AITech Trends

Netflix and Eyeline Labs Introduce ID-V2V for Identity-Preserving Video Restylization

Netflix Research and Eyeline Labs have introduced ID-V2V, a video-to-video framework for identity-preserving video restylization in production environments. Starting from a single edited keyframe, the system propagates scene, lighting, and style changes across a source video while strictly preserving the subject’s facial likeness and performance, including expressions, eye gaze, and lip synchronization. The paper is accepted at SIGGRAPH Asia 2026.

Key Takeaways

  • ID-V2V uses one edited keyframe to restyle an entire video while preserving facial likeness and fine-grained performance
  • The framework decouples identity preservation from edit-driven synthesis, building training pairs from a single video without paired footage
  • It supports single- and multi-subject scenes and tolerates imperfect keyframes; research by Netflix Research and Eyeline Labs

Official explainer and results video (source: Eyeline Labs)

The challenge

Human performances are central to visual storytelling, but keeping identity and performance intact while allowing flexible visual edits has been difficult for generative video models. ID-V2V formalizes this as identity-preserving video restylization: applying scene, lighting, and style changes defined by an edited keyframe across a source video, without altering facial likeness or the original performance.

Decoupling identity from synthesis

The main obstacle is data. Paired videos of the same person performing identical actions under different styles are rarely available, and prior self-reconstruction approaches using abstract control signals such as facial landmarks or low-dimensional identity embeddings are too coarse to retain fine-grained facial performance.

The team’s key insight is that facial appearance and expression should remain invariant, with illumination as the primary permissible variation. Identity preservation is therefore cast as a video relighting problem, while visual edit propagation becomes controlled video synthesis guided by the edited keyframe. ID-V2V combines relit facial regions and facial normal maps to tightly constrain likeness and performance, and edited keyframes with depth sequences for flexible, temporally coherent generation.

Training without paired data

During training, face regions are kept as input while everything outside them is masked out. Starting from an ordinary training video shot with cinematic lighting, the team applies an image relighting model to the faces, converting them to regular lighting that approximates the source video’s lighting at inference. The model then learns to invert the process, restoring the original cinematic lighting while preserving likeness and subtle expression.

Because the training target is simply the original video, every pair is perfectly aligned, giving direct supervision for both identity preservation and scene generation without real-world paired footage.

Imperfect keyframes and applications

Image-editing models such as NanoBanana often cannot hold the exact pose or expression when restyling a frame, leaving the stylized first frame misaligned with the source video. ID-V2V follows the imperfect keyframe on the first frame, then re-aligns to the source video’s identity and performance from the second frame onward.

Video relighting is also supported as a special case. In the demos, the team relights the first frame with NanoBanana Pro and propagates the new lighting across the video, preserving motion, expression, and objects. Experiments show ID-V2V significantly outperforms existing methods in facial likeness and fine-grained facial performance, and robustly supports single- and multi-subject scenarios.

Cover image generated by AI for illustrative purposes only.

Source: ID-V2V Project Page · arXiv · GitHub