LAION Big Video Dataset

LAION has released LAION-BVD, an open-source video dataset totaling 10 million hours, by downloading 80 million of the 1.3 billion video URLs collected from CommonCrawl. This dataset is designed for multimodal pre-training spanning video, audio, and image modalities, with clips extracted via content-based scene detection and synthetic captions generated.

Benchmark Performance

• ViCLIP: Models trained on LAION-BVD achieved up to 2.1% higher performance than InternVid-based models on standard video-text benchmarks.
• CLAP: Demonstrated competitive performance on audio-text benchmarks utilizing diverse field sounds.
• CLIP: Image-text data extracted from video frames showed visual distribution differences from web image corpora, proving strong retrieval performance.

Release Purpose and Restrictions

LAION-BVD is released for research purposes only to improve accessibility to large-scale multimodal data and facilitate independent scientific investigation. Commercial use is prohibited, and researchers are responsible for evaluating any biases or copyright issues contained within the dataset.