# Artisan mesh generation

Generated meshes often carry holes, stray faces and duplicated faces; MeshRepair reads the draft as a prefix and autoregressively emits only a deletion-then-addition edit sequence, raising quad-equivalent face F1 on LATO2 drafts from 0.294 to 0.340 where rule-based repair tools leave it flat or lower.

Generated meshes often carry holes, stray faces and duplicated faces; MeshRepair reads the draft as a prefix and autoregressively emits only a deletion-then-addition edit sequence, raising quad-equivalent face F1 on LATO2 drafts from 0.294 to 0.340 where rule-based repair tools leave it flat or lower.

MeshCarve introduces a flow matching method that runs entirely in spatially compressed latents: a VAE generates coarse vertex occupancy in one forward pass as an anchor, then a vertex flow and an edge flow generate vertex positions and edge connections separately, with both latents learned by a VertexVAE and an EdgeVAE sharing a hierarchical sparse transformer backbone with spatial-aware compression and vertex-link encoding; its latent holds 0.13 tokens per face on average, achieving the lowest Chamfer distance and highest normal consistency on Objaverse and generalizing to Toys4K.

MeshCarve introduces a flow matching method that runs entirely in spatially compressed latents: a VAE generates coarse vertex occupancy in one forward pass as an anchor, then a vertex flow and an edge flow generate vertex positions and edge connections separately, with both latents learned by a VertexVAE and an EdgeVAE sharing a hierarchical sparse transformer backbone with spatial-aware compression and vertex-link encoding; its latent holds 0.13 tokens per face on average, achieving the lowest Chamfer distance and highest normal consistency on Objaverse and generalizing to Toys4K.