The input mesh is analyzed directly on the triangle surface — no voxelization, no signed distance field, no volumetric intermediate. Curvature tensors, principal directions, and topology features are extracted in a single pass at 6.3 MB working memory regardless of mesh complexity.

A trained architecture predicts quad edge flow across every surface point. The model operates on actual mesh connectivity — not a rasterized proxy. It sees edges, not pixels. Trained on artist-produced ground truth, validated against production topology standards.

The predicted field is corrected using the mesh's own edge connectivity structure. A dual-representation architecture separates curvature alignment from topology smoothing — two concerns every other method conflates. Thin walls, sharp creases, and concavities are preserved without heuristic pre-passes.

Topologically required irregular vertices are identified from a first-class conflict energy signal — not detected after the fact but predicted as part of the field itself. Placement is automatic. No manual pole management, no artist intervention for genus-zero through high-genus inputs.

A GPU-native solver produces the parameterization that becomes the quad grid. The solve respects the field's singularity structure exactly — no post-hoc fixup. Density is spatially controllable. The solver's working format is a novel sparse structure designed for GPU cache hierarchy, not ported from a CPU library.

Clean quads are traced along the parameterization isolines. T-junctions are resolved. Boundary edges align to input creases. The output is a standard indexed quad mesh ready for subdivision, rigging, or further editing. No cleanup pass required.