What you’ll be able to do

By the end of this guide you’ll know how to describe a picture in your head clearly enough that an AI actually produces something close to it. Not a perfect mind-read, that doesn’t exist, but a proper attempt you can refine. You’ll be able to generate an illustration, tweak it, export it at the right size, and know when to stop tinkering and just use the thing.

That’s it, really. No design degree needed.

Before you start

Right, a few honesty checks before you open any tool.

AI image generators are pattern machines. They’ve learned from huge amounts of existing imagery, and when you type a description, they predict what pixels should go where based on that learning. They’re genuinely good at style, mood, colour, composition, lighting, all the stuff that used to take years to learn. They’re less good at exact text in images (though some tools, like Ideogram, are built specifically to handle text better than most). They’re also inconsistent with hands, and with keeping the same character looking identical across multiple images, though this keeps improving.

On rights: check the terms of service for whichever tool you use, because commercial use rules differ between them and between plans. Don’t type a living artist’s name into a prompt hoping to copy their style, that’s a genuine ethical problem, not just a technical shortcut. And don’t try to generate images of real, identifiable people, most mainstream tools restrict this anyway, and it’s not a grey area worth wandering into.

The other thing worth knowing early: how you describe something matters more than which button you press. Vague prompts get vague results. Specific ones don’t guarantee brilliance, but they get you much closer, much faster.

Get set up

You don’t need five tools. You need one that fits what you’re making.

If you want something conversational, where you can describe an idea in normal sentences and refine it by chatting back and forth, ChatGPT’s image generation is a sensible starting point. You’re likely already used to typing into it, so there’s no new interface to learn.

Google Gemini also generates images and sits inside a chat interface, similar idea, and worth trying if you’re already using Gemini for other things.

Midjourney is popular for illustration specifically, it has a strong reputation for stylised, painterly results, and while it started life inside Discord (and still works there), there is now a web app too, which is the easier place for a beginner to start.

Adobe Firefly is worth a look if you already use Adobe products, it’s built with commercial use and licensing clarity in mind, which matters if you’re making something for work rather than for fun.

Ideogram is the one to reach for if your illustration needs actual readable text in it, a poster, a logo mockup, a sign in a scene. Most tools struggle with spelling out words correctly; Ideogram was built to handle this better.

And if you want full control, running things on your own machine, adjusting endlessly, no cloud dependency, Stable Diffusion is the option, though it asks more of you technically. Not a first stop for a complete beginner, but worth knowing it exists.

Pick one. Make an account. Don’t overthink the choice, you can always try another later.

Try it yourself

Here’s the bit that actually matters: describing what you want.

A weak prompt says “a cat in a garden.” A strong prompt says what the cat looks like, what style you want, how it’s lit, what mood you’re after. Think of five things: subject, style, composition, lighting, mood. You don’t need all five every time, but the more you include, the less guessing the AI has to do.

Try this as a first attempt:

A ginger cat sitting on a stone wall in a cottage garden, illustrated in a soft watercolour style, warm afternoon light, gentle and cosy mood, loose brushwork, pastel colour palette

See how that’s doing more work than “a cat in a garden”? Subject (ginger cat, stone wall, cottage garden), style (watercolour, loose brushwork), lighting (warm afternoon), mood (cosy). That’s the formula.

Once you’ve got a result you like but not quite, iterate. Don’t start again from scratch, build on what worked:

Keep everything the same but make it evening instead of afternoon, with warm lamplight glowing from a nearby window

That phrase, “keep everything but change X,” is one of the most useful things you can type. It tells the tool to preserve what’s working and adjust one variable, rather than rolling the dice on a whole new image.

If you’re making something with a recurring character or object, say a mascot for a small project, keep a reference image handy and upload it when the tool allows, then describe the new scene around it. Not every tool supports this the same way, so check what your chosen one offers, but where it exists, it’s the difference between a coherent set of illustrations and a pile of unrelated ones.

Here’s a third prompt, this time for something with actual text in it, which is where Ideogram tends to do better than most:

A hand-lettered illustrated poster that says "Garden Party" in bold whimsical script, surrounded by illustrated flowers and vines, warm pastel colours, flat vector illustration style

Notice the quote marks around the actual text you want rendered. Being explicit about what words should appear, and where roughly they sit in the composition, helps.

Check the result

Right, you’ve got an image. Now don’t just accept it, test it.

First, the obvious one: does it match what you asked for? Read your prompt back and tick off each part. Subject, style, lighting, mood, all present? If you asked for watercolour and got something that looks like a photo, that’s a sign to be more explicit about medium next time, actually say “watercolour painting” or “digital illustration” rather than assuming the AI will infer it.

Second, look closely at detail. Zoom in on hands, faces, any text. This is where generators still slip up. A sixth finger, warped text, an object that doesn’t quite make physical sense. These flaws are common enough that checking for them should be automatic, not optional.

Third, check the size and format actually work for where you’re putting it. If you need a square image for social media and you’ve got a landscape one, either regenerate with the right dimensions specified in your prompt, or export and crop, most tools give you an option to set aspect ratio before you even generate.

Fourth, and this one’s easy to skip: does it look finished, or does it look like a rough draft? There’s a difference between “technically matches the prompt” and “actually good enough to share.” If it’s the former but not the latter, that’s not failure, that’s just where iteration comes in.

Last check: read the tool’s terms again if this is for anything beyond personal use. Confirm you’re within what’s allowed for commercial work before you build a whole campaign around one image.

If it doesn’t work

Sometimes the result is just wrong, or weird, or nothing like what you pictured. A few common causes.

Too vague. If your prompt was three words, that’s most of the problem. Go back and add style, lighting, mood, whatever’s missing.

Conflicting instructions. Asking for “photorealistic cartoon style” confuses the model because those two things pull in different directions. Pick a lane.

Overloaded prompt. Cramming ten ideas into one sentence often means the AI drops some of them. Simplify, then add detail back in stages.

On privacy: don’t upload images containing other people’s faces without their consent, don’t hand over anything confidential (unreleased product designs, internal documents, personal identifying information) as a reference image, and don’t assume anything you upload disappears. Read the tool’s data policy, particularly around whether your prompts or uploads are used for training. If you wouldn’t want it public, don’t put it in.

Also worth saying plainly: don’t try to generate images of real, named people, and don’t type a living artist’s name to mimic their style. Beyond the ethical issue, most mainstream tools now restrict or block this anyway, so you’ll likely just get a refusal or a filtered result.

If none of that’s the issue and it’s still not working, try a different tool. They’re trained differently, and one that struggles with your idea might handle it fine elsewhere.

Keep exploring

Once the basics feel comfortable, the natural next step is building a small set of consistent images, same character, same palette, across several scenes, which is where reference images and careful prompt iteration really start paying off.

If you’re curious about turning static images into something interactive, our guide on how to get AI to build Three.js assets is a reasonable next stop. And if the illustration you’ve made is heading somewhere specific, like a personal project or portfolio, our guide on how to get AI to build me a website walks through getting that built.

Sources and review notes

Review date: 15 September 2026 — every source above was opened and checked against the text on that date.