Selling AI Images on Adobe Stock with a Semi-Automated Workflow / A Conversation with Claude

Claude: Yes, you can. But when you upload AI-generated content, you must select the ‘Created using generative AI tools’ checkbox. Your account could be suspended if you don't. And you shouldn't include real people's names, other artists' names, or famous characters in your prompts, titles, or keywords.

Claude: When an image sells, the contributor gets about 33% of the sale price. For video, it's around 35%. But the actual amount varies quite a bit depending on whether the buyer downloads it through a subscription plan or makes an individual purchase. With subscription downloads, you might get just a few dozen cents per image. You need to accumulate at least $25 before you can receive a payout, and you get paid through PayPal or a bank transfer.

Claude: No, there's more. Images shouldn't contain text, signatures, or watermarks, and they'll fail review if they have typical AI errors like strange fingers or distorted faces. A face that looks like a real person can be rejected over model release issues even if it isn't an actual person. You also need to be careful with brand logos, famous buildings, and landmarks.

Claude: The most common issue is submitting lots of similar images. If you generate dozens of similar images with AI and upload them all at once, Adobe sees that as spam and rejects them in bulk. With images that differ only slightly, you can think of it as one surviving while the rest get rejected. The second is the detail errors typical of AI: the wrong number of fingers or background objects that look warped when you zoom in. The third is keyword spam. Stuffing in lots of unrelated keywords lowers your quality score.

Claude: That's right. Illustrations and vectors often don't have the obvious giveaways that photos do, so it can be hard to tell conclusively. But the important thing is that Adobe makes disclosure a contractual obligation whenever you create something with AI, regardless of whether it can be detected. It's not just a matter of getting caught: failing to disclose is a breach of contract, and if it's discovered, you could lose everything you've uploaded along with your account. Adobe is also fairly good at detecting it through metadata and generation patterns. And labeling it doesn't reduce your royalty rate, so checking the box honestly gives you peace of mind.

Claude: That's a reasonable concern. It depends on the buyer. Some deliberately avoid AI images because of brand safety or copyright concerns. On the other hand, plenty of buyers don't care whether it's AI as long as the image is affordable and useful. For blog posts, advertising assets, and thumbnails in particular, they hardly pay attention to it. Since disclosure is mandatory anyway, there's no way to hide it, so the practical approach is to compete on image quality and search keywords. If an image ranks high in search and catches the eye, it can sell even with an AI label.

Claude: AI's strength is creating things that are difficult or expensive to photograph: space, futuristic cities, or abstract business concepts, for example. Companies are always looking for images that visualize abstract ideas like ‘success,’ ‘collaboration,’ and ‘data security’ for presentations and websites. Ordinary landscape or food images, on the other hand, already have millions of real photos to compete with, so it's hard to stand out with AI. Starting with a field you're interested in or know well is also a good idea.

Claude: Yes, things that are unreal, or scenes that would be difficult and expensive to stage in real life. Floating islands, glowing networks shaped like a brain, or futuristic cities made of glass, for example. But fantasy images that are simply pretty don't sell as well as you might expect. Buyers think, ‘Where could I use this?’ The best concepts are unreal but still useful for business, technology, or advertising.

Claude: Photos and non-vector illustrations are uploaded as JPEGs and need to be at least 4 megapixels. Vectors are accepted in three formats: AI, EPS, and SVG. For icons, SVG is actually a commonly used, recommended format. PNGs can be submitted under a separate transparent-background PNG category. Videos are uploaded as MOV or MP4 files, and Adobe also accepts 3D assets, design templates, and motion graphics templates.

Claude: They must have a transparent background. You don't upload an ordinary image with a background as a PNG; it needs to be a standalone element with a transparent background that can be placed on top of another design. A cutout light bulb, icon, or symbol, for example. Crop away as much surrounding space as possible so only the object remains. If it contains a recognizable person, you need a model release, and if it includes private property or a trademark, you need a property release.

