SadTalker

SadTalker is an open-source research project that turns a single portrait photo plus an audio file into a realistic talking-head video. The recipe from its README is simple: one portrait image plus audio equals a talking head video. It learns 3D motion coefficients from the audio and applies them to the face, producing natural speech and head movement. The project was published at CVPR 2023 by researchers from Xi'an Jiaotong University, Tencent AI Lab, and Ant Group. The code is Apache 2.0 licensed and runs fully offline.

Key Features

1. Audio-Driven Face Animation

Feed it a portrait and an audio file through the command line. It outputs an MP4 of the face speaking with synced lips and motion.

2. Full-Body Video Mode

The --still flag with full preprocessing generates natural full-body video from a single image, not just a talking head.

3. Face Enhancement Pipeline

An optional GFPGAN enhancer sharpens faces in the output. An offline patch avoids model downloads at generation time.

4. Higher-Resolution Face Model

A 512x512 face model released in beta sits alongside the default 256 model. The changelog notes clear performance improvements.

5. Local Web Interface

A Gradio WebUI runs locally on Windows, Linux, and Mac. An official extension also plugs into Stable Diffusion WebUI.

6. Free Online Demos

Try it without installing through the Hugging Face Spaces demo, Google Colab notebooks, or the project's Discord server.

Frequently Asked Questions (FAQ)

What do I need to install SadTalker?

Anaconda, Python 3.8, PyTorch, and ffmpeg, plus checkpoints from the releases page. Setup guides cover Linux, Windows, macOS, and Docker.

Can I use SadTalker commercially?

The code is Apache 2.0, but bundled model weights like Wav2Lip carry research or non-commercial licenses. Check every model license before commercial use.

Is this an official Tencent product?

No. The README states clearly that it is not an official Tencent product, even though Tencent AI Lab researchers co-authored it.

Does it collect my data?

No. The code runs completely offline on your machine and does not collect any personal information.

How popular is the project?

Very: roughly 14,000 GitHub stars and 2,700 forks, with community demos across Bilibili, YouTube, and X.

Conclusion

SadTalker is a landmark open-source talking-head project that still holds up for research and experimentation. The offline code, free demos, and strong community make it easy to try today. Watch the model-weight licenses if you plan commercial work, since some restrict it. For animating portraits with audio, it remains a top open choice for developers.