The idea in plain language.
Different points on the same trajectory should map to the same clean result.
How it works
The model is trained so that inputs at different noise levels along a common trajectory produce consistent outputs. It can learn from a pretrained diffusion model or be trained directly with an appropriate objective. This creates a path toward generation with fewer evaluations, while retaining the option to use multiple steps when that improves sample quality.
What to keep in mind
Fewer steps trade off against sample quality and training choices. The reported results are task-specific and do not eliminate the cost of training the generative model.
Source: Consistency Models. The original manuscript contains the methods, experiments, figures, and references. An arXiv posting date may follow an earlier conference publication. Read the linked record for version history.