SynTalker: Enabling Synergistic Full-Body Control in Prompt-Based Co-Speech Motion Generation

Project Page • Arxiv Paper • Demo Video • Web Gradio Demo • Colab • Citation

📝 Release Plans

• A simple and powerful cospeech model (corespond to paper Table2:SynTalker (w/o both) )
• Training scripts (include training rvqvae and diffusion)
• A web demo (We strongly suggest you to try it!)
• Our syntalker can recieve both speech and text input simultaneously
• Training scripts (include data preprocessing, training rvqvae, text-motion alignspace and diffusion)

💖 Online Demo

Thank Hugging Face🤗 for providing us GPU! Feel free to exprience our online web demo!

⚒️ Installation

Build Environtment

conda create -n syntalker python=3.12
conda activate syntalker
pip install -r requirements.txt
bash demo/install_mfa.sh

Download Model

gdown https://drive.google.com/drive/folders/1tGTB40jF7v0RBXYU-VGRDsDOZp__Gd0_?usp=drive_link -O ./ckpt --folder
gdown https://drive.google.com/drive/folders/1MCks7CMNBtAzU2XihYezNmiGT_6pWex8?usp=drive_link -O ./datasets/hub --folder

Download Dataset

For evaluation and training, not necessary for running a web demo or inference.

• Download the original raw data

bash preprocess/bash_raw_cospeech_download.sh

🚩 Running

Run a web demo

python demo.py -c ./configs/diffusion_rvqvae_128_hf.yaml

Notice: If you use ssh to conect and run code in a headless computer, you may encounter an error pyglet.canvas.xlib.NoSuchDisplayException: Cannot connect to "None". Here, we recommend a method to solve it.

sudo apt-get install libegl1-mesa-dev libgles2-mesa-dev
PYOPENGL_PLATFORM='egl' python demo.py -c ./configs/diffusion_rvqvae_128_hf.yaml

Eval

Require download dataset

python test.py -c configs/diffusion_rvqvae_128.yaml

We also provide a colab notebook for you to evaluate it.

📺 Visualize

Following EMAGE, you can download SMPLX blender addon, and install it in your blender 3.x or 4.x. Click the button Add Animation to visualize the generated smplx file (like xxx.npz).

🔥 Training from scratch

1. Train RVQVAE

Well, if your multiple gpus, we can parellel run these three commands.

python rvq_beatx_train.py --batch-size 256 --lr 2e-4 --total-iter 300000 --lr-scheduler 200000 --nb-code 512 --code-dim 512 --down-t 2 --depth 3 --dilation-growth-rate 3 --out-dir outputs/rvqvae --vq-act relu --quantizer ema_reset --loss-vel 0.5 --recons-loss l1_smooth --exp-name RVQVAE --body_part upper

python rvq_beatx_train.py --batch-size 256 --lr 2e-4 --total-iter 300000 --lr-scheduler 200000 --nb-code 512 --code-dim 512 --down-t 2 --depth 3 --dilation-growth-rate 3 --out-dir outputs/rvqvae --vq-act relu --quantizer ema_reset --loss-vel 0.5 --recons-loss l1_smooth --exp-name RVQVAE --body_part hands

python rvq_beatx_train.py --batch-size 256 --lr 2e-4 --total-iter 300000 --lr-scheduler 200000 --nb-code 512 --code-dim 512 --down-t 2 --depth 3 --dilation-growth-rate 3 --out-dir outputs/rvqvae --vq-act relu --quantizer ema_reset --loss-vel 0.5 --recons-loss l1_smooth --exp-name RVQVAE --body_part lower_trans

2. Train Diffusion Model

python train.py -c configs/diffusion_rvqvae_128.yaml

💭 Enabling Text Control

Download models and datasets

python preprocess/download_hf.py

Generate

It will generate motions using the prompt you input and the speech audio from test datasets

python test_h3d.py -c configs/diffusion_h3d.yaml --upper_prompt "your prompt" --lower_prompt "your prompt"

For example:

python test_h3d.py -c configs/diffusion_h3d.yaml --upper_prompt "A person raises up right hand" --lower_prompt "A man is kneel down"

The path of the generated motion npy file, the rendered video file, and the audio used, will be automatically printed in the console.

Training (TO DO)

1. Datapreprocess

We need to download AMASS datasets(both smlph and smplx format) from https://amass.is.tue.mpg.de/ into ./datasets.

We also need to download BEATX datasets as same as previous.

bash preprocess/bash_raw_cospeech_download.sh

2. FPS Correction

The FPS for all motions in AMASS SMPLX is currently set to 120, which is incorrect and needs to be adjusted. Meanwhile, the FPS for SMPLH is correct, so we need to align all SMPLX FPS values with the corresponding SMPLH FPS values.

Following SMPLX_FPS_Correction.ipynb to correct it.

🙏 Acknowledgments

Thanks to EMAGE, DiffuseStyleGesture, MDM, T2M-GPT, MoMask, MotionCLIP, TMR, OpenTMA, HumanML3D, human_body_prior, our code is partially borrowing from them. Please check these useful repos.

📖 Citation

If you find our code or paper helps, please consider citing:

@inproceedings{chen2024syntalker,
  author = {Bohong Chen and Yumeng Li and Yao-Xiang Ding and Tianjia Shao and Kun Zhou},
  title = {Enabling Synergistic Full-Body Control in Prompt-Based Co-Speech Motion Generation},
  booktitle = {Proceedings of the 32nd ACM International Conference on Multimedia},
  year = {2024},
  publisher = {ACM},
  address = {New York, NY, USA},
  pages = {10},
  doi = {10.1145/3664647.3680847}
}