Home · Datasets · LAION-5B
DATASET

LAION-5B

The largest openly available image-text dataset used to train Stable Diffusion and other multimodal models.

TARGET QUERY laion 5b dataset · ~8K/mo
SIZE
5.85B image-text pairs
CREATOR
LAION
MODALITY
multimodal
LICENSE
CC-BY-4.0
RELEASED
2022-03
OVERVIEW Updated 2026-05-17

Overview

LAION-5B is the largest publicly available image-text dataset, containing 5.85 billion image-text pairs filtered from Common Crawl. Created by the LAION non-profit collective, it enabled the training of Stable Diffusion and made large-scale multimodal research accessible to the open-source community.

What’s In It

The dataset comprises 5.85 billion image-URL and alt-text pairs, split into 2.32 billion English pairs and 2.26 billion multilingual pairs. Each pair includes CLIP similarity scores enabling quality filtering. The data spans virtually every visual domain from photography to artwork to diagrams.

How It’s Used

LAION-5B is primarily used for training text-to-image models and CLIP-style vision-language models. Stable Diffusion was trained on a filtered aesthetic subset. OpenCLIP uses it for reproducing CLIP at scale. Researchers also use it for studying web-scale data distributions.

Controversies

In December 2023, Stanford researchers found CSAM material in the dataset, leading to its temporary removal. The dataset also contains copyrighted material, raising legal questions about training generative AI on scraped web data. LAION conducted extensive filtering before re-releasing cleaned versions.