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

Forum topic · 小凯 · 2026-08-27

Summary

LAION-BVD is a large-scale open video dataset built for multimodal pre-training, presented in an arXiv paper (2608.24845) by Andreas Hochlehnert, Marianna Nezhurina, Mehdi Cherti, and colleagues. The dataset contains 1.3 billion platform-specific video URLs collected from CommonCrawl, from which 80 million videos totaling 10 million hours were downloaded. It targets pre-training across video, audio, and image modalities. Content-aware scene detection extracts clips, and video and audio captions are synthetically generated for each. Models trained on LAION-BVD achieve competitive performance on standard video-text and audio-text benchmarks, improving consistently with more training and larger model scale. The authors also explore scene-changing video frames as an alternative image-text data source; these frames show a visual distribution distinct from standard web image corpora, and models trained on them achieve strong image-text retrieval performance. The dataset is released openly to the research community.

Overview

Research area: Computer Vision Authors: Andreas Hochlehnert, Marianna Nezhurina, Mehdi Cherti, Andrej Radonjic, Thaddäus Wiedemer, et al. arXiv: 2608.24845

Key Points

Scale: LAION-BVD contains 1.3 billion platform-specific video URLs harvested from CommonCrawl, from which 80 million videos were downloaded, totaling 10 million hours of footage.
Multimodal design: The dataset is built for pre-training across video, audio, and image modalities.
Curation pipeline: Content-aware scene detection extracts clips, and video and audio captions are synthetically generated for each clip.
Benchmark results: Models trained on these data achieve competitive performance on standard video-text and audio-text benchmarks, with consistent improvements as training duration or model scale increases.
Frames as image data: The authors extract scene-changing frames as an alternative source of image-text data. These frames exhibit a visual distribution distinct from standard web image corpora, and models trained on them achieve strong image-text retrieval performance.
Open release: LAION-BVD is released openly to the research community.

Original Abstract (excerpt)

> We present LAION-BVD, a large-scale open video dataset for multimodal learning, which contains 1.3B platform-specific video URLs collected from CommonCrawl. From these, we download 80M videos with a total duration of 10 million hours. The dataset is designed for multimodal pre-training across the video, audio, and image modalities...

*Source: arXiv preprint 2608.24845*