Selling photos and image archives for AI training

You can sell photos for AI training, but rarely one at a time: AI companies buy images by the library, from stock agencies, data catalogues and the holders of large archives, and they pay a few cents an image for everyday stock and much more for specialist, captioned and released collections. Across the sales and payouts on record, the median is $0.029 per image, against $0.35 per image in what buyers publicly offer holders and $1.71 per image in what data catalogues ask. If your photos already sit with a stock agency, the agency does the selling: Shutterstock passes on a 20% average share of what it receives for data licences, Adobe pays a discretionary yearly bonus, and neither publishes a price per image.

This guide is for anyone who holds a body of images: a photographer with a back catalogue, a studio or agency with years of commissioned shoots, a publisher, brand, museum or archive with its own collection, and the contributors whose work already sits with a stock agency. It covers who buys images for AI training and what for, what the agencies pay their contributors, what archives sell for on each side of the market, how each agency handles consent, what model and property releases do and do not cover, why stock licences exclude training, how to prepare an archive, and how a sale through AIData.Broker works. Every number links to its record.

AIData.Broker keeps the price board, the buyer board and the images page this guide draws on, and it also acts as the agent of holders who license or sell their data, so we are a party in this market. Check any figure you plan to rely on against its source.

This guide is research from public sources, not legal, tax or investment advice.

Contents

1. Who buys images for AI training, and for what
2. What stock agencies pay contributors
3. What image archives sell for
4. Consent, agency by agency
5. Model and property releases
6. Why stock licences exclude AI training
7. How to prepare an image archive
8. How a sale through AIData.Broker works
9. Questions people ask
10. What this means for you

Who buys images for AI training, and for what

Three kinds of model learn from photographs, and each wants something different from an archive.

• Image generators (systems that turn a text prompt into a picture) learn from pairs of images and descriptions. They want breadth: every subject, style and lighting condition, each picture with an accurate caption.
• Vision-language models, the assistants that answer questions about a photo, a chart or a scanned page, learn to read images the same way, and they also want breadth and good descriptions.
• Computer vision systems learn narrower tasks: finding objects, reading licence plates, grading skin conditions, spotting faults on a production line, reading an X-ray or a satellite scene. They want specific subjects, captured and labelled precisely, and they pay most for what is hard to photograph.

The first image generators learned from pictures taken from the open web. LAION-5B, a public index of web images and their captions, lists 5.85 billion image-text pairs gathered that way. That route now comes with litigation: Getty Images' suit against Stability AI is in the fact-discovery stage in California, and Disney and Universal sued Midjourney in 2025. A buyer that wants to be sure it may use what it trains on now licenses it, and images were among the first kinds of data to be licensed this way.

Who pays, and how

Labs that license whole libraries

At the top of the market, the deals are with whole libraries. Shutterstock earned $104 million of AI licensing revenue in 2023 across all its AI customers, and its first deals with large technology companies were later expanded. Freepik licensed the majority of its 200-million-image archive to two large technology companies at 2 to 4 cents per image, its chief executive told Reuters. Getty Images disclosed two multiyear deals with a combined value of approximately $65,000,000 in early 2026, one with a social platform and one with a large AI company; it recognised about $40M of that in the quarter the deals were signed. Shutterstock also licensed its visual media to the model developer Reka AI, and Wirestock, which began as a stock distributor, now supplies six of the largest foundation model makers and has paid out $15 million to contributors.

These are agencies with hundreds of millions of images, most of them captioned and keyworded. They show what a very large library can negotiate. They are not a benchmark for a photographer's hard drive, and the price per image inside them is low: a few cents an image in Freepik's case.

Demand is shifting from bulk to specific

The agencies' own numbers suggest the first wave of bulk stock licences has been sold. Shutterstock's Data, Distribution, and Services revenue, the line that holds its AI data licensing, rose to $203.3 million in 2025, but the fourth quarter of that year was down 19% on the year before, and the second quarter of 2026 came in at $56.1 million, below the same quarter of 2025. Getty's newer agreements with OpenAI and Perplexity are display agreements that show Getty images with credits inside answers, rather than training licences, and Getty's own guidance treats big archive deals as one-off sales that do not repeat each year.

What buyers still look for is what they lack: subjects stock agencies hold little of, precise captions, and documented consent from the people in the picture. That is the part of the market where an independent archive can still compete.

