % self.patch_size != 0: + pad_length = self.patch_size - (x.shape[1] % self.patch_size) + x = F.pad(x, (0, 0, 0, pad_length), mode='constant', value=0) + + x = x.transpose(1, 2) + x = self.proj_in(x) + x = x.transpose(1, 2) + + encoder_hidden_states = self.condition_embedder(encoder_hidden_states) + + seq_len = x.shape[1] + cos, sin = self.rotary_emb(x, seq_len=seq_len) + + for layer in self.layers: + x = layer( + hidden_states=x, + temb=timestep_proj, + encoder_hidden_states=encoder_hidden_states, + position_embeddings=(cos, sin), + attention_mask=None, + encoder_attention_mask=None + ) + + shift, scale = (comfy.model_management.cast_to(self.scale_shift_table, dtype=temb.dtype, device=temb.device) + temb.unsqueeze(1)).chunk(2, dim=1) + x = self.norm_out(x) * (1 + scale) + shift + + x = x.transpose(1, 2) + x = self.proj_out(x) + x = x.transpose(1, 2) + + x = x[:, :original_seq_len, :] + return x + + +class AttentionPooler(nn.Module): + def __init__(self, hidden_size, num_layers, head_dim, rms_norm_eps, dtype=None, device=None, operations=None): + super().__init__() + Linear = get_layer_class(operations, "Linear") + self.embed_tokens = Linear(hidden_size, hidden_size, dtype=dtype, device=device) + self.norm = operations.RMSNorm(hidden_size, eps=rms_norm_eps, dtype=dtype, device=device) + self.rotary_emb = RotaryEmbedding(head_dim, dtype=dtype, device=device, operations=operations) + self.special_token = nn.Parameter(torch.empty(1, 1, hidden_size, dtype=dtype, device=device)) + + self.layers = nn.ModuleList([ + AceStepEncoderLayer( + hidden_size, 16, 8, head_dim, hidden_size * 3, rms_norm_eps, + dtype=dtype, device=device, operations=operations + ) + for _ in range(num_layers) + ]) + + def forward(self, x): + B, T, P, D = x.shape + x = self.embed_tokens(x) + special = self.special_token.expand(B, T, 1, -1) + x = torch.cat([special, x], dim=2) + x = x.view(B * T, P + 1, D) + + cos, sin = self.rotary_emb(x, seq_len=P + 1) + for layer in self.layers: + x = layer(x, (cos, sin)) + + x = self.norm(x) + return x[:, 0, :].view(B, T, D) + + +class FSQ(nn.Module): + def __init__( + self, + levels, + dim=None, + device=None, + dtype=None, + operations=None + ): + super().__init__() + + _levels = torch.tensor(levels, dtype=torch.int32, device=device) + self.register_buffer('_levels', _levels, persistent=False) + + _basis = torch.cumprod(torch.tensor([1] + levels[:-1], dtype=torch.int32, device=device), dim=0) + self.register_buffer('_basis', _basis, persistent=False) + + self.codebook_dim = len(levels) + self.dim = dim if dim is not None else self.codebook_dim + + requires_projection = self.dim != self.codebook_dim + if requires_projection: + self.project_in = operations.Linear(self.dim, self.codebook_dim, device=device, dtype=dtype) + self.project_out = operations.Linear(self.codebook_dim, self.dim, device=device, dtype=dtype) + else: + self.project_in = nn.Identity() + self.project_out = nn.Identity() + + self.codebook_size = self._levels.prod().item() + + indices = torch.arange(self.codebook_size, device=device) + implicit_codebook = self._indices_to_codes(indices) + + if dtype is not None: + implicit_codebook = implicit_codebook.to(dtype) + + self.register_buffer('implicit_codebook', implicit_codebook, persistent=False) + + def bound(self, z): + levels_minus_1 = (self._levels - 1).to(z.dtype) + scale = 2. / levels_minus_1 + bracket = (levels_minus_1 * (torch.tanh(z) + 1) / 2.) + 0.5 + + zhat = bracket.floor() + bracket_ste = bracket + (zhat - bracket).detach() + + return scale * bracket_ste - 1. + + def _indices_to_codes(self, indices): + indices = indices.unsqueeze(-1) + codes_non_centered = (indices // self._basis) % self._levels + return codes_non_centered.float() * (2. / (self._levels.float() - 1)) - 1. + + def codes_to_indices(self, zhat): + zhat_normalized = (zhat + 1.) / (2. / (self._levels.to(zhat.dtype) - 1)) + return (zhat_normalized * self._basis.to(zhat.dtype)).sum(dim=-1).round().to(torch.int32) + + def forward(self, z): + orig_dtype = z.dtype + z = self.project_in(z) + + codes = self.bound(z) + indices = self.codes_to_indices(codes) + + out = self.project_out(codes) + return out.to(orig_dtype), indices + + +class ResidualFSQ(nn.Module): + def __init__( + self, + levels, + num_quantizers, + dim=None, + bound_hard_clamp=True, + device=None, + dtype=None, + operations=None, + **kwargs + ): + super().__init__() + + codebook_dim = len(levels) + dim = dim if dim is not None else codebook_dim + + requires_projection = codebook_dim != dim + if requires_projection: + self.project_in = operations.Linear(dim, codebook_dim, device=device, dtype=dtype) + self.project_out = operations.Linear(codebook_dim, dim, device=device, dtype=dtype) + else: + self.project_in = nn.Identity() + self.project_out = nn.Identity() + + self.layers = nn.ModuleList() + levels_tensor = torch.tensor(levels, device=device) + scales = [] + + for ind in range(num_quantizers): + scale_val = levels_tensor.float() ** -ind + scales.append(scale_val) + + self.layers.append(FSQ( + levels=levels, + dim=codebook_dim, + device=device, + dtype=dtype, + operations=operations + )) + + scales_tensor = torch.stack(scales) + if dtype is not None: + scales_tensor = scales_tensor.to(dtype) + self.register_buffer('scales', scales_tensor, persistent=False) + + if bound_hard_clamp: + val = 1 + (1 / (levels_tensor.float() - 1)) + if dtype is not None: + val = val.to(dtype) + self.register_buffer('soft_clamp_input_value', val, persistent=False) + + def get_output_from_indices(self, indices, dtype=torch.float32): + if indices.dim() == 2: + indices = indices.unsqueeze(-1) + + all_codes = [] + for i, layer in enumerate(self.layers): + idx = indices[..., i].long() + codes = F.embedding(idx, comfy.model_management.cast_to(layer.implicit_codebook, device=idx.device, dtype=dtype)) + all_codes.append(codes * comfy.model_management.cast_to(self.scales[i], device=idx.device, dtype=dtype)) + + codes_summed = torch.stack(all_codes).sum(dim=0) + return self.project_out(codes_summed) + + def forward(self, x): + x = self.project_in(x) + + if hasattr(self, 'soft_clamp_input_value'): + sc_val = self.soft_clamp_input_value.to(x.dtype) + x = (x / sc_val).tanh() * sc_val + + quantized_out = torch.tensor(0., device=x.device, dtype=x.dtype) + residual = x + all_indices = [] + + for layer, scale in zip(self.layers, self.scales): + scale = scale.to(residual.dtype) + + quantized, indices = layer(residual / scale) + quantized = quantized * scale + + residual = residual - quantized.detach() + quantized_out = quantized_out + quantized + all_indices.append(indices) + + quantized_out = self.project_out(quantized_out) + all_indices = torch.stack(all_indices, dim=-1) + + return quantized_out, all_indices + + +class AceStepAudioTokenizer(nn.Module): + def __init__( + self, + audio_acoustic_hidden_dim, + hidden_size, + pool_window_size, + fsq_dim, + fsq_levels, + fsq_input_num_quantizers, + num_layers, + head_dim, + rms_norm_eps, + dtype=None, + device=None, + operations=None + ): + super().__init__() + Linear = get_layer_class(operations, "Linear") + self.audio_acoustic_proj = Linear(audio_acoustic_hidden_dim, hidden_size, dtype=dtype, device=device) + self.attention_pooler = AttentionPooler( + hidden_size, num_layers, head_dim, rms_norm_eps, dtype=dtype, device=device, operations=operations + ) + self.pool_window_size = pool_window_size + self.fsq_dim = fsq_dim + self.quantizer = ResidualFSQ( + dim=fsq_dim, + levels=fsq_levels, + num_quantizers=fsq_input_num_quantizers, + bound_hard_clamp=True, + dtype=dtype, device=device, operations=operations + ) + + def forward(self, hidden_states): + hidden_states = self.audio_acoustic_proj(hidden_states) + hidden_states = self.attention_pooler(hidden_states) + quantized, indices = self.quantizer(hidden_states) + return quantized, indices + + def tokenize(self, x): + B, T, D = x.shape + P = self.pool_window_size + + if T % P != 0: + pad = P - (T % P) + x = F.pad(x, (0, 0, 0, pad)) + T = x.shape[1] + + T_patch = T // P + x = x.view(B, T_patch, P, D) + + quantized, indices = self.forward(x) + return quantized, indices + + +class AudioTokenDetokenizer(nn.Module): + def __init__( + self, + hidden_size, + pool_window_size, + audio_acoustic_hidden_dim, + num_layers, + head_dim, + dtype=None, + device=None, + operations=None + ): + super().__init__() + Linear = get_layer_class(operations, "Linear") + self.pool_window_size = pool_window_size + self.embed_tokens = Linear(hidden_size, hidden_size, dtype=dtype, device=device) + self.special_tokens = nn.Parameter(torch.empty(1, pool_window_size, hidden_size, dtype=dtype, device=device)) + self.rotary_emb = RotaryEmbedding(head_dim, dtype=dtype, device=device, operations=operations) + self.layers = nn.ModuleList([ + AceStepEncoderLayer( + hidden_size, 16, 8, head_dim, hidden_size * 3, 1e-6, + dtype=dtype, device=device, operations=operations + ) + for _ in range(num_layers) + ]) + self.norm = operations.RMSNorm(hidden_size, dtype=dtype, device=device) + self.proj_out = Linear(hidden_size, audio_acoustic_hidden_dim, dtype=dtype, device=device) + + def forward(self, x): + B, T, D = x.shape + x = self.embed_tokens(x) + x = x.unsqueeze(2).repeat(1, 1, self.pool_window_size, 1) + x = x + comfy.model_management.cast_to(self.special_tokens.expand(B, T, -1, -1), device=x.device, dtype=x.dtype) + x = x.view(B * T, self.pool_window_size, D) + + cos, sin = self.rotary_emb(x, seq_len=self.pool_window_size) + for layer in self.layers: + x = layer(x, (cos, sin)) + + x = self.norm(x) + x = self.proj_out(x) + return x.view(B, T * self.pool_window_size, -1) + + +class AceStepConditionGenerationModel(nn.Module): + def __init__( + self, + in_channels=192, + hidden_size=2048, + text_hidden_dim=1024, + timbre_hidden_dim=64, + audio_acoustic_hidden_dim=64, + num_dit_layers=24, + num_lyric_layers=8, + num_timbre_layers=4, + num_tokenizer_layers=2, + num_heads=16, + num_kv_heads=8, + head_dim=128, + intermediate_size=6144, + patch_size=2, + pool_window_size=5, + rms_norm_eps=1e-06, + timestep_mu=-0.4, + timestep_sigma=1.0, + data_proportion=0.5, + sliding_window=128, + layer_types=None, + fsq_dim=2048, + fsq_levels=[8, 8, 8, 5, 5, 5], + fsq_input_num_quantizers=1, + audio_model=None, + dtype=None, + device=None, + operations=None + ): + super().__init__() + self.dtype = dtype + self.timestep_mu = timestep_mu + self.timestep_sigma = timestep_sigma + self.data_proportion = data_proportion + self.pool_window_size = pool_window_size + + if layer_types is None: + layer_types = [] + for i in range(num_dit_layers): + layer_types.append("sliding_attention" if i % 2 == 0 else "full_attention") + + self.decoder = AceStepDiTModel( + in_channels, hidden_size, num_dit_layers, num_heads, num_kv_heads, head_dim, + intermediate_size, patch_size, audio_acoustic_hidden_dim, + layer_types=layer_types, sliding_window=sliding_window, rms_norm_eps=rms_norm_eps, + dtype=dtype, device=device, operations=operations + ) + self.encoder = AceStepConditionEncoder( + text_hidden_dim, timbre_hidden_dim, hidden_size, num_lyric_layers, num_timbre_layers, + num_heads, num_kv_heads, head_dim, intermediate_size, rms_norm_eps, + dtype=dtype, device=device, operations=operations + ) + self.tokenizer = AceStepAudioTokenizer( + audio_acoustic_hidden_dim, hidden_size, pool_window_size, fsq_dim=fsq_dim, fsq_levels=fsq_levels, fsq_input_num_quantizers=fsq_input_num_quantizers, num_layers=num_tokenizer_layers, head_dim=head_dim, rms_norm_eps=rms_norm_eps, + dtype=dtype, device=device, operations=operations + ) + self.detokenizer = AudioTokenDetokenizer( + hidden_size, pool_window_size, audio_acoustic_hidden_dim, num_layers=2, head_dim=head_dim, + dtype=dtype, device=device, operations=operations + ) + self.null_condition_emb = nn.Parameter(torch.empty(1, 1, hidden_size, dtype=dtype, device=device)) + + def prepare_condition( + self, + text_hidden_states, text_attention_mask, + lyric_hidden_states, lyric_attention_mask, + refer_audio_acoustic_hidden_states_packed, refer_audio_order_mask, + src_latents, chunk_masks, is_covers, + precomputed_lm_hints_25Hz=None, + audio_codes=None + ): + encoder_hidden, encoder_mask = self.encoder( + text_hidden_states, text_attention_mask, + lyric_hidden_states, lyric_attention_mask, + refer_audio_acoustic_hidden_states_packed, refer_audio_order_mask + ) + + if precomputed_lm_hints_25Hz is not None: + lm_hints = precomputed_lm_hints_25Hz + else: + if audio_codes is not None: + if audio_codes.shape[1] * 5 < src_latents.shape[1]: + audio_codes = torch.nn.functional.pad(audio_codes, (0, math.ceil(src_latents.shape[1] / 5) - audio_codes.shape[1]), "constant", 35847) + lm_hints_5Hz = self.tokenizer.quantizer.get_output_from_indices(audio_codes, dtype=text_hidden_states.dtype) + else: + assert False + # TODO ? + + lm_hints = self.detokenizer(lm_hints_5Hz) + + lm_hints = lm_hints[:, :src_latents.shape[1], :] + if is_covers is None: + src_latents = lm_hints + else: + src_latents = torch.where(is_covers.unsqueeze(-1).unsqueeze(-1) > 0, lm_hints, src_latents) + + context_latents = torch.cat([src_latents, chunk_masks.to(src_latents.dtype)], dim=-1) + + return encoder_hidden, encoder_mask, context_latents + + def forward(self, x, timestep, context, lyric_embed=None, refer_audio=None, audio_codes=None, **kwargs): + text_attention_mask = None + lyric_attention_mask = None + refer_audio_order_mask = None + attention_mask = None + chunk_masks = None + is_covers = None + src_latents = None + precomputed_lm_hints_25Hz = None + lyric_hidden_states = lyric_embed + text_hidden_states = context + refer_audio_acoustic_hidden_states_packed = refer_audio.movedim(-1, -2) + + x = x.movedim(-1, -2) + + if refer_audio_order_mask is None: + refer_audio_order_mask = torch.zeros((x.shape[0],), device=x.device, dtype=torch.long) + + if src_latents is None and is_covers is None: + src_latents = x + + if chunk_masks is None: + chunk_masks = torch.ones_like(x) + + enc_hidden, enc_mask, context_latents = self.prepare_condition( + text_hidden_states, text_attention_mask, + lyric_hidden_states, lyric_attention_mask, + refer_audio_acoustic_hidden_states_packed, refer_audio_order_mask, + src_latents, chunk_masks, is_covers, precomputed_lm_hints_25Hz=precomputed_lm_hints_25Hz, audio_codes=audio_codes + ) + + out = self.decoder(hidden_states=x, + timestep=timestep, + timestep_r=timestep, + attention_mask=attention_mask, + encoder_hidden_states=enc_hidden, + encoder_attention_mask=enc_mask, + context_latents=context_latents + ) + + return out.movedim(-1, -2) diff --git a/comfy/model_base.py b/comfy/model_base.py index 85acdb66a..89944548c 100644 --- a/comfy/model_base.py +++ b/comfy/model_base.py @@ -50,6 +50,7 @@ import comfy.ldm.omnigen.omnigen2 import comfy.ldm.qwen_image.model import comfy.ldm.kandinsky5.model import comfy.ldm.anima.model +import comfy.ldm.ace.ace_step15 import comfy.model_management import comfy.patcher_extension @@ -1540,6 +1541,47 @@ class ACEStep(BaseModel): out['lyrics_strength'] = comfy.conds.CONDConstant(kwargs.get("lyrics_strength", 1.0)) return out +class ACEStep15(BaseModel): + def __init__(self, model_config, model_type=ModelType.FLOW, device=None): + super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.ace.ace_step15.AceStepConditionGenerationModel) + + def extra_conds(self, **kwargs): + out = super().extra_conds(**kwargs) + device = kwargs["device"] + + cross_attn = kwargs.get("cross_attn", None) + if cross_attn is not None: + out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn) + + conditioning_lyrics = kwargs.get("conditioning_lyrics", None) + if cross_attn is not None: + out['lyric_embed'] = comfy.conds.CONDRegular(conditioning_lyrics) + + refer_audio = kwargs.get("reference_audio_timbre_latents", None) + if refer_audio is None or len(refer_audio) == 0: + refer_audio = torch.tensor([[[-1.3672e-01, -1.5820e-01, 5.8594e-01, -5.7422e-01, 3.0273e-02, + 2.7930e-01, -2.5940e-03, -2.0703e-01, -1.6113e-01, -1.4746e-01, + -2.7710e-02, -1.8066e-01, -2.9688e-01, 1.6016e+00, -2.6719e+00, + 7.7734e-01, -1.3516e+00, -1.9434e-01, -7.1289e-02, -5.0938e+00, + 2.4316e-01, 4.7266e-01, 4.6387e-02, -6.6406e-01, -2.1973e-01, + -6.7578e-01, -1.5723e-01, 9.5312e-01, -2.0020e-01, -1.7109e+00, + 5.8984e-01, -5.7422e-01, 5.1562e-01, 2.8320e-01, 1.4551e-01, + -1.8750e-01, -5.9814e-02, 3.6719e-01, -1.0059e-01, -1.5723e-01, + 2.0605e-01, -4.3359e-01, -8.2812e-01, 4.5654e-02, -6.6016e-01, + 1.4844e-01, 9.4727e-02, 3.8477e-01, -1.2578e+00, -3.3203e-01, + -8.5547e-01, 4.3359e-01, 4.2383e-01, -8.9453e-01, -5.0391e-01, + -5.6152e-02, -2.9219e+00, -2.4658e-02, 5.0391e-01, 9.8438e-01, + 7.2754e-02, -2.1582e-01, 6.3672e-01, 1.0000e+00]]], device=device).movedim(-1, 1).repeat(1, 1, 750) + else: + refer_audio = refer_audio[-1] + out['refer_audio'] = comfy.conds.CONDRegular(refer_audio) + + audio_codes = kwargs.get("audio_codes", None) + if audio_codes is not None: + out['audio_codes'] = comfy.conds.CONDRegular(torch.tensor(audio_codes, device=device)) + + return out + class Omnigen2(BaseModel): def __init__(self, model_config, model_type=ModelType.FLOW, device=None): super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.omnigen.omnigen2.OmniGen2Transformer2DModel) diff --git a/comfy/model_detection.py b/comfy/model_detection.py index 8cea16e50..e8ad725df 100644 --- a/comfy/model_detection.py +++ b/comfy/model_detection.py @@ -655,6 +655,11 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): dit_config["num_visual_blocks"] = count_blocks(state_dict_keys, '{}visual_transformer_blocks.'.format(key_prefix) + '{}.') return dit_config + if '{}encoder.lyric_encoder.layers.0.input_layernorm.weight'.format(key_prefix) in state_dict_keys: + dit_config = {} + dit_config["audio_model"] = "ace1.5" + return dit_config + if '{}input_blocks.0.0.weight'.format(key_prefix) not in state_dict_keys: return None diff --git a/comfy/sd.py b/comfy/sd.py index fd0ac85e7..722c0c154 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -59,6 +59,7 @@ import comfy.text_encoders.kandinsky5 import comfy.text_encoders.jina_clip_2 import comfy.text_encoders.newbie import comfy.text_encoders.anima +import comfy.text_encoders.ace15 import comfy.model_patcher import comfy.lora @@ -452,6 +453,8 @@ class VAE: self.extra_1d_channel = None self.crop_input = True + self.audio_sample_rate = 44100 + if config is None: if "decoder.mid.block_1.mix_factor" in sd: encoder_config = {'double_z': True, 'z_channels': 4, 'resolution': 256, 'in_channels': 3, 'out_ch': 3, 'ch': 128, 'ch_mult': [1, 2, 4, 4], 'num_res_blocks': 2, 'attn_resolutions': [], 'dropout': 0.0} @@ -549,14 +552,25 @@ class VAE: encoder_config={'target': "comfy.ldm.modules.diffusionmodules.model.Encoder", 'params': ddconfig}, decoder_config={'target': "comfy.ldm.modules.diffusionmodules.model.Decoder", 'params': ddconfig}) elif "decoder.layers.1.layers.0.beta" in sd: - self.first_stage_model = AudioOobleckVAE() + config = {} + param_key = None + if "decoder.layers.2.layers.1.weight_v" in sd: + param_key = "decoder.layers.2.layers.1.weight_v" + if "decoder.layers.2.layers.1.parametrizations.weight.original1" in sd: + param_key = "decoder.layers.2.layers.1.parametrizations.weight.original1" + if param_key is not None: + if sd[param_key].shape[-1] == 12: + config["strides"] = [2, 4, 4, 6, 10] + self.audio_sample_rate = 48000 + + self.first_stage_model = AudioOobleckVAE(**config) self.memory_used_encode = lambda shape, dtype: (1000 * shape[2]) * model_management.dtype_size(dtype) self.memory_used_decode = lambda shape, dtype: (1000 * shape[2] * 2048) * model_management.dtype_size(dtype) self.latent_channels = 64 self.output_channels = 2 self.pad_channel_value = "replicate" self.upscale_ratio = 2048 - self.downscale_ratio = 2048 + self.downscale_ratio = 2048 self.latent_dim = 1 self.process_output = lambda audio: audio self.process_input = lambda audio: audio @@ -1427,6 +1441,9 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip clip_data_jina = clip_data[0] tokenizer_data["gemma_spiece_model"] = clip_data_gemma.get("spiece_model", None) tokenizer_data["jina_spiece_model"] = clip_data_jina.get("spiece_model", None) + elif clip_type == CLIPType.ACE: + clip_target.clip = comfy.text_encoders.ace15.te(**llama_detect(clip_data)) + clip_target.tokenizer = comfy.text_encoders.ace15.ACE15Tokenizer else: clip_target.clip = sdxl_clip.SDXLClipModel clip_target.tokenizer = sdxl_clip.SDXLTokenizer diff --git a/comfy/sd1_clip.py b/comfy/sd1_clip.py index 9ecfc9c55..4c817d468 100644 --- a/comfy/sd1_clip.py +++ b/comfy/sd1_clip.py @@ -155,6 +155,8 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder): self.execution_device = options.get("execution_device", self.execution_device) if isinstance(self.layer, list) or self.layer == "all": pass + elif isinstance(layer_idx, list): + self.layer = layer_idx elif layer_idx is None or abs(layer_idx) > self.num_layers: self.layer = "last" else: diff --git a/comfy/supported_models.py b/comfy/supported_models.py index d25271d6e..6b7d831cb 100644 --- a/comfy/supported_models.py +++ b/comfy/supported_models.py @@ -24,6 +24,7 @@ import comfy.text_encoders.hunyuan_image import comfy.text_encoders.kandinsky5 import comfy.text_encoders.z_image import comfy.text_encoders.anima +import comfy.text_encoders.ace15 from . import supported_models_base from . import latent_formats @@ -1596,6 +1597,38 @@ class Kandinsky5Image(Kandinsky5): return supported_models_base.ClipTarget(comfy.text_encoders.kandinsky5.Kandinsky5TokenizerImage, comfy.text_encoders.kandinsky5.te(**hunyuan_detect)) -models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, LTXAV, HunyuanVideo15_SR_Distilled, HunyuanVideo15, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, ZImage, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, Omnigen2, QwenImage, Flux2, Kandinsky5Image, Kandinsky5, Anima] +class ACEStep15(supported_models_base.BASE): + unet_config = { + "audio_model": "ace1.5", + } + + unet_extra_config = { + } + + sampling_settings = { + "multiplier": 1.0, + "shift": 3.0, + } + + latent_format = comfy.latent_formats.ACEAudio15 + + memory_usage_factor = 4.7 + + supported_inference_dtypes = [torch.bfloat16, torch.float32] + + vae_key_prefix = ["vae."] + text_encoder_key_prefix = ["text_encoders."] + + def get_model(self, state_dict, prefix="", device=None): + out = model_base.ACEStep15(self, device=device) + return out + + def clip_target(self, state_dict={}): + pref = self.text_encoder_key_prefix[0] + hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen3_2b.transformer.".format(pref)) + return supported_models_base.ClipTarget(comfy.text_encoders.ace15.ACE15Tokenizer, comfy.text_encoders.ace15.te(**hunyuan_detect)) + + +models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, LTXAV, HunyuanVideo15_SR_Distilled, HunyuanVideo15, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, ZImage, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, ACEStep15, Omnigen2, QwenImage, Flux2, Kandinsky5Image, Kandinsky5, Anima] models += [SVD_img2vid] diff --git a/comfy/text_encoders/ace15.py b/comfy/text_encoders/ace15.py new file mode 100644 index 000000000..9070cb577 --- /dev/null +++ b/comfy/text_encoders/ace15.py @@ -0,0 +1,218 @@ +from .anima import Qwen3Tokenizer +import comfy.text_encoders.llama +from comfy import sd1_clip +import torch +import math + + +def sample_manual_loop_no_classes( + model, + ids=None, + paddings=[], + execution_dtype=None, + cfg_scale: float = 2.0, + temperature: float = 0.85, + top_p: float = 0.9, + top_k: int = None, + seed: int = 1, + min_tokens: int = 1, + max_new_tokens: int = 2048, + audio_start_id: int = 151669, # The cutoff ID for audio codes + eos_token_id: int = 151645, +): + device = model.execution_device + + if execution_dtype is None: + if comfy.model_management.should_use_bf16(device): + execution_dtype = torch.bfloat16 + else: + execution_dtype = torch.float32 + + embeds, attention_mask, num_tokens, embeds_info = model.process_tokens(ids, device) + for i, t in enumerate(paddings): + attention_mask[i, :t] = 0 + attention_mask[i, t:] = 1 + + output_audio_codes = [] + past_key_values = [] + generator = torch.Generator(device=device) + generator.manual_seed(seed) + model_config = model.transformer.model.config + + for x in range(model_config.num_hidden_layers): + past_key_values.append((torch.empty([embeds.shape[0], model_config.num_key_value_heads, embeds.shape[1] + min_tokens, model_config.head_dim], device=device, dtype=execution_dtype), torch.empty([embeds.shape[0], model_config.num_key_value_heads, embeds.shape[1] + min_tokens, model_config.head_dim], device=device, dtype=execution_dtype), 0)) + + for step in range(max_new_tokens): + outputs = model.transformer(None, attention_mask, embeds=embeds.to(execution_dtype), num_tokens=num_tokens, intermediate_output=None, dtype=execution_dtype, embeds_info=embeds_info, past_key_values=past_key_values) + next_token_logits = model.transformer.logits(outputs[0])[:, -1] + past_key_values = outputs[2] + + cond_logits = next_token_logits[0:1] + uncond_logits = next_token_logits[1:2] + cfg_logits = uncond_logits + cfg_scale * (cond_logits - uncond_logits) + + if eos_token_id is not None and eos_token_id < audio_start_id and min_tokens < step: + eos_score = cfg_logits[:, eos_token_id].clone() + + # Only generate audio tokens + cfg_logits[:, :audio_start_id] = float('-inf') + + if eos_token_id is not None and eos_token_id < audio_start_id and min_tokens < step: + cfg_logits[:, eos_token_id] = eos_score + + if top_k is not None and top_k > 0: + top_k_vals, _ = torch.topk(cfg_logits, top_k) + min_val = top_k_vals[..., -1, None] + cfg_logits[cfg_logits < min_val] = float('-inf') + + if top_p is not None and top_p < 1.0: + sorted_logits, sorted_indices = torch.sort(cfg_logits, descending=True) + cumulative_probs = torch.cumsum(torch.softmax(sorted_logits, dim=-1), dim=-1) + sorted_indices_to_remove = cumulative_probs > top_p + sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone() + sorted_indices_to_remove[..., 0] = 0 + indices_to_remove = sorted_indices_to_remove.scatter(1, sorted_indices, sorted_indices_to_remove) + cfg_logits[indices_to_remove] = float('-inf') + + if temperature > 0: + cfg_logits = cfg_logits / temperature + next_token = torch.multinomial(torch.softmax(cfg_logits, dim=-1), num_samples=1, generator=generator).squeeze(1) + else: + next_token = torch.argmax(cfg_logits, dim=-1) + + token = next_token.item() + + if token == eos_token_id: + break + + embed, _, _, _ = model.process_tokens([[token]], device) + embeds = embed.repeat(2, 1, 1) + attention_mask = torch.cat([attention_mask, torch.ones((2, 1), device=device, dtype=attention_mask.dtype)], dim=1) + + output_audio_codes.append(token - audio_start_id) + + return output_audio_codes + + +def generate_audio_codes(model, positive, negative, min_tokens=1, max_tokens=1024, seed=0): + cfg_scale = 2.0 + + positive = [[token for token, _ in inner_list] for inner_list in positive] + negative = [[token for token, _ in inner_list] for inner_list in negative] + positive = positive[0] + negative = negative[0] + + neg_pad = 0 + if len(negative) < len(positive): + neg_pad = (len(positive) - len(negative)) + negative = [model.special_tokens["pad"]] * neg_pad + negative + + pos_pad = 0 + if len(negative) > len(positive): + pos_pad = (len(negative) - len(positive)) + positive = [model.special_tokens["pad"]] * pos_pad + positive + + paddings = [pos_pad, neg_pad] + return sample_manual_loop_no_classes(model, [positive, negative], paddings, cfg_scale=cfg_scale, seed=seed, min_tokens=min_tokens, max_new_tokens=max_tokens) + + +class ACE15Tokenizer(sd1_clip.SD1Tokenizer): + def __init__(self, embedding_directory=None, tokenizer_data={}): + super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, name="qwen3_06b", tokenizer=Qwen3Tokenizer) + + def tokenize_with_weights(self, text, return_word_ids=False, **kwargs): + out = {} + lyrics = kwargs.get("lyrics", "") + bpm = kwargs.get("bpm", 120) + duration = kwargs.get("duration", 120) + keyscale = kwargs.get("keyscale", "C major") + timesignature = kwargs.get("timesignature", 2) + language = kwargs.get("language", "en") + seed = kwargs.get("seed", 0) + + duration = math.ceil(duration) + meta_lm = 'bpm: {}\nduration: {}\nkeyscale: {}\ntimesignature: {}'.format(bpm, duration, keyscale, timesignature) + lm_template = "<|im_start|>system\n# Instruction\nGenerate audio semantic tokens based on the given conditions:\n\n<|im_end|>\n<|im_start|>user\n# Caption\n{}\n{}\n<|im_end|>\n<|im_start|>assistant\n\n{}\n\n\n<|im_end|>\n" + + meta_cap = '- bpm: {}\n- timesignature: {}\n- keyscale: {}\n- duration: {}\n'.format(bpm, timesignature, keyscale, duration) + out["lm_prompt"] = self.qwen3_06b.tokenize_with_weights(lm_template.format(text, lyrics, meta_lm), disable_weights=True) + out["lm_prompt_negative"] = self.qwen3_06b.tokenize_with_weights(lm_template.format(text, lyrics, ""), disable_weights=True) + + out["lyrics"] = self.qwen3_06b.tokenize_with_weights("# Languages\n{}\n\n# Lyric{}<|endoftext|><|endoftext|>".format(language, lyrics), return_word_ids, disable_weights=True, **kwargs) + out["qwen3_06b"] = self.qwen3_06b.tokenize_with_weights("# Instruction\nGenerate audio semantic tokens based on the given conditions:\n\n# Caption\n{}# Metas\n{}<|endoftext|>\n<|endoftext|>".format(text, meta_cap), return_word_ids, **kwargs) + out["lm_metadata"] = {"min_tokens": duration * 5, "seed": seed} + return out + + +class Qwen3_06BModel(sd1_clip.SDClipModel): + def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, attention_mask=True, model_options={}): + super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"pad": 151643}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Qwen3_06B_ACE15, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options) + +class Qwen3_2B_ACE15(sd1_clip.SDClipModel): + def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, attention_mask=True, model_options={}): + llama_quantization_metadata = model_options.get("llama_quantization_metadata", None) + if llama_quantization_metadata is not None: + model_options = model_options.copy() + model_options["quantization_metadata"] = llama_quantization_metadata + + super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"pad": 151643}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Qwen3_2B_ACE15_lm, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options) + +class ACE15TEModel(torch.nn.Module): + def __init__(self, device="cpu", dtype=None, dtype_llama=None, model_options={}): + super().__init__() + if dtype_llama is None: + dtype_llama = dtype + + self.qwen3_06b = Qwen3_06BModel(device=device, dtype=dtype, model_options=model_options) + self.qwen3_2b = Qwen3_2B_ACE15(device=device, dtype=dtype_llama, model_options=model_options) + self.dtypes = set([dtype, dtype_llama]) + + def encode_token_weights(self, token_weight_pairs): + token_weight_pairs_base = token_weight_pairs["qwen3_06b"] + token_weight_pairs_lyrics = token_weight_pairs["lyrics"] + + self.qwen3_06b.set_clip_options({"layer": None}) + base_out, _, extra = self.qwen3_06b.encode_token_weights(token_weight_pairs_base) + self.qwen3_06b.set_clip_options({"layer": [0]}) + lyrics_embeds, _, extra_l = self.qwen3_06b.encode_token_weights(token_weight_pairs_lyrics) + + lm_metadata = token_weight_pairs["lm_metadata"] + audio_codes = generate_audio_codes(self.qwen3_2b, token_weight_pairs["lm_prompt"], token_weight_pairs["lm_prompt_negative"], min_tokens=lm_metadata["min_tokens"], max_tokens=lm_metadata["min_tokens"], seed=lm_metadata["seed"]) + + return base_out, None, {"conditioning_lyrics": lyrics_embeds[:, 0], "audio_codes": [audio_codes]} + + def set_clip_options(self, options): + self.qwen3_06b.set_clip_options(options) + self.qwen3_2b.set_clip_options(options) + + def reset_clip_options(self): + self.qwen3_06b.reset_clip_options() + self.qwen3_2b.reset_clip_options() + + def load_sd(self, sd): + if "model.layers.0.post_attention_layernorm.weight" in sd: + shape = sd["model.layers.0.post_attention_layernorm.weight"].shape + if shape[0] == 1024: + return self.qwen3_06b.load_sd(sd) + else: + return self.qwen3_2b.load_sd(sd) + + def memory_estimation_function(self, token_weight_pairs, device=None): + lm_metadata = token_weight_pairs["lm_metadata"] + constant = 0.4375 + if comfy.model_management.should_use_bf16(device): + constant *= 0.5 + + token_weight_pairs = token_weight_pairs.get("lm_prompt", []) + num_tokens = sum(map(lambda a: len(a), token_weight_pairs)) + num_tokens += lm_metadata['min_tokens'] + return num_tokens * constant * 1024 * 1024 + +def te(dtype_llama=None, llama_quantization_metadata=None): + class ACE15TEModel_(ACE15TEModel): + def __init__(self, device="cpu", dtype=None, model_options={}): + if llama_quantization_metadata is not None: + model_options = model_options.copy() + model_options["llama_quantization_metadata"] = llama_quantization_metadata + super().__init__(device=device, dtype_llama=dtype_llama, dtype=dtype, model_options=model_options) + return ACE15TEModel_ diff --git a/comfy/text_encoders/llama.py b/comfy/text_encoders/llama.py index 68ac1e804..d2324ffc5 100644 --- a/comfy/text_encoders/llama.py +++ b/comfy/text_encoders/llama.py @@ -103,6 +103,52 @@ class Qwen3_06BConfig: final_norm: bool = True lm_head: bool = False +@dataclass +class Qwen3_06B_ACE15_Config: + vocab_size: int = 151669 + hidden_size: int = 1024 + intermediate_size: int = 3072 + num_hidden_layers: int = 28 + num_attention_heads: int = 16 + num_key_value_heads: int = 8 + max_position_embeddings: int = 32768 + rms_norm_eps: float = 1e-6 + rope_theta: float = 1000000.0 + transformer_type: str = "llama" + head_dim = 128 + rms_norm_add = False + mlp_activation = "silu" + qkv_bias = False + rope_dims = None + q_norm = "gemma3" + k_norm = "gemma3" + rope_scale = None + final_norm: bool = True + lm_head: bool = False + +@dataclass +class Qwen3_2B_ACE15_lm_Config: + vocab_size: int = 217204 + hidden_size: int = 2048 + intermediate_size: int = 6144 + num_hidden_layers: int = 28 + num_attention_heads: int = 16 + num_key_value_heads: int = 8 + max_position_embeddings: int = 40960 + rms_norm_eps: float = 1e-6 + rope_theta: float = 1000000.0 + transformer_type: str = "llama" + head_dim = 128 + rms_norm_add = False + mlp_activation = "silu" + qkv_bias = False + rope_dims = None + q_norm = "gemma3" + k_norm = "gemma3" + rope_scale = None + final_norm: bool = True + lm_head: bool = False + @dataclass class Qwen3_4BConfig: vocab_size: int = 151936 @@ -729,6 +775,27 @@ class Qwen3_06B(BaseLlama, torch.nn.Module): self.model = Llama2_(config, device=device, dtype=dtype, ops=operations) self.dtype = dtype +class Qwen3_06B_ACE15(BaseLlama, torch.nn.Module): + def __init__(self, config_dict, dtype, device, operations): + super().__init__() + config = Qwen3_06B_ACE15_Config(**config_dict) + self.num_layers = config.num_hidden_layers + + self.model = Llama2_(config, device=device, dtype=dtype, ops=operations) + self.dtype = dtype + +class Qwen3_2B_ACE15_lm(BaseLlama, torch.nn.Module): + def __init__(self, config_dict, dtype, device, operations): + super().__init__() + config = Qwen3_2B_ACE15_lm_Config(**config_dict) + self.num_layers = config.num_hidden_layers + + self.model = Llama2_(config, device=device, dtype=dtype, ops=operations) + self.dtype = dtype + + def logits(self, x): + return torch.nn.functional.linear(x[:, -1:], self.model.embed_tokens.weight.to(x), None) + class Qwen3_4B(BaseLlama, torch.nn.Module): def __init__(self, config_dict, dtype, device, operations): super().__init__() diff --git a/comfy_extras/nodes_ace.py b/comfy_extras/nodes_ace.py index 1409233c9..376584e5c 100644 --- a/comfy_extras/nodes_ace.py +++ b/comfy_extras/nodes_ace.py @@ -28,12 +28,39 @@ class TextEncodeAceStepAudio(io.ComfyNode): conditioning = node_helpers.conditioning_set_values(conditioning, {"lyrics_strength": lyrics_strength}) return io.NodeOutput(conditioning) +class TextEncodeAceStepAudio15(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="TextEncodeAceStepAudio1.5", + category="conditioning", + inputs=[ + io.Clip.Input("clip"), + io.String.Input("tags", multiline=True, dynamic_prompts=True), + io.String.Input("lyrics", multiline=True, dynamic_prompts=True), + io.Int.Input("seed", default=0, min=0, max=0xffffffffffffffff, control_after_generate=True), + io.Int.Input("bpm", default=120, min=10, max=300), + io.Float.Input("duration", default=120.0, min=0.0, max=2000.0, step=0.1), + io.Combo.Input("timesignature", options=['2', '3', '4', '6']), + io.Combo.Input("language", options=["en", "ja", "zh", "es", "de", "fr", "pt", "ru", "it", "nl", "pl", "tr", "vi", "cs", "fa", "id", "ko", "uk", "hu", "ar", "sv", "ro", "el"]), + io.Combo.Input("keyscale", options=[f"{root} {quality}" for quality in ["major", "minor"] for root in ["C", "C#", "Db", "D", "D#", "Eb", "E", "F", "F#", "Gb", "G", "G#", "Ab", "A", "A#", "Bb", "B"]]), + ], + outputs=[io.Conditioning.Output()], + ) + + @classmethod + def execute(cls, clip, tags, lyrics, seed, bpm, duration, timesignature, language, keyscale) -> io.NodeOutput: + tokens = clip.tokenize(tags, lyrics=lyrics, bpm=bpm, duration=duration, timesignature=int(timesignature), language=language, keyscale=keyscale, seed=seed) + conditioning = clip.encode_from_tokens_scheduled(tokens) + return io.NodeOutput(conditioning) + class EmptyAceStepLatentAudio(io.ComfyNode): @classmethod def define_schema(cls): return io.Schema( node_id="EmptyAceStepLatentAudio", + display_name="Empty Ace Step 1.0 Latent Audio", category="latent/audio", inputs=[ io.Float.Input("seconds", default=120.0, min=1.0, max=1000.0, step=0.1), @@ -51,12 +78,60 @@ class EmptyAceStepLatentAudio(io.ComfyNode): return io.NodeOutput({"samples": latent, "type": "audio"}) +class EmptyAceStep15LatentAudio(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="EmptyAceStep1.5LatentAudio", + display_name="Empty Ace Step 1.5 Latent Audio", + category="latent/audio", + inputs=[ + io.Float.Input("seconds", default=120.0, min=1.0, max=1000.0, step=0.01), + io.Int.Input( + "batch_size", default=1, min=1, max=4096, tooltip="The number of latent images in the batch." + ), + ], + outputs=[io.Latent.Output()], + ) + + @classmethod + def execute(cls, seconds, batch_size) -> io.NodeOutput: + length = round((seconds * 48000 / 1920)) + latent = torch.zeros([batch_size, 64, length], device=comfy.model_management.intermediate_device()) + return io.NodeOutput({"samples": latent, "type": "audio"}) + +class ReferenceTimbreAudio(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="ReferenceTimbreAudio", + category="advanced/conditioning/audio", + is_experimental=True, + description="This node sets the reference audio for timbre (for ace step 1.5)", + inputs=[ + io.Conditioning.Input("conditioning"), + io.Latent.Input("latent", optional=True), + ], + outputs=[ + io.Conditioning.Output(), + ] + ) + + @classmethod + def execute(cls, conditioning, latent=None) -> io.NodeOutput: + if latent is not None: + conditioning = node_helpers.conditioning_set_values(conditioning, {"reference_audio_timbre_latents": [latent["samples"]]}, append=True) + return io.NodeOutput(conditioning) + class AceExtension(ComfyExtension): @override async def get_node_list(self) -> list[type[io.ComfyNode]]: return [ TextEncodeAceStepAudio, EmptyAceStepLatentAudio, + TextEncodeAceStepAudio15, + EmptyAceStep15LatentAudio, + ReferenceTimbreAudio, ] async def comfy_entrypoint() -> AceExtension: diff --git a/comfy_extras/nodes_audio.py b/comfy_extras/nodes_audio.py index 271b75fbd..bef723dce 100644 --- a/comfy_extras/nodes_audio.py +++ b/comfy_extras/nodes_audio.py @@ -82,13 +82,14 @@ class VAEEncodeAudio(IO.ComfyNode): @classmethod def execute(cls, vae, audio) -> IO.NodeOutput: sample_rate = audio["sample_rate"] - if 44100 != sample_rate: - waveform = torchaudio.functional.resample(audio["waveform"], sample_rate, 44100) + vae_sample_rate = getattr(vae, "audio_sample_rate", 44100) + if vae_sample_rate != sample_rate: + waveform = torchaudio.functional.resample(audio["waveform"], sample_rate, vae_sample_rate) else: waveform = audio["waveform"] t = vae.encode(waveform.movedim(1, -1)) - return IO.NodeOutput({"samples":t}) + return IO.NodeOutput({"samples": t}) encode = execute # TODO: remove @@ -114,7 +115,8 @@ class VAEDecodeAudio(IO.ComfyNode): std = torch.std(audio, dim=[1,2], keepdim=True) * 5.0 std[std < 1.0] = 1.0 audio /= std - return IO.NodeOutput({"waveform": audio, "sample_rate": 44100 if "sample_rate" not in samples else samples["sample_rate"]}) + vae_sample_rate = getattr(vae, "audio_sample_rate", 44100) + return IO.NodeOutput({"waveform": audio, "sample_rate": vae_sample_rate if "sample_rate" not in samples else samples["sample_rate"]}) decode = execute # TODO: remove diff --git a/nodes.py b/nodes.py index 1cb43d9e2..e11a8ed80 100644 --- a/nodes.py +++ b/nodes.py @@ -1001,7 +1001,7 @@ class DualCLIPLoader: def INPUT_TYPES(s): return {"required": { "clip_name1": (folder_paths.get_filename_list("text_encoders"), ), "clip_name2": (folder_paths.get_filename_list("text_encoders"), ), - "type": (["sdxl", "sd3", "flux", "hunyuan_video", "hidream", "hunyuan_image", "hunyuan_video_15", "kandinsky5", "kandinsky5_image", "ltxv", "newbie"], ), + "type": (["sdxl", "sd3", "flux", "hunyuan_video", "hidream", "hunyuan_image", "hunyuan_video_15", "kandinsky5", "kandinsky5_image", "ltxv", "newbie", "ace"], ), }, "optional": { "device": (["default", "cpu"], {"advanced": True}), From be4345d1c9327c7a9292bc8f4153d35717020ef3 Mon Sep 17 00:00:00 2001 From: ComfyUI Wiki Date: Tue, 3 Feb 2026 15:08:43 +0800 Subject: [PATCH 558/647] chore: update workflow templates to v0.8.31 (#12239) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 3ca417dd8..0c401873a 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,5 +1,5 @@ comfyui-frontend-package==1.37.11 -comfyui-workflow-templates==0.8.27 +comfyui-workflow-templates==0.8.31 comfyui-embedded-docs==0.4.0 torch torchsde From 66e1b07402a84363b174c35d09acfcab0af7b0e5 Mon Sep 17 00:00:00 2001 From: comfyanonymous Date: Tue, 3 Feb 2026 02:20:59 -0500 Subject: [PATCH 559/647] ComfyUI v0.12.0 --- comfyui_version.py | 2 +- pyproject.toml | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/comfyui_version.py b/comfyui_version.py index b1ebaa115..bc6076b67 100644 --- a/comfyui_version.py +++ b/comfyui_version.py @@ -1,3 +1,3 @@ # This file is automatically generated by the build process when version is # updated in pyproject.toml. -__version__ = "0.11.1" +__version__ = "0.12.0" diff --git a/pyproject.toml b/pyproject.toml index 042f124e4..28aa03067 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -1,6 +1,6 @@ [project] name = "ComfyUI" -version = "0.11.1" +version = "0.12.0" readme = "README.md" license = { file = "LICENSE" } requires-python = ">=3.10" From f5030e26fd72bdf19ab2b6463dc502fa0b2e34ba Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 3 Feb 2026 01:09:30 -0800 Subject: [PATCH 560/647] Add progress bar to ace step. (#12242) --- comfy/text_encoders/ace15.py | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/comfy/text_encoders/ace15.py b/comfy/text_encoders/ace15.py index 9070cb577..75e77151d 100644 --- a/comfy/text_encoders/ace15.py +++ b/comfy/text_encoders/ace15.py @@ -3,6 +3,7 @@ import comfy.text_encoders.llama from comfy import sd1_clip import torch import math +import comfy.utils def sample_manual_loop_no_classes( @@ -42,6 +43,8 @@ def sample_manual_loop_no_classes( for x in range(model_config.num_hidden_layers): past_key_values.append((torch.empty([embeds.shape[0], model_config.num_key_value_heads, embeds.shape[1] + min_tokens, model_config.head_dim], device=device, dtype=execution_dtype), torch.empty([embeds.shape[0], model_config.num_key_value_heads, embeds.shape[1] + min_tokens, model_config.head_dim], device=device, dtype=execution_dtype), 0)) + progress_bar = comfy.utils.ProgressBar(max_new_tokens) + for step in range(max_new_tokens): outputs = model.transformer(None, attention_mask, embeds=embeds.to(execution_dtype), num_tokens=num_tokens, intermediate_output=None, dtype=execution_dtype, embeds_info=embeds_info, past_key_values=past_key_values) next_token_logits = model.transformer.logits(outputs[0])[:, -1] @@ -90,6 +93,7 @@ def sample_manual_loop_no_classes( attention_mask = torch.cat([attention_mask, torch.ones((2, 1), device=device, dtype=attention_mask.dtype)], dim=1) output_audio_codes.append(token - audio_start_id) + progress_bar.update_absolute(step) return output_audio_codes From affe881354d1e41df9ac9394b2d8257faebfdba1 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 3 Feb 2026 08:07:04 -0800 Subject: [PATCH 561/647] Fix some issues with mac. (#12247) --- comfy/text_encoders/ace15.py | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-) diff --git a/comfy/text_encoders/ace15.py b/comfy/text_encoders/ace15.py index 75e77151d..48c29fa9a 100644 --- a/comfy/text_encoders/ace15.py +++ b/comfy/text_encoders/ace15.py @@ -57,8 +57,9 @@ def sample_manual_loop_no_classes( if eos_token_id is not None and eos_token_id < audio_start_id and min_tokens < step: eos_score = cfg_logits[:, eos_token_id].clone() + remove_logit_value = torch.finfo(cfg_logits.dtype).min # Only generate audio tokens - cfg_logits[:, :audio_start_id] = float('-inf') + cfg_logits[:, :audio_start_id] = remove_logit_value if eos_token_id is not None and eos_token_id < audio_start_id and min_tokens < step: cfg_logits[:, eos_token_id] = eos_score @@ -66,7 +67,7 @@ def sample_manual_loop_no_classes( if top_k is not None and top_k > 0: top_k_vals, _ = torch.topk(cfg_logits, top_k) min_val = top_k_vals[..., -1, None] - cfg_logits[cfg_logits < min_val] = float('-inf') + cfg_logits[cfg_logits < min_val] = remove_logit_value if top_p is not None and top_p < 1.0: sorted_logits, sorted_indices = torch.sort(cfg_logits, descending=True) @@ -75,7 +76,7 @@ def sample_manual_loop_no_classes( sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone() sorted_indices_to_remove[..., 0] = 0 indices_to_remove = sorted_indices_to_remove.scatter(1, sorted_indices, sorted_indices_to_remove) - cfg_logits[indices_to_remove] = float('-inf') + cfg_logits[indices_to_remove] = remove_logit_value if temperature > 0: cfg_logits = cfg_logits / temperature From 223364743c35d6e1dc4f5ebc3796234d0f8484cf Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 3 Feb 2026 08:31:36 -0800 Subject: [PATCH 562/647] llama: cast logits as a comfy-weight (#12248) This is using a different layers weight with .to(). Change it to use the ops caster if the original layer is a comfy weight so that it picks up dynamic_vram and async_offload functionality in full. Co-authored-by: Rattus --- comfy/text_encoders/llama.py | 15 ++++++++++++++- 1 file changed, 14 insertions(+), 1 deletion(-) diff --git a/comfy/text_encoders/llama.py b/comfy/text_encoders/llama.py index d2324ffc5..d1c628d20 100644 --- a/comfy/text_encoders/llama.py +++ b/comfy/text_encoders/llama.py @@ -6,6 +6,7 @@ import math from comfy.ldm.modules.attention import optimized_attention_for_device import comfy.model_management +import comfy.ops import comfy.ldm.common_dit import comfy.clip_model @@ -794,7 +795,19 @@ class Qwen3_2B_ACE15_lm(BaseLlama, torch.nn.Module): self.dtype = dtype def logits(self, x): - return torch.nn.functional.linear(x[:, -1:], self.model.embed_tokens.weight.to(x), None) + input = x[:, -1:] + module = self.model.embed_tokens + + offload_stream = None + if module.comfy_cast_weights: + weight, _, offload_stream = comfy.ops.cast_bias_weight(module, input, offloadable=True) + else: + weight = self.model.embed_tokens.weight.to(x) + + x = torch.nn.functional.linear(input, weight, None) + + comfy.ops.uncast_bias_weight(module, weight, None, offload_stream) + return x class Qwen3_4B(BaseLlama, torch.nn.Module): def __init__(self, config_dict, dtype, device, operations): From 85fc35e8fa44c6174425acb4f9167792bcc903a8 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 3 Feb 2026 09:19:39 -0800 Subject: [PATCH 563/647] Fix mac issue. (#12250) --- comfy/text_encoders/llama.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/comfy/text_encoders/llama.py b/comfy/text_encoders/llama.py index d1c628d20..f4ac224c6 100644 --- a/comfy/text_encoders/llama.py +++ b/comfy/text_encoders/llama.py @@ -628,10 +628,10 @@ class Llama2_(nn.Module): mask = None if attention_mask is not None: mask = 1.0 - attention_mask.to(x.dtype).reshape((attention_mask.shape[0], 1, -1, attention_mask.shape[-1])).expand(attention_mask.shape[0], 1, seq_len, attention_mask.shape[-1]) - mask = mask.masked_fill(mask.to(torch.bool), float("-inf")) + mask = mask.masked_fill(mask.to(torch.bool), torch.finfo(x.dtype).min) if seq_len > 1: - causal_mask = torch.empty(past_len + seq_len, past_len + seq_len, dtype=x.dtype, device=x.device).fill_(float("-inf")).triu_(1) + causal_mask = torch.empty(past_len + seq_len, past_len + seq_len, dtype=x.dtype, device=x.device).fill_(torch.finfo(x.dtype).min).triu_(1) if mask is not None: mask += causal_mask else: From fb23935c1139875d47e7b6e239f657bae73bda5b Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Tue, 3 Feb 2026 20:31:46 +0200 Subject: [PATCH 564/647] feat(comfy_api): add basic 3D Model file types (#12129) * feat(comfy_api): add basic 3D Model file types * update Tripo nodes to use File3DGLB * update Rodin3D nodes to use File3DGLB * address PR review feedback: - Rename File3D parameter 'path' to 'source' - Convert File3D.data property to get_data() - Make .glb extension check case-insensitive in nodes_rodin.py - Restrict SaveGLB node to only accept File3DGLB * Fixed a bug in the Meshy Rig and Animation nodes * Fix backward compatability --- comfy_api/latest/__init__.py | 3 +- comfy_api/latest/_io.py | 52 +++++++++- comfy_api/latest/_util/__init__.py | 3 +- comfy_api/latest/_util/geometry_types.py | 77 ++++++++++++++ comfy_api_nodes/apis/meshy.py | 5 + comfy_api_nodes/nodes_hunyuan3d.py | 45 ++++---- comfy_api_nodes/nodes_meshy.py | 116 ++++++++++++++------- comfy_api_nodes/nodes_rodin.py | 67 ++++++++---- comfy_api_nodes/nodes_tripo.py | 126 +++++++++++------------ comfy_api_nodes/util/__init__.py | 2 + comfy_api_nodes/util/download_helpers.py | 38 ++++++- comfy_extras/nodes_hunyuan3d.py | 27 ++++- comfy_extras/nodes_load_3d.py | 26 ++++- 13 files changed, 427 insertions(+), 160 deletions(-) diff --git a/comfy_api/latest/__init__.py b/comfy_api/latest/__init__.py index b0fa14ff6..8542a1dbc 100644 --- a/comfy_api/latest/__init__.py +++ b/comfy_api/latest/__init__.py @@ -7,7 +7,7 @@ from comfy_api.internal.singleton import ProxiedSingleton from comfy_api.internal.async_to_sync import create_sync_class from ._input import ImageInput, AudioInput, MaskInput, LatentInput, VideoInput from ._input_impl import VideoFromFile, VideoFromComponents -from ._util import VideoCodec, VideoContainer, VideoComponents, MESH, VOXEL +from ._util import VideoCodec, VideoContainer, VideoComponents, MESH, VOXEL, File3D from . import _io_public as io from . import _ui_public as ui from comfy_execution.utils import get_executing_context @@ -105,6 +105,7 @@ class Types: VideoComponents = VideoComponents MESH = MESH VOXEL = VOXEL + File3D = File3D ComfyAPI = ComfyAPI_latest diff --git a/comfy_api/latest/_io.py b/comfy_api/latest/_io.py index eeea9781a..93cf482ca 100644 --- a/comfy_api/latest/_io.py +++ b/comfy_api/latest/_io.py @@ -27,7 +27,7 @@ if TYPE_CHECKING: from comfy_api.internal import (_ComfyNodeInternal, _NodeOutputInternal, classproperty, copy_class, first_real_override, is_class, prune_dict, shallow_clone_class) from comfy_execution.graph_utils import ExecutionBlocker -from ._util import MESH, VOXEL, SVG as _SVG +from ._util import MESH, VOXEL, SVG as _SVG, File3D class FolderType(str, Enum): @@ -667,6 +667,49 @@ class Voxel(ComfyTypeIO): class Mesh(ComfyTypeIO): Type = MESH + +@comfytype(io_type="FILE_3D") +class File3DAny(ComfyTypeIO): + """General 3D file type - accepts any supported 3D format.""" + Type = File3D + + +@comfytype(io_type="FILE_3D_GLB") +class File3DGLB(ComfyTypeIO): + """GLB format 3D file - binary glTF, best for web and cross-platform.""" + Type = File3D + + +@comfytype(io_type="FILE_3D_GLTF") +class File3DGLTF(ComfyTypeIO): + """GLTF format 3D file - JSON-based glTF with external resources.""" + Type = File3D + + +@comfytype(io_type="FILE_3D_FBX") +class File3DFBX(ComfyTypeIO): + """FBX format 3D file - best for game engines and animation.""" + Type = File3D + + +@comfytype(io_type="FILE_3D_OBJ") +class File3DOBJ(ComfyTypeIO): + """OBJ format 3D file - simple geometry format.""" + Type = File3D + + +@comfytype(io_type="FILE_3D_STL") +class File3DSTL(ComfyTypeIO): + """STL format 3D file - best for 3D printing.""" + Type = File3D + + +@comfytype(io_type="FILE_3D_USDZ") +class File3DUSDZ(ComfyTypeIO): + """USDZ format 3D file - Apple AR format.""" + Type = File3D + + @comfytype(io_type="HOOKS") class Hooks(ComfyTypeIO): if TYPE_CHECKING: @@ -2037,6 +2080,13 @@ __all__ = [ "LossMap", "Voxel", "Mesh", + "File3DAny", + "File3DGLB", + "File3DGLTF", + "File3DFBX", + "File3DOBJ", + "File3DSTL", + "File3DUSDZ", "Hooks", "HookKeyframes", "TimestepsRange", diff --git a/comfy_api/latest/_util/__init__.py b/comfy_api/latest/_util/__init__.py index 6313eb01b..115baf392 100644 --- a/comfy_api/latest/_util/__init__.py +++ b/comfy_api/latest/_util/__init__.py @@ -1,5 +1,5 @@ from .video_types import VideoContainer, VideoCodec, VideoComponents -from .geometry_types import VOXEL, MESH +from .geometry_types import VOXEL, MESH, File3D from .image_types import SVG __all__ = [ @@ -9,5 +9,6 @@ __all__ = [ "VideoComponents", "VOXEL", "MESH", + "File3D", "SVG", ] diff --git a/comfy_api/latest/_util/geometry_types.py b/comfy_api/latest/_util/geometry_types.py index 385122778..b586fceb3 100644 --- a/comfy_api/latest/_util/geometry_types.py +++ b/comfy_api/latest/_util/geometry_types.py @@ -1,3 +1,8 @@ +import shutil +from io import BytesIO +from pathlib import Path +from typing import IO + import torch @@ -10,3 +15,75 @@ class MESH: def __init__(self, vertices: torch.Tensor, faces: torch.Tensor): self.vertices = vertices self.faces = faces + + +class File3D: + """Class representing a 3D file from a file path or binary stream. + + Supports both disk-backed (file path) and memory-backed (BytesIO) storage. + """ + + def __init__(self, source: str | IO[bytes], file_format: str = ""): + self._source = source + self._format = file_format or self._infer_format() + + def _infer_format(self) -> str: + if isinstance(self._source, str): + return Path(self._source).suffix.lstrip(".").lower() + return "" + + @property + def format(self) -> str: + return self._format + + @format.setter + def format(self, value: str) -> None: + self._format = value.lstrip(".").lower() if value else "" + + @property + def is_disk_backed(self) -> bool: + return isinstance(self._source, str) + + def get_source(self) -> str | IO[bytes]: + if isinstance(self._source, str): + return self._source + if hasattr(self._source, "seek"): + self._source.seek(0) + return self._source + + def get_data(self) -> BytesIO: + if isinstance(self._source, str): + with open(self._source, "rb") as f: + result = BytesIO(f.read()) + return result + if hasattr(self._source, "seek"): + self._source.seek(0) + if isinstance(self._source, BytesIO): + return self._source + return BytesIO(self._source.read()) + + def save_to(self, path: str) -> str: + dest = Path(path) + dest.parent.mkdir(parents=True, exist_ok=True) + + if isinstance(self._source, str): + if Path(self._source).resolve() != dest.resolve(): + shutil.copy2(self._source, dest) + else: + if hasattr(self._source, "seek"): + self._source.seek(0) + with open(dest, "wb") as f: + f.write(self._source.read()) + return str(dest) + + def get_bytes(self) -> bytes: + if isinstance(self._source, str): + return Path(self._source).read_bytes() + if hasattr(self._source, "seek"): + self._source.seek(0) + return self._source.read() + + def __repr__(self) -> str: + if isinstance(self._source, str): + return f"File3D(source={self._source!r}, format={self._format!r})" + return f"File3D(, format={self._format!r})" diff --git a/comfy_api_nodes/apis/meshy.py b/comfy_api_nodes/apis/meshy.py index be46d0d58..7d72e6e91 100644 --- a/comfy_api_nodes/apis/meshy.py +++ b/comfy_api_nodes/apis/meshy.py @@ -109,14 +109,19 @@ class MeshyTextureRequest(BaseModel): class MeshyModelsUrls(BaseModel): glb: str = Field("") + fbx: str = Field("") + usdz: str = Field("") + obj: str = Field("") class MeshyRiggedModelsUrls(BaseModel): rigged_character_glb_url: str = Field("") + rigged_character_fbx_url: str = Field("") class MeshyAnimatedModelsUrls(BaseModel): animation_glb_url: str = Field("") + animation_fbx_url: str = Field("") class MeshyResultTextureUrls(BaseModel): diff --git a/comfy_api_nodes/nodes_hunyuan3d.py b/comfy_api_nodes/nodes_hunyuan3d.py index