FlashMesh: Faster and Better Autoregressive Mesh Synthesis via Structured Speculation

TL;DR

FlashMesh tackles the slow inference of autoregressive mesh generation by rethinking decoding as a predict–correct–verify process. By leveraging a three-level hourglass Transformer and specialized SP-Block/HF-Block modules, it enables parallel speculative predictions across faces, vertices, and coordinates while enforcing geometric consistency through correction and verification steps. The approach yields up to $2\times$ faster generation with improved geometric fidelity and topology, validated across multiple meshes and baselines. These results demonstrate that incorporating structural mesh priors into speculative decoding can significantly accelerate and enhance autoregressive 3D synthesis. This work offers a practical pathway toward scalable, high-quality mesh generation for interactive and large-scale 3D content creation.

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 2 x 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.

Figures (10)

• Figure 1: Overall architecture of the proposed predict-correct-verify framework. Predict: the original Hourglass Transformer generates main tokens, while the lightweight SP-Block and HF-Block parallelly produce draft tokens. Correct: a correction mechanism enforces vertex-sharing consistency. Verify: the backbone re-evaluates main and corrected draft tokens in a single forward pass and accepts the verified ones. Bottom right: Point-level pipeline of multi-layer multi-head speculative decoding.
• Figure 2: (a) The Speculative Prediction Block (SP-Block) predicts multiple draft tokens in parallel from the current hidden state. (b) The Hierarchical Fusion Block (HF-Block) refines speculative embeddings by fusing them with cached local context for accurate token prediction.
• Figure 3: Example of vertex misalignment in parallel face generation and our correction mechanism.
• Figure 4: Example of the verify mechanism with $D=2$. The backbone verifies draft tokens in parallel, accepts consistent tokens up to $S^{*}$, and reprocesses the subsequent segment.
• Figure 5: Qualitative comparison of mesh generation results. We compare FlashMesh against baseline methods including BPT and DeepMesh. Besides, in the top three samples, we also show the results of Ours (Meshtron-2B) and Meshtron-2B, while in the bottom three samples, we also present Ours (Mesh-RFT) and Mesh-RFT. Our method, FlashMesh, achieving high geometric fidelity while significantly accelerating the generation process.