CLIP-Actor (ECCV 2022)

Project Page | Paper

This repository contains a pytorch implementation for the ECCV 2022 paper, CLIP-Actor: Text-Driven Recommendation and Stylization for Animating Human Meshes. CLIP-Actor is a novel text-driven motion recommendation and neural mesh stylization system for human mesh animation.

https://user-images.githubusercontent.com/55628873/173112861-93e07ac6-5303-44db-940c-68b75a947085.mp4

Getting Started

This code was developed on Ubuntu 18.04 with Python 3.7, CUDA 10.2 and PyTorch 1.9.0. Later versions should work, but have not been tested.

System Requirements

• Python 3.7
• CUDA 10.2
• Single GPU w/ minimum 24 GB RAM

Environment Setup

Create and activate a virtual environment to work in, e.g. using Conda:

conda create -n clip_actor python=3.7
conda activate clip_actor

Install PyTorch and PyTorch3D. For CUDA 10.2, this would look like:

conda install -c pytorch pytorch=1.9.0 torchvision=0.10.0 cudatoolkit=10.2
conda install -c fvcore -c iopath -c conda-forge fvcore iopath
conda install pytorch3d -c pytorch3d

Install the remaining requirements with pip:

pip install -r requirements.txt

You must also have ffmpeg installed on your system to save visualizations.

Download Body Models and Datasets

To run CLIP-Actor, you need to download relevant body models and datasets.

Check DOWNLOAD.md for details.

Running CLIP-Actor

Run below commands to generate whatever stylized 4D human avatar you want!

python clip_actor.py --prompt "a scuba diver is scuba diving" --exp_name scuba_diving
python clip_actor.py --prompt "Freddie Mercury is dancing" --exp_name mercury_dancing

The outputs will be the final .mp4 video, stylized .obj files, colored render views, and screenshots during training.

Citation

If you find our code or paper helps, please consider citing:

@inproceedings{youwang2022clipactor,
      title={CLIP-Actor: Text-Driven Recommendation and Stylization for Animating Human Meshes},
      author={Kim Youwang and Kim Ji-Yeon and Tae-Hyun Oh},
      year={2022},
      booktitle={ECCV}
}

Acknowledgement

This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No.2022-00164860, Development of Human Digital Twin Technology Based on Dynamic Behavior Modeling and Human-Object-Space Interaction; and No.2021-0-02068, Artificial Intelligence Innovation Hub).

The implementation of CLIP-Actor is largely inspired and fine-tuned from the seminal prior work, Text2Mesh (Michael et al.). We thank the authors of Text2Mesh who made their code public. Also If you find our work helpful, please consider citing them as well.