lalala125/DiscreteSDF

Gaussian Splatting with Discretized SDF for Relightable Assets

This repository contains the official checkpoints for DiscretizedSDF, a novel method presented in the paper Gaussian Splatting with Discretized SDF for Relightable Assets. This work will be presented at ICCV 2025.

DiscretizedSDF significantly advances 3D Gaussian splatting (3DGS) for inverse rendering, enabling the creation of high-quality relightable 3D assets. It addresses common challenges in applying geometry constraints to discrete Gaussian primitives and the high memory costs associated with continuous Signed Distance Fields (SDFs).

Key innovations include:

• Discretized SDF: A novel approach to represent continuous SDFs discretely, encoding sampled values within each Gaussian. This allows for efficient SDF rendering via splatting, bypassing costly ray tracing.
• Projection-based Consistency Loss: A unique regularization technique that projects Gaussians onto the SDF's zero-level set, ensuring geometric alignment with the splatted surface.

Thanks to these innovations, DiscretizedSDF achieves superior relighting quality without additional memory overhead compared to standard 3DGS, and it simplifies the training process by avoiding complex manual optimizations.

Links

• 📄 Paper: Gaussian Splatting with Discretized SDF for Relightable Assets
• 🌐 Project Page: https://nk-cs-zzl.github.io/projects/dsdf/index.html
• 💻 GitHub Repository: https://github.com/NK-CS-ZZL/DiscretizedSDF

Usage

For detailed installation instructions, environment setup, and information on training and evaluation, please refer to the official GitHub repository.

To run a quick relighting video demo with the provided checkpoints:

1. Clone the repository:

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

2. Follow the installation steps on the GitHub repository's "Dependencies and Installation" section to set up the environment and dependencies.
3. Download pretrained models (e.g., from HuggingFace as mentioned in the GitHub README) and place them in the pretrained folder.
4. Run the demo script:

   sh demo.sh
   

Citation

If you find our work 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}
}