Weight: 1.0 - Cooperates with concept LoRAs, Finicky with other character LoRAs
Samper: Euler_Ancestral Arbitrary
Scheduler: Beta Arbitrary
Behaves consistently between Base Krea 2 Turbo and Realia Krea 2 (I only tested it on these two). I prefer the expressions that can be achieved with Realia. Using a non-heretic Qwen text encoder will result in dulled expressions. "Fingerless gloves" does work to make her sleeve things.
There are inconsistencies with the outfit when compared to the anime, but it remains consistent between generations and the narrow scope of checkpoints I tested it with. When she starts appearing different, just toss in the trigger word and mention the specific part that is inconsistent.
All training data for this LoRA was SFW content. Trained on Krea 2 Raw. I do not really care what you do with this. I accept payment in the form of pictures of hairy feet from obese men.
Description
V1, don't expect a V2.
FAQ
Comments (6)
Let me know if there are issues, I've never made a LoRA before.
No glaring issues yet.
@AI_DK If you're willing to share any of the images you create and if you think they'd be good for a second training pass, I would bow down at your feet.
@AI_DK Especially if it is a form fitting outfit that is free of artifacts from her standard outfit, which is a current limitation on the dataset. I've decided I want to do a V2 with a larger dataset and a lower learning rate. Probably will shift back to the 750 step save and build back to maybe 1500 steps.
@Vermodia Her outfit is close enough, and I like the thighdentation. it tends towards the semi-realistic style you trained on, so getting another style requires some prompt, lora, or prompt-weighting.
Very nice