Wanted to try training a character LoRA, so here's Alice Thymefield, upcoming S rank from Zenless Zone Zero.The keyword is "zzzatf." There were already some noticeable associations that Illustrious has with "alice_thymefield," so I picked an acronym for "zenless zone zero alice thyme-field." For simplicity, that's the only tag associated with the character, so the outfit might vary a little bit, but it can usually be recovered by gently nudging the model in the right direction with terms like "skirt."This LoRA was trained on Illustrious-XL-v1.0, but I've found it works best with models like <a rel="ugc" href="https://civarchive.com/models/140272/hassaku-xl-illustrious">Hassaku XL Illustrious</a> and <a rel="ugc" href="https://civarchive.com/models/1307857?modelVersionId=1723898">Unholy Desire Mix</a>.The LoRA was trained on a dataset of 78 manually selected images limited to a training resolution of 1024,1024 and run for 40 epochs and 1160 steps in total with a batch size of 4, with 29 steps per epoch (I don't know where kohya gets that epoch step count from, it's about 10 higher than the naive calculation of 19.5). The learning rate for the was 10e-4 for the unet and 5e-5 for the text encoder. The AdamW optimizer was used with a cosine scheduler.

Files with SHA256: 2db94d184781b1eac1bf375245632fa97643ca9ffdb226f1a361bc939fc30c90