b3a736643..813a7c809 100644 --- a/comfy_api_nodes/nodes_hunyuan3d.py +++ b/comfy_api_nodes/nodes_hunyuan3d.py @@ -1,5 +1,3 @@ -import os - from typing_extensions import override from comfy_api.latest import IO, ComfyExtension, Input @@ -14,7 +12,7 @@ from comfy_api_nodes.apis.hunyuan3d import ( ) from comfy_api_nodes.util import ( ApiEndpoint, - download_url_to_bytesio, + download_url_to_file_3d, downscale_image_tensor_by_max_side, poll_op, sync_op, @@ -22,14 +20,13 @@ from comfy_api_nodes.util import ( validate_image_dimensions, validate_string, ) -from folder_paths import get_output_directory -def get_glb_obj_from_response(response_objs: list[ResultFile3D]) -> ResultFile3D: +def get_file_from_response(response_objs: list[ResultFile3D], file_type: str) -> ResultFile3D | None: for i in response_objs: - if i.Type.lower() == "glb": + if i.Type.lower() == file_type.lower(): return i - raise ValueError("No GLB file found in response. Please report this to the developers.") + return None class TencentTextToModelNode(IO.ComfyNode): @@ -74,7 +71,9 @@ class TencentTextToModelNode(IO.ComfyNode): ), ], outputs=[ - IO.String.Output(display_name="model_file"), + IO.String.Output(display_name="model_file"), # for backward compatibility only + IO.File3DGLB.Output(display_name="GLB"), + IO.File3DOBJ.Output(display_name="OBJ"), ], hidden=[ IO.Hidden.auth_token_comfy_org, @@ -124,19 +123,20 @@ class TencentTextToModelNode(IO.ComfyNode): ) if response.Error: raise ValueError(f"Task creation failed with code {response.Error.Code}: {response.Error.Message}") + task_id = response.JobId result = await poll_op( cls, ApiEndpoint(path="/proxy/tencent/hunyuan/3d-pro/query", method="POST"), - data=To3DProTaskQueryRequest(JobId=response.JobId), + data=To3DProTaskQueryRequest(JobId=task_id), response_model=To3DProTaskResultResponse, status_extractor=lambda r: r.Status, ) - model_file = f"hunyuan_model_{response.JobId}.glb" - await download_url_to_bytesio( - get_glb_obj_from_response(result.ResultFile3Ds).Url, - os.path.join(get_output_directory(), model_file), + glb_result = get_file_from_response(result.ResultFile3Ds, "glb") + obj_result = get_file_from_response(result.ResultFile3Ds, "obj") + file_glb = await download_url_to_file_3d(glb_result.Url, "glb", task_id=task_id) if glb_result else None + return IO.NodeOutput( + file_glb, file_glb, await download_url_to_file_3d(obj_result.Url, "obj", task_id=task_id) if obj_result else None ) - return IO.NodeOutput(model_file) class TencentImageToModelNode(IO.ComfyNode): @@ -184,7 +184,9 @@ class TencentImageToModelNode(IO.ComfyNode): ), ], outputs=[ - IO.String.Output(display_name="model_file"), + IO.String.Output(display_name="model_file"), # for backward compatibility only + IO.File3DGLB.Output(display_name="GLB"), + IO.File3DOBJ.Output(display_name="OBJ"), ], hidden=[ IO.Hidden.auth_token_comfy_org, @@ -269,19 +271,20 @@ class TencentImageToModelNode(IO.ComfyNode): ) if response.Error: raise ValueError(f"Task creation failed with code {response.Error.Code}: {response.Error.Message}") + task_id = response.JobId result = await poll_op( cls, ApiEndpoint(path="/proxy/tencent/hunyuan/3d-pro/query", method="POST"), - data=To3DProTaskQueryRequest(JobId=response.JobId), + data=To3DProTaskQueryRequest(JobId=task_id), response_model=To3DProTaskResultResponse, status_extractor=lambda r: r.Status, ) - model_file = f"hunyuan_model_{response.JobId}.glb" - await download_url_to_bytesio( - get_glb_obj_from_response(result.ResultFile3Ds).Url, - os.path.join(get_output_directory(), model_file), + glb_result = get_file_from_response(result.ResultFile3Ds, "glb") + obj_result = get_file_from_response(result.ResultFile3Ds, "obj") + file_glb = await download_url_to_file_3d(glb_result.Url, "glb", task_id=task_id) if glb_result else None + return IO.NodeOutput( + file_glb, file_glb, await download_url_to_file_3d(obj_result.Url, "obj", task_id=task_id) if obj_result else None ) - return IO.NodeOutput(model_file) class TencentHunyuan3DExtension(ComfyExtension): diff --git a/comfy_api_nodes/nodes_meshy.py b/comfy_api_nodes/nodes_meshy.py index 740607983..65f6f0d2d 100644 --- a/comfy_api_nodes/nodes_meshy.py +++ b/comfy_api_nodes/nodes_meshy.py @@ -1,5 +1,3 @@ -import os - from typing_extensions import override from comfy_api.latest import IO, ComfyExtension, Input @@ -20,13 +18,12 @@ from comfy_api_nodes.apis.meshy import ( ) from comfy_api_nodes.util import ( ApiEndpoint, - download_url_to_bytesio, + download_url_to_file_3d, poll_op, sync_op, upload_images_to_comfyapi, validate_string, ) -from folder_paths import get_output_directory class MeshyTextToModelNode(IO.ComfyNode): @@ -79,8 +76,10 @@ class MeshyTextToModelNode(IO.ComfyNode): ), ], outputs=[ - IO.String.Output(display_name="model_file"), + IO.String.Output(display_name="model_file"), # for backward compatibility only IO.Custom("MESHY_TASK_ID").Output(display_name="meshy_task_id"), + IO.File3DGLB.Output(display_name="GLB"), + IO.File3DFBX.Output(display_name="FBX"), ], hidden=[ IO.Hidden.auth_token_comfy_org, @@ -122,16 +121,20 @@ class MeshyTextToModelNode(IO.ComfyNode): seed=seed, ), ) + task_id = response.result result = await poll_op( cls, - ApiEndpoint(path=f"/proxy/meshy/openapi/v2/text-to-3d/{response.result}"), + ApiEndpoint(path=f"/proxy/meshy/openapi/v2/text-to-3d/{task_id}"), response_model=MeshyModelResult, status_extractor=lambda r: r.status, progress_extractor=lambda r: r.progress, ) - model_file = f"meshy_model_{response.result}.glb" - await download_url_to_bytesio(result.model_urls.glb, os.path.join(get_output_directory(), model_file)) - return IO.NodeOutput(model_file, response.result) + return IO.NodeOutput( + f"{task_id}.glb", + task_id, + await download_url_to_file_3d(result.model_urls.glb, "glb", task_id=task_id), + await download_url_to_file_3d(result.model_urls.fbx, "fbx", task_id=task_id), + ) class MeshyRefineNode(IO.ComfyNode): @@ -167,8 +170,10 @@ class MeshyRefineNode(IO.ComfyNode): ), ], outputs=[ - IO.String.Output(display_name="model_file"), + IO.String.Output(display_name="model_file"), # for backward compatibility only IO.Custom("MESHY_TASK_ID").Output(display_name="meshy_task_id"), + IO.File3DGLB.Output(display_name="GLB"), + IO.File3DFBX.Output(display_name="FBX"), ], hidden=[ IO.Hidden.auth_token_comfy_org, @@ -210,16 +215,20 @@ class MeshyRefineNode(IO.ComfyNode): ai_model=model, ), ) + task_id = response.result result = await poll_op( cls, - ApiEndpoint(path=f"/proxy/meshy/openapi/v2/text-to-3d/{response.result}"), + ApiEndpoint(path=f"/proxy/meshy/openapi/v2/text-to-3d/{task_id}"), response_model=MeshyModelResult, status_extractor=lambda r: r.status, progress_extractor=lambda r: r.progress, ) - model_file = f"meshy_model_{response.result}.glb" - await download_url_to_bytesio(result.model_urls.glb, os.path.join(get_output_directory(), model_file)) - return IO.NodeOutput(model_file, response.result) + return IO.NodeOutput( + f"{task_id}.glb", + task_id, + await download_url_to_file_3d(result.model_urls.glb, "glb", task_id=task_id), + await download_url_to_file_3d(result.model_urls.fbx, "fbx", task_id=task_id), + ) class MeshyImageToModelNode(IO.ComfyNode): @@ -303,8 +312,10 @@ class MeshyImageToModelNode(IO.ComfyNode): ), ], outputs=[ - IO.String.Output(display_name="model_file"), + IO.String.Output(display_name="model_file"), # for backward compatibility only IO.Custom("MESHY_TASK_ID").Output(display_name="meshy_task_id"), + IO.File3DGLB.Output(display_name="GLB"), + IO.File3DFBX.Output(display_name="FBX"), ], hidden=[ IO.Hidden.auth_token_comfy_org, @@ -368,16 +379,20 @@ class MeshyImageToModelNode(IO.ComfyNode): seed=seed, ), ) + task_id = response.result result = await poll_op( cls, - ApiEndpoint(path=f"/proxy/meshy/openapi/v1/image-to-3d/{response.result}"), + ApiEndpoint(path=f"/proxy/meshy/openapi/v1/image-to-3d/{task_id}"), response_model=MeshyModelResult, status_extractor=lambda r: r.status, progress_extractor=lambda r: r.progress, ) - model_file = f"meshy_model_{response.result}.glb" - await download_url_to_bytesio(result.model_urls.glb, os.path.join(get_output_directory(), model_file)) - return IO.NodeOutput(model_file, response.result) + return IO.NodeOutput( + f"{task_id}.glb", + task_id, + await download_url_to_file_3d(result.model_urls.glb, "glb", task_id=task_id), + await download_url_to_file_3d(result.model_urls.fbx, "fbx", task_id=task_id), + ) class MeshyMultiImageToModelNode(IO.ComfyNode): @@ -464,8 +479,10 @@ class MeshyMultiImageToModelNode(IO.ComfyNode): ), ], outputs=[ - IO.String.Output(display_name="model_file"), + IO.String.Output(display_name="model_file"), # for backward compatibility only IO.Custom("MESHY_TASK_ID").Output(display_name="meshy_task_id"), + IO.File3DGLB.Output(display_name="GLB"), + IO.File3DFBX.Output(display_name="FBX"), ], hidden=[ IO.Hidden.auth_token_comfy_org, @@ -531,16 +548,20 @@ class MeshyMultiImageToModelNode(IO.ComfyNode): seed=seed, ), ) + task_id = response.result result = await poll_op( cls, - ApiEndpoint(path=f"/proxy/meshy/openapi/v1/multi-image-to-3d/{response.result}"), + ApiEndpoint(path=f"/proxy/meshy/openapi/v1/multi-image-to-3d/{task_id}"), response_model=MeshyModelResult, status_extractor=lambda r: r.status, progress_extractor=lambda r: r.progress, ) - model_file = f"meshy_model_{response.result}.glb" - await download_url_to_bytesio(result.model_urls.glb, os.path.join(get_output_directory(), model_file)) - return IO.NodeOutput(model_file, response.result) + return IO.NodeOutput( + f"{task_id}.glb", + task_id, + await download_url_to_file_3d(result.model_urls.glb, "glb", task_id=task_id), + await download_url_to_file_3d(result.model_urls.fbx, "fbx", task_id=task_id), + ) class MeshyRigModelNode(IO.ComfyNode): @@ -571,8 +592,10 @@ class MeshyRigModelNode(IO.ComfyNode): ), ], outputs=[ - IO.String.Output(display_name="model_file"), + IO.String.Output(display_name="model_file"), # for backward compatibility only IO.Custom("MESHY_RIGGED_TASK_ID").Output(display_name="rig_task_id"), + IO.File3DGLB.Output(display_name="GLB"), + IO.File3DFBX.Output(display_name="FBX"), ], hidden=[ IO.Hidden.auth_token_comfy_org, @@ -606,18 +629,20 @@ class MeshyRigModelNode(IO.ComfyNode): texture_image_url=texture_image_url, ), ) + task_id = response.result result = await poll_op( cls, - ApiEndpoint(path=f"/proxy/meshy/openapi/v1/rigging/{response.result}"), + ApiEndpoint(path=f"/proxy/meshy/openapi/v1/rigging/{task_id}"), response_model=MeshyRiggedResult, status_extractor=lambda r: r.status, progress_extractor=lambda r: r.progress, ) - model_file = f"meshy_model_{response.result}.glb" - await download_url_to_bytesio( - result.result.rigged_character_glb_url, os.path.join(get_output_directory(), model_file) + return IO.NodeOutput( + f"{task_id}.glb", + task_id, + await download_url_to_file_3d(result.result.rigged_character_glb_url, "glb", task_id=task_id), + await download_url_to_file_3d(result.result.rigged_character_fbx_url, "fbx", task_id=task_id), ) - return IO.NodeOutput(model_file, response.result) class MeshyAnimateModelNode(IO.ComfyNode): @@ -640,7 +665,9 @@ class MeshyAnimateModelNode(IO.ComfyNode): ), ], outputs=[ - IO.String.Output(display_name="model_file"), + IO.String.Output(display_name="model_file"), # for backward compatibility only + IO.File3DGLB.Output(display_name="GLB"), + IO.File3DFBX.Output(display_name="FBX"), ], hidden=[ IO.Hidden.auth_token_comfy_org, @@ -669,16 +696,19 @@ class MeshyAnimateModelNode(IO.ComfyNode): action_id=action_id, ), ) + task_id = response.result result = await poll_op( cls, - ApiEndpoint(path=f"/proxy/meshy/openapi/v1/animations/{response.result}"), + ApiEndpoint(path=f"/proxy/meshy/openapi/v1/animations/{task_id}"), response_model=MeshyAnimationResult, status_extractor=lambda r: r.status, progress_extractor=lambda r: r.progress, ) - model_file = f"meshy_model_{response.result}.glb" - await download_url_to_bytesio(result.result.animation_glb_url, os.path.join(get_output_directory(), model_file)) - return IO.NodeOutput(model_file, response.result) + return IO.NodeOutput( + f"{task_id}.glb", + await download_url_to_file_3d(result.result.animation_glb_url, "glb", task_id=task_id), + await download_url_to_file_3d(result.result.animation_fbx_url, "fbx", task_id=task_id), + ) class MeshyTextureNode(IO.ComfyNode): @@ -715,8 +745,10 @@ class MeshyTextureNode(IO.ComfyNode): ), ], outputs=[ - IO.String.Output(display_name="model_file"), + IO.String.Output(display_name="model_file"), # for backward compatibility only IO.Custom("MODEL_TASK_ID").Output(display_name="meshy_task_id"), + IO.File3DGLB.Output(display_name="GLB"), + IO.File3DFBX.Output(display_name="FBX"), ], hidden=[ IO.Hidden.auth_token_comfy_org, @@ -760,16 +792,20 @@ class MeshyTextureNode(IO.ComfyNode): image_style_url=image_style_url, ), ) + task_id = response.result result = await poll_op( cls, - ApiEndpoint(path=f"/proxy/meshy/openapi/v1/retexture/{response.result}"), + ApiEndpoint(path=f"/proxy/meshy/openapi/v1/retexture/{task_id}"), response_model=MeshyModelResult, status_extractor=lambda r: r.status, progress_extractor=lambda r: r.progress, ) - model_file = f"meshy_model_{response.result}.glb" - await download_url_to_bytesio(result.model_urls.glb, os.path.join(get_output_directory(), model_file)) - return IO.NodeOutput(model_file, response.result) + return IO.NodeOutput( + f"{task_id}.glb", + task_id, + await download_url_to_file_3d(result.model_urls.glb, "glb", task_id=task_id), + await download_url_to_file_3d(result.model_urls.fbx, "fbx", task_id=task_id), + ) class MeshyExtension(ComfyExtension): diff --git a/comfy_api_nodes/nodes_rodin.py b/comfy_api_nodes/nodes_rodin.py index 3ffdc8b90..f9cff121f 100644 --- a/comfy_api_nodes/nodes_rodin.py +++ b/comfy_api_nodes/nodes_rodin.py @@ -10,7 +10,6 @@ import folder_paths as comfy_paths import os import logging import math -from typing import Optional from io import BytesIO from typing_extensions import override from PIL import Image @@ -28,8 +27,9 @@ from comfy_api_nodes.util import ( poll_op, ApiEndpoint, download_url_to_bytesio, + download_url_to_file_3d, ) -from comfy_api.latest import ComfyExtension, IO +from comfy_api.latest import ComfyExtension, IO, Types COMMON_PARAMETERS = [ @@ -177,7 +177,7 @@ def check_rodin_status(response: Rodin3DCheckStatusResponse) -> str: return "DONE" return "Generating" -def extract_progress(response: Rodin3DCheckStatusResponse) -> Optional[int]: +def extract_progress(response: Rodin3DCheckStatusResponse) -> int | None: if not response.jobs: return None completed_count = sum(1 for job in response.jobs if job.status == JobStatus.Done) @@ -207,17 +207,25 @@ async def get_rodin_download_list(uuid: str, cls: type[IO.ComfyNode]) -> Rodin3D ) -async def download_files(url_list, task_uuid: str): +async def download_files(url_list, task_uuid: str) -> tuple[str | None, Types.File3D | None]: result_folder_name = f"Rodin3D_{task_uuid}" save_path = os.path.join(comfy_paths.get_output_directory(), result_folder_name) os.makedirs(save_path, exist_ok=True) model_file_path = None + file_3d = None + for i in url_list.list: file_path = os.path.join(save_path, i.name) - if file_path.endswith(".glb"): + if i.name.lower().endswith(".glb"): model_file_path = os.path.join(result_folder_name, i.name) - await download_url_to_bytesio(i.url, file_path) - return model_file_path + file_3d = await download_url_to_file_3d(i.url, "glb") + # Save to disk for backward compatibility + with open(file_path, "wb") as f: + f.write(file_3d.get_bytes()) + else: + await download_url_to_bytesio(i.url, file_path) + + return model_file_path, file_3d class Rodin3D_Regular(IO.ComfyNode): @@ -234,7 +242,10 @@ class Rodin3D_Regular(IO.ComfyNode): IO.Image.Input("Images"), *COMMON_PARAMETERS, ], - outputs=[IO.String.Output(display_name="3D Model Path")], + outputs=[ + IO.String.Output(display_name="3D Model Path"), # for backward compatibility only + IO.File3DGLB.Output(display_name="GLB"), + ], hidden=[ IO.Hidden.auth_token_comfy_org, IO.Hidden.api_key_comfy_org, @@ -271,9 +282,9 @@ class Rodin3D_Regular(IO.ComfyNode): ) await poll_for_task_status(subscription_key, cls) download_list = await get_rodin_download_list(task_uuid, cls) - model = await download_files(download_list, task_uuid) + model_path, file_3d = await download_files(download_list, task_uuid) - return IO.NodeOutput(model) + return IO.NodeOutput(model_path, file_3d) class Rodin3D_Detail(IO.ComfyNode): @@ -290,7 +301,10 @@ class Rodin3D_Detail(IO.ComfyNode): IO.Image.Input("Images"), *COMMON_PARAMETERS, ], - outputs=[IO.String.Output(display_name="3D Model Path")], + outputs=[ + IO.String.Output(display_name="3D Model Path"), # for backward compatibility only + IO.File3DGLB.Output(display_name="GLB"), + ], hidden=[ IO.Hidden.auth_token_comfy_org, IO.Hidden.api_key_comfy_org, @@ -327,9 +341,9 @@ class Rodin3D_Detail(IO.ComfyNode): ) await poll_for_task_status(subscription_key, cls) download_list = await get_rodin_download_list(task_uuid, cls) - model = await download_files(download_list, task_uuid) + model_path, file_3d = await download_files(download_list, task_uuid) - return IO.NodeOutput(model) + return IO.NodeOutput(model_path, file_3d) class Rodin3D_Smooth(IO.ComfyNode): @@ -346,7 +360,10 @@ class Rodin3D_Smooth(IO.ComfyNode): IO.Image.Input("Images"), *COMMON_PARAMETERS, ], - outputs=[IO.String.Output(display_name="3D Model Path")], + outputs=[ + IO.String.Output(display_name="3D Model Path"), # for backward compatibility only + IO.File3DGLB.Output(display_name="GLB"), + ], hidden=[ IO.Hidden.auth_token_comfy_org, IO.Hidden.api_key_comfy_org, @@ -382,9 +399,9 @@ class Rodin3D_Smooth(IO.ComfyNode): ) await poll_for_task_status(subscription_key, cls) download_list = await get_rodin_download_list(task_uuid, cls) - model = await download_files(download_list, task_uuid) + model_path, file_3d = await download_files(download_list, task_uuid) - return IO.NodeOutput(model) + return IO.NodeOutput(model_path, file_3d) class Rodin3D_Sketch(IO.ComfyNode): @@ -408,7 +425,10 @@ class Rodin3D_Sketch(IO.ComfyNode): optional=True, ), ], - outputs=[IO.String.Output(display_name="3D Model Path")], + outputs=[ + IO.String.Output(display_name="3D Model Path"), # for backward compatibility only + IO.File3DGLB.Output(display_name="GLB"), + ], hidden=[ IO.Hidden.auth_token_comfy_org, IO.Hidden.api_key_comfy_org, @@ -441,9 +461,9 @@ class Rodin3D_Sketch(IO.ComfyNode): ) await poll_for_task_status(subscription_key, cls) download_list = await get_rodin_download_list(task_uuid, cls) - model = await download_files(download_list, task_uuid) + model_path, file_3d = await download_files(download_list, task_uuid) - return IO.NodeOutput(model) + return IO.NodeOutput(model_path, file_3d) class Rodin3D_Gen2(IO.ComfyNode): @@ -475,7 +495,10 @@ class Rodin3D_Gen2(IO.ComfyNode): ), IO.Boolean.Input("TAPose", default=False), ], - outputs=[IO.String.Output(display_name="3D Model Path")], + outputs=[ + IO.String.Output(display_name="3D Model Path"), # for backward compatibility only + IO.File3DGLB.Output(display_name="GLB"), + ], hidden=[ IO.Hidden.auth_token_comfy_org, IO.Hidden.api_key_comfy_org, @@ -511,9 +534,9 @@ class Rodin3D_Gen2(IO.ComfyNode): ) await poll_for_task_status(subscription_key, cls) download_list = await get_rodin_download_list(task_uuid, cls) - model = await download_files(download_list, task_uuid) + model_path, file_3d = await download_files(download_list, task_uuid) - return IO.NodeOutput(model) + return IO.NodeOutput(model_path, file_3d) class Rodin3DExtension(ComfyExtension): diff --git a/comfy_api_nodes/nodes_tripo.py b/comfy_api_nodes/nodes_tripo.py index 5abf27b4d..67c7f59fc 100644 --- a/comfy_api_nodes/nodes_tripo.py +++ b/comfy_api_nodes/nodes_tripo.py @@ -1,10 +1,6 @@ -import os -from typing import Optional - -import torch from typing_extensions import override -from comfy_api.latest import IO, ComfyExtension +from comfy_api.latest import IO, ComfyExtension, Input from comfy_api_nodes.apis.tripo import ( TripoAnimateRetargetRequest, TripoAnimateRigRequest, @@ -26,12 +22,11 @@ from comfy_api_nodes.apis.tripo import ( ) from comfy_api_nodes.util import ( ApiEndpoint, - download_url_as_bytesio, + download_url_to_file_3d, poll_op, sync_op, upload_images_to_comfyapi, ) -from folder_paths import get_output_directory def get_model_url_from_response(response: TripoTaskResponse) -> str: @@ -45,7 +40,7 @@ def get_model_url_from_response(response: TripoTaskResponse) -> str: async def poll_until_finished( node_cls: type[IO.ComfyNode], response: TripoTaskResponse, - average_duration: Optional[int] = None, + average_duration: int | None = None, ) -> IO.NodeOutput: """Polls the Tripo API endpoint until the task reaches a terminal state, then returns the response.""" if response.code != 0: @@ -69,12 +64,8 @@ async def poll_until_finished( ) if response_poll.data.status == TripoTaskStatus.SUCCESS: url = get_model_url_from_response(response_poll) - bytesio = await download_url_as_bytesio(url) - # Save the downloaded model file - model_file = f"tripo_model_{task_id}.glb" - with open(os.path.join(get_output_directory(), model_file), "wb") as f: - f.write(bytesio.getvalue()) - return IO.NodeOutput(model_file, task_id) + file_glb = await download_url_to_file_3d(url, "glb", task_id=task_id) + return IO.NodeOutput(f"{task_id}.glb", task_id, file_glb) raise RuntimeError(f"Failed to generate mesh: {response_poll}") @@ -107,8 +98,9 @@ class TripoTextToModelNode(IO.ComfyNode): IO.Combo.Input("geometry_quality", default="standard", options=["standard", "detailed"], optional=True), ], outputs=[ - IO.String.Output(display_name="model_file"), + IO.String.Output(display_name="model_file"), # for backward compatibility only IO.Custom("MODEL_TASK_ID").Output(display_name="model task_id"), + IO.File3DGLB.Output(display_name="GLB"), ], hidden=[ IO.Hidden.auth_token_comfy_org, @@ -155,18 +147,18 @@ class TripoTextToModelNode(IO.ComfyNode): async def execute( cls, prompt: str, - negative_prompt: Optional[str] = None, + negative_prompt: str | None = None, model_version=None, - style: Optional[str] = None, - texture: Optional[bool] = None, - pbr: Optional[bool] = None, - image_seed: Optional[int] = None, - model_seed: Optional[int] = None, - texture_seed: Optional[int] = None, - texture_quality: Optional[str] = None, - geometry_quality: Optional[str] = None, - face_limit: Optional[int] = None, - quad: Optional[bool] = None, + style: str | None = None, + texture: bool | None = None, + pbr: bool | None = None, + image_seed: int | None = None, + model_seed: int | None = None, + texture_seed: int | None = None, + texture_quality: str | None = None, + geometry_quality: str | None = None, + face_limit: int | None = None, + quad: bool | None = None, ) -> IO.NodeOutput: style_enum = None if style == "None" else style if not prompt: @@ -232,8 +224,9 @@ class TripoImageToModelNode(IO.ComfyNode): IO.Combo.Input("geometry_quality", default="standard", options=["standard", "detailed"], optional=True), ], outputs=[ - IO.String.Output(display_name="model_file"), + IO.String.Output(display_name="model_file"), # for backward compatibility only IO.Custom("MODEL_TASK_ID").Output(display_name="model task_id"), + IO.File3DGLB.Output(display_name="GLB"), ], hidden=[ IO.Hidden.auth_token_comfy_org, @@ -279,19 +272,19 @@ class TripoImageToModelNode(IO.ComfyNode): @classmethod async def execute( cls, - image: torch.Tensor, - model_version: Optional[str] = None, - style: Optional[str] = None, - texture: Optional[bool] = None, - pbr: Optional[bool] = None, - model_seed: Optional[int] = None, + image: Input.Image, + model_version: str | None = None, + style: str | None = None, + texture: bool | None = None, + pbr: bool | None = None, + model_seed: int | None = None, orientation=None, - texture_seed: Optional[int] = None, - texture_quality: Optional[str] = None, - geometry_quality: Optional[str] = None, - texture_alignment: Optional[str] = None, - face_limit: Optional[int] = None, - quad: Optional[bool] = None, + texture_seed: int | None = None, + texture_quality: str | None = None, + geometry_quality: str | None = None, + texture_alignment: str | None = None, + face_limit: int | None = None, + quad: bool | None = None, ) -> IO.NodeOutput: style_enum = None if style == "None" else style if image is None: @@ -368,8 +361,9 @@ class TripoMultiviewToModelNode(IO.ComfyNode): IO.Combo.Input("geometry_quality", default="standard", options=["standard", "detailed"], optional=True), ], outputs=[ - IO.String.Output(display_name="model_file"), + IO.String.Output(display_name="model_file"), # for backward compatibility only IO.Custom("MODEL_TASK_ID").Output(display_name="model task_id"), + IO.File3DGLB.Output(display_name="GLB"), ], hidden=[ IO.Hidden.auth_token_comfy_org, @@ -411,21 +405,21 @@ class TripoMultiviewToModelNode(IO.ComfyNode): @classmethod async def execute( cls, - image: torch.Tensor, - image_left: Optional[torch.Tensor] = None, - image_back: Optional[torch.Tensor] = None, - image_right: Optional[torch.Tensor] = None, - model_version: Optional[str] = None, - orientation: Optional[str] = None, - texture: Optional[bool] = None, - pbr: Optional[bool] = None, - model_seed: Optional[int] = None, - texture_seed: Optional[int] = None, - texture_quality: Optional[str] = None, - geometry_quality: Optional[str] = None, - texture_alignment: Optional[str] = None, - face_limit: Optional[int] = None, - quad: Optional[bool] = None, + image: Input.Image, + image_left: Input.Image | None = None, + image_back: Input.Image | None = None, + image_right: Input.Image | None = None, + model_version: str | None = None, + orientation: str | None = None, + texture: bool | None = None, + pbr: bool | None = None, + model_seed: int | None = None, + texture_seed: int | None = None, + texture_quality: str | None = None, + geometry_quality: str | None = None, + texture_alignment: str | None = None, + face_limit: int | None = None, + quad: bool | None = None, ) -> IO.NodeOutput: if image is None: raise RuntimeError("front image for multiview is required") @@ -487,8 +481,9 @@ class TripoTextureNode(IO.ComfyNode): ), ], outputs=[ - IO.String.Output(display_name="model_file"), + IO.String.Output(display_name="model_file"), # for backward compatibility only IO.Custom("MODEL_TASK_ID").Output(display_name="model task_id"), + IO.File3DGLB.Output(display_name="GLB"), ], hidden=[ IO.Hidden.auth_token_comfy_org, @@ -512,11 +507,11 @@ class TripoTextureNode(IO.ComfyNode): async def execute( cls, model_task_id, - texture: Optional[bool] = None, - pbr: Optional[bool] = None, - texture_seed: Optional[int] = None, - texture_quality: Optional[str] = None, - texture_alignment: Optional[str] = None, + texture: bool | None = None, + pbr: bool | None = None, + texture_seed: int | None = None, + texture_quality: str | None = None, + texture_alignment: str | None = None, ) -> IO.NodeOutput: response = await sync_op( cls, @@ -547,8 +542,9 @@ class TripoRefineNode(IO.ComfyNode): IO.Custom("MODEL_TASK_ID").Input("model_task_id", tooltip="Must be a v1.4 Tripo model"), ], outputs=[ - IO.String.Output(display_name="model_file"), + IO.String.Output(display_name="model_file"), # for backward compatibility only IO.Custom("MODEL_TASK_ID").Output(display_name="model task_id"), + IO.File3DGLB.Output(display_name="GLB"), ], hidden=[ IO.Hidden.auth_token_comfy_org, @@ -583,8 +579,9 @@ class TripoRigNode(IO.ComfyNode): category="api node/3d/Tripo", inputs=[IO.Custom("MODEL_TASK_ID").Input("original_model_task_id")], outputs=[ - IO.String.Output(display_name="model_file"), + IO.String.Output(display_name="model_file"), # for backward compatibility only IO.Custom("RIG_TASK_ID").Output(display_name="rig task_id"), + IO.File3DGLB.Output(display_name="GLB"), ], hidden=[ IO.Hidden.auth_token_comfy_org, @@ -642,8 +639,9 @@ class TripoRetargetNode(IO.ComfyNode): ), ], outputs=[ - IO.String.Output(display_name="model_file"), + IO.String.Output(display_name="model_file"), # for backward compatibility only IO.Custom("RETARGET_TASK_ID").Output(display_name="retarget task_id"), + IO.File3DGLB.Output(display_name="GLB"), ], hidden=[ IO.Hidden.auth_token_comfy_org, diff --git a/comfy_api_nodes/util/__init__.py b/comfy_api_nodes/util/__init__.py index c3c9ff4bf..18b020eef 100644 --- a/comfy_api_nodes/util/__init__.py +++ b/comfy_api_nodes/util/__init__.py @@ -28,6 +28,7 @@ from .conversions import ( from .download_helpers import ( download_url_as_bytesio, download_url_to_bytesio, + download_url_to_file_3d, download_url_to_image_tensor, download_url_to_video_output, ) @@ -69,6 +70,7 @@ __all__ = [ # Download helpers "download_url_as_bytesio", "download_url_to_bytesio", + "download_url_to_file_3d", "download_url_to_image_tensor", "download_url_to_video_output", # Conversions diff --git a/comfy_api_nodes/util/download_helpers.py b/comfy_api_nodes/util/download_helpers.py index 4668d14a9..78bcf1fa1 100644 --- a/comfy_api_nodes/util/download_helpers.py +++ b/comfy_api_nodes/util/download_helpers.py @@ -11,7 +11,8 @@ import torch from aiohttp.client_exceptions import ClientError, ContentTypeError from comfy_api.latest import IO as COMFY_IO -from comfy_api.latest import InputImpl +from comfy_api.latest import InputImpl, Types +from folder_paths import get_output_directory from . import request_logger from ._helpers import ( @@ -261,3 +262,38 @@ def _generate_operation_id(method: str, url: str, attempt: int) -> str: except Exception: slug = "download" return f"{method}_{slug}_try{attempt}_{uuid.uuid4().hex[:8]}" + + +async def download_url_to_file_3d( + url: str, + file_format: str, + *, + task_id: str | None = None, + timeout: float | None = None, + max_retries: int = 5, + cls: type[COMFY_IO.ComfyNode] = None, +) -> Types.File3D: + """Downloads a 3D model file from a URL into memory as BytesIO. + + If task_id is provided, also writes the file to disk in the output directory + for backward compatibility with the old save-to-disk behavior. + """ + file_format = file_format.lstrip(".").lower() + data = BytesIO() + await download_url_to_bytesio( + url, + data, + timeout=timeout, + max_retries=max_retries, + cls=cls, + ) + + if task_id is not None: + # This is only for backward compatability with current behavior when every 3D node is output node + # All new API nodes should not use "task_id" and instead users should use "SaveGLB" node to save results + output_dir = Path(get_output_directory()) + output_path = output_dir / f"{task_id}.{file_format}" + output_path.write_bytes(data.getvalue()) + data.seek(0) + + return Types.File3D(source=data, file_format=file_format) diff --git a/comfy_extras/nodes_hunyuan3d.py b/comfy_extras/nodes_hunyuan3d.py index 5bb5df48e..eda1639ab 100644 --- a/comfy_extras/nodes_hunyuan3d.py +++ b/comfy_extras/nodes_hunyuan3d.py @@ -622,14 +622,20 @@ class SaveGLB(IO.ComfyNode): category="3d", is_output_node=True, inputs=[ - IO.Mesh.Input("mesh"), + IO.MultiType.Input( + IO.Mesh.Input("mesh"), + types=[ + IO.File3DGLB, + ], + tooltip="Mesh or GLB file to save", + ), IO.String.Input("filename_prefix", default="mesh/ComfyUI"), ], hidden=[IO.Hidden.prompt, IO.Hidden.extra_pnginfo] ) @classmethod - def execute(cls, mesh, filename_prefix) -> IO.NodeOutput: + def execute(cls, mesh: Types.MESH | Types.File3D, filename_prefix: str) -> IO.NodeOutput: full_output_folder, filename, counter, subfolder, filename_prefix = folder_paths.get_save_image_path(filename_prefix, folder_paths.get_output_directory()) results = [] @@ -641,15 +647,26 @@ class SaveGLB(IO.ComfyNode): for x in cls.hidden.extra_pnginfo: metadata[x] = json.dumps(cls.hidden.extra_pnginfo[x]) - for i in range(mesh.vertices.shape[0]): + if isinstance(mesh, Types.File3D): + # Handle File3D input - save BytesIO data to output folder f = f"{filename}_{counter:05}_.glb" - save_glb(mesh.vertices[i], mesh.faces[i], os.path.join(full_output_folder, f), metadata) + mesh.save_to(os.path.join(full_output_folder, f)) results.append({ "filename": f, "subfolder": subfolder, "type": "output" }) - counter += 1 + else: + # Handle Mesh input - save vertices and faces as GLB + for i in range(mesh.vertices.shape[0]): + f = f"{filename}_{counter:05}_.glb" + save_glb(mesh.vertices[i], mesh.faces[i], os.path.join(full_output_folder, f), metadata) + results.append({ + "filename": f, + "subfolder": subfolder, + "type": "output" + }) + counter += 1 return IO.NodeOutput(ui={"3d": results}) diff --git a/comfy_extras/nodes_load_3d.py b/comfy_extras/nodes_load_3d.py index 4b8d950ae..f29510488 100644 --- a/comfy_extras/nodes_load_3d.py +++ b/comfy_extras/nodes_load_3d.py @@ -1,9 +1,10 @@ import nodes import folder_paths import os +import uuid from typing_extensions import override -from comfy_api.latest import IO, ComfyExtension, InputImpl, UI +from comfy_api.latest import IO, UI, ComfyExtension, InputImpl, Types from pathlib import Path @@ -81,7 +82,19 @@ class Preview3D(IO.ComfyNode): is_experimental=True, is_output_node=True, inputs=[ - IO.String.Input("model_file", default="", multiline=False), + IO.MultiType.Input( + IO.String.Input("model_file", default="", multiline=False), + types=[ + IO.File3DGLB, + IO.File3DGLTF, + IO.File3DFBX, + IO.File3DOBJ, + IO.File3DSTL, + IO.File3DUSDZ, + IO.File3DAny, + ], + tooltip="3D model file or path string", + ), IO.Load3DCamera.Input("camera_info", optional=True), IO.Image.Input("bg_image", optional=True), ], @@ -89,10 +102,15 @@ class Preview3D(IO.ComfyNode): ) @classmethod - def execute(cls, model_file, **kwargs) -> IO.NodeOutput: + def execute(cls, model_file: str | Types.File3D, **kwargs) -> IO.NodeOutput: + if isinstance(model_file, Types.File3D): + filename = f"preview3d_{uuid.uuid4().hex}.{model_file.format}" + model_file.save_to(os.path.join(folder_paths.get_output_directory(), filename)) + else: + filename = model_file camera_info = kwargs.get("camera_info", None) bg_image = kwargs.get("bg_image", None) - return IO.NodeOutput(ui=UI.PreviewUI3D(model_file, camera_info, bg_image=bg_image)) + return IO.NodeOutput(ui=UI.PreviewUI3D(filename, camera_info, bg_image=bg_image)) process = execute # TODO: remove From ab1050bec3ecb89d1bc06c13e67626805bd8e1ee Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 3 Feb 2026 10:54:23 -0800 Subject: [PATCH 565/647] Support ace step 1.5 base model loras. (#12252) --- comfy/lora.py | 6 ++++++ 1 file changed, 6 insertions(+) diff --git a/comfy/lora.py b/comfy/lora.py index 7b31d055c..44030bcab 100644 --- a/comfy/lora.py +++ b/comfy/lora.py @@ -332,6 +332,12 @@ def model_lora_keys_unet(model, key_map={}): key_map["{}".format(key_lora)] = k key_map["transformer.{}".format(key_lora)] = k + if isinstance(model, comfy.model_base.ACEStep15): + for k in sdk: + if k.startswith("diffusion_model.decoder.") and k.endswith(".weight"): + key_lora = k[len("diffusion_model.decoder."):-len(".weight")] + key_map["base_model.model.{}".format(key_lora)] = k # Official base model loras + return key_map From b8315e66cbcf675d683ed83892314cbd3a3d1bf7 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 3 Feb 2026 11:40:45 -0800 Subject: [PATCH 566/647] Fix tiled vae for ace step 1.5 (#12253) --- comfy/sd.py | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-) diff --git a/comfy/sd.py b/comfy/sd.py index 722c0c154..aa93e101d 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -554,6 +554,8 @@ class VAE: elif "decoder.layers.1.layers.0.beta" in sd: config = {} param_key = None + self.upscale_ratio = 2048 + self.downscale_ratio = 2048 if "decoder.layers.2.layers.1.weight_v" in sd: param_key = "decoder.layers.2.layers.1.weight_v" if "decoder.layers.2.layers.1.parametrizations.weight.original1" in sd: @@ -562,6 +564,8 @@ class VAE: if sd[param_key].shape[-1] == 12: config["strides"] = [2, 4, 4, 6, 10] self.audio_sample_rate = 48000 + self.upscale_ratio = 1920 + self.downscale_ratio = 1920 self.first_stage_model = AudioOobleckVAE(**config) self.memory_used_encode = lambda shape, dtype: (1000 * shape[2]) * model_management.dtype_size(dtype) @@ -569,8 +573,6 @@ class VAE: self.latent_channels = 64 self.output_channels = 2 self.pad_channel_value = "replicate" - self.upscale_ratio = 2048 - self.downscale_ratio = 2048 self.latent_dim = 1 self.process_output = lambda audio: audio self.process_input = lambda audio: audio @@ -870,7 +872,7 @@ class VAE: / 3.0) return output - def decode_tiled_1d(self, samples, tile_x=128, overlap=32): + def decode_tiled_1d(self, samples, tile_x=256, overlap=32): if samples.ndim == 3: decode_fn = lambda a: self.first_stage_model.decode(a.to(self.vae_dtype).to(self.device)).float() else: From 3be017516643d9a8d8ed85904512ed4b240ef8c6 Mon Sep 17 00:00:00 2001 From: comfyanonymous Date: Tue, 3 Feb 2026 14:13:18 -0500 Subject: [PATCH 567/647] ComfyUI v0.12.1 --- comfyui_version.py | 2 +- pyproject.toml | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/comfyui_version.py b/comfyui_version.py index bc6076b67..2e2c12ced 100644 --- a/comfyui_version.py +++ b/comfyui_version.py @@ -1,3 +1,3 @@ # This file is automatically generated by the build process when version is # updated in pyproject.toml. -__version__ = "0.12.0" +__version__ = "0.12.1" diff --git a/pyproject.toml b/pyproject.toml index 28aa03067..c21ee03f1 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -1,6 +1,6 @@ [project] name = "ComfyUI" -version = "0.12.0" +version = "0.12.1" readme = "README.md" license = { file = "LICENSE" } requires-python = ">=3.10" From fe2511468d6b3f7a1581a711ae480f3313f2b39d Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 3 Feb 2026 16:01:38 -0800 Subject: [PATCH 568/647] Support the 4B ace step 1.5 lm model. (#12257) Can be used as an alternative to the 1.7B --- comfy/sd.py | 7 +++- comfy/supported_models.py | 12 +++++- comfy/text_encoders/ace15.py | 42 ++++++++++++++++----- comfy/text_encoders/llama.py | 72 ++++++++++++++++++++++++++---------- 4 files changed, 101 insertions(+), 32 deletions(-) diff --git a/comfy/sd.py b/comfy/sd.py index aa93e101d..bc63d6ced 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -1444,7 +1444,12 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip tokenizer_data["gemma_spiece_model"] = clip_data_gemma.get("spiece_model", None) tokenizer_data["jina_spiece_model"] = clip_data_jina.get("spiece_model", None) elif clip_type == CLIPType.ACE: - clip_target.clip = comfy.text_encoders.ace15.te(**llama_detect(clip_data)) + te_models = [detect_te_model(clip_data[0]), detect_te_model(clip_data[1])] + if TEModel.QWEN3_4B in te_models: + model_type = "qwen3_4b" + else: + model_type = "qwen3_2b" + clip_target.clip = comfy.text_encoders.ace15.te(lm_model=model_type, **llama_detect(clip_data)) clip_target.tokenizer = comfy.text_encoders.ace15.ACE15Tokenizer else: clip_target.clip = sdxl_clip.SDXLClipModel diff --git a/comfy/supported_models.py b/comfy/supported_models.py index 6b7d831cb..77264ed28 100644 --- a/comfy/supported_models.py +++ b/comfy/supported_models.py @@ -1625,8 +1625,16 @@ class ACEStep15(supported_models_base.BASE): def clip_target(self, state_dict={}): pref = self.text_encoder_key_prefix[0] - hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen3_2b.transformer.".format(pref)) - return supported_models_base.ClipTarget(comfy.text_encoders.ace15.ACE15Tokenizer, comfy.text_encoders.ace15.te(**hunyuan_detect)) + detect_2b = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen3_2b.transformer.".format(pref)) + detect_4b = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen3_4b.transformer.".format(pref)) + if "dtype_llama" in detect_2b: + detect = detect_2b + detect["lm_model"] = "qwen3_2b" + elif "dtype_llama" in detect_4b: + detect = detect_4b + detect["lm_model"] = "qwen3_4b" + + return supported_models_base.ClipTarget(comfy.text_encoders.ace15.ACE15Tokenizer, comfy.text_encoders.ace15.te(**detect)) models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, LTXAV, HunyuanVideo15_SR_Distilled, HunyuanVideo15, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, ZImage, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, ACEStep15, Omnigen2, QwenImage, Flux2, Kandinsky5Image, Kandinsky5, Anima] diff --git a/comfy/text_encoders/ace15.py b/comfy/text_encoders/ace15.py index 48c29fa9a..73d710671 100644 --- a/comfy/text_encoders/ace15.py +++ b/comfy/text_encoders/ace15.py @@ -162,14 +162,34 @@ class Qwen3_2B_ACE15(sd1_clip.SDClipModel): super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"pad": 151643}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Qwen3_2B_ACE15_lm, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options) +class Qwen3_4B_ACE15(sd1_clip.SDClipModel): + def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, attention_mask=True, model_options={}): + llama_quantization_metadata = model_options.get("llama_quantization_metadata", None) + if llama_quantization_metadata is not None: + model_options = model_options.copy() + model_options["quantization_metadata"] = llama_quantization_metadata + + super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"pad": 151643}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Qwen3_4B_ACE15_lm, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options) + class ACE15TEModel(torch.nn.Module): - def __init__(self, device="cpu", dtype=None, dtype_llama=None, model_options={}): + def __init__(self, device="cpu", dtype=None, dtype_llama=None, lm_model=None, model_options={}): super().__init__() if dtype_llama is None: dtype_llama = dtype + model = None + self.constant = 0.4375 + if lm_model == "qwen3_4b": + model = Qwen3_4B_ACE15 + self.constant = 0.5625 + elif lm_model == "qwen3_2b": + model = Qwen3_2B_ACE15 + + self.lm_model = lm_model self.qwen3_06b = Qwen3_06BModel(device=device, dtype=dtype, model_options=model_options) - self.qwen3_2b = Qwen3_2B_ACE15(device=device, dtype=dtype_llama, model_options=model_options) + if model is not None: + setattr(self, self.lm_model, model(device=device, dtype=dtype_llama, model_options=model_options)) + self.dtypes = set([dtype, dtype_llama]) def encode_token_weights(self, token_weight_pairs): @@ -182,17 +202,21 @@ class ACE15TEModel(torch.nn.Module): lyrics_embeds, _, extra_l = self.qwen3_06b.encode_token_weights(token_weight_pairs_lyrics) lm_metadata = token_weight_pairs["lm_metadata"] - audio_codes = generate_audio_codes(self.qwen3_2b, token_weight_pairs["lm_prompt"], token_weight_pairs["lm_prompt_negative"], min_tokens=lm_metadata["min_tokens"], max_tokens=lm_metadata["min_tokens"], seed=lm_metadata["seed"]) + audio_codes = generate_audio_codes(getattr(self, self.lm_model, self.qwen3_06b), token_weight_pairs["lm_prompt"], token_weight_pairs["lm_prompt_negative"], min_tokens=lm_metadata["min_tokens"], max_tokens=lm_metadata["min_tokens"], seed=lm_metadata["seed"]) return base_out, None, {"conditioning_lyrics": lyrics_embeds[:, 0], "audio_codes": [audio_codes]} def set_clip_options(self, options): self.qwen3_06b.set_clip_options(options) - self.qwen3_2b.set_clip_options(options) + lm_model = getattr(self, self.lm_model, None) + if lm_model is not None: + lm_model.set_clip_options(options) def reset_clip_options(self): self.qwen3_06b.reset_clip_options() - self.qwen3_2b.reset_clip_options() + lm_model = getattr(self, self.lm_model, None) + if lm_model is not None: + lm_model.reset_clip_options() def load_sd(self, sd): if "model.layers.0.post_attention_layernorm.weight" in sd: @@ -200,11 +224,11 @@ class ACE15TEModel(torch.nn.Module): if shape[0] == 1024: return self.qwen3_06b.load_sd(sd) else: - return self.qwen3_2b.load_sd(sd) + return getattr(self, self.lm_model).load_sd(sd) def memory_estimation_function(self, token_weight_pairs, device=None): lm_metadata = token_weight_pairs["lm_metadata"] - constant = 0.4375 + constant = self.constant if comfy.model_management.should_use_bf16(device): constant *= 0.5 @@ -213,11 +237,11 @@ class ACE15TEModel(torch.nn.Module): num_tokens += lm_metadata['min_tokens'] return num_tokens * constant * 1024 * 1024 -def te(dtype_llama=None, llama_quantization_metadata=None): +def te(dtype_llama=None, llama_quantization_metadata=None, lm_model="qwen3_2b"): class ACE15TEModel_(ACE15TEModel): def __init__(self, device="cpu", dtype=None, model_options={}): if llama_quantization_metadata is not None: model_options = model_options.copy() model_options["llama_quantization_metadata"] = llama_quantization_metadata - super().__init__(device=device, dtype_llama=dtype_llama, dtype=dtype, model_options=model_options) + super().__init__(device=device, dtype_llama=dtype_llama, lm_model=lm_model, dtype=dtype, model_options=model_options) return ACE15TEModel_ diff --git a/comfy/text_encoders/llama.py b/comfy/text_encoders/llama.py index f4ac224c6..3afd094d1 100644 --- a/comfy/text_encoders/llama.py +++ b/comfy/text_encoders/llama.py @@ -150,6 +150,29 @@ class Qwen3_2B_ACE15_lm_Config: final_norm: bool = True lm_head: bool = False +@dataclass +class Qwen3_4B_ACE15_lm_Config: + vocab_size: int = 217204 + hidden_size: int = 2560 + intermediate_size: int = 9728 + num_hidden_layers: int = 36 + num_attention_heads: int = 32 + num_key_value_heads: int = 8 + max_position_embeddings: int = 40960 + rms_norm_eps: float = 1e-6 + rope_theta: float = 1000000.0 + transformer_type: str = "llama" + head_dim = 128 + rms_norm_add = False + mlp_activation = "silu" + qkv_bias = False + rope_dims = None + q_norm = "gemma3" + k_norm = "gemma3" + rope_scale = None + final_norm: bool = True + lm_head: bool = False + @dataclass class Qwen3_4BConfig: vocab_size: int = 151936 @@ -739,6 +762,21 @@ class BaseLlama: def forward(self, input_ids, *args, **kwargs): return self.model(input_ids, *args, **kwargs) +class BaseQwen3: + def logits(self, x): + input = x[:, -1:] + module = self.model.embed_tokens + + offload_stream = None + if module.comfy_cast_weights: + weight, _, offload_stream = comfy.ops.cast_bias_weight(module, input, offloadable=True) + else: + weight = self.model.embed_tokens.weight.to(x) + + x = torch.nn.functional.linear(input, weight, None) + + comfy.ops.uncast_bias_weight(module, weight, None, offload_stream) + return x class Llama2(BaseLlama, torch.nn.Module): def __init__(self, config_dict, dtype, device, operations): @@ -767,7 +805,7 @@ class Qwen25_3B(BaseLlama, torch.nn.Module): self.model = Llama2_(config, device=device, dtype=dtype, ops=operations) self.dtype = dtype -class Qwen3_06B(BaseLlama, torch.nn.Module): +class Qwen3_06B(BaseLlama, BaseQwen3, torch.nn.Module): def __init__(self, config_dict, dtype, device, operations): super().__init__() config = Qwen3_06BConfig(**config_dict) @@ -776,7 +814,7 @@ class Qwen3_06B(BaseLlama, torch.nn.Module): self.model = Llama2_(config, device=device, dtype=dtype, ops=operations) self.dtype = dtype -class Qwen3_06B_ACE15(BaseLlama, torch.nn.Module): +class Qwen3_06B_ACE15(BaseLlama, BaseQwen3, torch.nn.Module): def __init__(self, config_dict, dtype, device, operations): super().__init__() config = Qwen3_06B_ACE15_Config(**config_dict) @@ -785,7 +823,7 @@ class Qwen3_06B_ACE15(BaseLlama, torch.nn.Module): self.model = Llama2_(config, device=device, dtype=dtype, ops=operations) self.dtype = dtype -class Qwen3_2B_ACE15_lm(BaseLlama, torch.nn.Module): +class Qwen3_2B_ACE15_lm(BaseLlama, BaseQwen3, torch.nn.Module): def __init__(self, config_dict, dtype, device, operations): super().__init__() config = Qwen3_2B_ACE15_lm_Config(**config_dict) @@ -794,22 +832,7 @@ class Qwen3_2B_ACE15_lm(BaseLlama, torch.nn.Module): self.model = Llama2_(config, device=device, dtype=dtype, ops=operations) self.dtype = dtype - def logits(self, x): - input = x[:, -1:] - module = self.model.embed_tokens - - offload_stream = None - if module.comfy_cast_weights: - weight, _, offload_stream = comfy.ops.cast_bias_weight(module, input, offloadable=True) - else: - weight = self.model.embed_tokens.weight.to(x) - - x = torch.nn.functional.linear(input, weight, None) - - comfy.ops.uncast_bias_weight(module, weight, None, offload_stream) - return x - -class Qwen3_4B(BaseLlama, torch.nn.Module): +class Qwen3_4B(BaseLlama, BaseQwen3, torch.nn.Module): def __init__(self, config_dict, dtype, device, operations): super().__init__() config = Qwen3_4BConfig(**config_dict) @@ -818,7 +841,16 @@ class Qwen3_4B(BaseLlama, torch.nn.Module): self.model = Llama2_(config, device=device, dtype=dtype, ops=operations) self.dtype = dtype -class Qwen3_8B(BaseLlama, torch.nn.Module): +class Qwen3_4B_ACE15_lm(BaseLlama, BaseQwen3, torch.nn.Module): + def __init__(self, config_dict, dtype, device, operations): + super().__init__() + config = Qwen3_4B_ACE15_lm_Config(**config_dict) + self.num_layers = config.num_hidden_layers + + self.model = Llama2_(config, device=device, dtype=dtype, ops=operations) + self.dtype = dtype + +class Qwen3_8B(BaseLlama, BaseQwen3, torch.nn.Module): def __init__(self, config_dict, dtype, device, operations): super().__init__() config = Qwen3_8BConfig(**config_dict) From 855849c6588180fec88186127aae1a3299387fa6 Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Tue, 3 Feb 2026 18:39:19 -0800 Subject: [PATCH 569/647] mm: Remove Aimdo exemption for empty_cache (#12260) Its more important to get the torch caching allocator GC up and running than supporting the pyt2.7 bug. Switch it on. Defeature dynamic_vram + pyt2.7. --- comfy/model_management.py | 8 +++----- main.py | 7 +++++-- 2 files changed, 8 insertions(+), 7 deletions(-) diff --git a/comfy/model_management.py b/comfy/model_management.py index 72348258b..b6291f340 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -1724,11 +1724,9 @@ def soft_empty_cache(force=False): elif is_mlu(): torch.mlu.empty_cache() elif torch.cuda.is_available(): - if comfy.memory_management.aimdo_allocator is None: - #Pytorch 2.7 and earlier crashes if you try and empty_cache when mempools exist - torch.cuda.synchronize() - torch.cuda.empty_cache() - torch.cuda.ipc_collect() + torch.cuda.synchronize() + torch.cuda.empty_cache() + torch.cuda.ipc_collect() def unload_all_models(): free_memory(1e30, get_torch_device()) diff --git a/main.py b/main.py index b8c951375..92d705b4d 100644 --- a/main.py +++ b/main.py @@ -192,7 +192,10 @@ import comfy_aimdo.control import comfy_aimdo.torch if enables_dynamic_vram(): - if comfy_aimdo.control.init_device(comfy.model_management.get_torch_device().index): + if comfy.model_management.torch_version_numeric < (2, 8): + logging.warning("Unsupported Pytorch detected. DynamicVRAM support requires Pytorch version 2.8 or later. Falling back to legacy ModelPatcher. VRAM estimates may be unreliable especially on Windows") + comfy.memory_management.aimdo_allocator = None + elif comfy_aimdo.control.init_device(comfy.model_management.get_torch_device().index): if args.verbose == 'DEBUG': comfy_aimdo.control.set_log_debug() elif args.verbose == 'CRITICAL': @@ -208,7 +211,7 @@ if enables_dynamic_vram(): comfy.memory_management.aimdo_allocator = comfy_aimdo.torch.get_torch_allocator() logging.info("DynamicVRAM support detected and enabled") else: - logging.info("No working comfy-aimdo install detected. DynamicVRAM support disabled. Falling back to legacy ModelPatcher. VRAM estimates may be unreliable especially on Windows") + logging.warning("No working comfy-aimdo install detected. DynamicVRAM support disabled. Falling back to legacy ModelPatcher. VRAM estimates may be unreliable especially on Windows") comfy.memory_management.aimdo_allocator = None From a31681564d9bd49530df593d87048232acecd842 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 3 Feb 2026 21:03:21 -0800 Subject: [PATCH 570/647] Fix crash with ace step 1.5 (#12264) --- comfy/text_encoders/ace15.py | 2 ++ 1 file changed, 2 insertions(+) diff --git a/comfy/text_encoders/ace15.py b/comfy/text_encoders/ace15.py index 73d710671..fce2b67ce 100644 --- a/comfy/text_encoders/ace15.py +++ b/comfy/text_encoders/ace15.py @@ -19,6 +19,7 @@ def sample_manual_loop_no_classes( min_tokens: int = 1, max_new_tokens: int = 2048, audio_start_id: int = 151669, # The cutoff ID for audio codes + audio_end_id: int = 215669, eos_token_id: int = 151645, ): device = model.execution_device @@ -60,6 +61,7 @@ def sample_manual_loop_no_classes( remove_logit_value = torch.finfo(cfg_logits.dtype).min # Only generate audio tokens cfg_logits[:, :audio_start_id] = remove_logit_value + cfg_logits[:, audio_end_id:] = remove_logit_value if eos_token_id is not None and eos_token_id < audio_start_id and min_tokens < step: cfg_logits[:, eos_token_id] = eos_score From 5087f1d497c5b615fbb5d1ff03fcc1df308bd025 Mon Sep 17 00:00:00 2001 From: comfyanonymous Date: Wed, 4 Feb 2026 00:08:59 -0500 Subject: [PATCH 571/647] ComfyUI v0.12.2 --- comfyui_version.py | 2 +- pyproject.toml | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/comfyui_version.py b/comfyui_version.py index 2e2c12ced..5d296cd1b 100644 --- a/comfyui_version.py +++ b/comfyui_version.py @@ -1,3 +1,3 @@ # This file is automatically generated by the build process when version is # updated in pyproject.toml. -__version__ = "0.12.1" +__version__ = "0.12.2" diff --git a/pyproject.toml b/pyproject.toml index c21ee03f1..1ddcc3596 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -1,6 +1,6 @@ [project] name = "ComfyUI" -version = "0.12.1" +version = "0.12.2" readme = "README.md" license = { file = "LICENSE" } requires-python = ">=3.10" From d30c609f5aa1d0cc75852815308adbb6ae21c644 Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Tue, 3 Feb 2026 22:48:47 -0800 Subject: [PATCH 572/647] utils: safetensors: dont slice data on torch level (#12266) Torch has alignment enforcement when viewing with data type changes but only relative to itself. Do all tensor constructions straight off the memory-view individually so pytorch doesnt see an alignment problem. The is needed for handling misaligned safetensors weights, which are reasonably common in third party models. This limits usage of this safetensors loader to GPU compute only as CPUs kernnel are very likely to bus error. But it works for dynamic_vram, where we really dont want to take a deep copy and we always use GPU copy_ which disentangles the misalignment. --- comfy/utils.py | 14 +++++++++----- 1 file changed, 9 insertions(+), 5 deletions(-) diff --git a/comfy/utils.py b/comfy/utils.py index c1b536833..1337e2205 100644 --- a/comfy/utils.py +++ b/comfy/utils.py @@ -82,14 +82,12 @@ _TYPES = { def load_safetensors(ckpt): f = open(ckpt, "rb") mapping = mmap.mmap(f.fileno(), 0, access=mmap.ACCESS_READ) + mv = memoryview(mapping) header_size = struct.unpack(" Date: Tue, 3 Feb 2026 23:08:45 -0800 Subject: [PATCH 573/647] mp: Fix checkpoint saving (#12268) Fix regression in the recent model saving refactor. Pass the non unet pieces down the layers so that checkpoints are complete. --- comfy/model_patcher.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index cdf289395..d888dbcfb 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -1400,7 +1400,7 @@ class ModelPatcher: continue key = "diffusion_model." + k unet_state_dict[k] = LazyCastingParam(self, key, comfy.utils.get_attr(self.model, key)) - return self.model.state_dict_for_saving(unet_state_dict) + return self.model.state_dict_for_saving(unet_state_dict, clip_state_dict=clip_state_dict, vae_state_dict=vae_state_dict, clip_vision_state_dict=clip_vision_state_dict) def __del__(self): self.unpin_all_weights() From e77b34dfead9758cf5e32b4cffeada0d0c56ab7d Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Wed, 4 Feb 2026 21:35:38 +0200 Subject: [PATCH 574/647] add File3DAny output to Load3D node; extend SaveGLB to accept File3DAny as input (#12276) * add File3DAny output to Load3D node; extend SaveGLB node to accept File3DAny as input * fix(grammar): capitalize letter --- comfy_extras/nodes_hunyuan3d.py | 12 ++++++++++-- comfy_extras/nodes_load_3d.py | 4 +++- 2 files changed, 13 insertions(+), 3 deletions(-) diff --git a/comfy_extras/nodes_hunyuan3d.py b/comfy_extras/nodes_hunyuan3d.py index eda1639ab..c2df3e859 100644 --- a/comfy_extras/nodes_hunyuan3d.py +++ b/comfy_extras/nodes_hunyuan3d.py @@ -618,6 +618,7 @@ class SaveGLB(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="SaveGLB", + display_name="Save 3D Model", search_aliases=["export 3d model", "save mesh"], category="3d", is_output_node=True, @@ -626,8 +627,14 @@ class SaveGLB(IO.ComfyNode): IO.Mesh.Input("mesh"), types=[ IO.File3DGLB, + IO.File3DGLTF, + IO.File3DOBJ, + IO.File3DFBX, + IO.File3DSTL, + IO.File3DUSDZ, + IO.File3DAny, ], - tooltip="Mesh or GLB file to save", + tooltip="Mesh or 3D file to save", ), IO.String.Input("filename_prefix", default="mesh/ComfyUI"), ], @@ -649,7 +656,8 @@ class SaveGLB(IO.ComfyNode): if isinstance(mesh, Types.File3D): # Handle File3D input - save BytesIO data to output folder - f = f"{filename}_{counter:05}_.glb" + ext = mesh.format or "glb" + f = f"{filename}_{counter:05}_.