IMPORTANT NOTE: This LoRA was trained for my personal use. Its only goal was to work with my (unpublished) models, so I don't offer any kind of "guarantee" it will work in its full potential with any other model. I'm sharing this with you all, so if you want to use it, test it, merge it... be my guest... and post your (good) results please :)

The main purpose of StabilyXXer is to make realistic models better and it's trained to improve the following features:

• model stability (less problems, less seed struggles, less hallucinations...)

• improved details (anatomy, facial features, posing, textures, backgorunds...)

• lighting and shadows (improving overall realism)

Use it carefully, because it's VERY strong and can easily affect your model too much!

I suggest LoRA values between 0.1-0.4 for gens, and up to 0.2 for merging with a base model. I've tried it with a few "well known" models, and the results were from "meh" to "very good", but I didn't test any model thoroughly with XYZ's or fine tuning... just a few quick examples.

StabilyXXer training: 200+ real photos (no AI gens), 100+ different and precise hand-written tags (each image was tagged independently), 13 epochs (this version is e-7, which worked the best with my models), no watermarks, signatures or any kind of text. The source images were 90% NSFW and 10% SFW (for the balance, but also for ancoring proper clothing, setup and backgrounds)

Example images were made during the testing of my latest realistic model, the LoRA was merged with 0.1 weight, which visibly improved most of what I wanted. They are not strictly NSFW or X+, because I don't like posting porn slop here, but the LoRA knows that stuff :)