TetraDiffusion: Tetrahedral Diffusion Models for 3D Shape Generation

Nikolai Kalischek^*,1, Torben Peters^*,1, Jan Dirk Wegner²,Konrad Schindler¹

¹ ETH Zürich ² University of Zürich

^* equal contribution

TetraDiffusion is a 3D denoising diffusion model that operates on a tetrahedral grid to enable the generation of high-resolution 3D shapes in seconds. All depicted meshes are shown without any postprocessing, hole-filling or smoothing.

🛠️ Setup

To set up the environment, follow these steps:

1. Install Conda from Anaconda website.

2. Clone the repository and change to directory.

   git clone [email protected]:PeterTor/TetraDiffusion.git
   cd TetraDiffusion/
   

3. Create the environment using environment.yml.

   conda env create -f environment.yml
   conda activate TetraDiffusion
   

🚀 Inference

Pre-trained models for inference will be available in a future update. To run inference:

1. Place weights in folder ./results/{class_resolution}/
2. Run inference, e.g.

   python inference.py --config_path="results/bike_128/"
   

⚙️ Dataset (generation)

We will provide processed datasets and scripts for generating tetrahedralized meshes (optimized within the tetrahedral grid) in a future update. The dataset will be available for download. To create your own dataset, follow the instructions in the preprocessing folder.

🏋️ Training

Citation

@article{kalischek2023tetradiffusion,
  title={TetraDiffusion: Tetrahedral Diffusion Models for 3D Shape Generation},
  author={Kalischek, Nikolai and Peters, Torben and Wegner, Jan D and Schindler, Konrad},
  journal={arXiv preprint arXiv:2211.13220v3},
  year={2023}
}

Acknowledgment

Parts of this repository are adpated from nvdiffrec.