there's really nothing magical about bokeh that requires machine learning imo. Lens blur can be approximated perfectly (by human perceptual standards) given a g

there's really nothing magical about bokeh that requires machine learning imo. Lens blur can be approximated perfectly (by human perceptual standards) given a good depth map, you just have to use a disk-shaped kernel instead of a gaussian distribution, then vary the size of the disk based on aperture size and distance to the object. Simulating lens blur is a pretty well researched topic at this point. There are other things that go into making these images more "realistic", like adding chromatic aberration, barrel distortion and halation, but by and large these things are affectations borne of nostalgia and not actual improvements to image quality.