TL;DR: We over-parameterize cameras as bundles of rays and propose a diffusion-based framework to predict this representations probabilistically.