LAION releases LAION-5B image-text dataset

The German nonprofit LAION released LAION-5B, a dataset of 5.85 billion image-and-caption pairs scraped from the public web, which it described as roughly fourteen times larger than LAION-400M, until then the largest openly accessible image-text dataset. The set was split into 2.3 billion English-language pairs, 2.2 billion pairs spanning more than a hundred other languages, and a further 1.2 billion pairs the pipeline could not confidently assign a language. The metadata — image URLs, captions and CLIP-derived similarity scores — was released under a Creative Commons licence; the underlying images themselves remained under their original, unaltered copyrights, since LAION distributed links and metadata rather than the images.

LAION said it built the dataset using CLIP models to filter for image-caption alignment and included nearest-neighbour indices, embeddings and watermark-detection scores to support downstream research. It disclosed that automated safety filtering flagged an estimated 2.9 to 3.3% of the dataset’s subsets as potentially unsafe, while cautioning that even this filtering could not “entirely exclude the possibility for harmful content” given the scale and unmoderated nature of a web crawl, and recommended the dataset for research use rather than for building deployed commercial products directly.

LAION-5B’s practical significance came from what got trained on it. It became the training corpus for Stable Diffusion, released by Stability AI later that year, and for a wide range of subsequent open image-generation and vision-language models — making it, for several years, one of the most consequential pieces of shared infrastructure in the open-model ecosystem, despite consisting of nothing more than links and captions rather than images LAION itself hosted.

That same scale and lack of curation became a liability: in December 2023, Stanford researchers reported finding child sexual abuse material referenced within LAION-5B, prompting LAION to withdraw the dataset and later release a filtered successor.