Step Distillation

An acceleration technique that compresses a generative model requiring multi-step sampling into one that produces samples in one or two steps. The student model directly fits the teacher model's output trajectory to skip intermediate steps, at the cost of typically losing some diversity or distributional accuracy.