AI Image Generation: One Year of Production, Unfiltered

The StartingUp Summer continues its inventory of a year of experiments, with the same ground rule as for the questions AI made obsolete: telling you what we actually did, costs and limits included. Today, AI image generation, which we've been running in production for months.

Might as well say it upfront: this blog's editorial illustrations, including the one on this very article, are AI-generated. For several months now, every publication has received its visual through Ideogram's API, with a house style enforced by the prompt: premium flat illustration, limited palette, no text in the image. This isn't a test or a demo, it's our actual production pipeline.

This debrief tells you what we learned along the way: what the three main families of tools we tested are actually worth, what the law really says about these images, what visible and invisible watermarks hide, and above all when to generate... and when not to.

Summary for decision-makers in a hurry: AI image generation is mature enough for professional use in production, for a few cents per visual, provided you pick the tool for the job: graphic styles with Ideogram, versatility with OpenAI, photorealism with Google. On the other hand, a purely generated image isn't protected by any copyright, and you should treat it as detectable as such.

Our need is modest and tightly scoped: one conceptual illustration per article, in a consistent graphic style, produced in a few minutes at publication time. Before AI, this need was solved with a stock photo library (result: interchangeable visuals everyone has already seen) or not solved at all. Generation changed the equation: for a few cents per image, we get a unique visual, consistent with the previous ones, that illustrates the article's actual idea rather than a vague set of keywords.

The real work isn't in clicking generate, it's in the style prompt, written once and reused everywhere: locked-down art direction, limited palette, no text allowed in the image (models write better than they used to, but a slightly-off headline ruins a visual), and a visual metaphor decided by a human for each article. The creative part stays with us; the machine executes.

This is our production choice, for down-to-earth reasons: a simple API, readable pricing (3 to 10 cents of a dollar per image depending on quality tier, as we write this), and above all a genuine feel for graphic styles, flat design, poster, pictogram, where other models pull everything toward photorealism. Ironic, but true: Ideogram's historical specialty is rendering text inside images, the best on the market, and our prompt forbids it from doing that. We use it for its stylistic consistency, not its typography.

The gpt-image lineage (DALL·E's heir) shines through its versatility and its integration in ChatGPT: you describe, you look, you correct in plain language, in the same conversation. It's the tool we recommend to marketing teams just starting out, precisely because it requires no infrastructure. On the API side, pricing ranges from under a cent to about twenty cents per image depending on model, quality and resolution. Our reservation: a visual signature recognizable from a mile away, which has been homogenizing the web's visuals since 2025.

The Gemini and Imagen family aims for photorealism, and hits it: for simulating a believable scene, natural light, a texture, it's the current benchmark. API pricing ranges from about 2 cents per image to roughly twenty cents for high-end, high-resolution models, and the lineup moves fast: models get regularly deprecated in favor of newer ones, which means keeping an eye on your dependency. For our editorial use, that realism is exactly what we avoid: a fake photo that's too convincing raises more questions than it answers.

This is the first question our clients ask about AI-generated images, and the honest answer has two layers. First layer, copyright: the US Copyright Office set its doctrine in a January 2025 report, and it's unambiguous.

A work entirely generated by AI is not eligible for copyright protection: merely writing prompts, however detailed, does not constitute sufficient human contribution. Only identifiable human contributions (selection, arrangement, substantial editing) may be protected, on a case-by-case basis.

"Copyright and Artificial Intelligence, Part 2: Copyrightability", US Copyright Office, January 2025 (summarized and translated by us)

European law hasn't ruled quite as clearly yet, but it rests on the same requirement of a human creative imprint. The practical consequence: a purely generated image is probably in the public domain, in effect. You can use it, but you can't stop a competitor from reusing it. For an article illustration, that's an acceptable risk. For a logo or a mascot, it's a dealbreaker.

Second layer, the platforms' terms of service, which settle not intellectual property but the contract between you and the tool. The major platforms (Ideogram, OpenAI, Google) assign you whatever rights they hold on the images produced, commercial use included, but with nuances that matter: on Ideogram, the free tier excludes commercial use and publishes your images in a public gallery by default; on OpenAI, commercial use is open at every tier, but opting out of training on your content depends on the plan. Read the line that concerns your subscription, not the marketing page.

And two red zones we impose on ourselves: never a "in the style of" prompt naming a living artist (legally uncertain, humanly indefensible), and never a protected brand, logo or character in an image meant for publication. Trademark law, meanwhile, applies fully to generated images.

Two families of marking need distinguishing. The visible watermark first, the one on free tiers: a logo in a corner of the image, like the Gemini star on images from Google's free plans. That's the price of free, and an unprofessional-looking signal on a company site: if a generated visual deserves to be published, it deserves the paid tier that removes it.

The invisible marking next, and this is the more important one: SynthID, Google DeepMind's technology, embeds an imperceptible signature directly into the pixels, designed to survive cropping and compression. Google applies it to all of its generated images, paid tiers included. C2PA metadata, meanwhile, attaches a signed manifest to the file (model used, date, producer); more fragile, it often disappears when a platform re-encodes the image. In spring 2026, OpenAI joined C2PA and adopted SynthID for all of its images: the three families we use now all mark their output.

The implication is simple: assume every generated image is detectable as such, today or in two years. The EU's AI regulation points the same way: its transparency obligations on generated content apply from August 2026. Our position hasn't changed since the first visual: we never pass off a generated image as a photograph, and owning it publicly (as in this article) costs far less than being caught out.

After a year of practice, here's the framework we actually apply, for ourselves as much as for our clients:

• Generate: editorial illustrations, abstract concepts, blog and social media visuals, mockup iterations. High volume, replicable style, no need for legal exclusivity: this is generation's home turf.
• Buy a photo (or organize a shoot): as soon as it's about embodied reality. Your team, your premises, your products, a real place. A convincing fake photo is a debt of trust: the day it's spotted, your entire communication becomes suspect.
• Pay an illustrator: for anything that must belong to you and last. Logo, mascot, visual identity, a homepage's signature illustration. You're buying a point of view, and a transfer of rights that lets you defend the work: exactly what generation can't sell you.

Image generation hasn't replaced photography or illustration for us: it filled the gap that used to stay empty, the visuals nobody would have budgeted for. That's perhaps the most honest conclusion of this debrief: used well, this technology doesn't lower the visual bar, it extends it to content that never had a claim on it.