FlashMesh: Efficient 3D Mesh Synthesis

• FlashMesh is a high-throughput autoregressive framework for 3D mesh synthesis that restructures token-wise decoding into a speculative, hierarchy-aware process.
• It integrates SP-Block, HF-Block, and a label-prediction head to predict, correct, and verify tokens, ensuring geometric fidelity and consistency.
• The approach doubles generation speed and improves mesh quality on benchmarks like ShapeNetV2 by leveraging strong structural priors in mesh data.

FlashMesh is a high-throughput, high-fidelity autoregressive framework for 3D mesh synthesis that restructures traditional token-wise decoding into a speculative, hierarchy-aware paradigm. Conventional autoregressive mesh models generate meshes one token at a time (vertex, face, or coordinate), leading to slow inference and limiting scalability. FlashMesh introduces a three-stage Predict–Correct–Verify mechanism combined with structured multi-token speculation and architectural enhancements to the hourglass transformer, doubling generation speed and improving geometric fidelity by leveraging strong structural priors present in mesh data (Shen et al., 19 Nov 2025).

1. Predict–Correct–Verify Paradigm

The core innovation of FlashMesh lies in replacing single-step decoding with a three-stage, multi-token decoding scheme:

1. Predict: At each split node in the hourglass transformer architecture, two lightweight modules—the Speculative Prediction Block (SP-Block) and the Hierarchical Fusion Block (HF-Block)—speculatively emit $D$ future tokens in parallel. Prediction occurs across mesh hierarchies: face, point, and coordinate levels.
2. Correct: To ensure geometric and topological consistency in the speculated tokens, a correction step is performed. A label-prediction head classifies each new point (as historical, new, or intra-batch), aligning misaligned intra-batch points to the nearest correct point and reordering them in $z$–$y$–$x$ order.
3. Verify: The backbone refines and verifies up to $D+1$ tokens in a single forward pass with causal masking. The outputs are compared to the draft speculations, accepting all consecutive tokens until the last match $s^*$, which advances the inference position for the next speculative batch.

This approach capitalizes on mesh data's hierarchical and strongly correlated structure; point, face, and coordinate tokens exhibit high mutual information, which enables safe multi-token speculation and acceptance.

2. Structured Multi-Token Speculation

FlashMesh designs its speculative generation to match the hierarchical structure of meshes. At decoding position $s$, the process unfolds as follows:

1. The hourglass transformer backbone produces the next main token $x_{s+1}$ after $N$ layers.
2. At the penultimate layer ("split node"), $D$ parallel SP-Block heads generate high-level speculative embeddings for positions $s+d$:

$$h_{s+d}^{(d)} = \mathrm{Linear}\Bigl( \mathrm{CrossAttn}^{(d)}(\mathrm{SelfAttn}^{(d)}(h_s),\,c) \Bigr) + h_s$$

where $c$ is an optional condition.

1. The high-level speculative embeddings are upsampled into lower-level token features by an Upsample operator:

$$[h_{s+3d}^{(3d)'},\,h_{s+3d+1}^{(3d+1)'},\,h_{s+3d+2}^{(3d+2)'}] = \mathrm{Upsample}(h_{s+d}^{(d)})$$

2. Each upsampled feature undergoes context integration via the HF-Block, attending to cached key/value states $X_{<s}$ from previous tokens:

$$Q_{s+t}=W_q h_{s+t}',\quad K_{<s}=W_k X^{k}_{<s},\quad V_{<s}=W_v X^{v}_{<s}$$

$$\tilde h_{s+t} = h_{s+t}' + \mathrm{FFN}(\mathrm{Attn}(Q_{s+t},K_{<s},V_{<s}))$$

3. SP-Block and HF-Block(s) are composed in series at each split node, with draft token distributions averaged across members.

This multi-layer, multi-head speculative decoding approach exploits mesh adjacency and local structure, improving both efficiency and coherence.

3. Architectural Enhancements to Hourglass Transformer

FlashMesh augments the baseline hourglass transformer (as in Meshtron) by inserting speculative decoding and correction mechanisms at key upsampling points:

• SP-Block: Parallel heads at split nodes that each hypothesize $D$ future high-level tokens in parallel.
• HF-Block(s): Each speculative token is contextually fused with prior tokens for coherence.
• Label Head: Added at point-level upsampling to classify candidate points for geometric correction.

