Netflix’s ID-V2V preserves identity while restyling videos from one source clip

netflix · hf · 2026-07-28

Netflix researchers present ID-V2V, a video-to-video framework for identity-preserving video restylization.

Goal: propagate scene, lighting, and style edits from an edited keyframe across a source video while preserving facial likeness, expressions, eye gaze, and lip sync.
The main challenge is the lack of paired training data, so the method reconstructs training pairs from a single video.
Their key idea is to decouple identity preservation from edit-driven synthesis: preserve faces via a relighting-style formulation, and use edited keyframes plus depth sequences to guide temporally coherent generation.
Controls include relit facial regions and facial normal maps for likeness/performance, plus edited keyframes and depth for propagation.
The authors say the system outperforms prior work on facial likeness and fine-grained performance in both single- and multi-subject scenes, and they release code on GitHub.

More from Multimodal