What’s Out There
← Back to swiping
Images & videoFreeNonprofit

LAION-5B

LAION e.V.

5.85 billion image-text links that trained Stable Diffusion, then got pulled over illegal content.

A list of image URLs and their alt-text captions, scraped from Common Crawl and filtered with a CLIP model so the text matches the image. It made open text-to-image models possible. In December 2023 LAION took it down after Stanford researchers found links to child sexual abuse material; a cleaned Re-LAION-5B followed in 2024.

Access

Cost
FreeOriginal release withdrawn; a cleaned successor (Re-LAION-5B) is distributed under LAION's terms.
License
Open licenseMetadata under CC BY 4.0; images keep their own copyrights
Sign-up
No sign-up
How you get it
Through the provider's siteParquet

What's in it

One row is
One row per image URL with its caption and similarity metadata
Covers
Global web; about 2.3 billion English and 2.2 billion other-language pairs
Years
Released March 2022; withdrawn December 2023
Updated
One-time release
Size
5.85 billion image-text pairs (metadata only; images must be fetched from their URLs)
image URLcaption textCLIP similaritywidth and heightlanguageNSFW and watermark scores

Quality

2 of 5

Landmark in scale and openness, but uncurated: it contained illegal material, personal photos, and copyrighted work, and many URLs are now dead.

  • Taken down in December 2023 after CSAM links were found
  • Subject of copyright litigation, including a German case brought by a photographer
  • Link rot means a large share of images can no longer be fetched

Connects to

  • Common Crawl: Source of the web pages from which image links and alt text were extracted
  • Open Images V7: A human-verified, openly licensed alternative at much smaller scale

Questions it could answer

  • What share of the original URLs still resolve today?
  • Which websites contributed the most image-text pairs?

Similar data sets

  • ImageNet

    The 14-million-photo set whose 2012 contest kicked off the deep-learning boom.

    Images & videoStanford University and Princeton University (ImageNet project)Free

  • Dolma

    The three-trillion-token corpus behind OLMo, with the recipe and tooling published alongside it.

    Language & textAllen Institute for AIFree

  • Open Images V7

    About 9 million Creative Commons photos with 16 million boxes and 61 million labels.

    Images & videoGoogleFree

  • The Pile

    The 825 GB open training set that taught the field how LLM data is assembled, and got sued over it.

    Language & textEleutherAIFree

  • Low Altitude Disaster Imagery (LADI) Dataset

    Aerial photos taken by Civil Air Patrol volunteers after U.S. disasters, labeled for damage and features.

    Images & videoMIT Lincoln Laboratory Humanitarian Assistance and Disaster Relief Systems GroupFree

  • Objaverse

    More than 800,000 annotated 3D objects, nearly all under Creative Commons licenses.

    Images & videoAllen Institute for AIFree