SadTalker
Open-source audio-driven talking-head animation from a single image (CVPR 2023).
About
SadTalker learns realistic 3D motion coefficients from audio to animate a single portrait image into a talking-head video, handling stylized and cartoon faces as well as photorealistic ones. Published at CVPR 2023 by researchers from Xi'an Jiaotong University, Tencent AI Lab, and Ant Group, it became one of the most-starred open talking-face projects (10.5k+ stars) and a staple of ComfyUI avatar workflows.
The pipeline separates 3DMM coefficient prediction (ExpNet for expression, PoseVAE for head pose) from face rendering, which keeps identity stable while the mouth, eyes, and head move with the speech. It remains a go-to for local, free talking-avatar generation, though the upstream repo is now quiet and active development has moved into ComfyUI nodes and community forks. Note: some derivative weights carry non-commercial terms — check the model card before commercial use.
What it can do
- Audio-driven facial animation from a single image
- 3DMM-based expression and head-pose control
- Stylized and photorealistic face support
- ComfyUI community nodes for no-code use
Setup
Clone https://github.com/OpenTalker/SadTalker, install Python dependencies, download checkpoints, run inference.py with an image and an audio file.
Prerequisites
Python 3.8+, CUDA GPU recommended, ~10GB disk for checkpoints.
Use cases
Works with
Trust score — 7/10
Scored 7/10 (metrics tier, metrics-v1). Evidence: Verified via vendor site/docs/GitHub and reputable coverage, wave 63; Trending score 72/100 from researcher assessment.
★ 10,500 GitHub stars
Sources
Every fact above was checked against these sources. Where a claim comes from the provider and couldn't be independently verified, the text says so.