• Real-time Audio-Driven, latency less than 1 sencond
• Generalize pretty well for chinese and other languages
• Generalize pretty well for different metahuman character

Create conda environment

conda create -n talking_face python=3.9.18
conda activate talking_face
pip install torch==1.12.1+cu113 torchvision==0.13.1+cu113 torchaudio==0.12.1 --extra-index-url https://download.pytorch.org/whl/cu113
pip install -r requirements.txt

Due to the license issue of VOCASET, we cannot distribute BlendVOCA directly. Instead, you can preprocess data/blendshape_residuals.pickle after constructing BlendVOCA directory as follows for the simple execution of the script.

mkdir BlendVOCA
BlendVOCA
   └─ templates
      ├─ ...
      └─ FaceTalk_170915_00223_TA.ply

• templates: Download the template meshes from VOCASET.

python preprocess_blendvoca.py

If you want to generate coefficients by yourself, we recommend constructing the BlendVOCA directory as follows for the simple execution of the script.

BlendVOCA
  ├─ blendshapes_head
  │  ├─ ...
  │  └─ FaceTalk_170915_00223_TA
  │     ├─ ...
  │     └─ noseSneerRight.obj
  ├─ templates_head
  │  ├─ ...
  │  └─ FaceTalk_170915_00223_TA.obj
  └─ unposedcleaneddata
     ├─ ...
     └─ FaceTalk_170915_00223_TA
        ├─ ...
        └─ sentence40

• blendshapes_head: Place the constructed blendshape meshes (head).
• templates_head: Place the template meshes (head).
• unposedcleaneddata: Download the mesh sequences (unposed cleaned data) from VOCASET.

And then, run the following command:

python optimize_blendshape_coeffs.py

This step will take about 2 hours.

We recommend constructing the BlendVOCA directory as follows for the simple execution of scripts.

BlendVOCA
  ├─ audio
  │  ├─ ...
  │  └─ FaceTalk_170915_00223_TA
  │     ├─ ...
  │     └─ sentence40.wav
  ├─ bs_npy
  │  ├─ ...
  │  └─ FaceTalk_170915_00223_TA01.npy
  │    
  ├─ blendshapes_head
  │  ├─ ...
  │  └─ FaceTalk_170915_00223_TA
  │     ├─ ...
  │     └─ noseSneerRight.obj
  └─ templates_head
     ├─ ...
     └─ FaceTalk_170915_00223_TA.obj

• audio: Download the audio from VOCASET.
• bs_npy: Place the constructed blendshape coefficients.
• blendshapes_head: Place the constructed blendshape meshes (head).
• templates_head: Place the template meshes (head).

 python main.py

1. Prepare Unreal Engine5(test on UE5.1 and UE5.3) metahuman project

   • Create default metahuman project in UE5
   • Move jsonlivelink plugin into the Plugins of UE5 Animation
   • Revise the blueprint of the face animation to cancel the default animation and rebuild
   • Start jsonlivelink
   • Run the level

2. Start the audio2face server, you can train and check your model under BlendVOCA, or download the model here:

   python audio2face_server.py --model_name save_512_xx_xx_xx_xx/100_model

3. Drive the metahuman Unreal Engine:

   cd metahuman_demo
   python demo.py --audio2face_url http://0.0.0.0:8000 --wav_path ../test/wav/speech_long.wav --livelink_host 0.0.0.0 --livelink_port 1234

Since I deploy the metahuman project on my windows PC, so the livelink_host should be my PC's IP.

• SAiD

@misc{park2023said,
      title={SAiD: Speech-driven Blendshape Facial Animation with Diffusion},
      author={Inkyu Park and Jaewoong Cho},
      year={2023},
      eprint={2401.08655},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

• SelfTalk_Release

  @inproceedings{peng2023selftalk,
    title={SelfTalk: A Self-Supervised Commutative Training Diagram to Comprehend 3D Talking Faces}, 
    author={Ziqiao Peng and Yihao Luo and Yue Shi and Hao Xu and Xiangyu Zhu and Hongyan Liu and Jun He and Zhaoxin Fan},
    journal={arXiv preprint arXiv:2306.10799},
    year={2023}
  }