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: 4000Steps, AdamW8bit, 1Batches, [768], LR 3e-5, WeightDecay 1e-2, Rank32, Repeats 3, No Axis Flip
[Trained with both Danbooru Tags and Natural Language. Tags weight1.2, Caption Weight 0.8]
FAQ
Comments (8)
What tool did you use to clean the image?
To clean them up I use Klein 9B in ComfyUI.
Load the image, mask the target area and a simple "remove this/that" prompt.
I use a detailer node to only edit the masked zone and avoid the awful Flux VAE color shift on the whole image.
I tired automatic masking with SAM3.1 but IMO it is too sloppy and require too much manual intervention to be worth it.
For Upscale is just SeedVR2 7B at 1280p.
@Weird0 those are lot of model for my laptop to handle and i only have 6gb vram. But thanks bro for replying and giving me the steps. Also is it possible for me to get the workflow for the cleaning if possible.
@ArtificalEverything Workflow is hella messy and I had to vibecode a custom node for it.
Im very willing to clean it up and share it but I dont really know what to do since it needs the niche node I patched together (Mask to SEGS).
@ArtificalEverything
Nevermind. Im a dumbass and fixed it in 10minutes. Im adding an article with the workflow, you should find it on my profile under articles.
Just needs impact pack if im not mistaken.
👍
@Weird0 Thanks so much. this will help me so much
Holy shit! The madman actually did it! V0.3 is a HUGE improvement!! And I think you can take it even further. For example, correcting typos like "burised body" on the Tomura picture and "burised" on the Toga one for better prompt adherence, it's "bruised". Otherwise I think we're getting there. I've got a lot of free time and I know how to clean up images so if you need any help with the dataset hit me up, I'm more than willing to help
10/10 Comment. Snorted it like cocaine.
Yea Im a dumbass and typo'd the prompt apparently, If/when I make a V0.4 I'll fix that lol.
Dataset cleaning I think my sloppy Klein-Edit job does the trick.
For me the LoRA is rly good rn. The only things left are one that are hard to fix or I dont know how:
-NSFW performance, rn its not great and the only way to try to fix it would be adding AI generated NSFW images made with V0.3 cranked into the dataset.
-Better compat with finetunes, lowkey no idea how to do that while avoiding overtraing for base/aesthetic folks.
Thanks again for your very helpful advice and encouragement. 👍🎉🎉