LAION review: the best open AI datasets for vision and video
Answer in brief: LAION publishes the largest open multimodal datasets. Re-LAION-5B (August 2024) has 5,526,641,167 text–image-link pairs, cleaned of known links to suspected CSAM, in research and research-safe versions. LAION-BVD (August 2026) adds 80 million downloaded videos (10 million hours), 55 million captioned clips and 300 million captioned frames, plus audio subsets. LAION-BVD is for research only, not commercial use, and Re-LAION-5B contains links, not images, so check terms and plan your downloads.
Best for: pretraining vision-language, video-language and audio-language models at web scale.
Independent research from LAION's official projects page and release posts, checked 2026-11-09. Not hands-on testing.
What it gives you, by task
- Image-text pretraining: Re-LAION-5B research and research-safe, both available as Hugging Face collections. LAION says its metadata can be used to clean existing LAION-5B derivatives.
- Video-text: LAION-BVD-V-55M, with 55M scene-level clips and captions generated by Qwen3-VL.
- Audio-text: LAION-BVD-A-10M and A-1.7M audio subsets, captioned with Audio Flamingo 3. LAION also lists a separate audio dataset project for training CLAP models.
- Image-text from video: LAION-BVD-I-300M, with 300M captioned frames taken at scene changes.
- 3D: LAION-3D is listed as an effort to build a large set of 3D models paired with descriptions.
How LAION-BVD was built
LAION extracted 1.3B video URLs for specific platforms (YouTube, Vimeo, Dailymotion) from Common Crawl WAT files using yt-dlp extractors, then downloaded about 80M videos. LAION says 43% of the videos are non-English. Its posted experiments report that ViCLIP models trained on BVD beat InternVid-trained baselines by up to 3.3 percentage points on aggregate benchmarks. The release includes the URLs, captions, subset metadata, and training and evaluation code.
Who it's built for
Academic and open-source teams training CLIP-style, video and audio foundation models, and safety researchers who need reproducible reference datasets.
Where it shines
- Scale that other open image and video datasets don't match.
- Safety revisions. Re-LAION-5B removed 2,236 links after matching against hash lists from IWF and C3P, together with the Stanford Internet Observatory.
- Open tools: img2dataset, CLIP Retrieval, CLIP Benchmark and OpenCLIP.
Honest drawbacks
- Re-LAION-5B contains links, not images, so you have to download and filter the content yourself, and some links will be dead.
- LAION-BVD is research-only and explicitly not for commercial use.
- Older releases such as the original LAION-5B have been superseded.
- LAION warns that the data may contain biases and uneven coverage of languages and regions.
Pricing and licenses
Free. LAION's release post says Re-LAION-5B is released under Apache-2.0; the linked images belong to third parties. LAION-BVD has its own terms of use and privacy policy and is for research only.
Best alternatives
- Hugging Face Hub: hosts the LAION collections and many other vision datasets.
- Common Crawl: the source LAION-BVD's URLs were extracted from.
- Kaggle: smaller labeled image datasets with free GPU notebooks.
Need human-labeled detection or segmentation data instead of web-scale pairs? Google's Open Images V7 has about 9M images with 16M bounding boxes on 600 classes; its annotations are CC BY 4.0 and its images are listed as CC BY 2.0, which Google says you should verify for each image.
FAQ
How many pairs are in Re-LAION-5B?
5,526,641,167 text–image-link pairs, in research and research-safe versions.
Can I use LAION-BVD commercially?
No. LAION says LAION-BVD is released exclusively for research and not for commercial use.
Does LAION host the images?
Re-LAION-5B provides links to images, not the images themselves. You download them with tools such as img2dataset.
Sources
- LAION projects
- Releasing Re-LAION-5B
- LAION-BVD announcement
- LAION-BVD on Hugging Face
← Back to the ranked list of the best AI datasets
Named on this page as a provider? Join the network for a linked profile, Visit buttons and featured placement.