Rectified Flow (Liu et al. 2022) straightens transport trajectories through iterated "reflow." Starting from random couplings (π₀ × π₁), the first flow learns velocity v_t(x) = E[X₁ - X₀ | X_t = x] — this may have crossing paths. Reflow re-couples endpoints: generate (X̂₀, X̂₁) pairs from the learned flow, then train a new flow on these paired samples. Each reflow step straightens trajectories, reducing crossings and enabling fewer integration steps. After k reflows, the k-rectified flow increasingly supports single-step generation. This is the mechanism behind Stable Diffusion 3 and FLUX models.