ComfyUI version 0.3.47

* [moonvalley] Update V2V node to match API specification

- Add exact resolution validation for supported resolutions (1920x1080, 1080x1920, 1152x1152, 1536x1152, 1152x1536)
- Change frame count validation from divisible by 32 to 16
- Add MP4 container format validation
- Remove internal parameters (steps, guidance_scale) from V2V inference params
- Update video duration handling to support only 5 seconds (auto-trim if longer)
- Add motion_intensity parameter (0-100) for Motion Transfer control type
- Add get_container_format() method to VideoInput classes

* update negative prompt

.ci/windows_base_files/README_VERY_IMPORTANT.txt

@@ -4,6 +4,9 @@ if you have a NVIDIA gpu:
 
 run_nvidia_gpu.bat
 
+if you want to enable the fast fp16 accumulation (faster for fp16 models with slightly less quality):
+
+run_nvidia_gpu_fast_fp16_accumulation.bat
 
 
 To run it in slow CPU mode:

.github/workflows/check-line-endings.yml

@@ -0,0 +1,40 @@
+name: Check for Windows Line Endings
+
+on:
+  pull_request:
+    branches: ['*'] # Trigger on all pull requests to any branch
+
+jobs:
+  check-line-endings:
+    runs-on: ubuntu-latest
+
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0 # Fetch all history to compare changes
+
+      - name: Check for Windows line endings (CRLF)
+        run: |
+          # Get the list of changed files in the PR
+          CHANGED_FILES=$(git diff --name-only ${{ github.event.pull_request.base.sha }}..${{ github.event.pull_request.head.sha }})
+
+          # Flag to track if CRLF is found
+          CRLF_FOUND=false
+
+          # Loop through each changed file
+          for FILE in $CHANGED_FILES; do
+            # Check if the file exists and is a text file
+            if [ -f "$FILE" ] && file "$FILE" | grep -q "text"; then
+              # Check for CRLF line endings
+              if grep -UP '\r$' "$FILE"; then
+                echo "Error: Windows line endings (CRLF) detected in $FILE"
+                CRLF_FOUND=true
+              fi
+            fi
+          done
+
+          # Exit with error if CRLF was found
+          if [ "$CRLF_FOUND" = true ]; then
+            exit 1
+          fi

.github/workflows/windows_release_nightly_pytorch.yml

@@ -7,7 +7,7 @@ on:
         description: 'cuda version'
         required: true
         type: string
-        default: "128"
+        default: "129"
 
       python_minor:
         description: 'python minor version'
@@ -19,7 +19,7 @@ on:
         description: 'python patch version'
         required: true
         type: string
-        default: "2"
+        default: "5"
 #  push:
 #    branches:
 #      - master
@@ -53,6 +53,8 @@ jobs:
             ls ../temp_wheel_dir
             ./python.exe -s -m pip install --pre ../temp_wheel_dir/*
             sed -i '1i../ComfyUI' ./python3${{ inputs.python_minor }}._pth
+
+            rm ./Lib/site-packages/torch/lib/dnnl.lib #I don't think this is actually used and I need the space
             cd ..
 
             git clone --depth 1 https://github.com/comfyanonymous/taesd

README.md

@@ -55,7 +55,7 @@ See what ComfyUI can do with the [example workflows](https://comfyanonymous.gith
 ## Features
 - Nodes/graph/flowchart interface to experiment and create complex Stable Diffusion workflows without needing to code anything.
 - Image Models
-   - SD1.x, SD2.x,
+   - SD1.x, SD2.x ([unCLIP](https://comfyanonymous.github.io/ComfyUI_examples/unclip/))
    - [SDXL](https://comfyanonymous.github.io/ComfyUI_examples/sdxl/), [SDXL Turbo](https://comfyanonymous.github.io/ComfyUI_examples/sdturbo/)
    - [Stable Cascade](https://comfyanonymous.github.io/ComfyUI_examples/stable_cascade/)
    - [SD3 and SD3.5](https://comfyanonymous.github.io/ComfyUI_examples/sd3/)
@@ -69,6 +69,7 @@ See what ComfyUI can do with the [example workflows](https://comfyanonymous.gith
 - Image Editing Models
    - [Omnigen 2](https://comfyanonymous.github.io/ComfyUI_examples/omnigen/)
    - [Flux Kontext](https://comfyanonymous.github.io/ComfyUI_examples/flux/#flux-kontext-image-editing-model)
+   - [HiDream E1.1](https://comfyanonymous.github.io/ComfyUI_examples/hidream/#hidream-e11)
 - Video Models
    - [Stable Video Diffusion](https://comfyanonymous.github.io/ComfyUI_examples/video/)
    - [Mochi](https://comfyanonymous.github.io/ComfyUI_examples/mochi/)
@@ -76,6 +77,7 @@ See what ComfyUI can do with the [example workflows](https://comfyanonymous.gith
    - [Hunyuan Video](https://comfyanonymous.github.io/ComfyUI_examples/hunyuan_video/)
    - [Nvidia Cosmos](https://comfyanonymous.github.io/ComfyUI_examples/cosmos/) and [Cosmos Predict2](https://comfyanonymous.github.io/ComfyUI_examples/cosmos_predict2/)
    - [Wan 2.1](https://comfyanonymous.github.io/ComfyUI_examples/wan/)
+   - [Wan 2.2](https://comfyanonymous.github.io/ComfyUI_examples/wan22/)
 - Audio Models
    - [Stable Audio](https://comfyanonymous.github.io/ComfyUI_examples/audio/)
    - [ACE Step](https://comfyanonymous.github.io/ComfyUI_examples/audio/)
@@ -83,9 +85,10 @@ See what ComfyUI can do with the [example workflows](https://comfyanonymous.gith
    - [Hunyuan3D 2.0](https://docs.comfy.org/tutorials/3d/hunyuan3D-2)
 - Asynchronous Queue system
 - Many optimizations: Only re-executes the parts of the workflow that changes between executions.
-- Smart memory management: can automatically run models on GPUs with as low as 1GB vram.
+- Smart memory management: can automatically run large models on GPUs with as low as 1GB vram with smart offloading.
 - Works even if you don't have a GPU with: ```--cpu``` (slow)
-- Can load ckpt, safetensors and diffusers models/checkpoints. Standalone VAEs and CLIP models.
+- Can load ckpt and safetensors: All in one checkpoints or standalone diffusion models, VAEs and CLIP models.
+- Safe loading of ckpt, pt, pth, etc.. files.
 - Embeddings/Textual inversion
 - [Loras (regular, locon and loha)](https://comfyanonymous.github.io/ComfyUI_examples/lora/)
 - [Hypernetworks](https://comfyanonymous.github.io/ComfyUI_examples/hypernetworks/)
@@ -96,12 +99,10 @@ See what ComfyUI can do with the [example workflows](https://comfyanonymous.gith
 - [Inpainting](https://comfyanonymous.github.io/ComfyUI_examples/inpaint/) with both regular and inpainting models.
 - [ControlNet and T2I-Adapter](https://comfyanonymous.github.io/ComfyUI_examples/controlnet/)
 - [Upscale Models (ESRGAN, ESRGAN variants, SwinIR, Swin2SR, etc...)](https://comfyanonymous.github.io/ComfyUI_examples/upscale_models/)
-- [unCLIP Models](https://comfyanonymous.github.io/ComfyUI_examples/unclip/)
 - [GLIGEN](https://comfyanonymous.github.io/ComfyUI_examples/gligen/)
 - [Model Merging](https://comfyanonymous.github.io/ComfyUI_examples/model_merging/)
 - [LCM models and Loras](https://comfyanonymous.github.io/ComfyUI_examples/lcm/)
 - Latent previews with [TAESD](#how-to-show-high-quality-previews)
-- Starts up very fast.
 - Works fully offline: core will never download anything unless you want to.
 - Optional API nodes to use paid models from external providers through the online [Comfy API](https://docs.comfy.org/tutorials/api-nodes/overview).
 - [Config file](extra_model_paths.yaml.example) to set the search paths for models.
@@ -178,10 +179,6 @@ If you have trouble extracting it, right click the file -> properties -> unblock
 
 See the [Config file](extra_model_paths.yaml.example) to set the search paths for models. In the standalone windows build you can find this file in the ComfyUI directory. Rename this file to extra_model_paths.yaml and edit it with your favorite text editor.
 
-## Jupyter Notebook
-
-To run it on services like paperspace, kaggle or colab you can use my [Jupyter Notebook](notebooks/comfyui_colab.ipynb)
-
 
 ## [comfy-cli](https://docs.comfy.org/comfy-cli/getting-started)
 
@@ -243,7 +240,7 @@ Nvidia users should install stable pytorch using this command:
 
 This is the command to install pytorch nightly instead which might have performance improvements.
 
-```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128```
+```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu129```
 
 #### Troubleshooting
 
@@ -297,6 +294,13 @@ For models compatible with Cambricon Extension for PyTorch (torch_mlu). Here's a
 2. Next, install the PyTorch(torch_mlu) following the instructions on the [Installation](https://www.cambricon.com/docs/sdk_1.15.0/cambricon_pytorch_1.17.0/user_guide_1.9/index.html)
 3. Launch ComfyUI by running `python main.py`
 
