Simplifying and Improving
On-Policy Distillation for Few-Step Diffusion

Simplified Matching Distillation (SMD) enables few-step generation without an auxiliary fake-score model, adversarial training, or the need for the student to represent dense teacher predictions.

We validate SMD across image, 3D, video, and audio–video generation, including editing and multimodal conditioning, and extend it to cross-model, multi-teacher, and causal video distillation.