How to Write AI Image Prompts (2026 Guide)

Type a one-line idea into any AI image generator and you will usually get something competent, generic, and nothing like what you imagined. The gap is rarely the model — it is the prompt. This guide teaches a six-layer formula for writing prompts from scratch, then proves it on a single subject rendered three times with our own free generator: same engine, same settings, one layer of the formula added at each stage, every prompt and image shown exactly as produced. (Working backwards from an image you already like is a different skill — that one is covered in our image to prompt guide.)

An image model has to make hundreds of decisions to produce a picture: the light, the palette, the camera position, the era, the weather, the material of every surface. Your prompt makes some of those decisions; the model averages its training data for all the rest. A vague prompt is not punished — it is averaged. That is why one-line prompts produce stock-photo blandness: you made one decision and the model made the other few hundred. Writing a good prompt is simply making more of the decisions that matter to you, in words the model can act on.

Almost every strong prompt, on every engine, is some arrangement of the same six layers. You will not always need all six — but when a render disappoints, the fix is almost always a missing layer, and this table tells you which one.

To show what each layer actually buys you, we rendered the same subject three times with the free AI image generator on this site — identical engine and settings each time, so the only variable is the prompt. The prompts are quoted verbatim and the renders are untouched.

A perfectly fine image, and none of it is yours. Daytime, calm sea, white tower, mid-distance view — every one of those was the model averaging its idea of "lighthouse". If this is what you wanted, stop here. It almost never is.

Now the image contains your decisions: the stone is weathered, the lantern room is red, the cliff and the dawn and the waves are all there. Layers 1–3 control CONTENT, and content is now solved. But look at the treatment — the light is pretty in an unremarkable way, the framing is conventional. That is because style, lighting and camera are still unwritten, so the model is still averaging them.

This is the jump people pay for, and it cost fourteen words: the mood, the mist, the silky long-exposure water, the sun burning through the clouds, the dramatic grade. Layers 4–6 control TREATMENT, and treatment is what separates a render you scroll past from one you save.

Two honest observations from this run, because they are lessons in themselves. First, the prompt asked for a "wide-angle shot from below" and got a wide but slightly elevated view — camera instructions are suggestions, not commands, and the fix is to keep the instruction and render again rather than to pile on more camera words. Second, compare the stages closely: the door colour and the exact shade of the lantern room drift between images. Each render is independent — a prompt buys you a distribution of images, not one repeatable picture. Both of these are normal, and knowing they are normal is half the craft.

Vocabulary is what turns the formula from a checklist into a tool. These are the words that reliably move each layer, on every mainstream engine:

Engines weigh early words more heavily than late ones, so front-load the subject and its defining detail, and let treatment layers trail. Beyond ordering, some engines offer explicit emphasis: Stable Diffusion can weight any term with (term:1.3) syntax — covered properly in our Stable Diffusion prompts guide — and Midjourney shapes results with parameters like --stylize, covered in the Midjourney prompts guide. On engines with no weighting (Flux, DALL·E 3), word order and repetition are your only emphasis tools, which is one more reason to keep prompts lean.

The six layers are universal; the packaging is not. The same layered lighthouse prompt would be phrased differently per engine:

Each has a dedicated deep-dive here: Midjourney, Stable Diffusion, Flux and DALL·E 3.

Once you understand the layers, you can also delegate them. Our free text to prompt enhancer takes a one-line idea and expands it into a fully layered prompt — no signup — which is a fast way to see the formula applied to your own subjects. If you are starting from an image rather than an idea, the image to prompt tool reverse-engineers a layered prompt from any picture. And the AI image generator that produced every render on this page gives you free daily images to practise with.

Write the subject, decide the details, place it in time, choose its medium, light it, frame it. Six decisions — and the difference between the first lighthouse on this page and the last one.