Google Cloud Vision and Amazon Rekognition will label a photograph accurately and cheaply, and neither will get that photograph onto Adobe Stock. General-purpose vision APIs return the objects they detect. Tools built for stock contributors return the title, description and ranked keyword set an agency actually stores, then move the file there. That gap is the entire comparison.
Two categories solving two different problems
A vision API answers a computer vision question: what is in this picture? It was built for developers who need to sort user uploads, moderate content, or make an internal photo library searchable. The output is a list of concepts with confidence scores, and it is genuinely good at producing that.
A stock contributor tool answers a commercial question: what does this agency need in order to accept, index and sell this file? That is a different specification. It includes a readable title in the agency house style, a keyword set ordered by importance, a category drawn from a fixed list, release flags, and a delivery route that puts the file and its metadata where the agency expects to find them. None of that is a vision problem.
So the two categories are rarely a real either-or decision. They sit at different points in the same pipeline, and the useful question is which one you should be paying for given what you actually do with your files once they are tagged.
What a general-purpose vision API actually returns
Google Cloud Vision
Label Detection returns broad descriptive concepts, and separate features cover object localisation, text, logos, landmarks and safe-search flags. Each feature applied to an image is billed as its own unit. Google's published pricing puts the first 1,000 units per month at no cost, then $1.50 per 1,000 units for label detection up to five million a month, with object localisation at $2.25 and web detection at $3.50 per thousand. Accuracy on everyday subjects is high, and the concepts come back clean and consistent.
Amazon Rekognition
Rekognition groups its image APIs by type, with DetectLabels sitting in the group that AWS's own pricing examples bill at $0.0010 per image for the first million each month, which works out to about $1.00 per thousand files, dropping at higher volume. New accounts get 1,000 free images a month for the first year. Running a second API against the same image counts as a second image, so multi-feature analysis multiplies the cost.
Clarifai and the wider tagging market
Beyond the two hyperscalers sits a layer of specialist vision platforms that sell general tagging alongside custom model training. Their strength is that you can teach a model your own vocabulary, which matters if you work in a narrow niche where generic labels are useless. Their pricing tends to follow a usage or seat model rather than a flat per-thousand rate, so it is worth reading the current terms on the vendor's own page rather than trusting a comparison article.
Where that output falls short of an agency submission
Point a vision API at a photograph of a woman laughing at a kitchen table with a laptop and you will get back something like: person, table, laptop, indoor, furniture, smile. Every one of those is correct. None of them is what a designer types into a stock search bar. Four specific gaps show up every time.
- No title or description. Agencies want a sentence, not a bag of nouns. Adobe's guidance asks for a short title, ideally under 70 characters, that describes what is happening in the frame. A label list is not a sentence and cannot be turned into one by concatenation.
- No concepts, only objects. Buyers search for remote work, burnout, small business, freedom, candid. Those are not visible objects, so a detector will not name them, and they are frequently the terms that sell the file.
- No ordering and no ceiling. Adobe caps keywords at roughly fifty per submission and treats their order as a ranking signal. A vision API returns whatever crosses the confidence threshold, sorted by confidence, which is not the same as sorted by commercial importance.
- No policy awareness. Detectors will happily return brand names, celebrity look-alikes and trademarked landmarks. Submitting those terms is one of the faster routes to a rejection, and nothing in the API is going to warn you.
There is also the part that has nothing to do with metadata at all: the file still has to reach the agency, in the right format, with releases attached and its keyword row correctly matched. The rules that govern what moderators accept are their own subject, covered separately in our guide to writing titles that comply with stock moderation policies.
What contributor-specific tools add, one by one
These tools are not better at seeing. They are built around the submission, which means they carry agency vocabularies, field limits, category lists and delivery routes. They differ sharply in how much of the pipeline they take over, and in what they ask for in return.
Xpiks
Best for contributors who want a desktop metadata editor they control completely. It is open source, writes IPTC and XMP directly into your files, and has strong keyword editing and spell-checking. The limitation is that it is an editor rather than a pipeline: the writing and the review are excellent, but the heavier automation and the multi-agency delivery side are things you assemble around it.
StockSubmitter
Best for breadth of destinations. It is the long-standing answer to the distribution problem and reaches an unusually wide list of agencies, which is exactly why it keeps appearing in contributor recommendations. It is a Windows desktop application with a dense interface and a real learning curve, and its pricing works on a submission-volume model rather than a flat seat fee.
PhotoTag.ai
Best for metadata only, with the least setup. It generates titles, descriptions and keywords from an upload and exports them, and it is a clean fit if you already have a delivery method you like. The trade-off is scope: it stops at the metadata, so uploading and tracking across agencies remain your job.
Wirestock
Best for contributors who would rather not run the pipeline at all. It is a managed service: you send files, it handles keywording and distribution. That genuinely removes work, and for people who dislike the admin it is a fair trade. The structural cost is that a managed model takes a share of what your files earn, and you give up direct control of how each submission is presented.
