Extensions/ComfyUI-DCW-Diffusion-Color-Wavelets-Node
ComfyUI Extension

ComfyUI-DCW-Diffusion-Color-Wavelets-Node

A ComfyUI node implementation of the Dynamic Consistency Weighting (DCW) technique that enables high-quality image generation with drastically reduced inference steps.

By DawnW0lf·Created 4 months ago·Updated 4 months ago· 1
DawnW0lf/ComfyUI-DCW-Diffusion-Color-Wavelets-Node
Nodes1
On cloudLocal install
Categorymodel_patches/DCW
Stars1
Updated4 months ago

Nodes (1)

Apply DCW (Wavelet Patch)
model_patches/DCW
Readme

DCW ComfyUI Wrapper

A ComfyUI node implementation of the Dynamic Consistency Weighting (DCW) technique that enables high-quality image generation with drastically reduced inference steps.

What is DCW?

DCW is a post-processing technique that improves the quality of diffusion model without requiring retraining by leveraging wavelet-domain analysis.

Quick Start

Installation

  1. Clone this repository into your ComfyUI custom_nodes folder:
cd ComfyUI/custom_nodes
git clone https://github.com/DawnW0lf/ComfyUI-DCW-Diffusion-Color-Wavelets-Node.git
  1. Install dependencies:
cd comfyui-dcw
pip install -r requirements.txt
  1. Restart ComfyUI

Usage

  1. Load your favorite model
  2. Add the "Apply DCW (Wavelet Patch)" node to your workflow
  3. Connect your model to the node's input

How It Works

DCW applies a wavelet-domain correction to the diffusion model's predictions:

  1. Decompose: Split the input and output into wavelet components (low and high frequency)
  2. Correct: Blend the low-frequency components based on the current timestep
  3. Reconstruct: Inverse wavelet transform to get the corrected output
  4. Scale: Apply adaptive scaling based on the diffusion timestep

The technique is particularly effective because it:

  • Preserves high-frequency details (textures, edges) from the model's natural output
  • Intelligently refines low-frequency components (color, composition) using temporal information
  • Adapts the correction strength based on where you are in the diffusion process

Parameters

  • Model: The diffusion model to patch (required)
  • Strength: Controls the intensity of DCW correction (0.0-4.0)
    • 0.0: No correction, standard generation
    • 0.3: Recommended for ultra-fast 8-step generation
    • 1.0: Medium correction
    • 2.0+: Heavy correction (may introduce artifacts at very low steps)
  • Wavelet: The wavelet basis to use
    • haar: Fastest, good general-purpose choice
    • db2: Finer detail preservation
    • db3: Maximum detail (slightly slower)

Latent Shape Handling

  • Images: [batch, channels, height, width] (4D)
  • Videos: [batch, channels, frames, height, width] (5D)

The implementation automatically detects and handles both formats.

Credits

This implementation is based on the research from:

Original Paper & Code: AMAP-ML/DCW

  • Licensed under CC-BY-NC-SA 4.0

ComfyUI Wrapper: Independent implementation for ComfyUI integration

The DCW technique was developed by the AMAP research team at NVIDIA.

License

This project is licensed under CC-BY-NC-SA 4.0 (Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International).

This work is based on the DCW technique from AMAP-ML/DCW, which is also licensed under CC-BY-NC-SA 4.0.

What this means:

  • ✅ You can use, modify, and distribute this code
  • ✅ You must attribute the original DCW work
  • ❌ You cannot use this for commercial purposes without permission from NVIDIA/AMAP-ML
  • ✅ Any modifications must also be shared under CC-BY-NC-SA 4.0

Contributing

Contributions are welcome! Please feel free to:

  • Report bugs and issues
  • Suggest improvements
  • Submit pull requests for enhancements

Citation

If you use this in research, please cite the original DCW paper:

@article{yu2026eluci,
  title={Elucidating the SNR-t Bias of Diffusion Probabilistic Models},
  author={Meng Yu and Lei Sun and Jianhao Zeng and Xiangxiang Chu and Kun Zhan},
  journal={arXiv preprint arXiv:2604.16044},
  year={2026}
}

Acknowledgments