A separate market: photos taken to order

Many buyers do not buy archives at all. They order new photos to a brief, through collection projects that pay per accepted image: DataForce pays $1.50 per accepted photo, Mindy Teams $0.30 per approved photo and Firstsource $0.28 per approved image of everyday objects. Across the collection side of the market, the median paid to the people who take the photos is $0.54 per image. A Japanese data exchange publishes open calls that show how much the subject matters: 20 points per photo of a television display, 70 points for building exteriors and 500 points for parked motorbikes, a point being a yen. If you are thinking of taking new photos rather than licensing what you have, those projects are the market; this guide is about the images you already hold.

What stock agencies pay contributors

The question contributors ask most is what an agency pays per image when it licenses their work for AI training. The plain answer is that neither Shutterstock nor Adobe publishes a per-image price on its own contributor pages. Shutterstock publishes a share of revenue, split in proportion to what each contributor's work made up of the datasets sold. Adobe publishes a bonus, paid at its discretion, with no formula. Getty states that contributors are compensated and gives no rate. Per-image figures exist only where contributors have reported their own payments, and every such figure below is labelled as a contributor's report.

Shutterstock: a share of a pooled fund

Shutterstock's contributor help page sets out how money from data licences reaches contributors: they receive a share of the whole contract value, proportionate to the volume of their content and metadata included in the datasets bought, at a 20% average corporate royalty rate of what Shutterstock receives. The money does not arrive per image. Earnings from datasets are "pooled in a collective fund and will be distributed periodically as the fund accumulates significant revenue for distribution", and appear in a Contributor Fund column of the earnings summary. A dataset inclusion never shows up as a download, so a contributor cannot read a price per image from the account.

Three more facts from Shutterstock's own pages shape what contributors can know. Payouts started with the OpenAI agreements: dataset earnings were "first issued in Q4 2022". Its contributor terms treat "royalty rates, royalty payments and earnings data" as confidential information, one reason contributors rarely publish what they receive. And since mid-2023 it accepts content that fails its stock review but is good enough for datasets, collected in a "Data Catalog" in the contributor's account, so a rejected image can still earn as training data.

What contributors have reported is small. A survey of contributors run by the photographer Robert Kneschke found a median of $0.0069 per portfolio image for one half-yearly payout, and the same record notes a contributor who told the UK competition authority in 2026 that their only fund payment was $5.35 on a portfolio of 16,960 assets. At the level of the whole company, the trade analyst Paul Melcher estimated Shutterstock's 2023 AI revenue at an average return of $0.135 per image across its archive, before the contributors' share. These are reports, not Shutterstock's figures, and they describe early payouts.

Adobe: a discretionary yearly bonus

Adobe trains its Firefly models on Adobe Stock content and pays contributors a yearly Firefly Contributor bonus. Its contributor FAQ says the latest was paid on 17 September 2026 and that "Bonus amounts will vary for each eligible Contributor and is paid at Adobe’s discretion". The 2026 bonus was based on two things: the content "considered for training between June 3, 2025, and June 2, 2026", and the licences those assets earned in the same twelve months. So both the size of a portfolio and how well it sells count. Adobe will not say what comes next: "We cannot share details about future bonus payments at this time."

The bonus is voluntary. Adobe says its existing terms "permit Adobe to create and train Firefly models", and the Adobe Stock Contributor Agreement grants a royalty-free licence for purposes that include "developing new features and services". Contributors cannot opt out of training (more on that under consent).

The per-image figures again come from contributors. Two who commented on Kneschke's blog after the first bonus in 2023 reported $12.60 for 1,406 photos online, about $0.009 per image, and $117.19 for 3,813 approved images that had sold well, about $0.031 per image. Kneschke later pooled the reports of 23 contributors and found an average of about 6.7 cents per sale in the bonus period, per licence sold rather than per image. Separately, Bloomberg reported in 2024 that Adobe was paying between 6 cents and 16 cents for each photo; Adobe's own pages publish no such figure.

Getty Images and iStock

Getty Images' 2025 annual report says contributors "are compensated for any inclusion of their content in AI data training sets and may share in the revenue generated by AI tools and services trained with their content", and gives no rate. For ordinary licensing, the same filing says suppliers under contract receive royalties of 20% to 50% of each licence fee. What Getty pays for inclusion in a training set has not been published.

