"reverse rendering" derives a 3d surface from flat image. Instead of throwing a neural network at the problem it is based on theory of light, so it obtains physically meaningful parameters, and can guide further research.
(on gh: mitsuba-renderer/mitsuba3)