>>937229312
I think you fundamentally misunderstand what AI does. Strap in.
AI works by taking a blank image filled with random static and turning that into the image you requested. It knows how to do that because it has previously taken thousands if not millions of images and broken them down from a finished image into static. But since you want to take an existing image and turn it into a similar image we have to give the AI a lot of specialized instructions.
Some people have things like body mapping where the AI can attempt to detect specific body parts based on similar images it's studied from. I don't have an independent body map like that since I haven't taken the time yet to learn how to set one up. I literally don't do this for a living, it's just a hobby.
I *do* have a specialized lora that can detect faces, but that's not specific for tits or overhead selfies or whatever.
And that's where the problem lies. Without building something that you can manually control how the body is positioned, you need to rely on the AIs more general knowledge of how things *should* look. Problem is the AI doesn't know the actual difference between a tit, a thigh, or a random block of flesh. That's why so many times you'll get back weird cronenberg shit unless you know what you're doing.
I can try describing the picture through the prompts, and I can try carefully telling it what needs to be changed, how, and what should remain the same, but if it gets confused because it wasn't trained on, in this case, a lot of overhead selfies of plus sized women in cars, then the best it can do is make an educated guess based on your instructions.
I'm sure someone out there has a massive collection of plus sized women taking overhead selfies in cars both with their tops on and with their tits out, but unless that person also took the time to make a specialized model for it, I'm not going to be able to use that.
