Nautilus: Locality-aware Autoencoder for Scalable Mesh Generation

Yuxuan Wang^*1, Xuanyu Yi^*1, Haohan Weng^*2, Qingshan Xu¹, Xiaokang Wei³,
Xianghui Yang², Chunchao Guo², Long Chen⁴, Hanwang Zhang^1
*Equal contribution
¹Nanyang Technological University, ²Tencent Hunyuan,
³The Hong Kong Polytechnic University, ⁴Hong Kong University of Science and Technology,

Preparation

Trained Model

We plan to release the checkpoints upon the acceptance of the paper.

Due to company confidentiality policies, we are unable to release the model trained on the full dataset with the 1024-dimension Michelangelo encoder. Instead, we provide a version trained with the 256-dimension Michelangelo encoder and a filtered dataset solely based on Objaverse. The performance of this model is moderately lower compared to the full version.

Installation

Install the packages in requirements.txt. The code is tested under CUDA version 11.8.

# clone the repository
git clone https://github.com/Yuxuan-W/nautilus.git
cd nautilus
# create a new conda environment
conda create -n nautilus python=3.9 -y
conda activate nautilus
# install pytorch, we use cuda 11.8
conda install pytorch==2.4.0 torchvision==0.19.0 torchaudio==2.4.0  pytorch-cuda=11.8 -c pytorch -c nvidia -y
# install other dependencies
pip install -r requirements.txt
# install torch-scatter and torch-cluster based on your cuda version
pip install torch-scatter torch-cluster -f https://data.pyg.org/whl/torch-2.4.0+cu118.html

Inference

The generation inference will be available upon the release of our checkpoints. Generating a 5000-face mesh asset typically takes around 3 to 4 minutes on a single A100 GPU.

bash infer.sh /your/path/to/checkpoint /your/path/to/pointcloud

Citation

If you find our work helps, please cite our paper:

@article{wang2025nautilus,
  title={Nautilus: Locality-aware Autoencoder for Scalable Mesh Generation},
  author={Wang, Yuxuan and Yi, Xuanyu and Weng, Haohan and Xu, Qingshan and Wei, Xiaokang and Yang, Xianghui and Guo, Chunchao and Chen, Long and Zhang, Hanwang},
  journal={arXiv preprint arXiv:2501.14317},
  year={2025}
}

Acknowledgement

Our code is based on the wonderful repo of meshgpt-pytorch. We thank the authors for their great work.