4D-Animal: Freely Reconstructing Animatable 3D Animals from Videos

| ArXiv | Code |

This repository contains the official implementation of the paper:
4D-Animal: Freely Reconstructing Animatable 3D Animals from Videos

Authors:
Shanshan Zhong, Jiawei Peng, Zehan Zheng, Zhongzhan Huang, Wufei Ma, Guofeng Zhang, Qihao Liu, Alan Yuille, Jieneng Chen

News

• 2025-11-11: Our paper has been accepted to WACV 2026 🎉🎉
• 2025-07-15: Initial code release 🎉

Prepraration

Clone the repository:

git clone https://github.com/zhongshsh/4D-Animal
cd 4D-Animal

Create and activate the conda environment:

conda create -n animal python=3.10 -y
conda activate animal

Install required dependencies:

pip uninstall torch torchvision torchaudio
pip install torch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 --index-url https://download.pytorch.org/whl/cu121

pip install git+https://github.com/facebookresearch/detectron2@main#subdirectory=projects/DensePose

pip install cogapp triton plotly
git clone https://github.com/facebookresearch/lightplane
cd lightplane
pip install -e .

pip install --no-index --no-cache-dir pytorch3d -f https://dl.fbaipublicfiles.com/pytorch3d/packaging/wheels/py310_cu121_pyt241/download.html

pip install pandas sqlalchemy plotly hydra-core tensorboard lpips opencv-python imageio[ffmpeg] pyyaml Pillow natsort

Download the external_data.tar.gz and unzip the file.

wget --header "Authorization: Bearer your_hf_token" https://huggingface.co/datasets/zhongshsh/4D-Animal/resolve/main/external_data.tar.gz 
tar -xzvf external_data.tar.gz

💡 Replace your_hf_token with your HuggingFace token from https://huggingface.co/settings/tokens.

Optimize a CoP3D scene

To optimize a CoP3D scene (e.g., 1030_23106_17099) and save results in the experiments directory:

python main_optimize_scene.py 'exp.sequence_index="1030_23106_17099"' 'exp.experiment_folder="experiments"'

Hyperparameters for reconstruction can be modified in config/config.yaml.

Visualize Reconstruction

To visualize reconstruction of a trained model:

python main_visualize_reconstruction.py --archive_path experiments/1030_23106_17099

Acknowledgement

We would like to express our sincere gratitude to the authors of Animal Avatar for their well-structured and inspiring codebase, which served as a valuable reference for our implementation.

We also thank the developers of the following projects DINO, CSE, PartGLEE, and BootsTAP for contributing such impressive models to our community.

Citation

If you find our work useful, please consider citing:

@article{zhong20254danimal,
  title={4D-Animal: Freely Reconstructing Animatable 3D Animals from Videos},
  author={Zhong, Shanshan and Peng, Jiawei and Zheng, Zehan and Huang, Zhongzhan and Ma, Wufei and Zhang, Guofeng and Liu, Qihao and Yuille, Alan and Chen, Jieneng},
  journal={ArXiv},
  year={2025},
  volume={abs/2507.10437},
}