MeshCrafter: Structurally Robust Autoregressive Mesh Generation

Autoregressive mesh generation becomes increasingly challenging as mesh complexity and sequence length grow. Although recent methods improve scalability through compact representations or localized generation, geometric structure is typically imposed through discrete partitioning and fixed serialization, which discards ambiguous transitions between neighboring regions. Moreover, standard next-token training optimizes local prediction likelihood rather than the geometric and structural quality of the complete generated mesh. We present MeshCrafter, a structure-aware autoregressive framework that incorporates geometric organization throughout region modeling, serialization, and post-training. MeshCrafter first constructs latent geometric regions from pretrained geometry features and refines their organization with a triplet objective, while explicitly estimating a continuous interface-proximity field to preserve structural transitions discarded by hard assignments. Based on this representation, we introduce soft-interface-guided hierarchical mesh tokenization, where discrete region ownership constrains the traversal space and continuous interface proximity determines the traversal priority. A hierarchical coarse-to-fine spatial encoding further provides a compact representation of 3D vertices. Finally, we introduce structure-aware reinforcement post-training, which compares groups of complete mesh trajectories using geometric and latent structural feedback and directly optimizes the generation policy at the sequence level. Experiments on a 1k-mesh benchmark demonstrate consistent improvements over representative autoregressive mesh generators. MeshCrafter achieves state-of-art results compared among strongest evaluated baselines, while presenting its robustness and effectiveness.