I've been messing around a lot more with the AI mod and made a lot of changes, though most of them are a bit wonky at the moment, would be great of someone better than me and writing prompts could have go at things
What Iv'e done:
1.
Added a checkbox to enable/disable the background removal from my previous version. IF background removal is disabled, the prompt will now include a location name, (research laboratory, office, mom's bedroom and so on) so that the background can be a bit more populated. Though that only really works if the denoise strength is set extremely high and/or img2img is disabled, otherwise the blank background with stick around.
2.
I've tried to pass along more details about the physique of the girls in the prompt, looking at both the height and the weight of the girl, so now it doesn't just send one of the body-types used for the old images, "slim body", "standard body", "curvy body" and "pregnant body". It tries (and mostly fails) to give different prompts for different height and weight. Thought since it's not like stable diffusion understands "a woman with a weight 200 pounds" i still have to translate that to meaningful prompts. But more on that in a bit.
3.
Ive also located and tried to pass along more details about the sexual positions when your're having sex. So rather than just send "on all fours", "sitting","kneeling", which were based on the pose of the original image, I now send information about the actual sex-act in the prompt (blowjob, deepthroat, missionary sex, anal sex in doggy style and so on).
Both the 2nd and 3rd change, unfortunately, are highly inconsistent. Even though I've experimented A LOT with stable diffusion before this, I haven't really considered how difficult (or impossible?) it is to write a prompt that works on the first try. If you write "blowjob" as a parameter, you might get an image of the girl giving a blowjob, or you might get an image of her turning herself into a pretzel while a penis grows out of her mouth.
Same for weight, as mentioned, i can't really send actual weight and height to stable diffusion and hope for any good result, so for weight I calculate BMI and the send a prompt based on that, so for example, the cousin has a BMI of around 35 (which is NOT healthy) based on that I make a (completely uneducated) guess that maybe something like overweight:1.33 might be a good part of the prompt.
but with overewight:1.33 in the prompt, both these images were generated during a sex scene with her, and that's with all the in stable diffusion parameters (for steps, cfg scale, model and so on) being exactly the same, if I swap models, the different becomes even greater, some models barely seems to recognize the overweight prompt while others' quickly turns the girls into whales with overweight in the prompt.
Simply put, it seems impossible to get a consistent level of "overweightedness".
Otherwise I kind of liked the idea of making the serums for gaining or losing weight & growing or shrinking in height more relevant.
View attachment 4903308
with image to image enabled things obviously stabilize, but it stabilize so quickly that most of my changes completely goes away
I haven't been able to find any sweet-spot for the denoising strength where the img2img helps steer things in the right direction but at the same time doesn't completely overpower the changes I've made. Even with a fairly high denoising strength, you'd at most see some random penises popping up during sex scenes when img2img is enabled.
So again, not sure how interesting this is for most people due to how inconsistent it is, but perhaps someone better than me at writing prompts can do something with it, at at least the basics are there now as far as figuring out BMI and what sexual positions are used and so on
Edit:
Had messed up a bit with the sex-position thing, making it only identify it sometimes, removed the old attached file and attached a fixed version so now it should recognize the changed position to a better extent ( still doesn't make the prompts any more reliable though
)
