Anima Style Lora based on the manhwa Sexercise. No trigger word, recommended strength from 0.8 - 1.0. I was quite surprised there are no censored artifacts as almost everything was censored expect for nipples from the raws/scanlations.
I plan to expand the dataset and retag for the next update to try and squeeze a little more fidelity out of the model.
If you like models like these, please like and follow!
Description
==First Version==
Trained on 337 images
FAQ
Comments (9)
New work bro. Are you able to do Moby di*k manhwa style?
Hmmm...
I probably can, but collecting datasets from manhwa is rather hard as you have to do tons on manual cropping and find good raws/scanlations at high resolutions to get a good lora.
I'll update you if I start working on the dataset and prepare for training
@donotpurge oh yh it is hard to get a dataset. I understand if you cant but it was just a suggestion. Thank you bro
@donotpurge what do you do about the speech bubbles? Do you just blank them out, or do you tag everything said?
@ImagineBreaker to maintain anima’s text gen, you keep both text and also remove text (keeping only the speech bubble). Then you tag those images accordingly, giving the model the best chance to learn without spitting out random gibberish text.
@donotpurge I'm not sure I fully understood what do you mean you keep text and also remove text? Which is it? Do you tag everything said in the speech boxes to the letter, or do you remove the text by erasing it so you don't have a bunch of untagged text confusing the model?
@ImagineBreaker I would keep both sets of images; the one with text and no text. Right now I'm working on the next version of this lora and I'm using an panelCleaner to scrub the text so I can add it back to the dataset and tag as new images.
And for your second question, I don't tag everything in the speech bubble. Its still experimental and unknown if tagging everything in the speech bubble is worth doing as the text generation of Anima is rather weak and can't handle too much. I will probably do that sometime down the line just to experiment and see.
@donotpurge I see so you use a panel cleaner to remove the text from the text boxes. Mind linking which one you're using?
@ImagineBreaker Sure thing, I'm using this https://github.com/VoxelCubes/PanelCleaner