I’ve been trying to do something like this for ages. I’m not quite happy with how it’s come out so far, so I plan to keep trying. I used a “canny” controlnet and used PhotoShop to produce the control image, which is just outlines. Then when I got something that started to look close I switched to inpainting of just the crotch region. That resulted in some artifacts around the mouth, so I used img2img, which caused it to start looking more like a painting, so I had to mess with the prompt. Not much use in sharing generation info on this one because of the number of different steps.