I'm quite the novice, but I created a Lora on Anima Base for Gyokai's signature style, specifically from the game Dohna Dohna. Due to the nature of the game and its artwork it is far more well balanced for girl based outputs than boys. I ran a total of 12 epochs, however epoch 10 functions the best at around 1.0 weight compared to epoch 12 having a stronger expression of the style but breaks down anatomy more for keywords like "chainsaw". Epoch 10 is what I chose to publish as it responds way better to raising and lowering the weight of the Lora, so if you find your output lacking the style or having too much of it feel free to lower/push that value. I did not tag the dataset with character names, so unless base anima model understands who "kirakira" is this lora wont help with that. This was done on purpose because I wanted the style not a dohna dohna character set for the lora. However I may simply have been mistaken and I could update this for an eventual V2 where I retrain and use character tagging for the recognizable characters from the game. All my tests were run at 40 steps 5cfg on the Anima Base model without any speed up Lora. I'm not an expert at this at all, so any feedback is appreciated. tldr: @gyokai tag (lowercase) Start at Lora weight 1.0, adjust as needed Tested at CFG 5 with 40 steps. Sampler: er_sde recommended (Anima's documented default). Be sure to prompt using the correct anima tagging order. Can optionally verbosely describe the scene after the tags, this is not lora specific but rather Anima specific. Anima's tag order: <code>[quality/meta/year/safety] [1girl/1boy] [character] [series] [artist] [general]</code>. The <code>@gyokai</code> token goes in the artist position, not at the start of the prompt.

Files with SHA256: 4a96bc5587109ee008a51d84102b21fb0ae2f68953eb2d53cb1c8055fe0de60e