Editorial illustration for LAION Releases 10-Million-Hour Video Dataset for AI Training
LAION Releases 10M-Hour Video Dataset for AI
LAION has pulled 80 million videos off the open web, adding up to 10 million hours of footage, and turned them into a new training set called the Big Video Dataset. The German nonprofit built it by trawling 1.3 billion video URLs indexed in CommonCrawl, then cutting the downloaded material into 55 million clips. Each clip comes with auto-generated descriptions of both the video and its audio track, and the release also includes 300 million still images pulled from the footage.
The scale puts BVD among the largest open video collections available for AI research, at a moment when most video training data used by major labs stays locked behind closed doors. LAION's own benchmark tests show the payoff: models trained on BVD beat comparable models trained on InternVid by as much as 2.1 percentage points on standard video-to-text tasks. The dataset is built to link video, audio, and text together, so a model learns to match a sound or a scene to the words that describe it.
LAION is restricting the release to research use only, and the legal footing behind that decision is worth a closer look.
According to the paper, models trained on BVD outperform comparable models trained on InternVid by up to 2.1 percentage points on common video-to-text benchmarks.
Why this matters For teams building video-native models, BVD is the kind of resource that used to require a Google DeepMind-sized budget: 10 million hours of footage, 55 million captioned clips, 300 million stills, all pulled from CommonCrawl and released openly. That's a real shift in who can train serious video-language models, not just labs with proprietary scraping pipelines.
But the 2.1 percentage point edge over InternVid is modest, and it's worth asking how much of that gain comes from scale versus caption quality, since the descriptions are auto-generated rather than human-verified. Anyone building on BVD should stress-test those labels before trusting them for fine-tuning.
There's also the provenance question LAION knows well from its image datasets: scraping 80 million videos out of 1.3 billion URLs found in CommonCrawl doesn't resolve who owns that footage or whether it was scraped with consent. Expect the same licensing scrutiny that followed LAION-5B. For researchers, BVD is a genuine unlock on scale. For anyone shipping products on top of it, the legal groundwork still needs checking before it becomes infrastructure.
Related Reading
Research & Benchmarks Google's FACTS benchmark shows 70% factuality ceiling across four tests
Research & Benchmarks Databricks finds multi-step agents beat single-turn RAG by 21% to 38% on STaRK
Research & Benchmarks Nvidia's DLSS 4.5 beta adds 6x Multi Frame Generation for RTX 50 GPUs
From Archives Anthropic's Self-Improving AI System Replicates Research Process
From Archives Google DeepMind's AI Co-Scientist Writes Plausible but Inaccurate Methods in Papers
Common Questions Answered
How many videos and hours of footage does LAION's Big Video Dataset contain?
LAION's Big Video Dataset contains 80 million videos totaling 10 million hours of footage, which were sourced from 1.3 billion video URLs indexed in CommonCrawl. The dataset was then processed into 55 million clips, each with auto-generated descriptions of both video and audio content, plus 300 million still images extracted from the footage.
How does the Big Video Dataset perform compared to InternVid on video-to-text benchmarks?
According to the paper, models trained on the Big Video Dataset outperform comparable models trained on InternVid by up to 2.1 percentage points on common video-to-text benchmarks. This performance advantage demonstrates the quality and effectiveness of LAION's dataset for training video-language models.
What makes LAION's Big Video Dataset significant for AI research teams?
The Big Video Dataset represents a democratization of video model training resources that previously required massive budgets comparable to organizations like Google DeepMind. By releasing 10 million hours of captioned video footage openly, LAION enables teams without proprietary scraping pipelines to train serious video-language models, fundamentally shifting accessibility in the field.
What type of descriptions are included with each clip in the Big Video Dataset?
Each of the 55 million clips in the Big Video Dataset comes with auto-generated descriptions that capture both the visual content of the video and the audio track information. This dual annotation approach provides comprehensive context for training video-language models.
Further Reading
- A 10-Million-Hour Open Video Dataset for Multimodal Pre-training - arXiv
- LAION-BVD: A 10-Million-Hour Open Video Dataset for Multimodal Research - LAION Blog
- LAION-BVD — 10 million hours of open video for multimodal training - ai-tldr.dev
- BVD: Big Video Dataset - a laion Collection - Hugging Face
- LAION releases 10-million-hour open video dataset with 1.3B URLs - Zeli
