ECCV 2026: Generative Relightable Avatars Combines Physics-Based and Diffusion Rendering for Full-Body Human Relighting

Researchers at the Max Planck Institute for Informatics have introduced Generative Relightable Avatars (GRA), a person-specific method for photorealistic free-view rendering and dynamic environment-map relighting of full-body human avatars. The hybrid pipeline starts with a tracked animated mesh, optimizes material parameters in UV-space using a microfacet BRDF model, and produces a coarse relit appearance — which is then refined by a feed-forward network (RelightNet) and finally enhanced by a fine-tuned video-to-video diffusion model that adds temporally coherent, high-detail output while preserving 3D control.

The approach is designed to address the inherently one-to-many nature of fine-grained appearance modeling, distinguishing it from fully regressive relightable avatar methods. An error-recycling strategy supports long-video generation, and the method generalizes to out-of-distribution lighting conditions — including single-light (OLAT) and near-field setups — without retraining. The paper is accepted to ECCV 2026.