In a collaboration between Netflix and Hugging Face, researchers have introduced ID-V2V, a novel video-to-video generative framework designed for identity-preserving video restylization. The work will be presented at SIGGRAPH Asia 2026 and addresses a key challenge in visual storytelling: how to edit the environment, lighting, and style of a video after capture while keeping the original human identity and performance intact.
The core problem is the lack of paired training data, as identity-preserving restylized video pairs are rare. To overcome this, the researchers propose decoupling source-grounded identity preservation from edit-driven video synthesis. Their key insight is that facial appearance and expression should remain invariant, with illumination being the primary permissible variation. Thus, they cast identity preservation as a video relighting problem and model visual edit propagation as controlled video synthesis guided by edited keyframes.
ID-V2V integrates complementary control signals: relit facial regions and facial normal maps tightly constrain facial likeness and performance, while edited keyframes and depth sequences enable flexible and temporally coherent generation. This design allows training pairs to be constructed from a single video, eliminating the need for scarce paired data.
Extensive experiments show that ID-V2V significantly outperforms existing methods in preserving facial likeness and fine-grained facial performance, supports both single- and multi-subject scenarios, and delivers high visual quality. The code is available on GitHub at Eyeline-Labs/ID-V2V.