>>78574323>I just followed a guide that used autismmixThe problem is that those are always derivative of their source models and tend to get extremely shittier the more complex the mixes get - It was the case for the NAI leak and it remains so for Pony. Unless you're baking a lora for something extremely specific (like the HL+EFF models for 1.5) it's really not worth training on mixes.
If you need to train on a given mix for whatever reason it's always good to prompt a dozen or so images on the base model to check if they got overbaked.
