Should only need "fkwarning" as trigger word(?) Trained with both tags and NL appended after the tags. Described the character(s) features in detail in the NL while keeping the tags as usual, omitting those features that were supposed to be baked. Interestingly, the lora seemed to still absorb the features just fine. Tagged some artists in the dataset, but not fluff-kevlar. This version seems to be the most accurate on Anima yet, not sure if overfit, but it's leaking more than previous version... Use strength 1.0 or close to it.