Back to feed
News Story
APriority80
THE DECODER
1 sources

LAION releases massive open video dataset with 10 million hours for AI research

LAION has released its Big Video Dataset (BVD), a massive open video dataset containing 80 million videos and 10 million hours of footage for AI research. The dataset outperforms the previous benchmark InternVid and may benefit from a 2024 Hamburg court ruling that permits collecting copyrighted content for non-commercial research.

SynthePulse Insight · AI deep reading

LAION Releases BVD: A 10-Million-Hour Open Video Dataset Pushing Scale and Legal Boundaries

Version 1 · 1 source

LAION introduces the Big Video Dataset (BVD), containing 10 million hours of video, sourced from 1.3 billion video URLs in CommonCrawl, and claims it outperforms InternVid on video-text benchmarks. The dataset is research-only, with a legal basis possibly stemming from a 2024 Hamburg court ruling.

  • LAION releases the Big Video Dataset (BVD), comprising 80 million videos, 10 million hours of footage, 55 million auto-captioned clips, and 300 million still images.
  • The dataset is derived from 1.3 billion video URLs found in CommonCrawl, with most videos sourced from YouTube and predominantly in English.
  • The paper claims models trained on BVD outperform those trained on InternVid by up to 2.1 percentage points on common video-text benchmarks.
  • BVD is restricted to research purposes, and LAION may rely on a 2024 Hamburg regional court ruling to legally collect copyrighted content.
  • The dataset and code are freely available, but LAION asks users to respect the rights of original content creators.
Open section navigationDataset Scale and Composition

Dataset Scale and Composition

LAION has released the Big Video Dataset (BVD), touted as one of the largest open video datasets for AI research. The dataset was filtered from 1.3 billion video URLs discovered in CommonCrawl, with 80 million videos actually downloaded, totaling 10 million hours of footage.

The team further extracted 55 million video clips with auto-generated video and audio descriptions, along with 300 million still images. Most videos in the dataset come from YouTube and are predominantly in English.

Performance Claims and Training Approach

According to the paper, models trained on BVD perform better on common video-text benchmarks than those trained on InternVid, by up to 2.1 percentage points. The training process binds video, audio, and text, learning the correspondence between visual content and descriptions or sounds.

It is important to note that this performance improvement is a claim from the paper and has not been independently verified.

Legal and Usage Restrictions

LAION is releasing the dataset for research purposes only. Legally, the organization may rely on a 2024 Hamburg regional court ruling that permits the collection of copyrighted content for non-commercial research.

LAION asks users to respect the rights of original content creators. Both the dataset and code are freely available.

Credibility boundary

This report is based on secondary coverage from THE DECODER, which sourced information from LAION's paper and official release. All specific figures and performance claims originate from LAION's paper and are source claims, not independently verified. The citation of the legal ruling is also a source claim.

Insight takeaway

The release of BVD marks a significant leap in the scale of open video datasets, but the research-only restriction and copyright legal basis warrant attention. Its performance advantage claims await independent verification.

Primary report

THE DECODER

Primary source