FactorPortrait: Controllable Portrait Animation via
Disentangled Expression, Pose, and Viewpoint

1Meta Reality Labs 2Technical University of Munich


DynamicSweep Dataset

This is a synthetic dataset created from Animatable Gaussian Avatars. For each identity, we randomly select a sequence of continuous expressions and poses to animate head Gaussians, and then render dynamic Gaussians along a camera trajectory. In this way, we can obtain a video with joint changes of viewpoint, pose, and expression along time.


Self Driving

Input GAGA CAP4D HunyuanPortrait Ours GT
Input GAGA CAP4D HunyuanPortrait Ours GT

Cross Driving

We use the single image from Phone dataset as Source ID image, and use the video from DynamicSweep dataset as Driving video, and also its camera trajectory.

Input GAGA CAP4D HunyuanPortrait Ours Reference/Driving
Input GAGA CAP4D HunyuanPortrait Ours Reference/Driving