Coverage as a design objective
A derivative-free search (TPE) tunes generator motion to cover the target's fingerprint. Geometry-only search: no RGB rendering or model training for each candidate. Only the selected configuration is ever rendered.
+10.7
peak PCK, GLU-Net on KITTI-2015: tuned MOVi-F vs. generic MOVi-F
74 → 96
PCK@3%, GLU-Net on KITTI-2012: generic → coverage-tuned SDF source
| Architecture | @5% generic | @5% tuned | @3% generic | @3% tuned |
|---|---|---|---|---|
| CATs++ | 97.2 | 98.7 | 93.8 | 97.4 |
| GLU-Net | 89.0 | 98.0 | 74.2 | 96.2 |
| FlowFormer | 95.4 | 96.8 | 84.8 | 91.9 |
Scale context, not a ranking: SD-DINO zero-shot scores 83.6 and pretrained UFM-Base 98.8 at PCK@3% on KITTI-2012 under a separate harness, neither budget-matched. At @1% the same GLU-Net rows read 22.3 → 89.1.
One objective improves two generators with nothing in common, photorealistic Kubric and asset-free fractals. The signal is a property of the data, not the renderer.