{ext}" mesh.save_to(os.path.join(full_output_folder, f)) results.append({ "filename": f, diff --git a/comfy_extras/nodes_load_3d.py b/comfy_extras/nodes_load_3d.py index f29510488..edbb5cd40 100644 --- a/comfy_extras/nodes_load_3d.py +++ b/comfy_extras/nodes_load_3d.py @@ -45,6 +45,7 @@ class Load3D(IO.ComfyNode): IO.Image.Output(display_name="normal"), IO.Load3DCamera.Output(display_name="camera_info"), IO.Video.Output(display_name="recording_video"), + IO.File3DAny.Output(display_name="model_3d"), ], ) @@ -66,7 +67,8 @@ class Load3D(IO.ComfyNode): video = InputImpl.VideoFromFile(recording_video_path) - return IO.NodeOutput(output_image, output_mask, model_file, normal_image, image['camera_info'], video) + file_3d = Types.File3D(folder_paths.get_annotated_filepath(model_file)) + return IO.NodeOutput(output_image, output_mask, model_file, normal_image, image['camera_info'], video, file_3d) process = execute # TODO: remove From 26dd7eb42180fb57c9da47e60d0a2bac659e47ad Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 4 Feb 2026 15:25:06 -0800 Subject: [PATCH 575/647] Fix ace step nan issue on some hardware/pytorch configs. (#12289) --- comfy/text_encoders/llama.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/comfy/text_encoders/llama.py b/comfy/text_encoders/llama.py index 3afd094d1..b6735d210 100644 --- a/comfy/text_encoders/llama.py +++ b/comfy/text_encoders/llama.py @@ -651,10 +651,10 @@ class Llama2_(nn.Module): mask = None if attention_mask is not None: mask = 1.0 - attention_mask.to(x.dtype).reshape((attention_mask.shape[0], 1, -1, attention_mask.shape[-1])).expand(attention_mask.shape[0], 1, seq_len, attention_mask.shape[-1]) - mask = mask.masked_fill(mask.to(torch.bool), torch.finfo(x.dtype).min) + mask = mask.masked_fill(mask.to(torch.bool), torch.finfo(x.dtype).min / 4) if seq_len > 1: - causal_mask = torch.empty(past_len + seq_len, past_len + seq_len, dtype=x.dtype, device=x.device).fill_(torch.finfo(x.dtype).min).triu_(1) + causal_mask = torch.empty(past_len + seq_len, past_len + seq_len, dtype=x.dtype, device=x.device).fill_(torch.finfo(x.dtype).min / 4).triu_(1) if mask is not None: mask += causal_mask else: From c8fcbd66eef0ab48d9fe7e4ee35c683a193af46b Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 4 Feb 2026 16:37:05 -0800 Subject: [PATCH 576/647] Try to fix ace text encoder slowness on some configs. (#12290) --- comfy/ops.py | 2 ++ 1 file changed, 2 insertions(+) diff --git a/comfy/ops.py b/comfy/ops.py index 53c5e4dc3..0f4eca7c7 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -54,6 +54,8 @@ try: SDPA_BACKEND_PRIORITY.insert(0, SDPBackend.CUDNN_ATTENTION) def scaled_dot_product_attention(q, k, v, *args, **kwargs): + if q.nelement() < 1024 * 128: # arbitrary number, for small inputs cudnn attention seems slower + return torch.nn.functional.scaled_dot_product_attention(q, k, v, *args, **kwargs) with sdpa_kernel(SDPA_BACKEND_PRIORITY, set_priority=True): return torch.nn.functional.scaled_dot_product_attention(q, k, v, *args, **kwargs) else: From 6125b8097952a374009af39639ff45da85f65500 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 4 Feb 2026 18:29:22 -0800 Subject: [PATCH 577/647] Add llm sampling options and make reference audio work on ace step 1.5 (#12295) --- comfy/ldm/ace/ace_step15.py | 3 +-- comfy/model_base.py | 19 ++++++++++++------- comfy/text_encoders/ace15.py | 31 +++++++++++++++++++++++-------- comfy_extras/nodes_ace.py | 16 +++++++++++----- 4 files changed, 47 insertions(+), 22 deletions(-) diff --git a/comfy/ldm/ace/ace_step15.py b/comfy/ldm/ace/ace_step15.py index d90549658..17a37e573 100644 --- a/comfy/ldm/ace/ace_step15.py +++ b/comfy/ldm/ace/ace_step15.py @@ -1035,8 +1035,7 @@ class AceStepConditionGenerationModel(nn.Module): audio_codes = torch.nn.functional.pad(audio_codes, (0, math.ceil(src_latents.shape[1] / 5) - audio_codes.shape[1]), "constant", 35847) lm_hints_5Hz = self.tokenizer.quantizer.get_output_from_indices(audio_codes, dtype=text_hidden_states.dtype) else: - assert False - # TODO ? + lm_hints_5Hz, indices = self.tokenizer.tokenize(refer_audio_acoustic_hidden_states_packed) lm_hints = self.detokenizer(lm_hints_5Hz) diff --git a/comfy/model_base.py b/comfy/model_base.py index 89944548c..a2a34f191 100644 --- a/comfy/model_base.py +++ b/comfy/model_base.py @@ -1548,6 +1548,7 @@ class ACEStep15(BaseModel): def extra_conds(self, **kwargs): out = super().extra_conds(**kwargs) device = kwargs["device"] + noise = kwargs["noise"] cross_attn = kwargs.get("cross_attn", None) if cross_attn is not None: @@ -1571,15 +1572,19 @@ class ACEStep15(BaseModel): 1.4844e-01, 9.4727e-02, 3.8477e-01, -1.2578e+00, -3.3203e-01, -8.5547e-01, 4.3359e-01, 4.2383e-01, -8.9453e-01, -5.0391e-01, -5.6152e-02, -2.9219e+00, -2.4658e-02, 5.0391e-01, 9.8438e-01, - 7.2754e-02, -2.1582e-01, 6.3672e-01, 1.0000e+00]]], device=device).movedim(-1, 1).repeat(1, 1, 750) + 7.2754e-02, -2.1582e-01, 6.3672e-01, 1.0000e+00]]], device=device).movedim(-1, 1).repeat(1, 1, noise.shape[2]) + pass_audio_codes = True else: - refer_audio = refer_audio[-1] + refer_audio = refer_audio[-1][:, :, :noise.shape[2]] + pass_audio_codes = False + + if pass_audio_codes: + audio_codes = kwargs.get("audio_codes", None) + if audio_codes is not None: + out['audio_codes'] = comfy.conds.CONDRegular(torch.tensor(audio_codes, device=device)) + refer_audio = refer_audio[:, :, :750] + out['refer_audio'] = comfy.conds.CONDRegular(refer_audio) - - audio_codes = kwargs.get("audio_codes", None) - if audio_codes is not None: - out['audio_codes'] = comfy.conds.CONDRegular(torch.tensor(audio_codes, device=device)) - return out class Omnigen2(BaseModel): diff --git a/comfy/text_encoders/ace15.py b/comfy/text_encoders/ace15.py index fce2b67ce..74e62733e 100644 --- a/comfy/text_encoders/ace15.py +++ b/comfy/text_encoders/ace15.py @@ -101,9 +101,7 @@ def sample_manual_loop_no_classes( return output_audio_codes -def generate_audio_codes(model, positive, negative, min_tokens=1, max_tokens=1024, seed=0): - cfg_scale = 2.0 - +def generate_audio_codes(model, positive, negative, min_tokens=1, max_tokens=1024, seed=0, cfg_scale=2.0, temperature=0.85, top_p=0.9, top_k=0): positive = [[token for token, _ in inner_list] for inner_list in positive] negative = [[token for token, _ in inner_list] for inner_list in negative] positive = positive[0] @@ -120,7 +118,7 @@ def generate_audio_codes(model, positive, negative, min_tokens=1, max_tokens=102 positive = [model.special_tokens["pad"]] * pos_pad + positive paddings = [pos_pad, neg_pad] - return sample_manual_loop_no_classes(model, [positive, negative], paddings, cfg_scale=cfg_scale, seed=seed, min_tokens=min_tokens, max_new_tokens=max_tokens) + return sample_manual_loop_no_classes(model, [positive, negative], paddings, cfg_scale=cfg_scale, temperature=temperature, top_p=top_p, top_k=top_k, seed=seed, min_tokens=min_tokens, max_new_tokens=max_tokens) class ACE15Tokenizer(sd1_clip.SD1Tokenizer): @@ -137,6 +135,12 @@ class ACE15Tokenizer(sd1_clip.SD1Tokenizer): language = kwargs.get("language", "en") seed = kwargs.get("seed", 0) + generate_audio_codes = kwargs.get("generate_audio_codes", True) + cfg_scale = kwargs.get("cfg_scale", 2.0) + temperature = kwargs.get("temperature", 0.85) + top_p = kwargs.get("top_p", 0.9) + top_k = kwargs.get("top_k", 0.0) + duration = math.ceil(duration) meta_lm = 'bpm: {}\nduration: {}\nkeyscale: {}\ntimesignature: {}'.format(bpm, duration, keyscale, timesignature) lm_template = "<|im_start|>system\n# Instruction\nGenerate audio semantic tokens based on the given conditions:\n\n<|im_end|>\n<|im_start|>user\n# Caption\n{}\n{}\n<|im_end|>\n<|im_start|>assistant\n\n{}\n\n\n<|im_end|>\n" @@ -147,7 +151,14 @@ class ACE15Tokenizer(sd1_clip.SD1Tokenizer): out["lyrics"] = self.qwen3_06b.tokenize_with_weights("# Languages\n{}\n\n# Lyric{}<|endoftext|><|endoftext|>".format(language, lyrics), return_word_ids, disable_weights=True, **kwargs) out["qwen3_06b"] = self.qwen3_06b.tokenize_with_weights("# Instruction\nGenerate audio semantic tokens based on the given conditions:\n\n# Caption\n{}# Metas\n{}<|endoftext|>\n<|endoftext|>".format(text, meta_cap), return_word_ids, **kwargs) - out["lm_metadata"] = {"min_tokens": duration * 5, "seed": seed} + out["lm_metadata"] = {"min_tokens": duration * 5, + "seed": seed, + "generate_audio_codes": generate_audio_codes, + "cfg_scale": cfg_scale, + "temperature": temperature, + "top_p": top_p, + "top_k": top_k, + } return out @@ -203,10 +214,14 @@ class ACE15TEModel(torch.nn.Module): self.qwen3_06b.set_clip_options({"layer": [0]}) lyrics_embeds, _, extra_l = self.qwen3_06b.encode_token_weights(token_weight_pairs_lyrics) - lm_metadata = token_weight_pairs["lm_metadata"] - audio_codes = generate_audio_codes(getattr(self, self.lm_model, self.qwen3_06b), token_weight_pairs["lm_prompt"], token_weight_pairs["lm_prompt_negative"], min_tokens=lm_metadata["min_tokens"], max_tokens=lm_metadata["min_tokens"], seed=lm_metadata["seed"]) + out = {"conditioning_lyrics": lyrics_embeds[:, 0]} - return base_out, None, {"conditioning_lyrics": lyrics_embeds[:, 0], "audio_codes": [audio_codes]} + lm_metadata = token_weight_pairs["lm_metadata"] + if lm_metadata["generate_audio_codes"]: + audio_codes = generate_audio_codes(getattr(self, self.lm_model, self.qwen3_06b), token_weight_pairs["lm_prompt"], token_weight_pairs["lm_prompt_negative"], min_tokens=lm_metadata["min_tokens"], max_tokens=lm_metadata["min_tokens"], seed=lm_metadata["seed"], cfg_scale=lm_metadata["cfg_scale"], temperature=lm_metadata["temperature"], top_p=lm_metadata["top_p"], top_k=lm_metadata["top_k"]) + out["audio_codes"] = [audio_codes] + + return base_out, None, out def set_clip_options(self, options): self.qwen3_06b.set_clip_options(options) diff --git a/comfy_extras/nodes_ace.py b/comfy_extras/nodes_ace.py index 376584e5c..dde5bbd2a 100644 --- a/comfy_extras/nodes_ace.py +++ b/comfy_extras/nodes_ace.py @@ -44,13 +44,18 @@ class TextEncodeAceStepAudio15(io.ComfyNode): io.Combo.Input("timesignature", options=['2', '3', '4', '6']), io.Combo.Input("language", options=["en", "ja", "zh", "es", "de", "fr", "pt", "ru", "it", "nl", "pl", "tr", "vi", "cs", "fa", "id", "ko", "uk", "hu", "ar", "sv", "ro", "el"]), io.Combo.Input("keyscale", options=[f"{root} {quality}" for quality in ["major", "minor"] for root in ["C", "C#", "Db", "D", "D#", "Eb", "E", "F", "F#", "Gb", "G", "G#", "Ab", "A", "A#", "Bb", "B"]]), + io.Boolean.Input("generate_audio_codes", default=True, tooltip="Enable the LLM that generates audio codes. This can be slow but will increase the quality of the generated audio. Turn this off if you are giving the model an audio reference.", advanced=True), + io.Float.Input("cfg_scale", default=2.0, min=0.0, max=100.0, step=0.1, advanced=True), + io.Float.Input("temperature", default=0.85, min=0.0, max=2.0, step=0.01, advanced=True), + io.Float.Input("top_p", default=0.9, min=0.0, max=2000.0, step=0.01, advanced=True), + io.Int.Input("top_k", default=0, min=0, max=100, advanced=True), ], outputs=[io.Conditioning.Output()], ) @classmethod - def execute(cls, clip, tags, lyrics, seed, bpm, duration, timesignature, language, keyscale) -> io.NodeOutput: - tokens = clip.tokenize(tags, lyrics=lyrics, bpm=bpm, duration=duration, timesignature=int(timesignature), language=language, keyscale=keyscale, seed=seed) + def execute(cls, clip, tags, lyrics, seed, bpm, duration, timesignature, language, keyscale, generate_audio_codes, cfg_scale, temperature, top_p, top_k) -> io.NodeOutput: + tokens = clip.tokenize(tags, lyrics=lyrics, bpm=bpm, duration=duration, timesignature=int(timesignature), language=language, keyscale=keyscale, seed=seed, generate_audio_codes=generate_audio_codes, cfg_scale=cfg_scale, temperature=temperature, top_p=top_p, top_k=top_k) conditioning = clip.encode_from_tokens_scheduled(tokens) return io.NodeOutput(conditioning) @@ -100,14 +105,15 @@ class EmptyAceStep15LatentAudio(io.ComfyNode): latent = torch.zeros([batch_size, 64, length], device=comfy.model_management.intermediate_device()) return io.NodeOutput({"samples": latent, "type": "audio"}) -class ReferenceTimbreAudio(io.ComfyNode): +class ReferenceAudio(io.ComfyNode): @classmethod def define_schema(cls): return io.Schema( node_id="ReferenceTimbreAudio", + display_name="Reference Audio", category="advanced/conditioning/audio", is_experimental=True, - description="This node sets the reference audio for timbre (for ace step 1.5)", + description="This node sets the reference audio for ace step 1.5", inputs=[ io.Conditioning.Input("conditioning"), io.Latent.Input("latent", optional=True), @@ -131,7 +137,7 @@ class AceExtension(ComfyExtension): EmptyAceStepLatentAudio, TextEncodeAceStepAudio15, EmptyAceStep15LatentAudio, - ReferenceTimbreAudio, + ReferenceAudio, ] async def comfy_entrypoint() -> AceExtension: From a50c32d63fe55d073edd7af2242f0536f50b362e Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 4 Feb 2026 19:15:30 -0800 Subject: [PATCH 578/647] Disable sage attention on ace step 1.5 (#12297) --- comfy/ldm/ace/ace_step15.py | 2 +- comfy/ldm/modules/attention.py | 3 +++ 2 files changed, 4 insertions(+), 1 deletion(-) diff --git a/comfy/ldm/ace/ace_step15.py b/comfy/ldm/ace/ace_step15.py index 17a37e573..f2b130bc1 100644 --- a/comfy/ldm/ace/ace_step15.py +++ b/comfy/ldm/ace/ace_step15.py @@ -183,7 +183,7 @@ class AceStepAttention(nn.Module): else: attn_bias = window_bias - attn_output = optimized_attention(query_states, key_states, value_states, self.num_heads, attn_bias, skip_reshape=True) + attn_output = optimized_attention(query_states, key_states, value_states, self.num_heads, attn_bias, skip_reshape=True, low_precision_attention=False) attn_output = self.o_proj(attn_output) return attn_output diff --git a/comfy/ldm/modules/attention.py b/comfy/ldm/modules/attention.py index ccf690945..10d051325 100644 --- a/comfy/ldm/modules/attention.py +++ b/comfy/ldm/modules/attention.py @@ -524,6 +524,9 @@ def attention_pytorch(q, k, v, heads, mask=None, attn_precision=None, skip_resha @wrap_attn def attention_sage(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False, **kwargs): + if kwargs.get("low_precision_attention", True) is False: + return attention_pytorch(q, k, v, heads, mask=mask, skip_reshape=skip_reshape, skip_output_reshape=skip_output_reshape, **kwargs) + exception_fallback = False if skip_reshape: b, _, _, dim_head = q.shape From a246cc02b274104d5f656b68ce505354c164aef8 Mon Sep 17 00:00:00 2001 From: blepping <157360029+blepping@users.noreply.github.com> Date: Wed, 4 Feb 2026 22:17:37 -0700 Subject: [PATCH 579/647] Improvements to ACE-Steps 1.5 text encoding (#12283) --- comfy/text_encoders/ace15.py | 56 +++++++++++++++++++++++++++++------- 1 file changed, 45 insertions(+), 11 deletions(-) diff --git a/comfy/text_encoders/ace15.py b/comfy/text_encoders/ace15.py index 74e62733e..00dd5ba90 100644 --- a/comfy/text_encoders/ace15.py +++ b/comfy/text_encoders/ace15.py @@ -3,6 +3,7 @@ import comfy.text_encoders.llama from comfy import sd1_clip import torch import math +import yaml import comfy.utils @@ -125,14 +126,43 @@ class ACE15Tokenizer(sd1_clip.SD1Tokenizer): def __init__(self, embedding_directory=None, tokenizer_data={}): super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, name="qwen3_06b", tokenizer=Qwen3Tokenizer) + def _metas_to_cot(self, *, return_yaml: bool = False, **kwargs) -> str: + user_metas = { + k: kwargs.pop(k) + for k in ("bpm", "duration", "keyscale", "timesignature", "language", "caption") + if k in kwargs + } + timesignature = user_metas.get("timesignature") + if isinstance(timesignature, str) and timesignature.endswith("/4"): + user_metas["timesignature"] = timesignature.rsplit("/", 1)[0] + user_metas = { + k: v if not isinstance(v, str) or not v.isdigit() else int(v) + for k, v in user_metas.items() + if v not in {"unspecified", None} + } + if len(user_metas): + meta_yaml = yaml.dump(user_metas, allow_unicode=True, sort_keys=True).strip() + else: + meta_yaml = "" + return f"\n{meta_yaml}\n" if not return_yaml else meta_yaml + + def _metas_to_cap(self, **kwargs) -> str: + use_keys = ("bpm", "duration", "keyscale", "timesignature") + user_metas = { k: kwargs.pop(k, "N/A") for k in use_keys } + duration = user_metas["duration"] + if duration == "N/A": + user_metas["duration"] = "30 seconds" + elif isinstance(duration, (str, int, float)): + user_metas["duration"] = f"{math.ceil(float(duration))} seconds" + else: + raise TypeError("Unexpected type for duration key, must be str, int or float") + return "\n".join(f"- {k}: {user_metas[k]}" for k in use_keys) + def tokenize_with_weights(self, text, return_word_ids=False, **kwargs): out = {} lyrics = kwargs.get("lyrics", "") - bpm = kwargs.get("bpm", 120) duration = kwargs.get("duration", 120) - keyscale = kwargs.get("keyscale", "C major") - timesignature = kwargs.get("timesignature", 2) - language = kwargs.get("language", "en") + language = kwargs.get("language") seed = kwargs.get("seed", 0) generate_audio_codes = kwargs.get("generate_audio_codes", True) @@ -141,16 +171,20 @@ class ACE15Tokenizer(sd1_clip.SD1Tokenizer): top_p = kwargs.get("top_p", 0.9) top_k = kwargs.get("top_k", 0.0) + duration = math.ceil(duration) - meta_lm = 'bpm: {}\nduration: {}\nkeyscale: {}\ntimesignature: {}'.format(bpm, duration, keyscale, timesignature) - lm_template = "<|im_start|>system\n# Instruction\nGenerate audio semantic tokens based on the given conditions:\n\n<|im_end|>\n<|im_start|>user\n# Caption\n{}\n{}\n<|im_end|>\n<|im_start|>assistant\n\n{}\n\n\n<|im_end|>\n" + kwargs["duration"] = duration - meta_cap = '- bpm: {}\n- timesignature: {}\n- keyscale: {}\n- duration: {}\n'.format(bpm, timesignature, keyscale, duration) - out["lm_prompt"] = self.qwen3_06b.tokenize_with_weights(lm_template.format(text, lyrics, meta_lm), disable_weights=True) - out["lm_prompt_negative"] = self.qwen3_06b.tokenize_with_weights(lm_template.format(text, lyrics, ""), disable_weights=True) + cot_text = self._metas_to_cot(caption = text, **kwargs) + meta_cap = self._metas_to_cap(**kwargs) - out["lyrics"] = self.qwen3_06b.tokenize_with_weights("# Languages\n{}\n\n# Lyric{}<|endoftext|><|endoftext|>".format(language, lyrics), return_word_ids, disable_weights=True, **kwargs) - out["qwen3_06b"] = self.qwen3_06b.tokenize_with_weights("# Instruction\nGenerate audio semantic tokens based on the given conditions:\n\n# Caption\n{}# Metas\n{}<|endoftext|>\n<|endoftext|>".format(text, meta_cap), return_word_ids, **kwargs) + lm_template = "<|im_start|>system\n# Instruction\nGenerate audio semantic tokens based on the given conditions:\n\n<|im_end|>\n<|im_start|>user\n# Caption\n{}\n# Lyric\n{}\n<|im_end|>\n<|im_start|>assistant\n{}\n<|im_end|>\n" + + out["lm_prompt"] = self.qwen3_06b.tokenize_with_weights(lm_template.format(text, lyrics, cot_text), disable_weights=True) + out["lm_prompt_negative"] = self.qwen3_06b.tokenize_with_weights(lm_template.format(text, lyrics, "\n"), disable_weights=True) + + out["lyrics"] = self.qwen3_06b.tokenize_with_weights("# Languages\n{}\n\n# Lyric\n{}<|endoftext|><|endoftext|>".format(language if language is not None else "", lyrics), return_word_ids, disable_weights=True, **kwargs) + out["qwen3_06b"] = self.qwen3_06b.tokenize_with_weights("# Instruction\nGenerate audio semantic tokens based on the given conditions:\n\n# Caption\n{}\n# Metas\n{}\n<|endoftext|>\n<|endoftext|>".format(text, meta_cap), return_word_ids, **kwargs) out["lm_metadata"] = {"min_tokens": duration * 5, "seed": seed, "generate_audio_codes": generate_audio_codes, From 35183543e004d8b7509c043e7a680bee07171622 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 4 Feb 2026 22:12:04 -0800 Subject: [PATCH 580/647] Add VAE tiled decode node for audio. (#12299) --- comfy/sd.py | 2 +- comfy_extras/nodes_audio.py | 43 +++++++++++++++++++++++++++++++------ 2 files changed, 38 insertions(+), 7 deletions(-) diff --git a/comfy/sd.py b/comfy/sd.py index bc63d6ced..bc9407405 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -976,7 +976,7 @@ class VAE: if overlap is not None: args["overlap"] = overlap - if dims == 1: + if dims == 1 or self.extra_1d_channel is not None: args.pop("tile_y") output = self.decode_tiled_1d(samples, **args) elif dims == 2: diff --git a/comfy_extras/nodes_audio.py b/comfy_extras/nodes_audio.py index bef723dce..b63dd8e97 100644 --- a/comfy_extras/nodes_audio.py +++ b/comfy_extras/nodes_audio.py @@ -94,6 +94,19 @@ class VAEEncodeAudio(IO.ComfyNode): encode = execute # TODO: remove +def vae_decode_audio(vae, samples, tile=None, overlap=None): + if tile is not None: + audio = vae.decode_tiled(samples["samples"], tile_y=tile, overlap=overlap).movedim(-1, 1) + else: + audio = vae.decode(samples["samples"]).movedim(-1, 1) + + std = torch.std(audio, dim=[1, 2], keepdim=True) * 5.0 + std[std < 1.0] = 1.0 + audio /= std + vae_sample_rate = getattr(vae, "audio_sample_rate", 44100) + return {"waveform": audio, "sample_rate": vae_sample_rate if "sample_rate" not in samples else samples["sample_rate"]} + + class VAEDecodeAudio(IO.ComfyNode): @classmethod def define_schema(cls): @@ -111,16 +124,33 @@ class VAEDecodeAudio(IO.ComfyNode): @classmethod def execute(cls, vae, samples) -> IO.NodeOutput: - audio = vae.decode(samples["samples"]).movedim(-1, 1) - std = torch.std(audio, dim=[1,2], keepdim=True) * 5.0 - std[std < 1.0] = 1.0 - audio /= std - vae_sample_rate = getattr(vae, "audio_sample_rate", 44100) - return IO.NodeOutput({"waveform": audio, "sample_rate": vae_sample_rate if "sample_rate" not in samples else samples["sample_rate"]}) + return IO.NodeOutput(vae_decode_audio(vae, samples)) decode = execute # TODO: remove +class VAEDecodeAudioTiled(IO.ComfyNode): + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="VAEDecodeAudioTiled", + search_aliases=["latent to audio"], + display_name="VAE Decode Audio (Tiled)", + category="latent/audio", + inputs=[ + IO.Latent.Input("samples"), + IO.Vae.Input("vae"), + IO.Int.Input("tile_size", default=512, min=32, max=8192, step=8), + IO.Int.Input("overlap", default=64, min=0, max=1024, step=8), + ], + outputs=[IO.Audio.Output()], + ) + + @classmethod + def execute(cls, vae, samples, tile_size, overlap) -> IO.NodeOutput: + return IO.NodeOutput(vae_decode_audio(vae, samples, tile_size, overlap)) + + class SaveAudio(IO.ComfyNode): @classmethod def define_schema(cls): @@ -675,6 +705,7 @@ class AudioExtension(ComfyExtension): EmptyLatentAudio, VAEEncodeAudio, VAEDecodeAudio, + VAEDecodeAudioTiled, SaveAudio, SaveAudioMP3, SaveAudioOpus, From cb459573c8fa025bbf9ecf312f6af376d659f567 Mon Sep 17 00:00:00 2001 From: comfyanonymous Date: Thu, 5 Feb 2026 01:13:35 -0500 Subject: [PATCH 581/647] ComfyUI v0.12.3 --- comfyui_version.py | 2 +- pyproject.toml | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/comfyui_version.py b/comfyui_version.py index 5d296cd1b..706b37763 100644 --- a/comfyui_version.py +++ b/comfyui_version.py @@ -1,3 +1,3 @@ # This file is automatically generated by the build process when version is # updated in pyproject.toml. -__version__ = "0.12.2" +__version__ = "0.12.3" diff --git a/pyproject.toml b/pyproject.toml index 1ddcc3596..f7925b92a 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -1,6 +1,6 @@ [project] name = "ComfyUI" -version = "0.12.2" +version = "0.12.3" readme = "README.md" license = { file = "LICENSE" } requires-python = ">=3.10" From 00efcc6cd028206ad81a90dec177c9a470a20a2a Mon Sep 17 00:00:00 2001 From: Comfy Org PR Bot Date: Thu, 5 Feb 2026 15:17:37 +0900 Subject: [PATCH 582/647] Bump comfyui-frontend-package to 1.38.13 (#12238) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 0c401873a..41cc9174b 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,4 +1,4 @@ -comfyui-frontend-package==1.37.11 +comfyui-frontend-package==1.38.13 comfyui-workflow-templates==0.8.31 comfyui-embedded-docs==0.4.0 torch From 2b70ab9ad0fd6a38b11546a18c546ce40cc176a1 Mon Sep 17 00:00:00 2001 From: AustinMroz Date: Wed, 4 Feb 2026 22:18:21 -0800 Subject: [PATCH 583/647] Add a Create List node (#12173) --- comfy_extras/nodes_toolkit.py | 47 +++++++++++++++++++++++++++++++++++ nodes.py | 3 ++- 2 files changed, 49 insertions(+), 1 deletion(-) create mode 100644 comfy_extras/nodes_toolkit.py diff --git a/comfy_extras/nodes_toolkit.py b/comfy_extras/nodes_toolkit.py new file mode 100644 index 000000000..71faf7226 --- /dev/null +++ b/comfy_extras/nodes_toolkit.py @@ -0,0 +1,47 @@ +from __future__ import annotations +from typing_extensions import override +from comfy_api.latest import ComfyExtension, io + + +class CreateList(io.ComfyNode): + @classmethod + def define_schema(cls): + template_matchtype = io.MatchType.Template("type") + template_autogrow = io.Autogrow.TemplatePrefix( + input=io.MatchType.Input("input", template=template_matchtype), + prefix="input", + ) + return io.Schema( + node_id="CreateList", + display_name="Create List", + category="logic", + is_input_list=True, + search_aliases=["Image Iterator", "Text Iterator", "Iterator"], + inputs=[io.Autogrow.Input("inputs", template=template_autogrow)], + outputs=[ + io.MatchType.Output( + template=template_matchtype, + is_output_list=True, + display_name="list", + ), + ], + ) + + @classmethod + def execute(cls, inputs: io.Autogrow.Type) -> io.NodeOutput: + output_list = [] + for input in inputs.values(): + output_list += input + return io.NodeOutput(output_list) + + +class ToolkitExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[io.ComfyNode]]: + return [ + CreateList, + ] + + +async def comfy_entrypoint() -> ToolkitExtension: + return ToolkitExtension() diff --git a/nodes.py b/nodes.py index e11a8ed80..91de7a9d7 100644 --- a/nodes.py +++ b/nodes.py @@ -2433,7 +2433,8 @@ async def init_builtin_extra_nodes(): "nodes_image_compare.py", "nodes_zimage.py", "nodes_lora_debug.py", - "nodes_color.py" + "nodes_color.py", + "nodes_toolkit.py", ] import_failed = [] From 6555dc65b82c5f072dcad87f0dbccb4fc5f85e6b Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Thu, 5 Feb 2026 13:43:45 -0800 Subject: [PATCH 584/647] Make ace step 1.5 work without the llm. (#12311) --- comfy/ldm/ace/ace_step15.py | 72 +++++++++++++++++++++++++++++++++---- comfy/model_base.py | 17 +++------ 2 files changed, 70 insertions(+), 19 deletions(-) diff --git a/comfy/ldm/ace/ace_step15.py b/comfy/ldm/ace/ace_step15.py index f2b130bc1..7fc7f1e8e 100644 --- a/comfy/ldm/ace/ace_step15.py +++ b/comfy/ldm/ace/ace_step15.py @@ -7,6 +7,67 @@ from comfy.ldm.modules.attention import optimized_attention import comfy.model_management from comfy.ldm.flux.layers import timestep_embedding +def get_silence_latent(length, device): + head = torch.tensor([[[ 0.5707, 0.0982, 0.6909, -0.5658, 0.6266, 0.6996, -0.1365, -0.1291, + -0.0776, -0.1171, -0.2743, -0.8422, -0.1168, 1.5539, -4.6936, 0.7436, + -1.1846, -0.2637, 0.6933, -6.7266, 0.0966, -0.1187, -0.3501, -1.1736, + 0.0587, -2.0517, -1.3651, 0.7508, -0.2490, -1.3548, -0.1290, -0.7261, + 1.1132, -0.3249, 0.2337, 0.3004, 0.6605, -0.0298, -0.1989, -0.4041, + 0.2843, -1.0963, -0.5519, 0.2639, -1.0436, -0.1183, 0.0640, 0.4460, + -1.1001, -0.6172, -1.3241, 1.1379, 0.5623, -0.1507, -0.1963, -0.4742, + -2.4697, 0.5302, 0.5381, 0.4636, -0.1782, -0.0687, 1.0333, 0.4202], + [ 0.3040, -0.1367, 0.6200, 0.0665, -0.0642, 0.4655, -0.1187, -0.0440, + 0.2941, -0.2753, 0.0173, -0.2421, -0.0147, 1.5603, -2.7025, 0.7907, + -0.9736, -0.0682, 0.1294, -5.0707, -0.2167, 0.3302, -0.1513, -0.8100, + -0.3894, -0.2884, -0.3149, 0.8660, -0.3817, -1.7061, 0.5824, -0.4840, + 0.6938, 0.1859, 0.1753, 0.3081, 0.0195, 0.1403, -0.0754, -0.2091, + 0.1251, -0.1578, -0.4968, -0.1052, -0.4554, -0.0320, 0.1284, 0.4974, + -1.1889, -0.0344, -0.8313, 0.2953, 0.5445, -0.6249, -0.1595, -0.0682, + -3.1412, 0.0484, 0.4153, 0.8260, -0.1526, -0.0625, 0.5366, 0.8473], + [ 5.3524e-02, -1.7534e-01, 5.4443e-01, -4.3501e-01, -2.1317e-03, + 3.7200e-01, -4.0143e-03, -1.5516e-01, -1.2968e-01, -1.5375e-01, + -7.7107e-02, -2.0593e-01, -3.2780e-01, 1.5142e+00, -2.6101e+00, + 5.8698e-01, -1.2716e+00, -2.4773e-01, -2.7933e-02, -5.0799e+00, + 1.1601e-01, 4.0987e-01, -2.2030e-02, -6.6495e-01, -2.0995e-01, + -6.3474e-01, -1.5893e-01, 8.2745e-01, -2.2992e-01, -1.6816e+00, + 5.4440e-01, -4.9579e-01, 5.5128e-01, 3.0477e-01, 8.3052e-02, + -6.1782e-02, 5.9036e-03, 2.9553e-01, -8.0645e-02, -1.0060e-01, + 1.9144e-01, -3.8124e-01, -7.2949e-01, 2.4520e-02, -5.0814e-01, + 2.3977e-01, 9.2943e-02, 3.9256e-01, -1.1993e+00, -3.2752e-01, + -7.2707e-01, 2.9476e-01, 4.3542e-01, -8.8597e-01, -4.1686e-01, + -8.5390e-02, -2.9018e+00, 6.4988e-02, 5.3945e-01, 9.1988e-01, + 5.8762e-02, -7.0098e-02, 6.4772e-01, 8.9118e-01], + [-3.2225e-02, -1.3195e-01, 5.6411e-01, -5.4766e-01, -5.2170e-03, + 3.1425e-01, -5.4367e-02, -1.9419e-01, -1.3059e-01, -1.3660e-01, + -9.0984e-02, -1.9540e-01, -2.5590e-01, 1.5440e+00, -2.6349e+00, + 6.8273e-01, -1.2532e+00, -1.9810e-01, -2.2793e-02, -5.0506e+00, + 1.8818e-01, 5.0109e-01, 7.3546e-03, -6.8771e-01, -3.0676e-01, + -7.3257e-01, -1.6687e-01, 9.2232e-01, -1.8987e-01, -1.7267e+00, + 5.3355e-01, -5.3179e-01, 4.4953e-01, 2.8820e-01, 1.3012e-01, + -2.0943e-01, -1.1348e-01, 3.3929e-01, -1.5069e-01, -1.2919e-01, + 1.8929e-01, -3.6166e-01, -8.0756e-01, 6.6387e-02, -5.8867e-01, + 1.6978e-01, 1.0134e-01, 3.3877e-01, -1.2133e+00, -3.2492e-01, + -8.1237e-01, 3.8101e-01, 4.3765e-01, -8.0596e-01, -4.4531e-01, + -4.7513e-02, -2.9266e+00, 1.1741e-03, 4.5123e-01, 9.3075e-01, + 5.3688e-02, -1.9621e-01, 6.4530e-01, 9.3870e-01]]], device=device).movedim(-1, 1) + + silence_latent = torch.tensor([[[-1.3672e-01, -1.5820e-01, 5.8594e-01, -5.7422e-01, 3.0273e-02, + 2.7930e-01, -2.5940e-03, -2.0703e-01, -1.6113e-01, -1.4746e-01, + -2.7710e-02, -1.8066e-01, -2.9688e-01, 1.6016e+00, -2.6719e+00, + 7.7734e-01, -1.3516e+00, -1.9434e-01, -7.1289e-02, -5.0938e+00, + 2.4316e-01, 4.7266e-01, 4.6387e-02, -6.6406e-01, -2.1973e-01, + -6.7578e-01, -1.5723e-01, 9.5312e-01, -2.0020e-01, -1.7109e+00, + 5.8984e-01, -5.7422e-01, 5.1562e-01, 2.8320e-01, 1.4551e-01, + -1.8750e-01, -5.9814e-02, 3.6719e-01, -1.0059e-01, -1.5723e-01, + 2.0605e-01, -4.3359e-01, -8.2812e-01, 4.5654e-02, -6.6016e-01, + 1.4844e-01, 9.4727e-02, 3.8477e-01, -1.2578e+00, -3.3203e-01, + -8.5547e-01, 4.3359e-01, 4.2383e-01, -8.9453e-01, -5.0391e-01, + -5.6152e-02, -2.9219e+00, -2.4658e-02, 5.0391e-01, 9.8438e-01, + 7.2754e-02, -2.1582e-01, 6.3672e-01, 1.0000e+00]]], device=device).movedim(-1, 1).repeat(1, 1, length) + silence_latent[:, :, :head.shape[-1]] = head + return silence_latent + + def get_layer_class(operations, layer_name): if operations is not None and hasattr(operations, layer_name): return getattr(operations, layer_name) @@ -1040,22 +1101,21 @@ class AceStepConditionGenerationModel(nn.Module): lm_hints = self.detokenizer(lm_hints_5Hz) lm_hints = lm_hints[:, :src_latents.shape[1], :] - if is_covers is None: + if is_covers is None or is_covers is True: src_latents = lm_hints - else: - src_latents = torch.where(is_covers.unsqueeze(-1).unsqueeze(-1) > 0, lm_hints, src_latents) + elif is_covers is False: + src_latents = refer_audio_acoustic_hidden_states_packed context_latents = torch.cat([src_latents, chunk_masks.to(src_latents.dtype)], dim=-1) return encoder_hidden, encoder_mask, context_latents - def forward(self, x, timestep, context, lyric_embed=None, refer_audio=None, audio_codes=None, **kwargs): + def forward(self, x, timestep, context, lyric_embed=None, refer_audio=None, audio_codes=None, is_covers=None, **kwargs): text_attention_mask = None lyric_attention_mask = None refer_audio_order_mask = None attention_mask = None chunk_masks = None - is_covers = None src_latents = None precomputed_lm_hints_25Hz = None lyric_hidden_states = lyric_embed @@ -1067,7 +1127,7 @@ class AceStepConditionGenerationModel(nn.Module): if refer_audio_order_mask is None: refer_audio_order_mask = torch.zeros((x.shape[0],), device=x.device, dtype=torch.long) - if src_latents is None and is_covers is None: + if src_latents is None: src_latents = x if chunk_masks is None: diff --git a/comfy/model_base.py b/comfy/model_base.py index a2a34f191..dcbf12074 100644 --- a/comfy/model_base.py +++ b/comfy/model_base.py @@ -1560,22 +1560,11 @@ class ACEStep15(BaseModel): refer_audio = kwargs.get("reference_audio_timbre_latents", None) if refer_audio is None or len(refer_audio) == 0: - refer_audio = torch.tensor([[[-1.3672e-01, -1.5820e-01, 5.8594e-01, -5.7422e-01, 3.0273e-02, - 2.7930e-01, -2.5940e-03, -2.0703e-01, -1.6113e-01, -1.4746e-01, - -2.7710e-02, -1.8066e-01, -2.9688e-01, 1.6016e+00, -2.6719e+00, - 7.7734e-01, -1.3516e+00, -1.9434e-01, -7.1289e-02, -5.0938e+00, - 2.4316e-01, 4.7266e-01, 4.6387e-02, -6.6406e-01, -2.1973e-01, - -6.7578e-01, -1.5723e-01, 9.5312e-01, -2.0020e-01, -1.7109e+00, - 5.8984e-01, -5.7422e-01, 5.1562e-01, 2.8320e-01, 1.4551e-01, - -1.8750e-01, -5.9814e-02, 3.6719e-01, -1.0059e-01, -1.5723e-01, - 2.0605e-01, -4.3359e-01, -8.2812e-01, 4.5654e-02, -6.6016e-01, - 1.4844e-01, 9.4727e-02, 3.8477e-01, -1.2578e+00, -3.3203e-01, - -8.5547e-01, 4.3359e-01, 4.2383e-01, -8.9453e-01, -5.0391e-01, - -5.6152e-02, -2.9219e+00, -2.4658e-02, 5.0391e-01, 9.8438e-01, - 7.2754e-02, -2.1582e-01, 6.3672e-01, 1.0000e+00]]], device=device).movedim(-1, 1).repeat(1, 1, noise.shape[2]) + refer_audio = comfy.ldm.ace.ace_step15.get_silence_latent(noise.shape[2], device) pass_audio_codes = True else: refer_audio = refer_audio[-1][:, :, :noise.shape[2]] + out['is_covers'] = comfy.conds.CONDConstant(True) pass_audio_codes = False if pass_audio_codes: @@ -1583,6 +1572,8 @@ class ACEStep15(BaseModel): if audio_codes is not None: out['audio_codes'] = comfy.conds.CONDRegular(torch.tensor(audio_codes, device=device)) refer_audio = refer_audio[:, :, :750] + else: + out['is_covers'] = comfy.conds.CONDConstant(False) out['refer_audio'] = comfy.conds.CONDRegular(refer_audio) return out From 458292fef0077470f5675ba52555e7bb4c28102e Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Thu, 5 Feb 2026 16:15:04 -0800 Subject: [PATCH 585/647] Fix some lowvram stuff with ace step 1.5 (#12312) --- comfy/ldm/ace/ace_step15.py | 12 ++++++------ 1 file changed, 6 insertions(+), 6 deletions(-) diff --git a/comfy/ldm/ace/ace_step15.py b/comfy/ldm/ace/ace_step15.py index 7fc7f1e8e..69338336d 100644 --- a/comfy/ldm/ace/ace_step15.py +++ b/comfy/ldm/ace/ace_step15.py @@ -738,7 +738,7 @@ class AttentionPooler(nn.Module): def forward(self, x): B, T, P, D = x.shape x = self.embed_tokens(x) - special = self.special_token.expand(B, T, 1, -1) + special = comfy.model_management.cast_to(self.special_token, device=x.device, dtype=x.dtype).expand(B, T, 1, -1) x = torch.cat([special, x], dim=2) x = x.view(B * T, P + 1, D) @@ -789,7 +789,7 @@ class FSQ(nn.Module): self.register_buffer('implicit_codebook', implicit_codebook, persistent=False) def bound(self, z): - levels_minus_1 = (self._levels - 1).to(z.dtype) + levels_minus_1 = (comfy.model_management.cast_to(self._levels, device=z.device, dtype=z.dtype) - 1) scale = 2. / levels_minus_1 bracket = (levels_minus_1 * (torch.tanh(z) + 1) / 2.) + 0.5 @@ -804,8 +804,8 @@ class FSQ(nn.Module): return codes_non_centered.float() * (2. / (self._levels.float() - 1)) - 1. def codes_to_indices(self, zhat): - zhat_normalized = (zhat + 1.) / (2. / (self._levels.to(zhat.dtype) - 1)) - return (zhat_normalized * self._basis.to(zhat.dtype)).sum(dim=-1).round().to(torch.int32) + zhat_normalized = (zhat + 1.) / (2. / (comfy.model_management.cast_to(self._levels, device=zhat.device, dtype=zhat.dtype) - 1)) + return (zhat_normalized * comfy.model_management.cast_to(self._basis, device=zhat.device, dtype=zhat.dtype)).sum(dim=-1).round().to(torch.int32) def forward(self, z): orig_dtype = z.dtype @@ -887,7 +887,7 @@ class ResidualFSQ(nn.Module): x = self.project_in(x) if hasattr(self, 'soft_clamp_input_value'): - sc_val = self.soft_clamp_input_value.to(x.dtype) + sc_val = comfy.model_management.cast_to(self.soft_clamp_input_value, device=x.device, dtype=x.dtype) x = (x / sc_val).tanh() * sc_val quantized_out = torch.tensor(0., device=x.device, dtype=x.dtype) @@ -895,7 +895,7 @@ class ResidualFSQ(nn.Module): all_indices = [] for layer, scale in zip(self.layers, self.scales): - scale = scale.to(residual.dtype) + scale = comfy.model_management.cast_to(scale, device=x.device, dtype=x.dtype) quantized, indices = layer(residual / scale) quantized = quantized * scale From c2d7f07dbf312ef9034c65102f1a45c4a3355c1a Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Thu, 5 Feb 2026 16:24:09 -0800 Subject: [PATCH 586/647] Fix issue when using disable_unet_model_creation (#12315) --- comfy/model_base.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/comfy/model_base.py b/comfy/model_base.py index dcbf12074..3bb54f59e 100644 --- a/comfy/model_base.py +++ b/comfy/model_base.py @@ -147,11 +147,11 @@ class BaseModel(torch.nn.Module): self.diffusion_model.to(memory_format=torch.channels_last) logging.debug("using channels last mode for diffusion model") logging.info("model weight dtype {}, manual cast: {}".format(self.get_dtype(), self.manual_cast_dtype)) + comfy.model_management.archive_model_dtypes(self.diffusion_model) + self.model_type = model_type self.model_sampling = model_sampling(model_config, model_type) - comfy.model_management.archive_model_dtypes(self.diffusion_model) - self.adm_channels = unet_config.get("adm_in_channels", None) if self.adm_channels is None: self.adm_channels = 0 From a1c101f861681ff18df5bdb0605e63c1ba9e8a96 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Jukka=20Sepp=C3=A4nen?= <40791699+kijai@users.noreply.github.com> Date: Fri, 6 Feb 2026 07:43:09 +0200 Subject: [PATCH 587/647] EasyCache: Support LTX2 (#12231) --- comfy_extras/nodes_easycache.py | 50 ++++++++++++++++++++++++--------- 1 file changed, 37 insertions(+), 13 deletions(-) diff --git a/comfy_extras/nodes_easycache.py b/comfy_extras/nodes_easycache.py index 90d730df6..51d1e5b9c 100644 --- a/comfy_extras/nodes_easycache.py +++ b/comfy_extras/nodes_easycache.py @@ -9,6 +9,14 @@ if TYPE_CHECKING: from uuid import UUID +def _extract_tensor(data, output_channels): + """Extract tensor from data, handling both single tensors and lists.""" + if isinstance(data, list): + # LTX2 AV tensors: [video, audio] + return data[0][:, :output_channels], data[1][:, :output_channels] + return data[:, :output_channels], None + + def easycache_forward_wrapper(executor, *args, **kwargs): # get values from args transformer_options: dict[str] = args[-1] @@ -17,7 +25,7 @@ def easycache_forward_wrapper(executor, *args, **kwargs): if not transformer_options: transformer_options = args[-2] easycache: EasyCacheHolder = transformer_options["easycache"] - x: torch.Tensor = args[0][:, :easycache.output_channels] + x, ax = _extract_tensor(args[0], easycache.output_channels) sigmas = transformer_options["sigmas"] uuids = transformer_options["uuids"] if sigmas is not None and easycache.is_past_end_timestep(sigmas): @@ -35,7 +43,11 @@ def easycache_forward_wrapper(executor, *args, **kwargs): if easycache.skip_current_step and can_apply_cache_diff: if easycache.verbose: logging.info(f"EasyCache [verbose] - was marked to skip this step by {easycache.first_cond_uuid}. Present uuids: {uuids}") - return easycache.apply_cache_diff(x, uuids) + result = easycache.apply_cache_diff(x, uuids) + if ax is not None: + result_audio = easycache.apply_cache_diff(ax, uuids, is_audio=True) + return [result, result_audio] + return result if easycache.initial_step: easycache.first_cond_uuid = uuids[0] has_first_cond_uuid = easycache.has_first_cond_uuid(uuids) @@ -51,13 +63,18 @@ def easycache_forward_wrapper(executor, *args, **kwargs): logging.info(f"EasyCache [verbose] - skipping step; cumulative_change_rate: {easycache.cumulative_change_rate}, reuse_threshold: {easycache.reuse_threshold}") # other conds should also skip this step, and instead use their cached values easycache.skip_current_step = True - return easycache.apply_cache_diff(x, uuids) + result = easycache.apply_cache_diff(x, uuids) + if ax is not None: + result_audio = easycache.apply_cache_diff(ax, uuids, is_audio=True) + return [result, result_audio] + return result else: if easycache.verbose: logging.info(f"EasyCache [verbose] - NOT skipping step; cumulative_change_rate: {easycache.cumulative_change_rate}, reuse_threshold: {easycache.reuse_threshold}") easycache.cumulative_change_rate = 0.0 - output: torch.Tensor = executor(*args, **kwargs) + full_output: torch.Tensor = executor(*args, **kwargs) + output, audio_output = _extract_tensor(full_output, easycache.output_channels) if has_first_cond_uuid and easycache.has_output_prev_norm(): output_change = (easycache.subsample(output, uuids, clone=False) - easycache.output_prev_subsampled).flatten().abs().mean() if easycache.verbose: @@ -74,13 +91,15 @@ def easycache_forward_wrapper(executor, *args, **kwargs): logging.info(f"EasyCache [verbose] - output_change_rate: {output_change_rate}") # TODO: allow cache_diff to be offloaded easycache.update_cache_diff(output, next_x_prev, uuids) + if audio_output is not None: + easycache.update_cache_diff(audio_output, ax, uuids, is_audio=True) if has_first_cond_uuid: easycache.x_prev_subsampled = easycache.subsample(next_x_prev, uuids) easycache.output_prev_subsampled = easycache.subsample(output, uuids) easycache.output_prev_norm = output.flatten().abs().mean() if easycache.verbose: logging.info(f"EasyCache [verbose] - x_prev_subsampled: {easycache.x_prev_subsampled.shape}") - return output + return full_output def lazycache_predict_noise_wrapper(executor, *args, **kwargs): # get values from args @@ -89,8 +108,8 @@ def lazycache_predict_noise_wrapper(executor, *args, **kwargs): easycache: LazyCacheHolder = model_options["transformer_options"]["easycache"] if easycache.is_past_end_timestep(timestep): return executor(*args, **kwargs) + x: torch.Tensor = _extract_tensor(args[0], easycache.output_channels) # prepare next x_prev - x: torch.Tensor = args[0][:, :easycache.output_channels] next_x_prev = x input_change = None do_easycache = easycache.should_do_easycache(timestep) @@ -197,6 +216,7 @@ class EasyCacheHolder: self.output_prev_subsampled: torch.Tensor = None self.output_prev_norm: torch.Tensor = None self.uuid_cache_diffs: dict[UUID, torch.Tensor] = {} + self.uuid_cache_diffs_audio: dict[UUID, torch.Tensor] = {} self.output_change_rates = [] self.approx_output_change_rates = [] self.total_steps_skipped = 0 @@ -245,20 +265,21 @@ class EasyCacheHolder: def can_apply_cache_diff(self, uuids: list[UUID]) -> bool: return all(uuid in self.uuid_cache_diffs for uuid in uuids) - def apply_cache_diff(self, x: torch.Tensor, uuids: list[UUID]): - if self.first_cond_uuid in uuids: + def apply_cache_diff(self, x: torch.Tensor, uuids: list[UUID], is_audio: bool = False): + if self.first_cond_uuid in uuids and not is_audio: self.total_steps_skipped += 1 + cache_diffs = self.uuid_cache_diffs_audio if is_audio else self.uuid_cache_diffs batch_offset = x.shape[0] // len(uuids) for i, uuid in enumerate(uuids): # slice out only what is relevant to this cond batch_slice = [slice(i*batch_offset,(i+1)*batch_offset)] # if cached dims don't match x dims, cut off excess and hope for the best (cosmos world2video) - if x.shape[1:] != self.uuid_cache_diffs[uuid].shape[1:]: + if x.shape[1:] != cache_diffs[uuid].shape[1:]: if not self.allow_mismatch: raise ValueError(f"Cached dims {self.uuid_cache_diffs[uuid].shape} don't match x dims {x.shape} - this is no good") slicing = [] skip_this_dim = True - for dim_u, dim_x in zip(self.uuid_cache_diffs[uuid].shape, x.shape): + for dim_u, dim_x in zip(cache_diffs[uuid].shape, x.shape): if skip_this_dim: skip_this_dim = False continue @@ -270,10 +291,11 @@ class EasyCacheHolder: else: slicing.append(slice(None)) batch_slice = batch_slice + slicing - x[tuple(batch_slice)] += self.uuid_cache_diffs[uuid].to(x.device) + x[tuple(batch_slice)] += cache_diffs[uuid].to(x.device) return x - def update_cache_diff(self, output: torch.Tensor, x: torch.Tensor, uuids: list[UUID]): + def update_cache_diff(self, output: torch.Tensor, x: torch.Tensor, uuids: list[UUID], is_audio: bool = False): + cache_diffs = self.uuid_cache_diffs_audio if is_audio else self.uuid_cache_diffs # if output dims don't match x dims, cut off excess and hope for the best (cosmos world2video) if output.shape[1:] != x.shape[1:]: if not self.allow_mismatch: @@ -293,7 +315,7 @@ class EasyCacheHolder: diff = output - x batch_offset = diff.shape[0] // len(uuids) for i, uuid in enumerate(uuids): - self.uuid_cache_diffs[uuid] = diff[i*batch_offset:(i+1)*batch_offset, ...] + cache_diffs[uuid] = diff[i*batch_offset:(i+1)*batch_offset, ...] def has_first_cond_uuid(self, uuids: list[UUID]) -> bool: return self.first_cond_uuid in uuids @@ -324,6 +346,8 @@ class EasyCacheHolder: self.output_prev_norm = None del self.uuid_cache_diffs self.uuid_cache_diffs = {} + del self.uuid_cache_diffs_audio + self.uuid_cache_diffs_audio = {} self.total_steps_skipped = 0 self.state_metadata = None return self From eba6c940fd04483fedec6b47bb93fa669e77fe8a Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Fri, 6 Feb 2026 16:14:56 -0800 Subject: [PATCH 588/647] Make ace step 1.5 base model work properly with default workflow. (#12337) --- comfy/ldm/ace/ace_step15.py | 5 ++++- comfy/model_base.py | 2 ++ 2 files changed, 6 insertions(+), 1 deletion(-) diff --git a/comfy/ldm/ace/ace_step15.py b/comfy/ldm/ace/ace_step15.py index 69338336d..1d7dc59a8 100644 --- a/comfy/ldm/ace/ace_step15.py +++ b/comfy/ldm/ace/ace_step15.py @@ -1110,7 +1110,7 @@ class AceStepConditionGenerationModel(nn.Module): return encoder_hidden, encoder_mask, context_latents - def forward(self, x, timestep, context, lyric_embed=None, refer_audio=None, audio_codes=None, is_covers=None, **kwargs): + def forward(self, x, timestep, context, lyric_embed=None, refer_audio=None, audio_codes=None, is_covers=None, replace_with_null_embeds=False, **kwargs): text_attention_mask = None lyric_attention_mask = None refer_audio_order_mask = None @@ -1140,6 +1140,9 @@ class AceStepConditionGenerationModel(nn.Module): src_latents, chunk_masks, is_covers, precomputed_lm_hints_25Hz=precomputed_lm_hints_25Hz, audio_codes=audio_codes ) + if replace_with_null_embeds: + enc_hidden[:] = self.null_condition_emb.to(enc_hidden) + out = self.decoder(hidden_states=x, timestep=timestep, timestep_r=timestep, diff --git a/comfy/model_base.py b/comfy/model_base.py index 3bb54f59e..3aa345254 100644 --- a/comfy/model_base.py +++ b/comfy/model_base.py @@ -1552,6 +1552,8 @@ class ACEStep15(BaseModel): cross_attn = kwargs.get("cross_attn", None) if cross_attn is not None: + if torch.count_nonzero(cross_attn) == 0: + out['replace_with_null_embeds'] = comfy.conds.CONDConstant(True) out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn) conditioning_lyrics = kwargs.get("conditioning_lyrics", None) From a831c19b703693f561e32780248514eeaa9e832e Mon Sep 17 00:00:00 2001 From: asagi4 <130366179+asagi4@users.noreply.github.com> Date: Sat, 7 Feb 2026 02:38:04 +0200 Subject: [PATCH 589/647] Fix return_word_ids=True with Anima tokenizer (#12328) --- comfy/text_encoders/anima.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy/text_encoders/anima.py b/comfy/text_encoders/anima.py index b6f58cb25..fcba097cb 100644 --- a/comfy/text_encoders/anima.py +++ b/comfy/text_encoders/anima.py @@ -23,7 +23,7 @@ class AnimaTokenizer: def tokenize_with_weights(self, text:str, return_word_ids=False, **kwargs): out = {} qwen_ids = self.qwen3_06b.tokenize_with_weights(text, return_word_ids, **kwargs) - out["qwen3_06b"] = [[(token, 1.0) for token, _ in inner_list] for inner_list in qwen_ids] # Set weights to 1.0 + out["qwen3_06b"] = [[(token, 1.0, id) if return_word_ids else (token, 1.0) for token, _, id in inner_list] for inner_list in qwen_ids] # Set weights to 1.0 out["t5xxl"] = self.t5xxl.tokenize_with_weights(text, return_word_ids, **kwargs) return out From 204e65b8dcb2db2014f40e1b4c8def3a00150cde Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Fri, 6 Feb 2026 16:48:20 -0800 Subject: [PATCH 590/647] Fix bug with last pr (#12338) --- comfy/text_encoders/anima.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy/text_encoders/anima.py b/comfy/text_encoders/anima.py index fcba097cb..d8c5a6f92 100644 --- a/comfy/text_encoders/anima.py +++ b/comfy/text_encoders/anima.py @@ -23,7 +23,7 @@ class AnimaTokenizer: def tokenize_with_weights(self, text:str, return_word_ids=False, **kwargs): out = {} qwen_ids = self.qwen3_06b.tokenize_with_weights(text, return_word_ids, **kwargs) - out["qwen3_06b"] = [[(token, 1.0, id) if return_word_ids else (token, 1.0) for token, _, id in inner_list] for inner_list in qwen_ids] # Set weights to 1.0 + out["qwen3_06b"] = [[(k[0], 1.0, k[2]) if return_word_ids else (k[0], 1.0) for k in inner_list] for inner_list in qwen_ids] # Set weights to 1.0 out["t5xxl"] = self.t5xxl.tokenize_with_weights(text, return_word_ids, **kwargs) return out From 6a263288427a9998086603db0e7078ebcb56f0c4 Mon Sep 17 00:00:00 2001 From: tdrussell Date: Fri, 6 Feb 2026 19:12:15 -0600 Subject: [PATCH 591/647] Support fp16 for Cosmos-Predict2 and Anima (#12249) --- comfy/ldm/cosmos/predict2.py | 24 +++++++++++++++++------- comfy/supported_models.py | 4 ++-- 2 files changed, 19 insertions(+), 9 deletions(-) diff --git a/comfy/ldm/cosmos/predict2.py b/comfy/ldm/cosmos/predict2.py index c270e6333..6491e486b 100644 --- a/comfy/ldm/cosmos/predict2.py +++ b/comfy/ldm/cosmos/predict2.py @@ -335,7 +335,7 @@ class FinalLayer(nn.Module): device=None, dtype=None, operations=None ): super().__init__() - self.layer_norm = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6) + self.layer_norm = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6) self.linear = operations.Linear( hidden_size, spatial_patch_size * spatial_patch_size * temporal_patch_size * out_channels, bias=False, device=device, dtype=dtype ) @@ -463,6 +463,8 @@ class Block(nn.Module): extra_per_block_pos_emb: Optional[torch.Tensor] = None, transformer_options: Optional[dict] = {}, ) -> torch.Tensor: + residual_dtype = x_B_T_H_W_D.dtype + compute_dtype = emb_B_T_D.dtype if extra_per_block_pos_emb is not None: x_B_T_H_W_D = x_B_T_H_W_D + extra_per_block_pos_emb @@ -512,7 +514,7 @@ class Block(nn.Module): result_B_T_H_W_D = rearrange( self.self_attn( # normalized_x_B_T_HW_D, - rearrange(normalized_x_B_T_H_W_D, "b t h w d -> b (t h w) d"), + rearrange(normalized_x_B_T_H_W_D.to(compute_dtype), "b t h w d -> b (t h w) d"), None, rope_emb=rope_emb_L_1_1_D, transformer_options=transformer_options, @@ -522,7 +524,7 @@ class Block(nn.Module): h=H, w=W, ) - x_B_T_H_W_D = x_B_T_H_W_D + gate_self_attn_B_T_1_1_D * result_B_T_H_W_D + x_B_T_H_W_D = x_B_T_H_W_D + gate_self_attn_B_T_1_1_D.to(residual_dtype) * result_B_T_H_W_D.to(residual_dtype) def _x_fn( _x_B_T_H_W_D: torch.Tensor, @@ -536,7 +538,7 @@ class Block(nn.Module): ) _result_B_T_H_W_D = rearrange( self.cross_attn( - rearrange(_normalized_x_B_T_H_W_D, "b t h w d -> b (t h w) d"), + rearrange(_normalized_x_B_T_H_W_D.to(compute_dtype), "b t h w d -> b (t h w) d"), crossattn_emb, rope_emb=rope_emb_L_1_1_D, transformer_options=transformer_options, @@ -555,7 +557,7 @@ class Block(nn.Module): shift_cross_attn_B_T_1_1_D, transformer_options=transformer_options, ) - x_B_T_H_W_D = result_B_T_H_W_D * gate_cross_attn_B_T_1_1_D + x_B_T_H_W_D + x_B_T_H_W_D = result_B_T_H_W_D.to(residual_dtype) * gate_cross_attn_B_T_1_1_D.to(residual_dtype) + x_B_T_H_W_D normalized_x_B_T_H_W_D = _fn( x_B_T_H_W_D, @@ -563,8 +565,8 @@ class Block(nn.Module): scale_mlp_B_T_1_1_D, shift_mlp_B_T_1_1_D, ) - result_B_T_H_W_D = self.mlp(normalized_x_B_T_H_W_D) - x_B_T_H_W_D = x_B_T_H_W_D + gate_mlp_B_T_1_1_D * result_B_T_H_W_D + result_B_T_H_W_D = self.mlp(normalized_x_B_T_H_W_D.to(compute_dtype)) + x_B_T_H_W_D = x_B_T_H_W_D + gate_mlp_B_T_1_1_D.to(residual_dtype) * result_B_T_H_W_D.to(residual_dtype) return x_B_T_H_W_D @@ -876,6 +878,14 @@ class MiniTrainDIT(nn.Module): "extra_per_block_pos_emb": extra_pos_emb_B_T_H_W_D_or_T_H_W_B_D, "transformer_options": kwargs.get("transformer_options", {}), } + + # The residual stream for this model has large values. To make fp16 compute_dtype work, we keep the residual stream + # in fp32, but run attention and MLP modules in fp16. + # An alternate method that clamps fp16 values "works" in the sense that it makes coherent images, but there is noticeable + # quality degradation and visual artifacts. + if x_B_T_H_W_D.dtype == torch.float16: + x_B_T_H_W_D = x_B_T_H_W_D.float() + for block in self.blocks: x_B_T_H_W_D = block( x_B_T_H_W_D, diff --git a/comfy/supported_models.py b/comfy/supported_models.py index 77264ed28..56a21b0ef 100644 --- a/comfy/supported_models.py +++ b/comfy/supported_models.py @@ -993,7 +993,7 @@ class CosmosT2IPredict2(supported_models_base.BASE): memory_usage_factor = 1.0 - supported_inference_dtypes = [torch.bfloat16, torch.float32] + supported_inference_dtypes = [torch.bfloat16, torch.float16, torch.float32] def __init__(self, unet_config): super().__init__(unet_config) @@ -1023,7 +1023,7 @@ class Anima(supported_models_base.BASE): memory_usage_factor = 1.0 - supported_inference_dtypes = [torch.bfloat16, torch.float32] + supported_inference_dtypes = [torch.bfloat16, torch.float16, torch.float32] def __init__(self, unet_config): super().__init__(unet_config) From 039955c52744909107dc68ade0698a55d81a8886 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Fri, 6 Feb 2026 17:14:52 -0800 Subject: [PATCH 592/647] Some fixes to previous pr. (#12339) --- comfy/ldm/cosmos/predict2.py | 2 +- comfy/supported_models.py | 10 ++++++---- 2 files changed, 7 insertions(+), 5 deletions(-) diff --git a/comfy/ldm/cosmos/predict2.py b/comfy/ldm/cosmos/predict2.py index 6491e486b..2268bff38 100644 --- a/comfy/ldm/cosmos/predict2.py +++ b/comfy/ldm/cosmos/predict2.py @@ -894,6 +894,6 @@ class MiniTrainDIT(nn.Module): **block_kwargs, ) - x_B_T_H_W_O = self.final_layer(x_B_T_H_W_D, t_embedding_B_T_D, adaln_lora_B_T_3D=adaln_lora_B_T_3D) + x_B_T_H_W_O = self.final_layer(x_B_T_H_W_D.to(crossattn_emb.dtype), t_embedding_B_T_D, adaln_lora_B_T_3D=adaln_lora_B_T_3D) x_B_C_Tt_Hp_Wp = self.unpatchify(x_B_T_H_W_O)[:, :, :orig_shape[-3], :orig_shape[-2], :orig_shape[-1]] return x_B_C_Tt_Hp_Wp diff --git a/comfy/supported_models.py b/comfy/supported_models.py index 56a21b0ef..d33db7507 100644 --- a/comfy/supported_models.py +++ b/comfy/supported_models.py @@ -1025,10 +1025,6 @@ class Anima(supported_models_base.BASE): supported_inference_dtypes = [torch.bfloat16, torch.float16, torch.float32] - def __init__(self, unet_config): - super().__init__(unet_config) - self.memory_usage_factor = (unet_config.get("model_channels", 2048) / 2048) * 0.95 - def get_model(self, state_dict, prefix="", device=None): out = model_base.Anima(self, device=device) return out @@ -1038,6 +1034,12 @@ class Anima(supported_models_base.BASE): detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen3_06b.transformer.".format(pref)) return supported_models_base.ClipTarget(comfy.text_encoders.anima.AnimaTokenizer, comfy.text_encoders.anima.te(**detect)) + def set_inference_dtype(self, dtype, manual_cast_dtype, **kwargs): + self.memory_usage_factor = (self.unet_config.get("model_channels", 2048) / 2048) * 0.95 + if dtype is torch.float16: + self.memory_usage_factor *= 1.4 + return super().set_inference_dtype(dtype, manual_cast_dtype, **kwargs) + class CosmosI2VPredict2(CosmosT2IPredict2): unet_config = { "image_model": "cosmos_predict2", From 17e7df43d19bde49efa46a32b89f5153b9cb0ded Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Fri, 6 Feb 2026 21:02:11 -0800 Subject: [PATCH 593/647] Pad ace step 1.5 ref audio if not long enough. (#12341) --- comfy/model_base.py | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/comfy/model_base.py b/comfy/model_base.py index 3aa345254..858789b30 100644 --- a/comfy/model_base.py +++ b/comfy/model_base.py @@ -1577,6 +1577,10 @@ class ACEStep15(BaseModel): else: out['is_covers'] = comfy.conds.CONDConstant(False) + if refer_audio.shape[2] < noise.shape[2]: + pad = comfy.ldm.ace.ace_step15.get_silence_latent(noise.shape[2], device) + refer_audio = torch.cat([refer_audio.to(pad), pad[:, :, refer_audio.shape[2]:]], dim=2) + out['refer_audio'] = comfy.conds.CONDRegular(refer_audio) return out From 5ff4fdedba2e72ffabf2948799a1c656d9002b52 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Jukka=20Sepp=C3=A4nen?= <40791699+kijai@users.noreply.github.com> Date: Sat, 7 Feb 2026 21:25:30 +0200 Subject: [PATCH 594/647] Fix LazyCache (#12344) --- comfy_extras/nodes_easycache.