FlashMesh

<h1 align="center" style="line-height:1.3; margin-bottom:0.6rem;"> <!-- FlashMesh 标题图片 --> <img src="./assets/flashmesh_title.png" alt="FlashMesh" style="display:block; margin:0 auto 0.4rem auto; max-width:100%;"> <!-- 副标题正常文字，单独一行 --> <span style="font-size:1.8rem; font-weight:600;"> Faster and Better Autoregressive Mesh Synthesis via Structured Speculation </span> </h1> <h4 align="center" style="line-height:1.4; margin-top:0.6rem"> <a href="https://openreview.net/profile?id=~Tingrui_Shen1">Tingrui Shen</a>^1*, <a href="https://github.com/Graphic-Kiliani">Yiheng Zhang</a>^2*, <a href="https://openreview.net/profile?id=~Chen_Tang1">Chen Tang</a>^1*, <a href="https://openreview.net/profile?id=~Chuan_Ping1">Chuan Ping</a>³, <a href="https://openreview.net/profile?id=~Zixing_Zhao1">Zixing Zhao</a>⁴, <a href="https://openreview.net/profile?id=~Le_Wan1">Le Wan</a>⁴, <a href="https://openreview.net/profile?id=~Yuwang_Wang1">Yuwang Wang</a>², <a href="https://openreview.net/profile?id=~Ronggang_Wang1">Ronggang Wang</a>⁵, <a href="https://openreview.net/profile?id=~Shengfeng_He1">Shengfeng He</a>^6† </h4> <p align="center" style="margin:0.2rem 0 0.6rem 0;"> ¹ South China University of Technology &nbsp;&nbsp;|&nbsp;&nbsp; ² Tsinghua University &nbsp;&nbsp;|&nbsp;&nbsp; ³ Zhejiang University &nbsp;&nbsp;|&nbsp;&nbsp; ⁴ Tencent VISVISE &nbsp;&nbsp;|&nbsp;&nbsp; ⁵ Peking University &nbsp;&nbsp;|&nbsp;&nbsp; ⁶ Singapore Management University </p> <p align="center" style="font-size:0.95em; color:#666; margin-top:0;"> * Equal contribution &nbsp;&nbsp;|&nbsp;&nbsp; † Corresponding author </p> <p align="center"> <a href="https://zazexy.github.io/flashmesh.github.io/"> <img src="https://img.shields.io/badge/Project%20Page-blue.svg" alt="Project Page" height="22"> </a> <a href="https://arxiv.org/abs/2511.15618"> <img src="https://img.shields.io/badge/arXiv-b31b1b.svg?logo=arXiv&logoColor=white" alt="arXiv height="22"> </a> </p> <h1 align="center" style="line-height:1.3; margin-bottom:0.6rem;"> <!-- FlashMesh 标题图片 --> <img src="./assets/teaser_lightning.jpg" alt="FlashMesh" style="display:block; margin:0 auto 0.4rem auto; max-width:100%;"> </h1> <!-- <p align="center"> <img width="90%" alt="pipeline", src="./assets/Teaser.png"> </p> --> </h4>

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 <b>FlashMesh</b>, 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 <b>a 2x speedup⚡</b> over standard autoregressive models while also <b>improving generation fidelity👍</b>. Our results demonstrate that structural priors in mesh data can be systematically harnessed to accelerate and enhance autoregressive generation.

TODO

• [ ] Release inference & training code of Hourglass tarnsformers
• [ ] Release inference code for FlashMesh
• [ ] Release training code for FlashMesh

Citation

If you find our work helpful, please consider citing:

@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}, 
}