ComfyUI version 0.3.51

* Bump template to 0.1.61

* Bump template to 0.1.62

.github/ISSUE_TEMPLATE/bug-report.yml

@@ -22,7 +22,7 @@ body:
       description: Please confirm you have tried to reproduce the issue with all custom nodes disabled.
       options:
         - label: I have tried disabling custom nodes and the issue persists (see [how to disable custom nodes](https://docs.comfy.org/troubleshooting/custom-node-issues#step-1%3A-test-with-all-custom-nodes-disabled) if you need help)
-          required: true
+          required: false
   - type: textarea
     attributes:
       label: Expected Behavior

.github/ISSUE_TEMPLATE/user-support.yml

@@ -18,7 +18,7 @@ body:
         description: Please confirm you have tried to reproduce the issue with all custom nodes disabled.
         options:
           - label: I have tried disabling custom nodes and the issue persists (see [how to disable custom nodes](https://docs.comfy.org/troubleshooting/custom-node-issues#step-1%3A-test-with-all-custom-nodes-disabled) if you need help)
-            required: true
+            required: false
     - type: textarea
       attributes:
             label: Your question

.github/workflows/stable-release.yml

@@ -90,7 +90,7 @@ jobs:
 
           cd ..
 
-          "C:\Program Files\7-Zip\7z.exe" a -t7z -m0=lzma2 -mx=9 -mfb=128 -md=512m -ms=on -mf=BCJ2 ComfyUI_windows_portable.7z ComfyUI_windows_portable
+          "C:\Program Files\7-Zip\7z.exe" a -t7z -m0=lzma2 -mx=9 -mfb=128 -md=768m -ms=on -mf=BCJ2 ComfyUI_windows_portable.7z ComfyUI_windows_portable
           mv ComfyUI_windows_portable.7z ComfyUI/ComfyUI_windows_portable_nvidia.7z
 
           cd ComfyUI_windows_portable

.github/workflows/windows_release_package.yml

@@ -86,7 +86,7 @@ jobs:
 
             cd ..
 
-            "C:\Program Files\7-Zip\7z.exe" a -t7z -m0=lzma2 -mx=9 -mfb=128 -md=512m -ms=on -mf=BCJ2 ComfyUI_windows_portable.7z ComfyUI_windows_portable
+            "C:\Program Files\7-Zip\7z.exe" a -t7z -m0=lzma2 -mx=9 -mfb=128 -md=768m -ms=on -mf=BCJ2 ComfyUI_windows_portable.7z ComfyUI_windows_portable
             mv ComfyUI_windows_portable.7z ComfyUI/new_ComfyUI_windows_portable_nvidia_cu${{ inputs.cu }}_or_cpu.7z
 
             cd ComfyUI_windows_portable

CODEOWNERS

@@ -5,20 +5,21 @@
 # Inlined the team members for now.
 
 # Maintainers
-*.md @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne
-/tests/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne
-/tests-unit/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne
-/notebooks/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne
-/script_examples/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne
-/.github/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne
-/requirements.txt @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne
-/pyproject.toml @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne
+*.md @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne @guill
+/tests/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne @guill
+/tests-unit/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne @guill
+/notebooks/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne @guill
+/script_examples/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne @guill
+/.github/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne @guill
+/requirements.txt @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne @guill
+/pyproject.toml @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne @guill
 
 # Python web server
-/api_server/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @christian-byrne
-/app/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @christian-byrne
-/utils/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @christian-byrne
+/api_server/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @christian-byrne @guill
+/app/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @christian-byrne @guill
+/utils/ @yoland68 @robinjhuang @webfiltered @pythongosssss @ltdrdata @christian-byrne @guill
 
 # Node developers
-/comfy_extras/ @yoland68 @robinjhuang @pythongosssss @ltdrdata @Kosinkadink @webfiltered @christian-byrne
-/comfy/comfy_types/ @yoland68 @robinjhuang @pythongosssss @ltdrdata @Kosinkadink @webfiltered @christian-byrne
+/comfy_extras/ @yoland68 @robinjhuang @pythongosssss @ltdrdata @Kosinkadink @webfiltered @christian-byrne @guill
+/comfy/comfy_types/ @yoland68 @robinjhuang @pythongosssss @ltdrdata @Kosinkadink @webfiltered @christian-byrne @guill
+/comfy_api_nodes/ @yoland68 @robinjhuang @pythongosssss @ltdrdata @Kosinkadink @webfiltered @christian-byrne @guill

README.md

@@ -39,7 +39,7 @@ ComfyUI lets you design and execute advanced stable diffusion pipelines using a
 ## Get Started
 
 #### [Desktop Application](https://www.comfy.org/download)
-- The easiest way to get started. 
+- The easiest way to get started.
 - Available on Windows & macOS.
 
 #### [Windows Portable Package](#installing)
@@ -211,27 +211,19 @@ This is the command to install the nightly with ROCm 6.4 which might have some p
 
