Research & Innovation
32 Tokens Per Object: RADIANT Renders Global Illumination Without Caring How Dense Your Mesh Is
Feed-forward neural renderers produce global illumination on new scenes without per-scene optimisation or runtime ray tracing. The ones built on transformers choke as meshes get denser, because they tokenize individual triangles. RADIANT gives each object a compact budget of 32 surface-patch tokens and decouples the attention sequence from polygon count entirely.