LAION Releases 10 Million Hours of Video for Open AI Research

LAION, the German non-profit best known for the image dataset behind Stable Diffusion, has released the Big Video Dataset, one of the largest open video collections ever published for AI research. The team started from 1.3 billion video URLs found in CommonCrawl, a public archive of the web, downloaded 80 million of those videos totalling around 10 million hours, then cut them into 55 million clips and extracted 300 million still images. Most of the material comes from YouTube, and most of it is in English.

The interesting engineering is in the labels. Rather than relying on whatever description a human happened to write, the team used content-aware scene detection to find clip boundaries, then generated captions for both the visuals and the audio automatically. Training on that combination teaches a model which pictures go with which words and which sounds, all at once. According to the accompanying paper, models trained on BVD beat comparable models trained on the older InternVid dataset by up to 2.1 percentage points on standard video-to-text benchmarks. The dataset and code are freely available, released for research use only.

What’s actually going on here: every capable AI model rests on a pile of training data, and for video that pile has mostly been private. Google, OpenAI and ByteDance train video models on material nobody outside those companies can inspect. That makes it impossible for independent researchers to check what went in, reproduce a result, or study bias. Open datasets fix that, and they are also the only way a university lab competes at all. The uncomfortable part is the legal footing. LAION points to a 2024 ruling by the Hamburg Regional Court that allowed it to gather copyrighted images for non-commercial research, and it asks users to respect creators’ rights. That is a research exemption, not permission, and it is why the release is research-only. Creators whose videos are in here were not asked, and there is no practical way to ask 80 million times.

What this means for you: if you are learning about AI, this is a rare chance to look at the raw material instead of the finished product. The dataset composition charts alone are instructive: seeing how heavily one platform and one language dominate explains a lot about why video models behave the way they do. If you are a researcher, this is a serious resource and the paper is worth reading for the captioning method. If you make videos for a living, the honest summary is that your work may well be in here, that the release is limited to non-commercial research, and that the boundary between research corpus and commercial training set has historically been thin. Nothing changes for you today, but it is worth knowing the ground you are standing on.

Sources

• LAION Big Video Dataset (LAION)
• LAION-BVD: A 10-Million-Hour Open Video Dataset for Multimodal Pre-training (arXiv)
• LAION drops massive open video dataset with 10 million hours of footage (The Decoder)