ID-V2V: Identity-Preserving Video Restylization
Yuancheng Xu, Mingming He, Pablo Salamanca, Li Ma, Yash Kant, Emmett Steven, Paul Debevec, Ning Yu
SIGGRAPH Asia 2026
cs.CV
2026-07-25
ID-V2V (Netflix, Eyeline Labs, Adobe; SIGGRAPH Asia 2026) does identity-preserving video restylization: it propagates an edited first frame's style and lighting changes across a whole video while strictly preserving facial likeness, expression, gaze and lip sync. The key move is recasting identity preservation as a video relighting problem (only illumination varies), with relit faces plus normal maps constraining identity and edited keyframe plus depth driving generation, which lets it build training pairs from a single video. Single- and two-subject identity similarity is 0.70 and 0.64, with user-study win rates of 75% and 86%.
In video creation there is a common paradigm: capture the human performance first, then design the visual look (background, lighting, style) in post. Generative video models can already synthesize high-quality, temporally coherent video, so a natural question is whether they can faithfully preserve a human performance (expressions, gaze, lip sync) while flexibly changing the visuals. Existing methods cannot: they either inject identity from one or two reference images and miss the temporal detail of a real performance, or they use abstract control signals like facial landmarks or low-dimensional identity embeddings, losing micro-expressions and lip sync.
Harder still, there is no paired training data: in the real world it is nearly impossible to get two videos of the same person doing the identical action under two different visual styles. ID-V2V (Netflix, Eyeline Labs, Adobe) tackles both the modeling and the data side.
The core is to split the task into two orthogonal goals, with one key observation. The observation: under identity-preserving restylization, facial structure and expression should stay invariant, and the only permissible variation is illumination. So "preserve identity" is recast as a video relighting problem, and "change visuals" as edited-keyframe-guided conditional image-to-video generation.
That yields complementary control signals:
The payoff is on data: relit faces, normal maps and depth can all be derived from a single training video, so paired supervision is synthesized from one video, sidestepping scarce real pairs. A depth-dropout trick lets the model optionally use depth at inference, avoiding color casts.
Table 1 compares single- and two-subject restylization against AnimateAnyone, FantasyPortrait, VACE, WanAnimate and SteadyDancer:
| Metric (single-subject) | ID-V2V | Strongest baseline |
| Identity similarity (AdaFace) | 0.701 | WanAnimate 0.529 |
| Expression similarity (EMOCA) | 0.899 | WanAnimate 0.841 |
| User study (likeness/perf/quality win rate) | 75.0%/74.3%/70.1% | WanAnimate 10.5%/11.4%/12.6% |
In the two-subject setting, identity similarity is 0.637 (VACE 0.401, WanAnimate 0.432) with user-study win rates of 85.6%/86.1%/81.1%. Ablations show both control signals matter: removing face video drops identity similarity from 0.701 to 0.574. ID-V2V does not degrade on general VBench and VBench2.0 quality, while face-specific baselines degrade noticeably from single- to two-subject.
For teams in cinematic digital humans, virtual production and human-centric content, ID-V2V offers a practical path: first-frame edits propagate automatically across the whole video, decoupling performance capture from visual design, so the final lighting and set do not need to be built on location. Formalizing "preserve identity" as relighting gives both a clean constraint (only illumination varies) and, as a byproduct, a fix for paired-data scarcity (single-video self-supervision). It also holds identity and interaction for each person in multi-subject scenes, where existing methods broadly struggle.
The authors give their own failure case: results break under irregular source lighting, so the relighting formulation is not robust under extreme original illumination. The user study has high win rates but is a preference vote, not an objective metric, and sample size and reviewer background are not clearly reported. The baseline set does not include the latest strong identity-preserving video methods under a unified comparison, so the lead may be overstated. On the VBench-style metrics in Table 1 the gap to baselines is small; identity and expression metrics are where ID-V2V really pulls ahead, and on general frame quality it may have no clear edge.