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

LAION-BVD is an open multimodal pre-training resource built from 1.3B platform-specific URLs collected through Common Crawl, yielding 80M successfully downloaded videos totaling 10M hours from YouTube, Vimeo, and Dailymotion. Its scalable pipeline uses PySceneDetect for content-aware scene segmentation, removes effectively static segments, and generates modality-specific captions with Qwen3-VL-2B-Instruct for video, Audio Flamingo 3 for sound, and DeepSeek-VL2-tiny for extracted frames. From 2.4M sampled videos, the release provides 55M scene-level video and audio clips plus 300M scene-changing frame-caption pairs. ViCLIP trained on BVD-V-50M achieved a 62.6 aggregate score across Kinetics-400, UCF-101, HMDB51, MSR-VTT, and MSVD after WiSE-FT checkpoint merging, exceeding the matched InternVid-10M-FLT result of 60.2. For audio, CLAP trained solely on BVD-A-10M reached a best aggregate score of 48.7 with a 431M-parameter model after 110M samples seen, demonstrating scaling with model and compute size. Frame-based OpenAI CLIP training produced strong retrieval but weaker ImageNet classification: a ViT-B/16 trained on 300M BVD frames reached 0.80 COCO image-to-text Recall@5. The dataset’s video-derived frame distribution also differs substantially from Re-LAION, with an FID of 33.92 versus 0.16 between independent Re-LAION samples, making BVD complementary to conventional web-image corpora. Its main limitations are automatically generated short captions, minimal safety filtering, and separate rather than joint audio-visual training.