PhyProbe: Rethinking Physical Consistency Evaluation in Generated Videos
PhyProbe scores generated videos for physical violations using frozen spatio-temporal features plus a lightweight calibrated head.
The paper argues that current evaluators either miss dynamics or overfit human-annotation datasets. PhyProbe trains on mixed supervision with ranking, noisy scalar regression, and anchor calibration to produce a stable 0–1 violation score. The authors report better results than prior methods on most pairwise benchmarks, especially generated-generated and no-correspondence cases. HF Daily Papers' note
The paper argues that current evaluators either miss dynamics or overfit human-annotation datasets. PhyProbe trains on mixed supervision with ranking, noisy scalar regression, and anchor calibration to produce a stable 0–1 violation score. The authors report better results than prior methods on most pairwise benchmarks, especially generated-generated and no-correspondence cases. HF Daily Papers' note
score 5