Alamy: an opt-in share

Alamy licenses for machine learning only through its opt-in Novel Use scheme, which pays up to 40% of novel use sales. A Novel Use sale pays the same split as a direct sale, so a new contributor on the starting rate receives 20% of each sale, and the ceiling for images exclusive to Alamy is up to 50%. For machine learning in particular, its contributor terms say the payment is at Alamy's discretion and "calculated based on a reasonable pro-rata of the total amount of Content licensed to the Customer", and trials for which Alamy charges nothing pay nothing. Alamy's Novel Use collection trained Bria's image model, and Alamy says that "Once the generated images start being used commercially the attribution framework will kick in and Alamy contributors will receive their share". Its AI collection is sold to model builders on enquiry, and it says "we always share revenue from data training deals with our contributors".

Canva, EyeEm, Depositphotos and Wirestock

Canva committed $200 million in content and AI royalties over three years to its creators, paid through a Creator Compensation Program to those who consent to training; it is a pool for all creators, not a rate per image. EyeEm, now part of Freepik, paid 50% of the licence fees it received on sales through its market, under terms that also made uploading the grant of training rights. Depositphotos runs its generator on Bria and says "Bria’s patented attribution engine ensures that creators whose work contributes to the AI training datasets receive fair compensation", with no rate stated. Wirestock set its fee on dataset deals at 50 percent, so a contributor keeps half of what its work earns in them.

The chart sets the per-image figures against what images fetch elsewhere. Its point is that what reaches a contributor through an agency sits at the bottom of the market, well below what the same agencies' libraries fetch per image and far below what curated collections are quoted at.

Sources: Shutterstock fund median, Adobe bonus, small portfolio, Adobe bonus, selling portfolio, Freepik archive, Adobe per photo, Shutterstock AI revenue per image, bulk photo vendor, photo archive rates discussed, curated images.

What image archives sell for

Per-image prices for images span a very wide range, because "an image" covers everything from a phone snapshot to a pathology slide. The useful way to read them is by side of the market (who pays whom) and by kind of image.

The market rates, from every record

The rates page computes the market's own figures from every public price per image at once. Today the median sale or payout that happened is $0.029 per image per image, a figure pulled down by the contributor payouts above; the median offer buyers publish to holders is $0.35 per image; the median asking price in data catalogues is $1.71 per image; and the middle half of all asking prices per image runs $0.27 to $5.27 per image. On the commissioning side, buyers pay vendors a median of $0.44 per image per image collected or labelled to order, and photographers on collection projects receive $0.54 per image. Whole-library licences, priced per contract rather than per image, sit at $378K to $35.4M per contract.

Bulk stock: cents per image

Everyday stock photography is the most plentiful kind of image, and its price shows it. Freepik's archive licensed at 2 to 4 cents per image; a bulk photo vendor publishes rates between $0.01 and $0.25 per image, varying with resolution and exclusivity; and Photobucket, the photo-hosting service, discussed rates of 5 cents to $1 a photo with buyers in 2024, though no deal was reported and it has since said it does not sell photos for AI training. A consumer app that buys people's camera rolls pays at the bottom of that range: users report about $7 for 1,000 photos.

Curated and specialist: dollars per image

Prices rise with curation. Defined.ai, a data vendor, told Reuters that curated images with metadata sell for $1 to $2 an image, and that sensitive content needing special handling sells for $5 to $7 per image. The top of the curated rate is about 100x the bottom of the bulk archive rate. Wirestock's chief executive said bulk sales pay contributors small amounts while specific briefs pay up to $15-20 per artwork.

Specialist collections sit higher still, because nobody can photograph them on demand. A pathology data provider lists stained whole-slide images from $2.00 per slide in volume to $5.00 per slide for small orders, and specially stained slides at $40.00 per slide. A collection of 187K+ X-ray images is listed at $914,500, and a satellite imagery provider sells a single archive image starting at $15. At the other end, an illustrator lists a collection of hand-drawn illustrations at $78,900, and a set of 5,000 jewellery product photos asks $130. Catalogue prices are asking prices, not sales; they show what holders hope for, and the range is the point.

