LivePortrait is a computer vision framework designed for portrait animation and generative video synthesis. It functions as a deep learning system that transfers facial expressions and head movements from a driving video source onto a static image or an existing portrait video, effectively decoupling the subject's identity from the dynamic motion patterns.
The framework utilizes keypoint-based motion retargeting and implicit 3D latent representations to map movements across different subjects, including both human and animal portraits. By employing canonical motion normalization and feature-space non-rigid warping, the system ensures that driving inputs are standardized and aligned with the target geometry before synthesis occurs.
The process includes generative adversarial refinement to maintain visual fidelity and skin texture, alongside mask-based stitching to integrate the animated facial region into the original background. The framework also supports inference optimization through pre-computed motion templates and input scaling to facilitate the production of consistent, high-quality video output.
