- Jun 17, 2025: Our work has been submitted and is under review.
- Jun 19, 2025: We have released the results of compared methods and testing code.
- Sep 3, 2025: Our work has been accepted by TIM 2025.
- Sep 28, 2025: Datasets, codes and pretrained weights are available.
Simplified abstract: This paper has proposed a staged framework for low-light image enhancement (LLIE) featuring a low-cost offline control strategy with natural transition effects. Extensive comparisons and ablations are conducted to demonstrate the efficiency of its designs. Eight benchmarks are included for comparisons, including LOL-v1-real, LSRW-huawei, LSRW-nikon, LIME, DICM, MEF, NPE, and VV. For the first three paired datasets, PSNR, SSIM, and LPIPS scores are measured. For the last five unpaired datasets, we employ three reference-free metrics, i.e. NIQE, ILNIQE, and BRISQUE.
More details can be found in the configuration files from the envs folder.
python==3.8.20
torch==1.11.0+cu113
torchvision==0.12.0+cu113
visdom==0.2.4
...
Pretrained weights have been already included in the checkpoints folder. Other resouces are all shared in: 百度网盘 or Google Drive (not ready)
More details can be found in readme.txt from the scripts folder
Our work is built upon the codebase of EnlightenGAN and PairLIE, and we sincerely thank them for their contribution.