RelightAnyone: A Generalized Relightable 3D Gaussian Head Model

Disney Research Studios is presenting a generalized relightable 3D Gaussian head model at CVPR 2026 that targets one of the annoying constraints in avatar capture: subject-specific OLAT data. Instead of requiring every performer to be shot under one-light-at-a-time lighting, the method learns from standard single- or multi-view images and maps flat-lit 3DGS avatars to relightable Gaussian parameters.

The practical trick is a two-stage pipeline. Stage one trains a flat-lit 3DGS representation across diverse multi-view datasets without OLAT, using a dataset-specific lighting code to align lighting in a self-supervised way. Stage two learns the mapping to physically based reflectance parameters from a much smaller OLAT dataset, which is then reused to generalize relighting to new subjects.

For game teams, the appeal is obvious: relightable digital humans are usually expensive because capture rigs, lighting setups, and per-subject data collection are hard to scale. If this holds up outside the paper, it could reduce the amount of bespoke capture needed for character pipelines, virtual production, and lookdev workflows that want believable heads under arbitrary lighting.

The other notable point is that the model can fit unseen subjects from as little as a single image. That is still research-grade, but it suggests a future where relightable avatars are less dependent on controlled studio sessions and more compatible with the kind of messy source material teams actually have.