Video walkthrough · Talk

One-step generative models: the MeanFlow line, paper by paper

A four-paper walkthrough of He Kaiming's group compressing diffusion's hundreds of steps down to one — MeanFlow, improved iMF, drift-based modeling, and FID-as-a-loss.

20:54 15 slides Jun 28, 2026

What the talk covers

  • MeanFlow — an average-velocity field plus an identity and a JVP turn many-step sampling into a single step (FID 3.43).
  • iMF fixes the target bias of MeanFlow and pushes quality much further (FID 1.72).
  • Drift-based generation drops the ODE and trades inference iterations for training iterations (FID 1.54).
  • FD-Loss trains FID itself as the objective, closing the loop between metric and loss (FID 0.72).