Skip to content

SUPIR

Restore images with text-guided generative priors

Visit official site

What SUPIR is

The method is described as an image restoration approach that uses generative priors, multimodal techniques and model scaling. It was trained on a data set of 20 million high-resolution images with descriptive text annotations and supports text-prompt-guided restoration.

What you can do with it

  • Restore images with generative priors and model scaling
  • Guide restoration with text prompts
  • Apply negative quality prompts to raise perceptual quality

Application scenarios

  • A photographer restores old photos to their original quality
  • An advertising team repairs product images for visual effect
  • A museum restores ancient artwork images to protect cultural heritage

Method notes

  • Multimodal techniques and generative priors are combined with model scaling
  • The training set holds 20 million high-resolution images with descriptive text annotations
  • A restoration-guided sampling method is used to suppress fidelity issues

Pros & cons

✓ What we like

  • Each of the 20 million high-resolution images carries a descriptive text annotation
  • Restoration can be controlled through text prompts

! What to watch out for

  • Negative quality prompts are part of the described method
  • Restoration-guided sampling addresses fidelity issues in generative restoration

FAQ

How large is the training data set?

The material states 20 million high-resolution, high-quality images, each with a descriptive text annotation.

What guides the restoration?

Text prompts, negative quality prompts and a restoration-guided sampling method are described.

Last reviewed: 2026-11-15

More AI image generation tools

View all →

How we review