This is a baseline and toolbox for wide-range low-light image enhancement. This repo supports over 15 benchmarks and extremely high-resolution (up to 4000x6000) low-light enhancement. Our method Retinexformer won the second place in the NTIRE 2024 Challenge on Low Light Enhancement. If you find this repo useful, please give it a star ⭐ and consider citing our paper in your research. Thank you.
- 2024.05.12 : RetinexMamba based on our Retinexformer framework and this repo has been released. The first Mamba work on low-light enhancement. Thanks to the efforts of the authors.
- 2024.03.22 : We release
distributed data parallel (DDP)andmix-precisiontraining strategies to help you train larger models. We releaseself-ensembletesting strategy to help you derive better results. In addition, we also release an adaptivesplit-and-testtesting strategy for high-resolution up to 4000x6000 low-light image enhancement. Feel free to use them. 🚀 - 2024.03.21 : Our methods Retinexformer and MST++ (NTIRE 2022 Spectral Reconstruction Challenge Winner) ranked top-2 in the NTIRE 2024 Challenge on Low Light Enhancement. Code, pre-trained weights, training logs, and enhancement results have been released in this repo. Feel free to use them! 🚀
- 2024.02.15 : NTIRE 2024 Challenge on Low Light Enhancement begins. Welcome to use our Retinexformer or MST++ (NTIRE 2022 Spectral Reconstruction Challenge Winner) to participate in this challenge! 🏆
- 2023.11.03 : The test setting of KinD, LLFlow, and recent diffusion models and the corresponding results on LOL are provided. Please note that we do not suggest this test setting because it uses the mean of the ground truth to obtain better results. But, if you want to follow KinD, LLFlow, and recent diffusion-based works for fair comparison, it is your choice to use this test setting. Please refer to the
Testingpart for details. - 2023.11.02 : Retinexformer is added to the Awesome-Transformer-Attention collection. 💫
- 2023.10.20 : Params and FLOPS evaluating function is provided. Feel free to check and use it.
- 2023.10.12 : Retinexformer is added to the ICCV-2023-paper collection. 🚀
- 2023.10.10 : Retinexformer is added to the low-level-vision-paper-record collection. ⭐
- 2023.10.06 : Retinexformer is added to the awesome-low-light-image-enhancement collection. 🎉
- 2023.09.20 : Some results on ExDark nighttime object detection are released.
- 2023.09.20 : Code, models, results, and training logs have been released. Feel free to use them. ⭐
- 2023.07.14 : Our paper has been accepted by ICCV 2023. Code and Models will be released. 🚀
-
Results on LOL-v1, LOL-v2-real, LOL-v2-synthetic, SID, SMID, SDSD-in, SDSD-out, and MIT Adobe FiveK datasets can be downloaded from Baidu Disk (code:
cyh2) or Google Drive -
Results on LOL-v1, LOL-v2-real, and LOL-v2-synthetic datasets with the same test setting as KinD, LLFlow, and recent diffusion models can be downloaded from Baidu Disk (code:
cyh2) or Google Drive. -
Results on the NTIRE 2024 low-light enhancement dataset can be downloaded from Baidu Disk (code:
cyh2) or Google Drive -
Results on LIME, NPE, MEF, DICM, and VV datasets can be downloaded from Baidu Disk (code:
cyh2) or Google Drive -
Results on ExDark nighttime object detection can be downloaded from Baidu Disk (code:
cyh2) or Google Drive. Please use this repo to run experiments on the ExDark dataset
Performance on LOL-v1, LOL-v2-real, LOL-v2-synthetic, SID, SMID, SDSD-in, and SDSD-out:
Performance on LOL with the same test setting as KinD, LLFlow, and diffusion models:
| Metric | LOL-v1 | LOL-v2-real | LOL-v2-synthetic |
|---|---|---|---|
| PSNR | 27.18 | 27.71 | 29.04 |
| SSIM | 0.850 | 0.856 | 0.939 |
Please note that we do not suggest this test setting because it uses the mean of the ground truth to obtain better results. But, if you want to follow KinD, LLFlow, and recent diffusion-based works, it is your choice to use this test setting. Please refer to the Testing part for details.
