Glonce

LAION releases LAION-BVD, a 10 million hour open video dataset

researchThe Glonce editorsHacker News🔥72Glonce rating72Impact72Novelty68Relevance98Surprise52Players65Credibility87

LAION has published LAION-BVD (Big Video Dataset), an open video dataset built for multimodal pre-training across video, audio and image data. The team started from 1.3B platform-specific video URLs collected from CommonCrawl and downloaded 80M of those videos, totaling 10 million hours of combined footage. Using content-aware scene detection, they split the videos into clips and generated synthetic video and audio captions for them. They also pulled scene-changing frames out of the videos as a separate source of image-text pairs, arguing these frames carry a visual distribution distinct from typical web image corpora. On the benchmark side, ViCLIP models trained on LAION-BVD match or exceed InternVid-trained models by up to 2.1% on standard video-text benchmarks, and the gains hold up consistently as training scale grows from 10M to 50M clips. CLAP models trained on the dataset's audio reach competitive performance against other large-scale uncurated audio datasets, drawing on the in-the-wild soundscapes embedded in the videos. The frame-based CLIP models trained on the extracted scene-change frames achieve strong image-text retrieval results, which the authors attribute to that distinct visual distribution complementing standard web image pre-training data. LAION frames the release as a response to large-scale video datasets and the models trained on them being increasingly concentrated inside a handful of mostly proprietary companies, and says the goal is to broaden access to multimodal training data and make evaluation of large-scale video, audio and image models more transparent. The dataset is released exclusively for research purposes, not for commercial use, and LAION asks users to respect content creators' rights and follow applicable laws and platform terms. LAION also flags that, like other large-scale web datasets, LAION-BVD may contain biases, stereotypes and uneven representation across languages, regions and topics, and that models trained on it may inherit those biases. The dataset is documented in an accompanying paper, "LAION-BVD: A 10-Million-Hour Open Video Dataset for Multimodal Pre-training," by Andreas Hochlehnert and eleven co-authors including Christoph Schuhmann, Wieland Brendel, Bernhard Scholkopf and Jenia Jitsev.

Key facts

  • LAION-BVD starts from 1.3B platform-specific video URLs collected from CommonCrawl and downloads 80M of them, totaling 10 million hours of video.
  • Content-aware scene detection splits the downloaded videos into clips with synthetically generated video and audio captions.
  • ViCLIP models trained on the dataset match or exceed InternVid-trained models by up to 2.1% on standard video-text benchmarks, with gains holding as training scale grows from 10M to 50M clips.
  • Scene-changing video frames are extracted as a separate image-text source and produce strong image-text retrieval results, thanks to a visual distribution distinct from typical web image corpora.
  • The dataset is released for research use only, not commercial use, and LAION warns it may carry the biases and uneven representation typical of large-scale web data.

Why it matters

Open, large-scale video datasets for multimodal training have been scarce because collecting and processing video at this scale is expensive, and the datasets that exist tend to sit inside a small number of mostly proprietary companies. LAION-BVD is an attempt to change that balance by putting a video corpus of comparable scale, 80M videos and 10 million hours, into the open for academic research, alongside the audio and image-text data derived from it.

Who it affects

Researchers working on multimodal foundation models, video-text retrieval, audio-text learning and image-text pre-training gain a new open corpus to train and benchmark against, in place of relying on proprietary datasets they cannot inspect or reproduce. The release also affects anyone studying dataset bias and reproducibility in large-scale video models, since LAION explicitly frames transparency and independent evaluation as part of the motivation.

How to use it

LAION-BVD is available for download from LAION's project page, but strictly for research purposes; commercial use is excluded. Users are expected to respect the rights and copyright of the original content creators and to follow applicable laws and platform terms when working with the data. No specific license name, such as CC BY or MIT, is given beyond that prose restriction.

How solid is it

The dataset is documented in a paper with twelve named authors from LAION and academic collaborators, including Christoph Schuhmann, Wieland Brendel, Bernhard Scholkopf and Jenia Jitsev, and is posted on arXiv. The reported benchmark gains, up to 2.1% over InternVid-trained ViCLIP models and consistent improvement from 10M to 50M training clips, are specific and directly tied to standard video-text benchmarks, which supports the claim of competitive performance rather than leaving it as an unverified assertion.

Risks and caveats

LAION itself cautions that, like other large-scale web-scraped datasets, LAION-BVD may contain biases, stereotypes and uneven representation across languages, regions and topics, and that models trained on it can inherit those biases. The dataset's own release notes do not give specific benchmark numbers for the CLAP audio-text or frame-based image-text results beyond qualitative descriptions, and no exact release date beyond the paper's 2026 publication year is stated.

“Large-scale video datasets and the models trained on them are increasingly concentrated within a small number of predominantly proprietary technology companies, limiting independent scientific investigation and reproducibility.”

— LAION-BVD release page