Authors: Yuyang Yin, Dejia Xu, Chuangchuang Tan, Ping Liu, Yao Zhao, Yunchao Wei

Institutions: Beijing Jiaotong University, UT Austin, A*Star

[2308.06725] CLE Diffusion: Controllable Light Enhancement Diffusion Model (arxiv.org)

Project Page：

CLE Diffusion: Controllable Light Enhancement Diffusion Model(ACM MM 2023) (yuyangyin.github.io)

Download LOL dataset from LOL. Put the dataset under '/data/LOL' file.

The code also supports other dataset.

Download Pretrianed model on LOL dataset and put it under 'ckpt' file.

Download Mask CLE Diffusion Pretrianed model and put it under 'ckpt' file.

pip install -r requirements.txt

Our diffusion code structure is based on the original implementation of DDPM. Increasing the size of the U-Net may lead to better results.

About training iteration. The training with 5000 iterations has converged quite well. We recommend training for 10,000 iterations to achieve better performance, and you can select the best-performing training iterations.

We test code on one RTX 3090 GPU. The training time is about 1-2 days.

python train.py   #train from scratch, you can change setting in modelConfig 
python train.py --pretrained_path ckpt/lol.pt  
python test.py --pretrained_path ckpt/lol.pt  

Mask CLE Diffusion finetunes lol checkpoint. In our experiments, lol checkpoint is better than mit-adobe-5K checkpoint.

We show some inference cases in 'data/Mask_CLE_cases'. Welcome to use your cases to test the performance.

python mask_generation.py   #generate masks for training
python train_mask.py --pretrained_path ckpt/lol.pt  #finetune Mask CLE Diffusion
python test_mask.py --pretrained_path ckpt/Mask_CLE.pt --input_path data/Mask_CLE_cases/opera.png --mask_path data/Mask_CLE_cases/opera_mask.png --data_name opera

• LOL dataset outputs

• release training and testing code

• release Mask CLE Diffusion code

• release mit-5k training code and checkpoints

• update lpips and LI-lpips metrics

This work is mainly built on DenoisingDiffusionProbabilityModel-ddpm. Thanks a lot to authors for sharing!

@inproceedings{yin2023cle,
  title={CLE Diffusion: Controllable Light Enhancement Diffusion Model},
  author={Yin, Yuyang and Xu, Dejia and Tan, Chuangchuang and Liu, Ping and Zhao, Yao and Wei, Yunchao},
  booktitle={Proceedings of the 31st ACM International Conference on Multimedia},
  pages={8145--8156},
  year={2023}
}

If you have any problems, please feel free to create a new issue or email me(yuyangyin@bjtu.edu.cn)..