ID-V2V introduces a video-to-video generative framework for identity-preserving restylization. It treats identity preservation as video relighting, while using edited keyframes and depth sequences to propagate scene, lighting, and style changes. Relit facial regions and facial normal maps constrain facial likeness and performance, including expressions, gaze, and lip synchronization. Because the method can construct training pairs from a single source video, it avoids relying on scarce paired restylized-video data. The authors report improvements over existing methods and support for both single- and multi-subject scenarios.
No heat snapshots are available in the last 24 hours.