AD-NeRF: Audio Driven Neural Radiance Fields for Talking Head Synthesis

PyTorch implementation for the paper "AD-NeRF: Audio Driven Neural Radiance Fields for Talking Head Synthesis (ICCV 2021)".
Authors: Yudong Guo, Keyu Chen, Sen Liang, Yong-Jin Liu, Hujun Bao and Juyong Zhang.

Prerequisites

• You can create an anaconda environment called adnerf with:

  conda env create -f environment.yml
  conda activate adnerf
  

• PyTorch3D

  Recommend install from a local clone

  git clone https://github.com/facebookresearch/pytorch3d.git
  cd pytorch3d && pip install -e .
  

• Basel Face Model 2009

  Put "01_MorphableModel.mat" to data_util/face_tracking/3DMM/; cd data_util/face_tracking; run

  python convert_BFM.py
  

Train AD-NeRF

• Data Preprocess ($id Obama for example)

  bash process_data.sh Obama
  

  • Input: A portrait video at 25fps containing voice audio. (dataset/vids/$id.mp4)
  • Output: folder dataset/$id that contains all files for training

• Train Two NeRFs (Head-NeRF and Torso-NeRF)

  • Train Head-NeRF with command

    python NeRFs/HeadNeRF/run_nerf.py --config dataset/$id/HeadNeRF_config.txt
    

  • Copy latest trainied model from dataset/$id/logs/$id_head to dataset/$id/logs/$id_com
  • Train Torso-NeRF with command

    python NeRFs/TorsoNeRF/run_nerf.py --config dataset/$id/TorsoNeRF_config.txt
    

  • You may need the pretrained models to avoid bad initialization. #3

Run AD-NeRF for rendering

• Reconstruct original video with audio input

  python NeRFs/TorsoNeRF/run_nerf.py --config dataset/$id/TorsoNeRFTest_config.txt --aud_file=dataset/$id/aud.npy --test_size=300
  

• Drive the target person with another audio input

  python NeRFs/TorsoNeRF/run_nerf.py --config dataset/$id/TorsoNeRFTest_config.txt --aud_file=${deepspeechfile.npy} --test_size=-1
  

Citation

If you find our work useful in your research, please consider citing our paper:

@inproceedings{guo2021adnerf,
  title={AD-NeRF: Audio Driven Neural Radiance Fields for Talking Head Synthesis},
  author={Yudong Guo and Keyu Chen and Sen Liang and Yongjin Liu and Hujun Bao and Juyong Zhang},
  booktitle={IEEE/CVF International Conference on Computer Vision (ICCV)},
  year={2021}
}

If you have questions, feel free to contact gyd2011@mail.ustc.edu.cn.

Acknowledgments

We use face-parsing.PyTorch for parsing head and torso maps, and DeepSpeech for audio feature extraction. The NeRF model is implemented based on NeRF-pytorch.