/GeneFacePlusPlus

GeneFace++: Generalized and Stable Real-Time 3D Talking Face Generation; Official Code

Primary LanguagePython

GeneFace++: Generalized and Stable Real-Time 3D Talking Face Generation

arXiv| GitHub Stars | 中文文档

This is the official implementation of GeneFace++ Paper with Pytorch, which enables high lip-sync, high video-reality and high system-efficiency 3D talking face generation. You can visit our Demo Page to watch demo videos and learn more details.



You may also interested in

Quick Start!

We provide a guide for a quick start in GeneFace++.

  • Step 1: Follow the steps in docs/prepare_env/install_guide.md, create a new python environment named geneface, and download 3DMM files into deep_3drecib/BFM.

  • Step 2: Download pre-processed dataset of May(Google Drive or BaiduYun Disk with password 98n4), and place it here data/binary/videos/May/trainval_dataset.npy

  • Step 3: Download pre-trained audio-to-motino model audio2motion_vae.zip (Google Drive or BaiduYun Disk with password 9cqp) and motion-to-video checkpoint motion2video_nerf.zip, which is specific to May (in this Google Drive or in thisBaiduYun Disk with password 98n4), and unzip them to ./checkpoints/

After these steps，your directories checkpoints and data should be like this：

> checkpoints
    > audio2motion_vae
    > motion2video_nerf
        > may_head
        > may_torso
> data
    > binary
        > videos
            > May
                trainval_dataset.npy
  • Step 4: activate geneface Python environment, and execute:
export PYTHONPATH=./
python inference/genefacepp_infer.py --a2m_ckpt=checkpoints/audio2motion_vae --head_ckpt= --torso_ckpt=checkpoints/motion2video_nerf/may_torso --drv_aud=data/raw/val_wavs/MacronSpeech.wav --out_name=may_demo.mp4

Or you can play with our Gradio WebUI:

export PYTHONPATH=./
python inference/app_genefacepp.py --a2m_ckpt=checkpoints/audio2motion_vae --head_ckpt= --torso_ckpt=checkpoints/motion2video_nerf/may_torso

Train GeneFace++ with your own videos

Please refer to details in docs/process_data and docs/train_and_infer.

Below are answers to frequently asked questions when training GeneFace++ on custom videos:

  • Please make sure that the head segment occupies a relatively large region in the video (e.g., similar to the provided May.mp4). Or you need to hand-crop your training video. issue
  • Make sure that the talking person appears in every frame of the video, otherwise the data preprocessing pipeline may be failed.
  • We only tested our code on Liunx (Ubuntu/CentOS). It is welcome that someone who are willing to share their installation guide on Windows/MacOS.

ToDo

  • Release Inference Code of Audio2Motion and Motion2Video.
  • Release Pre-trained weights of Audio2Motion and Motion2Video.
  • Release Training Code of Motino2Video Renderer.
  • Release Gradio Demo.
  • **Release Training Code of Audio2Motion and Post-Net. (Maybe 2024.06.01) **

Citation

If you found this repo helpful to your work, please consider cite us:

@article{ye2023geneface,
  title={GeneFace: Generalized and High-Fidelity Audio-Driven 3D Talking Face Synthesis},
  author={Ye, Zhenhui and Jiang, Ziyue and Ren, Yi and Liu, Jinglin and He, Jinzheng and Zhao, Zhou},
  journal={arXiv preprint arXiv:2301.13430},
  year={2023}
}
@article{ye2023geneface++,
  title={GeneFace++: Generalized and Stable Real-Time Audio-Driven 3D Talking Face Generation},
  author={Ye, Zhenhui and He, Jinzheng and Jiang, Ziyue and Huang, Rongjie and Huang, Jiawei and Liu, Jinglin and Ren, Yi and Yin, Xiang and Ma, Zejun and Zhao, Zhou},
  journal={arXiv preprint arXiv:2305.00787},
  year={2023}
}