+#### Iluvatar Corex
+
+For models compatible with Iluvatar Extension for PyTorch. Here's a step-by-step guide tailored to your platform and installation method:
+
+1. Install the Iluvatar Corex Toolkit by adhering to the platform-specific instructions on the [Installation](https://support.iluvatar.com/#/DocumentCentre?id=1&nameCenter=2&productId=520117912052801536)
+2. Launch ComfyUI by running `python main.py`
+
 # Running
 
 ```python main.py```

app/frontend_management.py

@@ -29,18 +29,48 @@ def frontend_install_warning_message():
 This error is happening because the ComfyUI frontend is no longer shipped as part of the main repo but as a pip package instead.
 """.strip()
 
+def parse_version(version: str) -> tuple[int, int, int]:
+        return tuple(map(int, version.split(".")))
+
+def is_valid_version(version: str) -> bool:
+    """Validate if a string is a valid semantic version (X.Y.Z format)."""
+    pattern = r"^(\d+)\.(\d+)\.(\d+)$"
+    return bool(re.match(pattern, version))
+
+def get_installed_frontend_version():
+    """Get the currently installed frontend package version."""
+    frontend_version_str = version("comfyui-frontend-package")
+    return frontend_version_str
+
+def get_required_frontend_version():
+    """Get the required frontend version from requirements.txt."""
+    try:
+        with open(requirements_path, "r", encoding="utf-8") as f:
+            for line in f:
+                line = line.strip()
+                if line.startswith("comfyui-frontend-package=="):
+                    version_str = line.split("==")[-1]
+                    if not is_valid_version(version_str):
+                        logging.error(f"Invalid version format in requirements.txt: {version_str}")
+                        return None
+                    return version_str
+            logging.error("comfyui-frontend-package not found in requirements.txt")
+            return None
+    except FileNotFoundError:
+        logging.error("requirements.txt not found. Cannot determine required frontend version.")
+        return None
+    except Exception as e:
+        logging.error(f"Error reading requirements.txt: {e}")
+        return None
 
 def check_frontend_version():
     """Check if the frontend version is up to date."""
 
-    def parse_version(version: str) -> tuple[int, int, int]:
-        return tuple(map(int, version.split(".")))
-
     try:
-        frontend_version_str = version("comfyui-frontend-package")
+        frontend_version_str = get_installed_frontend_version()
         frontend_version = parse_version(frontend_version_str)
-        with open(requirements_path, "r", encoding="utf-8") as f:
-            required_frontend = parse_version(f.readline().split("=")[-1])
+        required_frontend_str = get_required_frontend_version()
+        required_frontend = parse_version(required_frontend_str)
         if frontend_version < required_frontend:
             app.logger.log_startup_warning(
                 f"""
@@ -168,6 +198,11 @@ def download_release_asset_zip(release: Release, destination_path: str) -> None:
 class FrontendManager:
     CUSTOM_FRONTENDS_ROOT = str(Path(__file__).parents[1] / "web_custom_versions")
 
+    @classmethod
+    def get_required_frontend_version(cls) -> str:
+        """Get the required frontend package version."""
+        return get_required_frontend_version()
+
     @classmethod
     def default_frontend_path(cls) -> str:
         try:

comfy/cli_args.py

@@ -49,7 +49,8 @@ parser.add_argument("--temp-directory", type=str, default=None, help="Set the Co
 parser.add_argument("--input-directory", type=str, default=None, help="Set the ComfyUI input directory. Overrides --base-directory.")
 parser.add_argument("--auto-launch", action="store_true", help="Automatically launch ComfyUI in the default browser.")
 parser.add_argument("--disable-auto-launch", action="store_true", help="Disable auto launching the browser.")
-parser.add_argument("--cuda-device", type=int, default=None, metavar="DEVICE_ID", help="Set the id of the cuda device this instance will use.")
+parser.add_argument("--cuda-device", type=int, default=None, metavar="DEVICE_ID", help="Set the id of the cuda device this instance will use. All other devices will not be visible.")
+parser.add_argument("--default-device", type=int, default=None, metavar="DEFAULT_DEVICE_ID", help="Set the id of the default device, all other devices will stay visible.")
 cm_group = parser.add_mutually_exclusive_group()
 cm_group.add_argument("--cuda-malloc", action="store_true", help="Enable cudaMallocAsync (enabled by default for torch 2.0 and up).")
 cm_group.add_argument("--disable-cuda-malloc", action="store_true", help="Disable cudaMallocAsync.")
@@ -144,6 +145,7 @@ class PerformanceFeature(enum.Enum):
 parser.add_argument("--fast", nargs="*", type=PerformanceFeature, help="Enable some untested and potentially quality deteriorating optimizations. --fast with no arguments enables everything. You can pass a list specific optimizations if you only want to enable specific ones. Current valid optimizations: fp16_accumulation fp8_matrix_mult cublas_ops")
 
 parser.add_argument("--mmap-torch-files", action="store_true", help="Use mmap when loading ckpt/pt files.")
+parser.add_argument("--disable-mmap", action="store_true", help="Don't use mmap when loading safetensors.")
 
 parser.add_argument("--dont-print-server", action="store_true", help="Don't print server output.")
 parser.add_argument("--quick-test-for-ci", action="store_true", help="Quick test for CI.")

comfy/gligen.py

@@ -1,55 +1,10 @@
 import math
 import torch
 from torch import nn
-from .ldm.modules.attention import CrossAttention
-from inspect import isfunction
+from .ldm.modules.attention import CrossAttention, FeedForward
 import comfy.ops
 ops = comfy.ops.manual_cast
 
-def exists(val):
-    return val is not None
-
-
-def uniq(arr):
-    return{el: True for el in arr}.keys()
-
-
-def default(val, d):
-    if exists(val):
-        return val
-    return d() if isfunction(d) else d
-
-
-# feedforward
-class GEGLU(nn.Module):
-    def __init__(self, dim_in, dim_out):
-        super().__init__()
-        self.proj = ops.Linear(dim_in, dim_out * 2)
-
-    def forward(self, x):
-        x, gate = self.proj(x).chunk(2, dim=-1)
-        return x * torch.nn.functional.gelu(gate)
-
-
-class FeedForward(nn.Module):
-    def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.):
-        super().__init__()
-        inner_dim = int(dim * mult)
-        dim_out = default(dim_out, dim)
-        project_in = nn.Sequential(
-            ops.Linear(dim, inner_dim),
-            nn.GELU()
-        ) if not glu else GEGLU(dim, inner_dim)
-
-        self.net = nn.Sequential(
-            project_in,
-            nn.Dropout(dropout),
-            ops.Linear(inner_dim, dim_out)
-        )
-
-    def forward(self, x):
-        return self.net(x)
-
 
 class GatedCrossAttentionDense(nn.Module):
     def __init__(self, query_dim, context_dim, n_heads, d_head):

comfy/k_diffusion/sa_solver.py

@@ -0,0 +1,121 @@
+# SA-Solver: Stochastic Adams Solver (NeurIPS 2023, arXiv:2309.05019)
+# Conference: https://proceedings.neurips.cc/paper_files/paper/2023/file/f4a6806490d31216a3ba667eb240c897-Paper-Conference.pdf
+# Codebase ref: https://github.com/scxue/SA-Solver
+
+import math
+from typing import Union, Callable
+import torch
+
+
+def compute_exponential_coeffs(s: torch.Tensor, t: torch.Tensor, solver_order: int, tau_t: float) -> torch.Tensor:
+    """Compute (1 + tau^2) * integral of exp((1 + tau^2) * x) * x^p dx from s to t with exp((1 + tau^2) * t) factored out, using integration by parts.
+
+    Integral of exp((1 + tau^2) * x) * x^p dx
+        = product_terms[p] - (p / (1 + tau^2)) * integral of exp((1 + tau^2) * x) * x^(p-1) dx,
+    with base case p=0 where integral equals product_terms[0].
+
+    where
+        product_terms[p] = x^p * exp((1 + tau^2) * x) / (1 + tau^2).
+
+    Construct a recursive coefficient matrix following the above recursive relation to compute all integral terms up to p = (solver_order - 1).
+    Return coefficients used by the SA-Solver in data prediction mode.
+
+    Args:
+        s: Start time s.
+        t: End time t.
+        solver_order: Current order of the solver.
+        tau_t: Stochastic strength parameter in the SDE.
+
+    Returns:
+        Exponential coefficients used in data prediction, with exp((1 + tau^2) * t) factored out, ordered from p=0 to p=solver_order−1, shape (solver_order,).
+    """
+    tau_mul = 1 + tau_t ** 2
+    h = t - s
+    p = torch.arange(solver_order, dtype=s.dtype, device=s.device)
+
+    # product_terms after factoring out exp((1 + tau^2) * t)
+    # Includes (1 + tau^2) factor from outside the integral
+    product_terms_factored = (t ** p - s ** p * (-tau_mul * h).exp())
+
+    # Lower triangular recursive coefficient matrix
+    # Accumulates recursive coefficients based on p / (1 + tau^2)
+    recursive_depth_mat = p.unsqueeze(1) - p.unsqueeze(0)
+    log_factorial = (p + 1).lgamma()
+    recursive_coeff_mat = log_factorial.unsqueeze(1) - log_factorial.unsqueeze(0)
+    if tau_t > 0:
+        recursive_coeff_mat = recursive_coeff_mat - (recursive_depth_mat * math.log(tau_mul))
+    signs = torch.where(recursive_depth_mat % 2 == 0, 1.0, -1.0)
+    recursive_coeff_mat = (recursive_coeff_mat.exp() * signs).tril()
+
+    return recursive_coeff_mat @ product_terms_factored
+
+
+def compute_simple_stochastic_adams_b_coeffs(sigma_next: torch.Tensor, curr_lambdas: torch.Tensor, lambda_s: torch.Tensor, lambda_t: torch.Tensor, tau_t: float, is_corrector_step: bool = False) -> torch.Tensor:
+    """Compute simple order-2 b coefficients from SA-Solver paper (Appendix D. Implementation Details)."""
+    tau_mul = 1 + tau_t ** 2
+    h = lambda_t - lambda_s
+    alpha_t = sigma_next * lambda_t.exp()
+    if is_corrector_step:
+        # Simplified 1-step (order-2) corrector
+        b_1 = alpha_t * (0.5 * tau_mul * h)
+        b_2 = alpha_t * (-h * tau_mul).expm1().neg() - b_1
+    else:
+        # Simplified 2-step predictor
+        b_2 = alpha_t * (0.5 * tau_mul * h ** 2) / (curr_lambdas[-2] - lambda_s)
+        b_1 = alpha_t * (-h * tau_mul).expm1().neg() - b_2
+    return torch.stack([b_2, b_1])
+
+
+def compute_stochastic_adams_b_coeffs(sigma_next: torch.Tensor, curr_lambdas: torch.Tensor, lambda_s: torch.Tensor, lambda_t: torch.Tensor, tau_t: float, simple_order_2: bool = False, is_corrector_step: bool = False) -> torch.Tensor:
+    """Compute b_i coefficients for the SA-Solver (see eqs. 15 and 18).
+
+    The solver order corresponds to the number of input lambdas (half-logSNR points).
+
+    Args:
+        sigma_next: Sigma at end time t.
+        curr_lambdas: Lambda time points used to construct the Lagrange basis, shape (N,).
+        lambda_s: Lambda at start time s.
+        lambda_t: Lambda at end time t.
+        tau_t: Stochastic strength parameter in the SDE.
+        simple_order_2: Whether to enable the simple order-2 scheme.
+        is_corrector_step: Flag for corrector step in simple order-2 mode.
+
+    Returns:
+        b_i coefficients for the SA-Solver, shape (N,), where N is the solver order.
+    """
+    num_timesteps = curr_lambdas.shape[0]
+
+    if simple_order_2 and num_timesteps == 2:
+        return compute_simple_stochastic_adams_b_coeffs(sigma_next, curr_lambdas, lambda_s, lambda_t, tau_t, is_corrector_step)
+
+    # Compute coefficients by solving a linear system from Lagrange basis interpolation
+    exp_integral_coeffs = compute_exponential_coeffs(lambda_s, lambda_t, num_timesteps, tau_t)
+    vandermonde_matrix_T = torch.vander(curr_lambdas, num_timesteps, increasing=True).T
+    lagrange_integrals = torch.linalg.solve(vandermonde_matrix_T, exp_integral_coeffs)
+
+    # (sigma_t * exp(-tau^2 * lambda_t)) * exp((1 + tau^2) * lambda_t)
+    # = sigma_t * exp(lambda_t) = alpha_t
+    # exp((1 + tau^2) * lambda_t) is extracted from the integral
+    alpha_t = sigma_next * lambda_t.exp()
+    return alpha_t * lagrange_integrals
+
+
+def get_tau_interval_func(start_sigma: float, end_sigma: float, eta: float = 1.0) -> Callable[[Union[torch.Tensor, float]], float]:
+    """Return a function that controls the stochasticity of SA-Solver.
+
+    When eta = 0, SA-Solver runs as ODE. The official approach uses
+    time t to determine the SDE interval, while here we use sigma instead.
+
+    See:
+        https://github.com/scxue/SA-Solver/blob/main/README.md
+    """
+
+    def tau_func(sigma: Union[torch.Tensor, float]) -> float:
+        if eta <= 0:
+            return 0.0  # ODE
+
+        if isinstance(sigma, torch.Tensor):
+            sigma = sigma.item()
+        return eta if start_sigma >= sigma >= end_sigma else 0.0
+
+    return tau_func

comfy/k_diffusion/sampling.py

@@ -9,6 +9,7 @@ from tqdm.auto import trange, tqdm
 
 from . import utils
 from . import deis
+from . import sa_solver
 import comfy.model_patcher
 import comfy.model_sampling
 
@@ -412,9 +413,13 @@ def sample_lms(model, x, sigmas, extra_args=None, callback=None, disable=None, o
             ds.pop(0)
         if callback is not None:
             callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
-        cur_order = min(i + 1, order)
-        coeffs = [linear_multistep_coeff(cur_order, sigmas_cpu, i, j) for j in range(cur_order)]
-        x = x + sum(coeff * d for coeff, d in zip(coeffs, reversed(ds)))
+        if sigmas[i + 1] == 0:
+            # Denoising step
+            x = denoised
+        else:
+            cur_order = min(i + 1, order)
+            coeffs = [linear_multistep_coeff(cur_order, sigmas_cpu, i, j) for j in range(cur_order)]
+            x = x + sum(coeff * d for coeff, d in zip(coeffs, reversed(ds)))
     return x
 
 
@@ -1067,7 +1072,9 @@ def sample_ipndm(model, x, sigmas, extra_args=None, callback=None, disable=None,
         d_cur = (x_cur - denoised) / t_cur
 
         order = min(max_order, i+1)
-        if order == 1:      # First Euler step.
+        if t_next == 0:     # Denoising step
+            x_next = denoised
+        elif order == 1:    # First Euler step.
             x_next = x_cur + (t_next - t_cur) * d_cur
         elif order == 2:    # Use one history point.
             x_next = x_cur + (t_next - t_cur) * (3 * d_cur - buffer_model[-1]) / 2
@@ -1085,6 +1092,7 @@ def sample_ipndm(model, x, sigmas, extra_args=None, callback=None, disable=None,
 
     return x_next
 
+
 #From https://github.com/zju-pi/diff-sampler/blob/main/diff-solvers-main/solvers.py
 #under Apache 2 license
 def sample_ipndm_v(model, x, sigmas, extra_args=None, callback=None, disable=None, max_order=4):
@@ -1108,7 +1116,9 @@ def sample_ipndm_v(model, x, sigmas, extra_args=None, callback=None, disable=Non
         d_cur = (x_cur - denoised) / t_cur
 
         order = min(max_order, i+1)
-        if order == 1:      # First Euler step.
+        if t_next == 0:     # Denoising step
+            x_next = denoised
+        elif order == 1:    # First Euler step.
             x_next = x_cur + (t_next - t_cur) * d_cur
         elif order == 2:    # Use one history point.
             h_n = (t_next - t_cur)
@@ -1148,6 +1158,7 @@ def sample_ipndm_v(model, x, sigmas, extra_args=None, callback=None, disable=Non
 
     return x_next
 
+
 #From https://github.com/zju-pi/diff-sampler/blob/main/diff-solvers-main/solvers.py
 #under Apache 2 license
 @torch.no_grad()
@@ -1198,39 +1209,22 @@ def sample_deis(model, x, sigmas, extra_args=None, callback=None, disable=None,
 
     return x_next
 
-@torch.no_grad()
-def sample_euler_cfg_pp(model, x, sigmas, extra_args=None, callback=None, disable=None):
-    extra_args = {} if extra_args is None else extra_args
-
-    temp = [0]
-    def post_cfg_function(args):
-        temp[0] = args["uncond_denoised"]
-        return args["denoised"]
-
-    model_options = extra_args.get("model_options", {}).copy()
-    extra_args["model_options"] = comfy.model_patcher.set_model_options_post_cfg_function(model_options, post_cfg_function, disable_cfg1_optimization=True)
-
-    s_in = x.new_ones([x.shape[0]])
-    for i in trange(len(sigmas) - 1, disable=disable):
-        sigma_hat = sigmas[i]
-        denoised = model(x, sigma_hat * s_in, **extra_args)
-        d = to_d(x, sigma_hat, temp[0])
-        if callback is not None:
-            callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigma_hat, 'denoised': denoised})
-        # Euler method
-        x = denoised + d * sigmas[i + 1]
-    return x
 
 @torch.no_grad()
 def sample_euler_ancestral_cfg_pp(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
-    """Ancestral sampling with Euler method steps."""
+    """Ancestral sampling with Euler method steps (CFG++)."""
     extra_args = {} if extra_args is None else extra_args
     seed = extra_args.get("seed", None)
     noise_sampler = default_noise_sampler(x, seed=seed) if noise_sampler is None else noise_sampler
 
-    temp = [0]
+    model_sampling = model.inner_model.model_patcher.get_model_object("model_sampling")
+    lambda_fn = partial(sigma_to_half_log_snr, model_sampling=model_sampling)
+
+    uncond_denoised = None
+
     def post_cfg_function(args):
-        temp[0] = args["uncond_denoised"]
+        nonlocal uncond_denoised
+        uncond_denoised = args["uncond_denoised"]
         return args["denoised"]
 
     model_options = extra_args.get("model_options", {}).copy()
@@ -1239,15 +1233,33 @@ def sample_euler_ancestral_cfg_pp(model, x, sigmas, extra_args=None, callback=No
     s_in = x.new_ones([x.shape[0]])
     for i in trange(len(sigmas) - 1, disable=disable):
         denoised = model(x, sigmas[i] * s_in, **extra_args)
-        sigma_down, sigma_up = get_ancestral_step(sigmas[i], sigmas[i + 1], eta=eta)
         if callback is not None:
             callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
-        d = to_d(x, sigmas[i], temp[0])
-        # Euler method
-        x = denoised + d * sigma_down
-        if sigmas[i + 1] > 0:
-            x = x + noise_sampler(sigmas[i], sigmas[i + 1]) * s_noise * sigma_up
+        if sigmas[i + 1] == 0:
+            # Denoising step
+            x = denoised
+        else:
+            alpha_s = sigmas[i] * lambda_fn(sigmas[i]).exp()
+            alpha_t = sigmas[i + 1] * lambda_fn(sigmas[i + 1]).exp()
+            d = to_d(x, sigmas[i], alpha_s * uncond_denoised)   # to noise
+
+            # DDIM stochastic sampling
+            sigma_down, sigma_up = get_ancestral_step(sigmas[i] / alpha_s, sigmas[i + 1] / alpha_t, eta=eta)
+            sigma_down = alpha_t * sigma_down
+
+            # Euler method
+            x = alpha_t * denoised + sigma_down * d
+            if eta > 0 and s_noise > 0:
+                x = x + alpha_t * noise_sampler(sigmas[i], sigmas[i + 1]) * s_noise * sigma_up
     return x
+
+
+@torch.no_grad()
+def sample_euler_cfg_pp(model, x, sigmas, extra_args=None, callback=None, disable=None):
+    """Euler method steps (CFG++)."""
+    return sample_euler_ancestral_cfg_pp(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=0.0, s_noise=0.0, noise_sampler=None)
+
+
 @torch.no_grad()
 def sample_dpmpp_2s_ancestral_cfg_pp(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
     """Ancestral sampling with DPM-Solver++(2S) second-order steps."""
@@ -1404,6 +1416,7 @@ def sample_res_multistep_ancestral(model, x, sigmas, extra_args=None, callback=N
 def sample_res_multistep_ancestral_cfg_pp(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
     return res_multistep(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, s_noise=s_noise, noise_sampler=noise_sampler, eta=eta, cfg_pp=True)
 
+
 @torch.no_grad()
 def sample_gradient_estimation(model, x, sigmas, extra_args=None, callback=None, disable=None, ge_gamma=2., cfg_pp=False):
     """Gradient-estimation sampler. Paper: https://openreview.net/pdf?id=o2ND9v0CeK"""
@@ -1430,19 +1443,19 @@ def sample_gradient_estimation(model, x, sigmas, extra_args=None, callback=None,
         if callback is not None:
             callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
         dt = sigmas[i + 1] - sigmas[i]
-        if i == 0:
+        if sigmas[i + 1] == 0:
+            # Denoising step
+            x = denoised
+        else:
             # Euler method
             if cfg_pp:
                 x = denoised + d * sigmas[i + 1]
             else:
                 x = x + d * dt
-        else:
-            # Gradient estimation
-            if cfg_pp:
+
+            if i >= 1:
+                # Gradient estimation
                 d_bar = (ge_gamma - 1) * (d - old_d)
-                x = denoised + d * sigmas[i + 1] + d_bar * dt
-            else:
-                d_bar = ge_gamma * d + (1 - ge_gamma) * old_d
                 x = x + d_bar * dt
         old_d = d
     return x
@@ -1636,3 +1649,113 @@ def sample_seeds_3(model, x, sigmas, extra_args=None, callback=None, disable=Non
             if inject_noise:
                 x = x + sigmas[i + 1] * (noise_coeff_3 * noise_1 + noise_coeff_2 * noise_2 + noise_coeff_1 * noise_3) * s_noise
     return x
+
+
+@torch.no_grad()
+def sample_sa_solver(model, x, sigmas, extra_args=None, callback=None, disable=False, tau_func=None, s_noise=1.0, noise_sampler=None, predictor_order=3, corrector_order=4, use_pece=False, simple_order_2=False):
+    """Stochastic Adams Solver with predictor-corrector method (NeurIPS 2023)."""
+    if len(sigmas) <= 1:
+        return x
+    extra_args = {} if extra_args is None else extra_args
+    seed = extra_args.get("seed", None)
+    noise_sampler = default_noise_sampler(x, seed=seed) if noise_sampler is None else noise_sampler
+    s_in = x.new_ones([x.shape[0]])
+
+    model_sampling = model.inner_model.model_patcher.get_model_object("model_sampling")
+    sigmas = offset_first_sigma_for_snr(sigmas, model_sampling)
+    lambdas = sigma_to_half_log_snr(sigmas, model_sampling=model_sampling)
+
+    if tau_func is None:
+        # Use default interval for stochastic sampling
+        start_sigma = model_sampling.percent_to_sigma(0.2)
+        end_sigma = model_sampling.percent_to_sigma(0.8)
+        tau_func = sa_solver.get_tau_interval_func(start_sigma, end_sigma, eta=1.0)
+
+    max_used_order = max(predictor_order, corrector_order)
+    x_pred = x  # x: current state, x_pred: predicted next state
+
+    h = 0.0
+    tau_t = 0.0
+    noise = 0.0
+    pred_list = []
+
+    # Lower order near the end to improve stability
+    lower_order_to_end = sigmas[-1].item() == 0
+
+    for i in trange(len(sigmas) - 1, disable=disable):
+        # Evaluation
+        denoised = model(x_pred, sigmas[i] * s_in, **extra_args)
+        if callback is not None:
+            callback({"x": x_pred, "i": i, "sigma": sigmas[i], "sigma_hat": sigmas[i], "denoised": denoised})
+        pred_list.append(denoised)
+        pred_list = pred_list[-max_used_order:]
+
+        predictor_order_used = min(predictor_order, len(pred_list))
+        if i == 0 or (sigmas[i + 1] == 0 and not use_pece):
+            corrector_order_used = 0
+        else:
+            corrector_order_used = min(corrector_order, len(pred_list))
+
+        if lower_order_to_end:
+            predictor_order_used = min(predictor_order_used, len(sigmas) - 2 - i)
+            corrector_order_used = min(corrector_order_used, len(sigmas) - 1 - i)
+
+        # Corrector
+        if corrector_order_used == 0:
+            # Update by the predicted state
+            x = x_pred
+        else:
+            curr_lambdas = lambdas[i - corrector_order_used + 1:i + 1]
+            b_coeffs = sa_solver.compute_stochastic_adams_b_coeffs(
+                sigmas[i],
+                curr_lambdas,
+                lambdas[i - 1],
+                lambdas[i],
+                tau_t,
+                simple_order_2,
+                is_corrector_step=True,
+            )
+            pred_mat = torch.stack(pred_list[-corrector_order_used:], dim=1)    # (B, K, ...)
+            corr_res = torch.tensordot(pred_mat, b_coeffs, dims=([1], [0]))  # (B, ...)
+            x = sigmas[i] / sigmas[i - 1] * (-(tau_t ** 2) * h).exp() * x + corr_res
+
+            if tau_t > 0 and s_noise > 0:
+                # The noise from the previous predictor step
+                x = x + noise
+
+            if use_pece:
+                # Evaluate the corrected state
+                denoised = model(x, sigmas[i] * s_in, **extra_args)
+                pred_list[-1] = denoised
+
+        # Predictor
+        if sigmas[i + 1] == 0:
+            # Denoising step
+            x = denoised
+        else:
+            tau_t = tau_func(sigmas[i + 1])
+            curr_lambdas = lambdas[i - predictor_order_used + 1:i + 1]
+            b_coeffs = sa_solver.compute_stochastic_adams_b_coeffs(
+                sigmas[i + 1],
+                curr_lambdas,
+                lambdas[i],
+                lambdas[i + 1],
+                tau_t,
+                simple_order_2,
+                is_corrector_step=False,
+            )
+            pred_mat = torch.stack(pred_list[-predictor_order_used:], dim=1)    # (B, K, ...)
+            pred_res = torch.tensordot(pred_mat, b_coeffs, dims=([1], [0]))  # (B, ...)
+            h = lambdas[i + 1] - lambdas[i]
+            x_pred = sigmas[i + 1] / sigmas[i] * (-(tau_t ** 2) * h).exp() * x + pred_res
+
+            if tau_t > 0 and s_noise > 0:
+                noise = noise_sampler(sigmas[i], sigmas[i + 1]) * sigmas[i + 1] * (-2 * tau_t ** 2 * h).expm1().neg().sqrt() * s_noise
+                x_pred = x_pred + noise
+    return x
+
+
+@torch.no_grad()
+def sample_sa_solver_pece(model, x, sigmas, extra_args=None, callback=None, disable=False, tau_func=None, s_noise=1.0, noise_sampler=None, predictor_order=3, corrector_order=4, simple_order_2=False):
+    """Stochastic Adams Solver with PECE (Predict–Evaluate–Correct–Evaluate) mode (NeurIPS 2023)."""
+    return sample_sa_solver(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, tau_func=tau_func, s_noise=s_noise, noise_sampler=noise_sampler, predictor_order=predictor_order, corrector_order=corrector_order, use_pece=True, simple_order_2=simple_order_2)

comfy/latent_formats.py

@@ -457,6 +457,82 @@ class Wan21(LatentFormat):
         latents_std = self.latents_std.to(latent.device, latent.dtype)
         return latent * latents_std / self.scale_factor + latents_mean
 
+class Wan22(Wan21):
+    latent_channels = 48
+    latent_dimensions = 3
+
+    latent_rgb_factors = [
+            [ 0.0119,  0.0103,  0.0046],
+            [-0.1062, -0.0504,  0.0165],
+            [ 0.0140,  0.0409,  0.0491],
+            [-0.0813, -0.0677,  0.0607],
+            [ 0.0656,  0.0851,  0.0808],
+            [ 0.0264,  0.0463,  0.0912],
+            [ 0.0295,  0.0326,  0.0590],
+            [-0.0244, -0.0270,  0.0025],
+            [ 0.0443, -0.0102,  0.0288],
+            [-0.0465, -0.0090, -0.0205],
+            [ 0.0359,  0.0236,  0.0082],
+            [-0.0776,  0.0854,  0.1048],
+            [ 0.0564,  0.0264,  0.0561],
+            [ 0.0006,  0.0594,  0.0418],
+            [-0.0319, -0.0542, -0.0637],
+            [-0.0268,  0.0024,  0.0260],
+            [ 0.0539,  0.0265,  0.0358],
+            [-0.0359, -0.0312, -0.0287],
+            [-0.0285, -0.1032, -0.1237],
+            [ 0.1041,  0.0537,  0.0622],
+            [-0.0086, -0.0374, -0.0051],
+            [ 0.0390,  0.0670,  0.2863],
+            [ 0.0069,  0.0144,  0.0082],
+            [ 0.0006, -0.0167,  0.0079],
+            [ 0.0313, -0.0574, -0.0232],
+            [-0.1454, -0.0902, -0.0481],
+            [ 0.0714,  0.0827,  0.0447],
+            [-0.0304, -0.0574, -0.0196],
+            [ 0.0401,  0.0384,  0.0204],
+            [-0.0758, -0.0297, -0.0014],
+            [ 0.0568,  0.1307,  0.1372],
+            [-0.0055, -0.0310, -0.0380],
+            [ 0.0239, -0.0305,  0.0325],
+            [-0.0663, -0.0673, -0.0140],
+            [-0.0416, -0.0047, -0.0023],
+            [ 0.0166,  0.0112, -0.0093],
+            [-0.0211,  0.0011,  0.0331],
+            [ 0.1833,  0.1466,  0.2250],
+            [-0.0368,  0.0370,  0.0295],
+            [-0.3441, -0.3543, -0.2008],
+            [-0.0479, -0.0489, -0.0420],
+            [-0.0660, -0.0153,  0.0800],
+            [-0.0101,  0.0068,  0.0156],
+            [-0.0690, -0.0452, -0.0927],
+            [-0.0145,  0.0041,  0.0015],
+            [ 0.0421,  0.0451,  0.0373],
+            [ 0.0504, -0.0483, -0.0356],
+            [-0.0837,  0.0168,  0.0055]
+        ]
+
+    latent_rgb_factors_bias = [0.0317, -0.0878, -0.1388]
+
+    def __init__(self):
+        self.scale_factor = 1.0
+        self.latents_mean = torch.tensor([
+                -0.2289, -0.0052, -0.1323, -0.2339, -0.2799, 0.0174, 0.1838, 0.1557,
+                -0.1382, 0.0542, 0.2813, 0.0891, 0.1570, -0.0098, 0.0375, -0.1825,
+                -0.2246, -0.1207, -0.0698, 0.5109, 0.2665, -0.2108, -0.2158, 0.2502,
+                -0.2055, -0.0322, 0.1109, 0.1567, -0.0729, 0.0899, -0.2799, -0.1230,
+                -0.0313, -0.1649, 0.0117, 0.0723, -0.2839, -0.2083, -0.0520, 0.3748,
+                0.0152, 0.1957, 0.1433, -0.2944, 0.3573, -0.0548, -0.1681, -0.0667,
+            ]).view(1, self.latent_channels, 1, 1, 1)
+        self.latents_std = torch.tensor([
+                0.4765, 1.0364, 0.4514, 1.1677, 0.5313, 0.4990, 0.4818, 0.5013,
+                0.8158, 1.0344, 0.5894, 1.0901, 0.6885, 0.6165, 0.8454, 0.4978,
+                0.5759, 0.3523, 0.7135, 0.6804, 0.5833, 1.4146, 0.8986, 0.5659,
+                0.7069, 0.5338, 0.4889, 0.4917, 0.4069, 0.4999, 0.6866, 0.4093,
+                0.5709, 0.6065, 0.6415, 0.4944, 0.5726, 1.2042, 0.5458, 1.6887,
+                0.3971, 1.0600, 0.3943, 0.5537, 0.5444, 0.4089, 0.7468, 0.7744
+            ]).view(1, self.latent_channels, 1, 1, 1)
+
 class Hunyuan3Dv2(LatentFormat):
     latent_channels = 64
     latent_dimensions = 1

comfy/ldm/chroma/model.py

@@ -254,13 +254,12 @@ class Chroma(nn.Module):
 
     def forward(self, x, timestep, context, guidance, control=None, transformer_options={}, **kwargs):
         bs, c, h, w = x.shape
-        patch_size = 2
-        x = comfy.ldm.common_dit.pad_to_patch_size(x, (patch_size, patch_size))
+        x = comfy.ldm.common_dit.pad_to_patch_size(x, (self.patch_size, self.patch_size))
 
-        img = rearrange(x, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=patch_size, pw=patch_size)
+        img = rearrange(x, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=self.patch_size, pw=self.patch_size)
 
-        h_len = ((h + (patch_size // 2)) // patch_size)
-        w_len = ((w + (patch_size // 2)) // patch_size)
+        h_len = ((h + (self.patch_size // 2)) // self.patch_size)
+        w_len = ((w + (self.patch_size // 2)) // self.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)
@@ -268,4 +267,4 @@ class Chroma(nn.Module):
 
         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, guidance, control, transformer_options, attn_mask=kwargs.get("attention_mask", None))
-        return rearrange(out, "b (h w) (c ph pw) -> b c (h ph) (w pw)", h=h_len, w=w_len, ph=2, pw=2)[:,:,:h,:w]
+        return rearrange(out, "b (h w) (c ph pw) -> b c (h ph) (w pw)", h=h_len, w=w_len, ph=self.patch_size, pw=self.patch_size)[:,:,:h,:w]

comfy/ldm/lightricks/vae/causal_video_autoencoder.py

@@ -973,7 +973,7 @@ class VideoVAE(nn.Module):
             norm_layer=config.get("norm_layer", "group_norm"),
             causal=config.get("causal_decoder", False),
             timestep_conditioning=self.timestep_conditioning,
-            spatial_padding_mode=config.get("spatial_padding_mode", "zeros"),
+            spatial_padding_mode=config.get("spatial_padding_mode", "reflect"),
         )
 
         self.per_channel_statistics = processor()

comfy/ldm/pixart/pixartms.py

@@ -1,256 +1,256 @@
-# Based on:
-# https://github.com/PixArt-alpha/PixArt-alpha [Apache 2.0 license]
-# https://github.com/PixArt-alpha/PixArt-sigma [Apache 2.0 license]
-import torch
-import torch.nn as nn
-
-from .blocks import (
-    t2i_modulate,
-    CaptionEmbedder,
-    AttentionKVCompress,
-    MultiHeadCrossAttention,
-    T2IFinalLayer,
-    SizeEmbedder,
-)
-from comfy.ldm.modules.diffusionmodules.mmdit import TimestepEmbedder, PatchEmbed, Mlp, get_1d_sincos_pos_embed_from_grid_torch
-
-
-def get_2d_sincos_pos_embed_torch(embed_dim, w, h, pe_interpolation=1.0, base_size=16, device=None, dtype=torch.float32):
-    grid_h, grid_w = torch.meshgrid(
-        torch.arange(h, device=device, dtype=dtype) / (h/base_size) / pe_interpolation,
-        torch.arange(w, device=device, dtype=dtype) / (w/base_size) / pe_interpolation,
-        indexing='ij'
-    )
-    emb_h = get_1d_sincos_pos_embed_from_grid_torch(embed_dim // 2, grid_h, device=device, dtype=dtype)
-    emb_w = get_1d_sincos_pos_embed_from_grid_torch(embed_dim // 2, grid_w, device=device, dtype=dtype)
-    emb = torch.cat([emb_w, emb_h], dim=1)  # (H*W, D)
-    return emb
-
-class PixArtMSBlock(nn.Module):
-    """
-    A PixArt block with adaptive layer norm zero (adaLN-Zero) conditioning.
-    """
-    def __init__(self, hidden_size, num_heads, mlp_ratio=4.0, drop_path=0., input_size=None,
-                 sampling=None, sr_ratio=1, qk_norm=False, dtype=None, device=None, operations=None, **block_kwargs):
-        super().__init__()
-        self.hidden_size = hidden_size
-        self.norm1 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
-        self.attn = AttentionKVCompress(
-            hidden_size, num_heads=num_heads, qkv_bias=True, sampling=sampling, sr_ratio=sr_ratio,
-            qk_norm=qk_norm, dtype=dtype, device=device, operations=operations, **block_kwargs
-        )
-        self.cross_attn = MultiHeadCrossAttention(
-            hidden_size, num_heads, dtype=dtype, device=device, operations=operations, **block_kwargs
-        )
-        self.norm2 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
-        # to be compatible with lower version pytorch
-        approx_gelu = lambda: nn.GELU(approximate="tanh")
-        self.mlp = Mlp(
-            in_features=hidden_size, hidden_features=int(hidden_size * mlp_ratio), act_layer=approx_gelu,
-            dtype=dtype, device=device, operations=operations
-        )
-        self.scale_shift_table = nn.Parameter(torch.randn(6, hidden_size) / hidden_size ** 0.5)
-
-    def forward(self, x, y, t, mask=None, HW=None, **kwargs):
-        B, N, C = x.shape
-
-        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (self.scale_shift_table[None].to(dtype=x.dtype, device=x.device) + t.reshape(B, 6, -1)).chunk(6, dim=1)
-        x = x + (gate_msa * self.attn(t2i_modulate(self.norm1(x), shift_msa, scale_msa), HW=HW))
-        x = x + self.cross_attn(x, y, mask)
-        x = x + (gate_mlp * self.mlp(t2i_modulate(self.norm2(x), shift_mlp, scale_mlp)))
-
-        return x
-
-
-### Core PixArt Model ###
-class PixArtMS(nn.Module):
-    """
-    Diffusion model with a Transformer backbone.
-    """
-    def __init__(
-            self,
-            input_size=32,
-            patch_size=2,
-            in_channels=4,
-            hidden_size=1152,
-            depth=28,
-            num_heads=16,
-            mlp_ratio=4.0,
-            class_dropout_prob=0.1,
-            learn_sigma=True,
-            pred_sigma=True,
-            drop_path: float = 0.,
-            caption_channels=4096,
-            pe_interpolation=None,
-            pe_precision=None,
-            config=None,
-            model_max_length=120,
-            micro_condition=True,
-            qk_norm=False,
-            kv_compress_config=None,
-            dtype=None,
-            device=None,
-            operations=None,
-            **kwargs,
-    ):
-        nn.Module.__init__(self)
-        self.dtype = dtype
-        self.pred_sigma = pred_sigma
-        self.in_channels = in_channels
-        self.out_channels = in_channels * 2 if pred_sigma else in_channels
-        self.patch_size = patch_size
-        self.num_heads = num_heads
-        self.pe_interpolation = pe_interpolation
-        self.pe_precision = pe_precision
-        self.hidden_size = hidden_size
-        self.depth = depth
-
-        approx_gelu = lambda: nn.GELU(approximate="tanh")
-        self.t_block = nn.Sequential(
-            nn.SiLU(),
-            operations.Linear(hidden_size, 6 * hidden_size, bias=True, dtype=dtype, device=device)
-        )
-        self.x_embedder = PatchEmbed(
-            patch_size=patch_size,
-            in_chans=in_channels,
-            embed_dim=hidden_size,
-            bias=True,
-            dtype=dtype,
-            device=device,
-            operations=operations
-        )
-        self.t_embedder = TimestepEmbedder(
-            hidden_size, dtype=dtype, device=device, operations=operations,
-        )
-        self.y_embedder = CaptionEmbedder(
-            in_channels=caption_channels, hidden_size=hidden_size, uncond_prob=class_dropout_prob,
-            act_layer=approx_gelu, token_num=model_max_length,
-            dtype=dtype, device=device, operations=operations,
-        )
-
-        self.micro_conditioning = micro_condition
-        if self.micro_conditioning:
-            self.csize_embedder = SizeEmbedder(hidden_size//3, dtype=dtype, device=device, operations=operations)
-            self.ar_embedder = SizeEmbedder(hidden_size//3, dtype=dtype, device=device, operations=operations)
-
-        # For fixed sin-cos embedding:
-        # num_patches = (input_size // patch_size) * (input_size // patch_size)
-        # self.base_size = input_size // self.patch_size
-        # self.register_buffer("pos_embed", torch.zeros(1, num_patches, hidden_size))
-
-        drop_path = [x.item() for x in torch.linspace(0, drop_path, depth)]  # stochastic depth decay rule
-        if kv_compress_config is None:
-            kv_compress_config = {
-                'sampling': None,
-                'scale_factor': 1,
-                'kv_compress_layer': [],
-            }
-        self.blocks = nn.ModuleList([
-            PixArtMSBlock(
-                hidden_size, num_heads, mlp_ratio=mlp_ratio, drop_path=drop_path[i],
-                sampling=kv_compress_config['sampling'],
-                sr_ratio=int(kv_compress_config['scale_factor']) if i in kv_compress_config['kv_compress_layer'] else 1,
-                qk_norm=qk_norm,
-                dtype=dtype,
-                device=device,
-                operations=operations,
-            )
-            for i in range(depth)
-        ])
-        self.final_layer = T2IFinalLayer(
-            hidden_size, patch_size, self.out_channels, dtype=dtype, device=device, operations=operations
-        )
-
-    def forward_orig(self, x, timestep, y, mask=None, c_size=None, c_ar=None, **kwargs):
-        """
-        Original forward pass of PixArt.
-        x: (N, C, H, W) tensor of spatial inputs (images or latent representations of images)
-        t: (N,) tensor of diffusion timesteps
-        y: (N, 1, 120, C) conditioning
-        ar: (N, 1): aspect ratio
-        cs: (N ,2) size conditioning for height/width
-        """
-        B, C, H, W = x.shape
-        c_res = (H + W) // 2
-        pe_interpolation = self.pe_interpolation
-        if pe_interpolation is None or self.pe_precision is not None:
-            # calculate pe_interpolation on-the-fly
-            pe_interpolation = round(c_res / (512/8.0), self.pe_precision or 0)
-
-        pos_embed = get_2d_sincos_pos_embed_torch(
-            self.hidden_size,
-            h=(H // self.patch_size),
-            w=(W // self.patch_size),
-            pe_interpolation=pe_interpolation,
-            base_size=((round(c_res / 64) * 64) // self.patch_size),
-            device=x.device,
-            dtype=x.dtype,
-        ).unsqueeze(0)
-
-        x = self.x_embedder(x) + pos_embed  # (N, T, D), where T = H * W / patch_size ** 2
-        t = self.t_embedder(timestep, x.dtype)  # (N, D)
-
-        if self.micro_conditioning and (c_size is not None and c_ar is not None):
-            bs = x.shape[0]
-            c_size = self.csize_embedder(c_size, bs)  # (N, D)
-            c_ar = self.ar_embedder(c_ar, bs)  # (N, D)
-            t = t + torch.cat([c_size, c_ar], dim=1)
-
-        t0 = self.t_block(t)
-        y = self.y_embedder(y, self.training)  # (N, D)
-
-        if mask is not None:
-            if mask.shape[0] != y.shape[0]:
-                mask = mask.repeat(y.shape[0] // mask.shape[0], 1)
-            mask = mask.squeeze(1).squeeze(1)
-            y = y.squeeze(1).masked_select(mask.unsqueeze(-1) != 0).view(1, -1, x.shape[-1])
-            y_lens = mask.sum(dim=1).tolist()
-        else:
-            y_lens = None
-            y = y.squeeze(1).view(1, -1, x.shape[-1])
-        for block in self.blocks:
-            x = block(x, y, t0, y_lens, (H, W), **kwargs)  # (N, T, D)
-
-        x = self.final_layer(x, t)  # (N, T, patch_size ** 2 * out_channels)
-        x = self.unpatchify(x, H, W)  # (N, out_channels, H, W)
-
-        return x
-
-    def forward(self, x, timesteps, context, c_size=None, c_ar=None, **kwargs):
-        B, C, H, W = x.shape
-
-        # Fallback for missing microconds
-        if self.micro_conditioning:
-            if c_size is None:
-                c_size = torch.tensor([H*8, W*8], dtype=x.dtype, device=x.device).repeat(B, 1)
-
-            if c_ar is None:
-                c_ar = torch.tensor([H/W], dtype=x.dtype, device=x.device).repeat(B, 1)
-
-        ## Still accepts the input w/o that dim but returns garbage
-        if len(context.shape) == 3:
-            context = context.unsqueeze(1)
-
-        ## run original forward pass
-        out = self.forward_orig(x, timesteps, context, c_size=c_size, c_ar=c_ar)
-
-        ## only return EPS
-        if self.pred_sigma:
-            return out[:, :self.in_channels]
-        return out
-
-    def unpatchify(self, x, h, w):
-        """
-        x: (N, T, patch_size**2 * C)
-        imgs: (N, H, W, C)
-        """
-        c = self.out_channels
-        p = self.x_embedder.patch_size[0]
-        h = h // self.patch_size
-        w = w // self.patch_size
-        assert h * w == x.shape[1]
-
-        x = x.reshape(shape=(x.shape[0], h, w, p, p, c))
-        x = torch.einsum('nhwpqc->nchpwq', x)
-        imgs = x.reshape(shape=(x.shape[0], c, h * p, w * p))
-        return imgs
+# Based on:
+# https://github.com/PixArt-alpha/PixArt-alpha [Apache 2.0 license]
+# https://github.com/PixArt-alpha/PixArt-sigma [Apache 2.0 license]
+import torch
+import torch.nn as nn
+
+from .blocks import (
+    t2i_modulate,
+    CaptionEmbedder,
+    AttentionKVCompress,
+    MultiHeadCrossAttention,
+    T2IFinalLayer,
+    SizeEmbedder,
+)
+from comfy.ldm.modules.diffusionmodules.mmdit import TimestepEmbedder, PatchEmbed, Mlp, get_1d_sincos_pos_embed_from_grid_torch
+
+
+def get_2d_sincos_pos_embed_torch(embed_dim, w, h, pe_interpolation=1.0, base_size=16, device=None, dtype=torch.float32):
+    grid_h, grid_w = torch.meshgrid(
+        torch.arange(h, device=device, dtype=dtype) / (h/base_size) / pe_interpolation,
+        torch.arange(w, device=device, dtype=dtype) / (w/base_size) / pe_interpolation,
+        indexing='ij'
+    )
+    emb_h = get_1d_sincos_pos_embed_from_grid_torch(embed_dim // 2, grid_h, device=device, dtype=dtype)
+    emb_w = get_1d_sincos_pos_embed_from_grid_torch(embed_dim // 2, grid_w, device=device, dtype=dtype)
+    emb = torch.cat([emb_w, emb_h], dim=1)  # (H*W, D)
+    return emb
+
+class PixArtMSBlock(nn.Module):
+    """
+    A PixArt block with adaptive layer norm zero (adaLN-Zero) conditioning.
+    """
+    def __init__(self, hidden_size, num_heads, mlp_ratio=4.0, drop_path=0., input_size=None,
+                 sampling=None, sr_ratio=1, qk_norm=False, dtype=None, device=None, operations=None, **block_kwargs):
+        super().__init__()
+        self.hidden_size = hidden_size
+        self.norm1 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        self.attn = AttentionKVCompress(
+            hidden_size, num_heads=num_heads, qkv_bias=True, sampling=sampling, sr_ratio=sr_ratio,
+            qk_norm=qk_norm, dtype=dtype, device=device, operations=operations, **block_kwargs
+        )
+        self.cross_attn = MultiHeadCrossAttention(
+            hidden_size, num_heads, dtype=dtype, device=device, operations=operations, **block_kwargs
+        )
+        self.norm2 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        # to be compatible with lower version pytorch
+        approx_gelu = lambda: nn.GELU(approximate="tanh")
+        self.mlp = Mlp(
+            in_features=hidden_size, hidden_features=int(hidden_size * mlp_ratio), act_layer=approx_gelu,
+            dtype=dtype, device=device, operations=operations
+        )
+        self.scale_shift_table = nn.Parameter(torch.randn(6, hidden_size) / hidden_size ** 0.5)
+
+    def forward(self, x, y, t, mask=None, HW=None, **kwargs):
+        B, N, C = x.shape
+
+        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (self.scale_shift_table[None].to(dtype=x.dtype, device=x.device) + t.reshape(B, 6, -1)).chunk(6, dim=1)
+        x = x + (gate_msa * self.attn(t2i_modulate(self.norm1(x), shift_msa, scale_msa), HW=HW))
+        x = x + self.cross_attn(x, y, mask)
+        x = x + (gate_mlp * self.mlp(t2i_modulate(self.norm2(x), shift_mlp, scale_mlp)))
+
+        return x
+
+
+### Core PixArt Model ###
+class PixArtMS(nn.Module):
+    """
+    Diffusion model with a Transformer backbone.
+    """
+    def __init__(
+            self,
+            input_size=32,
+            patch_size=2,
+            in_channels=4,
+            hidden_size=1152,
+            depth=28,
+            num_heads=16,
+            mlp_ratio=4.0,
+            class_dropout_prob=0.1,
+            learn_sigma=True,
+            pred_sigma=True,
+            drop_path: float = 0.,
+            caption_channels=4096,
+            pe_interpolation=None,
+            pe_precision=None,
+            config=None,
+            model_max_length=120,
+            micro_condition=True,
+            qk_norm=False,
+            kv_compress_config=None,
+            dtype=None,
+            device=None,
+            operations=None,
+            **kwargs,
+    ):
+        nn.Module.__init__(self)
+        self.dtype = dtype
+        self.pred_sigma = pred_sigma
+        self.in_channels = in_channels
+        self.out_channels = in_channels * 2 if pred_sigma else in_channels
+        self.patch_size = patch_size
+        self.num_heads = num_heads
+        self.pe_interpolation = pe_interpolation
+        self.pe_precision = pe_precision
+        self.hidden_size = hidden_size
+        self.depth = depth
+
+        approx_gelu = lambda: nn.GELU(approximate="tanh")
+        self.t_block = nn.Sequential(
+            nn.SiLU(),
+            operations.Linear(hidden_size, 6 * hidden_size, bias=True, dtype=dtype, device=device)
+        )
+        self.x_embedder = PatchEmbed(
+            patch_size=patch_size,
+            in_chans=in_channels,
+            embed_dim=hidden_size,
+            bias=True,
+            dtype=dtype,
+            device=device,
+            operations=operations
+        )
+        self.t_embedder = TimestepEmbedder(
+            hidden_size, dtype=dtype, device=device, operations=operations,
+        )
+        self.y_embedder = CaptionEmbedder(
+            in_channels=caption_channels, hidden_size=hidden_size, uncond_prob=class_dropout_prob,
+            act_layer=approx_gelu, token_num=model_max_length,
+            dtype=dtype, device=device, operations=operations,
+        )
+
+        self.micro_conditioning = micro_condition
+        if self.micro_conditioning:
+            self.csize_embedder = SizeEmbedder(hidden_size//3, dtype=dtype, device=device, operations=operations)
+            self.ar_embedder = SizeEmbedder(hidden_size//3, dtype=dtype, device=device, operations=operations)
+
+        # For fixed sin-cos embedding:
+        # num_patches = (input_size // patch_size) * (input_size // patch_size)
+        # self.base_size = input_size // self.patch_size
+        # self.register_buffer("pos_embed", torch.zeros(1, num_patches, hidden_size))
+
+        drop_path = [x.item() for x in torch.linspace(0, drop_path, depth)]  # stochastic depth decay rule
+        if kv_compress_config is None:
+            kv_compress_config = {
+                'sampling': None,
+                'scale_factor': 1,
+                'kv_compress_layer': [],
+            }
+        self.blocks = nn.ModuleList([
+            PixArtMSBlock(
