It's "magic" in a sense to be able to create part of a picture that was hidden, but not magic because it can't guess it perfectly right as soon as it's not one object over a consistent background. but the editor always offers 4 variants, and usually one is better/"less worse" than the others so I guess Google can also progressively refine it's training set. what I would like is either a way to describe what you want with natural language like with chat got or offer a way to zoom in and mock what you want. in the case of the beach picture.l, draw some yellow where it should put sand, blue where it's water, and let the AI render it the most realistically possible. AI is not ready to do all the editing autonomously but with some guidance it can be a great tool