RelightAnyone: 3D Faces with Freely Changeable Lighting

RelightAnyone is a model for relighting 3D Gaussian head avatars built from single- or multi-view images under different scene illumination. It proposes a two-stage approach that learns relightable pa

RelightAnyone is a study that aims to create a 3D Gaussian-based head avatar from one or more images of a person and re-render it under new environmental lighting. Existing high-quality relighting typically required OLAT (one-light-at-a-time) capture for each person, with one light illuminated at a time. This work presents an approach that does not require separate OLAT capture for a new subject.

The key is to learn the correspondence between a 3DGS avatar built under flat lighting and relightable Gaussian parameters. In the first stage, a fixed-lighting avatar is modeled from multiview data without OLAT, using learnable lighting codes to accommodate different lighting conditions across datasets. In the following stage, the method learns a transformation into physically based reflectance parameters using relatively small amounts of OLAT data.

This separation is designed to achieve identity generalization across broad human datasets while reducing the amount of complex lighting-capture data required. According to the official description, the trained model can be adapted to unseen people and aims to create avatars for novel-view synthesis and relighting even when the input is limited, such as a single portrait taken outdoors. DisneyResearch|Studios released this work as a CVPR 2026 paper.

Maru