>>175271) simplify your prompts. Natural language can be useful at the beginning, but rather than giving SD instructions like you're talking to a graphic artist, think of it like tagging a photo for someone with vision issues. Just off the cuff, here's how I would have prompted that second image:
"an extremely obese woman stands in her home office, obese, huge belly, belly overhang, fat arms, brown hair, long hair, t-shirt, pajama bottoms, comic posters, computer, desk, messy room"
Once you have a pretty reasonable image, you can use a simple photo editor to pull out sections of the image and reprocess them with img2img if the fine details don't work. That'll give you a lot of control once the overall composition is how you like it.
2) add a weight-based LoRA to the mix. Symix, BGV5EX, Obese Girls, Chubby Girls just to name a few. Those LoRAs tend to 'understand' obese bodies better than any particular model will, especially if you're aiming for realism. Just keep the weights a little lower than normal (like 0.5-0.8) and it'll go a long way to keeping a large belly just a belly.
3) add "linea nigra" to your negatives. That's the vertical line pregnant women develop that a lot of the older models try to add to a large belly. They tend to 'think' that a really fat gut must be a pregnant person. It's not a magic bullet, but it's helped keep a lot of the larger bellies in check in my experience.
And I wrote a guide that details my process. It's not perfect, and I don't claim it's the only way to do this stuff, but you might find it useful.
https://www.deviantart.com/theguywhodidathing/art/How-did-you-DO-that-989204336
