One-step Generation

Producing a result directly from noise with just one network forward pass, solving diffusion models' slow sampling.

One-step generation means a generative model produces a sample with just a single network forward computation (1 NFE, or number of function evaluations = 1). GANs, VAEs, and normalizing flows are naturally one-step generators; diffusion models and flow matching, by contrast, need tens to thousands of iterative steps starting from noise, giving high quality but at a slow pace. For robots, every extra step in an action head adds control latency, so compressing a diffusion or flow-based policy down to one step is an important direction. There are three common approaches: distillation, training a one-step student from a multi-step teacher, as in consistency distillation; changing the training objective so the model learns to 'jump straight to the endpoint' in one step, as in consistency models and MeanFlow; and straightening the generation path, as in rectified flow. The usual cost is a small quality drop or a more complicated training setup.

ExampleConsistency Policy distills Diffusion Policy into a consistency model, running an order of magnitude faster than the fastest existing methods and fitting on a laptop GPU; MP1 uses MeanFlow to generate an action in one step, at about 6.8 ms of inference time.

Also called

1-NFE Generation, One-step Sampling

Related

Denoising Steps · Consistency Model · MeanFlow · Rectified Flow · Consistency Policy · Inference Latency

Sources

Consistency Models (arXiv:2303.01469)
Consistency Policy: Accelerated Visuomotor Policies via Consistency Distillation (arXiv:2405.07503)
Mean Flows for One-step Generative Modeling (arXiv:2505.13447)