Learning the Integral of a Diffusion Model

The article explores flow maps as a method to accelerate sampling from diffusion models by training neural networks to directly predict paths through data space rather than iteratively estimating tangent directions. Flow maps can predict any point along a path from any other point on that path, offering benefits beyond faster sampling including more efficient reward-based learning and improved sampling control. The author aims to clarify the various approaches, formalisms, and terminology surrounding flow maps, which have recently become a popular research topic despite the confusing landscape of different implementations.