LAION Releases LAION-BVD, a 10M-Hour Open Video Dataset

Eighty million videos, ten million hours. That is the size of LAION-BVD, an open video corpus released by LAION and collaborators and posted to arXiv. The videos were pulled from 1.3 billion platform-specific URLs collected via CommonCrawl, then split into clips using content-aware scene detection and paired with synthetic captions across video, audio, and image modalities.

The paper's own claim about model quality is careful. Systems trained on the data reach "competitive performance on standard video-text and audio-text benchmarks, with consistent improvements as training or model scale increases." The extracted frames do useful double duty: the authors note that scene-changing frames show "a visual distribution distinct from standard web image corpora," making them a fresh source of image-text pairs.

The abstract does not publish per-benchmark numbers or name the teacher models used to write the captions. Release is under CC-BY 4.0, with code and metadata on the LAION-AI/BVD GitHub repo. The author list includes Andreas Hochlehnert, Christoph Schuhmann, Jenia Jitsev, Wieland Brendel, Bernhard Schölkopf, and Matthias Bethge.