# AI images vs stock photos — cost, consistency and the limits > Ten credits an image against a stock library, compared on specificity, catalog consistency and cost, with the commercial limits you should plan around. Published 2026-09-13 · eroq.ai — canonical: https://eroq.ai/blog/ai-images-vs-stock-photos The image version of this question is easier to answer than the video one, because the cost gap is wider and the limits are sharper. Generated images are cheap enough that cost stops deciding anything almost immediately — which is useful, because it moves the decision onto the two things that matter commercially: whether the picture can be specific enough, and whether it is allowed to be the picture. ## Cost per image, and what that changes Every image model on eroq costs the same: **10 credits**, whether you use the photoreal [Image One](/models/eroq-image-one), the anime model, or the painterly one. One price, three registers — the differences are covered in [Image One vs Anime vs Art](/blog/image-one-vs-anime-vs-art). Ten credits is 10 cents at the entry pack rate, about 8 cents on a Creator plan and about 6 on Agency. The 50 credits a new account starts with cover 5 images. At that price, volume changes shape. A forty-product catalog, four variations of each so you can pick: **40 × 4 = 160 renders × 10 credits = 1,600 credits** — $16.00 at the entry pack rate, about $13.07 on a Creator plan. The batch control renders one to four at a time, so a batch of four is 40 credits and arrives as a set you choose from. Rejecting three of four costs 30 credits, which is the point: **at this price, iteration is free enough to be your default method.** You stop composing the perfect prompt and start generating four, reading them, and adjusting. That is the real difference from a licensed library, where each extra option costs the same as the first. ## Specificity — the catalog problem A stock library contains photographs that were already taken. That is an enormous set, and for generic needs it is fine. It fails on the same axis every time: it has a mug on a desk, not *your* mug on *your* desk in your brand's colors at the angle your grid needs. Generation starts from the description instead, which means the constraint moves from "does this shot exist" to "can I describe it precisely". Here is what precision looks like as a prompt: > A matte black travel mug stands on a pale concrete ledge beside a folded wool scarf, morning light raking across from the left and catching the brushed lid, city rooftops soft and out of focus behind. Shot on 35mm, shallow depth of field, cool neutral palette, quiet and unstyled. Add a photograph of the actual mug as a [reference image](/glossary/reference-image) and the render keeps the object rather than inventing a plausible one. That combination — your object, your described scene — is the thing a library structurally cannot do. It is also why concept work leans on generation: a product that does not exist yet has no stock photography by definition, a workflow covered in [concept art](/use-cases/concept-art). ## Consistency across a set Buy twelve stock photos for a category page and you get twelve photographers, twelve lighting setups and twelve color temperatures, then spend an afternoon grading them into something that looks deliberate. Generated sets invert that. The same palette, lighting and lens language can be applied to every render, and references carry the specifics. A recurring face is a Character with reference photos — the method is in [image generation for characters](/blog/image-generation-for-characters). A recurring object or location is an Element. Negative prompts and CFG control let you push a set toward the same register rather than curating it into one. For a catalog that is the whole argument. Two hundred SKUs photographed to one lighting standard is a studio day with a budget. Two hundred renders sharing a rack is an afternoon and 2,000 credits. ## The limits that matter commercially These are the ones that end up in a legal review or a print check, so plan around them before the shoot, not after. - **No real identifiable people.** Generating a real, identifiable person without consent is not allowed. There is no version of "the famous one holding our product" available here, and model-released stock exists exactly because this problem is old. - **Fine print and exact logos are still unreliable.** Small text drifts. A label that must read correctly, a legally required disclosure, a competitor's trademark rendered accurately — do not trust a render for any of them. Composite the real artwork in post or photograph the packaging. - **There is no upscaler.** Image One returns 1024×1024 on the square ratio. At 300 dpi that is about 3.4 inches, which is fine for web, social and small print, and not fine for a poster. Large-format print is a real constraint, not a workflow detail. - **There is no editing inside the studio.** No retouching, compositing or color work — you export and finish in your own tools, as you would a licensed file. - **You cannot train a model on your own catalog.** You steer with prompts, references and negative prompts, which is enough for consistency across a set but not the same as a model that has memorized your product line. - **A render is not evidence.** Anything presented as a record of a real thing — your actual premises, your actual team, a real event — should be photographed. That is an honesty line, not a technical one. Five aspect ratios are available on images: 1:1, 3:4, 4:3, 9:16 and 16:9. There is no ultrawide image ratio, so a 21:9 banner is a crop. ## Where stock photos still win - **Editorial and news.** Real events, real people, real context. Not a generation problem. - **Real places.** A named landmark, a specific skyline, an actual store interior. Generation gives a convincing lookalike, which is worse than useless when the point is that it is that place. - **People with releases.** When you need a human face in a commercial context and cannot generate a real one, a model-released photograph is the clean answer. - **Third-party products in frame.** Anything where another company's branding must appear accurately. - **Large-format print.** Until an upscaler exists, resolution decides this one. ## The mixed library most teams end up with In practice nobody picks a side. A working setup is a licensed library for real places and released people, generated images for products, scenes, backgrounds and concepts, and real photography for the shots that must be literally true. The test on your own work is short. Sort your last twenty images into three piles — had to be real, had to be specific, could have been anything. The third pile is where a library was already fine. The second is where four attempts at 10 credits beats an afternoon of searching. The first is why you do not cancel the stock subscription. Write a prompt with the shape above and render four variations at [the image studio](/studio/image) — 40 credits, and you will know which pile your work is in. Prompt patterns that hold up commercially are in [photoreal AI image prompts](/blog/photoreal-ai-image-prompts), and the per-plan credit rates are on the [pricing page](/pricing). ## FAQ ### How much does one AI image cost? It is 10 credits on every eroq image model — about 10 cents at the entry pack rate and roughly 8 cents on a Creator plan. A batch of four is 40 credits, which is what makes generating options and discarding most of them a reasonable default. ### Can AI images replace stock photos for a product catalog? For product and scene imagery, usually yes, and consistency is the reason more than cost — a shared palette, lighting and lens language applied across every render beats grading twelve photographers into agreement. Keep a library for real places, released people and anything where another brand's artwork must appear correctly. ### Are AI images good enough for print? For web, social and small print, yes. The square output is 1024×1024, which is about 3.4 inches at 300 dpi, and there is no upscaler, so large-format print and posters remain a photography or licensing job. ### What are the legal limits on commercial AI images? The hard rule is no real identifiable people without consent, which rules out likeness-based advertising. Beyond that, treat unreliable fine print and logos as a production risk — composite real artwork rather than trusting a render — and never present a generated image as a record of something that actually happened.