LAION Releases 80-Million-Video Open Dataset for Multimodal AI Research

LAION has released LAION-BVD (Big Video Dataset), a large-scale open video dataset built for multimodal learning. The dataset contains 80 million downloaded videos with a combined duration of 10 million hours, along with 1.3 billion platform-specific video URLs (projects.laion.ai/bvd).

The dataset is published in the LAION-AI/BVD repository on GitHub and is accompanied by a dedicated project page at projects.laion.ai/bvd. LAION is a non-profit organization that makes machine learning resources available to the general public (laion.ai).

LAION-BVD is released exclusively for research purposes and is not licensed for commercial use. Its stated objectives are to support scientific research, reproducibility, safety analysis, and the study of multimodal foundation models — systems that can process and connect information across video, audio, and images.

The scale is notable. At 10 million hours of video content, BVD gives researchers working on video-language models, audio-visual learning, and cross-modal representation a substantial corpus to draw from. The 1.3 billion platform-specific URLs serve as pointers to source content, while 80 million videos have been downloaded and packaged for direct use. This structure — an index of URLs plus downloadable video files — mirrors the approach LAION took with its image datasets, where index files point to source locations and researchers fetch the subsets they need.

The tooling behind the dataset relies on video2dataset, an open-source tool LAION introduced in July 2023 for curating video and audio datasets efficiently and at scale (laion.ai/blog/video2dataset). The tool handles parallelized downloading, processing, and packaging of video content into shards — smaller chunks of data that can be distributed across multiple machines for training. For researchers familiar with LAION's img2dataset for image curation, video2dataset extends the same architecture to the video and audio domain.

LAION is upfront about the dataset's limitations. BVD may contain biases, stereotypes, and uneven representation across languages, regions, and topics. The project page documents these caveats directly rather than burying them in fine print, which matters because researchers need to account for distributional skew when evaluating models trained on or benchmarked against this corpus. The dataset's composition reflects whatever is present on the source platforms, and downstream model behavior will inherit those properties.

The research-only licensing restriction is a deliberate boundary. By excluding commercial use, LAION positions BVD as a resource for the open research community: academic labs, independent researchers, and non-profit institutions that may lack the budget to license proprietary video corpora at comparable scale. This is consistent with LAION's organizational mission of broadening access to machine learning resources.

For the multimodal research community, the practical significance is straightforward. Video pre-training has been a compute- and data-intensive undertaking, and open video datasets at this scale have been scarce. Researchers building video-language models, audio-visual transformers, or unified multimodal foundation models now have a large, openly indexed corpus to work from, subject to the research-use terms.

The reproducibility angle deserves attention. One of LAION-BVD's stated goals is to support reproducibility in multimodal research. The ability to point to a fixed, publicly indexed dataset with documented contents and known limitations addresses a persistent problem in the field: results that cannot be independently verified because training data is proprietary or undocumented. Whether BVD becomes a standard benchmark or training reference will depend on adoption and on how the dataset holds up under scrutiny, but the structural conditions for that role are present.

The safety-analysis use case is also explicitly named. Access to a large, open video corpus allows researchers to probe foundation models for harmful outputs, distributional biases, and failure modes without negotiating access to closed datasets. This is a function that open datasets are uniquely positioned to serve, and BVD's scale makes it relevant for that work.

LAION has not released detailed documentation of the dataset's content distribution across source platforms, languages, or domains beyond the general caveat about uneven representation. Researchers evaluating BVD for specific use cases will likely need to conduct their own distributional analysis on the subsets they download.

The dataset is available now through the GitHub repository and the project page.