EricRollei-Eric_Image_Processing_Nodes/Docs/MODEL_WEIGHTS.md

Model Weights Guide

This document provides information on obtaining pretrained model weights for Eric's Image Processing Nodes.

Automatic Downloads

Most models will download automatically when you first use a node. The download happens only once, and weights are cached for future use.

Cache Location: ComfyUI/custom_nodes/Eric_Image_Processing_Nodes/models/pretrained_weights/

Model Weights by Node

1. DnCNN Denoise Node

What it does: Classic CNN denoiser for various noise levels

Pretrained Models:

Model	Noise Level	Channels	Auto-Download	Size
color_blind	Unknown (blind)	RGB	✅ Yes	~670 KB
grayscale_blind	Unknown (blind)	Grayscale	✅ Yes	~670 KB
grayscale_sigma15	σ=15	Grayscale	✅ Yes	~670 KB
grayscale_sigma25	σ=25	Grayscale	✅ Yes	~670 KB
grayscale_sigma50	σ=50	Grayscale	✅ Yes	~670 KB

Source: https://github.com/cszn/DnCNN

Citation:

@article{zhang2017dncnn,
  title={Beyond a Gaussian Denoiser: Residual Learning of Deep CNN for Image Denoising},
  author={Zhang, Kai and Zuo, Wangmeng and Chen, Yunjin and Meng, Deyu and Zhang, Lei},
  journal={IEEE Transactions on Image Processing},
  volume={26},
  number={7},
  pages={3142--3155},
  year={2017},
  publisher={IEEE}
}

License: Academic research use (no explicit license stated)

---

2. SCUNet Node

What it does: Practical blind denoising with Swin Transformer + CNN

Pretrained Models:

Model	Type	Purpose	Auto-Download	Size
scunet_color_real_psnr	PSNR-optimized	Best quality	✅ Yes	~18 MB
scunet_color_real_gan	GAN-trained	Perceptual quality	✅ Yes	~18 MB

Source: https://github.com/cszn/SCUNet

HuggingFace Mirror: https://huggingface.co/datasets/eugenesiow/SCUNet/

Citation:

@inproceedings{zhang2022scunet,
  title={Practical Blind Denoising via Swin-Conv-UNet and Data Synthesis},
  author={Zhang, Kai and Li, Yawei and Liang, Jingyun and Cao, Jiezhang and Zhang, Yulun and Tang, Hao and Timofte, Radu and Van Gool, Luc},
  booktitle={European Conference on Computer Vision},
  pages={593--610},
  year={2022},
  organization={Springer}
}

License: Not explicitly stated (academic research)

---

3. SwinIR Node

What it does: Image restoration using Swin Transformer

Pretrained Models:

Model	Task	Scale	Auto-Download	Size
001_classicalSR_DF2K_s64w8_SwinIR-M_x2	Super-resolution	2x	✅ Yes	~11 MB
001_classicalSR_DF2K_s64w8_SwinIR-M_x4	Super-resolution	4x	✅ Yes	~11 MB
001_classicalSR_DIV2K_s48w8_SwinIR-M_x8	Super-resolution	8x	✅ Yes	~11 MB
002_lightweightSR_DIV2K_s64w8_SwinIR-S_x4	Lightweight SR	4x	✅ Yes	~1 MB
003_realSR_BSRGAN_DFO_s64w8_SwinIR-M_x4_GAN	Real-world SR (GAN)	4x	✅ Yes	~12 MB
003_realSR_BSRGAN_DFO_s64w8_SwinIR-M_x4_PSNR	Real-world SR (PSNR)	4x	✅ Yes	~12 MB
004_grayDN_DFWB_s128w8_SwinIR-M_noise15	Grayscale denoise	σ=15	✅ Yes	~11 MB
004_grayDN_DFWB_s128w8_SwinIR-M_noise25	Grayscale denoise	σ=25	✅ Yes	~11 MB
004_grayDN_DFWB_s128w8_SwinIR-M_noise50	Grayscale denoise	σ=50	✅ Yes	~11 MB
005_colorDN_DFWB_s128w8_SwinIR-M_noise15	Color denoise	σ=15	✅ Yes	~11 MB
005_colorDN_DFWB_s128w8_SwinIR-M_noise25	Color denoise	σ=25	✅ Yes	~11 MB
005_colorDN_DFWB_s128w8_SwinIR-M_noise50	Color denoise	σ=50	✅ Yes	~11 MB

