>>2841621It is very hard to get good results on pony. There seems to be this weird effect where before we can make a lora that looks good or accurate, we need our base (be it either base model, or a base prompt/lora) to be looking reasonably aesthetic. Doesn't make a lot of sense but that is my experience.
I use kezime lora at low strength to 'stabilise' other loras. Why kezime lora turned out so great compared to others I have no idea. The dataset isn't particularly big or diverse and there is nothing different about the hyperparameters of the training.
I still gen on pony, I've tried autism a bit and I don't like the way it pulls the style into more generic anime, even if it can give more accurate results for artist styles. It also seems to lose its stylistic influence when not prompting 1girls, which is fine for /e/ but not for /h/.
>>2841683Hmm... 9 8 7 source anime seems to get a fair bit closer to the style. I'll run this through some other loras and see if it applies broadly.
