From Understanding to Erasing: Towards Complete and Stable Video Object Removal

Dingming Liu¹  Wenjing Wang²  Chen Li²  Jing LYU² 

¹Peking University  ²WeChat Vision, Tencent Inc. 

---

Results

Video&Mask	Output

Dependencies

Create Conda Environment and Install Dependencies

# create new anaconda env
conda create -n undereraser python=3.12 -y
conda activate undereraser

# install python dependencies
pip3 install -r requirements.txt

Get Started

Pre-trained models

Download the weights from this link. Put the two files under the folder weight.

We use pretrained Wan2.1-Fun-V1.1-14B-InP as our base model. You can download the Wan2.1-Fun-14B-InP base model from this link. Put the whole folder under the folder models.

The models will be arranged like this:

models
 ├── Wan2.1-Fun-V1.1-14B-InP
   ├── google
     ├── umt5-xxl
       ├── spiece.model
           ...
   ├── xlm-roberta-large
     ├── sentencepiece.bpe.model
         ...
   ├── config.json
   ├── configuration.json
   ├── diffusion_pytorch_model.safetensors
   ├── models_clip_open-clip-xlm-roberta-large-vit-huge-14.pth
   ├── models_t5_umt5-xxl-enc-bf16.pth
   ├── Wan2.1_VAE.pth

Inference

We provide some examples in the data folder. Run the following commands to try it out:

python infer.py 

You can also prepare and test your own data following the same format.

Dataset

The test datasets are available at this link, including our constructed Camera-Bench and Scene-Bench.

TODO

• Release training code.

Citation

If you find our repo useful for your research, please consider citing our paper:

@article{liu2026eraser,
   title={From Understanding to Erasing: Towards Complete and Stable Video Object Removal}, 
   author={Liu, Dingming and Wang, Wenjing and Li, Chen and LYU, Jing},
   journal={arXiv preprint},
   year={2026}
}

Acknowledgement

This code is based on VideoX-Fun and LightX2V. Thanks for their awesome works！