Animate Face (face vid2vid)

@kaiwenzheng1 VersionGenerate

Reenacts a fixed source face with the head pose and expression of each incoming driving Image frame using the face-vid2vid keypoint-and-warp stack, emitting one animated Image per frame so a downstream Image stream becomes the animated video.

How it fits

Typical backends

Drive the bound source face from a recorded video and publish the animated frames over HTTP for a downstream consumer.

Reenact the source face live from a remote video URL and stream the result to a browser viewer.

Caveats

Hard constraintThe source face and the model checkpoint are bound at startup from their configured file inputs; when either points at a directory the first image or bundle found is used, and a missing source image fails startup.
I/O contractFrames are emitted at the generator's native square resolution, and with split enabled each output is a side-by-side of the detected driver face and the animated result rather than the animation alone.
FallbackWhen no face is detected in a driving frame a fixed placeholder is emitted at the same frame size, so the output stream never changes dimensions even as detection flickers.
Compatibilitydevice set to its automatic option prefers a GPU and falls back to CPU, but the generator and keypoint networks are impractical on CPU, so a GPU host is effectively required for real-time use.
State lifetimeThe source identity is loaded once and reused for every frame for the life of the deployment; changing the driven face requires a redeploy with a new source image input.

Versions

7e22babdlatestdefault
Automated release
6/29/2026