●builder2-step high-fidelity generation cuts inference cost roughly 4x versus 8-step baselines, directly relevant for latency-sensitive image generation APIs.
●researcherTeacher-aligned GAN targets and step-decoupled parameters are concrete techniques addressing the known quality cliff below 4-step distillation.