Performance on NTIRE 2024 test-challenge:
| Method | Retinexformer | MST++ | Ensemble |
|---|---|---|---|
| PSNR | 24.61 | 24.59 | 25.30 |
| SSIM | 0.85 | 0.85 | 0.85 |
Feel free to check the Codalab leaderboard. Our method ranks second.
Performance on MIT Adobe FiveK:
Performance on LIME, NPE, MEF, DICM, and VV:
Performance on ExDark Nighttime object detection:
| NTIRE - dev - 2000x3000 | NTIRE - challenge - 4000x6000 |
|---|---|
 |
 |
We suggest you use pytorch 1.11 to re-implement the results in our ICCV 2023 paper and pytorch 2 to re-implement the results in NTIRE 2024 Challenge because pytorch 2 can save more memory in mix-precision training.
- Make Conda Environment
conda create -n Retinexformer python=3.7
conda activate Retinexformer
- Install Dependencies
conda install pytorch=1.11 torchvision cudatoolkit=11.3 -c pytorch
pip install matplotlib scikit-learn scikit-image opencv-python yacs joblib natsort h5py tqdm tensorboard
pip install einops gdown addict future lmdb numpy pyyaml requests scipy yapf lpips
- Install BasicSR
python setup.py develop --no_cuda_ext
- Make Conda Environment
conda create -n torch2 python=3.9 -y
conda activate torch2
- Install Dependencies
conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia
pip install matplotlib scikit-learn scikit-image opencv-python yacs joblib natsort h5py tqdm tensorboard
pip install einops gdown addict future lmdb numpy pyyaml requests scipy yapf lpips thop timm
- Install BasicSR
python setup.py develop --no_cuda_ext
Download the following datasets:
LOL-v1 Baidu Disk (code: cyh2), Google Drive
LOL-v2 Baidu Disk (code: cyh2), Google Drive
SID Baidu Disk (code: gplv), Google Drive
SMID Baidu Disk (code: btux), Google Drive
SDSD-indoor Baidu Disk (code: jo1v), Google Drive
SDSD-outdoor Baidu Disk (code: uibk), Google Drive
MIT-Adobe FiveK Baidu Disk (code:cyh2), Google Drive, Official
NTIRE 2024 Baidu Disk (code:cyh2), Google Drive links for training input, training GT, and mini-val set.
Note:
(1) Please use bandizip to jointly unzip the .zip and .z01 files of SMID, SDSD-indoor, and SDSD-outdoor
(2) Please process the raw images of the MIT Adobe FiveK dataset following the sRGB output mode or directly download and use the sRGB image pairs processed by us in the Baidu Disk (code:cyh2) and Google Drive
(3) Please download the text_list.txt from Google Drive or Baidu Disk (code: ggbh) and then put it into the folder data/SMID/SMID_Long_np/
Then organize these datasets as follows:
|--data
| |--LOLv1
| | |--Train
| | | |--input
| | | | |--100.png
| | | | |--101.png
| | | | ...
| | | |--target
| | | | |--100.png
| | | | |--101.png
| | | | ...
| | |--Test
| | | |--input
| | | | |--111.png
| | | | |--146.png
| | | | ...
| | | |--target
| | | | |--111.png
| | | | |--146.png
| | | | ...
| |--LOLv2
| | |--Real_captured
| | | |--Train
| | | | |--Low
| | | | | |--00001.png
| | | | | |--00002.png
| | | | | ...
| | | | |--Normal
| | | | | |--00001.png
| | | | | |--00002.png
| | | | | ...
| | | |--Test
| | | | |--Low
| | | | | |--00690.png
| | | | | |--00691.png
| | | | | ...