 ### Intel GPUs (Windows and Linux)
 
-(Option 1) Intel Arc GPU users can install native PyTorch with torch.xpu support using pip (currently available in PyTorch nightly builds). More information can be found [here](https://pytorch.org/docs/main/notes/get_start_xpu.html)
-  
-1. To install PyTorch nightly, use the following command:
+(Option 1) Intel Arc GPU users can install native PyTorch with torch.xpu support using pip. More information can be found [here](https://pytorch.org/docs/main/notes/get_start_xpu.html)
+
+1. To install PyTorch xpu, use the following command:
+
+```pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/xpu```
+
+This is the command to install the Pytorch xpu nightly which might have some performance improvements:
 
 ```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/xpu```
 
-2. Launch ComfyUI by running `python main.py`
-
-
 (Option 2) Alternatively, Intel GPUs supported by Intel Extension for PyTorch (IPEX) can leverage IPEX for improved performance.
 
-1. For Intel® Arc™ A-Series Graphics utilizing IPEX, create a conda environment and use the commands below:
-
-```
-conda install libuv
-pip install torch==2.3.1.post0+cxx11.abi torchvision==0.18.1.post0+cxx11.abi torchaudio==2.3.1.post0+cxx11.abi intel-extension-for-pytorch==2.3.110.post0+xpu --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/ --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/cn/
-```
-
-For other supported Intel GPUs with IPEX, visit [Installation](https://intel.github.io/intel-extension-for-pytorch/index.html#installation?platform=gpu) for more information.
-
-Additional discussion and help can be found [here](https://github.com/comfyanonymous/ComfyUI/discussions/476).
+1. visit [Installation](https://intel.github.io/intel-extension-for-pytorch/index.html#installation?platform=gpu) for more information.
 
 ### NVIDIA
 
@@ -352,7 +344,7 @@ Generate a self-signed certificate (not appropriate for shared/production use) a
 
 Use `--tls-keyfile key.pem --tls-certfile cert.pem` to enable TLS/SSL, the app will now be accessible with `https://...` instead of `http://...`.
 
-> Note: Windows users can use [alexisrolland/docker-openssl](https://github.com/alexisrolland/docker-openssl) or one of the [3rd party binary distributions](https://wiki.openssl.org/index.php/Binaries) to run the command example above. 
+> Note: Windows users can use [alexisrolland/docker-openssl](https://github.com/alexisrolland/docker-openssl) or one of the [3rd party binary distributions](https://wiki.openssl.org/index.php/Binaries) to run the command example above.
 <br/><br/>If you use a container, note that the volume mount `-v` can be a relative path so `... -v ".\:/openssl-certs" ...` would create the key & cert files in the current directory of your command prompt or powershell terminal.
 
 ## Support and dev channel

comfy/cli_args.py

@@ -132,6 +132,8 @@ parser.add_argument("--reserve-vram", type=float, default=None, help="Set the am
 
 parser.add_argument("--async-offload", action="store_true", help="Use async weight offloading.")
 
+parser.add_argument("--force-non-blocking", action="store_true", help="Force ComfyUI to use non-blocking operations for all applicable tensors. This may improve performance on some non-Nvidia systems but can cause issues with some workflows.")
+
 parser.add_argument("--default-hashing-function", type=str, choices=['md5', 'sha1', 'sha256', 'sha512'], default='sha256', help="Allows you to choose the hash function to use for duplicate filename / contents comparison. Default is sha256.")
 
 parser.add_argument("--disable-smart-memory", action="store_true", help="Force ComfyUI to agressively offload to regular ram instead of keeping models in vram when it can.")

comfy/clip_model.py

@@ -97,7 +97,7 @@ class CLIPTextModel_(torch.nn.Module):
         self.encoder = CLIPEncoder(num_layers, embed_dim, heads, intermediate_size, intermediate_activation, dtype, device, operations)
         self.final_layer_norm = operations.LayerNorm(embed_dim, dtype=dtype, device=device)
 
-    def forward(self, input_tokens=None, attention_mask=None, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=torch.float32):
+    def forward(self, input_tokens=None, attention_mask=None, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=torch.float32, embeds_info=[]):
         if embeds is not None:
             x = embeds + comfy.ops.cast_to(self.embeddings.position_embedding.weight, dtype=dtype, device=embeds.device)
         else:

comfy/context_windows.py

@@ -0,0 +1,540 @@
+from __future__ import annotations
+from typing import TYPE_CHECKING, Callable
+import torch
+import numpy as np
+import collections
+from dataclasses import dataclass
+from abc import ABC, abstractmethod
+import logging
+import comfy.model_management
+import comfy.patcher_extension
+if TYPE_CHECKING:
+    from comfy.model_base import BaseModel
+    from comfy.model_patcher import ModelPatcher
+    from comfy.controlnet import ControlBase
+
+
+class ContextWindowABC(ABC):
+    def __init__(self):
+        ...
+
+    @abstractmethod
+    def get_tensor(self, full: torch.Tensor) -> torch.Tensor:
+        """
+        Get torch.Tensor applicable to current window.
+        """
+        raise NotImplementedError("Not implemented.")
+
+    @abstractmethod
+    def add_window(self, full: torch.Tensor, to_add: torch.Tensor) -> torch.Tensor:
+        """
+        Apply torch.Tensor of window to the full tensor, in place. Returns reference to updated full tensor, not a copy.
+        """
+        raise NotImplementedError("Not implemented.")
+
+class ContextHandlerABC(ABC):
+    def __init__(self):
+        ...
+
+    @abstractmethod
+    def should_use_context(self, model: BaseModel, conds: list[list[dict]], x_in: torch.Tensor, timestep: torch.Tensor, model_options: dict[str]) -> bool:
+        raise NotImplementedError("Not implemented.")
+
+    @abstractmethod
+    def get_resized_cond(self, cond_in: list[dict], x_in: torch.Tensor, window: ContextWindowABC, device=None) -> list:
+        raise NotImplementedError("Not implemented.")
+
+    @abstractmethod
+    def execute(self, calc_cond_batch: Callable, model: BaseModel, conds: list[list[dict]], x_in: torch.Tensor, timestep: torch.Tensor, model_options: dict[str]):
+        raise NotImplementedError("Not implemented.")
+
+
+
+class IndexListContextWindow(ContextWindowABC):
+    def __init__(self, index_list: list[int], dim: int=0):
+        self.index_list = index_list
+        self.context_length = len(index_list)
+        self.dim = dim
+
+    def get_tensor(self, full: torch.Tensor, device=None, dim=None) -> torch.Tensor:
+        if dim is None:
+            dim = self.dim
+        if dim == 0 and full.shape[dim] == 1:
+            return full
+        idx = [slice(None)] * dim + [self.index_list]
+        return full[idx].to(device)
+
+    def add_window(self, full: torch.Tensor, to_add: torch.Tensor, dim=None) -> torch.Tensor:
+        if dim is None:
+            dim = self.dim
+        idx = [slice(None)] * dim + [self.index_list]
+        full[idx] += to_add
+        return full
+
+
+class IndexListCallbacks:
+    EVALUATE_CONTEXT_WINDOWS = "evaluate_context_windows"
+    COMBINE_CONTEXT_WINDOW_RESULTS = "combine_context_window_results"
+    EXECUTE_START = "execute_start"
+    EXECUTE_CLEANUP = "execute_cleanup"
+
+    def init_callbacks(self):
+        return {}
+
+
+@dataclass
+class ContextSchedule:
+    name: str
+    func: Callable
+
+@dataclass
+class ContextFuseMethod:
+    name: str
+    func: Callable
+
+ContextResults = collections.namedtuple("ContextResults", ['window_idx', 'sub_conds_out', 'sub_conds', 'window'])
+class IndexListContextHandler(ContextHandlerABC):
+    def __init__(self, context_schedule: ContextSchedule, fuse_method: ContextFuseMethod, context_length: int=1, context_overlap: int=0, context_stride: int=1, closed_loop=False, dim=0):
+        self.context_schedule = context_schedule
+        self.fuse_method = fuse_method
+        self.context_length = context_length
+        self.context_overlap = context_overlap
+        self.context_stride = context_stride
+        self.closed_loop = closed_loop
+        self.dim = dim
+        self._step = 0
+
+        self.callbacks = {}
+
+    def should_use_context(self, model: BaseModel, conds: list[list[dict]], x_in: torch.Tensor, timestep: torch.Tensor, model_options: dict[str]) -> bool:
+        # for now, assume first dim is batch - should have stored on BaseModel in actual implementation
+        if x_in.size(self.dim) > self.context_length:
+            logging.info(f"Using context windows {self.context_length} for {x_in.size(self.dim)} frames.")
+            return True
+        return False
+
+    def prepare_control_objects(self, control: ControlBase, device=None) -> ControlBase:
+        if control.previous_controlnet is not None:
+            self.prepare_control_objects(control.previous_controlnet, device)
+        return control
+
+    def get_resized_cond(self, cond_in: list[dict], x_in: torch.Tensor, window: IndexListContextWindow, device=None) -> list:
+        if cond_in is None:
+            return None
+        # reuse or resize cond items to match context requirements
+        resized_cond = []
+        # cond object is a list containing a dict - outer list is irrelevant, so just loop through it
+        for actual_cond in cond_in:
+            resized_actual_cond = actual_cond.copy()
+            # now we are in the inner dict - "pooled_output" is a tensor, "control" is a ControlBase object, "model_conds" is dictionary
+            for key in actual_cond:
+                try:
+                    cond_item = actual_cond[key]
+                    if isinstance(cond_item, torch.Tensor):
+                        # check that tensor is the expected length - x.size(0)
+                        if self.dim < cond_item.ndim and cond_item.size(self.dim) == x_in.size(self.dim):
+                            # if so, it's subsetting time - tell controls the expected indeces so they can handle them
+                            actual_cond_item = window.get_tensor(cond_item)
+                            resized_actual_cond[key] = actual_cond_item.to(device)
+                        else:
+                            resized_actual_cond[key] = cond_item.to(device)
+                    # look for control
+                    elif key == "control":
+                        resized_actual_cond[key] = self.prepare_control_objects(cond_item, device)
+                    elif isinstance(cond_item, dict):
+                        new_cond_item = cond_item.copy()
+                        # when in dictionary, look for tensors and CONDCrossAttn [comfy/conds.py] (has cond attr that is a tensor)
+                        for cond_key, cond_value in new_cond_item.items():
+                            if isinstance(cond_value, torch.Tensor):
+                                if cond_value.ndim < self.dim and cond_value.size(0) == x_in.size(self.dim):
+                                    new_cond_item[cond_key] = window.get_tensor(cond_value, device)
+                            # if has cond that is a Tensor, check if needs to be subset
+                            elif hasattr(cond_value, "cond") and isinstance(cond_value.cond, torch.Tensor):
+                                if cond_value.cond.ndim < self.dim and cond_value.cond.size(0) == x_in.size(self.dim):
+                                    new_cond_item[cond_key] = cond_value._copy_with(window.get_tensor(cond_value.cond, device))
+                            elif cond_key == "num_video_frames": # for SVD
+                                new_cond_item[cond_key] = cond_value._copy_with(cond_value.cond)
+                                new_cond_item[cond_key].cond = window.context_length
+                        resized_actual_cond[key] = new_cond_item
+                    else:
+                        resized_actual_cond[key] = cond_item
+                finally:
+                    del cond_item  # just in case to prevent VRAM issues
+            resized_cond.append(resized_actual_cond)
+        return resized_cond
+
+    def set_step(self, timestep: torch.Tensor, model_options: dict[str]):
+        mask = torch.isclose(model_options["transformer_options"]["sample_sigmas"], timestep, rtol=0.0001)
+        matches = torch.nonzero(mask)
+        if torch.numel(matches) == 0:
+            raise Exception("No sample_sigmas matched current timestep; something went wrong.")
+        self._step = int(matches[0].item())
+
+    def get_context_windows(self, model: BaseModel, x_in: torch.Tensor, model_options: dict[str]) -> list[IndexListContextWindow]:
+        full_length = x_in.size(self.dim) # TODO: choose dim based on model
+        context_windows = self.context_schedule.func(full_length, self, model_options)
+        context_windows = [IndexListContextWindow(window, dim=self.dim) for window in context_windows]
+        return context_windows
+
+    def execute(self, calc_cond_batch: Callable, model: BaseModel, conds: list[list[dict]], x_in: torch.Tensor, timestep: torch.Tensor, model_options: dict[str]):
+        self.set_step(timestep, model_options)
+        context_windows = self.get_context_windows(model, x_in, model_options)
+        enumerated_context_windows = list(enumerate(context_windows))
+
+        conds_final = [torch.zeros_like(x_in) for _ in conds]
+        if self.fuse_method.name == ContextFuseMethods.RELATIVE:
+            counts_final = [torch.ones(get_shape_for_dim(x_in, self.dim), device=x_in.device) for _ in conds]
+        else:
+            counts_final = [torch.zeros(get_shape_for_dim(x_in, self.dim), device=x_in.device) for _ in conds]
+        biases_final = [([0.0] * x_in.shape[self.dim]) for _ in conds]
+
+        for callback in comfy.patcher_extension.get_all_callbacks(IndexListCallbacks.EXECUTE_START, self.callbacks):
+            callback(self, model, x_in, conds, timestep, model_options)
+
+        for enum_window in enumerated_context_windows:
+            results = self.evaluate_context_windows(calc_cond_batch, model, x_in, conds, timestep, [enum_window], model_options)
+            for result in results:
+                self.combine_context_window_results(x_in, result.sub_conds_out, result.sub_conds, result.window, result.window_idx, len(enumerated_context_windows), timestep,
+                                            conds_final, counts_final, biases_final)
+        try:
+            # finalize conds
+            if self.fuse_method.name == ContextFuseMethods.RELATIVE:
+                # relative is already normalized, so return as is
+                del counts_final
+                return conds_final
+            else:
+                # normalize conds via division by context usage counts
+                for i in range(len(conds_final)):
+                    conds_final[i] /= counts_final[i]
+                del counts_final
+                return conds_final
+        finally:
+            for callback in comfy.patcher_extension.get_all_callbacks(IndexListCallbacks.EXECUTE_CLEANUP, self.callbacks):
+                callback(self, model, x_in, conds, timestep, model_options)
+
+    def evaluate_context_windows(self, calc_cond_batch: Callable, model: BaseModel, x_in: torch.Tensor, conds, timestep: torch.Tensor, enumerated_context_windows: list[tuple[int, IndexListContextWindow]],
+                                model_options, device=None, first_device=None):
+        results: list[ContextResults] = []
+        for window_idx, window in enumerated_context_windows:
+            # allow processing to end between context window executions for faster Cancel
+            comfy.model_management.throw_exception_if_processing_interrupted()
+
+            for callback in comfy.patcher_extension.get_all_callbacks(IndexListCallbacks.EVALUATE_CONTEXT_WINDOWS, self.callbacks):
+                callback(self, model, x_in, conds, timestep, model_options, window_idx, window, model_options, device, first_device)
+
+            # update exposed params
+            model_options["transformer_options"]["context_window"] = window
+            # get subsections of x, timestep, conds
+            sub_x = window.get_tensor(x_in, device)
+            sub_timestep = window.get_tensor(timestep, device, dim=0)
+            sub_conds = [self.get_resized_cond(cond, x_in, window, device) for cond in conds]
+
+            sub_conds_out = calc_cond_batch(model, sub_conds, sub_x, sub_timestep, model_options)
+            if device is not None:
+                for i in range(len(sub_conds_out)):
+                    sub_conds_out[i] = sub_conds_out[i].to(x_in.device)
+            results.append(ContextResults(window_idx, sub_conds_out, sub_conds, window))
+        return results
+
+
+    def combine_context_window_results(self, x_in: torch.Tensor, sub_conds_out, sub_conds, window: IndexListContextWindow, window_idx: int, total_windows: int, timestep: torch.Tensor,
+                                    conds_final: list[torch.Tensor], counts_final: list[torch.Tensor], biases_final: list[torch.Tensor]):
+        if self.fuse_method.name == ContextFuseMethods.RELATIVE:
+            for pos, idx in enumerate(window.index_list):
+                # bias is the influence of a specific index in relation to the whole context window
+                bias = 1 - abs(idx - (window.index_list[0] + window.index_list[-1]) / 2) / ((window.index_list[-1] - window.index_list[0] + 1e-2) / 2)
+                bias = max(1e-2, bias)
+                # take weighted average relative to total bias of current idx
+                for i in range(len(sub_conds_out)):
+                    bias_total = biases_final[i][idx]
+                    prev_weight = (bias_total / (bias_total + bias))
+                    new_weight = (bias / (bias_total + bias))
+                    # account for dims of tensors
+                    idx_window = [slice(None)] * self.dim + [idx]
+                    pos_window = [slice(None)] * self.dim + [pos]
+                    # apply new values
+                    conds_final[i][idx_window] = conds_final[i][idx_window] * prev_weight + sub_conds_out[i][pos_window] * new_weight
+                    biases_final[i][idx] = bias_total + bias
+        else:
+            # add conds and counts based on weights of fuse method
+            weights = get_context_weights(window.context_length, x_in.shape[self.dim], window.index_list, self, sigma=timestep)
+            weights_tensor = match_weights_to_dim(weights, x_in, self.dim, device=x_in.device)
+            for i in range(len(sub_conds_out)):
+                window.add_window(conds_final[i], sub_conds_out[i] * weights_tensor)
+                window.add_window(counts_final[i], weights_tensor)
+
+        for callback in comfy.patcher_extension.get_all_callbacks(IndexListCallbacks.COMBINE_CONTEXT_WINDOW_RESULTS, self.callbacks):
+            callback(self, x_in, sub_conds_out, sub_conds, window, window_idx, total_windows, timestep, conds_final, counts_final, biases_final)
+
+
+def _prepare_sampling_wrapper(executor, model, noise_shape: torch.Tensor, *args, **kwargs):
+    # limit noise_shape length to context_length for more accurate vram use estimation
+    model_options = kwargs.get("model_options", None)
+    if model_options is None:
+        raise Exception("model_options not found in prepare_sampling_wrapper; this should never happen, something went wrong.")
+    handler: IndexListContextHandler = model_options.get("context_handler", None)
+    if handler is not None:
+        noise_shape = list(noise_shape)
+        noise_shape[handler.dim] = min(noise_shape[handler.dim], handler.context_length)
+    return executor(model, noise_shape, *args, **kwargs)
+
+
+def create_prepare_sampling_wrapper(model: ModelPatcher):
+    model.add_wrapper_with_key(
+        comfy.patcher_extension.WrappersMP.PREPARE_SAMPLING,
+        "ContextWindows_prepare_sampling",
+        _prepare_sampling_wrapper
+    )
+
+
+def match_weights_to_dim(weights: list[float], x_in: torch.Tensor, dim: int, device=None) -> torch.Tensor:
+    total_dims = len(x_in.shape)
+    weights_tensor = torch.Tensor(weights).to(device=device)
+    for _ in range(dim):
+        weights_tensor = weights_tensor.unsqueeze(0)
+    for _ in range(total_dims - dim - 1):
+        weights_tensor = weights_tensor.unsqueeze(-1)
+    return weights_tensor
+
+def get_shape_for_dim(x_in: torch.Tensor, dim: int) -> list[int]:
+    total_dims = len(x_in.shape)
+    shape = []
+    for _ in range(dim):
+        shape.append(1)
+    shape.append(x_in.shape[dim])
+    for _ in range(total_dims - dim - 1):
+        shape.append(1)
+    return shape
+
+class ContextSchedules:
+    UNIFORM_LOOPED = "looped_uniform"
+    UNIFORM_STANDARD = "standard_uniform"
+    STATIC_STANDARD = "standard_static"
+    BATCHED = "batched"
+
+
+# from https://github.com/neggles/animatediff-cli/blob/main/src/animatediff/pipelines/context.py
+def create_windows_uniform_looped(num_frames: int, handler: IndexListContextHandler, model_options: dict[str]):
+    windows = []
+    if num_frames < handler.context_length:
+        windows.append(list(range(num_frames)))
+        return windows
+
+    context_stride = min(handler.context_stride, int(np.ceil(np.log2(num_frames / handler.context_length))) + 1)
+    # obtain uniform windows as normal, looping and all
+    for context_step in 1 << np.arange(context_stride):
+        pad = int(round(num_frames * ordered_halving(handler._step)))
+        for j in range(
+            int(ordered_halving(handler._step) * context_step) + pad,
+            num_frames + pad + (0 if handler.closed_loop else -handler.context_overlap),
+            (handler.context_length * context_step - handler.context_overlap),
+        ):
+            windows.append([e % num_frames for e in range(j, j + handler.context_length * context_step, context_step)])
+
+    return windows
+
+def create_windows_uniform_standard(num_frames: int, handler: IndexListContextHandler, model_options: dict[str]):
+    # unlike looped, uniform_straight does NOT allow windows that loop back to the beginning;
+    # instead, they get shifted to the corresponding end of the frames.
+    # in the case that a window (shifted or not) is identical to the previous one, it gets skipped.
+    windows = []
+    if num_frames <= handler.context_length:
+        windows.append(list(range(num_frames)))
+        return windows
+
+    context_stride = min(handler.context_stride, int(np.ceil(np.log2(num_frames / handler.context_length))) + 1)
+    # first, obtain uniform windows as normal, looping and all
+    for context_step in 1 << np.arange(context_stride):
+        pad = int(round(num_frames * ordered_halving(handler._step)))
+        for j in range(
+            int(ordered_halving(handler._step) * context_step) + pad,
+            num_frames + pad + (-handler.context_overlap),
+            (handler.context_length * context_step - handler.context_overlap),
+        ):
+            windows.append([e % num_frames for e in range(j, j + handler.context_length * context_step, context_step)])
+
+    # now that windows are created, shift any windows that loop, and delete duplicate windows
+    delete_idxs = []
+    win_i = 0
+    while win_i < len(windows):
+        # if window is rolls over itself, need to shift it
+        is_roll, roll_idx = does_window_roll_over(windows[win_i], num_frames)
+        if is_roll:
+            roll_val = windows[win_i][roll_idx]  # roll_val might not be 0 for windows of higher strides
+            shift_window_to_end(windows[win_i], num_frames=num_frames)
+            # check if next window (cyclical) is missing roll_val
+            if roll_val not in windows[(win_i+1) % len(windows)]:
+                # need to insert new window here - just insert window starting at roll_val
+                windows.insert(win_i+1, list(range(roll_val, roll_val + handler.context_length)))
+        # delete window if it's not unique
+        for pre_i in range(0, win_i):
+            if windows[win_i] == windows[pre_i]:
+                delete_idxs.append(win_i)
+                break
+        win_i += 1
+
+    # reverse delete_idxs so that they will be deleted in an order that doesn't break idx correlation
+    delete_idxs.reverse()
+    for i in delete_idxs:
+        windows.pop(i)
+
+    return windows
+
+
+def create_windows_static_standard(num_frames: int, handler: IndexListContextHandler, model_options: dict[str]):
+    windows = []
+    if num_frames <= handler.context_length:
+        windows.append(list(range(num_frames)))
+        return windows
+    # always return the same set of windows
+    delta = handler.context_length - handler.context_overlap
+    for start_idx in range(0, num_frames, delta):
+        # if past the end of frames, move start_idx back to allow same context_length
+        ending = start_idx + handler.context_length
+        if ending >= num_frames:
+            final_delta = ending - num_frames
+            final_start_idx = start_idx - final_delta
+            windows.append(list(range(final_start_idx, final_start_idx + handler.context_length)))
+            break
+        windows.append(list(range(start_idx, start_idx + handler.context_length)))
+    return windows
+
+
+def create_windows_batched(num_frames: int, handler: IndexListContextHandler, model_options: dict[str]):
+    windows = []
+    if num_frames <= handler.context_length:
+        windows.append(list(range(num_frames)))
+        return windows
+    # always return the same set of windows;
+    # no overlap, just cut up based on context_length;
+    # last window size will be different if num_frames % opts.context_length != 0
+    for start_idx in range(0, num_frames, handler.context_length):
+        windows.append(list(range(start_idx, min(start_idx + handler.context_length, num_frames))))
