preliminary
To make photos "dance" you need to prepare two core materials: a still portrait photo and a driving video of the target action.
Detailed steps
- Environment deployment
- Clone the project repository: git clone https://github.com/bytedance/X-Dyna.git
- Install dependencies: pip install -r requirements.txt
- Configure PyTorch 2.0 environment: bash env_torch2_install.sh
- Prepare material
- Select a clear front facing portrait photo (.jpg/.png)
- Prepare a reference video containing standard dance moves (.mp4)
- Execution generation
- Run command: python inference_xdyna.py -input_image photo path -driving_video video path
- You can adjust the motion smoothness with the -num_mix parameter.
Optimization Tips
When encountering incoherent movements, it is recommended to 1) increase ddim_steps to 50-100, 2) use LCM LoRA to accelerate the model, and 3) make sure that the movements in the drive video are complete and coherent. The project also recommends using the best_frame parameter to select the most suitable start frame.
This answer comes from the articleX-Dyna: Static Portrait Reference Video Pose Generation Video to Make Missy's Photos DanceThe