+                hidden_size, num_heads, mlp_ratio=mlp_ratio, drop_path=drop_path[i],
+                sampling=kv_compress_config['sampling'],
+                sr_ratio=int(kv_compress_config['scale_factor']) if i in kv_compress_config['kv_compress_layer'] else 1,
+                qk_norm=qk_norm,
+                dtype=dtype,
+                device=device,
+                operations=operations,
+            )
+            for i in range(depth)
+        ])
+        self.final_layer = T2IFinalLayer(
+            hidden_size, patch_size, self.out_channels, dtype=dtype, device=device, operations=operations
+        )
+
+    def forward_orig(self, x, timestep, y, mask=None, c_size=None, c_ar=None, **kwargs):
+        """
+        Original forward pass of PixArt.
+        x: (N, C, H, W) tensor of spatial inputs (images or latent representations of images)
+        t: (N,) tensor of diffusion timesteps
+        y: (N, 1, 120, C) conditioning
+        ar: (N, 1): aspect ratio
+        cs: (N ,2) size conditioning for height/width
+        """
+        B, C, H, W = x.shape
+        c_res = (H + W) // 2
+        pe_interpolation = self.pe_interpolation
+        if pe_interpolation is None or self.pe_precision is not None:
+            # calculate pe_interpolation on-the-fly
+            pe_interpolation = round(c_res / (512/8.0), self.pe_precision or 0)
+
+        pos_embed = get_2d_sincos_pos_embed_torch(
+            self.hidden_size,
+            h=(H // self.patch_size),
+            w=(W // self.patch_size),
+            pe_interpolation=pe_interpolation,
+            base_size=((round(c_res / 64) * 64) // self.patch_size),
+            device=x.device,
+            dtype=x.dtype,
+        ).unsqueeze(0)
+
+        x = self.x_embedder(x) + pos_embed  # (N, T, D), where T = H * W / patch_size ** 2
+        t = self.t_embedder(timestep, x.dtype)  # (N, D)
+
+        if self.micro_conditioning and (c_size is not None and c_ar is not None):
+            bs = x.shape[0]
+            c_size = self.csize_embedder(c_size, bs)  # (N, D)
+            c_ar = self.ar_embedder(c_ar, bs)  # (N, D)
+            t = t + torch.cat([c_size, c_ar], dim=1)
+
+        t0 = self.t_block(t)
+        y = self.y_embedder(y, self.training)  # (N, D)
+
+        if mask is not None:
+            if mask.shape[0] != y.shape[0]:
+                mask = mask.repeat(y.shape[0] // mask.shape[0], 1)
+            mask = mask.squeeze(1).squeeze(1)
+            y = y.squeeze(1).masked_select(mask.unsqueeze(-1) != 0).view(1, -1, x.shape[-1])
+            y_lens = mask.sum(dim=1).tolist()
+        else:
+            y_lens = None
+            y = y.squeeze(1).view(1, -1, x.shape[-1])
+        for block in self.blocks:
+            x = block(x, y, t0, y_lens, (H, W), **kwargs)  # (N, T, D)
+
+        x = self.final_layer(x, t)  # (N, T, patch_size ** 2 * out_channels)
+        x = self.unpatchify(x, H, W)  # (N, out_channels, H, W)
+
+        return x
+
+    def forward(self, x, timesteps, context, c_size=None, c_ar=None, **kwargs):
+        B, C, H, W = x.shape
+
+        # Fallback for missing microconds
+        if self.micro_conditioning:
+            if c_size is None:
+                c_size = torch.tensor([H*8, W*8], dtype=x.dtype, device=x.device).repeat(B, 1)
+
+            if c_ar is None:
+                c_ar = torch.tensor([H/W], dtype=x.dtype, device=x.device).repeat(B, 1)
+
+        ## Still accepts the input w/o that dim but returns garbage
+        if len(context.shape) == 3:
+            context = context.unsqueeze(1)
+
+        ## run original forward pass
+        out = self.forward_orig(x, timesteps, context, c_size=c_size, c_ar=c_ar)
+
+        ## only return EPS
+        if self.pred_sigma:
+            return out[:, :self.in_channels]
+        return out
+
+    def unpatchify(self, x, h, w):
+        """
+        x: (N, T, patch_size**2 * C)
+        imgs: (N, H, W, C)
+        """
+        c = self.out_channels
+        p = self.x_embedder.patch_size[0]
+        h = h // self.patch_size
+        w = w // self.patch_size
+        assert h * w == x.shape[1]
+
+        x = x.reshape(shape=(x.shape[0], h, w, p, p, c))
+        x = torch.einsum('nhwpqc->nchpwq', x)
+        imgs = x.reshape(shape=(x.shape[0], c, h * p, w * p))
+        return imgs

comfy/ldm/wan/model.py

@@ -146,6 +146,15 @@ WAN_CROSSATTENTION_CLASSES = {
 }
 
 
+def repeat_e(e, x):
+    repeats = 1
+    if e.shape[1] > 1:
+        repeats = x.shape[1] // e.shape[1]
+    if repeats == 1:
+        return e
+    return torch.repeat_interleave(e, repeats, dim=1)
+
+
 class WanAttentionBlock(nn.Module):
 
     def __init__(self,
@@ -202,20 +211,23 @@ class WanAttentionBlock(nn.Module):
         """
         # assert e.dtype == torch.float32
 
-        e = (comfy.model_management.cast_to(self.modulation, dtype=x.dtype, device=x.device) + e).chunk(6, dim=1)
+        if e.ndim < 4:
+            e = (comfy.model_management.cast_to(self.modulation, dtype=x.dtype, device=x.device) + e).chunk(6, dim=1)
+        else:
+            e = (comfy.model_management.cast_to(self.modulation, dtype=x.dtype, device=x.device).unsqueeze(0) + e).unbind(2)
         # assert e[0].dtype == torch.float32
 
         # self-attention
         y = self.self_attn(
-            self.norm1(x) * (1 + e[1]) + e[0],
+            self.norm1(x) * (1 + repeat_e(e[1], x)) + repeat_e(e[0], x),
             freqs)
 
-        x = x + y * e[2]
+        x = x + y * repeat_e(e[2], x)
 
         # cross-attention & ffn
         x = x + self.cross_attn(self.norm3(x), context, context_img_len=context_img_len)
-        y = self.ffn(self.norm2(x) * (1 + e[4]) + e[3])
-        x = x + y * e[5]
+        y = self.ffn(self.norm2(x) * (1 + repeat_e(e[4], x)) + repeat_e(e[3], x))
+        x = x + y * repeat_e(e[5], x)
         return x
 
 
@@ -325,8 +337,12 @@ class Head(nn.Module):
             e(Tensor): Shape [B, C]
         """
         # assert e.dtype == torch.float32
-        e = (comfy.model_management.cast_to(self.modulation, dtype=x.dtype, device=x.device) + e.unsqueeze(1)).chunk(2, dim=1)
-        x = (self.head(self.norm(x) * (1 + e[1]) + e[0]))
+        if e.ndim < 3:
+            e = (comfy.model_management.cast_to(self.modulation, dtype=x.dtype, device=x.device) + e.unsqueeze(1)).chunk(2, dim=1)
+        else:
+            e = (comfy.model_management.cast_to(self.modulation, dtype=x.dtype, device=x.device).unsqueeze(0) + e.unsqueeze(2)).unbind(2)
+
+        x = (self.head(self.norm(x) * (1 + repeat_e(e[1], x)) + repeat_e(e[0], x)))
         return x
 
 
@@ -506,8 +522,9 @@ class WanModel(torch.nn.Module):
 
         # time embeddings
         e = self.time_embedding(
-            sinusoidal_embedding_1d(self.freq_dim, t).to(dtype=x[0].dtype))
-        e0 = self.time_projection(e).unflatten(1, (6, self.dim))
+            sinusoidal_embedding_1d(self.freq_dim, t.flatten()).to(dtype=x[0].dtype))
+        e = e.reshape(t.shape[0], -1, e.shape[-1])
+        e0 = self.time_projection(e).unflatten(2, (6, self.dim))
 
         # context
         context = self.text_embedding(context)

comfy/ldm/wan/vae.py

@@ -52,15 +52,6 @@ class RMS_norm(nn.Module):
             x, dim=(1 if self.channel_first else -1)) * self.scale * self.gamma.to(x) + (self.bias.to(x) if self.bias is not None else 0)
 
 
-class Upsample(nn.Upsample):
-
-    def forward(self, x):
-        """
-        Fix bfloat16 support for nearest neighbor interpolation.
-        """
-        return super().forward(x.float()).type_as(x)
-
-
 class Resample(nn.Module):
 
     def __init__(self, dim, mode):
@@ -73,11 +64,11 @@ class Resample(nn.Module):
         # layers
         if mode == 'upsample2d':
             self.resample = nn.Sequential(
-                Upsample(scale_factor=(2., 2.), mode='nearest-exact'),
+                nn.Upsample(scale_factor=(2., 2.), mode='nearest-exact'),
                 ops.Conv2d(dim, dim // 2, 3, padding=1))
         elif mode == 'upsample3d':
             self.resample = nn.Sequential(
-                Upsample(scale_factor=(2., 2.), mode='nearest-exact'),
+                nn.Upsample(scale_factor=(2., 2.), mode='nearest-exact'),
                 ops.Conv2d(dim, dim // 2, 3, padding=1))
             self.time_conv = CausalConv3d(
                 dim, dim * 2, (3, 1, 1), padding=(1, 0, 0))
@@ -157,29 +148,6 @@ class Resample(nn.Module):
                     feat_idx[0] += 1
         return x
 
-    def init_weight(self, conv):
-        conv_weight = conv.weight
-        nn.init.zeros_(conv_weight)
-        c1, c2, t, h, w = conv_weight.size()
-        one_matrix = torch.eye(c1, c2)
-        init_matrix = one_matrix
-        nn.init.zeros_(conv_weight)
-        #conv_weight.data[:,:,-1,1,1] = init_matrix * 0.5
-        conv_weight.data[:, :, 1, 0, 0] = init_matrix  #* 0.5
-        conv.weight.data.copy_(conv_weight)
-        nn.init.zeros_(conv.bias.data)
-
-    def init_weight2(self, conv):
-        conv_weight = conv.weight.data
-        nn.init.zeros_(conv_weight)
-        c1, c2, t, h, w = conv_weight.size()
-        init_matrix = torch.eye(c1 // 2, c2)
-        #init_matrix = repeat(init_matrix, 'o ... -> (o 2) ...').permute(1,0,2).contiguous().reshape(c1,c2)
-        conv_weight[:c1 // 2, :, -1, 0, 0] = init_matrix
-        conv_weight[c1 // 2:, :, -1, 0, 0] = init_matrix
-        conv.weight.data.copy_(conv_weight)
-        nn.init.zeros_(conv.bias.data)
-
 
 class ResidualBlock(nn.Module):
 
@@ -494,12 +462,6 @@ class WanVAE(nn.Module):
         self.decoder = Decoder3d(dim, z_dim, dim_mult, num_res_blocks,
                                  attn_scales, self.temperal_upsample, dropout)
 
-    def forward(self, x):
-        mu, log_var = self.encode(x)
-        z = self.reparameterize(mu, log_var)
-        x_recon = self.decode(z)
-        return x_recon, mu, log_var
-
     def encode(self, x):
         self.clear_cache()
         ## cache
@@ -545,18 +507,6 @@ class WanVAE(nn.Module):
         self.clear_cache()
         return out
 
-    def reparameterize(self, mu, log_var):
-        std = torch.exp(0.5 * log_var)
-        eps = torch.randn_like(std)
-        return eps * std + mu
-
-    def sample(self, imgs, deterministic=False):
-        mu, log_var = self.encode(imgs)
-        if deterministic:
-            return mu
-        std = torch.exp(0.5 * log_var.clamp(-30.0, 20.0))
-        return mu + std * torch.randn_like(std)
-
     def clear_cache(self):
         self._conv_num = count_conv3d(self.decoder)
         self._conv_idx = [0]

comfy/ldm/wan/vae2_2.py

@@ -0,0 +1,726 @@
+# original version: https://github.com/Wan-Video/Wan2.2/blob/main/wan/modules/vae2_2.py
+# Copyright 2024-2025 The Alibaba Wan Team Authors. All rights reserved.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange
+from .vae import AttentionBlock, CausalConv3d, RMS_norm
+
+import comfy.ops
+ops = comfy.ops.disable_weight_init
+
+CACHE_T = 2
+
+
+class Resample(nn.Module):
+
+    def __init__(self, dim, mode):
+        assert mode in (
+            "none",
+            "upsample2d",
+            "upsample3d",
+            "downsample2d",
+            "downsample3d",
+        )
+        super().__init__()
+        self.dim = dim
+        self.mode = mode
+
+        # layers
+        if mode == "upsample2d":
+            self.resample = nn.Sequential(
+                nn.Upsample(scale_factor=(2.0, 2.0), mode="nearest-exact"),
+                ops.Conv2d(dim, dim, 3, padding=1),
+            )
+        elif mode == "upsample3d":
+            self.resample = nn.Sequential(
+                nn.Upsample(scale_factor=(2.0, 2.0), mode="nearest-exact"),
+                ops.Conv2d(dim, dim, 3, padding=1),
+                # ops.Conv2d(dim, dim//2, 3, padding=1)
+            )
+            self.time_conv = CausalConv3d(
+                dim, dim * 2, (3, 1, 1), padding=(1, 0, 0))
+        elif mode == "downsample2d":
+            self.resample = nn.Sequential(
+                nn.ZeroPad2d((0, 1, 0, 1)),
+                ops.Conv2d(dim, dim, 3, stride=(2, 2)))
+        elif mode == "downsample3d":
+            self.resample = nn.Sequential(
+                nn.ZeroPad2d((0, 1, 0, 1)),
+                ops.Conv2d(dim, dim, 3, stride=(2, 2)))
+            self.time_conv = CausalConv3d(
+                dim, dim, (3, 1, 1), stride=(2, 1, 1), padding=(0, 0, 0))
+        else:
+            self.resample = nn.Identity()
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        b, c, t, h, w = x.size()
+        if self.mode == "upsample3d":
+            if feat_cache is not None:
+                idx = feat_idx[0]
+                if feat_cache[idx] is None:
+                    feat_cache[idx] = "Rep"
+                    feat_idx[0] += 1
+                else:
+                    cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                    if (cache_x.shape[2] < 2 and feat_cache[idx] is not None and
+                            feat_cache[idx] != "Rep"):
+                        # cache last frame of last two chunk
+                        cache_x = torch.cat(
+                            [
+                                feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                                    cache_x.device),
+                                cache_x,
+                            ],
+                            dim=2,
+                        )
+                    if (cache_x.shape[2] < 2 and feat_cache[idx] is not None and
+                            feat_cache[idx] == "Rep"):
+                        cache_x = torch.cat(
+                            [
+                                torch.zeros_like(cache_x).to(cache_x.device),
+                                cache_x
+                            ],
+                            dim=2,
+                        )
+                    if feat_cache[idx] == "Rep":
+                        x = self.time_conv(x)
+                    else:
+                        x = self.time_conv(x, feat_cache[idx])
+                    feat_cache[idx] = cache_x
+                    feat_idx[0] += 1
+                    x = x.reshape(b, 2, c, t, h, w)
+                    x = torch.stack((x[:, 0, :, :, :, :], x[:, 1, :, :, :, :]),
+                                    3)
+                    x = x.reshape(b, c, t * 2, h, w)
+        t = x.shape[2]
+        x = rearrange(x, "b c t h w -> (b t) c h w")
+        x = self.resample(x)
+        x = rearrange(x, "(b t) c h w -> b c t h w", t=t)
+
+        if self.mode == "downsample3d":
+            if feat_cache is not None:
+                idx = feat_idx[0]
+                if feat_cache[idx] is None:
+                    feat_cache[idx] = x.clone()
+                    feat_idx[0] += 1
+                else:
+                    cache_x = x[:, :, -1:, :, :].clone()
+                    x = self.time_conv(
+                        torch.cat([feat_cache[idx][:, :, -1:, :, :], x], 2))
+                    feat_cache[idx] = cache_x
+                    feat_idx[0] += 1
+        return x
+
+
+class ResidualBlock(nn.Module):
+
+    def __init__(self, in_dim, out_dim, dropout=0.0):
+        super().__init__()
+        self.in_dim = in_dim
+        self.out_dim = out_dim
+
+        # layers
+        self.residual = nn.Sequential(
+            RMS_norm(in_dim, images=False),
+            nn.SiLU(),
+            CausalConv3d(in_dim, out_dim, 3, padding=1),
+            RMS_norm(out_dim, images=False),
+            nn.SiLU(),
+            nn.Dropout(dropout),
+            CausalConv3d(out_dim, out_dim, 3, padding=1),
+        )
+        self.shortcut = (
+            CausalConv3d(in_dim, out_dim, 1)
+            if in_dim != out_dim else nn.Identity())
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        old_x = x
+        for layer in self.residual:
+            if isinstance(layer, CausalConv3d) and feat_cache is not None:
+                idx = feat_idx[0]
+                cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                    # cache last frame of last two chunk
+                    cache_x = torch.cat(
+                        [
+                            feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                                cache_x.device),
+                            cache_x,
+                        ],
+                        dim=2,
+                    )
+                x = layer(x, feat_cache[idx])
+                feat_cache[idx] = cache_x
+                feat_idx[0] += 1
+            else:
+                x = layer(x)
+        return x + self.shortcut(old_x)
+
+
+def patchify(x, patch_size):
+    if patch_size == 1:
+        return x
+    if x.dim() == 4:
+        x = rearrange(
+            x, "b c (h q) (w r) -> b (c r q) h w", q=patch_size, r=patch_size)
+    elif x.dim() == 5:
+        x = rearrange(
+            x,
+            "b c f (h q) (w r) -> b (c r q) f h w",
+            q=patch_size,
+            r=patch_size,
+        )
+    else:
+        raise ValueError(f"Invalid input shape: {x.shape}")
+
+    return x
+
+
+def unpatchify(x, patch_size):
+    if patch_size == 1:
+        return x
+
+    if x.dim() == 4:
+        x = rearrange(
+            x, "b (c r q) h w -> b c (h q) (w r)", q=patch_size, r=patch_size)
+    elif x.dim() == 5:
+        x = rearrange(
+            x,
+            "b (c r q) f h w -> b c f (h q) (w r)",
+            q=patch_size,
+            r=patch_size,
+        )
+    return x
+
+
+class AvgDown3D(nn.Module):
+
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        factor_t,
+        factor_s=1,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.factor_t = factor_t
+        self.factor_s = factor_s
+        self.factor = self.factor_t * self.factor_s * self.factor_s
+
+        assert in_channels * self.factor % out_channels == 0
+        self.group_size = in_channels * self.factor // out_channels
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        pad_t = (self.factor_t - x.shape[2] % self.factor_t) % self.factor_t
+        pad = (0, 0, 0, 0, pad_t, 0)
+        x = F.pad(x, pad)
+        B, C, T, H, W = x.shape
+        x = x.view(
+            B,
+            C,
+            T // self.factor_t,
+            self.factor_t,
+            H // self.factor_s,
+            self.factor_s,
+            W // self.factor_s,
+            self.factor_s,
+        )
+        x = x.permute(0, 1, 3, 5, 7, 2, 4, 6).contiguous()
+        x = x.view(
+            B,
+            C * self.factor,
+            T // self.factor_t,
+            H // self.factor_s,
+            W // self.factor_s,
+        )
+        x = x.view(
+            B,
+            self.out_channels,
+            self.group_size,
+            T // self.factor_t,
+            H // self.factor_s,
+            W // self.factor_s,
+        )
+        x = x.mean(dim=2)
+        return x
+
+
+class DupUp3D(nn.Module):
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        factor_t,
+        factor_s=1,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+
+        self.factor_t = factor_t
+        self.factor_s = factor_s
+        self.factor = self.factor_t * self.factor_s * self.factor_s
+
+        assert out_channels * self.factor % in_channels == 0
+        self.repeats = out_channels * self.factor // in_channels
+
+    def forward(self, x: torch.Tensor, first_chunk=False) -> torch.Tensor:
+        x = x.repeat_interleave(self.repeats, dim=1)
+        x = x.view(
+            x.size(0),
+            self.out_channels,
+            self.factor_t,
+            self.factor_s,
+            self.factor_s,
+            x.size(2),
+            x.size(3),
+            x.size(4),
+        )
+        x = x.permute(0, 1, 5, 2, 6, 3, 7, 4).contiguous()
+        x = x.view(
+            x.size(0),
+            self.out_channels,
+            x.size(2) * self.factor_t,
+            x.size(4) * self.factor_s,
+            x.size(6) * self.factor_s,
+        )
+        if first_chunk:
+            x = x[:, :, self.factor_t - 1:, :, :]
+        return x
+
+
+class Down_ResidualBlock(nn.Module):
+
+    def __init__(self,
+                 in_dim,
+                 out_dim,
+                 dropout,
+                 mult,
+                 temperal_downsample=False,
+                 down_flag=False):
+        super().__init__()
+
+        # Shortcut path with downsample
+        self.avg_shortcut = AvgDown3D(
+            in_dim,
+            out_dim,
+            factor_t=2 if temperal_downsample else 1,
+            factor_s=2 if down_flag else 1,
+        )
+
+        # Main path with residual blocks and downsample
+        downsamples = []
+        for _ in range(mult):
+            downsamples.append(ResidualBlock(in_dim, out_dim, dropout))
+            in_dim = out_dim
+
+        # Add the final downsample block
+        if down_flag:
+            mode = "downsample3d" if temperal_downsample else "downsample2d"
+            downsamples.append(Resample(out_dim, mode=mode))
+
+        self.downsamples = nn.Sequential(*downsamples)
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        x_copy = x
+        for module in self.downsamples:
+            x = module(x, feat_cache, feat_idx)
+
+        return x + self.avg_shortcut(x_copy)
+
+
+class Up_ResidualBlock(nn.Module):
+
+    def __init__(self,
+                 in_dim,
+                 out_dim,
+                 dropout,
+                 mult,
+                 temperal_upsample=False,
+                 up_flag=False):
+        super().__init__()
+        # Shortcut path with upsample
+        if up_flag:
+            self.avg_shortcut = DupUp3D(
+                in_dim,
+                out_dim,
+                factor_t=2 if temperal_upsample else 1,
+                factor_s=2 if up_flag else 1,
+            )
+        else:
+            self.avg_shortcut = None
+
+        # Main path with residual blocks and upsample
+        upsamples = []
+        for _ in range(mult):
+            upsamples.append(ResidualBlock(in_dim, out_dim, dropout))
+            in_dim = out_dim
+
+        # Add the final upsample block
+        if up_flag:
+            mode = "upsample3d" if temperal_upsample else "upsample2d"
+            upsamples.append(Resample(out_dim, mode=mode))
+
+        self.upsamples = nn.Sequential(*upsamples)
+
+    def forward(self, x, feat_cache=None, feat_idx=[0], first_chunk=False):
+        x_main = x
+        for module in self.upsamples:
+            x_main = module(x_main, feat_cache, feat_idx)
+        if self.avg_shortcut is not None:
+            x_shortcut = self.avg_shortcut(x, first_chunk)
+            return x_main + x_shortcut
+        else:
+            return x_main
+
+
+class Encoder3d(nn.Module):
+
+    def __init__(
+        self,
+        dim=128,
+        z_dim=4,
+        dim_mult=[1, 2, 4, 4],
+        num_res_blocks=2,
+        attn_scales=[],
+        temperal_downsample=[True, True, False],
+        dropout=0.0,
+    ):
+        super().__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_downsample = temperal_downsample
+
+        # dimensions
+        dims = [dim * u for u in [1] + dim_mult]
+        scale = 1.0
+
+        # init block
+        self.conv1 = CausalConv3d(12, dims[0], 3, padding=1)
+
+        # downsample blocks
+        downsamples = []
+        for i, (in_dim, out_dim) in enumerate(zip(dims[:-1], dims[1:])):
+            t_down_flag = (
+                temperal_downsample[i]
+                if i < len(temperal_downsample) else False)
+            downsamples.append(
+                Down_ResidualBlock(
+                    in_dim=in_dim,
+                    out_dim=out_dim,
+                    dropout=dropout,
+                    mult=num_res_blocks,
+                    temperal_downsample=t_down_flag,
+                    down_flag=i != len(dim_mult) - 1,
+                ))
+            scale /= 2.0
+        self.downsamples = nn.Sequential(*downsamples)
+
+        # middle blocks
+        self.middle = nn.Sequential(
+            ResidualBlock(out_dim, out_dim, dropout),
+            AttentionBlock(out_dim),
+            ResidualBlock(out_dim, out_dim, dropout),
+        )
+
+        # # output blocks
+        self.head = nn.Sequential(
+            RMS_norm(out_dim, images=False),
+            nn.SiLU(),
+            CausalConv3d(out_dim, z_dim, 3, padding=1),
+        )
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+
+        if feat_cache is not None:
+            idx = feat_idx[0]
+            cache_x = x[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                cache_x = torch.cat(
+                    [
+                        feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                            cache_x.device),
+                        cache_x,
+                    ],
+                    dim=2,
+                )
+            x = self.conv1(x, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            x = self.conv1(x)
+
+        ## downsamples
+        for layer in self.downsamples:
+            if feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        ## middle
+        for layer in self.middle:
+            if isinstance(layer, ResidualBlock) and feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        ## head
+        for layer in self.head:
+            if isinstance(layer, CausalConv3d) and feat_cache is not None:
+                idx = feat_idx[0]
+                cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                    cache_x = torch.cat(
+                        [
+                            feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                                cache_x.device),
+                            cache_x,
+                        ],
+                        dim=2,
+                    )
+                x = layer(x, feat_cache[idx])
+                feat_cache[idx] = cache_x
+                feat_idx[0] += 1
+            else:
+                x = layer(x)
+
+        return x
+
+
+class Decoder3d(nn.Module):
+
+    def __init__(
+        self,
+        dim=128,
+        z_dim=4,
+        dim_mult=[1, 2, 4, 4],
+        num_res_blocks=2,
+        attn_scales=[],
+        temperal_upsample=[False, True, True],
+        dropout=0.0,
+    ):
+        super().__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_upsample = temperal_upsample
+
+        # dimensions
+        dims = [dim * u for u in [dim_mult[-1]] + dim_mult[::-1]]
+        # init block
+        self.conv1 = CausalConv3d(z_dim, dims[0], 3, padding=1)
+
+        # middle blocks
+        self.middle = nn.Sequential(
+            ResidualBlock(dims[0], dims[0], dropout),
+            AttentionBlock(dims[0]),
+            ResidualBlock(dims[0], dims[0], dropout),
+        )
+
+        # upsample blocks
+        upsamples = []
+        for i, (in_dim, out_dim) in enumerate(zip(dims[:-1], dims[1:])):
+            t_up_flag = temperal_upsample[i] if i < len(
+                temperal_upsample) else False
+            upsamples.append(
+                Up_ResidualBlock(
+                    in_dim=in_dim,
+                    out_dim=out_dim,
+                    dropout=dropout,
+                    mult=num_res_blocks + 1,
+                    temperal_upsample=t_up_flag,
+                    up_flag=i != len(dim_mult) - 1,
+                ))
+        self.upsamples = nn.Sequential(*upsamples)
+
+        # output blocks
+        self.head = nn.Sequential(
+            RMS_norm(out_dim, images=False),
+            nn.SiLU(),
+            CausalConv3d(out_dim, 12, 3, padding=1),
+        )
+
+    def forward(self, x, feat_cache=None, feat_idx=[0], first_chunk=False):
+        if feat_cache is not None:
+            idx = feat_idx[0]
+            cache_x = x[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                cache_x = torch.cat(
+                    [
+                        feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                            cache_x.device),
+                        cache_x,
+                    ],
+                    dim=2,
+                )
+            x = self.conv1(x, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            x = self.conv1(x)
+
+        for layer in self.middle:
+            if isinstance(layer, ResidualBlock) and feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        ## upsamples
+        for layer in self.upsamples:
+            if feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx, first_chunk)
+            else:
+                x = layer(x)
+
+        ## head
+        for layer in self.head:
+            if isinstance(layer, CausalConv3d) and feat_cache is not None:
+                idx = feat_idx[0]
+                cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                    cache_x = torch.cat(
+                        [
+                            feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                                cache_x.device),
+                            cache_x,
+                        ],
+                        dim=2,
+                    )
+                x = layer(x, feat_cache[idx])
+                feat_cache[idx] = cache_x
+                feat_idx[0] += 1
+            else:
+                x = layer(x)
+        return x
+
+
+def count_conv3d(model):
+    count = 0
+    for m in model.modules():
+        if isinstance(m, CausalConv3d):
+            count += 1
+    return count
+
+
+class WanVAE(nn.Module):
+
+    def __init__(
+        self,
+        dim=160,
+        dec_dim=256,
+        z_dim=16,
+        dim_mult=[1, 2, 4, 4],
+        num_res_blocks=2,
+        attn_scales=[],
+        temperal_downsample=[True, True, False],
+        dropout=0.0,
+    ):
+        super().__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_downsample = temperal_downsample
+        self.temperal_upsample = temperal_downsample[::-1]
+
+        # modules
+        self.encoder = Encoder3d(
+            dim,
+            z_dim * 2,
+            dim_mult,
+            num_res_blocks,
+            attn_scales,
+            self.temperal_downsample,
+            dropout,
+        )
+        self.conv1 = CausalConv3d(z_dim * 2, z_dim * 2, 1)
+        self.conv2 = CausalConv3d(z_dim, z_dim, 1)
+        self.decoder = Decoder3d(
+            dec_dim,
+            z_dim,
+            dim_mult,
+            num_res_blocks,
+            attn_scales,
+            self.temperal_upsample,
+            dropout,
+        )
+
+    def encode(self, x):
+        self.clear_cache()
+        x = patchify(x, patch_size=2)
+        t = x.shape[2]
+        iter_ = 1 + (t - 1) // 4
+        for i in range(iter_):
+            self._enc_conv_idx = [0]
+            if i == 0:
+                out = self.encoder(
+                    x[:, :, :1, :, :],
+                    feat_cache=self._enc_feat_map,
+                    feat_idx=self._enc_conv_idx,
+                )
+            else:
+                out_ = self.encoder(
+                    x[:, :, 1 + 4 * (i - 1):1 + 4 * i, :, :],
+                    feat_cache=self._enc_feat_map,
+                    feat_idx=self._enc_conv_idx,
+                )
+                out = torch.cat([out, out_], 2)
+        mu, log_var = self.conv1(out).chunk(2, dim=1)
+        self.clear_cache()
+        return mu
+
+    def decode(self, z):
+        self.clear_cache()
+        iter_ = z.shape[2]
+        x = self.conv2(z)
+        for i in range(iter_):
+            self._conv_idx = [0]
+            if i == 0:
+                out = self.decoder(
+                    x[:, :, i:i + 1, :, :],
+                    feat_cache=self._feat_map,
+                    feat_idx=self._conv_idx,
+                    first_chunk=True,
+                )
+            else:
+                out_ = self.decoder(
+                    x[:, :, i:i + 1, :, :],
+                    feat_cache=self._feat_map,
+                    feat_idx=self._conv_idx,
+                )
+                out = torch.cat([out, out_], 2)
+        out = unpatchify(out, patch_size=2)
+        self.clear_cache()
+        return out
+
+    def reparameterize(self, mu, log_var):
+        std = torch.exp(0.5 * log_var)
+        eps = torch.randn_like(std)
+        return eps * std + mu
+
+    def sample(self, imgs, deterministic=False):
+        mu, log_var = self.encode(imgs)
+        if deterministic:
+            return mu
+        std = torch.exp(0.5 * log_var.clamp(-30.0, 20.0))
+        return mu + std * torch.randn_like(std)
+
+    def clear_cache(self):
+        self._conv_num = count_conv3d(self.decoder)
+        self._conv_idx = [0]
+        self._feat_map = [None] * self._conv_num
+        # cache encode
+        self._enc_conv_num = count_conv3d(self.encoder)
+        self._enc_conv_idx = [0]
+        self._enc_feat_map = [None] * self._enc_conv_num

comfy/model_base.py

@@ -1097,8 +1097,9 @@ class WAN21(BaseModel):
                 image[:, i: i + 16] = self.process_latent_in(image[:, i: i + 16])
             image = utils.resize_to_batch_size(image, noise.shape[0])
 
-        if not self.image_to_video or extra_channels == image.shape[1]:
-            return image
+        if extra_channels != image.shape[1] + 4:
+            if not self.image_to_video or extra_channels == image.shape[1]:
+                return image
 
         if image.shape[1] > (extra_channels - 4):
             image = image[:, :(extra_channels - 4)]
@@ -1182,6 +1183,31 @@ class WAN21_Camera(WAN21):
             out['camera_conditions'] = comfy.conds.CONDRegular(camera_conditions)
         return out
 
+class WAN22(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLOW, image_to_video=False, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.wan.model.WanModel)
+        self.image_to_video = image_to_video
+
+    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)
+
+        denoise_mask = kwargs.get("concat_mask", kwargs.get("denoise_mask", None))
+        if denoise_mask is not None:
+            out["denoise_mask"] = comfy.conds.CONDRegular(denoise_mask)
+        return out
+
+    def process_timestep(self, timestep, x, denoise_mask=None, **kwargs):
+        if denoise_mask is None:
+            return timestep
+        temp_ts = (torch.mean(denoise_mask[:, :, :, :, :], dim=(1, 3, 4), keepdim=True) * timestep.view([timestep.shape[0]] + [1] * (denoise_mask.ndim - 1))).reshape(timestep.shape[0], -1)
+        return temp_ts
+
+    def scale_latent_inpaint(self, sigma, noise, latent_image, **kwargs):
+        return latent_image
+
 class Hunyuan3Dv2(BaseModel):
     def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
         super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.hunyuan3d.model.Hunyuan3Dv2)

comfy/model_detection.py

@@ -346,7 +346,9 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
         dit_config = {}
         dit_config["image_model"] = "wan2.1"
         dim = state_dict['{}head.modulation'.format(key_prefix)].shape[-1]
+        out_dim = state_dict['{}head.head.weight'.format(key_prefix)].shape[0] // 4
         dit_config["dim"] = dim
+        dit_config["out_dim"] = out_dim
         dit_config["num_heads"] = dim // 128
         dit_config["ffn_dim"] = state_dict['{}blocks.0.ffn.0.weight'.format(key_prefix)].shape[0]
         dit_config["num_layers"] = count_blocks(state_dict_keys, '{}blocks.'.format(key_prefix) + '{}.')

comfy/model_management.py

@@ -101,7 +101,7 @@ if args.directml is not None:
     lowvram_available = False #TODO: need to find a way to get free memory in directml before this can be enabled by default.
 
 try:
-    import intel_extension_for_pytorch as ipex
+    import intel_extension_for_pytorch as ipex  # noqa: F401
     _ = torch.xpu.device_count()
     xpu_available = xpu_available or torch.xpu.is_available()
 except:
@@ -128,6 +128,11 @@ try:
 except:
     mlu_available = False
 
+try:
+    ixuca_available = hasattr(torch, "corex")
+except:
+    ixuca_available = False
+
 if args.cpu:
     cpu_state = CPUState.CPU
 
@@ -151,6 +156,12 @@ def is_mlu():
         return True
     return False
 
+def is_ixuca():
+    global ixuca_available
+    if ixuca_available:
+        return True
+    return False
+
 def get_torch_device():
     global directml_enabled
     global cpu_state
@@ -186,8 +197,9 @@ def get_total_memory(dev=None, torch_total_too=False):
         elif is_intel_xpu():
             stats = torch.xpu.memory_stats(dev)
             mem_reserved = stats['reserved_bytes.all.current']
+            mem_total_xpu = torch.xpu.get_device_properties(dev).total_memory
             mem_total_torch = mem_reserved
-            mem_total = torch.xpu.get_device_properties(dev).total_memory
+            mem_total = mem_total_xpu
         elif is_ascend_npu():
             stats = torch.npu.memory_stats(dev)
             mem_reserved = stats['reserved_bytes.all.current']
@@ -288,7 +300,7 @@ try:
         if torch_version_numeric[0] >= 2:
             if ENABLE_PYTORCH_ATTENTION == False and args.use_split_cross_attention == False and args.use_quad_cross_attention == False:
                 ENABLE_PYTORCH_ATTENTION = True
-    if is_intel_xpu() or is_ascend_npu() or is_mlu():
+    if is_intel_xpu() or is_ascend_npu() or is_mlu() or is_ixuca():
         if args.use_split_cross_attention == False and args.use_quad_cross_attention == False:
             ENABLE_PYTORCH_ATTENTION = True
 except:
@@ -307,7 +319,10 @@ try:
         logging.info("ROCm version: {}".format(rocm_version))
         if args.use_split_cross_attention == False and args.use_quad_cross_attention == False:
             if torch_version_numeric >= (2, 7):  # works on 2.6 but doesn't actually seem to improve much
-                if any((a in arch) for a in ["gfx90a", "gfx942", "gfx1100", "gfx1101", "gfx1151"]):  # TODO: more arches, TODO: gfx1201 and gfx950
+                if any((a in arch) for a in ["gfx90a", "gfx942", "gfx1100", "gfx1101", "gfx1151"]):  # TODO: more arches, TODO: gfx950
+                    ENABLE_PYTORCH_ATTENTION = True
+            if torch_version_numeric >= (2, 8):
+                if any((a in arch) for a in ["gfx1201"]):
                     ENABLE_PYTORCH_ATTENTION = True
         if torch_version_numeric >= (2, 7) and rocm_version >= (6, 4):
             if any((a in arch) for a in ["gfx1201", "gfx942", "gfx950"]):  # TODO: more arches
@@ -377,6 +392,8 @@ def get_torch_device_name(device):
             except:
                 allocator_backend = ""
             return "{} {} : {}".format(device, torch.cuda.get_device_name(device), allocator_backend)
+        elif device.type == "xpu":
+            return "{} {}".format(device, torch.xpu.get_device_name(device))
         else:
             return "{}".format(device.type)
     elif is_intel_xpu():
@@ -512,6 +529,8 @@ WINDOWS = any(platform.win32_ver())
 EXTRA_RESERVED_VRAM = 400 * 1024 * 1024
 if WINDOWS:
     EXTRA_RESERVED_VRAM = 600 * 1024 * 1024 #Windows is higher because of the shared vram issue
+    if total_vram > (15 * 1024):  # more extra reserved vram on 16GB+ cards
+        EXTRA_RESERVED_VRAM += 100 * 1024 * 1024
 
 if args.reserve_vram is not None:
     EXTRA_RESERVED_VRAM = args.reserve_vram * 1024 * 1024 * 1024
@@ -876,6 +895,7 @@ def vae_dtype(device=None, allowed_dtypes=[]):
             return d
 
         # NOTE: bfloat16 seems to work on AMD for the VAE but is extremely slow in some cases compared to fp32
+        # slowness still a problem on pytorch nightly 2.9.0.dev20250720+rocm6.4 tested on RDNA3
         if d == torch.bfloat16 and (not is_amd()) and should_use_bf16(device):
             return d
 
@@ -929,7 +949,7 @@ def device_supports_non_blocking(device):
     if is_device_mps(device):
         return False #pytorch bug? mps doesn't support non blocking
     if is_intel_xpu():
-        return False
+        return True
     if args.deterministic: #TODO: figure out why deterministic breaks non blocking from gpu to cpu (previews)
         return False
     if directml_enabled:
@@ -968,6 +988,8 @@ def get_offload_stream(device):
         stream_counter = (stream_counter + 1) % len(ss)
         if is_device_cuda(device):
             ss[stream_counter].wait_stream(torch.cuda.current_stream())
+        elif is_device_xpu(device):
+            ss[stream_counter].wait_stream(torch.xpu.current_stream())
         stream_counters[device] = stream_counter
         return s
     elif is_device_cuda(device):
@@ -979,6 +1001,15 @@ def get_offload_stream(device):
         stream_counter = (stream_counter + 1) % len(ss)
         stream_counters[device] = stream_counter
         return s
+    elif is_device_xpu(device):
+        ss = []
+        for k in range(NUM_STREAMS):
+            ss.append(torch.xpu.Stream(device=device, priority=0))
+        STREAMS[device] = ss
+        s = ss[stream_counter]
+        stream_counter = (stream_counter + 1) % len(ss)
+        stream_counters[device] = stream_counter
+        return s
     return None
 
 def sync_stream(device, stream):
@@ -986,6 +1017,8 @@ def sync_stream(device, stream):
         return
     if is_device_cuda(device):
         torch.cuda.current_stream().wait_stream(stream)
+    elif is_device_xpu(device):
+        torch.xpu.current_stream().wait_stream(stream)
 
 def cast_to(weight, dtype=None, device=None, non_blocking=False, copy=False, stream=None):
     if device is None or weight.device == device:
@@ -1027,6 +1060,8 @@ def xformers_enabled():
         return False
     if is_mlu():
         return False
+    if is_ixuca():
+        return False
     if directml_enabled:
         return False
     return XFORMERS_IS_AVAILABLE
@@ -1062,6 +1097,8 @@ def pytorch_attention_flash_attention():
             return True
         if is_amd():
             return True #if you have pytorch attention enabled on AMD it probably supports at least mem efficient attention
+        if is_ixuca():
+            return True
     return False
 
 def force_upcast_attention_dtype():
@@ -1092,8 +1129,8 @@ def get_free_memory(dev=None, torch_free_too=False):
             stats = torch.xpu.memory_stats(dev)
             mem_active = stats['active_bytes.all.current']
             mem_reserved = stats['reserved_bytes.all.current']
-            mem_free_torch = mem_reserved - mem_active
             mem_free_xpu = torch.xpu.get_device_properties(dev).total_memory - mem_reserved
+            mem_free_torch = mem_reserved - mem_active
             mem_free_total = mem_free_xpu + mem_free_torch
         elif is_ascend_npu():
             stats = torch.npu.memory_stats(dev)
@@ -1142,6 +1179,9 @@ def is_device_cpu(device):
 def is_device_mps(device):
     return is_device_type(device, 'mps')
 
+def is_device_xpu(device):
+    return is_device_type(device, 'xpu')
+
 def is_device_cuda(device):
     return is_device_type(device, 'cuda')
 
@@ -1173,7 +1213,10 @@ def should_use_fp16(device=None, model_params=0, prioritize_performance=True, ma
         return False
 
     if is_intel_xpu():
-        return True
+        if torch_version_numeric < (2, 3):
+            return True
+        else:
+            return torch.xpu.get_device_properties(device).has_fp16
 
     if is_ascend_npu():
         return True
@@ -1181,6 +1224,9 @@ def should_use_fp16(device=None, model_params=0, prioritize_performance=True, ma
     if is_mlu():
         return True
 
+    if is_ixuca():
+        return True
+
     if torch.version.hip:
         return True
 
@@ -1236,11 +1282,17 @@ def should_use_bf16(device=None, model_params=0, prioritize_performance=True, ma
         return False
 
     if is_intel_xpu():
-        return True
+        if torch_version_numeric < (2, 6):
+            return True
+        else:
+            return torch.xpu.get_device_capability(device)['has_bfloat16_conversions']
 
     if is_ascend_npu():
         return True
 
+    if is_ixuca():
+        return True
+
     if is_amd():
         arch = torch.cuda.get_device_properties(device).gcnArchName
         if any((a in arch) for a in ["gfx1030", "gfx1031", "gfx1010", "gfx1011", "gfx1012", "gfx906", "gfx900", "gfx803"]):  # RDNA2 and older don't support bf16

comfy/model_patcher.py

@@ -379,6 +379,9 @@ class ModelPatcher:
     def set_model_sampler_pre_cfg_function(self, pre_cfg_function, disable_cfg1_optimization=False):
         self.model_options = set_model_options_pre_cfg_function(self.model_options, pre_cfg_function, disable_cfg1_optimization)
 
+    def set_model_sampler_calc_cond_batch_function(self, sampler_calc_cond_batch_function):
+        self.model_options["sampler_calc_cond_batch_function"] = sampler_calc_cond_batch_function
+
     def set_model_unet_function_wrapper(self, unet_wrapper_function: UnetWrapperFunction):
         self.model_options["model_function_wrapper"] = unet_wrapper_function
 

comfy/ops.py

@@ -336,9 +336,12 @@ class fp8_ops(manual_cast):
             return None
 
         def forward_comfy_cast_weights(self, input):
-            out = fp8_linear(self, input)
-            if out is not None:
-                return out
+            try:
+                out = fp8_linear(self, input)
+                if out is not None:
+                    return out
+            except Exception as e:
+                logging.info("Exception during fp8 op: {}".format(e))
 
             weight, bias = cast_bias_weight(self, input)
             return torch.nn.functional.linear(input, weight, bias)

comfy/samplers.py

@@ -373,7 +373,11 @@ def sampling_function(model, x, timestep, uncond, cond, cond_scale, model_option
         uncond_ = uncond
 
     conds = [cond, uncond_]
-    out = calc_cond_batch(model, conds, x, timestep, model_options)
+    if "sampler_calc_cond_batch_function" in model_options:
+        args = {"conds": conds, "input": x, "sigma": timestep, "model": model, "model_options": model_options}
+        out = model_options["sampler_calc_cond_batch_function"](args)
+    else:
+        out = calc_cond_batch(model, conds, x, timestep, model_options)
 
     for fn in model_options.get("sampler_pre_cfg_function", []):
         args = {"conds":conds, "conds_out": out, "cond_scale": cond_scale, "timestep": timestep,
@@ -716,7 +720,7 @@ KSAMPLER_NAMES = ["euler", "euler_cfg_pp", "euler_ancestral", "euler_ancestral_c
                   "lms", "dpm_fast", "dpm_adaptive", "dpmpp_2s_ancestral", "dpmpp_2s_ancestral_cfg_pp", "dpmpp_sde", "dpmpp_sde_gpu",
                   "dpmpp_2m", "dpmpp_2m_cfg_pp", "dpmpp_2m_sde", "dpmpp_2m_sde_gpu", "dpmpp_3m_sde", "dpmpp_3m_sde_gpu", "ddpm", "lcm",
                   "ipndm", "ipndm_v", "deis", "res_multistep", "res_multistep_cfg_pp", "res_multistep_ancestral", "res_multistep_ancestral_cfg_pp",
-                  "gradient_estimation", "gradient_estimation_cfg_pp", "er_sde", "seeds_2", "seeds_3"]
+                  "gradient_estimation", "gradient_estimation_cfg_pp", "er_sde", "seeds_2", "seeds_3", "sa_solver", "sa_solver_pece"]
 
 class KSAMPLER(Sampler):
     def __init__(self, sampler_function, extra_options={}, inpaint_options={}):

comfy/sd.py

@@ -14,10 +14,12 @@ import comfy.ldm.genmo.vae.model
 import comfy.ldm.lightricks.vae.causal_video_autoencoder
 import comfy.ldm.cosmos.vae
 import comfy.ldm.wan.vae
+import comfy.ldm.wan.vae2_2
 import comfy.ldm.hunyuan3d.vae
 import comfy.ldm.ace.vae.music_dcae_pipeline
 import yaml
 import math
+import os
 
 import comfy.utils
 
@@ -419,17 +421,30 @@ class VAE:
                 self.memory_used_encode = lambda shape, dtype: (50 * (round((shape[2] + 7) / 8) * 8) * shape[3] * shape[4]) * model_management.dtype_size(dtype)
                 self.working_dtypes = [torch.bfloat16, torch.float32]
             elif "decoder.middle.0.residual.0.gamma" in sd:
-                self.upscale_ratio = (lambda a: max(0, a * 4 - 3), 8, 8)
-                self.upscale_index_formula = (4, 8, 8)
-                self.downscale_ratio = (lambda a: max(0, math.floor((a + 3) / 4)), 8, 8)
-                self.downscale_index_formula = (4, 8, 8)
-                self.latent_dim = 3
-                self.latent_channels = 16
-                ddconfig = {"dim": 96, "z_dim": self.latent_channels, "dim_mult": [1, 2, 4, 4], "num_res_blocks": 2, "attn_scales": [], "temperal_downsample": [False, True, True], "dropout": 0.0}
-                self.first_stage_model = comfy.ldm.wan.vae.WanVAE(**ddconfig)
-                self.working_dtypes = [torch.bfloat16, torch.float16, torch.float32]
-                self.memory_used_encode = lambda shape, dtype: 6000 * shape[3] * shape[4] * model_management.dtype_size(dtype)
-                self.memory_used_decode = lambda shape, dtype: 7000 * shape[3] * shape[4] * (8 * 8) * model_management.dtype_size(dtype)
