LAION Releases LAION-BVD, a 10M-Hour Open Video Dataset

In late August, LAION released LAION-BVD (Big Video Dataset), a large-scale fully open video dataset for multimodal (video, audio, image) pretraining, aiming to fill the video-data gap in open multimodal research. According to LAION member Christoph Schuhmann (JJitsev) and contributor Andreas Hochlehnert, the dataset is already public and shows strong benchmark performance.

Confirmed

• Sourced from CommonCrawl: 1.3 billion platform-specific video URLs collected
• 80 million videos actually downloaded, totaling 10 million hours
• 55 million captioned (synthetic-subtitle) video clips
• 300 million frame-text pairs
• Fully open, targeting multimodal pretraining

Why it matters

• Video data has long been a weak spot for open multimodal research; LAION-BVD provides an open corpus at a scale previously only matched by proprietary closed datasets
• It continues LAION's open-data lineage (e.g., LAION-5B) and gives the community a reusable foundation for training open multimodal models
• The synthetic captions and frame-text pairs lower the engineering barrier for downstream multimodal alignment tasks

2026-08-26 ~ 2026-08-28 · 5 related posts

• [source] LAION-BVD: A 10-Million-Hour Open Video Dataset for Multimodal Pre-training — Andreas Hochlehnert · 2026-08-26
• LAION releases 10M-hour open video dataset BVD for multimodal training — HirokatuKataoka · 2026-08-27

3 near-duplicate retellings: HirokatuKataoka · JJitsev · const_reborn