LAION releases LAION-5B image-text metadata

A public index for multimodal training data

LAION's March 31, 2022 announcement released LAION-5B, a collection of metadata for 5.85 billion image-text pairs filtered using CLIP. Researchers could use URLs, captions and associated representations to select material for multimodal experiments. The collection was not a curated guarantee that every pair was accurate or appropriate for every use. [1]

The release licensed its metadata under CC BY 4.0 while explicitly preserving the images' own copyright. The exact day comes from the primary announcement and a same-day unaffiliated community discussion; InfoQ's May 17 report is later corroboration of the release, rather than evidence that it happened in May. [1][3][2]