OSH reconstruction failure — GT vs OSH vs vanilla
25-step Euler reverse sampling (rectified flow, from pure noise t=1 → data t=0) on the GT sparse structure, decoded with the released SLAT mesh decoder, normal-map render. The sparse-structure coords are teacher-forced to GT occupancy, so each render isolates Stage-2 SLAT fine-geometry. Same object, three columns: ground truth, OSH-conditioned reconstruction, vanilla reconstruction.
What to look for. In-distribution: OSH ≈ GT, clean. On OOD (held-out categories & ArtiCraft primitives): the fine geometry collapses for BOTH arms, and OSH is no better (often worse) than vanilla — the visual face of the flow-MSE numbers (in-dist OSH 0.035; OOD OSH 5.28 > vanilla 2.89).
GT
OSH
vanilla
in-distribution
Dishwasher
Table
held-out PNM (category-OOD)
Chair
Faucet
Toilet
ArtiCraft (representation-OOD)
cannon
