AI Photo Prompts

Write it the way a photographer would brief a shoot

«A realistic photo of a woman in a café» leaves almost every decision to the model, and the model answers each one with the average of everything it has seen. That average is what makes AI pictures look like stock photos. A photographer arriving at a shoot would want six things settled first, and a prompt that settles them gets a picture with a point of view.

1. Who or what, doing what. An age range, a job, an object, an animal, and one action or expression. «Hands still dirty from the work» says more than «a craftsman».
2. Where and when. The place and the time of day, because the time decides the light before you've described any.
3. The light. Its source, its direction and whether it's hard or soft: a window on the left, sun behind the subject, one softbox high at forty-five degrees.
4. The camera. Lens, aperture, the height of the camera and, when something moves, the shutter speed.
5. The look. A film stock or a grade, the amount of grain, and «natural skin texture» for any face.
6. The frame. Where the subject sits, the aspect ratio, and a closing line of things to keep out.

Put together, a brief you can type from memory:

A [kind of photograph] of [subject, doing what] in [place], [time].
[Lens] at [aperture], camera at [height or angle].
Light: [source], from [direction], [hard or soft].
Look: [film stock or grade], [grain], natural skin texture.
Subject on the [left or right] third, aspect ratio [ratio].
Photorealistic. No text, no watermark, no extra people.

Every block on this page is that brief filled in for one kind of photograph, and every one builds the picture from nothing. You don't upload anything.

Which brief for which picture

The blocks below fall into four groups: portraits, commercial shots, places, and moments. Pick by what the picture is for.

When you want your own face in the picture

A brief from scratch invents a person. If the picture has to show you, your partner or your family, you need a prompt that edits an uploaded photo and holds every face exactly as it is. Those live on four pages of their own:

• Headshot prompts turn a selfie into a LinkedIn or corporate portrait with the backdrop, clothes and light changed.
• Couple photo prompts put the two of you into an engagement portrait, a film look or a travel scene.
• Christmas photo prompts make holiday cards from a family photo, plus winter scenes from scratch.
• Halloween photo prompts handle costumes and spooky settings, fun rather than gory.

For other edits of a photo of yourself, such as a new background or a retro look, see the ChatGPT prompts for photos.

One brief, before and after

Here's the café request from the top of this page as it's usually typed:

A realistic photo of a woman in a cafe, beautiful, 8k, ultra detailed, masterpiece.

«8k», «ultra detailed» and «masterpiece» don't describe anything a camera does, so they add nothing the model can act on. We rewrote it as a brief for this page. It's ours, written to show the method:

An environmental portrait of a barista in her forties behind the
counter of a small neighborhood café, mid-morning. 35mm lens at f/2,
camera at counter height. Soft window light from the right, a warm
pendant lamp behind her, the far wall in shadow. She's wiping the
steam wand and looking just past the camera, half smiling. Documentary
color, light grain, natural skin texture. Subject on the left third,
aspect ratio 4:5. Photorealistic. No readable signs, no text, no
watermark, no other people in focus.

Each line answers a question the first version left open. «Barista in her forties» replaces a generic woman with someone who belongs there. The window and the pendant lamp give the face one side lit and one side dark. The steam wand gives her hands something to do, which beats any pose. And «no readable signs» heads off the fake lettering that cafés in AI pictures are full of.

Follow-up lines that fix a photo

In a chat-based tool you can send these as a reply. In Midjourney, add them to the prompt and run it again.

Where AI photos go wrong

A detailed brief fixes most problems, but a few come from the models themselves, and it helps to know them before you spend an afternoon rewording.

• Contradictions. «f/11, everything sharp» and «creamy background blur» can't both happen, and neither can «harsh midday sun» and «soft moody light». The model picks one or blends them badly. Read the brief once for clashes before you run it.
• Hands and text. Both are better than they were and still the first things to check. Give hands a simple job, such as holding a cup or resting on a bench, and keep words out unless the tool is good at them.
• Named landmarks. Ask for a real building and you'll get something close to it with the details invented. That's fine for a mood picture and wrong for anything that claims to show the place.
• Pictures passed off as real. A generated interior isn't the house you're selling, and a generated market isn't a photo of an event that happened. Use the Bright Interior Listing Photo for templates and mock-ups, and say a picture is AI-made wherever someone might take it for a record.
• Too much in one brief. Past a certain length the model starts dropping details. If a line keeps being ignored, move it closer to the start or cut something less important.

Gemini, ChatGPT and Midjourney

The blocks are written to run in all three. What changes is how each tool reads the last line. As of October 2026:

• Gemini generates images with Nano Banana 2 when the chat model is set to Flash or Pro. Paid subscribers can rerun an image with Nano Banana Pro through «Redo with Pro». Google lists downloads at 2K with a Google AI plan and 1K without. Our Nano Banana prompts are written for that model.
• ChatGPT uses ChatGPT Images 2.5, which OpenAI lists for every tier, free included. Pick the ratio from the aspect-ratio picker or leave it in the prompt, as these blocks do. The editor lets you select one area and describe a change to just that part.
• Midjourney runs V8.2 by default and makes square images unless told otherwise. Replace «aspect ratio 4:5» with --ar 4:5 at the end, and add --raw, which Midjourney says gives more photo-like results from simple prompts. Move the «No…» line into --no text, watermark. Midjourney's docs warn that moderation reads each word after --no separately, so keep that list to single, harmless words. Our Midjourney prompts cover its other parameters.

For prices and plan limits, see our reviews of Gemini, ChatGPT and Midjourney. If you want illustration or design work instead of photographs, the AI image prompts page is the place to start.