Abstract

Current auto-regressive mesh generation methods suffer from issues such as incompleteness, insufficient detail, and poor generalization. In this paper, we propose an Auto-regressive Auto-encoder (ArAE) model capable of generating high-quality 3D meshes with up to 4,000 faces at a spatial resolution of 5123. We introduce a novel mesh tokenization algorithm that efficiently compresses triangular meshes into 1D token sequences, significantly enhancing training efficiency. Furthermore, our model compresses variable-length triangular meshes into a fixed-length latent space, enabling training latent diffusion models for better generalization. Extensive experiments demonstrate the superior quality, diversity, and generalization capabilities of our model in both point cloud and image-conditioned mesh generation tasks.

Pipeline

EdgeRunner pipeline overview

Mesh Tokenizer

Tokenization

Our tokenizer uses a modified EdgeBreaker algorithm to traverse and sequentialize a triangular mesh. The colors of triangles denote different face tokens: L (visit left), R (visit right), E (end of sequence).

De-tokenization

We can de-tokenize our generated mesh token sequences to triangular meshes. Here are some visualizations of the generation progress:

Conditioned Mesh Generation

PointCloud-Conditioned

Drag with left mouse button to move the camera, right mouse button to move the object.

More Results

1 2 3 4 5 6 7 8

Image-Conditioned

Drag with left mouse button to move the camera, right mouse button to move the object.

More Results

1 2 3 4

Diversity of Generation

Given the same input condition (point cloud) but different random seeds, our model can generate diverse meshes:

Diversity of generation results

Click the image to view meshes in 3D.

Citation

@inproceedings{tang2025edgerunner,
  title={EdgeRunner: Auto-regressive Auto-encoder for Artistic Mesh Generation},
  author={Tang, Jiaxiang and Li, Zhaoshuo and Hao, Zekun and Liu, Xian and Zeng, Gang and Liu, Ming-Yu and Zhang, Qinsheng},
  booktitle={International Conference on Learning Representations (ICLR)},
  year={2025}
}