資料集與多模態研究
LAION-BVD Releases 1.3 Billion Video URLs, With 80 Million Videos Processed and 10 Million Total Hours
Built from Common Crawl, LAION-BVD provides open research data spanning video, audio, and frames, including subsets containing 55 million clips with synthetic captions and 300 million frames. ViCLIP outperformed the InternVid baseline by as much as 2.1%, but the data is restricted to research use, while source rights, link rot, and synthetic annotation errors still require separate governance.
Teams from LAION, the University of Tübingen, and other institutions have released LAION-BVD in an effort to narrow the scale gap between open multimodal research datasets and the proprietary video corpora held by major technology companies. The researchers first extracted 1.3 billion platform-specific video URLs from Common Crawl, successfully downloading and processing approximately 80 million videos totaling 10 million hours. They then divided the videos into clips using content-aware scene detection, generated synthetic text descriptions for the video and audio, and extracted frames at points of visual change.
The released artifacts are not packaged as a single archive but as a layered data product. Available on Hugging Face are the complete URL index, BVD-V-55M with approximately 55 million clips and synthetic video captions, 10 million audio samples, and BVD-I-300M with approximately 300 million frames. This design allows researchers to obtain only the URLs or a specific modality, lowering the barrier associated with directly copying all of the original media. However, it also means reproducibility will be affected by source-video deletions, regional restrictions, and platform download policies.
The team evaluated the data’s utility using ViCLIP, CLAP, and CLIP. According to the official summary, ViCLIP continued to improve as training scaled from 10 million to 50 million clips, outperforming the InternVid FLT baseline by as much as 2.1% on standard video-text benchmarks. The audio model performed comparably to models trained on other large, uncurated corpora, while the distribution of video frames also differed from that of typical web images. These results demonstrate that the data can be used to train representation models, but they do not directly establish the quality of large video generation models. The current results were also produced by the dataset’s authors and have yet to be independently verified.
LAION explicitly restricts the data to research and non-commercial use and advises users to comply independently with copyright law, platform terms, and applicable legislation. Web video may also introduce linguistic and geographic imbalances, stereotypes, harmful content, and hallucinations in synthetic captions. Key issues to watch include URL survival rates, deduplication, content-safety audits, and whether third parties can reproduce the reported scaling trends under a fixed compute budget.