V0.3 Important Notes:
Trained with Danboru tags and Natural language, pain in the ass to manually write/edit but holly smokes, WAY better prompt adherence.
Upped the dataset size from 46 to a nice 69 (scraped the bottom of the web for valid images)
Bumped to 4000 Steps with lower Learning rate, no flipping and less repeats.
Fun bonus, example dataset images before fixing (Editing, cropping, upscaling):
A style LoRA for Anima (by extension its finetunes) by CircleStone Labs.
This LoRA was trained on a curated dataset of 69 Illustrations from Kohei Horikoshi. All images were treated to remove text, any censoring and most symbols, then batch tagged with ConvNextV2 Huge Full and manually tagged for fixing/adding tags.
This Lora is best used at 1.0 strength, imo. Trigger word is very optional, but hey, its H0RIKOSHI.
Also this was trained on my chef's kiss AMD 7900XT, (under Linux with Ostris's Ai-Toolkit❤️)
Description
Main Settings: 2000Steps, AdamW8bit, 4Batches, [768,1024], LR 5e-5, WeightDecay 1e-2, Rank32, Repeats 3
FAQ
Comments (3)
Needs just a bit more steps/repeats. Perhaps more images in the dataset. Have you considered using flipping or cropping to artificially increase the number of images? cuz otherwise this has mad potential, the Shigaraki pic specifically got the painterly style down but it's not cooked enough
I am going to try experiment a bit with more repeats/steps, maybe LR as well.
I find getting hands on recent, colored illustrations from Horikoshi to be tricky, I am thinking of volumes cover scans if i end up needing a bigger dataset.
I will look into the dataset settings (crop, flip, etc) thank you for the feedback 👍
Same feelings, maybe it needs a bigger dataset? Getting mixed results, some generation fine, other not so much