Skip to content

Align Your Steps

Optimise sampling schedules for diffusion models

Visit official site
Category AI models

What Align Your Steps is

The method is described as a way to optimise sampling schedules for diffusion models, using stochastic calculus to find schedules suited to specific solvers, trained models and datasets.

What you can do with it

  • Optimise sampling schedules with stochastic calculus
  • Tailor schedules to a chosen solver, DM and dataset
  • Minimise the KLUB term over time discretisation
  • Use the provided optimised schedules in plug-and-play form

Application scenarios

  • Raise image quality on CIFAR10, FFHQ and ImageNet benchmarks
  • Generate finer text-to-image results in Stable Diffusion 1.5
  • Reduce colour distortion in video generation

Who it is for

  • Researchers and developers raising output quality while keeping real-time use
  • Professionals seeking efficient sampling in image, video and 3D synthesis
  • AI enterprises and research institutions looking for a route to higher model performance

Pros & cons

✓ What we like

  • Effectiveness is validated on image and video generation benchmarks
  • Optimised schedules are available in plug-and-play form

! What to watch out for

  • Each optimised schedule is tied to a specific solver, DM and dataset
  • The method works by minimising the KLUB term over time discretisation

FAQ

What does the method optimise?

The material describes optimising the sampling schedule, or time discretisation, of diffusion models.

How is it started?

The steps name a quick-start guide, Colab notebooks and a comparison with hand-crafted schedules.

Last reviewed: 2026-11-15

More LLM API platform tools

View all →

How we review