+    return windows
+
+
+def create_windows_default(num_frames: int, handler: IndexListContextHandler):
+    return [list(range(num_frames))]
+
+
+CONTEXT_MAPPING = {
+    ContextSchedules.UNIFORM_LOOPED: create_windows_uniform_looped,
+    ContextSchedules.UNIFORM_STANDARD: create_windows_uniform_standard,
+    ContextSchedules.STATIC_STANDARD: create_windows_static_standard,
+    ContextSchedules.BATCHED: create_windows_batched,
+}
+
+
+def get_matching_context_schedule(context_schedule: str) -> ContextSchedule:
+    func = CONTEXT_MAPPING.get(context_schedule, None)
+    if func is None:
+        raise ValueError(f"Unknown context_schedule '{context_schedule}'.")
+    return ContextSchedule(context_schedule, func)
+
+
+def get_context_weights(length: int, full_length: int, idxs: list[int], handler: IndexListContextHandler, sigma: torch.Tensor=None):
+    return handler.fuse_method.func(length, sigma=sigma, handler=handler, full_length=full_length, idxs=idxs)
+
+
+def create_weights_flat(length: int, **kwargs) -> list[float]:
+    # weight is the same for all
+    return [1.0] * length
+
+def create_weights_pyramid(length: int, **kwargs) -> list[float]:
+    # weight is based on the distance away from the edge of the context window;
+    # based on weighted average concept in FreeNoise paper
+    if length % 2 == 0:
+        max_weight = length // 2
+        weight_sequence = list(range(1, max_weight + 1, 1)) + list(range(max_weight, 0, -1))
+    else:
+        max_weight = (length + 1) // 2
+        weight_sequence = list(range(1, max_weight, 1)) + [max_weight] + list(range(max_weight - 1, 0, -1))
+    return weight_sequence
+
+def create_weights_overlap_linear(length: int, full_length: int, idxs: list[int], handler: IndexListContextHandler, **kwargs):
+    # based on code in Kijai's WanVideoWrapper: https://github.com/kijai/ComfyUI-WanVideoWrapper/blob/dbb2523b37e4ccdf45127e5ae33e31362f755c8e/nodes.py#L1302
+    # only expected overlap is given different weights
+    weights_torch = torch.ones((length))
+    # blend left-side on all except first window
+    if min(idxs) > 0:
+        ramp_up = torch.linspace(1e-37, 1, handler.context_overlap)
+        weights_torch[:handler.context_overlap] = ramp_up
+    # blend right-side on all except last window
+    if max(idxs) < full_length-1:
+        ramp_down = torch.linspace(1, 1e-37, handler.context_overlap)
+        weights_torch[-handler.context_overlap:] = ramp_down
+    return weights_torch
+
+class ContextFuseMethods:
+    FLAT = "flat"
+    PYRAMID = "pyramid"
+    RELATIVE = "relative"
+    OVERLAP_LINEAR = "overlap-linear"
+
+    LIST = [PYRAMID, FLAT, OVERLAP_LINEAR]
+    LIST_STATIC = [PYRAMID, RELATIVE, FLAT, OVERLAP_LINEAR]
+
+
+FUSE_MAPPING = {
+    ContextFuseMethods.FLAT: create_weights_flat,
+    ContextFuseMethods.PYRAMID: create_weights_pyramid,
+    ContextFuseMethods.RELATIVE: create_weights_pyramid,
+    ContextFuseMethods.OVERLAP_LINEAR: create_weights_overlap_linear,
+}
+
+def get_matching_fuse_method(fuse_method: str) -> ContextFuseMethod:
+    func = FUSE_MAPPING.get(fuse_method, None)
+    if func is None:
+        raise ValueError(f"Unknown fuse_method '{fuse_method}'.")
+    return ContextFuseMethod(fuse_method, func)
+
+# Returns fraction that has denominator that is a power of 2
+def ordered_halving(val):
+    # get binary value, padded with 0s for 64 bits
+    bin_str = f"{val:064b}"
+    # flip binary value, padding included
+    bin_flip = bin_str[::-1]
+    # convert binary to int
+    as_int = int(bin_flip, 2)
+    # divide by 1 << 64, equivalent to 2**64, or 18446744073709551616,
+    # or b10000000000000000000000000000000000000000000000000000000000000000 (1 with 64 zero's)
+    return as_int / (1 << 64)
+
+
+def get_missing_indexes(windows: list[list[int]], num_frames: int) -> list[int]:
+    all_indexes = list(range(num_frames))
+    for w in windows:
+        for val in w:
+            try:
+                all_indexes.remove(val)
+            except ValueError:
+                pass
+    return all_indexes
+
+
+def does_window_roll_over(window: list[int], num_frames: int) -> tuple[bool, int]:
+    prev_val = -1
+    for i, val in enumerate(window):
+        val = val % num_frames
+        if val < prev_val:
+            return True, i
+        prev_val = val
+    return False, -1
+
+
+def shift_window_to_start(window: list[int], num_frames: int):
+    start_val = window[0]
+    for i in range(len(window)):
+        # 1) subtract each element by start_val to move vals relative to the start of all frames
+        # 2) add num_frames and take modulus to get adjusted vals
+        window[i] = ((window[i] - start_val) + num_frames) % num_frames
+
+
+def shift_window_to_end(window: list[int], num_frames: int):
+    # 1) shift window to start
+    shift_window_to_start(window, num_frames)
+    end_val = window[-1]
+    end_delta = num_frames - end_val - 1
+    for i in range(len(window)):
+        # 2) add end_delta to each val to slide windows to end
+        window[i] = window[i] + end_delta

comfy/ldm/flux/model.py

@@ -224,19 +224,27 @@ class Flux(nn.Module):
         if ref_latents is not None:
             h = 0
             w = 0
+            index = 0
+            index_ref_method = kwargs.get("ref_latents_method", "offset") == "index"
             for ref in ref_latents:
-                h_offset = 0
-                w_offset = 0
-                if ref.shape[-2] + h > ref.shape[-1] + w:
-                    w_offset = w
+                if index_ref_method:
+                    index += 1
+                    h_offset = 0
+                    w_offset = 0
                 else:
-                    h_offset = h
+                    index = 1
+                    h_offset = 0
+                    w_offset = 0
+                    if ref.shape[-2] + h > ref.shape[-1] + w:
+                        w_offset = w
+                    else:
+                        h_offset = h
+                    h = max(h, ref.shape[-2] + h_offset)
+                    w = max(w, ref.shape[-1] + w_offset)
 
-                kontext, kontext_ids = self.process_img(ref, index=1, h_offset=h_offset, w_offset=w_offset)
+                kontext, kontext_ids = self.process_img(ref, index=index, h_offset=h_offset, w_offset=w_offset)
                 img = torch.cat([img, kontext], dim=1)
                 img_ids = torch.cat([img_ids, kontext_ids], dim=1)
-                h = max(h, ref.shape[-2] + h_offset)
-                w = max(w, ref.shape[-1] + w_offset)
 
         txt_ids = torch.zeros((bs, context.shape[1], 3), device=x.device, dtype=x.dtype)
         out = self.forward_orig(img, img_ids, context, txt_ids, timestep, y, guidance, control, transformer_options, attn_mask=kwargs.get("attention_mask", None))

comfy/ldm/hunyuan3d/vae.py

@@ -178,7 +178,7 @@ class FourierEmbedder(nn.Module):
 
 class CrossAttentionProcessor:
     def __call__(self, attn, q, k, v):
-        out = F.scaled_dot_product_attention(q, k, v)
+        out = comfy.ops.scaled_dot_product_attention(q, k, v)
         return out
 
 

comfy/ldm/modules/attention.py

@@ -448,7 +448,7 @@ def attention_pytorch(q, k, v, heads, mask=None, attn_precision=None, skip_resha
             mask = mask.unsqueeze(1)
 
     if SDP_BATCH_LIMIT >= b:
-        out = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=mask, dropout_p=0.0, is_causal=False)
+        out = comfy.ops.scaled_dot_product_attention(q, k, v, attn_mask=mask, dropout_p=0.0, is_causal=False)
         if not skip_output_reshape:
             out = (
                 out.transpose(1, 2).reshape(b, -1, heads * dim_head)
@@ -461,7 +461,7 @@ def attention_pytorch(q, k, v, heads, mask=None, attn_precision=None, skip_resha
                 if mask.shape[0] > 1:
                     m = mask[i : i + SDP_BATCH_LIMIT]
 
-            out[i : i + SDP_BATCH_LIMIT] = torch.nn.functional.scaled_dot_product_attention(
+            out[i : i + SDP_BATCH_LIMIT] = comfy.ops.scaled_dot_product_attention(
                 q[i : i + SDP_BATCH_LIMIT],
                 k[i : i + SDP_BATCH_LIMIT],
                 v[i : i + SDP_BATCH_LIMIT],

comfy/ldm/modules/diffusionmodules/model.py

@@ -285,7 +285,7 @@ def pytorch_attention(q, k, v):
     )
 
     try:
-        out = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=None, dropout_p=0.0, is_causal=False)
+        out = comfy.ops.scaled_dot_product_attention(q, k, v, attn_mask=None, dropout_p=0.0, is_causal=False)
         out = out.transpose(2, 3).reshape(orig_shape)
     except model_management.OOM_EXCEPTION:
         logging.warning("scaled_dot_product_attention OOMed: switched to slice attention")

comfy/ldm/qwen_image/model.py

@@ -333,21 +333,25 @@ class QwenImageTransformer2DModel(nn.Module):
         self.proj_out = operations.Linear(self.inner_dim, patch_size * patch_size * self.out_channels, bias=True, dtype=dtype, device=device)
         self.gradient_checkpointing = False
 
-    def pos_embeds(self, x, context):
+    def process_img(self, x, index=0, h_offset=0, w_offset=0):
         bs, c, t, h, w = x.shape
         patch_size = self.patch_size
+        hidden_states = comfy.ldm.common_dit.pad_to_patch_size(x, (1, self.patch_size, self.patch_size))
+        orig_shape = hidden_states.shape
+        hidden_states = hidden_states.view(orig_shape[0], orig_shape[1], orig_shape[-2] // 2, 2, orig_shape[-1] // 2, 2)
+        hidden_states = hidden_states.permute(0, 2, 4, 1, 3, 5)
+        hidden_states = hidden_states.reshape(orig_shape[0], (orig_shape[-2] // 2) * (orig_shape[-1] // 2), orig_shape[1] * 4)
         h_len = ((h + (patch_size // 2)) // patch_size)
         w_len = ((w + (patch_size // 2)) // patch_size)
 
-        img_ids = torch.zeros((h_len, w_len, 3), device=x.device, dtype=x.dtype)
-        img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(0, h_len - 1, steps=h_len, device=x.device, dtype=x.dtype).unsqueeze(1)
-        img_ids[:, :, 2] = img_ids[:, :, 2] + torch.linspace(0, w_len - 1, steps=w_len, device=x.device, dtype=x.dtype).unsqueeze(0)
-        img_ids = repeat(img_ids, "h w c -> b (h w) c", b=bs)
+        h_offset = ((h_offset + (patch_size // 2)) // patch_size)
+        w_offset = ((w_offset + (patch_size // 2)) // patch_size)
 
-        txt_start = round(max(h_len, w_len))
-        txt_ids = torch.linspace(txt_start, txt_start + context.shape[1], steps=context.shape[1], device=x.device, dtype=x.dtype).reshape(1, -1, 1).repeat(bs, 1, 3)
-        ids = torch.cat((txt_ids, img_ids), dim=1)
-        return self.pe_embedder(ids).squeeze(1).unsqueeze(2).to(x.dtype)
+        img_ids = torch.zeros((h_len, w_len, 3), device=x.device)
+        img_ids[:, :, 0] = img_ids[:, :, 1] + index
+        img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(h_offset, h_len - 1 + h_offset, steps=h_len, device=x.device, dtype=x.dtype).unsqueeze(1) - (h_len // 2)
+        img_ids[:, :, 2] = img_ids[:, :, 2] + torch.linspace(w_offset, w_len - 1 + w_offset, steps=w_len, device=x.device, dtype=x.dtype).unsqueeze(0) - (w_len // 2)
+        return hidden_states, repeat(img_ids, "h w c -> b (h w) c", b=bs), orig_shape
 
     def forward(
         self,
@@ -356,19 +360,47 @@ class QwenImageTransformer2DModel(nn.Module):
         context,
         attention_mask=None,
         guidance: torch.Tensor = None,
+        ref_latents=None,
+        transformer_options={},
         **kwargs
     ):
         timestep = timesteps
         encoder_hidden_states = context
         encoder_hidden_states_mask = attention_mask
 
-        image_rotary_emb = self.pos_embeds(x, context)
+        hidden_states, img_ids, orig_shape = self.process_img(x)
+        num_embeds = hidden_states.shape[1]
 
-        hidden_states = comfy.ldm.common_dit.pad_to_patch_size(x, (1, self.patch_size, self.patch_size))
-        orig_shape = hidden_states.shape
-        hidden_states = hidden_states.view(orig_shape[0], orig_shape[1], orig_shape[-2] // 2, 2, orig_shape[-1] // 2, 2)
-        hidden_states = hidden_states.permute(0, 2, 4, 1, 3, 5)
-        hidden_states = hidden_states.reshape(orig_shape[0], (orig_shape[-2] // 2) * (orig_shape[-1] // 2), orig_shape[1] * 4)
+        if ref_latents is not None:
+            h = 0
+            w = 0
+            index = 0
+            index_ref_method = kwargs.get("ref_latents_method", "index") == "index"
+            for ref in ref_latents:
+                if index_ref_method:
+                    index += 1
+                    h_offset = 0
+                    w_offset = 0
+                else:
+                    index = 1
+                    h_offset = 0
+                    w_offset = 0
+                    if ref.shape[-2] + h > ref.shape[-1] + w:
+                        w_offset = w
+                    else:
+                        h_offset = h
+                    h = max(h, ref.shape[-2] + h_offset)
+                    w = max(w, ref.shape[-1] + w_offset)
+
+                kontext, kontext_ids, _ = self.process_img(ref, index=index, h_offset=h_offset, w_offset=w_offset)
+                hidden_states = torch.cat([hidden_states, kontext], dim=1)
+                img_ids = torch.cat([img_ids, kontext_ids], dim=1)
+
+        txt_start = round(max(((x.shape[-1] + (self.patch_size // 2)) // self.patch_size) // 2, ((x.shape[-2] + (self.patch_size // 2)) // self.patch_size) // 2))
+        txt_ids = torch.arange(txt_start, txt_start + context.shape[1], device=x.device).reshape(1, -1, 1).repeat(x.shape[0], 1, 3)
+        ids = torch.cat((txt_ids, img_ids), dim=1)
+        image_rotary_emb = self.pe_embedder(ids).squeeze(1).unsqueeze(2).to(x.dtype)
+        del ids, txt_ids, img_ids
 
         hidden_states = self.img_in(hidden_states)
         encoder_hidden_states = self.txt_norm(encoder_hidden_states)
@@ -383,18 +415,30 @@ class QwenImageTransformer2DModel(nn.Module):
             else self.time_text_embed(timestep, guidance, hidden_states)
         )
 
-        for block in self.transformer_blocks:
-            encoder_hidden_states, hidden_states = block(
-                hidden_states=hidden_states,
-                encoder_hidden_states=encoder_hidden_states,
-                encoder_hidden_states_mask=encoder_hidden_states_mask,
-                temb=temb,
-                image_rotary_emb=image_rotary_emb,
-            )
+        patches_replace = transformer_options.get("patches_replace", {})
+        blocks_replace = patches_replace.get("dit", {})
+
+        for i, block in enumerate(self.transformer_blocks):
+            if ("double_block", i) in blocks_replace:
+                def block_wrap(args):
+                    out = {}
+                    out["txt"], out["img"] = block(hidden_states=args["img"], encoder_hidden_states=args["txt"], encoder_hidden_states_mask=encoder_hidden_states_mask, temb=args["vec"], image_rotary_emb=args["pe"])
+                    return out
+                out = blocks_replace[("double_block", i)]({"img": hidden_states, "txt": encoder_hidden_states, "vec": temb, "pe": image_rotary_emb}, {"original_block": block_wrap})
+                hidden_states = out["img"]
+                encoder_hidden_states = out["txt"]
+            else:
+                encoder_hidden_states, hidden_states = block(
+                    hidden_states=hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    encoder_hidden_states_mask=encoder_hidden_states_mask,
+                    temb=temb,
+                    image_rotary_emb=image_rotary_emb,
+                )
 
         hidden_states = self.norm_out(hidden_states, temb)
         hidden_states = self.proj_out(hidden_states)
 
-        hidden_states = hidden_states.view(orig_shape[0], orig_shape[-2] // 2, orig_shape[-1] // 2, orig_shape[1], 2, 2)
+        hidden_states = hidden_states[:, :num_embeds].view(orig_shape[0], orig_shape[-2] // 2, orig_shape[-1] // 2, orig_shape[1], 2, 2)
         hidden_states = hidden_states.permute(0, 3, 1, 4, 2, 5)
         return hidden_states.reshape(orig_shape)[:, :, :, :x.shape[-2], :x.shape[-1]]

comfy/ldm/wan/model.py

@@ -768,7 +768,12 @@ class CameraWanModel(WanModel):
                  operations=None,
                  ):
 
-        super().__init__(model_type='i2v', patch_size=patch_size, text_len=text_len, in_dim=in_dim, dim=dim, ffn_dim=ffn_dim, freq_dim=freq_dim, text_dim=text_dim, out_dim=out_dim, num_heads=num_heads, num_layers=num_layers, window_size=window_size, qk_norm=qk_norm, cross_attn_norm=cross_attn_norm, eps=eps, flf_pos_embed_token_number=flf_pos_embed_token_number, image_model=image_model, device=device, dtype=dtype, operations=operations)
+        if model_type == 'camera':
+            model_type = 'i2v'
+        else:
+            model_type = 't2v'
+
+        super().__init__(model_type=model_type, patch_size=patch_size, text_len=text_len, in_dim=in_dim, dim=dim, ffn_dim=ffn_dim, freq_dim=freq_dim, text_dim=text_dim, out_dim=out_dim, num_heads=num_heads, num_layers=num_layers, window_size=window_size, qk_norm=qk_norm, cross_attn_norm=cross_attn_norm, eps=eps, flf_pos_embed_token_number=flf_pos_embed_token_number, image_model=image_model, device=device, dtype=dtype, operations=operations)
         operation_settings = {"operations": operations, "device": device, "dtype": dtype}
 
         self.control_adapter = WanCamAdapter(in_dim_control_adapter, dim, kernel_size=patch_size[1:], stride=patch_size[1:], operation_settings=operation_settings)

comfy/model_base.py

@@ -890,6 +890,10 @@ class Flux(BaseModel):
             for lat in ref_latents:
                 latents.append(self.process_latent_in(lat))
             out['ref_latents'] = comfy.conds.CONDList(latents)
+
+            ref_latents_method = kwargs.get("reference_latents_method", None)
+            if ref_latents_method is not None:
+                out['ref_latents_method'] = comfy.conds.CONDConstant(ref_latents_method)
         return out
 
     def extra_conds_shapes(self, **kwargs):
@@ -1321,10 +1325,28 @@ class Omnigen2(BaseModel):
 class QwenImage(BaseModel):
     def __init__(self, model_config, model_type=ModelType.FLUX, device=None):
         super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.qwen_image.model.QwenImageTransformer2DModel)
+        self.memory_usage_factor_conds = ("ref_latents",)
 
     def extra_conds(self, **kwargs):
         out = super().extra_conds(**kwargs)
         cross_attn = kwargs.get("cross_attn", None)
         if cross_attn is not None:
             out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+        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_latents_method = kwargs.get("reference_latents_method", None)
+            if ref_latents_method is not None:
+                out['ref_latents_method'] = comfy.conds.CONDConstant(ref_latents_method)
+        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()), ref_latents)) // 16])
         return out

comfy/model_detection.py

@@ -364,7 +364,10 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
             dit_config["vace_in_dim"] = state_dict['{}vace_patch_embedding.weight'.format(key_prefix)].shape[1]
             dit_config["vace_layers"] = count_blocks(state_dict_keys, '{}vace_blocks.'.format(key_prefix) + '{}.')
         elif '{}control_adapter.conv.weight'.format(key_prefix) in state_dict_keys:
-            dit_config["model_type"] = "camera"
+            if '{}img_emb.proj.0.bias'.format(key_prefix) in state_dict_keys:
+                dit_config["model_type"] = "camera"
+            else:
+                dit_config["model_type"] = "camera_2.2"
         else:
             if '{}img_emb.proj.0.bias'.format(key_prefix) in state_dict_keys:
                 dit_config["model_type"] = "i2v"

comfy/model_management.py

@@ -78,7 +78,6 @@ try:
     torch_version = torch.version.__version__
     temp = torch_version.split(".")
     torch_version_numeric = (int(temp[0]), int(temp[1]))
-    xpu_available = (torch_version_numeric[0] < 2 or (torch_version_numeric[0] == 2 and torch_version_numeric[1] <= 4)) and torch.xpu.is_available()
 except:
     pass
 
@@ -102,10 +101,14 @@ if args.directml is not None:
 
 try:
     import intel_extension_for_pytorch as ipex  # noqa: F401
-    _ = torch.xpu.device_count()
-    xpu_available = xpu_available or torch.xpu.is_available()
 except:
-    xpu_available = xpu_available or (hasattr(torch, "xpu") and torch.xpu.is_available())
+    pass
+
+try:
+    _ = torch.xpu.device_count()
+    xpu_available = torch.xpu.is_available()
+except:
+    xpu_available = False
 
 try:
     if torch.backends.mps.is_available():
@@ -946,10 +949,12 @@ def pick_weight_dtype(dtype, fallback_dtype, device=None):
     return dtype
 
 def device_supports_non_blocking(device):
+    if args.force_non_blocking:
+        return True
     if is_device_mps(device):
         return False #pytorch bug? mps doesn't support non blocking
-    if is_intel_xpu():
-        return True
+    if is_intel_xpu(): #xpu does support non blocking but it is slower on iGPUs for some reason so disable by default until situation changes
+        return False
     if args.deterministic: #TODO: figure out why deterministic breaks non blocking from gpu to cpu (previews)
         return False
     if directml_enabled:
@@ -1282,10 +1287,10 @@ def should_use_bf16(device=None, model_params=0, prioritize_performance=True, ma
         return False
 
     if is_intel_xpu():
-        if torch_version_numeric < (2, 6):
+        if torch_version_numeric < (2, 3):
             return True
         else:
-            return torch.xpu.get_device_capability(device)['has_bfloat16_conversions']
+            return torch.xpu.is_bf16_supported()
 
     if is_ascend_npu():
         return True

comfy/ops.py

@@ -24,6 +24,32 @@ import comfy.float
 import comfy.rmsnorm
 import contextlib
 
+
+def scaled_dot_product_attention(q, k, v, *args, **kwargs):
+    return torch.nn.functional.scaled_dot_product_attention(q, k, v, *args, **kwargs)
+
+
+try:
+    if torch.cuda.is_available():
+        from torch.nn.attention import SDPBackend, sdpa_kernel
+        import inspect
+        if "set_priority" in inspect.signature(sdpa_kernel).parameters:
+            SDPA_BACKEND_PRIORITY = [
+                SDPBackend.FLASH_ATTENTION,
+                SDPBackend.EFFICIENT_ATTENTION,
+                SDPBackend.MATH,
+            ]
+
+            SDPA_BACKEND_PRIORITY.insert(0, SDPBackend.CUDNN_ATTENTION)
+
+            def scaled_dot_product_attention(q, k, v, *args, **kwargs):
+                with sdpa_kernel(SDPA_BACKEND_PRIORITY, set_priority=True):
+                    return torch.nn.functional.scaled_dot_product_attention(q, k, v, *args, **kwargs)
+        else:
+            logging.warning("Torch version too old to set sdpa backend priority.")
+except (ModuleNotFoundError, TypeError):
+    logging.warning("Could not set sdpa backend priority.")
+
 cast_to = comfy.model_management.cast_to #TODO: remove once no more references
 
 def cast_to_input(weight, input, non_blocking=False, copy=True):

comfy/rmsnorm.py

@@ -1,6 +1,7 @@
 import torch
 import comfy.model_management
 import numbers
+import logging
 
 RMSNorm = None
 
@@ -9,6 +10,7 @@ try:
     RMSNorm = torch.nn.RMSNorm
 except:
     rms_norm_torch = None
+    logging.warning("Please update pytorch to use native RMSNorm")
 
 
 def rms_norm(x, weight=None, eps=1e-6):

comfy/sampler_helpers.py

@@ -149,7 +149,7 @@ def cleanup_models(conds, models):
 
     cleanup_additional_models(set(control_cleanup))
 
-def prepare_model_patcher(model: 'ModelPatcher', conds, model_options: dict):
+def prepare_model_patcher(model: ModelPatcher, conds, model_options: dict):
     '''
     Registers hooks from conds.
     '''
@@ -158,8 +158,8 @@ def prepare_model_patcher(model: 'ModelPatcher', conds, model_options: dict):
     for k in conds:
         get_hooks_from_cond(conds[k], hooks)
     # add wrappers and callbacks from ModelPatcher to transformer_options
-    model_options["transformer_options"]["wrappers"] = comfy.patcher_extension.copy_nested_dicts(model.wrappers)
-    model_options["transformer_options"]["callbacks"] = comfy.patcher_extension.copy_nested_dicts(model.callbacks)
+    comfy.patcher_extension.merge_nested_dicts(model_options["transformer_options"].setdefault("wrappers", {}), model.wrappers, copy_dict1=False)
+    comfy.patcher_extension.merge_nested_dicts(model_options["transformer_options"].setdefault("callbacks", {}), model.callbacks, copy_dict1=False)
     # begin registering hooks
     registered = comfy.hooks.HookGroup()
     target_dict = comfy.hooks.create_target_dict(comfy.hooks.EnumWeightTarget.Model)

comfy/samplers.py

@@ -16,6 +16,7 @@ import comfy.sampler_helpers
 import comfy.model_patcher
 import comfy.patcher_extension
 import comfy.hooks
+import comfy.context_windows
 import scipy.stats
 import numpy
 
@@ -198,14 +199,20 @@ def finalize_default_conds(model: 'BaseModel', hooked_to_run: dict[comfy.hooks.H
             hooked_to_run.setdefault(p.hooks, list())
             hooked_to_run[p.hooks] += [(p, i)]
 
-def calc_cond_batch(model: 'BaseModel', conds: list[list[dict]], x_in: torch.Tensor, timestep, model_options):
+def calc_cond_batch(model: BaseModel, conds: list[list[dict]], x_in: torch.Tensor, timestep, model_options: dict[str]):
+    handler: comfy.context_windows.ContextHandlerABC = model_options.get("context_handler", None)
+    if handler is None or not handler.should_use_context(model, conds, x_in, timestep, model_options):
+        return _calc_cond_batch_outer(model, conds, x_in, timestep, model_options)
+    return handler.execute(_calc_cond_batch_outer, model, conds, x_in, timestep, model_options)
+
+def _calc_cond_batch_outer(model: BaseModel, conds: list[list[dict]], x_in: torch.Tensor, timestep, model_options):
     executor = comfy.patcher_extension.WrapperExecutor.new_executor(
         _calc_cond_batch,
         comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.CALC_COND_BATCH, model_options, is_model_options=True)
     )
     return executor.execute(model, conds, x_in, timestep, model_options)
 
-def _calc_cond_batch(model: 'BaseModel', conds: list[list[dict]], x_in: torch.Tensor, timestep, model_options):
+def _calc_cond_batch(model: BaseModel, conds: list[list[dict]], x_in: torch.Tensor, timestep, model_options):
     out_conds = []
     out_counts = []
     # separate conds by matching hooks

comfy/sd1_clip.py

@@ -204,17 +204,19 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
             tokens_embed = self.transformer.get_input_embeddings()(tokens_embed, out_dtype=torch.float32)
             index = 0
             pad_extra = 0
+            embeds_info = []
             for o in other_embeds:
                 emb = o[1]
                 if torch.is_tensor(emb):
                     emb = {"type": "embedding", "data": emb}
 
+                extra = None
                 emb_type = emb.get("type", None)
                 if emb_type == "embedding":
                     emb = emb.get("data", None)
                 else:
                     if hasattr(self.transformer, "preprocess_embed"):
-                        emb = self.transformer.preprocess_embed(emb, device=device)
+                        emb, extra = self.transformer.preprocess_embed(emb, device=device)
                     else:
                         emb = None
 
@@ -229,6 +231,7 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
                     tokens_embed = torch.cat([tokens_embed[:, :ind], emb, tokens_embed[:, ind:]], dim=1)
                     attention_mask = attention_mask[:ind] + [1] * emb_shape + attention_mask[ind:]
                     index += emb_shape - 1
+                    embeds_info.append({"type": emb_type, "index": ind, "size": emb_shape, "extra": extra})
                 else:
                     index += -1
                     pad_extra += emb_shape
@@ -243,11 +246,11 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
             attention_masks.append(attention_mask)
             num_tokens.append(sum(attention_mask))
 
-        return torch.cat(embeds_out), torch.tensor(attention_masks, device=device, dtype=torch.long), num_tokens
+        return torch.cat(embeds_out), torch.tensor(attention_masks, device=device, dtype=torch.long), num_tokens, embeds_info
 
     def forward(self, tokens):
         device = self.transformer.get_input_embeddings().weight.device
-        embeds, attention_mask, num_tokens = self.process_tokens(tokens, device)
+        embeds, attention_mask, num_tokens, embeds_info = self.process_tokens(tokens, device)
 
         attention_mask_model = None
         if self.enable_attention_masks:
@@ -258,7 +261,7 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
         else:
             intermediate_output = self.layer_idx
 
-        outputs = self.transformer(None, attention_mask_model, embeds=embeds, num_tokens=num_tokens, intermediate_output=intermediate_output, final_layer_norm_intermediate=self.layer_norm_hidden_state, dtype=torch.float32)
+        outputs = self.transformer(None, attention_mask_model, embeds=embeds, num_tokens=num_tokens, intermediate_output=intermediate_output, final_layer_norm_intermediate=self.layer_norm_hidden_state, dtype=torch.float32, embeds_info=embeds_info)
 
         if self.layer == "last":
             z = outputs[0].float()
@@ -531,7 +534,10 @@ class SDTokenizer:
         min_padding = tokenizer_options.get("{}_min_padding".format(self.embedding_key), self.min_padding)
 
         text = escape_important(text)
-        parsed_weights = token_weights(text, 1.0)
+        if kwargs.get("disable_weights", False):
+            parsed_weights = [(text, 1.0)]
+        else:
+            parsed_weights = token_weights(text, 1.0)
 
         # tokenize words
         tokens = []

comfy/supported_models.py

@@ -1046,6 +1046,18 @@ class WAN21_Camera(WAN21_T2V):
     def get_model(self, state_dict, prefix="", device=None):
         out = model_base.WAN21_Camera(self, image_to_video=False, device=device)
         return out
+
+class WAN22_Camera(WAN21_T2V):
+    unet_config = {
+        "image_model": "wan2.1",
+        "model_type": "camera_2.2",
+        "in_dim": 36,
+    }
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.WAN21_Camera(self, image_to_video=False, device=device)
+        return out
+
 class WAN21_Vace(WAN21_T2V):
     unet_config = {
         "image_model": "wan2.1",
@@ -1260,6 +1272,6 @@ class QwenImage(supported_models_base.BASE):
         return supported_models_base.ClipTarget(comfy.text_encoders.qwen_image.QwenImageTokenizer, comfy.text_encoders.qwen_image.te(**hunyuan_detect))
 
 
-models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, Hunyuan3Dv2mini, Hunyuan3Dv2, HiDream, Chroma, ACEStep, Omnigen2, QwenImage]
+models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, Hunyuan3Dv2mini, Hunyuan3Dv2, HiDream, Chroma, ACEStep, Omnigen2, QwenImage]
 
 models += [SVD_img2vid]

comfy/text_encoders/bert.py

@@ -116,7 +116,7 @@ class BertModel_(torch.nn.Module):
         self.embeddings = BertEmbeddings(config_dict["vocab_size"], config_dict["max_position_embeddings"], config_dict["type_vocab_size"], config_dict["pad_token_id"], embed_dim, layer_norm_eps, dtype, device, operations)
         self.encoder = BertEncoder(config_dict["num_hidden_layers"], embed_dim, config_dict["intermediate_size"], config_dict["num_attention_heads"], layer_norm_eps, dtype, device, operations)
 
-    def forward(self, input_tokens, attention_mask=None, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None):
+    def forward(self, input_tokens, attention_mask=None, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None, embeds_info=[]):
         x = self.embeddings(input_tokens, embeds=embeds, dtype=dtype)
         mask = None
         if attention_mask is not None:

comfy/text_encoders/llama.py

@@ -2,12 +2,14 @@ import torch
 import torch.nn as nn
 from dataclasses import dataclass
 from typing import Optional, Any
+import math
 
 from comfy.ldm.modules.attention import optimized_attention_for_device
 import comfy.model_management
 import comfy.ldm.common_dit
 
 import comfy.model_management
+from . import qwen_vl
 
 @dataclass
 class Llama2Config:
@@ -25,6 +27,7 @@ class Llama2Config:
     rms_norm_add = False
     mlp_activation = "silu"
     qkv_bias = False
+    rope_dims = None
 
 @dataclass
 class Qwen25_3BConfig:
@@ -42,6 +45,7 @@ class Qwen25_3BConfig:
     rms_norm_add = False
     mlp_activation = "silu"
     qkv_bias = True
+    rope_dims = None
 
 @dataclass
 class Qwen25_7BVLI_Config:
@@ -59,6 +63,7 @@ class Qwen25_7BVLI_Config:
     rms_norm_add = False
     mlp_activation = "silu"
     qkv_bias = True
+    rope_dims = [16, 24, 24]
 
 @dataclass
 class Gemma2_2B_Config:
@@ -76,6 +81,7 @@ class Gemma2_2B_Config:
     rms_norm_add = True
     mlp_activation = "gelu_pytorch_tanh"
     qkv_bias = False
+    rope_dims = None
 
 class RMSNorm(nn.Module):
     def __init__(self, dim: int, eps: float = 1e-5, add=False, device=None, dtype=None):
@@ -100,24 +106,30 @@ def rotate_half(x):
     return torch.cat((-x2, x1), dim=-1)
 
 
-def precompute_freqs_cis(head_dim, seq_len, theta, device=None):
+def precompute_freqs_cis(head_dim, position_ids, theta, rope_dims=None, device=None):
     theta_numerator = torch.arange(0, head_dim, 2, device=device).float()
     inv_freq = 1.0 / (theta ** (theta_numerator / head_dim))
 
-    position_ids = torch.arange(0, seq_len, device=device).unsqueeze(0)
-
     inv_freq_expanded = inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
     position_ids_expanded = position_ids[:, None, :].float()
     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()
+    if rope_dims is not None and position_ids.shape[0] > 1:
+        mrope_section = rope_dims * 2
+        cos = torch.cat([m[i % 3] for i, m in enumerate(cos.split(mrope_section, dim=-1))], dim=-1).unsqueeze(0)
+        sin = torch.cat([m[i % 3] for i, m in enumerate(sin.split(mrope_section, dim=-1))], dim=-1).unsqueeze(0)
+    else:
+        cos = cos.unsqueeze(1)
+        sin = sin.unsqueeze(1)
+
     return (cos, sin)
 
 
 def apply_rope(xq, xk, freqs_cis):
-    cos = freqs_cis[0].unsqueeze(1)
-    sin = freqs_cis[1].unsqueeze(1)
+    cos = freqs_cis[0]
+    sin = freqs_cis[1]
     q_embed = (xq * cos) + (rotate_half(xq) * sin)
     k_embed = (xk * cos) + (rotate_half(xk) * sin)
     return q_embed, k_embed
@@ -277,7 +289,7 @@ class Llama2_(nn.Module):
         self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps, add=config.rms_norm_add, device=device, dtype=dtype)
         # self.lm_head = ops.Linear(config.hidden_size, config.vocab_size, bias=False, device=device, dtype=dtype)
 
-    def forward(self, x, attention_mask=None, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None):
+    def forward(self, x, attention_mask=None, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None, position_ids=None, embeds_info=[]):
         if embeds is not None:
             x = embeds
         else:
@@ -286,9 +298,13 @@ class Llama2_(nn.Module):
         if self.normalize_in:
             x *= self.config.hidden_size ** 0.5
 
+        if position_ids is None:
+            position_ids = torch.arange(0, x.shape[1], device=x.device).unsqueeze(0)
+
         freqs_cis = precompute_freqs_cis(self.config.head_dim,
-                                         x.shape[1],
+                                         position_ids,
                                          self.config.rope_theta,
+                                         self.config.rope_dims,
                                          device=x.device)
 
         mask = None
@@ -372,8 +388,38 @@ class Qwen25_7BVLI(BaseLlama, torch.nn.Module):
         self.num_layers = config.num_hidden_layers
 
         self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
+        self.visual = qwen_vl.Qwen2VLVisionTransformer(hidden_size=1280, output_hidden_size=config.hidden_size, device=device, dtype=dtype, ops=operations)
         self.dtype = dtype
 
+    def preprocess_embed(self, embed, device):
+        if embed["type"] == "image":
+            image, grid = qwen_vl.process_qwen2vl_images(embed["data"])
+            return self.visual(image.to(device, dtype=torch.float32), grid), grid
+        return None, None
+
+    def forward(self, x, attention_mask=None, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None, embeds_info=[]):
+        grid = None
+        for e in embeds_info:
+            if e.get("type") == "image":
+                grid = e.get("extra", None)
+                position_ids = torch.zeros((3, embeds.shape[1]), device=embeds.device)
+                start = e.get("index")
+                position_ids[:, :start] = torch.arange(0, start, device=embeds.device)
+                end = e.get("size") + start
+                len_max = int(grid.max()) // 2
+                start_next = len_max + start
+                position_ids[:, end:] = torch.arange(start_next, start_next + (embeds.shape[1] - end), device=embeds.device)
+                position_ids[0, start:end] = start
+                max_d = int(grid[0][1]) // 2
+                position_ids[1, start:end] = torch.arange(start, start + max_d, device=embeds.device).unsqueeze(1).repeat(1, math.ceil((end - start) / max_d)).flatten(0)[:end - start]
+                max_d = int(grid[0][2]) // 2
+                position_ids[2, start:end] = torch.arange(start, start + max_d, device=embeds.device).unsqueeze(0).repeat(math.ceil((end - start) / max_d), 1).flatten(0)[:end - start]
+
+        if grid is None:
+            position_ids = None
+
+        return super().forward(x, attention_mask=attention_mask, embeds=embeds, num_tokens=num_tokens, intermediate_output=intermediate_output, final_layer_norm_intermediate=final_layer_norm_intermediate, dtype=dtype, position_ids=position_ids)
+
 class Gemma2_2B(BaseLlama, torch.nn.Module):
     def __init__(self, config_dict, dtype, device, operations):
         super().__init__()

comfy/text_encoders/qwen_image.py

@@ -15,13 +15,27 @@ class QwenImageTokenizer(sd1_clip.SD1Tokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, name="qwen25_7b", tokenizer=Qwen25_7BVLITokenizer)
         self.llama_template = "<|im_start|>system\nDescribe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background:<|im_end|>\n<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n"
+        self.llama_template_images = "<|im_start|>system\nDescribe the key features of the input image (color, shape, size, texture, objects, background), then explain how the user's text instruction should alter or modify the image. Generate a new image that meets the user's requirements while maintaining consistency with the original input where appropriate.<|im_end|>\n<|im_start|>user\n<|vision_start|><|image_pad|><|vision_end|>{}<|im_end|>\n<|im_start|>assistant\n"
 
-    def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None,**kwargs):
+    def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None, images=[], **kwargs):
         if llama_template is None:
-            llama_text = self.llama_template.format(text)
+            if len(images) > 0:
+                llama_text = self.llama_template_images.format(text)
+            else:
+                llama_text = self.llama_template.format(text)
         else:
             llama_text = llama_template.format(text)
-        return super().tokenize_with_weights(llama_text, return_word_ids=return_word_ids, **kwargs)
+        tokens = super().tokenize_with_weights(llama_text, return_word_ids=return_word_ids, disable_weights=True, **kwargs)
+        key_name = next(iter(tokens))
+        embed_count = 0
+        qwen_tokens = tokens[key_name]
+        for r in qwen_tokens:
+            for i in range(len(r)):
+                if r[i][0] == 151655:
+                    if len(images) > embed_count:
+                        r[i] = ({"type": "image", "data": images[embed_count], "original_type": "image"},) + r[i][1:]
+                        embed_count += 1
+        return tokens
 
 
 class Qwen25_7BVLIModel(sd1_clip.SDClipModel):

comfy/text_encoders/qwen_vl.py

@@ -0,0 +1,428 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from typing import Optional, Tuple
+import math
+from comfy.ldm.modules.attention import optimized_attention_for_device
+
+
+def process_qwen2vl_images(
+    images: torch.Tensor,
+    min_pixels: int = 3136,
+    max_pixels: int = 12845056,
+    patch_size: int = 14,
+    temporal_patch_size: int = 2,
+    merge_size: int = 2,
+    image_mean: list = None,
+    image_std: list = None,
+):
+    if image_mean is None:
+        image_mean = [0.48145466, 0.4578275, 0.40821073]
+    if image_std is None:
+        image_std = [0.26862954, 0.26130258, 0.27577711]
+
+    batch_size, height, width, channels = images.shape
+    device = images.device
+    # dtype = images.dtype
+
+    images = images.permute(0, 3, 1, 2)
+
+    grid_thw_list = []
+    img = images[0]
+
+    factor = patch_size * merge_size
+
+    h_bar = round(height / factor) * factor
+    w_bar = round(width / factor) * factor
+
+    if h_bar * w_bar > max_pixels:
+        beta = math.sqrt((height * width) / max_pixels)
+        h_bar = max(factor, math.floor(height / beta / factor) * factor)
+        w_bar = max(factor, math.floor(width / beta / factor) * factor)
+    elif h_bar * w_bar < min_pixels:
+        beta = math.sqrt(min_pixels / (height * width))
+        h_bar = math.ceil(height * beta / factor) * factor
+        w_bar = math.ceil(width * beta / factor) * factor
+
+    img_resized = F.interpolate(
+        img.unsqueeze(0),
+        size=(h_bar, w_bar),
+        mode='bilinear',
+        align_corners=False
+    ).squeeze(0)
+
+    normalized = img_resized.clone()
+    for c in range(3):
+        normalized[c] = (img_resized[c] - image_mean[c]) / image_std[c]
+
+    grid_h = h_bar // patch_size
+    grid_w = w_bar // patch_size
+    grid_thw = torch.tensor([1, grid_h, grid_w], device=device, dtype=torch.long)
+
+    pixel_values = normalized
+    grid_thw_list.append(grid_thw)
+    image_grid_thw = torch.stack(grid_thw_list)
+
+    grid_t = 1
+    channel = pixel_values.shape[0]
+    pixel_values = pixel_values.unsqueeze(0).repeat(2, 1, 1, 1)
+
+    patches = pixel_values.reshape(
+        grid_t,
+        temporal_patch_size,
+        channel,
+        grid_h // merge_size,
+        merge_size,
+        patch_size,
+        grid_w // merge_size,
+        merge_size,
+        patch_size,
+    )
+
+    patches = patches.permute(0, 3, 6, 4, 7, 2, 1, 5, 8)
+    flatten_patches = patches.reshape(
+        grid_t * grid_h * grid_w,
+        channel * temporal_patch_size * patch_size * patch_size
+    )
+
+    return flatten_patches, image_grid_thw
+
+
+class VisionPatchEmbed(nn.Module):
+    def __init__(
+        self,
+        patch_size: int = 14,
+        temporal_patch_size: int = 2,
+        in_channels: int = 3,
+        embed_dim: int = 3584,
+        device=None,
+        dtype=None,
+        ops=None,
+    ):
+        super().__init__()
+        self.patch_size = patch_size
+        self.temporal_patch_size = temporal_patch_size
+        self.in_channels = in_channels
+        self.embed_dim = embed_dim
+
+        kernel_size = [temporal_patch_size, patch_size, patch_size]
+        self.proj = ops.Conv3d(
+            in_channels,
+            embed_dim,
+            kernel_size=kernel_size,
+            stride=kernel_size,
+            bias=False,
+            device=device,
+            dtype=dtype
+        )
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = hidden_states.view(
+            -1, self.in_channels, self.temporal_patch_size, self.patch_size, self.patch_size
+        )
+        hidden_states = self.proj(hidden_states)
+        return hidden_states.view(-1, self.embed_dim)
+
+
+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_vision(q, k, cos, sin):
+    cos, sin = cos.unsqueeze(-2).float(), sin.unsqueeze(-2).float()
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+class VisionRotaryEmbedding(nn.Module):
+    def __init__(self, dim: int, theta: float = 10000.0):
+        super().__init__()
+        self.dim = dim
+        self.theta = theta
+
+    def forward(self, seqlen: int, device) -> torch.Tensor:
+        inv_freq = 1.0 / (self.theta ** (torch.arange(0, self.dim, 2, dtype=torch.float, device=device) / self.dim))
+        seq = torch.arange(seqlen, device=inv_freq.device, dtype=inv_freq.dtype)
+        freqs = torch.outer(seq, inv_freq)
+        return freqs
+
+
+class PatchMerger(nn.Module):
+    def __init__(self, dim: int, context_dim: int, spatial_merge_size: int = 2, device=None, dtype=None, ops=None):
+        super().__init__()
+        self.hidden_size = context_dim * (spatial_merge_size ** 2)
+        self.ln_q = ops.RMSNorm(context_dim, eps=1e-6, device=device, dtype=dtype)
+        self.mlp = nn.Sequential(
+            ops.Linear(self.hidden_size, self.hidden_size, device=device, dtype=dtype),
+            nn.GELU(),
+            ops.Linear(self.hidden_size, dim, device=device, dtype=dtype),
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.ln_q(x).reshape(-1, self.hidden_size)
+        x = self.mlp(x)
+        return x
+
+
+class VisionAttention(nn.Module):
+    def __init__(self, hidden_size: int, num_heads: int, device=None, dtype=None, ops=None):
+        super().__init__()
+        self.hidden_size = hidden_size
+        self.num_heads = num_heads
+        self.head_dim = hidden_size // num_heads
+        self.scaling = self.head_dim ** -0.5
+
+        self.qkv = ops.Linear(hidden_size, hidden_size * 3, bias=True, device=device, dtype=dtype)
+        self.proj = ops.Linear(hidden_size, hidden_size, bias=True, device=device, dtype=dtype)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        cu_seqlens=None,
+        optimized_attention=None,
+    ) -> torch.Tensor:
+        if hidden_states.dim() == 2:
+            seq_length, _ = hidden_states.shape
+            batch_size = 1
+            hidden_states = hidden_states.unsqueeze(0)
+        else:
+            batch_size, seq_length, _ = hidden_states.shape
+
+        qkv = self.qkv(hidden_states)
+        qkv = qkv.reshape(batch_size, seq_length, 3, self.num_heads, self.head_dim)
+        query_states, key_states, value_states = qkv.reshape(seq_length, 3, self.num_heads, -1).permute(1, 0, 2, 3).unbind(0)
+
+        if position_embeddings is not None:
+            cos, sin = position_embeddings
+            query_states, key_states = apply_rotary_pos_emb_vision(query_states, key_states, cos, sin)
+
+        query_states = query_states.transpose(0, 1).unsqueeze(0)
+        key_states = key_states.transpose(0, 1).unsqueeze(0)
+        value_states = value_states.transpose(0, 1).unsqueeze(0)
+
+        lengths = cu_seqlens[1:] - cu_seqlens[:-1]
+        splits = [
+            torch.split(tensor, lengths.tolist(), dim=2) for tensor in (query_states, key_states, value_states)
+        ]
+
+        attn_outputs = [
+            optimized_attention(q, k, v, self.num_heads, skip_reshape=True)
+            for q, k, v in zip(*splits)
+        ]
+        attn_output = torch.cat(attn_outputs, dim=1)
+        attn_output = attn_output.reshape(seq_length, -1)
+        attn_output = self.proj(attn_output)
+
+        return attn_output
+
+
+class VisionMLP(nn.Module):
+    def __init__(self, hidden_size: int, intermediate_size: int, device=None, dtype=None, ops=None):
+        super().__init__()
+        self.gate_proj = ops.Linear(hidden_size, intermediate_size, bias=True, device=device, dtype=dtype)
+        self.up_proj = ops.Linear(hidden_size, intermediate_size, bias=True, device=device, dtype=dtype)
+        self.down_proj = ops.Linear(intermediate_size, hidden_size, bias=True, device=device, dtype=dtype)
+        self.act_fn = nn.SiLU()
+
+    def forward(self, hidden_state):
+        return self.down_proj(self.act_fn(self.gate_proj(hidden_state)) * self.up_proj(hidden_state))
+
+
+class VisionBlock(nn.Module):
+    def __init__(self, hidden_size: int, intermediate_size: int, num_heads: int, device=None, dtype=None, ops=None):
+        super().__init__()
+        self.norm1 = ops.RMSNorm(hidden_size, eps=1e-6, device=device, dtype=dtype)
+        self.norm2 = ops.RMSNorm(hidden_size, eps=1e-6, device=device, dtype=dtype)
+        self.attn = VisionAttention(hidden_size, num_heads, device=device, dtype=dtype, ops=ops)
+        self.mlp = VisionMLP(hidden_size, intermediate_size, device=device, dtype=dtype, ops=ops)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        cu_seqlens=None,
+        optimized_attention=None,
+    ) -> torch.Tensor:
+        residual = hidden_states
+        hidden_states = self.norm1(hidden_states)
+        hidden_states = self.attn(hidden_states, position_embeddings, cu_seqlens, optimized_attention)
+        hidden_states = residual + hidden_states
+
+        residual = hidden_states
+        hidden_states = self.norm2(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        return hidden_states
+
+
+class Qwen2VLVisionTransformer(nn.Module):
+    def __init__(
+        self,
+        hidden_size: int = 3584,
+        output_hidden_size: int = 3584,
+        intermediate_size: int = 3420,
+        num_heads: int = 16,
+        num_layers: int = 32,
+        patch_size: int = 14,
+        temporal_patch_size: int = 2,
+        spatial_merge_size: int = 2,
+        window_size: int = 112,
+        device=None,
+        dtype=None,
+        ops=None
+    ):
+        super().__init__()
+        self.hidden_size = hidden_size
+        self.patch_size = patch_size
+        self.spatial_merge_size = spatial_merge_size
+        self.window_size = window_size
+        self.fullatt_block_indexes = [7, 15, 23, 31]
+
+        self.patch_embed = VisionPatchEmbed(
+            patch_size=patch_size,
+            temporal_patch_size=temporal_patch_size,
+            in_channels=3,
+            embed_dim=hidden_size,
+            device=device,
+            dtype=dtype,
+            ops=ops,
+        )
+
+        head_dim = hidden_size // num_heads
+        self.rotary_pos_emb = VisionRotaryEmbedding(head_dim // 2)
+
+        self.blocks = nn.ModuleList([
+            VisionBlock(hidden_size, intermediate_size, num_heads, device, dtype, ops)
+            for _ in range(num_layers)
+        ])
+
+        self.merger = PatchMerger(
+            dim=output_hidden_size,
+            context_dim=hidden_size,
+            spatial_merge_size=spatial_merge_size,
+            device=device,
+            dtype=dtype,
+            ops=ops,
+        )
+
+    def get_window_index(self, grid_thw):
+        window_index = []
+        cu_window_seqlens = [0]
+        window_index_id = 0
+        vit_merger_window_size = self.window_size // self.spatial_merge_size // self.patch_size
+
+        for grid_t, grid_h, grid_w in grid_thw:
+            llm_grid_h = grid_h // self.spatial_merge_size
+            llm_grid_w = grid_w // self.spatial_merge_size
+
+            index = torch.arange(grid_t * llm_grid_h * llm_grid_w).reshape(grid_t, llm_grid_h, llm_grid_w)
+
+            pad_h = vit_merger_window_size - llm_grid_h % vit_merger_window_size
+            pad_w = vit_merger_window_size - llm_grid_w % vit_merger_window_size
+            num_windows_h = (llm_grid_h + pad_h) // vit_merger_window_size
+            num_windows_w = (llm_grid_w + pad_w) // vit_merger_window_size
+
+            index_padded = F.pad(index, (0, pad_w, 0, pad_h), "constant", -100)
+            index_padded = index_padded.reshape(
+                grid_t,
+                num_windows_h,
+                vit_merger_window_size,
+                num_windows_w,
+                vit_merger_window_size,
+            )
