V0.2 Important Notes: PSA: This version (0.2) is a bit wonky and likes straight-on shots too much, I am aware of the issue but fixing it means reworking and re-labeling the dataset, but man im tired boss. Trained with Danboru tags and Natural language, real pain in the ass to manually write/edit but, better prompt adherence. Bumped to 4000 Steps. A style LoRA for Anima (by extension its finetunes) by <a target="_blank" rel="ugc" href="https://huggingface.co/circlestone-labs">CircleStone Labs</a>. This LoRA was trained on a curated dataset of 76 Illustrations from <a target="_blank" rel="ugc" href="https://www.pixiv.net/en/users/1095515">Pepe Jonasan</a>. All images were treated to remove text, any censoring and most symbols, then batch tagged with <a target="_blank" rel="ugc" href="https://huggingface.co/Mooshie/convnextv2_huge.dbv4-full">ConvNextV2 Huge Full</a> 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 P3PEJONASAN. Also this was trained on my chef's kiss AMD 7900XT, (under Linux with <a target="_blank" rel="ugc" href="https://github.com/ostris/ai-toolkit/">Ostris's Ai-Toolkit</a>❤️)

PEPEJONASAN-STYLE-ANIMA

Main Settings: 4000Steps, AdamW8bit, 1Batches, [768], LR 3e-5, WeightDecay 1e-2, Rank32, Repeats 2(TAG) and 2(CAPTION), Tags X Axis Flip [Trained with both Danbooru Tags and Natural Language. Tags weight1.2, Caption Weight 0.8]

Files with SHA256: 144d148d20036f1abd199fedc864575aa9190c24887d1282bf95e75727ed0f51