| | | | |--Normal
| | | | | |--00690.png
| | | | | |--00691.png
| | | | | ...
| | |--Synthetic
| | | |--Train
| | | | |--Low
| | | | | |--r000da54ft.png
| | | | | |--r02e1abe2t.png
| | | | | ...
| | | | |--Normal
| | | | | |--r000da54ft.png
| | | | | |--r02e1abe2t.png
| | | | | ...
| | | |--Test
| | | | |--Low
| | | | | |--r00816405t.png
| | | | | |--r02189767t.png
| | | | | ...
| | | | |--Normal
| | | | | |--r00816405t.png
| | | | | |--r02189767t.png
| | | | | ...
| |--SDSD
| | |--indoor_static_np
| | | |--input
| | | | |--pair1
| | | | | |--0001.npy
| | | | | |--0002.npy
| | | | | ...
| | | | |--pair2
| | | | | |--0001.npy
| | | | | |--0002.npy
| | | | | ...
| | | | ...
| | | |--GT
| | | | |--pair1
| | | | | |--0001.npy
| | | | | |--0002.npy
| | | | | ...
| | | | |--pair2
| | | | | |--0001.npy
| | | | | |--0002.npy
| | | | | ...
| | | | ...
| | |--outdoor_static_np
| | | |--input
| | | | |--MVI_0898
| | | | | |--0001.npy
| | | | | |--0002.npy
| | | | | ...
| | | | |--MVI_0918
| | | | | |--0001.npy
| | | | | |--0002.npy
| | | | | ...
| | | | ...
| | | |--GT
| | | | |--MVI_0898
| | | | | |--0001.npy
| | | | | |--0002.npy
| | | | | ...
| | | | |--MVI_0918
| | | | | |--0001.npy
| | | | | |--0002.npy
| | | | | ...
| | | | ...
| |--SID
| | |--short_sid2
| | | |--00001
| | | | |--00001_00_0.04s.npy
| | | | |--00001_00_0.1s.npy
| | | | |--00001_01_0.04s.npy
| | | | |--00001_01_0.1s.npy
| | | | ...
| | | |--00002
| | | | |--00002_00_0.04s.npy
| | | | |--00002_00_0.1s.npy
| | | | |--00002_01_0.04s.npy
| | | | |--00002_01_0.1s.npy
| | | | ...
| | | ...
| | |--long_sid2
| | | |--00001
| | | | |--00001_00_0.04s.npy
| | | | |--00001_00_0.1s.npy
| | | | |--00001_01_0.04s.npy
| | | | |--00001_01_0.1s.npy
| | | | ...
| | | |--00002
| | | | |--00002_00_0.04s.npy
| | | | |--00002_00_0.1s.npy
| | | | |--00002_01_0.04s.npy
| | | | |--00002_01_0.1s.npy
| | | | ...
| | | ...
| |--SMID
| | |--SMID_LQ_np
| | | |--0001
| | | | |--0001.npy
| | | | |--0002.npy
| | | | ...
| | | |--0002
| | | | |--0001.npy
| | | | |--0002.npy
| | | | ...
| | | ...
| | |--SMID_Long_np
| | | |--text_list.txt
| | | |--0001
| | | | |--0001.npy
| | | | |--0002.npy
| | | | ...
| | | |--0002
| | | | |--0001.npy
| | | | |--0002.npy
| | | | ...
| | | ...
| |--FiveK
| | |--train
| | | |--input
| | | | |--a0099-kme_264.jpg
| | | | |--a0101-kme_610.jpg
| | | | ...
| | | |--target
| | | | |--a0099-kme_264.jpg
| | | | |--a0101-kme_610.jpg
| | | | ...
| | |--test
| | | |--input
| | | | |--a4574-DSC_0038.jpg
| | | | |--a4576-DSC_0217.jpg
| | | | ...
| | | |--target
| | | | |--a4574-DSC_0038.jpg
| | | | |--a4576-DSC_0217.jpg
| | | | ...
| |--NTIRE
| | |--train
| | | |--input
| | | | |--1.png
| | | | |--3.png
| | | | ...
| | | |--target
| | | | |--1.png
| | | | |--3.png
| | | | ...
| | |--minival
| | | |--input
| | | | |--1.png
| | | | |--31.png
| | | | ...
| | | |--target
| | | | |--1.png
| | | | |--31.png
| | | | ...
