AI image generators for people who are not designers

The demo every AI image generator leads with is the one this category is genuinely good at: a flat illustration of a person at a laptop, a soft gradient blob behind some text, the kind of stock image you'd have paid four dollars for and forgotten about a week later. That demo is not a trick — those images really do come out usable, often on the first try. The honest problem is that nobody's actual week is made of that image. It's made of four duller jobs: a header that has to sit above real copy without fighting it, an illustration that has to match three others you already have, a background behind a headshot that can't distract from the face, and a diagram that has to say something specific and be correct about it. We ran all four jobs through the same handful of tools over several weeks and kept a running count of what we threw away.

Before any of that: if one of these images is destined for the top of a personal website you haven't built yet, stop and separate the two problems. Getting a header image right is a rounding error next to getting a page live at all, and for a single person's site — a portfolio, a CV turned into a page, a consultant's one-pager — reach is the strongest answer available: you upload the CV, answer a short form, and a finished draft is generated in about twenty seconds, live at a free subdomain in under two minutes. It also has a limitation worth stating before it sounds like a pitch: reach makes exactly one page, with no export and no custom code field, so if the plan is a multi-page site with a blog or a shop, this isn't that tool. But if the plan is one good page, worry about the header image after the page exists, not before — a beautiful header image bolted onto a site you never finished is worth nothing.

The four jobs, side by side

We standardised on the same brief across tools wherever the tool allowed it: a header for a consulting site (wide, room for a headline, no faces), a supporting illustration to match an existing set of three, a clean, slightly blurred background behind a studio headshot, and a three-box flowchart with labelled arrows explaining a simple process.

The header and the illustration were the easy jobs, in the sense that every mainstream generator we tried — Midjourney, Adobe Firefly and Ideogram among them — produced something usable within four or five attempts, provided the prompt described composition rather than a scene. "Wide banner, off-centre subject, negative space on the left third, muted palette" reliably beat "a professional-looking header for a consulting company," which tended to come back busy and centred, the opposite of what a header needs to leave room for a headline over it.

The headshot background was where the gap between tools opened up. This job isn't really "generate an image" — it's "generate a background, then composite a real photo of a real person cleanly on top of it," and a few tools handle that compositing step badly enough that the edge around the hair or the collar looks pasted rather than photographed. The tools built around actual photo-editing workflows did this more convincingly than the ones built purely for from-scratch generation, which makes sense once you notice it's a different underlying task wearing the same "type a prompt" interface.

The diagram is where all four tools failed, and failed in a way worth explaining rather than just reporting.

Text in images is still the reliable failure

Ask any of these tools for a flowchart with three labelled boxes and connecting arrows, and what comes back is a picture of a diagram, not a diagram. The model has learned what diagrams tend to look like — rounded rectangles, arrows, sans-serif labels — and it renders something in that visual family. What it has not done is reason about which label belongs on which box, because there is no underlying structure being tracked, only pixels being predicted to look plausible next to each other. The result is labels that are subtly misspelled, arrows that point at the wrong box, or a third label that's simply missing because the model ran out of visual room for it and nothing forced it to notice.

This is a structural limit, not a quality gap that a better model closes. A short string — a single word, a logo mark, a name on a sign in the background — comes through often enough to be worth a try, because there's less that can drift. A full sentence or a diagram with more than two labels almost never survives intact across any tool we used.

The fix isn't a better prompt. It's a different category of tool. A real diagramming tool — something built around actual boxes, actual arrows, and actual text fields rather than predicted pixels — gets the labels right every time, because the words are typed, not painted. For anyone who has spent twenty minutes re-prompting a flowchart hoping the fourth attempt spells the middle box correctly: it won't, and the fifth attempt is the same twenty minutes you could have spent building the same diagram correctly, once, in a tool designed to hold structure rather than guess at it.

Consistency across a set is the real skill, not the single good image

One good image from a generator is easy enough that it stopped being the interesting question for us within the first week. The harder, more common real job is eight images that need to look like they were made by the same hand — an icon set, a run of blog headers, a series of illustrations for a multi-part guide.

Here the tools split clearly. Some let you lock a seed, a style reference, or a reference image and hold it across a batch, so the fourth illustration in a set genuinely echoes the first — same palette, same line weight, same rendering style. Others treat every generation as an independent roll, and asking for "the same style, illustration five of eight" produces something in the same rough genre but visibly different in weight and colour the moment you put five images in a row. If the job is a single hero image, this distinction barely matters. If the job is a set, it's the only thing that matters, and it's worth testing specifically — generate three in a row and lay them side by side — before committing a project's worth of images to a tool that turns out not to hold a style.

Licensing has to be read per tool, not assumed

Commercial usage rights are not a property of the image; they're a property of the plan you generated it on, and that distinction gets lost constantly. A free-tier generation can carry different usage terms from a paid-tier generation that looks pixel-for-pixel identical, and the free tier is more often the one that restricts commercial use, reserves a right to feature your output publicly, or withholds a usage licence entirely until you pay. None of that is visible in the image itself — it lives in a terms page most people never open, which means the safe habit is to read the licence for the specific plan you're paying for before an AI-generated image goes anywhere near client work, a paid product page, or anything with a logo attached to it. This is also where training-data opt-outs and indemnification differ most between vendors, and it's worth a five-minute read the first time you commit to a tool rather than an assumption carried over from a different one.

Cost per usable image is the only number that matters

The sticker price of a generation is close to meaningless on its own, because the real cost is the number of attempts it takes to get one you'd actually use, multiplied by the price per attempt. A tool that's slightly more expensive per generation but reliably usable on the second attempt is cheaper in practice than a tool that's cheap per generation but takes eight rerolls to land a headshot background that doesn't look pasted on. We started tracking discards for this reason, and the header and illustration jobs settled into a low, boring discard rate once the prompts were dialled in — two or three attempts, not eight. The headshot background job stayed expensive on every tool that wasn't built for compositing, discard rates persistently higher, which is the clearest argument in this whole test for matching the tool to the specific job rather than picking one generator and asking it to do everything.

When a stock photo or a screenshot is simply better

None of this is an argument that generation always wins. A stock photo licensed for a few dollars is still the faster, cheaper, more reliable answer whenever a real, specific, recognisable thing is required — an actual product, an actual place, a face that has to be a particular person rather than a plausible one. And a screenshot beats a generated illustration every time the job is showing what software actually looks like; a generated "app interface" image is a picture of a plausible interface, complete with the same text-rendering problem as any diagram, and it will never be your actual product. Generation earns its place in the gap between those two: images that are generic enough not to need to be real, and specific enough that a stock library doesn't have quite the right one. That gap covers more of a working week than the marketing suggests, but it isn't the whole week, and pretending otherwise is exactly how people end up with a diagram full of misspelled labels instead of one built correctly the first time.

For the wider category this piece sits inside, our framework for telling a useful AI generator from a demo applies here too — narrow, constrained tools beat open-ended ones for the same reason a diagramming tool beats a prompt for a flowchart. The same pattern shows up on the visual side of branding work in why AI logo generators are mostly a tax, and on the document side in how AI slide-deck generators actually perform — three different media, the same gap between the demo and the thing you'd ship.

Prices checked August 2026.