Rastock AI
Best for contributors who want the metadata and the delivery in one place without handing over a share of royalties. It writes agency-shaped titles, descriptions and ranked keywords, then delivers over FTP or SFTP to more than 10 agencies including Adobe Stock, Shutterstock, Alamy, Depositphotos, Pond5, Canva, 123RF, VectorStock and Vecteezy, with any other FTP-capable agency addable by hand. The control side is the part that distinguishes it from a tagger: banned and mandatory keyword rules, per-agency field limits, and rollback if a batch comes out wrong. It does not claim ownership of your files and IPTC, CSV and XML export stay open, so leaving is cheap. Where it is not the answer: if you only need raw labels inside your own software, an API is simpler and cheaper.
When to use a vision API, and when to use a contributor tool
Use a general-purpose vision API when the labels are an input to something you are building. You are indexing an internal archive, routing uploads, filtering for moderation, or enriching a database, and you have a developer who can call an endpoint and store the response. At that point $1.00 to $1.50 per thousand images is very hard to beat, and the absence of stock-specific formatting is irrelevant because you were never submitting anything.
Use a contributor tool when the output is a submission. The moment your metadata has to satisfy someone else's rules, keyword ceiling, category list and file transfer, you are paying for the rules rather than the recognition. Building that layer yourself on top of a raw API is possible, and a handful of technical contributors have done it, but it means maintaining agency vocabularies and upload plumbing forever.
Use both when you have a genuinely unusual niche. A custom-trained vision model can recognise the things generic detectors miss, such as specific equipment or regional subject matter, and its output can feed a contributor tool that turns those concepts into a compliant submission. This is a small minority of cases, but for medical, industrial or scientific libraries it is sometimes the only accurate route.
The cost comparison that misleads people
Set a vision API next to a subscription tool and the API looks dramatically cheaper. Tagging 5,000 images through label detection lands somewhere around five to eight dollars, which no subscription can match. The comparison is misleading because the two prices buy different amounts of finished work.
The API price covers labels. It does not cover the developer time to call it, the mapping from concepts to agency keywords, the title writing, the category assignment, the per-agency field limits, the FTP delivery or the retry logic when an upload fails at three in the morning. Once those are priced honestly, in hours rather than dollars, the gap closes and often reverses. The way to settle it is arithmetic on your own volume against a tool's published tiers, not on the headline per-thousand rate.
What to check before you commit to either
Run the same twenty files through whatever you are considering and read the output as a moderator would, not as a developer. Are the titles sentences a buyer would recognise? Do the keywords include the concepts as well as the objects? Is anything in there that would get the file rejected? Can you export the result in a format you own, and can you get your metadata out again if you leave?
Twenty files will tell you more than any comparison table, including this one. If you want the wider field rather than this single distinction, we keep a broader survey of tools for generating keywords for stock images, and a technical explainer on how AI image tagging works under the hood.
Frequently asked questions
Can I use Google Cloud Vision to keyword stock photos?
You can use it to generate raw labels, but the output is not a submission. It returns detected objects with confidence scores, not a title, a ranked keyword set, a category or the conceptual terms buyers actually search for. Contributors who go this route end up writing a translation layer between the API response and each agency's requirements, which is the work the API was never doing.
Which is cheaper, a vision API or a keywording tool?
Per image, the API, and by a wide margin. Label detection runs roughly one to one and a half dollars per thousand images at both major cloud providers. That price buys labels only. Once you add the engineering, the agency mapping and the upload handling that a contributor tool includes, the comparison stops being about the per-image rate and becomes a question of how you value your own hours.
Why do generic taggers miss the keywords that sell?
Because they are detectors. They name what is visibly present, so a photo returns person, laptop, table. Stock buyers search for concepts instead: remote work, burnout, freedom, candid. Those are interpretations of a scene rather than objects in it, and a model trained to detect will not produce them unless something else is asking it to think commercially.
Do contributor tools use the same underlying vision models?
Often something similar sits at the bottom of the stack, and that is fine. The difference is everything built on top: agency vocabularies, keyword ceilings and ordering rules, category lists, banned-term checks, release handling and file delivery. The value is in the layer that turns recognition into an accepted submission, not in the recognition itself.
Is a managed service better than doing it myself?
It depends on what you would rather give up. A managed pipeline removes the admin entirely, which is a real benefit if the workflow is what stops you uploading. In exchange it usually takes a share of earnings and control over how each file is presented. Self-directed tools keep both the revenue and the decisions with you, at the cost of running the process.
How do I test a tool before paying for it?
Take twenty files that represent your real work, including two or three awkward ones, and run them through. Read the titles aloud, check whether the keywords include concepts and not just objects, and look for anything that would trigger a rejection. Rastock offers a 14-day free trial with no card required, and most alternatives have a free tier large enough for a test of that size.