Source: https://github.com/JingyunLiang/SwinIR

Download Link: https://github.com/JingyunLiang/SwinIR/releases/

Citation:

@inproceedings{liang2021swinir,
  title={SwinIR: Image Restoration Using Swin Transformer},
  author={Liang, Jingyun and Cao, Jiezhang and Sun, Guolei and Zhang, Kai and Van Gool, Luc and Timofte, Radu},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops},
  pages={1833--1844},
  year={2021}
}

License: Apache 2.0

---

4. NAFNet Node

What it does: Simple baseline for image restoration

Pretrained Models:

Model	Task	Purpose	Auto-Download	Size
NAFNet-width32.pth	Denoising	General denoising	✅ Yes	~6 MB
NAFNet-width64.pth	Deblurring	Motion/defocus blur	✅ Yes	~23 MB

Source: https://github.com/megvii-research/NAFNet

Citation:

@inproceedings{chen2022nafnet,
  title={Simple Baselines for Image Restoration},
  author={Chen, Liangyu and Chu, Xiaojie and Zhang, Xiangyu and Sun, Jian},
  booktitle={European Conference on Computer Vision},
  pages={17--33},
  year={2022},
  organization={Springer}
}

License: MIT License

---

5. Noise-DA Node

What it does: Noise-aware diffusion for denoising and deblurring

Pretrained Models:

Model	Task	Status	Size
noise_da_denoise.pth	Denoising	Custom-trained	~45 MB
noise_da_deblur.pth	Deblurring	Custom-trained	~45 MB

Note: These models are trained specifically for this project. They are NOT automatically downloaded as they are not from a public source. You can train your own or request pre-trained weights from the author.

Training: See Docs/TRAINING_GUIDE.md for instructions on training your own Noise-DA models.

---

6. Film Grain Analysis NN

What it does: Neural network for film grain detection and classification

Pretrained Models:

Model	Purpose	Status	Size
fga_nn_model.pth	Grain analysis	Custom-trained	~12 MB

Note: This model is trained on a custom dataset of film grain samples. It's NOT automatically downloaded. You can train your own using the training script at scripts/train_film_grain_models.py.

Training Data Requirements:

• Authentic film grain samples
• Digital noise samples
• Synthetic overlay samples
• Clean reference images

---

7. BM3D Node (No Weights Required)

What it does: Block-matching 3D filtering algorithm

No pretrained weights needed! BM3D is an algorithmic approach that doesn't require neural network training.

Implementation: Uses pytorch-bm3d library (https://github.com/lizhihao6/pytorch-bm3d)

Citation:

@article{dabov2007bm3d,
  title={Image Denoising by Sparse 3-D Transform-Domain Collaborative Filtering},
  author={Dabov, Kostadin and Foi, Alessandro and Katkovnik, Vladimir and Egiazarian, Karen},
  journal={IEEE Transactions on Image Processing},
  volume={16},
  number={8},
  pages={2080--2095},
  year={2007},
  publisher={IEEE}
}

License: MIT License (pytorch-bm3d implementation)

---

8. DeepInv Service (External Models)

What it does: Diffusion-based denoising via external service

Models Used by Service:

• DiffUNet: Diffusion-based U-Net
• RAM: Recognize Anything Model (for image understanding)
• SwinIR: Swin Transformer for restoration
• DRUNet: Deep Residual U-Net
• DnCNN: Denoising CNN (via DeepInv wrapper)

Note: These models are loaded by the external DeepInv service, not by ComfyUI directly. See Docs/DEEPINV_SERVICE_GUIDE.md for setup instructions.