Whole libraries: negotiated licences

A library large and distinctive enough to negotiate is priced as a whole, not per image. The disclosed examples are agency-scale: Shutterstock's initial deals of $25 to $50 million, Getty's approximately $65,000,000 across two multiyear agreements. Two lessons from them carry down to smaller libraries. Most of a multi-year licence can be collected at the start, so the structure of payment is worth negotiating as much as the total. And an archive sold in bulk to one buyer is often a one-off, so a non-exclusive licence that leaves the holder free to license again is worth more over time; non-exclusive data sells more than once.

What the valuation uses

The valuation on this site prices everyday photography at $0.02 to $0.25 an image, from what one large archive licensed at per image to the top of a bulk photo vendor's rates, and professional libraries with metadata at $0.25 to $2 an image, up to a data vendor's quoted rate for curated images. On its starting answers, a small everyday library in which nobody is recognisable, or everyone has signed a release, comes out at $100 to $1,300. That is a useful number to know before anyone quotes one. It also explains why most individual photographers sell through agencies: a small everyday library is worth too little to justify a negotiated sale on its own.

Why the agency route pays so little per image

Three things push a contributor's figure down. Shares of a bulk price: the agency licenses its whole library at a bulk rate of a few cents an image, and the contributor receives a share of that, a fifth on average at Shutterstock. Pooling: when the money is split across hundreds of millions of images, each image's part is tiny, whatever the headline value of the deal. Volume of everyday stock: a buyer choosing between libraries of the same ordinary subjects pays the ordinary rate. What lifts an image above that is what a bulk library does not have: a specialist subject, a description that tells a model exactly what is in the frame, and consent it can rely on.

Consent, agency by agency

How an agency asks its contributors before their work is used for training is now one of the clearest differences between agencies. There are four models in use, plus agencies that name none, and the consent record keeps the sources for each.

Opt-out, in by default. Shutterstock added the setting in January 2023, and contributors can opt out of "having your content included in future datasets". Few have: its head of data licensing told VentureBeat that only about 1 percent chose to. Canva makes stock media available for training by default, and a contributor who opts out "will no longer be eligible for the Magic Creators Payment". In both cases an opt-out works forward, not back.

Opt-in. Alamy makes training a choice. A contributor who grants Novel Use rights lets Alamy "license your Content at any price, by any method and for any use we feel appropriate", including machine learning, trials and high-volume, low-price licences, without consulting the contributor. That is a broad grant, and it can be withdrawn "at any time during the month of April in each year". Opt-in is better for the contributor than the alternatives, but the grant itself is wide once given.

No choice. Adobe states plainly: "We currently do not offer an opt-out option for Adobe Stock content." EyeEm's terms went further, granting "the sublicensable and transferable right to use your Content for the training, development and improvement of software, algorithms and machine learning models" to anyone who uploaded; eyeem.com now redirects to Freepik. Freepik trained its own F Lite model on approximately 80 million images from its internal library, and the model card names no contributor opt-in or payment.

Reversed. Photobucket, which discussed licensing its users' photos for training in 2024 at rates of 5 cents to $1 a photo ( no deal was confirmed), now tells visitors: "We don't sell your photos or videos to train AI models."

What this means in practice: a photographer chooses the consent model by choosing the agency, and the general grants in agency contracts often already cover training. Shutterstock's contributor terms grant it rights in accepted content "to develop new features and products", and Adobe's agreement covers "developing new features and services". A holder that wants to decide each training licence itself, and set its price, keeps those images out of agencies whose grant already covers training and licenses them directly.

Model and property releases

A model release is a person's written permission to use their image; a property release is the same for a private building, artwork or recognisable object. Stock agencies require them before an image can be licensed commercially. For AI training they matter in a different way, and they cover less than most photographers assume.

What the agencies require. Shutterstock asks for valid model releases for content "that, in Shutterstock's judgment, contains an identifiable face or identifiable human figure" or other identifiable attribute, including voice and likeness. Canva requires releases "for all Stock Media you upload to Canva that, in our judgment, contains an identifiable face, identifiable person, other identifiable attribute, or property". EyeEm required consent documentation for anyone "recognizable or identifiable by certain characteristics or attributes (e.g. scars or tattoos)". Alamy restricts content showing people, property or intellectual property "to Editorial Use only" where no release has been indicated. Adobe requires that anyone generating images from photos of real people "may only prompt using images of people for whom you have a signed model release".

