>>938487173
Since you're using civit, lower the steps to 23 to save some buzz. Cfg is fine (maybe go up to 7 if you want, keep testing), change the sampler to euler a (costs less, decent result) or to dpm ++ 2m karras (costs more, better results). The safe bets for models that I personally use are wai nsfw illustrious and ntrmix
>>938487209
I've had one up for months and I'm surprised they haven't deleted it yet. Good lora