+            index_padded = index_padded.permute(0, 1, 3, 2, 4).reshape(
+                grid_t,
+                num_windows_h * num_windows_w,
+                vit_merger_window_size,
+                vit_merger_window_size,
+            )
+
+            seqlens = (index_padded != -100).sum([2, 3]).reshape(-1)
+            index_padded = index_padded.reshape(-1)
+            index_new = index_padded[index_padded != -100]
+            window_index.append(index_new + window_index_id)
+
+            cu_seqlens_tmp = seqlens.cumsum(0) * self.spatial_merge_size * self.spatial_merge_size + cu_window_seqlens[-1]
+            cu_window_seqlens.extend(cu_seqlens_tmp.tolist())
+            window_index_id += (grid_t * llm_grid_h * llm_grid_w).item()
+
+        window_index = torch.cat(window_index, dim=0)
+        return window_index, cu_window_seqlens
+
+    def get_position_embeddings(self, grid_thw, device):
+        pos_ids = []
+
+        for t, h, w in grid_thw:
+            hpos_ids = torch.arange(h, device=device).unsqueeze(1).expand(-1, w)
+            hpos_ids = hpos_ids.reshape(
+                h // self.spatial_merge_size,
+                self.spatial_merge_size,
+                w // self.spatial_merge_size,
+                self.spatial_merge_size,
+            )
+            hpos_ids = hpos_ids.permute(0, 2, 1, 3).flatten()
+
+            wpos_ids = torch.arange(w, device=device).unsqueeze(0).expand(h, -1)
+            wpos_ids = wpos_ids.reshape(
+                h // self.spatial_merge_size,
+                self.spatial_merge_size,
+                w // self.spatial_merge_size,
+                self.spatial_merge_size,
+            )
+            wpos_ids = wpos_ids.permute(0, 2, 1, 3).flatten()
+
+            pos_ids.append(torch.stack([hpos_ids, wpos_ids], dim=-1).repeat(t, 1))
+
+        pos_ids = torch.cat(pos_ids, dim=0)
+        max_grid_size = grid_thw[:, 1:].max()
+        rotary_pos_emb_full = self.rotary_pos_emb(max_grid_size, device)
+        return rotary_pos_emb_full[pos_ids].flatten(1)
+
+    def forward(
+        self,
+        pixel_values: torch.Tensor,
+        image_grid_thw: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        optimized_attention = optimized_attention_for_device(pixel_values.device, mask=False, small_input=True)
+
+        hidden_states = self.patch_embed(pixel_values)
+
+        window_index, cu_window_seqlens = self.get_window_index(image_grid_thw)
+        cu_window_seqlens = torch.tensor(cu_window_seqlens, device=hidden_states.device)
+        cu_window_seqlens = torch.unique_consecutive(cu_window_seqlens)
+
+        position_embeddings = self.get_position_embeddings(image_grid_thw, hidden_states.device)
+
+        seq_len, _ = hidden_states.size()
+        spatial_merge_unit = self.spatial_merge_size * self.spatial_merge_size
+
+        hidden_states = hidden_states.reshape(seq_len // spatial_merge_unit, spatial_merge_unit, -1)
+        hidden_states = hidden_states[window_index, :, :]
+        hidden_states = hidden_states.reshape(seq_len, -1)
+
+        position_embeddings = position_embeddings.reshape(seq_len // spatial_merge_unit, spatial_merge_unit, -1)
+        position_embeddings = position_embeddings[window_index, :, :]
+        position_embeddings = position_embeddings.reshape(seq_len, -1)
+        position_embeddings = torch.cat((position_embeddings, position_embeddings), dim=-1)
+        position_embeddings = (position_embeddings.cos(), position_embeddings.sin())
+
+        cu_seqlens = torch.repeat_interleave(image_grid_thw[:, 1] * image_grid_thw[:, 2], image_grid_thw[:, 0]).cumsum(
+            dim=0,
+            dtype=torch.int32,
+        )
+        cu_seqlens = F.pad(cu_seqlens, (1, 0), value=0)
+
+        for i, block in enumerate(self.blocks):
+            if i in self.fullatt_block_indexes:
+                cu_seqlens_now = cu_seqlens
+            else:
+                cu_seqlens_now = cu_window_seqlens
+            hidden_states = block(hidden_states, position_embeddings, cu_seqlens_now, optimized_attention=optimized_attention)
+
+        hidden_states = self.merger(hidden_states)
+        return hidden_states

comfy/text_encoders/t5.py

@@ -199,7 +199,7 @@ class T5Stack(torch.nn.Module):
         self.final_layer_norm = T5LayerNorm(model_dim, dtype=dtype, device=device, operations=operations)
         # self.dropout = nn.Dropout(config.dropout_rate)
 
-    def forward(self, x, attention_mask=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None):
+    def forward(self, x, attention_mask=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None, embeds_info=[]):
         mask = None
         if attention_mask is not None:
             mask = 1.0 - attention_mask.to(x.dtype).reshape((attention_mask.shape[0], 1, -1, attention_mask.shape[-1])).expand(attention_mask.shape[0], 1, attention_mask.shape[-1], attention_mask.shape[-1])

comfy_api_nodes/apis/__init__.py

@@ -1315,6 +1315,7 @@ class KlingTaskStatus(str, Enum):
 class KlingTextToVideoModelName(str, Enum):
     kling_v1 = 'kling-v1'
     kling_v1_6 = 'kling-v1-6'
+    kling_v2_1_master = 'kling-v2-1-master'
 
 
 class KlingVideoGenAspectRatio(str, Enum):
@@ -1347,6 +1348,8 @@ class KlingVideoGenModelName(str, Enum):
     kling_v1_5 = 'kling-v1-5'
     kling_v1_6 = 'kling-v1-6'
     kling_v2_master = 'kling-v2-master'
+    kling_v2_1 = 'kling-v2-1'
+    kling_v2_1_master = 'kling-v2-1-master'
 
 
 class KlingVideoResult(BaseModel):
@@ -1620,13 +1623,14 @@ class MinimaxTaskResultResponse(BaseModel):
     task_id: str = Field(..., description='The task ID being queried.')
 
 
-class Model(str, Enum):
+class MiniMaxModel(str, Enum):
     T2V_01_Director = 'T2V-01-Director'
     I2V_01_Director = 'I2V-01-Director'
     S2V_01 = 'S2V-01'
     I2V_01 = 'I2V-01'
     I2V_01_live = 'I2V-01-live'
     T2V_01 = 'T2V-01'
+    Hailuo_02 = 'MiniMax-Hailuo-02'
 
 
 class SubjectReferenceItem(BaseModel):
@@ -1648,7 +1652,7 @@ class MinimaxVideoGenerationRequest(BaseModel):
         None,
         description='URL or base64 encoding of the first frame image. Required when model is I2V-01, I2V-01-Director, or I2V-01-live.',
     )
-    model: Model = Field(
+    model: MiniMaxModel = Field(
         ...,
         description='Required. ID of model. Options: T2V-01-Director, I2V-01-Director, S2V-01, I2V-01, I2V-01-live, T2V-01',
     )
@@ -1665,6 +1669,14 @@ class MinimaxVideoGenerationRequest(BaseModel):
         None,
         description='Only available when model is S2V-01. The model will generate a video based on the subject uploaded through this parameter.',
     )
+    duration: Optional[int] = Field(
+        None,
+        description="The length of the output video in seconds."
+    )
+    resolution: Optional[str] = Field(
+        None,
+        description="The dimensions of the video display. 1080p corresponds to 1920 x 1080 pixels, 768p corresponds to 1366 x 768 pixels."
+    )
 
 
 class MinimaxVideoGenerationResponse(BaseModel):

comfy_api_nodes/nodes_gemini.py

@@ -46,6 +46,8 @@ class GeminiModel(str, Enum):
 
     gemini_2_5_pro_preview_05_06 = "gemini-2.5-pro-preview-05-06"
     gemini_2_5_flash_preview_04_17 = "gemini-2.5-flash-preview-04-17"
+    gemini_2_5_pro = "gemini-2.5-pro"
+    gemini_2_5_flash = "gemini-2.5-flash"
 
 
 def get_gemini_endpoint(
@@ -97,7 +99,7 @@ class GeminiNode(ComfyNodeABC):
                     {
                         "tooltip": "The Gemini model to use for generating responses.",
                         "options": [model.value for model in GeminiModel],
-                        "default": GeminiModel.gemini_2_5_pro_preview_05_06.value,
+                        "default": GeminiModel.gemini_2_5_pro.value,
                     },
                 ),
                 "seed": (

comfy_api_nodes/nodes_kling.py

@@ -421,6 +421,8 @@ class KlingTextToVideoNode(KlingNodeBase):
             "pro mode / 10s duration / kling-v2-master": ("pro", "10", "kling-v2-master"),
             "standard mode / 5s duration / kling-v2-master": ("std", "5", "kling-v2-master"),
             "standard mode / 10s duration / kling-v2-master": ("std", "10", "kling-v2-master"),
+            "pro mode / 5s duration / kling-v2-1-master": ("pro", "5", "kling-v2-1-master"),
+            "pro mode / 10s duration / kling-v2-1-master": ("pro", "10", "kling-v2-1-master"),
         }
 
     @classmethod

comfy_api_nodes/nodes_minimax.py

@@ -1,3 +1,4 @@
+from inspect import cleandoc
 from typing import Union
 import logging
 import torch
@@ -10,7 +11,7 @@ from comfy_api_nodes.apis import (
     MinimaxFileRetrieveResponse,
     MinimaxTaskResultResponse,
     SubjectReferenceItem,
-    Model
+    MiniMaxModel
 )
 from comfy_api_nodes.apis.client import (
     ApiEndpoint,
@@ -84,7 +85,6 @@ class MinimaxTextToVideoNode:
     FUNCTION = "generate_video"
     CATEGORY = "api node/video/MiniMax"
     API_NODE = True
-    OUTPUT_NODE = True
 
     async def generate_video(
         self,
@@ -121,7 +121,7 @@ class MinimaxTextToVideoNode:
                 response_model=MinimaxVideoGenerationResponse,
             ),
             request=MinimaxVideoGenerationRequest(
-                model=Model(model),
+                model=MiniMaxModel(model),
                 prompt=prompt_text,
                 callback_url=None,
                 first_frame_image=image_url,
@@ -251,7 +251,6 @@ class MinimaxImageToVideoNode(MinimaxTextToVideoNode):
     FUNCTION = "generate_video"
     CATEGORY = "api node/video/MiniMax"
     API_NODE = True
-    OUTPUT_NODE = True
 
 
 class MinimaxSubjectToVideoNode(MinimaxTextToVideoNode):
@@ -313,7 +312,181 @@ class MinimaxSubjectToVideoNode(MinimaxTextToVideoNode):
     FUNCTION = "generate_video"
     CATEGORY = "api node/video/MiniMax"
     API_NODE = True
-    OUTPUT_NODE = True
+
+
+class MinimaxHailuoVideoNode:
+    """Generates videos from prompt, with optional start frame using the new MiniMax Hailuo-02 model."""
+
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "prompt_text": (
+                    "STRING",
+                    {
+                        "multiline": True,
+                        "default": "",
+                        "tooltip": "Text prompt to guide the video generation.",
+                    },
+                ),
+            },
+            "optional": {
+                "seed": (
+                    IO.INT,
+                    {
+                        "default": 0,
+                        "min": 0,
+                        "max": 0xFFFFFFFFFFFFFFFF,
+                        "control_after_generate": True,
+                        "tooltip": "The random seed used for creating the noise.",
+                    },
+                ),
+                "first_frame_image": (
+                    IO.IMAGE,
+                    {
+                        "tooltip": "Optional image to use as the first frame to generate a video."
+                    },
+                ),
+                "prompt_optimizer": (
+                    IO.BOOLEAN,
+                    {
+                        "tooltip": "Optimize prompt to improve generation quality when needed.",
+                        "default": True,
+                    },
+                ),
+                "duration": (
+                    IO.COMBO,
+                    {
+                        "tooltip": "The length of the output video in seconds.",
+                        "default": 6,
+                        "options": [6, 10],
+                    },
+                ),
+                "resolution": (
+                    IO.COMBO,
+                    {
+                        "tooltip": "The dimensions of the video display. "
+                                   "1080p corresponds to 1920 x 1080 pixels, 768p corresponds to 1366 x 768 pixels.",
+                        "default": "768P",
+                        "options": ["768P", "1080P"],
+                    },
+                ),
+            },
+            "hidden": {
+                "auth_token": "AUTH_TOKEN_COMFY_ORG",
+                "comfy_api_key": "API_KEY_COMFY_ORG",
+                "unique_id": "UNIQUE_ID",
+            },
+        }
+
+    RETURN_TYPES = ("VIDEO",)
+    DESCRIPTION = cleandoc(__doc__ or "")
+    FUNCTION = "generate_video"
+    CATEGORY = "api node/video/MiniMax"
+    API_NODE = True
+
+    async def generate_video(
+        self,
+        prompt_text,
+        seed=0,
+        first_frame_image: torch.Tensor=None, # used for ImageToVideo
+        prompt_optimizer=True,
+        duration=6,
+        resolution="768P",
+        model="MiniMax-Hailuo-02",
+        unique_id: Union[str, None]=None,
+        **kwargs,
+    ):
+        if first_frame_image is None:
+            validate_string(prompt_text, field_name="prompt_text")
+
+        if model == "MiniMax-Hailuo-02" and resolution.upper() == "1080P" and duration != 6:
+            raise Exception(
+                "When model is MiniMax-Hailuo-02 and resolution is 1080P, duration is limited to 6 seconds."
+            )
+
+        # upload image, if passed in
+        image_url = None
+        if first_frame_image is not None:
+            image_url = (await upload_images_to_comfyapi(first_frame_image, max_images=1, auth_kwargs=kwargs))[0]
+
+        video_generate_operation = SynchronousOperation(
+            endpoint=ApiEndpoint(
+                path="/proxy/minimax/video_generation",
+                method=HttpMethod.POST,
+                request_model=MinimaxVideoGenerationRequest,
+                response_model=MinimaxVideoGenerationResponse,
+            ),
+            request=MinimaxVideoGenerationRequest(
+                model=MiniMaxModel(model),
+                prompt=prompt_text,
+                callback_url=None,
+                first_frame_image=image_url,
+                prompt_optimizer=prompt_optimizer,
+                duration=duration,
+                resolution=resolution,
+            ),
+            auth_kwargs=kwargs,
+        )
+        response = await video_generate_operation.execute()
+
+        task_id = response.task_id
+        if not task_id:
+            raise Exception(f"MiniMax generation failed: {response.base_resp}")
+
+        average_duration = 120 if resolution == "768P" else 240
+        video_generate_operation = PollingOperation(
+            poll_endpoint=ApiEndpoint(
+                path="/proxy/minimax/query/video_generation",
+                method=HttpMethod.GET,
+                request_model=EmptyRequest,
+                response_model=MinimaxTaskResultResponse,
+                query_params={"task_id": task_id},
+            ),
+            completed_statuses=["Success"],
+            failed_statuses=["Fail"],
+            status_extractor=lambda x: x.status.value,
+            estimated_duration=average_duration,
+            node_id=unique_id,
+            auth_kwargs=kwargs,
+        )
+        task_result = await video_generate_operation.execute()
+
+        file_id = task_result.file_id
+        if file_id is None:
+            raise Exception("Request was not successful. Missing file ID.")
+        file_retrieve_operation = SynchronousOperation(
+            endpoint=ApiEndpoint(
+                path="/proxy/minimax/files/retrieve",
+                method=HttpMethod.GET,
+                request_model=EmptyRequest,
+                response_model=MinimaxFileRetrieveResponse,
+                query_params={"file_id": int(file_id)},
+            ),
+            request=EmptyRequest(),
+            auth_kwargs=kwargs,
+        )
+        file_result = await file_retrieve_operation.execute()
+
+        file_url = file_result.file.download_url
+        if file_url is None:
+            raise Exception(
+                f"No video was found in the response. Full response: {file_result.model_dump()}"
+            )
+        logging.info(f"Generated video URL: {file_url}")
+        if unique_id:
+            if hasattr(file_result.file, "backup_download_url"):
+                message = f"Result URL: {file_url}\nBackup URL: {file_result.file.backup_download_url}"
+            else:
+                message = f"Result URL: {file_url}"
+            PromptServer.instance.send_progress_text(message, unique_id)
+
+        video_io = await download_url_to_bytesio(file_url)
+        if video_io is None:
+            error_msg = f"Failed to download video from {file_url}"
+            logging.error(error_msg)
+            raise Exception(error_msg)
+        return (VideoFromFile(video_io),)
 
 
 # A dictionary that contains all nodes you want to export with their names
@@ -322,6 +495,7 @@ NODE_CLASS_MAPPINGS = {
     "MinimaxTextToVideoNode": MinimaxTextToVideoNode,
     "MinimaxImageToVideoNode": MinimaxImageToVideoNode,
     # "MinimaxSubjectToVideoNode": MinimaxSubjectToVideoNode,
+    "MinimaxHailuoVideoNode": MinimaxHailuoVideoNode,
 }
 
 # A dictionary that contains the friendly/humanly readable titles for the nodes
@@ -329,4 +503,5 @@ NODE_DISPLAY_NAME_MAPPINGS = {
     "MinimaxTextToVideoNode": "MiniMax Text to Video",
     "MinimaxImageToVideoNode": "MiniMax Image to Video",
     "MinimaxSubjectToVideoNode": "MiniMax Subject to Video",
+    "MinimaxHailuoVideoNode": "MiniMax Hailuo Video",
 }

comfy_api_nodes/nodes_moonvalley.py

@@ -1,6 +1,5 @@
 import logging
 from typing import Any, Callable, Optional, TypeVar
-import random
 import torch
 from comfy_api_nodes.util.validation_utils import (
     get_image_dimensions,
@@ -208,20 +207,29 @@ def _get_video_dimensions(video: VideoInput) -> tuple[int, int]:
 def _validate_video_dimensions(width: int, height: int) -> None:
     """Validates video dimensions meet Moonvalley V2V requirements."""
     supported_resolutions = {
-        (1920, 1080), (1080, 1920), (1152, 1152),
-        (1536, 1152), (1152, 1536)
+        (1920, 1080),
+        (1080, 1920),
+        (1152, 1152),
+        (1536, 1152),
+        (1152, 1536),
     }
 
     if (width, height) not in supported_resolutions:
-        supported_list = ', '.join([f'{w}x{h}' for w, h in sorted(supported_resolutions)])
-        raise ValueError(f"Resolution {width}x{height} not supported. Supported: {supported_list}")
+        supported_list = ", ".join(
+            [f"{w}x{h}" for w, h in sorted(supported_resolutions)]
+        )
+        raise ValueError(
+            f"Resolution {width}x{height} not supported. Supported: {supported_list}"
+        )
 
 
 def _validate_container_format(video: VideoInput) -> None:
     """Validates video container format is MP4."""
     container_format = video.get_container_format()
-    if container_format not in ['mp4', 'mov,mp4,m4a,3gp,3g2,mj2']:
-        raise ValueError(f"Only MP4 container format supported. Got: {container_format}")
+    if container_format not in ["mp4", "mov,mp4,m4a,3gp,3g2,mj2"]:
+        raise ValueError(
+            f"Only MP4 container format supported. Got: {container_format}"
+        )
 
 
 def _validate_and_trim_duration(video: VideoInput) -> VideoInput:
@@ -244,7 +252,6 @@ def _trim_if_too_long(video: VideoInput, duration: float) -> VideoInput:
     return video
 
 
-
 def trim_video(video: VideoInput, duration_sec: float) -> VideoInput:
     """
     Returns a new VideoInput object trimmed from the beginning to the specified duration,
@@ -302,7 +309,9 @@ def trim_video(video: VideoInput, duration_sec: float) -> VideoInput:
         # Calculate target frame count that's divisible by 16
         fps = input_container.streams.video[0].average_rate
         estimated_frames = int(duration_sec * fps)
-        target_frames = (estimated_frames // 16) * 16  # Round down to nearest multiple of 16
+        target_frames = (
+            estimated_frames // 16
+        ) * 16  # Round down to nearest multiple of 16
 
         if target_frames == 0:
             raise ValueError("Video too short: need at least 16 frames for Moonvalley")
@@ -424,7 +433,7 @@ class BaseMoonvalleyVideoNode:
                     MoonvalleyTextToVideoInferenceParams,
                     "negative_prompt",
                     multiline=True,
-                    default="low-poly, flat shader, bad rigging, stiff animation, uncanny eyes, low-quality textures, looping glitch, cheap effect, overbloom, bloom spam, default lighting, game asset, stiff face, ugly specular, AI artifacts",
+                    default="<synthetic> <scene cut> gopro, bright, contrast, static, overexposed, vignette, artifacts, still, noise, texture, scanlines, videogame, 360 camera, VR, transition, flare, saturation, distorted, warped, wide angle, saturated, vibrant, glowing, cross dissolve, cheesy, ugly hands, mutated hands, mutant, disfigured, extra fingers, blown out, horrible, blurry, worst quality, bad, dissolve, melt, fade in, fade out, wobbly, weird, low quality, plastic, stock footage, video camera, boring",
                 ),
                 "resolution": (
                     IO.COMBO,
@@ -441,12 +450,11 @@ class BaseMoonvalleyVideoNode:
                         "tooltip": "Resolution of the output video",
                     },
                 ),
-                # "length": (IO.COMBO,{"options":['5s','10s'], "default": '5s'}),
                 "prompt_adherence": model_field_to_node_input(
                     IO.FLOAT,
                     MoonvalleyTextToVideoInferenceParams,
                     "guidance_scale",
-                    default=7.0,
+                    default=10.0,
                     step=1,
                     min=1,
                     max=20,
@@ -455,13 +463,12 @@ class BaseMoonvalleyVideoNode:
                     IO.INT,
                     MoonvalleyTextToVideoInferenceParams,
                     "seed",
-                    default=random.randint(0, 2**32 - 1),
+                    default=9,
                     min=0,
                     max=4294967295,
                     step=1,
                     display="number",
                     tooltip="Random seed value",
-                    control_after_generate=True,
                 ),
                 "steps": model_field_to_node_input(
                     IO.INT,
@@ -532,9 +539,11 @@ class MoonvalleyImg2VideoNode(BaseMoonvalleyVideoNode):
         # Get MIME type from tensor - assuming PNG format for image tensors
         mime_type = "image/png"
 
-        image_url = (await upload_images_to_comfyapi(
-            image, max_images=1, auth_kwargs=kwargs, mime_type=mime_type
-        ))[0]
+        image_url = (
+            await upload_images_to_comfyapi(
+                image, max_images=1, auth_kwargs=kwargs, mime_type=mime_type
+            )
+        )[0]
 
         request = MoonvalleyTextToVideoRequest(
             image_url=image_url, prompt_text=prompt, inference_params=inference_params
@@ -570,17 +579,39 @@ class MoonvalleyVideo2VideoNode(BaseMoonvalleyVideoNode):
         return {
             "required": {
                 "prompt": model_field_to_node_input(
-                    IO.STRING, MoonvalleyVideoToVideoRequest, "prompt_text",
-                    multiline=True
+                    IO.STRING,
+                    MoonvalleyVideoToVideoRequest,
+                    "prompt_text",
+                    multiline=True,
                 ),
                 "negative_prompt": model_field_to_node_input(
                     IO.STRING,
                     MoonvalleyVideoToVideoInferenceParams,
                     "negative_prompt",
                     multiline=True,
-                    default="low-poly, flat shader, bad rigging, stiff animation, uncanny eyes, low-quality textures, looping glitch, cheap effect, overbloom, bloom spam, default lighting, game asset, stiff face, ugly specular, AI artifacts"
+                    default="<synthetic> <scene cut> gopro, bright, contrast, static, overexposed, vignette, artifacts, still, noise, texture, scanlines, videogame, 360 camera, VR, transition, flare, saturation, distorted, warped, wide angle, saturated, vibrant, glowing, cross dissolve, cheesy, ugly hands, mutated hands, mutant, disfigured, extra fingers, blown out, horrible, blurry, worst quality, bad, dissolve, melt, fade in, fade out, wobbly, weird, low quality, plastic, stock footage, video camera, boring",
+                ),
+                "seed": model_field_to_node_input(
+                    IO.INT,
+                    MoonvalleyVideoToVideoInferenceParams,
+                    "seed",
+                    default=9,
+                    min=0,
+                    max=4294967295,
+                    step=1,
+                    display="number",
+                    tooltip="Random seed value",
+                    control_after_generate=False,
+                ),
+                "prompt_adherence": model_field_to_node_input(
+                    IO.FLOAT,
+                    MoonvalleyVideoToVideoInferenceParams,
+                    "guidance_scale",
+                    default=10.0,
+                    step=1,
+                    min=1,
+                    max=20,
                 ),
-                "seed": model_field_to_node_input(IO.INT,MoonvalleyVideoToVideoInferenceParams, "seed", default=random.randint(0, 2**32 - 1), min=0, max=4294967295, step=1, display="number", tooltip="Random seed value", control_after_generate=True),
             },
             "hidden": {
                 "auth_token": "AUTH_TOKEN_COMFY_ORG",
@@ -588,7 +619,14 @@ class MoonvalleyVideo2VideoNode(BaseMoonvalleyVideoNode):
                 "unique_id": "UNIQUE_ID",
             },
             "optional": {
-                "video": (IO.VIDEO, {"default": "", "multiline": False, "tooltip": "The reference video used to generate the output video. Must be at least 5 seconds long. Videos longer than 5s will be automatically trimmed. Only MP4 format supported."}),
+                "video": (
+                    IO.VIDEO,
+                    {
