LAION publishes 10-million-hour open video dataset LAION-BVD; processed subsets available

LAION has posted an arXiv paper and public dataset artifacts for what it describes as a 10-million-hour open video dataset for multimodal AI training, a release centered on 80 million downloaded videos gathered from a much larger pool of web links. In the Aug. 25 paper, “LAION-BVD: A 10-Million-Hour Open Video Dataset for Multimodal Pre-training,” researchers from LAION, the University of Tübingen and partner institutions wrote: “We present LAION-BVD, a large-scale open video dataset for multimodal learning.”

What is publicly available is more limited than the headline scale might suggest. LAION has released processed subsets including 55 million scene clips, called BVD-V-55M, and about 300 million scene-changing frames, called BVD-I-300M, along with URLs and captions for released subsets on Hugging Face and the processing and training pipeline on GitHub. But the full raw video pool is not being distributed on Hugging Face. Instead, the paper says research institutions can request raw-data access under LAION’s terms, and LAION’s GitHub README says, “LAION-BVD is released exclusively for research purposes and not for commercial use.”

The release matters because large, reusable video training sets have been much harder to assemble and host than image-text datasets, especially at internet scale. LAION is best known for earlier open image-text collections such as LAION-400M and LAION-5B, which became widely used in open AI research. By duration, LAION-BVD is far larger than a commonly used earlier open video-text dataset, InternVid, which is about 7 million videos and roughly 760,000 hours. LAION-BVD’s claimed 10 million hours positions it as a much bigger resource for training systems that learn across video, audio and images.

According to the paper, the project began with 1.3 billion platform-specific video URLs extracted from CommonCrawl, a public archive of web data. The authors say they downloaded 80 million videos totaling about 10 million hours. Most of the source material came from YouTube, at about 94%, with Vimeo at about 4% and Dailymotion at about 2%. The dataset is multilingual, with about 57% of the material in English, the paper says. To turn those videos into training data, the pipeline uses scene detection to split videos into clips and AI models to generate synthetic captions describing both video and audio. The paper says the experimental training subset was built from roughly 2.4 million videos, producing a recaptioned set of 55 million clips.

The authors say models trained on LAION-BVD produced competitive results on standard video-text and audio-text benchmarks, and that performance improved as training or model scale increased. In particular, the paper says ViCLIP models trained on LAION-BVD matched or exceeded InternVid-trained models by up to 2.1% on standard video-text benchmarks. Those are the paper’s reported results, not independent verification.

LAION, a nonprofit research group focused on open datasets and models, is presenting LAION-BVD as a video-and-audio counterpart to the image-text resources that helped fuel earlier waves of open AI development. The release also arrives with familiar scrutiny around scraped training data: LAION’s earlier image datasets faced content-safety and copyright criticism, including a 2023 Stanford Internet Observatory finding of known child sexual abuse material in LAION-5B, as well as later legal disputes over web-scraped training data.