Back to overview

Open Access

Dataset Versions

All versions of LAION-BVD are freely available on Hugging Face.
Choose the subset that fits your research needs.

URL-only Variants

Lightweight metadata-only versions — no media files, just URLs and additional metadata

URLs

BVD-URLs

The full 1.3B extracted platform-specific URLs from CommonCrawl. The foundation of the entire LAION-BVD collection.

1.3B URLs CommonCrawl
Hugging Face
URLs

BVD-V-55M-URLs

URL-only version of BVD-V-55M with additional metadata on the 2.4M source videos.

55M entries Video metadata
Hugging Face
URLs

BVD-A-1.7M-URLs

URL-only version of BVD-A-1.7M for lightweight access to the unique-source audio subset.

1.7M entries Audio metadata
Hugging Face
URLs

BVD-A-10M-URLs

URL-only version of BVD-A-10M for lightweight access to the large-scale audio sample.

10M entries Audio metadata
Hugging Face
URLs

BVD-I-300M-URLs

URL-only version of BVD-I-300M for lightweight access to the large-scale image frame collection.

300M entries Frame metadata
Hugging Face

Full Collections

Complete datasets including raw media files

Video

BVD-RAW

The raw pool of 80M downloaded videos totaling 10M video hours, sourced from URLs in BVD-URLs.

80M videos 10M hours Research collaboration
Request Access
Video

BVD-V-55M

55M clips extracted via content-aware scene detection from a randomly sampled subset of 2.4M videos from BVD-RAW.

55M clips Scene detection 2.4M source videos
Hugging Face
Audio

BVD-A-1.7M

1.7M audio clips from BVD-V-55M, sampled for uniqueness of source videos to maximize diversity.

1.7M clips Unique sources
Hugging Face
Audio

BVD-A-10M

10M audio clips randomly sampled from BVD-V-55M for large-scale audio pre-training.

10M clips Random sample
Hugging Face
Image

BVD-I-300M

300M extracted scene-changing frames randomly sampled from BVD-RAW for image-text pre-training.

300M frames Scene changes
Hugging Face