MiniMax H3 license agreement

This adapter is distributed under the MiniMax H3 Community License. Please read the LICENSE before downloading.

Log in or Sign Up to review the conditions and access this model content.

🎬 H3Remover: Reinforcing Video Object Removal without Paired Data

✨ Remove Anything with H3Remover.

Project Page   Paper on arXiv   Model Weights

GitHub Code  Β·  Watch eight comparisons  Β·  πŸƒ Run the Gradio demo locally  Β·  Quickstart  Β·  Citation

MOSE: selected vehicles disappear from a busy intersection DAVIS: removal in an indoor scene with a person lifting a bicycle

Left: clearing traffic from an intersection (MOSE). Right: a person lifting a bicycle indoors (DAVIS). Click either preview for the synchronized input and result, or watch the 100-example showcase.

πŸƒβ€β™‚οΈ Gradio Demo

Load the city traffic example, preview its red target mask, run removal, and play the edited video in Gradio.

From a mask to a finished edit in the local Gradio demo. The final video comes from this run; the inference wait is shortened.

🎞️ Video results

Eight clips from the released adapter. In every player, the left shows the input with the selected targets highlighted; the right shows the edit at the same instant. Watch the shadows on the dunes, the reflection on the lake, or the second camel that stays in the scene.

People and long shadows on sand

PROVE-H Β· Open video
A boat and its reflection

WIPER Β· Open video
One camel, not the other

DAVIS Β· Open video
Remove the players, keep the ball

PROVE-H Β· Open video
Traffic beneath an overpass

Waymo Β· Open video
Two people in a kitchen

PROVE-M Β· Open video
A deer against a snowy forest

WIPER Β· Open video
A mirrored ball and its shadow

WIPER Β· Open video

Browse all 88 comparisons on the project page Β· Watch the full showcase film

Overview

Give H3Remover a video and a mask of what should disappear. The mask can select one object or several at once. The edit targets not just the object itself, but also the shadows, reflections and light it casts, while preserving unrelated textures and camera motion. This release is a LoRA adapter for MiniMax H3 Ref2VA, trained with reinforcement learning without paired object-free targets.

No street mask yet? The optional street workflow selects people and vehicles with SAM 3.1 and uses Pi4 to propose other moving regions. Preview the combined mask before editing.

Quickstart

1. Install

Use Linux, an NVIDIA GPU, and FFmpeg. The reference setup uses Python 3.13 and PyTorch 2.10.0 with CUDA 13:

git clone https://github.com/chenkangjie1123/H3Remover.git
cd H3Remover
conda create -n h3remover python=3.13 -y
conda activate h3remover
python -m pip install torch==2.10.0 torchvision==0.25.0 --index-url https://download.pytorch.org/whl/cu130
python -m pip install -r requirements.txt

2. Download weights

python download_weights.py

This downloads the pinned adapter and full Ref2VA base to weights/ and verifies the adapter checksum. If access is required, accept the model licenses and run hf auth login first.

3. Remove objects

python infer.py \
  --video examples/street/source.mp4 --mask examples/street/mask.mp4 \
  --output outputs/street.mp4

Replace the example with your own video and synchronized binary mask. White selects objects to remove; black keeps the rest. The pipeline creates the visual mask reference automatically.

The output is a silent 124-frame, 24 FPS, 960 Γ— 544 MP4, using 50 inference steps. Provide at least 5.17 seconds; longer inputs use the first clip. Component CPU offloading and tiled VAE decoding are enabled. See the inference guide for input previews and settings.

Adapter

Property Value
Base MiniMax H3 Ref2VA
Rank / alpha 16 / 16
Format Safetensors, FP32
Size 160,613,888 bytes

SHA-256: 6fe5b881ec1a9804af005a3ed512e54a6c3341b0e5a2204323a2ed6482262c84

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}
}

License

The adapter is a modified derivative of MiniMax H3, distributed under the MiniMax H3 Community License. The base model is downloaded separately.

Downloads last month
1
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for chenkangjie1123/H3Remover

Adapter
(203)
this model

Paper for chenkangjie1123/H3Remover