Xiaomi’s MiLM Plus Releases PROVE: Perception-Aligned Object Removal Metrics RC-S and RC-T With a Real-World Video Benchmark
MarkTechPost Michal Sutter
Xiaomi’s MiLM Plus made new scores for object removal that judge the edited area, not the whole frame. That matters because old metrics kept praising blur and bad erasers.
Based on reporting by MarkTechPost, Michal Sutter — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Object removal has moved faster than the scores used to judge it. The edits look better now — shadows, reflections, even occluded structure can be rebuilt convincingly — but the usual metrics still get the ordering wrong because they expect a neat before-and-after answer where none exists.
MiLM Plus, the Xiaomi team, is trying to fix that with PROVE, short for Perceptual RemOVal cohErence. The work has been accepted at ACM MM 2026 and centers on two new metrics: RC-S for spatial coherence and RC-T for temporal consistency. Both work locally on the edited region instead of scoring the whole frame, and both avoid needing a reference video.
The core trick is sliding-window Maximum Mean Discrepancy over DINOv2 features. RC-S breaks a mask into connected pieces, expands each target, then compares masked features against nearby background features in a local crop. RC-T does something similar across adjacent frames, cropping them together under the union of their masks and then scoring only the overlap where the object was restored in both frames. PROVE also includes PROVE-Bench, which comes in two parts: 80 paired videos for PROVE-M and 100 harder, ground-truth-free clips for PROVE-H.
The paper’s argument is blunt. PSNR, SSIM, LPIPS, ReMOVE and CFD can all reward the wrong thing, especially when blur sneaks in or when the edited region is small compared with the full frame. On ROSE-Bench, the authors say ReMOVE and CFD could even improve as the masked region was blurred. RC-S, by contrast, matched human rankings far better, with an average Kendall’s tau of 0.59 and Spearman’s rho of 0.66.
This is not being sold as a product feature. It is an evaluation harness: Apache 2.0 code, a PyTorch repo, one CLI entry point, and a setup that needs Python 3.10+, PyTorch 2.6+, Transformers 4.51+ and DINOv2-giant weights. RC-S runs at 134.6 ms per frame on a single RTX 4090, which makes it plausible for nightly CI, not for live on-device scoring. In other words: finally, a metric that knows the hole is the whole point.
My take — AI-written commentary, not fact-checked reporting
This is the kind of boring infrastructure work the field keeps pretending it doesn’t need, right up until the leaderboard starts rewarding mush. Full-frame metrics have had a long, lazy run; local scoring is the more honest idea, and Xiaomi’s team seems to have noticed before everyone else finished arguing about blur. A rare case where the fix is less flashy than the problem, which is usually how the useful stuff arrives.
Read more about this at: MarkTechPost