+                if "decoder.upsamples.0.upsamples.0.residual.2.weight" in sd:  # Wan 2.2 VAE
+                    self.upscale_ratio = (lambda a: max(0, a * 4 - 3), 16, 16)
+                    self.upscale_index_formula = (4, 16, 16)
+                    self.downscale_ratio = (lambda a: max(0, math.floor((a + 3) / 4)), 16, 16)
+                    self.downscale_index_formula = (4, 16, 16)
+                    self.latent_dim = 3
+                    self.latent_channels = 48
+                    ddconfig = {"dim": 160, "z_dim": self.latent_channels, "dim_mult": [1, 2, 4, 4], "num_res_blocks": 2, "attn_scales": [], "temperal_downsample": [False, True, True], "dropout": 0.0}
+                    self.first_stage_model = comfy.ldm.wan.vae2_2.WanVAE(**ddconfig)
+                    self.working_dtypes = [torch.bfloat16, torch.float16, torch.float32]
+                    self.memory_used_encode = lambda shape, dtype: 3300 * shape[3] * shape[4] * model_management.dtype_size(dtype)
+                    self.memory_used_decode = lambda shape, dtype: 8000 * shape[3] * shape[4] * (16 * 16) * model_management.dtype_size(dtype)
+                else:  # Wan 2.1 VAE
+                    self.upscale_ratio = (lambda a: max(0, a * 4 - 3), 8, 8)
+                    self.upscale_index_formula = (4, 8, 8)
+                    self.downscale_ratio = (lambda a: max(0, math.floor((a + 3) / 4)), 8, 8)
+                    self.downscale_index_formula = (4, 8, 8)
+                    self.latent_dim = 3
+                    self.latent_channels = 16
+                    ddconfig = {"dim": 96, "z_dim": self.latent_channels, "dim_mult": [1, 2, 4, 4], "num_res_blocks": 2, "attn_scales": [], "temperal_downsample": [False, True, True], "dropout": 0.0}
+                    self.first_stage_model = comfy.ldm.wan.vae.WanVAE(**ddconfig)
+                    self.working_dtypes = [torch.bfloat16, torch.float16, torch.float32]
+                    self.memory_used_encode = lambda shape, dtype: 6000 * shape[3] * shape[4] * model_management.dtype_size(dtype)
+                    self.memory_used_decode = lambda shape, dtype: 7000 * shape[3] * shape[4] * (8 * 8) * model_management.dtype_size(dtype)
             elif "geo_decoder.cross_attn_decoder.ln_1.bias" in sd:
                 self.latent_dim = 1
                 ln_post = "geo_decoder.ln_post.weight" in sd
@@ -977,6 +992,12 @@ def load_gligen(ckpt_path):
         model = model.half()
     return comfy.model_patcher.ModelPatcher(model, load_device=model_management.get_torch_device(), offload_device=model_management.unet_offload_device())
 
+def model_detection_error_hint(path, state_dict):
+    filename = os.path.basename(path)
+    if 'lora' in filename.lower():
+        return "\nHINT: This seems to be a Lora file and Lora files should be put in the lora folder and loaded with a lora loader node.."
+    return ""
+
 def load_checkpoint(config_path=None, ckpt_path=None, output_vae=True, output_clip=True, embedding_directory=None, state_dict=None, config=None):
     logging.warning("Warning: The load checkpoint with config function is deprecated and will eventually be removed, please use the other one.")
     model, clip, vae, _ = load_checkpoint_guess_config(ckpt_path, output_vae=output_vae, output_clip=output_clip, output_clipvision=False, embedding_directory=embedding_directory, output_model=True)
@@ -1005,7 +1026,7 @@ def load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=True, o
     sd, metadata = comfy.utils.load_torch_file(ckpt_path, return_metadata=True)
     out = load_state_dict_guess_config(sd, output_vae, output_clip, output_clipvision, embedding_directory, output_model, model_options, te_model_options=te_model_options, metadata=metadata)
     if out is None:
-        raise RuntimeError("ERROR: Could not detect model type of: {}".format(ckpt_path))
+        raise RuntimeError("ERROR: Could not detect model type of: {}\n{}".format(ckpt_path, model_detection_error_hint(ckpt_path, sd)))
     return out
 
 def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_clipvision=False, embedding_directory=None, output_model=True, model_options={}, te_model_options={}, metadata=None):
@@ -1177,7 +1198,7 @@ def load_diffusion_model(unet_path, model_options={}):
     model = load_diffusion_model_state_dict(sd, model_options=model_options)
     if model is None:
         logging.error("ERROR UNSUPPORTED DIFFUSION MODEL {}".format(unet_path))
-        raise RuntimeError("ERROR: Could not detect model type of: {}".format(unet_path))
+        raise RuntimeError("ERROR: Could not detect model type of: {}\n{}".format(unet_path, model_detection_error_hint(unet_path, sd)))
     return model
 
 def load_unet(unet_path, dtype=None):

comfy/sd1_tokenizer/tokenizer_config.json

@@ -18,7 +18,7 @@
     "single_word": false
   },
   "errors": "replace",
-  "model_max_length": 77,
+  "model_max_length": 8192,
   "name_or_path": "openai/clip-vit-large-patch14",
   "pad_token": "<|endoftext|>",
   "special_tokens_map_file": "./special_tokens_map.json",

comfy/supported_models.py

@@ -1059,6 +1059,19 @@ class WAN21_Vace(WAN21_T2V):
         out = model_base.WAN21_Vace(self, image_to_video=False, device=device)
         return out
 
+class WAN22_T2V(WAN21_T2V):
+    unet_config = {
+        "image_model": "wan2.1",
+        "model_type": "t2v",
+        "out_dim": 48,
+    }
+
+    latent_format = latent_formats.Wan22
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.WAN22(self, image_to_video=True, device=device)
+        return out
+
 class Hunyuan3Dv2(supported_models_base.BASE):
     unet_config = {
         "image_model": "hunyuan3d2",
@@ -1214,9 +1227,9 @@ class Omnigen2(supported_models_base.BASE):
     def clip_target(self, state_dict={}):
         pref = self.text_encoder_key_prefix[0]
         hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen25_3b.transformer.".format(pref))
-        return supported_models_base.ClipTarget(comfy.text_encoders.omnigen2.LuminaTokenizer, comfy.text_encoders.omnigen2.te(**hunyuan_detect))
+        return supported_models_base.ClipTarget(comfy.text_encoders.omnigen2.Omnigen2Tokenizer, comfy.text_encoders.omnigen2.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, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, Hunyuan3Dv2mini, Hunyuan3Dv2, HiDream, Chroma, ACEStep, Omnigen2]
+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]
 
 models += [SVD_img2vid]

comfy/text_encoders/pixart_t5.py

@@ -1,42 +1,42 @@
-import os
-
-from comfy import sd1_clip
-import comfy.text_encoders.t5
-import comfy.text_encoders.sd3_clip
-from comfy.sd1_clip import gen_empty_tokens
-
-from transformers import T5TokenizerFast
-
-class T5XXLModel(comfy.text_encoders.sd3_clip.T5XXLModel):
-    def __init__(self, **kwargs):
-        super().__init__(**kwargs)
-
-    def gen_empty_tokens(self, special_tokens, *args, **kwargs):
-        # PixArt expects the negative to be all pad tokens
-        special_tokens = special_tokens.copy()
-        special_tokens.pop("end")
-        return gen_empty_tokens(special_tokens, *args, **kwargs)
-
-class PixArtT5XXL(sd1_clip.SD1ClipModel):
-    def __init__(self, device="cpu", dtype=None, model_options={}):
-        super().__init__(device=device, dtype=dtype, name="t5xxl", clip_model=T5XXLModel, model_options=model_options)
-
-class T5XXLTokenizer(sd1_clip.SDTokenizer):
-    def __init__(self, embedding_directory=None, tokenizer_data={}):
-        tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_tokenizer")
-        super().__init__(tokenizer_path, embedding_directory=embedding_directory, pad_with_end=False, embedding_size=4096, embedding_key='t5xxl', tokenizer_class=T5TokenizerFast, has_start_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, tokenizer_data=tokenizer_data) # no padding
-
-class PixArtTokenizer(sd1_clip.SD1Tokenizer):
-    def __init__(self, embedding_directory=None, tokenizer_data={}):
-        super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, clip_name="t5xxl", tokenizer=T5XXLTokenizer)
-
-def pixart_te(dtype_t5=None, t5xxl_scaled_fp8=None):
-    class PixArtTEModel_(PixArtT5XXL):
-        def __init__(self, device="cpu", dtype=None, model_options={}):
-            if t5xxl_scaled_fp8 is not None and "t5xxl_scaled_fp8" not in model_options:
-                model_options = model_options.copy()
-                model_options["t5xxl_scaled_fp8"] = t5xxl_scaled_fp8
-            if dtype is None:
-                dtype = dtype_t5
-            super().__init__(device=device, dtype=dtype, model_options=model_options)
-    return PixArtTEModel_
+import os
+
+from comfy import sd1_clip
+import comfy.text_encoders.t5
+import comfy.text_encoders.sd3_clip
+from comfy.sd1_clip import gen_empty_tokens
+
+from transformers import T5TokenizerFast
+
+class T5XXLModel(comfy.text_encoders.sd3_clip.T5XXLModel):
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+
+    def gen_empty_tokens(self, special_tokens, *args, **kwargs):
+        # PixArt expects the negative to be all pad tokens
+        special_tokens = special_tokens.copy()
+        special_tokens.pop("end")
+        return gen_empty_tokens(special_tokens, *args, **kwargs)
+
+class PixArtT5XXL(sd1_clip.SD1ClipModel):
+    def __init__(self, device="cpu", dtype=None, model_options={}):
+        super().__init__(device=device, dtype=dtype, name="t5xxl", clip_model=T5XXLModel, model_options=model_options)
+
+class T5XXLTokenizer(sd1_clip.SDTokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_tokenizer")
+        super().__init__(tokenizer_path, embedding_directory=embedding_directory, pad_with_end=False, embedding_size=4096, embedding_key='t5xxl', tokenizer_class=T5TokenizerFast, has_start_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, tokenizer_data=tokenizer_data) # no padding
+
+class PixArtTokenizer(sd1_clip.SD1Tokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, clip_name="t5xxl", tokenizer=T5XXLTokenizer)
+
+def pixart_te(dtype_t5=None, t5xxl_scaled_fp8=None):
+    class PixArtTEModel_(PixArtT5XXL):
+        def __init__(self, device="cpu", dtype=None, model_options={}):
+            if t5xxl_scaled_fp8 is not None and "t5xxl_scaled_fp8" not in model_options:
+                model_options = model_options.copy()
+                model_options["t5xxl_scaled_fp8"] = t5xxl_scaled_fp8
+            if dtype is None:
+                dtype = dtype_t5
+            super().__init__(device=device, dtype=dtype, model_options=model_options)
+    return PixArtTEModel_

comfy/utils.py

@@ -31,6 +31,7 @@ from einops import rearrange
 from comfy.cli_args import args
 
 MMAP_TORCH_FILES = args.mmap_torch_files
+DISABLE_MMAP = args.disable_mmap
 
 ALWAYS_SAFE_LOAD = False
 if hasattr(torch.serialization, "add_safe_globals"):  # TODO: this was added in pytorch 2.4, the unsafe path should be removed once earlier versions are deprecated
@@ -58,7 +59,10 @@ def load_torch_file(ckpt, safe_load=False, device=None, return_metadata=False):
             with safetensors.safe_open(ckpt, framework="pt", device=device.type) as f:
                 sd = {}
                 for k in f.keys():
-                    sd[k] = f.get_tensor(k)
+                    tensor = f.get_tensor(k)
+                    if DISABLE_MMAP:  # TODO: Not sure if this is the best way to bypass the mmap issues
+                        tensor = tensor.to(device=device, copy=True)
+                    sd[k] = tensor
                 if return_metadata:
                     metadata = f.metadata()
         except Exception as e:
@@ -77,6 +81,7 @@ def load_torch_file(ckpt, safe_load=False, device=None, return_metadata=False):
         if safe_load or ALWAYS_SAFE_LOAD:
             pl_sd = torch.load(ckpt, map_location=device, weights_only=True, **torch_args)
         else:
+            logging.warning("WARNING: loading {} unsafely, upgrade your pytorch to 2.4 or newer to load this file safely.".format(ckpt))
             pl_sd = torch.load(ckpt, map_location=device, pickle_module=comfy.checkpoint_pickle)
         if "state_dict" in pl_sd:
             sd = pl_sd["state_dict"]
@@ -693,6 +698,26 @@ def resize_to_batch_size(tensor, batch_size):
 
     return output
 
+def resize_list_to_batch_size(l, batch_size):
+    in_batch_size = len(l)
+    if in_batch_size == batch_size or in_batch_size == 0:
+        return l
+
+    if batch_size <= 1:
+        return l[:batch_size]
+
+    output = []
+    if batch_size < in_batch_size:
+        scale = (in_batch_size - 1) / (batch_size - 1)
+        for i in range(batch_size):
+            output.append(l[min(round(i * scale), in_batch_size - 1)])
+    else:
+        scale = in_batch_size / batch_size
+        for i in range(batch_size):
+           output.append(l[min(math.floor((i + 0.5) * scale), in_batch_size - 1)])
+
+    return output
+
 def convert_sd_to(state_dict, dtype):
     keys = list(state_dict.keys())
     for k in keys:
@@ -997,11 +1022,12 @@ def set_progress_bar_global_hook(function):
     PROGRESS_BAR_HOOK = function
 
 class ProgressBar:
-    def __init__(self, total):
+    def __init__(self, total, node_id=None):
         global PROGRESS_BAR_HOOK
         self.total = total
         self.current = 0
         self.hook = PROGRESS_BAR_HOOK
+        self.node_id = node_id
 
     def update_absolute(self, value, total=None, preview=None):
         if total is not None:
@@ -1010,7 +1036,7 @@ class ProgressBar:
             value = self.total
         self.current = value
         if self.hook is not None:
-            self.hook(self.current, self.total, preview)
+            self.hook(self.current, self.total, preview, node_id=self.node_id)
 
     def update(self, value):
         self.update_absolute(self.current + value)

comfy/weight_adapter/__init__.py

@@ -15,9 +15,20 @@ adapters: list[type[WeightAdapterBase]] = [
     OFTAdapter,
     BOFTAdapter,
 ]
+adapter_maps: dict[str, type[WeightAdapterBase]] = {
+    "LoRA": LoRAAdapter,
+    "LoHa": LoHaAdapter,
+    "LoKr": LoKrAdapter,
+    "OFT": OFTAdapter,
+    ## We disable not implemented algo for now
+    # "GLoRA": GLoRAAdapter,
+    # "BOFT": BOFTAdapter,
+}
+
 
 __all__ = [
     "WeightAdapterBase",
     "WeightAdapterTrainBase",
-    "adapters"
+    "adapters",
+    "adapter_maps",
 ] + [a.__name__ for a in adapters]

comfy/weight_adapter/base.py

@@ -133,3 +133,43 @@ def tucker_weight_from_conv(up, down, mid):
 def tucker_weight(wa, wb, t):
     temp = torch.einsum("i j ..., j r -> i r ...", t, wb)
     return torch.einsum("i j ..., i r -> r j ...", temp, wa)
+
+
+def factorization(dimension: int, factor: int = -1) -> tuple[int, int]:
+    """
+    return a tuple of two value of input dimension decomposed by the number closest to factor
+    second value is higher or equal than first value.
+
+    examples)
+    factor
+        -1               2                4               8               16               ...
+    127 -> 1, 127   127 -> 1, 127    127 -> 1, 127   127 -> 1, 127   127 -> 1, 127
+    128 -> 8, 16    128 -> 2, 64     128 -> 4, 32    128 -> 8, 16    128 -> 8, 16
+    250 -> 10, 25   250 -> 2, 125    250 -> 2, 125   250 -> 5, 50    250 -> 10, 25
+    360 -> 8, 45    360 -> 2, 180    360 -> 4, 90    360 -> 8, 45    360 -> 12, 30
+    512 -> 16, 32   512 -> 2, 256    512 -> 4, 128   512 -> 8, 64    512 -> 16, 32
+    1024 -> 32, 32  1024 -> 2, 512   1024 -> 4, 256  1024 -> 8, 128  1024 -> 16, 64
+    """
+
+    if factor > 0 and (dimension % factor) == 0 and dimension >= factor**2:
+        m = factor
+        n = dimension // factor
+        if m > n:
+            n, m = m, n
+        return m, n
+    if factor < 0:
+        factor = dimension
+    m, n = 1, dimension
+    length = m + n
+    while m < n:
+        new_m = m + 1
+        while dimension % new_m != 0:
+            new_m += 1
+        new_n = dimension // new_m
+        if new_m + new_n > length or new_m > factor:
+            break
+        else:
+            m, n = new_m, new_n
+    if m > n:
+        n, m = m, n
+    return m, n

comfy/weight_adapter/loha.py

@@ -3,7 +3,120 @@ from typing import Optional
 
 import torch
 import comfy.model_management
-from .base import WeightAdapterBase, weight_decompose
+from .base import WeightAdapterBase, WeightAdapterTrainBase, weight_decompose
+
+
+class HadaWeight(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, w1u, w1d, w2u, w2d, scale=torch.tensor(1)):
+        ctx.save_for_backward(w1d, w1u, w2d, w2u, scale)
+        diff_weight = ((w1u @ w1d) * (w2u @ w2d)) * scale
+        return diff_weight
+
+    @staticmethod
+    def backward(ctx, grad_out):
+        (w1d, w1u, w2d, w2u, scale) = ctx.saved_tensors
+        grad_out = grad_out * scale
+        temp = grad_out * (w2u @ w2d)
+        grad_w1u = temp @ w1d.T
+        grad_w1d = w1u.T @ temp
+
+        temp = grad_out * (w1u @ w1d)
+        grad_w2u = temp @ w2d.T
+        grad_w2d = w2u.T @ temp
+
+        del temp
+        return grad_w1u, grad_w1d, grad_w2u, grad_w2d, None
+
+
+class HadaWeightTucker(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, t1, w1u, w1d, t2, w2u, w2d, scale=torch.tensor(1)):
+        ctx.save_for_backward(t1, w1d, w1u, t2, w2d, w2u, scale)
+
+        rebuild1 = torch.einsum("i j ..., j r, i p -> p r ...", t1, w1d, w1u)
+        rebuild2 = torch.einsum("i j ..., j r, i p -> p r ...", t2, w2d, w2u)
+
+        return rebuild1 * rebuild2 * scale
+
+    @staticmethod
+    def backward(ctx, grad_out):
+        (t1, w1d, w1u, t2, w2d, w2u, scale) = ctx.saved_tensors
+        grad_out = grad_out * scale
+
+        temp = torch.einsum("i j ..., j r -> i r ...", t2, w2d)
+        rebuild = torch.einsum("i j ..., i r -> r j ...", temp, w2u)
+
+        grad_w = rebuild * grad_out
+        del rebuild
+
+        grad_w1u = torch.einsum("r j ..., i j ... -> r i", temp, grad_w)
+        grad_temp = torch.einsum("i j ..., i r -> r j ...", grad_w, w1u.T)
+        del grad_w, temp
+
+        grad_w1d = torch.einsum("i r ..., i j ... -> r j", t1, grad_temp)
+        grad_t1 = torch.einsum("i j ..., j r -> i r ...", grad_temp, w1d.T)
+        del grad_temp
+
+        temp = torch.einsum("i j ..., j r -> i r ...", t1, w1d)
+        rebuild = torch.einsum("i j ..., i r -> r j ...", temp, w1u)
+
+        grad_w = rebuild * grad_out
+        del rebuild
+
+        grad_w2u = torch.einsum("r j ..., i j ... -> r i", temp, grad_w)
+        grad_temp = torch.einsum("i j ..., i r -> r j ...", grad_w, w2u.T)
+        del grad_w, temp
+
+        grad_w2d = torch.einsum("i r ..., i j ... -> r j", t2, grad_temp)
+        grad_t2 = torch.einsum("i j ..., j r -> i r ...", grad_temp, w2d.T)
+        del grad_temp
+        return grad_t1, grad_w1u, grad_w1d, grad_t2, grad_w2u, grad_w2d, None
+
+
+class LohaDiff(WeightAdapterTrainBase):
+    def __init__(self, weights):
+        super().__init__()
+        # Unpack weights tuple from LoHaAdapter
+        w1a, w1b, alpha, w2a, w2b, t1, t2, _ = weights
+
+        # Create trainable parameters
+        self.hada_w1_a = torch.nn.Parameter(w1a)
+        self.hada_w1_b = torch.nn.Parameter(w1b)
+        self.hada_w2_a = torch.nn.Parameter(w2a)
+        self.hada_w2_b = torch.nn.Parameter(w2b)
+
+        self.use_tucker = False
+        if t1 is not None and t2 is not None:
+            self.use_tucker = True
+            self.hada_t1 = torch.nn.Parameter(t1)
+            self.hada_t2 = torch.nn.Parameter(t2)
+        else:
+            # Keep the attributes for consistent access
+            self.hada_t1 = None
+            self.hada_t2 = None
+
+        # Store rank and non-trainable alpha
+        self.rank = w1b.shape[0]
+        self.alpha = torch.nn.Parameter(torch.tensor(alpha), requires_grad=False)
+
+    def __call__(self, w):
+        org_dtype = w.dtype
+
+        scale = self.alpha / self.rank
+        if self.use_tucker:
+            diff_weight = HadaWeightTucker.apply(self.hada_t1, self.hada_w1_a, self.hada_w1_b, self.hada_t2, self.hada_w2_a, self.hada_w2_b, scale)
+        else:
+            diff_weight = HadaWeight.apply(self.hada_w1_a, self.hada_w1_b, self.hada_w2_a, self.hada_w2_b, scale)
+
+        # Add the scaled difference to the original weight
+        weight = w.to(diff_weight) + diff_weight.reshape(w.shape)
+
+        return weight.to(org_dtype)
+
+    def passive_memory_usage(self):
+        """Calculates memory usage of the trainable parameters."""
+        return sum(param.numel() * param.element_size() for param in self.parameters())
 
 
 class LoHaAdapter(WeightAdapterBase):
@@ -13,6 +126,25 @@ class LoHaAdapter(WeightAdapterBase):
         self.loaded_keys = loaded_keys
         self.weights = weights
 
+    @classmethod
+    def create_train(cls, weight, rank=1, alpha=1.0):
+        out_dim = weight.shape[0]
+        in_dim = weight.shape[1:].numel()
+        mat1 = torch.empty(out_dim, rank, device=weight.device, dtype=weight.dtype)
+        mat2 = torch.empty(rank, in_dim, device=weight.device, dtype=weight.dtype)
+        torch.nn.init.normal_(mat1, 0.1)
+        torch.nn.init.constant_(mat2, 0.0)
+        mat3 = torch.empty(out_dim, rank, device=weight.device, dtype=weight.dtype)
+        mat4 = torch.empty(rank, in_dim, device=weight.device, dtype=weight.dtype)
+        torch.nn.init.normal_(mat3, 0.1)
+        torch.nn.init.normal_(mat4, 0.01)
+        return LohaDiff(
+            (mat1, mat2, alpha, mat3, mat4, None, None, None)
+        )
+
+    def to_train(self):
+        return LohaDiff(self.weights)
+
     @classmethod
     def load(
         cls,

comfy/weight_adapter/lokr.py

@@ -3,7 +3,77 @@ from typing import Optional
 
 import torch
 import comfy.model_management
-from .base import WeightAdapterBase, weight_decompose
+from .base import (
+    WeightAdapterBase,
+    WeightAdapterTrainBase,
+    weight_decompose,
+    factorization,
+)
+
+
+class LokrDiff(WeightAdapterTrainBase):
+    def __init__(self, weights):
+        super().__init__()
+        (lokr_w1, lokr_w2, alpha, lokr_w1_a, lokr_w1_b, lokr_w2_a, lokr_w2_b, lokr_t2, dora_scale) = weights
+        self.use_tucker = False
+        if lokr_w1_a is not None:
+            _, rank_a = lokr_w1_a.shape[0], lokr_w1_a.shape[1]
+            rank_a, _ = lokr_w1_b.shape[0], lokr_w1_b.shape[1]
+            self.lokr_w1_a = torch.nn.Parameter(lokr_w1_a)
+            self.lokr_w1_b = torch.nn.Parameter(lokr_w1_b)
+            self.w1_rebuild = True
+            self.ranka = rank_a
+
+        if lokr_w2_a is not None:
+            _, rank_b = lokr_w2_a.shape[0], lokr_w2_a.shape[1]
+            rank_b, _ = lokr_w2_b.shape[0], lokr_w2_b.shape[1]
+            self.lokr_w2_a = torch.nn.Parameter(lokr_w2_a)
+            self.lokr_w2_b = torch.nn.Parameter(lokr_w2_b)
+            if lokr_t2 is not None:
+                self.use_tucker = True
+                self.lokr_t2 = torch.nn.Parameter(lokr_t2)
+            self.w2_rebuild = True
+            self.rankb = rank_b
+
+        if lokr_w1 is not None:
+            self.lokr_w1 = torch.nn.Parameter(lokr_w1)
+            self.w1_rebuild = False
+
+        if lokr_w2 is not None:
+            self.lokr_w2 = torch.nn.Parameter(lokr_w2)
+            self.w2_rebuild = False
+
+        self.alpha = torch.nn.Parameter(torch.tensor(alpha), requires_grad=False)
+
+    @property
+    def w1(self):
+        if self.w1_rebuild:
+            return (self.lokr_w1_a @ self.lokr_w1_b) * (self.alpha / self.ranka)
+        else:
+            return self.lokr_w1
+
+    @property
+    def w2(self):
+        if self.w2_rebuild:
+            if self.use_tucker:
+                w2 = torch.einsum(
+                    'i j k l, j r, i p -> p r k l',
+                    self.lokr_t2,
+                    self.lokr_w2_b,
+                    self.lokr_w2_a
+                )
+            else:
+                w2 = self.lokr_w2_a @ self.lokr_w2_b
+            return w2 * (self.alpha / self.rankb)
+        else:
+            return self.lokr_w2
+
+    def __call__(self, w):
+        diff = torch.kron(self.w1, self.w2)
+        return w + diff.reshape(w.shape).to(w)
+
+    def passive_memory_usage(self):
+        return sum(param.numel() * param.element_size() for param in self.parameters())
 
 
 class LoKrAdapter(WeightAdapterBase):
@@ -13,6 +83,20 @@ class LoKrAdapter(WeightAdapterBase):
         self.loaded_keys = loaded_keys
         self.weights = weights
 
+    @classmethod
+    def create_train(cls, weight, rank=1, alpha=1.0):
+        out_dim = weight.shape[0]
+        in_dim = weight.shape[1:].numel()
+        out1, out2 = factorization(out_dim, rank)
+        in1, in2 = factorization(in_dim, rank)
+        mat1 = torch.empty(out1, in1, device=weight.device, dtype=weight.dtype)
+        mat2 = torch.empty(out2, in2, device=weight.device, dtype=weight.dtype)
+        torch.nn.init.kaiming_uniform_(mat2, a=5**0.5)
+        torch.nn.init.constant_(mat1, 0.0)
+        return LokrDiff(
+            (mat1, mat2, alpha, None, None, None, None, None, None)
+        )
+
     @classmethod
     def load(
         cls,

comfy/weight_adapter/oft.py

@@ -3,7 +3,58 @@ from typing import Optional
 
 import torch
 import comfy.model_management
-from .base import WeightAdapterBase, weight_decompose
+from .base import WeightAdapterBase, WeightAdapterTrainBase, weight_decompose, factorization
+
+
+class OFTDiff(WeightAdapterTrainBase):
+    def __init__(self, weights):
+        super().__init__()
+        # Unpack weights tuple from LoHaAdapter
+        blocks, rescale, alpha, _ = weights
+
+        # Create trainable parameters
+        self.oft_blocks = torch.nn.Parameter(blocks)
+        if rescale is not None:
+            self.rescale = torch.nn.Parameter(rescale)
+            self.rescaled = True
+        else:
+            self.rescaled = False
+        self.block_num, self.block_size, _ = blocks.shape
+        self.constraint = float(alpha)
+        self.alpha = torch.nn.Parameter(torch.tensor(alpha), requires_grad=False)
+
+    def __call__(self, w):
+        org_dtype = w.dtype
+        I = torch.eye(self.block_size, device=self.oft_blocks.device)
+
+        ## generate r
+        # for Q = -Q^T
+        q = self.oft_blocks - self.oft_blocks.transpose(1, 2)
+        normed_q = q
+        if self.constraint:
+            q_norm = torch.norm(q) + 1e-8
+            if q_norm > self.constraint:
+                normed_q = q * self.constraint / q_norm
+        # use float() to prevent unsupported type
+        r = (I + normed_q) @ (I - normed_q).float().inverse()
+
+        ## Apply chunked matmul on weight
+        _, *shape = w.shape
+        org_weight = w.to(dtype=r.dtype)
+        org_weight = org_weight.unflatten(0, (self.block_num, self.block_size))
+        # Init R=0, so add I on it to ensure the output of step0 is original model output
+        weight = torch.einsum(
+            "k n m, k n ... -> k m ...",
+            r,
+            org_weight,
+        ).flatten(0, 1)
+        if self.rescaled:
+            weight = self.rescale * weight
+        return weight.to(org_dtype)
+
+    def passive_memory_usage(self):
+        """Calculates memory usage of the trainable parameters."""
+        return sum(param.numel() * param.element_size() for param in self.parameters())
 
 
 class OFTAdapter(WeightAdapterBase):
@@ -13,6 +64,18 @@ class OFTAdapter(WeightAdapterBase):
         self.loaded_keys = loaded_keys
         self.weights = weights
 
+    @classmethod
+    def create_train(cls, weight, rank=1, alpha=1.0):
+        out_dim = weight.shape[0]
+        block_size, block_num = factorization(out_dim, rank)
+        block = torch.zeros(block_num, block_size, block_size, device=weight.device, dtype=weight.dtype)
+        return OFTDiff(
+            (block, None, alpha, None)
+        )
+
+    def to_train(self):
+        return OFTDiff(self.weights)
+
     @classmethod
     def load(
         cls,
@@ -60,6 +123,8 @@ class OFTAdapter(WeightAdapterBase):
         blocks = v[0]
         rescale = v[1]
         alpha = v[2]
+        if alpha is None:
+            alpha = 0
         dora_scale = v[3]
 
         blocks = comfy.model_management.cast_to_device(blocks, weight.device, intermediate_dtype)

comfy_api/feature_flags.py

@@ -0,0 +1,69 @@
+"""
+Feature flags module for ComfyUI WebSocket protocol negotiation.
+
+This module handles capability negotiation between frontend and backend,
+allowing graceful protocol evolution while maintaining backward compatibility.
+"""
+
+from typing import Any, Dict
+
+from comfy.cli_args import args
+
+# Default server capabilities
+SERVER_FEATURE_FLAGS: Dict[str, Any] = {
+    "supports_preview_metadata": True,
+    "max_upload_size": args.max_upload_size * 1024 * 1024, # Convert MB to bytes
+}
+
+
+def get_connection_feature(
+    sockets_metadata: Dict[str, Dict[str, Any]],
+    sid: str,
+    feature_name: str,
+    default: Any = False
+) -> Any:
+    """
+    Get a feature flag value for a specific connection.
+
+    Args:
+        sockets_metadata: Dictionary of socket metadata
+        sid: Session ID of the connection
+        feature_name: Name of the feature to check
+        default: Default value if feature not found
+
+    Returns:
+        Feature value or default if not found
+    """
+    if sid not in sockets_metadata:
+        return default
+
+    return sockets_metadata[sid].get("feature_flags", {}).get(feature_name, default)
+
+
+def supports_feature(
+    sockets_metadata: Dict[str, Dict[str, Any]],
+    sid: str,
+    feature_name: str
+) -> bool:
+    """
+    Check if a connection supports a specific feature.
+
+    Args:
+        sockets_metadata: Dictionary of socket metadata
+        sid: Session ID of the connection
+        feature_name: Name of the feature to check
+
+    Returns:
+        Boolean indicating if feature is supported
+    """
+    return get_connection_feature(sockets_metadata, sid, feature_name, False) is True
+
+
+def get_server_features() -> Dict[str, Any]:
+    """
+    Get the server's feature flags.
+
+    Returns:
+        Dictionary of server feature flags
+    """
+    return SERVER_FEATURE_FLAGS.copy()

comfy_api/input/video_types.py

@@ -1,6 +1,8 @@
 from __future__ import annotations
 from abc import ABC, abstractmethod
-from typing import Optional
+from typing import Optional, Union
+import io
+import av
 from comfy_api.util import VideoContainer, VideoCodec, VideoComponents
 
 class VideoInput(ABC):
@@ -31,6 +33,22 @@ class VideoInput(ABC):
         """
         pass
 
+    def get_stream_source(self) -> Union[str, io.BytesIO]:
+        """
+        Get a streamable source for the video. This allows processing without
+        loading the entire video into memory.
+
+        Returns:
+            Either a file path (str) or a BytesIO object that can be opened with av.
+
+        Default implementation creates a BytesIO buffer, but subclasses should
+        override this for better performance when possible.
+        """
+        buffer = io.BytesIO()
+        self.save_to(buffer)
+        buffer.seek(0)
+        return buffer
+
     # Provide a default implementation, but subclasses can provide optimized versions
     # if possible.
     def get_dimensions(self) -> tuple[int, int]:
@@ -53,3 +71,15 @@ class VideoInput(ABC):
         components = self.get_components()
         frame_count = components.images.shape[0]
         return float(frame_count / components.frame_rate)
+
+    def get_container_format(self) -> str:
+        """
+        Returns the container format of the video (e.g., 'mp4', 'mov', 'avi').
+
+        Returns:
+            Container format as string
+        """
+        # Default implementation - subclasses should override for better performance
+        source = self.get_stream_source()
+        with av.open(source, mode="r") as container:
+            return container.format.name

comfy_api/input_impl/video_types.py

@@ -64,6 +64,15 @@ class VideoFromFile(VideoInput):
         """
         self.__file = file
 
+    def get_stream_source(self) -> str | io.BytesIO:
+        """
+        Return the underlying file source for efficient streaming.
+        This avoids unnecessary memory copies when the source is already a file path.
+        """
+        if isinstance(self.__file, io.BytesIO):
+            self.__file.seek(0)
+        return self.__file
+
     def get_dimensions(self) -> tuple[int, int]:
         """
         Returns the dimensions of the video input.
@@ -112,6 +121,18 @@ class VideoFromFile(VideoInput):
 
         raise ValueError(f"Could not determine duration for file '{self.__file}'")
 
+    def get_container_format(self) -> str:
+        """
+        Returns the container format of the video (e.g., 'mp4', 'mov', 'avi').
+
+        Returns:
+            Container format as string
+        """
+        if isinstance(self.__file, io.BytesIO):
+            self.__file.seek(0)
+        with av.open(self.__file, mode='r') as container:
+            return container.format.name
+
     def get_components_internal(self, container: InputContainer) -> VideoComponents:
         # Get video frames
         frames = []

comfy_api_nodes/README.md

@@ -2,7 +2,7 @@
 
 ## Introduction 
 
-Below are a collection of nodes that work by calling external APIs. More information available in our [docs](https://docs.comfy.org/tutorials/api-nodes/overview#api-nodes).
+Below are a collection of nodes that work by calling external APIs. More information available in our [docs](https://docs.comfy.org/tutorials/api-nodes/overview).
 
 ## Development
 

comfy_api_nodes/apis/__init__.py

@@ -1,6 +1,6 @@
 # generated by datamodel-codegen:
 #   filename:  filtered-openapi.yaml
-#   timestamp: 2025-05-19T21:38:55+00:00
+#   timestamp: 2025-07-06T09:47:31+00:00
 
 from __future__ import annotations
 
@@ -1355,6 +1355,158 @@ class ModelResponseProperties(BaseModel):
     )
 
 
+class Keyframes(BaseModel):
+    image_url: Optional[str] = None
+
+
+class MoonvalleyPromptResponse(BaseModel):
+    error: Optional[Dict[str, Any]] = None
+    frame_conditioning: Optional[Dict[str, Any]] = None
+    id: Optional[str] = None
+    inference_params: Optional[Dict[str, Any]] = None
+    meta: Optional[Dict[str, Any]] = None
+    model_params: Optional[Dict[str, Any]] = None
+    output_url: Optional[str] = None
+    prompt_text: Optional[str] = None
+    status: Optional[str] = None
+
+
+class MoonvalleyTextToVideoInferenceParams(BaseModel):
+    add_quality_guidance: Optional[bool] = Field(
+        True, description='Whether to add quality guidance'
+    )
+    caching_coefficient: Optional[float] = Field(
+        0.3, description='Caching coefficient for optimization'
+    )
+    caching_cooldown: Optional[int] = Field(
+        3, description='Number of caching cooldown steps'
+    )
+    caching_warmup: Optional[int] = Field(
+        3, description='Number of caching warmup steps'
+    )
+    clip_value: Optional[float] = Field(
+        3, description='CLIP value for generation control'
+    )
+    conditioning_frame_index: Optional[int] = Field(
+        0, description='Index of the conditioning frame'
+    )
+    cooldown_steps: Optional[int] = Field(
+        None, description='Number of cooldown steps (calculated based on num_frames)'
+    )
+    fps: Optional[int] = Field(
+        24, description='Frames per second of the generated video'
+    )
+    guidance_scale: Optional[float] = Field(
+        12.5, description='Guidance scale for generation control'
+    )
+    height: Optional[int] = Field(
+        1080, description='Height of the generated video in pixels'
+    )
+    negative_prompt: Optional[str] = Field(None, description='Negative prompt text')
+    num_frames: Optional[int] = Field(64, description='Number of frames to generate')
+    seed: Optional[int] = Field(
+        None, description='Random seed for generation (default: random)'
+    )
+    shift_value: Optional[float] = Field(
+        3, description='Shift value for generation control'
+    )
+    steps: Optional[int] = Field(80, description='Number of denoising steps')
+    use_guidance_schedule: Optional[bool] = Field(
+        True, description='Whether to use guidance scheduling'
+    )
+    use_negative_prompts: Optional[bool] = Field(
+        False, description='Whether to use negative prompts'
+    )
+    use_timestep_transform: Optional[bool] = Field(
+        True, description='Whether to use timestep transformation'
+    )
+    warmup_steps: Optional[int] = Field(
+        None, description='Number of warmup steps (calculated based on num_frames)'
+    )
+    width: Optional[int] = Field(
+        1920, description='Width of the generated video in pixels'
+    )
+
+
+class MoonvalleyTextToVideoRequest(BaseModel):
+    image_url: Optional[str] = None
+    inference_params: Optional[MoonvalleyTextToVideoInferenceParams] = None
+    prompt_text: Optional[str] = None
+    webhook_url: Optional[str] = None
+
+
+class MoonvalleyUploadFileRequest(BaseModel):
+    file: Optional[StrictBytes] = None
+
+
+class MoonvalleyUploadFileResponse(BaseModel):
+    access_url: Optional[str] = None
+
+
+class MoonvalleyVideoToVideoInferenceParams(BaseModel):
+    add_quality_guidance: Optional[bool] = Field(
+        True, description='Whether to add quality guidance'
+    )
+    caching_coefficient: Optional[float] = Field(
+        0.3, description='Caching coefficient for optimization'
+    )
+    caching_cooldown: Optional[int] = Field(
+        3, description='Number of caching cooldown steps'
+    )
+    caching_warmup: Optional[int] = Field(
+        3, description='Number of caching warmup steps'
+    )
+    clip_value: Optional[float] = Field(
+        3, description='CLIP value for generation control'
+    )
+    conditioning_frame_index: Optional[int] = Field(
+        0, description='Index of the conditioning frame'
+    )
+    cooldown_steps: Optional[int] = Field(
+        None, description='Number of cooldown steps (calculated based on num_frames)'
+    )
+    guidance_scale: Optional[float] = Field(
+        12.5, description='Guidance scale for generation control'
+    )
+    negative_prompt: Optional[str] = Field(None, description='Negative prompt text')
+    seed: Optional[int] = Field(
+        None, description='Random seed for generation (default: random)'
+    )
+    shift_value: Optional[float] = Field(
+        3, description='Shift value for generation control'
+    )
+    steps: Optional[int] = Field(80, description='Number of denoising steps')
+    use_guidance_schedule: Optional[bool] = Field(
+        True, description='Whether to use guidance scheduling'
+    )
+    use_negative_prompts: Optional[bool] = Field(
+        False, description='Whether to use negative prompts'
+    )
+    use_timestep_transform: Optional[bool] = Field(
+        True, description='Whether to use timestep transformation'
+    )
+    warmup_steps: Optional[int] = Field(
+        None, description='Number of warmup steps (calculated based on num_frames)'
+    )
+
+
+class ControlType(str, Enum):
+    motion_control = 'motion_control'
+    pose_control = 'pose_control'
+
+
+class MoonvalleyVideoToVideoRequest(BaseModel):
+    control_type: ControlType = Field(
+        ..., description='Supported types for video control'
+    )
+    inference_params: Optional[MoonvalleyVideoToVideoInferenceParams] = None
+    prompt_text: str = Field(..., description='Describes the video to generate')
+    video_url: str = Field(..., description='Url to control video')
+    webhook_url: Optional[str] = Field(
+        None, description='Optional webhook URL for notifications'
+    )
+
+
 class Moderation(str, Enum):
     low = 'low'
     auto = 'auto'
@@ -3107,6 +3259,23 @@ class LumaUpscaleVideoGenerationRequest(BaseModel):
     resolution: Optional[LumaVideoModelOutputResolution] = None
 
 
+class MoonvalleyImageToVideoRequest(MoonvalleyTextToVideoRequest):
+    keyframes: Optional[Dict[str, Keyframes]] = None
+
+
+class MoonvalleyResizeVideoRequest(MoonvalleyVideoToVideoRequest):
+    frame_position: Optional[List[int]] = Field(None, max_length=2, min_length=2)
+    frame_resolution: Optional[List[int]] = Field(None, max_length=2, min_length=2)
+    scale: Optional[List[int]] = Field(None, max_length=2, min_length=2)
+
+
+class MoonvalleyTextToImageRequest(BaseModel):
+    image_url: Optional[str] = None
+    inference_params: Optional[MoonvalleyTextToVideoInferenceParams] = None
+    prompt_text: Optional[str] = None
+    webhook_url: Optional[str] = None
+
+
 class OutputContent(RootModel[Union[OutputTextContent, OutputAudioContent]]):
     root: Union[OutputTextContent, OutputAudioContent]
 

comfy_api_nodes/nodes_gemini.py

@@ -406,7 +406,7 @@ class GeminiInputFiles(ComfyNodeABC):
 
     def create_file_part(self, file_path: str) -> GeminiPart:
         mime_type = (
-            GeminiMimeType.pdf
+            GeminiMimeType.application_pdf
             if file_path.endswith(".pdf")
             else GeminiMimeType.text_plain
         )

comfy_api_nodes/nodes_kling.py

@@ -132,6 +132,8 @@ def poll_until_finished(
         result_url_extractor=result_url_extractor,
         estimated_duration=estimated_duration,
         node_id=node_id,
+        poll_interval=16.0,
+        max_poll_attempts=256,
     ).execute()
 
 

comfy_api_nodes/nodes_moonvalley.py

@@ -0,0 +1,743 @@
+import logging
+from typing import Any, Callable, Optional, TypeVar
+import random
+import torch
+from comfy_api_nodes.util.validation_utils import (
+    get_image_dimensions,
+    validate_image_dimensions,
+)
+
+
+from comfy_api_nodes.apis import (
+    MoonvalleyTextToVideoRequest,
+    MoonvalleyTextToVideoInferenceParams,
+    MoonvalleyVideoToVideoInferenceParams,
+    MoonvalleyVideoToVideoRequest,
+    MoonvalleyPromptResponse,
+)
+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,
+    upload_video_to_comfyapi,
+)
+from comfy_api_nodes.mapper_utils import model_field_to_node_input
+
+from comfy_api.input.video_types import VideoInput
+from comfy.comfy_types.node_typing import IO
+from comfy_api.input_impl import VideoFromFile
+import av
+import io
+
+API_UPLOADS_ENDPOINT = "/proxy/moonvalley/uploads"
+API_PROMPTS_ENDPOINT = "/proxy/moonvalley/prompts"
+API_VIDEO2VIDEO_ENDPOINT = "/proxy/moonvalley/prompts/video-to-video"
+API_TXT2VIDEO_ENDPOINT = "/proxy/moonvalley/prompts/text-to-video"
+API_IMG2VIDEO_ENDPOINT = "/proxy/moonvalley/prompts/image-to-video"
+
+MIN_WIDTH = 300
+MIN_HEIGHT = 300
+
+MAX_WIDTH = 10000
+MAX_HEIGHT = 10000
+
+MIN_VID_WIDTH = 300
+MIN_VID_HEIGHT = 300
+
+MAX_VID_WIDTH = 10000
+MAX_VID_HEIGHT = 10000
+
+MAX_VIDEO_SIZE = 1024 * 1024 * 1024  # 1 GB max for in-memory video processing
+
+MOONVALLEY_MAREY_MAX_PROMPT_LENGTH = 5000
+R = TypeVar("R")
+
+
+class MoonvalleyApiError(Exception):
+    """Base exception for Moonvalley API errors."""
+
+    pass
+
+
+def is_valid_task_creation_response(response: MoonvalleyPromptResponse) -> bool:
+    """Verifies that the initial response contains a task ID."""
+    return bool(response.id)
+
+
+def validate_task_creation_response(response) -> None:
+    if not is_valid_task_creation_response(response):
+        error_msg = f"Moonvalley Marey API: Initial request failed. Code: {response.code}, Message: {response.message}, Data: {response}"
+        logging.error(error_msg)
+        raise MoonvalleyApiError(error_msg)
+
+
+def get_video_from_response(response):
+    video = response.output_url
+    logging.info(
+        "Moonvalley Marey API: Task %s succeeded. Video URL: %s", response.id, video
+    )
+    return video
+
+
+def get_video_url_from_response(response) -> Optional[str]:
+    """Returns the first video url from the Moonvalley video generation task result.
+    Will not raise an error if the response is not valid.
+    """
+    if response:
+        return str(get_video_from_response(response))
+    else:
+        return None
+
+
+def poll_until_finished(
+    auth_kwargs: dict[str, str],
+    api_endpoint: ApiEndpoint[Any, R],
+    result_url_extractor: Optional[Callable[[R], str]] = None,
+    node_id: Optional[str] = None,
+) -> R:
+    """Polls the Moonvalley API endpoint until the task reaches a terminal state, then returns the response."""
+    return PollingOperation(
+        poll_endpoint=api_endpoint,
+        completed_statuses=[
+            "completed",
+        ],
+        max_poll_attempts=240,  # 64 minutes with 16s interval
+        poll_interval=16.0,
+        failed_statuses=["error"],
+        status_extractor=lambda response: (
+            response.status if response and response.status else None
+        ),
+        auth_kwargs=auth_kwargs,
+        result_url_extractor=result_url_extractor,
+        node_id=node_id,
+    ).execute()
+
+
+def validate_prompts(
+    prompt: str, negative_prompt: str, max_length=MOONVALLEY_MAREY_MAX_PROMPT_LENGTH
+):
+    """Verifies that the prompt isn't empty and that neither prompt is too long."""
+    if not prompt:
+        raise ValueError("Positive prompt is empty")
+    if len(prompt) > max_length:
+        raise ValueError(f"Positive prompt is too long: {len(prompt)} characters")
+    if negative_prompt and len(negative_prompt) > max_length:
+        raise ValueError(
+            f"Negative prompt is too long: {len(negative_prompt)} characters"
+        )
+    return True
+
+
+def validate_input_media(width, height, with_frame_conditioning, num_frames_in=None):
+    # inference validation
+    # T = num_frames
+    # in all cases, the following must be true: T divisible by 16 and H,W by 8. in addition...
+    # with image conditioning: H*W must be divisible by 8192
+    # without image conditioning: T divisible by 32
+    if num_frames_in and not num_frames_in % 16 == 0:
+        return False, ("The input video total frame count must be divisible by 16!")
+
+    if height % 8 != 0 or width % 8 != 0:
+        return False, (
+            f"Height ({height}) and width ({width}) must be " "divisible by 8"
+        )
+
+    if with_frame_conditioning:
+        if (height * width) % 8192 != 0:
+            return False, (
+                f"Height * width ({height * width}) must be "
+                "divisible by 8192 for frame conditioning"
+            )
+    else:
+        if num_frames_in and not num_frames_in % 32 == 0:
+            return False, ("The input video total frame count must be divisible by 32!")
+
+
+def validate_input_image(
+    image: torch.Tensor, with_frame_conditioning: bool = False
+) -> None:
+    """
+    Validates the input image adheres to the expectations of the API:
+    - The image resolution should not be less than 300*300px
+    - The aspect ratio of the image should be between 1:2.5 ~ 2.5:1
+
+    """
+    height, width = get_image_dimensions(image)
+    validate_input_media(width, height, with_frame_conditioning)
+    validate_image_dimensions(
+        image, min_width=300, min_height=300, max_height=MAX_HEIGHT, max_width=MAX_WIDTH
+    )
+
+
+def validate_video_to_video_input(video: VideoInput) -> VideoInput:
+    """
+    Validates and processes video input for Moonvalley Video-to-Video generation.
+
+    Args:
+        video: Input video to validate
+
+    Returns:
+        Validated and potentially trimmed video
+
+    Raises:
+        ValueError: If video doesn't meet requirements
+        MoonvalleyApiError: If video duration is too short
+    """
+    width, height = _get_video_dimensions(video)
+    _validate_video_dimensions(width, height)
+    _validate_container_format(video)
+
+    return _validate_and_trim_duration(video)
+
+
+def _get_video_dimensions(video: VideoInput) -> tuple[int, int]:
+    """Extracts video dimensions with error handling."""
+    try:
+        return video.get_dimensions()
+    except Exception as e:
+        logging.error("Error getting dimensions of video: %s", e)
+        raise ValueError(f"Cannot get video dimensions: {e}") from e
+
+
+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)
+    }
+
+    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}")
+
+
+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}")
+
+
+def _validate_and_trim_duration(video: VideoInput) -> VideoInput:
+    """Validates video duration and trims to 5 seconds if needed."""
+    duration = video.get_duration()
+    _validate_minimum_duration(duration)
+    return _trim_if_too_long(video, duration)
+
+
+def _validate_minimum_duration(duration: float) -> None:
+    """Ensures video is at least 5 seconds long."""
+    if duration < 5:
+        raise MoonvalleyApiError("Input video must be at least 5 seconds long.")
