V6: almost the same but smaller and should be better, the longer i look for differences the less I see, less shiny skin and orange peel. I tried makin ILL version but it obliterates cutie marks and immediately overfits so eveone receives snout. If you use spikezilla use v5 or more detailed prompt for him.
V5: Battle on shiny skin, sparkle, glinting and lens flare is over. It took only 48 tries to understand that i was fighting the dataset.
V4: I don't want to talk about it.
V3: Last successfull version before V5.
V2: V1 is abysmal so there it is
V1: by hooves-art is actually not part of training data but it's style known to checkpoint, and if used with lora helps with generalization but watermark does not appear. But effect on stability is questionable. "by hooves-art" interferes with model knowledge.
Works on Illustrious models too.