Meant to reinforce NAI's lackluster Miko training, and reintroduce the concept to finetunes.
Trained on 445 images featuring solo miko girls
DIM 16 trained for 11250 steps
Description
Text encoder only, trained for 2515 steps on the same dataset as v1. (unet learn rate set to 0).
Seems to produce good results, but I haven't tested it much.