ID-V2V: Identity-Preserving Video Restylization
The paper frames face-preserving restyling as a relighting problem, then uses an edited keyframe to drive the rest of the video.
ID-V2V is built to carry scene, lighting, and style edits across a source video while keeping facial likeness, expressions, gaze, and lip sync intact. The authors say scarce paired training data is the core obstacle, so their method constructs training pairs from a single video instead. It combines relit facial regions and facial normal maps for identity control with keyframes and depth sequences for coherent video generation. The paper says experiments show stronger facial identity and performance preservation than existing methods, including single- and multi-subject cases. HF Daily Papers' note
ID-V2V is built to carry scene, lighting, and style edits across a source video while keeping facial likeness, expressions, gaze, and lip sync intact. The authors say scarce paired training data is the core obstacle, so their method constructs training pairs from a single video instead. It combines relit facial regions and facial normal maps for identity control with keyframes and depth sequences for coherent video generation. The paper says experiments show stronger facial identity and performance preservation than existing methods, including single- and multi-subject cases. HF Daily Papers' note
score 5