+                        "default": "",
+                        "multiline": False,
+                        "tooltip": "The reference video used to generate the output video. Must be at least 5 seconds long. Videos longer than 5s will be automatically trimmed. Only MP4 format supported.",
+                    },
+                ),
                 "control_type": (
                     ["Motion Transfer", "Pose Transfer"],
                     {"default": "Motion Transfer"},
@@ -602,8 +640,14 @@ class MoonvalleyVideo2VideoNode(BaseMoonvalleyVideoNode):
                         "max": 100,
                         "tooltip": "Only used if control_type is 'Motion Transfer'",
                     },
-                )
-            }
+                ),
+                "image": model_field_to_node_input(
+                    IO.IMAGE,
+                    MoonvalleyTextToVideoRequest,
+                    "image_url",
+                    tooltip="The reference image used to generate the video",
+                ),
+            },
         }
 
     RETURN_TYPES = ("VIDEO",)
@@ -613,6 +657,7 @@ class MoonvalleyVideo2VideoNode(BaseMoonvalleyVideoNode):
         self, prompt, negative_prompt, unique_id: Optional[str] = None, **kwargs
     ):
         video = kwargs.get("video")
+        image = kwargs.get("image", None)
 
         if not video:
             raise MoonvalleyApiError("video is required")
@@ -620,8 +665,16 @@ class MoonvalleyVideo2VideoNode(BaseMoonvalleyVideoNode):
         video_url = ""
         if video:
             validated_video = validate_video_to_video_input(video)
-            video_url = await upload_video_to_comfyapi(validated_video, auth_kwargs=kwargs)
+            video_url = await upload_video_to_comfyapi(
+                validated_video, auth_kwargs=kwargs
+            )
+        mime_type = "image/png"
 
+        if not image is None:
+            validate_input_image(image, with_frame_conditioning=True)
+            image_url = await upload_images_to_comfyapi(
+                image=image, auth_kwargs=kwargs, max_images=1, mime_type=mime_type
+            )
         control_type = kwargs.get("control_type")
         motion_intensity = kwargs.get("motion_intensity")
 
@@ -631,12 +684,12 @@ class MoonvalleyVideo2VideoNode(BaseMoonvalleyVideoNode):
         # Only include motion_intensity for Motion Transfer
         control_params = {}
         if control_type == "Motion Transfer" and motion_intensity is not None:
-            control_params['motion_intensity'] = motion_intensity
+            control_params["motion_intensity"] = motion_intensity
 
-        inference_params=MoonvalleyVideoToVideoInferenceParams(
+        inference_params = MoonvalleyVideoToVideoInferenceParams(
             negative_prompt=negative_prompt,
             seed=kwargs.get("seed"),
-            control_params=control_params
+            control_params=control_params,
         )
 
         control = self.parseControlParameter(control_type)
@@ -647,6 +700,7 @@ class MoonvalleyVideo2VideoNode(BaseMoonvalleyVideoNode):
             prompt_text=prompt,
             inference_params=inference_params,
         )
+        request.image_url = image_url if not image is None else None
 
         initial_operation = SynchronousOperation(
             endpoint=ApiEndpoint(
@@ -694,15 +748,15 @@ class MoonvalleyTxt2VideoNode(BaseMoonvalleyVideoNode):
         validate_prompts(prompt, negative_prompt, MOONVALLEY_MAREY_MAX_PROMPT_LENGTH)
         width_height = self.parseWidthHeightFromRes(kwargs.get("resolution"))
 
-        inference_params=MoonvalleyTextToVideoInferenceParams(
-                    negative_prompt=negative_prompt,
-                    steps=kwargs.get("steps"),
-                    seed=kwargs.get("seed"),
-                    guidance_scale=kwargs.get("prompt_adherence"),
-                    num_frames=128,
-                    width=width_height.get("width"),
-                    height=width_height.get("height"),
-                )
+        inference_params = MoonvalleyTextToVideoInferenceParams(
+            negative_prompt=negative_prompt,
+            steps=kwargs.get("steps"),
+            seed=kwargs.get("seed"),
+            guidance_scale=kwargs.get("prompt_adherence"),
+            num_frames=128,
+            width=width_height.get("width"),
+            height=width_height.get("height"),
+        )
         request = MoonvalleyTextToVideoRequest(
             prompt_text=prompt, inference_params=inference_params
         )

comfy_api_nodes/nodes_openai.py

@@ -80,6 +80,9 @@ class SupportedOpenAIModel(str, Enum):
     gpt_4_1 = "gpt-4.1"
     gpt_4_1_mini = "gpt-4.1-mini"
     gpt_4_1_nano = "gpt-4.1-nano"
+    gpt_5 = "gpt-5"
+    gpt_5_mini = "gpt-5-mini"
+    gpt_5_nano = "gpt-5-nano"
 
 
 class OpenAIDalle2(ComfyNodeABC):
@@ -464,8 +467,6 @@ class OpenAIGPTImage1(ComfyNodeABC):
         path = "/proxy/openai/images/generations"
         content_type = "application/json"
         request_class = OpenAIImageGenerationRequest
-        img_binaries = []
-        mask_binary = None
         files = []
 
         if image is not None:
@@ -484,14 +485,11 @@ class OpenAIGPTImage1(ComfyNodeABC):
                 img_byte_arr = io.BytesIO()
                 img.save(img_byte_arr, format="PNG")
                 img_byte_arr.seek(0)
-                img_binary = img_byte_arr
-                img_binary.name = f"image_{i}.png"
 
-                img_binaries.append(img_binary)
                 if batch_size == 1:
-                    files.append(("image", img_binary))
+                    files.append(("image", (f"image_{i}.png", img_byte_arr, "image/png")))
                 else:
-                    files.append(("image[]", img_binary))
+                    files.append(("image[]", (f"image_{i}.png", img_byte_arr, "image/png")))
 
         if mask is not None:
             if image is None:
@@ -511,9 +509,7 @@ class OpenAIGPTImage1(ComfyNodeABC):
             mask_img_byte_arr = io.BytesIO()
             mask_img.save(mask_img_byte_arr, format="PNG")
             mask_img_byte_arr.seek(0)
-            mask_binary = mask_img_byte_arr
-            mask_binary.name = "mask.png"
-            files.append(("mask", mask_binary))
+            files.append(("mask", ("mask.png", mask_img_byte_arr, "image/png")))
 
         # Build the operation
         operation = SynchronousOperation(

comfy_api_nodes/nodes_vidu.py

@@ -0,0 +1,622 @@
+import logging
+from enum import Enum
+from typing import Any, Callable, Optional, Literal, TypeVar
+from typing_extensions import override
+
+import torch
+from pydantic import BaseModel, Field
+
+from comfy_api.latest import ComfyExtension, io as comfy_io
+from comfy_api_nodes.util.validation_utils import (
+    validate_aspect_ratio_closeness,
+    validate_image_dimensions,
+    validate_image_aspect_ratio_range,
+    get_number_of_images,
+)
+from comfy_api_nodes.apis.client import (
+    ApiEndpoint,
+    HttpMethod,
+    SynchronousOperation,
+    PollingOperation,
+    EmptyRequest,
+)
+from comfy_api_nodes.apinode_utils import download_url_to_video_output, upload_images_to_comfyapi
+
+
+VIDU_TEXT_TO_VIDEO = "/proxy/vidu/text2video"
+VIDU_IMAGE_TO_VIDEO = "/proxy/vidu/img2video"
+VIDU_REFERENCE_VIDEO = "/proxy/vidu/reference2video"
+VIDU_START_END_VIDEO = "/proxy/vidu/start-end2video"
+VIDU_GET_GENERATION_STATUS = "/proxy/vidu/tasks/%s/creations"
+
+R = TypeVar("R")
+
+class VideoModelName(str, Enum):
+    vidu_q1 = 'viduq1'
+
+
+class AspectRatio(str, Enum):
+    r_16_9 = "16:9"
+    r_9_16 = "9:16"
+    r_1_1 = "1:1"
+
+
+class Resolution(str, Enum):
+    r_1080p = "1080p"
+
+
+class MovementAmplitude(str, Enum):
+    auto = "auto"
+    small = "small"
+    medium = "medium"
+    large = "large"
+
+
+class TaskCreationRequest(BaseModel):
+    model: VideoModelName = VideoModelName.vidu_q1
+    prompt: Optional[str] = Field(None, max_length=1500)
+    duration: Optional[Literal[5]] = 5
+    seed: Optional[int] = Field(0, ge=0, le=2147483647)
+    aspect_ratio: Optional[AspectRatio] = AspectRatio.r_16_9
+    resolution: Optional[Resolution] = Resolution.r_1080p
+    movement_amplitude: Optional[MovementAmplitude] = MovementAmplitude.auto
+    images: Optional[list[str]] = Field(None, description="Base64 encoded string or image URL")
+
+
+class TaskStatus(str, Enum):
+    created = "created"
+    queueing = "queueing"
+    processing = "processing"
+    success = "success"
+    failed = "failed"
+
+
+class TaskCreationResponse(BaseModel):
+    task_id: str = Field(...)
+    state: TaskStatus = Field(...)
+    created_at: str = Field(...)
+    code: Optional[int] = Field(None, description="Error code")
+
+
+class TaskResult(BaseModel):
+    id: str = Field(..., description="Creation id")
+    url: str = Field(..., description="The URL of the generated results, valid for one hour")
+    cover_url: str = Field(..., description="The cover URL of the generated results, valid for one hour")
+
+
+class TaskStatusResponse(BaseModel):
+    state: TaskStatus = Field(...)
+    err_code: Optional[str] = Field(None)
+    creations: list[TaskResult] = Field(..., description="Generated results")
+
+
+async def poll_until_finished(
+    auth_kwargs: dict[str, str],
+    api_endpoint: ApiEndpoint[Any, R],
+    result_url_extractor: Optional[Callable[[R], str]] = None,
+    estimated_duration: Optional[int] = None,
+    node_id: Optional[str] = None,
+) -> R:
+    return await PollingOperation(
+        poll_endpoint=api_endpoint,
+        completed_statuses=[TaskStatus.success.value],
+        failed_statuses=[TaskStatus.failed.value],
+        status_extractor=lambda response: response.state.value,
+        auth_kwargs=auth_kwargs,
+        result_url_extractor=result_url_extractor,
+        estimated_duration=estimated_duration,
+        node_id=node_id,
+        poll_interval=16.0,
+        max_poll_attempts=256,
+    ).execute()
+
+
+def get_video_url_from_response(response) -> Optional[str]:
+    if response.creations:
+        return response.creations[0].url
+    return None
+
+
+def get_video_from_response(response) -> TaskResult:
+    if not response.creations:
+        error_msg = f"Vidu request does not contain results. State: {response.state}, Error Code: {response.err_code}"
+        logging.info(error_msg)
+        raise RuntimeError(error_msg)
+    logging.info("Vidu task %s succeeded. Video URL: %s", response.creations[0].id, response.creations[0].url)
+    return response.creations[0]
+
+
+async def execute_task(
+    vidu_endpoint: str,
+    auth_kwargs: Optional[dict[str, str]],
+    payload: TaskCreationRequest,
+    estimated_duration: int,
+    node_id: str,
+) -> R:
+    response = await SynchronousOperation(
+        endpoint=ApiEndpoint(
+            path=vidu_endpoint,
+            method=HttpMethod.POST,
+            request_model=TaskCreationRequest,
+            response_model=TaskCreationResponse,
+        ),
+        request=payload,
+        auth_kwargs=auth_kwargs,
+    ).execute()
+    if response.state == TaskStatus.failed:
+        error_msg = f"Vidu request failed. Code: {response.code}"
+        logging.error(error_msg)
+        raise RuntimeError(error_msg)
+    return await poll_until_finished(
+        auth_kwargs,
+        ApiEndpoint(
+            path=VIDU_GET_GENERATION_STATUS % response.task_id,
+            method=HttpMethod.GET,
+            request_model=EmptyRequest,
+            response_model=TaskStatusResponse,
+        ),
+        result_url_extractor=get_video_url_from_response,
+        estimated_duration=estimated_duration,
+        node_id=node_id,
+    )
+
+
+class ViduTextToVideoNode(comfy_io.ComfyNode):
+
+    @classmethod
+    def define_schema(cls):
+        return comfy_io.Schema(
+            node_id="ViduTextToVideoNode",
+            display_name="Vidu Text To Video Generation",
+            category="api node/video/Vidu",
+            description="Generate video from text prompt",
+            inputs=[
+                comfy_io.Combo.Input(
+                    "model",
+                    options=[model.value for model in VideoModelName],
+                    default=VideoModelName.vidu_q1.value,
+                    tooltip="Model name",
+                ),
+                comfy_io.String.Input(
+                    "prompt",
+                    multiline=True,
+                    tooltip="A textual description for video generation",
+                ),
+                comfy_io.Int.Input(
+                    "duration",
+                    default=5,
+                    min=5,
+                    max=5,
+                    step=1,
+                    display_mode=comfy_io.NumberDisplay.number,
+                    tooltip="Duration of the output video in seconds",
+                    optional=True,
+                ),
+                comfy_io.Int.Input(
+                    "seed",
+                    default=0,
+                    min=0,
+                    max=2147483647,
+                    step=1,
+                    display_mode=comfy_io.NumberDisplay.number,
+                    control_after_generate=True,
+                    tooltip="Seed for video generation (0 for random)",
+                    optional=True,
+                ),
+                comfy_io.Combo.Input(
+                    "aspect_ratio",
+                    options=[model.value for model in AspectRatio],
+                    default=AspectRatio.r_16_9.value,
+                    tooltip="The aspect ratio of the output video",
+                    optional=True,
+                ),
+                comfy_io.Combo.Input(
+                    "resolution",
+                    options=[model.value for model in Resolution],
+                    default=Resolution.r_1080p.value,
+                    tooltip="Supported values may vary by model & duration",
+                    optional=True,
+                ),
+                comfy_io.Combo.Input(
+                    "movement_amplitude",
+                    options=[model.value for model in MovementAmplitude],
+                    default=MovementAmplitude.auto.value,
+                    tooltip="The movement amplitude of objects in the frame",
+                    optional=True,
+                ),
+            ],
+            outputs=[
+                comfy_io.Video.Output(),
+            ],
+            hidden=[
+                comfy_io.Hidden.auth_token_comfy_org,
+                comfy_io.Hidden.api_key_comfy_org,
+                comfy_io.Hidden.unique_id,
+            ],
+            is_api_node=True,
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        model: str,
+        prompt: str,
+        duration: int,
+        seed: int,
+        aspect_ratio: str,
+        resolution: str,
+        movement_amplitude: str,
+    ) -> comfy_io.NodeOutput:
+        if not prompt:
+            raise ValueError("The prompt field is required and cannot be empty.")
+        payload = TaskCreationRequest(
+            model_name=model,
+            prompt=prompt,
+            duration=duration,
+            seed=seed,
+            aspect_ratio=aspect_ratio,
+            resolution=resolution,
+            movement_amplitude=movement_amplitude,
+        )
+        auth = {
+            "auth_token": cls.hidden.auth_token_comfy_org,
+            "comfy_api_key": cls.hidden.api_key_comfy_org,
+        }
+        results = await execute_task(VIDU_TEXT_TO_VIDEO, auth, payload, 320, cls.hidden.unique_id)
+        return comfy_io.NodeOutput(await download_url_to_video_output(get_video_from_response(results).url))
+
+
+class ViduImageToVideoNode(comfy_io.ComfyNode):
+
+    @classmethod
+    def define_schema(cls):
+        return comfy_io.Schema(
+            node_id="ViduImageToVideoNode",
+            display_name="Vidu Image To Video Generation",
+            category="api node/video/Vidu",
+            description="Generate video from image and optional prompt",
+            inputs=[
+                comfy_io.Combo.Input(
+                    "model",
+                    options=[model.value for model in VideoModelName],
+                    default=VideoModelName.vidu_q1.value,
+                    tooltip="Model name",
+                ),
+                comfy_io.Image.Input(
+                    "image",
+                    tooltip="An image to be used as the start frame of the generated video",
+                ),
+                comfy_io.String.Input(
+                    "prompt",
+                    multiline=True,
+                    default="",
+                    tooltip="A textual description for video generation",
+                    optional=True,
+                ),
+                comfy_io.Int.Input(
+                    "duration",
+                    default=5,
+                    min=5,
+                    max=5,
+                    step=1,
+                    display_mode=comfy_io.NumberDisplay.number,
+                    tooltip="Duration of the output video in seconds",
+                    optional=True,
+                ),
+                comfy_io.Int.Input(
+                    "seed",
+                    default=0,
+                    min=0,
+                    max=2147483647,
+                    step=1,
+                    display_mode=comfy_io.NumberDisplay.number,
+                    control_after_generate=True,
+                    tooltip="Seed for video generation (0 for random)",
+                    optional=True,
+                ),
+                comfy_io.Combo.Input(
+                    "resolution",
+                    options=[model.value for model in Resolution],
+                    default=Resolution.r_1080p.value,
+                    tooltip="Supported values may vary by model & duration",
+                    optional=True,
+                ),
+                comfy_io.Combo.Input(
+                    "movement_amplitude",
+                    options=[model.value for model in MovementAmplitude],
+                    default=MovementAmplitude.auto.value,
+                    tooltip="The movement amplitude of objects in the frame",
+                    optional=True,
+                ),
+            ],
+            outputs=[
+                comfy_io.Video.Output(),
+            ],
+            hidden=[
+                comfy_io.Hidden.auth_token_comfy_org,
+                comfy_io.Hidden.api_key_comfy_org,
+                comfy_io.Hidden.unique_id,
+            ],
+            is_api_node=True,
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        model: str,
+        image: torch.Tensor,
+        prompt: str,
+        duration: int,
+        seed: int,
+        resolution: str,
+        movement_amplitude: str,
+    ) -> comfy_io.NodeOutput:
+        if get_number_of_images(image) > 1:
+            raise ValueError("Only one input image is allowed.")
+        validate_image_aspect_ratio_range(image, (1, 4), (4, 1))
+        payload = TaskCreationRequest(
+            model_name=model,
+            prompt=prompt,
+            duration=duration,
+            seed=seed,
+            resolution=resolution,
+            movement_amplitude=movement_amplitude,
+        )
+        auth = {
+            "auth_token": cls.hidden.auth_token_comfy_org,
+            "comfy_api_key": cls.hidden.api_key_comfy_org,
+        }
+        payload.images = await upload_images_to_comfyapi(
+            image,
+            max_images=1,
+            mime_type="image/png",
+            auth_kwargs=auth,
+        )
+        results = await execute_task(VIDU_IMAGE_TO_VIDEO, auth, payload, 120, cls.hidden.unique_id)
+        return comfy_io.NodeOutput(await download_url_to_video_output(get_video_from_response(results).url))
+
+
+class ViduReferenceVideoNode(comfy_io.ComfyNode):
+
+    @classmethod
+    def define_schema(cls):
+        return comfy_io.Schema(
+            node_id="ViduReferenceVideoNode",
+            display_name="Vidu Reference To Video Generation",
+            category="api node/video/Vidu",
+            description="Generate video from multiple images and prompt",
+            inputs=[
+                comfy_io.Combo.Input(
+                    "model",
+                    options=[model.value for model in VideoModelName],
+                    default=VideoModelName.vidu_q1.value,
+                    tooltip="Model name",
+                ),
+                comfy_io.Image.Input(
+                    "images",
+                    tooltip="Images to use as references to generate a video with consistent subjects (max 7 images).",
+                ),
+                comfy_io.String.Input(
+                    "prompt",
+                    multiline=True,
+                    tooltip="A textual description for video generation",
+                ),
+                comfy_io.Int.Input(
+                    "duration",
+                    default=5,
+                    min=5,
+                    max=5,
+                    step=1,
+                    display_mode=comfy_io.NumberDisplay.number,
+                    tooltip="Duration of the output video in seconds",
+                    optional=True,
+                ),
+                comfy_io.Int.Input(
+                    "seed",
+                    default=0,
+                    min=0,
+                    max=2147483647,
+                    step=1,
+                    display_mode=comfy_io.NumberDisplay.number,
+                    control_after_generate=True,
+                    tooltip="Seed for video generation (0 for random)",
+                    optional=True,
+                ),
+                comfy_io.Combo.Input(
+                    "aspect_ratio",
+                    options=[model.value for model in AspectRatio],
+                    default=AspectRatio.r_16_9.value,
+                    tooltip="The aspect ratio of the output video",
+                    optional=True,
+                ),
+                comfy_io.Combo.Input(
+                    "resolution",
+                    options=[model.value for model in Resolution],
+                    default=Resolution.r_1080p.value,
+                    tooltip="Supported values may vary by model & duration",
+                    optional=True,
+                ),
+                comfy_io.Combo.Input(
+                    "movement_amplitude",
+                    options=[model.value for model in MovementAmplitude],
+                    default=MovementAmplitude.auto.value,
+                    tooltip="The movement amplitude of objects in the frame",
+                    optional=True,
+                ),
+            ],
+            outputs=[
+                comfy_io.Video.Output(),
+            ],
+            hidden=[
+                comfy_io.Hidden.auth_token_comfy_org,
+                comfy_io.Hidden.api_key_comfy_org,
+                comfy_io.Hidden.unique_id,
+            ],
+            is_api_node=True,
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        model: str,
+        images: torch.Tensor,
+        prompt: str,
+        duration: int,
+        seed: int,
+        aspect_ratio: str,
+        resolution: str,
+        movement_amplitude: str,
+    ) -> comfy_io.NodeOutput:
+        if not prompt:
+            raise ValueError("The prompt field is required and cannot be empty.")
+        a = get_number_of_images(images)
+        if a > 7:
+            raise ValueError("Too many images, maximum allowed is 7.")
+        for image in images:
+            validate_image_aspect_ratio_range(image, (1, 4), (4, 1))
+            validate_image_dimensions(image, min_width=128, min_height=128)
+        payload = TaskCreationRequest(
+            model_name=model,
+            prompt=prompt,
+            duration=duration,
+            seed=seed,
+            aspect_ratio=aspect_ratio,
+            resolution=resolution,
+            movement_amplitude=movement_amplitude,
+        )
+        auth = {
+            "auth_token": cls.hidden.auth_token_comfy_org,
+            "comfy_api_key": cls.hidden.api_key_comfy_org,
+        }
+        payload.images = await upload_images_to_comfyapi(
+            images,
+            max_images=7,
+            mime_type="image/png",
+            auth_kwargs=auth,
+        )
+        results = await execute_task(VIDU_REFERENCE_VIDEO, auth, payload, 120, cls.hidden.unique_id)
+        return comfy_io.NodeOutput(await download_url_to_video_output(get_video_from_response(results).url))
+
+
+class ViduStartEndToVideoNode(comfy_io.ComfyNode):
+
+    @classmethod
+    def define_schema(cls):
+        return comfy_io.Schema(
+            node_id="ViduStartEndToVideoNode",
+            display_name="Vidu Start End To Video Generation",
+            category="api node/video/Vidu",
+            description="Generate a video from start and end frames and a prompt",
+            inputs=[
+                comfy_io.Combo.Input(
+                    "model",
+                    options=[model.value for model in VideoModelName],
+                    default=VideoModelName.vidu_q1.value,
+                    tooltip="Model name",
+                ),
+                comfy_io.Image.Input(
+                    "first_frame",
+                    tooltip="Start frame",
+                ),
+                comfy_io.Image.Input(
+                    "end_frame",
+                    tooltip="End frame",
+                ),
+                comfy_io.String.Input(
+                    "prompt",
+                    multiline=True,
+                    tooltip="A textual description for video generation",
+                    optional=True,
+                ),
+                comfy_io.Int.Input(
+                    "duration",
+                    default=5,
+                    min=5,
+                    max=5,
+                    step=1,
+                    display_mode=comfy_io.NumberDisplay.number,