+
+
+def _trim_if_too_long(video: VideoInput, duration: float) -> VideoInput:
+    """Trims video to 5 seconds if longer."""
+    if duration > 5:
+        return trim_video(video, 5)
+    return video
+
+
+
+def trim_video(video: VideoInput, duration_sec: float) -> VideoInput:
+    """
+    Returns a new VideoInput object trimmed from the beginning to the specified duration,
+    using av to avoid loading entire video into memory.
+
+    Args:
+        video: Input video to trim
+        duration_sec: Duration in seconds to keep from the beginning
+
+    Returns:
+        VideoFromFile object that owns the output buffer
+    """
+    output_buffer = io.BytesIO()
+
+    input_container = None
+    output_container = None
+
+    try:
+        # Get the stream source - this avoids loading entire video into memory
+        # when the source is already a file path
+        input_source = video.get_stream_source()
+
+        # Open containers
+        input_container = av.open(input_source, mode="r")
+        output_container = av.open(output_buffer, mode="w", format="mp4")
+
+        # Set up output streams for re-encoding
+        video_stream = None
+        audio_stream = None
+
+        for stream in input_container.streams:
+            logging.info(f"Found stream: type={stream.type}, class={type(stream)}")
+            if isinstance(stream, av.VideoStream):
+                # Create output video stream with same parameters
+                video_stream = output_container.add_stream(
+                    "h264", rate=stream.average_rate
+                )
+                video_stream.width = stream.width
+                video_stream.height = stream.height
+                video_stream.pix_fmt = "yuv420p"
+                logging.info(
+                    f"Added video stream: {stream.width}x{stream.height} @ {stream.average_rate}fps"
+                )
+            elif isinstance(stream, av.AudioStream):
+                # Create output audio stream with same parameters
+                audio_stream = output_container.add_stream(
+                    "aac", rate=stream.sample_rate
+                )
+                audio_stream.sample_rate = stream.sample_rate
+                audio_stream.layout = stream.layout
+                logging.info(
+                    f"Added audio stream: {stream.sample_rate}Hz, {stream.channels} channels"
+                )
+
+        # Calculate target frame count that's divisible by 16
+        fps = input_container.streams.video[0].average_rate
+        estimated_frames = int(duration_sec * fps)
+        target_frames = (estimated_frames // 16) * 16  # Round down to nearest multiple of 16
+
+        if target_frames == 0:
+            raise ValueError("Video too short: need at least 16 frames for Moonvalley")
+
+        frame_count = 0
+        audio_frame_count = 0
+
+        # Decode and re-encode video frames
+        if video_stream:
+            for frame in input_container.decode(video=0):
+                if frame_count >= target_frames:
+                    break
+
+                # Re-encode frame
+                for packet in video_stream.encode(frame):
+                    output_container.mux(packet)
+                frame_count += 1
+
+            # Flush encoder
+            for packet in video_stream.encode():
+                output_container.mux(packet)
+
+            logging.info(
+                f"Encoded {frame_count} video frames (target: {target_frames})"
+            )
+
+        # Decode and re-encode audio frames
+        if audio_stream:
+            input_container.seek(0)  # Reset to beginning for audio
+            for frame in input_container.decode(audio=0):
+                if frame.time >= duration_sec:
+                    break
+
+                # Re-encode frame
+                for packet in audio_stream.encode(frame):
+                    output_container.mux(packet)
+                audio_frame_count += 1
+
+            # Flush encoder
+            for packet in audio_stream.encode():
+                output_container.mux(packet)
+
+            logging.info(f"Encoded {audio_frame_count} audio frames")
+
+        # Close containers
+        output_container.close()
+        input_container.close()
+
+        # Return as VideoFromFile using the buffer
+        output_buffer.seek(0)
+        return VideoFromFile(output_buffer)
+
+    except Exception as e:
+        # Clean up on error
+        if input_container is not None:
+            input_container.close()
+        if output_container is not None:
+            output_container.close()
+        raise RuntimeError(f"Failed to trim video: {str(e)}") from e
+
+
+# --- BaseMoonvalleyVideoNode ---
+class BaseMoonvalleyVideoNode:
+    def parseWidthHeightFromRes(self, resolution: str):
+        # Accepts a string like "16:9 (1920 x 1080)" and returns width, height as a dict
+        res_map = {
+            "16:9 (1920 x 1080)": {"width": 1920, "height": 1080},
+            "9:16 (1080 x 1920)": {"width": 1080, "height": 1920},
+            "1:1 (1152 x 1152)": {"width": 1152, "height": 1152},
+            "4:3 (1536 x 1152)": {"width": 1536, "height": 1152},
+            "3:4 (1152 x 1536)": {"width": 1152, "height": 1536},
+            "21:9 (2560 x 1080)": {"width": 2560, "height": 1080},
+        }
+        if resolution in res_map:
+            return res_map[resolution]
+        else:
+            # Default to 1920x1080 if unknown
+            return {"width": 1920, "height": 1080}
+
+    def parseControlParameter(self, value):
+        control_map = {
+            "Motion Transfer": "motion_control",
+            "Canny": "canny_control",
+            "Pose Transfer": "pose_control",
+            "Depth": "depth_control",
+        }
+        if value in control_map:
+            return control_map[value]
+        else:
+            return control_map["Motion Transfer"]
+
+    def get_response(
+        self, task_id: str, auth_kwargs: dict[str, str], node_id: Optional[str] = None
+    ) -> MoonvalleyPromptResponse:
+        return poll_until_finished(
+            auth_kwargs,
+            ApiEndpoint(
+                path=f"{API_PROMPTS_ENDPOINT}/{task_id}",
+                method=HttpMethod.GET,
+                request_model=EmptyRequest,
+                response_model=MoonvalleyPromptResponse,
+            ),
+            result_url_extractor=get_video_url_from_response,
+            node_id=node_id,
+        )
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "prompt": model_field_to_node_input(
+                    IO.STRING,
+                    MoonvalleyTextToVideoRequest,
+                    "prompt_text",
+                    multiline=True,
+                ),
+                "negative_prompt": model_field_to_node_input(
+                    IO.STRING,
+                    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",
+                ),
+                "resolution": (
+                    IO.COMBO,
+                    {
+                        "options": [
+                            "16:9 (1920 x 1080)",
+                            "9:16 (1080 x 1920)",
+                            "1:1 (1152 x 1152)",
+                            "4:3 (1440 x 1080)",
+                            "3:4 (1080 x 1440)",
+                            "21:9 (2560 x 1080)",
+                        ],
+                        "default": "16:9 (1920 x 1080)",
+                        "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,
+                    step=1,
+                    min=1,
+                    max=20,
+                ),
+                "seed": model_field_to_node_input(
+                    IO.INT,
+                    MoonvalleyTextToVideoInferenceParams,
+                    "seed",
+                    default=random.randint(0, 2**32 - 1),
+                    min=0,
+                    max=4294967295,
+                    step=1,
+                    display="number",
+                    tooltip="Random seed value",
+                    control_after_generate=True,
+                ),
+                "steps": model_field_to_node_input(
+                    IO.INT,
+                    MoonvalleyTextToVideoInferenceParams,
+                    "steps",
+                    default=100,
+                    min=1,
+                    max=100,
+                ),
+            },
+            "hidden": {
+                "auth_token": "AUTH_TOKEN_COMFY_ORG",
+                "comfy_api_key": "API_KEY_COMFY_ORG",
+                "unique_id": "UNIQUE_ID",
+            },
+            "optional": {
+                "image": model_field_to_node_input(
+                    IO.IMAGE,
+                    MoonvalleyTextToVideoRequest,
+                    "image_url",
+                    tooltip="The reference image used to generate the video",
+                ),
+            },
+        }
+
+    RETURN_TYPES = ("STRING",)
+    FUNCTION = "generate"
+    CATEGORY = "api node/video/Moonvalley Marey"
+    API_NODE = True
+
+    def generate(self, **kwargs):
+        return None
+
+
+# --- MoonvalleyImg2VideoNode ---
+class MoonvalleyImg2VideoNode(BaseMoonvalleyVideoNode):
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return super().INPUT_TYPES()
+
+    RETURN_TYPES = ("VIDEO",)
+    RETURN_NAMES = ("video",)
+    DESCRIPTION = "Moonvalley Marey Image to Video Node"
+
+    def generate(
+        self, prompt, negative_prompt, unique_id: Optional[str] = None, **kwargs
+    ):
+        image = kwargs.get("image", None)
+        if image is None:
+            raise MoonvalleyApiError("image is required")
+
+        validate_input_image(image, True)
+        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"),
+            use_negative_prompts=True,
+        )
+        """Upload image to comfy backend to have a URL available for further processing"""
+        # Get MIME type from tensor - assuming PNG format for image tensors
+        mime_type = "image/png"
+
+        image_url = 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
+        )
+        initial_operation = SynchronousOperation(
+            endpoint=ApiEndpoint(
+                path=API_IMG2VIDEO_ENDPOINT,
+                method=HttpMethod.POST,
+                request_model=MoonvalleyTextToVideoRequest,
+                response_model=MoonvalleyPromptResponse,
+            ),
+            request=request,
+            auth_kwargs=kwargs,
+        )
+        task_creation_response = initial_operation.execute()
+        validate_task_creation_response(task_creation_response)
+        task_id = task_creation_response.id
+
+        final_response = self.get_response(
+            task_id, auth_kwargs=kwargs, node_id=unique_id
+        )
+        video = download_url_to_video_output(final_response.output_url)
+        return (video,)
+
+
+# --- MoonvalleyVid2VidNode ---
+class MoonvalleyVideo2VideoNode(BaseMoonvalleyVideoNode):
+    def __init__(self):
+        super().__init__()
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "prompt": model_field_to_node_input(
+                    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"
+                ),
+                "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",
+                "comfy_api_key": "API_KEY_COMFY_ORG",
+                "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."}),
+                "control_type": (
+                    ["Motion Transfer", "Pose Transfer"],
+                    {"default": "Motion Transfer"},
+                ),
+                "motion_intensity": (
+                    "INT",
+                    {
+                        "default": 100,
+                        "step": 1,
+                        "min": 0,
+                        "max": 100,
+                        "tooltip": "Only used if control_type is 'Motion Transfer'",
+                    },
+                )
+            }
+        }
+
+    RETURN_TYPES = ("VIDEO",)
+    RETURN_NAMES = ("video",)
+
+    def generate(
+        self, prompt, negative_prompt, unique_id: Optional[str] = None, **kwargs
+    ):
+        video = kwargs.get("video")
+
+        if not video:
+            raise MoonvalleyApiError("video is required")
+
+        video_url = ""
+        if video:
+            validated_video = validate_video_to_video_input(video)
+            video_url = upload_video_to_comfyapi(validated_video, auth_kwargs=kwargs)
+
+        control_type = kwargs.get("control_type")
+        motion_intensity = kwargs.get("motion_intensity")
+
+        """Validate prompts and inference input"""
+        validate_prompts(prompt, negative_prompt)
+
+        # 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
+
+        inference_params=MoonvalleyVideoToVideoInferenceParams(
+            negative_prompt=negative_prompt,
+            seed=kwargs.get("seed"),
+            control_params=control_params
+        )
+
+        control = self.parseControlParameter(control_type)
+
+        request = MoonvalleyVideoToVideoRequest(
+            control_type=control,
+            video_url=video_url,
+            prompt_text=prompt,
+            inference_params=inference_params,
+        )
+
+        initial_operation = SynchronousOperation(
+            endpoint=ApiEndpoint(
+                path=API_VIDEO2VIDEO_ENDPOINT,
+                method=HttpMethod.POST,
+                request_model=MoonvalleyVideoToVideoRequest,
+                response_model=MoonvalleyPromptResponse,
+            ),
+            request=request,
+            auth_kwargs=kwargs,
+        )
+        task_creation_response = initial_operation.execute()
+        validate_task_creation_response(task_creation_response)
+        task_id = task_creation_response.id
+
+        final_response = self.get_response(
+            task_id, auth_kwargs=kwargs, node_id=unique_id
+        )
+
+        video = download_url_to_video_output(final_response.output_url)
+
+        return (video,)
+
+
+# --- MoonvalleyTxt2VideoNode ---
+class MoonvalleyTxt2VideoNode(BaseMoonvalleyVideoNode):
+    def __init__(self):
+        super().__init__()
+
+    RETURN_TYPES = ("VIDEO",)
+    RETURN_NAMES = ("video",)
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        input_types = super().INPUT_TYPES()
+        # Remove image-specific parameters
+        for param in ["image"]:
+            if param in input_types["optional"]:
+                del input_types["optional"][param]
+        return input_types
+
+    def generate(
+        self, prompt, negative_prompt, unique_id: Optional[str] = None, **kwargs
+    ):
+        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"),
+                )
+        request = MoonvalleyTextToVideoRequest(
+            prompt_text=prompt, inference_params=inference_params
+        )
+
+        initial_operation = SynchronousOperation(
+            endpoint=ApiEndpoint(
+                path=API_TXT2VIDEO_ENDPOINT,
+                method=HttpMethod.POST,
+                request_model=MoonvalleyTextToVideoRequest,
+                response_model=MoonvalleyPromptResponse,
+            ),
+            request=request,
+            auth_kwargs=kwargs,
+        )
+        task_creation_response = initial_operation.execute()
+        validate_task_creation_response(task_creation_response)
+        task_id = task_creation_response.id
+
+        final_response = self.get_response(
+            task_id, auth_kwargs=kwargs, node_id=unique_id
+        )
+
+        video = download_url_to_video_output(final_response.output_url)
+        return (video,)
+
+
+NODE_CLASS_MAPPINGS = {
+    "MoonvalleyImg2VideoNode": MoonvalleyImg2VideoNode,
+    "MoonvalleyTxt2VideoNode": MoonvalleyTxt2VideoNode,
+    "MoonvalleyVideo2VideoNode": MoonvalleyVideo2VideoNode,
+}
+
+
+NODE_DISPLAY_NAME_MAPPINGS = {
+    "MoonvalleyImg2VideoNode": "Moonvalley Marey Image to Video",
+    "MoonvalleyTxt2VideoNode": "Moonvalley Marey Text to Video",
+    "MoonvalleyVideo2VideoNode": "Moonvalley Marey Video to Video",
+}

comfy_execution/caching.py

@@ -1,6 +1,7 @@
 import itertools
 from typing import Sequence, Mapping, Dict
 from comfy_execution.graph import DynamicPrompt
+from abc import ABC, abstractmethod
 
 import nodes
 
@@ -16,12 +17,13 @@ def include_unique_id_in_input(class_type: str) -> bool:
     NODE_CLASS_CONTAINS_UNIQUE_ID[class_type] = "UNIQUE_ID" in class_def.INPUT_TYPES().get("hidden", {}).values()
     return NODE_CLASS_CONTAINS_UNIQUE_ID[class_type]
 
-class CacheKeySet:
+class CacheKeySet(ABC):
     def __init__(self, dynprompt, node_ids, is_changed_cache):
         self.keys = {}
         self.subcache_keys = {}
 
-    def add_keys(self, node_ids):
+    @abstractmethod
+    async def add_keys(self, node_ids):
         raise NotImplementedError()
 
     def all_node_ids(self):
@@ -60,9 +62,8 @@ class CacheKeySetID(CacheKeySet):
     def __init__(self, dynprompt, node_ids, is_changed_cache):
         super().__init__(dynprompt, node_ids, is_changed_cache)
         self.dynprompt = dynprompt
-        self.add_keys(node_ids)
 
-    def add_keys(self, node_ids):
+    async def add_keys(self, node_ids):
         for node_id in node_ids:
             if node_id in self.keys:
                 continue
@@ -77,37 +78,36 @@ class CacheKeySetInputSignature(CacheKeySet):
         super().__init__(dynprompt, node_ids, is_changed_cache)
         self.dynprompt = dynprompt
         self.is_changed_cache = is_changed_cache
-        self.add_keys(node_ids)
 
     def include_node_id_in_input(self) -> bool:
         return False
 
-    def add_keys(self, node_ids):
+    async def add_keys(self, node_ids):
         for node_id in node_ids:
             if node_id in self.keys:
                 continue
             if not self.dynprompt.has_node(node_id):
                 continue
             node = self.dynprompt.get_node(node_id)
-            self.keys[node_id] = self.get_node_signature(self.dynprompt, node_id)
+            self.keys[node_id] = await self.get_node_signature(self.dynprompt, node_id)
             self.subcache_keys[node_id] = (node_id, node["class_type"])
 
-    def get_node_signature(self, dynprompt, node_id):
+    async def get_node_signature(self, dynprompt, node_id):
         signature = []
         ancestors, order_mapping = self.get_ordered_ancestry(dynprompt, node_id)
-        signature.append(self.get_immediate_node_signature(dynprompt, node_id, order_mapping))
+        signature.append(await self.get_immediate_node_signature(dynprompt, node_id, order_mapping))
         for ancestor_id in ancestors:
-            signature.append(self.get_immediate_node_signature(dynprompt, ancestor_id, order_mapping))
+            signature.append(await self.get_immediate_node_signature(dynprompt, ancestor_id, order_mapping))
         return to_hashable(signature)
 
-    def get_immediate_node_signature(self, dynprompt, node_id, ancestor_order_mapping):
+    async def get_immediate_node_signature(self, dynprompt, node_id, ancestor_order_mapping):
         if not dynprompt.has_node(node_id):
             # This node doesn't exist -- we can't cache it.
             return [float("NaN")]
         node = dynprompt.get_node(node_id)
         class_type = node["class_type"]
         class_def = nodes.NODE_CLASS_MAPPINGS[class_type]
-        signature = [class_type, self.is_changed_cache.get(node_id)]
+        signature = [class_type, await self.is_changed_cache.get(node_id)]
         if self.include_node_id_in_input() or (hasattr(class_def, "NOT_IDEMPOTENT") and class_def.NOT_IDEMPOTENT) or include_unique_id_in_input(class_type):
             signature.append(node_id)
         inputs = node["inputs"]
@@ -150,9 +150,10 @@ class BasicCache:
         self.cache = {}
         self.subcaches = {}
 
-    def set_prompt(self, dynprompt, node_ids, is_changed_cache):
+    async def set_prompt(self, dynprompt, node_ids, is_changed_cache):
         self.dynprompt = dynprompt
         self.cache_key_set = self.key_class(dynprompt, node_ids, is_changed_cache)
+        await self.cache_key_set.add_keys(node_ids)
         self.is_changed_cache = is_changed_cache
         self.initialized = True
 
@@ -201,13 +202,13 @@ class BasicCache:
         else:
             return None
 
-    def _ensure_subcache(self, node_id, children_ids):
+    async def _ensure_subcache(self, node_id, children_ids):
         subcache_key = self.cache_key_set.get_subcache_key(node_id)
         subcache = self.subcaches.get(subcache_key, None)
         if subcache is None:
             subcache = BasicCache(self.key_class)
             self.subcaches[subcache_key] = subcache
-        subcache.set_prompt(self.dynprompt, children_ids, self.is_changed_cache)
+        await subcache.set_prompt(self.dynprompt, children_ids, self.is_changed_cache)
         return subcache
 
     def _get_subcache(self, node_id):
@@ -259,10 +260,10 @@ class HierarchicalCache(BasicCache):
         assert cache is not None
         cache._set_immediate(node_id, value)
 
-    def ensure_subcache_for(self, node_id, children_ids):
+    async def ensure_subcache_for(self, node_id, children_ids):
         cache = self._get_cache_for(node_id)
         assert cache is not None
-        return cache._ensure_subcache(node_id, children_ids)
+        return await cache._ensure_subcache(node_id, children_ids)
 
 class LRUCache(BasicCache):
     def __init__(self, key_class, max_size=100):
@@ -273,8 +274,8 @@ class LRUCache(BasicCache):
         self.used_generation = {}
         self.children = {}
 
-    def set_prompt(self, dynprompt, node_ids, is_changed_cache):
-        super().set_prompt(dynprompt, node_ids, is_changed_cache)
+    async def set_prompt(self, dynprompt, node_ids, is_changed_cache):
+        await super().set_prompt(dynprompt, node_ids, is_changed_cache)
         self.generation += 1
         for node_id in node_ids:
             self._mark_used(node_id)
@@ -303,11 +304,11 @@ class LRUCache(BasicCache):
         self._mark_used(node_id)
         return self._set_immediate(node_id, value)
 
-    def ensure_subcache_for(self, node_id, children_ids):
+    async def ensure_subcache_for(self, node_id, children_ids):
         # Just uses subcaches for tracking 'live' nodes
-        super()._ensure_subcache(node_id, children_ids)
+        await super()._ensure_subcache(node_id, children_ids)
 
-        self.cache_key_set.add_keys(children_ids)
+        await self.cache_key_set.add_keys(children_ids)
         self._mark_used(node_id)
         cache_key = self.cache_key_set.get_data_key(node_id)
         self.children[cache_key] = []
@@ -337,7 +338,7 @@ class DependencyAwareCache(BasicCache):
         self.ancestors = {}    # Maps node_id -> set of ancestor node_ids
         self.executed_nodes = set()  # Tracks nodes that have been executed
 
-    def set_prompt(self, dynprompt, node_ids, is_changed_cache):
+    async def set_prompt(self, dynprompt, node_ids, is_changed_cache):
         """
         Clear the entire cache and rebuild the dependency graph.
 
@@ -354,7 +355,7 @@ class DependencyAwareCache(BasicCache):
         self.executed_nodes.clear()
 
         # Call the parent method to initialize the cache with the new prompt
-        super().set_prompt(dynprompt, node_ids, is_changed_cache)
+        await super().set_prompt(dynprompt, node_ids, is_changed_cache)
 
         # Rebuild the dependency graph
         self._build_dependency_graph(dynprompt, node_ids)
@@ -405,7 +406,7 @@ class DependencyAwareCache(BasicCache):
         """
         return self._get_immediate(node_id)
 
-    def ensure_subcache_for(self, node_id, children_ids):
+    async def ensure_subcache_for(self, node_id, children_ids):
         """
         Ensure a subcache exists for a node and update dependencies.
 
@@ -416,7 +417,7 @@ class DependencyAwareCache(BasicCache):
         Returns:
             The subcache object for the node.
         """
-        subcache = super()._ensure_subcache(node_id, children_ids)
+        subcache = await super()._ensure_subcache(node_id, children_ids)
         for child_id in children_ids:
             self.descendants[node_id].add(child_id)
             self.ancestors[child_id].add(node_id)

comfy_execution/graph.py

@@ -2,6 +2,8 @@ from __future__ import annotations
 from typing import Type, Literal
 
 import nodes
+import asyncio
+import inspect
 from comfy_execution.graph_utils import is_link
 from comfy.comfy_types.node_typing import ComfyNodeABC, InputTypeDict, InputTypeOptions
 
@@ -100,6 +102,8 @@ class TopologicalSort:
         self.pendingNodes = {}
         self.blockCount = {} # Number of nodes this node is directly blocked by
         self.blocking = {} # Which nodes are blocked by this node
+        self.externalBlocks = 0
+        self.unblockedEvent = asyncio.Event()
 
     def get_input_info(self, unique_id, input_name):
         class_type = self.dynprompt.get_node(unique_id)["class_type"]
@@ -153,6 +157,16 @@ class TopologicalSort:
         for link in links:
             self.add_strong_link(*link)
 
+    def add_external_block(self, node_id):
+        assert node_id in self.blockCount, "Can't add external block to a node that isn't pending"
+        self.externalBlocks += 1
+        self.blockCount[node_id] += 1
+        def unblock():
+            self.externalBlocks -= 1
+            self.blockCount[node_id] -= 1
+            self.unblockedEvent.set()
+        return unblock
+
     def is_cached(self, node_id):
         return False
 
@@ -181,11 +195,16 @@ class ExecutionList(TopologicalSort):
     def is_cached(self, node_id):
         return self.output_cache.get(node_id) is not None
 
-    def stage_node_execution(self):
+    async def stage_node_execution(self):
         assert self.staged_node_id is None
         if self.is_empty():
             return None, None, None
         available = self.get_ready_nodes()
+        while len(available) == 0 and self.externalBlocks > 0:
+            # Wait for an external block to be released
+            await self.unblockedEvent.wait()
+            self.unblockedEvent.clear()
+            available = self.get_ready_nodes()
         if len(available) == 0:
             cycled_nodes = self.get_nodes_in_cycle()
             # Because cycles composed entirely of static nodes are caught during initial validation,
@@ -221,8 +240,15 @@ class ExecutionList(TopologicalSort):
                 return True
             return False
 
+        # If an available node is async, do that first.
+        # This will execute the asynchronous function earlier, reducing the overall time.
+        def is_async(node_id):
+            class_type = self.dynprompt.get_node(node_id)["class_type"]
+            class_def = nodes.NODE_CLASS_MAPPINGS[class_type]
+            return inspect.iscoroutinefunction(getattr(class_def, class_def.FUNCTION))
+
         for node_id in node_list:
-            if is_output(node_id):
+            if is_output(node_id) or is_async(node_id):
                 return node_id
 
         #This should handle the VAEDecode -> preview case

comfy_execution/progress.py

@@ -0,0 +1,347 @@
+from typing import TypedDict, Dict, Optional
+from typing_extensions import override
+from PIL import Image
+from enum import Enum
+from abc import ABC
+from tqdm import tqdm
+from typing import TYPE_CHECKING
+if TYPE_CHECKING:
+    from comfy_execution.graph import DynamicPrompt
+from protocol import BinaryEventTypes
+from comfy_api import feature_flags
+
+
+class NodeState(Enum):
+    Pending = "pending"
+    Running = "running"
+    Finished = "finished"
+    Error = "error"
+
+
+class NodeProgressState(TypedDict):
+    """
+    A class to represent the state of a node's progress.
+    """
+
+    state: NodeState
+    value: float
+    max: float
+
+
+class ProgressHandler(ABC):
+    """
+    Abstract base class for progress handlers.
+    Progress handlers receive progress updates and display them in various ways.
+    """
+
+    def __init__(self, name: str):
+        self.name = name
+        self.enabled = True
+
+    def set_registry(self, registry: "ProgressRegistry"):
+        pass
+
+    def start_handler(self, node_id: str, state: NodeProgressState, prompt_id: str):
+        """Called when a node starts processing"""
+        pass
+
+    def update_handler(
+        self,
+        node_id: str,
+        value: float,
+        max_value: float,
+        state: NodeProgressState,
+        prompt_id: str,
+        image: Optional[Image.Image] = None,
+    ):
+        """Called when a node's progress is updated"""
+        pass
+
+    def finish_handler(self, node_id: str, state: NodeProgressState, prompt_id: str):
+        """Called when a node finishes processing"""
+        pass
+
+    def reset(self):
+        """Called when the progress registry is reset"""
+        pass
+
+    def enable(self):
+        """Enable this handler"""
+        self.enabled = True
+
+    def disable(self):
+        """Disable this handler"""
+        self.enabled = False
+
+
+class CLIProgressHandler(ProgressHandler):
+    """
+    Handler that displays progress using tqdm progress bars in the CLI.
+    """
+
+    def __init__(self):
+        super().__init__("cli")
+        self.progress_bars: Dict[str, tqdm] = {}
+
+    @override
+    def start_handler(self, node_id: str, state: NodeProgressState, prompt_id: str):
+        # Create a new tqdm progress bar
+        if node_id not in self.progress_bars:
+            self.progress_bars[node_id] = tqdm(
+                total=state["max"],
+                desc=f"Node {node_id}",
+                unit="steps",
+                leave=True,
+                position=len(self.progress_bars),
+            )
+
+    @override
+    def update_handler(
+        self,
+        node_id: str,
+        value: float,
+        max_value: float,
+        state: NodeProgressState,
+        prompt_id: str,
+        image: Optional[Image.Image] = None,
+    ):
+        # Handle case where start_handler wasn't called
+        if node_id not in self.progress_bars:
+            self.progress_bars[node_id] = tqdm(
+                total=max_value,
+                desc=f"Node {node_id}",
+                unit="steps",
+                leave=True,
+                position=len(self.progress_bars),
+            )
+            self.progress_bars[node_id].update(value)
+        else:
+            # Update existing progress bar
+            if max_value != self.progress_bars[node_id].total:
+                self.progress_bars[node_id].total = max_value
+            # Calculate the update amount (difference from current position)
+            current_position = self.progress_bars[node_id].n
+            update_amount = value - current_position
+            if update_amount > 0:
+                self.progress_bars[node_id].update(update_amount)
+
+    @override
+    def finish_handler(self, node_id: str, state: NodeProgressState, prompt_id: str):
+        # Complete and close the progress bar if it exists
+        if node_id in self.progress_bars:
+            # Ensure the bar shows 100% completion
+            remaining = state["max"] - self.progress_bars[node_id].n
+            if remaining > 0:
+                self.progress_bars[node_id].update(remaining)
+            self.progress_bars[node_id].close()
+            del self.progress_bars[node_id]
+
+    @override
+    def reset(self):
+        # Close all progress bars
+        for bar in self.progress_bars.values():
+            bar.close()
+        self.progress_bars.clear()
+
+
+class WebUIProgressHandler(ProgressHandler):
+    """
+    Handler that sends progress updates to the WebUI via WebSockets.
+    """
+
+    def __init__(self, server_instance):
+        super().__init__("webui")
+        self.server_instance = server_instance
+
+    def set_registry(self, registry: "ProgressRegistry"):
+        self.registry = registry
+
+    def _send_progress_state(self, prompt_id: str, nodes: Dict[str, NodeProgressState]):
+        """Send the current progress state to the client"""
+        if self.server_instance is None:
+            return
+
+        # Only send info for non-pending nodes
+        active_nodes = {
+            node_id: {
+                "value": state["value"],
+                "max": state["max"],
+                "state": state["state"].value,
+                "node_id": node_id,
+                "prompt_id": prompt_id,
+                "display_node_id": self.registry.dynprompt.get_display_node_id(node_id),
+                "parent_node_id": self.registry.dynprompt.get_parent_node_id(node_id),
+                "real_node_id": self.registry.dynprompt.get_real_node_id(node_id),
+            }
+            for node_id, state in nodes.items()
+            if state["state"] != NodeState.Pending
+        }
+
+        # Send a combined progress_state message with all node states
+        self.server_instance.send_sync(
+            "progress_state", {"prompt_id": prompt_id, "nodes": active_nodes}
+        )
+
+    @override
+    def start_handler(self, node_id: str, state: NodeProgressState, prompt_id: str):
+        # Send progress state of all nodes
+        if self.registry:
+            self._send_progress_state(prompt_id, self.registry.nodes)
+
+    @override
+    def update_handler(
+        self,
+        node_id: str,
+        value: float,
+        max_value: float,
+        state: NodeProgressState,
+        prompt_id: str,
+        image: Optional[Image.Image] = None,
+    ):
+        # Send progress state of all nodes
+        if self.registry:
+            self._send_progress_state(prompt_id, self.registry.nodes)
+        if image:
+            # Only send new format if client supports it
+            if feature_flags.supports_feature(
+                self.server_instance.sockets_metadata,
+                self.server_instance.client_id,
+                "supports_preview_metadata",
+            ):
+                metadata = {
+                    "node_id": node_id,
+                    "prompt_id": prompt_id,
+                    "display_node_id": self.registry.dynprompt.get_display_node_id(
+                        node_id
+                    ),
+                    "parent_node_id": self.registry.dynprompt.get_parent_node_id(
+                        node_id
+                    ),
+                    "real_node_id": self.registry.dynprompt.get_real_node_id(node_id),
+                }
+                self.server_instance.send_sync(
+                    BinaryEventTypes.PREVIEW_IMAGE_WITH_METADATA,
+                    (image, metadata),
+                    self.server_instance.client_id,
+                )
+
+    @override
+    def finish_handler(self, node_id: str, state: NodeProgressState, prompt_id: str):
+        # Send progress state of all nodes
+        if self.registry:
+            self._send_progress_state(prompt_id, self.registry.nodes)
+
+
+class ProgressRegistry:
+    """
+    Registry that maintains node progress state and notifies registered handlers.
+    """
+
+    def __init__(self, prompt_id: str, dynprompt: "DynamicPrompt"):
+        self.prompt_id = prompt_id
+        self.dynprompt = dynprompt
+        self.nodes: Dict[str, NodeProgressState] = {}
+        self.handlers: Dict[str, ProgressHandler] = {}
+
+    def register_handler(self, handler: ProgressHandler) -> None:
+        """Register a progress handler"""
+        self.handlers[handler.name] = handler
+
+    def unregister_handler(self, handler_name: str) -> None:
+        """Unregister a progress handler"""
+        if handler_name in self.handlers:
+            # Allow handler to clean up resources
+            self.handlers[handler_name].reset()
+            del self.handlers[handler_name]
+
+    def enable_handler(self, handler_name: str) -> None:
+        """Enable a progress handler"""
+        if handler_name in self.handlers:
+            self.handlers[handler_name].enable()
+
+    def disable_handler(self, handler_name: str) -> None:
+        """Disable a progress handler"""
+        if handler_name in self.handlers:
+            self.handlers[handler_name].disable()
+
+    def ensure_entry(self, node_id: str) -> NodeProgressState:
+        """Ensure a node entry exists"""
+        if node_id not in self.nodes:
+            self.nodes[node_id] = NodeProgressState(
+                state=NodeState.Pending, value=0, max=1
+            )
+        return self.nodes[node_id]
+
+    def start_progress(self, node_id: str) -> None:
+        """Start progress tracking for a node"""
+        entry = self.ensure_entry(node_id)
+        entry["state"] = NodeState.Running
+        entry["value"] = 0.0
+        entry["max"] = 1.0
+
+        # Notify all enabled handlers
+        for handler in self.handlers.values():
+            if handler.enabled:
+                handler.start_handler(node_id, entry, self.prompt_id)
+
+    def update_progress(
+        self, node_id: str, value: float, max_value: float, image: Optional[Image.Image]
+    ) -> None:
+        """Update progress for a node"""
+        entry = self.ensure_entry(node_id)
+        entry["state"] = NodeState.Running
+        entry["value"] = value
+        entry["max"] = max_value
+
+        # Notify all enabled handlers
+        for handler in self.handlers.values():
+            if handler.enabled:
+                handler.update_handler(
+                    node_id, value, max_value, entry, self.prompt_id, image
+                )
+
+    def finish_progress(self, node_id: str) -> None:
+        """Finish progress tracking for a node"""
+        entry = self.ensure_entry(node_id)
+        entry["state"] = NodeState.Finished
+        entry["value"] = entry["max"]
+
+        # Notify all enabled handlers
+        for handler in self.handlers.values():
+            if handler.enabled:
+                handler.finish_handler(node_id, entry, self.prompt_id)
+
+    def reset_handlers(self) -> None:
+        """Reset all handlers"""
+        for handler in self.handlers.values():
+            handler.reset()
+
+# Global registry instance
+global_progress_registry: ProgressRegistry = None
+
+def reset_progress_state(prompt_id: str, dynprompt: "DynamicPrompt") -> None:
+    global global_progress_registry
+
+    # Reset existing handlers if registry exists
+    if global_progress_registry is not None:
+        global_progress_registry.reset_handlers()
+
+    # Create new registry
+    global_progress_registry = ProgressRegistry(prompt_id, dynprompt)
+
+
+def add_progress_handler(handler: ProgressHandler) -> None:
+    registry = get_progress_state()
+    handler.set_registry(registry)
+    registry.register_handler(handler)
+
+
+def get_progress_state() -> ProgressRegistry:
+    global global_progress_registry
+    if global_progress_registry is None:
+        from comfy_execution.graph import DynamicPrompt
+
+        global_progress_registry = ProgressRegistry(
+            prompt_id="", dynprompt=DynamicPrompt({})
+        )
+    return global_progress_registry

comfy_execution/utils.py

@@ -0,0 +1,46 @@
+import contextvars
+from typing import Optional, NamedTuple
+
+class ExecutionContext(NamedTuple):
+    """
+    Context information about the currently executing node.
+
+    Attributes:
+        node_id: The ID of the currently executing node
+        list_index: The index in a list being processed (for operations on batches/lists)
+    """
+    prompt_id: str
+    node_id: str
+    list_index: Optional[int]
+
+current_executing_context: contextvars.ContextVar[Optional[ExecutionContext]] = contextvars.ContextVar("current_executing_context", default=None)
+
+def get_executing_context() -> Optional[ExecutionContext]:
+    return current_executing_context.get(None)
+
+class CurrentNodeContext:
+    """
+    Context manager for setting the current executing node context.
+
+    Sets the current_executing_context on enter and resets it on exit.
+
+    Example:
+        with CurrentNodeContext(node_id="123", list_index=0):
+            # Code that should run with the current node context set
+            process_image()
+    """
+    def __init__(self, prompt_id: str, node_id: str, list_index: Optional[int] = None):
+        self.context = ExecutionContext(
+            prompt_id= prompt_id,
+            node_id= node_id,
+            list_index= list_index
+        )
+        self.token = None
+
+    def __enter__(self):
+        self.token = current_executing_context.set(self.context)
+        return self
+
+    def __exit__(self, exc_type, exc_val, exc_tb):
+        if self.token is not None:
+            current_executing_context.reset(self.token)

comfy_extras/nodes_audio.py

@@ -133,14 +133,6 @@ def save_audio(self, audio, filename_prefix="ComfyUI", format="flac", prompt=Non
             if sample_rate != audio["sample_rate"]:
                 waveform = torchaudio.functional.resample(waveform, audio["sample_rate"], sample_rate)
 
-        # Create in-memory WAV buffer
-        wav_buffer = io.BytesIO()
-        torchaudio.save(wav_buffer, waveform, sample_rate, format="WAV")
-        wav_buffer.seek(0)  # Rewind for reading
-
-        # Use PyAV to convert and add metadata
-        input_container = av.open(wav_buffer)
-
         # Create output with specified format
         output_buffer = io.BytesIO()
         output_container = av.open(output_buffer, mode='w', format=format)
@@ -150,7 +142,6 @@ def save_audio(self, audio, filename_prefix="ComfyUI", format="flac", prompt=Non
             output_container.metadata[key] = value
 
         # Set up the output stream with appropriate properties
-        input_container.streams.audio[0]
         if format == "opus":
             out_stream = output_container.add_stream("libopus", rate=sample_rate)
             if quality == "64k":
@@ -175,18 +166,16 @@ def save_audio(self, audio, filename_prefix="ComfyUI", format="flac", prompt=Non
         else: #format == "flac":
             out_stream = output_container.add_stream("flac", rate=sample_rate)
 
-
-        # Copy frames from input to output
-        for frame in input_container.decode(audio=0):
-            frame.pts = None  # Let PyAV handle timestamps
-            output_container.mux(out_stream.encode(frame))
+        frame = av.AudioFrame.from_ndarray(waveform.movedim(0, 1).reshape(1, -1).float().numpy(), format='flt', layout='mono' if waveform.shape[0] == 1 else 'stereo')
+        frame.sample_rate = sample_rate
+        frame.pts = 0
+        output_container.mux(out_stream.encode(frame))
 
         # Flush encoder
         output_container.mux(out_stream.encode(None))
 
         # Close containers
         output_container.close()
-        input_container.close()
 
         # Write the output to file
         output_buffer.seek(0)
@@ -289,6 +278,42 @@ class PreviewAudio(SaveAudio):
                 "hidden": {"prompt": "PROMPT", "extra_pnginfo": "EXTRA_PNGINFO"},
                 }
 
+def f32_pcm(wav: torch.Tensor) -> torch.Tensor:
+    """Convert audio to float 32 bits PCM format."""
+    if wav.dtype.is_floating_point:
+        return wav
+    elif wav.dtype == torch.int16:
+        return wav.float() / (2 ** 15)
+    elif wav.dtype == torch.int32:
+        return wav.float() / (2 ** 31)
+    raise ValueError(f"Unsupported wav dtype: {wav.dtype}")
+
+def load(filepath: str) -> tuple[torch.Tensor, int]:
+    with av.open(filepath) as af:
+        if not af.streams.audio:
+            raise ValueError("No audio stream found in the file.")
+
+        stream = af.streams.audio[0]
+        sr = stream.codec_context.sample_rate
+        n_channels = stream.channels
+
+        frames = []
+        length = 0
+        for frame in af.decode(streams=stream.index):
+            buf = torch.from_numpy(frame.to_ndarray())
+            if buf.shape[0] != n_channels:
+                buf = buf.view(-1, n_channels).t()
+
+            frames.append(buf)
+            length += buf.shape[1]
+
+        if not frames:
+            raise ValueError("No audio frames decoded.")
+
+        wav = torch.cat(frames, dim=1)
+        wav = f32_pcm(wav)
+        return wav, sr
+
 class LoadAudio:
     @classmethod
     def INPUT_TYPES(s):
@@ -303,7 +328,7 @@ class LoadAudio:
 
     def load(self, audio):
         audio_path = folder_paths.get_annotated_filepath(audio)
-        waveform, sample_rate = torchaudio.load(audio_path)
+        waveform, sample_rate = load(audio_path)
         audio = {"waveform": waveform.unsqueeze(0), "sample_rate": sample_rate}
         return (audio, )
 

comfy_extras/nodes_cfg.py

@@ -40,6 +40,33 @@ class CFGZeroStar:
         m.set_model_sampler_post_cfg_function(cfg_zero_star)
         return (m, )
 
+class CFGNorm:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {"model": ("MODEL",),
+                             "strength": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 100.0, "step": 0.01}),
+                            }}
+    RETURN_TYPES = ("MODEL",)
+    RETURN_NAMES = ("patched_model",)
+    FUNCTION = "patch"
+    CATEGORY = "advanced/guidance"
+    EXPERIMENTAL = True
+
+    def patch(self, model, strength):
+        m = model.clone()
+        def cfg_norm(args):
+            cond_p = args['cond_denoised']
+            pred_text_ = args["denoised"]
+
+            norm_full_cond = torch.norm(cond_p, dim=1, keepdim=True)
+            norm_pred_text = torch.norm(pred_text_, dim=1, keepdim=True)
+            scale = (norm_full_cond / (norm_pred_text + 1e-8)).clamp(min=0.0, max=1.0)
+            return pred_text_ * scale * strength
+
+        m.set_model_sampler_post_cfg_function(cfg_norm)
+        return (m, )
+
 NODE_CLASS_MAPPINGS = {
-    "CFGZeroStar": CFGZeroStar
+    "CFGZeroStar": CFGZeroStar,
+    "CFGNorm": CFGNorm,
 }

comfy_extras/nodes_custom_sampler.py

@@ -2,6 +2,7 @@ import math
 import comfy.samplers
 import comfy.sample
 from comfy.k_diffusion import sampling as k_diffusion_sampling
+from comfy.k_diffusion import sa_solver
 from comfy.comfy_types import IO, ComfyNodeABC, InputTypeDict
 import latent_preview
 import torch
@@ -300,6 +301,35 @@ class ExtendIntermediateSigmas:
 
         return (extended_sigmas,)
 
+
+class SamplingPercentToSigma:
+    @classmethod
+    def INPUT_TYPES(cls) -> InputTypeDict:
+        return {
+            "required": {
+                "model": (IO.MODEL, {}),
+                "sampling_percent": (IO.FLOAT, {"default": 0.0, "min": 0.0, "max": 1.0, "step": 0.0001}),
+                "return_actual_sigma": (IO.BOOLEAN, {"default": False, "tooltip": "Return the actual sigma value instead of the value used for interval checks.\nThis only affects results at 0.0 and 1.0."}),
+            }
+        }
+
+    RETURN_TYPES = (IO.FLOAT,)
+    RETURN_NAMES = ("sigma_value",)
+    CATEGORY = "sampling/custom_sampling/sigmas"
+
+    FUNCTION = "get_sigma"
+
+    def get_sigma(self, model, sampling_percent, return_actual_sigma):
+        model_sampling = model.get_model_object("model_sampling")
+        sigma_val = model_sampling.percent_to_sigma(sampling_percent)
+        if return_actual_sigma:
+            if sampling_percent == 0.0:
+                sigma_val = model_sampling.sigma_max.item()
+            elif sampling_percent == 1.0:
+                sigma_val = model_sampling.sigma_min.item()
+        return (sigma_val,)
+
+
 class KSamplerSelect:
     @classmethod
     def INPUT_TYPES(s):
@@ -521,6 +551,49 @@ class SamplerER_SDE(ComfyNodeABC):
         return (sampler,)
 
 
+class SamplerSASolver(ComfyNodeABC):
+    @classmethod
+    def INPUT_TYPES(cls) -> InputTypeDict:
+        return {
+            "required": {
+                "model": (IO.MODEL, {}),
+                "eta": (IO.FLOAT, {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01, "round": False},),
+                "sde_start_percent": (IO.FLOAT, {"default": 0.2, "min": 0.0, "max": 1.0, "step": 0.001},),
+                "sde_end_percent": (IO.FLOAT, {"default": 0.8, "min": 0.0, "max": 1.0, "step": 0.001},),
+                "s_noise": (IO.FLOAT, {"default": 1.0, "min": 0.0, "max": 100.0, "step": 0.01, "round": False},),
+                "predictor_order": (IO.INT, {"default": 3, "min": 1, "max": 6}),
+                "corrector_order": (IO.INT, {"default": 4, "min": 0, "max": 6}),
+                "use_pece": (IO.BOOLEAN, {}),
+                "simple_order_2": (IO.BOOLEAN, {}),
+            }
+        }
+
+    RETURN_TYPES = (IO.SAMPLER,)
+    CATEGORY = "sampling/custom_sampling/samplers"
+
+    FUNCTION = "get_sampler"
+
+    def get_sampler(self, model, eta, sde_start_percent, sde_end_percent, s_noise, predictor_order, corrector_order, use_pece, simple_order_2):
