Meet Kitten.

A face/identity driven LoRA built off very tight crops on a very defined data set.

(Also my first "fully baked" non-test LoRA, and first released LoRA).

What does it reinforce?

Upturned nose, a defined bridge, pouty lips, long eyelashes, defined eye shape and brow, well manicured eyebrows, a defined jawline, and light freckling across the upper cheeks and top of the nose. It should blend well in to most Illustrious-XL models and merges, most SDXL models, and plays nicely with many Pony-XL models. Trained on Illustrious-XL as the base to maximize compatibility.

Many anime models will play nicely, but the less nose definition in the general model the more you may have to push the LoRA.

A note on Krea 2

The Krea 2 version of this LoRA is a true alpha state, and the first of any of the Kitten versions not trained locally. Still working on setting up local generation and training for Krea 2, so this was done via Civit AI using the same training dataset as the previous models (tight crop, face only for the same look but flexibility in body/usage). Note provided here just to say my normal train and test pipeline wasn't used here, so I'm learning as you are.

Usage

• Trigger Word: k1tt3nf4c3

• Secondary Trigger: freckles (helps to reinforce the facial features, and bring some identity back from other LoRAs that overpower the intentional subtlety behind it)

• Illustrious XL (trained on Illustrious-XL)

  • LoRA Strength: 0.6 to 1.2
    (you can get hints at 0.3-0.5, but just hints)

• Illustrious - Anime Focus (Trained on Nova Anime XL)

  • LoRA Strength: 0.6 to 1.2

• Pony - Realistic / Cyber Realistic (Trained on CyberRealism Pony)

  • LoRA Strength: 0.4 to 0.8
    (core images are based on illustrations, will start to pull skin to sketch quickly, great if it is what you want, but if you're trying to hold "realism" this will pull it out quickly)

• Z-Image-Turbo (Trained on Z-Image-Turbo + ZIT Trainer)

  • LoRA Strength: 0.5 to 0.7 (Realism)

  • LoRA Strength: 0.7 to 1.0 (Artistic)
    (much longer baked to push the original art style from the base images, higher weighting can make it harder to remove certain feature locks or specific hair/eye colors, but still has flexibility)

• Krea 2 (trained remotely on the same dataset)

  • LoRA Strength: 0.4 to 0.6 (Realism)

  • LoRA Strength: 0.7 to 1.0 (Artistic)
    (the Epoch for this initial alpha build is a bit heavier baked than I'd like, so the LoRA pulls pretty hard, especially on Krea 2 Turbo. I'll likely dial this back when I get local training for Krea 2 setup, but the "Kitten" artistic style can pull a lot harder than I'd normally like even at lower LoRA weights, the character is still clearly defined and visible at lower weights however)

As much as I hate the full l33tsp34k version of the trigger word, initial training output and Illustrious-XL models (many of them) leaned to heavily into cat girls, whiskers, or whisker tattoos, this prevents it.

Why does Kitten Exist?

Since starting AI generation coming from a background of photography and design work, the randomness has bugged me, especially when I find a specific character out of a large batch I like, then spend hours (or days) trying to chase that specific set of features again on other seeds, with other models.

Throwing mountains of prompt tuning at the problem, breaking model token limits constantly, only getting truly "close" a few times out of hundreds of images, but rarely ever finding "it" again.

So, I went all in on learning how to train my own LoRA, painstakingly worked through hundreds of Qwen Image Edit workflow runs that exceeded my VRAM, prompt tweaks, outputs, on and on. But in the end, I found her again, and again, and again.

Success.

Training Information

The dataset that drives this is very narrowly focused on the face, so it should play nicely across almost any body type, hair styles, hair colors, and most clothing/outfit LoRAs that don't overly pull on or push face characteristics. Trained on 40 tightly cropped face photos from the same source image, it should hold the character's primary identity while allowing for near complete customization everywhere else.

Cut at 300 steps to find the sweet spot on where safety prompts started to leak in slightly to allow for variation, it is intentionally a bit lightly tuned, the goal is to allow the character to show while having near full flexibility in all other aspects. Some models will play nicer than others, and some additional LoRAs may overpower if not balanced well.

Acknowledgement

This LoRA could not, and would not, exist without @purplelady's Elegence Niji LoRA (oddly enough, the Z-Image-Turbo version) creating the original style I found and loved, and the original version of the character that came out of it. You'll find as a result that due to that core structure it play's wonderfully together with the Illustrious version of Elegence Niji, and I highly recommend it if you're looking to amplify and reinforce this specific set of facial traits in your output.

The the source and training data is many steps removed from those original Z-Image outputs, the concepts that lead to this chase never left.

Next?

Her partner in crime, and raves, Minnie is in the works.
Update - Minnie is released, model found here