LumiTokens edges out physics-based 3D relighting via token-space transformation

A new arXiv paper from researchers at Adobe Research and Northeastern University presents LumiTokens, a framework that treats 3D relighting as a pure token-space operation rather than a rendering or material-decomposition problem. The authors — Yiwen Chen, Matheus Gadelha, and Huaizu Jiang — observed that recent latent scene representations encode multi-view images into compact tokens without fixed physical semantics, opening up a fresh design space. Their Scene Token Editor processes these scene tokens jointly with light-ray tokens through self-attention, producing updated tokens that decode directly into multi-view-consistent relit images. No 3D meshes, no BRDF estimation, no ray tracing — just learned latent transformations.

What makes LumiTokens particularly powerful is its unified lighting interface and progressive editing capability. All lighting signals — environment maps, point lights, and area lights — are parameterized as Plucker ray tokens, a representation with no explicit spatial structure but enough geometric information to support native 3D user interaction. Because the editor outputs tokens in the same latent space as its input, users can stack edits incrementally, adding one light source at a time and composing the effects directly in token space. Experiments show relighting quality comparable or superior to existing methods, while enabling a workflow that would normally require full recomputation for every new lighting arrangement. For 3D creators, this could make iterative lighting design dramatically faster.