FastMesh: Efficient Artistic Mesh Generation via Component Decoupling (3DV 2026)

Jeonghwan Kim, Yushi Lan, Armando Fortes, Yongwei Chen, Xingang Pan

S-lab, Nanyang Technological University

This repository is an official implementation of the paper FastMesh: Efficient Artistic Mesh Generation via Component Decoupling. Our approach efficiently produces 3D objects by substantially reducing the number of tokens required for generation. Here is our Project Page.

🚧 TODO

• Release Inference Code
• Release Model Weights
• Release Training Code
• Release HuggingFace Demo

⚙️ Installation

Our environment has been tested on CUDA 11.8 with A6000.
(It would be greatly appreciated if you report an issue when you find any errors.)

conda create -n fastmesh python=3.10
conda activate fastmesh
pip install torch==2.7.1 torchvision==0.22.1 torchaudio==2.7.1 --index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txt
pip install flash-attn --no-build-isolation

💡 Quick Start

Generate meshes from sampled point cloud with V1K variant:

python inference.py --mesh_path assets --variant V1K --batch_size 3

Generate meshes from sampled point cloud with V4K variant:

python inference.py --mesh_path assets --variant V4K --batch_size 1

😊 Acknowledgement

We thank all the authors who made their code public, which tremendously accelerates our project progress. (MeshAnything, MeshAnythingV2, BPT, Michelangelo, TreeMeshGPT, VAR)

📚 Citation

If you find our work helpful, please consider citing:

@misc{kim2025fastmesh,
      title={FastMesh: Efficient Artistic Mesh Generation via Component Decoupling}, 
      author={Jeonghwan Kim and Yushi Lan and Armando Fortes and Yongwei Chen and Xingang Pan},
      year={2025},
      eprint={2508.19188},
      archivePrefix={arXiv},
      url={https://arxiv.org/abs/2508.19188}, 
}