A high-level schematic:

…→ Transformer block → SP-Block → (Upsample) → HF-Block(s) → continue → …

This configuration enables effective speculative generation over mesh hierarchies and corrections to preserve mesh validity.

4. Mathematical Formulation and Training Losses

The mathematical operations for speculative decoding are detailed as follows:

• Main token prediction:

$$p(x_{s+1}\mid x_{<s+1},c),\quad \hat x_{s+1} = \arg\max p(\cdot)$$

• Draft tokens: Produced for $d=2\ldots D+1$ by SP+HF blocks.
• Training loss: Combines a coordinate loss for token prediction and a label loss for point classification:

$$\mathcal{L}_{\mathrm{coord}} = -\frac{1}{N_c} \sum_{t=1}^{N_c} \log p_t(x_t)$$

$$\mathcal{L}_{\mathrm{label}} = -\frac{1}{N_p} \sum_{t=1}^{N_p} \log p_t(y_t)$$

$$\mathcal{L}_{\mathrm{total}} = \mathcal{L}_{\mathrm{coord}} + \gamma \mathcal{L}_{\mathrm{label}}$$

• Verification: Using causal masking, the backbone re-computes predictions $x'_{s+2:s+D+2}$. Tokens $x_{s+t} = x'_{s+t}$ are accepted for all $t \le s^*$, where:

$s^* = \max\{t \le D+1 : x_{s+t} = x'_{s+t}\}$

Pseudocode for the decoding loop is provided in the data.

5. Quantitative Evaluation and Efficiency

FlashMesh achieves substantial gains in efficiency and maintains or improves fidelity across established mesh generation benchmarks (ShapeNetV2, gObjaverse):

Method	Params	CD↓	HD↓	BBox-IoU↑	TPS↑	Speed-up
Meshtron (1B)	1.1 B	0.121	0.269	0.901	98.6	1.00×
FlashMesh (Meshtron)	1.6 B	0.120	0.267	0.905	180.4	×1.83
Meshtron (2B)	2.3 B	0.092	0.206	0.942	67.3	1.00×
FlashMesh (Meshtron)	3.4 B	0.089	0.198	0.949	136.6	×2.03

Ablation studies reveal incremental gains in speed and accuracy as speculative prediction, hierarchical fusion, and correction modules are introduced:

Config	CD↓	HD↓	BBox-IoU↑	TPS↑
A: baseline Meshtron 1B	0.121	0.269	0.901	95.5
B: + SP-Block only	0.122	0.269	0.903	109.7
C: + SP + HF-Block	0.120	0.268	0.904	176.5
D: + SP + HF + Correction	0.120	0.267	0.905	180.4

Measured mean accepted speculative tokens approach 9.8/18 for faces and 10.0/15 for points, producing actual throughput ratio improvements close to 2× on both 1B and 2B parameter Meshtron backbones.

6. Exploiting Structural Priors in Mesh Data

Mesh data is characterized by strong structural and geometric priors:

• Adjacent faces share vertices, yielding high mutual information $I(\mathrm{face}_i; \mathrm{face}_{i+1})$.
• Each vertex comprises an $(x, y, z)$ coordinate triple, enforcing local continuity.
• The mesh’s triangular topology mandates $z$–$y$–$x$ ordering and shared-index constraints.

FlashMesh’s multi-token speculation and loss jointly maximize mutual information terms in the joint entropy:

$H(X, Y) = H(X) + H(Y|X)$

$H(X) + H(Y) = H(X|Y) + 2 I(X;Y) + H(Y|X)$

This approach increases the model’s sensitivity to geometric and topological relations, enhancing coherence and reducing error accumulation. Algorithmically, the correction module enforces shared-vertex consistency by merging intra-batch points as needed.

A plausible implication is that this structural exploitation not only accelerates decoding but also regularizes the sampling process, ensuring a lower incidence of degenerate or inconsistent mesh outputs under speculative decoding.

7. Significance and Context

FlashMesh demonstrates that hierarchical, structure-aware speculative decoding is practical and effective for mesh generation tasks. By integrating speculative prediction, mesh-specific correction, and verification stages, combined with multi-token objectives that emphasize inter-token dependencies, FlashMesh achieves both substantial acceleration and qualitative gains over previous autoregressive mesh frameworks. These results substantiate the feasibility of systematically leveraging structural mesh priors to improve both inference speed and output quality for interactive and large-scale 3D applications (Shen et al., 19 Nov 2025).