What Datasets Does LAION Offer?

Updated

LAION provides four main dataset resources: LAION-400M (400 million English image-text pairs), LAION-5B (5.85 billion multilingual CLIP-filtered image-text pairs), LAION-Aesthetics (an aesthetic-quality subset filtered from LAION-5B), and the Re-LAION 5B release dated 30 August 2024. All are offered by LAION as a non-profit organization that describes itself as "100% non-profit" and "100% free," with the stated goal of liberating machine learning research by providing datasets, tools, and models.

Dataset Overview

Dataset Scale What it contains
LAION-400M 400 million pairs Open dataset of English image-text pairs
LAION-5B 5.85 billion pairs Multilingual image-text pairs, filtered with CLIP
LAION-Aesthetics Subset of LAION-5B Images filtered by a model trained to score aesthetically pleasing images
Re-LAION 5B Released 30 Aug 2024 A re-release of the 5B dataset (see the project page for scope details)

How the Datasets Relate

LAION-5B is the broad foundation: 5.85 billion multilingual image-text pairs that have passed CLIP filtering. LAION-Aesthetics is not a separate crawl but a subset of LAION-5B, narrowed by an aesthetic scoring model. LAION-400M is the earlier, English-only collection at 400 million pairs. Re-LAION 5B, published on 30 August 2024, is the most recent release listed on the site.

This layering matters when choosing: if you need multilingual coverage at maximum scale, LAION-5B is the base resource; if you need a smaller, quality-filtered pool for image-generation or aesthetics-related work, LAION-Aesthetics is the derived option.

Beyond Datasets

LAION also lists CLIP H/14, described as the largest CLIP vision transformer model, alongside its datasets, tools, and models. The organization frames its work around open public education and reusing existing datasets and models to use resources more environment-friendly.

Practical Notes Before You Start

  • Check the release you need. The site lists Re-LAION 5B as a distinct release from the original LAION-5B; confirm which version fits your project before downloading.
  • Plan for scale. A 5.85 billion-pair dataset is not a small download; storage and processing capacity should be settled first.
  • Use the site's request channels. LAION provides a Dataset Requests entry point and an FAQ, which are the appropriate places to resolve access and usage questions rather than assuming terms.
  • Licensing and content review are your responsibility. The site's own pages (including its GEP Policy and Privacy Policy) are the reference points; verify conditions for your specific use case.

For the current status of any dataset, the LAION Projects page and Blog are the authoritative sources, since release details can change after publication.