The idea in plain language.
Resolve unstable training dynamics before scaling the model.
How it works
The paper analyzes parameterizations shared by diffusion and consistency models to identify causes of unstable continuous-time training. It then modifies the process parameterization, architecture, and objective. The experiments examine whether these changes allow larger models to generate high-quality images using a small number of sampling steps.
What to keep in mind
Scaling and image-quality results are measured in particular experimental settings. They do not imply that every continuous-time formulation trains reliably.
Source: Simplifying, Stabilizing and Scaling Continuous-Time 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.