← News

ResearchSources: arXiv · 19 Nov 2026

H3Remover tunes MiniMax-H3 video editing model for object removal without paired training data

Researchers report a reinforcement approach that trains video object removal on unpaired data, using rewards for depth, background geometry and appearance to avoid superficial edits.

In plain words

Removing an object from a video usually requires examples of the same scene with and without the object, which are hard to collect. H3Remover instead trains the MiniMax-H3 video editor using automatic scoring signals that reward convincing removal and penalize unwanted changes. The authors report stronger reconstruction results than the base model.

What was announced

The authors present H3Remover, a system that reinforces MiniMax-H3 Base Ref2VA for video object removal without paired object-free training targets. Pretrained video editors can retain targets or alter unrelated content, so the method applies three rewards: a target-removal reward favoring calibrated depth changes consistent with revealing farther content, a background-geometry reward preserving relative depth ordering, and an appearance reward constraining RGB changes in source-derived protected regions while exempting targets and estimated effects. Frozen HPSv3 adds perceptual-quality feedback for synthesized content.

Training uses DiffusionART to optimize LoRA adapters while the backbone stays frozen, and evaluation regions remain fixed across candidates. The authors report substantial reconstruction gains over the original H3, with qualitative comparisons showing more effective object removal and better preservation of surrounding scene content. The evidence comes from the paper's abstract, which states the reported results without independent verification.

What it does not show

  • The source is the paper's abstract only; full experimental details, datasets and metrics are not available in the supplied passage.
  • Reported gains over the original H3 come from the authors' own evaluation and have not been independently confirmed.

Sources

Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 20 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.