H3Remover

Remove Anything with H3Remover

Watch demo ↓Explore results ↓Code · GitHub ↗Model weights ↗Paper · PDF ↗
82 seconds · 1080pDownload video ↗

THE RESULTS

Many scenes. One editor.

From city traffic to people, animals and everyday objects.
Moving comparisons across real-world scenes.

THE IDEA

Unlock the priors.
Guide the edit.

A pretrained video editor already knows how to reconstruct a scene.
H3Remover reinforces when—and where—to make a change.

H3Remover (H3 Remover) adapts MiniMax-H3 Base Ref2VA for video object removal without paired object-free training targets. Source-conditioned rewards encourage target removal while constraining changes to the surrounding scene. DiffusionART updates lightweight LoRA adapters while the pretrained backbone remains frozen.

01 / GEOMETRY

Reveal what lies behind.

Target-depth feedback encourages farther content in removal regions, while relative depth ordering constrains the surrounding geometry.

02 / APPEARANCE

Keep the scene intact.

Binary protection masks preserve background appearance while allowing estimated object-associated effects to change.

03 / QUALITY

Encourage natural results.

Frozen HPSv3 supplies complementary perceptual feedback alongside the explicit removal and preservation rewards.

View the full method diagram +H3Remover training pipeline with geometry, appearance and perceptual quality rewards

THE PAPER

Reinforcing Video Object Removal
without Paired Data

Kangjie Chen1*, Xiangyu Li1*, Jinhao Deng1, Dongbin Zhang1, Chaoda Zheng1, Shijia Chen1,
Yuqi Zhang1, Choo Sin Wai3, Boyi Sun4, Puyi Wang2, Yu Zhang1, Xianming Liu1, Boyang Wang1†

1 XPeng Motors2 The Chinese University of Hong Kong3 Tsinghua University4 Institute of Automation, Chinese Academy of Sciences

* Equal contribution   † Corresponding author