Merge remote-tracking branch 'origin/master' into feat/cache-provider-api

.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"

.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:

.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 }}"

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/)
@@ -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:
 
@@ -208,11 +208,11 @@ 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
 
-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 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```
 
@@ -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):
 

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] = []

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))

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:

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
@@ -55,12 +57,16 @@ 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()
         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]
@@ -107,10 +113,14 @@ 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")

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]
+}

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."""
 

comfy/float.py

@@ -65,3 +65,147 @@ 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):
+    orig_shape = x.shape
+    sign = torch.signbit(x).to(torch.uint8)
+
+    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 = x.abs()
+    exp = torch.floor(torch.log2(x) + 1.1925).clamp(0, 3)
+
+    mantissa = torch.where(
+        exp > 0,
+        (x / (2.0 ** (exp - 1)) - 1.0) * 2.0,
+        (x * 2.0),
+        out=x
+    ).round().to(torch.uint8)
+    del x
+
+    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(orig_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_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
+
+    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
+
+    orig_shape = list(orig_shape)
+
+    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)
+
+    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)

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 =[
@@ -592,6 +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 = [
@@ -725,6 +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"
 
@@ -749,6 +753,7 @@ class ACEAudio(LatentFormat):
 
 class ChromaRadiance(LatentFormat):
     latent_channels = 3
+    spacial_downscale_ratio = 1
 
     def __init__(self):
         self.latent_rgb_factors = [

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

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), 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 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
+
+            # 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 = (
@@ -146,28 +212,12 @@ 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,
-        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)
@@ -177,144 +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
 
@@ -526,9 +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)
@@ -543,72 +562,82 @@ 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")
+        has_spatial_mask = kwargs.get("has_spatial_mask", None)
+        v_patches_per_frame = None
+        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]
+
+        # 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)
+            # 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 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]),
+            ]
+
             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 +796,11 @@ 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

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,
@@ -189,9 +179,12 @@ 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

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):

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):

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.
@@ -378,10 +446,12 @@ 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,
         operations=None,
+        **kwargs,
     ) -> None:
         super().__init__()
         self.dtype = dtype
@@ -491,6 +561,41 @@ 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 +636,168 @@ 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) or self.siglip_embedder is None:
+            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:
+                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,)
+                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]):
+                    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):]
+
+        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 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)
+            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 +807,11 @@ 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 +826,6 @@ 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 +837,7 @@ 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 +845,7 @@ 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 +854,7 @@ 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
 

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):
     """

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]
 

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

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

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)
 
@@ -472,10 +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]
@@ -495,10 +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]

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:
@@ -322,6 +323,7 @@ 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:

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
@@ -1147,9 +1148,31 @@ 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)
+        self.memory_usage_factor_conds = ("ref_latents",)
 
     def extra_conds(self, **kwargs):
         out = super().extra_conds(**kwargs)
@@ -1169,6 +1192,35 @@ 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):
@@ -1526,6 +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)

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
@@ -251,7 +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
@@ -442,8 +444,15 @@ 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)
+            if sig_weight is not None:
+                dit_config["siglip_feat_dim"] = sig_weight.shape[0]
 
         return dit_config
 
@@ -545,6 +554,8 @@ 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

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

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)
@@ -546,7 +548,8 @@ 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]
 
@@ -624,21 +627,29 @@ 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):
@@ -690,7 +701,7 @@ 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:

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_by_block(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
 

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

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
@@ -57,6 +58,7 @@ 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
@@ -100,6 +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:
@@ -635,14 +736,13 @@ 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
@@ -1014,6 +1114,7 @@ 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 +1147,8 @@ class TEModel(Enum):
     QWEN3_2B = 17
     GEMMA_3_12B = 18
     JINA_CLIP_2 = 19
+    QWEN3_8B = 20
+    QWEN3_06B = 21
 
 
 def detect_te_model(sd):
@@ -1059,9 +1162,9 @@ 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
@@ -1089,6 +1192,10 @@ 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
+            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
@@ -1214,14 +1321,24 @@ 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
+        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:

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:

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
@@ -763,17 +764,31 @@ 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)
         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 = {
@@ -845,7 +860,7 @@ 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)
@@ -992,6 +1007,36 @@ 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",
@@ -1042,13 +1087,13 @@ class ZImage(Lumina2):
         "shift": 3.0,
     }
 
