TalkSHOW: Generating Holistic 3D Human Motion from Speech [CVPR2023]

The official PyTorch implementation of the CVPR2023 paper "Generating Holistic 3D Human Motion from Speech".

Please visit our webpage for more details.

HighLight

We directly provide the input and our output for the demo data, you can find them in /demo/ and /demo_audio/. TalkSHOW can generalize well on English, French, Songs so far. Looking forward to more demos.

You can directly use the generated motion to animate your 3D character or your own digital avatar. We will provide more demos, please stay tuned. And we are quite looking forward to your pull request.

Notes

We are using 100 dimension parameters for SMPL-X facial expression, if you need other dimensions parameters, you can use this code to convert.

https://github.com/yhw-yhw/SHOW/blob/main/cvt_exp_dim_tool.py

TODO

• 🤗Hugging Face Demo
• Animated 2D videos by the generated motion from TalkSHOW.

Getting started

The training code was tested on Ubuntu 18.04.5 LTS and the visualization code was test on Windows 10, and it requires:

• Python 3.7
• conda3 or miniconda3
• CUDA capable GPU (one is enough)

1. Setup environment

Clone the repo:

git clone https://github.com/yhw-yhw/TalkSHOW
cd TalkSHOW

Create conda environment:

conda create --name talkshow python=3.7
conda activate talkshow

Please install pytorch (v1.10.1).

pip install -r requirements.txt

Please install MPI-Mesh.

2. Get data

Please note that if you only want to generate demo videos, you can skip this step and directly download the pretrained models.

Download SHOW_dataset_v1.0.zip from TalkSHOW download webpage, unzip using for i in $(ls *.tar.gz);do tar xvf $i;done.

Run python data_utils/dataset_preprocess.py to check and split dataset. Modify data_root in config/*.json to the dataset-path.

Modify data_root in data_utils/apply_split.py to the dataset path and run it to apply data_utils/split_more_than_2s.pkl to the dataset.

We will update the benchmark soon.

3. Download the pretrained models (Optional)

Download pretrained models, unzip and place it in the TalkSHOW folder, i.e. path-to-TalkSHOW/experiments.

4. Training

Please note that the process of loading data for the first time can be quite slow. If you have already completed the loading process, setting dataset_load_mode to pickle in config/[config_name].json will make the loading process much faster.

# 1. Train VQ-VAEs. 
bash train_body_vq.sh
# 2. Train PixelCNN. Please modify "Model:vq_path" in config/body_pixel.json to the path of VQ-VAEs.
bash train_body_pixel.sh
# 3. Train face generator.
bash train_face.sh

5. Testing

Modify the arguments in test_face.sh and test_body.sh. Then

bash test_face.sh
bash test_body.sh

5. Visualization

If you ssh into the linux machine, NotImplementedError might occur. In this case, please refer to issue for solving the error.

Download smplx model (Please register in the official SMPLX webpage before you use it.) and place it in path-to-TalkSHOW/visualise/smplx_model. To visualise the test set and generated result (in each video, left: generated result | right: ground truth). The videos and generated motion data are saved in ./visualise/video/body-pixel:

bash visualise.sh

If you ssh into the linux machine, there might be an error about OffscreenRenderer. In this case, please refer to issue for solving the error.

To reproduce the demo videos, run

# the whole body demo
python scripts/demo.py --config_file ./config/body_pixel.json --infer --audio_file ./demo_audio/1st-page.wav --id 0 --whole_body
# the face demo
python scripts/demo.py --config_file ./config/body_pixel.json --infer --audio_file ./demo_audio/style.wav --id 0 --only_face
# the identity-specific demo
python scripts/demo.py --config_file ./config/body_pixel.json --infer --audio_file ./demo_audio/style.wav --id 0
python scripts/demo.py --config_file ./config/body_pixel.json --infer --audio_file ./demo_audio/style.wav --id 1
python scripts/demo.py --config_file ./config/body_pixel.json --infer --audio_file ./demo_audio/style.wav --id 2
python scripts/demo.py --config_file ./config/body_pixel.json --infer --audio_file ./demo_audio/style.wav --id 3 --stand
# the diversity demo
python scripts/demo.py --config_file ./config/body_pixel.json --infer --audio_file ./demo_audio/style.wav --id 0 --num_samples 12
# the french demo
python scripts/demo.py --config_file ./config/body_pixel.json --infer --audio_file ./demo_audio/french.wav --id 0
# the synthetic speech demo
python scripts/demo.py --config_file ./config/body_pixel.json --infer --audio_file ./demo_audio/rich.wav --id 0
# the song demo
python scripts/demo.py --config_file ./config/body_pixel.json --infer --audio_file ./demo_audio/song.wav --id 0

6. Baseline

For training the reproducted "Learning Speech-driven 3D Conversational Gestures from Video" (Habibie et al.), you could run

python -W ignore scripts/train.py --speakers oliver seth conan chemistry --config_file ./config/LS3DCG.json

For visualization with the pretrained model, download the above pretrained models and run

python scripts/demo.py --config_file ./config/LS3DCG.json --infer --audio_file ./demo_audio/style.wav --body_model_name s2g_LS3DCG --body_model_path experiments/2022-10-19-smplx_S2G-LS3DCG/ckpt-99.pth --id 0

Citation

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

@inproceedings{yi2022generating,
  title={Generating Holistic 3D Human Motion from Speech},
  author={Yi, Hongwei and Liang, Hualin and Liu, Yifei and Cao, Qiong and Wen, Yandong and Bolkart, Timo and Tao, Dacheng and Black, Michael J},
  booktitle={CVPR},
  year={2023}
}

Acknowledgements

For functions or scripts that are based on external sources, we acknowledge the origin individually in each file.
Here are some great resources we benefit:

• Freeform for training pipeline
• MPI-Mesh, Pyrender, Smplx, VOCA for rendering
• Wav2Vec2 and Faceformer for audio encoder

Contact

For questions, please contact [email protected] or [email protected] or [email protected] or [email protected]

For commercial licensing, please contact [email protected]