Title: MeshWeaver: Sparse-Voxel-Guided Surface Weaving for Autoregressive Mesh Generation

Abstract: By converting meshes into sequential tokens and employing a language-modeling training strategy, autoregressive mesh generation has recently attracted significant interest. Nevertheless, current methods are hampered by two core issues: first, inefficient tokenization results in excessively long sequences, thereby hindering the scalability to high-polygon meshes; second, they lack geometry-aware guidance, as the generation process relies solely on global shape embeddings rather than local surface details. To address these challenges, we present MeshWeaver, an autoregressive framework that reimagines mesh generation as a surface weaving process. Instead of predicting independent coordinates, our approach directly forecasts the subsequent vertex. Central to this method is a multi-level sparse-voxel encoder that incorporates geometric context into the generation pipeline through three distinct mechanisms: utilizing voxel features as vertex representations, directing token prediction via cross-attention to these features, and acting as a structural scaffold that restricts generation around the input surface. This hierarchical architecture facilitates coarse-to-fine vertex prediction within a single decoding step, ensuring a tight integration between the generative model and 3D geometry. Comprehensive experiments reveal that MeshWeaver sets a new state-of-the-art compression ratio of 18%, supports the generation of meshes with up to 16K faces, and markedly enhances geometric fidelity compared to previous techniques.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC