Once every object in the frame is mapped to 3D coordinates, exaggerating the parallax is trivial (like rendering a video game for 3D glasses / VR), but to your point, figuring out what was behind those objects (to fill in the negative space that remains after translation) is a guessing game, equivalent to various smart fill / magic eraser features. Originally they used only context from the same photo, and nowadays they also use generative AI. Not always great results.