+        model_sampling = model.get_model_object("model_sampling")
+        start_sigma = model_sampling.percent_to_sigma(sde_start_percent)
+        end_sigma = model_sampling.percent_to_sigma(sde_end_percent)
+        tau_func = sa_solver.get_tau_interval_func(start_sigma, end_sigma, eta=eta)
+
+        sampler_name = "sa_solver"
+        sampler = comfy.samplers.ksampler(
+            sampler_name,
+            {
+                "tau_func": tau_func,
+                "s_noise": s_noise,
+                "predictor_order": predictor_order,
+                "corrector_order": corrector_order,
+                "use_pece": use_pece,
+                "simple_order_2": simple_order_2,
+            },
+        )
+        return (sampler,)
+
+
 class Noise_EmptyNoise:
     def __init__(self):
         self.seed = 0
@@ -639,9 +712,10 @@ class CFGGuider:
         return (guider,)
 
 class Guider_DualCFG(comfy.samplers.CFGGuider):
-    def set_cfg(self, cfg1, cfg2):
+    def set_cfg(self, cfg1, cfg2, nested=False):
         self.cfg1 = cfg1
         self.cfg2 = cfg2
+        self.nested = nested
 
     def set_conds(self, positive, middle, negative):
         middle = node_helpers.conditioning_set_values(middle, {"prompt_type": "negative"})
@@ -651,14 +725,20 @@ class Guider_DualCFG(comfy.samplers.CFGGuider):
         negative_cond = self.conds.get("negative", None)
         middle_cond = self.conds.get("middle", None)
         positive_cond = self.conds.get("positive", None)
-        if model_options.get("disable_cfg1_optimization", False) == False:
-            if math.isclose(self.cfg2, 1.0):
-                negative_cond = None
-                if math.isclose(self.cfg1, 1.0):
-                    middle_cond = None
 
-        out = comfy.samplers.calc_cond_batch(self.inner_model, [negative_cond, middle_cond, positive_cond], x, timestep, model_options)
-        return comfy.samplers.cfg_function(self.inner_model, out[1], out[0], self.cfg2, x, timestep, model_options=model_options, cond=middle_cond, uncond=negative_cond) + (out[2] - out[1]) * self.cfg1
+        if self.nested:
+            out = comfy.samplers.calc_cond_batch(self.inner_model, [negative_cond, middle_cond, positive_cond], x, timestep, model_options)
+            pred_text = comfy.samplers.cfg_function(self.inner_model, out[2], out[1], self.cfg1, x, timestep, model_options=model_options, cond=positive_cond, uncond=middle_cond)
+            return out[0] + self.cfg2 * (pred_text - out[0])
+        else:
+            if model_options.get("disable_cfg1_optimization", False) == False:
+                if math.isclose(self.cfg2, 1.0):
+                    negative_cond = None
+                    if math.isclose(self.cfg1, 1.0):
+                        middle_cond = None
+
+            out = comfy.samplers.calc_cond_batch(self.inner_model, [negative_cond, middle_cond, positive_cond], x, timestep, model_options)
+            return comfy.samplers.cfg_function(self.inner_model, out[1], out[0], self.cfg2, x, timestep, model_options=model_options, cond=middle_cond, uncond=negative_cond) + (out[2] - out[1]) * self.cfg1
 
 class DualCFGGuider:
     @classmethod
@@ -670,6 +750,7 @@ class DualCFGGuider:
                     "negative": ("CONDITIONING", ),
                     "cfg_conds": ("FLOAT", {"default": 8.0, "min": 0.0, "max": 100.0, "step":0.1, "round": 0.01}),
                     "cfg_cond2_negative": ("FLOAT", {"default": 8.0, "min": 0.0, "max": 100.0, "step":0.1, "round": 0.01}),
+                    "style": (["regular", "nested"],),
                      }
                 }
 
@@ -678,10 +759,10 @@ class DualCFGGuider:
     FUNCTION = "get_guider"
     CATEGORY = "sampling/custom_sampling/guiders"
 
-    def get_guider(self, model, cond1, cond2, negative, cfg_conds, cfg_cond2_negative):
+    def get_guider(self, model, cond1, cond2, negative, cfg_conds, cfg_cond2_negative, style):
         guider = Guider_DualCFG(model)
         guider.set_conds(cond1, cond2, negative)
-        guider.set_cfg(cfg_conds, cfg_cond2_negative)
+        guider.set_cfg(cfg_conds, cfg_cond2_negative, nested=(style == "nested"))
         return (guider,)
 
 class DisableNoise:
@@ -829,11 +910,13 @@ NODE_CLASS_MAPPINGS = {
     "SamplerDPMPP_2S_Ancestral": SamplerDPMPP_2S_Ancestral,
     "SamplerDPMAdaptative": SamplerDPMAdaptative,
     "SamplerER_SDE": SamplerER_SDE,
+    "SamplerSASolver": SamplerSASolver,
     "SplitSigmas": SplitSigmas,
     "SplitSigmasDenoise": SplitSigmasDenoise,
     "FlipSigmas": FlipSigmas,
     "SetFirstSigma": SetFirstSigma,
     "ExtendIntermediateSigmas": ExtendIntermediateSigmas,
+    "SamplingPercentToSigma": SamplingPercentToSigma,
 
     "CFGGuider": CFGGuider,
     "DualCFGGuider": DualCFGGuider,

comfy_extras/nodes_fresca.py

@@ -71,8 +71,11 @@ class FreSca:
     DESCRIPTION = "Applies frequency-dependent scaling to the guidance"
     def patch(self, model, scale_low, scale_high, freq_cutoff):
         def custom_cfg_function(args):
-            cond = args["conds_out"][0]
-            uncond = args["conds_out"][1]
+            conds_out = args["conds_out"]
+            if len(conds_out) <= 1 or None in args["conds"][:2]:
+                return conds_out
+            cond = conds_out[0]
+            uncond = conds_out[1]
 
             guidance = cond - uncond
             filtered_guidance = Fourier_filter(
@@ -83,7 +86,7 @@ class FreSca:
             )
             filtered_cond = filtered_guidance + uncond
 
-            return [filtered_cond, uncond]
+            return [filtered_cond, uncond] + conds_out[2:]
 
         m = model.clone()
         m.set_model_sampler_pre_cfg_function(custom_cfg_function)

comfy_extras/nodes_images.py

@@ -583,6 +583,49 @@ class GetImageSize:
 
         return width, height, batch_size
 
+class ImageRotate:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "image": (IO.IMAGE,),
+                              "rotation": (["none", "90 degrees", "180 degrees", "270 degrees"],),
+                              }}
+    RETURN_TYPES = (IO.IMAGE,)
+    FUNCTION = "rotate"
+
+    CATEGORY = "image/transform"
+
+    def rotate(self, image, rotation):
+        rotate_by = 0
+        if rotation.startswith("90"):
+            rotate_by = 1
+        elif rotation.startswith("180"):
+            rotate_by = 2
+        elif rotation.startswith("270"):
+            rotate_by = 3
+
+        image = torch.rot90(image, k=rotate_by, dims=[2, 1])
+        return (image,)
+
+class ImageFlip:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "image": (IO.IMAGE,),
+                              "flip_method": (["x-axis: vertically", "y-axis: horizontally"],),
+                              }}
+    RETURN_TYPES = (IO.IMAGE,)
+    FUNCTION = "flip"
+
+    CATEGORY = "image/transform"
+
+    def flip(self, image, flip_method):
+        if flip_method.startswith("x"):
+            image = torch.flip(image, dims=[1])
+        elif flip_method.startswith("y"):
+            image = torch.flip(image, dims=[2])
+
+        return (image,)
+
+
 NODE_CLASS_MAPPINGS = {
     "ImageCrop": ImageCrop,
     "RepeatImageBatch": RepeatImageBatch,
@@ -594,4 +637,6 @@ NODE_CLASS_MAPPINGS = {
     "ImageStitch": ImageStitch,
     "ResizeAndPadImage": ResizeAndPadImage,
     "GetImageSize": GetImageSize,
+    "ImageRotate": ImageRotate,
+    "ImageFlip": ImageFlip,
 }

comfy_extras/nodes_load_3d.py

@@ -5,6 +5,8 @@ import os
 from comfy.comfy_types import IO
 from comfy_api.input_impl import VideoFromFile
 
+from pathlib import Path
+
 
 def normalize_path(path):
     return path.replace('\\', '/')
@@ -16,7 +18,14 @@ class Load3D():
 
         os.makedirs(input_dir, exist_ok=True)
 
-        files = [normalize_path(os.path.join("3d", f)) for f in os.listdir(input_dir) if f.endswith(('.gltf', '.glb', '.obj', '.fbx', '.stl'))]
+        input_path = Path(input_dir)
+        base_path = Path(folder_paths.get_input_directory())
+
+        files = [
+            normalize_path(str(file_path.relative_to(base_path)))
+            for file_path in input_path.rglob("*")
+            if file_path.suffix.lower() in {'.gltf', '.glb', '.obj', '.fbx', '.stl'}
+        ]
 
         return {"required": {
             "model_file": (sorted(files), {"file_upload": True}),
@@ -61,7 +70,14 @@ class Load3DAnimation():
 
         os.makedirs(input_dir, exist_ok=True)
 
-        files = [normalize_path(os.path.join("3d", f)) for f in os.listdir(input_dir) if f.endswith(('.gltf', '.glb', '.fbx'))]
+        input_path = Path(input_dir)
+        base_path = Path(folder_paths.get_input_directory())
+
+        files = [
+            normalize_path(str(file_path.relative_to(base_path)))
+            for file_path in input_path.rglob("*")
+            if file_path.suffix.lower() in {'.gltf', '.glb', '.fbx'}
+        ]
 
         return {"required": {
             "model_file": (sorted(files), {"file_upload": True}),

comfy_extras/nodes_lt.py

@@ -134,8 +134,8 @@ class LTXVAddGuide:
         _, num_keyframes = get_keyframe_idxs(cond)
         latent_count = latent_length - num_keyframes
         frame_idx = frame_idx if frame_idx >= 0 else max((latent_count - 1) * time_scale_factor + 1 + frame_idx, 0)
-        if guide_length > 1:
-            frame_idx = frame_idx // time_scale_factor * time_scale_factor # frame index must be divisible by 8
+        if guide_length > 1 and frame_idx != 0:
+            frame_idx = (frame_idx - 1) // time_scale_factor * time_scale_factor + 1 # frame index - 1 must be divisible by 8 or frame_idx == 0
 
         latent_idx = (frame_idx + time_scale_factor - 1) // time_scale_factor
 
@@ -144,7 +144,7 @@ class LTXVAddGuide:
     def add_keyframe_index(self, cond, frame_idx, guiding_latent, scale_factors):
         keyframe_idxs, _ = get_keyframe_idxs(cond)
         _, latent_coords = self._patchifier.patchify(guiding_latent)
-        pixel_coords = latent_to_pixel_coords(latent_coords, scale_factors, True)
+        pixel_coords = latent_to_pixel_coords(latent_coords, scale_factors, causal_fix=frame_idx == 0)  # we need the causal fix only if we're placing the new latents at index 0
         pixel_coords[:, 0] += frame_idx
         if keyframe_idxs is None:
             keyframe_idxs = pixel_coords

comfy_extras/nodes_mask.py

@@ -152,7 +152,7 @@ class ImageColorToMask:
     def image_to_mask(self, image, color):
         temp = (torch.clamp(image, 0, 1.0) * 255.0).round().to(torch.int)
         temp = torch.bitwise_left_shift(temp[:,:,:,0], 16) + torch.bitwise_left_shift(temp[:,:,:,1], 8) + temp[:,:,:,2]
-        mask = torch.where(temp == color, 255, 0).float()
+        mask = torch.where(temp == color, 1.0, 0).float()
         return (mask,)
 
 class SolidMask:
@@ -247,7 +247,7 @@ class MaskComposite:
         visible_width, visible_height = (right - left, bottom - top,)
 
         source_portion = source[:, :visible_height, :visible_width]
-        destination_portion = destination[:, top:bottom, left:right]
+        destination_portion = output[:, top:bottom, left:right]
 
         if operation == "multiply":
             output[:, top:bottom, left:right] = destination_portion * source_portion

comfy_extras/nodes_pixart.py

@@ -1,24 +1,24 @@
-from nodes import MAX_RESOLUTION
-
-class CLIPTextEncodePixArtAlpha:
-    @classmethod
-    def INPUT_TYPES(s):
-        return {"required": {
-            "width": ("INT", {"default": 1024.0, "min": 0, "max": MAX_RESOLUTION}),
-            "height": ("INT", {"default": 1024.0, "min": 0, "max": MAX_RESOLUTION}),
-            # "aspect_ratio": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}),
-            "text": ("STRING", {"multiline": True, "dynamicPrompts": True}), "clip": ("CLIP", ),
-            }}
-
-    RETURN_TYPES = ("CONDITIONING",)
-    FUNCTION = "encode"
-    CATEGORY = "advanced/conditioning"
-    DESCRIPTION = "Encodes text and sets the resolution conditioning for PixArt Alpha. Does not apply to PixArt Sigma."
-
-    def encode(self, clip, width, height, text):
-        tokens = clip.tokenize(text)
-        return (clip.encode_from_tokens_scheduled(tokens, add_dict={"width": width, "height": height}),)
-
-NODE_CLASS_MAPPINGS = {
-    "CLIPTextEncodePixArtAlpha": CLIPTextEncodePixArtAlpha,
-}
+from nodes import MAX_RESOLUTION
+
+class CLIPTextEncodePixArtAlpha:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {
+            "width": ("INT", {"default": 1024.0, "min": 0, "max": MAX_RESOLUTION}),
+            "height": ("INT", {"default": 1024.0, "min": 0, "max": MAX_RESOLUTION}),
+            # "aspect_ratio": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}),
+            "text": ("STRING", {"multiline": True, "dynamicPrompts": True}), "clip": ("CLIP", ),
+            }}
+
+    RETURN_TYPES = ("CONDITIONING",)
+    FUNCTION = "encode"
+    CATEGORY = "advanced/conditioning"
+    DESCRIPTION = "Encodes text and sets the resolution conditioning for PixArt Alpha. Does not apply to PixArt Sigma."
+
+    def encode(self, clip, width, height, text):
+        tokens = clip.tokenize(text)
+        return (clip.encode_from_tokens_scheduled(tokens, add_dict={"width": width, "height": height}),)
+
+NODE_CLASS_MAPPINGS = {
+    "CLIPTextEncodePixArtAlpha": CLIPTextEncodePixArtAlpha,
+}

comfy_extras/nodes_slg.py

@@ -78,7 +78,75 @@ class SkipLayerGuidanceDiT:
 
         return (m, )
 
+class SkipLayerGuidanceDiTSimple:
+    '''
+    Simple version of the SkipLayerGuidanceDiT node that only modifies the uncond pass.
+    '''
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {"model": ("MODEL", ),
+                             "double_layers": ("STRING", {"default": "7, 8, 9", "multiline": False}),
+                             "single_layers": ("STRING", {"default": "7, 8, 9", "multiline": False}),
+                             "start_percent": ("FLOAT", {"default": 0.0, "min": 0.0, "max": 1.0, "step": 0.001}),
+                             "end_percent": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0, "step": 0.001}),
+                                }}
+    RETURN_TYPES = ("MODEL",)
+    FUNCTION = "skip_guidance"
+    EXPERIMENTAL = True
+
+    DESCRIPTION = "Simple version of the SkipLayerGuidanceDiT node that only modifies the uncond pass."
+
+    CATEGORY = "advanced/guidance"
+
+    def skip_guidance(self, model, start_percent, end_percent, double_layers="", single_layers=""):
+        def skip(args, extra_args):
+            return args
+
+        model_sampling = model.get_model_object("model_sampling")
+        sigma_start = model_sampling.percent_to_sigma(start_percent)
+        sigma_end = model_sampling.percent_to_sigma(end_percent)
+
+        double_layers = re.findall(r'\d+', double_layers)
+        double_layers = [int(i) for i in double_layers]
+
+        single_layers = re.findall(r'\d+', single_layers)
+        single_layers = [int(i) for i in single_layers]
+
+        if len(double_layers) == 0 and len(single_layers) == 0:
+            return (model, )
+
+        def calc_cond_batch_function(args):
+            x = args["input"]
+            model = args["model"]
+            conds = args["conds"]
+            sigma = args["sigma"]
+
+            model_options = args["model_options"]
+            slg_model_options = model_options.copy()
+
+            for layer in double_layers:
+                slg_model_options = comfy.model_patcher.set_model_options_patch_replace(slg_model_options, skip, "dit", "double_block", layer)
+
+            for layer in single_layers:
+                slg_model_options = comfy.model_patcher.set_model_options_patch_replace(slg_model_options, skip, "dit", "single_block", layer)
+
+            cond, uncond = conds
+            sigma_ = sigma[0].item()
+            if sigma_ >= sigma_end and sigma_ <= sigma_start and uncond is not None:
+                cond_out, _ = comfy.samplers.calc_cond_batch(model, [cond, None], x, sigma, model_options)
+                _, uncond_out = comfy.samplers.calc_cond_batch(model, [None, uncond], x, sigma, slg_model_options)
+                out = [cond_out, uncond_out]
+            else:
+                out = comfy.samplers.calc_cond_batch(model, conds, x, sigma, model_options)
+
+            return out
+
+        m = model.clone()
+        m.set_model_sampler_calc_cond_batch_function(calc_cond_batch_function)
+
+        return (m, )
 
 NODE_CLASS_MAPPINGS = {
     "SkipLayerGuidanceDiT": SkipLayerGuidanceDiT,
+    "SkipLayerGuidanceDiTSimple": SkipLayerGuidanceDiTSimple,
 }

comfy_extras/nodes_train.py

@@ -20,41 +20,82 @@ import folder_paths
 import node_helpers
 from comfy.cli_args import args
 from comfy.comfy_types.node_typing import IO
-from comfy.weight_adapter import adapters
+from comfy.weight_adapter import adapters, adapter_maps
+
+
+def make_batch_extra_option_dict(d, indicies, full_size=None):
+    new_dict = {}
+    for k, v in d.items():
+        newv = v
+        if isinstance(v, dict):
+            newv = make_batch_extra_option_dict(v, indicies, full_size=full_size)
+        elif isinstance(v, torch.Tensor):
+            if full_size is None or v.size(0) == full_size:
+                newv = v[indicies]
+        elif isinstance(v, (list, tuple)) and len(v) == full_size:
+            newv = [v[i] for i in indicies]
+        new_dict[k] = newv
+    return new_dict
 
 
 class TrainSampler(comfy.samplers.Sampler):
-
-    def __init__(self, loss_fn, optimizer, loss_callback=None):
+    def __init__(self, loss_fn, optimizer, loss_callback=None, batch_size=1, grad_acc=1, total_steps=1, seed=0, training_dtype=torch.bfloat16):
         self.loss_fn = loss_fn
         self.optimizer = optimizer
         self.loss_callback = loss_callback
+        self.batch_size = batch_size
+        self.total_steps = total_steps
+        self.grad_acc = grad_acc
+        self.seed = seed
+        self.training_dtype = training_dtype
 
     def sample(self, model_wrap, sigmas, extra_args, callback, noise, latent_image=None, denoise_mask=None, disable_pbar=False):
-        self.optimizer.zero_grad()
-        noise = model_wrap.inner_model.model_sampling.noise_scaling(sigmas, noise, latent_image, False)
-        latent = model_wrap.inner_model.model_sampling.noise_scaling(
-            torch.zeros_like(sigmas),
-            torch.zeros_like(noise, requires_grad=True),
-            latent_image,
-            False
-        )
+        cond = model_wrap.conds["positive"]
+        dataset_size = sigmas.size(0)
+        torch.cuda.empty_cache()
+        for i in (pbar:=tqdm.trange(self.total_steps, desc="Training LoRA", smoothing=0.01, disable=not comfy.utils.PROGRESS_BAR_ENABLED)):
+            noisegen = comfy_extras.nodes_custom_sampler.Noise_RandomNoise(self.seed + i * 1000)
+            indicies = torch.randperm(dataset_size)[:self.batch_size].tolist()
 
-        # Ensure model is in training mode and computing gradients
-        # x0 pred
-        denoised = model_wrap(noise, sigmas, **extra_args)
-        try:
-            loss = self.loss_fn(denoised, latent.clone())
-        except RuntimeError as e:
-            if "does not require grad and does not have a grad_fn" in str(e):
-                logging.info("WARNING: This is likely due to the model is loaded in inference mode.")
-        loss.backward()
-        if self.loss_callback:
-            self.loss_callback(loss.item())
+            batch_latent = torch.stack([latent_image[i] for i in indicies])
+            batch_noise = noisegen.generate_noise({"samples": batch_latent}).to(batch_latent.device)
+            batch_sigmas = [
+                model_wrap.inner_model.model_sampling.percent_to_sigma(
+                    torch.rand((1,)).item()
+                ) for _ in range(min(self.batch_size, dataset_size))
+            ]
+            batch_sigmas = torch.tensor(batch_sigmas).to(batch_latent.device)
 
-        self.optimizer.step()
-        # torch.cuda.memory._dump_snapshot("trainn.pickle")
-        # torch.cuda.memory._record_memory_history(enabled=None)
+            xt = model_wrap.inner_model.model_sampling.noise_scaling(
+                batch_sigmas,
+                batch_noise,
+                batch_latent,
+                False
+            )
+            x0 = model_wrap.inner_model.model_sampling.noise_scaling(
+                torch.zeros_like(batch_sigmas),
+                torch.zeros_like(batch_noise),
+                batch_latent,
+                False
+            )
+
+            model_wrap.conds["positive"] = [
+                cond[i] for i in indicies
+            ]
+            batch_extra_args = make_batch_extra_option_dict(extra_args, indicies, full_size=dataset_size)
+
+            with torch.autocast(xt.device.type, dtype=self.training_dtype):
+                x0_pred = model_wrap(xt, batch_sigmas, **batch_extra_args)
+                loss = self.loss_fn(x0_pred, x0)
+            loss.backward()
+            if self.loss_callback:
+                self.loss_callback(loss.item())
+            pbar.set_postfix({"loss": f"{loss.item():.4f}"})
+
+            if (i+1) % self.grad_acc == 0:
+                self.optimizer.step()
+                self.optimizer.zero_grad()
+        torch.cuda.empty_cache()
         return torch.zeros_like(latent_image)
 
 
@@ -75,7 +116,7 @@ class BiasDiff(torch.nn.Module):
         return self.passive_memory_usage()
 
 
-def load_and_process_images(image_files, input_dir, resize_method="None"):
+def load_and_process_images(image_files, input_dir, resize_method="None", w=None, h=None):
     """Utility function to load and process a list of images.
 
     Args:
@@ -90,7 +131,6 @@ def load_and_process_images(image_files, input_dir, resize_method="None"):
         raise ValueError("No valid images found in input")
 
     output_images = []
-    w, h = None, None
 
     for file in image_files:
         image_path = os.path.join(input_dir, file)
@@ -206,6 +246,103 @@ class LoadImageSetFromFolderNode:
         return (output_tensor,)
 
 
+class LoadImageTextSetFromFolderNode:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "folder": (folder_paths.get_input_subfolders(), {"tooltip": "The folder to load images from."}),
+                "clip": (IO.CLIP, {"tooltip": "The CLIP model used for encoding the text."}),
+            },
+            "optional": {
+                "resize_method": (
+                    ["None", "Stretch", "Crop", "Pad"],
+                    {"default": "None"},
+                ),
+                "width": (
+                    IO.INT,
+                    {
+                        "default": -1,
+                        "min": -1,
+                        "max": 10000,
+                        "step": 1,
+                        "tooltip": "The width to resize the images to. -1 means use the original width.",
+                    },
+                ),
+                "height": (
+                    IO.INT,
+                    {
+                        "default": -1,
+                        "min": -1,
+                        "max": 10000,
+                        "step": 1,
+                        "tooltip": "The height to resize the images to. -1 means use the original height.",
+                    },
+                )
+            },
+        }
+
+    RETURN_TYPES = ("IMAGE", IO.CONDITIONING,)
+    FUNCTION = "load_images"
+    CATEGORY = "loaders"
+    EXPERIMENTAL = True
+    DESCRIPTION = "Loads a batch of images and caption from a directory for training."
+
+    def load_images(self, folder, clip, resize_method, width=None, height=None):
+        if clip is None:
+            raise RuntimeError("ERROR: clip input is invalid: None\n\nIf the clip is from a checkpoint loader node your checkpoint does not contain a valid clip or text encoder model.")
+
+        logging.info(f"Loading images from folder: {folder}")
+
+        sub_input_dir = os.path.join(folder_paths.get_input_directory(), folder)
+        valid_extensions = [".png", ".jpg", ".jpeg", ".webp"]
+
+        image_files = []
+        for item in os.listdir(sub_input_dir):
+            path = os.path.join(sub_input_dir, item)
+            if any(item.lower().endswith(ext) for ext in valid_extensions):
+                image_files.append(path)
+            elif os.path.isdir(path):
+                # Support kohya-ss/sd-scripts folder structure
+                repeat = 1
+                if item.split("_")[0].isdigit():
+                    repeat = int(item.split("_")[0])
+                image_files.extend([
+                    os.path.join(path, f) for f in os.listdir(path) if any(f.lower().endswith(ext) for ext in valid_extensions)
+                ] * repeat)
+
+        caption_file_path = [
+            f.replace(os.path.splitext(f)[1], ".txt")
+            for f in image_files
+        ]
+        captions = []
+        for caption_file in caption_file_path:
+            caption_path = os.path.join(sub_input_dir, caption_file)
+            if os.path.exists(caption_path):
+                with open(caption_path, "r", encoding="utf-8") as f:
+                    caption = f.read().strip()
+                    captions.append(caption)
+            else:
+                captions.append("")
+
+        width = width if width != -1 else None
+        height = height if height != -1 else None
+        output_tensor = load_and_process_images(image_files, sub_input_dir, resize_method, width, height)
+
+        logging.info(f"Loaded {len(output_tensor)} images from {sub_input_dir}.")
+
+        logging.info(f"Encoding captions from {sub_input_dir}.")
+        conditions = []
+        empty_cond = clip.encode_from_tokens_scheduled(clip.tokenize(""))
+        for text in captions:
+            if text == "":
+                conditions.append(empty_cond)
+            tokens = clip.tokenize(text)
+            conditions.extend(clip.encode_from_tokens_scheduled(tokens))
+        logging.info(f"Encoded {len(conditions)} captions from {sub_input_dir}.")
+        return (output_tensor, conditions)
+
+
 def draw_loss_graph(loss_map, steps):
     width, height = 500, 300
     img = Image.new("RGB", (width, height), "white")
@@ -283,6 +420,16 @@ class TrainLoraNode:
                         "tooltip": "The batch size to use for training.",
                     },
                 ),
+                "grad_accumulation_steps": (
+                    IO.INT,
+                    {
+                        "default": 1,
+                        "min": 1,
+                        "max": 1024,
+                        "step": 1,
+                        "tooltip": "The number of gradient accumulation steps to use for training.",
+                    }
+                ),
                 "steps": (
                     IO.INT,
                     {
@@ -342,6 +489,17 @@ class TrainLoraNode:
                     ["bf16", "fp32"],
                     {"default": "bf16", "tooltip": "The dtype to use for lora."},
                 ),
+                "algorithm": (
+                    list(adapter_maps.keys()),
+                    {"default": list(adapter_maps.keys())[0], "tooltip": "The algorithm to use for training."},
+                ),
+                "gradient_checkpointing": (
+                    IO.BOOLEAN,
+                    {
+                        "default": True,
+                        "tooltip": "Use gradient checkpointing for training.",
+                    }
+                ),
                 "existing_lora": (
                     folder_paths.get_filename_list("loras") + ["[None]"],
                     {
@@ -365,6 +523,7 @@ class TrainLoraNode:
         positive,
         batch_size,
         steps,
+        grad_accumulation_steps,
         learning_rate,
         rank,
         optimizer,
@@ -372,6 +531,8 @@ class TrainLoraNode:
         seed,
         training_dtype,
         lora_dtype,
+        algorithm,
+        gradient_checkpointing,
         existing_lora,
     ):
         mp = model.clone()
@@ -381,6 +542,13 @@ class TrainLoraNode:
 
         latents = latents["samples"].to(dtype)
         num_images = latents.shape[0]
+        logging.info(f"Total Images: {num_images}, Total Captions: {len(positive)}")
+        if len(positive) == 1 and num_images > 1:
+            positive = positive * num_images
+        elif len(positive) != num_images:
+            raise ValueError(
+                f"Number of positive conditions ({len(positive)}) does not match number of images ({num_images})."
+            )
 
         with torch.inference_mode(False):
             lora_sd = {}
@@ -415,10 +583,8 @@ class TrainLoraNode:
                                 if existing_adapter is not None:
                                     break
                             else:
-                                # If no existing adapter found, use LoRA
-                                # We will add algo option in the future
                                 existing_adapter = None
-                                adapter_cls = adapters[0]
+                                adapter_cls = adapter_maps[algorithm]
 
                             if existing_adapter is not None:
                                 train_adapter = existing_adapter.to_train().to(lora_dtype)
@@ -472,45 +638,45 @@ class TrainLoraNode:
                 criterion = torch.nn.SmoothL1Loss()
 
             # setup models
-            for m in find_all_highest_child_module_with_forward(mp.model.diffusion_model):
-                patch(m)
+            if gradient_checkpointing:
+                for m in find_all_highest_child_module_with_forward(mp.model.diffusion_model):
+                    patch(m)
+            mp.model.requires_grad_(False)
             comfy.model_management.load_models_gpu([mp], memory_required=1e20, force_full_load=True)
 
             # Setup sampler and guider like in test script
             loss_map = {"loss": []}
             def loss_callback(loss):
                 loss_map["loss"].append(loss)
-                pbar.set_postfix({"loss": f"{loss:.4f}"})
             train_sampler = TrainSampler(
-                criterion, optimizer, loss_callback=loss_callback
+                criterion,
+                optimizer,
+                loss_callback=loss_callback,
+                batch_size=batch_size,
+                grad_acc=grad_accumulation_steps,
+                total_steps=steps*grad_accumulation_steps,
+                seed=seed,
+                training_dtype=dtype
             )
             guider = comfy_extras.nodes_custom_sampler.Guider_Basic(mp)
             guider.set_conds(positive)  # Set conditioning from input
-            ss = comfy_extras.nodes_custom_sampler.SamplerCustomAdvanced()
-
-            # yoland: this currently resize to the first image in the dataset
 
             # Training loop
-            torch.cuda.empty_cache()
             try:
-                for step in (pbar:=tqdm.trange(steps, desc="Training LoRA", smoothing=0.01, disable=not comfy.utils.PROGRESS_BAR_ENABLED)):
-                    # Generate random sigma
-                    sigma = mp.model.model_sampling.percent_to_sigma(
-                        torch.rand((1,)).item()
-                    )
-                    sigma = torch.tensor([sigma])
-
-                    noise = comfy_extras.nodes_custom_sampler.Noise_RandomNoise(step * 1000 + seed)
-
-                    indices = torch.randperm(num_images)[:batch_size]
-                    ss.sample(
-                        noise, guider, train_sampler, sigma, {"samples": latents[indices].clone()}
-                    )
+                # Generate dummy sigmas and noise
+                sigmas = torch.tensor(range(num_images))
+                noise = comfy_extras.nodes_custom_sampler.Noise_RandomNoise(seed)
+                guider.sample(
+                    noise.generate_noise({"samples": latents}),
+                    latents,
+                    train_sampler,
+                    sigmas,
+                    seed=noise.seed
+                )
             finally:
                 for m in mp.model.modules():
                     unpatch(m)
-            del ss, train_sampler, optimizer
-            torch.cuda.empty_cache()
+            del train_sampler, optimizer
 
             for adapter in all_weight_adapters:
                 adapter.requires_grad_(False)
@@ -697,6 +863,7 @@ NODE_CLASS_MAPPINGS = {
     "SaveLoRANode": SaveLoRA,
     "LoraModelLoader": LoraModelLoader,
     "LoadImageSetFromFolderNode": LoadImageSetFromFolderNode,
+    "LoadImageTextSetFromFolderNode": LoadImageTextSetFromFolderNode,
     "LossGraphNode": LossGraphNode,
 }
 
@@ -705,5 +872,6 @@ NODE_DISPLAY_NAME_MAPPINGS = {
     "SaveLoRANode": "Save LoRA Weights",
     "LoraModelLoader": "Load LoRA Model",
     "LoadImageSetFromFolderNode": "Load Image Dataset from Folder",
+    "LoadImageTextSetFromFolderNode": "Load Image and Text Dataset from Folder",
     "LossGraphNode": "Plot Loss Graph",
 }

comfy_extras/nodes_wan.py

@@ -1,3 +1,4 @@
+import math
 import nodes
 import node_helpers
 import torch
@@ -5,7 +6,9 @@ import comfy.model_management
 import comfy.utils
 import comfy.latent_formats
 import comfy.clip_vision
-
+import json
+import numpy as np
+from typing import Tuple
 
 class WanImageToVideo:
     @classmethod
@@ -383,7 +386,350 @@ class WanPhantomSubjectToVideo:
         out_latent["samples"] = latent
         return (positive, cond2, negative, out_latent)
 
+def parse_json_tracks(tracks):
+    """Parse JSON track data into a standardized format"""
+    tracks_data = []
+    try:
+        # If tracks is a string, try to parse it as JSON
+        if isinstance(tracks, str):
+            parsed = json.loads(tracks.replace("'", '"'))
+            tracks_data.extend(parsed)
+        else:
+            # If tracks is a list of strings, parse each one
+            for track_str in tracks:
+                parsed = json.loads(track_str.replace("'", '"'))
+                tracks_data.append(parsed)
+
+        # Check if we have a single track (dict with x,y) or a list of tracks
+        if tracks_data and isinstance(tracks_data[0], dict) and 'x' in tracks_data[0]:
+            # Single track detected, wrap it in a list
+            tracks_data = [tracks_data]
+        elif tracks_data and isinstance(tracks_data[0], list) and tracks_data[0] and isinstance(tracks_data[0][0], dict) and 'x' in tracks_data[0][0]:
+            # Already a list of tracks, nothing to do
+            pass
+        else:
+            # Unexpected format
+            pass
+
+    except json.JSONDecodeError:
+        tracks_data = []
+    return tracks_data
+
+def process_tracks(tracks_np: np.ndarray, frame_size: Tuple[int, int], num_frames, quant_multi: int = 8, **kwargs):
+    # tracks: shape [t, h, w, 3] => samples align with 24 fps, model trained with 16 fps.
+    # frame_size: tuple (W, H)
+    tracks = torch.from_numpy(tracks_np).float()
+
+    if tracks.shape[1] == 121:
+        tracks = torch.permute(tracks, (1, 0, 2, 3))
+
+    tracks, visibles = tracks[..., :2], tracks[..., 2:3]
+
+    short_edge = min(*frame_size)
+
+    frame_center = torch.tensor([*frame_size]).type_as(tracks) / 2
+    tracks = tracks - frame_center
+
+    tracks = tracks / short_edge * 2
+
+    visibles = visibles * 2 - 1
+
+    trange = torch.linspace(-1, 1, tracks.shape[0]).view(-1, 1, 1, 1).expand(*visibles.shape)
+
+    out_ = torch.cat([trange, tracks, visibles], dim=-1).view(121, -1, 4)
+
+    out_0 = out_[:1]
+
+    out_l = out_[1:] # 121 => 120 | 1
+    a = 120 // math.gcd(120, num_frames)
+    b = num_frames // math.gcd(120, num_frames)
+    out_l = torch.repeat_interleave(out_l, b, dim=0)[1::a]  # 120 => 120 * b => 120 * b / a == F
+
+    final_result = torch.cat([out_0, out_l], dim=0)
+
+    return final_result
+
+FIXED_LENGTH = 121
+def pad_pts(tr):
+    """Convert list of {x,y} to (FIXED_LENGTH,1,3) array, padding/truncating."""
+    pts = np.array([[p['x'], p['y'], 1] for p in tr], dtype=np.float32)
+    n = pts.shape[0]
+    if n < FIXED_LENGTH:
+        pad = np.zeros((FIXED_LENGTH - n, 3), dtype=np.float32)
+        pts = np.vstack((pts, pad))
+    else:
+        pts = pts[:FIXED_LENGTH]
+    return pts.reshape(FIXED_LENGTH, 1, 3)
+
+def ind_sel(target: torch.Tensor, ind: torch.Tensor, dim: int = 1):
+    """Index selection utility function"""
+    assert (
+        len(ind.shape) > dim
+    ), "Index must have the target dim, but get dim: %d, ind shape: %s" % (dim, str(ind.shape))
+
+    target = target.expand(
+        *tuple(
+            [ind.shape[k] if target.shape[k] == 1 else -1 for k in range(dim)]
+            + [
+                -1,
+            ]
+            * (len(target.shape) - dim)
+        )
+    )
+
+    ind_pad = ind
+
+    if len(target.shape) > dim + 1:
+        for _ in range(len(target.shape) - (dim + 1)):
+            ind_pad = ind_pad.unsqueeze(-1)
+        ind_pad = ind_pad.expand(*(-1,) * (dim + 1), *target.shape[(dim + 1) : :])
+
+    return torch.gather(target, dim=dim, index=ind_pad)
+
+def merge_final(vert_attr: torch.Tensor, weight: torch.Tensor, vert_assign: torch.Tensor):
+    """Merge vertex attributes with weights"""
+    target_dim = len(vert_assign.shape) - 1
+    if len(vert_attr.shape) == 2:
+        assert vert_attr.shape[0] > vert_assign.max()
+        new_shape = [1] * target_dim + list(vert_attr.shape)
+        tensor = vert_attr.reshape(new_shape)
+        sel_attr = ind_sel(tensor, vert_assign.type(torch.long), dim=target_dim)
+    else:
+        assert vert_attr.shape[1] > vert_assign.max()
+        new_shape = [vert_attr.shape[0]] + [1] * (target_dim - 1) + list(vert_attr.shape[1:])
+        tensor = vert_attr.reshape(new_shape)
+        sel_attr = ind_sel(tensor, vert_assign.type(torch.long), dim=target_dim)
+
+    final_attr = torch.sum(sel_attr * weight.unsqueeze(-1), dim=-2)
+    return final_attr
+
+
+def _patch_motion_single(
+    tracks: torch.FloatTensor,  # (B, T, N, 4)
+    vid: torch.FloatTensor,     # (C, T, H, W)
+    temperature: float,
+    vae_divide: tuple,
+    topk: int,
+):
+    """Apply motion patching based on tracks"""
+    _, T, H, W = vid.shape
+    N = tracks.shape[2]
+    _, tracks_xy, visible = torch.split(
+        tracks, [1, 2, 1], dim=-1
+    )  # (B, T, N, 2) | (B, T, N, 1)
+    tracks_n = tracks_xy / torch.tensor([W / min(H, W), H / min(H, W)], device=tracks_xy.device)
+    tracks_n = tracks_n.clamp(-1, 1)
+    visible = visible.clamp(0, 1)
+
+    xx = torch.linspace(-W / min(H, W), W / min(H, W), W)
+    yy = torch.linspace(-H / min(H, W), H / min(H, W), H)
+
+    grid = torch.stack(torch.meshgrid(yy, xx, indexing="ij")[::-1], dim=-1).to(
+        tracks_xy.device
+    )
+
+    tracks_pad = tracks_xy[:, 1:]
+    visible_pad = visible[:, 1:]
+
+    visible_align = visible_pad.view(T - 1, 4, *visible_pad.shape[2:]).sum(1)
+    tracks_align = (tracks_pad * visible_pad).view(T - 1, 4, *tracks_pad.shape[2:]).sum(
+        1
+    ) / (visible_align + 1e-5)
+    dist_ = (
+        (tracks_align[:, None, None] - grid[None, :, :, None]).pow(2).sum(-1)
+    )  # T, H, W, N
+    weight = torch.exp(-dist_ * temperature) * visible_align.clamp(0, 1).view(
+        T - 1, 1, 1, N
+    )
+    vert_weight, vert_index = torch.topk(
+        weight, k=min(topk, weight.shape[-1]), dim=-1
+    )
+
+    grid_mode = "bilinear"
+    point_feature = torch.nn.functional.grid_sample(
+        vid.permute(1, 0, 2, 3)[:1],
+        tracks_n[:, :1].type(vid.dtype),
+        mode=grid_mode,
+        padding_mode="zeros",
+        align_corners=False,
+    )
+    point_feature = point_feature.squeeze(0).squeeze(1).permute(1, 0) # N, C=16
+
+    out_feature = merge_final(point_feature, vert_weight, vert_index).permute(3, 0, 1, 2) # T - 1, H, W, C => C, T - 1, H, W
+    out_weight = vert_weight.sum(-1) # T - 1, H, W
+
+    # out feature -> already soft weighted
+    mix_feature = out_feature + vid[:, 1:] * (1 - out_weight.clamp(0, 1))
+
+    out_feature_full = torch.cat([vid[:, :1], mix_feature], dim=1) # C, T, H, W
+    out_mask_full = torch.cat([torch.ones_like(out_weight[:1]), out_weight], dim=0)  # T, H, W
+
+    return out_mask_full[None].expand(vae_divide[0], -1, -1, -1), out_feature_full
+
+
+def patch_motion(
+    tracks: torch.FloatTensor,  # (B, TB, T, N, 4)
+    vid: torch.FloatTensor,     # (C, T, H, W)
+    temperature: float = 220.0,
+    vae_divide: tuple = (4, 16),
+    topk: int = 2,
+):
+    B = len(tracks)
+
+    # Process each batch separately
+    out_masks = []
+    out_features = []
+
+    for b in range(B):
+        mask, feature = _patch_motion_single(
+            tracks[b],  # (T, N, 4)
+            vid[b],        # (C, T, H, W)
+            temperature,
+            vae_divide,
+            topk
+        )
+        out_masks.append(mask)
+        out_features.append(feature)
+
+    # Stack results: (B, C, T, H, W)
+    out_mask_full = torch.stack(out_masks, dim=0)
+    out_feature_full = torch.stack(out_features, dim=0)
+
+    return out_mask_full, out_feature_full
+
+class WanTrackToVideo:
+    @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", ),
+                }}
+
+    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):
+
+        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)
+
+        latent = torch.zeros([batch_size, 16, ((length - 1) // 4) + 1, height // 8, width // 8],
+                           device=comfy.model_management.intermediate_device())
+
+        if isinstance(tracks_data[0][0], dict):
+            tracks_data = [tracks_data]
+
+        processed_tracks = []
+        for batch in tracks_data:
+            arrs = []
+            for track in batch:
+                pts = pad_pts(track)
+                arrs.append(pts)
+
+            tracks_np = np.stack(arrs, axis=0)
+            processed_tracks.append(process_tracks(tracks_np, (width, height), length - 1).unsqueeze(0))
+
+        if start_image is not None:
+            start_image = comfy.utils.common_upscale(start_image[:batch_size].movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1)
+            videos = torch.ones((start_image.shape[0], length, height, width, start_image.shape[-1]), device=start_image.device, dtype=start_image.dtype) * 0.5
+            for i in range(start_image.shape[0]):
+                videos[i, 0] = start_image[i]
+
+            latent_videos = []
+            videos = comfy.utils.resize_to_batch_size(videos, batch_size)
+            for i in range(batch_size):
+                latent_videos += [vae.encode(videos[i, :, :, :, :3])]
+            y = torch.cat(latent_videos, dim=0)
+
+            # Scale latent since patch_motion is non-linear
+            y = comfy.latent_formats.Wan21().process_in(y)
+
+            processed_tracks = comfy.utils.resize_list_to_batch_size(processed_tracks, batch_size)
+            res = patch_motion(
+                processed_tracks, y, temperature=temperature, topk=topk, vae_divide=(4, 16)
+            )
+
+            mask, concat_latent_image = res
+            concat_latent_image = comfy.latent_formats.Wan21().process_out(concat_latent_image)
+            mask = -mask + 1.0  # Invert mask to match expected format
+            positive = node_helpers.conditioning_set_values(positive,
+                                                            {"concat_mask": mask,
+                                                            "concat_latent_image": concat_latent_image})
+            negative = node_helpers.conditioning_set_values(negative,
+                                                            {"concat_mask": mask,
+                                                            "concat_latent_image": concat_latent_image})
+
+        if clip_vision_output is not None:
+            positive = node_helpers.conditioning_set_values(positive, {"clip_vision_output": clip_vision_output})
+            negative = node_helpers.conditioning_set_values(negative, {"clip_vision_output": clip_vision_output})
+
+        out_latent = {}
+        out_latent["samples"] = latent
+        return (positive, negative, out_latent)
+
+
+class Wan22ImageToVideoLatent:
+    @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", ),
+                }}
+
+
+    RETURN_TYPES = ("LATENT",)
+    FUNCTION = "encode"
+
+    CATEGORY = "conditioning/inpaint"
+
+    def encode(self, vae, width, height, length, batch_size, start_image=None):
+        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,)
+
+        mask = torch.ones([latent.shape[0], 1, ((length - 1) // 4) + 1, latent.shape[-2], latent.shape[-1]], 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)
+            latent_temp = vae.encode(start_image)
+            latent[:, :, :latent_temp.shape[-3]] = latent_temp
+            mask[:, :, :latent_temp.shape[-3]] *= 0.0
+
+        out_latent = {}
+        latent_format = comfy.latent_formats.Wan22()
+        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,)
+
+
 NODE_CLASS_MAPPINGS = {
+    "WanTrackToVideo": WanTrackToVideo,
     "WanImageToVideo": WanImageToVideo,
     "WanFunControlToVideo": WanFunControlToVideo,
     "WanFunInpaintToVideo": WanFunInpaintToVideo,
@@ -392,4 +738,5 @@ NODE_CLASS_MAPPINGS = {
     "TrimVideoLatent": TrimVideoLatent,
     "WanCameraImageToVideo": WanCameraImageToVideo,
     "WanPhantomSubjectToVideo": WanPhantomSubjectToVideo,
+    "Wan22ImageToVideoLatent": Wan22ImageToVideoLatent,
 }

comfyui_version.py

@@ -1,3 +1,3 @@
 # This file is automatically generated by the build process when version is
 # updated in pyproject.toml.
-__version__ = "0.3.43"
+__version__ = "0.3.47"

cuda_malloc.py

@@ -74,7 +74,8 @@ if not args.cuda_malloc:
                 module = importlib.util.module_from_spec(spec)
                 spec.loader.exec_module(module)
                 version = module.__version__
-        if int(version[0]) >= 2: #enable by default for torch version 2.0 and up
+
+        if int(version[0]) >= 2 and "+cu" in version: #enable by default for torch version 2.0 and up only on cuda torch
             args.cuda_malloc = cuda_malloc_supported()
     except:
         pass

execution.py

@@ -8,12 +8,14 @@ import time
 import traceback
 from enum import Enum
 from typing import List, Literal, NamedTuple, Optional
+import asyncio
 
 import torch
 
 import comfy.model_management
 import nodes
 from comfy_execution.caching import (
+    BasicCache,
     CacheKeySetID,
     CacheKeySetInputSignature,
     DependencyAwareCache,
@@ -28,6 +30,8 @@ from comfy_execution.graph import (
 )
 from comfy_execution.graph_utils import GraphBuilder, is_link
 from comfy_execution.validation import validate_node_input
+from comfy_execution.progress import get_progress_state, reset_progress_state, add_progress_handler, WebUIProgressHandler
+from comfy_execution.utils import CurrentNodeContext
 
 
 class ExecutionResult(Enum):
@@ -39,12 +43,13 @@ class DuplicateNodeError(Exception):
     pass
 
 class IsChangedCache:
-    def __init__(self, dynprompt, outputs_cache):
+    def __init__(self, prompt_id: str, dynprompt: DynamicPrompt, outputs_cache: BasicCache):
+        self.prompt_id = prompt_id
         self.dynprompt = dynprompt
         self.outputs_cache = outputs_cache
         self.is_changed = {}
 
-    def get(self, node_id):
+    async def get(self, node_id):
         if node_id in self.is_changed:
             return self.is_changed[node_id]
 
@@ -62,7 +67,8 @@ class IsChangedCache:
         # Intentionally do not use cached outputs here. We only want constants in IS_CHANGED
         input_data_all, _ = get_input_data(node["inputs"], class_def, node_id, None)
         try:
-            is_changed = _map_node_over_list(class_def, input_data_all, "IS_CHANGED")
+            is_changed = await _async_map_node_over_list(self.prompt_id, node_id, class_def, input_data_all, "IS_CHANGED")
+            is_changed = await resolve_map_node_over_list_results(is_changed)
             node["is_changed"] = [None if isinstance(x, ExecutionBlocker) else x for x in is_changed]
         except Exception as e:
             logging.warning("WARNING: {}".format(e))
@@ -117,6 +123,8 @@ class CacheSet:
         }
         return result
 
+SENSITIVE_EXTRA_DATA_KEYS = ("auth_token_comfy_org", "api_key_comfy_org")
+
 def get_input_data(inputs, class_def, unique_id, outputs=None, dynprompt=None, extra_data={}):
     valid_inputs = class_def.INPUT_TYPES()
     input_data_all = {}
@@ -164,7 +172,19 @@ def get_input_data(inputs, class_def, unique_id, outputs=None, dynprompt=None, e
 
 map_node_over_list = None #Don't hook this please
 
-def _map_node_over_list(obj, input_data_all, func, allow_interrupt=False, execution_block_cb=None, pre_execute_cb=None):
