So I've recently asked by several peeps on my training regarding Anima models, since of course, it's different from Illustrious, and I've come with kind of a series of tips just from my experience.
Settings on the trainer:
I usually leave most of those on default, to be honest, since it uses ai toolkit it carries most of the procedure. The difference seems to be mostly the filesize of the LoRa itself, I'm sure you noticed that I try to keep my models on the small side.
When you get the previews of your epochs, if the images look really weird or off style, I discard the whole thing and re train from scratch, if the epoch is not accurate you're gonna have a real hard time trying to correct it, It is really expensive buzz wise but it will spare you the stress from a bad training. I run my stuff locally so I have the benefit of testing my models freely, it helps a lot.
The prompts in Anima are particularly important, since we're still running on early versions of the checkpoint it takes more effort to take distance from let's say, anime styles, just throw a bunch of (((anime))) in negatives for now, mix and match 'score_8' 'score_9' and so on.
'style parody' doesn't seem as effective as it was in Illustrious so remember that one, you can sometimes get away with using the same prompts as in Illustrious, it depends a lot on the style of the LoRa. Also at least locally the LoRa perform better as you keep generating, takes a couple of tries, might be just on my side tho. Don't get discouraged and try different things!
I do get the feeling that the models we've trained will work better as the checkpoints advance, without needing to re-train them
I share this musings with you freely because we're on the same boat, I mean I sure do love money donations, but as we learn from each other we improve the process overall, so there's that, I also do it out of selfishness, it does benefit me in the long run. True Greed ¯\_(ツ)_/¯