SadTalker
Turns one portrait image plus an audio clip into a talking-head video, from a CVPR 2023 research project.
One portrait and one audio file produce a talking-head video on your own machine.
SadTalker animates a single face image so it speaks an audio track, with head motion and expressions. It suits researchers and hobbyists comfortable with Python and a GPU, not anyone expecting a polished product with ongoing updates.
What it is
You run `inference.py` with a driven audio file and a source image or video, optionally adding the GFPGAN face enhancer. It has still, reference and resize modes plus a full-image mode for upper-body clips, and a Gradio WebUI or Stable Diffusion WebUI extension.
Who it's for
Fits technical users and researchers working on audio-driven animation who can install old PyTorch builds and download checkpoints. It does not suit non-technical creators, and the README's own changelog ends in 2023.
Key features
Single image to video
Provide one portrait and an audio file and it generates the talking video.
Face enhancer option
Add `--enhancer gfpgan` to improve quality of the generated face.
Still and full modes
`--still` with `--preprocess full` gives a more natural full-image result.
Reference and resize modes
Extra modes added in later versions for pose reference and resizing.
Gradio WebUI
Run `app_sadtalker.py` or webui.bat/webui.sh for a browser interface.
Stable Diffusion WebUI extension
An extension for stable-diffusion-webui is published with its own docs.
Up and running in 3 steps
Our honest verdict
What's great
- It runs offline; the disclaimer says the code does not collect personal information, so source photos stay on your machine.
- Only one image is needed, where many lip-sync tools expect source video.
- The README says the licence moved to Apache 2.0 and the non-commercial restriction was removed.
- A Windows start.bat and Linux scripts download models, so setup is documented for both systems.
Where it falls short
- The setup is old and heavy: the README asks for Python 3.8, PyTorch 1.12.1 with CUDA 11.3, ffmpeg and a long list of downloaded checkpoints.
- The README's changelog ends in 2023 and the latest tagged release is v0.0.2, so expect little active development.
- The LICENSE says Apache 2.0 applies except for listed third-party components, so you must check the bundled models such as wav2lip and 3DMM files for your own use.
- The README's disclaimer warns against using it for fraud, deception or infringing portrait rights; you are responsible for consent when animating someone's face.
Pricing & license
SadTalker is Apache-2.0 and free to run yourself. The README mentions a free Discord bot and Hugging Face and Colab demos, but shows no paid plan. Your cost is a GPU machine and the time to set it up.
Open remote AI-training roles
870 open remote AI-training roles, updated November 14, 2026; most pay $20–$170 an hour. Zekai lists them free; the application goes to the employer’s platform.
Alternatives
These reviewed AI tools do the same job instantly — from our scored directory.
Use it, or have us set it up
FAQ
Can I use SadTalker commercially?
The README says the licence was updated to Apache 2.0 and the non-commercial restriction removed. The LICENSE excepts third-party components, so verify the bundled models before commercial use.
Does it need a GPU?
The install guide pins a CUDA build of PyTorch and the project is built around GPU use; the README does not state a minimum card.
Is there a hosted version?
The README links a Hugging Face Space demo and Colab notebooks, but the project itself is code you run yourself.
