LAION-5B

LAION-5B is an openly released index of roughly 5.85 billion image-text pairs harvested from Common Crawl web data. Rather than hosting images, it provides image URLs paired with author-supplied alt-text captions, along with metadata such as CLIP similarity scores, image dimensions, language tags, and NSFW predictions. About 2.3 billion pairs are English, with a further large multilingual portion and a set of pairs whose language could not be resolved.

What’s in it

Candidate pairs were filtered by computing CLIP embedding similarity between image and caption and discarding low-agreement matches, leaving the retained set of ~5.85B examples; the collection is commonly accessed in shards via tools like img2dataset. The dataset was created by LAION (Large-scale Artificial Intelligence Open Network), a German non-profit, and became widely known as the training data behind open text-to-image models such as Stable Diffusion, as well as open CLIP reimplementations.

Tasks, license & access

LAION-5B is used to pretrain image-text contrastive models (CLIP-style), text-to-image generators, and multimodal retrieval systems. The metadata is released under a CC BY 4.0 license, but users must download the underlying images themselves and are responsible for the varied copyright status of each source. After the original release was withdrawn over CSAM concerns, a cleaned successor, Re-LAION-5B (~5.5B pairs), was published in 2024 with known illegal links removed.

• Modality: image-text pairs (URLs plus captions and metadata)
• Scale: ~5.85B pairs (Re-LAION-5B ~5.5B)
• Tasks: CLIP pretraining, text-to-image, multimodal retrieval
• License: CC BY 4.0 metadata; images under source terms
• Creator: LAION e.V.

As one of the first openly documented web-scale multimodal corpora, LAION-5B was pivotal in democratizing large image-text model training, while also surfacing hard questions about consent, copyright, and content safety at scale.