RFD 1121: Layers come from geometry, and the missing categories come from a multiview rebuild
pixal3d-image-to-textured-mesh, trellis2-image-to-textured-mesh, pose-consensus, voxhammer-image-mesh-editing, plus garment-retarget and ledger repositories
Problem
See-Through cuts one flat image into body-part layers, and a layer must show the whole part. An occluder hides the surface behind it, so the model invents those pixels, and that invention is the main source of error. Rendering each hm08 vertex group on its own and letting the depth buffer order them removes it, because the hidden surface is real geometry.
That route covers only what ANNY models, and ANNY models a body: hm08 has 12 named groups, ten tags have no geometry, and body is one 13,380-vertex block with no face, ears, nose or neck.
Decision
Split the corpus into two halves, each with its own truth condition. Half A is the modelled half. Render hm08 groups as depth-ordered layers. Nothing generates and no pixel is inpainted. It needs a body sub-partition into head and neck tags, proved in Lean. Clothing belongs here too: hm08 carries tights, skirt and hair as proxies, cloth-fit retargets an authored garment onto them without intersections under MIT, and a dressed render differenced against an undressed one gives the layer exactly. The five absent worn tags want a proxy each.
Half B is the missing half. Build the geometry first and render it second; the order is the whole decision. Qwen-Image supplies appearance under its Apache-2.0 depth ControlNet, a 3D lift turns the views into geometry, and VoxHammer adds one category at a time under a mask. Never inpaint in 2D, because each 2D inpaint runs once per view and disagrees with the next. Half B is generated synthetic, so all four conditions apply, and pose-consensus refits it to its pose.
See DETAILS.md for the hm08-partition correction, the ten absent tags, the frozen vertex order and its hash gate, the garment route, and the ledger protocol.