Distillation

The large model becomes the teacher. Instead of learning from raw data, the student learns to reproduce what the teacher produces — and can be built to reach a comparable result in far fewer steps, or with fewer parameters, or both. That is the entire mechanism behind a fast variant that costs a fraction of its parent's time.

What you lose is usually the tail. Common subjects and clean briefs come back nearly indistinguishable; the gap shows on unusual combinations, fine detail and prompts that need careful adherence, because those are precisely the cases the student had least opportunity to imitate.

Distilled models also behave differently on the controls. They are trained to work at low step counts and low guidance, so settings carried over from the full-size model tend to produce over-cooked, over-contrasted results rather than better ones.