+async def resolve_map_node_over_list_results(results):
+    remaining = [x for x in results if isinstance(x, asyncio.Task) and not x.done()]
+    if len(remaining) == 0:
+        return [x.result() if isinstance(x, asyncio.Task) else x for x in results]
+    else:
+        done, pending = await asyncio.wait(remaining)
+        for task in done:
+            exc = task.exception()
+            if exc is not None:
+                raise exc
+        return [x.result() if isinstance(x, asyncio.Task) else x for x in results]
+
+async def _async_map_node_over_list(prompt_id, unique_id, obj, input_data_all, func, allow_interrupt=False, execution_block_cb=None, pre_execute_cb=None):
     # check if node wants the lists
     input_is_list = getattr(obj, "INPUT_IS_LIST", False)
 
@@ -178,7 +198,7 @@ def _map_node_over_list(obj, input_data_all, func, allow_interrupt=False, execut
         return {k: v[i if len(v) > i else -1] for k, v in d.items()}
 
     results = []
-    def process_inputs(inputs, index=None, input_is_list=False):
+    async def process_inputs(inputs, index=None, input_is_list=False):
         if allow_interrupt:
             nodes.before_node_execution()
         execution_block = None
@@ -194,20 +214,37 @@ def _map_node_over_list(obj, input_data_all, func, allow_interrupt=False, execut
         if execution_block is None:
             if pre_execute_cb is not None and index is not None:
                 pre_execute_cb(index)
-            results.append(getattr(obj, func)(**inputs))
+            f = getattr(obj, func)
+            if inspect.iscoroutinefunction(f):
+                async def async_wrapper(f, prompt_id, unique_id, list_index, args):
+                    with CurrentNodeContext(prompt_id, unique_id, list_index):
+                        return await f(**args)
+                task = asyncio.create_task(async_wrapper(f, prompt_id, unique_id, index, args=inputs))
+                # Give the task a chance to execute without yielding
+                await asyncio.sleep(0)
+                if task.done():
+                    result = task.result()
+                    results.append(result)
+                else:
+                    results.append(task)
+            else:
+                with CurrentNodeContext(prompt_id, unique_id, index):
+                    result = f(**inputs)
+                results.append(result)
         else:
             results.append(execution_block)
 
     if input_is_list:
-        process_inputs(input_data_all, 0, input_is_list=input_is_list)
+        await process_inputs(input_data_all, 0, input_is_list=input_is_list)
     elif max_len_input == 0:
-        process_inputs({})
+        await process_inputs({})
     else:
         for i in range(max_len_input):
             input_dict = slice_dict(input_data_all, i)
-            process_inputs(input_dict, i)
+            await process_inputs(input_dict, i)
     return results
 
+
 def merge_result_data(results, obj):
     # check which outputs need concatenating
     output = []
@@ -229,11 +266,18 @@ def merge_result_data(results, obj):
             output.append([o[i] for o in results])
     return output
 
-def get_output_data(obj, input_data_all, execution_block_cb=None, pre_execute_cb=None):
+async def get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=None, pre_execute_cb=None):
+    return_values = await _async_map_node_over_list(prompt_id, unique_id, obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb)
+    has_pending_task = any(isinstance(r, asyncio.Task) and not r.done() for r in return_values)
+    if has_pending_task:
+        return return_values, {}, False, has_pending_task
+    output, ui, has_subgraph = get_output_from_returns(return_values, obj)
+    return output, ui, has_subgraph, False
+
+def get_output_from_returns(return_values, obj):
     results = []
     uis = []
     subgraph_results = []
-    return_values = _map_node_over_list(obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb)
     has_subgraph = False
     for i in range(len(return_values)):
         r = return_values[i]
@@ -267,6 +311,10 @@ def get_output_data(obj, input_data_all, execution_block_cb=None, pre_execute_cb
     else:
         output = []
     ui = dict()
+    # TODO: Think there's an existing bug here
+    # If we're performing a subgraph expansion, we probably shouldn't be returning UI values yet.
+    # They'll get cached without the completed subgraphs. It's an edge case and I'm not aware of
+    # any nodes that use both subgraph expansion and custom UI outputs, but might be a problem in the future.
     if len(uis) > 0:
         ui = {k: [y for x in uis for y in x[k]] for k in uis[0].keys()}
     return output, ui, has_subgraph
@@ -279,7 +327,7 @@ def format_value(x):
     else:
         return str(x)
 
-def execute(server, dynprompt, caches, current_item, extra_data, executed, prompt_id, execution_list, pending_subgraph_results):
+async def execute(server, dynprompt, caches, current_item, extra_data, executed, prompt_id, execution_list, pending_subgraph_results, pending_async_nodes):
     unique_id = current_item
     real_node_id = dynprompt.get_real_node_id(unique_id)
     display_node_id = dynprompt.get_display_node_id(unique_id)
@@ -291,11 +339,26 @@ def execute(server, dynprompt, caches, current_item, extra_data, executed, promp
         if server.client_id is not None:
             cached_output = caches.ui.get(unique_id) or {}
             server.send_sync("executed", { "node": unique_id, "display_node": display_node_id, "output": cached_output.get("output",None), "prompt_id": prompt_id }, server.client_id)
+        get_progress_state().finish_progress(unique_id)
         return (ExecutionResult.SUCCESS, None, None)
 
     input_data_all = None
     try:
-        if unique_id in pending_subgraph_results:
+        if unique_id in pending_async_nodes:
+            results = []
+            for r in pending_async_nodes[unique_id]:
+                if isinstance(r, asyncio.Task):
+                    try:
+                        results.append(r.result())
+                    except Exception as ex:
+                        # An async task failed - propagate the exception up
+                        del pending_async_nodes[unique_id]
+                        raise ex
+                else:
+                    results.append(r)
+            del pending_async_nodes[unique_id]
+            output_data, output_ui, has_subgraph = get_output_from_returns(results, class_def)
+        elif unique_id in pending_subgraph_results:
             cached_results = pending_subgraph_results[unique_id]
             resolved_outputs = []
             for is_subgraph, result in cached_results:
@@ -317,6 +380,7 @@ def execute(server, dynprompt, caches, current_item, extra_data, executed, promp
             output_ui = []
             has_subgraph = False
         else:
+            get_progress_state().start_progress(unique_id)
             input_data_all, missing_keys = get_input_data(inputs, class_def, unique_id, caches.outputs, dynprompt, extra_data)
             if server.client_id is not None:
                 server.last_node_id = display_node_id
@@ -328,7 +392,8 @@ def execute(server, dynprompt, caches, current_item, extra_data, executed, promp
                 caches.objects.set(unique_id, obj)
 
             if hasattr(obj, "check_lazy_status"):
-                required_inputs = _map_node_over_list(obj, input_data_all, "check_lazy_status", allow_interrupt=True)
+                required_inputs = await _async_map_node_over_list(prompt_id, unique_id, obj, input_data_all, "check_lazy_status", allow_interrupt=True)
+                required_inputs = await resolve_map_node_over_list_results(required_inputs)
                 required_inputs = set(sum([r for r in required_inputs if isinstance(r,list)], []))
                 required_inputs = [x for x in required_inputs if isinstance(x,str) and (
                     x not in input_data_all or x in missing_keys
@@ -357,8 +422,18 @@ def execute(server, dynprompt, caches, current_item, extra_data, executed, promp
                 else:
                     return block
             def pre_execute_cb(call_index):
+                # TODO - How to handle this with async functions without contextvars (which requires Python 3.12)?
                 GraphBuilder.set_default_prefix(unique_id, call_index, 0)
-            output_data, output_ui, has_subgraph = get_output_data(obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb)
+            output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb)
+            if has_pending_tasks:
+                pending_async_nodes[unique_id] = output_data
+                unblock = execution_list.add_external_block(unique_id)
+                async def await_completion():
+                    tasks = [x for x in output_data if isinstance(x, asyncio.Task)]
+                    await asyncio.gather(*tasks, return_exceptions=True)
+                    unblock()
+                asyncio.create_task(await_completion())
+                return (ExecutionResult.PENDING, None, None)
         if len(output_ui) > 0:
             caches.ui.set(unique_id, {
                 "meta": {
@@ -401,7 +476,8 @@ def execute(server, dynprompt, caches, current_item, extra_data, executed, promp
                     cached_outputs.append((True, node_outputs))
             new_node_ids = set(new_node_ids)
             for cache in caches.all:
-                cache.ensure_subcache_for(unique_id, new_node_ids).clean_unused()
+                subcache = await cache.ensure_subcache_for(unique_id, new_node_ids)
+                subcache.clean_unused()
             for node_id in new_output_ids:
                 execution_list.add_node(node_id)
             for link in new_output_links:
@@ -446,6 +522,7 @@ def execute(server, dynprompt, caches, current_item, extra_data, executed, promp
 
         return (ExecutionResult.FAILURE, error_details, ex)
 
+    get_progress_state().finish_progress(unique_id)
     executed.add(unique_id)
 
     return (ExecutionResult.SUCCESS, None, None)
@@ -500,6 +577,11 @@ class PromptExecutor:
             self.add_message("execution_error", mes, broadcast=False)
 
     def execute(self, prompt, prompt_id, extra_data={}, execute_outputs=[]):
+        asyncio_loop = asyncio.new_event_loop()
+        asyncio.set_event_loop(asyncio_loop)
+        asyncio.run(self.execute_async(prompt, prompt_id, extra_data, execute_outputs))
+
+    async def execute_async(self, prompt, prompt_id, extra_data={}, execute_outputs=[]):
         nodes.interrupt_processing(False)
 
         if "client_id" in extra_data:
@@ -512,9 +594,11 @@ class PromptExecutor:
 
         with torch.inference_mode():
             dynamic_prompt = DynamicPrompt(prompt)
-            is_changed_cache = IsChangedCache(dynamic_prompt, self.caches.outputs)
+            reset_progress_state(prompt_id, dynamic_prompt)
+            add_progress_handler(WebUIProgressHandler(self.server))
+            is_changed_cache = IsChangedCache(prompt_id, dynamic_prompt, self.caches.outputs)
             for cache in self.caches.all:
-                cache.set_prompt(dynamic_prompt, prompt.keys(), is_changed_cache)
+                await cache.set_prompt(dynamic_prompt, prompt.keys(), is_changed_cache)
                 cache.clean_unused()
 
             cached_nodes = []
@@ -527,6 +611,7 @@ class PromptExecutor:
                           { "nodes": cached_nodes, "prompt_id": prompt_id},
                           broadcast=False)
             pending_subgraph_results = {}
+            pending_async_nodes = {} # TODO - Unify this with pending_subgraph_results
             executed = set()
             execution_list = ExecutionList(dynamic_prompt, self.caches.outputs)
             current_outputs = self.caches.outputs.all_node_ids()
@@ -534,12 +619,13 @@ class PromptExecutor:
                 execution_list.add_node(node_id)
 
             while not execution_list.is_empty():
-                node_id, error, ex = execution_list.stage_node_execution()
+                node_id, error, ex = await execution_list.stage_node_execution()
                 if error is not None:
                     self.handle_execution_error(prompt_id, dynamic_prompt.original_prompt, current_outputs, executed, error, ex)
                     break
 
-                result, error, ex = execute(self.server, dynamic_prompt, self.caches, node_id, extra_data, executed, prompt_id, execution_list, pending_subgraph_results)
+                assert node_id is not None, "Node ID should not be None at this point"
+                result, error, ex = await execute(self.server, dynamic_prompt, self.caches, node_id, extra_data, executed, prompt_id, execution_list, pending_subgraph_results, pending_async_nodes)
                 self.success = result != ExecutionResult.FAILURE
                 if result == ExecutionResult.FAILURE:
                     self.handle_execution_error(prompt_id, dynamic_prompt.original_prompt, current_outputs, executed, error, ex)
@@ -569,7 +655,7 @@ class PromptExecutor:
                 comfy.model_management.unload_all_models()
 
 
-def validate_inputs(prompt, item, validated):
+async def validate_inputs(prompt_id, prompt, item, validated):
     unique_id = item
     if unique_id in validated:
         return validated[unique_id]
@@ -646,7 +732,7 @@ def validate_inputs(prompt, item, validated):
                 errors.append(error)
                 continue
             try:
-                r = validate_inputs(prompt, o_id, validated)
+                r = await validate_inputs(prompt_id, prompt, o_id, validated)
                 if r[0] is False:
                     # `r` will be set in `validated[o_id]` already
                     valid = False
@@ -771,7 +857,8 @@ def validate_inputs(prompt, item, validated):
             input_filtered['input_types'] = [received_types]
 
         #ret = obj_class.VALIDATE_INPUTS(**input_filtered)
-        ret = _map_node_over_list(obj_class, input_filtered, "VALIDATE_INPUTS")
+        ret = await _async_map_node_over_list(prompt_id, unique_id, obj_class, input_filtered, "VALIDATE_INPUTS")
+        ret = await resolve_map_node_over_list_results(ret)
         for x in input_filtered:
             for i, r in enumerate(ret):
                 if r is not True and not isinstance(r, ExecutionBlocker):
@@ -804,7 +891,7 @@ def full_type_name(klass):
         return klass.__qualname__
     return module + '.' + klass.__qualname__
 
-def validate_prompt(prompt):
+async def validate_prompt(prompt_id, prompt):
     outputs = set()
     for x in prompt:
         if 'class_type' not in prompt[x]:
@@ -847,7 +934,7 @@ def validate_prompt(prompt):
         valid = False
         reasons = []
         try:
-            m = validate_inputs(prompt, o, validated)
+            m = await validate_inputs(prompt_id, prompt, o, validated)
             valid = m[0]
             reasons = m[1]
         except Exception as ex:
@@ -960,6 +1047,11 @@ class PromptQueue:
             if status is not None:
                 status_dict = copy.deepcopy(status._asdict())
 
+            # Remove sensitive data from extra_data before storing in history
+            for sensitive_val in SENSITIVE_EXTRA_DATA_KEYS:
+                if sensitive_val in prompt[3]:
+                    prompt[3].pop(sensitive_val)
+
             self.history[prompt[1]] = {
                 "prompt": prompt,
                 "outputs": {},
@@ -1005,7 +1097,7 @@ class PromptQueue:
                     return True
         return False
 
-    def get_history(self, prompt_id=None, max_items=None, offset=-1):
+    def get_history(self, prompt_id=None, max_items=None, offset=-1, map_function=None):
         with self.mutex:
             if prompt_id is None:
                 out = {}
@@ -1014,13 +1106,21 @@ class PromptQueue:
                     offset = len(self.history) - max_items
                 for k in self.history:
                     if i >= offset:
-                        out[k] = self.history[k]
+                        p = self.history[k]
+                        if map_function is not None:
+                            p = map_function(p)
+                        out[k] = p
                         if max_items is not None and len(out) >= max_items:
                             break
                     i += 1
                 return out
             elif prompt_id in self.history:
-                return {prompt_id: copy.deepcopy(self.history[prompt_id])}
+                p = self.history[prompt_id]
+                if map_function is None:
+                    p = copy.deepcopy(p)
+                else:
+                    p = map_function(p)
+                return {prompt_id: p}
             else:
                 return {}
 

main.py

@@ -11,6 +11,9 @@ import itertools
 import utils.extra_config
 import logging
 import sys
+from comfy_execution.progress import get_progress_state
+from comfy_execution.utils import get_executing_context
+from comfy_api import feature_flags
 
 if __name__ == "__main__":
     #NOTE: These do not do anything on core ComfyUI, they are for custom nodes.
@@ -112,6 +115,15 @@ if os.name == "nt":
     logging.getLogger("xformers").addFilter(lambda record: 'A matching Triton is not available' not in record.getMessage())
 
 if __name__ == "__main__":
+    if args.default_device is not None:
+        default_dev = args.default_device
+        devices = list(range(32))
+        devices.remove(default_dev)
+        devices.insert(0, default_dev)
+        devices = ','.join(map(str, devices))
+        os.environ['CUDA_VISIBLE_DEVICES'] = str(devices)
+        os.environ['HIP_VISIBLE_DEVICES'] = str(devices)
+
     if args.cuda_device is not None:
         os.environ['CUDA_VISIBLE_DEVICES'] = str(args.cuda_device)
         os.environ['HIP_VISIBLE_DEVICES'] = str(args.cuda_device)
@@ -127,11 +139,14 @@ if __name__ == "__main__":
 
     import cuda_malloc
 
+if 'torch' in sys.modules:
+    logging.warning("WARNING: Potential Error in code: Torch already imported, torch should never be imported before this point.")
+
 import comfy.utils
 
 import execution
 import server
-from server import BinaryEventTypes
+from protocol import BinaryEventTypes
 import nodes
 import comfy.model_management
 import comfyui_version
@@ -227,15 +242,34 @@ async def run(server_instance, address='', port=8188, verbose=True, call_on_star
         server_instance.start_multi_address(addresses, call_on_start, verbose), server_instance.publish_loop()
     )
 
-
 def hijack_progress(server_instance):
-    def hook(value, total, preview_image):
+    def hook(value, total, preview_image, prompt_id=None, node_id=None):
+        executing_context = get_executing_context()
+        if prompt_id is None and executing_context is not None:
+            prompt_id = executing_context.prompt_id
+        if node_id is None and executing_context is not None:
+            node_id = executing_context.node_id
         comfy.model_management.throw_exception_if_processing_interrupted()
-        progress = {"value": value, "max": total, "prompt_id": server_instance.last_prompt_id, "node": server_instance.last_node_id}
+        if prompt_id is None:
+            prompt_id = server_instance.last_prompt_id
+        if node_id is None:
+            node_id = server_instance.last_node_id
+        progress = {"value": value, "max": total, "prompt_id": prompt_id, "node": node_id}
+        get_progress_state().update_progress(node_id, value, total, preview_image)
 
         server_instance.send_sync("progress", progress, server_instance.client_id)
         if preview_image is not None:
-            server_instance.send_sync(BinaryEventTypes.UNENCODED_PREVIEW_IMAGE, preview_image, server_instance.client_id)
+            # Only send old method if client doesn't support preview metadata
+            if not feature_flags.supports_feature(
+                server_instance.sockets_metadata,
+                server_instance.client_id,
+                "supports_preview_metadata",
+            ):
+                server_instance.send_sync(
+                    BinaryEventTypes.UNENCODED_PREVIEW_IMAGE,
+                    preview_image,
+                    server_instance.client_id,
+                )
 
     comfy.utils.set_progress_bar_global_hook(hook)
 

nodes.py

@@ -2310,6 +2310,7 @@ def init_builtin_api_nodes():
         "nodes_pika.py",
         "nodes_runway.py",
         "nodes_tripo.py",
+        "nodes_moonvalley.py",
         "nodes_rodin.py",
         "nodes_gemini.py",
     ]

notebooks/comfyui_colab.ipynb

@@ -1,322 +0,0 @@
-{
- "cells": [
-  {
-   "cell_type": "markdown",
-   "metadata": {
-    "id": "aaaaaaaaaa"
-   },
-   "source": [
-    "Git clone the repo and install the requirements. (ignore the pip errors about protobuf)"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {
-    "id": "bbbbbbbbbb"
-   },
-   "outputs": [],
-   "source": [
-    "#@title Environment Setup\n",
-    "\n",
-    "\n",
-    "OPTIONS = {}\n",
-    "\n",
-    "USE_GOOGLE_DRIVE = False  #@param {type:\"boolean\"}\n",
-    "UPDATE_COMFY_UI = True  #@param {type:\"boolean\"}\n",
-    "WORKSPACE = 'ComfyUI'\n",
-    "OPTIONS['USE_GOOGLE_DRIVE'] = USE_GOOGLE_DRIVE\n",
-    "OPTIONS['UPDATE_COMFY_UI'] = UPDATE_COMFY_UI\n",
-    "\n",
-    "if OPTIONS['USE_GOOGLE_DRIVE']:\n",
-    "    !echo \"Mounting Google Drive...\"\n",
-    "    %cd /\n",
-    "    \n",
-    "    from google.colab import drive\n",
-    "    drive.mount('/content/drive')\n",
-    "\n",
-    "    WORKSPACE = \"/content/drive/MyDrive/ComfyUI\"\n",
-    "    %cd /content/drive/MyDrive\n",
-    "\n",
-    "![ ! -d $WORKSPACE ] && echo -= Initial setup ComfyUI =- && git clone https://github.com/comfyanonymous/ComfyUI\n",
-    "%cd $WORKSPACE\n",
-    "\n",
-    "if OPTIONS['UPDATE_COMFY_UI']:\n",
-    "  !echo -= Updating ComfyUI =-\n",
-    "  !git pull\n",
-    "\n",
-    "!echo -= Install dependencies =-\n",
-    "!pip install xformers!=0.0.18 -r requirements.txt --extra-index-url https://download.pytorch.org/whl/cu121 --extra-index-url https://download.pytorch.org/whl/cu118 --extra-index-url https://download.pytorch.org/whl/cu117"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {
-    "id": "cccccccccc"
-   },
-   "source": [
-    "Download some models/checkpoints/vae or custom comfyui nodes (uncomment the commands for the ones you want)"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {
-    "id": "dddddddddd"
-   },
-   "outputs": [],
-   "source": [
-    "# Checkpoints\n",
-    "\n",
-    "### SDXL\n",
-    "### I recommend these workflow examples: https://comfyanonymous.github.io/ComfyUI_examples/sdxl/\n",
-    "\n",
-    "#!wget -c https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/resolve/main/sd_xl_base_1.0.safetensors -P ./models/checkpoints/\n",
-    "#!wget -c https://huggingface.co/stabilityai/stable-diffusion-xl-refiner-1.0/resolve/main/sd_xl_refiner_1.0.safetensors -P ./models/checkpoints/\n",
-    "\n",
-    "# SDXL ReVision\n",
-    "#!wget -c https://huggingface.co/comfyanonymous/clip_vision_g/resolve/main/clip_vision_g.safetensors -P ./models/clip_vision/\n",
-    "\n",
-    "# SD1.5\n",
-    "!wget -c https://huggingface.co/Comfy-Org/stable-diffusion-v1-5-archive/resolve/main/v1-5-pruned-emaonly-fp16.safetensors -P ./models/checkpoints/\n",
-    "\n",
-    "# SD2\n",
-    "#!wget -c https://huggingface.co/stabilityai/stable-diffusion-2-1-base/resolve/main/v2-1_512-ema-pruned.safetensors -P ./models/checkpoints/\n",
-    "#!wget -c https://huggingface.co/stabilityai/stable-diffusion-2-1/resolve/main/v2-1_768-ema-pruned.safetensors -P ./models/checkpoints/\n",
-    "\n",
-    "# Some SD1.5 anime style\n",
-    "#!wget -c https://huggingface.co/WarriorMama777/OrangeMixs/resolve/main/Models/AbyssOrangeMix2/AbyssOrangeMix2_hard.safetensors -P ./models/checkpoints/\n",
-    "#!wget -c https://huggingface.co/WarriorMama777/OrangeMixs/resolve/main/Models/AbyssOrangeMix3/AOM3A1_orangemixs.safetensors -P ./models/checkpoints/\n",
-    "#!wget -c https://huggingface.co/WarriorMama777/OrangeMixs/resolve/main/Models/AbyssOrangeMix3/AOM3A3_orangemixs.safetensors -P ./models/checkpoints/\n",
-    "#!wget -c https://huggingface.co/Linaqruf/anything-v3.0/resolve/main/anything-v3-fp16-pruned.safetensors -P ./models/checkpoints/\n",
-    "\n",
-    "# Waifu Diffusion 1.5 (anime style SD2.x 768-v)\n",
-    "#!wget -c https://huggingface.co/waifu-diffusion/wd-1-5-beta3/resolve/main/wd-illusion-fp16.safetensors -P ./models/checkpoints/\n",
-    "\n",
-    "\n",
-    "# unCLIP models\n",
-    "#!wget -c https://huggingface.co/comfyanonymous/illuminatiDiffusionV1_v11_unCLIP/resolve/main/illuminatiDiffusionV1_v11-unclip-h-fp16.safetensors -P ./models/checkpoints/\n",
-    "#!wget -c https://huggingface.co/comfyanonymous/wd-1.5-beta2_unCLIP/resolve/main/wd-1-5-beta2-aesthetic-unclip-h-fp16.safetensors -P ./models/checkpoints/\n",
-    "\n",
-    "\n",
-    "# VAE\n",
-    "!wget -c https://huggingface.co/stabilityai/sd-vae-ft-mse-original/resolve/main/vae-ft-mse-840000-ema-pruned.safetensors -P ./models/vae/\n",
-    "#!wget -c https://huggingface.co/WarriorMama777/OrangeMixs/resolve/main/VAEs/orangemix.vae.pt -P ./models/vae/\n",
-    "#!wget -c https://huggingface.co/hakurei/waifu-diffusion-v1-4/resolve/main/vae/kl-f8-anime2.ckpt -P ./models/vae/\n",
-    "\n",
-    "\n",
-    "# Loras\n",
-    "#!wget -c https://civitai.com/api/download/models/10350 -O ./models/loras/theovercomer8sContrastFix_sd21768.safetensors #theovercomer8sContrastFix SD2.x 768-v\n",
-    "#!wget -c https://civitai.com/api/download/models/10638 -O ./models/loras/theovercomer8sContrastFix_sd15.safetensors #theovercomer8sContrastFix SD1.x\n",
-    "#!wget -c https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/resolve/main/sd_xl_offset_example-lora_1.0.safetensors -P ./models/loras/ #SDXL offset noise lora\n",
-    "\n",
-    "\n",
-    "# T2I-Adapter\n",
-    "#!wget -c https://huggingface.co/TencentARC/T2I-Adapter/resolve/main/models/t2iadapter_depth_sd14v1.pth -P ./models/controlnet/\n",
-    "#!wget -c https://huggingface.co/TencentARC/T2I-Adapter/resolve/main/models/t2iadapter_seg_sd14v1.pth -P ./models/controlnet/\n",
-    "#!wget -c https://huggingface.co/TencentARC/T2I-Adapter/resolve/main/models/t2iadapter_sketch_sd14v1.pth -P ./models/controlnet/\n",
-    "#!wget -c https://huggingface.co/TencentARC/T2I-Adapter/resolve/main/models/t2iadapter_keypose_sd14v1.pth -P ./models/controlnet/\n",
-    "#!wget -c https://huggingface.co/TencentARC/T2I-Adapter/resolve/main/models/t2iadapter_openpose_sd14v1.pth -P ./models/controlnet/\n",
-    "#!wget -c https://huggingface.co/TencentARC/T2I-Adapter/resolve/main/models/t2iadapter_color_sd14v1.pth -P ./models/controlnet/\n",
-    "#!wget -c https://huggingface.co/TencentARC/T2I-Adapter/resolve/main/models/t2iadapter_canny_sd14v1.pth -P ./models/controlnet/\n",
-    "\n",
-    "# T2I Styles Model\n",
-    "#!wget -c https://huggingface.co/TencentARC/T2I-Adapter/resolve/main/models/t2iadapter_style_sd14v1.pth -P ./models/style_models/\n",
-    "\n",
-    "# CLIPVision model (needed for styles model)\n",
-    "#!wget -c https://huggingface.co/openai/clip-vit-large-patch14/resolve/main/pytorch_model.bin -O ./models/clip_vision/clip_vit14.bin\n",
-    "\n",
-    "\n",
-    "# ControlNet\n",
-    "#!wget -c https://huggingface.co/comfyanonymous/ControlNet-v1-1_fp16_safetensors/resolve/main/control_v11e_sd15_ip2p_fp16.safetensors -P ./models/controlnet/\n",
-    "#!wget -c https://huggingface.co/comfyanonymous/ControlNet-v1-1_fp16_safetensors/resolve/main/control_v11e_sd15_shuffle_fp16.safetensors -P ./models/controlnet/\n",
-    "#!wget -c https://huggingface.co/comfyanonymous/ControlNet-v1-1_fp16_safetensors/resolve/main/control_v11p_sd15_canny_fp16.safetensors -P ./models/controlnet/\n",
-    "#!wget -c https://huggingface.co/comfyanonymous/ControlNet-v1-1_fp16_safetensors/resolve/main/control_v11f1p_sd15_depth_fp16.safetensors -P ./models/controlnet/\n",
-    "#!wget -c https://huggingface.co/comfyanonymous/ControlNet-v1-1_fp16_safetensors/resolve/main/control_v11p_sd15_inpaint_fp16.safetensors -P ./models/controlnet/\n",
-    "#!wget -c https://huggingface.co/comfyanonymous/ControlNet-v1-1_fp16_safetensors/resolve/main/control_v11p_sd15_lineart_fp16.safetensors -P ./models/controlnet/\n",
-    "#!wget -c https://huggingface.co/comfyanonymous/ControlNet-v1-1_fp16_safetensors/resolve/main/control_v11p_sd15_mlsd_fp16.safetensors -P ./models/controlnet/\n",
-    "#!wget -c https://huggingface.co/comfyanonymous/ControlNet-v1-1_fp16_safetensors/resolve/main/control_v11p_sd15_normalbae_fp16.safetensors -P ./models/controlnet/\n",
-    "#!wget -c https://huggingface.co/comfyanonymous/ControlNet-v1-1_fp16_safetensors/resolve/main/control_v11p_sd15_openpose_fp16.safetensors -P ./models/controlnet/\n",
-    "#!wget -c https://huggingface.co/comfyanonymous/ControlNet-v1-1_fp16_safetensors/resolve/main/control_v11p_sd15_scribble_fp16.safetensors -P ./models/controlnet/\n",
-    "#!wget -c https://huggingface.co/comfyanonymous/ControlNet-v1-1_fp16_safetensors/resolve/main/control_v11p_sd15_seg_fp16.safetensors -P ./models/controlnet/\n",
-    "#!wget -c https://huggingface.co/comfyanonymous/ControlNet-v1-1_fp16_safetensors/resolve/main/control_v11p_sd15_softedge_fp16.safetensors -P ./models/controlnet/\n",
-    "#!wget -c https://huggingface.co/comfyanonymous/ControlNet-v1-1_fp16_safetensors/resolve/main/control_v11p_sd15s2_lineart_anime_fp16.safetensors -P ./models/controlnet/\n",
-    "#!wget -c https://huggingface.co/comfyanonymous/ControlNet-v1-1_fp16_safetensors/resolve/main/control_v11u_sd15_tile_fp16.safetensors -P ./models/controlnet/\n",
-    "\n",
-    "# ControlNet SDXL\n",
-    "#!wget -c https://huggingface.co/stabilityai/control-lora/resolve/main/control-LoRAs-rank256/control-lora-canny-rank256.safetensors -P ./models/controlnet/\n",
-    "#!wget -c https://huggingface.co/stabilityai/control-lora/resolve/main/control-LoRAs-rank256/control-lora-depth-rank256.safetensors -P ./models/controlnet/\n",
-    "#!wget -c https://huggingface.co/stabilityai/control-lora/resolve/main/control-LoRAs-rank256/control-lora-recolor-rank256.safetensors -P ./models/controlnet/\n",
-    "#!wget -c https://huggingface.co/stabilityai/control-lora/resolve/main/control-LoRAs-rank256/control-lora-sketch-rank256.safetensors -P ./models/controlnet/\n",
-    "\n",
-    "# Controlnet Preprocessor nodes by Fannovel16\n",
-    "#!cd custom_nodes && git clone https://github.com/Fannovel16/comfy_controlnet_preprocessors; cd comfy_controlnet_preprocessors && python install.py\n",
-    "\n",
-    "\n",
-    "# GLIGEN\n",
-    "#!wget -c https://huggingface.co/comfyanonymous/GLIGEN_pruned_safetensors/resolve/main/gligen_sd14_textbox_pruned_fp16.safetensors -P ./models/gligen/\n",
-    "\n",
-    "\n",
-    "# ESRGAN upscale model\n",
-    "#!wget -c https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.0/RealESRGAN_x4plus.pth -P ./models/upscale_models/\n",
-    "#!wget -c https://huggingface.co/sberbank-ai/Real-ESRGAN/resolve/main/RealESRGAN_x2.pth -P ./models/upscale_models/\n",
-    "#!wget -c https://huggingface.co/sberbank-ai/Real-ESRGAN/resolve/main/RealESRGAN_x4.pth -P ./models/upscale_models/\n",
-    "\n",
-    "\n"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {
-    "id": "kkkkkkkkkkkkkkk"
-   },
-   "source": [
-    "### Run ComfyUI with cloudflared (Recommended Way)\n",
-    "\n",
-    "\n"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {
-    "id": "jjjjjjjjjjjjjj"
-   },
-   "outputs": [],
-   "source": [
-    "!wget https://github.com/cloudflare/cloudflared/releases/latest/download/cloudflared-linux-amd64.deb\n",
-    "!dpkg -i cloudflared-linux-amd64.deb\n",
-    "\n",
-    "import subprocess\n",
-    "import threading\n",
-    "import time\n",
-    "import socket\n",
-    "import urllib.request\n",
-    "\n",
-    "def iframe_thread(port):\n",
-    "  while True:\n",
-    "      time.sleep(0.5)\n",
-    "      sock = socket.socket(socket.AF_INET, socket.SOCK_STREAM)\n",
-    "      result = sock.connect_ex(('127.0.0.1', port))\n",
-    "      if result == 0:\n",
-    "        break\n",
-    "      sock.close()\n",
-    "  print(\"\\nComfyUI finished loading, trying to launch cloudflared (if it gets stuck here cloudflared is having issues)\\n\")\n",
-    "\n",
-    "  p = subprocess.Popen([\"cloudflared\", \"tunnel\", \"--url\", \"http://127.0.0.1:{}\".format(port)], stdout=subprocess.PIPE, stderr=subprocess.PIPE)\n",
-    "  for line in p.stderr:\n",
-    "    l = line.decode()\n",
-    "    if \"trycloudflare.com \" in l:\n",
-    "      print(\"This is the URL to access ComfyUI:\", l[l.find(\"http\"):], end='')\n",
-    "    #print(l, end='')\n",
-    "\n",
-    "\n",
-    "threading.Thread(target=iframe_thread, daemon=True, args=(8188,)).start()\n",
-    "\n",
-    "!python main.py --dont-print-server"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {
-    "id": "kkkkkkkkkkkkkk"
-   },
-   "source": [
-    "### Run ComfyUI with localtunnel\n",
-    "\n",
-    "\n"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {
-    "id": "jjjjjjjjjjjjj"
-   },
-   "outputs": [],
-   "source": [
-    "!npm install -g localtunnel\n",
-    "\n",
-    "import threading\n",
-    "\n",
-    "def iframe_thread(port):\n",
-    "  while True:\n",
-    "      time.sleep(0.5)\n",
-    "      sock = socket.socket(socket.AF_INET, socket.SOCK_STREAM)\n",
-    "      result = sock.connect_ex(('127.0.0.1', port))\n",
-    "      if result == 0:\n",
-    "        break\n",
-    "      sock.close()\n",
-    "  print(\"\\nComfyUI finished loading, trying to launch localtunnel (if it gets stuck here localtunnel is having issues)\\n\")\n",
-    "\n",
-    "  print(\"The password/enpoint ip for localtunnel is:\", urllib.request.urlopen('https://ipv4.icanhazip.com').read().decode('utf8').strip(\"\\n\"))\n",
-    "  p = subprocess.Popen([\"lt\", \"--port\", \"{}\".format(port)], stdout=subprocess.PIPE)\n",
-    "  for line in p.stdout:\n",
-    "    print(line.decode(), end='')\n",
-    "\n",
-    "\n",
-    "threading.Thread(target=iframe_thread, daemon=True, args=(8188,)).start()\n",
-    "\n",
-    "!python main.py --dont-print-server"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {
-    "id": "gggggggggg"
-   },
-   "source": [
-    "### Run ComfyUI with colab iframe (use only in case the previous way with localtunnel doesn't work)\n",
-    "\n",
-    "You should see the ui appear in an iframe. If you get a 403 error, it's your firefox settings or an extension that's messing things up.\n",
-    "\n",
-    "If you want to open it in another window use the link.\n",
-    "\n",
-    "Note that some UI features like live image previews won't work because the colab iframe blocks websockets."
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {
-    "id": "hhhhhhhhhh"
-   },
-   "outputs": [],
-   "source": [
-    "import threading\n",
-    "def iframe_thread(port):\n",
-    "  while True:\n",
-    "      time.sleep(0.5)\n",
-    "      sock = socket.socket(socket.AF_INET, socket.SOCK_STREAM)\n",
-    "      result = sock.connect_ex(('127.0.0.1', port))\n",
-    "      if result == 0:\n",
-    "        break\n",
-    "      sock.close()\n",
-    "  from google.colab import output\n",
-    "  output.serve_kernel_port_as_iframe(port, height=1024)\n",
-    "  print(\"to open it in a window you can open this link here:\")\n",
-    "  output.serve_kernel_port_as_window(port)\n",
-    "\n",
-    "threading.Thread(target=iframe_thread, daemon=True, args=(8188,)).start()\n",
-    "\n",
-    "!python main.py --dont-print-server"
-   ]
-  }
- ],
- "metadata": {
-  "accelerator": "GPU",
-  "colab": {
-   "provenance": []
-  },
-  "gpuClass": "standard",
-  "kernelspec": {
-   "display_name": "Python 3",
-   "name": "python3"
-  },
-  "language_info": {
-   "name": "python"
-  }
- },
- "nbformat": 4,
- "nbformat_minor": 0
-}

protocol.py

@@ -0,0 +1,7 @@
+
+class BinaryEventTypes:
+    PREVIEW_IMAGE = 1
+    UNENCODED_PREVIEW_IMAGE = 2
+    TEXT = 3
+    PREVIEW_IMAGE_WITH_METADATA = 4
+

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "ComfyUI"
-version = "0.3.43"
+version = "0.3.47"
 readme = "README.md"
 license = { file = "LICENSE" }
 requires-python = ">=3.9"

requirements.txt

@@ -1,6 +1,6 @@
 comfyui-frontend-package==1.23.4
-comfyui-workflow-templates==0.1.31
-comfyui-embedded-docs==0.2.3
+comfyui-workflow-templates==0.1.41
+comfyui-embedded-docs==0.2.4
 torch
 torchsde
 torchvision

script_examples/websockets_api_example.py

@@ -10,11 +10,11 @@ import urllib.parse
 server_address = "127.0.0.1:8188"
 client_id = str(uuid.uuid4())
 
-def queue_prompt(prompt):
-    p = {"prompt": prompt, "client_id": client_id}
+def queue_prompt(prompt, prompt_id):
+    p = {"prompt": prompt, "client_id": client_id, "prompt_id": prompt_id}
     data = json.dumps(p).encode('utf-8')
-    req =  urllib.request.Request("http://{}/prompt".format(server_address), data=data)
-    return json.loads(urllib.request.urlopen(req).read())
+    req = urllib.request.Request("http://{}/prompt".format(server_address), data=data)
+    urllib.request.urlopen(req).read()
 
 def get_image(filename, subfolder, folder_type):
     data = {"filename": filename, "subfolder": subfolder, "type": folder_type}
@@ -27,7 +27,8 @@ def get_history(prompt_id):
         return json.loads(response.read())
 
 def get_images(ws, prompt):
-    prompt_id = queue_prompt(prompt)['prompt_id']
+    prompt_id = str(uuid.uuid4())
+    queue_prompt(prompt, prompt_id)
     output_images = {}
     while True:
         out = ws.recv()

server.py

@@ -26,6 +26,7 @@ import mimetypes
 from comfy.cli_args import args
 import comfy.utils
 import comfy.model_management
+from comfy_api import feature_flags
 import node_helpers
 from comfyui_version import __version__
 from app.frontend_management import FrontendManager
@@ -35,11 +36,7 @@ from app.model_manager import ModelFileManager
 from app.custom_node_manager import CustomNodeManager
 from typing import Optional, Union
 from api_server.routes.internal.internal_routes import InternalRoutes
-
-class BinaryEventTypes:
-    PREVIEW_IMAGE = 1
-    UNENCODED_PREVIEW_IMAGE = 2
-    TEXT = 3
+from protocol import BinaryEventTypes
 
 async def send_socket_catch_exception(function, message):
     try:
@@ -178,6 +175,7 @@ class PromptServer():
         max_upload_size = round(args.max_upload_size * 1024 * 1024)
         self.app = web.Application(client_max_size=max_upload_size, middlewares=middlewares)
         self.sockets = dict()
+        self.sockets_metadata = dict()
         self.web_root = (
             FrontendManager.init_frontend(args.front_end_version)
             if args.front_end_root is None
@@ -202,20 +200,53 @@ class PromptServer():
             else:
                 sid = uuid.uuid4().hex
 
+            # Store WebSocket for backward compatibility
             self.sockets[sid] = ws
+            # Store metadata separately
+            self.sockets_metadata[sid] = {"feature_flags": {}}
 
             try:
                 # Send initial state to the new client
-                await self.send("status", { "status": self.get_queue_info(), 'sid': sid }, sid)
+                await self.send("status", {"status": self.get_queue_info(), "sid": sid}, sid)
                 # On reconnect if we are the currently executing client send the current node
                 if self.client_id == sid and self.last_node_id is not None:
                     await self.send("executing", { "node": self.last_node_id }, sid)
 
+                # Flag to track if we've received the first message
+                first_message = True
+
                 async for msg in ws:
                     if msg.type == aiohttp.WSMsgType.ERROR:
                         logging.warning('ws connection closed with exception %s' % ws.exception())
+                    elif msg.type == aiohttp.WSMsgType.TEXT:
+                        try:
+                            data = json.loads(msg.data)
+                            # Check if first message is feature flags
+                            if first_message and data.get("type") == "feature_flags":
+                                # Store client feature flags
+                                client_flags = data.get("data", {})
+                                self.sockets_metadata[sid]["feature_flags"] = client_flags
+
+                                # Send server feature flags in response
+                                await self.send(
+                                    "feature_flags",
+                                    feature_flags.get_server_features(),
+                                    sid,
+                                )
+
+                                logging.info(
+                                    f"Feature flags negotiated for client {sid}: {client_flags}"
+                                )
+                            first_message = False
+                        except json.JSONDecodeError:
+                            logging.warning(
+                                f"Invalid JSON received from client {sid}: {msg.data}"
+                            )
+                        except Exception as e:
+                            logging.error(f"Error processing WebSocket message: {e}")
             finally:
                 self.sockets.pop(sid, None)
+                self.sockets_metadata.pop(sid, None)
             return ws
 
         @routes.get("/")
@@ -522,6 +553,7 @@ class PromptServer():
             ram_free = comfy.model_management.get_free_memory(cpu_device)
             vram_total, torch_vram_total = comfy.model_management.get_total_memory(device, torch_total_too=True)
             vram_free, torch_vram_free = comfy.model_management.get_free_memory(device, torch_free_too=True)
+            required_frontend_version = FrontendManager.get_required_frontend_version()
 
             system_stats = {
                 "system": {
@@ -529,6 +561,7 @@ class PromptServer():
                     "ram_total": ram_total,
                     "ram_free": ram_free,
                     "comfyui_version": __version__,
+                    "required_frontend_version": required_frontend_version,
                     "python_version": sys.version,
                     "pytorch_version": comfy.model_management.torch_version,
                     "embedded_python": os.path.split(os.path.split(sys.executable)[0])[1] == "python_embeded",
@@ -548,6 +581,10 @@ class PromptServer():
             }
             return web.json_response(system_stats)
 
+        @routes.get("/features")
+        async def get_features(request):
+            return web.json_response(feature_flags.get_server_features())
+
         @routes.get("/prompt")
         async def get_prompt(request):
             return web.json_response(self.get_queue_info())
@@ -643,7 +680,8 @@ class PromptServer():
 
             if "prompt" in json_data:
                 prompt = json_data["prompt"]
-                valid = execution.validate_prompt(prompt)
+                prompt_id = str(json_data.get("prompt_id", uuid.uuid4()))
+                valid = await execution.validate_prompt(prompt_id, prompt)
                 extra_data = {}
                 if "extra_data" in json_data:
                     extra_data = json_data["extra_data"]
@@ -651,7 +689,6 @@ class PromptServer():
                 if "client_id" in json_data:
                     extra_data["client_id"] = json_data["client_id"]
                 if valid[0]:
-                    prompt_id = str(uuid.uuid4())
                     outputs_to_execute = valid[2]
                     self.prompt_queue.put((number, prompt_id, prompt, extra_data, outputs_to_execute))
                     response = {"prompt_id": prompt_id, "number": number, "node_errors": valid[3]}
@@ -766,6 +803,10 @@ class PromptServer():
     async def send(self, event, data, sid=None):
         if event == BinaryEventTypes.UNENCODED_PREVIEW_IMAGE:
             await self.send_image(data, sid=sid)
+        elif event == BinaryEventTypes.PREVIEW_IMAGE_WITH_METADATA:
+            # data is (preview_image, metadata)
+            preview_image, metadata = data
+            await self.send_image_with_metadata(preview_image, metadata, sid=sid)
         elif isinstance(data, (bytes, bytearray)):
             await self.send_bytes(event, data, sid)
         else:
@@ -804,6 +845,43 @@ class PromptServer():
         preview_bytes = bytesIO.getvalue()
         await self.send_bytes(BinaryEventTypes.PREVIEW_IMAGE, preview_bytes, sid=sid)
 
+    async def send_image_with_metadata(self, image_data, metadata=None, sid=None):
+        image_type = image_data[0]
+        image = image_data[1]
+        max_size = image_data[2]
+        if max_size is not None:
+            if hasattr(Image, 'Resampling'):
+                resampling = Image.Resampling.BILINEAR
+            else:
+                resampling = Image.Resampling.LANCZOS
+
+            image = ImageOps.contain(image, (max_size, max_size), resampling)
+
+        mimetype = "image/png" if image_type == "PNG" else "image/jpeg"
+
+        # Prepare metadata
+        if metadata is None:
+            metadata = {}
+        metadata["image_type"] = mimetype
+
+        # Serialize metadata as JSON
+        import json
+        metadata_json = json.dumps(metadata).encode('utf-8')
+        metadata_length = len(metadata_json)
+
+        # Prepare image data
+        bytesIO = BytesIO()
+        image.save(bytesIO, format=image_type, quality=95, compress_level=1)
+        image_bytes = bytesIO.getvalue()
+
+        # Combine metadata and image
+        combined_data = bytearray()
+        combined_data.extend(struct.pack(">I", metadata_length))
+        combined_data.extend(metadata_json)
+        combined_data.extend(image_bytes)
+
+        await self.send_bytes(BinaryEventTypes.PREVIEW_IMAGE_WITH_METADATA, combined_data, sid=sid)
+
     async def send_bytes(self, event, data, sid=None):
         message = self.encode_bytes(event, data)
 
@@ -845,10 +923,10 @@ class PromptServer():
         ssl_ctx = None
         scheme = "http"
         if args.tls_keyfile and args.tls_certfile:
-                ssl_ctx = ssl.SSLContext(protocol=ssl.PROTOCOL_TLS_SERVER, verify_mode=ssl.CERT_NONE)
-                ssl_ctx.load_cert_chain(certfile=args.tls_certfile,
+            ssl_ctx = ssl.SSLContext(protocol=ssl.PROTOCOL_TLS_SERVER, verify_mode=ssl.CERT_NONE)
+            ssl_ctx.load_cert_chain(certfile=args.tls_certfile,
                                 keyfile=args.tls_keyfile)
-                scheme = "https"
+            scheme = "https"
 
         if verbose:
             logging.info("Starting server\n")

tests-unit/app_test/frontend_manager_test.py

@@ -1,7 +1,7 @@
 import argparse
 import pytest
 from requests.exceptions import HTTPError
-from unittest.mock import patch
+from unittest.mock import patch, mock_open
 
 from app.frontend_management import (
     FrontendManager,
@@ -172,3 +172,36 @@ def test_init_frontend_fallback_on_error():
     # Assert
     assert frontend_path == "/default/path"
     mock_check.assert_called_once()
+
+
+def test_get_frontend_version():
+    # Arrange
+    expected_version = "1.25.0"
+    mock_requirements_content = """torch
+torchsde
+comfyui-frontend-package==1.25.0
+other-package==1.0.0
+numpy"""
+
+    # Act
+    with patch("builtins.open", mock_open(read_data=mock_requirements_content)):
+        version = FrontendManager.get_required_frontend_version()
+
+    # Assert
+    assert version == expected_version
+
+
+def test_get_frontend_version_invalid_semver():
+    # Arrange
+    mock_requirements_content = """torch
+torchsde
+comfyui-frontend-package==1.29.3.75
+other-package==1.0.0
+numpy"""
+
+    # Act
+    with patch("builtins.open", mock_open(read_data=mock_requirements_content)):
+        version = FrontendManager.get_required_frontend_version()
+
+    # Assert
+    assert version is None

tests-unit/feature_flags_test.py

@@ -0,0 +1,98 @@
+"""Tests for feature flags functionality."""
+
+from comfy_api.feature_flags import (
+    get_connection_feature,
+    supports_feature,
+    get_server_features,
+    SERVER_FEATURE_FLAGS,
+)
+
+
+class TestFeatureFlags:
+    """Test suite for feature flags functions."""
+
+    def test_get_server_features_returns_copy(self):
