AI Image Prompts in 2026: Write a Brief, Not a Keyword Pile

I generate illustrations for posts like this one, and the prompts that work today look nothing like the ones from three years ago. The comma-separated pile of "8K, highly detailed, trending on artstation" was a Stable Diffusion 1.5 habit. The current models read a paragraph and want to be told what the picture is for.

This post is about what actually changes the output on the models people use in 2026: GPT Image in ChatGPT, Google's Nano Banana models in Gemini, Midjourney v7, and the Flux family for anyone running their own. Most of it transfers between them. Where it does not, I say so.

Write the brief you would give a photographer

A prompt that works has four parts, in roughly this order: subject, setting, look, and purpose.

Subject. Not "a cat" but "a ginger tabby, older, slightly overweight, sitting upright". Specific nouns and one or two adjectives that a person could check.

Setting. Where it is and what is happening. "On a windowsill, rain on the glass, a street lamp outside just coming on."

Look. Medium, light and lens in plain words. "Photograph, eye level, warm light from inside the room, background soft." If you want an illustration, say which kind: "flat vector illustration, two colours, no outlines".

Purpose. This is the part people skip and the part the 2026 models use most. "For a blog header, 16:9, leave the right third empty for a title." Tell it the aspect ratio in the prompt and in the tool's setting when there is one; the two do not always agree.

Written out: "A photograph of an older ginger tabby sitting upright on a windowsill, rain on the glass behind it and a street lamp just coming on outside. Eye level, warm light from inside the room, background soft. For a blog header, 16:9, right third empty for a title." Forty-odd words, one sentence per part. The AI Image Prompt Builder walks through the same four parts and assembles the paragraph, which is useful when you are writing a dozen of them for a series and want them structured the same way.

Length is not the lever it used to be. GPT Image and Nano Banana follow prompts of two or three hundred words without dropping the end. Midjourney still rewards brevity. If a long prompt is giving you a muddled picture, the problem is usually that it contains two pictures, not that it is long.

Old habits that now cost you

Quality tags. "4K", "8K", "masterpiece", "highly detailed", "professional quality". On current models these do nothing or push the image toward a generic glossy look. Say what you want detailed instead: "visible fabric weave on the jacket".

Artist names. "In the style of" a living artist is against the terms of most platforms, is filtered by several, and is the one thing about this technology that people have a right to be angry about. Describe the qualities instead: "muted palette, thick visible brushwork, flat perspective".

Negative prompts as the main tool. Stable Diffusion and Flux interfaces have a negative field and Midjourney has --no. GPT Image and Gemini have neither; you write "an empty street, no people" in the prose and the model handles it. On all of them, describing what should be there works better than listing what should not.

Contradictions. "Bright sunny day, dark moody atmosphere" gets you one of the two, chosen at random. Pick.

Choreographing every element. "Person on the left looking right, mountain top right, river from top right to bottom left" produces stiff pictures, because the composition you are dictating is rarely one that looks natural. Give the framing ("wide shot, subject small in the frame") and let the model place things. Then fix placement in the edit step, not in the first prompt.

Text in the image. For years the advice was "never ask for text". GPT Image and Nano Banana now render short text reliably, and Ideogram has done so for longer. Put the exact string in quotes in the prompt and keep it under six words. Longer than that, add it afterwards in an editor.

Iterate by editing, not by regenerating

The biggest change since 2024 is that the main models edit. You keep the picture and change one thing, instead of rolling the dice again with a longer prompt.

My sequence for a header image:

1. One short prompt for the concept, three or four candidates. Pick the composition, ignore everything else.
2. Edit the winner in follow-up turns: "same image, make the light warmer", "same image, remove the second chair", "same image, extend the right side and leave it empty". One change per turn, so you can see what each one did.
3. Only when the picture is right, ask for the final size and format.

Keep the prompts that worked. A text file with the subject line stripped out gives you a reusable style block; swap the subject and the series stays consistent. The Word Counter is enough to see when a template has grown past what Midjourney will honour. If you are cleaning up a prompt that has accumulated contradictions through iteration, the Prompt Improver rewrites it into one coherent paragraph.

For real consistency across many images, use the tool's reference feature rather than words: Midjourney's style and character references, an uploaded reference image in Gemini or ChatGPT, or a LoRA if you run Flux yourself. Words get you close; a reference image gets you the same.

Prompts for the jobs people actually have

Blog header. Something recognisably about the topic, calm, room for a title. "Photograph of a tidy desk with a laptop showing a spreadsheet, morning light from the left, nothing else on the desk, 16:9, right half empty." Avoid people's faces here; they date the image and draw the eye away from the title.

Social card. "Flat lay of a notebook, pen and coffee cup on pale marble, top down, soft shadows, large empty area on the right, 1:1." The empty area is where the text goes. Ask for it explicitly or the model will fill it.

Product mockup. Describe the scene and a blank stand-in for the product: "a plain white ceramic mug on a wooden cafe table, steam rising, blurred warm lights behind, product photography". Composite your real design onto it afterwards; asking the model to reproduce your logo will get you something close and wrong.

Icon. "Minimal icon of a mountain, single dark blue colour, flat, geometric, centred on white, app icon." Then generate the size variants yourself; the Favicon Generator turns one PNG into every size a site needs.

Series illustrations. Fix the style block once and only change the subject: "Isometric flat illustration, blue and purple palette, thin lines, of [subject]." Ten posts, one look.

What you owe the reader when you publish it

Disclosure is now partly a legal question in the EU. Article 50 of the EU AI Act has applied since 2 August 2026. Providers of image generators must mark their outputs as machine-readable AI content (a watermark or metadata), and anyone who publishes a deepfake, meaning a realistic image of real people, places or events, must make clear that it was generated or manipulated. A stylised illustration for a blog post is not a deepfake. A photorealistic street scene presented as a photograph is closer to the line than most people think. When in doubt, a caption costs nothing.

Copyright. The US Copyright Office concluded in January 2025 that a prompt alone does not make you the author; purely generated images are not protected there. Heavy human editing or compositing can be. The EU position is similar in practice. Assume you cannot stop someone reusing a generated image, and do not build a brand asset on one.

Commercial terms differ per platform. Midjourney grants commercial use on paid plans, with a carve-out for large companies. OpenAI and Google grant you the output but keep the right to use it in training unless you opt out. Read the current terms; they change more often than this post does.

Real people. Never generate a realistic image of an identifiable person without their consent. Every major platform blocks it for public figures and most jurisdictions now have a law against it for everyone else.

FAQ

Do longer prompts give better images?

On GPT Image and Nano Banana, a clear 150-word brief beats a vague 20-word one, and neither drops the end. On Midjourney, keep it under about 60 words. On all of them, a prompt that describes two different pictures gives you a bad blend; the fix is to remove one, not to add detail.

Should I use negative prompts?

Where the tool has them (Stable Diffusion, Flux, Midjourney's --no), use them for artefacts: "watermark, text, extra fingers". Where it does not (ChatGPT, Gemini), write the exclusion into the sentence. Either way, the positive description is doing the real work.

How do I keep a consistent style across many images?

A fixed style paragraph plus a variable subject line gets you most of the way. A reference image or a style reference feature gets you the rest.

Are the results good enough for professional marketing?

For blog and social imagery, presentation visuals and concept work, yes, and they have been for a while. For a campaign hero, a product photo that has to match the physical product, or anything with a face people will recognise, a photographer is still the safer choice.

Which model is best?

It changes every few months, so I will not rank them here. Run the same brief through two of them and pick the one that needs fewer edits for your kind of image. That answer is worth more than any comparison table.