RGFVR: Reference-Guided Face Video Restoration with Flow Matching
Researchers have developed RGFVR, a novel framework for face video restoration that uses reference guidance to preserve subject identity. This method conditions a pre-trained text-to-video generator with identity information and employs a two-stage training process. Experiments demonstrate RGFVR's effectiveness in improving fidelity, temporal consistency, and identity preservation across various degradation types. AI
IMPACT This research advances generative modeling techniques for video, potentially improving applications in media synthesis and restoration.