DetailGen3D: Generative 3D Geometry Enhancement via Data-Dependent Flow

🔥 Updates

📅 April 2025

• 🚀 Initial Release: Published code, pretrained models, and interactive demo.

🔨 Installation

Clone the repo:

git clone https://github.com/VAST-AI-Research/DetailGen3D.git
cd DetailGen3D

Create a conda environment (optional):

conda create -n detailgen3d python=3.10
conda activate detailgen3d

Install dependencies:

# pytorch (select correct CUDA version)
pip install torch torchvision --index-url https://download.pytorch.org/whl/{your-cuda-version}

# other dependencies
pip install -r requirements.txt

💡 Quick Start

Step 1: Prepare coarse mesh

Upload a mesh with less detail. We recommend using these 3d generation tools:

• TripoSR
• InstantMesh

Step 2: Generate geometry details:

python scripts/inference_detailgen3d.py \
  --mesh_input assets/model/cb7e6c4a-b4dd-483c-9789-3d4887ee7434.glb \
  --image_input assets/image/cb7e6c4a-b4dd-483c-9789-3d4887ee7434.png

The required model weights will be automatically downloaded:

• DetailGen3D model from VAST-AI/DetailGen3D → pretrained_weights/DetailGen3D

⭐ Acknowledgements

We would like to thank the following open-source projects and research works that made DetailGen3D possible:

• 🤗 Diffusers for their excellent diffusion model framework
• HunyuanDiT for DiT
• FlashVDM for their lightning vecset decoder
• 3DShape2VecSet for 3D shape representation
• TripoSG as our base model

We are grateful to the broader research community for their open exploration and contributions to the field of 3D generation.

📚 Citation

@misc{deng2025detailgen3dgenerative3dgeometry,
      title={DetailGen3D: Generative 3D Geometry Enhancement via Data-Dependent Flow}, 
      author={Ken Deng and Yuan-Chen Guo and Jingxiang Sun and Zi-Xin Zou and Yangguang Li and Xin Cai and Yan-Pei Cao and Yebin Liu and Ding Liang},
      year={2025},
      eprint={2411.16820},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2411.16820}, 
}