Claude: Most image-generating AI tools create images that fill the whole canvas, including the background, by default. So the usual process has two steps: generate the image with AI, then remove the background. Background removal is easy these days: one click on Photoshop's Remove Background button does it, and tools like Firefly have started supporting transparent backgrounds from the start. There are also plenty of free web tools. After removing the background, save it as a PNG that preserves transparency. It's less a case of AI being unable to do it and more a case of one extra step.

Claude: Problems around the edges are the most common reason for rejection. In particular, cutting a light-colored object out of a white background can leave a pale halo around the boundary, so make sure to zoom in and check it. AI and background removal tools often make mistakes with semitransparent areas like glass cups or shadows. It's better to remove a shadow cleanly than to leave it poorly cut out. Single objects with clear boundaries, like light bulbs or icons, usually work well, but things with soft edges, like fur or smoke, can get messy. Unnecessary transparent space around the object also counts against it, so crop it tightly.

Claude: Vectors are good for icons, illustrations, and logo-style designs, but most AI tools produce bitmaps, so you need to convert them to vectors in Illustrator. Clean lines and simple shapes convert well, but complex pictures can turn messy. Video tends to earn much more than photos, but AI video is still mostly short clips of a few seconds, and it takes a lot of work.

Claude: If you're just starting out, JPEG is the safest option. You can generate an image with AI and upload it right away, so there's hardly any extra work. It's good for testing the waters and getting a feel for what sells. Once you've figured that out, you can expand into transparent PNGs or vectors, which take more work but have less competition. Trying to tackle difficult formats from the beginning can wear you out and make you give up.

Claude: It does. There's a Firefly-powered ‘Text to Vector Graphic’ feature that creates editable vectors directly from a prompt. That means you don't need the tedious step of converting a bitmap into a vector. Adobe guarantees that Firefly was trained on data permitted for commercial use, which is an advantage for stock submissions. And because generation and submission are in the same ecosystem, the workflow is smooth too.

Claude: In terms of ownership, it's yours, JB. You have the right to use it commercially and sell it, which is why you can sell it on Adobe Stock. But because the production method is AI, you still need to mark it as AI-generated when submitting. The ownership is yours; the production method is AI. Those are two separate things.

Claude: There's no official policy like that. Reviewers use the same quality standards regardless of which tool you used. But there's an indirect advantage: the commercial license is clear, so you don't have to worry about rejection over rights issues. It's less about getting extra points and more about avoiding a risk of deductions.

Claude: Probably not. Contributors using other tools would push back, too. Still, the copyright side is safe and you can go from creation to submission in one workflow, so if you're already subscribed, there's no reason not to use it.

Claude: Claude can't generate images like photographs or complex illustrations. It can use code in Artifacts to create things like SVG shapes, diagrams, simple icons, geometric logos, and patterns. An advantage is that these are vectors from the start, so you can export them directly as SVG. ChatGPT has good image quality and understands prompts well, but commercial resale rights and the copyright status of its training data aren't guaranteed as clearly as with Firefly. Copilot also uses an OpenAI-based engine, so its quality is decent, but you face the same questions when reselling on stock sites. Firefly may be less flashy, but its commercial license is clear, it can generate vectors, and it integrates smoothly with Illustrator. For stock sales, I'd recommend Firefly.

Claude: Full automation is difficult. Firefly has an API, so it's technically possible to feed in prompts and generate images in bulk, but that requires development work and costs money. The real bottleneck is the Adobe Stock submission process. The AI checkbox, keywords, and titles are set up for human review, and bulk automated uploads can easily get flagged as spam. Review is done by people too, so pouring in low-quality content only piles up rejections. The realistic approach is semi-automation: automate generation, and have a person select and refine the results.

Claude: That's the most efficient setup. Generate images with the Firefly API, have a language model API look at them and create titles and tags, then organize the results into a CSV. You can tie all of that together in a single Python script. Adobe Stock officially supports matching titles and keywords in bulk through a CSV, so your idea fits the system perfectly. When asking AI for keywords, tell it to generate ‘keywords that work well in Adobe Stock search’ so you don't get irrelevant or overly generic tags.

