PROVE-Bench: A New Benchmark for Evaluating Object Removal in Visual Media

A new benchmark called PROVE-Bench has been unveiled by researchers to assess object removal in both images and videos. This two-tier system includes perception-aligned metrics known as RC (Removal Coherence). It aims to overcome the shortcomings of current evaluation techniques, which frequently conflict with human judgment. Full-reference metrics often favor copy-paste actions rather than true removal, while no-reference metrics can be biased towards blurry outcomes, and global temporal metrics overlook localized artifacts. The RC metrics feature RC-S for spatial coherence, utilizing sliding-window comparisons between masked and background areas, and RC-T for temporal consistency, monitoring distributions in shared restored regions across neighboring frames. PROVE-Bench also includes PROVE-M, an 80-video paired dataset enhanced with motion augmentation for community benchmarking. This research is documented in arXiv preprint 2605.14534.