Gaussian Splatting with Discretized SDF for Relightable Assets

This repository contains the official implementation of the following paper:

Gaussian Splatting with Discretized SDF for Relightable Assets
Zuo-Liang Zhu¹, Jian Yang¹, Beibei Wang^2
1Nankai University ²Nanjing University
In ICCV 2025

[Paper] [Project Page] [Video]

DiscretizedSDF is an efficient, robust solution for object relighting, aiming to produce decent decompositions of geometry, material, and lighting for multi-view observations.

News

• Jul. 21, 2025: Our code is publicly available.
• Jul. 22, 2025: Our paper is publicly available on ArXiv.
• Jul. 22, 2025: Release pretrained models.

Method Overview

For more technical details, please refer to our paper on arXiv.

Dependencies and Installation

1. Clone repo.

   git clone https://github.com/NK-CS-ZZL/DiscretizedSDF.git
   cd DiscretizedSDF

2. Create Conda environment and install dependencies

   conda create -n dsdf python=3.10
   conda activate dsdf
   pip install torch==2.2.1 torchvision==0.17.1 --index-url https://download.pytorch.org/whl/cu118
   pip install -r requirements.txt
   git clone https://github.com/NVlabs/nvdiffrast
   pip install ./nvdiffrast
   pip install ./submodules/fused-ssim
   pip install ./submodules/diff-surfel-sdf-rasterization
   pip install ./submodules/simple-knn

   Note that

   • Our code is verfied under CUDA11.8 runtime, so we recommend to use the same environment to guarantee reproductibility.
   • Please switch to the corresponding runtime if the NVCC version is higher than 11.8.

3. Download pretrained models for demos from Download and place them to the pretrained folder

Quick Demo

We provide a demo checkpoint and a environment map in the demo folder. You can simply run sh demo.sh to creating a relighting video demo in 3 minutes.

Download

Dataset	Training Bash	🔗 Source	🔗 Checkpoint	🔗 Result
Glossy Synthetic	train_glossy.sh	Images	Google Driver	Google Driver
Shiny Blender	train_shiny.sh	Images / Point Cloud	Google Driver	Google Driver
TensoIR Synthetic	train_tir.sh	Images / Env. maps	Google Driver	Google Driver

Update: Now you can also download our checkpoints from HuggingFace.

Training and Evaluation

Please refer to develop.md to learn how to benchmark the DiscretizedSDF and how to train yourself DiscretizedSDF model from the scratch.

Citation

If you find our repo useful for your research, please consider citing our paper:

@inproceedings{zhu_2025_dsdf,
   title={Gaussian Splatting with Discretized SDF for Relightable Assets},
   author={Zhu, Zuo-Liang and Yang, Jian and Wang, Beibei},
   booktitle={Proceedings of IEEE International Conference on Computer Vision (ICCV)},
   year={2025}
}

License

This code is licensed under the Creative Commons Attribution-NonCommercial 4.0 International for non-commercial use only. Please note that any commercial use of this code requires formal permission prior to use.

Contact

For technical questions, please contact nkuzhuzl[AT]gmail.com.

For commercial licensing, please contact beibei.wang[AT]nju.edu.cn。

Acknowledgement

We thank Zixiong Wang for his suggestions during the project.

Here are some great resources we benefit from: GaussianShader, 2DGS, NeRO, TensoSDF, Ref-NeuS

If you develop/use DiscretizedSDF in your projects, welcome to let us know. We will list your projects in this repository.