GS^3: Efficient Relighting with Triple Gaussian Splatting

Zoubin Bi · Yixin Zeng · Chong Zeng · Fan Pei · Xiang Feng · Kun Zhou · Hongzhi Wu

SIGGRAPH Asia 2024 Conference Papers

Project Page | Paper | arXiv | Data

We present a spatial and angular Gaussian based representation and a triple splatting process, for real-time, high-quality novel lighting-and-view synthesis from multi-view point-lit input images.

Setup

Environment

Conda

bash conda create --name gs3 python=3.10 pytorch==2.4.1 torchvision==0.19.1 pytorch-cuda=12.4 cuda-toolkit=12.4 cuda-cudart=12.4 -c pytorch -c "nvidia/label/cuda-12.4.0" conda activate gs3 pip install ninja # speedup torch cuda extensions compilation pip install -r requirements.txt

Docker

We also provide a docker container.

You can use our pre-built docker image.

bash docker run -it --gpus all --rm iamncj/gs3:241002

Or you can build your own docker image.

bash docker build -t gs3:latest .

Usage

We provide a few sample scripts from our paper.

Train

For real captured scenes, please use --cam_opt and --pl_opt to enable camera pose and light optimization.

bash bash real_train.sh # real captured scenes bash syn_train.sh # synthetic scenes

Test

For real captured scenes, we provide --valid and corresponding .json file to render a circle view. If you are going to run the test set of real captured scenes, please remember to add --opt_pose to use the calibrated poses.

bash bash real_render.sh # real captured scenes bash syn_render.sh # synthetic scenes

Data

We release our data and pretrained models at huggingface.

Note: For blender scene rendering, we use the script from NRHints. For pre-captured scene rendering, we use Asuna.

You can download the data by running the following command:

bash pip install huggingface_hub[hf_transfer] HF_HUB_ENABLE_HF_TRANSFER=1 huggingface-cli download --repo-type dataset gsrelight/gsrelight-data --local-dir /path/to/data

Citation

Cite as below if you find this repository is helpful to your project:

@inproceedings{bi2024rgs, title = {GS\textsuperscript{3}: Efficient Relighting with Triple Gaussian Splatting}, author = {Zoubin Bi and Yixin Zeng and Chong Zeng and Fan Pei and Xiang Feng and Kun Zhou and Hongzhi Wu}, booktitle = {SIGGRAPH Asia 2024 Conference Papers}, year = {2024} }

Some of our dataset are borrowed from NRHints. Please also cite NRHints if you use those data.

Acknowledgments

We have intensively borrow codes from gaussian splatting and gsplat. We also use tiny-cuda-nn for it's efficient MLP implementation. Many thanks to the authors for sharing their codes.