DeMapGS: Simultaneous Mesh Deformation and Surface Attribute

Introduction

DeMapGS is a structured Gaussian Splatting framework that jointly optimizes deformable surfaces and surface-attached 2D Gaussian splats. The unified representation in our method supports extraction of high-fidelity diffuse, normal, and displacement maps.

Hardware Requirements

• One CUDA-ready GPU. We have tested on RTX4090, L4, A100.
• Minimal VRAM 24 GB.

Installation

0. Clone this repo

   git clone --recursive https://github.com/CyberAgentAILab/DeMapGS.git

Using Docker

0. Have docker and nvidia-container-toolkit installed.

1. Build Docker image and set up environment (PyTorch3D installation may take ~30 minutes)

   sudo make setup

2. Get inside the docker container. Change DATA_PATH in Makefile

   sudo make run

Manual Setup

Check docs/ManualSetup.md

Data

1. Download Blender scenes from link. Unzip them and put them under rootpath

   rootpath
       ├── buddha
       │   ├── test/
       │   ├── train/
       │   ├── model.obj
       │   ├── template.obj
       │   └── transforms_train.json
       └── ...

2. ActorHQ data see include/modified_smplx

Example to run

Without Docker

1. Update rootpath in example_run.sh.

2. Run:

   bash example_run.sh

3. Results appear in output-blender/buddha.

Inside Docker

1. Run:

   xvfb-run -a bash example_run.sh

2. Results appear in output-blender/buddha.

License

Copyright (c) 2025 CyberAgent AI Lab

This project is licensed under CC BY-NC 4.0. This project builds upon prior work on 3D Gaussian Splatting 3DGS and 2D Gaussian Splatting 2DGS, which are licensed under the Gaussian Splatting License. The original license text is included in the licenses directory.

SMPL-X License Required

If using the ActorHQ dataset features, you must obtain a separate SMPL-X license from Max Planck Institute at their website

Citation

If you find this code useful for your research, please cite our paper:

@inproceedings{zhou2025demapgs,
    title={DeMapGS: Simultaneous Mesh Deformation and Surface Attribute Mapping via Gaussian Splatting},
    author={Zhou, Shuyi and Zhong, Shengze and Takayama, Kenshi and Taketomi, Takafumi and Oishi, Takeshi},
    booktitle={ACM SIGGRAPH 2025 conference papers},
    year={2025}
}