We also provide download links for LIME, NPE, MEF, DICM, and VV datasets that have no ground truth:
Baidu Disk (code: cyh2)
or Google Drive
Download our models from Baidu Disk (code: cyh2) or Google Drive. Put them in folder pretrained_weights
# activate the environment
conda activate Retinexformer
# LOL-v1
python3 Enhancement/test_from_dataset.py --opt Options/RetinexFormer_LOL_v1.yml --weights pretrained_weights/LOL_v1.pth --dataset LOL_v1
# LOL-v2-real
python3 Enhancement/test_from_dataset.py --opt Options/RetinexFormer_LOL_v2_real.yml --weights pretrained_weights/LOL_v2_real.pth --dataset LOL_v2_real
# LOL-v2-synthetic
python3 Enhancement/test_from_dataset.py --opt Options/RetinexFormer_LOL_v2_synthetic.yml --weights pretrained_weights/LOL_v2_synthetic.pth --dataset LOL_v2_synthetic
# SID
python3 Enhancement/test_from_dataset.py --opt Options/RetinexFormer_SID.yml --weights pretrained_weights/SID.pth --dataset SID
# SMID
python3 Enhancement/test_from_dataset.py --opt Options/RetinexFormer_SMID.yml --weights pretrained_weights/SMID.pth --dataset SMID
# SDSD-indoor
python3 Enhancement/test_from_dataset.py --opt Options/RetinexFormer_SDSD_indoor.yml --weights pretrained_weights/SDSD_indoor.pth --dataset SDSD_indoor
# SDSD-outdoor
python3 Enhancement/test_from_dataset.py --opt Options/RetinexFormer_SDSD_outdoor.yml --weights pretrained_weights/SDSD_outdoor.pth --dataset SDSD_outdoor
# FiveK
python3 Enhancement/test_from_dataset.py --opt Options/RetinexFormer_FiveK.yml --weights pretrained_weights/FiveK.pth --dataset FiveK
# NTIRE
python3 Enhancement/test_from_dataset.py --opt Options/RetinexFormer_NTIRE.yml --weights pretrained_weights/NTIRE.pth --dataset NTIRE --self_ensemble
# MST_Plus_Plus trained with 4 GPUs on NTIRE
python3 Enhancement/test_from_dataset.py --opt Options/MST_Plus_Plus_NTIRE_4x1800.yml --weights pretrained_weights/MST_Plus_Plus_4x1800.pth --dataset NTIRE --self_ensemble
# MST_Plus_Plus trained with 8 GPUs on NTIRE
python3 Enhancement/test_from_dataset.py --opt Options/MST_Plus_Plus_NTIRE_8x1150.yml --weights pretrained_weights/MST_Plus_Plus_8x1150.pth --dataset NTIRE --self_ensemble
We add the self-ensemble strategy in the testing code to derive better results. Just add a --self_ensemble action at the end of the above test command to use it.