What a release does not cover. Many releases in existing archives were written for advertising and editorial use, before photos were licensed for training. Getty's licence says so directly: it does not warrant that consent has been obtained for machine learning or identification uses "with respect to model-released content", which is the release caveat recorded with the licence terms. Shutterstock does not pass releases to dataset buyers: "full model releases are never shared, and the identity of models and contributors is never shared with computer vision partners", although broad demographic fields taken from releases (age, gender, race or ethnicity) can travel in dataset metadata. So a release proves a person agreed to be photographed and published. Whether it extends to training depends on its wording, and a buyer reading the release will look for AI training, or at least onward licensing for any purpose, named in it.

Faces are biometric data in some places. Illinois treats scans of face geometry as biometric identifiers, and scraping public photos to extract face geometry led to a nationwide class action against Clearview AI. The FTC, the US consumer protection regulator, has ordered a photo app developer to delete not only the photos it misused but the models trained on them. A buyer that has read those cases asks about faces before it asks about price. The standard a careful buyer looks for is the one in the biometrics rights question: images of consenting people with written releases naming AI training and sale, and no biometric templates extracted by the holder.

Property, art and brands. A property release covers a private building, a recognisable artwork or a product design. Logos, artworks, sculptures and other people's photographs inside the frame are someone else's work, and a buyer's third-party content question expects them to be identified and either removed, cropped or licensed.

For a holder, the practical split is three ways: images with nobody identifiable and no protected property, which can go to market at once; images with people or property covered by releases whose wording reaches AI training; and images with identifiable people and no adequate release, which need new releases, de-identification or leaving out.

Why stock licences exclude AI training

Each of the large stock agencies whose licences we read writes AI training out of its ordinary customer licence. Getty Images says that, unless an invoice or licence agreement explicitly authorises it, a customer "may not use content (including any caption information, keywords or other metadata associated with content) for any machine learning and/or artificial intelligence purposes"; iStock's licence has the same clause. Shutterstock's licence forbids using visual content "as training data for any artificial intelligence, machine learning, or generative AI system, tool, process, or dataset". Adobe's enterprise Stock terms bar using stock assets "to directly or indirectly create, train, test, or otherwise improve any machine learning algorithms or artificial intelligence systems". Alamy allows editing with AI tools on the condition that "the software must not use or incorporate the Content for the purposes of image generation or machine learning training purposes". Freepik's terms bar "any machine learning and/or artificial intelligence purposes".

The reason is commercial as much as legal. Training is a different use from a single-image licence and is priced differently, so the agencies sell it separately, as its own product. Shutterstock's dataset buyers "may only use them to train machine learning and computer vision models", with commercial use of the images themselves prohibited. Its data licensing page offers AI data from more than 640 million images on enquiry, Depositphotos offers more than 330 million assets alongside partner datasets, and Shutterstock also invites outside libraries to "license your content for AI training" through its Data Partner Network. Unsplash gives away a small sample dataset, "Free for commercial and non-commercial usage", and sells its full dataset on request.

The courts are part of the reason too. The UK High Court's judgment in Getty's case against Stability AI did not decide whether training infringes, because Getty dropped that claim for want of evidence that training happened in the UK; its secondary infringement claim was dismissed, and permission to appeal was granted. While the law is unsettled, a written licence is what lets a buyer show where its images came from.

Two consequences follow for anyone holding images. Stock images you licensed in are not yours to pass on. An agency, publisher or brand archive often contains stock photos bought for past campaigns; those must come out before an archive is offered for training, or the holder breaches the agency's licence and the buyer inherits the problem. Your own images can be licensed for training separately, even where you also sell them as stock: the agency's exclusion binds its customers, not you, unless your contributor agreement made the images exclusive to the agency or already granted it training rights. Read your contributor agreement before you offer anything you also sell through an agency.

How to prepare an image archive

Preparing an archive is mostly a matter of knowing exactly what is in it and what you may do with each part. The guide to preparing data for sale covers the general method; the points below are the ones particular to images.

1. Count and describe what you hold

Start with an inventory: how many images, of what subjects, from which years, at what resolution and in which formats (raw files, full-resolution JPEG or TIFF, edited versions). Separate originals from edits and published from unpublished. Unpublished originals, outtakes and alternates are often worth more than the published frame, because they are not already on the open web.

