Deep Kiss V2 (Illustrious)

READ!

Share any good gens you get, they get used for a future dataset and you get buzz

After 5 failed runs and thousands of buzz... I finally got this.

"What Does This LoRa Do?"

Kissing is notoriously difficult to generate correctly. My goal was to stomp out the body horror issues found in the old LoRa(Which to be fair, was based on pony which is a ton less stable than illustrious) while keeping the "aesthetic sauce" intact, and I mostly succeeded.

"Why Use This LoRa?"

Sure, modern models can generate kisses (mostly) fine, but they are boring, this LoRa aligns kisses to a nicer aesthetic. It also allows for greater variability in how kisses are depicted.

For best results, I suggest using WAI-NSFW-ILLUSTRIOUS-V11.0. However, any model based on Illustrious should work well.

How to use it (How These Tags Were Used in the Dataset, and what they will get you when prompted):

Kiss = Lips touching, no tongue.

French kiss = Lips touching + some tongue visible.

Tongue out = makes the tongues more visible, may mess with image coherency, generally paired with 'french kiss'

Tongue kiss = Tongues touching, but lips not touching.

Head tilt = One person’s head is tilted sideways.

Turning head = When the head is turned to the side to receive a kiss

Lips = this generally depends on the artstyle as some have lips and some dont, but if they dont this would be good to add if you want visible lips

Arched back + ??? from behind = A new position (see preview images). Recommended aspect ratio for this position: 832x1216 or similar.

Important notes:

1- Lowered tag bloat, no need to use "kiss" alongside "french kiss"—just "french kiss" is enough. "French kiss" already implies the presence of a tongue, so no need to add "tongue" separately, and also the "tongue" tag generally is a big cause of body horror.

2- If the image looks off, inpainting works wonders, just inpaint over the faces/tongues and retry. If that doesn’t fix it, reroll the inpainting with a different seed.

3- For lora strength 0.7 strength works best in most cases—it enhances the image with no detriment to style LoRas. 1.0 strength works too but may cause issues. Adjust based on your needs.

4- Quality tags may mess with the 'aesthetic' quality, 'masterpiece' is usually fine though

5- WAI-NSFW-ILLUSTRIOUS V11.0 has inbuilt artist tags like (blackwhiplash, kunaboto, aster crowley, etc). These are a lot easier to use than LoRas

6- Make use of Img2Img + Adetailer + Inpainting for top tier results

If you got spare buzz and would like to support me, I'd appreciate it, this took a whole lot of work from gathering a 140+ img dataset to proper tagging, to trying out different epochs, to finding out how to properly tag it, to... you get the gist😵‍💫

Deep Kiss V1 (Pony)

Hello. This is my first LoRa.

It helps generate 'deeper' kisses, for all the fellows who share this fixation(and there is quite a few of us), I felt the kisses generated were too 'Vanilla' so i got to work

This works best with another style LoRa in my tests, otherwise it may be prone to generate body horror, but it does work

If you got a good gen please do share it here