PhyProbe: Rethinking Physical Consistency Evaluation in Generated Videos

PhyProbe is a new evaluator that extracts features from a frozen pretrained spatio-temporal encoder and maps them to a scalar physical consistency violation score using a lightweight scoring head. It is trained with a unified objective combining pairwise ranking, regression on noisy scalar annotations, and anchor-based calibration over heterogeneous supervision sources. The approach aims to address limitations of existing methods that rely on off-the-shelf vision-language models or fine-tuned evaluators trained on human annotations.