[CVPR25] DnLUT: Ultra-Efficient Color Image Denoising via Channel-Aware Lookup Tables

:construction: This repo is still under construction. Thanks for your waiting!

:star: If DnLUT is helpful to your images or projects, please help star this repo. Thanks! :hugs:

TODO

• Add training codes and config files
• Add LUT transferring and inference codes

Installation

git clone https://github.com/Stephen0808/DnLUT.git
pip install -r requirements.txt

Dataset

We build our training and evaluation dataset by following Restormer. You could build other datasets with their repo as reference. If you want to download the correponding dataset for our task in huggingface, I will later upload them if you put forward your issue.

Gaussian Denoising

• Download training (DIV2K, Flickr2K, WED, BSD) and testing datasets, run

cd build_data
python download_data.py --data train-test --noise gaussian

• Generate image patches from full-resolution training images, run

python generate_patches_dfwb.py 

Real Image Denoising

• Download SIDD training data, run

cd build_data
python download_data.py --data train --noise real

• Generate image patches from full-resolution training images, run

python generate_patches_sidd.py 

Training

cd ~/DnLUT/dn
python 1_train_model_dnlut.py

Transferring

cd ~/DnLUT/dn
python 2_transfer_to_lut_dn.py

Citation

@article{yang2025dnlut,
  title={DnLUT: Ultra-Efficient Color Image Denoising via Channel-Aware Lookup Tables},
  author={Yang, Sidi and Huang, Binxiao and Zhang, Yulun and Yu, Dahai and Yang, Yujiu and Wong, Ngai},
  journal={arXiv preprint arXiv:2503.15931},
  year={2025}
}

Acknowledgement

Our codes are built upon SRLUT and MuLUT. Thanks for their great works.