The official implementation of AAAI24 paper Multi-Domain Multi-Scale Diffusion Model for Low-Light Image Enhancement.

Environment

create a new conda env, and run

$ pip install -r requirements.txt

torch/torchvision with CUDA version >= 11.3 should be fine.

Demo

1. Download pretrained model

Download the Pretrained MDMS model from Baidu NetDisk or Google Drive.

Put the downloaded ckpt in datasets/scratch/LLIE/ckpts.

2. Inference

# in {path_to_this_repo}/,
$ python eval_diffusion.py

Put the test input in datasets/scratch/LLIE/data/lowlight/test/input.

Output results will be saved in results/images/lowlight/lowlight.

Evaluation

Put the test GT in datasets/scratch/LLIE/data/lowlight/test/gt for paired evaluation.

# in {path_to_this_repo}/,
$ python evaluation.py

• Note that our evaluation metrics are slightly different from PyDiff (inherited from BasicSR).

Results

All results listed in our paper including the compared methods are available in Baidu Netdisk or Google Drive.

• Note that the provided model is trained on the LOLv1 training set, but generalizes well on other datasets.
• For SSIM, we directly calculate the performance on RGB channel rather than just grayscale images in PyDiff.
• For LPIPS, we use a different normalization method (NormA) compared to PyDiff (NormB).

Our method remains superior under the same setting as PyDiff.

1. Test results on LOLv1 test set.

2. Generalization results on LOLv2 syn and real test sets.

3. Generalization results on other unpaired datasets.

We will perform more training and tests on other datasets in the future.

Training

Put the training dataset in datasets/scratch/LLIE/data/lowlight/train.

# in {path_to_this_repo}/,
$ python train_diffusion.py

Detailed training instructions will be updated soon.

Citation

If you find this paper useful, please consider staring this repo and citing our paper:

@inproceedings{shang2024multi,
  title={Multi-Domain Multi-Scale Diffusion Model for Low-Light Image Enhancement},
  author={Shang, Kai and Shao, Mingwen and Wang, Chao and Cheng, Yuanshuo and Wang, Shuigen},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={38},
  number={5},
  pages={4722--4730},
  year={2024}
}