Sander Dieleman · 2026-05-06 · notable
Sander Dieleman: Learning the Integral of a Diffusion Model
DeepMind's Sander Dieleman maps the fast-moving flow-maps literature: networks that learn the integral of a diffusion ODE, jumping between any two noise levels in one step instead of denoising step-by-step.
A taxonomy of flow maps — the family of methods replacing iterative diffusion sampling with single-jump prediction.
What is it?
A long technical post by Google DeepMind research scientist Sander Dieleman covering flow maps, networks that predict any point on a diffusion sampling path from any other point, generalizing consistency models, MeanFlow, Terminal Velocity Matching, and similar few-step samplers.
How does it work?
Dieleman organizes training methods around three consistency rules — compositionality, Lagrangian, and Eulerian — and surveys both distillation approaches (training from a frozen diffusion teacher) and from-scratch approaches (marginal-from-conditional learning, self-distillation). He grounds each method in the underlying ODE formulation and notes which ones power production systems like FLUX.1 and Wan 2.2.
Why does it matter?
Flow maps are how modern image, video, and audio generators get to one- or two-step sampling without quality loss. This post is the first end-to-end map of a literature that was scattered across a dozen 2025–2026 papers, so it is the article anyone working on fast generative sampling will be sent.
Who is it for?
ML researchers and engineers working on diffusion models, generative modeling, or fast inference.