DiffPoseTalk: Speech-Driven Stylistic 3D Facial Animation and Head Pose Generation via Diffusion Models

• Release the TFHP dataset
  • Release the processed lmdb dataset
  • Release the raw dataset
• Release the training and inference code
• Release the pretrained models
• Release the data processing code

• Download FLAME-related files following this instruction and setup the enviornments. You can easily do this by running the following commands:

  bash ./setup/fetch_data.sh
  bash ./install_conda.sh

• (Optional) If you want to train with the HDTF_TFHP, please follow this instruction to download the processed dataset.

We provide two sets of pretrained models for inference. One predicts head motion and the other does not. The DiffPoseTalk system takes a speech clip, a template face shape parameter, and a style feature as input, and outputs diverse and stylistic lip-synced animations.

The style encoder can extract a style feature from an arbitray four-second motion sequence:

python extract_style.py --exp_name <STYLE_ENC_NAME> --iter <STYLE_ENC_ITER> -c <FLAME_MOTION_SEQ> -o <OUTPUT_NAME> -s <STARTING_FRAME>

Note that the <FLAME_MOTION_SEQ> should be a .npz file that has the exp and pose keys. The extracted style feature will be saved under the corresponding folder (<STYLE_ENC_NAME>/<STYLE_ENC_ITER>) under demo/input/style.

We have also provided some feature examples for the pretrained models.

python demo.py --exp_name <DENOISING_NETWORK_NAME> --iter <DENOISING_NETWORK_ITER> -a <AUDIO> -c <SHAPE_TEMPLATE> -s <STYLE_FEATURE> -o <OUTPUT>.mp4 -n <N_REPITIONS> -ss <CFG_SCALE_FOR_STYLE> -sa <CFG_SCALE_FOR_AUDIO>

The <SHAPE_TEMPLATE> should be a .npy file containing a frame of shape parameter. The <STYLE_ENC_NAME>/<STYLE_ENC_ITER> in <STYLE_FEATURE>'s path can be omitted.

You can also pass --dtr 0.99 to enable dynamic thresholding to obtain results with better quality but lower diversity.

Here are some examples:

python demo.py --exp_name head-SA-hubert-WM --iter 110000 -a demo/input/audio/FAST.flac -c demo/input/coef/TH217.npy demo/input/style/TH217.npy -o TH217-FAST-TH217.mp4 -n 3 -ss 3 -sa 1.15 -dtr 0.99
python demo.py --exp_name SA-hubert-WM --iter 100000 -a demo/input/audio/further_divide_our.flac -c demo/input/coef/TH050.npy -s demo/input/style/normal.npy -o TH050-further-normal.mp4 -n 3 -ss 3 -sa 1.15

Please note:

• The DiffPoseTalk system consists of a style encoder and a denoising network. You will need to train them one by one.
• To optimize I/O performance, we use lmdb to pack and handle the training data. Read "Preparing Your Dataset" to see how to generate your own lmdb dataset.

python main_se.py --exp_name <STYLE_ENC_NAME> --data_root <DATA_ROOT> [--no_head_pose]

The style encoder will be saved under the experiments/SE/<STYLE_ENC_NAME> folder. View the validation results in the TensorBoard to select the best model.

You will need to specify the path to the style encoder checkpoint using the --style_enc_ckpt argument. You can also experiment with different argument values and combinations.

python main_dpt.py --exp_name <DENOISING_NETWORK_NAME> --data_root <DATA_ROOT> --use_indicator --scheduler Warmup --audio_model hubert --style_enc_ckpt <PATH_TO_STYLE_ENC_CKPT> [--no_head_pose]

The denoising network will be saved under the experiments/DPT/<exp_name> folder.

[TODO]

@article{sun2024diffposetalk,
  title={DiffPoseTalk: Speech-Driven Stylistic 3D Facial Animation and Head Pose Generation via Diffusion Models},
  author={Sun, Zhiyao and Lv, Tian and Ye, Sheng and Lin, Matthieu and Sheng, Jenny and Wen, Yu-Hui and Yu, Minjing and Liu, Yong-Jin},
  doi={10.1145/3658221},
  journal={ACM Transactions on Graphics (TOG)},
  volume={43},
  number={4},
  articleno={46},
  numpages={9},
  year={2024},
  publisher={ACM New York, NY, USA}
}