Type | |
Stats | 1,795 |
Reviews | (394) |
Published | May 2, 2023 |
Base Model | |
Training | Steps: 900 Epochs: 5 |
Trigger Words | chloe beast |
Hash | AutoV2 D136941393 |
This LoRA is still a work in progress, your mileage may vary! Detailed explanation below.
A variation of the "dangerous beast" outfit from Fate/Grand Order originally worn by Chloe von Einzbern in Fate/Kaleid Liner Prisma Illya.
This is a costume LoRA.
This LoRA is part of a series I'm working on, check out the others here:
About the v0.6 Release
This one gave me a lot of trouble (and still is continuing to do so). tl;dr: Messes up the belly button a lot and is somewhat overfit to the training data!
I still wanted to publish this preliminary version however, since I have no idea when I'll be able to produce a better version than this. Adding something like "tattoo", "piercing" and/or "navel piercing" to the negative prompt may help.
Common issues are the o-ring turning into a piercing, the bellybutton turning out weird or multiple bellybuttons appearing (usually each having their own o-ring).
One of the most persistent issues is that the model really struggles specifically with the area framed by the o-ring (usually the bellybutton), this was worse initially but it got a lot better after I manually removed the stomach tattoo that is usually present in this area in most of the training images (which was a pain to do).
This variation (despite not being trained all that much) seems to be somewhat overfitted to the dataset, meaning that a lot of the properties that shouldn't be learned (especially style and color of the hair and eyes) may pollute your generated images. This model (and some of the ones I trained on this data afterwards) all trade between consistency of the outfit and generalization of the character features, which is annoying and made selection pretty hard. I've yet to find a "best of both worlds" model for this, so this one skews more on the consistency side.
My hunch is that this problem requires more samples and more variation among the samples to derive a more generalized and flexible model, so going forward my plan will be to take whatever images I can generate with this that look decent enough and extend my dataset using those. That might however take some time (and motivation), so I don't know how long it'll be until I manage to improve on this.