EdgeRunner: Auto-regressive Auto-encoder for Artistic Mesh Generation

TL;DR

EdgeRunner introduces an auto-regressive auto-encoder (ArAE) with a novel EdgeBreaker-based mesh tokenizer to generate artistic meshes up to 4,000 faces at a resolution of $512^3$. By mapping variable-length meshes into a fixed-length latent space, it enables latent diffusion conditioned on point clouds or single-view images, improving generalization and cross-modal generation. The approach achieves higher quality and diversity than prior autoregressive mesh methods, with efficient training and competitive inference times. This work advances scalable, topology-preserving 3D mesh generation for downstream creative applications.

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 $512^3$. 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.

Figures (15)

• Figure 1: EdgeRunner efficiently generates diverse, high-quality artistic meshes conditioned on point clouds or single-view images.
• Figure 2: Pipeline of our method. Our ArAE model compresses variable-length mesh into fixed-length latent code, which can be further used to train latent diffusion models conditioned on other input modalities, such as single-view images.
• Figure 3: Illustration of our mesh tokenizer. Our tokenizer traverses the 3D mesh triangle-by-triangle and converts it into a 1D token sequence. Through edge sharing, we reach a compression rate of 50% (4 or 5 tokens per face on average) compared to naïve tokenization of 9 tokens per face.
• Figure 4: Half-edge representation for triangular faces.
• Figure 5: Comparison on point cloud conditioned generation. We show the reference dense mesh and generated meshes conditioned on randomly sampled point cloud.