Laion Big Video Dataset

LAION-BVD comprises 1.3 billion platform-specific URLs from Common Crawl, of which 80 million videos (≈10 million hours) were downloaded and processed. Scene detection produced captions for video clips and audio, and extracted frames created an image-text corpus. Models trained on this data match or surpass baselines, improving up to 2.1% on video-text benchmarks as scale increases. The dataset is released for non-commercial research to support reproducible multimodal pre-training and bias analysis.