Relightable Hollywood-ported Characters

Relightable Holoported Characters (RHC) is a person-specific research method from the Max Planck Institute for Informatics, the VIA Research Center and Google (London and Zurich). It was presented at CVPR 2026 as an award candidate. It renders and relights full-body, highly dynamic humans from sparse-view RGB video at inference time.

Instead of classical one-light-at-a-time (OLAT) relighting, a transformer-based RelightNet predicts relit appearance in a single forward pass. It takes physics-informed features (geometry, albedo, shading and view) computed from a coarse human mesh proxy and the input views, and cross-attends them with an environment map. It outputs texel-aligned 3D Gaussian splats attached to the mesh. Training uses a lightstage dataset in which frames lit by random environment maps alternate with uniformly lit tracking frames.

The authors report better visual fidelity and lighting reproduction than state-of-the-art baselines, including on out-of-distribution OLAT lighting never seen in training. They also show telepresence use, with several avatars relit consistently in a shared digital environment. The project page lists no code, model weights or demo. A YouTube creator likewise noted that code is not yet released. The method is person-specific and depends on lightstage-captured training data.