HiFi-Mesh: High-Fidelity Efficient 3D Mesh Generation via Compact Autoregressive Dependence

TL;DR

HiFi-Mesh tackles the bottlenecks of autoregressive 3D mesh generation by introducing Latent Autoregressive Network (LANE) to create compact latent-space dependencies for subsequence autoregression and Adaptive Computation Graph Reconfiguration (AdaGraph) to enable parallel subgraph generation. LANE partitions the mesh sequence into M subsequences, builds hierarchical latent spaces, and uses learnable queries to recover details with reduced long-historical dependence; AdaGraph enables spatiotemporal decoupling and parallel inference, yielding about a $3\times$ speedup and a $6\times$ increase in maximum sequence length. On Objaverse data with 8 NVIDIA H20 GPUs, HiFi-Mesh achieves superior generation speed and geometric fidelity, while maintaining topology and point-cloud consistency compared to baselines. This work demonstrates a scalable, high-fidelity pipeline for automatic 3D mesh synthesis with practical potential for large-scale content creation.

Abstract

High-fidelity 3D meshes can be tokenized into one-dimension (1D) sequences and directly modeled using autoregressive approaches for faces and vertices. However, existing methods suffer from insufficient resource utilization, resulting in slow inference and the ability to handle only small-scale sequences, which severely constrains the expressible structural details. We introduce the Latent Autoregressive Network (LANE), which incorporates compact autoregressive dependencies in the generation process, achieving a $6\times$ improvement in maximum generatable sequence length compared to existing methods. To further accelerate inference, we propose the Adaptive Computation Graph Reconfiguration (AdaGraph) strategy, which effectively overcomes the efficiency bottleneck of traditional serial inference through spatiotemporal decoupling in the generation process. Experimental validation demonstrates that LANE achieves superior performance across generation speed, structural detail, and geometric consistency, providing an effective solution for high-quality 3D mesh generation.

Figures (7)

• Figure 1: Performance comparisons of EdgeRunner tang2024edgerunner, MeshAnythingV2 chen2024meshanything, TreeMeshGPT lionar2025treemeshgpt, and HiFi-Mesh (Ours) on Chamferi Distance $\downarrow$, Inference Speed $\uparrow$, and Maximum Generatable Sequence Length $\uparrow$.
• Figure 2: Visual Comparisons between TreeMeshGPT lionar2025treemeshgpt and HiFi-Mesh (Ours). A mesh can be tokenized into a 1D topological sequence. #Token represents sequence length with the density of vertices and faces proportional to it. $X$ Tok/s denotes the inference speed of sequence token. HiFi-Mesh enables detail-rich mesh generation with faster speed.
• Figure 3: The illustration of our proposed method. During training, the purpose of LANE model (left) uniformly divides the topological sequence into $M$ subsequences for sequential generation. The model takes point cloud $P$ and sequence length $L$ as inputs, employing a Latent Space Extractor and Autoregressive Block for sequence semantic reconstruction to build hierarchical latent space representations $\{sc_m\}_{m\in[1,M]}$ as compact autoregressive dependencies. Subsequently, $K$ LANE Blocks serve as generation components, utilizing Learnable Queries $Q$ as latent space queriers. Guided by subsequence index $I=\{m\} _{m\in[1,M]})$, each step generates the $m^{th}$ subsequence $\hat{s_m}$ from the only first $m$ latent spaces. During inference, the AdaGraph strategy (right) rapidly generates hierarchical latent space representations $\{sc_m\}_{m\in[1,M]}$ and instantiates $M$ independent pathways $\{lk_m\}_{m\in[1,M]}$ composed of LANE Blocks. Each pathway dynamically reconstructs a dedicated computational subgraph based on the target subsequence and selectively activates corresponding latent spaces, enabling fast parallel generation. Finally, mesh generation is completed through subsequence concatenation and detokenization operations.
• Figure 4: Qualitative comparisons between TreeMeshGPT lionar2025treemeshgpt, EdgeRunner tang2024edgerunner, MeshAnythingV2 chen2024meshanything and HiFi-Mesh (Ours). MeshAnythingV2 exhibits limitations in maximum sequence scale, causing most significant structural information loss. EdgeRunner shows theoretical improvements but maintains preference for small-scale sequences. TreeMeshGPT achieves some improvement in detail modeling but still suffers from insufficient sequence generation capability. Finally, HiFi-Mesh enables superior preservation of details and generation of detail-rich meshes.
• Figure 5: Functional presentation. Point clouds are sampled from imperfect meshes. HiFi-Mesh generates meshes with rich details guided by different sequence lengths, while EdgeRunner tang2024edgerunner faces local topological breaks.