Yes, I ran out of buzz, so some images are not even highres, but yeah, I was a dumb fucker that didn't know what he was doing and deleted the whole dataset without saving it, so I just have this Epoch (which came out good enough), but I don't have any other Epoch in order to test and see which one is better, I spent everything I had yesterday on this and this is the result, hopefully someone will enjoy it. This version is heavily trained on big booty instagram-like style, so if you prompt for huge ass (f.e), you'll get some gyat, also tattoos, I have a lot of tattooed images in the dataset, 7/10 times it comes out alright or good enough, but we all know A.I ain't on that level yet, but it does generate. It can also generate Futanari, and some "pov" dick nudes, I had some male pictures in but tbh, I don't think it captured it enough, anyways, I'm blabbing too much, there you go.
Alright, after spending last night entirely testing this LoRA and generating images with it, I can definitely say it is working properly, but since the dataset is fairly simple & small, it isn't as flexible as I'd like it to be, so I might be training a new version fairly soon.
Recommended
I recommend using DPM++ 2M Karras for a more pixelated style, and 20~25 Steps. If you want a more smooth/high definition image, then go with Euler A, 30 Steps. And y'all can test it against different Checkpoints, I've tested on almost all of them and some does a good job, others not so much. Strength 0.9 and/or 1.05, lower for more flexibility. As for your cfg, anything below 4.0 is good to go, like always, 4.0 being the max. I'd recommend lower for a more noisy/realistic result.
Some negatives I'd recommend if you'd want a different result:
tattoo, arm tattoo, shoulder tattoo, neck tattoo, full body tattoo, pale skin, dark skinned female, tan skin, huge ass, overweight, plump, belly, BBLThank you again for using my LoRA, you can support me through either donating Buzz ('cause I'm out completely), or you can use one of the links provided beneath my profile.
Description
Changelog 05/03/25:
So, version 1.1 is a bit more controlled and it was trained with a few changes in the parameters, definitely better at generating realistic images using standard illustrious models, but the moment you use it with a dedicated "Realistic" model, i.e I tried with PornMaster for example, the LoRA looses completely control and is (barely) noticeable, I'm still trying to figure out why it gives out a more 'porny' style instead of following the 'amateur photo' style I'm going for, but it's here, you can test it out for yourselves.