pls define "medium resolution"
preferably actual number of pixels
512x512 total pixels. But it doesn't matter, the preview version has been specifically trained to work at 1024x1024 total pixels. It's just that it wasn't trained very much at that resolution. The final model is going to be trained for a long time at 1024.
The final model is going to be trained for a long time at 1024.
Please consider doing buckets ranging from 512x to 1536x, so on average its 1024x but can go above and below more accurately.
512x512 total pixels. But it doesn't matter, the preview version has been specifically trained to work at 1024x1024 total pixels. It's just that it wasn't trained very much at that resolution. The final model is going to be trained for a long time at 1024.
Please, if you can, train on the entire dataset using 1024px (not just a subset). The model is really good but it's still learning small details like clothes logo or emblems and it probably needs higher resolution to learn them well, for example, it can't generate well the Tokiwadai school uniform emblem (there are 8200 "tokiwadai_school_uniform" tagged images on danbooru). The same with Suzumiya Haruhi uniform and Beatrice's butterfly pupils from Re:Zero, and these are really famous characters, so less famous characters are probably the same.
Tokiwadai school uniform example:
Anima:
Noob: