A character LoRA has one job: make the same face appear in situations it never saw. Overtrain and you get the dataset back, clothes and backgrounds included; undertrain and she drifts into "a woman who looks a bit like her". The settings below are what every Gmanski character is trained with, and they sit in the middle of that range on purpose.
Academy / Characters that stay themselves / Lesson 2
Pro lesson · 7 min
Training a Krea 2 LoRA that generalises
The settings we train every roster character with, how long it really takes on which GPU, how to read the result, and the two failure modes you can see in the first ten test renders.
The rest of this lesson is for Pro members
Pro is $29 for 3 months and includes every lesson, every premium workflow and the whole character roster.