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy_extras/nodes_easycache.py b/comfy_extras/nodes_easycache.py index 51d1e5b9c..b1912392c 100644 --- a/comfy_extras/nodes_easycache.py +++ b/comfy_extras/nodes_easycache.py @@ -108,7 +108,7 @@ def lazycache_predict_noise_wrapper(executor, *args, **kwargs): easycache: LazyCacheHolder = model_options["transformer_options"]["easycache"] if easycache.is_past_end_timestep(timestep): return executor(*args, **kwargs) - x: torch.Tensor = _extract_tensor(args[0], easycache.output_channels) + x: torch.Tensor = args[0][:, :easycache.output_channels] # prepare next x_prev next_x_prev = x input_change = None From 9bf5aa54dbda5b5de36812cfc10b123ae0930283 Mon Sep 17 00:00:00 2001 From: chaObserv <154517000+chaObserv@users.noreply.github.com> Date: Sun, 8 Feb 2026 06:38:51 +0800 Subject: [PATCH 595/647] Add search_aliases to sa-solver and seeds-2 node (#12327) --- comfy_extras/nodes_custom_sampler.py | 2 ++ 1 file changed, 2 insertions(+) diff --git a/comfy_extras/nodes_custom_sampler.py b/comfy_extras/nodes_custom_sampler.py index 8afd13acf..61a234634 100644 --- a/comfy_extras/nodes_custom_sampler.py +++ b/comfy_extras/nodes_custom_sampler.py @@ -622,6 +622,7 @@ class SamplerSASolver(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="SamplerSASolver", + search_aliases=["sde"], category="sampling/custom_sampling/samplers", inputs=[ io.Model.Input("model"), @@ -666,6 +667,7 @@ class SamplerSEEDS2(io.ComfyNode): def define_schema(cls): return io.Schema( node_id="SamplerSEEDS2", + search_aliases=["sde", "exp heun"], category="sampling/custom_sampling/samplers", inputs=[ io.Combo.Input("solver_type", options=["phi_1", "phi_2"]), From 3760d74005a6954f54657dc59d9e57fd4c44b3fd Mon Sep 17 00:00:00 2001 From: ComfyUI Wiki Date: Sun, 8 Feb 2026 07:34:52 +0800 Subject: [PATCH 596/647] chore: update embedded docs to v0.4.1 (#12346) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 41cc9174b..5e34a2a49 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,6 +1,6 @@ comfyui-frontend-package==1.38.13 comfyui-workflow-templates==0.8.31 -comfyui-embedded-docs==0.4.0 +comfyui-embedded-docs==0.4.1 torch torchsde torchvision From f350a842611f4d75da7104c2d2965f45989089b9 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sat, 7 Feb 2026 16:16:28 -0800 Subject: [PATCH 597/647] Disable prompt weights for ltxv2. (#12354) --- comfy/text_encoders/lt.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy/text_encoders/lt.py b/comfy/text_encoders/lt.py index 26573fb12..3f87dfd6a 100644 --- a/comfy/text_encoders/lt.py +++ b/comfy/text_encoders/lt.py @@ -25,7 +25,7 @@ def ltxv_te(*args, **kwargs): class Gemma3_12BTokenizer(sd1_clip.SDTokenizer): def __init__(self, embedding_directory=None, tokenizer_data={}): tokenizer = tokenizer_data.get("spiece_model", None) - super().__init__(tokenizer, pad_with_end=False, embedding_size=3840, embedding_key='gemma3_12b', tokenizer_class=SPieceTokenizer, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, tokenizer_args={"add_bos": True, "add_eos": False}, tokenizer_data=tokenizer_data) + super().__init__(tokenizer, pad_with_end=False, embedding_size=3840, embedding_key='gemma3_12b', tokenizer_class=SPieceTokenizer, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, disable_weights=True, tokenizer_args={"add_bos": True, "add_eos": False}, tokenizer_data=tokenizer_data) def state_dict(self): return {"spiece_model": self.tokenizer.serialize_model()} From a0302cc6a85dcb950a7308f7a31a224ef54f3d58 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sun, 8 Feb 2026 18:16:40 -0800 Subject: [PATCH 598/647] Make tonemap latent work on any dim latents. (#12363) --- comfy_extras/nodes_latent.py | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-) diff --git a/comfy_extras/nodes_latent.py b/comfy_extras/nodes_latent.py index 6aecf1561..8d2d7297a 100644 --- a/comfy_extras/nodes_latent.py +++ b/comfy_extras/nodes_latent.py @@ -391,8 +391,9 @@ class LatentOperationTonemapReinhard(io.ComfyNode): latent_vector_magnitude = (torch.linalg.vector_norm(latent, dim=(1)) + 0.0000000001)[:,None] normalized_latent = latent / latent_vector_magnitude - mean = torch.mean(latent_vector_magnitude, dim=(1,2,3), keepdim=True) - std = torch.std(latent_vector_magnitude, dim=(1,2,3), keepdim=True) + dims = list(range(1, latent_vector_magnitude.ndim)) + mean = torch.mean(latent_vector_magnitude, dim=dims, keepdim=True) + std = torch.std(latent_vector_magnitude, dim=dims, keepdim=True) top = (std * 5 + mean) * multiplier From 62315fbb15861e64b917d0a072dad5dc9a15173c Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Mon, 9 Feb 2026 13:16:08 -0800 Subject: [PATCH 599/647] Dynamic VRAM fixes - Ace 1.5 performance + a VRAM leak (#12368) * revert threaded model loader change This change was only needed to get around the pytorch 2.7 mempool bugs, and should have been reverted along with #12260. This fixes a different memory leak where pytorch gets confused about cache emptying. * load non comfy weights * MPDynamic: Pre-generate the tensors for vbars Apparently this is an expensive operation that slows down things. * bump to aimdo 1.8 New features: watermark limit feature logging enhancements -O2 build on linux --- comfy/model_management.py | 37 ++++++------------------------------- comfy/model_patcher.py | 7 ++++++- comfy/ops.py | 2 +- execution.py | 7 ++++++- requirements.txt | 2 +- 5 files changed, 20 insertions(+), 35 deletions(-) diff --git a/comfy/model_management.py b/comfy/model_management.py index b6291f340..6018c1ab6 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -19,7 +19,7 @@ import psutil import logging from enum import Enum -from comfy.cli_args import args, PerformanceFeature, enables_dynamic_vram +from comfy.cli_args import args, PerformanceFeature import threading import torch import sys @@ -651,7 +651,7 @@ def free_memory(memory_required, device, keep_loaded=[], for_dynamic=False, ram_ soft_empty_cache() return unloaded_models -def load_models_gpu_orig(models, memory_required=0, force_patch_weights=False, minimum_memory_required=None, force_full_load=False): +def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimum_memory_required=None, force_full_load=False): cleanup_models_gc() global vram_state @@ -747,26 +747,6 @@ def load_models_gpu_orig(models, memory_required=0, force_patch_weights=False, m current_loaded_models.insert(0, loaded_model) return -def load_models_gpu_thread(models, memory_required, force_patch_weights, minimum_memory_required, force_full_load): - with torch.inference_mode(): - load_models_gpu_orig(models, memory_required, force_patch_weights, minimum_memory_required, force_full_load) - soft_empty_cache() - -def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimum_memory_required=None, force_full_load=False): - #Deliberately load models outside of the Aimdo mempool so they can be retained accross - #nodes. Use a dummy thread to do it as pytorch documents that mempool contexts are - #thread local. So exploit that to escape context - if enables_dynamic_vram(): - t = threading.Thread( - target=load_models_gpu_thread, - args=(models, memory_required, force_patch_weights, minimum_memory_required, force_full_load) - ) - t.start() - t.join() - else: - load_models_gpu_orig(models, memory_required=memory_required, force_patch_weights=force_patch_weights, - minimum_memory_required=minimum_memory_required, force_full_load=force_full_load) - def load_model_gpu(model): return load_models_gpu([model]) @@ -1226,21 +1206,16 @@ def cast_to(weight, dtype=None, device=None, non_blocking=False, copy=False, str if dtype is None: dtype = weight._model_dtype - r = torch.empty_like(weight, dtype=dtype, device=device) - signature = comfy_aimdo.model_vbar.vbar_fault(weight._v) if signature is not None: - raw_tensor = comfy_aimdo.torch.aimdo_to_tensor(weight._v, device) - v_tensor = comfy.memory_management.interpret_gathered_like(cast_geometry, raw_tensor)[0] + v_tensor = comfy.memory_management.interpret_gathered_like(cast_geometry, weight._v_tensor)[0] if not comfy_aimdo.model_vbar.vbar_signature_compare(signature, weight._v_signature): weight._v_signature = signature #Send it over v_tensor.copy_(weight, non_blocking=non_blocking) - #always take a deep copy even if _v is good, as we have no reasonable point to unpin - #a non comfy weight - r.copy_(v_tensor) - comfy_aimdo.model_vbar.vbar_unpin(weight._v) - return r + return v_tensor.to(dtype=dtype) + + r = torch.empty_like(weight, dtype=dtype, device=device) if weight.dtype != r.dtype and weight.dtype != weight._model_dtype: #Offloaded casting could skip this, however it would make the quantizations diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index d888dbcfb..b9a117a7c 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -1492,7 +1492,9 @@ class ModelPatcherDynamic(ModelPatcher): if vbar is not None: vbar.prioritize() - #We have way more tools for acceleration on comfy weight offloading, so always + #We force reserve VRAM for the non comfy-weight so we dont have to deal + #with pin and unpin syncrhonization which can be expensive for small weights + #with a high layer rate (e.g. autoregressive LLMs). #prioritize the non-comfy weights (note the order reverse). loading = self._load_list(prio_comfy_cast_weights=True) loading.sort(reverse=True) @@ -1541,6 +1543,7 @@ class ModelPatcherDynamic(ModelPatcher): if vbar is not None and not hasattr(m, "_v"): m._v = vbar.alloc(v_weight_size) + m._v_tensor = comfy_aimdo.torch.aimdo_to_tensor(m._v, device_to) allocated_size += v_weight_size else: @@ -1555,8 +1558,10 @@ class ModelPatcherDynamic(ModelPatcher): weight_size = geometry.numel() * geometry.element_size() if vbar is not None and not hasattr(weight, "_v"): weight._v = vbar.alloc(weight_size) + weight._v_tensor = comfy_aimdo.torch.aimdo_to_tensor(weight._v, device_to) weight._model_dtype = model_dtype allocated_size += weight_size + vbar.set_watermark_limit(allocated_size) logging.info(f"Model {self.model.__class__.__name__} prepared for dynamic VRAM loading. {allocated_size // (1024 ** 2)}MB Staged. {num_patches} patches attached.") diff --git a/comfy/ops.py b/comfy/ops.py index 0f4eca7c7..ea0d70702 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -87,7 +87,7 @@ def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compu signature = comfy_aimdo.model_vbar.vbar_fault(s._v) if signature is not None: - xfer_dest = comfy_aimdo.torch.aimdo_to_tensor(s._v, device) + xfer_dest = s._v_tensor resident = comfy_aimdo.model_vbar.vbar_signature_compare(signature, s._v_signature) if not resident: diff --git a/execution.py b/execution.py index 3dbab82e6..896862c6b 100644 --- a/execution.py +++ b/execution.py @@ -13,8 +13,11 @@ from contextlib import nullcontext import torch +from comfy.cli_args import args import comfy.memory_management import comfy.model_management +import comfy_aimdo.model_vbar + from latent_preview import set_preview_method import nodes from comfy_execution.caching import ( @@ -527,8 +530,10 @@ async def execute(server, dynprompt, caches, current_item, extra_data, executed, output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data) finally: if allocator is not None: + if args.verbose == "DEBUG": + comfy_aimdo.model_vbar.vbars_analyze() comfy.model_management.reset_cast_buffers() - torch.cuda.synchronize() + comfy_aimdo.model_vbar.vbars_reset_watermark_limits() if has_pending_tasks: pending_async_nodes[unique_id] = output_data diff --git a/requirements.txt b/requirements.txt index 5e34a2a49..4fda07fde 100644 --- a/requirements.txt +++ b/requirements.txt @@ -22,7 +22,7 @@ alembic SQLAlchemy av>=14.2.0 comfy-kitchen>=0.2.7 -comfy-aimdo>=0.1.7 +comfy-aimdo>=0.1.8 requests #non essential dependencies: From baf8c874557f1522a99d47d94faad12b0257c8f1 Mon Sep 17 00:00:00 2001 From: blepping <157360029+blepping@users.noreply.github.com> Date: Mon, 9 Feb 2026 17:41:49 -0700 Subject: [PATCH 600/647] Iimprovements to ACE-Steps 1.5 text encoding (part 2) (#12350) --- comfy/text_encoders/ace15.py | 114 +++++++++++++++++++++++++---------- 1 file changed, 81 insertions(+), 33 deletions(-) diff --git a/comfy/text_encoders/ace15.py b/comfy/text_encoders/ace15.py index 00dd5ba90..5dac644c2 100644 --- a/comfy/text_encoders/ace15.py +++ b/comfy/text_encoders/ace15.py @@ -3,6 +3,7 @@ import comfy.text_encoders.llama from comfy import sd1_clip import torch import math +from tqdm.auto import trange import yaml import comfy.utils @@ -23,6 +24,8 @@ def sample_manual_loop_no_classes( audio_end_id: int = 215669, eos_token_id: int = 151645, ): + if ids is None: + return [] device = model.execution_device if execution_dtype is None: @@ -32,6 +35,7 @@ def sample_manual_loop_no_classes( execution_dtype = torch.float32 embeds, attention_mask, num_tokens, embeds_info = model.process_tokens(ids, device) + embeds_batch = embeds.shape[0] for i, t in enumerate(paddings): attention_mask[i, :t] = 0 attention_mask[i, t:] = 1 @@ -41,22 +45,27 @@ def sample_manual_loop_no_classes( generator = torch.Generator(device=device) generator.manual_seed(seed) model_config = model.transformer.model.config + past_kv_shape = [embeds_batch, model_config.num_key_value_heads, embeds.shape[1] + min_tokens, model_config.head_dim] for x in range(model_config.num_hidden_layers): - past_key_values.append((torch.empty([embeds.shape[0], model_config.num_key_value_heads, embeds.shape[1] + min_tokens, model_config.head_dim], device=device, dtype=execution_dtype), torch.empty([embeds.shape[0], model_config.num_key_value_heads, embeds.shape[1] + min_tokens, model_config.head_dim], device=device, dtype=execution_dtype), 0)) + past_key_values.append((torch.empty(past_kv_shape, device=device, dtype=execution_dtype), torch.empty(past_kv_shape, device=device, dtype=execution_dtype), 0)) progress_bar = comfy.utils.ProgressBar(max_new_tokens) - for step in range(max_new_tokens): + for step in trange(max_new_tokens, desc="LM sampling"): outputs = model.transformer(None, attention_mask, embeds=embeds.to(execution_dtype), num_tokens=num_tokens, intermediate_output=None, dtype=execution_dtype, embeds_info=embeds_info, past_key_values=past_key_values) next_token_logits = model.transformer.logits(outputs[0])[:, -1] past_key_values = outputs[2] - cond_logits = next_token_logits[0:1] - uncond_logits = next_token_logits[1:2] - cfg_logits = uncond_logits + cfg_scale * (cond_logits - uncond_logits) + if cfg_scale != 1.0: + cond_logits = next_token_logits[0:1] + uncond_logits = next_token_logits[1:2] + cfg_logits = uncond_logits + cfg_scale * (cond_logits - uncond_logits) + else: + cfg_logits = next_token_logits[0:1] - if eos_token_id is not None and eos_token_id < audio_start_id and min_tokens < step: + use_eos_score = eos_token_id is not None and eos_token_id < audio_start_id and min_tokens < step + if use_eos_score: eos_score = cfg_logits[:, eos_token_id].clone() remove_logit_value = torch.finfo(cfg_logits.dtype).min @@ -64,7 +73,7 @@ def sample_manual_loop_no_classes( cfg_logits[:, :audio_start_id] = remove_logit_value cfg_logits[:, audio_end_id:] = remove_logit_value - if eos_token_id is not None and eos_token_id < audio_start_id and min_tokens < step: + if use_eos_score: cfg_logits[:, eos_token_id] = eos_score if top_k is not None and top_k > 0: @@ -93,8 +102,8 @@ def sample_manual_loop_no_classes( break embed, _, _, _ = model.process_tokens([[token]], device) - embeds = embed.repeat(2, 1, 1) - attention_mask = torch.cat([attention_mask, torch.ones((2, 1), device=device, dtype=attention_mask.dtype)], dim=1) + embeds = embed.repeat(embeds_batch, 1, 1) + attention_mask = torch.cat([attention_mask, torch.ones((embeds_batch, 1), device=device, dtype=attention_mask.dtype)], dim=1) output_audio_codes.append(token - audio_start_id) progress_bar.update_absolute(step) @@ -104,22 +113,29 @@ def sample_manual_loop_no_classes( def generate_audio_codes(model, positive, negative, min_tokens=1, max_tokens=1024, seed=0, cfg_scale=2.0, temperature=0.85, top_p=0.9, top_k=0): positive = [[token for token, _ in inner_list] for inner_list in positive] - negative = [[token for token, _ in inner_list] for inner_list in negative] positive = positive[0] - negative = negative[0] - neg_pad = 0 - if len(negative) < len(positive): - neg_pad = (len(positive) - len(negative)) - negative = [model.special_tokens["pad"]] * neg_pad + negative + if cfg_scale != 1.0: + negative = [[token for token, _ in inner_list] for inner_list in negative] + negative = negative[0] - pos_pad = 0 - if len(negative) > len(positive): - pos_pad = (len(negative) - len(positive)) - positive = [model.special_tokens["pad"]] * pos_pad + positive + neg_pad = 0 + if len(negative) < len(positive): + neg_pad = (len(positive) - len(negative)) + negative = [model.special_tokens["pad"]] * neg_pad + negative - paddings = [pos_pad, neg_pad] - return sample_manual_loop_no_classes(model, [positive, negative], paddings, cfg_scale=cfg_scale, temperature=temperature, top_p=top_p, top_k=top_k, seed=seed, min_tokens=min_tokens, max_new_tokens=max_tokens) + pos_pad = 0 + if len(negative) > len(positive): + pos_pad = (len(negative) - len(positive)) + positive = [model.special_tokens["pad"]] * pos_pad + positive + + paddings = [pos_pad, neg_pad] + ids = [positive, negative] + else: + paddings = [] + ids = [positive] + + return sample_manual_loop_no_classes(model, ids, paddings, cfg_scale=cfg_scale, temperature=temperature, top_p=top_p, top_k=top_k, seed=seed, min_tokens=min_tokens, max_new_tokens=max_tokens) class ACE15Tokenizer(sd1_clip.SD1Tokenizer): @@ -129,12 +145,12 @@ class ACE15Tokenizer(sd1_clip.SD1Tokenizer): def _metas_to_cot(self, *, return_yaml: bool = False, **kwargs) -> str: user_metas = { k: kwargs.pop(k) - for k in ("bpm", "duration", "keyscale", "timesignature", "language", "caption") + for k in ("bpm", "duration", "keyscale", "timesignature", "language") if k in kwargs } timesignature = user_metas.get("timesignature") if isinstance(timesignature, str) and timesignature.endswith("/4"): - user_metas["timesignature"] = timesignature.rsplit("/", 1)[0] + user_metas["timesignature"] = timesignature[:-2] user_metas = { k: v if not isinstance(v, str) or not v.isdigit() else int(v) for k, v in user_metas.items() @@ -147,8 +163,11 @@ class ACE15Tokenizer(sd1_clip.SD1Tokenizer): return f"\n{meta_yaml}\n" if not return_yaml else meta_yaml def _metas_to_cap(self, **kwargs) -> str: - use_keys = ("bpm", "duration", "keyscale", "timesignature") + use_keys = ("bpm", "timesignature", "keyscale", "duration") user_metas = { k: kwargs.pop(k, "N/A") for k in use_keys } + timesignature = user_metas.get("timesignature") + if isinstance(timesignature, str) and timesignature.endswith("/4"): + user_metas["timesignature"] = timesignature[:-2] duration = user_metas["duration"] if duration == "N/A": user_metas["duration"] = "30 seconds" @@ -159,9 +178,13 @@ class ACE15Tokenizer(sd1_clip.SD1Tokenizer): return "\n".join(f"- {k}: {user_metas[k]}" for k in use_keys) def tokenize_with_weights(self, text, return_word_ids=False, **kwargs): - out = {} + text = text.strip() + text_negative = kwargs.get("caption_negative", text).strip() lyrics = kwargs.get("lyrics", "") + lyrics_negative = kwargs.get("lyrics_negative", lyrics) duration = kwargs.get("duration", 120) + if isinstance(duration, str): + duration = float(duration.split(None, 1)[0]) language = kwargs.get("language") seed = kwargs.get("seed", 0) @@ -171,21 +194,46 @@ class ACE15Tokenizer(sd1_clip.SD1Tokenizer): top_p = kwargs.get("top_p", 0.9) top_k = kwargs.get("top_k", 0.0) - duration = math.ceil(duration) kwargs["duration"] = duration + tokens_duration = duration * 5 + min_tokens = int(kwargs.get("min_tokens", tokens_duration)) + max_tokens = int(kwargs.get("max_tokens", tokens_duration)) + + metas_negative = { + k.rsplit("_", 1)[0]: kwargs.pop(k) + for k in ("bpm_negative", "duration_negative", "keyscale_negative", "timesignature_negative", "language_negative", "caption_negative") + if k in kwargs + } + if not kwargs.get("use_negative_caption"): + _ = metas_negative.pop("caption", None) cot_text = self._metas_to_cot(caption = text, **kwargs) + cot_text_negative = "\n" if not metas_negative else self._metas_to_cot(**metas_negative) meta_cap = self._metas_to_cap(**kwargs) - lm_template = "<|im_start|>system\n# Instruction\nGenerate audio semantic tokens based on the given conditions:\n\n<|im_end|>\n<|im_start|>user\n# Caption\n{}\n# Lyric\n{}\n<|im_end|>\n<|im_start|>assistant\n{}\n<|im_end|>\n" + lm_template = "<|im_start|>system\n# Instruction\nGenerate audio semantic tokens based on the given conditions:\n\n<|im_end|>\n<|im_start|>user\n# Caption\n{}\n\n# Lyric\n{}\n<|im_end|>\n<|im_start|>assistant\n{}\n\n<|im_end|>\n" + lyrics_template = "# Languages\n{}\n\n# Lyric\n{}<|endoftext|><|endoftext|>" + qwen3_06b_template = "# Instruction\nGenerate audio semantic tokens based on the given conditions:\n\n# Caption\n{}\n\n# Metas\n{}\n<|endoftext|>\n<|endoftext|>" - out["lm_prompt"] = self.qwen3_06b.tokenize_with_weights(lm_template.format(text, lyrics, cot_text), disable_weights=True) - out["lm_prompt_negative"] = self.qwen3_06b.tokenize_with_weights(lm_template.format(text, lyrics, "\n"), disable_weights=True) + llm_prompts = { + "lm_prompt": lm_template.format(text, lyrics.strip(), cot_text), + "lm_prompt_negative": lm_template.format(text_negative, lyrics_negative.strip(), cot_text_negative), + "lyrics": lyrics_template.format(language if language is not None else "", lyrics), + "qwen3_06b": qwen3_06b_template.format(text, meta_cap), + } - out["lyrics"] = self.qwen3_06b.tokenize_with_weights("# Languages\n{}\n\n# Lyric\n{}<|endoftext|><|endoftext|>".format(language if language is not None else "", lyrics), return_word_ids, disable_weights=True, **kwargs) - out["qwen3_06b"] = self.qwen3_06b.tokenize_with_weights("# Instruction\nGenerate audio semantic tokens based on the given conditions:\n\n# Caption\n{}\n# Metas\n{}\n<|endoftext|>\n<|endoftext|>".format(text, meta_cap), return_word_ids, **kwargs) - out["lm_metadata"] = {"min_tokens": duration * 5, + out = { + prompt_key: self.qwen3_06b.tokenize_with_weights( + prompt, + prompt_key == "qwen3_06b" and return_word_ids, + disable_weights = True, + **kwargs, + ) + for prompt_key, prompt in llm_prompts.items() + } + out["lm_metadata"] = {"min_tokens": min_tokens, + "max_tokens": max_tokens, "seed": seed, "generate_audio_codes": generate_audio_codes, "cfg_scale": cfg_scale, @@ -252,7 +300,7 @@ class ACE15TEModel(torch.nn.Module): lm_metadata = token_weight_pairs["lm_metadata"] if lm_metadata["generate_audio_codes"]: - audio_codes = generate_audio_codes(getattr(self, self.lm_model, self.qwen3_06b), token_weight_pairs["lm_prompt"], token_weight_pairs["lm_prompt_negative"], min_tokens=lm_metadata["min_tokens"], max_tokens=lm_metadata["min_tokens"], seed=lm_metadata["seed"], cfg_scale=lm_metadata["cfg_scale"], temperature=lm_metadata["temperature"], top_p=lm_metadata["top_p"], top_k=lm_metadata["top_k"]) + audio_codes = generate_audio_codes(getattr(self, self.lm_model, self.qwen3_06b), token_weight_pairs["lm_prompt"], token_weight_pairs["lm_prompt_negative"], min_tokens=lm_metadata["min_tokens"], max_tokens=lm_metadata["max_tokens"], seed=lm_metadata["seed"], cfg_scale=lm_metadata["cfg_scale"], temperature=lm_metadata["temperature"], top_p=lm_metadata["top_p"], top_k=lm_metadata["top_k"]) out["audio_codes"] = [audio_codes] return base_out, None, out From a4be04c5d750cc5d62256f7f86bb5a7c0a78e28d Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 9 Feb 2026 16:45:56 -0800 Subject: [PATCH 601/647] Ace step prompts match now. (#12376) --- comfy/text_encoders/ace15.py | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/comfy/text_encoders/ace15.py b/comfy/text_encoders/ace15.py index 5dac644c2..73697b3c1 100644 --- a/comfy/text_encoders/ace15.py +++ b/comfy/text_encoders/ace15.py @@ -145,7 +145,7 @@ class ACE15Tokenizer(sd1_clip.SD1Tokenizer): def _metas_to_cot(self, *, return_yaml: bool = False, **kwargs) -> str: user_metas = { k: kwargs.pop(k) - for k in ("bpm", "duration", "keyscale", "timesignature", "language") + for k in ("bpm", "duration", "keyscale", "timesignature") if k in kwargs } timesignature = user_metas.get("timesignature") @@ -208,8 +208,8 @@ class ACE15Tokenizer(sd1_clip.SD1Tokenizer): if not kwargs.get("use_negative_caption"): _ = metas_negative.pop("caption", None) - cot_text = self._metas_to_cot(caption = text, **kwargs) - cot_text_negative = "\n" if not metas_negative else self._metas_to_cot(**metas_negative) + cot_text = self._metas_to_cot(caption=text, **kwargs) + cot_text_negative = "\n\n" if not metas_negative else self._metas_to_cot(**metas_negative) meta_cap = self._metas_to_cap(**kwargs) lm_template = "<|im_start|>system\n# Instruction\nGenerate audio semantic tokens based on the given conditions:\n\n<|im_end|>\n<|im_start|>user\n# Caption\n{}\n\n# Lyric\n{}\n<|im_end|>\n<|im_start|>assistant\n{}\n\n<|im_end|>\n" From 349a636a2b0f15aba2930b9af905bb805d2fe30b Mon Sep 17 00:00:00 2001 From: ComfyUI Wiki Date: Tue, 10 Feb 2026 10:25:34 +0800 Subject: [PATCH 602/647] chore: update workflow templates to v0.8.37 (#12377) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 4fda07fde..4e2773f5d 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,5 +1,5 @@ comfyui-frontend-package==1.38.13 -comfyui-workflow-templates==0.8.31 +comfyui-workflow-templates==0.8.37 comfyui-embedded-docs==0.4.1 torch torchsde From c1b63a7e78b606bc14cd49a02e9338274db28a60 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Tue, 10 Feb 2026 04:58:27 +0200 Subject: [PATCH 603/647] fix(Moonvalley-API-Nodes): adjust "steps" parameter to not raise exception (#12370) --- comfy_api_nodes/nodes_moonvalley.py | 14 +++++++------- 1 file changed, 7 insertions(+), 7 deletions(-) diff --git a/comfy_api_nodes/nodes_moonvalley.py b/comfy_api_nodes/nodes_moonvalley.py index 08315fa2b..78a230529 100644 --- a/comfy_api_nodes/nodes_moonvalley.py +++ b/comfy_api_nodes/nodes_moonvalley.py @@ -219,8 +219,8 @@ class MoonvalleyImg2VideoNode(IO.ComfyNode): ), IO.Int.Input( "steps", - default=33, - min=1, + default=80, + min=75, # steps should be greater or equal to cooldown_steps(75) + warmup_steps(0) max=100, step=1, tooltip="Number of denoising steps", @@ -340,8 +340,8 @@ class MoonvalleyVideo2VideoNode(IO.ComfyNode): ), IO.Int.Input( "steps", - default=33, - min=1, + default=60, + min=60, # steps should be greater or equal to cooldown_steps(36) + warmup_steps(24) max=100, step=1, display_mode=IO.NumberDisplay.number, @@ -370,7 +370,7 @@ class MoonvalleyVideo2VideoNode(IO.ComfyNode): video: Input.Video | None = None, control_type: str = "Motion Transfer", motion_intensity: int | None = 100, - steps=33, + steps=60, prompt_adherence=4.5, ) -> IO.NodeOutput: validated_video = validate_video_to_video_input(video) @@ -465,8 +465,8 @@ class MoonvalleyTxt2VideoNode(IO.ComfyNode): ), IO.Int.Input( "steps", - default=33, - min=1, + default=80, + min=75, # steps should be greater or equal to cooldown_steps(75) + warmup_steps(0) max=100, step=1, tooltip="Inference steps", From 8ca842a8edb26006e730e631ec1153cd42f46d3b Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Tue, 10 Feb 2026 19:34:54 +0200 Subject: [PATCH 604/647] feat(api-nodes-Kling): add new models (V3, O3) (#12389) * feat(api-nodes-Kling): add new models (V3, O3) * remove storyboard from VideoToVideo node * added check for total duration of storyboards * fixed other small things * updated display name for nodes * added "fake" seed --- comfy_api_nodes/apis/__init__.py | 8 +- comfy_api_nodes/apis/kling.py | 46 +- comfy_api_nodes/nodes_kling.py | 764 ++++++++++++++++++++++++++++--- 3 files changed, 750 insertions(+), 68 deletions(-) diff --git a/comfy_api_nodes/apis/__init__.py b/comfy_api_nodes/apis/__init__.py index ee2aa1ce6..46a583b5e 100644 --- a/comfy_api_nodes/apis/__init__.py +++ b/comfy_api_nodes/apis/__init__.py @@ -1197,12 +1197,6 @@ class KlingImageGenImageReferenceType(str, Enum): face = 'face' -class KlingImageGenModelName(str, Enum): - kling_v1 = 'kling-v1' - kling_v1_5 = 'kling-v1-5' - kling_v2 = 'kling-v2' - - class KlingImageGenerationsRequest(BaseModel): aspect_ratio: Optional[KlingImageGenAspectRatio] = '16:9' callback_url: Optional[AnyUrl] = Field( @@ -1218,7 +1212,7 @@ class KlingImageGenerationsRequest(BaseModel): 0.5, description='Reference intensity for user-uploaded images', ge=0.0, le=1.0 ) image_reference: Optional[KlingImageGenImageReferenceType] = None - model_name: Optional[KlingImageGenModelName] = 'kling-v1' + model_name: str = Field(...) n: Optional[int] = Field(1, description='Number of generated images', ge=1, le=9) negative_prompt: Optional[str] = Field( None, description='Negative text prompt', max_length=200 diff --git a/comfy_api_nodes/apis/kling.py b/comfy_api_nodes/apis/kling.py index bf54ede3e..9c0446075 100644 --- a/comfy_api_nodes/apis/kling.py +++ b/comfy_api_nodes/apis/kling.py @@ -1,12 +1,22 @@ from pydantic import BaseModel, Field +class MultiPromptEntry(BaseModel): + index: int = Field(...) + prompt: str = Field(...) + duration: str = Field(...) + + class OmniProText2VideoRequest(BaseModel): model_name: str = Field(..., description="kling-video-o1") aspect_ratio: str = Field(..., description="'16:9', '9:16' or '1:1'") duration: str = Field(..., description="'5' or '10'") prompt: str = Field(...) mode: str = Field("pro") + multi_shot: bool | None = Field(None) + multi_prompt: list[MultiPromptEntry] | None = Field(None) + shot_type: str | None = Field(None) + sound: str = Field(..., description="'on' or 'off'") class OmniParamImage(BaseModel): @@ -26,6 +36,10 @@ class OmniProFirstLastFrameRequest(BaseModel): duration: str = Field(..., description="'5' or '10'") prompt: str = Field(...) mode: str = Field("pro") + sound: str | None = Field(None, description="'on' or 'off'") + multi_shot: bool | None = Field(None) + multi_prompt: list[MultiPromptEntry] | None = Field(None) + shot_type: str | None = Field(None) class OmniProReferences2VideoRequest(BaseModel): @@ -38,6 +52,10 @@ class OmniProReferences2VideoRequest(BaseModel): duration: str | None = Field(..., description="From 3 to 10.") prompt: str = Field(...) mode: str = Field("pro") + sound: str | None = Field(None, description="'on' or 'off'") + multi_shot: bool | None = Field(None) + multi_prompt: list[MultiPromptEntry] | None = Field(None) + shot_type: str | None = Field(None) class TaskStatusVideoResult(BaseModel): @@ -54,6 +72,7 @@ class TaskStatusImageResult(BaseModel): class TaskStatusResults(BaseModel): videos: list[TaskStatusVideoResult] | None = Field(None) images: list[TaskStatusImageResult] | None = Field(None) + series_images: list[TaskStatusImageResult] | None = Field(None) class TaskStatusResponseData(BaseModel): @@ -77,31 +96,42 @@ class OmniImageParamImage(BaseModel): class OmniProImageRequest(BaseModel): - model_name: str = Field(..., description="kling-image-o1") - resolution: str = Field(..., description="'1k' or '2k'") + model_name: str = Field(...) + resolution: str = Field(...) aspect_ratio: str | None = Field(...) prompt: str = Field(...) mode: str = Field("pro") n: int | None = Field(1, le=9) image_list: list[OmniImageParamImage] | None = Field(..., max_length=10) + result_type: str | None = Field(None, description="Set to 'series' for series generation") + series_amount: int | None = Field(None, ge=2, le=9, description="Number of images in a series") class TextToVideoWithAudioRequest(BaseModel): - model_name: str = Field(..., description="kling-v2-6") + model_name: str = Field(...) aspect_ratio: str = Field(..., description="'16:9', '9:16' or '1:1'") - duration: str = Field(..., description="'5' or '10'") - prompt: str = Field(...) + duration: str = Field(...) + prompt: str | None = Field(...) + negative_prompt: str | None = Field(None) mode: str = Field("pro") sound: str = Field(..., description="'on' or 'off'") + multi_shot: bool | None = Field(None) + multi_prompt: list[MultiPromptEntry] | None = Field(None) + shot_type: str | None = Field(None) class ImageToVideoWithAudioRequest(BaseModel): - model_name: str = Field(..., description="kling-v2-6") + model_name: str = Field(...) image: str = Field(...) - duration: str = Field(..., description="'5' or '10'") - prompt: str = Field(...) + image_tail: str | None = Field(None) + duration: str = Field(...) + prompt: str | None = Field(...) + negative_prompt: str | None = Field(None) mode: str = Field("pro") sound: str = Field(..., description="'on' or 'off'") + multi_shot: bool | None = Field(None) + multi_prompt: list[MultiPromptEntry] | None = Field(None) + shot_type: str | None = Field(None) class MotionControlRequest(BaseModel): diff --git a/comfy_api_nodes/nodes_kling.py b/comfy_api_nodes/nodes_kling.py index 739fe1855..b89c85561 100644 --- a/comfy_api_nodes/nodes_kling.py +++ b/comfy_api_nodes/nodes_kling.py @@ -38,7 +38,6 @@ from comfy_api_nodes.apis import ( KlingImageGenerationsRequest, KlingImageGenerationsResponse, KlingImageGenImageReferenceType, - KlingImageGenModelName, KlingImageGenAspectRatio, KlingVideoEffectsRequest, KlingVideoEffectsResponse, @@ -52,6 +51,7 @@ from comfy_api_nodes.apis import ( from comfy_api_nodes.apis.kling import ( ImageToVideoWithAudioRequest, MotionControlRequest, + MultiPromptEntry, OmniImageParamImage, OmniParamImage, OmniParamVideo, @@ -71,6 +71,7 @@ from comfy_api_nodes.util import ( sync_op, tensor_to_base64_string, upload_audio_to_comfyapi, + upload_image_to_comfyapi, upload_images_to_comfyapi, upload_video_to_comfyapi, validate_image_aspect_ratio, @@ -80,6 +81,31 @@ from comfy_api_nodes.util import ( validate_video_duration, ) + +def _generate_storyboard_inputs(count: int) -> list: + inputs = [] + for i in range(1, count + 1): + inputs.extend( + [ + IO.String.Input( + f"storyboard_{i}_prompt", + multiline=True, + default="", + tooltip=f"Prompt for storyboard segment {i}. Max 512 characters.", + ), + IO.Int.Input( + f"storyboard_{i}_duration", + default=4, + min=1, + max=15, + display_mode=IO.NumberDisplay.slider, + tooltip=f"Duration for storyboard segment {i} in seconds.", + ), + ] + ) + return inputs + + KLING_API_VERSION = "v1" PATH_TEXT_TO_VIDEO = f"/proxy/kling/{KLING_API_VERSION}/videos/text2video" PATH_IMAGE_TO_VIDEO = f"/proxy/kling/{KLING_API_VERSION}/videos/image2video" @@ -820,20 +846,48 @@ class OmniProTextToVideoNode(IO.ComfyNode): def define_schema(cls) -> IO.Schema: return IO.Schema( node_id="KlingOmniProTextToVideoNode", - display_name="Kling Omni Text to Video (Pro)", + display_name="Kling 3.0 Omni Text to Video", category="api node/video/Kling", description="Use text prompts to generate videos with the latest Kling model.", inputs=[ - IO.Combo.Input("model_name", options=["kling-video-o1"]), + IO.Combo.Input("model_name", options=["kling-v3-omni", "kling-video-o1"]), IO.String.Input( "prompt", multiline=True, tooltip="A text prompt describing the video content. " - "This can include both positive and negative descriptions.", + "This can include both positive and negative descriptions. " + "Ignored when storyboards are enabled.", ), IO.Combo.Input("aspect_ratio", options=["16:9", "9:16", "1:1"]), - IO.Combo.Input("duration", options=[5, 10]), + IO.Int.Input("duration", default=5, min=3, max=15, display_mode=IO.NumberDisplay.slider), IO.Combo.Input("resolution", options=["1080p", "720p"], optional=True), + IO.DynamicCombo.Input( + "storyboards", + options=[ + IO.DynamicCombo.Option("disabled", []), + IO.DynamicCombo.Option("1 storyboard", _generate_storyboard_inputs(1)), + IO.DynamicCombo.Option("2 storyboards", _generate_storyboard_inputs(2)), + IO.DynamicCombo.Option("3 storyboards", _generate_storyboard_inputs(3)), + IO.DynamicCombo.Option("4 storyboards", _generate_storyboard_inputs(4)), + IO.DynamicCombo.Option("5 storyboards", _generate_storyboard_inputs(5)), + IO.DynamicCombo.Option("6 storyboards", _generate_storyboard_inputs(6)), + ], + tooltip="Generate a series of video segments with individual prompts and durations. " + "Ignored for o1 model.", + optional=True, + ), + IO.Boolean.Input("generate_audio", default=False, optional=True), + IO.Int.Input( + "seed", + default=0, + min=0, + max=2147483647, + display_mode=IO.NumberDisplay.number, + control_after_generate=True, + tooltip="Seed controls whether the node should re-run; " + "results are non-deterministic regardless of seed.", + optional=True, + ), ], outputs=[ IO.Video.Output(), @@ -845,11 +899,15 @@ class OmniProTextToVideoNode(IO.ComfyNode): ], is_api_node=True, price_badge=IO.PriceBadge( - depends_on=IO.PriceBadgeDepends(widgets=["duration", "resolution"]), + depends_on=IO.PriceBadgeDepends(widgets=["duration", "resolution", "model_name", "generate_audio"]), expr=""" ( $mode := (widgets.resolution = "720p") ? "std" : "pro"; - $rates := {"std": 0.084, "pro": 0.112}; + $isV3 := $contains(widgets.model_name, "v3"); + $audio := $isV3 and widgets.generate_audio; + $rates := $audio + ? {"std": 0.112, "pro": 0.14} + : {"std": 0.084, "pro": 0.112}; {"type":"usd","usd": $lookup($rates, $mode) * widgets.duration} ) """, @@ -864,8 +922,45 @@ class OmniProTextToVideoNode(IO.ComfyNode): aspect_ratio: str, duration: int, resolution: str = "1080p", + storyboards: dict | None = None, + generate_audio: bool = False, + seed: int = 0, ) -> IO.NodeOutput: - validate_string(prompt, min_length=1, max_length=2500) + _ = seed + if model_name == "kling-video-o1": + if duration not in (5, 10): + raise ValueError("kling-video-o1 only supports durations of 5 or 10 seconds.") + if generate_audio: + raise ValueError("kling-video-o1 does not support audio generation.") + stories_enabled = storyboards is not None and storyboards["storyboards"] != "disabled" + if stories_enabled and model_name == "kling-video-o1": + raise ValueError("kling-video-o1 does not support storyboards.") + validate_string(prompt, strip_whitespace=True, min_length=0 if stories_enabled else 1, max_length=2500) + + multi_shot = None + multi_prompt_list = None + if stories_enabled: + count = int(storyboards["storyboards"].split()[0]) + multi_shot = True + multi_prompt_list = [] + for i in range(1, count + 1): + sb_prompt = storyboards[f"storyboard_{i}_prompt"] + sb_duration = storyboards[f"storyboard_{i}_duration"] + validate_string(sb_prompt, field_name=f"storyboard_{i}_prompt", min_length=1, max_length=512) + multi_prompt_list.append( + MultiPromptEntry( + index=i, + prompt=sb_prompt, + duration=str(sb_duration), + ) + ) + total_storyboard_duration = sum(int(e.duration) for e in multi_prompt_list) + if total_storyboard_duration != duration: + raise ValueError( + f"Total storyboard duration ({total_storyboard_duration}s) " + f"must equal the global duration ({duration}s)." + ) + response = await sync_op( cls, ApiEndpoint(path="/proxy/kling/v1/videos/omni-video", method="POST"), @@ -876,6 +971,10 @@ class OmniProTextToVideoNode(IO.ComfyNode): aspect_ratio=aspect_ratio, duration=str(duration), mode="pro" if resolution == "1080p" else "std", + multi_shot=multi_shot, + multi_prompt=multi_prompt_list, + shot_type="customize" if multi_shot else None, + sound="on" if generate_audio else "off", ), ) return await finish_omni_video_task(cls, response) @@ -887,24 +986,26 @@ class OmniProFirstLastFrameNode(IO.ComfyNode): def define_schema(cls) -> IO.Schema: return IO.Schema( node_id="KlingOmniProFirstLastFrameNode", - display_name="Kling Omni First-Last-Frame to Video (Pro)", + display_name="Kling 3.0 Omni First-Last-Frame to Video", category="api node/video/Kling", description="Use a start frame, an optional end frame, or reference images with the latest Kling model.", inputs=[ - IO.Combo.Input("model_name", options=["kling-video-o1"]), + IO.Combo.Input("model_name", options=["kling-v3-omni", "kling-video-o1"]), IO.String.Input( "prompt", multiline=True, tooltip="A text prompt describing the video content. " - "This can include both positive and negative descriptions.", + "This can include both positive and negative descriptions. " + "Ignored when storyboards are enabled.", ), - IO.Int.Input("duration", default=5, min=3, max=10, display_mode=IO.NumberDisplay.slider), + IO.Int.Input("duration", default=5, min=3, max=15, display_mode=IO.NumberDisplay.slider), IO.Image.Input("first_frame"), IO.Image.Input( "end_frame", optional=True, tooltip="An optional end frame for the video. " - "This cannot be used simultaneously with 'reference_images'.", + "This cannot be used simultaneously with 'reference_images'. " + "Does not work with storyboards.", ), IO.Image.Input( "reference_images", @@ -912,6 +1013,38 @@ class OmniProFirstLastFrameNode(IO.ComfyNode): tooltip="Up to 6 additional reference images.", ), IO.Combo.Input("resolution", options=["1080p", "720p"], optional=True), + IO.DynamicCombo.Input( + "storyboards", + options=[ + IO.DynamicCombo.Option("disabled", []), + IO.DynamicCombo.Option("1 storyboard", _generate_storyboard_inputs(1)), + IO.DynamicCombo.Option("2 storyboards", _generate_storyboard_inputs(2)), + IO.DynamicCombo.Option("3 storyboards", _generate_storyboard_inputs(3)), + IO.DynamicCombo.Option("4 storyboards", _generate_storyboard_inputs(4)), + IO.DynamicCombo.Option("5 storyboards", _generate_storyboard_inputs(5)), + IO.DynamicCombo.Option("6 storyboards", _generate_storyboard_inputs(6)), + ], + tooltip="Generate a series of video segments with individual prompts and durations. " + "Only supported for kling-v3-omni.", + optional=True, + ), + IO.Boolean.Input( + "generate_audio", + default=False, + optional=True, + tooltip="Generate audio for the video. Only supported for kling-v3-omni.", + ), + IO.Int.Input( + "seed", + default=0, + min=0, + max=2147483647, + display_mode=IO.NumberDisplay.number, + control_after_generate=True, + tooltip="Seed controls whether the node should re-run; " + "results are non-deterministic regardless of seed.", + optional=True, + ), ], outputs=[ IO.Video.Output(), @@ -923,11 +1056,15 @@ class OmniProFirstLastFrameNode(IO.ComfyNode): ], is_api_node=True, price_badge=IO.PriceBadge( - depends_on=IO.PriceBadgeDepends(widgets=["duration", "resolution"]), + depends_on=IO.PriceBadgeDepends(widgets=["duration", "resolution", "model_name", "generate_audio"]), expr=""" ( $mode := (widgets.resolution = "720p") ? "std" : "pro"; - $rates := {"std": 0.084, "pro": 0.112}; + $isV3 := $contains(widgets.model_name, "v3"); + $audio := $isV3 and widgets.generate_audio; + $rates := $audio + ? {"std": 0.112, "pro": 0.14} + : {"std": 0.084, "pro": 0.112}; {"type":"usd","usd": $lookup($rates, $mode) * widgets.duration} ) """, @@ -944,15 +1081,59 @@ class OmniProFirstLastFrameNode(IO.ComfyNode): end_frame: Input.Image | None = None, reference_images: Input.Image | None = None, resolution: str = "1080p", + storyboards: dict | None = None, + generate_audio: bool = False, + seed: int = 0, ) -> IO.NodeOutput: + _ = seed + if model_name == "kling-video-o1": + if duration > 10: + raise ValueError("kling-video-o1 does not support durations greater than 10 seconds.") + if generate_audio: + raise ValueError("kling-video-o1 does not support audio generation.") + stories_enabled = storyboards is not None and storyboards["storyboards"] != "disabled" + if stories_enabled and model_name == "kling-video-o1": + raise ValueError("kling-video-o1 does not support storyboards.") prompt = normalize_omni_prompt_references(prompt) - validate_string(prompt, min_length=1, max_length=2500) + validate_string(prompt, strip_whitespace=True, min_length=0 if stories_enabled else 1, max_length=2500) if end_frame is not None and reference_images is not None: raise ValueError("The 'end_frame' input cannot be used simultaneously with 'reference_images'.") - if duration not in (5, 10) and end_frame is None and reference_images is None: + if end_frame is not None and stories_enabled: + raise ValueError("The 'end_frame' input cannot be used simultaneously with storyboards.") + if ( + model_name == "kling-video-o1" + and duration not in (5, 10) + and end_frame is None + and reference_images is None + ): raise ValueError( "Duration is only supported for 5 or 10 seconds if there is no end frame or reference images." ) + + multi_shot = None + multi_prompt_list = None + if stories_enabled: + count = int(storyboards["storyboards"].split()[0]) + multi_shot = True + multi_prompt_list = [] + for i in range(1, count + 1): + sb_prompt = storyboards[f"storyboard_{i}_prompt"] + sb_duration = storyboards[f"storyboard_{i}_duration"] + validate_string(sb_prompt, field_name=f"storyboard_{i}_prompt", min_length=1, max_length=512) + multi_prompt_list.append( + MultiPromptEntry( + index=i, + prompt=sb_prompt, + duration=str(sb_duration), + ) + ) + total_storyboard_duration = sum(int(e.duration) for e in multi_prompt_list) + if total_storyboard_duration != duration: + raise ValueError( + f"Total storyboard duration ({total_storyboard_duration}s) " + f"must equal the global duration ({duration}s)." + ) + validate_image_dimensions(first_frame, min_width=300, min_height=300) validate_image_aspect_ratio(first_frame, (1, 2.5), (2.5, 1)) image_list: list[OmniParamImage] = [ @@ -988,6 +1169,10 @@ class OmniProFirstLastFrameNode(IO.ComfyNode): duration=str(duration), image_list=image_list, mode="pro" if resolution == "1080p" else "std", + sound="on" if generate_audio else "off", + multi_shot=multi_shot, + multi_prompt=multi_prompt_list, + shot_type="customize" if multi_shot else None, ), ) return await finish_omni_video_task(cls, response) @@ -999,24 +1184,57 @@ class OmniProImageToVideoNode(IO.ComfyNode): def define_schema(cls) -> IO.Schema: return IO.Schema( node_id="KlingOmniProImageToVideoNode", - display_name="Kling Omni Image to Video (Pro)", + display_name="Kling 3.0 Omni Image to Video", category="api node/video/Kling", description="Use up to 7 reference images to generate a video with the latest Kling model.", inputs=[ - IO.Combo.Input("model_name", options=["kling-video-o1"]), + IO.Combo.Input("model_name", options=["kling-v3-omni", "kling-video-o1"]), IO.String.Input( "prompt", multiline=True, tooltip="A text prompt describing the video content. " - "This can include both positive and negative descriptions.", + "This can include both positive and negative descriptions. " + "Ignored when storyboards are enabled.", ), IO.Combo.Input("aspect_ratio", options=["16:9", "9:16", "1:1"]), - IO.Int.Input("duration", default=3, min=3, max=10, display_mode=IO.NumberDisplay.slider), + IO.Int.Input("duration", default=5, min=3, max=15, display_mode=IO.NumberDisplay.slider), IO.Image.Input( "reference_images", tooltip="Up to 7 reference images.", ), IO.Combo.Input("resolution", options=["1080p", "720p"], optional=True), + IO.DynamicCombo.Input( + "storyboards", + options=[ + IO.DynamicCombo.Option("disabled", []), + IO.DynamicCombo.Option("1 storyboard", _generate_storyboard_inputs(1)), + IO.DynamicCombo.Option("2 storyboards", _generate_storyboard_inputs(2)), + IO.DynamicCombo.Option("3 storyboards", _generate_storyboard_inputs(3)), + IO.DynamicCombo.Option("4 storyboards", _generate_storyboard_inputs(4)), + IO.DynamicCombo.Option("5 storyboards", _generate_storyboard_inputs(5)), + IO.DynamicCombo.Option("6 storyboards", _generate_storyboard_inputs(6)), + ], + tooltip="Generate a series of video segments with individual prompts and durations. " + "Only supported for kling-v3-omni.", + optional=True, + ), + IO.Boolean.Input( + "generate_audio", + default=False, + optional=True, + tooltip="Generate audio for the video. Only supported for kling-v3-omni.", + ), + IO.Int.Input( + "seed", + default=0, + min=0, + max=2147483647, + display_mode=IO.NumberDisplay.number, + control_after_generate=True, + tooltip="Seed controls whether the node should re-run; " + "results are non-deterministic regardless of seed.", + optional=True, + ), ], outputs=[ IO.Video.Output(), @@ -1028,11 +1246,15 @@ class OmniProImageToVideoNode(IO.ComfyNode): ], is_api_node=True, price_badge=IO.PriceBadge( - depends_on=IO.PriceBadgeDepends(widgets=["duration", "resolution"]), + depends_on=IO.PriceBadgeDepends(widgets=["duration", "resolution", "model_name", "generate_audio"]), expr=""" ( $mode := (widgets.resolution = "720p") ? "std" : "pro"; - $rates := {"std": 0.084, "pro": 0.112}; + $isV3 := $contains(widgets.model_name, "v3"); + $audio := $isV3 and widgets.generate_audio; + $rates := $audio + ? {"std": 0.112, "pro": 0.14} + : {"std": 0.084, "pro": 0.112}; {"type":"usd","usd": $lookup($rates, $mode) * widgets.duration} ) """, @@ -1048,9 +1270,46 @@ class OmniProImageToVideoNode(IO.ComfyNode): duration: int, reference_images: Input.Image, resolution: str = "1080p", + storyboards: dict | None = None, + generate_audio: bool = False, + seed: int = 0, ) -> IO.NodeOutput: + _ = seed + if model_name == "kling-video-o1": + if duration > 10: + raise ValueError("kling-video-o1 does not support durations greater than 10 seconds.") + if generate_audio: + raise ValueError("kling-video-o1 does not support audio generation.") + stories_enabled = storyboards is not None and storyboards["storyboards"] != "disabled" + if stories_enabled and model_name == "kling-video-o1": + raise ValueError("kling-video-o1 does not support storyboards.") prompt = normalize_omni_prompt_references(prompt) - validate_string(prompt, min_length=1, max_length=2500) + validate_string(prompt, strip_whitespace=True, min_length=0 if stories_enabled else 1, max_length=2500) + + multi_shot = None + multi_prompt_list = None + if stories_enabled: + count = int(storyboards["storyboards"].split()[0]) + multi_shot = True + multi_prompt_list = [] + for i in range(1, count + 1): + sb_prompt = storyboards[f"storyboard_{i}_prompt"] + sb_duration = storyboards[f"storyboard_{i}_duration"] + validate_string(sb_prompt, field_name=f"storyboard_{i}_prompt", min_length=1, max_length=512) + multi_prompt_list.append( + MultiPromptEntry( + index=i, + prompt=sb_prompt, + duration=str(sb_duration), + ) + ) + total_storyboard_duration = sum(int(e.duration) for e in multi_prompt_list) + if total_storyboard_duration != duration: + raise ValueError( + f"Total storyboard duration ({total_storyboard_duration}s) " + f"must equal the global duration ({duration}s)." + ) + if get_number_of_images(reference_images) > 7: raise ValueError("The maximum number of reference images is 7.") for i in reference_images: @@ -1070,6 +1329,10 @@ class OmniProImageToVideoNode(IO.ComfyNode): duration=str(duration), image_list=image_list, mode="pro" if resolution == "1080p" else "std", + sound="on" if generate_audio else "off", + multi_shot=multi_shot, + multi_prompt=multi_prompt_list, + shot_type="customize" if multi_shot else None, ), ) return await finish_omni_video_task(cls, response) @@ -1081,11 +1344,11 @@ class OmniProVideoToVideoNode(IO.ComfyNode): def define_schema(cls) -> IO.Schema: return IO.Schema( node_id="KlingOmniProVideoToVideoNode", - display_name="Kling Omni Video to Video (Pro)", + display_name="Kling 3.0 Omni Video to Video", category="api node/video/Kling", description="Use a video and up to 4 reference images to generate a video with the latest Kling model.", inputs=[ - IO.Combo.Input("model_name", options=["kling-video-o1"]), + IO.Combo.Input("model_name", options=["kling-v3-omni", "kling-video-o1"]), IO.String.Input( "prompt", multiline=True, @@ -1102,6 +1365,17 @@ class OmniProVideoToVideoNode(IO.ComfyNode): optional=True, ), IO.Combo.Input("resolution", options=["1080p", "720p"], optional=True), + IO.Int.Input( + "seed", + default=0, + min=0, + max=2147483647, + display_mode=IO.NumberDisplay.number, + control_after_generate=True, + tooltip="Seed controls whether the node should re-run; " + "results are non-deterministic regardless of seed.", + optional=True, + ), ], outputs=[ IO.Video.Output(), @@ -1135,7 +1409,9 @@ class OmniProVideoToVideoNode(IO.ComfyNode): keep_original_sound: bool, reference_images: Input.Image | None = None, resolution: str = "1080p", + seed: int = 0, ) -> IO.NodeOutput: + _ = seed prompt = normalize_omni_prompt_references(prompt) validate_string(prompt, min_length=1, max_length=2500) validate_video_duration(reference_video, min_duration=3.0, max_duration=10.05) @@ -1179,11 +1455,11 @@ class OmniProEditVideoNode(IO.ComfyNode): def define_schema(cls) -> IO.Schema: return IO.Schema( node_id="KlingOmniProEditVideoNode", - display_name="Kling Omni Edit Video (Pro)", + display_name="Kling 3.0 Omni Edit Video", category="api node/video/Kling", description="Edit an existing video with the latest model from Kling.", inputs=[ - IO.Combo.Input("model_name", options=["kling-video-o1"]), + IO.Combo.Input("model_name", options=["kling-v3-omni", "kling-video-o1"]), IO.String.Input( "prompt", multiline=True, @@ -1198,6 +1474,17 @@ class OmniProEditVideoNode(IO.ComfyNode): optional=True, ), IO.Combo.Input("resolution", options=["1080p", "720p"], optional=True), + IO.Int.Input( + "seed", + default=0, + min=0, + max=2147483647, + display_mode=IO.NumberDisplay.number, + control_after_generate=True, + tooltip="Seed controls whether the node should re-run; " + "results are non-deterministic regardless of seed.", + optional=True, + ), ], outputs=[ IO.Video.Output(), @@ -1229,7 +1516,9 @@ class OmniProEditVideoNode(IO.ComfyNode): keep_original_sound: bool, reference_images: Input.Image | None = None, resolution: str = "1080p", + seed: int = 0, ) -> IO.NodeOutput: + _ = seed prompt = normalize_omni_prompt_references(prompt) validate_string(prompt, min_length=1, max_length=2500) validate_video_duration(video, min_duration=3.0, max_duration=10.05) @@ -1273,27 +1562,43 @@ class OmniProImageNode(IO.ComfyNode): def define_schema(cls) -> IO.Schema: return IO.Schema( node_id="KlingOmniProImageNode", - display_name="Kling Omni Image (Pro)", + display_name="Kling 3.0 Omni Image", category="api node/image/Kling", description="Create or edit images with the latest model from Kling.", inputs=[ - IO.Combo.Input("model_name", options=["kling-image-o1"]), + IO.Combo.Input("model_name", options=["kling-v3-omni", "kling-image-o1"]), IO.String.Input( "prompt", multiline=True, tooltip="A text prompt describing the image content. " "This can include both positive and negative descriptions.", ), - IO.Combo.Input("resolution", options=["1K", "2K"]), + IO.Combo.Input("resolution", options=["1K", "2K", "4K"]), IO.Combo.Input( "aspect_ratio", options=["16:9", "9:16", "1:1", "4:3", "3:4", "3:2", "2:3", "21:9"], ), + IO.Combo.Input( + "series_amount", + options=["disabled", "2", "3", "4", "5", "6", "7", "8", "9"], + tooltip="Generate a series of images. Not supported for kling-image-o1.", + ), IO.Image.Input( "reference_images", tooltip="Up to 10 additional reference images.", optional=True, ), + IO.Int.Input( + "seed", + default=0, + min=0, + max=2147483647, + display_mode=IO.NumberDisplay.number, + control_after_generate=True, + tooltip="Seed controls whether the node should re-run; " + "results are non-deterministic regardless of seed.", + optional=True, + ), ], outputs=[ IO.Image.Output(), @@ -1305,7 +1610,16 @@ class OmniProImageNode(IO.ComfyNode): ], is_api_node=True, price_badge=IO.PriceBadge( - expr="""{"type":"usd","usd":0.028}""", + depends_on=IO.PriceBadgeDepends(widgets=["resolution", "series_amount", "model_name"]), + expr=""" + ( + $prices := {"1k": 0.028, "2k": 0.028, "4k": 0.056}; + $base := $lookup($prices, widgets.resolution); + $isO1 := widgets.model_name = "kling-image-o1"; + $mult := ($isO1 or widgets.series_amount = "disabled") ? 1 : $number(widgets.series_amount); + {"type":"usd","usd": $base * $mult} + ) + """, ), ) @@ -1316,8 +1630,13 @@ class OmniProImageNode(IO.ComfyNode): prompt: str, resolution: str, aspect_ratio: str, + series_amount: str = "disabled", reference_images: Input.Image | None = None, + seed: int = 0, ) -> IO.NodeOutput: + _ = seed + if model_name == "kling-image-o1" and resolution == "4K": + raise ValueError("4K resolution is not supported for kling-image-o1 model.") prompt = normalize_omni_prompt_references(prompt) validate_string(prompt, min_length=1, max_length=2500) image_list: list[OmniImageParamImage] = [] @@ -1329,6 +1648,9 @@ class OmniProImageNode(IO.ComfyNode): validate_image_aspect_ratio(i, (1, 2.5), (2.5, 1)) for i in await upload_images_to_comfyapi(cls, reference_images, wait_label="Uploading reference image"): image_list.append(OmniImageParamImage(image=i)) + use_series = series_amount != "disabled" + if use_series and model_name == "kling-image-o1": + raise ValueError("kling-image-o1 does not support series generation.") response = await sync_op( cls, ApiEndpoint(path="/proxy/kling/v1/images/omni-image", method="POST"), @@ -1339,6 +1661,8 @@ class OmniProImageNode(IO.ComfyNode): resolution=resolution.lower(), aspect_ratio=aspect_ratio, image_list=image_list if image_list else None, + result_type="series" if use_series else None, + series_amount=int(series_amount) if use_series else None, ), ) if response.code: @@ -1351,7 +1675,9 @@ class OmniProImageNode(IO.ComfyNode): response_model=TaskStatusResponse, status_extractor=lambda r: (r.data.task_status if r.data else None), ) - return IO.NodeOutput(await download_url_to_image_tensor(final_response.data.task_result.images[0].url)) + images = final_response.data.task_result.series_images or final_response.data.task_result.images + tensors = [await download_url_to_image_tensor(img.url) for img in images] + return IO.NodeOutput(torch.cat(tensors, dim=0)) class KlingCameraControlT2VNode(IO.ComfyNode): @@ -2119,7 +2445,7 @@ class KlingImageGenerationNode(IO.ComfyNode): def define_schema(cls) -> IO.Schema: return IO.Schema( node_id="KlingImageGenerationNode", - display_name="Kling Image Generation", + display_name="Kling 3.0 Image", category="api node/image/Kling", description="Kling Image Generation Node. Generate an image from a text prompt with an optional reference image.", inputs=[ @@ -2147,11 +2473,7 @@ class KlingImageGenerationNode(IO.ComfyNode): display_mode=IO.NumberDisplay.slider, tooltip="Subject reference similarity", ), - IO.Combo.Input( - "model_name", - options=[i.value for i in KlingImageGenModelName], - default="kling-v2", - ), + IO.Combo.Input("model_name", options=["kling-v3", "kling-v2", "kling-v1-5"]), IO.Combo.Input( "aspect_ratio", options=[i.value for i in KlingImageGenAspectRatio], @@ -2165,6 +2487,17 @@ class KlingImageGenerationNode(IO.ComfyNode): tooltip="Number of generated images", ), IO.Image.Input("image", optional=True), + IO.Int.Input( + "seed", + default=0, + min=0, + max=2147483647, + display_mode=IO.NumberDisplay.number, + control_after_generate=True, + tooltip="Seed controls whether the node should re-run; " + "results are non-deterministic regardless of seed.", + optional=True, + ), ], outputs=[ IO.Image.Output(), @@ -2183,7 +2516,7 @@ class KlingImageGenerationNode(IO.ComfyNode): $base := $contains($m,"kling-v1-5") ? (inputs.image.connected ? 