+        """Test that get_server_features returns a copy of the server flags."""
+        features = get_server_features()
+        # Verify it's a copy by modifying it
+        features["test_flag"] = True
+        # Original should be unchanged
+        assert "test_flag" not in SERVER_FEATURE_FLAGS
+
+    def test_get_server_features_contains_expected_flags(self):
+        """Test that server features contain expected flags."""
+        features = get_server_features()
+        assert "supports_preview_metadata" in features
+        assert features["supports_preview_metadata"] is True
+        assert "max_upload_size" in features
+        assert isinstance(features["max_upload_size"], (int, float))
+
+    def test_get_connection_feature_with_missing_sid(self):
+        """Test getting feature for non-existent session ID."""
+        sockets_metadata = {}
+        result = get_connection_feature(sockets_metadata, "missing_sid", "some_feature")
+        assert result is False  # Default value
+
+    def test_get_connection_feature_with_custom_default(self):
+        """Test getting feature with custom default value."""
+        sockets_metadata = {}
+        result = get_connection_feature(
+            sockets_metadata, "missing_sid", "some_feature", default="custom_default"
+        )
+        assert result == "custom_default"
+
+    def test_get_connection_feature_with_feature_flags(self):
+        """Test getting feature from connection with feature flags."""
+        sockets_metadata = {
+            "sid1": {
+                "feature_flags": {
+                    "supports_preview_metadata": True,
+                    "custom_feature": "value",
+                },
+            }
+        }
+        result = get_connection_feature(sockets_metadata, "sid1", "supports_preview_metadata")
+        assert result is True
+
+        result = get_connection_feature(sockets_metadata, "sid1", "custom_feature")
+        assert result == "value"
+
+    def test_get_connection_feature_missing_feature(self):
+        """Test getting non-existent feature from connection."""
+        sockets_metadata = {
+            "sid1": {"feature_flags": {"existing_feature": True}}
+        }
+        result = get_connection_feature(sockets_metadata, "sid1", "missing_feature")
+        assert result is False
+
+    def test_supports_feature_returns_boolean(self):
+        """Test that supports_feature always returns boolean."""
+        sockets_metadata = {
+            "sid1": {
+                "feature_flags": {
+                    "bool_feature": True,
+                    "string_feature": "value",
+                    "none_feature": None,
+                },
+            }
+        }
+
+        # True boolean feature
+        assert supports_feature(sockets_metadata, "sid1", "bool_feature") is True
+
+        # Non-boolean values should return False
+        assert supports_feature(sockets_metadata, "sid1", "string_feature") is False
+        assert supports_feature(sockets_metadata, "sid1", "none_feature") is False
+        assert supports_feature(sockets_metadata, "sid1", "missing_feature") is False
+
+    def test_supports_feature_with_missing_connection(self):
+        """Test supports_feature with missing connection."""
+        sockets_metadata = {}
+        assert supports_feature(sockets_metadata, "missing_sid", "any_feature") is False
+
+    def test_empty_feature_flags_dict(self):
+        """Test connection with empty feature flags dictionary."""
+        sockets_metadata = {"sid1": {"feature_flags": {}}}
+        result = get_connection_feature(sockets_metadata, "sid1", "any_feature")
+        assert result is False
+        assert supports_feature(sockets_metadata, "sid1", "any_feature") is False

tests-unit/requirements.txt

@@ -1,3 +1,4 @@
 pytest>=7.8.0
 pytest-aiohttp
 pytest-asyncio
+websocket-client

tests-unit/websocket_feature_flags_test.py

@@ -0,0 +1,77 @@
+"""Simplified tests for WebSocket feature flags functionality."""
+from comfy_api import feature_flags
+
+
+class TestWebSocketFeatureFlags:
+    """Test suite for WebSocket feature flags integration."""
+
+    def test_server_feature_flags_response(self):
+        """Test server feature flags are properly formatted."""
+        features = feature_flags.get_server_features()
+
+        # Check expected server features
+        assert "supports_preview_metadata" in features
+        assert features["supports_preview_metadata"] is True
+        assert "max_upload_size" in features
+        assert isinstance(features["max_upload_size"], (int, float))
+
+    def test_progress_py_checks_feature_flags(self):
+        """Test that progress.py checks feature flags before sending metadata."""
+        # This simulates the check in progress.py
+        client_id = "test_client"
+        sockets_metadata = {"test_client": {"feature_flags": {}}}
+
+        # The actual check would be in progress.py
+        supports_metadata = feature_flags.supports_feature(
+            sockets_metadata, client_id, "supports_preview_metadata"
+        )
+
+        assert supports_metadata is False
+
+    def test_multiple_clients_different_features(self):
+        """Test handling multiple clients with different feature support."""
+        sockets_metadata = {
+            "modern_client": {
+                "feature_flags": {"supports_preview_metadata": True}
+            },
+            "legacy_client": {
+                "feature_flags": {}
+            }
+        }
+
+        # Check modern client
+        assert feature_flags.supports_feature(
+            sockets_metadata, "modern_client", "supports_preview_metadata"
+        ) is True
+
+        # Check legacy client
+        assert feature_flags.supports_feature(
+            sockets_metadata, "legacy_client", "supports_preview_metadata"
+        ) is False
+
+    def test_feature_negotiation_message_format(self):
+        """Test the format of feature negotiation messages."""
+        # Client message format
+        client_message = {
+            "type": "feature_flags",
+            "data": {
+                "supports_preview_metadata": True,
+                "api_version": "1.0.0"
+            }
+        }
+
+        # Verify structure
+        assert client_message["type"] == "feature_flags"
+        assert "supports_preview_metadata" in client_message["data"]
+
+        # Server response format (what would be sent)
+        server_features = feature_flags.get_server_features()
+        server_message = {
+            "type": "feature_flags",
+            "data": server_features
+        }
+
+        # Verify structure
+        assert server_message["type"] == "feature_flags"
+        assert "supports_preview_metadata" in server_message["data"]
+        assert server_message["data"]["supports_preview_metadata"] is True

tests/inference/extra_model_paths.yaml

@@ -1,4 +1,4 @@
 # Config for testing nodes
 testing:
-    custom_nodes: tests/inference/testing_nodes
+    custom_nodes: testing_nodes
 

tests/inference/test_async_nodes.py

@@ -0,0 +1,410 @@
+import pytest
+import time
+import torch
+import urllib.error
+import numpy as np
+import subprocess
+
+from pytest import fixture
+from comfy_execution.graph_utils import GraphBuilder
+from tests.inference.test_execution import ComfyClient
+
+
+@pytest.mark.execution
+class TestAsyncNodes:
+    @fixture(scope="class", autouse=True, params=[
+        (False, 0),
+        (True, 0),
+        (True, 100),
+    ])
+    def _server(self, args_pytest, request):
+        pargs = [
+            'python','main.py',
+            '--output-directory', args_pytest["output_dir"],
+            '--listen', args_pytest["listen"],
+            '--port', str(args_pytest["port"]),
+            '--extra-model-paths-config', 'tests/inference/extra_model_paths.yaml',
+        ]
+        use_lru, lru_size = request.param
+        if use_lru:
+            pargs += ['--cache-lru', str(lru_size)]
+        # Running server with args: pargs
+        p = subprocess.Popen(pargs)
+        yield
+        p.kill()
+        torch.cuda.empty_cache()
+
+    @fixture(scope="class", autouse=True)
+    def shared_client(self, args_pytest, _server):
+        client = ComfyClient()
+        n_tries = 5
+        for i in range(n_tries):
+            time.sleep(4)
+            try:
+                client.connect(listen=args_pytest["listen"], port=args_pytest["port"])
+            except ConnectionRefusedError:
+                # Retrying...
+                pass
+            else:
+                break
+        yield client
+        del client
+        torch.cuda.empty_cache()
+
+    @fixture
+    def client(self, shared_client, request):
+        shared_client.set_test_name(f"async_nodes[{request.node.name}]")
+        yield shared_client
+
+    @fixture
+    def builder(self, request):
+        yield GraphBuilder(prefix=request.node.name)
+
+    # Happy Path Tests
+
+    def test_basic_async_execution(self, client: ComfyClient, builder: GraphBuilder):
+        """Test that a basic async node executes correctly."""
+        g = builder
+        image = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1)
+        sleep_node = g.node("TestSleep", value=image.out(0), seconds=0.1)
+        output = g.node("SaveImage", images=sleep_node.out(0))
+
+        result = client.run(g)
+
+        # Verify execution completed
+        assert result.did_run(sleep_node), "Async sleep node should have executed"
+        assert result.did_run(output), "Output node should have executed"
+
+        # Verify the image passed through correctly
+        result_images = result.get_images(output)
+        assert len(result_images) == 1, "Should have 1 image"
+        assert np.array(result_images[0]).min() == 0 and np.array(result_images[0]).max() == 0, "Image should be black"
+
+    def test_multiple_async_parallel_execution(self, client: ComfyClient, builder: GraphBuilder):
+        """Test that multiple async nodes execute in parallel."""
+        g = builder
+        image = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1)
+
+        # Create multiple async sleep nodes with different durations
+        sleep1 = g.node("TestSleep", value=image.out(0), seconds=0.3)
+        sleep2 = g.node("TestSleep", value=image.out(0), seconds=0.4)
+        sleep3 = g.node("TestSleep", value=image.out(0), seconds=0.5)
+
+        # Add outputs for each
+        _output1 = g.node("PreviewImage", images=sleep1.out(0))
+        _output2 = g.node("PreviewImage", images=sleep2.out(0))
+        _output3 = g.node("PreviewImage", images=sleep3.out(0))
+
+        start_time = time.time()
+        result = client.run(g)
+        elapsed_time = time.time() - start_time
+
+        # Should take ~0.5s (max duration) not 1.2s (sum of durations)
+        assert elapsed_time < 0.8, f"Parallel execution took {elapsed_time}s, expected < 0.8s"
+
+        # Verify all nodes executed
+        assert result.did_run(sleep1) and result.did_run(sleep2) and result.did_run(sleep3)
+
+    def test_async_with_dependencies(self, client: ComfyClient, builder: GraphBuilder):
+        """Test async nodes with proper dependency handling."""
+        g = builder
+        image1 = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1)
+        image2 = g.node("StubImage", content="WHITE", height=512, width=512, batch_size=1)
+
+        # Chain of async operations
+        sleep1 = g.node("TestSleep", value=image1.out(0), seconds=0.2)
+        sleep2 = g.node("TestSleep", value=image2.out(0), seconds=0.2)
+
+        # Average depends on both async results
+        average = g.node("TestVariadicAverage", input1=sleep1.out(0), input2=sleep2.out(0))
+        output = g.node("SaveImage", images=average.out(0))
+
+        result = client.run(g)
+
+        # Verify execution order
+        assert result.did_run(sleep1) and result.did_run(sleep2)
+        assert result.did_run(average) and result.did_run(output)
+
+        # Verify averaged result
+        result_images = result.get_images(output)
+        avg_value = np.array(result_images[0]).mean()
+        assert abs(avg_value - 127.5) < 1, f"Average value {avg_value} should be ~127.5"
+
+    def test_async_validate_inputs(self, client: ComfyClient, builder: GraphBuilder):
+        """Test async VALIDATE_INPUTS function."""
+        g = builder
+        # Create a test node with async validation
+        validation_node = g.node("TestAsyncValidation", value=5.0, threshold=10.0)
+        g.node("SaveImage", images=validation_node.out(0))
+
+        # Should pass validation
+        result = client.run(g)
+        assert result.did_run(validation_node)
+
+        # Test validation failure
+        validation_node.inputs['threshold'] = 3.0  # Will fail since value > threshold
+        with pytest.raises(urllib.error.HTTPError):
+            client.run(g)
+
+    def test_async_lazy_evaluation(self, client: ComfyClient, builder: GraphBuilder):
+        """Test async nodes with lazy evaluation."""
+        g = builder
+        input1 = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1)
+        input2 = g.node("StubImage", content="WHITE", height=512, width=512, batch_size=1)
+        mask = g.node("StubMask", value=0.0, height=512, width=512, batch_size=1)
+
+        # Create async nodes that will be evaluated lazily
+        sleep1 = g.node("TestSleep", value=input1.out(0), seconds=0.3)
+        sleep2 = g.node("TestSleep", value=input2.out(0), seconds=0.3)
+
+        # Use lazy mix that only needs sleep1 (mask=0.0)
+        lazy_mix = g.node("TestLazyMixImages", image1=sleep1.out(0), image2=sleep2.out(0), mask=mask.out(0))
+        g.node("SaveImage", images=lazy_mix.out(0))
+
+        start_time = time.time()
+        result = client.run(g)
+        elapsed_time = time.time() - start_time
+
+        # Should only execute sleep1, not sleep2
+        assert elapsed_time < 0.5, f"Should skip sleep2, took {elapsed_time}s"
+        assert result.did_run(sleep1), "Sleep1 should have executed"
+        assert not result.did_run(sleep2), "Sleep2 should have been skipped"
+
+    def test_async_check_lazy_status(self, client: ComfyClient, builder: GraphBuilder):
+        """Test async check_lazy_status function."""
+        g = builder
+        # Create a node with async check_lazy_status
+        lazy_node = g.node("TestAsyncLazyCheck",
+                          input1="value1",
+                          input2="value2",
+                          condition=True)
+        g.node("SaveImage", images=lazy_node.out(0))
+
+        result = client.run(g)
+        assert result.did_run(lazy_node)
+
+    # Error Handling Tests
+
+    def test_async_execution_error(self, client: ComfyClient, builder: GraphBuilder):
+        """Test that async execution errors are properly handled."""
+        g = builder
+        image = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1)
+        # Create an async node that will error
+        error_node = g.node("TestAsyncError", value=image.out(0), error_after=0.1)
+        g.node("SaveImage", images=error_node.out(0))
+
+        try:
+            client.run(g)
+            assert False, "Should have raised an error"
+        except Exception as e:
+            assert 'prompt_id' in e.args[0], f"Did not get proper error message: {e}"
+            assert e.args[0]['node_id'] == error_node.id, "Error should be from async error node"
+
+    def test_async_validation_error(self, client: ComfyClient, builder: GraphBuilder):
+        """Test async validation error handling."""
+        g = builder
+        # Node with async validation that will fail
+        validation_node = g.node("TestAsyncValidationError", value=15.0, max_value=10.0)
+        g.node("SaveImage", images=validation_node.out(0))
+
+        with pytest.raises(urllib.error.HTTPError) as exc_info:
+            client.run(g)
+        # Verify it's a validation error
+        assert exc_info.value.code == 400
+
+    def test_async_timeout_handling(self, client: ComfyClient, builder: GraphBuilder):
+        """Test handling of async operations that timeout."""
+        g = builder
+        image = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1)
+        # Very long sleep that would timeout
+        timeout_node = g.node("TestAsyncTimeout", value=image.out(0), timeout=0.5, operation_time=2.0)
+        g.node("SaveImage", images=timeout_node.out(0))
+
+        try:
+            client.run(g)
+            assert False, "Should have raised a timeout error"
+        except Exception as e:
+            assert 'timeout' in str(e).lower(), f"Expected timeout error, got: {e}"
+
+    def test_concurrent_async_error_recovery(self, client: ComfyClient, builder: GraphBuilder):
+        """Test that workflow can recover after async errors."""
+        g = builder
+        image = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1)
+
+        # First run with error
+        error_node = g.node("TestAsyncError", value=image.out(0), error_after=0.1)
+        g.node("SaveImage", images=error_node.out(0))
+
+        try:
+            client.run(g)
+        except Exception:
+            pass  # Expected
+
+        # Second run should succeed
+        g2 = GraphBuilder(prefix="recovery_test")
+        image2 = g2.node("StubImage", content="WHITE", height=512, width=512, batch_size=1)
+        sleep_node = g2.node("TestSleep", value=image2.out(0), seconds=0.1)
+        g2.node("SaveImage", images=sleep_node.out(0))
+
+        result = client.run(g2)
+        assert result.did_run(sleep_node), "Should be able to run after error"
+
+    def test_sync_error_during_async_execution(self, client: ComfyClient, builder: GraphBuilder):
+        """Test handling when sync node errors while async node is executing."""
+        g = builder
+        image = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1)
+
+        # Async node that takes time
+        sleep_node = g.node("TestSleep", value=image.out(0), seconds=0.5)
+
+        # Sync node that will error immediately
+        error_node = g.node("TestSyncError", value=image.out(0))
+
+        # Both feed into output
+        g.node("PreviewImage", images=sleep_node.out(0))
+        g.node("PreviewImage", images=error_node.out(0))
+
+        try:
+            client.run(g)
+            assert False, "Should have raised an error"
+        except Exception as e:
+            # Verify the sync error was caught even though async was running
+            assert 'prompt_id' in e.args[0]
+
+    # Edge Cases
+
+    def test_async_with_execution_blocker(self, client: ComfyClient, builder: GraphBuilder):
+        """Test async nodes with execution blockers."""
+        g = builder
+        image1 = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1)
+        image2 = g.node("StubImage", content="WHITE", height=512, width=512, batch_size=1)
+
+        # Async sleep nodes
+        sleep1 = g.node("TestSleep", value=image1.out(0), seconds=0.2)
+        sleep2 = g.node("TestSleep", value=image2.out(0), seconds=0.2)
+
+        # Create list of images
+        image_list = g.node("TestMakeListNode", value1=sleep1.out(0), value2=sleep2.out(0))
+
+        # Create list of blocking conditions - [False, True] to block only the second item
+        int1 = g.node("StubInt", value=1)
+        int2 = g.node("StubInt", value=2)
+        block_list = g.node("TestMakeListNode", value1=int1.out(0), value2=int2.out(0))
+
+        # Compare each value against 2, so first is False (1 != 2) and second is True (2 == 2)
+        compare = g.node("TestIntConditions", a=block_list.out(0), b=2, operation="==")
+
+        # Block based on the comparison results
+        blocker = g.node("TestExecutionBlocker", input=image_list.out(0), block=compare.out(0), verbose=False)
+
+        output = g.node("PreviewImage", images=blocker.out(0))
+
+        result = client.run(g)
+        images = result.get_images(output)
+        assert len(images) == 1, "Should have blocked second image"
+
+    def test_async_caching_behavior(self, client: ComfyClient, builder: GraphBuilder):
+        """Test that async nodes are properly cached."""
+        g = builder
+        image = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1)
+        sleep_node = g.node("TestSleep", value=image.out(0), seconds=0.2)
+        g.node("SaveImage", images=sleep_node.out(0))
+
+        # First run
+        result1 = client.run(g)
+        assert result1.did_run(sleep_node), "Should run first time"
+
+        # Second run - should be cached
+        start_time = time.time()
+        result2 = client.run(g)
+        elapsed_time = time.time() - start_time
+
+        assert not result2.did_run(sleep_node), "Should be cached"
+        assert elapsed_time < 0.1, f"Cached run took {elapsed_time}s, should be instant"
+
+    def test_async_with_dynamic_prompts(self, client: ComfyClient, builder: GraphBuilder):
+        """Test async nodes within dynamically generated prompts."""
+        g = builder
+        image1 = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1)
+        image2 = g.node("StubImage", content="WHITE", height=512, width=512, batch_size=1)
+
+        # Node that generates async nodes dynamically
+        dynamic_async = g.node("TestDynamicAsyncGeneration",
+                              image1=image1.out(0),
+                              image2=image2.out(0),
+                              num_async_nodes=3,
+                              sleep_duration=0.2)
+        g.node("SaveImage", images=dynamic_async.out(0))
+
+        start_time = time.time()
+        result = client.run(g)
+        elapsed_time = time.time() - start_time
+
+        # Should execute async nodes in parallel within dynamic prompt
+        assert elapsed_time < 0.5, f"Dynamic async execution took {elapsed_time}s"
+        assert result.did_run(dynamic_async)
+
+    def test_async_resource_cleanup(self, client: ComfyClient, builder: GraphBuilder):
+        """Test that async resources are properly cleaned up."""
+        g = builder
+        image = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1)
+
+        # Create multiple async nodes that use resources
+        resource_nodes = []
+        for i in range(5):
+            node = g.node("TestAsyncResourceUser",
+                         value=image.out(0),
+                         resource_id=f"resource_{i}",
+                         duration=0.1)
+            resource_nodes.append(node)
+            g.node("PreviewImage", images=node.out(0))
+
+        result = client.run(g)
+
+        # Verify all nodes executed
+        for node in resource_nodes:
+            assert result.did_run(node)
+
+        # Run again to ensure resources were cleaned up
+        result2 = client.run(g)
+        # Should be cached but not error due to resource conflicts
+        for node in resource_nodes:
+            assert not result2.did_run(node), "Should be cached"
+
+    def test_async_cancellation(self, client: ComfyClient, builder: GraphBuilder):
+        """Test cancellation of async operations."""
+        # This would require implementing cancellation in the client
+        # For now, we'll test that long-running async operations can be interrupted
+        pass  # TODO: Implement when cancellation API is available
+
+    def test_mixed_sync_async_execution(self, client: ComfyClient, builder: GraphBuilder):
+        """Test workflows with both sync and async nodes."""
+        g = builder
+        image1 = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1)
+        image2 = g.node("StubImage", content="WHITE", height=512, width=512, batch_size=1)
+        mask = g.node("StubMask", value=0.5, height=512, width=512, batch_size=1)
+
+        # Mix of sync and async operations
+        # Sync: lazy mix images
+        sync_op1 = g.node("TestLazyMixImages", image1=image1.out(0), image2=image2.out(0), mask=mask.out(0))
+        # Async: sleep
+        async_op1 = g.node("TestSleep", value=sync_op1.out(0), seconds=0.2)
+        # Sync: custom validation
+        sync_op2 = g.node("TestCustomValidation1", input1=async_op1.out(0), input2=0.5)
+        # Async: sleep again
+        async_op2 = g.node("TestSleep", value=sync_op2.out(0), seconds=0.2)
+
+        output = g.node("SaveImage", images=async_op2.out(0))
+
+        result = client.run(g)
+
+        # Verify all nodes executed in correct order
+        assert result.did_run(sync_op1)
+        assert result.did_run(async_op1)
+        assert result.did_run(sync_op2)
+        assert result.did_run(async_op2)
+
+        # Image should be a mix of black and white (gray)
+        result_images = result.get_images(output)
+        avg_value = np.array(result_images[0]).mean()
+        assert abs(avg_value - 63.75) < 5, f"Average value {avg_value} should be ~63.75"

tests/inference/test_execution.py

@@ -252,7 +252,7 @@ class TestExecution:
 
     @pytest.mark.parametrize("test_type, test_value", [
         ("StubInt", 5),
-        ("StubFloat", 5.0)
+        ("StubMask", 5.0)
     ])
     def test_validation_error_edge1(self, test_type, test_value, client: ComfyClient, builder: GraphBuilder):
         g = builder
@@ -497,6 +497,69 @@ class TestExecution:
         assert numpy.array(images[0]).min() == 63 and numpy.array(images[0]).max() == 63, "Image should have value 0.25"
         assert not result.did_run(test_node), "The execution should have been cached"
 
+    def test_parallel_sleep_nodes(self, client: ComfyClient, builder: GraphBuilder):
+        g = builder
+        image = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1)
+
+        # Create sleep nodes for each duration
+        sleep_node1 = g.node("TestSleep", value=image.out(0), seconds=2.8)
+        sleep_node2 = g.node("TestSleep", value=image.out(0), seconds=2.9)
+        sleep_node3 = g.node("TestSleep", value=image.out(0), seconds=3.0)
+
+        # Add outputs to verify the execution
+        _output1 = g.node("PreviewImage", images=sleep_node1.out(0))
+        _output2 = g.node("PreviewImage", images=sleep_node2.out(0))
+        _output3 = g.node("PreviewImage", images=sleep_node3.out(0))
+
+        start_time = time.time()
+        result = client.run(g)
+        elapsed_time = time.time() - start_time
+
+        # The test should take around 0.4 seconds (the longest sleep duration)
+        # plus some overhead, but definitely less than the sum of all sleeps (0.9s)
+        # We'll allow for up to 0.8s total to account for overhead
+        assert elapsed_time < 4.0, f"Parallel execution took {elapsed_time}s, expected less than 0.8s"
+
+        # Verify that all nodes executed
+        assert result.did_run(sleep_node1), "Sleep node 1 should have run"
+        assert result.did_run(sleep_node2), "Sleep node 2 should have run"
+        assert result.did_run(sleep_node3), "Sleep node 3 should have run"
+
+    def test_parallel_sleep_expansion(self, client: ComfyClient, builder: GraphBuilder):
+        g = builder
+        # Create input images with different values
+        image1 = g.node("StubImage", content="BLACK", height=512, width=512, batch_size=1)
+        image2 = g.node("StubImage", content="WHITE", height=512, width=512, batch_size=1)
+        image3 = g.node("StubImage", content="WHITE", height=512, width=512, batch_size=1)
+
+        # Create a TestParallelSleep node that expands into multiple TestSleep nodes
+        parallel_sleep = g.node("TestParallelSleep",
+                                image1=image1.out(0),
+                                image2=image2.out(0),
+                                image3=image3.out(0),
+                                sleep1=0.4,
+                                sleep2=0.5,
+                                sleep3=0.6)
+        output = g.node("SaveImage", images=parallel_sleep.out(0))
+
+        start_time = time.time()
+        result = client.run(g)
+        elapsed_time = time.time() - start_time
+
+        # Similar to the previous test, expect parallel execution of the sleep nodes
+        # which should complete in less than the sum of all sleeps
+        assert elapsed_time < 0.8, f"Expansion execution took {elapsed_time}s, expected less than 0.8s"
+
+        # Verify the parallel sleep node executed
+        assert result.did_run(parallel_sleep), "ParallelSleep node should have run"
+
+        # Verify we get an image as output (blend of the three input images)
+        result_images = result.get_images(output)
+        assert len(result_images) == 1, "Should have 1 image"
+        # Average pixel value should be around 170 (255 * 2 // 3)
+        avg_value = numpy.array(result_images[0]).mean()
+        assert avg_value == 170, f"Image average value {avg_value} should be 170"
+
     # This tests that nodes with OUTPUT_IS_LIST function correctly when they receive an ExecutionBlocker
     # as input. We also test that when that list (containing an ExecutionBlocker) is passed to a node,
     # only that one entry in the list is blocked.

tests/inference/testing_nodes/testing-pack/__init__.py

@@ -1,23 +1,26 @@
-from .specific_tests import TEST_NODE_CLASS_MAPPINGS, TEST_NODE_DISPLAY_NAME_MAPPINGS
-from .flow_control import FLOW_CONTROL_NODE_CLASS_MAPPINGS, FLOW_CONTROL_NODE_DISPLAY_NAME_MAPPINGS
-from .util import UTILITY_NODE_CLASS_MAPPINGS, UTILITY_NODE_DISPLAY_NAME_MAPPINGS
-from .conditions import CONDITION_NODE_CLASS_MAPPINGS, CONDITION_NODE_DISPLAY_NAME_MAPPINGS
-from .stubs import TEST_STUB_NODE_CLASS_MAPPINGS, TEST_STUB_NODE_DISPLAY_NAME_MAPPINGS
-
-# NODE_CLASS_MAPPINGS = GENERAL_NODE_CLASS_MAPPINGS.update(COMPONENT_NODE_CLASS_MAPPINGS)
-# NODE_DISPLAY_NAME_MAPPINGS = GENERAL_NODE_DISPLAY_NAME_MAPPINGS.update(COMPONENT_NODE_DISPLAY_NAME_MAPPINGS)
-
-NODE_CLASS_MAPPINGS = {}
-NODE_CLASS_MAPPINGS.update(TEST_NODE_CLASS_MAPPINGS)
-NODE_CLASS_MAPPINGS.update(FLOW_CONTROL_NODE_CLASS_MAPPINGS)
-NODE_CLASS_MAPPINGS.update(UTILITY_NODE_CLASS_MAPPINGS)
-NODE_CLASS_MAPPINGS.update(CONDITION_NODE_CLASS_MAPPINGS)
-NODE_CLASS_MAPPINGS.update(TEST_STUB_NODE_CLASS_MAPPINGS)
-
-NODE_DISPLAY_NAME_MAPPINGS = {}
-NODE_DISPLAY_NAME_MAPPINGS.update(TEST_NODE_DISPLAY_NAME_MAPPINGS)
-NODE_DISPLAY_NAME_MAPPINGS.update(FLOW_CONTROL_NODE_DISPLAY_NAME_MAPPINGS)
-NODE_DISPLAY_NAME_MAPPINGS.update(UTILITY_NODE_DISPLAY_NAME_MAPPINGS)
-NODE_DISPLAY_NAME_MAPPINGS.update(CONDITION_NODE_DISPLAY_NAME_MAPPINGS)
-NODE_DISPLAY_NAME_MAPPINGS.update(TEST_STUB_NODE_DISPLAY_NAME_MAPPINGS)
-
+from .specific_tests import TEST_NODE_CLASS_MAPPINGS, TEST_NODE_DISPLAY_NAME_MAPPINGS
+from .flow_control import FLOW_CONTROL_NODE_CLASS_MAPPINGS, FLOW_CONTROL_NODE_DISPLAY_NAME_MAPPINGS
+from .util import UTILITY_NODE_CLASS_MAPPINGS, UTILITY_NODE_DISPLAY_NAME_MAPPINGS
+from .conditions import CONDITION_NODE_CLASS_MAPPINGS, CONDITION_NODE_DISPLAY_NAME_MAPPINGS
+from .stubs import TEST_STUB_NODE_CLASS_MAPPINGS, TEST_STUB_NODE_DISPLAY_NAME_MAPPINGS
+from .async_test_nodes import ASYNC_TEST_NODE_CLASS_MAPPINGS, ASYNC_TEST_NODE_DISPLAY_NAME_MAPPINGS
+
+# NODE_CLASS_MAPPINGS = GENERAL_NODE_CLASS_MAPPINGS.update(COMPONENT_NODE_CLASS_MAPPINGS)
+# NODE_DISPLAY_NAME_MAPPINGS = GENERAL_NODE_DISPLAY_NAME_MAPPINGS.update(COMPONENT_NODE_DISPLAY_NAME_MAPPINGS)
+
+NODE_CLASS_MAPPINGS = {}
+NODE_CLASS_MAPPINGS.update(TEST_NODE_CLASS_MAPPINGS)
+NODE_CLASS_MAPPINGS.update(FLOW_CONTROL_NODE_CLASS_MAPPINGS)
+NODE_CLASS_MAPPINGS.update(UTILITY_NODE_CLASS_MAPPINGS)
+NODE_CLASS_MAPPINGS.update(CONDITION_NODE_CLASS_MAPPINGS)
+NODE_CLASS_MAPPINGS.update(TEST_STUB_NODE_CLASS_MAPPINGS)
+NODE_CLASS_MAPPINGS.update(ASYNC_TEST_NODE_CLASS_MAPPINGS)
+
+NODE_DISPLAY_NAME_MAPPINGS = {}
+NODE_DISPLAY_NAME_MAPPINGS.update(TEST_NODE_DISPLAY_NAME_MAPPINGS)
+NODE_DISPLAY_NAME_MAPPINGS.update(FLOW_CONTROL_NODE_DISPLAY_NAME_MAPPINGS)
+NODE_DISPLAY_NAME_MAPPINGS.update(UTILITY_NODE_DISPLAY_NAME_MAPPINGS)
+NODE_DISPLAY_NAME_MAPPINGS.update(CONDITION_NODE_DISPLAY_NAME_MAPPINGS)
+NODE_DISPLAY_NAME_MAPPINGS.update(TEST_STUB_NODE_DISPLAY_NAME_MAPPINGS)
+NODE_DISPLAY_NAME_MAPPINGS.update(ASYNC_TEST_NODE_DISPLAY_NAME_MAPPINGS)
+

tests/inference/testing_nodes/testing-pack/async_test_nodes.py

@@ -0,0 +1,343 @@
+import torch
+import asyncio
+from typing import Dict
+from comfy.utils import ProgressBar
+from comfy_execution.graph_utils import GraphBuilder
+from comfy.comfy_types.node_typing import ComfyNodeABC
+from comfy.comfy_types import IO
+
+
+class TestAsyncValidation(ComfyNodeABC):
+    """Test node with async VALIDATE_INPUTS."""
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "value": ("FLOAT", {"default": 5.0}),
+                "threshold": ("FLOAT", {"default": 10.0}),
+            },
+        }
+
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "process"
+    CATEGORY = "_for_testing/async"
+
+    @classmethod
+    async def VALIDATE_INPUTS(cls, value, threshold):
+        # Simulate async validation (e.g., checking remote service)
+        await asyncio.sleep(0.05)
+
+        if value > threshold:
+            return f"Value {value} exceeds threshold {threshold}"
+        return True
+
+    def process(self, value, threshold):
+        # Create image based on value
+        intensity = value / 10.0
+        image = torch.ones([1, 512, 512, 3]) * intensity
+        return (image,)
+
+
+class TestAsyncError(ComfyNodeABC):
+    """Test node that errors during async execution."""
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "value": (IO.ANY, {}),
+                "error_after": ("FLOAT", {"default": 0.1, "min": 0.0, "max": 10.0}),
+            },
+        }
+
+    RETURN_TYPES = (IO.ANY,)
+    FUNCTION = "error_execution"
+    CATEGORY = "_for_testing/async"
+
+    async def error_execution(self, value, error_after):
+        await asyncio.sleep(error_after)
+        raise RuntimeError("Intentional async execution error for testing")
+
+
+class TestAsyncValidationError(ComfyNodeABC):
+    """Test node with async validation that always fails."""
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "value": ("FLOAT", {"default": 5.0}),
+                "max_value": ("FLOAT", {"default": 10.0}),
+            },
+        }
+
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "process"
+    CATEGORY = "_for_testing/async"
+
+    @classmethod
+    async def VALIDATE_INPUTS(cls, value, max_value):
+        await asyncio.sleep(0.05)
+        # Always fail validation for values > max_value
+        if value > max_value:
+            return f"Async validation failed: {value} > {max_value}"
+        return True
+
+    def process(self, value, max_value):
+        # This won't be reached if validation fails
+        image = torch.ones([1, 512, 512, 3]) * (value / max_value)
+        return (image,)
+
+
+class TestAsyncTimeout(ComfyNodeABC):
+    """Test node that simulates timeout scenarios."""
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "value": (IO.ANY, {}),
+                "timeout": ("FLOAT", {"default": 1.0, "min": 0.1, "max": 10.0}),
+                "operation_time": ("FLOAT", {"default": 2.0, "min": 0.1, "max": 10.0}),
+            },
+        }
+
+    RETURN_TYPES = (IO.ANY,)
+    FUNCTION = "timeout_execution"
+    CATEGORY = "_for_testing/async"
+
+    async def timeout_execution(self, value, timeout, operation_time):
+        try:
+            # This will timeout if operation_time > timeout
+            await asyncio.wait_for(asyncio.sleep(operation_time), timeout=timeout)
+            return (value,)
+        except asyncio.TimeoutError:
+            raise RuntimeError(f"Operation timed out after {timeout} seconds")
+
+
+class TestSyncError(ComfyNodeABC):
+    """Test node that errors synchronously (for mixed sync/async testing)."""
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "value": (IO.ANY, {}),
+            },
+        }
+
+    RETURN_TYPES = (IO.ANY,)
+    FUNCTION = "sync_error"
+    CATEGORY = "_for_testing/async"
+
+    def sync_error(self, value):
+        raise RuntimeError("Intentional sync execution error for testing")
+
+
+class TestAsyncLazyCheck(ComfyNodeABC):
+    """Test node with async check_lazy_status."""
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "input1": (IO.ANY, {"lazy": True}),
+                "input2": (IO.ANY, {"lazy": True}),
+                "condition": ("BOOLEAN", {"default": True}),
+            },
+        }
+
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "process"
+    CATEGORY = "_for_testing/async"
+
+    async def check_lazy_status(self, condition, input1, input2):
+        # Simulate async checking (e.g., querying remote service)
+        await asyncio.sleep(0.05)
+
+        needed = []
+        if condition and input1 is None:
+            needed.append("input1")
+        if not condition and input2 is None:
+            needed.append("input2")
+        return needed
+
+    def process(self, input1, input2, condition):
+        # Return a simple image
+        return (torch.ones([1, 512, 512, 3]),)
+
+
+class TestDynamicAsyncGeneration(ComfyNodeABC):
+    """Test node that dynamically generates async nodes."""
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "image1": ("IMAGE",),
+                "image2": ("IMAGE",),
+                "num_async_nodes": ("INT", {"default": 3, "min": 1, "max": 10}),
+                "sleep_duration": ("FLOAT", {"default": 0.2, "min": 0.1, "max": 1.0}),
+            },
+        }
+
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "generate_async_workflow"
+    CATEGORY = "_for_testing/async"
+
+    def generate_async_workflow(self, image1, image2, num_async_nodes, sleep_duration):
+        g = GraphBuilder()
+
+        # Create multiple async sleep nodes
+        sleep_nodes = []
+        for i in range(num_async_nodes):
+            image = image1 if i % 2 == 0 else image2
+            sleep_node = g.node("TestSleep", value=image, seconds=sleep_duration)
+            sleep_nodes.append(sleep_node)
+
+        # Average all results
+        if len(sleep_nodes) == 1:
+            final_node = sleep_nodes[0]
+        else:
+            avg_inputs = {"input1": sleep_nodes[0].out(0)}
+            for i, node in enumerate(sleep_nodes[1:], 2):
+                avg_inputs[f"input{i}"] = node.out(0)
+            final_node = g.node("TestVariadicAverage", **avg_inputs)
+
+        return {
+            "result": (final_node.out(0),),
+            "expand": g.finalize(),
+        }
+
+
+class TestAsyncResourceUser(ComfyNodeABC):
+    """Test node that uses resources during async execution."""
+
+    # Class-level resource tracking for testing
+    _active_resources: Dict[str, bool] = {}
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "value": (IO.ANY, {}),
+                "resource_id": ("STRING", {"default": "resource_0"}),
+                "duration": ("FLOAT", {"default": 0.1, "min": 0.0, "max": 1.0}),
+            },
+        }
+
+    RETURN_TYPES = (IO.ANY,)
+    FUNCTION = "use_resource"
+    CATEGORY = "_for_testing/async"
+
+    async def use_resource(self, value, resource_id, duration):
+        # Check if resource is already in use
+        if self._active_resources.get(resource_id, False):
+            raise RuntimeError(f"Resource {resource_id} is already in use!")
+
+        # Mark resource as in use
+        self._active_resources[resource_id] = True
+
+        try:
+            # Simulate resource usage
+            await asyncio.sleep(duration)
+            return (value,)
+        finally:
+            # Always clean up resource
+            self._active_resources[resource_id] = False
+
+
+class TestAsyncBatchProcessing(ComfyNodeABC):
+    """Test async processing of batched inputs."""
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "images": ("IMAGE",),
+                "process_time_per_item": ("FLOAT", {"default": 0.1, "min": 0.01, "max": 1.0}),
+            },
+            "hidden": {
+                "unique_id": "UNIQUE_ID",
+            },
+        }
+
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "process_batch"
+    CATEGORY = "_for_testing/async"
+
+    async def process_batch(self, images, process_time_per_item, unique_id):
+        batch_size = images.shape[0]
+        pbar = ProgressBar(batch_size, node_id=unique_id)
+
+        # Process each image in the batch
+        processed = []
+        for i in range(batch_size):
+            # Simulate async processing
+            await asyncio.sleep(process_time_per_item)
+
+            # Simple processing: invert the image
+            processed_image = 1.0 - images[i:i+1]
+            processed.append(processed_image)
+
+            pbar.update(1)
+
+        # Stack processed images
+        result = torch.cat(processed, dim=0)
+        return (result,)
+
+
+class TestAsyncConcurrentLimit(ComfyNodeABC):
+    """Test concurrent execution limits for async nodes."""
+
+    _semaphore = asyncio.Semaphore(2)  # Only allow 2 concurrent executions
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "value": (IO.ANY, {}),
+                "duration": ("FLOAT", {"default": 0.5, "min": 0.1, "max": 2.0}),
+                "node_id": ("INT", {"default": 0}),
+            },
+        }
+
+    RETURN_TYPES = (IO.ANY,)
+    FUNCTION = "limited_execution"
+    CATEGORY = "_for_testing/async"
+
+    async def limited_execution(self, value, duration, node_id):
+        async with self._semaphore:
+            # Node {node_id} acquired semaphore
+            await asyncio.sleep(duration)
+            # Node {node_id} releasing semaphore
+            return (value,)
+
+
+# Add node mappings
+ASYNC_TEST_NODE_CLASS_MAPPINGS = {
+    "TestAsyncValidation": TestAsyncValidation,
+    "TestAsyncError": TestAsyncError,
+    "TestAsyncValidationError": TestAsyncValidationError,
+    "TestAsyncTimeout": TestAsyncTimeout,
+    "TestSyncError": TestSyncError,
+    "TestAsyncLazyCheck": TestAsyncLazyCheck,
+    "TestDynamicAsyncGeneration": TestDynamicAsyncGeneration,
+    "TestAsyncResourceUser": TestAsyncResourceUser,
+    "TestAsyncBatchProcessing": TestAsyncBatchProcessing,
+    "TestAsyncConcurrentLimit": TestAsyncConcurrentLimit,
+}
+
+ASYNC_TEST_NODE_DISPLAY_NAME_MAPPINGS = {
+    "TestAsyncValidation": "Test Async Validation",
+    "TestAsyncError": "Test Async Error",
+    "TestAsyncValidationError": "Test Async Validation Error",
+    "TestAsyncTimeout": "Test Async Timeout",
+    "TestSyncError": "Test Sync Error",
+    "TestAsyncLazyCheck": "Test Async Lazy Check",
+    "TestDynamicAsyncGeneration": "Test Dynamic Async Generation",
+    "TestAsyncResourceUser": "Test Async Resource User",
+    "TestAsyncBatchProcessing": "Test Async Batch Processing",
+    "TestAsyncConcurrentLimit": "Test Async Concurrent Limit",
+}

tests/inference/testing_nodes/testing-pack/specific_tests.py

@@ -1,6 +1,11 @@
 import torch
+import time
+import asyncio
+from comfy.utils import ProgressBar
 from .tools import VariantSupport
 from comfy_execution.graph_utils import GraphBuilder
+from comfy.comfy_types.node_typing import ComfyNodeABC
+from comfy.comfy_types import IO
 
 class TestLazyMixImages:
     @classmethod
@@ -333,6 +338,131 @@ class TestMixedExpansionReturns:
                 "expand": g.finalize(),
             }
 
+class TestSamplingInExpansion:
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "model": ("MODEL",),
+                "clip": ("CLIP",),
+                "vae": ("VAE",),
+                "seed": ("INT", {"default": 0, "min": 0, "max": 0xffffffffffffffff}),
+                "steps": ("INT", {"default": 20, "min": 1, "max": 100}),
+                "cfg": ("FLOAT", {"default": 7.0, "min": 0.0, "max": 30.0}),
+                "prompt": ("STRING", {"multiline": True, "default": "a beautiful landscape with mountains and trees"}),
+                "negative_prompt": ("STRING", {"multiline": True, "default": "blurry, bad quality, worst quality"}),
+            },
+        }
+
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "sampling_in_expansion"
+
+    CATEGORY = "Testing/Nodes"
+
+    def sampling_in_expansion(self, model, clip, vae, seed, steps, cfg, prompt, negative_prompt):
+        g = GraphBuilder()
+
+        # Create a basic image generation workflow using the input model, clip and vae
+        # 1. Setup text prompts using the provided CLIP model
+        positive_prompt = g.node("CLIPTextEncode",
+                               text=prompt,
+                               clip=clip)
+        negative_prompt = g.node("CLIPTextEncode",
+                                text=negative_prompt,
+                                clip=clip)
+
+        # 2. Create empty latent with specified size
+        empty_latent = g.node("EmptyLatentImage", width=512, height=512, batch_size=1)
+
+        # 3. Setup sampler and generate image latent
+        sampler = g.node("KSampler",
+                        model=model,
+                        positive=positive_prompt.out(0),
+                        negative=negative_prompt.out(0),
+                        latent_image=empty_latent.out(0),
+                        seed=seed,
+                        steps=steps,
+                        cfg=cfg,
+                        sampler_name="euler_ancestral",
+                        scheduler="normal")
+
+        # 4. Decode latent to image using VAE
+        output = g.node("VAEDecode", samples=sampler.out(0), vae=vae)
+
+        return {
+            "result": (output.out(0),),
+            "expand": g.finalize(),
+        }
+
+class TestSleep(ComfyNodeABC):
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "value": (IO.ANY, {}),
+                "seconds": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 9999.0, "step": 0.01, "tooltip": "The amount of seconds to sleep."}),
+            },
+            "hidden": {
+                "unique_id": "UNIQUE_ID",
+            },
+        }
+    RETURN_TYPES = (IO.ANY,)
+    FUNCTION = "sleep"
+
+    CATEGORY = "_for_testing"
+
+    async def sleep(self, value, seconds, unique_id):
+        pbar = ProgressBar(seconds, node_id=unique_id)
+        start = time.time()
+        expiration = start + seconds
+        now = start
+        while now < expiration:
+            now = time.time()
+            pbar.update_absolute(now - start)
+            await asyncio.sleep(0.01)
+        return (value,)
+
+class TestParallelSleep(ComfyNodeABC):
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "image1": ("IMAGE", ),
+                "image2": ("IMAGE", ),
+                "image3": ("IMAGE", ),
+                "sleep1": ("FLOAT", {"default": 0.5, "min": 0.0, "max": 10.0, "step": 0.01}),
+                "sleep2": ("FLOAT", {"default": 0.5, "min": 0.0, "max": 10.0, "step": 0.01}),
+                "sleep3": ("FLOAT", {"default": 0.5, "min": 0.0, "max": 10.0, "step": 0.01}),
+            },
+            "hidden": {
+                "unique_id": "UNIQUE_ID",
+            },
+        }
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "parallel_sleep"
+    CATEGORY = "_for_testing"
+    OUTPUT_NODE = True
+
+    def parallel_sleep(self, image1, image2, image3, sleep1, sleep2, sleep3, unique_id):
+        # Create a graph dynamically with three TestSleep nodes
+        g = GraphBuilder()
+
+        # Create sleep nodes for each duration and image
+        sleep_node1 = g.node("TestSleep", value=image1, seconds=sleep1)
+        sleep_node2 = g.node("TestSleep", value=image2, seconds=sleep2)
+        sleep_node3 = g.node("TestSleep", value=image3, seconds=sleep3)
+
+        # Blend the results using TestVariadicAverage
+        blend = g.node("TestVariadicAverage",
+                       input1=sleep_node1.out(0),
+                       input2=sleep_node2.out(0),
+                       input3=sleep_node3.out(0))
+
+        return {
+            "result": (blend.out(0),),
+            "expand": g.finalize(),
+        }
+
 TEST_NODE_CLASS_MAPPINGS = {
     "TestLazyMixImages": TestLazyMixImages,
     "TestVariadicAverage": TestVariadicAverage,
@@ -345,6 +475,9 @@ TEST_NODE_CLASS_MAPPINGS = {
     "TestCustomValidation5": TestCustomValidation5,
     "TestDynamicDependencyCycle": TestDynamicDependencyCycle,
     "TestMixedExpansionReturns": TestMixedExpansionReturns,
+    "TestSamplingInExpansion": TestSamplingInExpansion,
+    "TestSleep": TestSleep,
+    "TestParallelSleep": TestParallelSleep,
 }
 
 TEST_NODE_DISPLAY_NAME_MAPPINGS = {
@@ -359,4 +492,7 @@ TEST_NODE_DISPLAY_NAME_MAPPINGS = {
     "TestCustomValidation5": "Custom Validation 5",
     "TestDynamicDependencyCycle": "Dynamic Dependency Cycle",
     "TestMixedExpansionReturns": "Mixed Expansion Returns",
+    "TestSamplingInExpansion": "Sampling In Expansion",
+    "TestSleep": "Test Sleep",
+    "TestParallelSleep": "Test Parallel Sleep",
 }