+                    tooltip="Duration of the output video in seconds",
+                    optional=True,
+                ),
+                comfy_io.Int.Input(
+                    "seed",
+                    default=0,
+                    min=0,
+                    max=2147483647,
+                    step=1,
+                    display_mode=comfy_io.NumberDisplay.number,
+                    control_after_generate=True,
+                    tooltip="Seed for video generation (0 for random)",
+                    optional=True,
+                ),
+                comfy_io.Combo.Input(
+                    "resolution",
+                    options=[model.value for model in Resolution],
+                    default=Resolution.r_1080p.value,
+                    tooltip="Supported values may vary by model & duration",
+                    optional=True,
+                ),
+                comfy_io.Combo.Input(
+                    "movement_amplitude",
+                    options=[model.value for model in MovementAmplitude],
+                    default=MovementAmplitude.auto.value,
+                    tooltip="The movement amplitude of objects in the frame",
+                    optional=True,
+                ),
+            ],
+            outputs=[
+                comfy_io.Video.Output(),
+            ],
+            hidden=[
+                comfy_io.Hidden.auth_token_comfy_org,
+                comfy_io.Hidden.api_key_comfy_org,
+                comfy_io.Hidden.unique_id,
+            ],
+            is_api_node=True,
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        model: str,
+        first_frame: torch.Tensor,
+        end_frame: torch.Tensor,
+        prompt: str,
+        duration: int,
+        seed: int,
+        resolution: str,
+        movement_amplitude: str,
+    ) -> comfy_io.NodeOutput:
+        validate_aspect_ratio_closeness(first_frame, end_frame, min_rel=0.8, max_rel=1.25, strict=False)
+        payload = TaskCreationRequest(
+            model_name=model,
+            prompt=prompt,
+            duration=duration,
+            seed=seed,
+            resolution=resolution,
+            movement_amplitude=movement_amplitude,
+        )
+        auth = {
+            "auth_token": cls.hidden.auth_token_comfy_org,
+            "comfy_api_key": cls.hidden.api_key_comfy_org,
+        }
+        payload.images = [
+            (await upload_images_to_comfyapi(frame, max_images=1, mime_type="image/png", auth_kwargs=auth))[0]
+            for frame in (first_frame, end_frame)
+        ]
+        results = await execute_task(VIDU_START_END_VIDEO, auth, payload, 96, cls.hidden.unique_id)
+        return comfy_io.NodeOutput(await download_url_to_video_output(get_video_from_response(results).url))
+
+
+class ViduExtension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[comfy_io.ComfyNode]]:
+        return [
+            ViduTextToVideoNode,
+            ViduImageToVideoNode,
+            ViduReferenceVideoNode,
+            ViduStartEndToVideoNode,
+        ]
+
+async def comfy_entrypoint() -> ViduExtension:
+    return ViduExtension()

comfy_api_nodes/util/validation_utils.py

@@ -53,6 +53,53 @@ def validate_image_aspect_ratio(
         )
 
 
+def validate_image_aspect_ratio_range(
+    image: torch.Tensor,
+    min_ratio: tuple[float, float],  # e.g. (1, 4)
+    max_ratio: tuple[float, float],  # e.g. (4, 1)
+    *,
+    strict: bool = True,             # True -> (min, max); False -> [min, max]
+) -> float:
+    a1, b1 = min_ratio
+    a2, b2 = max_ratio
+    if a1 <= 0 or b1 <= 0 or a2 <= 0 or b2 <= 0:
+        raise ValueError("Ratios must be positive, like (1, 4) or (4, 1).")
+    lo, hi = (a1 / b1), (a2 / b2)
+    if lo > hi:
+        lo, hi = hi, lo
+        a1, b1, a2, b2 = a2, b2, a1, b1  # swap only for error text
+    w, h = get_image_dimensions(image)
+    if w <= 0 or h <= 0:
+        raise ValueError(f"Invalid image dimensions: {w}x{h}")
+    ar = w / h
+    ok = (lo < ar < hi) if strict else (lo <= ar <= hi)
+    if not ok:
+        op = "<" if strict else "≤"
+        raise ValueError(f"Image aspect ratio {ar:.6g} is outside allowed range: {a1}:{b1} {op} ratio {op} {a2}:{b2}")
+    return ar
+
+
+def validate_aspect_ratio_closeness(
+    start_img,
+    end_img,
+    min_rel: float,
+    max_rel: float,
+    *,
+    strict: bool = False,   # True => exclusive, False => inclusive
+) -> None:
+    w1, h1 = get_image_dimensions(start_img)
+    w2, h2 = get_image_dimensions(end_img)
+    if min(w1, h1, w2, h2) <= 0:
+        raise ValueError("Invalid image dimensions")
+    ar1 = w1 / h1
+    ar2 = w2 / h2
+    # Normalize so it is symmetric (no need to check both ar1/ar2 and ar2/ar1)
+    closeness = max(ar1, ar2) / min(ar1, ar2)
+    limit = max(max_rel, 1.0 / min_rel)  # for 0.8..1.25 this is 1.25
+    if (closeness >= limit) if strict else (closeness > limit):
+        raise ValueError(f"Aspect ratios must be close: start/end={ar1/ar2:.4f}, allowed range {min_rel}–{max_rel}.")
+
+
 def validate_video_dimensions(
     video: VideoInput,
     min_width: Optional[int] = None,
@@ -98,3 +145,9 @@ def validate_video_duration(
         raise ValueError(
             f"Video duration must be at most {max_duration}s, got {duration}s"
         )
+
+
+def get_number_of_images(images):
+    if isinstance(images, torch.Tensor):
+        return images.shape[0] if images.ndim >= 4 else 1
+    return len(images)

comfy_extras/nodes_audio.py

@@ -346,6 +346,24 @@ class LoadAudio:
             return "Invalid audio file: {}".format(audio)
         return True
 
+class RecordAudio:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {"audio": ("AUDIO_RECORD", {})}}
+
+    CATEGORY = "audio"
+
+    RETURN_TYPES = ("AUDIO", )
+    FUNCTION = "load"
+
+    def load(self, audio):
+        audio_path = folder_paths.get_annotated_filepath(audio)
+
+        waveform, sample_rate = torchaudio.load(audio_path)
+        audio = {"waveform": waveform.unsqueeze(0), "sample_rate": sample_rate}
+        return (audio, )
+
+
 NODE_CLASS_MAPPINGS = {
     "EmptyLatentAudio": EmptyLatentAudio,
     "VAEEncodeAudio": VAEEncodeAudio,
@@ -356,6 +374,7 @@ NODE_CLASS_MAPPINGS = {
     "LoadAudio": LoadAudio,
     "PreviewAudio": PreviewAudio,
     "ConditioningStableAudio": ConditioningStableAudio,
+    "RecordAudio": RecordAudio,
 }
 
 NODE_DISPLAY_NAME_MAPPINGS = {
@@ -367,4 +386,5 @@ NODE_DISPLAY_NAME_MAPPINGS = {
     "SaveAudio": "Save Audio (FLAC)",
     "SaveAudioMP3": "Save Audio (MP3)",
     "SaveAudioOpus": "Save Audio (Opus)",
+    "RecordAudio": "Record Audio",
 }

comfy_extras/nodes_context_windows.py

@@ -0,0 +1,89 @@
+from __future__ import annotations
+from comfy_api.latest import ComfyExtension, io
+import comfy.context_windows
+import nodes
+
+
+class ContextWindowsManualNode(io.ComfyNode):
+    @classmethod
+    def define_schema(cls) -> io.Schema:
+        return io.Schema(
+            node_id="ContextWindowsManual",
+            display_name="Context Windows (Manual)",
+            category="context",
+            description="Manually set context windows.",
+            inputs=[
+                io.Model.Input("model", tooltip="The model to apply context windows to during sampling."),
+                io.Int.Input("context_length", min=1, default=16, tooltip="The length of the context window."),
+                io.Int.Input("context_overlap", min=0, default=4, tooltip="The overlap of the context window."),
+                io.Combo.Input("context_schedule", options=[
+                    comfy.context_windows.ContextSchedules.STATIC_STANDARD,
+                    comfy.context_windows.ContextSchedules.UNIFORM_STANDARD,
+                    comfy.context_windows.ContextSchedules.UNIFORM_LOOPED,
+                    comfy.context_windows.ContextSchedules.BATCHED,
+                    ], tooltip="The stride of the context window."),
+                io.Int.Input("context_stride", min=1, default=1, tooltip="The stride of the context window; only applicable to uniform schedules."),
+                io.Boolean.Input("closed_loop", default=False, tooltip="Whether to close the context window loop; only applicable to looped schedules."),
+                io.Combo.Input("fuse_method", options=comfy.context_windows.ContextFuseMethods.LIST_STATIC, default=comfy.context_windows.ContextFuseMethods.PYRAMID, tooltip="The method to use to fuse the context windows."),
+                io.Int.Input("dim", min=0, max=5, default=0, tooltip="The dimension to apply the context windows to."),
+            ],
+            outputs=[
+                io.Model.Output(tooltip="The model with context windows applied during sampling."),
+            ],
+            is_experimental=True,
+        )
+
+    @classmethod
+    def execute(cls, model: io.Model.Type, context_length: int, context_overlap: int, context_schedule: str, context_stride: int, closed_loop: bool, fuse_method: str, dim: int) -> io.Model:
+        model = model.clone()
+        model.model_options["context_handler"] = comfy.context_windows.IndexListContextHandler(
+            context_schedule=comfy.context_windows.get_matching_context_schedule(context_schedule),
+            fuse_method=comfy.context_windows.get_matching_fuse_method(fuse_method),
+            context_length=context_length,
+            context_overlap=context_overlap,
+            context_stride=context_stride,
+            closed_loop=closed_loop,
+            dim=dim)
+        # make memory usage calculation only take into account the context window latents
+        comfy.context_windows.create_prepare_sampling_wrapper(model)
+        return io.NodeOutput(model)
+
+class WanContextWindowsManualNode(ContextWindowsManualNode):
+    @classmethod
+    def define_schema(cls) -> io.Schema:
+        schema = super().define_schema()
+        schema.node_id = "WanContextWindowsManual"
+        schema.display_name = "WAN Context Windows (Manual)"
+        schema.description = "Manually set context windows for WAN-like models (dim=2)."
+        schema.inputs = [
+            io.Model.Input("model", tooltip="The model to apply context windows to during sampling."),
+                io.Int.Input("context_length", min=1, max=nodes.MAX_RESOLUTION, step=4, default=81, tooltip="The length of the context window."),
+                io.Int.Input("context_overlap", min=0, default=30, tooltip="The overlap of the context window."),
+                io.Combo.Input("context_schedule", options=[
+                    comfy.context_windows.ContextSchedules.STATIC_STANDARD,
+                    comfy.context_windows.ContextSchedules.UNIFORM_STANDARD,
+                    comfy.context_windows.ContextSchedules.UNIFORM_LOOPED,
+                    comfy.context_windows.ContextSchedules.BATCHED,
+                    ], tooltip="The stride of the context window."),
+                io.Int.Input("context_stride", min=1, default=1, tooltip="The stride of the context window; only applicable to uniform schedules."),
+                io.Boolean.Input("closed_loop", default=False, tooltip="Whether to close the context window loop; only applicable to looped schedules."),
+                io.Combo.Input("fuse_method", options=comfy.context_windows.ContextFuseMethods.LIST_STATIC, default=comfy.context_windows.ContextFuseMethods.PYRAMID, tooltip="The method to use to fuse the context windows."),
+        ]
+        return schema
+
+    @classmethod
+    def execute(cls, model: io.Model.Type, context_length: int, context_overlap: int, context_schedule: str, context_stride: int, closed_loop: bool, fuse_method: str) -> io.Model:
+        context_length = max(((context_length - 1) // 4) + 1, 1)  # at least length 1
+        context_overlap = max(((context_overlap - 1) // 4) + 1, 0)  # at least overlap 0
+        return super().execute(model, context_length, context_overlap, context_schedule, context_stride, closed_loop, fuse_method, dim=2)
+
+
+class ContextWindowsExtension(ComfyExtension):
+    async def get_node_list(self) -> list[type[io.ComfyNode]]:
+        return [
+            ContextWindowsManualNode,
+            WanContextWindowsManualNode,
+        ]
+
+def comfy_entrypoint():
+    return ContextWindowsExtension()

comfy_extras/nodes_flux.py

@@ -100,9 +100,28 @@ class FluxKontextImageScale:
         return (image, )
 
 
+class FluxKontextMultiReferenceLatentMethod:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {
+            "conditioning": ("CONDITIONING", ),
+            "reference_latents_method": (("offset", "index"), ),
+            }}
+
+    RETURN_TYPES = ("CONDITIONING",)
+    FUNCTION = "append"
+    EXPERIMENTAL = True
+
+    CATEGORY = "advanced/conditioning/flux"
+
+    def append(self, conditioning, reference_latents_method):
+        c = node_helpers.conditioning_set_values(conditioning, {"reference_latents_method": reference_latents_method})
+        return (c, )
+
 NODE_CLASS_MAPPINGS = {
     "CLIPTextEncodeFlux": CLIPTextEncodeFlux,
     "FluxGuidance": FluxGuidance,
     "FluxDisableGuidance": FluxDisableGuidance,
     "FluxKontextImageScale": FluxKontextImageScale,
+    "FluxKontextMultiReferenceLatentMethod": FluxKontextMultiReferenceLatentMethod,
 }

comfy_extras/nodes_qwen.py

@@ -0,0 +1,48 @@
+import node_helpers
+import comfy.utils
+import math
+
+
+class TextEncodeQwenImageEdit:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {
+            "clip": ("CLIP", ),
+            "prompt": ("STRING", {"multiline": True, "dynamicPrompts": True}),
+            },
+            "optional": {"vae": ("VAE", ),
+                         "image": ("IMAGE", ),}}
+
+    RETURN_TYPES = ("CONDITIONING",)
+    FUNCTION = "encode"
+
+    CATEGORY = "advanced/conditioning"
+
+    def encode(self, clip, prompt, vae=None, image=None):
+        ref_latent = None
+        if image is None:
+            images = []
+        else:
+            samples = image.movedim(-1, 1)
+            total = int(1024 * 1024)
+
+            scale_by = math.sqrt(total / (samples.shape[3] * samples.shape[2]))
+            width = round(samples.shape[3] * scale_by)
+            height = round(samples.shape[2] * scale_by)
+
+            s = comfy.utils.common_upscale(samples, width, height, "area", "disabled")
+            image = s.movedim(1, -1)
+            images = [image[:, :, :, :3]]
+            if vae is not None:
+                ref_latent = vae.encode(image[:, :, :, :3])
+
+        tokens = clip.tokenize(prompt, images=images)
+        conditioning = clip.encode_from_tokens_scheduled(tokens)
+        if ref_latent is not None:
+            conditioning = node_helpers.conditioning_set_values(conditioning, {"reference_latents": [ref_latent]}, append=True)
+        return (conditioning, )
+
+
+NODE_CLASS_MAPPINGS = {
+    "TextEncodeQwenImageEdit": TextEncodeQwenImageEdit,
+}

comfy_extras/nodes_wan.py

@@ -9,29 +9,35 @@ import comfy.clip_vision
 import json
 import numpy as np
 from typing import Tuple
+from typing_extensions import override
+from comfy_api.latest import ComfyExtension, io
 
-class WanImageToVideo:
+class WanImageToVideo(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {"required": {"positive": ("CONDITIONING", ),
-                             "negative": ("CONDITIONING", ),
-                             "vae": ("VAE", ),
-                             "width": ("INT", {"default": 832, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
-                             "height": ("INT", {"default": 480, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
-                             "length": ("INT", {"default": 81, "min": 1, "max": nodes.MAX_RESOLUTION, "step": 4}),
-                             "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096}),
-                },
-                "optional": {"clip_vision_output": ("CLIP_VISION_OUTPUT", ),
-                             "start_image": ("IMAGE", ),
-                }}
+    def define_schema(cls):
+        return io.Schema(
+            node_id="WanImageToVideo",
+            category="conditioning/video_models",
+            inputs=[
+                io.Conditioning.Input("positive"),
+                io.Conditioning.Input("negative"),
+                io.Vae.Input("vae"),
+                io.Int.Input("width", default=832, min=16, max=nodes.MAX_RESOLUTION, step=16),
+                io.Int.Input("height", default=480, min=16, max=nodes.MAX_RESOLUTION, step=16),
+                io.Int.Input("length", default=81, min=1, max=nodes.MAX_RESOLUTION, step=4),
+                io.Int.Input("batch_size", default=1, min=1, max=4096),
+                io.ClipVisionOutput.Input("clip_vision_output", optional=True),
+                io.Image.Input("start_image", optional=True),
+            ],
+            outputs=[
+                io.Conditioning.Output(display_name="positive"),
+                io.Conditioning.Output(display_name="negative"),
+                io.Latent.Output(display_name="latent"),
+            ],
+        )
 
-    RETURN_TYPES = ("CONDITIONING", "CONDITIONING", "LATENT")
-    RETURN_NAMES = ("positive", "negative", "latent")
-    FUNCTION = "encode"
-
-    CATEGORY = "conditioning/video_models"
-
-    def encode(self, positive, negative, vae, width, height, length, batch_size, start_image=None, clip_vision_output=None):
+    @classmethod
+    def execute(cls, positive, negative, vae, width, height, length, batch_size, start_image=None, clip_vision_output=None) -> io.NodeOutput:
         latent = torch.zeros([batch_size, 16, ((length - 1) // 4) + 1, height // 8, width // 8], device=comfy.model_management.intermediate_device())
         if start_image is not None:
             start_image = comfy.utils.common_upscale(start_image[:length].movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1)
@@ -51,32 +57,36 @@ class WanImageToVideo:
 
         out_latent = {}
         out_latent["samples"] = latent
-        return (positive, negative, out_latent)
+        return io.NodeOutput(positive, negative, out_latent)
 
 
-class WanFunControlToVideo:
+class WanFunControlToVideo(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {"required": {"positive": ("CONDITIONING", ),
-                             "negative": ("CONDITIONING", ),
-                             "vae": ("VAE", ),
-                             "width": ("INT", {"default": 832, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
-                             "height": ("INT", {"default": 480, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
-                             "length": ("INT", {"default": 81, "min": 1, "max": nodes.MAX_RESOLUTION, "step": 4}),
-                             "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096}),
-                },
-                "optional": {"clip_vision_output": ("CLIP_VISION_OUTPUT", ),
-                             "start_image": ("IMAGE", ),
-                             "control_video": ("IMAGE", ),
-                }}
+    def define_schema(cls):
+        return io.Schema(
+            node_id="WanFunControlToVideo",
+            category="conditioning/video_models",
+            inputs=[
+                io.Conditioning.Input("positive"),
+                io.Conditioning.Input("negative"),
+                io.Vae.Input("vae"),
+                io.Int.Input("width", default=832, min=16, max=nodes.MAX_RESOLUTION, step=16),
+                io.Int.Input("height", default=480, min=16, max=nodes.MAX_RESOLUTION, step=16),
+                io.Int.Input("length", default=81, min=1, max=nodes.MAX_RESOLUTION, step=4),
+                io.Int.Input("batch_size", default=1, min=1, max=4096),
+                io.ClipVisionOutput.Input("clip_vision_output", optional=True),
+                io.Image.Input("start_image", optional=True),
+                io.Image.Input("control_video", optional=True),
+            ],
+            outputs=[
+                io.Conditioning.Output(display_name="positive"),
+                io.Conditioning.Output(display_name="negative"),
+                io.Latent.Output(display_name="latent"),
+            ],
+        )
 
-    RETURN_TYPES = ("CONDITIONING", "CONDITIONING", "LATENT")
-    RETURN_NAMES = ("positive", "negative", "latent")
-    FUNCTION = "encode"
-
-    CATEGORY = "conditioning/video_models"
-
-    def encode(self, positive, negative, vae, width, height, length, batch_size, start_image=None, clip_vision_output=None, control_video=None):
+    @classmethod
+    def execute(cls, positive, negative, vae, width, height, length, batch_size, start_image=None, clip_vision_output=None, control_video=None) -> io.NodeOutput:
         latent = torch.zeros([batch_size, 16, ((length - 1) // 4) + 1, height // 8, width // 8], device=comfy.model_management.intermediate_device())
         concat_latent = torch.zeros([batch_size, 16, ((length - 1) // 4) + 1, height // 8, width // 8], device=comfy.model_management.intermediate_device())
         concat_latent = comfy.latent_formats.Wan21().process_out(concat_latent)
@@ -101,31 +111,34 @@ class WanFunControlToVideo:
 
         out_latent = {}
         out_latent["samples"] = latent
-        return (positive, negative, out_latent)
+        return io.NodeOutput(positive, negative, out_latent)
 
-class Wan22FunControlToVideo:
+class Wan22FunControlToVideo(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {"required": {"positive": ("CONDITIONING", ),
-                             "negative": ("CONDITIONING", ),
-                             "vae": ("VAE", ),
-                             "width": ("INT", {"default": 832, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
-                             "height": ("INT", {"default": 480, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
-                             "length": ("INT", {"default": 81, "min": 1, "max": nodes.MAX_RESOLUTION, "step": 4}),
-                             "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096}),
-                },
-                "optional": {"ref_image": ("IMAGE", ),
-                             "control_video": ("IMAGE", ),
-                             # "start_image": ("IMAGE", ),
-                }}
+    def define_schema(cls):
+        return io.Schema(
+            node_id="Wan22FunControlToVideo",
+            category="conditioning/video_models",
+            inputs=[
+                io.Conditioning.Input("positive"),
+                io.Conditioning.Input("negative"),
+                io.Vae.Input("vae"),
+                io.Int.Input("width", default=832, min=16, max=nodes.MAX_RESOLUTION, step=16),
+                io.Int.Input("height", default=480, min=16, max=nodes.MAX_RESOLUTION, step=16),
+                io.Int.Input("length", default=81, min=1, max=nodes.MAX_RESOLUTION, step=4),
+                io.Int.Input("batch_size", default=1, min=1, max=4096),
+                io.Image.Input("ref_image", optional=True),
+                io.Image.Input("control_video", optional=True),
+            ],
+            outputs=[
+                io.Conditioning.Output(display_name="positive"),
+                io.Conditioning.Output(display_name="negative"),
+                io.Latent.Output(display_name="latent"),
+            ],
+        )
 
-    RETURN_TYPES = ("CONDITIONING", "CONDITIONING", "LATENT")
-    RETURN_NAMES = ("positive", "negative", "latent")
-    FUNCTION = "encode"
-
-    CATEGORY = "conditioning/video_models"
-
-    def encode(self, positive, negative, vae, width, height, length, batch_size, ref_image=None, start_image=None, control_video=None):
+    @classmethod
+    def execute(cls, positive, negative, vae, width, height, length, batch_size, ref_image=None, start_image=None, control_video=None) -> io.NodeOutput:
         latent = torch.zeros([batch_size, 16, ((length - 1) // 4) + 1, height // 8, width // 8], device=comfy.model_management.intermediate_device())
         concat_latent = torch.zeros([batch_size, 16, ((length - 1) // 4) + 1, height // 8, width // 8], device=comfy.model_management.intermediate_device())
         concat_latent = comfy.latent_formats.Wan21().process_out(concat_latent)
@@ -158,32 +171,36 @@ class Wan22FunControlToVideo:
 
         out_latent = {}
         out_latent["samples"] = latent
-        return (positive, negative, out_latent)
+        return io.NodeOutput(positive, negative, out_latent)
 
-class WanFirstLastFrameToVideo:
+class WanFirstLastFrameToVideo(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {"required": {"positive": ("CONDITIONING", ),
-                             "negative": ("CONDITIONING", ),
-                             "vae": ("VAE", ),
-                             "width": ("INT", {"default": 832, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
-                             "height": ("INT", {"default": 480, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
-                             "length": ("INT", {"default": 81, "min": 1, "max": nodes.MAX_RESOLUTION, "step": 4}),
-                             "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096}),
-                },
-                "optional": {"clip_vision_start_image": ("CLIP_VISION_OUTPUT", ),
-                             "clip_vision_end_image": ("CLIP_VISION_OUTPUT", ),
-                             "start_image": ("IMAGE", ),
-                             "end_image": ("IMAGE", ),
-                }}
+    def define_schema(cls):
+        return io.Schema(
+            node_id="WanFirstLastFrameToVideo",
+            category="conditioning/video_models",
+            inputs=[
+                io.Conditioning.Input("positive"),
+                io.Conditioning.Input("negative"),
+                io.Vae.Input("vae"),
+                io.Int.Input("width", default=832, min=16, max=nodes.MAX_RESOLUTION, step=16),
+                io.Int.Input("height", default=480, min=16, max=nodes.MAX_RESOLUTION, step=16),
