We introduce a text-to-intrinsic (albedo, rough, metallic, normal) model, which can be used rerendered under arbitrary lighting condition or even distilled into room-scale PBR textures.

Native-Resolution Image Synthesis

Novel visually prompted diffusion model generates spatially controlled histopathology images, extends to unannotated TCGA, and produces synthetic training data that outperforms real data for segmentation models.