Flow Maps: Predicting Integral Paths for Faster Diffusion Sampling

Flow maps are a recent extension of diffusion models that allow a neural network to predict any point on a diffusion path from any other point on the same path, bypassing the iterative denoiser steps that normally drive sampling. The concept builds on the bijective mapping between noise and data samples that deterministic sampling algorithms exploit, and it is formalized in the taxonomy proposed by Boffi et al. Instead of predicting the clean input \(\hat{x}_0\) at each timestep, flow maps predict the integral of the tangent direction, enabling a single forward pass to jump from a noise sample to a target data point. This approach can accelerate sampling, support reward‑based learning, and improve sampling steerability, and it can be trained from scratch or distilled from existing diffusion models. The article also contrasts stochastic sampling, which draws random samples at each step, with deterministic sampling, which follows a smooth, deterministic trajectory guided by the denoiser’s tangent predictions. Flow Matching, a popular deterministic sampler, uses a velocity field \(v_t\) that is linearly related to \(\hat{x}_0\) and \(x_t\), preserving marginal distributions while collapsing conditional distributions to delta functions. Training flow maps involves back‑propagating through the integral of the tangent direction or using denoiser predictions, and the literature contains many formalisms and terminology that can be confusing for newcomers. The blog highlights that flow maps are gaining traction for their speed, steerability, and potential for efficient reward‑based learning, but also warns that the field is still evolving and requires careful implementation.