It's a pretty crude model, so for now I recommend using LoRA, created by msw. Yes, I'm breaking the rules a bit, because no one asked me to create LoRA, but nevertheless. I haven't created LoRA for a long time, and the topic with the bug caused by Google was marked as "not planned" by the author of the Colab trainer for creating LoRA, so I'll have to put up with it.
I also decided this way: if there aren't many images with the ninja form, then I'll combine them into one version.