One-step Generation in the Post Diffusion Era

A review of one-step generative modeling beyond diffusion-time iteration, covering Consistency Models, CTM, MeanFlow, DMD, and the 2026 Drifting Models framework.

Overview

This post reviews the shift from iterative diffusion and flow models toward one-step generation.
After summarizing prior approaches such as Consistency Models, CTM, MeanFlow, and DMD, it focuses on Drifting Models, which replace inference-time dynamics with training-time distribution evolution and learn an explicit drift field to align generated and data distributions. —