@RapidFire , in the animation I posted to compare how this model behaves to my others, and without, it's that four image grind animation, the bottom right is using my other model.

It's not too complicated, but there are a decent number of steps and tools to share. When I have time, I'll compile it. A lot of the stuff I'm using, I downloaded a while ago, so I'm not actually sure where I got all the tools. I'd have to do an easter egg hunt to find out.

In short though, I use kohya_ss for the actual training, I use imgbrd-grabber to grab the images and their tags from gelbooru, I tag the images like so. (note, I don't include <> in the actual tag files, that's just to show what's in that section.

<trigger tags>, score_9, source_<source type, maybe anime>, source_<rating, ie explicit, general, etc>, by <author name>, <ip sources, ie, maybe Final Fantasy>, <characters, ie, maybe Tifa Lockheart>, <general tags>

I drop any meta tags like absurdres and whatnot.

I'll do from 25 to 100 images for each concept and enough repeats to make it 200 total steps for each concept per epoch, then I'll do 12 to 24 epochs depending on how complicated the concepts are, and how many concpets I do.

I usually use a prodigy optimizer.

If I have time, I'll figure out the specific commit of kohya that I use, and post a copy of my json. The json I use doesn't work with the newest version, otherwise I'd post it now.