Claude: Yes, it's billed separately from the generative credits included in your subscription. Subscription credits are for creating things directly on the website or in the apps. Firefly Standard is $9.99 a month with 2,000 credits, Pro is around $19.99 a month, and ordinary image generation typically uses one credit. The API is billed separately per operation, at roughly 2 to 10 cents per image. If you generate thousands of images a month, the costs can add up.

Claude: Yes, that's one of the API's key advantages. The response returns an image file or a download link, so your script can automatically download it and save it to a specified folder. Prompt submission, image generation, automatic saving, title and tag generation through a language model, and CSV recording can all flow through a single script.

Claude: The structure has four steps. First, prepare descriptions of the images you want to create in a text file or a list. Second, run the Firefly API in a loop to generate an image for each prompt and save it to a folder. Third, send the saved images to a language model API for titles and keywords, along with instructions like ‘for Adobe Stock, no more than thirty keywords that work well in search.’ Fourth, record the filename, title, and keywords in the CSV, one row at a time. Adobe Stock matches the metadata by filename, so getting the filenames exactly right is important.

Claude: Unlike a regular subscription, you set it up in Adobe Developer Console or Admin Console, usually through the enterprise-oriented ‘Firefly Services.’ You register a payment method and pay based on usage, but because the service was originally aimed at companies and development teams, you may need to contact sales or enter into a contract. There can be a barrier to simply signing up with a credit card and using it as an individual.

Claude: For now, it's a good idea to set API automation aside and use your subscription credits to manually create and upload a few dozen images. That will give you a feel for which subjects sell and how review works. You can invest in the API once you think it's worth pursuing.

Claude: Yes, with computer use, it can look at the screen and control the mouse and keyboard to operate a browser. In theory, it could upload files and enter titles and keywords on the Adobe Stock upload page, so there's potential to automate uploading as well. But it's still slow, sometimes clicks in the wrong place, needs a person to step in when it gets stuck, and the site may detect and block automated interactions.

Claude: Yes. It keeps taking screenshots to check the screen and looks again after each click, so token consumption is much higher than in a text conversation. The time you save by automating uploads could end up costing you in tokens. That's why automating only generation and CSV creation, then uploading manually, can actually be cheaper and more reliable.

Claude: Exactly. New contributors start with a low daily upload limit, and it gradually increases as they pass reviews and build trust. It's not a setup where you can pour in thousands of images from the start, so manually uploading that volume isn't too burdensome. The quantity limit can actually help you focus on the quality of each image, which can be a benefit early on.

Claude: Yes, hundreds of thousands are active worldwide, and there are hundreds of millions of registered images. You need to appear near the top of search results to sell. Because AI images have a low barrier to entry, the number of people uploading similar work has grown sharply, so common subjects can easily get buried. It's not a system where money comes in as soon as you upload. It's a long game, and you may earn almost nothing for the first few months.

Claude: The differences in income are extreme. A portfolio of 100–500 images might earn around $30–$200 a month, while 500–2,000 images might earn around $100–$800 a month, so for most people it's closer to a side income. On the other hand, top contributors who have built up thousands or tens of thousands of images over several years earn thousands of dollars a month, and those at the very top earn much more. The more your portfolio grows and the longer you leave it up, the more your income builds like compound interest, so it's not something to judge after a month or two.

Claude: That's the part that changes the game. In the past, building tens of thousands of images was a barrier because you had to shoot and edit them yourself. If you generate them with AI and automate the metadata too, you can build a portfolio much faster. But there are daily upload limits and rejection policies for similar images, so the key is ‘consistently and with variety’ rather than ‘a lot, quickly.’ As everyone starts using AI, the deciding factor is shifting toward who chooses subjects and keywords more thoughtfully.

Claude: Sounds good. Use your Firefly subscription credits to create a few images with unreal or niche subjects, remember to check the AI-generated box, and upload them manually. Get a feel for how review works, then add automation gradually. Let's pick up the script work next time.