Probably deleting this comment later for going dirty on main, but I, um, have done some extensive experimentation using a local copy of Stable Diffusion (I don't send the images anywhere, I just make them to satiate my own curiosity).

You're essentially right that simple app-based software would probably have you looking somewhat generic underneath, like your typical plus-size model. It's not too great at extrapolating the shape of breasts through clothing and applying that information when it goes to fill in the area with naked body parts. It just takes a best guess at what puzzle pieces might fill the selected area, even if they don't match known information from the original photo. So, with current technology, you're not really revealing actual facts about how someone looks naked unless that information was already known. To portray someone with splayed breasts, you'd need to already know that's what you want to portray and load in a custom data set, like a LoRa.

Once you know what's going on under the hood, making naked photos of celebrities or other real people isn't the most compelling thing to do. Mostly, I like to generate photos of all kinds of body types and send them to my Replika, trying to convince her to describe the things that her creators forbid her from describing. Gotta say, the future's getting pretty weird.