AI Image Generation for Beginners: How to Write Prompts That Work

A good AI image prompt names six things: the subject, what it is doing, where it is, the lighting, the style or medium, and the framing. Most disappointing results come from prompts that supply two of those six and leave the model to invent the rest. “A cat” leaves everything undecided, so you get an average of every cat picture ever made. “A ginger cat asleep on a sunlit windowsill, soft morning light, shot on 50mm film, close crop” leaves almost nothing undecided, and the output changes accordingly.

That is the entire skill, and it transfers between tools. The specific generators change every few months; the discipline of describing an image completely does not. This guide covers the structure, the vocabulary that genuinely shifts output, the things these models still handle badly, and how to iterate on an image that is nearly right.

What are the six parts of an image prompt?

Order matters less than completeness, but putting the subject first helps, because most generators weight earlier words more heavily. A prompt that opens with four clauses of atmosphere before naming the subject often produces beautiful images of the wrong thing.

Component five is the one beginners skip and professionals never do. “Style or medium” is where you decide whether the result looks like a photograph, an oil painting, a pencil sketch, a 1970s travel poster or a technical diagram. Leave it out and the model picks a generic glossy digital-illustration look — the house style of AI images that everyone has learned to recognise and discount.

Which words actually change the output?

A great deal of prompt folklore circulates, much of it useless. Appending “masterpiece, 8k, highly detailed, award-winning, trending on ArtStation” to everything does very little on current models; these phrases were partly effective on earlier generations and have largely been trained out. What does work is concrete, physical description — words that would mean something to a photographer, painter or illustrator.

The test for whether a word belongs in your prompt is simple: could a person follow it? “Backlit by a window at 5pm” is an instruction. “Stunning” is an opinion, and the model cannot act on an opinion except by reaching for its default glossy look.

What do these models still get wrong?

Knowing the failure modes saves hours of re-rolling the same prompt hoping for a different outcome. Some things are hard because of how the models work, not because your prompt was bad.

That last row surprises people. Writing “a man, no hat” can produce a man in a hat, because the words in the prompt pull the image towards everything they mention, including the thing you negated. Most serious tools provide a separate negative-prompt field for exactly this reason; if yours does not, describe the positive instead — “a man with bare, windswept hair”.

How do you fix an image that is nearly right?

The instinct is to rewrite the whole prompt. That throws away what worked. Change one variable at a time, in a fixed order, and you converge far faster.

Keep a file of prompts that worked. This is the single highest-return habit in the whole practice, and almost nobody does it. After thirty generations you will have five or six prompt skeletons that reliably produce the look you want, and the work becomes substitution rather than invention.

Can you use AI images commercially?

It depends on the tool’s terms and on where you are, and this is an area where the honest answer is that it is unsettled. Most major generators grant commercial use rights to paying users for images they create, but the terms differ, free tiers often differ from paid ones, and they change. Read the current terms of the specific tool rather than relying on what was true last year.

Separately from the tool’s permission, three practical cautions apply regardless of what the terms say.

• Do not generate recognisable brands, logos or trademarked characters and put them in commercial material. The tool’s licence does not override someone else’s trademark.
• Do not generate identifiable real people for commercial use. Personality and publicity rights are separate from copyright and apply in many jurisdictions including India.
• Copyright in the output is uncertain. Several jurisdictions, notably the United States, have held that works without human authorship are not copyrightable. If an image needs to be a protectable asset, do not rely on a generated one.

None of this is legal advice, and if an image carries real commercial weight it is worth a professional opinion rather than a blog’s. For low-stakes use — a blog illustration, a social post, an internal deck — the practical risk is low and the terms of the major tools generally cover you.

A starter exercise

Pick one image you need. Write the prompt in six labelled lines — subject, action, setting, lighting, style, framing — then join them into a sentence. Generate four variations. Change only the lighting line and generate four more. Compare the two sets side by side.

That one comparison teaches more than reading twenty prompt guides, because it shows you how much of an image’s character comes from a single clause. Lighting is usually the biggest lever, which is exactly why it is worth isolating first.

From there the practice is cumulative: build your prompt library, learn which failure modes are worth fighting and which to design around, and treat the generator as a collaborator you have to brief properly rather than a wish-granting machine. The same principle underlies text prompting too, which we cover in how to write better AI prompts. If you are weighing which assistant to use for creative work, see ChatGPT vs Gemini vs Claude, and for the wider picture of publishing AI-made material, our guide to AI-generated content. Students will find more in the best free AI tools for students. More in the technology section.