A research team including the nonprofit LAION has released a large-scale video research dataset called "LAION-BVD." Starting from 1.3 billion video URLs extracted from Common Crawl, the team gathered roughly 80 million videos totaling 10 million hours, and also provides code on GitHub for processing this material into training data for video, audio, and still images. According to the paper's comparisons, this is roughly 13 times the total duration of InternVid's 760,300 hours, previously the largest publicly available video dataset. However, what anyone can obtain immediately is only the URLs and metadata—the raw videos and training subsets require an application. It's important to distinguish what "open" actually refers to and how much of the 10 million hours has actually been quality-checked.
Building 10 million hours from 1.3 billion URLs
The collection process began with WAT files from Common Crawl published up through March 2024. The research team extracted approximately 4.7 billion candidate URLs, narrowed them down to links pointing to YouTube, Vimeo, and Dailymotion, and compiled a ledger of 1.3 billion entries. From there, they attempted to retrieve 130 million videos and successfully saved 80 million—about 60%. The team used 2,000 virtual servers coordinated with Celery, along with yt-dlp and a residential proxy network.
While the volume is enormous, the sources are far from evenly distributed. YouTube accounts for 94% of the videos, Vimeo for 4%, and Dailymotion for 2%. English is the most common language at 57%, while Japanese makes up 4%. The average video length is 7.7 minutes, with a median of 3.7 minutes; about 6% run longer than 30 minutes. The most recent uploads date to early 2024.
The research team applied no additional safety filtering during collection. Their reasoning was that the platforms themselves moderate harmful or dangerous content—but passing the public standards of three companies doesn't mean the footage has been confirmed safe for use as training data. A large volume of data doesn't necessarily mean it has been carefully curated.
"Open" access comes in three distinct tiers
Under the name LAION-BVD sit multiple releases that differ in content and access requirements: the URL ledger, processed media, and raw video—three separate layers.
| Layer | Contents | Access conditions |
|---|---|---|
| Public URL ledger | 1,352,648,614 rows in BVD-URLs. Contains no video itself—just URLs and crawl-time metadata, totaling 190GB | Available via Hugging Face |
| Processed media | 55M video clips, 10M audio clips, 300M still images, and more | Requires application through a central form; limited to academic, non-commercial research |
| Raw video | 80 million videos in BVD-RAW, totaling 10 million hours | Requires approval as a research collaborator, then transfer to the user's own S3-compatible bucket |
What's immediately available is the URL and metadata layer. The media data—including the 10 million hours of raw video and the 55M clips—requires academic, non-commercial researchers to apply and receive individual approval. BVD-V-55M is listed as 41.1TB on Hugging Face, but unapproved users cannot access the files. BVD-RAW must be transferred to the user's own S3-compatible storage via GRASS, and users may be charged transfer fees of up to $2.50 per TB. "Free" doesn't mean LAION covers the storage destination and transfer costs.
The June 2026 terms of use draw an even stricter boundary. Raw videos and annotations are provided to designated users within a "closed research data environment," and commercial AI training, validation, or fine-tuning is prohibited. Access does not transfer any copyright usage rights over the videos in question, and each user must independently establish the legal basis for downloading or analyzing them. The CC BY 4.0 license attached to the URL ledger on Hugging Face should not be mistaken for the license governing the linked videos themselves.
The 55M clips represent just 0.56% of the total
The 55M clips used to validate ViCLIP and CLAP in the paper come from 2.4 million videos, totaling approximately 56,000 hours—about 0.56% of the full 10-million-hour raw dataset. This is a time-based ratio, not a ratio of clip count or information content. Still, it serves as a useful marker for distinguishing the headline figure of 10 million hours from the actual scope of model evaluation. The video and audio experiments used a derived subset built from 55M clips, created by scene-splitting a randomly sampled parent set of 2.4 million videos. The still images came from a separate process: 300 million frames extracted from the raw video pool, evaluated separately using ViCLIP, CLAP, and CLIP. This is not an integrated training approach that time-synchronizes video, audio, and stills together.