-    memory_usage_factor = 2.0
+    memory_usage_factor = 2.8
 
     supported_inference_dtypes = [torch.bfloat16, torch.float32]
 
     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)
 
@@ -1551,6 +1596,6 @@ 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]

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:

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_

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_

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
@@ -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}
@@ -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<think>\n\n</think>\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_

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_

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:

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
@@ -99,6 +121,28 @@ 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
@@ -619,6 +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__()
@@ -628,6 +681,15 @@ 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__()

comfy/text_encoders/lt.py

@@ -119,7 +119,17 @@ class LTXAVTEModel(torch.nn.Module):
             if len(sdo) == 0:
                 sdo = sd
 
-            return self.load_state_dict(sdo, strict=False)
+            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

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_

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_

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_

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
@@ -610,6 +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",
@@ -638,6 +647,8 @@ 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:
@@ -928,7 +939,9 @@ 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)
@@ -1097,6 +1110,10 @@ 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 +1121,8 @@ 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 +1131,29 @@ 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)

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]

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)

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

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)

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

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)

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

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

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

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))

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")
@@ -751,7 +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")
@@ -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):
@@ -997,20 +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")
@@ -1119,8 +1140,8 @@ 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()
@@ -1148,6 +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).'
 
@@ -1224,7 +1247,10 @@ 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
 
 @dataclass
 class NodeInfoV3:
@@ -1234,11 +1260,78 @@ 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
+    dev_only: 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
@@ -1256,6 +1349,8 @@ 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.
 
@@ -1282,12 +1377,18 @@ 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
+    """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
     """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:
@@ -1314,6 +1415,8 @@ 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."""
@@ -1386,8 +1489,11 @@ 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")
+            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
 
@@ -1418,8 +1524,10 @@ 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")
+            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
 
@@ -1428,6 +1536,7 @@ 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
@@ -1437,8 +1546,12 @@ 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:
@@ -1475,11 +1588,16 @@ 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)
@@ -1493,6 +1611,11 @@ 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
@@ -1674,6 +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
@@ -1738,6 +1869,14 @@ 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]:
@@ -1768,6 +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:
@@ -1776,6 +1917,8 @@ 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 = []
@@ -1923,6 +2066,7 @@ __all__ = [
     "ControlNet",
     "Vae",
     "Model",
+    "ModelPatch",
     "ClipVision",
     "ClipVisionOutput",
     "AudioEncoder",
@@ -1971,4 +2115,6 @@ __all__ = [
     "add_to_dict_v3",
     "V3Data",
     "ImageCompare",
+    "PriceBadgeDepends",
+    "PriceBadge",
 ]

comfy_api_nodes/README.md

@@ -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
-
-```

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)

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)
 
@@ -65,11 +54,13 @@ 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 +132,9 @@ VIDEO_TASKS_EXECUTION_TIME = {
         "720p": 65,
         "1080p": 100,
     },
+    "seedance-1-5-pro-251215": {
+        "480p": 80,
+        "720p": 100,
+        "1080p": 150,
+    },
 }

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(...)

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.'
+    )

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

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)

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'
+    )

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)

comfy_api_nodes/apis/openai_api.py

@@ -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)")

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')

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))

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")

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)

comfy_api_nodes/canary.py

@@ -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 = {}

comfy_api_nodes/mapper_utils.py

@@ -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

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,
@@ -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):

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 FIBO 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()

comfy_api_nodes/nodes_bytedance.py

@@ -5,11 +5,10 @@ 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,
-    Image2ImageTaskCreationRequest,
     Image2VideoTaskCreationRequest,
     ImageTaskCreationResponse,
     Seedream4Options,
@@ -126,6 +125,9 @@ class ByteDanceImageNode(IO.ComfyNode):
                 IO.Hidden.unique_id,
             ],
             is_api_node=True,
+            price_badge=IO.PriceBadge(
+                expr="""{"type":"usd","usd":0.03}""",
+            ),
         )
 
     @classmethod
@@ -171,99 +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
@@ -367,6 +276,19 @@ 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
@@ -461,7 +383,12 @@ 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(
@@ -512,6 +439,12 @@ 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(),
@@ -522,6 +455,7 @@ class ByteDanceTextToVideoNode(IO.ComfyNode):
                 IO.Hidden.unique_id,
             ],
             is_api_node=True,
+            price_badge=PRICE_BADGE_VIDEO,
         )
 