Source: https://github.com/deepinv/deepinv

License: BSD 3-Clause License

---

Manual Download Instructions

If automatic download fails, you can manually download weights:

Step 1: Download the Model

Visit the source repository and download the .pth or .pt file.

Step 2: Place in Correct Directory

ComfyUI/
└── custom_nodes/
    └── Eric_Image_Processing_Nodes/
        └── models/
            └── pretrained_weights/
                ├── dncnn/
                │   ├── color_blind.pth
                │   ├── grayscale_blind.pth
                │   └── ...
                ├── scunet/
                │   ├── scunet_color_real_psnr.pth
                │   └── scunet_color_real_gan.pth
                ├── swinir/
                │   ├── 001_classicalSR_DF2K_s64w8_SwinIR-M_x2.pth
                │   └── ...
                └── nafnet/
                    ├── NAFNet-width32.pth
                    └── NAFNet-width64.pth

Step 3: Restart ComfyUI

The nodes will detect the manually placed weights on next launch.

---

Troubleshooting

Download Fails

Symptoms: Error message about failed download, timeout, or connection refused.

Solutions:

1. Check internet connection: Ensure you can access GitHub and HuggingFace
2. Check firewall: Some corporate firewalls block GitHub raw content
3. Use VPN: If in a country with restricted access
4. Manual download: Follow instructions above
5. Check disk space: Ensure enough space in models/ directory

Model Not Found

Symptoms: Error message about missing model file.

Solutions:

1. Check file path: Ensure model is in correct subdirectory
2. Check filename: Must match expected name exactly (case-sensitive on Linux)
3. Check file corruption: Re-download if file size is wrong
4. Check permissions: Ensure ComfyUI can read the file

Out of Memory

Symptoms: CUDA out of memory error when loading model.

Solutions:

1. Use smaller model: Choose a lighter variant (e.g., NAFNet-width32 instead of width64)
2. Reduce batch size: Process images one at a time
3. Close other GPU applications: Free up VRAM
4. Use CPU mode: Set device to CPU in node settings (slower but uses system RAM)

---

Storage Requirements

Minimum Storage: ~200 MB (basic models)

Recommended Storage: 1-2 GB (all models)

Full Collection: 3-5 GB (all models + variants)

Breakdown by Model Family:

• DnCNN: ~3 MB total (5 models)
• SCUNet: ~36 MB (2 models)
• SwinIR: ~140 MB (12 models)
• NAFNet: ~30 MB (2 models)
• Noise-DA: ~90 MB (2 models, if trained)
• Film Grain NN: ~12 MB (if trained)

---

License Compliance

When using these pretrained models, please ensure you:

1. ✅ Cite the original papers in any research publications
2. ✅ Respect academic use restrictions (most models are research-only)
3. ✅ Check commercial use permissions before using in commercial products
4. ✅ Provide attribution in derivative works
5. ✅ Comply with individual model licenses (see each model's source repository)

For commercial use, contact the original model authors for licensing:

• DnCNN/SCUNet/SwinIR: Kai Zhang (cskaizhang@gmail.com)
• NAFNet: Megvii Research (see their GitHub)
• DeepInv models: Check individual model sources

---

Additional Resources

• Main Documentation: Docs/README.md
• DeepInv Service Setup: Docs/DEEPINV_SERVICE_GUIDE.md
• Training Custom Models: Docs/TRAINING_GUIDE.md
• GPU Acceleration: Docs/GPU_BM3D_QUICK_START.md
• BM3D Technical Details: Docs/BM3D_Resolution_Scaling_Guide.md

---

Questions? Contact Eric Hiss at eric@historic.camera or eric@rollei.us