0.028 : 0.014) - : ($contains($m,"kling-v1") ? 0.0035 : 0.014); + : $contains($m,"kling-v3") ? 0.028 : 0.014; {"type":"usd","usd": $base * widgets.n} ) """, @@ -2193,7 +2526,7 @@ class KlingImageGenerationNode(IO.ComfyNode): @classmethod async def execute( cls, - model_name: KlingImageGenModelName, + model_name: str, prompt: str, negative_prompt: str, image_type: KlingImageGenImageReferenceType, @@ -2202,17 +2535,11 @@ class KlingImageGenerationNode(IO.ComfyNode): n: int, aspect_ratio: KlingImageGenAspectRatio, image: torch.Tensor | None = None, + seed: int = 0, ) -> IO.NodeOutput: + _ = seed validate_string(prompt, field_name="prompt", min_length=1, max_length=MAX_PROMPT_LENGTH_IMAGE_GEN) validate_string(negative_prompt, field_name="negative_prompt", max_length=MAX_PROMPT_LENGTH_IMAGE_GEN) - - if image is None: - image_type = None - elif model_name == KlingImageGenModelName.kling_v1: - raise ValueError(f"The model {KlingImageGenModelName.kling_v1.value} does not support reference images.") - else: - image = tensor_to_base64_string(image) - task_creation_response = await sync_op( cls, ApiEndpoint(path=PATH_IMAGE_GENERATIONS, method="POST"), @@ -2221,8 +2548,8 @@ class KlingImageGenerationNode(IO.ComfyNode): model_name=model_name, prompt=prompt, negative_prompt=negative_prompt, - image=image, - image_reference=image_type, + image=tensor_to_base64_string(image) if image is not None else None, + image_reference=image_type if image is not None else None, image_fidelity=image_fidelity, human_fidelity=human_fidelity, n=n, @@ -2252,7 +2579,7 @@ class TextToVideoWithAudio(IO.ComfyNode): def define_schema(cls) -> IO.Schema: return IO.Schema( node_id="KlingTextToVideoWithAudio", - display_name="Kling Text to Video with Audio", + display_name="Kling 2.6 Text to Video with Audio", category="api node/video/Kling", inputs=[ IO.Combo.Input("model_name", options=["kling-v2-6"]), @@ -2320,7 +2647,7 @@ class ImageToVideoWithAudio(IO.ComfyNode): def define_schema(cls) -> IO.Schema: return IO.Schema( node_id="KlingImageToVideoWithAudio", - display_name="Kling Image(First Frame) to Video with Audio", + display_name="Kling 2.6 Image(First Frame) to Video with Audio", category="api node/video/Kling", inputs=[ IO.Combo.Input("model_name", options=["kling-v2-6"]), @@ -2478,6 +2805,335 @@ class MotionControl(IO.ComfyNode): return IO.NodeOutput(await download_url_to_video_output(final_response.data.task_result.videos[0].url)) +class KlingVideoNode(IO.ComfyNode): + + @classmethod + def define_schema(cls) -> IO.Schema: + return IO.Schema( + node_id="KlingVideoNode", + display_name="Kling 3.0 Video", + category="api node/video/Kling", + description="Generate videos with Kling V3. " + "Supports text-to-video and image-to-video with optional storyboard multi-prompt and audio generation.", + inputs=[ + IO.DynamicCombo.Input( + "multi_shot", + options=[ + IO.DynamicCombo.Option( + "disabled", + [ + IO.String.Input("prompt", multiline=True, default=""), + IO.String.Input("negative_prompt", multiline=True, default=""), + IO.Int.Input( + "duration", + default=5, + min=3, + max=15, + display_mode=IO.NumberDisplay.slider, + ), + ], + ), + IO.DynamicCombo.Option("1 storyboard", _generate_storyboard_inputs(1)), + IO.DynamicCombo.Option("2 storyboards", _generate_storyboard_inputs(2)), + IO.DynamicCombo.Option("3 storyboards", _generate_storyboard_inputs(3)), + IO.DynamicCombo.Option("4 storyboards", _generate_storyboard_inputs(4)), + IO.DynamicCombo.Option("5 storyboards", _generate_storyboard_inputs(5)), + IO.DynamicCombo.Option("6 storyboards", _generate_storyboard_inputs(6)), + ], + tooltip="Generate a series of video segments with individual prompts and durations.", + ), + IO.Boolean.Input("generate_audio", default=True), + IO.DynamicCombo.Input( + "model", + options=[ + IO.DynamicCombo.Option( + "kling-v3", + [ + IO.Combo.Input("resolution", options=["1080p", "720p"]), + IO.Combo.Input( + "aspect_ratio", + options=["16:9", "9:16", "1:1"], + tooltip="Ignored in image-to-video mode.", + ), + ], + ), + ], + tooltip="Model and generation settings.", + ), + IO.Int.Input( + "seed", + default=0, + min=0, + max=2147483647, + display_mode=IO.NumberDisplay.number, + control_after_generate=True, + tooltip="Seed controls whether the node should re-run; " + "results are non-deterministic regardless of seed.", + ), + IO.Image.Input( + "start_frame", + optional=True, + tooltip="Optional start frame image. When connected, switches to image-to-video mode.", + ), + ], + outputs=[ + IO.Video.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends( + widgets=[ + "model.resolution", + "generate_audio", + "multi_shot", + "multi_shot.duration", + "multi_shot.storyboard_1_duration", + "multi_shot.storyboard_2_duration", + "multi_shot.storyboard_3_duration", + "multi_shot.storyboard_4_duration", + "multi_shot.storyboard_5_duration", + "multi_shot.storyboard_6_duration", + ], + ), + expr=""" + ( + $rates := {"1080p": {"off": 0.112, "on": 0.168}, "720p": {"off": 0.084, "on": 0.126}}; + $res := $lookup(widgets, "model.resolution"); + $audio := widgets.generate_audio ? "on" : "off"; + $rate := $lookup($lookup($rates, $res), $audio); + $ms := widgets.multi_shot; + $isSb := $ms != "disabled"; + $n := $isSb ? $number($substring($ms, 0, 1)) : 0; + $d1 := $lookup(widgets, "multi_shot.storyboard_1_duration"); + $d2 := $n >= 2 ? $lookup(widgets, "multi_shot.storyboard_2_duration") : 0; + $d3 := $n >= 3 ? $lookup(widgets, "multi_shot.storyboard_3_duration") : 0; + $d4 := $n >= 4 ? $lookup(widgets, "multi_shot.storyboard_4_duration") : 0; + $d5 := $n >= 5 ? $lookup(widgets, "multi_shot.storyboard_5_duration") : 0; + $d6 := $n >= 6 ? $lookup(widgets, "multi_shot.storyboard_6_duration") : 0; + $dur := $isSb ? $d1 + $d2 + $d3 + $d4 + $d5 + $d6 : $lookup(widgets, "multi_shot.duration"); + {"type":"usd","usd": $rate * $dur} + ) + """, + ), + ) + + @classmethod + async def execute( + cls, + multi_shot: dict, + generate_audio: bool, + model: dict, + seed: int, + start_frame: Input.Image | None = None, + ) -> IO.NodeOutput: + _ = seed + mode = "pro" if model["resolution"] == "1080p" else "std" + custom_multi_shot = False + if multi_shot["multi_shot"] == "disabled": + shot_type = None + else: + shot_type = "customize" + custom_multi_shot = True + + multi_prompt_list = None + if shot_type == "customize": + count = int(multi_shot["multi_shot"].split()[0]) + multi_prompt_list = [] + for i in range(1, count + 1): + sb_prompt = multi_shot[f"storyboard_{i}_prompt"] + sb_duration = multi_shot[f"storyboard_{i}_duration"] + validate_string(sb_prompt, field_name=f"storyboard_{i}_prompt", min_length=1, max_length=512) + multi_prompt_list.append( + MultiPromptEntry( + index=i, + prompt=sb_prompt, + duration=str(sb_duration), + ) + ) + duration = sum(int(e.duration) for e in multi_prompt_list) + if duration < 3 or duration > 15: + raise ValueError( + f"Total storyboard duration ({duration}s) must be between 3 and 15 seconds." + ) + else: + duration = multi_shot["duration"] + validate_string(multi_shot["prompt"], min_length=1, max_length=2500) + + if start_frame is not None: + validate_image_dimensions(start_frame, min_width=300, min_height=300) + validate_image_aspect_ratio(start_frame, (1, 2.5), (2.5, 1)) + image_url = await upload_image_to_comfyapi(cls, start_frame, wait_label="Uploading start frame") + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/kling/v1/videos/image2video", method="POST"), + response_model=TaskStatusResponse, + data=ImageToVideoWithAudioRequest( + model_name=model["model"], + image=image_url, + prompt=None if custom_multi_shot else multi_shot["prompt"], + negative_prompt=None if custom_multi_shot else multi_shot["negative_prompt"], + mode=mode, + duration=str(duration), + sound="on" if generate_audio else "off", + multi_shot=True if shot_type else None, + multi_prompt=multi_prompt_list, + shot_type=shot_type, + ), + ) + poll_path = f"/proxy/kling/v1/videos/image2video/{response.data.task_id}" + else: + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/kling/v1/videos/text2video", method="POST"), + response_model=TaskStatusResponse, + data=TextToVideoWithAudioRequest( + model_name=model["model"], + aspect_ratio=model["aspect_ratio"], + prompt=None if custom_multi_shot else multi_shot["prompt"], + negative_prompt=None if custom_multi_shot else multi_shot["negative_prompt"], + mode=mode, + duration=str(duration), + sound="on" if generate_audio else "off", + multi_shot=True if shot_type else None, + multi_prompt=multi_prompt_list, + shot_type=shot_type, + ), + ) + poll_path = f"/proxy/kling/v1/videos/text2video/{response.data.task_id}" + + if response.code: + raise RuntimeError( + f"Kling request failed. Code: {response.code}, Message: {response.message}, Data: {response.data}" + ) + final_response = await poll_op( + cls, + ApiEndpoint(path=poll_path), + response_model=TaskStatusResponse, + status_extractor=lambda r: (r.data.task_status if r.data else None), + ) + return IO.NodeOutput(await download_url_to_video_output(final_response.data.task_result.videos[0].url)) + + +class KlingFirstLastFrameNode(IO.ComfyNode): + + @classmethod + def define_schema(cls) -> IO.Schema: + return IO.Schema( + node_id="KlingFirstLastFrameNode", + display_name="Kling 3.0 First-Last-Frame to Video", + category="api node/video/Kling", + description="Generate videos with Kling V3 using first and last frames.", + inputs=[ + IO.String.Input("prompt", multiline=True, default=""), + IO.Int.Input( + "duration", + default=5, + min=3, + max=15, + display_mode=IO.NumberDisplay.slider, + ), + IO.Image.Input("first_frame"), + IO.Image.Input("end_frame"), + IO.Boolean.Input("generate_audio", default=True), + IO.DynamicCombo.Input( + "model", + options=[ + IO.DynamicCombo.Option( + "kling-v3", + [ + IO.Combo.Input("resolution", options=["1080p", "720p"]), + ], + ), + ], + tooltip="Model and generation settings.", + ), + IO.Int.Input( + "seed", + default=0, + min=0, + max=2147483647, + display_mode=IO.NumberDisplay.number, + control_after_generate=True, + tooltip="Seed controls whether the node should re-run; " + "results are non-deterministic regardless of seed.", + ), + ], + outputs=[ + IO.Video.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + price_badge=IO.PriceBadge( + depends_on=IO.PriceBadgeDepends( + widgets=["model.resolution", "generate_audio", "duration"], + ), + expr=""" + ( + $rates := {"1080p": {"off": 0.112, "on": 0.168}, "720p": {"off": 0.084, "on": 0.126}}; + $res := $lookup(widgets, "model.resolution"); + $audio := widgets.generate_audio ? "on" : "off"; + $rate := $lookup($lookup($rates, $res), $audio); + {"type":"usd","usd": $rate * widgets.duration} + ) + """, + ), + ) + + @classmethod + async def execute( + cls, + prompt: str, + duration: int, + first_frame: Input.Image, + end_frame: Input.Image, + generate_audio: bool, + model: dict, + seed: int, + ) -> IO.NodeOutput: + _ = seed + validate_string(prompt, min_length=1, max_length=2500) + validate_image_dimensions(first_frame, min_width=300, min_height=300) + validate_image_aspect_ratio(first_frame, (1, 2.5), (2.5, 1)) + validate_image_dimensions(end_frame, min_width=300, min_height=300) + validate_image_aspect_ratio(end_frame, (1, 2.5), (2.5, 1)) + image_url = await upload_image_to_comfyapi(cls, first_frame, wait_label="Uploading first frame") + image_tail_url = await upload_image_to_comfyapi(cls, end_frame, wait_label="Uploading end frame") + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/kling/v1/videos/image2video", method="POST"), + response_model=TaskStatusResponse, + data=ImageToVideoWithAudioRequest( + model_name=model["model"], + image=image_url, + image_tail=image_tail_url, + prompt=prompt, + mode="pro" if model["resolution"] == "1080p" else "std", + duration=str(duration), + sound="on" if generate_audio else "off", + ), + ) + if response.code: + raise RuntimeError( + f"Kling request failed. Code: {response.code}, Message: {response.message}, Data: {response.data}" + ) + final_response = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/kling/v1/videos/image2video/{response.data.task_id}"), + response_model=TaskStatusResponse, + status_extractor=lambda r: (r.data.task_status if r.data else None), + ) + return IO.NodeOutput(await download_url_to_video_output(final_response.data.task_result.videos[0].url)) + + class KlingExtension(ComfyExtension): @override async def get_node_list(self) -> list[type[IO.ComfyNode]]: @@ -2504,6 +3160,8 @@ class KlingExtension(ComfyExtension): TextToVideoWithAudio, ImageToVideoWithAudio, MotionControl, + KlingVideoNode, + KlingFirstLastFrameNode, ] From 6615db925c9f84843e29db118852e14b643a1a03 Mon Sep 17 00:00:00 2001 From: ComfyUI Wiki Date: Wed, 11 Feb 2026 02:24:56 +0800 Subject: [PATCH 605/647] chore: update workflow templates to v0.8.38 (#12394) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 4e2773f5d..7de6a413c 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,5 +1,5 @@ comfyui-frontend-package==1.38.13 -comfyui-workflow-templates==0.8.37 +comfyui-workflow-templates==0.8.38 comfyui-embedded-docs==0.4.1 torch torchsde From 6648ab68bc934a185c90a2a872c87dc64d093751 Mon Sep 17 00:00:00 2001 From: comfyanonymous Date: Tue, 10 Feb 2026 13:26:29 -0500 Subject: [PATCH 606/647] ComfyUI v0.13.0 --- comfyui_version.py | 2 +- pyproject.toml | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/comfyui_version.py b/comfyui_version.py index 706b37763..cf4e89816 100644 --- a/comfyui_version.py +++ b/comfyui_version.py @@ -1,3 +1,3 @@ # This file is automatically generated by the build process when version is # updated in pyproject.toml. -__version__ = "0.12.3" +__version__ = "0.13.0" diff --git a/pyproject.toml b/pyproject.toml index f7925b92a..9dab9a50c 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -1,6 +1,6 @@ [project] name = "ComfyUI" -version = "0.12.3" +version = "0.13.0" readme = "README.md" license = { file = "LICENSE" } requires-python = ">=3.10" From fe053ba5eb34c8abcc5d17a25c114340af1833aa Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Tue, 10 Feb 2026 10:37:17 -0800 Subject: [PATCH 607/647] mp: dont deep-clone objects from model_options (#12382) If there are non-trivial python objects nested in the model_options, this causes all sorts of issues. Traverse lists and dicts so clones can safely overide settings and BYO objects but stop there on the deepclone. --- comfy/model_patcher.py | 3 +-- comfy/utils.py | 18 ++++++++++++++++++ 2 files changed, 19 insertions(+), 2 deletions(-) diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index b9a117a7c..19c9031ea 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -19,7 +19,6 @@ from __future__ import annotations import collections -import copy import inspect import logging import math @@ -317,7 +316,7 @@ class ModelPatcher: n.object_patches = self.object_patches.copy() n.weight_wrapper_patches = self.weight_wrapper_patches.copy() - n.model_options = copy.deepcopy(self.model_options) + n.model_options = comfy.utils.deepcopy_list_dict(self.model_options) n.backup = self.backup n.object_patches_backup = self.object_patches_backup n.parent = self diff --git a/comfy/utils.py b/comfy/utils.py index 1337e2205..edd80cebe 100644 --- a/comfy/utils.py +++ b/comfy/utils.py @@ -1376,3 +1376,21 @@ def string_to_seed(data): else: crc >>= 1 return crc ^ 0xFFFFFFFF + +def deepcopy_list_dict(obj, memo=None): + if memo is None: + memo = {} + + obj_id = id(obj) + if obj_id in memo: + return memo[obj_id] + + if isinstance(obj, dict): + res = {deepcopy_list_dict(k, memo): deepcopy_list_dict(v, memo) for k, v in obj.items()} + elif isinstance(obj, list): + res = [deepcopy_list_dict(i, memo) for i in obj] + else: + res = obj + + memo[obj_id] = res + return res From f719f9c06266e7944683009b403e995d4c61d5f0 Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Tue, 10 Feb 2026 10:37:46 -0800 Subject: [PATCH 608/647] sd: delay VAE dtype archive until after override (#12388) VAEs have host specific dtype logic that should override the dynamic _model_dtype. Defer the archiving of model dtypes until after. --- comfy/sd.py | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-) diff --git a/comfy/sd.py b/comfy/sd.py index bc9407405..f65e7cadd 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -793,8 +793,6 @@ class VAE: self.first_stage_model = AutoencoderKL(**(config['params'])) self.first_stage_model = self.first_stage_model.eval() - model_management.archive_model_dtypes(self.first_stage_model) - if device is None: device = model_management.vae_device() self.device = device @@ -803,6 +801,7 @@ class VAE: dtype = model_management.vae_dtype(self.device, self.working_dtypes) self.vae_dtype = dtype self.first_stage_model.to(self.vae_dtype) + model_management.archive_model_dtypes(self.first_stage_model) self.output_device = model_management.intermediate_device() mp = comfy.model_patcher.CoreModelPatcher From 123a7874a97c4a8b8f06d4b7c2b1a566b8f0d057 Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Tue, 10 Feb 2026 10:38:28 -0800 Subject: [PATCH 609/647] ops: Fix vanilla-fp8 loaded lora quality (#12390) This was missing the stochastic rounding required for fp8 downcast to be consistent with model_patcher.patch_weight_to_device. Missed in testing as I spend too much time with quantized tensors and overlooked the simpler ones. --- comfy/ops.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/comfy/ops.py b/comfy/ops.py index ea0d70702..33803b223 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -169,8 +169,8 @@ def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compu if orig.dtype == dtype and len(fns) == 0: #The layer actually wants our freshly saved QT x = y - else: - y = x + elif update_weight: + y = comfy.float.stochastic_rounding(x, orig.dtype, seed = comfy.utils.string_to_seed(s.seed_key)) if update_weight: orig.copy_(y) for f in fns: From 00fff6019ecf0f4306005579e93cef0cd51a3a1c Mon Sep 17 00:00:00 2001 From: guill Date: Tue, 10 Feb 2026 14:37:14 -0800 Subject: [PATCH 610/647] feat(jobs): add 3d to PREVIEWABLE_MEDIA_TYPES for first-class 3D output support (#12381) Co-authored-by: Jedrzej Kosinski --- comfy_execution/jobs.py | 79 +++++++++++-- tests/execution/test_jobs.py | 208 ++++++++++++++++++++++++++++++++++- 2 files changed, 271 insertions(+), 16 deletions(-) diff --git a/comfy_execution/jobs.py b/comfy_execution/jobs.py index bf091a448..370014fb6 100644 --- a/comfy_execution/jobs.py +++ b/comfy_execution/jobs.py @@ -20,10 +20,60 @@ class JobStatus: # Media types that can be previewed in the frontend -PREVIEWABLE_MEDIA_TYPES = frozenset({'images', 'video', 'audio'}) +PREVIEWABLE_MEDIA_TYPES = frozenset({'images', 'video', 'audio', '3d'}) # 3D file extensions for preview fallback (no dedicated media_type exists) -THREE_D_EXTENSIONS = frozenset({'.obj', '.fbx', '.gltf', '.glb'}) +THREE_D_EXTENSIONS = frozenset({'.obj', '.fbx', '.gltf', '.glb', '.usdz'}) + + +def has_3d_extension(filename: str) -> bool: + lower = filename.lower() + return any(lower.endswith(ext) for ext in THREE_D_EXTENSIONS) + + +def normalize_output_item(item): + """Normalize a single output list item for the jobs API. + + Returns the normalized item, or None to exclude it. + String items with 3D extensions become {filename, type, subfolder} dicts. + """ + if item is None: + return None + if isinstance(item, str): + if has_3d_extension(item): + return {'filename': item, 'type': 'output', 'subfolder': '', 'mediaType': '3d'} + return None + if isinstance(item, dict): + return item + return None + + +def normalize_outputs(outputs: dict) -> dict: + """Normalize raw node outputs for the jobs API. + + Transforms string 3D filenames into file output dicts and removes + None items. All other items (non-3D strings, dicts, etc.) are + preserved as-is. + """ + normalized = {} + for node_id, node_outputs in outputs.items(): + if not isinstance(node_outputs, dict): + normalized[node_id] = node_outputs + continue + normalized_node = {} + for media_type, items in node_outputs.items(): + if media_type == 'animated' or not isinstance(items, list): + normalized_node[media_type] = items + continue + normalized_items = [] + for item in items: + if item is None: + continue + norm = normalize_output_item(item) + normalized_items.append(norm if norm is not None else item) + normalized_node[media_type] = normalized_items + normalized[node_id] = normalized_node + return normalized def _extract_job_metadata(extra_data: dict) -> tuple[Optional[int], Optional[str]]: @@ -45,9 +95,9 @@ def is_previewable(media_type: str, item: dict) -> bool: Maintains backwards compatibility with existing logic. Priority: - 1. media_type is 'images', 'video', or 'audio' + 1. media_type is 'images', 'video', 'audio', or '3d' 2. format field starts with 'video/' or 'audio/' - 3. filename has a 3D extension (.obj, .fbx, .gltf, .glb) + 3. filename has a 3D extension (.obj, .fbx, .gltf, .glb, .usdz) """ if media_type in PREVIEWABLE_MEDIA_TYPES: return True @@ -139,7 +189,7 @@ def normalize_history_item(prompt_id: str, history_item: dict, include_outputs: }) if include_outputs: - job['outputs'] = outputs + job['outputs'] = normalize_outputs(outputs) job['execution_status'] = status_info job['workflow'] = { 'prompt': prompt, @@ -171,18 +221,23 @@ def get_outputs_summary(outputs: dict) -> tuple[int, Optional[dict]]: continue for item in items: - count += 1 - - if not isinstance(item, dict): + normalized = normalize_output_item(item) + if normalized is None: continue - if preview_output is None and is_previewable(media_type, item): + count += 1 + + if preview_output is not None: + continue + + if isinstance(normalized, dict) and is_previewable(media_type, normalized): enriched = { - **item, + **normalized, 'nodeId': node_id, - 'mediaType': media_type } - if item.get('type') == 'output': + if 'mediaType' not in normalized: + enriched['mediaType'] = media_type + if normalized.get('type') == 'output': preview_output = enriched elif fallback_preview is None: fallback_preview = enriched diff --git a/tests/execution/test_jobs.py b/tests/execution/test_jobs.py index 4d2f9ed36..83c36fe48 100644 --- a/tests/execution/test_jobs.py +++ b/tests/execution/test_jobs.py @@ -5,8 +5,11 @@ from comfy_execution.jobs import ( is_previewable, normalize_queue_item, normalize_history_item, + normalize_output_item, + normalize_outputs, get_outputs_summary, apply_sorting, + has_3d_extension, ) @@ -35,8 +38,8 @@ class TestIsPreviewable: """Unit tests for is_previewable()""" def test_previewable_media_types(self): - """Images, video, audio media types should be previewable.""" - for media_type in ['images', 'video', 'audio']: + """Images, video, audio, 3d media types should be previewable.""" + for media_type in ['images', 'video', 'audio', '3d']: assert is_previewable(media_type, {}) is True def test_non_previewable_media_types(self): @@ -46,7 +49,7 @@ class TestIsPreviewable: def test_3d_extensions_previewable(self): """3D file extensions should be previewable regardless of media_type.""" - for ext in ['.obj', '.fbx', '.gltf', '.glb']: + for ext in ['.obj', '.fbx', '.gltf', '.glb', '.usdz']: item = {'filename': f'model{ext}'} assert is_previewable('files', item) is True @@ -160,7 +163,7 @@ class TestGetOutputsSummary: def test_3d_files_previewable(self): """3D file extensions should be previewable.""" - for ext in ['.obj', '.fbx', '.gltf', '.glb']: + for ext in ['.obj', '.fbx', '.gltf', '.glb', '.usdz']: outputs = { 'node1': { 'files': [{'filename': f'model{ext}', 'type': 'output'}] @@ -192,6 +195,64 @@ class TestGetOutputsSummary: assert preview['mediaType'] == 'images' assert preview['subfolder'] == 'outputs' + def test_string_3d_filename_creates_preview(self): + """String items with 3D extensions should synthesize a preview (Preview3D node output). + Only the .glb counts — nulls and non-file strings are excluded.""" + outputs = { + 'node1': { + 'result': ['preview3d_abc123.glb', None, None] + } + } + count, preview = get_outputs_summary(outputs) + assert count == 1 + assert preview is not None + assert preview['filename'] == 'preview3d_abc123.glb' + assert preview['mediaType'] == '3d' + assert preview['nodeId'] == 'node1' + assert preview['type'] == 'output' + + def test_string_non_3d_filename_no_preview(self): + """String items without 3D extensions should not create a preview.""" + outputs = { + 'node1': { + 'result': ['data.json', None] + } + } + count, preview = get_outputs_summary(outputs) + assert count == 0 + assert preview is None + + def test_string_3d_filename_used_as_fallback(self): + """String 3D preview should be used when no dict items are previewable.""" + outputs = { + 'node1': { + 'latents': [{'filename': 'latent.safetensors'}], + }, + 'node2': { + 'result': ['model.glb', None] + } + } + count, preview = get_outputs_summary(outputs) + assert preview is not None + assert preview['filename'] == 'model.glb' + assert preview['mediaType'] == '3d' + + +class TestHas3DExtension: + """Unit tests for has_3d_extension()""" + + def test_recognized_extensions(self): + for ext in ['.obj', '.fbx', '.gltf', '.glb', '.usdz']: + assert has_3d_extension(f'model{ext}') is True + + def test_case_insensitive(self): + assert has_3d_extension('MODEL.GLB') is True + assert has_3d_extension('Scene.GLTF') is True + + def test_non_3d_extensions(self): + for name in ['photo.png', 'video.mp4', 'data.json', 'model']: + assert has_3d_extension(name) is False + class TestApplySorting: """Unit tests for apply_sorting()""" @@ -395,3 +456,142 @@ class TestNormalizeHistoryItem: 'prompt': {'nodes': {'1': {}}}, 'extra_data': {'create_time': 1234567890, 'client_id': 'abc'}, } + + def test_include_outputs_normalizes_3d_strings(self): + """Detail view should transform string 3D filenames into file output dicts.""" + history_item = { + 'prompt': ( + 5, + 'prompt-3d', + {'nodes': {}}, + {'create_time': 1234567890}, + ['node1'], + ), + 'status': {'status_str': 'success', 'completed': True, 'messages': []}, + 'outputs': { + 'node1': { + 'result': ['preview3d_abc123.glb', None, None] + } + }, + } + job = normalize_history_item('prompt-3d', history_item, include_outputs=True) + + assert job['outputs_count'] == 1 + result_items = job['outputs']['node1']['result'] + assert len(result_items) == 1 + assert result_items[0] == { + 'filename': 'preview3d_abc123.glb', + 'type': 'output', + 'subfolder': '', + 'mediaType': '3d', + } + + def test_include_outputs_preserves_dict_items(self): + """Detail view normalization should pass dict items through unchanged.""" + history_item = { + 'prompt': ( + 5, + 'prompt-img', + {'nodes': {}}, + {'create_time': 1234567890}, + ['node1'], + ), + 'status': {'status_str': 'success', 'completed': True, 'messages': []}, + 'outputs': { + 'node1': { + 'images': [ + {'filename': 'photo.png', 'type': 'output', 'subfolder': ''}, + ] + } + }, + } + job = normalize_history_item('prompt-img', history_item, include_outputs=True) + + assert job['outputs_count'] == 1 + assert job['outputs']['node1']['images'] == [ + {'filename': 'photo.png', 'type': 'output', 'subfolder': ''}, + ] + + +class TestNormalizeOutputItem: + """Unit tests for normalize_output_item()""" + + def test_none_returns_none(self): + assert normalize_output_item(None) is None + + def test_string_3d_extension_synthesizes_dict(self): + result = normalize_output_item('model.glb') + assert result == {'filename': 'model.glb', 'type': 'output', 'subfolder': '', 'mediaType': '3d'} + + def test_string_non_3d_extension_returns_none(self): + assert normalize_output_item('data.json') is None + + def test_string_no_extension_returns_none(self): + assert normalize_output_item('camera_info_string') is None + + def test_dict_passes_through(self): + item = {'filename': 'test.png', 'type': 'output'} + assert normalize_output_item(item) is item + + def test_other_types_return_none(self): + assert normalize_output_item(42) is None + assert normalize_output_item(True) is None + + +class TestNormalizeOutputs: + """Unit tests for normalize_outputs()""" + + def test_empty_outputs(self): + assert normalize_outputs({}) == {} + + def test_dict_items_pass_through(self): + outputs = { + 'node1': { + 'images': [{'filename': 'a.png', 'type': 'output'}], + } + } + result = normalize_outputs(outputs) + assert result == outputs + + def test_3d_string_synthesized(self): + outputs = { + 'node1': { + 'result': ['model.glb', None, None], + } + } + result = normalize_outputs(outputs) + assert result == { + 'node1': { + 'result': [ + {'filename': 'model.glb', 'type': 'output', 'subfolder': '', 'mediaType': '3d'}, + ], + } + } + + def test_animated_key_preserved(self): + outputs = { + 'node1': { + 'images': [{'filename': 'a.png', 'type': 'output'}], + 'animated': [True], + } + } + result = normalize_outputs(outputs) + assert result['node1']['animated'] == [True] + + def test_non_dict_node_outputs_preserved(self): + outputs = {'node1': 'unexpected_value'} + result = normalize_outputs(outputs) + assert result == {'node1': 'unexpected_value'} + + def test_none_items_filtered_but_other_types_preserved(self): + outputs = { + 'node1': { + 'result': ['data.json', None, [1, 2, 3]], + } + } + result = normalize_outputs(outputs) + assert result == { + 'node1': { + 'result': ['data.json', [1, 2, 3]], + } + } From dbe70b6821994ce92d9cf211cc685862d0b6c0ca Mon Sep 17 00:00:00 2001 From: AustinMroz Date: Tue, 10 Feb 2026 14:42:21 -0800 Subject: [PATCH 611/647] Add a VideoSlice node (#12107) * Base TrimVideo implementation * Raise error if as_trimmed call fails * Bigger max start_time, tooltips, and formatting * Count packets unless codec has subframes * Remove incorrect nested decode * Add null check for audio streams * Support non-strict duration * Added strict_duration bool to node definition * Empty commit for approval * Fix duration * Support 5.1 audio layout on save --------- Co-authored-by: Jedrzej Kosinski --- comfy_api/latest/_input/video_types.py | 15 ++ comfy_api/latest/_input_impl/video_types.py | 201 ++++++++++++++------ comfy_extras/nodes_video.py | 51 +++++ 3 files changed, 207 insertions(+), 60 deletions(-) diff --git a/comfy_api/latest/_input/video_types.py b/comfy_api/latest/_input/video_types.py index e634a0311..451e9526e 100644 --- a/comfy_api/latest/_input/video_types.py +++ b/comfy_api/latest/_input/video_types.py @@ -34,6 +34,21 @@ class VideoInput(ABC): """ pass + @abstractmethod + def as_trimmed( + self, + start_time: float | None = None, + duration: float | None = None, + strict_duration: bool = False, + ) -> VideoInput | None: + """ + Create a new VideoInput which is trimmed to have the corresponding start_time and duration + + Returns: + A new VideoInput, or None if the result would have negative duration + """ + pass + def get_stream_source(self) -> Union[str, io.BytesIO]: """ Get a streamable source for the video. This allows processing without diff --git a/comfy_api/latest/_input_impl/video_types.py b/comfy_api/latest/_input_impl/video_types.py index 1405d0b81..3463ed1c9 100644 --- a/comfy_api/latest/_input_impl/video_types.py +++ b/comfy_api/latest/_input_impl/video_types.py @@ -6,6 +6,7 @@ from typing import Optional from .._input import AudioInput, VideoInput import av import io +import itertools import json import numpy as np import math @@ -29,7 +30,6 @@ def container_to_output_format(container_format: str | None) -> str | None: formats = container_format.split(",") return formats[0] - def get_open_write_kwargs( dest: str | io.BytesIO, container_format: str, to_format: str | None ) -> dict: @@ -57,12 +57,14 @@ class VideoFromFile(VideoInput): Class representing video input from a file. """ - def __init__(self, file: str | io.BytesIO): + def __init__(self, file: str | io.BytesIO, *, start_time: float=0, duration: float=0): """ Initialize the VideoFromFile object based off of either a path on disk or a BytesIO object containing the file contents. """ self.__file = file + self.__start_time = start_time + self.__duration = duration def get_stream_source(self) -> str | io.BytesIO: """ @@ -96,6 +98,16 @@ class VideoFromFile(VideoInput): Returns: Duration in seconds """ + raw_duration = self._get_raw_duration() + if self.__start_time < 0: + duration_from_start = min(raw_duration, -self.__start_time) + else: + duration_from_start = raw_duration - self.__start_time + if self.__duration: + return min(self.__duration, duration_from_start) + return duration_from_start + + def _get_raw_duration(self) -> float: if isinstance(self.__file, io.BytesIO): self.__file.seek(0) with av.open(self.__file, mode="r") as container: @@ -113,9 +125,13 @@ class VideoFromFile(VideoInput): if video_stream and video_stream.average_rate: frame_count = 0 container.seek(0) - for packet in container.demux(video_stream): - for _ in packet.decode(): - frame_count += 1 + frame_iterator = ( + container.decode(video_stream) + if video_stream.codec.capabilities & 0x100 + else container.demux(video_stream) + ) + for packet in frame_iterator: + frame_count += 1 if frame_count > 0: return float(frame_count / video_stream.average_rate) @@ -131,36 +147,54 @@ class VideoFromFile(VideoInput): with av.open(self.__file, mode="r") as container: video_stream = self._get_first_video_stream(container) - # 1. Prefer the frames field if available - if video_stream.frames and video_stream.frames > 0: + # 1. Prefer the frames field if available and usable + if ( + video_stream.frames + and video_stream.frames > 0 + and not self.__start_time + and not self.__duration + ): return int(video_stream.frames) # 2. Try to estimate from duration and average_rate using only metadata - if container.duration is not None and video_stream.average_rate: - duration_seconds = float(container.duration / av.time_base) - estimated_frames = int(round(duration_seconds * float(video_stream.average_rate))) - if estimated_frames > 0: - return estimated_frames - if ( getattr(video_stream, "duration", None) is not None and getattr(video_stream, "time_base", None) is not None and video_stream.average_rate ): - duration_seconds = float(video_stream.duration * video_stream.time_base) + raw_duration = float(video_stream.duration * video_stream.time_base) + if self.__start_time < 0: + duration_from_start = min(raw_duration, -self.__start_time) + else: + duration_from_start = raw_duration - self.__start_time + duration_seconds = min(self.__duration, duration_from_start) estimated_frames = int(round(duration_seconds * float(video_stream.average_rate))) if estimated_frames > 0: return estimated_frames # 3. Last resort: decode frames and count them (streaming) - frame_count = 0 - container.seek(0) - for packet in container.demux(video_stream): - for _ in packet.decode(): - frame_count += 1 - - if frame_count == 0: - raise ValueError(f"Could not determine frame count for file '{self.__file}'") + if self.__start_time < 0: + start_time = max(self._get_raw_duration() + self.__start_time, 0) + else: + start_time = self.__start_time + frame_count = 1 + start_pts = int(start_time / video_stream.time_base) + end_pts = int((start_time + self.__duration) / video_stream.time_base) + container.seek(start_pts, stream=video_stream) + frame_iterator = ( + container.decode(video_stream) + if video_stream.codec.capabilities & 0x100 + else container.demux(video_stream) + ) + for frame in frame_iterator: + if frame.pts >= start_pts: + break + else: + raise ValueError(f"Could not determine frame count for file '{self.__file}'\nNo frames exist for start_time {self.__start_time}") + for frame in frame_iterator: + if frame.pts >= end_pts: + break + frame_count += 1 return frame_count def get_frame_rate(self) -> Fraction: @@ -199,9 +233,21 @@ class VideoFromFile(VideoInput): return container.format.name def get_components_internal(self, container: InputContainer) -> VideoComponents: + video_stream = self._get_first_video_stream(container) + if self.__start_time < 0: + start_time = max(self._get_raw_duration() + self.__start_time, 0) + else: + start_time = self.__start_time # Get video frames frames = [] - for frame in container.decode(video=0): + start_pts = int(start_time / video_stream.time_base) + end_pts = int((start_time + self.__duration) / video_stream.time_base) + container.seek(start_pts, stream=video_stream) + for frame in container.decode(video_stream): + if frame.pts < start_pts: + continue + if self.__duration and frame.pts >= end_pts: + break img = frame.to_ndarray(format='rgb24') # shape: (H, W, 3) img = torch.from_numpy(img) / 255.0 # shape: (H, W, 3) frames.append(img) @@ -209,31 +255,44 @@ class VideoFromFile(VideoInput): images = torch.stack(frames) if len(frames) > 0 else torch.zeros(0, 3, 0, 0) # Get frame rate - video_stream = next(s for s in container.streams if s.type == 'video') - frame_rate = Fraction(video_stream.average_rate) if video_stream and video_stream.average_rate else Fraction(1) + frame_rate = Fraction(video_stream.average_rate) if video_stream.average_rate else Fraction(1) # Get audio if available audio = None - try: - container.seek(0) # Reset the container to the beginning - for stream in container.streams: - if stream.type != 'audio': - continue - assert isinstance(stream, av.AudioStream) - audio_frames = [] - for packet in container.demux(stream): - for frame in packet.decode(): - assert isinstance(frame, av.AudioFrame) - audio_frames.append(frame.to_ndarray()) # shape: (channels, samples) - if len(audio_frames) > 0: - audio_data = np.concatenate(audio_frames, axis=1) # shape: (channels, total_samples) - audio_tensor = torch.from_numpy(audio_data).unsqueeze(0) # shape: (1, channels, total_samples) - audio = AudioInput({ - "waveform": audio_tensor, - "sample_rate": int(stream.sample_rate) if stream.sample_rate else 1, - }) - except StopIteration: - pass # No audio stream + container.seek(start_pts, stream=video_stream) + # Use last stream for consistency + if len(container.streams.audio): + audio_stream = container.streams.audio[-1] + audio_frames = [] + resample = av.audio.resampler.AudioResampler(format='fltp').resample + frames = itertools.chain.from_iterable( + map(resample, container.decode(audio_stream)) + ) + + has_first_frame = False + for frame in frames: + offset_seconds = start_time - frame.pts * audio_stream.time_base + to_skip = int(offset_seconds * audio_stream.sample_rate) + if to_skip < frame.samples: + has_first_frame = True + break + if has_first_frame: + audio_frames.append(frame.to_ndarray()[..., to_skip:]) + + for frame in frames: + if frame.time > start_time + self.__duration: + break + audio_frames.append(frame.to_ndarray()) # shape: (channels, samples) + if len(audio_frames) > 0: + audio_data = np.concatenate(audio_frames, axis=1) # shape: (channels, total_samples) + if self.__duration: + audio_data = audio_data[..., :int(self.__duration * audio_stream.sample_rate)] + + audio_tensor = torch.from_numpy(audio_data).unsqueeze(0) # shape: (1, channels, total_samples) + audio = AudioInput({ + "waveform": audio_tensor, + "sample_rate": int(audio_stream.sample_rate) if audio_stream.sample_rate else 1, + }) metadata = container.metadata return VideoComponents(images=images, audio=audio, frame_rate=frame_rate, metadata=metadata) @@ -250,7 +309,7 @@ class VideoFromFile(VideoInput): path: str | io.BytesIO, format: VideoContainer = VideoContainer.AUTO, codec: VideoCodec = VideoCodec.AUTO, - metadata: Optional[dict] = None + metadata: Optional[dict] = None, ): if isinstance(self.__file, io.BytesIO): self.__file.seek(0) # Reset the BytesIO object to the beginning @@ -262,15 +321,14 @@ class VideoFromFile(VideoInput): reuse_streams = False if codec != VideoCodec.AUTO and codec != video_encoding and video_encoding is not None: reuse_streams = False + if self.__start_time or self.__duration: + reuse_streams = False if not reuse_streams: components = self.get_components_internal(container) video = VideoFromComponents(components) return video.save_to( - path, - format=format, - codec=codec, - metadata=metadata + path, format=format, codec=codec, metadata=metadata ) streams = container.streams @@ -304,10 +362,21 @@ class VideoFromFile(VideoInput): output_container.mux(packet) def _get_first_video_stream(self, container: InputContainer): - video_stream = next((s for s in container.streams if s.type == "video"), None) - if video_stream is None: - raise ValueError(f"No video stream found in file '{self.__file}'") - return video_stream + if len(container.streams.video): + return container.streams.video[0] + raise ValueError(f"No video stream found in file '{self.__file}'") + + def as_trimmed( + self, start_time: float = 0, duration: float = 0, strict_duration: bool = True + ) -> VideoInput | None: + trimmed = VideoFromFile( + self.get_stream_source(), + start_time=start_time + self.__start_time, + duration=duration, + ) + if trimmed.get_duration() < duration and strict_duration: + return None + return trimmed class VideoFromComponents(VideoInput): @@ -322,7 +391,7 @@ class VideoFromComponents(VideoInput): return VideoComponents( images=self.__components.images, audio=self.__components.audio, - frame_rate=self.__components.frame_rate + frame_rate=self.__components.frame_rate, ) def save_to( @@ -330,7 +399,7 @@ class VideoFromComponents(VideoInput): path: str, format: VideoContainer = VideoContainer.AUTO, codec: VideoCodec = VideoCodec.AUTO, - metadata: Optional[dict] = None + metadata: Optional[dict] = None, ): if format != VideoContainer.AUTO and format != VideoContainer.MP4: raise ValueError("Only MP4 format is supported for now") @@ -357,7 +426,10 @@ class VideoFromComponents(VideoInput): audio_stream: Optional[av.AudioStream] = None if self.__components.audio: audio_sample_rate = int(self.__components.audio['sample_rate']) - audio_stream = output.add_stream('aac', rate=audio_sample_rate) + waveform = self.__components.audio['waveform'] + waveform = waveform[0, :, :math.ceil((audio_sample_rate / frame_rate) * self.__components.images.shape[0])] + layout = {1: 'mono', 2: 'stereo', 6: '5.1'}.get(waveform.shape[0], 'stereo') + audio_stream = output.add_stream('aac', rate=audio_sample_rate, layout=layout) # Encode video for i, frame in enumerate(self.__components.images): @@ -372,12 +444,21 @@ class VideoFromComponents(VideoInput): output.mux(packet) if audio_stream and self.__components.audio: - waveform = self.__components.audio['waveform'] - waveform = waveform[:, :, :math.ceil((audio_sample_rate / frame_rate) * self.__components.images.shape[0])] - frame = av.AudioFrame.from_ndarray(waveform.movedim(2, 1).reshape(1, -1).float().cpu().numpy(), format='flt', layout='mono' if waveform.shape[1] == 1 else 'stereo') + frame = av.AudioFrame.from_ndarray(waveform.float().cpu().numpy(), format='fltp', layout=layout) frame.sample_rate = audio_sample_rate frame.pts = 0 output.mux(audio_stream.encode(frame)) # Flush encoder output.mux(audio_stream.encode(None)) + + def as_trimmed( + self, + start_time: float | None = None, + duration: float | None = None, + strict_duration: bool = True, + ) -> VideoInput | None: + if self.get_duration() < start_time + duration: + return None + #TODO Consider tracking duration and trimming at time of save? + return VideoFromFile(self.get_stream_source(), start_time=start_time, duration=duration) diff --git a/comfy_extras/nodes_video.py b/comfy_extras/nodes_video.py index ccf7b63d3..cd765a7c1 100644 --- a/comfy_extras/nodes_video.py +++ b/comfy_extras/nodes_video.py @@ -202,6 +202,56 @@ class LoadVideo(io.ComfyNode): return True +class VideoSlice(io.ComfyNode): + @classmethod + def define_schema(cls): + return io.Schema( + node_id="Video Slice", + display_name="Video Slice", + search_aliases=[ + "trim video duration", + "skip first frames", + "frame load cap", + "start time", + ], + category="image/video", + inputs=[ + io.Video.Input("video"), + io.Float.Input( + "start_time", + default=0.0, + max=1e5, + min=-1e5, + step=0.001, + tooltip="Start time in seconds", + ), + io.Float.Input( + "duration", + default=0.0, + min=0.0, + step=0.001, + tooltip="Duration in seconds, or 0 for unlimited duration", + ), + io.Boolean.Input( + "strict_duration", + default=False, + tooltip="If True, when the specified duration is not possible, an error will be raised.", + ), + ], + outputs=[ + io.Video.Output(), + ], + ) + + @classmethod + def execute(cls, video: io.Video.Type, start_time: float, duration: float, strict_duration: bool) -> io.NodeOutput: + trimmed = video.as_trimmed(start_time, duration, strict_duration=strict_duration) + if trimmed is not None: + return io.NodeOutput(trimmed) + raise ValueError( + f"Failed to slice video:\nSource duration: {video.get_duration()}\nStart time: {start_time}\nTarget duration: {duration}" + ) + class VideoExtension(ComfyExtension): @override @@ -212,6 +262,7 @@ class VideoExtension(ComfyExtension): CreateVideo, GetVideoComponents, LoadVideo, + VideoSlice, ] async def comfy_entrypoint() -> VideoExtension: From cdcf4119b3e826bd69fa986772485fb5b44a54cd Mon Sep 17 00:00:00 2001 From: Kohaku-Blueleaf <59680068+KohakuBlueleaf@users.noreply.github.com> Date: Wed, 11 Feb 2026 10:45:19 +0800 Subject: [PATCH 612/647] [Trainer] training with proper offloading (#12189) * Fix bypass dtype/device moving * Force offloading mode for training * training context var * offloading implementation in training node * fix wrong input type * Support bypass load lora model, correct adapter/offloading handling --- comfy/ldm/flux/math.py | 39 +++++--- comfy/model_management.py | 5 + comfy/sampler_helpers.py | 16 +++- comfy/weight_adapter/bypass.py | 20 ++-- comfy_extras/nodes_train.py | 162 ++++++++++++++++++++++++++++----- 5 files changed, 196 insertions(+), 46 deletions(-) diff --git a/comfy/ldm/flux/math.py b/comfy/ldm/flux/math.py index f9597de5b..5e764bb46 100644 --- a/comfy/ldm/flux/math.py +++ b/comfy/ldm/flux/math.py @@ -29,19 +29,34 @@ def rope(pos: Tensor, dim: int, theta: int) -> Tensor: return out.to(dtype=torch.float32, device=pos.device) +def _apply_rope1(x: Tensor, freqs_cis: Tensor): + x_ = x.to(dtype=freqs_cis.dtype).reshape(*x.shape[:-1], -1, 1, 2) + + x_out = freqs_cis[..., 0] * x_[..., 0] + x_out.addcmul_(freqs_cis[..., 1], x_[..., 1]) + + return x_out.reshape(*x.shape).type_as(x) + + +def _apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor): + return apply_rope1(xq, freqs_cis), apply_rope1(xk, freqs_cis) + + try: import comfy.quant_ops - apply_rope = comfy.quant_ops.ck.apply_rope - apply_rope1 = comfy.quant_ops.ck.apply_rope1 + q_apply_rope = comfy.quant_ops.ck.apply_rope + q_apply_rope1 = comfy.quant_ops.ck.apply_rope1 + def apply_rope(xq, xk, freqs_cis): + if comfy.model_management.in_training: + return _apply_rope(xq, xk, freqs_cis) + else: + return apply_rope1(xq, freqs_cis), apply_rope1(xk, freqs_cis) + def apply_rope1(x, freqs_cis): + if comfy.model_management.in_training: + return _apply_rope1(x, freqs_cis) + else: + return q_apply_rope1(x, freqs_cis) except: logging.warning("No comfy kitchen, using old apply_rope functions.") - def apply_rope1(x: Tensor, freqs_cis: Tensor): - x_ = x.to(dtype=freqs_cis.dtype).reshape(*x.shape[:-1], -1, 1, 2) - - x_out = freqs_cis[..., 0] * x_[..., 0] - x_out.addcmul_(freqs_cis[..., 1], x_[..., 1]) - - return x_out.reshape(*x.shape).type_as(x) - - def apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor): - return apply_rope1(xq, freqs_cis), apply_rope1(xk, freqs_cis) + apply_rope = _apply_rope + apply_rope1 = _apply_rope1 diff --git a/comfy/model_management.py b/comfy/model_management.py index 6018c1ab6..304931eb0 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -55,6 +55,11 @@ cpu_state = CPUState.GPU total_vram = 0 + +# Training Related State +in_training = False + + def get_supported_float8_types(): float8_types = [] try: diff --git a/comfy/sampler_helpers.py b/comfy/sampler_helpers.py index 9134e6d71..1f75f2ba7 100644 --- a/comfy/sampler_helpers.py +++ b/comfy/sampler_helpers.py @@ -122,20 +122,26 @@ def estimate_memory(model, noise_shape, conds): minimum_memory_required = model.model.memory_required([noise_shape[0]] + list(noise_shape[1:]), cond_shapes=cond_shapes_min) return memory_required, minimum_memory_required -def prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None, force_full_load=False): +def prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None, force_full_load=False, force_offload=False): executor = comfy.patcher_extension.WrapperExecutor.new_executor( _prepare_sampling, comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.PREPARE_SAMPLING, model_options, is_model_options=True) ) - return executor.execute(model, noise_shape, conds, model_options=model_options, force_full_load=force_full_load) + return executor.execute(model, noise_shape, conds, model_options=model_options, force_full_load=force_full_load, force_offload=force_offload) -def _prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None, force_full_load=False): +def _prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None, force_full_load=False, force_offload=False): real_model: BaseModel = None models, inference_memory = get_additional_models(conds, model.model_dtype()) models += get_additional_models_from_model_options(model_options) models += model.get_nested_additional_models() # TODO: does this require inference_memory update? - memory_required, minimum_memory_required = estimate_memory(model, noise_shape, conds) - comfy.model_management.load_models_gpu([model] + models, memory_required=memory_required + inference_memory, minimum_memory_required=minimum_memory_required + inference_memory, force_full_load=force_full_load) + if force_offload: # In training + offload enabled, we want to force prepare sampling to trigger partial load + memory_required = 1e20 + minimum_memory_required = None + else: + memory_required, minimum_memory_required = estimate_memory(model, noise_shape, conds) + memory_required += inference_memory + minimum_memory_required += inference_memory + comfy.model_management.load_models_gpu([model] + models, memory_required=memory_required, minimum_memory_required=minimum_memory_required, force_full_load=force_full_load) real_model = model.model return real_model, conds, models diff --git a/comfy/weight_adapter/bypass.py b/comfy/weight_adapter/bypass.py index d4aaf98ca..b9d5ec7d9 100644 --- a/comfy/weight_adapter/bypass.py +++ b/comfy/weight_adapter/bypass.py @@ -21,6 +21,7 @@ from typing import Optional, Union import torch import torch.nn as nn +import comfy.model_management from .base import WeightAdapterBase, WeightAdapterTrainBase from comfy.patcher_extension import PatcherInjection @@ -181,18 +182,21 @@ class BypassForwardHook: ) return # Already injected - # Move adapter weights to module's device to avoid CPU-GPU transfer on every forward - device = None + # Move adapter weights to compute device (GPU) + # Use get_torch_device() instead of module.weight.device because + # with offloading, module weights may be on CPU while compute happens on GPU + device = comfy.model_management.get_torch_device() + + # Get dtype from module weight if available dtype = None if hasattr(self.module, "weight") and self.module.weight is not None: - device = self.module.weight.device dtype = self.module.weight.dtype - elif hasattr(self.module, "W_q"): # Quantized layers might use different attr - device = self.module.W_q.device - dtype = self.module.W_q.dtype - if device is not None: - self._move_adapter_weights_to_device(device, dtype) + # Only use dtype if it's a standard float type, not quantized + if dtype is not None and dtype not in (torch.float32, torch.float16, torch.bfloat16): + dtype = None + + self._move_adapter_weights_to_device(device, dtype) self.original_forward = self.module.forward self.module.forward = self._bypass_forward diff --git a/comfy_extras/nodes_train.py b/comfy_extras/nodes_train.py index 024a89391..630eedc9f 100644 --- a/comfy_extras/nodes_train.py +++ b/comfy_extras/nodes_train.py @@ -4,6 +4,7 @@ import os import numpy as np import safetensors import torch +import torch.nn as nn import torch.utils.checkpoint from tqdm.auto import trange from PIL import Image, ImageDraw, ImageFont @@ -27,6 +28,11 @@ class TrainGuider(comfy_extras.nodes_custom_sampler.Guider_Basic): """ CFGGuider with modifications for training specific logic """ + + def __init__(self, *args, offloading=False, **kwargs): + super().__init__(*args, **kwargs) + self.offloading = offloading + def outer_sample( self, noise, @@ -45,9 +51,11 @@ class TrainGuider(comfy_extras.nodes_custom_sampler.Guider_Basic): noise.shape, self.conds, self.model_options, - force_full_load=True, # mirror behavior in TrainLoraNode.execute() to keep model loaded + force_full_load=not self.offloading, + force_offload=self.offloading, ) ) + torch.cuda.empty_cache() device = self.model_patcher.load_device if denoise_mask is not None: @@ -404,16 +412,97 @@ def find_all_highest_child_module_with_forward( return result -def patch(m): +def find_modules_at_depth( + model: nn.Module, depth: int = 1, result=None, current_depth=0, name=None +) -> list[nn.Module]: + """ + Find modules at a specific depth level for gradient checkpointing. + + Args: + model: The model to search + depth: Target depth level (1 = top-level blocks, 2 = their children, etc.) + result: Accumulator for results + current_depth: Current recursion depth + name: Current module name for logging + + Returns: + List of modules at the target depth + """ + if result is None: + result = [] + name = name or "root" + + # Skip container modules (they don't have meaningful forward) + is_container = isinstance(model, (nn.ModuleList, nn.Sequential, nn.ModuleDict)) + has_forward = hasattr(model, "forward") and not is_container + + if has_forward: + current_depth += 1 + if current_depth == depth: + result.append(model) + logging.debug(f"Found module at depth {depth}: {name} ({model.__class__.__name__})") + return result + + # Recurse into children + for next_name, child in model.named_children(): + find_modules_at_depth(child, depth, result, current_depth, f"{name}.