Netflix drops ID-V2V to restyle videos without messing up faces

Netflix's Eyeline Labs has open-sourced ID-V2V, a generative video framework built with one specific job in mind: letting you restyle lighting, clothing, and backgrounds across an entire video using a single edited keyframe, while keeping the actor's facial identity, micro-expressions, gaze, and lip sync intact.

Traditional video restyling tools notoriously warp human faces into unrecognisable caricatures. ID-V2V supports both single and multi-subject shots, and because it treats full-scene restyling as a superset of lighting control, it also works seamlessly as a standalone video relighting engine without touching the background.

Forgiving keyframes and clever training

In practice, artists usually prep their restyled first frame using image editors like NanoBanana. These tools often tweak the actor's head angle or mouth shape, creating an immediate mismatch with the original footage.

ID-V2V handles this elegantly. While the output's first frame reflects your imperfect keyframe, the very next frame snaps right back to the actor's original pose, gaze, and performance. You never have to waste hours crafting a pixel-perfect first frame.

The real breakthrough lies in how it was trained. In production, collecting paired video clips of the exact same human performance across different styles is virtually impossible. Previous attempts relied on facial landmarks or low-dimensional embeddings, resulting in stiff, lifeless faces.

Eyeline Labs solved this by decoupling scene synthesis from facial preservation. They isolated faces from ordinary video clips, stripped away the cinematic lighting using an image relighting tool, and trained the model to invert that lighting while hallucinating the masked-out background from scratch. This turned identity preservation into an inverse relighting task, creating flawless training pairs out of single video files.

720p limits and a hefty 80GB footprint

ID-V2V currently caps out at 720p resolution. That is sharp enough for look-development, pre-visualization, and concept reels, but it still falls short of final theatrical mastering.

The bigger hurdle is local compute. The core model weighs in at nearly 80GB, putting it well out of reach for standard consumer setups. You will need top-tier enthusiast GPUs or workstation-class hardware to run inference locally.

Eyeline Labs has published complete setup instructions and code documentation on GitHub, letting interested builders deploy it directly on local machines.

Why it matters

For VFX artists, commercial editors, and indie filmmakers, ID-V2V represents the kind of surgical control generative video desperately needs. Instead of rolling the dice on random text-to-video outputs, creators can now restyle complex scenes or alter lighting setups without discarding a single frame of an actor's nuanced performance.

That said, an 80GB model tied to 720p output makes this a bleeding-edge sandbox rather than an everyday production driver. We think the real breakthrough will arrive once the open-source community quantizes this architecture down to practical VRAM levels and scales output clarity past 1080p.