They're not 1 dim. They're trained on less blocks of the model. This one is trained on two

It's mostly just flux being as good as it is. Just take any 15-20ish decent images. Most of these only have the trigger word as the caption, only my Sweet_Anita one has regularization images. Learning rate is 0.0002. Pretty much anyone can do the same if they're a little bored. I would suggest running training on selective blocks though, I've found those ones are a decent bit better.