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

2026-01-29

AI Summary

This work addresses the limitations of existing autoregressive approaches in high-fidelity 3D mesh generation, which suffer from inefficient resource utilization, slow sequential inference, and difficulty in modeling long sequences—hindering fine-grained geometric detail. To overcome these challenges, the authors propose the Latent Autoregressive Network (LANE), which encodes 3D meshes into compact one-dimensional sequences and introduces an adaptive computation graph reconstruction strategy (AdaGraph) to efficiently model autoregressive dependencies between vertices and faces. This approach substantially alleviates bottlenecks in sequence length and inference efficiency, outperforming current methods in generation speed, geometric consistency, and structural detail. Notably, LANE achieves up to a sixfold increase in the maximum generable sequence length.

Technology Category

• Machine Learning: Deep Generative Models & Autoencoders
• Natural Language Processing: Generation
• Computer Vision: 3D Computer Vision

Application Category

• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Responsible Web: Machine-in-the-loop, human agency and autonomy
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

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.

Problem

• 3D mesh generation
• autoregressive modeling
• sequence length limitation
• inference efficiency
• structural detail

Innovation

• autoregressive generation
• 3D mesh generation
• compact dependence
• adaptive computation graph
• high-fidelity modeling