Strips as Tokens: Artist Mesh Generation with Native UV Segmentation
This presentation explores SATO, a breakthrough autoregressive framework that generates professional-quality 3D meshes with artist-aligned topology and native UV segmentation. Unlike previous systems that treat mesh generation as a purely geometric problem, SATO encodes meshes as sequences of triangle strips—mirroring how artists naturally think about edge flow and surface structure. The framework uniquely generates both geometry and UV chart partitions in a single unified sequence, eliminating traditional post-processing bottlenecks and producing meshes ready for texturing, animation, and rendering workflows.Script
Most neural mesh generators produce tangled topology that professional artists must painstakingly clean up by hand. The disconnect isn't just aesthetic—it's a fundamental mismatch between how models tokenize geometry and how production pipelines actually work.
SATO replaces arbitrary coordinate tokenization with strip-based encoding that mirrors how artists conceptualize surface structure. Each strip follows organized edge flows, creating topological coherence that earlier methods had to learn implicitly. The same token stream decodes into triangles or quads depending solely on stride, allowing the model to transfer knowledge between both representations and generating UV chart boundaries as part of the autoregressive process itself.
This architectural shift translates into measurable quality gains across every dimension that matters for production use.
Against leading baselines including MeshAnythingV2 and DeepMesh, SATO achieves a 33 percent F1-score improvement on ShapeNet while maintaining normal consistency above 0.97. User studies with professional 3D artists confirm the difference—SATO meshes exhibit cleaner topology, lower UV distortion, and significantly less manual cleanup required for production workflows.
The unified decoder architecture enables a training strategy impossible for previous methods. SATO pretrains on abundant triangle meshes to learn geometric priors, then fine-tunes on scarce but high-quality quad datasets. Because quads decompose naturally into triangle pairs within the strip representation, improvements in one domain directly transfer to the other—quad fine-tuning even refines triangle mesh organization.
By encoding the structural priors artists rely on directly into the tokenization, SATO eliminates the gap between what models generate and what production demands. The framework produces meshes with clean UV layouts, organized edge flows, and topology ready for downstream tasks—all from a single autoregressive pass.
SATO proves that how you tokenize geometry isn't just an implementation detail—it fundamentally determines whether generated meshes align with the workflows professionals actually use. Visit EmergentMind.com to explore more research and create your own video presentations.
