H3Remover

Reinforcing Video Object Removal
without Paired Data

Remove selected objects, their shadows and reflections. Keep the rest of the scene.

See the difference

Input · selected objectsH3Remover
MOSE · Multiple vehicles crossing a city street.Browse the full gallery

Abstract

H3Remover adapts a pretrained video editor using reinforcement learning, without paired object-free training targets. Depth rewards encourage the editor to reveal what lies behind the target and preserve surrounding geometry. Appearance and preference rewards help retain unrelated scene details.

Previews loop as you scroll. Select a scene for the full-quality video. Left: input with target masks. Right: H3Remover.

Citation

@misc{chen2026h3removerreinforcingvideoobject,
  title         = {H3Remover: Reinforcing Video Object Removal without Paired Data},
  author        = {Kangjie Chen and Xiangyu Li and Jinhao Deng and Dongbin Zhang and
                   Chaoda Zheng and Shijia Chen and Yuqi Zhang and Choo Sin Wai and
                   Boyi Sun and Puyi Wang and Yu Zhang and Xianming Liu and Boyang Wang},
  year          = {2026},
  eprint        = {2611.24633},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2611.24633}
}
Download .bib