ComfyUI-NAFNet

ComfyUI custom nodes for NAFNet - Simple Baselines for Image Restoration.

This node pack provides state-of-the-art image denoising, deblurring, and stereo super-resolution capabilities using NAFNet and NAFSSR models from MEGVII Research.

Features

• Image Denoising - Remove noise from images using SIDD-trained models
• Image Deblurring - Remove motion blur using GoPro/REDS-trained models
• Stereo Super-Resolution - Upscale stereo image pairs (2x/4x) using NAFSSR
• Tiled Processing - Handle large images with limited VRAM
• Model Caching - Efficient memory usage with automatic model caching

Nodes

| Node | Description | |------|-------------| | NAFNet Load Model | Load any NAFNet/NAFSSR model for use with NAFNet Restore | | NAFNet Restore | Generic restoration with optional tiled processing | | NAFNet Denoise (SIDD) | Quick denoise using SIDD-trained models | | NAFNet Deblur (GoPro) | Quick deblur using GoPro-trained models | | NAFSSR Stereo Super-Resolution | Upscale stereo image pairs with cross-attention |

Installation

Manual Installation:

cd ComfyUI/custom_nodes
git clone https://github.com/marduk191/ComfyUI-NAFNet.git
cd ComfyUI-NAFNet
pip install -r requirements.txt

Note: Models (1.3 GB) are included via Git LFS. If you don't have Git LFS installed, run the fallback downloader:

python download_models.py

Models

| Model | Task | Size | Best For | |-------|------|------|----------| | NAFNet-SIDD-width32.pth | Denoising | 111 MB | Smartphone photos, general sensor noise | | NAFNet-SIDD-width64.pth | Denoising | 443 MB | Smartphone photos, general sensor noise | | NAFNet-GoPro-width32.pth | Deblurring | 66 MB | GoPro motion blur only | | NAFNet-GoPro-width64.pth | Deblurring | 259 MB | GoPro motion blur only | | NAFNet-REDS-width64.pth | Video Deblurring | 259 MB | Video frames with compression artifacts | | NAFSSR-L_2x.pth | Stereo SR 2x | 92 MB | Stereo image pairs | | NAFSSR-L_4x.pth | Stereo SR 4x | 92 MB | Stereo image pairs |

Important: Model Domain Specificity

NAFNet models are domain-specific - they only work well on images similar to their training data:

• SIDD (Denoising): Trained on smartphone camera noise. Works well on most photos with sensor noise.
• GoPro (Deblurring): Trained specifically on GoPro camera motion blur. Will produce artifacts on other blur types.
• REDS (Video Deblurring): Trained on video frames with specific blur/compression patterns. Not for photos.

If you apply the wrong model to your image, you'll get colorful noise/artifacts instead of restoration.

Usage Examples

Denoise an Image

1. Add Load Image node
2. Add NAFNet Denoise (SIDD) node
3. Connect image output to NAFNet input
4. Choose width64 for best quality or width32 for speed

Deblur an Image

1. Add Load Image node
2. Add NAFNet Deblur (GoPro) node
3. Connect and run

Process Large Images (Tiled)

1. Add Load Image node
2. Add NAFNet Load Model node (select any model)
3. Add NAFNet Restore node
4. Set tile_size to 512 and tile_overlap to 64
5. Connect and run

Stereo Super-Resolution

1. Add two Load Image nodes (left and right views)
2. Add NAFSSR Stereo Super-Resolution node
3. Connect left/right images to corresponding inputs
4. Choose 2x or 4x upscaling
5. Both outputs are upscaled with cross-view attention

Sample Workflows

Sample workflow JSON files are included in the workflows/ folder:

• nafnet_denoise.json - Basic denoising
• nafnet_deblur.json - Basic deblurring
• nafnet_loader_restore.json - Using model loader
• nafnet_tiled_restore.json - Tiled processing for large images
• nafssr_stereo_sr.json - Stereo super-resolution

Performance Tips

• Use width32 models for faster processing with slightly lower quality
• Use width64 models for best quality
• Enable tiled processing for images larger than 1024x1024
• NAFSSR works best with properly aligned stereo pairs

Requirements

• ComfyUI
• PyTorch (CUDA recommended)
• gdown (only needed if Git LFS models fail to download)

Credits

• NAFNet by MEGVII Research
• NAFSSR - Stereo Image Super-Resolution Using NAFNet

Citation

@inproceedings{chen2022simple,
  title={Simple Baselines for Image Restoration},
  author={Chen, Liangyu and Chu, Xiaojie and Zhang, Xiangyu and Sun, Jian},
  booktitle={European Conference on Computer Vision (ECCV)},
  year={2022}
}

@inproceedings{chu2022nafssr,
  title={NAFSSR: Stereo Image Super-Resolution Using NAFNet},
  author={Chu, Xiaojie and Chen, Liangyu and Yu, Wenqing},
  booktitle={CVPR Workshop},
  year={2022}
}

License

This project is released under the MIT License. The NAFNet model weights are subject to their original license from MEGVII Research.