訓練資料
LAION-BVD Releases a 10-Million-Hour Video Corpus, but Media Files Are Limited to Approved Noncommercial Research
LAION-BVD compiled 1.3 billion video URLs from Common Crawl and downloaded approximately 80 million videos totaling 10 million hours. The data pipeline and URL lists are public, but access to the main media subset is restricted, and researchers remain responsible for compliance.
LAION announced the Big Video Dataset (BVD) on August 25, aiming to bring large-scale video, audio, and image pretraining data into an inspectable research environment. The team first extracted 1.3 billion platform-specific video URLs from Common Crawl, then successfully downloaded approximately 80 million videos with a combined duration of about 10 million hours. These figures should not be conflated: the former is merely the candidate URL collection, while the latter represents the raw corpus actually obtained.
The public pipeline covers downloading, content-aware scene segmentation, frame extraction, synthetic captioning, and model training. Its main derived datasets include 55 million scene clips with video captions and timestamps, 10 million audio clips, and 300 million scene-change frames. The video and audio descriptions are generated by models rather than taken directly from webpages. This helps address missing descriptive text, but it may also propagate recognition errors and biases from the captioning models into downstream training.
LAION also released training pipelines for ViCLIP, CLAP, and CLIP. The authors report that ViCLIP trained on BVD outperformed the InternVid baseline by as much as 2.1 percentage points on standard video–text benchmarks, with performance continuing to improve as the dataset grew from 10 million to 50 million clips. This provides rare public experimental material for studying scaling behavior across dataset size, model size, and modality combinations.
“Open” still requires an important qualification: the full media collection is not available for unrestricted anonymous download. The 55-million-clip subset is approximately 41.1 TB and requires applicants to submit institutional affiliation, intended use, and authorization details for manual review. Access is limited to academic and noncommercial research; only the URLs and some metadata are directly available. LAION also explicitly warns that the corpus may present geographic and linguistic biases, stereotypes, and content-rights issues. Its release does not mean that copyright obligations or platform terms have been cleared on users’ behalf. Key issues to watch next include data-removal mechanisms, source traceability, errors in synthetic captions, and whether third parties can reproduce the official results under restricted-access conditions.