Yuxue Yang^1,2, Lue Fan², Zuzeng Lin³, Feng Wang⁴, Zhaoxiang Zhang^1,2†

¹UCAS  ²CASIA  ³TJU  ⁴CreateAI  ^†Corresponding author

Official implementation of LayerAnimate: Layer-level Control for Animation, ICCV 2025

Videos on the project website vividly introduces our work and presents qualitative results for an enhanced view experience.

• [25-08-22] Release the Layer Curation Pipeline, including the demo and comprehensive usage guidance.
• [25-06-26] Our work is accepted by ICCV 2025! 🎉
• [25-05-29] We have extended LayerAnimate to the DiT (Wan2.1 1.3B) variant, enabling the generation of 81 frames at 480 × 832 resolution. It performs surprisingly well in the Real-World Domain shown in the project website.
• [25-03-31] Release the online demo on Hugging Face.
• [25-03-30] Release a gradio script app.py to run the demo locally. Please raise an issue if you encounter any problems.
• [25-03-22] Release the checkpoint and the inference script. We update layer curation pipeline and support trajectory control for a flexible composition of various layer-level controls.
• [25-01-15] Release the project page and the arXiv preprint.

We have released a comprehensive pipeline for extracting motion-based layers from video sequences. The layer curation pipeline automatically decomposes videos into different layers based on motion patterns, where you can control the number of extracted layers by adjusting the layer capacity parameter to obtain varying levels of motion granularity.

More details can be found in the repo.

git clone [email protected]:IamCreateAI/LayerAnimate.git
conda create -n layeranimate python=3.10 -y
conda activate layeranimate
pip install -r requirements.txt
pip install wan@git+https://github.com/Wan-Video/Wan2.1  # If you want to use DiT variant.

Download the pretrained weights and put them in checkpoints/ directory as follows:

checkpoints/
├─ LayerAnimate-Mix (UNet variant)
└─ LayerAnimate-DiT

Run the following command to generate a video from input images:

python scripts/animate_Layer.py --config scripts/demo1.yaml --savedir outputs/sample1

python scripts/animate_Layer.py --config scripts/demo2.yaml --savedir outputs/sample2

python scripts/animate_Layer.py --config scripts/demo3.yaml --savedir outputs/sample3

python scripts/animate_Layer.py --config scripts/demo4.yaml --savedir outputs/sample4

python scripts/animate_Layer.py --config scripts/demo5.yaml --savedir outputs/sample5

Note that the layer-level controls are prepared in __assets__/demos.

You can run the demo locally by executing the following command:

python scripts/app.py --savedir outputs/gradio

Then, open the link in your browser to access the demo interface. The output video and the video with trajectory will be saved in the outputs/gradio directory.

Run the following command to generate a video from input images:

python scripts/infer_DiT.py --config __assets__/demos/realworld/config.yaml --savedir outputs/realworld

We take the config.yaml in demos/realworld/ as an example. You can also modify the config file to suit your needs.

• Release the code and checkpoint of LayerAnimate.
• Upload a gradio script to run the demo locally.
• Create a online demo in the huggingface space.
• DiT-based LayerAnimate.
• Release layer curation pipeline.
• Training script for LayerAnimate.

We sincerely thank the great work ToonCrafter, LVCD, AniDoc, and Wan-Video for their inspiring work and contributions to the AIGC community.

Please consider citing our work as follows if it is helpful.

@article{yang2025layeranimate,
  author    = {Yang, Yuxue and Fan, Lue and Lin, Zuzeng and Wang, Feng and Zhang, Zhaoxiang},
  title     = {LayerAnimate: Layer-level Control for Animation},
  journal   = {arXiv preprint arXiv:2501.08295},
  year      = {2025},
}