We provide the same test setting as LLFlow, KinD, and recent diffusion models. Please note that we do not suggest this test setting because it uses the mean of ground truth to enhance the output of the model. But if you want to follow this test setting, just add a --GT_mean action at the end of the above test command as
# LOL-v1
python3 Enhancement/test_from_dataset.py --opt Options/RetinexFormer_LOL_v1.yml --weights pretrained_weights/LOL_v1.pth --dataset LOL_v1 --GT_mean
# LOL-v2-real
python3 Enhancement/test_from_dataset.py --opt Options/RetinexFormer_LOL_v2_real.yml --weights pretrained_weights/LOL_v2_real.pth --dataset LOL_v2_real --GT_mean
# LOL-v2-synthetic
python3 Enhancement/test_from_dataset.py --opt Options/RetinexFormer_LOL_v2_synthetic.yml --weights pretrained_weights/LOL_v2_synthetic.pth --dataset LOL_v2_synthetic --GT_meanWe have provided a function my_summary() in Enhancement/utils.py, please use this function to evaluate the parameters and computational complexity of the models, especially the Transformers as
from utils import my_summary
my_summary(RetinexFormer(), 256, 256, 3, 1)
Feel free to check our training logs from Baidu Disk (code: cyh2) or Google Drive
We suggest you use the environment with pytorch 2 to train our model on the NTIRE 2024 dataset and the environment with pytorch 1.11 to train our model on other datasets.
# activate the enviroment
conda activate Retinexformer
# LOL-v1
python3 basicsr/train.py --opt Options/RetinexFormer_LOL_v1.yml
# LOL-v2-real
python3 basicsr/train.py --opt Options/RetinexFormer_LOL_v2_real.yml
# LOL-v2-synthetic
python3 basicsr/train.py --opt Options/RetinexFormer_LOL_v2_synthetic.yml
# SID
python3 basicsr/train.py --opt Options/RetinexFormer_SID.yml
# SMID
python3 basicsr/train.py --opt Options/RetinexFormer_SMID.yml
# SDSD-indoor
python3 basicsr/train.py --opt Options/RetinexFormer_SDSD_indoor.yml
# SDSD-outdoor
python3 basicsr/train.py --opt Options/RetinexFormer_SDSD_outdoor.yml
# FiveK
python3 basicsr/train.py --opt Options/RetinexFormer_FiveK.ymlTrain our Retinexformer and MST++ with the distributed data parallel (DDP) strategy of pytorch on the NTIRE 2024 Low-Light Enhancement dataset. Please note that we use the mix-precision strategy in the training process, which is controlled by the bool hyperparameter use_amp in the config file.
# activate the enviroment
conda activate torch2
# Train Retinexformer with 8 GPUs on NTIRE
bash train_multigpu.sh Options/RetinexFormer_NTIRE_8x2000.yml 0,1,2,3,4,5,6,7 4321
# Train MST++ with 4 GPUs on NTIRE
bash train_multigpu.sh Options/RetinexFormer_NTIRE_4x1800.yml 0,1,2,3,4,5,6,7 4329
# Train MST++ with 8 GPUs on NTIRE
bash train_multigpu.sh Options/MST_Plus_Plus_NTIRE_8x1150.yml 0,1,2,3,4,5,6,7 4343
@InProceedings{Cai_2023_ICCV,
author = {Cai, Yuanhao and Bian, Hao and Lin, Jing and Wang, Haoqian and Timofte, Radu and Zhang, Yulun},
title = {Retinexformer: One-stage Retinex-based Transformer for Low-light Image Enhancement},
booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
month = {October},
year = {2023},
pages = {12504-12513}
}
@inproceedings{retinexformer,
title={Retinexformer: One-stage Retinex-based Transformer for Low-light Image Enhancement},
author={Yuanhao Cai and Hao Bian and Jing Lin and Haoqian Wang and Radu Timofte and Yulun Zhang},
booktitle={ICCV},
year={2023}
}
# MST++
@inproceedings{mst,
title={Mask-guided Spectral-wise Transformer for Efficient Hyperspectral Image Reconstruction},
author={Yuanhao Cai and Jing Lin and Xiaowan Hu and Haoqian Wang and Xin Yuan and Yulun Zhang and Radu Timofte and Luc Van Gool},
booktitle={CVPR},
year={2022}
}