Consistency models (Song et al. 2023) learn a function f_θ(x_t, t) that maps any noisy point on a diffusion trajectory directly to its clean endpoint x_0. The self-consistency property requires f_θ(x_t, t) = f_θ(x_s, s) for all t, s on the same PF-ODE trajectory — every noisy version of the same image maps to the same clean output. This enables one-step sampling (unlike diffusion models requiring 10-1000 steps), with quality improvable via multi-step refinement. Training uses either consistency distillation (from a pretrained diffusion model) or consistency training (from scratch). Here trajectories converge to fixed points in one shot.