{next_name}") + + return result + + +class OffloadCheckpointFunction(torch.autograd.Function): + """ + Gradient checkpointing that works with weight offloading. + + Forward: no_grad -> compute -> weights can be freed + Backward: enable_grad -> recompute -> backward -> weights can be freed + + For single input, single output modules (Linear, Conv*). + """ + + @staticmethod + def forward(ctx, x: torch.Tensor, forward_fn): + ctx.save_for_backward(x) + ctx.forward_fn = forward_fn + with torch.no_grad(): + return forward_fn(x) + + @staticmethod + def backward(ctx, grad_out: torch.Tensor): + x, = ctx.saved_tensors + forward_fn = ctx.forward_fn + + # Clear context early + ctx.forward_fn = None + + with torch.enable_grad(): + x_detached = x.detach().requires_grad_(True) + y = forward_fn(x_detached) + y.backward(grad_out) + grad_x = x_detached.grad + + # Explicit cleanup + del y, x_detached, forward_fn + + return grad_x, None + + +def patch(m, offloading=False): if not hasattr(m, "forward"): return org_forward = m.forward - def fwd(args, kwargs): - return org_forward(*args, **kwargs) + # Branch 1: Linear/Conv* -> offload-compatible checkpoint (single input/output) + if offloading and isinstance(m, (nn.Linear, nn.Conv1d, nn.Conv2d, nn.Conv3d)): + def checkpointing_fwd(x): + return OffloadCheckpointFunction.apply(x, org_forward) + # Branch 2: Others -> standard checkpoint + else: + def fwd(args, kwargs): + return org_forward(*args, **kwargs) - def checkpointing_fwd(*args, **kwargs): - return torch.utils.checkpoint.checkpoint(fwd, args, kwargs, use_reentrant=False) + def checkpointing_fwd(*args, **kwargs): + return torch.utils.checkpoint.checkpoint(fwd, args, kwargs, use_reentrant=False) m.org_forward = org_forward m.forward = checkpointing_fwd @@ -936,6 +1025,18 @@ class TrainLoraNode(io.ComfyNode): default=True, tooltip="Use gradient checkpointing for training.", ), + io.Int.Input( + "checkpoint_depth", + default=1, + min=1, + max=5, + tooltip="Depth level for gradient checkpointing.", + ), + io.Boolean.Input( + "offloading", + default=False, + tooltip="Depth level for gradient checkpointing.", + ), io.Combo.Input( "existing_lora", options=folder_paths.get_filename_list("loras") + ["[None]"], @@ -982,6 +1083,8 @@ class TrainLoraNode(io.ComfyNode): lora_dtype, algorithm, gradient_checkpointing, + checkpoint_depth, + offloading, existing_lora, bucket_mode, bypass_mode, @@ -1000,6 +1103,8 @@ class TrainLoraNode(io.ComfyNode): lora_dtype = lora_dtype[0] algorithm = algorithm[0] gradient_checkpointing = gradient_checkpointing[0] + offloading = offloading[0] + checkpoint_depth = checkpoint_depth[0] existing_lora = existing_lora[0] bucket_mode = bucket_mode[0] bypass_mode = bypass_mode[0] @@ -1054,16 +1159,18 @@ class TrainLoraNode(io.ComfyNode): # Setup gradient checkpointing if gradient_checkpointing: - for m in find_all_highest_child_module_with_forward( - mp.model.diffusion_model - ): - patch(m) + modules_to_patch = find_modules_at_depth( + mp.model.diffusion_model, depth=checkpoint_depth + ) + logging.info(f"Gradient checkpointing: patching {len(modules_to_patch)} modules at depth {checkpoint_depth}") + for m in modules_to_patch: + patch(m, offloading=offloading) torch.cuda.empty_cache() # With force_full_load=False we should be able to have offloading # But for offloading in training we need custom AutoGrad hooks for fwd/bwd comfy.model_management.load_models_gpu( - [mp], memory_required=1e20, force_full_load=True + [mp], memory_required=1e20, force_full_load=not offloading ) torch.cuda.empty_cache() @@ -1100,7 +1207,7 @@ class TrainLoraNode(io.ComfyNode): ) # Setup guider - guider = TrainGuider(mp) + guider = TrainGuider(mp, offloading=offloading) guider.set_conds(positive) # Inject bypass hooks if bypass mode is enabled @@ -1113,6 +1220,7 @@ class TrainLoraNode(io.ComfyNode): # Run training loop try: + comfy.model_management.in_training = True _run_training_loop( guider, train_sampler, @@ -1123,6 +1231,7 @@ class TrainLoraNode(io.ComfyNode): multi_res, ) finally: + comfy.model_management.in_training = False # Eject bypass hooks if they were injected if bypass_injections is not None: for injection in bypass_injections: @@ -1132,19 +1241,20 @@ class TrainLoraNode(io.ComfyNode): unpatch(m) del train_sampler, optimizer - # Finalize adapters + for param in lora_sd: + lora_sd[param] = lora_sd[param].to(lora_dtype).detach() + for adapter in all_weight_adapters: adapter.requires_grad_(False) - - for param in lora_sd: - lora_sd[param] = lora_sd[param].to(lora_dtype) + del adapter + del all_weight_adapters # mp in train node is highly specialized for training # use it in inference will result in bad behavior so we don't return it return io.NodeOutput(lora_sd, loss_map, steps + existing_steps) -class LoraModelLoader(io.ComfyNode):# +class LoraModelLoader(io.ComfyNode): @classmethod def define_schema(cls): return io.Schema( @@ -1166,6 +1276,11 @@ class LoraModelLoader(io.ComfyNode):# max=100.0, tooltip="How strongly to modify the diffusion model. This value can be negative.", ), + io.Boolean.Input( + "bypass", + default=False, + tooltip="When enabled, applies LoRA in bypass mode without modifying base model weights. Useful for training and when model weights are offloaded.", + ), ], outputs=[ io.Model.Output( @@ -1175,13 +1290,18 @@ class LoraModelLoader(io.ComfyNode):# ) @classmethod - def execute(cls, model, lora, strength_model): + def execute(cls, model, lora, strength_model, bypass=False): if strength_model == 0: return io.NodeOutput(model) - model_lora, _ = comfy.sd.load_lora_for_models( - model, None, lora, strength_model, 0 - ) + if bypass: + model_lora, _ = comfy.sd.load_bypass_lora_for_models( + model, None, lora, strength_model, 0 + ) + else: + model_lora, _ = comfy.sd.load_lora_for_models( + model, None, lora, strength_model, 0 + ) return io.NodeOutput(model_lora) From 76a7fa96dbdc2eda89218601fe3aed5997df055f Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Tue, 10 Feb 2026 19:04:32 -0800 Subject: [PATCH 613/647] Make built in lora training work on anima. (#12402) --- comfy/ldm/anima/model.py | 16 ++++++++++++++-- comfy/model_base.py | 12 ++++++++---- 2 files changed, 22 insertions(+), 6 deletions(-) diff --git a/comfy/ldm/anima/model.py b/comfy/ldm/anima/model.py index 2e6ed58fa..6fb51c4a4 100644 --- a/comfy/ldm/anima/model.py +++ b/comfy/ldm/anima/model.py @@ -195,8 +195,20 @@ class Anima(MiniTrainDIT): super().__init__(*args, **kwargs) self.llm_adapter = LLMAdapter(device=kwargs.get("device"), dtype=kwargs.get("dtype"), operations=kwargs.get("operations")) - def preprocess_text_embeds(self, text_embeds, text_ids): + def preprocess_text_embeds(self, text_embeds, text_ids, t5xxl_weights=None): if text_ids is not None: - return self.llm_adapter(text_embeds, text_ids) + out = self.llm_adapter(text_embeds, text_ids) + if t5xxl_weights is not None: + out = out * t5xxl_weights + + if out.shape[1] < 512: + out = torch.nn.functional.pad(out, (0, 0, 0, 512 - out.shape[1])) + return out else: return text_embeds + + def forward(self, x, timesteps, context, **kwargs): + t5xxl_ids = kwargs.pop("t5xxl_ids", None) + if t5xxl_ids is not None: + context = self.preprocess_text_embeds(context, t5xxl_ids, t5xxl_weights=kwargs.pop("t5xxl_weights", None)) + return super().forward(x, timesteps, context, **kwargs) diff --git a/comfy/model_base.py b/comfy/model_base.py index 858789b30..4a74cb1ce 100644 --- a/comfy/model_base.py +++ b/comfy/model_base.py @@ -1160,12 +1160,16 @@ class Anima(BaseModel): device = kwargs["device"] if cross_attn is not None: if t5xxl_ids is not None: - cross_attn = self.diffusion_model.preprocess_text_embeds(cross_attn.to(device=device, dtype=self.get_dtype()), t5xxl_ids.unsqueeze(0).to(device=device)) if t5xxl_weights is not None: - cross_attn *= t5xxl_weights.unsqueeze(0).unsqueeze(-1).to(cross_attn) + t5xxl_weights = t5xxl_weights.unsqueeze(0).unsqueeze(-1).to(cross_attn) + t5xxl_ids = t5xxl_ids.unsqueeze(0) + + if torch.is_inference_mode_enabled(): # if not we are training + cross_attn = self.diffusion_model.preprocess_text_embeds(cross_attn.to(device=device, dtype=self.get_dtype()), t5xxl_ids.to(device=device), t5xxl_weights=t5xxl_weights.to(device=device, dtype=self.get_dtype())) + else: + out['t5xxl_ids'] = comfy.conds.CONDRegular(t5xxl_ids) + out['t5xxl_weights'] = comfy.conds.CONDRegular(t5xxl_weights) - if cross_attn.shape[1] < 512: - cross_attn = torch.nn.functional.pad(cross_attn, (0, 0, 0, 512 - cross_attn.shape[1])) out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn) return out From 2c7cef4a23c08e3f02a33c693d927158a15a11f6 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Wed, 11 Feb 2026 20:51:49 +0200 Subject: [PATCH 614/647] fix(api-nodes): retry on connection errors during polling instead of aborting (#12393) --- comfy_api_nodes/util/client.py | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/comfy_api_nodes/util/client.py b/comfy_api_nodes/util/client.py index 8a1259506..391748e7a 100644 --- a/comfy_api_nodes/util/client.py +++ b/comfy_api_nodes/util/client.py @@ -143,9 +143,9 @@ async def poll_op( poll_interval: float = 5.0, max_poll_attempts: int = 160, timeout_per_poll: float = 120.0, - max_retries_per_poll: int = 3, + max_retries_per_poll: int = 10, retry_delay_per_poll: float = 1.0, - retry_backoff_per_poll: float = 2.0, + retry_backoff_per_poll: float = 1.4, estimated_duration: int | None = None, cancel_endpoint: ApiEndpoint | None = None, cancel_timeout: float = 10.0, @@ -240,9 +240,9 @@ async def poll_op_raw( poll_interval: float = 5.0, max_poll_attempts: int = 160, timeout_per_poll: float = 120.0, - max_retries_per_poll: int = 3, + max_retries_per_poll: int = 10, retry_delay_per_poll: float = 1.0, - retry_backoff_per_poll: float = 2.0, + retry_backoff_per_poll: float = 1.4, estimated_duration: int | None = None, cancel_endpoint: ApiEndpoint | None = None, cancel_timeout: float = 10.0, From 4993411fd9a43d642971925272c3748d9e058131 Mon Sep 17 00:00:00 2001 From: Benjamin Lu Date: Wed, 11 Feb 2026 11:15:13 -0800 Subject: [PATCH 615/647] Dispatch desktop auto-bump when a ComfyUI release is published (#12398) * Dispatch desktop auto-bump on ComfyUI release publish * Fix release webhook secret checks in step conditions * Require desktop dispatch token in release webhook * Apply suggestion from @Copilot Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com> --------- Co-authored-by: Luke Mino-Altherr Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com> Co-authored-by: Jedrzej Kosinski --- .github/workflows/release-webhook.yml | 36 +++++++++++++++++++++++++++ 1 file changed, 36 insertions(+) diff --git a/.github/workflows/release-webhook.yml b/.github/workflows/release-webhook.yml index 6fceb7560..737e4c488 100644 --- a/.github/workflows/release-webhook.yml +++ b/.github/workflows/release-webhook.yml @@ -7,6 +7,8 @@ on: jobs: send-webhook: runs-on: ubuntu-latest + env: + DESKTOP_REPO_DISPATCH_TOKEN: ${{ secrets.DESKTOP_REPO_DISPATCH_TOKEN }} steps: - name: Send release webhook env: @@ -106,3 +108,37 @@ jobs: --fail --silent --show-error echo "✅ Release webhook sent successfully" + + - name: Send repository dispatch to desktop + env: + DISPATCH_TOKEN: ${{ env.DESKTOP_REPO_DISPATCH_TOKEN }} + RELEASE_TAG: ${{ github.event.release.tag_name }} + RELEASE_URL: ${{ github.event.release.html_url }} + run: | + set -euo pipefail + + if [ -z "${DISPATCH_TOKEN:-}" ]; then + echo "::error::DESKTOP_REPO_DISPATCH_TOKEN is required but not set." + exit 1 + fi + + PAYLOAD="$(jq -n \ + --arg release_tag "$RELEASE_TAG" \ + --arg release_url "$RELEASE_URL" \ + '{ + event_type: "comfyui_release_published", + client_payload: { + release_tag: $release_tag, + release_url: $release_url + } + }')" + + curl -fsSL \ + -X POST \ + -H "Accept: application/vnd.github+json" \ + -H "Content-Type: application/json" \ + -H "Authorization: Bearer ${DISPATCH_TOKEN}" \ + https://api.github.com/repos/Comfy-Org/desktop/dispatches \ + -d "$PAYLOAD" + + echo "✅ Dispatched ComfyUI release ${RELEASE_TAG} to Comfy-Org/desktop" From 2b7cc7e3b69127a81b9232d4e8305eb678fa3d0c Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Wed, 11 Feb 2026 21:30:19 +0200 Subject: [PATCH 616/647] [API Nodes] enable Magnific Upscalers (#12179) * feat(api-nodes): enable Magnific Upscalers * update price badges --------- Co-authored-by: Jedrzej Kosinski --- comfy_api_nodes/nodes_magnific.py | 62 +++++++++++++++++++++++++++---- 1 file changed, 55 insertions(+), 7 deletions(-) diff --git a/comfy_api_nodes/nodes_magnific.py b/comfy_api_nodes/nodes_magnific.py index 013e71cc8..83a581c5d 100644 --- a/comfy_api_nodes/nodes_magnific.py +++ b/comfy_api_nodes/nodes_magnific.py @@ -30,6 +30,30 @@ from comfy_api_nodes.util import ( validate_image_dimensions, ) +_EUR_TO_USD = 1.19 + + +def _tier_price_eur(megapixels: float) -> float: + """Price in EUR for a single Magnific upscaling step based on input megapixels.""" + if megapixels <= 1.3: + return 0.143 + if megapixels <= 3.0: + return 0.286 + if megapixels <= 6.4: + return 0.429 + return 1.716 + + +def _calculate_magnific_upscale_price_usd(width: int, height: int, scale: int) -> float: + """Calculate total Magnific upscale price in USD for given input dimensions and scale factor.""" + num_steps = int(math.log2(scale)) + total_eur = 0.0 + pixels = width * height + for _ in range(num_steps): + total_eur += _tier_price_eur(pixels / 1_000_000) + pixels *= 4 + return round(total_eur * _EUR_TO_USD, 2) + class MagnificImageUpscalerCreativeNode(IO.ComfyNode): @classmethod @@ -103,11 +127,20 @@ class MagnificImageUpscalerCreativeNode(IO.ComfyNode): ], is_api_node=True, price_badge=IO.PriceBadge( - depends_on=IO.PriceBadgeDepends(widgets=["scale_factor"]), + depends_on=IO.PriceBadgeDepends(widgets=["scale_factor", "auto_downscale"]), expr=""" ( - $max := widgets.scale_factor = "2x" ? 1.326 : 1.657; - {"type": "range_usd", "min_usd": 0.11, "max_usd": $max} + $ad := widgets.auto_downscale; + $mins := $ad + ? {"2x": 0.172, "4x": 0.343, "8x": 0.515, "16x": 0.515} + : {"2x": 0.172, "4x": 0.343, "8x": 0.515, "16x": 0.844}; + $maxs := {"2x": 0.515, "4x": 0.844, "8x": 1.015, "16x": 1.187}; + { + "type": "range_usd", + "min_usd": $lookup($mins, widgets.scale_factor), + "max_usd": $lookup($maxs, widgets.scale_factor), + "format": { "approximate": true } + } ) """, ), @@ -168,6 +201,10 @@ class MagnificImageUpscalerCreativeNode(IO.ComfyNode): f"Use a smaller input image or lower scale factor." ) + final_height, final_width = get_image_dimensions(image) + actual_scale = int(scale_factor.rstrip("x")) + price_usd = _calculate_magnific_upscale_price_usd(final_width, final_height, actual_scale) + initial_res = await sync_op( cls, ApiEndpoint(path="/proxy/freepik/v1/ai/image-upscaler", method="POST"), @@ -189,6 +226,7 @@ class MagnificImageUpscalerCreativeNode(IO.ComfyNode): ApiEndpoint(path=f"/proxy/freepik/v1/ai/image-upscaler/{initial_res.task_id}"), response_model=TaskResponse, status_extractor=lambda x: x.status, + price_extractor=lambda _: price_usd, poll_interval=10.0, max_poll_attempts=480, ) @@ -257,8 +295,14 @@ class MagnificImageUpscalerPreciseV2Node(IO.ComfyNode): depends_on=IO.PriceBadgeDepends(widgets=["scale_factor"]), expr=""" ( - $max := widgets.scale_factor = "2x" ? 1.326 : 1.657; - {"type": "range_usd", "min_usd": 0.11, "max_usd": $max} + $mins := {"2x": 0.172, "4x": 0.343, "8x": 0.515, "16x": 0.844}; + $maxs := {"2x": 2.045, "4x": 2.545, "8x": 2.889, "16x": 3.06}; + { + "type": "range_usd", + "min_usd": $lookup($mins, widgets.scale_factor), + "max_usd": $lookup($maxs, widgets.scale_factor), + "format": { "approximate": true } + } ) """, ), @@ -321,6 +365,9 @@ class MagnificImageUpscalerPreciseV2Node(IO.ComfyNode): f"Use a smaller input image or lower scale factor." ) + final_height, final_width = get_image_dimensions(image) + price_usd = _calculate_magnific_upscale_price_usd(final_width, final_height, requested_scale) + initial_res = await sync_op( cls, ApiEndpoint(path="/proxy/freepik/v1/ai/image-upscaler-precision-v2", method="POST"), @@ -339,6 +386,7 @@ class MagnificImageUpscalerPreciseV2Node(IO.ComfyNode): ApiEndpoint(path=f"/proxy/freepik/v1/ai/image-upscaler-precision-v2/{initial_res.task_id}"), response_model=TaskResponse, status_extractor=lambda x: x.status, + price_extractor=lambda _: price_usd, poll_interval=10.0, max_poll_attempts=480, ) @@ -877,8 +925,8 @@ class MagnificExtension(ComfyExtension): @override async def get_node_list(self) -> list[type[IO.ComfyNode]]: return [ - # MagnificImageUpscalerCreativeNode, - # MagnificImageUpscalerPreciseV2Node, + MagnificImageUpscalerCreativeNode, + MagnificImageUpscalerPreciseV2Node, MagnificImageStyleTransferNode, MagnificImageRelightNode, MagnificImageSkinEnhancerNode, From d297a749a2fa3a34ebff898797feef161bcd64c6 Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Wed, 11 Feb 2026 11:50:16 -0800 Subject: [PATCH 617/647] dynamic_vram: Fix windows Aimdo crash + Fix LLM performance (#12408) * model_management: lazy-cache aimdo_tensor These tensors cosntructed from aimdo-allocations are CPU expensive to make on the pytorch side. Add a cache version that will be valid with signature match to fast path past whatever torch is doing. * dynamic_vram: Minimize fast path CPU work Move as much as possible inside the not resident if block and cache the formed weight and bias rather than the flat intermediates. In extreme layer weight rates this adds up. --- comfy/model_management.py | 8 ++++++-- comfy/model_patcher.py | 2 -- comfy/ops.py | 21 ++++++++++++++------- 3 files changed, 20 insertions(+), 11 deletions(-) diff --git a/comfy/model_management.py b/comfy/model_management.py index 304931eb0..38c3e482b 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -1213,8 +1213,12 @@ def cast_to(weight, dtype=None, device=None, non_blocking=False, copy=False, str signature = comfy_aimdo.model_vbar.vbar_fault(weight._v) if signature is not None: - v_tensor = comfy.memory_management.interpret_gathered_like(cast_geometry, weight._v_tensor)[0] - if not comfy_aimdo.model_vbar.vbar_signature_compare(signature, weight._v_signature): + if comfy_aimdo.model_vbar.vbar_signature_compare(signature, weight._v_signature): + v_tensor = weight._v_tensor + else: + raw_tensor = comfy_aimdo.torch.aimdo_to_tensor(weight._v, device) + v_tensor = comfy.memory_management.interpret_gathered_like(cast_geometry, raw_tensor)[0] + weight._v_tensor = v_tensor weight._v_signature = signature #Send it over v_tensor.copy_(weight, non_blocking=non_blocking) diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index 19c9031ea..224e218e3 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -1542,7 +1542,6 @@ class ModelPatcherDynamic(ModelPatcher): if vbar is not None and not hasattr(m, "_v"): m._v = vbar.alloc(v_weight_size) - m._v_tensor = comfy_aimdo.torch.aimdo_to_tensor(m._v, device_to) allocated_size += v_weight_size else: @@ -1557,7 +1556,6 @@ class ModelPatcherDynamic(ModelPatcher): weight_size = geometry.numel() * geometry.element_size() if vbar is not None and not hasattr(weight, "_v"): weight._v = vbar.alloc(weight_size) - weight._v_tensor = comfy_aimdo.torch.aimdo_to_tensor(weight._v, device_to) weight._model_dtype = model_dtype allocated_size += weight_size vbar.set_watermark_limit(allocated_size) diff --git a/comfy/ops.py b/comfy/ops.py index 33803b223..688937e43 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -83,14 +83,18 @@ def cast_to_input(weight, input, non_blocking=False, copy=True): def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compute_dtype): offload_stream = None xfer_dest = None - cast_geometry = comfy.memory_management.tensors_to_geometries([ s.weight, s.bias ]) signature = comfy_aimdo.model_vbar.vbar_fault(s._v) - if signature is not None: - xfer_dest = s._v_tensor resident = comfy_aimdo.model_vbar.vbar_signature_compare(signature, s._v_signature) + if signature is not None: + if resident: + weight = s._v_weight + bias = s._v_bias + else: + xfer_dest = comfy_aimdo.torch.aimdo_to_tensor(s._v, device) if not resident: + cast_geometry = comfy.memory_management.tensors_to_geometries([ s.weight, s.bias ]) cast_dest = None xfer_source = [ s.weight, s.bias ] @@ -140,9 +144,13 @@ def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compu post_cast.copy_(pre_cast) xfer_dest = cast_dest - params = comfy.memory_management.interpret_gathered_like(cast_geometry, xfer_dest) - weight = params[0] - bias = params[1] + params = comfy.memory_management.interpret_gathered_like(cast_geometry, xfer_dest) + weight = params[0] + bias = params[1] + if signature is not None: + s._v_weight = weight + s._v_bias = bias + s._v_signature=signature def post_cast(s, param_key, x, dtype, resident, update_weight): lowvram_fn = getattr(s, param_key + "_lowvram_function", None) @@ -182,7 +190,6 @@ def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compu weight = post_cast(s, "weight", weight, dtype, resident, update_weight) if s.bias is not None: bias = post_cast(s, "bias", bias, bias_dtype, resident, update_weight) - s._v_signature=signature #FIXME: weird offload return protocol return weight, bias, (offload_stream, device if signature is not None else None, None) From 2a4328d639810858aa625c7bfedb974a13a57abe Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Wed, 11 Feb 2026 11:53:42 -0800 Subject: [PATCH 618/647] ace15: Use dynamic_vram friendly trange (#12409) Factor out the ksampler trange and use it in ACE LLM to prevent the silent stall at 0 and rate distortion due to first-step model load. --- comfy/k_diffusion/sampling.py | 32 ++------------------------------ comfy/text_encoders/ace15.py | 3 +-- comfy/utils.py | 27 +++++++++++++++++++++++++++ 3 files changed, 30 insertions(+), 32 deletions(-) diff --git a/comfy/k_diffusion/sampling.py b/comfy/k_diffusion/sampling.py index c0c51d51a..6978eb717 100644 --- a/comfy/k_diffusion/sampling.py +++ b/comfy/k_diffusion/sampling.py @@ -1,12 +1,11 @@ import math -import time from functools import partial from scipy import integrate import torch from torch import nn import torchsde -from tqdm.auto import trange as trange_, tqdm +from tqdm.auto import tqdm from . import utils from . import deis @@ -15,34 +14,7 @@ import comfy.model_patcher import comfy.model_sampling import comfy.memory_management - - -def trange(*args, **kwargs): - if comfy.memory_management.aimdo_allocator is None: - return trange_(*args, **kwargs) - - pbar = trange_(*args, **kwargs, smoothing=1.0) - pbar._i = 0 - pbar.set_postfix_str(" Model Initializing ... ") - - _update = pbar.update - - def warmup_update(n=1): - pbar._i += 1 - if pbar._i == 1: - pbar.i1_time = time.time() - pbar.set_postfix_str(" Model Initialization complete! ") - elif pbar._i == 2: - #bring forward the effective start time based the the diff between first and second iteration - #to attempt to remove load overhead from the final step rate estimate. - pbar.start_t = pbar.i1_time - (time.time() - pbar.i1_time) - pbar.set_postfix_str("") - - _update(n) - - pbar.update = warmup_update - return pbar - +from comfy.utils import model_trange as trange def append_zero(x): return torch.cat([x, x.new_zeros([1])]) diff --git a/comfy/text_encoders/ace15.py b/comfy/text_encoders/ace15.py index 73697b3c1..b8198a820 100644 --- a/comfy/text_encoders/ace15.py +++ b/comfy/text_encoders/ace15.py @@ -3,7 +3,6 @@ import comfy.text_encoders.llama from comfy import sd1_clip import torch import math -from tqdm.auto import trange import yaml import comfy.utils @@ -52,7 +51,7 @@ def sample_manual_loop_no_classes( progress_bar = comfy.utils.ProgressBar(max_new_tokens) - for step in trange(max_new_tokens, desc="LM sampling"): + for step in comfy.utils.model_trange(max_new_tokens, desc="LM sampling"): outputs = model.transformer(None, attention_mask, embeds=embeds.to(execution_dtype), num_tokens=num_tokens, intermediate_output=None, dtype=execution_dtype, embeds_info=embeds_info, past_key_values=past_key_values) next_token_logits = model.transformer.logits(outputs[0])[:, -1] past_key_values = outputs[2] diff --git a/comfy/utils.py b/comfy/utils.py index edd80cebe..e0a94e2e1 100644 --- a/comfy/utils.py +++ b/comfy/utils.py @@ -27,6 +27,7 @@ from PIL import Image import logging import itertools from torch.nn.functional import interpolate +from tqdm.auto import trange from einops import rearrange from comfy.cli_args import args, enables_dynamic_vram import json @@ -1155,6 +1156,32 @@ def tiled_scale_multidim(samples, function, tile=(64, 64), overlap=8, upscale_am def tiled_scale(samples, function, tile_x=64, tile_y=64, overlap = 8, upscale_amount = 4, out_channels = 3, output_device="cpu", pbar = None): return tiled_scale_multidim(samples, function, (tile_y, tile_x), overlap=overlap, upscale_amount=upscale_amount, out_channels=out_channels, output_device=output_device, pbar=pbar) +def model_trange(*args, **kwargs): + if comfy.memory_management.aimdo_allocator is None: + return trange(*args, **kwargs) + + pbar = trange(*args, **kwargs, smoothing=1.0) + pbar._i = 0 + pbar.set_postfix_str(" Model Initializing ... ") + + _update = pbar.update + + def warmup_update(n=1): + pbar._i += 1 + if pbar._i == 1: + pbar.i1_time = time.time() + pbar.set_postfix_str(" Model Initialization complete! ") + elif pbar._i == 2: + #bring forward the effective start time based the the diff between first and second iteration + #to attempt to remove load overhead from the final step rate estimate. + pbar.start_t = pbar.i1_time - (time.time() - pbar.i1_time) + pbar.set_postfix_str("") + + _update(n) + + pbar.update = warmup_update + return pbar + PROGRESS_BAR_ENABLED = True def set_progress_bar_enabled(enabled): global PROGRESS_BAR_ENABLED From 3fe61cedda090c744dcf6f579ed48744fa66ef5f Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Wed, 11 Feb 2026 11:54:02 -0800 Subject: [PATCH 619/647] model_patcher: guard against none model_dtype (#12410) Handle the case where the _model_dtype exists but is none with the intended fallback. --- comfy/model_patcher.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index 224e218e3..f278fccac 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -1525,7 +1525,7 @@ class ModelPatcherDynamic(ModelPatcher): setattr(m, param_key + "_function", weight_function) geometry = weight if not isinstance(weight, QuantizedTensor): - model_dtype = getattr(m, param_key + "_comfy_model_dtype", weight.dtype) + model_dtype = getattr(m, param_key + "_comfy_model_dtype", None) or weight.dtype weight._model_dtype = model_dtype geometry = comfy.memory_management.TensorGeometry(shape=weight.shape, dtype=model_dtype) return comfy.memory_management.vram_aligned_size(geometry) @@ -1551,7 +1551,7 @@ class ModelPatcherDynamic(ModelPatcher): weight.seed_key = key set_dirty(weight, dirty) geometry = weight - model_dtype = getattr(m, param + "_comfy_model_dtype", weight.dtype) + model_dtype = getattr(m, param + "_comfy_model_dtype", None) or weight.dtype geometry = comfy.memory_management.TensorGeometry(shape=weight.shape, dtype=model_dtype) weight_size = geometry.numel() * geometry.element_size() if vbar is not None and not hasattr(weight, "_v"): From e5ae670a4016d3698a806e7f840fcecc50639848 Mon Sep 17 00:00:00 2001 From: askmyteapot <62238146+askmyteapot@users.noreply.github.com> Date: Thu, 12 Feb 2026 11:28:48 +1000 Subject: [PATCH 620/647] Update ace15.py to allow min_p sampling (#12373) --- comfy/text_encoders/ace15.py | 15 ++++++++++++--- comfy_extras/nodes_ace.py | 5 +++-- 2 files changed, 15 insertions(+), 5 deletions(-) diff --git a/comfy/text_encoders/ace15.py b/comfy/text_encoders/ace15.py index b8198a820..0fdd4669f 100644 --- a/comfy/text_encoders/ace15.py +++ b/comfy/text_encoders/ace15.py @@ -16,6 +16,7 @@ def sample_manual_loop_no_classes( temperature: float = 0.85, top_p: float = 0.9, top_k: int = None, + min_p: float = 0.000, seed: int = 1, min_tokens: int = 1, max_new_tokens: int = 2048, @@ -80,6 +81,12 @@ def sample_manual_loop_no_classes( min_val = top_k_vals[..., -1, None] cfg_logits[cfg_logits < min_val] = remove_logit_value + if min_p is not None and min_p > 0: + probs = torch.softmax(cfg_logits, dim=-1) + p_max = probs.max(dim=-1, keepdim=True).values + indices_to_remove = probs < (min_p * p_max) + cfg_logits[indices_to_remove] = remove_logit_value + if top_p is not None and top_p < 1.0: sorted_logits, sorted_indices = torch.sort(cfg_logits, descending=True) cumulative_probs = torch.cumsum(torch.softmax(sorted_logits, dim=-1), dim=-1) @@ -110,7 +117,7 @@ def sample_manual_loop_no_classes( return output_audio_codes -def generate_audio_codes(model, positive, negative, min_tokens=1, max_tokens=1024, seed=0, cfg_scale=2.0, temperature=0.85, top_p=0.9, top_k=0): +def generate_audio_codes(model, positive, negative, min_tokens=1, max_tokens=1024, seed=0, cfg_scale=2.0, temperature=0.85, top_p=0.9, top_k=0, min_p=0.000): positive = [[token for token, _ in inner_list] for inner_list in positive] positive = positive[0] @@ -134,7 +141,7 @@ def generate_audio_codes(model, positive, negative, min_tokens=1, max_tokens=102 paddings = [] ids = [positive] - return sample_manual_loop_no_classes(model, ids, paddings, cfg_scale=cfg_scale, temperature=temperature, top_p=top_p, top_k=top_k, seed=seed, min_tokens=min_tokens, max_new_tokens=max_tokens) + return sample_manual_loop_no_classes(model, ids, paddings, cfg_scale=cfg_scale, temperature=temperature, top_p=top_p, top_k=top_k, min_p=min_p, seed=seed, min_tokens=min_tokens, max_new_tokens=max_tokens) class ACE15Tokenizer(sd1_clip.SD1Tokenizer): @@ -192,6 +199,7 @@ class ACE15Tokenizer(sd1_clip.SD1Tokenizer): temperature = kwargs.get("temperature", 0.85) top_p = kwargs.get("top_p", 0.9) top_k = kwargs.get("top_k", 0.0) + min_p = kwargs.get("min_p", 0.000) duration = math.ceil(duration) kwargs["duration"] = duration @@ -239,6 +247,7 @@ class ACE15Tokenizer(sd1_clip.SD1Tokenizer): "temperature": temperature, "top_p": top_p, "top_k": top_k, + "min_p": min_p, } return out @@ -299,7 +308,7 @@ class ACE15TEModel(torch.nn.Module): lm_metadata = token_weight_pairs["lm_metadata"] if lm_metadata["generate_audio_codes"]: - audio_codes = generate_audio_codes(getattr(self, self.lm_model, self.qwen3_06b), token_weight_pairs["lm_prompt"], token_weight_pairs["lm_prompt_negative"], min_tokens=lm_metadata["min_tokens"], max_tokens=lm_metadata["max_tokens"], seed=lm_metadata["seed"], cfg_scale=lm_metadata["cfg_scale"], temperature=lm_metadata["temperature"], top_p=lm_metadata["top_p"], top_k=lm_metadata["top_k"]) + audio_codes = generate_audio_codes(getattr(self, self.lm_model, self.qwen3_06b), token_weight_pairs["lm_prompt"], token_weight_pairs["lm_prompt_negative"], min_tokens=lm_metadata["min_tokens"], max_tokens=lm_metadata["min_tokens"], seed=lm_metadata["seed"], cfg_scale=lm_metadata["cfg_scale"], temperature=lm_metadata["temperature"], top_p=lm_metadata["top_p"], top_k=lm_metadata["top_k"], min_p=lm_metadata["min_p"]) out["audio_codes"] = [audio_codes] return base_out, None, out diff --git a/comfy_extras/nodes_ace.py b/comfy_extras/nodes_ace.py index dde5bbd2a..9cf84ab4d 100644 --- a/comfy_extras/nodes_ace.py +++ b/comfy_extras/nodes_ace.py @@ -49,13 +49,14 @@ class TextEncodeAceStepAudio15(io.ComfyNode): io.Float.Input("temperature", default=0.85, min=0.0, max=2.0, step=0.01, advanced=True), io.Float.Input("top_p", default=0.9, min=0.0, max=2000.0, step=0.01, advanced=True), io.Int.Input("top_k", default=0, min=0, max=100, advanced=True), + io.Float.Input("min_p", default=0.000, min=0.0, max=1.0, step=0.001, advanced=True), ], outputs=[io.Conditioning.Output()], ) @classmethod - def execute(cls, clip, tags, lyrics, seed, bpm, duration, timesignature, language, keyscale, generate_audio_codes, cfg_scale, temperature, top_p, top_k) -> io.NodeOutput: - tokens = clip.tokenize(tags, lyrics=lyrics, bpm=bpm, duration=duration, timesignature=int(timesignature), language=language, keyscale=keyscale, seed=seed, generate_audio_codes=generate_audio_codes, cfg_scale=cfg_scale, temperature=temperature, top_p=top_p, top_k=top_k) + def execute(cls, clip, tags, lyrics, seed, bpm, duration, timesignature, language, keyscale, generate_audio_codes, cfg_scale, temperature, top_p, top_k, min_p) -> io.NodeOutput: + tokens = clip.tokenize(tags, lyrics=lyrics, bpm=bpm, duration=duration, timesignature=int(timesignature), language=language, keyscale=keyscale, seed=seed, generate_audio_codes=generate_audio_codes, cfg_scale=cfg_scale, temperature=temperature, top_p=top_p, top_k=top_k, min_p=min_p) conditioning = clip.encode_from_tokens_scheduled(tokens) return io.NodeOutput(conditioning) From 66c18522fbcde5b62731e3fb080a84b14e3dacfc Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 11 Feb 2026 19:12:16 -0800 Subject: [PATCH 621/647] Add a tip for common error. (#12414) --- execution.py | 2 ++ 1 file changed, 2 insertions(+) diff --git a/execution.py b/execution.py index 896862c6b..f549a2f0f 100644 --- a/execution.py +++ b/execution.py @@ -623,6 +623,8 @@ async def execute(server, dynprompt, caches, current_item, extra_data, executed, logging.info("Memory summary: {}".format(comfy.model_management.debug_memory_summary())) logging.error("Got an OOM, unloading all loaded models.") comfy.model_management.unload_all_models() + elif isinstance(ex, RuntimeError) and ("mat1 and mat2 shapes" in str(ex)) and "Sampler" in class_type: + tips = "\n\nTIPS: If you have any \"Load CLIP\" or \"*CLIP Loader\" nodes in your workflow connected to this sampler node make sure the correct file(s) and type is selected." error_details = { "node_id": real_node_id, From 4a93a62371b64f9d11a140a09faf985c48902d2e Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Thu, 12 Feb 2026 11:38:51 +0200 Subject: [PATCH 622/647] fix(api-nodes): add separate retry budget for 429 rate limit responses (#12421) --- comfy_api_nodes/util/client.py | 207 ++++++++++++----------- comfy_api_nodes/util/download_helpers.py | 30 ++-- comfy_api_nodes/util/request_logger.py | 66 ++++---- comfy_api_nodes/util/upload_helpers.py | 55 +++--- 4 files changed, 177 insertions(+), 181 deletions(-) diff --git a/comfy_api_nodes/util/client.py b/comfy_api_nodes/util/client.py index 391748e7a..94886af7b 100644 --- a/comfy_api_nodes/util/client.py +++ b/comfy_api_nodes/util/client.py @@ -57,6 +57,7 @@ class _RequestConfig: files: dict[str, Any] | list[tuple[str, Any]] | None multipart_parser: Callable | None max_retries: int + max_retries_on_rate_limit: int retry_delay: float retry_backoff: float wait_label: str = "Waiting" @@ -65,6 +66,7 @@ class _RequestConfig: final_label_on_success: str | None = "Completed" progress_origin_ts: float | None = None price_extractor: Callable[[dict[str, Any]], float | None] | None = None + is_rate_limited: Callable[[int, Any], bool] | None = None @dataclass @@ -78,7 +80,7 @@ class _PollUIState: active_since: float | None = None # start time of current active interval (None if queued) -_RETRY_STATUS = {408, 429, 500, 502, 503, 504} +_RETRY_STATUS = {408, 500, 502, 503, 504} # status 429 is handled separately COMPLETED_STATUSES = ["succeeded", "succeed", "success", "completed", "finished", "done", "complete"] FAILED_STATUSES = ["cancelled", "canceled", "canceling", "fail", "failed", "error"] QUEUED_STATUSES = ["created", "queued", "queueing", "submitted", "initializing"] @@ -103,6 +105,8 @@ async def sync_op( final_label_on_success: str | None = "Completed", progress_origin_ts: float | None = None, monitor_progress: bool = True, + max_retries_on_rate_limit: int = 16, + is_rate_limited: Callable[[int, Any], bool] | None = None, ) -> M: raw = await sync_op_raw( cls, @@ -122,6 +126,8 @@ async def sync_op( final_label_on_success=final_label_on_success, progress_origin_ts=progress_origin_ts, monitor_progress=monitor_progress, + max_retries_on_rate_limit=max_retries_on_rate_limit, + is_rate_limited=is_rate_limited, ) if not isinstance(raw, dict): raise Exception("Expected JSON response to validate into a Pydantic model, got non-JSON (binary or text).") @@ -194,6 +200,8 @@ async def sync_op_raw( final_label_on_success: str | None = "Completed", progress_origin_ts: float | None = None, monitor_progress: bool = True, + max_retries_on_rate_limit: int = 16, + is_rate_limited: Callable[[int, Any], bool] | None = None, ) -> dict[str, Any] | bytes: """ Make a single network request. @@ -222,6 +230,8 @@ async def sync_op_raw( final_label_on_success=final_label_on_success, progress_origin_ts=progress_origin_ts, price_extractor=price_extractor, + max_retries_on_rate_limit=max_retries_on_rate_limit, + is_rate_limited=is_rate_limited, ) return await _request_base(cfg, expect_binary=as_binary) @@ -506,7 +516,7 @@ def _friendly_http_message(status: int, body: Any) -> str: if status == 409: return "There is a problem with your account. Please contact support@comfy.org." if status == 429: - return "Rate Limit Exceeded: Please try again later." + return "Rate Limit Exceeded: The server returned 429 after all retry attempts. Please wait and try again." try: if isinstance(body, dict): err = body.get("error") @@ -586,6 +596,8 @@ async def _request_base(cfg: _RequestConfig, expect_binary: bool): start_time = cfg.progress_origin_ts if cfg.progress_origin_ts is not None else time.monotonic() attempt = 0 delay = cfg.retry_delay + rate_limit_attempts = 0 + rate_limit_delay = cfg.retry_delay operation_succeeded: bool = False final_elapsed_seconds: int | None = None extracted_price: float | None = None @@ -653,17 +665,14 @@ async def _request_base(cfg: _RequestConfig, expect_binary: bool): payload_headers["Content-Type"] = "application/json" payload_kw["json"] = cfg.data or {} - try: - request_logger.log_request_response( - operation_id=operation_id, - request_method=method, - request_url=url, - request_headers=dict(payload_headers) if payload_headers else None, - request_params=dict(params) if params else None, - request_data=request_body_log, - ) - except Exception as _log_e: - logging.debug("[DEBUG] request logging failed: %s", _log_e) + request_logger.log_request_response( + operation_id=operation_id, + request_method=method, + request_url=url, + request_headers=dict(payload_headers) if payload_headers else None, + request_params=dict(params) if params else None, + request_data=request_body_log, + ) req_coro = sess.request(method, url, params=params, **payload_kw) req_task = asyncio.create_task(req_coro) @@ -688,41 +697,33 @@ async def _request_base(cfg: _RequestConfig, expect_binary: bool): body = await resp.json() except (ContentTypeError, json.JSONDecodeError): body = await resp.text() - if resp.status in _RETRY_STATUS and attempt <= cfg.max_retries: + should_retry = False + wait_time = 0.0 + retry_label = "" + is_rl = resp.status == 429 or ( + cfg.is_rate_limited is not None and cfg.is_rate_limited(resp.status, body) + ) + if is_rl and rate_limit_attempts < cfg.max_retries_on_rate_limit: + rate_limit_attempts += 1 + wait_time = min(rate_limit_delay, 30.0) + rate_limit_delay *= cfg.retry_backoff + retry_label = f"rate-limit retry {rate_limit_attempts} of {cfg.max_retries_on_rate_limit}" + should_retry = True + elif resp.status in _RETRY_STATUS and (attempt - rate_limit_attempts) <= cfg.max_retries: + wait_time = delay + delay *= cfg.retry_backoff + retry_label = f"retry {attempt - rate_limit_attempts} of {cfg.max_retries}" + should_retry = True + + if should_retry: logging.warning( - "HTTP %s %s -> %s. Retrying in %.2fs (retry %d of %d).", + "HTTP %s %s -> %s. Waiting %.2fs (%s).", method, url, resp.status, - delay, - attempt, - cfg.max_retries, + wait_time, + retry_label, ) - try: - request_logger.log_request_response( - operation_id=operation_id, - request_method=method, - request_url=url, - response_status_code=resp.status, - response_headers=dict(resp.headers), - response_content=body, - error_message=_friendly_http_message(resp.status, body), - ) - except Exception as _log_e: - logging.debug("[DEBUG] response logging failed: %s", _log_e) - - await sleep_with_interrupt( - delay, - cfg.node_cls, - cfg.wait_label if cfg.monitor_progress else None, - start_time if cfg.monitor_progress else None, - cfg.estimated_total, - display_callback=_display_time_progress if cfg.monitor_progress else None, - ) - delay *= cfg.retry_backoff - continue - msg = _friendly_http_message(resp.status, body) - try: request_logger.log_request_response( operation_id=operation_id, request_method=method, @@ -730,10 +731,27 @@ async def _request_base(cfg: _RequestConfig, expect_binary: bool): response_status_code=resp.status, response_headers=dict(resp.headers), response_content=body, - error_message=msg, + error_message=f"HTTP {resp.status} ({retry_label}, will retry in {wait_time:.1f}s)", ) - except Exception as _log_e: - logging.debug("[DEBUG] response logging failed: %s", _log_e) + await sleep_with_interrupt( + wait_time, + cfg.node_cls, + cfg.wait_label if cfg.monitor_progress else None, + start_time if cfg.monitor_progress else None, + cfg.estimated_total, + display_callback=_display_time_progress if cfg.monitor_progress else None, + ) + continue + msg = _friendly_http_message(resp.status, body) + request_logger.log_request_response( + operation_id=operation_id, + request_method=method, + request_url=url, + response_status_code=resp.status, + response_headers=dict(resp.headers), + response_content=body, + error_message=msg, + ) raise Exception(msg) if expect_binary: @@ -753,17 +771,14 @@ async def _request_base(cfg: _RequestConfig, expect_binary: bool): bytes_payload = bytes(buff) operation_succeeded = True final_elapsed_seconds = int(time.monotonic() - start_time) - try: - request_logger.log_request_response( - operation_id=operation_id, - request_method=method, - request_url=url, - response_status_code=resp.status, - response_headers=dict(resp.headers), - response_content=bytes_payload, - ) - except Exception as _log_e: - logging.debug("[DEBUG] response logging failed: %s", _log_e) + request_logger.log_request_response( + operation_id=operation_id, + request_method=method, + request_url=url, + response_status_code=resp.status, + response_headers=dict(resp.headers), + response_content=bytes_payload, + ) return bytes_payload else: try: @@ -780,45 +795,39 @@ async def _request_base(cfg: _RequestConfig, expect_binary: bool): extracted_price = cfg.price_extractor(payload) if cfg.price_extractor else None operation_succeeded = True final_elapsed_seconds = int(time.monotonic() - start_time) - try: - request_logger.log_request_response( - operation_id=operation_id, - request_method=method, - request_url=url, - response_status_code=resp.status, - response_headers=dict(resp.headers), - response_content=response_content_to_log, - ) - except Exception as _log_e: - logging.debug("[DEBUG] response logging failed: %s", _log_e) + request_logger.log_request_response( + operation_id=operation_id, + request_method=method, + request_url=url, + response_status_code=resp.status, + response_headers=dict(resp.headers), + response_content=response_content_to_log, + ) return payload except ProcessingInterrupted: logging.debug("Polling was interrupted by user") raise except (ClientError, OSError) as e: - if attempt <= cfg.max_retries: + if (attempt - rate_limit_attempts) <= cfg.max_retries: logging.warning( "Connection error calling %s %s. Retrying in %.2fs (%d/%d): %s", method, url, delay, - attempt, + attempt - rate_limit_attempts, cfg.max_retries, str(e), ) - try: - request_logger.log_request_response( - operation_id=operation_id, - request_method=method, - request_url=url, - request_headers=dict(payload_headers) if payload_headers else None, - request_params=dict(params) if params else None, - request_data=request_body_log, - error_message=f"{type(e).__name__}: {str(e)} (will retry)", - ) - except Exception as _log_e: - logging.debug("[DEBUG] request error logging failed: %s", _log_e) + request_logger.log_request_response( + operation_id=operation_id, + request_method=method, + request_url=url, + request_headers=dict(payload_headers) if payload_headers else None, + request_params=dict(params) if params else None, + request_data=request_body_log, + error_message=f"{type(e).__name__}: {str(e)} (will retry)", + ) await sleep_with_interrupt( delay, cfg.node_cls, @@ -831,23 +840,6 @@ async def _request_base(cfg: _RequestConfig, expect_binary: bool): continue diag = await _diagnose_connectivity() if not diag["internet_accessible"]: - try: - request_logger.log_request_response( - operation_id=operation_id, - request_method=method, - request_url=url, - request_headers=dict(payload_headers) if payload_headers else None, - request_params=dict(params) if params else None, - request_data=request_body_log, - error_message=f"LocalNetworkError: {str(e)}", - ) - except Exception as _log_e: - logging.debug("[DEBUG] final error logging failed: %s", _log_e) - raise LocalNetworkError( - "Unable to connect to the API server due to local network issues. " - "Please check your internet connection and try again." - ) from e - try: request_logger.log_request_response( operation_id=operation_id, request_method=method, @@ -855,10 +847,21 @@ async def _request_base(cfg: _RequestConfig, expect_binary: bool): request_headers=dict(payload_headers) if payload_headers else None, request_params=dict(params) if params else None, request_data=request_body_log, - error_message=f"ApiServerError: {str(e)}", + error_message=f"LocalNetworkError: {str(e)}", ) - except Exception as _log_e: - logging.debug("[DEBUG] final error logging failed: %s", _log_e) + raise LocalNetworkError( + "Unable to connect to the API server due to local network issues. " + "Please check your internet connection and try again." + ) from e + request_logger.log_request_response( + operation_id=operation_id, + request_method=method, + request_url=url, + request_headers=dict(payload_headers) if payload_headers else None, + request_params=dict(params) if params else None, + request_data=request_body_log, + error_message=f"ApiServerError: {str(e)}", + ) raise ApiServerError( f"The API server at {default_base_url()} is currently unreachable. " f"The service may be experiencing issues." diff --git a/comfy_api_nodes/util/download_helpers.py b/comfy_api_nodes/util/download_helpers.py index 78bcf1fa1..aa588d038 100644 --- a/comfy_api_nodes/util/download_helpers.py +++ b/comfy_api_nodes/util/download_helpers.py @@ -167,27 +167,25 @@ async def download_url_to_bytesio( with contextlib.suppress(Exception): dest.seek(0) - with contextlib.suppress(Exception): - request_logger.log_request_response( - operation_id=op_id, - request_method="GET", - request_url=url, - response_status_code=resp.status, - response_headers=dict(resp.headers), - response_content=f"[streamed {written} bytes to dest]", - ) + request_logger.log_request_response( + operation_id=op_id, + request_method="GET", + request_url=url, + response_status_code=resp.status, + response_headers=dict(resp.headers), + response_content=f"[streamed {written} bytes to dest]", + ) return except asyncio.CancelledError: raise ProcessingInterrupted("Task cancelled") from None except (ClientError, OSError) as e: if attempt <= max_retries: - with contextlib.suppress(Exception): - request_logger.log_request_response( - operation_id=op_id, - request_method="GET", - request_url=url, - error_message=f"{type(e).__name__}: {str(e)} (will retry)", - ) + request_logger.log_request_response( + operation_id=op_id, + request_method="GET", + request_url=url, + error_message=f"{type(e).__name__}: {str(e)} (will retry)", + ) await sleep_with_interrupt(delay, cls, None, None, None) delay *= retry_backoff continue diff --git a/comfy_api_nodes/util/request_logger.py b/comfy_api_nodes/util/request_logger.py index e0cb4428d..fe0543d9b 100644 --- a/comfy_api_nodes/util/request_logger.py +++ b/comfy_api_nodes/util/request_logger.py @@ -8,7 +8,6 @@ from typing import Any import folder_paths -# Get the logger instance logger = logging.getLogger(__name__) @@ -91,38 +90,41 @@ def log_request_response( Filenames are sanitized and length-limited for cross-platform safety. If we still fail to write, we fall back to appending into api.log. """ - log_dir = get_log_directory() - filepath = _build_log_filepath(log_dir, operation_id, request_url) - - log_content: list[str] = [] - log_content.append(f"Timestamp: {datetime.datetime.now().isoformat()}") - log_content.append(f"Operation ID: {operation_id}") - log_content.append("-" * 30 + " REQUEST " + "-" * 30) - log_content.append(f"Method: {request_method}") - log_content.append(f"URL: {request_url}") - if request_headers: - log_content.append(f"Headers:\n{_format_data_for_logging(request_headers)}") - if request_params: - log_content.append(f"Params:\n{_format_data_for_logging(request_params)}") - if request_data is not None: - log_content.append(f"Data/Body:\n{_format_data_for_logging(request_data)}") - - log_content.append("\n" + "-" * 30 + " RESPONSE " + "-" * 30) - if response_status_code is not None: - log_content.append(f"Status Code: {response_status_code}") - if response_headers: - log_content.append(f"Headers:\n{_format_data_for_logging(response_headers)}") - if response_content is not None: - log_content.append(f"Content:\n{_format_data_for_logging(response_content)}") - if error_message: - log_content.append(f"Error:\n{error_message}") - try: - with open(filepath, "w", encoding="utf-8") as f: - f.write("\n".join(log_content)) - logger.debug("API log saved to: %s", filepath) - except Exception as e: - logger.error("Error writing API log to %s: %s", filepath, str(e)) + log_dir = get_log_directory() + filepath = _build_log_filepath(log_dir, operation_id, request_url) + + log_content: list[str] = [] + log_content.append(f"Timestamp: {datetime.datetime.now().isoformat()}") + log_content.append(f"Operation ID: {operation_id}") + log_content.append("-" * 30 + " REQUEST " + "-" * 30) + log_content.append(f"Method: {request_method}") + log_content.append(f"URL: {request_url}") + if request_headers: + log_content.append(f"Headers:\n{_format_data_for_logging(request_headers)}") + if request_params: + log_content.append(f"Params:\n{_format_data_for_logging(request_params)}") + if request_data is not None: + log_content.append(f"Data/Body:\n{_format_data_for_logging(request_data)}") + + log_content.append("\n" + "-" * 30 + " RESPONSE " + "-" * 30) + if response_status_code is not None: + log_content.append(f"Status Code: {response_status_code}") + if response_headers: + log_content.append(f"Headers:\n{_format_data_for_logging(response_headers)}") + if response_content is not None: + log_content.append(f"Content:\n{_format_data_for_logging(response_content)}") + if error_message: + log_content.append(f"Error:\n{error_message}") + + try: + with open(filepath, "w", encoding="utf-8") as f: + f.write("\n".join(log_content)) + logger.debug("API log saved to: %s", filepath) + except Exception as e: + logger.error("Error writing API log to %s: %s", filepath, str(e)) + except Exception as _log_e: + logging.debug("[DEBUG] log_request_response failed: %s", _log_e) if __name__ == '__main__': diff --git a/comfy_api_nodes/util/upload_helpers.py b/comfy_api_nodes/util/upload_helpers.py index 83d936ce1..7cc565263 100644 --- a/comfy_api_nodes/util/upload_helpers.py +++ b/comfy_api_nodes/util/upload_helpers.py @@ -255,17 +255,14 @@ async def upload_file( monitor_task = asyncio.create_task(_monitor()) sess: aiohttp.ClientSession | None = None try: - try: - request_logger.log_request_response( - operation_id=operation_id, - request_method="PUT", - request_url=upload_url, - request_headers=headers or None, - request_params=None, - request_data=f"[File data {len(data)} bytes]", - ) - except Exception as e: - logging.debug("[DEBUG] upload request logging failed: %s", e) + request_logger.log_request_response( + operation_id=operation_id, + request_method="PUT", + request_url=upload_url, + request_headers=headers or None, + request_params=None, + request_data=f"[File data {len(data)} bytes]", + ) sess = aiohttp.ClientSession(timeout=timeout) req = sess.put(upload_url, data=data, headers=headers, skip_auto_headers=skip_auto_headers) @@ -311,31 +308,27 @@ async def upload_file( delay *= retry_backoff continue raise Exception(f"Failed to upload (HTTP {resp.status}).") - try: - request_logger.log_request_response( - operation_id=operation_id, - request_method="PUT", - request_url=upload_url, - response_status_code=resp.status, - response_headers=dict(resp.headers), - response_content="File uploaded successfully.", - ) - except Exception as e: - logging.debug("[DEBUG] upload response logging failed: %s", e) + request_logger.log_request_response( + operation_id=operation_id, + request_method="PUT", + request_url=upload_url, + response_status_code=resp.status, + response_headers=dict(resp.headers), + response_content="File uploaded successfully.", + ) return except asyncio.CancelledError: raise ProcessingInterrupted("Task cancelled") from None except (aiohttp.ClientError, OSError) as e: if attempt <= max_retries: - with contextlib.suppress(Exception): - request_logger.log_request_response( - operation_id=operation_id, - request_method="PUT", - request_url=upload_url, - request_headers=headers or None, - request_data=f"[File data {len(data)} bytes]", - error_message=f"{type(e).__name__}: {str(e)} (will retry)", - ) + request_logger.log_request_response( + operation_id=operation_id, + request_method="PUT", + request_url=upload_url, + request_headers=headers or None, + request_data=f"[File data {len(data)} bytes]", + error_message=f"{type(e).__name__}: {str(e)} (will retry)", + ) await sleep_with_interrupt( delay, cls, From 117e2143543dd649d47345e183748a82d48d12d3 Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Thu, 12 Feb 2026 16:51:50 -0800 Subject: [PATCH 623/647] ModelPatcherDynamic: force load non leaf weights (#12433) The current behaviour of the default ModelPatcher is to .to a model only if its fully loaded, which is how random non-leaf weights get loaded in non-LowVRAM conditions. The however means they never get loaded in dynamic_vram. In the dynamic_vram case, force load them to the GPU. --- comfy/model_patcher.py | 19 ++++++++++--------- 1 file changed, 10 insertions(+), 9 deletions(-) diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index f278fccac..b1d907ba4 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -679,18 +679,19 @@ class ModelPatcher: for key in list(self.pinned): self.unpin_weight(key) - def _load_list(self, prio_comfy_cast_weights=False): + def _load_list(self, prio_comfy_cast_weights=False, default_device=None): loading = [] for n, m in self.model.named_modules(): - params = [] - skip = False - for name, param in m.named_parameters(recurse=False): - params.append(name) + default = False + params = { name: param for name, param in m.named_parameters(recurse=False) } for name, param in m.named_parameters(recurse=True): if name not in params: - skip = True # skip random weights in non leaf modules + default = True # default random weights in non leaf modules break - if not skip and (hasattr(m, "comfy_cast_weights") or len(params) > 0): + if default and default_device is not None: + for param in params.values(): + param.data = param.data.to(device=default_device) + if not default and (hasattr(m, "comfy_cast_weights") or len(params) > 0): module_mem = comfy.model_management.module_size(m) module_offload_mem = module_mem if hasattr(m, "comfy_cast_weights"): @@ -1495,7 +1496,7 @@ class ModelPatcherDynamic(ModelPatcher): #with pin and unpin syncrhonization which can be expensive for small weights #with a high layer rate (e.g. autoregressive LLMs). #prioritize the non-comfy weights (note the order reverse). - loading = self._load_list(prio_comfy_cast_weights=True) + loading = self._load_list(prio_comfy_cast_weights=True, default_device=device_to) loading.sort(reverse=True) for x in loading: @@ -1579,7 +1580,7 @@ class ModelPatcherDynamic(ModelPatcher): return 0 if vbar is None else vbar.free_memory(memory_to_free) def partially_unload_ram(self, ram_to_unload): - loading = self._load_list(prio_comfy_cast_weights=True) + loading = self._load_list(prio_comfy_cast_weights=True, default_device=self.offload_device) for x in loading: _, _, _, _, m, _ = x ram_to_unload -= comfy.pinned_memory.unpin_memory(m) From ae79e33345dc893f8e0632c380c0e91dc09ac6e8 Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Thu, 12 Feb 2026 16:56:42 -0800 Subject: [PATCH 624/647] llama: use a more efficient rope implementation (#12434) Get rid of the cat and unary negation and inplace add-cmul the two halves of the rope. Precompute -sin once at the start of the model rather than every transformer block. This is slightly faster on both GPU and CPU bound setups. --- comfy/text_encoders/llama.py | 25 ++++++++++++++----------- 1 file changed, 14 insertions(+), 11 deletions(-) diff --git a/comfy/text_encoders/llama.py b/comfy/text_encoders/llama.py index b6735d210..54f3d5595 100644 --- a/comfy/text_encoders/llama.py +++ b/comfy/text_encoders/llama.py @@ -355,13 +355,6 @@ class RMSNorm(nn.Module): -def rotate_half(x): - """Rotates half the hidden dims of the input.""" - x1 = x[..., : x.shape[-1] // 2] - x2 = x[..., x.shape[-1] // 2 :] - return torch.cat((-x2, x1), dim=-1) - - def precompute_freqs_cis(head_dim, position_ids, theta, rope_scale=None, rope_dims=None, device=None): if not isinstance(theta, list): theta = [theta] @@ -390,20 +383,30 @@ def precompute_freqs_cis(head_dim, position_ids, theta, rope_scale=None, rope_di else: cos = cos.unsqueeze(1) sin = sin.unsqueeze(1) - out.append((cos, sin)) + sin_split = sin.shape[-1] // 2 + out.append((cos, sin[..., : sin_split], -sin[..., sin_split :])) if len(out) == 1: return out[0] return out - def apply_rope(xq, xk, freqs_cis): org_dtype = xq.dtype cos = freqs_cis[0] sin = freqs_cis[1] - q_embed = (xq * cos) + (rotate_half(xq) * sin) - k_embed = (xk * cos) + (rotate_half(xk) * sin) + nsin = freqs_cis[2] + + q_embed = (xq * cos) + q_split = q_embed.shape[-1] // 2 + q_embed[..., : q_split].addcmul_(xq[..., q_split :], nsin) + q_embed[..., q_split :].addcmul_(xq[..., : q_split], sin) + + k_embed = (xk * cos) + k_split = k_embed.shape[-1] // 2 + k_embed[..., : k_split].addcmul_(xk[..., k_split :], nsin) + k_embed[..., k_split :].addcmul_(xk[..., : k_split], sin) + return q_embed.to(org_dtype), k_embed.to(org_dtype) From e03fe8b5919a23a473cea6e53f916f7403c082a5 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Thu, 12 Feb 2026 20:29:12 -0800 Subject: [PATCH 625/647] Update command to install AMD stable linux pytorch. (#12437) --- README.