2. Captions, keywords and metadata

Metadata is part of what is sold. Shutterstock splits its fund by each contributor's content and metadata included in a dataset, and Getty's licence protects "caption information, keywords or other metadata" alongside the image itself. A caption that says exactly what is in the frame (the subject, the setting, the action, the lighting, the lens) is what turns a picture into training data for an image generator or a vision-language model. Existing stock keywords, shoot notes and catalogue records are worth converting into one consistent file. For computer vision buyers, labels on the image (what is where in the frame) raise the value further.

3. Rights: chain of title, agencies and licensed-in images

For each body of images, establish who holds the copyright. Your own photos are yours; work you shot for a client may belong to that client under its contract with you; work by staff photographers usually belongs to the employer; work by freelancers belongs to them unless it was assigned in writing. Then check every agency agreement: images exclusive to an agency, or covered by a grant that already includes training, may not be free to license elsewhere. Finally, find and remove stock images and other third-party content you licensed in. The rights questions set out what a buyer will ask, and the answers become your rights file.

4. Duplicates and near-duplicates

Archives hold many near-identical frames: bursts, bracketed exposures, the same shot in several edits and sizes. Buyers count unique images and pay for those; duplicates inflate the count without adding value and can count against a dataset in a buyer's quality review. Remove exact duplicates, and group near-duplicates so the buyer can choose.

5. People, faces and personal data

Sort images by whether anyone is identifiable, and for those who are, whether a release covers AI use. Where it does not, the choice is new releases, leaving the images out, or masking what identifies a person: faces, tattoos, licence plates, house numbers and screens. Strip embedded metadata such as GPS coordinates and camera serial numbers from every file either way. Our de-identification standard lists pictures of faces among what we do not sell unless the people in them consented in writing to AI training and sale. Masked images are worth less to a buyer who needs faces, but they can be licensed where the alternative is leaving them out altogether.

6. A sample and a datasheet

Buyers ask for a sample before they price, and they want it to be representative: the same mix of subjects, quality and metadata as the whole. Prepare a sample large enough to show the whole range, with its captions, and a one-document description of the whole collection. The datasheet tool structures it.

How a sale through us works

AIData.Broker represents holders selling their image libraries to AI companies. We work for the holder only: we value the library against public prices, settle its rights with the holder, and put it before every buyer that fits at once, so the holder sees competing offers rather than the first one. The holder itself licenses or sells to the buyer, and we do not take ownership of the images.

1. A valuation first. The valuation asks a few questions about the library (how many images, what kind, whether people in them have signed releases) and returns a range from the public prices on the board, with no upload. Files are optional and come later.
2. The rights file and the datasheet. We go through the rights questions with the holder, record the answers, and write the datasheet that describes the library; the holder approves it before any buyer sees it.
3. The round. A blind profile of the library goes to every buyer that fits, without the holder's name, less any buyer the holder leaves out. Buyers who want to go further sign an undertaking and see the datasheet and, where the holder agrees, a sample.
4. Offers and the licence. Buyers make offers, which we set side by side; the offer check reads any written offer against market terms. The holder chooses, and signs the licence with the buyer directly.

The steps and documents are on how it works. Our fee is a success fee on what the holder is actually paid, band by band: 15% of the first $250,000, 10% from $250,000 to $1 million, and 6% above $1 million, and nothing if nothing sells; on a $300,000 sale it comes to $42,500. The mandate makes us the holder's only agent for the library it names for 2 months, and a sale to a buyer we introduced, made within 12 months after it ends, still carries the fee if that buyer is on the list we send in writing when it ends. We run sales only for holders based in the United States, the United Kingdom and Canada.

Which libraries fit. The smallest sale we take on is $26,000. At the valuation's everyday rate of $0.02 to $0.25 an image, a general library has to run to hundreds of thousands of images or more to clear that; at the professional rate of $0.25 to $2 an image, a well-described professional library gets there with far fewer. A photographer with a small everyday portfolio is usually better served by a stock agency or a data catalogue, and the images page names the routes for bulk libraries. The libraries where a round makes the most difference are specialist and professional ones: medical, scientific, industrial, aerial or satellite imagery, historical collections with good records, and large professional archives with captions and releases, where buyers' needs and prices differ widely. Pictures of faces go into a sale only where the people in them consented in writing to AI training and sale, as our de-identification standard sets out. The page on going direct or through us sets out the trade-off.

