>>14416295This is rapidly dropping in impressiveness as I come up with algorithms to explain it: geographical prompts will almost always refer to backgrounds (exception being where they are the main part, obviously), so one just needs to generate a convincing "Victorian" backdrop. The slicing into distinct layers or theaters of operation is central.
The rabbit is the depicted main person (and it's clearly a portrait requesting portrait, rather than e.g. of an object), so the AI can not only profit from rabbit data drawn from its database, but just main personalities of paintings in general. (note: I wonder what the AI would produce on a prompt that has less than a significant number of training data. What about a "binturong"? I guess it just extrapolates a best guess from the few as such tagged pictures and the "mammal" connection).
The same applies to bench and other objects. They can all be generated via independent routines, before you run a couple routines that analyze the geometry and position everything in such a way that reference data agrees with as being natural or naturalistic.
