# LAION-5B

LAION-5B is a large-scale open dataset of 5.85 billion image-text pairs, released in 2022 by the non-profit LAION. It is widely used to train generative models such as Stable Diffusion.

LAION-5B is a large-scale open dataset of 5.85 billion image-text pairs, created by the non-profit organization LAION (Large-scale Artificial Intelligence Open Network). Released in 2022, it is one of the largest publicly available datasets of its kind and has been widely used to train generative models such as Stable Diffusion. It is a key resource in the field of [artificial intelligence](https://www.wikiprompt.org/wiki/artificial-intelligence) and [machine learning](https://www.wikiprompt.org/wiki/machine-learning).

## History and creation

LAION-5B was preceded by LAION-400M, a dataset of 400 million image-text pairs released in August 2021. The larger LAION-5B was assembled by crawling the Common Crawl web corpus, which archives billions of web pages. The team used a CLIP model, developed by [OpenAI](https://www.wikiprompt.org/wiki/openai), to compute embeddings of images and their associated alt-text or captions. Pairs with high cosine similarity between image and text embeddings were retained, while low-quality or mismatched pairs were discarded. The resulting dataset was published in October 2022 in a paper titled "LAION-5B: An open large-scale dataset for training next generation image-text models." The creation process relied on techniques from [deep learning](https://www.wikiprompt.org/wiki/deep-learning) and [neural networks](https://www.wikiprompt.org/wiki/neural-network), particularly the [transformer](https://www.wikiprompt.org/wiki/transformer) architecture used in CLIP.

## Composition and subsets

LAION-5B consists of three subsets: LAION-2B-en (2.32 billion English pairs), LAION-2B-multi (2.27 billion multilingual pairs), and LAION-1B-nolang (1.26 billion pairs without language filtering). The dataset does not store images directly; instead, it provides metadata such as image URLs, alt-text, dimensions, and a similarity score. This design allows researchers to download images on demand, but it also means that some URLs may become inaccessible over time. The subsets are organized to support research in multilingual and cross-lingual learning, which is relevant to [large language models](https://www.wikiprompt.org/wiki/large-language-model) and multimodal systems.

## Applications

LAION-5B has been used to train a variety of [generative AI](https://www.wikiprompt.org/wiki/generative-ai) models. The most notable example is Stable Diffusion, a text-to-image model released in August 2022 by Stability AI. Stable Diffusion uses a latent diffusion architecture with a [U-Net](https://www.wikiprompt.org/wiki/u-net) and [cross-attention](https://www.wikiprompt.org/wiki/cross-attention) layers to generate images from text prompts. The model was trained on a subset of LAION-5B and demonstrated that high-quality image generation could be achieved with open data and open-source methods. LAION-5B has also been used for training image-text models, contrastive models, and other multimodal systems. It serves as a benchmark for evaluating the performance of such models on large-scale data. The dataset's scale and diversity are particularly valuable for [multi-head attention](https://www.wikiprompt.org/wiki/multi-head-attention)-based architectures, which benefit from large amounts of paired data.

## Controversies and limitations

The dataset has raised concerns regarding privacy, copyright, and bias. Because it is derived from web crawls, it contains personal photos, copyrighted artwork, and potentially offensive or harmful content. LAION has attempted to filter out known problematic material, but the sheer scale of the dataset makes complete removal difficult. In 2023, the dataset was temporarily taken offline after a legal complaint, though it was later restored with additional filtering. Researchers have also noted that the dataset reflects the biases present in the web, including gender and racial stereotypes. These issues are common to large web-scraped datasets and remain an active area of research.

## Impact and legacy

LAION-5B has become a standard resource in the open-source AI community. It demonstrated that large-scale datasets could be built and shared by a volunteer organization, in contrast to proprietary datasets held by private companies. The dataset has influenced subsequent efforts to create open multimodal datasets and has enabled a wave of research in generative AI and machine learning. Its release also highlighted the importance of data governance and the need for responsible dataset curation.

---
Source: https://www.wikiprompt.org/wiki/laion-5b
License: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/)
Last updated: 2026-09-12T16:26:52.976747+00:00