Generating outputs in a small number of inference steps rather than many, enabling faster generation at the cost of potential quality trade-offs.