Faster and Better Autoregressive Mesh Synthesis via Structured Speculation

1South China University of Technology, 2Tsinghua University, 3Zhejiang University, 4Tencent VISVISE, 5Peking University, 6Singapore Management University
*Equal contribution,
Corresponding author
FlashMesh Teaser Image

Abstract

Autoregressive models can generate high-quality 3D meshes by sequentially producing vertices and faces, but their token-by-token decoding results in slow inference, limiting practical use in interactive and large-scale applications. We present FlashMesh, a fast and high-fidelity mesh generation framework that rethinks autoregressive decoding through a predict-correct-verify paradigm. The key insight is that mesh tokens exhibit strong structural and geometric correlations that enable confident multi-token speculation. FlashMesh leverages this by introducing a speculative decoding scheme tailored to the commonly used hourglass transformer architecture, enabling parallel prediction across face, point, and coordinate levels. Extensive experiments show that FlashMesh achieves up to a 2x speedup⚡ over standard autoregressive models while also improving generation fidelity👍. Our results demonstrate that structural priors in mesh data can be systematically harnessed to accelerate and enhance autoregressive generation.

Architecture

FlashMesh Architecture Image

⚡2x Speed up

FlashMesh Demo Case 1

Meshes generated by FlashMesh

Point Cloud

Drag to rotate • Scroll to zoom

input

Mesh

Drag to rotate • Scroll to zoom

face count: 9307

Point Cloud

Drag to rotate • Scroll to zoom

input

Mesh

Drag to rotate • Scroll to zoom

face count: 8440

Point Cloud

Drag to rotate • Scroll to zoom

input

Mesh

Drag to rotate • Scroll to zoom

face count: 8662

Point Cloud

Drag to rotate • Scroll to zoom

input

Mesh

Drag to rotate • Scroll to zoom

face count: 8029

BibTeX

@misc{shen2025flashmeshfasterbetterautoregressive,
      title={FlashMesh: Faster and Better Autoregressive Mesh Synthesis via Structured Speculation}, 
      author={Tingrui Shen and Yiheng Zhang and Chen Tang and Chuan Ping and Zixing Zhao and Le Wan and Yuwang Wang and Ronggang Wang and Shengfeng He},
      year={2025},
      eprint={2511.15618},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2511.15618}, 
    }