Skip to content

Generative Powers of Ten

Generate consistent content across extreme semantic zoom

Visit official site

What Generative Powers of Ten is

The method is described as generating consistent multi-scale content with text to image models, allowing extreme semantic zoom from a forest landscape to a macro shot of an insect on a branch.

What you can do with it

  • Produce content that stays consistent across scales
  • Render a continuous zoom video
  • Explore the scales of a scene interactively
  • Guide the zoom level to match an input image

Method notes

  • Joint multi-scale diffusion sampling encourages consistency
  • Each scale follows a different text prompt
  • Varying the seed changes results for one prompt
  • Comparisons with super-resolution and outpainting are described

Pros & cons

✓ What we like

  • Zoom levels can be guided to match an input image
  • Changing the seed gives different results for one prompt

! What to watch out for

  • Consistency is described across jointly sampled scales
  • Each scale needs its own text prompt

FAQ

What does it produce?

Consistent multi-scale content and continuous zoom videos.

How are scales kept consistent?

Through joint multi-scale diffusion sampling that keeps each sampling process intact.

Last reviewed: 2026-11-15

More AI image generation tools

View all →

How we review