Home

Awesome

PD-FGC: Progressive Disentangled Representation Learning for Fine-Grained Controllable Talking Head Synthesis (CVPR 2023)

Duomin Wang, Yu Deng, Zixin Yin, Heung-Yeung Shum, and Baoyuan Wang.

Project | Paper | Demo

<img src='./data/misc/teaser.png' width=1000>

We present a novel one-shot talking head synthesis method that achieves disentangled and fine-grained control over lip motion, eye gaze&blink, head pose, and emotional expression. We represent different motions via disentangled latent representations and leverage an image generator to synthesize talking heads from them. To effectively disentangle each motion factor, we propose a progressive disentangled representation learning strategy by separating the factors in a coarse-to-fine manner, where we first extract unified motion feature from the driving signal, and then isolate each fine-grained motion from the unified feature. We introduce motion-specific contrastive learning and regressing for non-emotional motions, and feature-level decorrelation and self-reconstruction for emotional expression, to fully utilize the inherent properties of each motion factor in unstructured video data to achieve disentanglement. Experiments show that our method provides high quality speech&lip-motion synchronization along with precise and disentangled control over multiple extra facial motions, which can hardly be achieved by previous methods.

<img src='./data/misc/overview.png' width=1000>

Requirements

pip install -r requirements.txt

Inference

tar -xzf checkpoints.tar

in . directory. There should be 4 models in checkpoints directory, including appearance_model.pth, motion_mode.pth, audio_model.pth and generator_model.pth.

bash scripts/inference.sh

*There are 5 arguments for input,

  1. --audio_path specifies input audio file path;
  2. --app_img_path specifies input appearance image path;
  3. --pose_path specifies input pose control image dir;
  4. --exp_path specifies input emotional expression control image dir;
  5. --eye_path specifies input eye motion control image dir.

The generated results will be save at test folder <img src='./data/misc/test1.gif' width=800/> <img src='./data/misc/test2.gif' width=800/> <img src='./data/misc/test3.gif' width=800/> <img src='./data/misc/test4.gif' width=800/>

From left to right are the appearance, the generated result, the pose motion video, the exp motion video and the eye motion video.

You can specify any video as motion control source from preprocessed Voxceleb2 dataset. For you own image and video, you can take PC-AVS's pre-processing as reference.

Disentangled control

  1. --mouth specifies lip motion control is enabled;
  2. --blink specifies blink control is enabled;
  3. --gaze specifies gaze control is enabled;
  4. --emo specifies emotional expression control is enabled;
  5. --headpose specifies headpose control is enabled;
  6. --all specifies all control factors are enabled.

License and Citation

The usage of this software is under CC-BY-4.0.

@inproceedings{wang2022pdfgc,
    title={Progressive Disentangled Representation Learning for Fine-Grained Controllable Talking Head Synthesis},
    author={Wang, Duomin and Deng, Yu and Yin, Zixin and Shum, Heung-Yeung and Wang, Baoyuan},
    booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    year={2023}
}

Acknowledgement