During processing, videos under 10 seconds or over 30 minutes were excluded, and PySceneDetect was run with a threshold of 30 to split content by scene. Motion between frames was measured at low resolution, and near-static segments were also discarded. Video captions were generated by feeding up to 32 frames into Qwen3-VL-2B-Instruct, capped at 20 words, while audio was captioned by Audio Flamingo 3 at 10 words or fewer. This allowed separate training labels for video and audio to be generated from the same scene.
The still-image pipeline was handled separately. Black frames were removed using ffmpeg, and up to 300 million frames exceeding a scene-change threshold of 0.1 were extracted. Frames captioned with DeepSeek-VL2-tiny show a distribution different from typical web images: the FID on CLIP embeddings was 33.92 between this set and ReLAION, compared to just 0.16 between two independent samples drawn from ReLAION itself. However, this gap is evidence of a distributional difference, not evidence of superior image quality or safety.
Don't extrapolate the 2.1-point gain to overall performance
The "up to 2.1 points" figure refers to applying WiSE-FT to ViCLIP L/14 and averaging across three classification metrics and four retrieval metrics—it does not mean uniform improvement across all benchmarks. The comparison was between a model trained on InternVid-10M-FLT, which achieved an aggregate average of 60.2, and a model trained on BVD-V-10M at the same scale of 10 million samples, which achieved 62.3. WiSE-FT is a post-processing step that blends the trained model's weights with the original CLIP; the paper matched these conditions to arrive at the 2.1-point gap.
The benefit of increasing data volume also varies by metric. When training BVD-V-10M and BVD-V-50M five times each while exposing the models to 50 million samples, the difference in aggregate average was 0.58 points, with a 95% confidence interval of 0.40 to 0.76. Clear improvements appeared on the HMDB action recognition test and some retrieval metrics, but the difference on Kinetics-400 was just 0.06 points and on UCF-101 just 0.14 points—with confidence intervals straddling zero in both cases. While the benefit of exposing models to more distinct clips is confirmed, there's no basis for expecting the same magnitude of improvement across every task.
With web-derived data, overlap between training material and evaluation data can also inflate scores. The research team cross-checked YouTube video IDs and found overlap rates of 0.26% for Kinetics-400, 5.5% for MSR-VTT, and 6.3% for MSVD. They report that performance was largely maintained even after re-evaluating with overlapping IDs removed. That said, the paper is a preprint submitted on August 25, 2026, and independent replication has yet to occur.
Safety and rights responsibilities fall on researchers
Even the synthetic captions contain measurable errors. When the research team audited 134 video captions, 106 (79.1%) were accurate, 25 (18.7%) contained minor errors, and 3 (2.2%) contained significant errors. They also examined 134 audio captions: 106 (79.1%) were accurate, but significant errors rose to 8 cases (6.0%). These are small samples relative to the 55 million total captions, and error rates broken down by person, region, or language are not provided.
Users also bear responsibility for storing and disposing of the data. The terms require encrypting local copies, restricting access to the user themselves, and prohibiting third-party sharing; they also mandate deleting dataset copies once research concludes. However, trained models, weights, and embeddings produced through legitimate use are exempt from this deletion requirement. Re-identification of individuals, contact, profiling, and use in prohibited biometric identification or surveillance systems are all disallowed.
What LAION-BVD has demonstrated is a processing pipeline capable of collecting massive volumes of web video and converting it into training material for video, audio, and still images, along with its usefulness in limited contrastive learning settings. Video generation, diffusion models, and integrated models that time-synchronize video and audio have not yet been evaluated. Three questions remain to be answered: Can applicant researchers actually gain access to the large-scale portion in practice? Can the caption quality be reproduced through larger, more stratified human audits? And does the advantage persist in generative tasks? Only once these three points are confirmed can the 10-million-hour scale be meaningfully connected to model evaluations that third parties can reproduce.