Towards Scale-Aware Low-Light Enhancement via Structure-Guided Transformer Design

This repo is our solution for NTIRE 2025 Low Light Image Enhancement Challenge.

Our paper has been accepted to the CVPR NTIRE Workshop. The arXiv version is available here.

🏆 Final Results

Our solution placed 2nd in the final testing phase of the competition!
This result highlights the effectiveness and robustness of our approach under rigorous evaluation.
Thank you to the organizers and all participants for an exciting and competitive challenge!

Description

Installation, Training, and Testing

Create Environment

1. Create Conda Environment

conda create --name imagine python=3.10
conda activate imagine

2. 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
pip install numpy==1.26.1

3. Install BasicSR

python setup.py develop --no_cuda_ext

Pre-trained Model

• Pre-trained Model for NTIRE 2025 Low Light Image Enhancement Challenge.

Our Submission on Test Sever

• Our Test Output.

Testing

Download above saved models and put it into the folder ./Enhancement/weights. To test the model, you need to specify the input image path (args.input_dir), the input structure prior path(args.input_dir_s), and pre-trained model path (args.weights) in ./Enhancement/test.py. Then run

cd Enhancement
python test.py 

You can check the output in test-results-ntire25.

Contact

If you have any question, please feel free to contact us via [email protected].