consistency models

Diffusion produces gorgeous samples but pays for it with dozens to hundreds of network evaluations, because it follows a long trajectory from noise to data. A consistency model collapses that whole trajectory into a single function: from any point at any noise level it predicts, in one shot, the clean datum at the end of the trajectory passing through it. If every point on a given trajectory maps to the same endpoint, then in particular pure noise maps directly to a sample, and you generate in one step.

The defining constraint is self-consistency: the model's output must be invariant along each solution trajectory of the probability flow ODE, with a boundary condition pinning the output at the smallest noise level to the identity. Training enforces this either by distillation, matching the model at adjacent points along an ODE trajectory generated by a pretrained diffusion teacher, or by consistency training from scratch using an unbiased estimate of the trajectory. The result still supports a few-step refinement loop that trades a little compute for higher quality, interpolating between one-step and full diffusion.

Consistency models are a leading answer to diffusion's latency problem, delivering near-diffusion quality at one or a few evaluations, and they sit alongside rectified flow and progressive distillation as the main families of fast samplers.