The setup: why flow matching, and where it hurts

In standard flow matching, we train a network \(v_\theta(x_t, t)\) to match a target instantaneous velocity field \(u(x_t, t)\) along an interpolation between data and noise:

At inference, we integrate the learned ODE \(\dot{x}_t = v_\theta(x_t, t)\) from noise to data. The quality of samples depends on how many ODE steps we can afford — more steps reduce discretization error, but cost more.

This creates a tension. Few-step or one-step generation is hugely valuable for deployment, but the model was trained to be an instantaneous derivative, not a finite-displacement operator. Distillation approaches (consistency models, progressive distillation, rectified flow) try to bridge that gap after the fact.

The mean-flow idea

The paper’s key move is to define and learn the average velocity:

i.e., the displacement-per-unit-time over a finite interval \([r, t]\), rather than the derivative at a single instant. The point: if you have \(\bar{u}\) exactly, then one Euler step over the whole interval is exact:

So one-step generation is no longer an approximation — it’s the natural inference rule for a model that learned the right object.

The identity that makes training tractable

The trick is that you can’t differentiate through an integral target cheaply. The authors derive a self-referential identity by differentiating the defining equation with respect to \(t\):

The right-hand side mixes the instantaneous velocity (from a standard flow-matching teacher) with a time-derivative of the mean-flow network itself. That makes training feasible: you optimize the network to satisfy this consistency identity via JVPs.

This is structurally reminiscent of consistency-model training — both ask a network to satisfy a self-referential invariant — but the invariant here is grounded in calculus, not in a discretization scheme.

Why I find this interesting

A few reasons it stood out at our journal club:

1. It reframes one-step generation as the primary object, not a distillation artifact. The hierarchy “train a continuous model, then distill to a few steps” is replaced by “train the few-step model directly.”
2. It is geometric. The mean-flow \(\bar{u}(x_t, r, t)\) is the natural finite-displacement operator along a transport path. This connects to my interest in flow-based generative geometry — what do the curvature, density concentration, and asymmetry of transport paths tell us about representation quality?
3. It hints at a state-modeling reading. If you think of generative modeling as inferring a latent state and rolling it forward, an instantaneous-velocity model is doing infinitesimal state updates, while a mean-flow model is doing calibrated finite-horizon predictions — closer to how world models behave.

Caveats and open questions

• Training stability with JVP-based self-referential losses is delicate; the paper reports tricks that matter.
• It is unclear (to me) how mean flows interact with conditional generation, especially with strong conditioning signals where the velocity changes rapidly near the data manifold.
• The connection to consistency models and to rectified flow deserves a careful theoretical comparison — they each impose different invariants on the same underlying object.

Why I wrote this up

I run Tech Talks as an AI journal club for critical discussion of recent generative-modeling, vision, and medical AI papers. These posts are an attempt to write up sessions in a form useful to others — short, opinionated, and focused on what I think matters in the paper rather than reproducing the table of results.

If you’d like to suggest a paper or join a session, reach out.

References

Geng, Z. et al. (2025). Mean Flows for One-Step Generative Modeling. [arXiv]