Skip to content

SPRIGHT

A dataset and model for spatial relations in images

Visit official site

What SPRIGHT is

The project is described as a large-scale vision-language dataset and model focused on spatial relations in text-to-image generation.

What you can do with it

  • Re-caption images with more spatial phrases
  • Fine-tune on images containing many objects
  • Generate images that follow spatial relations

Example uses

  • A living room with a sofa right of the fireplace
  • A fruit basket with apples left and bananas right
  • A street with tall buildings on both sides and a fountain

Model notes

  • The dataset was built from 6 million re-captioned images
  • Fine-tuning used 444 images with many objects
  • FID and CMMD scores are stated to improve

Pros & cons

✓ What we like

  • Spatial phrases are increased in the captions
  • A reported high level of spatial consistency is stated

! What to watch out for

  • Fine-tuning is described on 444 object-rich images
  • The stated results concern spatial consistency benchmarks

FAQ

How large is the dataset?

The material states 6 million images were re-captioned.

Which quality scores are named?

FID and CMMD are listed as improved image quality scores.

Last reviewed: 2026-11-15

More AI image generation tools

View all →

How we review