🌲Introducing our paper A generalizable light transport 3D embedding for Global Illumination arxiv.org/pdf/2510.18189. Just as Transformers learn long-range relationships between words or pixels, our new paper shows they can also learn how light interacts and bounces around a 3D scene. Building on the analogy of light transport operator and Attention, our fully scalable model learns a generalizable light transport 3D embedding that captures global illumination. It takes as input the 3D scene assets—geometry, materials, and lighting represented as a point cloud—and encodes the complex light interaction into a 3D latent space, with codes anchored at the sampled scene points. Remarkably, the model is independent of both viewpoint and resolution, enabling view-consistent rendering across diverse scenes. (See teaser and attention figure.) (Our team work: with my mentor Marco @marcosalvi, labmates Mukund @mukundvermar, Cheng, NV folks Bart @BartWronsk, Lifan @winmad4869, and Tzu-Mao @tzumaoli & Ravi.) 🧵

We demonstrate the versatility of our framework by reusing a shared pre-trained encoder while introducing task-specific decoders. Using the same model, we accurately estimate diffuse global illumination that generalizes across diverse indoor scenes. With limited fine-tuning on a switched decoder, we further demonstrate preliminary results on glossy materials by predicting spatial–directional incident radiance fields, and explore extensions to jumpstart path guiding for unbiased global illumination.