receiving an input image with first lighting properties;

processing the input image, comprising a first plurality of pixels, using a convolutional neural network to generate a second plurality of pixels corresponding to the first plurality of pixels, the second plurality of pixels comprising an estimate of the first lighting properties, the estimate comprising estimates of at least one of: hue values, saturation values, and brightness values of the first plurality of pixels;

modifying the input image within a region with an augmentation to generate a modified input image, the augmentation having second lighting properties; and

changing the second lighting properties of the augmentation in the modified input image based on pixels of the second plurality of pixels corresponding to a same position within the first plurality of pixels as the region within the modified input image.