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/README.md b/README.md index 96dc2904b..3ccdc9c19 100644 --- a/README.md +++ b/README.md @@ -227,7 +227,7 @@ Put your VAE in: models/vae AMD users can install rocm and pytorch with pip if you don't have it already installed, this is the command to install the stable version: -```pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm6.4``` +```pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm7.1``` This is the command to install the nightly with ROCm 7.1 which might have some performance improvements: From 8902907d7ab949ce42dd9b658b4a4582ed9fb630 Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Fri, 13 Feb 2026 12:29:37 -0800 Subject: [PATCH 626/647] dynamic_vram: Training fixes (#12442) --- comfy/model_patcher.py | 4 ++++ comfy_extras/nodes_train.py | 11 ++++++++++- 2 files changed, 14 insertions(+), 1 deletion(-) diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index b1d907ba4..67dce088e 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -1561,6 +1561,8 @@ class ModelPatcherDynamic(ModelPatcher): allocated_size += weight_size vbar.set_watermark_limit(allocated_size) + move_weight_functions(m, device_to) + logging.info(f"Model {self.model.__class__.__name__} prepared for dynamic VRAM loading. {allocated_size // (1024 ** 2)}MB Staged. {num_patches} patches attached.") self.model.device = device_to @@ -1601,6 +1603,8 @@ class ModelPatcherDynamic(ModelPatcher): if unpatch_weights: self.partially_unload_ram(1e32) self.partially_unload(None, 1e32) + for m in self.model.modules(): + move_weight_functions(m, device_to) def partially_load(self, device_to, extra_memory=0, force_patch_weights=False): assert not force_patch_weights #See above diff --git a/comfy_extras/nodes_train.py b/comfy_extras/nodes_train.py index 630eedc9f..aa2d88673 100644 --- a/comfy_extras/nodes_train.py +++ b/comfy_extras/nodes_train.py @@ -1035,7 +1035,7 @@ class TrainLoraNode(io.ComfyNode): io.Boolean.Input( "offloading", default=False, - tooltip="Depth level for gradient checkpointing.", + tooltip="Offload the Model to RAM. Requires Bypass Mode.", ), io.Combo.Input( "existing_lora", @@ -1124,6 +1124,15 @@ class TrainLoraNode(io.ComfyNode): lora_dtype = node_helpers.string_to_torch_dtype(lora_dtype) mp.set_model_compute_dtype(dtype) + if mp.is_dynamic(): + if not bypass_mode: + logging.info("Training MP is Dynamic - forcing bypass mode. Start comfy with --highvram to force weight diff mode") + bypass_mode = True + offloading = True + elif offloading: + if not bypass_mode: + logging.info("Training Offload selected - forcing bypass mode. Set bypass = True to remove this message") + # Prepare latents and compute counts latents, num_images, multi_res = _prepare_latents_and_count( latents, dtype, bucket_mode From e1add563f9e89026e8c4e8825a2b279fbd67d23a Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Fri, 13 Feb 2026 12:35:13 -0800 Subject: [PATCH 627/647] Use torch RMSNorm for flux models and refactor hunyuan video code. (#12432) --- comfy/controlnet.py | 1 + comfy/ldm/chroma/layers.py | 3 +- comfy/ldm/chroma_radiance/layers.py | 8 ++--- comfy/ldm/flux/layers.py | 53 +++++++---------------------- comfy/ldm/flux/model.py | 3 +- comfy/ldm/hunyuan_video/model.py | 11 +++--- comfy/lora_convert.py | 2 +- comfy/model_detection.py | 18 +++++++--- comfy/supported_models.py | 32 +++++++++++++++-- comfy/utils.py | 12 +++---- 10 files changed, 74 insertions(+), 69 deletions(-) diff --git a/comfy/controlnet.py b/comfy/controlnet.py index 9e1e704e0..8336412f2 100644 --- a/comfy/controlnet.py +++ b/comfy/controlnet.py @@ -560,6 +560,7 @@ def load_controlnet_hunyuandit(controlnet_data, model_options={}): def load_controlnet_flux_xlabs_mistoline(sd, mistoline=False, model_options={}): model_config, operations, load_device, unet_dtype, manual_cast_dtype, offload_device = controlnet_config(sd, model_options=model_options) control_model = comfy.ldm.flux.controlnet.ControlNetFlux(mistoline=mistoline, operations=operations, device=offload_device, dtype=unet_dtype, **model_config.unet_config) + sd = model_config.process_unet_state_dict(sd) control_model = controlnet_load_state_dict(control_model, sd) extra_conds = ['y', 'guidance'] control = ControlNet(control_model, load_device=load_device, manual_cast_dtype=manual_cast_dtype, extra_conds=extra_conds) diff --git a/comfy/ldm/chroma/layers.py b/comfy/ldm/chroma/layers.py index 2d5684348..df348a8ed 100644 --- a/comfy/ldm/chroma/layers.py +++ b/comfy/ldm/chroma/layers.py @@ -3,7 +3,6 @@ from torch import Tensor, nn from comfy.ldm.flux.layers import ( MLPEmbedder, - RMSNorm, ModulationOut, ) @@ -29,7 +28,7 @@ class Approximator(nn.Module): super().__init__() self.in_proj = operations.Linear(in_dim, hidden_dim, bias=True, dtype=dtype, device=device) self.layers = nn.ModuleList([MLPEmbedder(hidden_dim, hidden_dim, dtype=dtype, device=device, operations=operations) for x in range( n_layers)]) - self.norms = nn.ModuleList([RMSNorm(hidden_dim, dtype=dtype, device=device, operations=operations) for x in range( n_layers)]) + self.norms = nn.ModuleList([operations.RMSNorm(hidden_dim, dtype=dtype, device=device) for x in range( n_layers)]) self.out_proj = operations.Linear(hidden_dim, out_dim, dtype=dtype, device=device) @property diff --git a/comfy/ldm/chroma_radiance/layers.py b/comfy/ldm/chroma_radiance/layers.py index 3c7bc9b6b..08d31e0ba 100644 --- a/comfy/ldm/chroma_radiance/layers.py +++ b/comfy/ldm/chroma_radiance/layers.py @@ -4,8 +4,6 @@ from functools import lru_cache import torch from torch import nn -from comfy.ldm.flux.layers import RMSNorm - class NerfEmbedder(nn.Module): """ @@ -145,7 +143,7 @@ class NerfGLUBlock(nn.Module): # We now need to generate parameters for 3 matrices. total_params = 3 * hidden_size_x**2 * mlp_ratio self.param_generator = operations.Linear(hidden_size_s, total_params, dtype=dtype, device=device) - self.norm = RMSNorm(hidden_size_x, dtype=dtype, device=device, operations=operations) + self.norm = operations.RMSNorm(hidden_size_x, dtype=dtype, device=device) self.mlp_ratio = mlp_ratio @@ -178,7 +176,7 @@ class NerfGLUBlock(nn.Module): class NerfFinalLayer(nn.Module): def __init__(self, hidden_size, out_channels, dtype=None, device=None, operations=None): super().__init__() - self.norm = RMSNorm(hidden_size, dtype=dtype, device=device, operations=operations) + self.norm = operations.RMSNorm(hidden_size, dtype=dtype, device=device) self.linear = operations.Linear(hidden_size, out_channels, dtype=dtype, device=device) def forward(self, x: torch.Tensor) -> torch.Tensor: @@ -190,7 +188,7 @@ class NerfFinalLayer(nn.Module): class NerfFinalLayerConv(nn.Module): def __init__(self, hidden_size: int, out_channels: int, dtype=None, device=None, operations=None): super().__init__() - self.norm = RMSNorm(hidden_size, dtype=dtype, device=device, operations=operations) + self.norm = operations.RMSNorm(hidden_size, dtype=dtype, device=device) self.conv = operations.Conv2d( in_channels=hidden_size, out_channels=out_channels, diff --git a/comfy/ldm/flux/layers.py b/comfy/ldm/flux/layers.py index 60f2bdae2..1f2975fb1 100644 --- a/comfy/ldm/flux/layers.py +++ b/comfy/ldm/flux/layers.py @@ -5,8 +5,6 @@ import torch from torch import Tensor, nn from .math import attention, rope -import comfy.ops -import comfy.ldm.common_dit class EmbedND(nn.Module): @@ -87,20 +85,12 @@ def build_mlp(hidden_size, mlp_hidden_dim, mlp_silu_act=False, yak_mlp=False, dt operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device), ) -class RMSNorm(torch.nn.Module): - def __init__(self, dim: int, dtype=None, device=None, operations=None): - super().__init__() - self.scale = nn.Parameter(torch.empty((dim), dtype=dtype, device=device)) - - def forward(self, x: Tensor): - return comfy.ldm.common_dit.rms_norm(x, self.scale, 1e-6) - class QKNorm(torch.nn.Module): def __init__(self, dim: int, dtype=None, device=None, operations=None): super().__init__() - self.query_norm = RMSNorm(dim, dtype=dtype, device=device, operations=operations) - self.key_norm = RMSNorm(dim, dtype=dtype, device=device, operations=operations) + self.query_norm = operations.RMSNorm(dim, dtype=dtype, device=device) + self.key_norm = operations.RMSNorm(dim, dtype=dtype, device=device) def forward(self, q: Tensor, k: Tensor, v: Tensor) -> tuple: q = self.query_norm(q) @@ -169,7 +159,7 @@ class SiLUActivation(nn.Module): class DoubleStreamBlock(nn.Module): - def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False, flipped_img_txt=False, modulation=True, mlp_silu_act=False, proj_bias=True, yak_mlp=False, dtype=None, device=None, operations=None): + def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False, modulation=True, mlp_silu_act=False, proj_bias=True, yak_mlp=False, dtype=None, device=None, operations=None): super().__init__() mlp_hidden_dim = int(hidden_size * mlp_ratio) @@ -197,8 +187,6 @@ class DoubleStreamBlock(nn.Module): self.txt_mlp = build_mlp(hidden_size, mlp_hidden_dim, mlp_silu_act=mlp_silu_act, yak_mlp=yak_mlp, dtype=dtype, device=device, operations=operations) - self.flipped_img_txt = flipped_img_txt - def forward(self, img: Tensor, txt: Tensor, vec: Tensor, pe: Tensor, attn_mask=None, modulation_dims_img=None, modulation_dims_txt=None, transformer_options={}): if self.modulation: img_mod1, img_mod2 = self.img_mod(vec) @@ -224,32 +212,17 @@ class DoubleStreamBlock(nn.Module): del txt_qkv txt_q, txt_k = self.txt_attn.norm(txt_q, txt_k, txt_v) - if self.flipped_img_txt: - q = torch.cat((img_q, txt_q), dim=2) - del img_q, txt_q - k = torch.cat((img_k, txt_k), dim=2) - del img_k, txt_k - v = torch.cat((img_v, txt_v), dim=2) - del img_v, txt_v - # run actual attention - attn = attention(q, k, v, - pe=pe, mask=attn_mask, transformer_options=transformer_options) - del q, k, v + q = torch.cat((txt_q, img_q), dim=2) + del txt_q, img_q + k = torch.cat((txt_k, img_k), dim=2) + del txt_k, img_k + v = torch.cat((txt_v, img_v), dim=2) + del txt_v, img_v + # run actual attention + attn = attention(q, k, v, pe=pe, mask=attn_mask, transformer_options=transformer_options) + del q, k, v - img_attn, txt_attn = attn[:, : img.shape[1]], attn[:, img.shape[1]:] - else: - q = torch.cat((txt_q, img_q), dim=2) - del txt_q, img_q - k = torch.cat((txt_k, img_k), dim=2) - del txt_k, img_k - v = torch.cat((txt_v, img_v), dim=2) - del txt_v, img_v - # run actual attention - attn = attention(q, k, v, - pe=pe, mask=attn_mask, transformer_options=transformer_options) - del q, k, v - - txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1]:] + txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1]:] # calculate the img bloks img += apply_mod(self.img_attn.proj(img_attn), img_mod1.gate, None, modulation_dims_img) diff --git a/comfy/ldm/flux/model.py b/comfy/ldm/flux/model.py index f40c2a7a9..260ccad7e 100644 --- a/comfy/ldm/flux/model.py +++ b/comfy/ldm/flux/model.py @@ -16,7 +16,6 @@ from .layers import ( SingleStreamBlock, timestep_embedding, Modulation, - RMSNorm ) @dataclass @@ -81,7 +80,7 @@ class Flux(nn.Module): self.txt_in = operations.Linear(params.context_in_dim, self.hidden_size, bias=params.ops_bias, dtype=dtype, device=device) if params.txt_norm: - self.txt_norm = RMSNorm(params.context_in_dim, dtype=dtype, device=device, operations=operations) + self.txt_norm = operations.RMSNorm(params.context_in_dim, dtype=dtype, device=device) else: self.txt_norm = None diff --git a/comfy/ldm/hunyuan_video/model.py b/comfy/ldm/hunyuan_video/model.py index 55ab550f8..563f28f6b 100644 --- a/comfy/ldm/hunyuan_video/model.py +++ b/comfy/ldm/hunyuan_video/model.py @@ -241,7 +241,6 @@ class HunyuanVideo(nn.Module): self.num_heads, mlp_ratio=params.mlp_ratio, qkv_bias=params.qkv_bias, - flipped_img_txt=True, dtype=dtype, device=device, operations=operations ) for _ in range(params.depth) @@ -378,14 +377,14 @@ class HunyuanVideo(nn.Module): extra_txt_ids = torch.zeros((txt_ids.shape[0], txt_vision_states.shape[1], txt_ids.shape[-1]), device=txt_ids.device, dtype=txt_ids.dtype) txt_ids = torch.cat((txt_ids, extra_txt_ids), dim=1) - ids = torch.cat((img_ids, txt_ids), dim=1) + ids = torch.cat((txt_ids, img_ids), dim=1) pe = self.pe_embedder(ids) img_len = img.shape[1] if txt_mask is not None: attn_mask_len = img_len + txt.shape[1] attn_mask = torch.zeros((1, 1, attn_mask_len), dtype=img.dtype, device=img.device) - attn_mask[:, 0, img_len:] = txt_mask + attn_mask[:, 0, :txt.shape[1]] = txt_mask else: attn_mask = None @@ -413,7 +412,7 @@ class HunyuanVideo(nn.Module): if add is not None: img += add - img = torch.cat((img, txt), 1) + img = torch.cat((txt, img), 1) transformer_options["total_blocks"] = len(self.single_blocks) transformer_options["block_type"] = "single" @@ -435,9 +434,9 @@ class HunyuanVideo(nn.Module): if i < len(control_o): add = control_o[i] if add is not None: - img[:, : img_len] += add + img[:, txt.shape[1]: img_len + txt.shape[1]] += add - img = img[:, : img_len] + img = img[:, txt.shape[1]: img_len + txt.shape[1]] if ref_latent is not None: img = img[:, ref_latent.shape[1]:] diff --git a/comfy/lora_convert.py b/comfy/lora_convert.py index 9d8d21efe..749e81df3 100644 --- a/comfy/lora_convert.py +++ b/comfy/lora_convert.py @@ -5,7 +5,7 @@ import comfy.utils def convert_lora_bfl_control(sd): #BFL loras for Flux sd_out = {} for k in sd: - k_to = "diffusion_model.{}".format(k.replace(".lora_B.bias", ".diff_b").replace("_norm.scale", "_norm.scale.set_weight")) + k_to = "diffusion_model.{}".format(k.replace(".lora_B.bias", ".diff_b").replace("_norm.scale", "_norm.set_weight")) sd_out[k_to] = sd[k] sd_out["diffusion_model.img_in.reshape_weight"] = torch.tensor([sd["img_in.lora_B.weight"].shape[0], sd["img_in.lora_A.weight"].shape[1]]) diff --git a/comfy/model_detection.py b/comfy/model_detection.py index e8ad725df..30ea03e8e 100644 --- a/comfy/model_detection.py +++ b/comfy/model_detection.py @@ -19,6 +19,12 @@ def count_blocks(state_dict_keys, prefix_string): count += 1 return count +def any_suffix_in(keys, prefix, main, suffix_list=[]): + for x in suffix_list: + if "{}{}{}".format(prefix, main, x) in keys: + return True + return False + def calculate_transformer_depth(prefix, state_dict_keys, state_dict): context_dim = None use_linear_in_transformer = False @@ -186,7 +192,7 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): dit_config["meanflow_sum"] = False return dit_config - if '{}double_blocks.0.img_attn.norm.key_norm.scale'.format(key_prefix) in state_dict_keys and ('{}img_in.weight'.format(key_prefix) in state_dict_keys or f"{key_prefix}distilled_guidance_layer.norms.0.scale" in state_dict_keys): #Flux, Chroma or Chroma Radiance (has no img_in.weight) + if any_suffix_in(state_dict_keys, key_prefix, 'double_blocks.0.img_attn.norm.key_norm.', ["weight", "scale"]) and ('{}img_in.weight'.format(key_prefix) in state_dict_keys or any_suffix_in(state_dict_keys, key_prefix, 'distilled_guidance_layer.norms.0.', ["weight", "scale"])): #Flux, Chroma or Chroma Radiance (has no img_in.weight) dit_config = {} if '{}double_stream_modulation_img.lin.weight'.format(key_prefix) in state_dict_keys: dit_config["image_model"] = "flux2" @@ -241,7 +247,8 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): dit_config["depth"] = count_blocks(state_dict_keys, '{}double_blocks.'.format(key_prefix) + '{}.') dit_config["depth_single_blocks"] = count_blocks(state_dict_keys, '{}single_blocks.'.format(key_prefix) + '{}.') - if '{}distilled_guidance_layer.0.norms.0.scale'.format(key_prefix) in state_dict_keys or '{}distilled_guidance_layer.norms.0.scale'.format(key_prefix) in state_dict_keys: #Chroma + + if any_suffix_in(state_dict_keys, key_prefix, 'distilled_guidance_layer.0.norms.0.', ["weight", "scale"]) or any_suffix_in(state_dict_keys, key_prefix, 'distilled_guidance_layer.norms.0.', ["weight", "scale"]): #Chroma dit_config["image_model"] = "chroma" dit_config["in_channels"] = 64 dit_config["out_channels"] = 64 @@ -249,7 +256,8 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): dit_config["out_dim"] = 3072 dit_config["hidden_dim"] = 5120 dit_config["n_layers"] = 5 - if f"{key_prefix}nerf_blocks.0.norm.scale" in state_dict_keys: #Chroma Radiance + + if any_suffix_in(state_dict_keys, key_prefix, 'nerf_blocks.0.norm.', ["weight", "scale"]): #Chroma Radiance dit_config["image_model"] = "chroma_radiance" dit_config["in_channels"] = 3 dit_config["out_channels"] = 3 @@ -259,7 +267,7 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): dit_config["nerf_depth"] = 4 dit_config["nerf_max_freqs"] = 8 dit_config["nerf_tile_size"] = 512 - dit_config["nerf_final_head_type"] = "conv" if f"{key_prefix}nerf_final_layer_conv.norm.scale" in state_dict_keys else "linear" + dit_config["nerf_final_head_type"] = "conv" if any_suffix_in(state_dict_keys, key_prefix, 'nerf_final_layer_conv.norm.', ["weight", "scale"]) else "linear" dit_config["nerf_embedder_dtype"] = torch.float32 if "{}__x0__".format(key_prefix) in state_dict_keys: # x0 pred dit_config["use_x0"] = True @@ -268,7 +276,7 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): else: dit_config["guidance_embed"] = "{}guidance_in.in_layer.weight".format(key_prefix) in state_dict_keys dit_config["yak_mlp"] = '{}double_blocks.0.img_mlp.gate_proj.weight'.format(key_prefix) in state_dict_keys - dit_config["txt_norm"] = "{}txt_norm.scale".format(key_prefix) in state_dict_keys + dit_config["txt_norm"] = any_suffix_in(state_dict_keys, key_prefix, 'txt_norm.', ["weight", "scale"]) if dit_config["yak_mlp"] and dit_config["txt_norm"]: # Ovis model dit_config["txt_ids_dims"] = [1, 2] diff --git a/comfy/supported_models.py b/comfy/supported_models.py index d33db7507..c28be1716 100644 --- a/comfy/supported_models.py +++ b/comfy/supported_models.py @@ -710,6 +710,15 @@ class Flux(supported_models_base.BASE): supported_inference_dtypes = [torch.bfloat16, torch.float16, torch.float32] + def process_unet_state_dict(self, state_dict): + out_sd = {} + for k in list(state_dict.keys()): + key_out = k + if key_out.endswith("_norm.scale"): + key_out = "{}.weight".format(key_out[:-len(".scale")]) + out_sd[key_out] = state_dict[k] + return out_sd + vae_key_prefix = ["vae."] text_encoder_key_prefix = ["text_encoders."] @@ -898,11 +907,13 @@ class HunyuanVideo(supported_models_base.BASE): key_out = key_out.replace("txt_in.c_embedder.linear_1.", "txt_in.c_embedder.in_layer.").replace("txt_in.c_embedder.linear_2.", "txt_in.c_embedder.out_layer.") key_out = key_out.replace("_mod.linear.", "_mod.lin.").replace("_attn_qkv.", "_attn.qkv.") key_out = key_out.replace("mlp.fc1.", "mlp.0.").replace("mlp.fc2.", "mlp.2.") - key_out = key_out.replace("_attn_q_norm.weight", "_attn.norm.query_norm.scale").replace("_attn_k_norm.weight", "_attn.norm.key_norm.scale") - key_out = key_out.replace(".q_norm.weight", ".norm.query_norm.scale").replace(".k_norm.weight", ".norm.key_norm.scale") + key_out = key_out.replace("_attn_q_norm.weight", "_attn.norm.query_norm.weight").replace("_attn_k_norm.weight", "_attn.norm.key_norm.weight") + key_out = key_out.replace(".q_norm.weight", ".norm.query_norm.weight").replace(".k_norm.weight", ".norm.key_norm.weight") key_out = key_out.replace("_attn_proj.", "_attn.proj.") key_out = key_out.replace(".modulation.linear.", ".modulation.lin.") key_out = key_out.replace("_in.mlp.2.", "_in.out_layer.").replace("_in.mlp.0.", "_in.in_layer.") + if key_out.endswith(".scale"): + key_out = "{}.weight".format(key_out[:-len(".scale")]) out_sd[key_out] = state_dict[k] return out_sd @@ -1264,6 +1275,15 @@ class Hunyuan3Dv2(supported_models_base.BASE): latent_format = latent_formats.Hunyuan3Dv2 + def process_unet_state_dict(self, state_dict): + out_sd = {} + for k in list(state_dict.keys()): + key_out = k + if key_out.endswith(".scale"): + key_out = "{}.weight".format(key_out[:-len(".scale")]) + out_sd[key_out] = state_dict[k] + return out_sd + def process_unet_state_dict_for_saving(self, state_dict): replace_prefix = {"": "model."} return utils.state_dict_prefix_replace(state_dict, replace_prefix) @@ -1341,6 +1361,14 @@ class Chroma(supported_models_base.BASE): supported_inference_dtypes = [torch.bfloat16, torch.float16, torch.float32] + def process_unet_state_dict(self, state_dict): + out_sd = {} + for k in list(state_dict.keys()): + key_out = k + if key_out.endswith(".scale"): + key_out = "{}.weight".format(key_out[:-len(".scale")]) + out_sd[key_out] = state_dict[k] + return out_sd def get_model(self, state_dict, prefix="", device=None): out = model_base.Chroma(self, device=device) diff --git a/comfy/utils.py b/comfy/utils.py index e0a94e2e1..d553a7c1b 100644 --- a/comfy/utils.py +++ b/comfy/utils.py @@ -675,10 +675,10 @@ def flux_to_diffusers(mmdit_config, output_prefix=""): "ff_context.linear_in.bias": "txt_mlp.0.bias", "ff_context.linear_out.weight": "txt_mlp.2.weight", "ff_context.linear_out.bias": "txt_mlp.2.bias", - "attn.norm_q.weight": "img_attn.norm.query_norm.scale", - "attn.norm_k.weight": "img_attn.norm.key_norm.scale", - "attn.norm_added_q.weight": "txt_attn.norm.query_norm.scale", - "attn.norm_added_k.weight": "txt_attn.norm.key_norm.scale", + "attn.norm_q.weight": "img_attn.norm.query_norm.weight", + "attn.norm_k.weight": "img_attn.norm.key_norm.weight", + "attn.norm_added_q.weight": "txt_attn.norm.query_norm.weight", + "attn.norm_added_k.weight": "txt_attn.norm.key_norm.weight", } for k in block_map: @@ -701,8 +701,8 @@ def flux_to_diffusers(mmdit_config, output_prefix=""): "norm.linear.bias": "modulation.lin.bias", "proj_out.weight": "linear2.weight", "proj_out.bias": "linear2.bias", - "attn.norm_q.weight": "norm.query_norm.scale", - "attn.norm_k.weight": "norm.key_norm.scale", + "attn.norm_q.weight": "norm.query_norm.weight", + "attn.norm_k.weight": "norm.key_norm.weight", "attn.to_qkv_mlp_proj.weight": "linear1.weight", # Flux 2 "attn.to_out.weight": "linear2.weight", # Flux 2 } From 831351a29e91ea758437227c2f3c915a6be6d1a6 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Fri, 13 Feb 2026 17:15:23 -0800 Subject: [PATCH 628/647] Support generating attention masks for left padded text encoders. (#12454) --- comfy/sd1_clip.py | 15 +++++++++++---- comfy/text_encoders/ace15.py | 8 +------- 2 files changed, 12 insertions(+), 11 deletions(-) diff --git a/comfy/sd1_clip.py b/comfy/sd1_clip.py index 4c817d468..b564d1529 100644 --- a/comfy/sd1_clip.py +++ b/comfy/sd1_clip.py @@ -171,8 +171,9 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder): def process_tokens(self, tokens, device): end_token = self.special_tokens.get("end", None) + pad_token = self.special_tokens.get("pad", -1) if end_token is None: - cmp_token = self.special_tokens.get("pad", -1) + cmp_token = pad_token else: cmp_token = end_token @@ -186,15 +187,21 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder): other_embeds = [] eos = False index = 0 + left_pad = False for y in x: if isinstance(y, numbers.Integral): - if eos: + token = int(y) + if index == 0 and token == pad_token: + left_pad = True + + if eos or (left_pad and token == pad_token): attention_mask.append(0) else: attention_mask.append(1) - token = int(y) + left_pad = False + tokens_temp += [token] - if not eos and token == cmp_token: + if not eos and token == cmp_token and not left_pad: if end_token is None: attention_mask[-1] = 0 eos = True diff --git a/comfy/text_encoders/ace15.py b/comfy/text_encoders/ace15.py index 0fdd4669f..f135d74c1 100644 --- a/comfy/text_encoders/ace15.py +++ b/comfy/text_encoders/ace15.py @@ -10,7 +10,6 @@ import comfy.utils def sample_manual_loop_no_classes( model, ids=None, - paddings=[], execution_dtype=None, cfg_scale: float = 2.0, temperature: float = 0.85, @@ -36,9 +35,6 @@ def sample_manual_loop_no_classes( embeds, attention_mask, num_tokens, embeds_info = model.process_tokens(ids, device) embeds_batch = embeds.shape[0] - for i, t in enumerate(paddings): - attention_mask[i, :t] = 0 - attention_mask[i, t:] = 1 output_audio_codes = [] past_key_values = [] @@ -135,13 +131,11 @@ def generate_audio_codes(model, positive, negative, min_tokens=1, max_tokens=102 pos_pad = (len(negative) - len(positive)) positive = [model.special_tokens["pad"]] * pos_pad + positive - paddings = [pos_pad, neg_pad] ids = [positive, negative] else: - paddings = [] ids = [positive] - return sample_manual_loop_no_classes(model, ids, paddings, cfg_scale=cfg_scale, temperature=temperature, top_p=top_p, top_k=top_k, min_p=min_p, seed=seed, min_tokens=min_tokens, max_new_tokens=max_tokens) + return sample_manual_loop_no_classes(model, ids, cfg_scale=cfg_scale, temperature=temperature, top_p=top_p, top_k=top_k, min_p=min_p, seed=seed, min_tokens=min_tokens, max_new_tokens=max_tokens) class ACE15Tokenizer(sd1_clip.SD1Tokenizer): From 726af73867c18c5ca8b980a2c28401d77e5b365a Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Fri, 13 Feb 2026 17:21:10 -0800 Subject: [PATCH 629/647] Fix some custom nodes. (#12455) --- comfy/ldm/flux/layers.py | 2 ++ 1 file changed, 2 insertions(+) diff --git a/comfy/ldm/flux/layers.py b/comfy/ldm/flux/layers.py index 1f2975fb1..3518a1922 100644 --- a/comfy/ldm/flux/layers.py +++ b/comfy/ldm/flux/layers.py @@ -6,6 +6,8 @@ from torch import Tensor, nn from .math import attention, rope +# Fix import for some custom nodes, TODO: delete eventually. +RMSNorm = None class EmbedND(nn.Module): def __init__(self, dim: int, theta: int, axes_dim: list): From 712efb466b9379e6761802c44027783d37d96a87 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Fri, 13 Feb 2026 18:56:54 -0800 Subject: [PATCH 630/647] Add left padding to LTXAV text encoder. (#12456) --- comfy/text_encoders/lt.py | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/comfy/text_encoders/lt.py b/comfy/text_encoders/lt.py index 3f87dfd6a..9cf87c0b2 100644 --- a/comfy/text_encoders/lt.py +++ b/comfy/text_encoders/lt.py @@ -25,7 +25,7 @@ def ltxv_te(*args, **kwargs): class Gemma3_12BTokenizer(sd1_clip.SDTokenizer): def __init__(self, embedding_directory=None, tokenizer_data={}): tokenizer = tokenizer_data.get("spiece_model", None) - super().__init__(tokenizer, pad_with_end=False, embedding_size=3840, embedding_key='gemma3_12b', tokenizer_class=SPieceTokenizer, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, disable_weights=True, tokenizer_args={"add_bos": True, "add_eos": False}, tokenizer_data=tokenizer_data) + super().__init__(tokenizer, pad_with_end=False, embedding_size=3840, embedding_key='gemma3_12b', tokenizer_class=SPieceTokenizer, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=512, pad_left=True, disable_weights=True, tokenizer_args={"add_bos": True, "add_eos": False}, tokenizer_data=tokenizer_data) def state_dict(self): return {"spiece_model": self.tokenizer.serialize_model()} @@ -97,6 +97,7 @@ class LTXAVTEModel(torch.nn.Module): token_weight_pairs = token_weight_pairs["gemma3_12b"] out, pooled, extra = self.gemma3_12b.encode_token_weights(token_weight_pairs) + out = out[:, :, -torch.sum(extra["attention_mask"]).item():] out_device = out.device if comfy.model_management.should_use_bf16(self.execution_device): out = out.to(device=self.execution_device, dtype=torch.bfloat16) @@ -138,6 +139,7 @@ class LTXAVTEModel(torch.nn.Module): token_weight_pairs = token_weight_pairs.get("gemma3_12b", []) num_tokens = sum(map(lambda a: len(a), token_weight_pairs)) + num_tokens = max(num_tokens, 64) return num_tokens * constant * 1024 * 1024 def ltxav_te(dtype_llama=None, llama_quantization_metadata=None): From dc9822b7df4785e77690e93b0e09feaff01e2e12 Mon Sep 17 00:00:00 2001 From: krigeta <75309361+krigeta@users.noreply.github.com> Date: Sat, 14 Feb 2026 08:53:52 +0530 Subject: [PATCH 631/647] Add working Qwen 2512 ControlNet (Fun ControlNet) support (#12359) --- comfy/controlnet.py | 73 +++++++++++ comfy/ldm/qwen_image/controlnet.py | 190 +++++++++++++++++++++++++++++ 2 files changed, 263 insertions(+) diff --git a/comfy/controlnet.py b/comfy/controlnet.py index 8336412f2..ba670b16d 100644 --- a/comfy/controlnet.py +++ b/comfy/controlnet.py @@ -297,6 +297,30 @@ class ControlNet(ControlBase): self.model_sampling_current = None super().cleanup() + +class QwenFunControlNet(ControlNet): + def get_control(self, x_noisy, t, cond, batched_number, transformer_options): + # Fun checkpoints are more sensitive to high strengths in the generic + # ControlNet merge path. Use a soft response curve so strength=1.0 stays + # unchanged while >1 grows more gently. + original_strength = self.strength + self.strength = math.sqrt(max(self.strength, 0.0)) + try: + return super().get_control(x_noisy, t, cond, batched_number, transformer_options) + finally: + self.strength = original_strength + + def pre_run(self, model, percent_to_timestep_function): + super().pre_run(model, percent_to_timestep_function) + self.set_extra_arg("base_model", model.diffusion_model) + + def copy(self): + c = QwenFunControlNet(None, global_average_pooling=self.global_average_pooling, load_device=self.load_device, manual_cast_dtype=self.manual_cast_dtype) + c.control_model = self.control_model + c.control_model_wrapped = self.control_model_wrapped + self.copy_to(c) + return c + class ControlLoraOps: class Linear(torch.nn.Module, comfy.ops.CastWeightBiasOp): def __init__(self, in_features: int, out_features: int, bias: bool = True, @@ -606,6 +630,53 @@ def load_controlnet_qwen_instantx(sd, model_options={}): control = ControlNet(control_model, compression_ratio=1, latent_format=latent_format, concat_mask=concat_mask, load_device=load_device, manual_cast_dtype=manual_cast_dtype, extra_conds=extra_conds) return control + +def load_controlnet_qwen_fun(sd, model_options={}): + load_device = comfy.model_management.get_torch_device() + weight_dtype = comfy.utils.weight_dtype(sd) + unet_dtype = model_options.get("dtype", weight_dtype) + manual_cast_dtype = comfy.model_management.unet_manual_cast(unet_dtype, load_device) + + operations = model_options.get("custom_operations", None) + if operations is None: + operations = comfy.ops.pick_operations(unet_dtype, manual_cast_dtype, disable_fast_fp8=True) + + in_features = sd["control_img_in.weight"].shape[1] + inner_dim = sd["control_img_in.weight"].shape[0] + + block_weight = sd["control_blocks.0.attn.to_q.weight"] + attention_head_dim = sd["control_blocks.0.attn.norm_q.weight"].shape[0] + num_attention_heads = max(1, block_weight.shape[0] // max(1, attention_head_dim)) + + model = comfy.ldm.qwen_image.controlnet.QwenImageFunControlNetModel( + control_in_features=in_features, + inner_dim=inner_dim, + num_attention_heads=num_attention_heads, + attention_head_dim=attention_head_dim, + num_control_blocks=5, + main_model_double=60, + injection_layers=(0, 12, 24, 36, 48), + operations=operations, + device=comfy.model_management.unet_offload_device(), + dtype=unet_dtype, + ) + model = controlnet_load_state_dict(model, sd) + + latent_format = comfy.latent_formats.Wan21() + control = QwenFunControlNet( + model, + compression_ratio=1, + latent_format=latent_format, + # Fun checkpoints already expect their own 33-channel context handling. + # Enabling generic concat_mask injects an extra mask channel at apply-time + # and breaks the intended fallback packing path. + concat_mask=False, + load_device=load_device, + manual_cast_dtype=manual_cast_dtype, + extra_conds=[], + ) + return control + def convert_mistoline(sd): return comfy.utils.state_dict_prefix_replace(sd, {"single_controlnet_blocks.": "controlnet_single_blocks."}) @@ -683,6 +754,8 @@ def load_controlnet_state_dict(state_dict, model=None, model_options={}): return load_controlnet_qwen_instantx(controlnet_data, model_options=model_options) elif "controlnet_x_embedder.weight" in controlnet_data: return load_controlnet_flux_instantx(controlnet_data, model_options=model_options) + elif "control_blocks.0.after_proj.weight" in controlnet_data and "control_img_in.weight" in controlnet_data: + return load_controlnet_qwen_fun(controlnet_data, model_options=model_options) elif "controlnet_blocks.0.linear.weight" in controlnet_data: #mistoline flux return load_controlnet_flux_xlabs_mistoline(convert_mistoline(controlnet_data), mistoline=True, model_options=model_options) diff --git a/comfy/ldm/qwen_image/controlnet.py b/comfy/ldm/qwen_image/controlnet.py index a6d408104..c0aae9240 100644 --- a/comfy/ldm/qwen_image/controlnet.py +++ b/comfy/ldm/qwen_image/controlnet.py @@ -2,6 +2,196 @@ import torch import math from .model import QwenImageTransformer2DModel +from .model import QwenImageTransformerBlock + + +class QwenImageFunControlBlock(QwenImageTransformerBlock): + def __init__(self, dim, num_attention_heads, attention_head_dim, has_before_proj=False, dtype=None, device=None, operations=None): + super().__init__( + dim=dim, + num_attention_heads=num_attention_heads, + attention_head_dim=attention_head_dim, + dtype=dtype, + device=device, + operations=operations, + ) + self.has_before_proj = has_before_proj + if has_before_proj: + self.before_proj = operations.Linear(dim, dim, device=device, dtype=dtype) + self.after_proj = operations.Linear(dim, dim, device=device, dtype=dtype) + + +class QwenImageFunControlNetModel(torch.nn.Module): + def __init__( + self, + control_in_features=132, + inner_dim=3072, + num_attention_heads=24, + attention_head_dim=128, + num_control_blocks=5, + main_model_double=60, + injection_layers=(0, 12, 24, 36, 48), + dtype=None, + device=None, + operations=None, + ): + super().__init__() + self.dtype = dtype + self.main_model_double = main_model_double + self.injection_layers = tuple(injection_layers) + # Keep base hint scaling at 1.0 so user-facing strength behaves similarly + # to the reference Gen2/VideoX implementation around strength=1. + self.hint_scale = 1.0 + self.control_img_in = operations.Linear(control_in_features, inner_dim, device=device, dtype=dtype) + + self.control_blocks = torch.nn.ModuleList([]) + for i in range(num_control_blocks): + self.control_blocks.append( + QwenImageFunControlBlock( + dim=inner_dim, + num_attention_heads=num_attention_heads, + attention_head_dim=attention_head_dim, + has_before_proj=(i == 0), + dtype=dtype, + device=device, + operations=operations, + ) + ) + + def _process_hint_tokens(self, hint): + if hint is None: + return None + if hint.ndim == 4: + hint = hint.unsqueeze(2) + + # Fun checkpoints are trained with 33 latent channels before 2x2 packing: + # [control_latent(16), mask(1), inpaint_latent(16)] -> 132 features. + # Default behavior (no inpaint input in stock Apply ControlNet) should use + # zeros for mask/inpaint branches, matching VideoX fallback semantics. + expected_c = self.control_img_in.weight.shape[1] // 4 + if hint.shape[1] == 16 and expected_c == 33: + zeros_mask = torch.zeros_like(hint[:, :1]) + zeros_inpaint = torch.zeros_like(hint) + hint = torch.cat([hint, zeros_mask, zeros_inpaint], dim=1) + + bs, c, t, h, w = hint.shape + hidden_states = torch.nn.functional.pad(hint, (0, w % 2, 0, h % 2)) + orig_shape = hidden_states.shape + hidden_states = hidden_states.view( + orig_shape[0], + orig_shape[1], + orig_shape[-3], + orig_shape[-2] // 2, + 2, + orig_shape[-1] // 2, + 2, + ) + hidden_states = hidden_states.permute(0, 2, 3, 5, 1, 4, 6) + hidden_states = hidden_states.reshape( + bs, + t * ((h + 1) // 2) * ((w + 1) // 2), + c * 4, + ) + + expected_in = self.control_img_in.weight.shape[1] + cur_in = hidden_states.shape[-1] + if cur_in < expected_in: + pad = torch.zeros( + (hidden_states.shape[0], hidden_states.shape[1], expected_in - cur_in), + device=hidden_states.device, + dtype=hidden_states.dtype, + ) + hidden_states = torch.cat([hidden_states, pad], dim=-1) + elif cur_in > expected_in: + hidden_states = hidden_states[:, :, :expected_in] + + return hidden_states + + def forward( + self, + x, + timesteps, + context, + attention_mask=None, + guidance: torch.Tensor = None, + hint=None, + transformer_options={}, + base_model=None, + **kwargs, + ): + if base_model is None: + raise RuntimeError("Qwen Fun ControlNet requires a QwenImage base model at runtime.") + + encoder_hidden_states_mask = attention_mask + # Keep attention mask disabled inside Fun control blocks to mirror + # VideoX behavior (they rely on seq lengths for RoPE, not masked attention). + encoder_hidden_states_mask = None + + hidden_states, img_ids, _ = base_model.process_img(x) + hint_tokens = self._process_hint_tokens(hint) + if hint_tokens is None: + raise RuntimeError("Qwen Fun ControlNet requires a control hint image.") + + if hint_tokens.shape[1] != hidden_states.shape[1]: + max_tokens = min(hint_tokens.shape[1], hidden_states.shape[1]) + hint_tokens = hint_tokens[:, :max_tokens] + hidden_states = hidden_states[:, :max_tokens] + img_ids = img_ids[:, :max_tokens] + + txt_start = round( + max( + ((x.shape[-1] + (base_model.patch_size // 2)) // base_model.patch_size) // 2, + ((x.shape[-2] + (base_model.patch_size // 2)) // base_model.patch_size) // 2, + ) + ) + txt_ids = torch.arange(txt_start, txt_start + context.shape[1], device=x.device).reshape(1, -1, 1).repeat(x.shape[0], 1, 3) + ids = torch.cat((txt_ids, img_ids), dim=1) + image_rotary_emb = base_model.pe_embedder(ids).to(x.dtype).contiguous() + + hidden_states = base_model.img_in(hidden_states) + encoder_hidden_states = base_model.txt_norm(context) + encoder_hidden_states = base_model.txt_in(encoder_hidden_states) + + if guidance is not None: + guidance = guidance * 1000 + + temb = ( + base_model.time_text_embed(timesteps, hidden_states) + if guidance is None + else base_model.time_text_embed(timesteps, guidance, hidden_states) + ) + + c = self.control_img_in(hint_tokens) + + for i, block in enumerate(self.control_blocks): + if i == 0: + c_in = block.before_proj(c) + hidden_states + all_c = [] + else: + all_c = list(torch.unbind(c, dim=0)) + c_in = all_c.pop(-1) + + encoder_hidden_states, c_out = block( + hidden_states=c_in, + encoder_hidden_states=encoder_hidden_states, + encoder_hidden_states_mask=encoder_hidden_states_mask, + temb=temb, + image_rotary_emb=image_rotary_emb, + transformer_options=transformer_options, + ) + + c_skip = block.after_proj(c_out) * self.hint_scale + all_c += [c_skip, c_out] + c = torch.stack(all_c, dim=0) + + hints = torch.unbind(c, dim=0)[:-1] + + controlnet_block_samples = [None] * self.main_model_double + for local_idx, base_idx in enumerate(self.injection_layers): + if local_idx < len(hints) and base_idx < len(controlnet_block_samples): + controlnet_block_samples[base_idx] = hints[local_idx] + + return {"input": controlnet_block_samples} class QwenImageControlNetModel(QwenImageTransformer2DModel): From df1e5e85142746a745a56572b705406b273a594c Mon Sep 17 00:00:00 2001 From: Christian Byrne Date: Sat, 14 Feb 2026 11:01:10 -0800 Subject: [PATCH 632/647] Update frontend package to 1.38.14 (#12469) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 7de6a413c..e939e486a 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,4 +1,4 @@ -comfyui-frontend-package==1.38.13 +comfyui-frontend-package==1.38.14 comfyui-workflow-templates==0.8.38 comfyui-embedded-docs==0.4.1 torch From e1ede29d827d573262caede8aeb6cbc98c323c81 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sat, 14 Feb 2026 19:53:52 -0800 Subject: [PATCH 633/647] Remove unsafe pickle loading code that was used on pytorch older than 2.4 (#12473) ComfyUI hasn't started on pytorch 2.4 since last month. --- comfy/checkpoint_pickle.py | 13 ------------- comfy/utils.py | 25 +++++++++++-------------- 2 files changed, 11 insertions(+), 27 deletions(-) delete mode 100644 comfy/checkpoint_pickle.py diff --git a/comfy/checkpoint_pickle.py b/comfy/checkpoint_pickle.py deleted file mode 100644 index 206551d3c..000000000 --- a/comfy/checkpoint_pickle.py +++ /dev/null @@ -1,13 +0,0 @@ -import pickle - -load = pickle.load - -class Empty: - pass - -class Unpickler(pickle.Unpickler): - def find_class(self, module, name): - #TODO: safe unpickle - if module.startswith("pytorch_lightning"): - return Empty - return super().find_class(module, name) diff --git a/comfy/utils.py b/comfy/utils.py index d553a7c1b..c1ce540b5 100644 --- a/comfy/utils.py +++ b/comfy/utils.py @@ -20,7 +20,7 @@ import torch import math import struct -import comfy.checkpoint_pickle +import comfy.memory_management import safetensors.torch import numpy as np from PIL import Image @@ -38,26 +38,26 @@ import warnings MMAP_TORCH_FILES = args.mmap_torch_files DISABLE_MMAP = args.disable_mmap -ALWAYS_SAFE_LOAD = False -if hasattr(torch.serialization, "add_safe_globals"): # TODO: this was added in pytorch 2.4, the unsafe path should be removed once earlier versions are deprecated + +if True: # ckpt/pt file whitelist for safe loading of old sd files class ModelCheckpoint: pass ModelCheckpoint.__module__ = "pytorch_lightning.callbacks.model_checkpoint" def scalar(*args, **kwargs): - from numpy.core.multiarray import scalar as sc - return sc(*args, **kwargs) + return None scalar.__module__ = "numpy.core.multiarray" from numpy import dtype from numpy.dtypes import Float64DType - from _codecs import encode + + def encode(*args, **kwargs): # no longer necessary on newer torch + return None + encode.__module__ = "_codecs" torch.serialization.add_safe_globals([ModelCheckpoint, scalar, dtype, Float64DType, encode]) - ALWAYS_SAFE_LOAD = True logging.info("Checkpoint files will always be loaded safely.") -else: - logging.warning("Warning, you are using an old pytorch version and some ckpt/pt files might be loaded unsafely. Upgrading to 2.4 or above is recommended as older versions of pytorch are no longer supported.") + # Current as of safetensors 0.7.0 _TYPES = { @@ -140,11 +140,8 @@ def load_torch_file(ckpt, safe_load=False, device=None, return_metadata=False): if MMAP_TORCH_FILES: torch_args["mmap"] = True - if safe_load or ALWAYS_SAFE_LOAD: - pl_sd = torch.load(ckpt, map_location=device, weights_only=True, **torch_args) - else: - logging.warning("WARNING: loading {} unsafely, upgrade your pytorch to 2.4 or newer to load this file safely.".format(ckpt)) - pl_sd = torch.load(ckpt, map_location=device, pickle_module=comfy.checkpoint_pickle) + pl_sd = torch.load(ckpt, map_location=device, weights_only=True, **torch_args) + if "state_dict" in pl_sd: sd = pl_sd["state_dict"] else: From ce4a1ab48d9f723eeaac37f88dde55086b1f233f Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Sun, 15 Feb 2026 11:31:59 +0200 Subject: [PATCH 634/647] chore(api-nodes): remove "gpt-4o" model (#12467) --- comfy_api_nodes/nodes_openai.py | 6 ------ 1 file changed, 6 deletions(-) diff --git a/comfy_api_nodes/nodes_openai.py b/comfy_api_nodes/nodes_openai.py index f05aaab7b..332107a82 100644 --- a/comfy_api_nodes/nodes_openai.py +++ b/comfy_api_nodes/nodes_openai.py @@ -43,7 +43,6 @@ class SupportedOpenAIModel(str, Enum): o1 = "o1" o3 = "o3" o1_pro = "o1-pro" - gpt_4o = "gpt-4o" gpt_4_1 = "gpt-4.1" gpt_4_1_mini = "gpt-4.1-mini" gpt_4_1_nano = "gpt-4.1-nano" @@ -649,11 +648,6 @@ class OpenAIChatNode(IO.ComfyNode): "usd": [0.01, 0.04], "format": { "approximate": true, "separator": "-", "suffix": " per 1K tokens" } } - : $contains($m, "gpt-4o") ? { - "type": "list_usd", - "usd": [0.0025, 0.01], - "format": { "approximate": true, "separator": "-", "suffix": " per 1K tokens" } - } : $contains($m, "gpt-4.1-nano") ? { "type": "list_usd", "usd": [0.0001, 0.0004], From 596ed686919f11f75be3cf9a79977a07d64002c5 Mon Sep 17 00:00:00 2001 From: Jedrzej Kosinski Date: Sun, 15 Feb 2026 02:12:30 -0800 Subject: [PATCH 635/647] Node Replacement API (#12014) --- app/node_replace_manager.py | 105 ++++++++++++++++++++++++++ comfy_api/feature_flags.py | 1 + comfy_api/latest/__init__.py | 13 +++- comfy_api/latest/_io.py | 63 ++++++++++++++++ comfy_extras/nodes_post_processing.py | 1 + comfy_extras/nodes_replacements.py | 103 +++++++++++++++++++++++++ nodes.py | 2 + server.py | 5 ++ 8 files changed, 291 insertions(+), 2 deletions(-) create mode 100644 app/node_replace_manager.py create mode 100644 comfy_extras/nodes_replacements.py diff --git a/app/node_replace_manager.py b/app/node_replace_manager.py new file mode 100644 index 000000000..03b603c70 --- /dev/null +++ b/app/node_replace_manager.py @@ -0,0 +1,105 @@ +from __future__ import annotations + +from aiohttp import web + +from typing import TYPE_CHECKING, TypedDict +if TYPE_CHECKING: + from comfy_api.latest._io_public import NodeReplace + +from comfy_execution.graph_utils import is_link +import nodes + +class NodeStruct(TypedDict): + inputs: dict[str, str | int | float | bool | tuple[str, int]] + class_type: str + _meta: dict[str, str] + +def copy_node_struct(node_struct: NodeStruct, empty_inputs: bool = False) -> NodeStruct: + new_node_struct = node_struct.copy() + if empty_inputs: + new_node_struct["inputs"] = {} + else: + new_node_struct["inputs"] = node_struct["inputs"].copy() + new_node_struct["_meta"] = node_struct["_meta"].copy() + return new_node_struct + + +class NodeReplaceManager: + """Manages node replacement registrations.""" + + def __init__(self): + self._replacements: dict[str, list[NodeReplace]] = {} + + def register(self, node_replace: NodeReplace): + """Register a node replacement mapping.""" + self._replacements.setdefault(node_replace.old_node_id, []).append(node_replace) + + def get_replacement(self, old_node_id: str) -> list[NodeReplace] | None: + """Get replacements for an old node ID.""" + return self._replacements.get(old_node_id) + + def has_replacement(self, old_node_id: str) -> bool: + """Check if a replacement exists for an old node ID.""" + return old_node_id in self._replacements + + def apply_replacements(self, prompt: dict[str, NodeStruct]): + connections: dict[str, list[tuple[str, str, int]]] = {} + need_replacement: set[str] = set() + for node_number, node_struct in prompt.items(): + class_type = node_struct["class_type"] + # need replacement if not in NODE_CLASS_MAPPINGS and has replacement + if class_type not in nodes.NODE_CLASS_MAPPINGS.keys() and self.has_replacement(class_type): + need_replacement.add(node_number) + # keep track of connections + for input_id, input_value in node_struct["inputs"].items(): + if is_link(input_value): + conn_number = input_value[0] + connections.setdefault(conn_number, []).append((node_number, input_id, input_value[1])) + for node_number in need_replacement: + node_struct = prompt[node_number] + class_type = node_struct["class_type"] + replacements = self.get_replacement(class_type) + if replacements is None: + continue + # just use the first replacement + replacement = replacements[0] + new_node_id = replacement.new_node_id + # if replacement is not a valid node, skip trying to replace it as will only cause confusion + if new_node_id not in nodes.NODE_CLASS_MAPPINGS.keys(): + continue + # first, replace node id (class_type) + new_node_struct = copy_node_struct(node_struct, empty_inputs=True) + new_node_struct["class_type"] = new_node_id + # TODO: consider replacing display_name in _meta as well for error reporting purposes; would need to query node schema + # second, replace inputs + if replacement.input_mapping is not None: + for input_map in replacement.input_mapping: + if "set_value" in input_map: + new_node_struct["inputs"][input_map["new_id"]] = input_map["set_value"] + elif "old_id" in input_map: + new_node_struct["inputs"][input_map["new_id"]] = node_struct["inputs"][input_map["old_id"]] + # finalize input replacement + prompt[node_number] = new_node_struct + # third, replace outputs + if replacement.output_mapping is not None: + # re-mapping outputs requires changing the input values of nodes that receive connections from this one + if node_number in connections: + for conns in connections[node_number]: + conn_node_number, conn_input_id, old_output_idx = conns + for output_map in replacement.output_mapping: + if output_map["old_idx"] == old_output_idx: + new_output_idx = output_map["new_idx"] + previous_input = prompt[conn_node_number]["inputs"][conn_input_id] + previous_input[1] = new_output_idx + + def as_dict(self): + """Serialize all replacements to dict.""" + return { + k: [v.as_dict() for v in v_list] + for k, v_list in self._replacements.items() + } + + def add_routes(self, routes): + @routes.get("/node_replacements") + async def get_node_replacements(request): + return web.json_response(self.as_dict()) diff --git a/comfy_api/feature_flags.py b/comfy_api/feature_flags.py index de167f037..a90a5ca40 100644 --- a/comfy_api/feature_flags.py +++ b/comfy_api/feature_flags.py @@ -14,6 +14,7 @@ SERVER_FEATURE_FLAGS: dict[str, Any] = { "supports_preview_metadata": True, "max_upload_size": args.max_upload_size * 1024 * 1024, # Convert MB to bytes "extension": {"manager": {"supports_v4": True}}, + "node_replacements": True, } diff --git a/comfy_api/latest/__init__.py b/comfy_api/latest/__init__.py index 8542a1dbc..f2399422b 100644 --- a/comfy_api/latest/__init__.py +++ b/comfy_api/latest/__init__.py @@ -21,6 +21,17 @@ class ComfyAPI_latest(ComfyAPIBase): VERSION = "latest" STABLE = False + def __init__(self): + super().