RADIANT: Object-Centric Feed-Forward Global Illumination

RADIANT is a feed-forward neural renderer that synthesizes global illumination by representing each object with a fixed budget of 32 surface-patch tokens, so attention cost depends on object count rather than mesh resolution. It uses an invariant form-factor attention to model light exchange between objects and a cross-attention view decoder to render images. On a 1.03M-scene benchmark, it reaches 32.2 dB PSNR at 600k training steps, 3.3 dB above RenderFormer.