Recently, significant progress has been made in generating expressive and temporally coherent portrait animations. Given a single portrait image, methods such as LivePortrait [GZL*24], Follow-Your-Emoji [MLW*24] or proprietary Act-One [Run25] can animate it using motion extracted from a driving video. The resulting animation supports a wide range of head movements while being able to preserve many aspects of facial expressions. Despite these methods’ impressive results, there are obstacles to their use in scenarios where computational budget is limited (e.g., on mobile devices) or when the aim is to retain the identity of the person in the driving video. In this paper, we present a novel approach to identity-preserving animation of a given portrait image that can run in real time even on a mobile device, inject the user’s identity, and reproduce detailed facial expressions. Our key contribution is an algorithmic solution that retains the appearance of the given portrait image while transferring coarse motions as well as detailed facial expressions of the user who drives the animation. The method runs in real time on a mobile GPU, achieving visual quality that sometimes outperforms even approaches that have notably higher computational overhead.

Full Text     BibTeX