LAION released LAION-BVD, an open video dataset for multimodal learning. It contains 1.3B platform-specific video URLs from CommonCrawl. Researchers downloaded 80M videos totaling 10 million hours. The dataset supports video, audio and image modality pre-training. Content-aware scene detection extracts clips with synthetic captions. Models trained on it perform competitively on video-text and audio-text benchmarks. Performance improves with more training or larger models. Scene-changing frames serve as an alternative image-text data source. These frames differ visually from standard web image corpora. Models trained on them achieve strong image-text retrieval results. LAION-BVD is now available to the research community.

• 1.3B video URLs collected
• 80M videos downloaded
• 10M hours total duration