>>60900918kinda true yeah
way too many people just accept dogshit loras because it's all that is available since people don't want to take the time to train shit
so you end up with anime char loras with a sub 20 image dataset, tagged with natural language and trained on some shitmix, which ends up barely working on most models and quickly fries lora style stacks
the old fat loras used to destroy poses and background, which was mostly fixed on bgvex, but style issue was introduced with symix
though with more overpowering styles it's masked a bit
I just use the three at 0.3-0.4 to cope
