Made a TI for a Ballgag, not perfect at all, but you´ll get some images :D
best would be to render a batch 4, so you get at least 1 or 2 images to use :D it does not work with every checkpoint, as i tested ... don´t know what to change on the training data or settings, to make it better ... i have a set images with different people wearing a ballgag, all ballgags have the same color, i have multiple images of each person, so about 5 images per person by about 5 to 6 persons! maybe i´ll try with one image per person later!?
trained with 2500 steps
Description
trained with less images and only one image per person
still not close to perfect, problem is probably, that the model will not just recognize the gag, but the person partially ...!?
FAQ
Comments (14)
Try a few checkpoint and outputs not good...T^T If you train this on some popular model, or something like half AbyssOrange half some realistic models, would that be better? @u@
I´ll give it a try, later, atm i´m runing a training on less images with only one image per person :D
I saw her in the corner of my eye and mistook her for Cr1tikal lol
What's up everybody, it's Critical. Today I'm gonna be hmgg hngg gumm.
Works pretty well. Thanks for training such specific things. Too bad the color is mandatory though ;)
I'd love to see a nose hook Lora or embedding someday. That's a naughty tool ;)
nose hooks are nice, and evil, at the same time ;) yeah, well, i had best resources for the black one, maybe I´ll do a red one or another color sometime, soon :D
@zerberuszero There's no need to do another color. ANYthing that you want to be a variable : you name it in the .txt files of your dataset. Name what you see : black latex, etc. And because you wrote "black" before the training, then it's a variable. And people will be able to generate your objects, while still being able to modify their color.
For more flexibility, everything that should be possible to be changed by the final user, has to be written down in your dataset.
Cheers.
@hansolocambo thanks :D I´ll try that out someday soon :D
train it 1.5 SD and then it will works
it´s trained on stable diffusion 1.5 :D
@zerberuszero then it doesn't work
damn, that thing eats tokens as hell, is that normal? costs me like 30 tokens
just saying because OP (tks, btw) talked about the training, so try to check your txt files and make them extremely specific, for example instead of "woman wearing ballgag" add "22 year old brunette woman with curly hair, long nose, .. etc etc" too, because every detail the checkpoint is able to recognize is then removed from the 'ballgag' concept, otherwise it starts thinking that every ballgag also implies everything else that is not named. that's why variation of images is good (so the training can flush out this bias), but since Loras and TI use fewer images to train it helps a lot to be specific
thanks for the hint, I´ll try that someday, soon :D