v3:
Steps: from 12 - more is better
For quality you can use dpm++ 2s ancestral. Something else if you want better speed. Default euler ancestral should work well.
Description
My attempt to train (targeted distillation) a model on my 12GB GPU (yes, it's possible)