Language-conditioned virtual photography on VPhotoBench 90 tasks (held-out)
0.578MqsPhotoFlow
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| PhotoFlowBudget=6-round2026.05 | 0.578 | 0.622 | 0.55 | 0.564 | 0.614 | |
| Single-Chain Reflection2026.05 | 0.567 | 0.567 | 0.53 | 0.545 | 0.616 | |
| Random Search2026.05 | 0.527 | 0.4 | 0.483 | 0.492 | 0.589 | |
| Anchor Bank Best-of-NSelection=Best-of-N2026.05 | 0.519 | 0.378 | 0.464 | 0.481 | 0.593 | |
| Single-Step LLM2026.05 | 0.514 | 0.4 | 0.447 | 0.47 | 0.603 |