     @classmethod
@@ -535,7 +469,10 @@ 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"])
 
@@ -550,7 +487,11 @@ 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))),
         )
 
@@ -567,7 +508,12 @@ 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(
@@ -622,6 +568,12 @@ 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(),
@@ -632,6 +584,7 @@ class ByteDanceImageToVideoNode(IO.ComfyNode):
                 IO.Hidden.unique_id,
             ],
             is_api_node=True,
+            price_badge=PRICE_BADGE_VIDEO,
         )
 
     @classmethod
@@ -646,7 +599,10 @@ 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)
@@ -668,6 +624,7 @@ 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))),
         )
@@ -685,7 +642,7 @@ 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(
@@ -744,6 +701,12 @@ 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(),
@@ -754,6 +717,7 @@ class ByteDanceFirstLastFrameNode(IO.ComfyNode):
                 IO.Hidden.unique_id,
             ],
             is_api_node=True,
+            price_badge=PRICE_BADGE_VIDEO,
         )
 
     @classmethod
@@ -769,7 +733,10 @@ 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):
@@ -802,6 +769,7 @@ 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))),
         )
@@ -877,6 +845,41 @@ class ByteDanceImageReferenceNode(IO.ComfyNode):
                 IO.Hidden.unique_id,
             ],
             is_api_node=True,
+            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
@@ -946,12 +949,64 @@ 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", "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],
+          "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-1-5-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";
+      $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;
+      $audioMultiplier := ($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, "format": { "approximate": true }}
+        : {"type":"range_usd","min_usd": $minCost, "max_usd": $maxCost, "format": { "approximate": true }}
+    )
+    """,
+)
+
+
 class ByteDanceExtension(ComfyExtension):
     @override
     async def get_node_list(self) -> list[type[IO.ComfyNode]]:
         return [
             ByteDanceImageNode,
-            ByteDanceImageEditNode,
             ByteDanceSeedreamNode,
             ByteDanceTextToVideoNode,
             ByteDanceImageToVideoNode,

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,
@@ -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
 
 
@@ -309,6 +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
@@ -570,6 +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
@@ -700,6 +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

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()

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,
@@ -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

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,
@@ -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))
 
@@ -567,7 +566,7 @@ 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:
@@ -764,6 +763,33 @@ 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 +844,16 @@ 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 +922,16 @@ 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 +1027,16 @@ 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 +1112,16 @@ 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 +1208,16 @@ 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 +1304,9 @@ 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 +1392,9 @@ 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 +1457,33 @@ 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 +1557,9 @@ 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 +1630,33 @@ 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 +1722,9 @@ 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 +1806,29 @@ 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 +1893,16 @@ 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 +1957,9 @@ 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 +2020,9 @@ 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 +2073,9 @@ 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 +2175,19 @@ 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 +2271,10 @@ 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 +2339,10 @@ 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 +2423,15 @@ 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

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

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,
@@ -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]]:

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()

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()

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,
@@ -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

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,
@@ -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

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,
@@ -160,6 +154,23 @@ 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
@@ -249,7 +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,
@@ -287,6 +298,25 @@ 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
@@ -334,9 +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",
@@ -348,7 +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,
@@ -399,6 +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,
                 ),
             ],
@@ -411,6 +442,28 @@ 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
@@ -442,8 +495,8 @@ 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)
@@ -465,7 +518,7 @@ 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)
@@ -566,32 +619,95 @@ 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
-    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,
@@ -605,9 +721,9 @@ 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:
@@ -619,13 +735,9 @@ 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(
@@ -635,7 +747,7 @@ 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)
 
@@ -646,36 +758,28 @@ 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)))
 
 
@@ -796,7 +900,7 @@ class OpenAIChatConfig(IO.ComfyNode):
             remove depending on model choice.
         """
         return IO.NodeOutput(
-            CreateModelResponseProperties(
+            ModelResponseProperties(
                 instructions=instructions,
                 truncation=truncation,
                 max_output_tokens=max_output_tokens,

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,
@@ -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]]:

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,
@@ -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}""",
+            ),
         )