>>2840531Now I'd really recommend to mess with weight decay too, I've been getting vastly better results cranking weight decay up. The standard is .1 I've been playing with it at .3 and it feels way better.
We spent all of 1.5 toe'ing the line on worrying about burning lora's but I don't think I've fried a lora yet in SDXL, it just needs a lot more. And also be mindful with your sets since the base resolution is 1024 it picks up way fucking more in your dataset, so any text stuff, watermarks, etc you don't like it will almost surely pick it up.
