Naughty to Nice Slider

XL VERSION AVAILABLE!

Usage

The biggest value to this LoRA (in my opinion) is that it bakes in around 50 tokens worth of data into the LoRA, and only uses single digits.

Obviously, use it to adjust characters to be more or less lewd. The newest version is much less lossy across it's range, so even without highres it produces decent images.

Lets be honest, 80% of use cases, you're probably using this in the naughty range. Why THINK about how to make a character lewd, when you're going to delete the image afterwards anyway? Just crank down the slider into the negatives and start generating.

The other 20% well, say you've got an image or an idea, but you're worried it might be too risque for what you're intended purpose is. Move that slider into the positives, and wipe away those features!

Prompt Advice

You may need to emphasize clothing in some prompts to keep the clothing theme!

Slider works best between -5 and 5. The negative can go lower and produce viable images, as can the positive. Use non-anime models at your own risk. This was trained on anime, so consider that a warning as results can vary.

I've noticed that V3's positives still suffer somewhat at 4 and 5, but the equivalent negatives are drastically better than before.

History

V1 - Slider designed to move from a very safe and pure (5) character to a nughty/sexy version (-5).

V2 - Fixes a lot of issues that V1 had, but has a slightly different feel, and it's own issues, but they're less severe than V1. Slider moves from (-4) to (4) (can go further, but results deteriorate depending on seed)

V2.5 - Follows methodology of V2, but fixes multiple issues of V2 (some new, some brought over from V1). Range back up to (-5) to (5).

Note: I've kept the internal metadata name as just naughty_to_nice between both versions, so you don't need to update your prompts with V2. If you leave V1 in the folder however, it will override control of the prompt for anything with _V2

V2.5 Has it's own name for generation.

V3 - Completely reworked methodology using new parameters. Much higher quality images across the full range (-5 to 5) of values. Works better with multiple different models, and produces better images even outside intended range (use at your own risk).

Training History

V1 - Trained using Ostris's AI-Toolkit for sliders for 400 steps over AbyssOrangeMix2.

V2 - Trained using Ostris's AI-Toolkit for sliders for 200 steps over DivineEleganceMixv6. Fine tuned many parameters and compared over 20 viable outputs to pick the best one!

V2.5 - Trained using Ostris's AI-Toolkit for sliders for 720 steps over DivineEleganceMixv6. Fixed many issues with V2, but follows same methodology.

V3 - Trained using Ostris's AI-Toolkit for sliders for 700 steps over DivineEleganceMixv6. Used new parameters, compared over 40 slightly different tuning parameter versions, and trained at a lower learn rate to prevent lossy behavior.

Other Notes

I started working on a version 3 of this LoRA in preparation for trying out XL training. I can't say when or if I'll actually get to it, but I wanted to start with a higher quality version based on new things I've learned in the past few months.