Licking Lesbians

LORA trigger word is lesbian cunnilingus

Lora strength is best at 0.4-0.7

IMPORTANT INFO : PLEASE READ

This LORA was trained on Uber Realistic Porn Merge (URPM) - Cunnilingus Edition (ALPHA1)

This LORA WILL NOT WORK WELL WITHOUT One of the above models, either Alpha1, 2, or 3. I get the best results with #1. This LORA just gives the model a little push, narrows it's focus, and frequently gives the look of a professional porn set.

Version 2 was created as a fine-tune of URPM-CE-A1 then extracted as a LORA. It is big. I know. But the quality jump is also big. I extracted a smaller version, but the quality didn't make it worth it over Version 1. If you think the Version 2 LORA is good, you should see the raw fine-tune. But, it's 4gb. Finetune to LORA extraction is definately the way to go. Will use this method for everything moving foward.

OK, here's some tips for Version 2. It is really best at 0.7 Strength. Low CFG as you can go and keep things consistent. The Bad news. It doesn't work well with other models. Believe me, i tried. Most models just don't have enough girl-girl data to work with, which is why we're mostly stuck with URPM for now. But, I'm working on that. :)

This LORA is VERY specific. It only generates images of a women performing cunnilingus on a woman lying on her back. It only works well at 512x512 but it can be pushed to about 680 portrait or landscape, your milage may vary.

For best results, start your prompt with:

lesbian cunnilingus, a photo of a woman licking a woman's pussy,

the following words in the training data can produce results:

pale skin, dark skin, small nipples, small breasts, pubic hair, finger in pussy, hand on another's head, hand on thigh, hand on stomach, tongue out, hand on breast, hand on another's breast

All sample images heavily use ADetailer

Version 2: All images upscaled 1.5 and used adetailer on faces and breasts.

Version 1: I didn't use any inpainting or upscaling to demonstrate the model's consistency, or lack of. It's a 1.5 model, so hands suck, and there's odd extremities regularly.

This LORA produces images that have better lighting and skin effects than URPM alone, but it lacks flexibility. It was trained using 115 carefully manicured images of women in a very specific pose.