ComfyUI-NAFNet_plus
Image restoration nodes for ComfyUI: NAFNet, SCUNet, Restormer, SwinIR, Real-ESRGAN, BSRGAN, HINet, MIRNet-v2, DnCNN and FFDNet denoising, deblurring and upscaling. Great for astrophotography. Models download automatically on first use.
ComfyUI-NAFNet_plus
Comprehensive ComfyUI custom nodes for state-of-the-art image restoration - especially good for astrophotography and telescope images.
Supported Models
| Model | Type | Best For | |-------|------|----------| | SCUNet | Blind Denoising | Astrophotography, unknown noise | | Restormer | Transformer Denoising | High-quality restoration | | SwinIR | Swin Transformer | Denoising & Super-Resolution | | Real-ESRGAN | Upscaling + Enhancement | General image enhancement | | BSRGAN | Blind Super-Resolution | Unknown degradation | | MIRNet-v2 | Multi-scale Denoising | Diverse noise types | | HINet | Instance Norm Denoising | Fast denoising | | NAFNet | Domain-Specific | Smartphone noise, GoPro blur | | DnCNN | Classic CNN | Fast Gaussian denoising | | FFDNet | Flexible Denoising | Adjustable noise level |
Installation
ComfyUI Manager / Comfy Registry (recommended) - search for ComfyUI-NAFNet_plus in the Manager, or:
comfy node install comfyui-nafnet-plus
Manual
cd ComfyUI/custom_nodes
git clone https://github.com/marduk191/ComfyUI-NAFNet_plus.git
cd ComfyUI-NAFNet_plus
pip install -r requirements.txt
Models
Models are not included in the repo. Each node downloads the weights it needs automatically the first time it runs, into ComfyUI/models/nafnet_plus/.
To pre-download everything (~2.4GB) or just specific models:
python download_models.py
python download_models.py scunet_color_real_gan.pth RealESRGAN_x4plus.pth
Models already in this pack's own models/ folder (older installs) are still found and used.
Nodes
Recommended for Astrophotography
- SCUNet Denoise (Blind/Real) - Best for unknown noise types
- Restormer Denoise (Real) - State-of-the-art transformer
- SwinIR Denoise - Handles various noise levels (15/25/50)
Super-Resolution / Upscaling
- SwinIR Super-Resolution (4x) - Real-world 4x upscaling
- Real-ESRGAN Upscale - Practical 2x/4x enhancement
- BSRGAN Upscale (Blind SR) - Handles unknown degradation
Fast / Lightweight Denoising
- DnCNN Denoise - Classic fast denoiser
- FFDNet Denoise (Adjustable) - Flexible noise level control
- HINet Denoise (SIDD) - Half Instance Normalization
Domain-Specific
- NAFNet Denoise (SIDD) - Smartphone camera noise (works best on mild, real sensor noise; heavy synthetic noise is out of its training domain)
- NAFNet Deblur (GoPro) - GoPro motion blur only
Model Downloads
Weights are fetched from:
- GitHub Releases (KAIR, SwinIR, Real-ESRGAN, Restormer, MIRNet-v2)
- Google Drive (NAFNet, HINet)
Example Workflows
Ready-made workflows are in example_workflows/ and also show up in ComfyUI's Templates browser under ComfyUI-NAFNet_plus.
Usage Tips
For Astrophotography / Telescope Images
- Try SCUNet (real_gan) first - good balance of sharpness
- Try Restormer for highest quality
- Use SwinIR Denoise (noise level 50) for very noisy images
- Chain with Real-ESRGAN or SwinIR SR for upscaling
Tiling for Large Images
All nodes support tiling for memory efficiency (tiled runs show progress and can be cancelled):
tile_size=0- Auto-tile images >1 megapixeltile_size=512- Manual tile size (adjust based on VRAM)
FFDNet Noise Level
The FFDNet node allows adjusting noise level (0-75):
- Lower values (5-15): Light denoising, preserves detail
- Medium values (25): Balanced
- Higher values (50+): Strong denoising
Credits
- NAFNet - MEGVII Research
- SCUNet - Kai Zhang
- Restormer - Syed Waqas Zamir
- SwinIR - Jingyun Liang
- Real-ESRGAN - Xintao Wang
- BSRGAN - Kai Zhang
- HINet - MEGVII
- MIRNet-v2 - Syed Waqas Zamir
- KAIR - Kai Zhang (DnCNN, FFDNet)
License
MIT License. Model weights are subject to their original licenses.