V0.3 Important Notes: PSA: As you can see in the Vlad image, this version sucks balls with dark scenes. Trained with Danboru tags and Natural language, pain in the ass to manually write/edit but holly smokes, WAY better prompt adherence. Bumped to 4000 Steps with lower Learning rate. 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 108 Illustrations from <a target="_blank" rel="ugc" href="https://www.pixiv.net/en/users/17151277">mamimi</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 M4MIMI. 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>❤️)

MAMIMI-STYLE-ANIMA

Main Settings: 4000Steps, AdamW8bit, 1Batches, [768], LR 4e-5, WeightDecay 1e-2, Rank32, Repeats 2, No Axis Flip [Trained with both Danbooru Tags and Natural Language. Tags weight1.2, Caption Weight 0.8]

Files with SHA256: ff6f811bfbe24db6a008723d44ff5219a276dffb40dbd5996ec682e7f1553a6a