OSH in-distribution gain (Stage-2, PNM)
How much does the OpenStateHead conditioning improve in-distribution articulated reconstruction? Matched checkpoints, identical eval, identical PNM held-out samples — the only difference is the OSH mask conditioning.
2.3×
lower SLAT flow-MSE @40k
OSH 0.035 vs vanilla 0.082
4.2×
lower articulation-L2 @40k
OSH 5.8e-6 vs vanilla 2.4e-5
whole run
OSH stays below vanilla the entire 40k
not a late-training artifact
| checkpoint | flow OSH | flow van | gain | art-L2 OSH | art-L2 van | gain |
|---|---|---|---|---|---|---|
| step 40000 | 0.0354 | 0.0815 | 2.30× | 5.84e-06 | 2.44e-05 | 4.17× |
Flow-MSE over training — OSH (blue) consistently below vanilla (orange).
Articulation-L2 (24-D joint regression) — the OSH mask prior helps joint-param recovery most.