[CVPR 2025] LITA-GS: Illumination-Agnostic Novel View Synthesis via Reference-Free 3D Gaussian Splatting and Physical Priors [[Paper]](https://arxiv.org/pdf/2504.00219)

August 27, 2025 · View on GitHub

Han Zhou, Wei Dong^†, Jun Chen

McMaster University, ^†Corresponding Author

Introduction

This repository represents the official implementation of our CVPR 2025 paper titled LITA-GS: Illumination-Agnostic Novel View Synthesis via Reference-Free 3D Gaussian Splatting and Physical Priors. If you find this repo useful, please give it a star ⭐ and consider citing our paper in your research. Thank you for your interest.

📢 News

2025-08-26 The code for underexposed scenes has been released！

Overall Framework

teaser

🛠️ Setup

The code was tested on:

RTX 5090, Python 3.9, CUDA 12.8, PyTorch 2.8 + cu12.8.

📦 Repository

Clone the repository (requires git):

git clone https://github.com/LowLevelAI/LITA-GS.git
cd LITA-GS

💻 Dependencies

Create the Conda environment:

conda create -n litags python=3.9
conda activate litags

Then install dependencies:

Install Pytorch

pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128

Set Cudatoolkit to 12.8

export PATH=/usr/local/cuda-12.8/bin:$PATH
export LD_LIBRARY_PATH=/usr/local/cuda-12.8/lib64:$LD_LIBRARY_PATH

Install dependencies

pip install trimesh tqdm mmcv==1.6.0 scipy scikit-image
pip install submodules/diff-gaussian-rasterization
pip install submodules/simple-knn

🏋️ Training

For underexposed scenes, we recommend using our generated colmap reuslts, where the overall brightness of underexposed images are adjusted to 0.2 or 0.49 before utilizing COLMAP software. Take the bike scene as an example, the training command could be：

python train_underexposed.py -s data/LOM_low_bike_colmap -m output/LOM_low_bike --config arguments/LOM/bike_low.py

The point clouds and testing outputs can be found in output/LOM_low_bike.

✏️ Contributing

Please refer to this instruction.

🎓 Citation

If you find this repo and our paper useful, please consider citing our paper:

@inproceedings{zhou2025litags,
  title={LITA-GS: Illumination-Agnostic Novel View Synthesis via Reference-Free 3D Gaussian Splatting and Physical Priors},
  author={Zhou, Han and Dong, Wei and Chen, Jun},
  booktitle={Proceedings of the Computer Vision and Pattern Recognition Conference},
  pages={21580--21589},
  year={2025}
}

Acknowledgement

This repo is built upon Gaussian DK. We thank the authors for their great job!

Han Zhou, Wei Dong†, Jun Chen

McMaster University, †Corresponding Author

Han Zhou, Wei Dong^†, Jun Chen

McMaster University, ^†Corresponding Author