General:

After a break and some experimental loras, I finally got around to doing something I could share. I rewatched SAO, and thought it would be cool to try and replicate the style of the first season (in this case the first half), so the old style of A-1 Pictures, that I find quite nice.

I don't usually train models for this many steps (over 2200), but I really wanted to make it work. It still isn't quite what I wanted, but I'll at least share this, since I'm not gonna train this version more.

Tags and information

Use tags:

sao_style, anime screenshot

It wasn't really trained with characters as main focus, but obviously some of them did appear in the dataset. They work decently well, though it's mostly due to existing knowledge from the model, you'll have to see for yourself.

* additional info *

It appears that some changes I made in the workflow made it save metadata differently than what I used to have, the negative prompt being the obvious difference. I also started using the turbo lora, so that's why the samples have lower steps and cfg. With that in mind, I think you'll get more detail by not using a turbo lora.

Small issue I noticed is that it likes the french braid quite a lot, so if it appears when you don't want it to, just put it into negative prompt

Support:

If you like what you find here, please consider making a donation https://ko-fi.com/hat0ri, thanks!