Questions people ask

Can I sell my photos for AI training? Yes, if you hold the copyright and the rights to what is in them. Most individual photographers do it through a stock agency, which licenses its whole library to AI companies and passes on a share: Shutterstock's is a 20% average of what it receives for data licences. Holders of large or specialist archives can license directly, to labs or through data catalogues, under a written licence.

How much do AI companies pay per photo? Everyday stock licenses for a few cents an image ( 2 to 4 cents in Freepik's case), curated images with metadata for up to 100x the lowest bulk rate, and specialist images such as medical slides for more. Across the public records, the median sale or payout is $0.029 per image per image and the median asking price $1.71 per image.

How much does Shutterstock pay contributors for AI training? Shutterstock publishes a share, not a price: contributors receive a 20% average of data licence revenue, split by how much of their content and metadata went into each dataset sold and paid periodically from a pooled fund. A contributor survey found a median of $0.0069 per image for one payout; that is a report, not Shutterstock's figure.

How much is the Adobe Firefly bonus per image? Adobe does not say. The bonus is paid yearly at its discretion, based on content considered for training and the licences it earned, and Adobe publishes no per-image amount or formula. Contributors who published their 2023 bonus reported about $0.009 and $0.031 per image.

Can I opt out of AI training on stock sites? It depends on the agency. Shutterstock and Canva let you opt out of future use (at Canva, at the cost of the creator payment); Alamy uses training only if you opt in; Adobe offers no opt-out, so the only way out is to remove your content. An opt-out does not reach datasets already sold. The consent record keeps each agency's terms.

Can I sell images I bought from a stock site? No. Getty, iStock, Shutterstock, Adobe, Alamy and Freepik all exclude machine learning and AI use from their customer licences unless separately authorised, so images licensed in from them cannot be passed on as training data. Remove them from an archive before you put it on the market.

Do I need model releases to sell photos for AI training? For identifiable people, a careful buyer wants releases whose wording reaches AI training or onward licensing for any purpose. Older releases written for advertising may not: Getty does not warrant that model-released content covers AI uses. Images without adequate releases sell at a discount, sell only where nobody can be identified, or need faces masked.

Is it better to sell through an agency or directly? For a small everyday portfolio, an agency or catalogue is usually the only practical route, and its share is the price of not running a sale yourself. For a large, specialist or well-described archive, a direct licence lets you keep the choice of buyer, the terms and most of the price; agencies and platforms keep up to about half of a data deal.

Should I give a buyer exclusive rights to my archive? Only for a price that reflects it: exclusive licences fetch about 2.5 to 5 times non-exclusive ones. Keep any exclusivity to a defined field and a defined window, and remember that a non-exclusive archive can be licensed again.

What this means for you

Selling images for AI training is a rights exercise with a price attached. The holders who earn the most per image have pictures nobody else has, described precisely, with every right and release in order, and they decide their own terms rather than accepting an agency's default. The steps below turn that into a plan.

1. Know your number before anyone quotes one. The price board and the rates page hold every public price per image, and the valuation puts a range on your own library from them. The guide to AI training data prices shows where images sit against other kinds of data, and what labs buy shows where demand is now.
2. Read your agency agreements. Find out which of your images sit with which agency, whether any are exclusive, and what each agency's grant already allows. Choose your agencies by their consent model if training matters to you.
3. Build the rights file. Chain of title, releases, licensed-in images removed, faces de-identified where needed. The guides on fair use and AI training and scraping for AI explain why buyers now pay for images they are sure they may use.
4. Describe the library. Captions, keywords and labels are part of what is sold. Convert what you have into one file, and prepare a representative sample and a datasheet.
5. Match the route to the library. A small everyday portfolio belongs with an agency or a catalogue; a specialist or professional archive is where competing offers make a difference. Compare buyers on the images page and the buyer board, read the guide to data licensing agreements for the clauses that matter, and check any written offer with the offer check. The sister guide on selling video to AI companies covers footage.

AIData.Broker acts as the holder's agent: the holder itself licenses or sells its images to the buyer, and we do not take ownership of them. The holder pays us a fee fixed in the mandate and paid only if a sale closes, and we run sales only for holders based in the United States, the United Kingdom and Canada.