‹ back to 2026-08-26

LAION-BVD: A 10-Million-Hour Open Video Dataset for Multimodal Pre-training

arXiv cs.CV Multimodal & Generative Andreas Hochlehnert, Marianna Nezhurina, Mehdi Cherti, Andrej Radonjic, Thaddäus Wiedemer, Christoph Schuhmann, Romain Beaumont, Wieland Brendel, Bernhard Schölkopf, A. Sophia Koepke, Jenia Jitsev, Matthias Bethge 2026-08-25

TL;DR - LAION-BVD is an open multimodal pre-training dataset comprising 80 million downloaded videos totaling 10 million hours, sourced from 1.3 billion CommonCrawl-discovered URLs. It substantially expands publicly accessible data for training video, audio, and image models.

Content-aware scene detection produces clips with synthetically generated video and audio captions.
Models trained on the data show competitive video-text and audio-text benchmark performance, improving consistently with training and model scale.
Scene-changing video frames provide image-text data with a different visual distribution from conventional web-image corpora.
Models trained on the extracted frames achieve strong image-text retrieval performance.

view merged work →