Many industries, such as film production, architecture, advertising or product design, benefit from the synthesis of highly realistic images from virtual 3D models. They resort to a set of algorithms, broadly termed as physically based rendering (PBR), which can generate synthetic images indistinguishable from real photographs. This is achieved through an accurate simulation of the physical behaviour of light and its interaction with the surface materials (light transport simulation). However, the visual realism of PBR comes at a high computational cost, hence limiting its application to cases where long offline rendering times are admissible, or to the high-end film industry which can afford render farms with lots of computers and graphics cards. This project aims at achieving a drastic reduction of the computational load required to synthesize photo-realistic images. To this end, our goal is to enrich the PBR field with a new set of powerful theoretical tools offered by recent advances in several disciplines, namely, machine learning, numerical analysis and, more importantly, Bayesian statistics. The proposed approach challenges decades of frequentist views in the PBR literature hence opening a new methodological path for photo realistic image synthesis. At the same time, the challenges of the application to PBR will push the very limits of these theoretical tools. Indeed, their application to PBR entails significant re-design and improvement efforts which will take the contribution of this proposal beyond the field of PBR. In the end, PBR will become affordable for most industries mentioned above and open doors for new industrial applications, for instance, virtual, mixed, and augmented reality, increasingly used by consumers and in fields such as Medicine. 

Project CNS2022-135480 funded by the Spanish Ministry of Science and Education and the Spanish State Agency for Research MCIN/AEI /10.13039/501100011033 and by the European Union NextGenerationEU/PRTR.