LAION just dropped one of the largest open video datasets ever assembled for AI research, and the scale is genuinely staggering. The Big Video Dataset, or BVD, starts with 1.3 billion video URLs pulled from CommonCrawl. From there, the team actually downloaded 80 million videos — roughly 10 million hours of footage — and distilled that into 55 million clips paired with auto-generated video and audio descriptions. They also extracted 300 million still images from the same pool. That’s not a dataset. That’s an infrastructure project.
The headline number that matters for researchers: models trained on BVD outperform comparable models trained on InternVid by up to 2.1 percentage points on standard video-to-text benchmarks. That gap sounds modest, but in the world of multimodal model evaluation, two points is the difference between a paper getting cited and a paper getting ignored. The training approach ties video, audio, and text together, forcing the model to learn which visual content corresponds to which descriptions and sounds. It’s a genuinely multimodal signal, not just video with captions bolted on.
Legally, LAION is standing on reasonably solid ground. A 2024 Hamburg Regional Court ruling allowed the organization to collect copyrighted content for non-commercial scientific research, which is exactly what BVD is positioned as. LAION is releasing the dataset for research only and asking users to respect the rights of original content creators. That’s a softer ask than an enforceable license, but it reflects the reality of open datasets built from scraped web content.
What’s interesting here is the timing. Open video datasets have lagged behind text and image corpora for years, largely because downloading and processing video at this scale is expensive and legally murky. InternVid has been the default choice for many academic labs, and BVD now gives those labs a bigger, better-performing alternative without licensing fees. Whether commercial labs adopt it is another question entirely — most will stick with proprietary data for competitive reasons. But for the academic and open-source research community, this removes a real bottleneck. The dataset and code are freely available now.
