laion/BVD-I-300M-URLs

LAION-BVD - 300M Video Frame URLs

This repository contains the URLs for ~300 million keyframes extracted from publicly available web videos. No image data is included, only the source video URL and the frame timestamp needed to reproduce each frame.

Frames were extracted from BVD-RAW and cover YouTube, Dailymotion, and Vimeo content.

Dataset structure

Column	Type	Description
webpage_url	string	URL of the source video
frame_pts_time	float	Presentation timestamp of the frame in seconds

Repository structure

frames/
  0/
    0000.parquet
    ...
  1/
    ...

Loading the data

import datasets

ds = datasets.load_dataset("laion/BVD-I-300M-URLs", data_dir="frames", split="train")

Or with pandas:

import pandas as pd

df = pd.read_parquet("frames/", engine="pyarrow")

Reproducing frames

Each frame can be reproduced from its source video using yt-dlp and ffmpeg:

yt-dlp -o video.mp4 "<webpage_url>"
ffmpeg -ss <frame_pts_time> -i video.mp4 -frames:v 1 frame.png

Relation to LAION-BVD

Release	Description
laion/BVD-URLs	1.3B video URLs extracted from CommonCrawl
laion/BVD-RAW-Index	File index of the ~80M downloaded raw videos (media by request)
laion/BVD-V-55M / -URLs	55M scene-level video clips with captions and timestamps
laion/BVD-A-10M / -URLs	10M audio clips randomly sampled from BVD-V-55M
laion/BVD-A-1.7M / -URLs	1.7M audio clips from BVD-V-55M with unique source videos
laion/BVD-I-300M / -URLs (this dataset)	300M keyframes sampled from BVD-RAW

BVD-RAW itself (the raw pool of 80M videos totalling 10M hours) is not distributed through Hugging Face. The gated subsets and BVD-RAW are available to academic and non-commercial researchers through a single central access request form; the BVD-RAW access guide describes the raw-corpus transfer procedure.

Links

• Project page: https://projects.laion.ai/bvd/
• Paper: https://arxiv.org/abs/2608.24845
• Code: https://github.com/LAION-AI/bvd
• Terms of Use: bvd_terms_of_use.pdf
• Access request form: https://forms.gle/5tXPS4gmBwSGYYbV8