LuSh-NeRF

[NeurIPS 2024] The official implementation of the Paper: "LuSh-NeRF: Lighting up and Sharpening NeRFs for Low-light Scenes"

⚙️ Setup

conda create -n lushnerf
conda activate lushnerf
git clone https://github.com/quzefan/LuSh-NeRF
cd LuSh-NeRF
pip install -r requirements.txt

Please get the GIM pretrained model gim_dkm_100h.ckpt and put it in the ./gim/weights folder.

🗂️ Data Preparing

For our data:

You can find our data at here. Please put the LOL-BlurNeRF folder in the ./data folder.

The images, images_1_preprocess, HL (Synthetic Only) folder in each scene contain the original, preprocess and GT images.

For your own data:

Please make sure your data is in LLFF format and put it in the ./data folder.

🚝 Training

For example, to train Poster scene,

python run_lushnerf.py --config ./configs/poster_lushnerf 

The training and tensorboard results will be save in <basedir>/<expname>.

When you training on your own data, please modify the scaleup-gamma and the scaleup-clahe hyper-parameters to make sure the images in the images_1_preprocess have a satisfactory color distribution.

🖌️ Rendering

For example, to render Poster scene,

python run_lushnerf.py --config ./configs/poster_lushnerf --render_only --render_radius_scale 2.0

Note

Training Speed

Since our method need to optimize mutiple MLP network, it's training speed may be a bit slow. If your GPU memory is not sufficient or need a quick test, you can lower the N_samples and N_importance (64 to 32).

For the other notes, please refer to the Deblur-NeRF.

Citation

If you find our work is helpful to your research, please cite the papers as follows:

@article{qu2024lush,
  title={LuSh-NeRF: Lighting up and Sharpening NeRFs for Low-light Scenes},
  author={Qu, Zefan and Xu, Ke and Hancke, Gerhard Petrus and Lau, Rynson WH},
  journal={arXiv preprint arXiv:2411.06757},
  year={2024}
}

Acknowledgements

Our codebase builds on Deblur-NeRF, DP-NeRF and GIM. Thanks the authors for sharing their awesome codebases!