MeshOctave treats mesh generation as a coarse-to-fine cascade where resolution is defined by spatial grid levels, not by serial simplification steps. Coarsening merges vertices in the same voxel and inherits connectivity; the inverse split-and-rewire operation is a per-face local decision encoded as discrete structural tokens. A scale-conditioned masked-uniform discrete learns these tokens and generates each scale in parallel.
Compared with autoregressive mesh generators, there is no token-by-token serial pass over faces. Compared with continuous flow models, topology is not resolved by thresholding continuous fields. MeshOctave reports non-trivial gains over strong autoregressive and flow-matching baselines on geometric fidelity and topological validity, and it supports adaptive resolution refinement and subdivision.