Can LAION-BVD be used for commercial projects?

No. LAION states that BVD is released exclusively for research purposes and not for commercial use, and the gated Hugging Face subsets restrict access to academic and non-commercial research. LAION also asks users to respect the rights and copyright of the original content creators and to follow the source platforms' terms of use.

Do you need approval to download LAION-BVD?

It depends on the subset. The URL-and-metadata releases such as laion/BVD-URLs and laion/BVD-V-55M-URLs are open on Hugging Face. The subsets that carry actual media, including BVD-V-55M and the full 80M-video BVD-RAW collection, are gated: you submit a central access request form and wait for manual approval, which LAION says usually takes several days.

How much storage does LAION-BVD need?

A lot. The BVD-V-55M clip subset alone is 41.1 TB of files on Hugging Face, and that is only the 55 million scene-level clips cut from 2.4 million source videos. The audio and image-frame subsets are separate downloads on top of that. Teams that cannot host it can start from the open URL lists and fetch only the slice they need.

How well does a model trained on LAION-BVD perform?

LAION published ViCLIP-L-14-BVD-V-55M-s50M-b32K, a 0.4B-parameter video-language model trained on 50M samples from the clip subset, as a reference point. It reports 63.3% top-1 on Kinetics-400, 80.5% on UCF-101 and 60.4% on HMDB51, plus 42.7% recall@1 on MSR-VTT video retrieval. The model is MIT-licensed.

LAION · 2026-08-25 · major

LAION-BVD — 10 million hours of open video for multimodal training

LAION-BVD is an open video dataset of 80 million videos totalling 10 million hours, pulled from CommonCrawl. It ships 55M captioned clips, 10M audio clips and 300M image frames for research use.

Hugging Face dataset card banner for LAION BVD-V-55M video clips

LAION released BVD, an open video dataset of 80 million videos and 10 million hours, built for multimodal pre-training.

Quick facts

Maker	LAION e.V.
Scale	80M videos, 10M hours
Source	1.3B video URLs from CommonCrawl
Subsets	55M video clips, 10M audio clips, 300M frames
Allowed use	Academic and non-commercial research only
Access	URL lists open on Hugging Face; media subsets gated
Baseline model	ViCLIP-L/14, 0.4B params, MIT

What is it?

LAION-BVD (Big Video Dataset) opens up video pre-training data at a scale that was previously private. LAION extracted 1.3 billion platform-specific video URLs from CommonCrawl and downloaded 80 million of the videos, totalling 10 million hours. From that pool it publishes ready-to-train subsets: 55 million captioned video clips, 10 million audio clips and 300 million extracted image frames.

How does it work?

Content-aware scene detection splits each source video in LAION-BVD into scene-level clips, and captions for both the video and the audio are then generated synthetically rather than scraped. Each clip record carries its caption, language and the timestamps marking where it sits in the source video. The frame subset takes scene-changing frames, whose visual distribution the paper reports as different from standard web image corpora.

Why does it matter?

Open video-language research has been stuck training on small public sets while frontier video models learn from data nobody outside the lab can inspect. LAION-BVD gives academics a corpus in the same size class, plus the URL lists needed to audit or re-derive it. LAION also published ViCLIP-L/14 trained on the 55M clip subset as an MIT-licensed baseline to build against.

Who is it for?

multimodal and video-language researchers

Frequently asked questions

Can LAION-BVD be used for commercial projects?: No. LAION states that BVD is released exclusively for research purposes and not for commercial use, and the gated Hugging Face subsets restrict access to academic and non-commercial research. LAION also asks users to respect the rights and copyright of the original content creators and to follow the source platforms' terms of use.
Do you need approval to download LAION-BVD?: It depends on the subset. The URL-and-metadata releases such as laion/BVD-URLs and laion/BVD-V-55M-URLs are open on Hugging Face. The subsets that carry actual media, including BVD-V-55M and the full 80M-video BVD-RAW collection, are gated: you submit a central access request form and wait for manual approval, which LAION says usually takes several days.
How much storage does LAION-BVD need?: A lot. The BVD-V-55M clip subset alone is 41.1 TB of files on Hugging Face, and that is only the 55 million scene-level clips cut from 2.4 million source videos. The audio and image-frame subsets are separate downloads on top of that. Teams that cannot host it can start from the open URL lists and fetch only the slice they need.
How well does a model trained on LAION-BVD perform?: LAION published ViCLIP-L-14-BVD-V-55M-s50M-b32K, a 0.4B-parameter video-language model trained on 50M samples from the clip subset, as a reference point. It reports 63.3% top-1 on Kinetics-400, 80.5% on UCF-101 and 60.4% on HMDB51, plus 42.7% recall@1 on MSR-VTT video retrieval. The model is MIT-licensed.

Try it

Request access at https://projects.laion.ai/bvd/download.html