Real-video supervision for physical consistency

Determine whether real-video supervision improves the physical consistency of videos generated through exemplar-guided optimization, using suitable video–text pairing, data selection, and evaluation.

Background

The experiments use generated videos as exemplars, so they do not establish whether real-world recordings would provide better supervision for object interactions and temporal behavior. The paper identifies physical consistency as a specific unresolved empirical issue and notes that answering it requires appropriate video–text pairings, data-selection procedures, and evaluation methods. GraftOOD provides a potentially suitable regression interface because it needs terminal exemplar videos rather than their generating denoising trajectories, although this design alone does not guarantee successful transfer to real-video data.

References

Whether such supervision improves physical consistency remains an empirical question, requiring suitable video–text pairing, data selection, and evaluation.

— GraftOOD: Grafting Out-of-Distribution Exemplars into Advantage-Weighted Rectification for Video Reward Optimization  (2610.13470 - Fu et al., 9 Oct 2026) in Appendix C, “Limitations and Future Work,” p. 17

More broadly, advances in rewards and evaluation remain necessary to determine whether exemplar-guided optimization improves physical plausibility rather than merely increasing proxy scores.

— GraftOOD: Grafting Out-of-Distribution Exemplars into Advantage-Weighted Rectification for Video Reward Optimization  (2610.13470 - Fu et al., 9 Oct 2026) in Appendix C, “Limitations and Future Work,” p. 17