About this version Not much new, I trained it from 17K steps up to 60K steps the low model. I'm not sure if anything got better but give it a try. I cleaned up a few captions but otherwise dataset is the same. Maybe around a 30% improvement on the loss based on the tensor graph. I wanted to train the images from 512 to 1024 but for some reason I cannot do block swap anymore on my pc...maybe I will try it with runpod later this week.