DiffuseStyleGesture: Stylized Audio-Driven Co-Speech Gesture Generation with Diffusion Models

| Demo | Presentation Video | Conference archive

Further Work

📢 QPGesture - Based on motion matching, the upper body gesture.

📢 UnifiedGesture - Training on multiple gesture datasets, refine the gestures.

News

📢 9/Oct/23 - We obtained the REPRODUCIBILITY AWARD by GENEA Committee, so we strongly recommend trying DiffuseStyleGesture+ in advance compared to code of DiffuseStyleGesture is partially optimized.

📢 29/Aug/23 - Release the paper of DiffuseStyleGesture+, refer to the official paper of GENEA Challenge 2023 to get more.

📢 5/Aug/23 - Release code and pre-trained models of DiffuseStyleGesture+ on BEAT and TWH.

📢 31/Jul/23 - Upload a tutorial video on visualizing gestures.

📢 25/Jun/23 - Upload presentation video.

📢 9/May/23 - First release - arxiv, demo, code, pre-trained models on ZEGGS and issue.

1. Getting started

This code was tested on NVIDIA GeForce RTX 2080 Ti and requires:

conda3 or miniconda3

conda create -n DiffuseStyleGesture python=3.7
conda activate DiffuseStyleGesture
pip install -r requirements.txt

2. Quick Start

Download pre-trained model from Tsinghua Cloud or Google Cloud and put it into ./main/mydiffusion_zeggs/.
Download the WavLM Large and put it into ./main/mydiffusion_zeggs/WavLM/.
cd ./main/mydiffusion_zeggs/ and run
```
python sample.py --config=./configs/DiffuseStyleGesture.yml --no_cuda 0 --gpu 0 --model_path './model000450000.pt' --audiowavlm_path "./015_Happy_4_x_1_0.wav" --max_len 320
```
You will get the .bvh file named yyyymmdd_hhmmss_smoothing_SG_minibatch_320_[1, 0, 0, 0, 0, 0]_123456.bvh in the sample_dir folder, which can then be visualized using Blender with the following result (To visualize bvh with Blender see this issue and this tutorial video):

https://github.com/YoungSeng/DiffuseStyleGesture/assets/37477030/2ef7aa70-69e0-4fd9-a551-6b8a5d075d17

The parameter no_cuda and gpu need to be the same, i.e. the GPU you want to use; max_len is the length you want to generate, this parameter should be 0 if you want to generate the whole length; if you want to use your own audio, you should rename your audio file name as xxx_style_xxx.wav, e.g. 000_Neutral_xxx.wav (Happy, Sad, ...). please refer to this issue to set the style and intensity you want.

3. Train your own model

(1) Get ZEGGS dataset

Same as ZEGGS.

An example is as follows. Download original ZEGGS datasets from here and put it in ./ubisoft-laforge-ZeroEGGS-main/data/ folder. Then cd ./ubisoft-laforge-ZeroEGGS-main/ZEGGS and run python data_pipeline.py to process the dataset. You will get ./ubisoft-laforge-ZeroEGGS-main/data/processed_v1/trimmed/train/ and ./ubisoft-laforge-ZeroEGGS-main/data/processed_v1/trimmed/test/ folders.

If you find it difficult to obtain and process the data, you can download the data after it has been processed by ZEGGS from Tsinghua Cloud or Baidu Cloud. And put it in ./ubisoft-laforge-ZeroEGGS-main/data/processed_v1/trimmed/ folder.

(2) Process ZEGGS dataset

cd ./main/mydiffusion_zeggs/
python zeggs_data_to_lmdb.py

(3) Train

python end2end.py --config=./configs/DiffuseStyleGesture.yml --no_cuda 0 --gpu 0

The model will save in ./main/mydiffusion_zeggs/zeggs_mymodel3_wavlm/ folder.

Reference

Our work mainly inspired by: MDM, Text2Gesture, Listen, denoise, action!

Citation

If you find this code useful in your research, please cite:

@inproceedings{ijcai2023p650,
  title     = {DiffuseStyleGesture: Stylized Audio-Driven Co-Speech Gesture Generation with Diffusion Models},
  author    = {Yang, Sicheng and Wu, Zhiyong and Li, Minglei and Zhang, Zhensong and Hao, Lei and Bao, Weihong and Cheng, Ming and Xiao, Long},
  booktitle = {Proceedings of the Thirty-Second International Joint Conference on
               Artificial Intelligence, {IJCAI-23}},
  publisher = {International Joint Conferences on Artificial Intelligence Organization},
  pages     = {5860--5868},
  year      = {2023},
  month     = {8},
  doi       = {10.24963/ijcai.2023/650},
  url       = {https://doi.org/10.24963/ijcai.2023/650},
}

Please feel free to contact us (yangsc21@mails.tsinghua.edu.cn) with any question or concerns.

YoungSeng / DiffuseStyleGesture

readme