PD12M
Spawning
12.4 million public-domain and CC0 image-caption pairs for training image models.
A collection of 12.4 million CC0 or public-domain image-caption pairs assembled to train generative image models without copyright concerns. Data may be adjusted as infringing works are discovered. Distributed under a permissive data license.
Access
- Cost
- FreeNo charge to download; hosted in an AWS S3 bucket.
- License
- Open licenseCommunity Data License Agreement, Permissive 2.0
- Sign-up
- No sign-up
- How you get it
- API and Bulk downloadimages, Parquet metadata
What's in it
- One row is
- One image with a caption
- Covers
- Not geographic
- Years
- Not stated
- Updated
- Irregular
- Size
- 12.4 million image-caption pairs
Quality
Valuable as a clean-license corpus. Captions are machine generated and quality varies.
- Generated captions are imperfect
- Content may be removed over time
Connects to
- Common Crawl: The broader web crawl behind many image-text datasets
- Biodiversity Heritage Library Metadata and Page Images: Public-domain book images from a different source
Questions it could answer
- What subjects are over- or under-represented in public-domain imagery?
- How do caption lengths differ by source?