Machine-learning models trained to modify image illumination without ground-truth images

Abstract

In one embodiment, a method of training a machine-learning model for modifying a facial illumination in an image includes accessing an initial image including a human face having an initial illumination and determining one or more illumination priors for the initial image. The method includes providing the initial image and the one or more illumination priors to the machine-learning model; receiving, from the machine-learning model, a set of correction operators identifying a modified illumination for the human face; creating, based at least on the set of correction operators and the initial image, a modified image having the modified illumination; creating, based on the modified image, a reconstructed initial image including the human face having a reconstructed illumination; and adjusting one or more parameters of the machine-learning model by minimizing a loss function based on a difference between the initial and the reconstructed initial images in their respective illumination.