DeepMesh-v2: Auto-Regressive Artist-Mesh Creation
With Reinforcement Learning

Junliang Ye^1*, Ruowen Zhao^1*, Zhengyi Wang^1*, Yikai Wang¹, Jun Zhu^1,2†
*Equal Contribution.
^†Corresponding authors.
¹Tsinghua University, ²ShengShu,

                   

All of the meshes above are generated by DeepMesh-v2. DeepMesh can generate high-quality meshes conditioned on the given point cloud by auto-regressive transformer.

Contents

• Result
• Important Notes
• Acknowledgement
• BibTeX

Results

Table 1: Comparison of different methods on artist-mesh test dataset

Metric	BPT	TreeMeshGPT	DeepMesh (1B)	DeepMeshv2 (2B)
Success Rate	51%	38%	74%	95%
Average Faces	1920	6620	12235	8898
HD	0.29744	0.45681	0.17975	0.14191
CD	0.15089	0.23764	0.08474	0.07574

Table 2: Comparison of different methods on dense mesh test dataset

Metric	BPT	TreeMeshGPT	DeepMesh (1B)	DeepMeshv2 (2B)
Success Rate	45%	39%	69%	90%
Average Faces	2257	7571	15328	9063
HD	0.32891	0.44925	0.24112	0.16171
CD	0.29743	0.25639	0.14295	0.08184

Important Notes

• Please refer to our project_page for more examples.

Acknowledgement

Our code is based on these wonderful repos:

• BPT
• LLaMA-Mesh
• Michelangelo

BibTeX

@article{zhao2025deepmesh,
  title={DeepMesh: Auto-Regressive Artist-mesh Creation with Reinforcement Learning},
  author={Zhao, Ruowen and Ye, Junliang and Wang, Zhengyi and Liu, Guangce and Chen, Yiwen and Wang, Yikai and Zhu, Jun},
  journal={arXiv preprint arXiv:2503.15265},
  year={2025}
}