+                io.Int.Input("length", default=81, min=1, max=nodes.MAX_RESOLUTION, step=4),
+                io.Int.Input("batch_size", default=1, min=1, max=4096),
+                io.ClipVisionOutput.Input("clip_vision_start_image", optional=True),
+                io.ClipVisionOutput.Input("clip_vision_end_image", optional=True),
+                io.Image.Input("start_image", optional=True),
+                io.Image.Input("end_image", optional=True),
+            ],
+            outputs=[
+                io.Conditioning.Output(display_name="positive"),
+                io.Conditioning.Output(display_name="negative"),
+                io.Latent.Output(display_name="latent"),
+            ],
+        )
 
-    RETURN_TYPES = ("CONDITIONING", "CONDITIONING", "LATENT")
-    RETURN_NAMES = ("positive", "negative", "latent")
-    FUNCTION = "encode"
-
-    CATEGORY = "conditioning/video_models"
-
-    def encode(self, positive, negative, vae, width, height, length, batch_size, start_image=None, end_image=None, clip_vision_start_image=None, clip_vision_end_image=None):
+    @classmethod
+    def execute(cls, positive, negative, vae, width, height, length, batch_size, start_image=None, end_image=None, clip_vision_start_image=None, clip_vision_end_image=None) -> io.NodeOutput:
         latent = torch.zeros([batch_size, 16, ((length - 1) // 4) + 1, height // 8, width // 8], device=comfy.model_management.intermediate_device())
         if start_image is not None:
             start_image = comfy.utils.common_upscale(start_image[:length].movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1)
@@ -224,62 +241,70 @@ class WanFirstLastFrameToVideo:
 
         out_latent = {}
         out_latent["samples"] = latent
-        return (positive, negative, out_latent)
+        return io.NodeOutput(positive, negative, out_latent)
 
 
-class WanFunInpaintToVideo:
+class WanFunInpaintToVideo(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {"required": {"positive": ("CONDITIONING", ),
-                             "negative": ("CONDITIONING", ),
-                             "vae": ("VAE", ),
-                             "width": ("INT", {"default": 832, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
-                             "height": ("INT", {"default": 480, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
-                             "length": ("INT", {"default": 81, "min": 1, "max": nodes.MAX_RESOLUTION, "step": 4}),
-                             "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096}),
-                },
-                "optional": {"clip_vision_output": ("CLIP_VISION_OUTPUT", ),
-                             "start_image": ("IMAGE", ),
-                             "end_image": ("IMAGE", ),
-                }}
+    def define_schema(cls):
+        return io.Schema(
+            node_id="WanFunInpaintToVideo",
+            category="conditioning/video_models",
+            inputs=[
+                io.Conditioning.Input("positive"),
+                io.Conditioning.Input("negative"),
+                io.Vae.Input("vae"),
+                io.Int.Input("width", default=832, min=16, max=nodes.MAX_RESOLUTION, step=16),
+                io.Int.Input("height", default=480, min=16, max=nodes.MAX_RESOLUTION, step=16),
+                io.Int.Input("length", default=81, min=1, max=nodes.MAX_RESOLUTION, step=4),
+                io.Int.Input("batch_size", default=1, min=1, max=4096),
+                io.ClipVisionOutput.Input("clip_vision_output", optional=True),
+                io.Image.Input("start_image", optional=True),
+                io.Image.Input("end_image", optional=True),
+            ],
+            outputs=[
+                io.Conditioning.Output(display_name="positive"),
+                io.Conditioning.Output(display_name="negative"),
+                io.Latent.Output(display_name="latent"),
+            ],
+        )
 
-    RETURN_TYPES = ("CONDITIONING", "CONDITIONING", "LATENT")
-    RETURN_NAMES = ("positive", "negative", "latent")
-    FUNCTION = "encode"
-
-    CATEGORY = "conditioning/video_models"
-
-    def encode(self, positive, negative, vae, width, height, length, batch_size, start_image=None, end_image=None, clip_vision_output=None):
+    @classmethod
+    def execute(cls, positive, negative, vae, width, height, length, batch_size, start_image=None, end_image=None, clip_vision_output=None) -> io.NodeOutput:
         flfv = WanFirstLastFrameToVideo()
-        return flfv.encode(positive, negative, vae, width, height, length, batch_size, start_image=start_image, end_image=end_image, clip_vision_start_image=clip_vision_output)
+        return flfv.execute(positive, negative, vae, width, height, length, batch_size, start_image=start_image, end_image=end_image, clip_vision_start_image=clip_vision_output)
 
 
-class WanVaceToVideo:
+class WanVaceToVideo(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {"required": {"positive": ("CONDITIONING", ),
-                             "negative": ("CONDITIONING", ),
-                             "vae": ("VAE", ),
-                             "width": ("INT", {"default": 832, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
-                             "height": ("INT", {"default": 480, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
-                             "length": ("INT", {"default": 81, "min": 1, "max": nodes.MAX_RESOLUTION, "step": 4}),
-                             "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096}),
-                             "strength": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1000.0, "step": 0.01}),
-                },
-                "optional": {"control_video": ("IMAGE", ),
-                             "control_masks": ("MASK", ),
-                             "reference_image": ("IMAGE", ),
-                }}
+    def define_schema(cls):
+        return io.Schema(
+            node_id="WanVaceToVideo",
+            category="conditioning/video_models",
+            is_experimental=True,
+            inputs=[
+                io.Conditioning.Input("positive"),
+                io.Conditioning.Input("negative"),
+                io.Vae.Input("vae"),
+                io.Int.Input("width", default=832, min=16, max=nodes.MAX_RESOLUTION, step=16),
+                io.Int.Input("height", default=480, min=16, max=nodes.MAX_RESOLUTION, step=16),
+                io.Int.Input("length", default=81, min=1, max=nodes.MAX_RESOLUTION, step=4),
+                io.Int.Input("batch_size", default=1, min=1, max=4096),
+                io.Float.Input("strength", default=1.0, min=0.0, max=1000.0, step=0.01),
+                io.Image.Input("control_video", optional=True),
+                io.Mask.Input("control_masks", optional=True),
+                io.Image.Input("reference_image", optional=True),
+            ],
+            outputs=[
+                io.Conditioning.Output(display_name="positive"),
+                io.Conditioning.Output(display_name="negative"),
+                io.Latent.Output(display_name="latent"),
+                io.Int.Output(display_name="trim_latent"),
+            ],
+        )
 
-    RETURN_TYPES = ("CONDITIONING", "CONDITIONING", "LATENT", "INT")
-    RETURN_NAMES = ("positive", "negative", "latent", "trim_latent")
-    FUNCTION = "encode"
-
-    CATEGORY = "conditioning/video_models"
-
-    EXPERIMENTAL = True
-
-    def encode(self, positive, negative, vae, width, height, length, batch_size, strength, control_video=None, control_masks=None, reference_image=None):
+    @classmethod
+    def execute(cls, positive, negative, vae, width, height, length, batch_size, strength, control_video=None, control_masks=None, reference_image=None) -> io.NodeOutput:
         latent_length = ((length - 1) // 4) + 1
         if control_video is not None:
             control_video = comfy.utils.common_upscale(control_video[:length].movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1)
@@ -336,52 +361,59 @@ class WanVaceToVideo:
         latent = torch.zeros([batch_size, 16, latent_length, height // 8, width // 8], device=comfy.model_management.intermediate_device())
         out_latent = {}
         out_latent["samples"] = latent
-        return (positive, negative, out_latent, trim_latent)
+        return io.NodeOutput(positive, negative, out_latent, trim_latent)
 
-class TrimVideoLatent:
+class TrimVideoLatent(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {"required": { "samples": ("LATENT",),
-                              "trim_amount": ("INT", {"default": 0, "min": 0, "max": 99999}),
-                             }}
+    def define_schema(cls):
+        return io.Schema(
+            node_id="TrimVideoLatent",
+            category="latent/video",
+            is_experimental=True,
+            inputs=[
+                io.Latent.Input("samples"),
+                io.Int.Input("trim_amount", default=0, min=0, max=99999),
+            ],
+            outputs=[
+                io.Latent.Output(),
+            ],
+        )
 
-    RETURN_TYPES = ("LATENT",)
-    FUNCTION = "op"
-
-    CATEGORY = "latent/video"
-
-    EXPERIMENTAL = True
-
-    def op(self, samples, trim_amount):
+    @classmethod
+    def execute(cls, samples, trim_amount) -> io.NodeOutput:
         samples_out = samples.copy()
 
         s1 = samples["samples"]
         samples_out["samples"] = s1[:, :, trim_amount:]
-        return (samples_out,)
+        return io.NodeOutput(samples_out)
 
-class WanCameraImageToVideo:
+class WanCameraImageToVideo(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {"required": {"positive": ("CONDITIONING", ),
-                             "negative": ("CONDITIONING", ),
-                             "vae": ("VAE", ),
-                             "width": ("INT", {"default": 832, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
-                             "height": ("INT", {"default": 480, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
-                             "length": ("INT", {"default": 81, "min": 1, "max": nodes.MAX_RESOLUTION, "step": 4}),
-                             "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096}),
-                },
-                "optional": {"clip_vision_output": ("CLIP_VISION_OUTPUT", ),
-                             "start_image": ("IMAGE", ),
-                             "camera_conditions": ("WAN_CAMERA_EMBEDDING", ),
-                }}
+    def define_schema(cls):
+        return io.Schema(
+            node_id="WanCameraImageToVideo",
+            category="conditioning/video_models",
+            inputs=[
+                io.Conditioning.Input("positive"),
+                io.Conditioning.Input("negative"),
+                io.Vae.Input("vae"),
+                io.Int.Input("width", default=832, min=16, max=nodes.MAX_RESOLUTION, step=16),
+                io.Int.Input("height", default=480, min=16, max=nodes.MAX_RESOLUTION, step=16),
+                io.Int.Input("length", default=81, min=1, max=nodes.MAX_RESOLUTION, step=4),
+                io.Int.Input("batch_size", default=1, min=1, max=4096),
+                io.ClipVisionOutput.Input("clip_vision_output", optional=True),
+                io.Image.Input("start_image", optional=True),
+                io.WanCameraEmbedding.Input("camera_conditions", optional=True),
+            ],
+            outputs=[
+                io.Conditioning.Output(display_name="positive"),
+                io.Conditioning.Output(display_name="negative"),
+                io.Latent.Output(display_name="latent"),
+            ],
+        )
 
-    RETURN_TYPES = ("CONDITIONING", "CONDITIONING", "LATENT")
-    RETURN_NAMES = ("positive", "negative", "latent")
-    FUNCTION = "encode"
-
-    CATEGORY = "conditioning/video_models"
-
-    def encode(self, positive, negative, vae, width, height, length, batch_size, start_image=None, clip_vision_output=None, camera_conditions=None):
+    @classmethod
+    def execute(cls, positive, negative, vae, width, height, length, batch_size, start_image=None, clip_vision_output=None, camera_conditions=None) -> io.NodeOutput:
         latent = torch.zeros([batch_size, 16, ((length - 1) // 4) + 1, height // 8, width // 8], device=comfy.model_management.intermediate_device())
         concat_latent = torch.zeros([batch_size, 16, ((length - 1) // 4) + 1, height // 8, width // 8], device=comfy.model_management.intermediate_device())
         concat_latent = comfy.latent_formats.Wan21().process_out(concat_latent)
@@ -390,9 +422,12 @@ class WanCameraImageToVideo:
             start_image = comfy.utils.common_upscale(start_image[:length].movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1)
             concat_latent_image = vae.encode(start_image[:, :, :, :3])
             concat_latent[:,:,:concat_latent_image.shape[2]] = concat_latent_image[:,:,:concat_latent.shape[2]]
+            mask = torch.ones((1, 1, latent.shape[2] * 4, latent.shape[-2], latent.shape[-1]))
+            mask[:, :, :start_image.shape[0] + 3] = 0.0
+            mask = mask.view(1, mask.shape[2] // 4, 4, mask.shape[3], mask.shape[4]).transpose(1, 2)
 
-            positive = node_helpers.conditioning_set_values(positive, {"concat_latent_image": concat_latent})
-            negative = node_helpers.conditioning_set_values(negative, {"concat_latent_image": concat_latent})
+            positive = node_helpers.conditioning_set_values(positive, {"concat_latent_image": concat_latent, "concat_mask": mask})
+            negative = node_helpers.conditioning_set_values(negative, {"concat_latent_image": concat_latent, "concat_mask": mask})
 
         if camera_conditions is not None:
             positive = node_helpers.conditioning_set_values(positive, {'camera_conditions': camera_conditions})
@@ -404,29 +439,34 @@ class WanCameraImageToVideo:
 
         out_latent = {}
         out_latent["samples"] = latent
-        return (positive, negative, out_latent)
+        return io.NodeOutput(positive, negative, out_latent)
 
-class WanPhantomSubjectToVideo:
+class WanPhantomSubjectToVideo(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {"required": {"positive": ("CONDITIONING", ),
-                             "negative": ("CONDITIONING", ),
-                             "vae": ("VAE", ),
-                             "width": ("INT", {"default": 832, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
-                             "height": ("INT", {"default": 480, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
-                             "length": ("INT", {"default": 81, "min": 1, "max": nodes.MAX_RESOLUTION, "step": 4}),
-                             "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096}),
-                },
-                "optional": {"images": ("IMAGE", ),
-                }}
+    def define_schema(cls):
+        return io.Schema(
+            node_id="WanPhantomSubjectToVideo",
+            category="conditioning/video_models",
+            inputs=[
+                io.Conditioning.Input("positive"),
+                io.Conditioning.Input("negative"),
+                io.Vae.Input("vae"),
+                io.Int.Input("width", default=832, min=16, max=nodes.MAX_RESOLUTION, step=16),
+                io.Int.Input("height", default=480, min=16, max=nodes.MAX_RESOLUTION, step=16),
+                io.Int.Input("length", default=81, min=1, max=nodes.MAX_RESOLUTION, step=4),
+                io.Int.Input("batch_size", default=1, min=1, max=4096),
+                io.Image.Input("images", optional=True),
+            ],
+            outputs=[
+                io.Conditioning.Output(display_name="positive"),
+                io.Conditioning.Output(display_name="negative_text"),
+                io.Conditioning.Output(display_name="negative_img_text"),
+                io.Latent.Output(display_name="latent"),
+            ],
+        )
 
-    RETURN_TYPES = ("CONDITIONING", "CONDITIONING", "CONDITIONING", "LATENT")
-    RETURN_NAMES = ("positive", "negative_text", "negative_img_text", "latent")
-    FUNCTION = "encode"
-
-    CATEGORY = "conditioning/video_models"
-
-    def encode(self, positive, negative, vae, width, height, length, batch_size, images):
+    @classmethod
+    def execute(cls, positive, negative, vae, width, height, length, batch_size, images) -> io.NodeOutput:
         latent = torch.zeros([batch_size, 16, ((length - 1) // 4) + 1, height // 8, width // 8], device=comfy.model_management.intermediate_device())
         cond2 = negative
         if images is not None:
@@ -442,7 +482,7 @@ class WanPhantomSubjectToVideo:
 
         out_latent = {}
         out_latent["samples"] = latent
-        return (positive, cond2, negative, out_latent)
+        return io.NodeOutput(positive, cond2, negative, out_latent)
 
 def parse_json_tracks(tracks):
     """Parse JSON track data into a standardized format"""
@@ -655,39 +695,41 @@ def patch_motion(
 
     return out_mask_full, out_feature_full
 
-class WanTrackToVideo:
+class WanTrackToVideo(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {"required": {
-                    "positive": ("CONDITIONING", ),
-                    "negative": ("CONDITIONING", ),
-                    "vae": ("VAE", ),
-                    "tracks": ("STRING", {"multiline": True, "default": "[]"}),
-                    "width": ("INT", {"default": 832, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
-                    "height": ("INT", {"default": 480, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
-                    "length": ("INT", {"default": 81, "min": 1, "max": nodes.MAX_RESOLUTION, "step": 4}),
-                    "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096}),
-                    "temperature": ("FLOAT", {"default": 220.0, "min": 1.0, "max": 1000.0, "step": 0.1}),
-                    "topk": ("INT", {"default": 2, "min": 1, "max": 10}),
-                    "start_image": ("IMAGE", ),
-                },
-                "optional": {
-                    "clip_vision_output": ("CLIP_VISION_OUTPUT", ),
-                }}
+    def define_schema(cls):
+        return io.Schema(
+            node_id="WanTrackToVideo",
+            category="conditioning/video_models",
+            inputs=[
+                io.Conditioning.Input("positive"),
+                io.Conditioning.Input("negative"),
+                io.Vae.Input("vae"),
+                io.String.Input("tracks", multiline=True, default="[]"),
+                io.Int.Input("width", default=832, min=16, max=nodes.MAX_RESOLUTION, step=16),
+                io.Int.Input("height", default=480, min=16, max=nodes.MAX_RESOLUTION, step=16),
+                io.Int.Input("length", default=81, min=1, max=nodes.MAX_RESOLUTION, step=4),
+                io.Int.Input("batch_size", default=1, min=1, max=4096),
+                io.Float.Input("temperature", default=220.0, min=1.0, max=1000.0, step=0.1),
+                io.Int.Input("topk", default=2, min=1, max=10),
+                io.Image.Input("start_image"),
+                io.ClipVisionOutput.Input("clip_vision_output", optional=True),
+            ],
+            outputs=[
+                io.Conditioning.Output(display_name="positive"),
+                io.Conditioning.Output(display_name="negative"),
+                io.Latent.Output(display_name="latent"),
+            ],
+        )
 
-    RETURN_TYPES = ("CONDITIONING", "CONDITIONING", "LATENT")
-    RETURN_NAMES = ("positive", "negative", "latent")
-    FUNCTION = "encode"
-
-    CATEGORY = "conditioning/video_models"
-
-    def encode(self, positive, negative, vae, tracks, width, height, length, batch_size,
-               temperature, topk, start_image=None, clip_vision_output=None):
+    @classmethod
+    def execute(cls, positive, negative, vae, tracks, width, height, length, batch_size,
+               temperature, topk, start_image=None, clip_vision_output=None) -> io.NodeOutput:
 
         tracks_data = parse_json_tracks(tracks)
 
         if not tracks_data:
-            return WanImageToVideo().encode(positive, negative, vae, width, height, length, batch_size, start_image=start_image, clip_vision_output=clip_vision_output)
+            return WanImageToVideo().execute(positive, negative, vae, width, height, length, batch_size, start_image=start_image, clip_vision_output=clip_vision_output)
 
         latent = torch.zeros([batch_size, 16, ((length - 1) // 4) + 1, height // 8, width // 8],
                            device=comfy.model_management.intermediate_device())
@@ -741,34 +783,36 @@ class WanTrackToVideo:
 
         out_latent = {}
         out_latent["samples"] = latent
-        return (positive, negative, out_latent)
+        return io.NodeOutput(positive, negative, out_latent)
 
 
-class Wan22ImageToVideoLatent:
+class Wan22ImageToVideoLatent(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {"required": {"vae": ("VAE", ),
-                             "width": ("INT", {"default": 1280, "min": 32, "max": nodes.MAX_RESOLUTION, "step": 32}),
-                             "height": ("INT", {"default": 704, "min": 32, "max": nodes.MAX_RESOLUTION, "step": 32}),
-                             "length": ("INT", {"default": 49, "min": 1, "max": nodes.MAX_RESOLUTION, "step": 4}),
-                             "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096}),
-                },
-                "optional": {"start_image": ("IMAGE", ),
-                }}
+    def define_schema(cls):
+        return io.Schema(
+            node_id="Wan22ImageToVideoLatent",
+            category="conditioning/inpaint",
+            inputs=[
+                io.Vae.Input("vae"),
+                io.Int.Input("width", default=1280, min=32, max=nodes.MAX_RESOLUTION, step=32),
+                io.Int.Input("height", default=704, min=32, max=nodes.MAX_RESOLUTION, step=32),
+                io.Int.Input("length", default=49, min=1, max=nodes.MAX_RESOLUTION, step=4),
+                io.Int.Input("batch_size", default=1, min=1, max=4096),
+                io.Image.Input("start_image", optional=True),
+            ],
+            outputs=[
+                io.Latent.Output(),
+            ],
+        )
 
-
-    RETURN_TYPES = ("LATENT",)
-    FUNCTION = "encode"
-
-    CATEGORY = "conditioning/inpaint"
-
-    def encode(self, vae, width, height, length, batch_size, start_image=None):
+    @classmethod
+    def execute(cls, vae, width, height, length, batch_size, start_image=None) -> io.NodeOutput:
         latent = torch.zeros([1, 48, ((length - 1) // 4) + 1, height // 16, width // 16], device=comfy.model_management.intermediate_device())
 
         if start_image is None:
             out_latent = {}
             out_latent["samples"] = latent
-            return (out_latent,)
+            return io.NodeOutput(out_latent)
 
         mask = torch.ones([latent.shape[0], 1, ((length - 1) // 4) + 1, latent.shape[-2], latent.shape[-1]], device=comfy.model_management.intermediate_device())
 
@@ -783,19 +827,25 @@ class Wan22ImageToVideoLatent:
         latent = latent_format.process_out(latent) * mask + latent * (1.0 - mask)
         out_latent["samples"] = latent.repeat((batch_size, ) + (1,) * (latent.ndim - 1))
         out_latent["noise_mask"] = mask.repeat((batch_size, ) + (1,) * (mask.ndim - 1))
-        return (out_latent,)
+        return io.NodeOutput(out_latent)
 
 
-NODE_CLASS_MAPPINGS = {
-    "WanTrackToVideo": WanTrackToVideo,
-    "WanImageToVideo": WanImageToVideo,
-    "WanFunControlToVideo": WanFunControlToVideo,
-    "Wan22FunControlToVideo": Wan22FunControlToVideo,
-    "WanFunInpaintToVideo": WanFunInpaintToVideo,
-    "WanFirstLastFrameToVideo": WanFirstLastFrameToVideo,
-    "WanVaceToVideo": WanVaceToVideo,
-    "TrimVideoLatent": TrimVideoLatent,
-    "WanCameraImageToVideo": WanCameraImageToVideo,
-    "WanPhantomSubjectToVideo": WanPhantomSubjectToVideo,
-    "Wan22ImageToVideoLatent": Wan22ImageToVideoLatent,
-}
+class WanExtension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[io.ComfyNode]]:
+        return [
+            WanTrackToVideo,
+            WanImageToVideo,
+            WanFunControlToVideo,
+            Wan22FunControlToVideo,
+            WanFunInpaintToVideo,
+            WanFirstLastFrameToVideo,
+            WanVaceToVideo,
+            TrimVideoLatent,
+            WanCameraImageToVideo,
+            WanPhantomSubjectToVideo,
+            Wan22ImageToVideoLatent,
+        ]
+
+async def comfy_entrypoint() -> WanExtension:
+    return WanExtension()

comfyui_version.py

@@ -1,3 +1,3 @@
 # This file is automatically generated by the build process when version is
 # updated in pyproject.toml.
-__version__ = "0.3.50"
+__version__ = "0.3.51"

nodes.py

@@ -2320,6 +2320,8 @@ async def init_builtin_extra_nodes():
         "nodes_camera_trajectory.py",
         "nodes_edit_model.py",
         "nodes_tcfg.py",
+        "nodes_context_windows.py",
+        "nodes_qwen.py",
     ]
 
     import_failed = []
@@ -2349,6 +2351,7 @@ async def init_builtin_api_nodes():
         "nodes_moonvalley.py",
         "nodes_rodin.py",
         "nodes_gemini.py",
+        "nodes_vidu.py",
     ]
 
     if not await load_custom_node(os.path.join(api_nodes_dir, "canary.py"), module_parent="comfy_api_nodes"):

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "ComfyUI"
-version = "0.3.50"
+version = "0.3.51"
 readme = "README.md"
 license = { file = "LICENSE" }
 requires-python = ">=3.9"

requirements.txt

@@ -1,5 +1,5 @@
-comfyui-frontend-package==1.23.4
-comfyui-workflow-templates==0.1.59
+comfyui-frontend-package==1.25.9
+comfyui-workflow-templates==0.1.62
 comfyui-embedded-docs==0.2.6
 torch
 torchsde
@@ -20,11 +20,11 @@ tqdm
 psutil
 alembic
 SQLAlchemy
+av>=14.2.0
 
 #non essential dependencies:
 kornia>=0.7.1
 spandrel
 soundfile
-av>=14.2.0
 pydantic~=2.0
 pydantic-settings~=2.0