I hadn't actually seen the demo when I commented! I'd assumed it was using an AI model to estimate distance from camera of the elements of a photograph and then reproducing a shallower depth-of-field. So my comment isn't all that relevant.