LAION just published LAION-BVD, a large-scale open video dataset aimed at multimodal pretraining research. The numbers: 1.3 billion video URLs collected from CommonCrawl, 80 million successfully downloaded, 10 million hours of total video content, and 55 million clips with synthetically generated captions. On top of that, they extracted 300 million frames as an image-text corpus with a visual distribution deliberately different from standard web image scrapes.
The pipeline has four stages: download from CommonCrawl URLs, apply content-aware scene detection to find meaningful clip boundaries, generate captions for clips and audio tracks separately using vision-language and audio-language models, and validate with reference models (ViCLIP for video-text, CLAP for audio-text, CLIP for image-text). The resulting dataset covers three modalities at once — video, audio, and static frames — with independent caption tracks for each, which is useful for training models that need to reason about temporal content and audio together rather than just treating video as a stack of images.
The context here is LAION-5B. That dataset, released in 2022, had a catalytic effect on open image-text models: it gave research teams a training corpus at a scale that had previously been accessible only to a handful of well-funded labs. DALL-E and CLIP-style models were democratized in large part because of it. LAION-BVD is an attempt to do something similar for video. The problem it addresses is real: nearly every capable video generation or video-understanding model you can name — Sora-class systems, video-enabled multimodal LLMs — was trained on proprietary or tightly licensed video corpora. Open-weight video models have existed, but open training data at this scale has not.
The license is research-only, not commercial, which is a meaningful constraint. You can’t use BVD to train a model you plan to ship in a product, at least not without separate clearances. For academic researchers and labs working on benchmarks, evaluation, or foundational architecture work, this is likely a non-issue. For anyone wanting to replicate a commercial video-multimodal pipeline from first principles, it’s a blocker.
The arXiv paper accompanying the release reports competitive performance on video-text retrieval and audio-text benchmarks, with the usual scaling behavior: consistent improvement as training data volume and model size increase, without saturation visible at the scales they tested. The extracted frame corpus achieves “strong image-text retrieval performance” in their experiments, which suggests the scene-detection-based sampling is producing genuinely varied visual content rather than redundant near-duplicates.
Ten million hours is a serious number. It’s roughly 70 petabytes of raw video at typical streaming quality — the kind of scale that used to require either a major platform’s internal data moat or a video provider partnership. Getting it from CommonCrawl, processed and captioned as a research artifact under a single license, is a concrete contribution to making open multimodal work viable outside the four or five labs that have historically owned this tier of training data.