__init__() + self.node_replacement = self.NodeReplacement() + self.execution = self.Execution() + + class NodeReplacement(ProxiedSingleton): + async def register(self, node_replace: io.NodeReplace) -> None: + """Register a node replacement mapping.""" + from server import PromptServer + PromptServer.instance.node_replace_manager.register(node_replace) + class Execution(ProxiedSingleton): async def set_progress( self, @@ -73,8 +84,6 @@ class ComfyAPI_latest(ComfyAPIBase): image=to_display, ) - execution: Execution - class ComfyExtension(ABC): async def on_load(self) -> None: """ diff --git a/comfy_api/latest/_io.py b/comfy_api/latest/_io.py index 93cf482ca..95d79c035 100644 --- a/comfy_api/latest/_io.py +++ b/comfy_api/latest/_io.py @@ -2030,6 +2030,68 @@ class _UIOutput(ABC): ... +class InputMapOldId(TypedDict): + """Map an old node input to a new node input by ID.""" + new_id: str + old_id: str + +class InputMapSetValue(TypedDict): + """Set a specific value for a new node input.""" + new_id: str + set_value: Any + +InputMap = InputMapOldId | InputMapSetValue +""" +Input mapping for node replacement. Type is inferred by dictionary keys: +- {"new_id": str, "old_id": str} - maps old input to new input +- {"new_id": str, "set_value": Any} - sets a specific value for new input +""" + +class OutputMap(TypedDict): + """Map outputs of node replacement via indexes.""" + new_idx: int + old_idx: int + +class NodeReplace: + """ + Defines a possible node replacement, mapping inputs and outputs of the old node to the new node. + + Also supports assigning specific values to the input widgets of the new node. + + Args: + new_node_id: The class name of the new replacement node. + old_node_id: The class name of the deprecated node. + old_widget_ids: Ordered list of input IDs for widgets that may not have an input slot + connected. The workflow JSON stores widget values by their relative position index, + not by ID. This list maps those positional indexes to input IDs, enabling the + replacement system to correctly identify widget values during node migration. + input_mapping: List of input mappings from old node to new node. + output_mapping: List of output mappings from old node to new node. + """ + def __init__(self, + new_node_id: str, + old_node_id: str, + old_widget_ids: list[str] | None=None, + input_mapping: list[InputMap] | None=None, + output_mapping: list[OutputMap] | None=None, + ): + self.new_node_id = new_node_id + self.old_node_id = old_node_id + self.old_widget_ids = old_widget_ids + self.input_mapping = input_mapping + self.output_mapping = output_mapping + + def as_dict(self): + """Create serializable representation of the node replacement.""" + return { + "new_node_id": self.new_node_id, + "old_node_id": self.old_node_id, + "old_widget_ids": self.old_widget_ids, + "input_mapping": list(self.input_mapping) if self.input_mapping else None, + "output_mapping": list(self.output_mapping) if self.output_mapping else None, + } + + __all__ = [ "FolderType", "UploadType", @@ -2121,4 +2183,5 @@ __all__ = [ "ImageCompare", "PriceBadgeDepends", "PriceBadge", + "NodeReplace", ] diff --git a/comfy_extras/nodes_post_processing.py b/comfy_extras/nodes_post_processing.py index a52a90e2c..66dac10b1 100644 --- a/comfy_extras/nodes_post_processing.py +++ b/comfy_extras/nodes_post_processing.py @@ -655,6 +655,7 @@ class BatchImagesMasksLatentsNode(io.ComfyNode): batched = batch_masks(values) return io.NodeOutput(batched) + class PostProcessingExtension(ComfyExtension): @override async def get_node_list(self) -> list[type[io.ComfyNode]]: diff --git a/comfy_extras/nodes_replacements.py b/comfy_extras/nodes_replacements.py new file mode 100644 index 000000000..7684e854c --- /dev/null +++ b/comfy_extras/nodes_replacements.py @@ -0,0 +1,103 @@ +from comfy_api.latest import ComfyExtension, io, ComfyAPI + +api = ComfyAPI() + + +async def register_replacements(): + """Register all built-in node replacements.""" + await register_replacements_longeredge() + await register_replacements_batchimages() + await register_replacements_upscaleimage() + await register_replacements_controlnet() + await register_replacements_load3d() + await register_replacements_preview3d() + await register_replacements_svdimg2vid() + await register_replacements_conditioningavg() + +async def register_replacements_longeredge(): + # No dynamic inputs here + await api.node_replacement.register(io.NodeReplace( + new_node_id="ImageScaleToMaxDimension", + old_node_id="ResizeImagesByLongerEdge", + old_widget_ids=["longer_edge"], + input_mapping=[ + {"new_id": "image", "old_id": "images"}, + {"new_id": "largest_size", "old_id": "longer_edge"}, + {"new_id": "upscale_method", "set_value": "lanczos"}, + ], + # just to test the frontend output_mapping code, does nothing really here + output_mapping=[{"new_idx": 0, "old_idx": 0}], + )) + +async def register_replacements_batchimages(): + # BatchImages node uses Autogrow + await api.node_replacement.register(io.NodeReplace( + new_node_id="BatchImagesNode", + old_node_id="ImageBatch", + input_mapping=[ + {"new_id": "images.image0", "old_id": "image1"}, + {"new_id": "images.image1", "old_id": "image2"}, + ], + )) + +async def register_replacements_upscaleimage(): + # ResizeImageMaskNode uses DynamicCombo + await api.node_replacement.register(io.NodeReplace( + new_node_id="ResizeImageMaskNode", + old_node_id="ImageScaleBy", + old_widget_ids=["upscale_method", "scale_by"], + input_mapping=[ + {"new_id": "input", "old_id": "image"}, + {"new_id": "resize_type", "set_value": "scale by multiplier"}, + {"new_id": "resize_type.multiplier", "old_id": "scale_by"}, + {"new_id": "scale_method", "old_id": "upscale_method"}, + ], + )) + +async def register_replacements_controlnet(): + # T2IAdapterLoader → ControlNetLoader + await api.node_replacement.register(io.NodeReplace( + new_node_id="ControlNetLoader", + old_node_id="T2IAdapterLoader", + input_mapping=[ + {"new_id": "control_net_name", "old_id": "t2i_adapter_name"}, + ], + )) + +async def register_replacements_load3d(): + # Load3DAnimation merged into Load3D + await api.node_replacement.register(io.NodeReplace( + new_node_id="Load3D", + old_node_id="Load3DAnimation", + )) + +async def register_replacements_preview3d(): + # Preview3DAnimation merged into Preview3D + await api.node_replacement.register(io.NodeReplace( + new_node_id="Preview3D", + old_node_id="Preview3DAnimation", + )) + +async def register_replacements_svdimg2vid(): + # Typo fix: SDV → SVD + await api.node_replacement.register(io.NodeReplace( + new_node_id="SVD_img2vid_Conditioning", + old_node_id="SDV_img2vid_Conditioning", + )) + +async def register_replacements_conditioningavg(): + # Typo fix: trailing space in node name + await api.node_replacement.register(io.NodeReplace( + new_node_id="ConditioningAverage", + old_node_id="ConditioningAverage ", + )) + +class NodeReplacementsExtension(ComfyExtension): + async def on_load(self) -> None: + await register_replacements() + + async def get_node_list(self) -> list[type[io.ComfyNode]]: + return [] + +async def comfy_entrypoint() -> NodeReplacementsExtension: + return NodeReplacementsExtension() diff --git a/nodes.py b/nodes.py index 91de7a9d7..db5f98408 100644 --- a/nodes.py +++ b/nodes.py @@ -2264,6 +2264,7 @@ async def load_custom_node(module_path: str, ignore=set(), module_parent="custom if not isinstance(extension, ComfyExtension): logging.warning(f"comfy_entrypoint in {module_path} did not return a ComfyExtension, skipping.") return False + await extension.on_load() node_list = await extension.get_node_list() if not isinstance(node_list, list): logging.warning(f"comfy_entrypoint in {module_path} did not return a list of nodes, skipping.") @@ -2435,6 +2436,7 @@ async def init_builtin_extra_nodes(): "nodes_lora_debug.py", "nodes_color.py", "nodes_toolkit.py", + "nodes_replacements.py", ] import_failed = [] diff --git a/server.py b/server.py index 2300393b2..8882e43c4 100644 --- a/server.py +++ b/server.py @@ -40,6 +40,7 @@ from app.user_manager import UserManager from app.model_manager import ModelFileManager from app.custom_node_manager import CustomNodeManager from app.subgraph_manager import SubgraphManager +from app.node_replace_manager import NodeReplaceManager from typing import Optional, Union from api_server.routes.internal.internal_routes import InternalRoutes from protocol import BinaryEventTypes @@ -204,6 +205,7 @@ class PromptServer(): self.model_file_manager = ModelFileManager() self.custom_node_manager = CustomNodeManager() self.subgraph_manager = SubgraphManager() + self.node_replace_manager = NodeReplaceManager() self.internal_routes = InternalRoutes(self) self.supports = ["custom_nodes_from_web"] self.prompt_queue = execution.PromptQueue(self) @@ -887,6 +889,8 @@ class PromptServer(): if "partial_execution_targets" in json_data: partial_execution_targets = json_data["partial_execution_targets"] + self.node_replace_manager.apply_replacements(prompt) + valid = await execution.validate_prompt(prompt_id, prompt, partial_execution_targets) extra_data = {} if "extra_data" in json_data: @@ -995,6 +999,7 @@ class PromptServer(): self.model_file_manager.add_routes(self.routes) self.custom_node_manager.add_routes(self.routes, self.app, nodes.LOADED_MODULE_DIRS.items()) self.subgraph_manager.add_routes(self.routes, nodes.LOADED_MODULE_DIRS.items()) + self.node_replace_manager.add_routes(self.routes) self.app.add_subapp('/internal', self.internal_routes.get_app()) # Prefix every route with /api for easier matching for delegation. From e2c71ceb0004da0d8a33dc9e79b31c2324241173 Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Sun, 15 Feb 2026 15:33:18 +0200 Subject: [PATCH 636/647] feat(api-nodes-Tencent): add ModelTo3DUV, 3DTextureEdit, 3DParts nodes (#12428) * feat(api-nodes-Tencent): add ModelTo3DUV, 3DTextureEdit, 3DParts nodes * add image output to TencentModelTo3DUV node * commented out two nodes * added rate_limit check to other hunyuan3d nodes --- comfy_api_nodes/apis/hunyuan3d.py | 20 ++ comfy_api_nodes/nodes_hunyuan3d.py | 295 ++++++++++++++++++++++++- comfy_api_nodes/util/__init__.py | 2 + comfy_api_nodes/util/upload_helpers.py | 21 ++ 4 files changed, 326 insertions(+), 12 deletions(-) diff --git a/comfy_api_nodes/apis/hunyuan3d.py b/comfy_api_nodes/apis/hunyuan3d.py index 6421c9bd5..e84eba31e 100644 --- a/comfy_api_nodes/apis/hunyuan3d.py +++ b/comfy_api_nodes/apis/hunyuan3d.py @@ -64,3 +64,23 @@ class To3DProTaskResultResponse(BaseModel): class To3DProTaskQueryRequest(BaseModel): JobId: str = Field(...) + + +class To3DUVFileInput(BaseModel): + Type: str = Field(..., description="File type: GLB, OBJ, or FBX") + Url: str = Field(...) + + +class To3DUVTaskRequest(BaseModel): + File: To3DUVFileInput = Field(...) + + +class TextureEditImageInfo(BaseModel): + Url: str = Field(...) + + +class TextureEditTaskRequest(BaseModel): + File3D: To3DUVFileInput = Field(...) + Image: TextureEditImageInfo | None = Field(None) + Prompt: str | None = Field(None) + EnablePBR: bool | None = Field(None) diff --git a/comfy_api_nodes/nodes_hunyuan3d.py b/comfy_api_nodes/nodes_hunyuan3d.py index 813a7c809..ca002cc60 100644 --- a/comfy_api_nodes/nodes_hunyuan3d.py +++ b/comfy_api_nodes/nodes_hunyuan3d.py @@ -1,31 +1,48 @@ from typing_extensions import override -from comfy_api.latest import IO, ComfyExtension, Input +from comfy_api.latest import IO, ComfyExtension, Input, Types from comfy_api_nodes.apis.hunyuan3d import ( Hunyuan3DViewImage, InputGenerateType, ResultFile3D, + TextureEditTaskRequest, To3DProTaskCreateResponse, To3DProTaskQueryRequest, To3DProTaskRequest, To3DProTaskResultResponse, + To3DUVFileInput, + To3DUVTaskRequest, ) from comfy_api_nodes.util import ( ApiEndpoint, download_url_to_file_3d, + download_url_to_image_tensor, downscale_image_tensor_by_max_side, poll_op, sync_op, + upload_3d_model_to_comfyapi, upload_image_to_comfyapi, validate_image_dimensions, validate_string, ) -def get_file_from_response(response_objs: list[ResultFile3D], file_type: str) -> ResultFile3D | None: +def _is_tencent_rate_limited(status: int, body: object) -> bool: + return ( + status == 400 + and isinstance(body, dict) + and "RequestLimitExceeded" in str(body.get("Response", {}).get("Error", {}).get("Code", "")) + ) + + +def get_file_from_response( + response_objs: list[ResultFile3D], file_type: str, raise_if_not_found: bool = True +) -> ResultFile3D | None: for i in response_objs: if i.Type.lower() == file_type.lower(): return i + if raise_if_not_found: + raise ValueError(f"'{file_type}' file type is not found in the response.") return None @@ -35,7 +52,7 @@ class TencentTextToModelNode(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="TencentTextToModelNode", - display_name="Hunyuan3D: Text to Model (Pro)", + display_name="Hunyuan3D: Text to Model", category="api node/3d/Tencent", inputs=[ IO.Combo.Input( @@ -120,6 +137,7 @@ class TencentTextToModelNode(IO.ComfyNode): EnablePBR=generate_type.get("pbr", None), PolygonType=generate_type.get("polygon_type", None), ), + is_rate_limited=_is_tencent_rate_limited, ) if response.Error: raise ValueError(f"Task creation failed with code {response.Error.Code}: {response.Error.Message}") @@ -131,11 +149,14 @@ class TencentTextToModelNode(IO.ComfyNode): response_model=To3DProTaskResultResponse, status_extractor=lambda r: r.Status, ) - glb_result = get_file_from_response(result.ResultFile3Ds, "glb") - obj_result = get_file_from_response(result.ResultFile3Ds, "obj") - file_glb = await download_url_to_file_3d(glb_result.Url, "glb", task_id=task_id) if glb_result else None return IO.NodeOutput( - file_glb, file_glb, await download_url_to_file_3d(obj_result.Url, "obj", task_id=task_id) if obj_result else None + f"{task_id}.glb", + await download_url_to_file_3d( + get_file_from_response(result.ResultFile3Ds, "glb").Url, "glb", task_id=task_id + ), + await download_url_to_file_3d( + get_file_from_response(result.ResultFile3Ds, "obj").Url, "obj", task_id=task_id + ), ) @@ -145,7 +166,7 @@ class TencentImageToModelNode(IO.ComfyNode): def define_schema(cls): return IO.Schema( node_id="TencentImageToModelNode", - display_name="Hunyuan3D: Image(s) to Model (Pro)", + display_name="Hunyuan3D: Image(s) to Model", category="api node/3d/Tencent", inputs=[ IO.Combo.Input( @@ -268,6 +289,7 @@ class TencentImageToModelNode(IO.ComfyNode): EnablePBR=generate_type.get("pbr", None), PolygonType=generate_type.get("polygon_type", None), ), + is_rate_limited=_is_tencent_rate_limited, ) if response.Error: raise ValueError(f"Task creation failed with code {response.Error.Code}: {response.Error.Message}") @@ -279,11 +301,257 @@ class TencentImageToModelNode(IO.ComfyNode): response_model=To3DProTaskResultResponse, status_extractor=lambda r: r.Status, ) - glb_result = get_file_from_response(result.ResultFile3Ds, "glb") - obj_result = get_file_from_response(result.ResultFile3Ds, "obj") - file_glb = await download_url_to_file_3d(glb_result.Url, "glb", task_id=task_id) if glb_result else None return IO.NodeOutput( - file_glb, file_glb, await download_url_to_file_3d(obj_result.Url, "obj", task_id=task_id) if obj_result else None + f"{task_id}.glb", + await download_url_to_file_3d( + get_file_from_response(result.ResultFile3Ds, "glb").Url, "glb", task_id=task_id + ), + await download_url_to_file_3d( + get_file_from_response(result.ResultFile3Ds, "obj").Url, "obj", task_id=task_id + ), + ) + + +class TencentModelTo3DUVNode(IO.ComfyNode): + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="TencentModelTo3DUVNode", + display_name="Hunyuan3D: Model to UV", + category="api node/3d/Tencent", + description="Perform UV unfolding on a 3D model to generate UV texture. " + "Input model must have less than 30000 faces.", + inputs=[ + IO.MultiType.Input( + "model_3d", + types=[IO.File3DGLB, IO.File3DOBJ, IO.File3DFBX, IO.File3DAny], + tooltip="Input 3D model (GLB, OBJ, or FBX)", + ), + IO.Int.Input( + "seed", + default=1, + min=0, + max=2147483647, + display_mode=IO.NumberDisplay.number, + control_after_generate=True, + tooltip="Seed controls whether the node should re-run; " + "results are non-deterministic regardless of seed.", + ), + ], + outputs=[ + IO.File3DOBJ.Output(display_name="OBJ"), + IO.File3DFBX.Output(display_name="FBX"), + IO.Image.Output(), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + price_badge=IO.PriceBadge(expr='{"type":"usd","usd":0.2}'), + ) + + SUPPORTED_FORMATS = {"glb", "obj", "fbx"} + + @classmethod + async def execute( + cls, + model_3d: Types.File3D, + seed: int, + ) -> IO.NodeOutput: + _ = seed + file_format = model_3d.format.lower() + if file_format not in cls.SUPPORTED_FORMATS: + raise ValueError( + f"Unsupported file format: '{file_format}'. " + f"Supported formats: {', '.join(sorted(cls.SUPPORTED_FORMATS))}." + ) + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/tencent/hunyuan/3d-uv", method="POST"), + response_model=To3DProTaskCreateResponse, + data=To3DUVTaskRequest( + File=To3DUVFileInput( + Type=file_format.upper(), + Url=await upload_3d_model_to_comfyapi(cls, model_3d, file_format), + ) + ), + is_rate_limited=_is_tencent_rate_limited, + ) + if response.Error: + raise ValueError(f"Task creation failed with code {response.Error.Code}: {response.Error.Message}") + result = await poll_op( + cls, + ApiEndpoint(path="/proxy/tencent/hunyuan/3d-uv/query", method="POST"), + data=To3DProTaskQueryRequest(JobId=response.JobId), + response_model=To3DProTaskResultResponse, + status_extractor=lambda r: r.Status, + ) + return IO.NodeOutput( + await download_url_to_file_3d(get_file_from_response(result.ResultFile3Ds, "obj").Url, "obj"), + await download_url_to_file_3d(get_file_from_response(result.ResultFile3Ds, "fbx").Url, "fbx"), + await download_url_to_image_tensor(get_file_from_response(result.ResultFile3Ds, "image").Url), + ) + + +class Tencent3DTextureEditNode(IO.ComfyNode): + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="Tencent3DTextureEditNode", + display_name="Hunyuan3D: 3D Texture Edit", + category="api node/3d/Tencent", + description="After inputting the 3D model, perform 3D model texture redrawing.", + inputs=[ + IO.MultiType.Input( + "model_3d", + types=[IO.File3DFBX, IO.File3DAny], + tooltip="3D model in FBX format. Model should have less than 100000 faces.", + ), + IO.String.Input( + "prompt", + multiline=True, + default="", + tooltip="Describes texture editing. Supports up to 1024 UTF-8 characters.", + ), + IO.Int.Input( + "seed", + default=0, + min=0, + max=2147483647, + display_mode=IO.NumberDisplay.number, + control_after_generate=True, + tooltip="Seed controls whether the node should re-run; " + "results are non-deterministic regardless of seed.", + ), + ], + outputs=[ + IO.File3DGLB.Output(display_name="GLB"), + IO.File3DFBX.Output(display_name="FBX"), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd": 0.6}""", + ), + ) + + @classmethod + async def execute( + cls, + model_3d: Types.File3D, + prompt: str, + seed: int, + ) -> IO.NodeOutput: + _ = seed + file_format = model_3d.format.lower() + if file_format != "fbx": + raise ValueError(f"Unsupported file format: '{file_format}'. Only FBX format is supported.") + validate_string(prompt, field_name="prompt", min_length=1, max_length=1024) + model_url = await upload_3d_model_to_comfyapi(cls, model_3d, file_format) + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/tencent/hunyuan/3d-texture-edit", method="POST"), + response_model=To3DProTaskCreateResponse, + data=TextureEditTaskRequest( + File3D=To3DUVFileInput(Type=file_format.upper(), Url=model_url), + Prompt=prompt, + EnablePBR=True, + ), + is_rate_limited=_is_tencent_rate_limited, + ) + if response.Error: + raise ValueError(f"Task creation failed with code {response.Error.Code}: {response.Error.Message}") + + result = await poll_op( + cls, + ApiEndpoint(path="/proxy/tencent/hunyuan/3d-texture-edit/query", method="POST"), + data=To3DProTaskQueryRequest(JobId=response.JobId), + response_model=To3DProTaskResultResponse, + status_extractor=lambda r: r.Status, + ) + return IO.NodeOutput( + await download_url_to_file_3d(get_file_from_response(result.ResultFile3Ds, "glb").Url, "glb"), + await download_url_to_file_3d(get_file_from_response(result.ResultFile3Ds, "fbx").Url, "fbx"), + ) + + +class Tencent3DPartNode(IO.ComfyNode): + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="Tencent3DPartNode", + display_name="Hunyuan3D: 3D Part", + category="api node/3d/Tencent", + description="Automatically perform component identification and generation based on the model structure.", + inputs=[ + IO.MultiType.Input( + "model_3d", + types=[IO.File3DFBX, IO.File3DAny], + tooltip="3D model in FBX format. Model should have less than 30000 faces.", + ), + IO.Int.Input( + "seed", + default=0, + min=0, + max=2147483647, + display_mode=IO.NumberDisplay.number, + control_after_generate=True, + tooltip="Seed controls whether the node should re-run; " + "results are non-deterministic regardless of seed.", + ), + ], + outputs=[ + IO.File3DFBX.Output(display_name="FBX"), + ], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + price_badge=IO.PriceBadge(expr='{"type":"usd","usd":0.6}'), + ) + + @classmethod + async def execute( + cls, + model_3d: Types.File3D, + seed: int, + ) -> IO.NodeOutput: + _ = seed + file_format = model_3d.format.lower() + if file_format != "fbx": + raise ValueError(f"Unsupported file format: '{file_format}'. Only FBX format is supported.") + model_url = await upload_3d_model_to_comfyapi(cls, model_3d, file_format) + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/tencent/hunyuan/3d-part", method="POST"), + response_model=To3DProTaskCreateResponse, + data=To3DUVTaskRequest( + File=To3DUVFileInput(Type=file_format.upper(), Url=model_url), + ), + is_rate_limited=_is_tencent_rate_limited, + ) + if response.Error: + raise ValueError(f"Task creation failed with code {response.Error.Code}: {response.Error.Message}") + result = await poll_op( + cls, + ApiEndpoint(path="/proxy/tencent/hunyuan/3d-part/query", method="POST"), + data=To3DProTaskQueryRequest(JobId=response.JobId), + response_model=To3DProTaskResultResponse, + status_extractor=lambda r: r.Status, + ) + return IO.NodeOutput( + await download_url_to_file_3d(get_file_from_response(result.ResultFile3Ds, "fbx").Url, "fbx"), ) @@ -293,6 +561,9 @@ class TencentHunyuan3DExtension(ComfyExtension): return [ TencentTextToModelNode, TencentImageToModelNode, + # TencentModelTo3DUVNode, + # Tencent3DTextureEditNode, + Tencent3DPartNode, ] diff --git a/comfy_api_nodes/util/__init__.py b/comfy_api_nodes/util/__init__.py index 18b020eef..f8a0ba8af 100644 --- a/comfy_api_nodes/util/__init__.py +++ b/comfy_api_nodes/util/__init__.py @@ -33,6 +33,7 @@ from .download_helpers import ( download_url_to_video_output, ) from .upload_helpers import ( + upload_3d_model_to_comfyapi, upload_audio_to_comfyapi, upload_file_to_comfyapi, upload_image_to_comfyapi, @@ -62,6 +63,7 @@ __all__ = [ "sync_op", "sync_op_raw", # Upload helpers + "upload_3d_model_to_comfyapi", "upload_audio_to_comfyapi", "upload_file_to_comfyapi", "upload_image_to_comfyapi", diff --git a/comfy_api_nodes/util/upload_helpers.py b/comfy_api_nodes/util/upload_helpers.py index 7cc565263..6d1d107a1 100644 --- a/comfy_api_nodes/util/upload_helpers.py +++ b/comfy_api_nodes/util/upload_helpers.py @@ -164,6 +164,27 @@ async def upload_video_to_comfyapi( return await upload_file_to_comfyapi(cls, video_bytes_io, filename, upload_mime_type, wait_label) +_3D_MIME_TYPES = { + "glb": "model/gltf-binary", + "obj": "model/obj", + "fbx": "application/octet-stream", +} + + +async def upload_3d_model_to_comfyapi( + cls: type[IO.ComfyNode], + model_3d: Types.File3D, + file_format: str, +) -> str: + """Uploads a 3D model file to ComfyUI API and returns its download URL.""" + return await upload_file_to_comfyapi( + cls, + model_3d.get_data(), + f"{uuid.uuid4()}.{file_format}", + _3D_MIME_TYPES.get(file_format, "application/octet-stream"), + ) + + async def upload_file_to_comfyapi( cls: type[IO.ComfyNode], file_bytes_io: BytesIO, From 2c1d06a4e32900c260bcb3d0888f20edc1e3e5ab Mon Sep 17 00:00:00 2001 From: Alexander Piskun <13381981+bigcat88@users.noreply.github.com> Date: Mon, 16 Feb 2026 03:22:30 +0200 Subject: [PATCH 637/647] feat(api-nodes): add Bria RMBG nodes (#12465) Co-authored-by: Jedrzej Kosinski --- comfy_api_nodes/apis/bria.py | 44 +++++- comfy_api_nodes/nodes_bria.py | 220 ++++++++++++++++++++++------ comfy_api_nodes/util/conversions.py | 2 +- 3 files changed, 218 insertions(+), 48 deletions(-) diff --git a/comfy_api_nodes/apis/bria.py b/comfy_api_nodes/apis/bria.py index 9119cacc6..8c496b56c 100644 --- a/comfy_api_nodes/apis/bria.py +++ b/comfy_api_nodes/apis/bria.py @@ -45,17 +45,55 @@ class BriaEditImageRequest(BaseModel): ) +class BriaRemoveBackgroundRequest(BaseModel): + image: str = Field(...) + sync: bool = Field(False) + visual_input_content_moderation: bool = Field( + False, description="If true, returns 422 on input image moderation failure." + ) + visual_output_content_moderation: bool = Field( + False, description="If true, returns 422 on visual output moderation failure." + ) + seed: int = Field(...) + + class BriaStatusResponse(BaseModel): request_id: str = Field(...) status_url: str = Field(...) warning: str | None = Field(None) -class BriaResult(BaseModel): +class BriaRemoveBackgroundResult(BaseModel): + image_url: str = Field(...) + + +class BriaRemoveBackgroundResponse(BaseModel): + status: str = Field(...) + result: BriaRemoveBackgroundResult | None = Field(None) + + +class BriaImageEditResult(BaseModel): structured_prompt: str = Field(...) image_url: str = Field(...) -class BriaResponse(BaseModel): +class BriaImageEditResponse(BaseModel): status: str = Field(...) - result: BriaResult | None = Field(None) + result: BriaImageEditResult | None = Field(None) + + +class BriaRemoveVideoBackgroundRequest(BaseModel): + video: str = Field(...) + background_color: str = Field(default="transparent", description="Background color for the output video.") + output_container_and_codec: str = Field(...) + preserve_audio: bool = Field(True) + seed: int = Field(...) + + +class BriaRemoveVideoBackgroundResult(BaseModel): + video_url: str = Field(...) + + +class BriaRemoveVideoBackgroundResponse(BaseModel): + status: str = Field(...) + result: BriaRemoveVideoBackgroundResult | None = Field(None) diff --git a/comfy_api_nodes/nodes_bria.py b/comfy_api_nodes/nodes_bria.py index d3a52bc1b..4044ee3ea 100644 --- a/comfy_api_nodes/nodes_bria.py +++ b/comfy_api_nodes/nodes_bria.py @@ -3,7 +3,11 @@ from typing_extensions import override from comfy_api.latest import IO, ComfyExtension, Input from comfy_api_nodes.apis.bria import ( BriaEditImageRequest, - BriaResponse, + BriaRemoveBackgroundRequest, + BriaRemoveBackgroundResponse, + BriaRemoveVideoBackgroundRequest, + BriaRemoveVideoBackgroundResponse, + BriaImageEditResponse, BriaStatusResponse, InputModerationSettings, ) @@ -11,10 +15,12 @@ from comfy_api_nodes.util import ( ApiEndpoint, convert_mask_to_image, download_url_to_image_tensor, - get_number_of_images, + download_url_to_video_output, poll_op, sync_op, - upload_images_to_comfyapi, + upload_image_to_comfyapi, + upload_video_to_comfyapi, + validate_video_duration, ) @@ -73,21 +79,15 @@ class BriaImageEditNode(IO.ComfyNode): IO.DynamicCombo.Input( "moderation", options=[ + IO.DynamicCombo.Option("false", []), IO.DynamicCombo.Option( "true", [ - IO.Boolean.Input( - "prompt_content_moderation", default=False - ), - IO.Boolean.Input( - "visual_input_moderation", default=False - ), - IO.Boolean.Input( - "visual_output_moderation", default=True - ), + IO.Boolean.Input("prompt_content_moderation", default=False), + IO.Boolean.Input("visual_input_moderation", default=False), + IO.Boolean.Input("visual_output_moderation", default=True), ], ), - IO.DynamicCombo.Option("false", []), ], tooltip="Moderation settings", ), @@ -127,50 +127,26 @@ class BriaImageEditNode(IO.ComfyNode): mask: Input.Image | None = None, ) -> IO.NodeOutput: if not prompt and not structured_prompt: - raise ValueError( - "One of prompt or structured_prompt is required to be non-empty." - ) - if get_number_of_images(image) != 1: - raise ValueError("Exactly one input image is required.") + raise ValueError("One of prompt or structured_prompt is required to be non-empty.") mask_url = None if mask is not None: - mask_url = ( - await upload_images_to_comfyapi( - cls, - convert_mask_to_image(mask), - max_images=1, - mime_type="image/png", - wait_label="Uploading mask", - ) - )[0] + mask_url = await upload_image_to_comfyapi(cls, convert_mask_to_image(mask), wait_label="Uploading mask") response = await sync_op( cls, ApiEndpoint(path="proxy/bria/v2/image/edit", method="POST"), data=BriaEditImageRequest( instruction=prompt if prompt else None, structured_instruction=structured_prompt if structured_prompt else None, - images=await upload_images_to_comfyapi( - cls, - image, - max_images=1, - mime_type="image/png", - wait_label="Uploading image", - ), + images=[await upload_image_to_comfyapi(cls, image, wait_label="Uploading image")], mask=mask_url, negative_prompt=negative_prompt if negative_prompt else None, guidance_scale=guidance_scale, seed=seed, model_version=model, steps_num=steps, - prompt_content_moderation=moderation.get( - "prompt_content_moderation", False - ), - visual_input_content_moderation=moderation.get( - "visual_input_moderation", False - ), - visual_output_content_moderation=moderation.get( - "visual_output_moderation", False - ), + prompt_content_moderation=moderation.get("prompt_content_moderation", False), + visual_input_content_moderation=moderation.get("visual_input_moderation", False), + visual_output_content_moderation=moderation.get("visual_output_moderation", False), ), response_model=BriaStatusResponse, ) @@ -178,7 +154,7 @@ class BriaImageEditNode(IO.ComfyNode): cls, ApiEndpoint(path=f"/proxy/bria/v2/status/{response.request_id}"), status_extractor=lambda r: r.status, - response_model=BriaResponse, + response_model=BriaImageEditResponse, ) return IO.NodeOutput( await download_url_to_image_tensor(response.result.image_url), @@ -186,11 +162,167 @@ class BriaImageEditNode(IO.ComfyNode): ) +class BriaRemoveImageBackground(IO.ComfyNode): + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="BriaRemoveImageBackground", + display_name="Bria Remove Image Background", + category="api node/image/Bria", + description="Remove the background from an image using Bria RMBG 2.0.", + inputs=[ + IO.Image.Input("image"), + IO.DynamicCombo.Input( + "moderation", + options=[ + IO.DynamicCombo.Option("false", []), + IO.DynamicCombo.Option( + "true", + [ + IO.Boolean.Input("visual_input_moderation", default=False), + IO.Boolean.Input("visual_output_moderation", default=True), + ], + ), + ], + tooltip="Moderation settings", + ), + IO.Int.Input( + "seed", + default=0, + min=0, + max=2147483647, + display_mode=IO.NumberDisplay.number, + control_after_generate=True, + tooltip="Seed controls whether the node should re-run; " + "results are non-deterministic regardless of seed.", + ), + ], + outputs=[IO.Image.Output()], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.018}""", + ), + ) + + @classmethod + async def execute( + cls, + image: Input.Image, + moderation: dict, + seed: int, + ) -> IO.NodeOutput: + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/bria/v2/image/edit/remove_background", method="POST"), + data=BriaRemoveBackgroundRequest( + image=await upload_image_to_comfyapi(cls, image, wait_label="Uploading image"), + sync=False, + visual_input_content_moderation=moderation.get("visual_input_moderation", False), + visual_output_content_moderation=moderation.get("visual_output_moderation", False), + seed=seed, + ), + response_model=BriaStatusResponse, + ) + response = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/bria/v2/status/{response.request_id}"), + status_extractor=lambda r: r.status, + response_model=BriaRemoveBackgroundResponse, + ) + return IO.NodeOutput(await download_url_to_image_tensor(response.result.image_url)) + + +class BriaRemoveVideoBackground(IO.ComfyNode): + + @classmethod + def define_schema(cls): + return IO.Schema( + node_id="BriaRemoveVideoBackground", + display_name="Bria Remove Video Background", + category="api node/video/Bria", + description="Remove the background from a video using Bria. ", + inputs=[ + IO.Video.Input("video"), + IO.Combo.Input( + "background_color", + options=[ + "Black", + "White", + "Gray", + "Red", + "Green", + "Blue", + "Yellow", + "Cyan", + "Magenta", + "Orange", + ], + tooltip="Background color for the output video.", + ), + IO.Int.Input( + "seed", + default=0, + min=0, + max=2147483647, + display_mode=IO.NumberDisplay.number, + control_after_generate=True, + tooltip="Seed controls whether the node should re-run; " + "results are non-deterministic regardless of seed.", + ), + ], + outputs=[IO.Video.Output()], + hidden=[ + IO.Hidden.auth_token_comfy_org, + IO.Hidden.api_key_comfy_org, + IO.Hidden.unique_id, + ], + is_api_node=True, + price_badge=IO.PriceBadge( + expr="""{"type":"usd","usd":0.14,"format":{"suffix":"/second"}}""", + ), + ) + + @classmethod + async def execute( + cls, + video: Input.Video, + background_color: str, + seed: int, + ) -> IO.NodeOutput: + validate_video_duration(video, max_duration=60.0) + response = await sync_op( + cls, + ApiEndpoint(path="/proxy/bria/v2/video/edit/remove_background", method="POST"), + data=BriaRemoveVideoBackgroundRequest( + video=await upload_video_to_comfyapi(cls, video), + background_color=background_color, + output_container_and_codec="mp4_h264", + seed=seed, + ), + response_model=BriaStatusResponse, + ) + response = await poll_op( + cls, + ApiEndpoint(path=f"/proxy/bria/v2/status/{response.request_id}"), + status_extractor=lambda r: r.status, + response_model=BriaRemoveVideoBackgroundResponse, + ) + return IO.NodeOutput(await download_url_to_video_output(response.result.video_url)) + + class BriaExtension(ComfyExtension): @override async def get_node_list(self) -> list[type[IO.ComfyNode]]: return [ BriaImageEditNode, + BriaRemoveImageBackground, + BriaRemoveVideoBackground, ] diff --git a/comfy_api_nodes/util/conversions.py b/comfy_api_nodes/util/conversions.py index 3e37e8a8c..82b6d22a5 100644 --- a/comfy_api_nodes/util/conversions.py +++ b/comfy_api_nodes/util/conversions.py @@ -57,7 +57,7 @@ def tensor_to_bytesio( image: torch.Tensor, *, total_pixels: int | None = 2048 * 2048, - mime_type: str = "image/png", + mime_type: str | None = "image/png", ) -> BytesIO: """Converts a torch.Tensor image to a named BytesIO object. From ecd2a19661ecccd96e26f111af21781f3e613f59 Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Sun, 15 Feb 2026 17:28:51 -0800 Subject: [PATCH 638/647] Fix lora Extraction in offload conditions (+ dynamic_vram mode) (#12479) * lora_extract: Add a trange If you bite off more than your GPU can chew, this kinda just hangs. Give a rough indication of progress counting the weights in a trange. * lora_extract: Support on-the-fly patching Use the on-the-fly approach from the regular model saving logic for lora extraction too. Switch off force_cast_weights accordingly. This gets extraction working in dynamic vram while also supporting extraction on GPU offloaded. --- comfy_extras/nodes_lora_extract.py | 21 ++++++++++++++++----- 1 file changed, 16 insertions(+), 5 deletions(-) diff --git a/comfy_extras/nodes_lora_extract.py b/comfy_extras/nodes_lora_extract.py index fb89e03f4..1542d0a88 100644 --- a/comfy_extras/nodes_lora_extract.py +++ b/comfy_extras/nodes_lora_extract.py @@ -7,6 +7,7 @@ import logging from enum import Enum from typing_extensions import override from comfy_api.latest import ComfyExtension, io +from tqdm.auto import trange CLAMP_QUANTILE = 0.99 @@ -49,12 +50,22 @@ LORA_TYPES = {"standard": LORAType.STANDARD, "full_diff": LORAType.FULL_DIFF} def calc_lora_model(model_diff, rank, prefix_model, prefix_lora, output_sd, lora_type, bias_diff=False): - comfy.model_management.load_models_gpu([model_diff], force_patch_weights=True) + comfy.model_management.load_models_gpu([model_diff]) sd = model_diff.model_state_dict(filter_prefix=prefix_model) - for k in sd: - if k.endswith(".weight"): + sd_keys = list(sd.keys()) + for index in trange(len(sd_keys), unit="weight"): + k = sd_keys[index] + op_keys = sd_keys[index].rsplit('.', 1) + if len(op_keys) < 2 or op_keys[1] not in ["weight", "bias"] or (op_keys[1] == "bias" and not bias_diff): + continue + op = comfy.utils.get_attr(model_diff.model, op_keys[0]) + if hasattr(op, "comfy_cast_weights") and not getattr(op, "comfy_patched_weights", False): + weight_diff = model_diff.patch_weight_to_device(k, model_diff.load_device, return_weight=True) + else: weight_diff = sd[k] + + if op_keys[1] == "weight": if lora_type == LORAType.STANDARD: if weight_diff.ndim < 2: if bias_diff: @@ -69,8 +80,8 @@ def calc_lora_model(model_diff, rank, prefix_model, prefix_lora, output_sd, lora elif lora_type == LORAType.FULL_DIFF: output_sd["{}{}.diff".format(prefix_lora, k[len(prefix_model):-7])] = weight_diff.contiguous().half().cpu() - elif bias_diff and k.endswith(".bias"): - output_sd["{}{}.diff_b".format(prefix_lora, k[len(prefix_model):-5])] = sd[k].contiguous().half().cpu() + elif bias_diff and op_keys[1] == "bias": + output_sd["{}{}.diff_b".format(prefix_lora, k[len(prefix_model):-5])] = weight_diff.contiguous().half().cpu() return output_sd class LoraSave(io.ComfyNode): From c0370044cd467b92f4db63b88029ebc700388d36 Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Sun, 15 Feb 2026 17:30:09 -0800 Subject: [PATCH 639/647] MPDynamic: force load flux img_in weight (Fixes flux1 canny+depth lora crash) (#12446) * lora: add weight shape calculations. This lets the loader know if a lora will change the shape of a weight so it can take appropriate action. * MPDynamic: force load flux img_in weight This weight is a bit special, in that the lora changes its geometry. This is rather unique, not handled by existing estimate and doesn't work for either offloading or dynamic_vram. Fix for dynamic_vram as a special case. Ideally we can fully precalculate these lora geometry changes at load time, but just get these models working first. --- comfy/lora.py | 25 +++++++++++++++++++++++++ comfy/model_patcher.py | 35 +++++++++++++++++++++++++++-------- comfy/weight_adapter/base.py | 6 ++++++ comfy/weight_adapter/lora.py | 7 +++++++ 4 files changed, 65 insertions(+), 8 deletions(-) diff --git a/comfy/lora.py b/comfy/lora.py index 44030bcab..279cf38bb 100644 --- a/comfy/lora.py +++ b/comfy/lora.py @@ -374,6 +374,31 @@ def pad_tensor_to_shape(tensor: torch.Tensor, new_shape: list[int]) -> torch.Ten return padded_tensor +def calculate_shape(patches, weight, key, original_weights=None): + current_shape = weight.shape + + for p in patches: + v = p[1] + offset = p[3] + + # Offsets restore the old shape; lists force a diff without metadata + if offset is not None or isinstance(v, list): + continue + + if isinstance(v, weight_adapter.WeightAdapterBase): + adapter_shape = v.calculate_shape(key) + if adapter_shape is not None: + current_shape = adapter_shape + continue + + # Standard diff logic with padding + if len(v) == 2: + patch_type, patch_data = v[0], v[1] + if patch_type == "diff" and len(patch_data) > 1 and patch_data[1]['pad_weight']: + current_shape = patch_data[0].shape + + return current_shape + def calculate_weight(patches, weight, key, intermediate_dtype=torch.float32, original_weights=None): for p in patches: strength = p[0] diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index 67dce088e..f01818f50 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -1514,8 +1514,10 @@ class ModelPatcherDynamic(ModelPatcher): weight, _, _ = get_key_weight(self.model, key) if weight is None: - return 0 + return (False, 0) if key in self.patches: + if comfy.lora.calculate_shape(self.patches[key], weight, key) != weight.shape: + return (True, 0) setattr(m, param_key + "_lowvram_function", LowVramPatch(key, self.patches)) num_patches += 1 else: @@ -1529,7 +1531,13 @@ class ModelPatcherDynamic(ModelPatcher): model_dtype = getattr(m, param_key + "_comfy_model_dtype", None) or weight.dtype weight._model_dtype = model_dtype geometry = comfy.memory_management.TensorGeometry(shape=weight.shape, dtype=model_dtype) - return comfy.memory_management.vram_aligned_size(geometry) + return (False, comfy.memory_management.vram_aligned_size(geometry)) + + def force_load_param(self, param_key, device_to): + key = key_param_name_to_key(n, param_key) + if key in self.backup: + comfy.utils.set_attr_param(self.model, key, self.backup[key].weight) + self.patch_weight_to_device(key, device_to=device_to) if hasattr(m, "comfy_cast_weights"): m.comfy_cast_weights = True @@ -1537,13 +1545,19 @@ class ModelPatcherDynamic(ModelPatcher): m.seed_key = n set_dirty(m, dirty) - v_weight_size = 0 - v_weight_size += setup_param(self, m, n, "weight") - v_weight_size += setup_param(self, m, n, "bias") + force_load, v_weight_size = setup_param(self, m, n, "weight") + force_load_bias, v_weight_bias = setup_param(self, m, n, "bias") + force_load = force_load or force_load_bias + v_weight_size += v_weight_bias - if vbar is not None and not hasattr(m, "_v"): - m._v = vbar.alloc(v_weight_size) - allocated_size += v_weight_size + if force_load: + logging.info(f"Module {n} has resizing Lora - force loading") + force_load_param(self, "weight", device_to) + force_load_param(self, "bias", device_to) + else: + if vbar is not None and not hasattr(m, "_v"): + m._v = vbar.alloc(v_weight_size) + allocated_size += v_weight_size else: for param in params: @@ -1606,6 +1620,11 @@ class ModelPatcherDynamic(ModelPatcher): for m in self.model.modules(): move_weight_functions(m, device_to) + keys = list(self.backup.keys()) + for k in keys: + bk = self.backup[k] + comfy.utils.set_attr_param(self.model, k, bk.weight) + def partially_load(self, device_to, extra_memory=0, force_patch_weights=False): assert not force_patch_weights #See above with self.use_ejected(skip_and_inject_on_exit_only=True): diff --git a/comfy/weight_adapter/base.py b/comfy/weight_adapter/base.py index bce89a0e2..d352e066b 100644 --- a/comfy/weight_adapter/base.py +++ b/comfy/weight_adapter/base.py @@ -49,6 +49,12 @@ class WeightAdapterBase: """ raise NotImplementedError + def calculate_shape( + self, + key + ): + return None + def calculate_weight( self, weight, diff --git a/comfy/weight_adapter/lora.py b/comfy/weight_adapter/lora.py index bc4260a8f..8e1261a12 100644 --- a/comfy/weight_adapter/lora.py +++ b/comfy/weight_adapter/lora.py @@ -214,6 +214,13 @@ class LoRAAdapter(WeightAdapterBase): else: return None + def calculate_shape( + self, + key + ): + reshape = self.weights[5] + return tuple(reshape) if reshape is not None else None + def calculate_weight( self, weight, From 88e6370527dbd602851de07d957a8f17b3ca9447 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sun, 15 Feb 2026 17:43:53 -0800 Subject: [PATCH 640/647] Remove workaround for old pytorch. (#12480) --- comfy/ldm/modules/diffusionmodules/model.py | 14 +------------- 1 file changed, 1 insertion(+), 13 deletions(-) diff --git a/comfy/ldm/modules/diffusionmodules/model.py b/comfy/ldm/modules/diffusionmodules/model.py index 5a22ef030..805592aa5 100644 --- a/comfy/ldm/modules/diffusionmodules/model.py +++ b/comfy/ldm/modules/diffusionmodules/model.py @@ -102,19 +102,7 @@ class VideoConv3d(nn.Module): return self.conv(x) def interpolate_up(x, scale_factor): - try: - return torch.nn.functional.interpolate(x, scale_factor=scale_factor, mode="nearest") - except: #operation not implemented for bf16 - orig_shape = list(x.shape) - out_shape = orig_shape[:2] - for i in range(len(orig_shape) - 2): - out_shape.append(round(orig_shape[i + 2] * scale_factor[i])) - out = torch.empty(out_shape, dtype=x.dtype, layout=x.layout, device=x.device) - split = 8 - l = out.shape[1] // split - for i in range(0, out.shape[1], l): - out[:,i:i+l] = torch.nn.functional.interpolate(x[:,i:i+l].to(torch.float32), scale_factor=scale_factor, mode="nearest").to(x.dtype) - return out + return torch.nn.functional.interpolate(x, scale_factor=scale_factor, mode="nearest") class Upsample(nn.Module): def __init__(self, in_channels, with_conv, conv_op=ops.Conv2d, scale_factor=2.0): From 1978f59ffdf242389ded3eec76274a4cbed9cc3d Mon Sep 17 00:00:00 2001 From: ComfyUI Wiki Date: Tue, 17 Feb 2026 06:33:43 +0800 Subject: [PATCH 641/647] chore: update workflow templates to v0.8.42 (#12491) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index e939e486a..0930bbbb8 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,5 +1,5 @@ comfyui-frontend-package==1.38.14 -comfyui-workflow-templates==0.8.38 +comfyui-workflow-templates==0.8.42 comfyui-embedded-docs==0.4.1 torch torchsde From 4454fab7f003c655e07f059c315e2aae0e5fb087 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 16 Feb 2026 17:09:24 -0800 Subject: [PATCH 642/647] Remove code to support RMSNorm on old pytorch. (#12499) --- comfy/ops.py | 6 ++---- comfy/rmsnorm.py | 55 ++++-------------------------------------------- 2 files changed, 6 insertions(+), 55 deletions(-) diff --git a/comfy/ops.py b/comfy/ops.py index 688937e43..026062f56 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -21,7 +21,6 @@ import logging import comfy.model_management from comfy.cli_args import args, PerformanceFeature, enables_dynamic_vram import comfy.float -import comfy.rmsnorm import json import comfy.memory_management import comfy.pinned_memory @@ -463,7 +462,7 @@ class disable_weight_init: else: return super().forward(*args, **kwargs) - class RMSNorm(comfy.rmsnorm.RMSNorm, CastWeightBiasOp): + class RMSNorm(torch.nn.RMSNorm, CastWeightBiasOp): def reset_parameters(self): self.bias = None return None @@ -475,8 +474,7 @@ class disable_weight_init: weight = None bias = None offload_stream = None - x = comfy.rmsnorm.rms_norm(input, weight, self.eps) # TODO: switch to commented out line when old torch is deprecated - # x = torch.nn.functional.rms_norm(input, self.normalized_shape, weight, self.eps) + x = torch.nn.functional.rms_norm(input, self.normalized_shape, weight, self.eps) uncast_bias_weight(self, weight, bias, offload_stream) return x diff --git a/comfy/rmsnorm.py b/comfy/rmsnorm.py index 555542a46..ab7cf14fa 100644 --- a/comfy/rmsnorm.py +++ b/comfy/rmsnorm.py @@ -1,57 +1,10 @@ import torch import comfy.model_management -import numbers -import logging - -RMSNorm = None - -try: - rms_norm_torch = torch.nn.functional.rms_norm - RMSNorm = torch.nn.RMSNorm -except: - rms_norm_torch = None - logging.warning("Please update pytorch to use native RMSNorm") +RMSNorm = torch.nn.RMSNorm def rms_norm(x, weight=None, eps=1e-6): - if rms_norm_torch is not None and not (torch.jit.is_tracing() or torch.jit.is_scripting()): - if weight is None: - return rms_norm_torch(x, (x.shape[-1],), eps=eps) - else: - return rms_norm_torch(x, weight.shape, weight=comfy.model_management.cast_to(weight, dtype=x.dtype, device=x.device), eps=eps) + if weight is None: + return torch.nn.functional.rms_norm(x, (x.shape[-1],), eps=eps) else: - r = x * torch.rsqrt(torch.mean(x**2, dim=-1, keepdim=True) + eps) - if weight is None: - return r - else: - return r * comfy.model_management.cast_to(weight, dtype=x.dtype, device=x.device) - - -if RMSNorm is None: - class RMSNorm(torch.nn.Module): - def __init__( - self, - normalized_shape, - eps=1e-6, - elementwise_affine=True, - device=None, - dtype=None, - ): - factory_kwargs = {"device": device, "dtype": dtype} - super().__init__() - if isinstance(normalized_shape, numbers.Integral): - # mypy error: incompatible types in assignment - normalized_shape = (normalized_shape,) # type: ignore[assignment] - self.normalized_shape = tuple(normalized_shape) # type: ignore[arg-type] - self.eps = eps - self.elementwise_affine = elementwise_affine - if self.elementwise_affine: - self.weight = torch.nn.Parameter( - torch.empty(self.normalized_shape, **factory_kwargs) - ) - else: - self.register_parameter("weight", None) - self.bias = None - - def forward(self, x): - return rms_norm(x, self.weight, self.eps) + return torch.nn.functional.rms_norm(x, weight.shape, weight=comfy.model_management.cast_to(weight, dtype=x.dtype, device=x.device), eps=eps) From b44fc4c589c66e39686239d6eff7d6088668c9a8 Mon Sep 17 00:00:00 2001 From: Alex Butler Date: Tue, 17 Feb 2026 03:16:19 +0000 Subject: [PATCH 643/647] add venv* to gitignore (#12431) --- .gitignore | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/.gitignore b/.gitignore index 4e8cea71e..2700ad5c2 100644 --- a/.gitignore +++ b/.gitignore @@ -11,7 +11,7 @@ extra_model_paths.yaml /.vs .vscode/ .idea/ -venv/ +venv*/ .venv/ /web/extensions/* !/web/extensions/logging.js.example From 8a6fbc2dc29d0b15c1e9655c24e7501829249995 Mon Sep 17 00:00:00 2001 From: Jedrzej Kosinski Date: Mon, 16 Feb 2026 19:20:21 -0800 Subject: [PATCH 644/647] Allow control_after_generate to be type ControlAfterGenerate in v3 schema (#12187) --- comfy_api/latest/_io.py | 13 ++++++++++--- 1 file changed, 10 insertions(+), 3 deletions(-) diff --git a/comfy_api/latest/_io.py b/comfy_api/latest/_io.py index 95d79c035..d18330d0b 100644 --- a/comfy_api/latest/_io.py +++ b/comfy_api/latest/_io.py @@ -75,6 +75,12 @@ class NumberDisplay(str, Enum): slider = "slider" +class ControlAfterGenerate(str, Enum): + fixed = "fixed" + increment = "increment" + decrement = "decrement" + randomize = "randomize" + class _ComfyType(ABC): Type = Any io_type: str = None @@ -263,7 +269,7 @@ class Int(ComfyTypeIO): class Input(WidgetInput): '''Integer input.''' def __init__(self, id: str, display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, - default: int=None, min: int=None, max: int=None, step: int=None, control_after_generate: bool=None, + default: int=None, min: int=None, max: int=None, step: int=None, control_after_generate: bool | ControlAfterGenerate=None, display_mode: NumberDisplay=None, socketless: bool=None, force_input: bool=None, extra_dict=None, raw_link: bool=None, advanced: bool=None): super().__init__(id, display_name, optional, tooltip, lazy, default, socketless, None, force_input, extra_dict, raw_link, advanced) self.min = min @@ -345,7 +351,7 @@ class Combo(ComfyTypeIO): tooltip: str=None, lazy: bool=None, default: str | int | Enum = None, - control_after_generate: bool=None, + control_after_generate: bool | ControlAfterGenerate=None, upload: UploadType=None, image_folder: FolderType=None, remote: RemoteOptions=None, @@ -389,7 +395,7 @@ class MultiCombo(ComfyTypeI): Type = list[str] class Input(Combo.Input): def __init__(self, id: str, options: list[str], display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None, - default: list[str]=None, placeholder: str=None, chip: bool=None, control_after_generate: bool=None, + default: list[str]=None, placeholder: str=None, chip: bool=None, control_after_generate: bool | ControlAfterGenerate=None, socketless: bool=None, extra_dict=None, raw_link: bool=None, advanced: bool=None): super().__init__(id, options, display_name, optional, tooltip, lazy, default, control_after_generate, socketless=socketless, extra_dict=extra_dict, raw_link=raw_link, advanced=advanced) self.multiselect = True @@ -2097,6 +2103,7 @@ __all__ = [ "UploadType", "RemoteOptions", "NumberDisplay", + "ControlAfterGenerate", "comfytype", "Custom", From 18927538a15d44c734653513e9fdbbe1e79a9f0c Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 16 Feb 2026 20:30:34 -0800 Subject: [PATCH 645/647] Implement NAG on all the models based on the Flux code. (#12500) Use the Normalized Attention Guidance node. Flux, Flux2, Klein, Chroma, Chroma radiance, Hunyuan Video, etc.. --- comfy/ldm/chroma/model.py | 2 + comfy/ldm/flux/layers.py | 18 ++++++ comfy/ldm/flux/model.py | 2 + comfy/ldm/hunyuan_video/model.py | 2 + comfy/model_patcher.py | 5 +- comfy_extras/nodes_nag.py | 99 ++++++++++++++++++++++++++++++++ nodes.py | 1 + 7 files changed, 128 insertions(+), 1 deletion(-) create mode 100644 comfy_extras/nodes_nag.py diff --git a/comfy/ldm/chroma/model.py b/comfy/ldm/chroma/model.py index 2e8ef0687..9fd865f20 100644 --- a/comfy/ldm/chroma/model.py +++ b/comfy/ldm/chroma/model.py @@ -152,6 +152,7 @@ class Chroma(nn.Module): transformer_options={}, attn_mask: Tensor = None, ) -> Tensor: + transformer_options = transformer_options.copy() patches_replace = transformer_options.get("patches_replace", {}) # running on sequences img @@ -228,6 +229,7 @@ class Chroma(nn.Module): transformer_options["total_blocks"] = len(self.single_blocks) transformer_options["block_type"] = "single" + transformer_options["img_slice"] = [txt.shape[1], img.shape[1]] for i, block in enumerate(self.single_blocks): transformer_options["block_index"] = i if i not in self.skip_dit: diff --git a/comfy/ldm/flux/layers.py b/comfy/ldm/flux/layers.py index 3518a1922..8b3f500d7 100644 --- a/comfy/ldm/flux/layers.py +++ b/comfy/ldm/flux/layers.py @@ -196,6 +196,9 @@ class DoubleStreamBlock(nn.Module): else: (img_mod1, img_mod2), (txt_mod1, txt_mod2) = vec + transformer_patches = transformer_options.get("patches", {}) + extra_options = transformer_options.copy() + # prepare image for attention img_modulated = self.img_norm1(img) img_modulated = apply_mod(img_modulated, (1 + img_mod1.scale), img_mod1.shift, modulation_dims_img) @@ -224,6 +227,12 @@ class DoubleStreamBlock(nn.Module): attn = attention(q, k, v, pe=pe, mask=attn_mask, transformer_options=transformer_options) del q, k, v + if "attn1_output_patch" in transformer_patches: + extra_options["img_slice"] = [txt.shape[1], attn.shape[1]] + patch = transformer_patches["attn1_output_patch"] + for p in patch: + attn = p(attn, extra_options) + txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1]:] # calculate the img bloks @@ -303,6 +312,9 @@ class SingleStreamBlock(nn.Module): else: mod = vec + transformer_patches = transformer_options.get("patches", {}) + extra_options = transformer_options.copy() + qkv, mlp = torch.split(self.linear1(apply_mod(self.pre_norm(x), (1 + mod.scale), mod.shift, modulation_dims)), [3 * self.hidden_size, self.mlp_hidden_dim_first], dim=-1) q, k, v = qkv.view(qkv.shape[0], qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4) @@ -312,6 +324,12 @@ class SingleStreamBlock(nn.Module): # compute attention attn = attention(q, k, v, pe=pe, mask=attn_mask, transformer_options=transformer_options) del q, k, v + + if "attn1_output_patch" in transformer_patches: + patch = transformer_patches["attn1_output_patch"] + for p in patch: + attn = p(attn, extra_options) + # compute activation in mlp stream, cat again and run second linear layer if self.yak_mlp: mlp = self.mlp_act(mlp[..., self.mlp_hidden_dim_first // 2:]) * mlp[..., :self.mlp_hidden_dim_first // 2] diff --git a/comfy/ldm/flux/model.py b/comfy/ldm/flux/model.py index 260ccad7e..ef4dcf7c5 100644 --- a/comfy/ldm/flux/model.py +++ b/comfy/ldm/flux/model.py @@ -142,6 +142,7 @@ class Flux(nn.Module): attn_mask: Tensor = None, ) -> Tensor: + transformer_options = transformer_options.copy() patches = transformer_options.get("patches", {}) patches_replace = transformer_options.get("patches_replace", {}) if img.ndim != 3 or txt.ndim != 3: @@ -231,6 +232,7 @@ class Flux(nn.Module): transformer_options["total_blocks"] = len(self.single_blocks) transformer_options["block_type"] = "single" + transformer_options["img_slice"] = [txt.shape[1], img.shape[1]] for i, block in enumerate(self.single_blocks): transformer_options["block_index"] = i if ("single_block", i) in blocks_replace: diff --git a/comfy/ldm/hunyuan_video/model.py b/comfy/ldm/hunyuan_video/model.py index 563f28f6b..b94cdfa87 100644 --- a/comfy/ldm/hunyuan_video/model.py +++ b/comfy/ldm/hunyuan_video/model.py @@ -304,6 +304,7 @@ class HunyuanVideo(nn.Module): control=None, transformer_options={}, ) -> Tensor: + transformer_options = transformer_options.copy() patches_replace = transformer_options.get("patches_replace", {}) initial_shape = list(img.shape) @@ -416,6 +417,7 @@ class HunyuanVideo(nn.Module): transformer_options["total_blocks"] = len(self.single_blocks) transformer_options["block_type"] = "single" + transformer_options["img_slice"] = [txt.shape[1], img.shape[1]] for i, block in enumerate(self.single_blocks): transformer_options["block_index"] = i if ("single_block", i) in blocks_replace: diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index f01818f50..21b4ce53e 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -406,13 +406,16 @@ class ModelPatcher: def memory_required(self, input_shape): return self.model.memory_required(input_shape=input_shape) + def disable_model_cfg1_optimization(self): + self.model_options["disable_cfg1_optimization"] = True + def set_model_sampler_cfg_function(self, sampler_cfg_function, disable_cfg1_optimization=False): if len(inspect.signature(sampler_cfg_function).parameters) == 3: self.model_options["sampler_cfg_function"] = lambda args: sampler_cfg_function(args["cond"], args["uncond"], args["cond_scale"]) #Old way else: self.model_options["sampler_cfg_function"] = sampler_cfg_function if disable_cfg1_optimization: - self.model_options["disable_cfg1_optimization"] = True + self.disable_model_cfg1_optimization() def set_model_sampler_post_cfg_function(self, post_cfg_function, disable_cfg1_optimization=False): self.model_options = set_model_options_post_cfg_function(self.model_options, post_cfg_function, disable_cfg1_optimization) diff --git a/comfy_extras/nodes_nag.py b/comfy_extras/nodes_nag.py new file mode 100644 index 000000000..033e40eb9 --- /dev/null +++ b/comfy_extras/nodes_nag.py @@ -0,0 +1,99 @@ +import torch +from comfy_api.latest import ComfyExtension, io +from typing_extensions import override + + +class NAGuidance(io.ComfyNode): + @classmethod + def define_schema(cls) -> io.Schema: + return io.Schema( + node_id="NAGuidance", + display_name="Normalized Attention Guidance", + description="Applies Normalized Attention Guidance to models, enabling negative prompts on distilled/schnell models.", + category="", + is_experimental=True, + inputs=[ + io.Model.Input("model", tooltip="The model to apply NAG to."), + io.Float.Input("nag_scale", min=0.0, default=5.0, max=50.0, step=0.1, tooltip="The guidance scale factor. Higher values push further from the negative prompt."), + io.Float.Input("nag_alpha", min=0.0, default=0.5, max=1.0, step=0.01, tooltip="Blending factor for the normalized attention. 1.0 is full replacement, 0.0 is no effect."), + io.Float.Input("nag_tau", min=1.0, default=1.5, max=10.0, step=0.01), + # io.Float.Input("start_percent", min=0.0, default=0.0, max=1.0, step=0.01, tooltip="The relative sampling step to begin applying NAG."), + # io.Float.Input("end_percent", min=0.0, default=1.0, max=1.0, step=0.01, tooltip="The relative sampling step to stop applying NAG."), + ], + outputs=[ + io.Model.Output(tooltip="The patched model with NAG enabled."), + ], + ) + + @classmethod + def execute(cls, model: io.Model.Type, nag_scale: float, nag_alpha: float, nag_tau: float) -> io.NodeOutput: + m = model.clone() + + # sigma_start = m.get_model_object("model_sampling").percent_to_sigma(start_percent) + # sigma_end = m.get_model_object("model_sampling").percent_to_sigma(end_percent) + + def nag_attention_output_patch(out, extra_options): + cond_or_uncond = extra_options.get("cond_or_uncond", None) + if cond_or_uncond is None: + return out + + if not (1 in cond_or_uncond and 0 in cond_or_uncond): + return out + + # sigma = extra_options.get("sigmas", None) + # if sigma is not None and len(sigma) > 0: + # sigma = sigma[0].item() + # if sigma > sigma_start or sigma < sigma_end: + # return out + + img_slice = extra_options.get("img_slice", None) + + if img_slice is not None: + orig_out = out + out = out[:, img_slice[0]:img_slice[1]] # only apply on img part + + batch_size = out.shape[0] + half_size = batch_size // len(cond_or_uncond) + + ind_neg = cond_or_uncond.index(1) + ind_pos = cond_or_uncond.index(0) + z_pos = out[half_size * ind_pos:half_size * (ind_pos + 1)] + z_neg = out[half_size * ind_neg:half_size * (ind_neg + 1)] + + guided = z_pos * nag_scale - z_neg * (nag_scale - 1.0) + + eps = 1e-6 + norm_pos = torch.norm(z_pos, p=1, dim=-1, keepdim=True).clamp_min(eps) + norm_guided = torch.norm(guided, p=1, dim=-1, keepdim=True).clamp_min(eps) + + ratio = norm_guided / norm_pos + scale_factor = torch.minimum(ratio, torch.full_like(ratio, nag_tau)) / ratio + + guided_normalized = guided * scale_factor + + z_final = guided_normalized * nag_alpha + z_pos * (1.0 - nag_alpha) + + if img_slice is not None: + orig_out[half_size * ind_neg:half_size * (ind_neg + 1), img_slice[0]:img_slice[1]] = z_final + orig_out[half_size * ind_pos:half_size * (ind_pos + 1), img_slice[0]:img_slice[1]] = z_final + return orig_out + else: + out[half_size * ind_pos:half_size * (ind_pos + 1)] = z_final + return out + + m.set_model_attn1_output_patch(nag_attention_output_patch) + m.disable_model_cfg1_optimization() + + return io.NodeOutput(m) + + +class NagExtension(ComfyExtension): + @override + async def get_node_list(self) -> list[type[io.ComfyNode]]: + return [ + NAGuidance, + ] + + +async def comfy_entrypoint() -> NagExtension: + return NagExtension() diff --git a/nodes.py b/nodes.py index db5f98408..dff56b79c 100644 --- a/nodes.py +++ b/nodes.py @@ -2437,6 +2437,7 @@ async def init_builtin_extra_nodes(): "nodes_color.py", "nodes_toolkit.py", "nodes_replacements.py", + "nodes_nag.py", ] import_failed = [] From c39653163d77161b2df2d57419129a4d6d081aa1 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 16 Feb 2026 21:29:20 -0800 Subject: [PATCH 646/647] Fix anima preprocess text embeds not using right inference dtype. (#12501) --- comfy/model_base.py | 18 ++++++++++-------- 1 file changed, 10 insertions(+), 8 deletions(-) diff --git a/comfy/model_base.py b/comfy/model_base.py index 4a74cb1ce..9dcef8741 100644 --- a/comfy/model_base.py +++ b/comfy/model_base.py @@ -178,10 +178,7 @@ class BaseModel(torch.nn.Module): xc = torch.cat([xc] + [comfy.model_management.cast_to_device(c_concat, xc.device, xc.dtype)], dim=1) context = c_crossattn - dtype = self.get_dtype() - - if self.manual_cast_dtype is not None: - dtype = self.manual_cast_dtype + dtype = self.get_dtype_inference() xc = xc.to(dtype) device = xc.device @@ -218,6 +215,13 @@ class BaseModel(torch.nn.Module): def get_dtype(self): return self.diffusion_model.dtype + def get_dtype_inference(self): + dtype = self.get_dtype() + + if self.manual_cast_dtype is not None: + dtype = self.manual_cast_dtype + return dtype + def encode_adm(self, **kwargs): return None @@ -372,9 +376,7 @@ class BaseModel(torch.nn.Module): input_shapes += shape if comfy.model_management.xformers_enabled() or comfy.model_management.pytorch_attention_flash_attention(): - dtype = self.get_dtype() - if self.manual_cast_dtype is not None: - dtype = self.manual_cast_dtype + dtype = self.get_dtype_inference() #TODO: this needs to be tweaked area = sum(map(lambda input_shape: input_shape[0] * math.prod(input_shape[2:]), input_shapes)) return (area * comfy.model_management.dtype_size(dtype) * 0.01 * self.memory_usage_factor) * (1024 * 1024) @@ -1165,7 +1167,7 @@ class Anima(BaseModel): t5xxl_ids = t5xxl_ids.unsqueeze(0) if torch.is_inference_mode_enabled(): # if not we are training - cross_attn = self.diffusion_model.preprocess_text_embeds(cross_attn.to(device=device, dtype=self.get_dtype()), t5xxl_ids.to(device=device), t5xxl_weights=t5xxl_weights.to(device=device, dtype=self.get_dtype())) + cross_attn = self.diffusion_model.preprocess_text_embeds(cross_attn.to(device=device, dtype=self.get_dtype_inference()), t5xxl_ids.to(device=device), t5xxl_weights=t5xxl_weights.to(device=device, dtype=self.get_dtype_inference())) else: out['t5xxl_ids'] = comfy.conds.CONDRegular(t5xxl_ids) out['t5xxl_weights'] = comfy.conds.CONDRegular(t5xxl_weights) From fe52843fe55b92dedaabff684294dd7a115d2204 Mon Sep 17 00:00:00 2001 From: comfyanonymous Date: Tue, 17 Feb 2026 00:39:54 -0500 Subject: [PATCH 647/647] ComfyUI v0.14.0 --- comfyui_version.py | 2 +- pyproject.toml | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/comfyui_version.py b/comfyui_version.py index cf4e89816..8f7f3228e 100644 --- a/comfyui_version.py +++ b/comfyui_version.py @@ -1,3 +1,3 @@ # This file is automatically generated by the build process when version is # updated in pyproject.toml. -__version__ = "0.13.0" +__version__ = "0.14.0" diff --git a/pyproject.toml b/pyproject.toml index 9dab9a50c..b132bb9c4 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -1,6 +1,6 @@ [project] name = "ComfyUI" -version = "0.13.0" +version = "0.14.0" readme = "README.md" license = { file = "LICENSE" } requires-python = ">=3.10"