Buy licensed image datasets for AI training
Last updated How we research
In short
Image datasets are licensed collections of photographs, illustrations or scans with captions and rights records, used to train image generation and computer vision models. Licensed archives are now a business: Shutterstock's chief executive told Bloomberg that its AI licensing operation produced $104 million in revenue in 2023. Scraped sets carry real risk, as LAION-5B's withdrawal in December 2023 showed.
Key takeaways
- AI teams use image datasets for (1) text-to-image generation, (2) image classification and object detection and (3) vision-language models and captioning.
- Usually priced per image, or per collection for large archives.
- Common delivery formats: Original-resolution JPEG, PNG or TIFF, captions and tags in JSON or Parquet and bounding boxes or masks in COCO format.
- Check before you license image datasets: (1) US copyright: owner of every image, with signed transfers where rights were acquired, (2) US: BIPA if face geometry will be scanned, since photographs alone are excluded and (3) model and property releases.
- Every seller we shortlist passes 8 vetting checks before you see a sample.
Image datasets: buyer's spec
- Used for
- Text-to-image generation
- Image classification and object detection
- Vision-language models and captioning
- Document and specialist imaging models
- Delivery formats
- Original-resolution JPEG, PNG or TIFF
- Captions and tags in JSON or Parquet
- Bounding boxes or masks in COCO format
- Per-image rights and release records
- Usually priced
- Per image, or per collection for large archives
- Check before you sign
- US copyright: owner of every image, with signed transfers where rights were acquired
- US: BIPA if face geometry will be scanned, since photographs alone are excluded
- Model and property releases
- Illegal and harmful content screened
- Machine-generated captions disclosed
- Duplicates and near-duplicates removed
- Location metadata stripped where it identifies people
Why AI teams work with us
- 1
Every seller of image datasets we shortlist has passed eight vetting checks, from authority to sign through to a re-identification test.
- 2
Each shortlist comes with a vetting file per seller: rights evidence, provenance record, de-identification method, data manifest and proposed license terms.
- 3
Samples are drawn at random across the archive and labeled with the sampling method, so you can judge the whole package rather than its best rows.
- 4
You contract directly with the seller, and we help agree acceptance tests, staged payments and remedies before any data is delivered.
What is an image dataset for AI training?
An image dataset is a collection of images paired with text or labels, plus the records that prove the seller can license each one. Captions drive text-to-image models, and labels such as bounding boxes drive detection models.
Specialist images, such as medical scans, satellite imagery or industrial inspection photos, are sold with expert labels and carry the rules of their own sector, including health data law for medical images.
Who sells licensed image datasets?
Licensed image datasets come from stock libraries, photo agencies, publishers and companies with large in-house image archives. Shutterstock is the best documented: its chief executive Paul Hennessy told Bloomberg in June 2024 that its AI licensing business produced $104 million in revenue the previous year. See our Shutterstock data licensing profile.
Company archives count too. Mercor's published listing includes site photos, alongside project records and drawings, in construction company deals worth $115,000 to $1 million for firms of 50 to 2,000 employees, as compiled by ddiy.co in October 2026.
Why does provenance matter so much for images?
Provenance matters for images because scraped collections have contained illegal material that no buyer can lawfully hold. In December 2023 the Stanford Internet Observatory reported more than 3,200 images of suspected child sexual abuse in LAION-5B, and LAION took its datasets down to make them safe before republishing.
A licensed archive with a known owner, a documented upload history and moderation records reduces that risk sharply. Our guide to data provenance lists the records to ask for.
What does an image rights record contain?
An image rights record shows who made each image and why the seller can license it. Ask for these fields for every image, and check them in full for a random subset.
- Creator and date of creation
- License or assignment giving the seller the right to license it
- Model release for identifiable people
- Property release for private property or artworks where needed
- Moderation history showing the image was reviewed
How are image datasets priced?
Image datasets are priced per image for smaller sets and per collection for large archives. Published per-image training rates are rare, so most prices are negotiated from the type of license.
| License type | Reported range |
|---|---|
| Research | $5,000 to $15,000 |
| Commercial | $25,000 to $500,000 and above |
| Professional, annual | $25,000 to $50,000 a year |
| Enterprise, annual | $50,000 to $500,000 a year |
| Exclusive or custom | $1 million and above |
Getphyllo notes that a research license rarely permits production training, so check the license type before comparing prices.
How should an image dataset be delivered?
An image dataset should be delivered as original-resolution files with a manifest listing each image's ID, dimensions, source, license record and caption. Labels for detection or segmentation should use a standard format such as COCO, with the labeling guidelines attached.
Ask for a duplicate report or perceptual hash values, since a stock archive can hold the same photograph at several crops and sizes. Our guide on how to evaluate a dataset sample covers the tests to run.
What should you check before buying an image dataset?
Before buying an image dataset, audit rights for a random subset, screen the full set with hash-matching tools for known illegal content and check captions against the images.
Strip location data from EXIF metadata where it could identify people's homes, and confirm model releases for identifiable faces. Illinois' Biometric Information Privacy Act lists a scan of face geometry as a biometric identifier but excludes photographs themselves, so the risk rises when face templates are extracted. California's CCPA counts face imagery from which a faceprint can be extracted as biometric information when it is used or intended to be used to establish identity.
If you hold UK or EU images, UK GDPR Article 4 gives facial images as an example of biometric data when they are processed to identify a person.
Our due diligence checklist covers the contract. To source vetted image archives, book a free consultation.
Frequently asked questions
Is it safe to train on scraped image datasets?
How much do licensed image datasets cost?
Do images of people need model releases for AI training?
Are captions included with licensed image datasets?
Sources
- Shutterstock's AI-licensing business generated $104 million, Bloomberg Law, accessed November 2026
- AI image dataset is pulled after child sex abuse pictures discovered, PetaPixel, accessed November 2026
- Best public data providers for AI training and grounding, Getphyllo, accessed November 2026
- How to sell company data to AI companies (2026 guide), ddiy.co, accessed November 2026
- Biometric Information Privacy Act, 740 ILCS 14, Illinois General Assembly, accessed November 2026
- California Civil Code section 1798.140 (CCPA definitions), California Legislative Information, accessed November 2026
- UK GDPR Article 4: definitions, legislation.gov.uk, accessed November 2026
For AI teams
Ready to source?
Book a consultation to brief us. We shortlist vetted sellers, each with a vetting file covering rights evidence, provenance and a representative sample.
Related guides
- Data provenance explained for AI teams
Buying data
- How to buy AI training data
Buying data
- AI training data due diligence checklist
Buying data
- AI training data pricing models
Pricing and value
- How to evaluate a dataset sample before you buy
Buying data