AI photo prompt guide: how to write image prompts in 5 simple steps

A good AI photo prompt describes the finished picture in plain sentences: what kind of image it is, who or what is in it, where it happens, how it's lit and how it's framed. Write that in one to three sentences, generate a few versions, then change one thing at a time until the image looks right. That's the whole method, and it probably works across today's main image models. I'd say most people overcomplicate it.

By the end of this guide you'll have a reusable five-part prompt formula and one finished prompt you built yourself, in about 15 minutes. The advice comes from the official prompting guides for Google's Nano Banana models, OpenAI's GPT Image models, Black Forest Labs' FLUX 3 Image and Midjourney, checked in late October 2026. We compiled it from those docs; nobody at the site ran hands-on tests for this piece, so treat every example prompt below as a starting point to adapt, not a tested result.

Before you start

• One image generator. Gemini, ChatGPT, Midjourney or any FLUX-based app will do. If you haven't picked one, our list of the best AI image generators compares free and paid options.
• A clear idea of the use. A blog header, a product shot, a phone wallpaper. You'll put this in the prompt.
• A notes file. Save every prompt that works. Small edits to a good prompt beat starting from scratch.
• Photos you have the right to use (optional). For editing, upload only your own photos or images you have a licence or permission for. Google's own docs remind you to make sure you have the necessary rights to any images you upload.

If you prefer to watch first, this beginner walkthrough from Taylor Bay Studios explains a prompt formula for Google's Nano Banana model.

Step 1: Name the image type and medium

Start with one short opening phrase that tells the model what kind of picture to make. Black Forest Labs suggests opening with the image type and medium, because it sets the look before you describe anything else.

Pick one: photograph, watercolour painting, pencil sketch, 3D render, flat vector illustration, film still. For photos, OpenAI's guide says to include the word "photorealistic" directly in the prompt to push its models toward a realistic result.

Example to adapt: A photorealistic close-up photograph of...

Tip: if you're learning, start with a simple type like "portrait". BFL notes this should make the effect of later additions easier to see.

Step 2: Describe the subject clearly

Now say who or what is in the picture and what it is doing. Be concrete. Google's guide puts it simply: "Be hyper-specific". Its own example swaps "fantasy armor" for a detailed description of the materials, collar and shoulder pieces.

Use exact numbers, too. Midjourney's docs point out that "cats" is vague and suggest "three cats" instead.

Example to adapt: A photorealistic close-up photograph of an elderly baker with flour-dusted hands, shaping a round loaf of sourdough on a wooden counter.

You should now have one sentence that someone could sketch from. If a friend would ask "what does it look like?", add that detail.

Step 3: Set the scene and the light

Add where it happens and how it's lit. BFL calls location one of the fastest ways to explore variations, since changing only the setting changes the mood. Lighting does the same job: soft window light, golden hour, overcast, neon, studio softbox.

Also say what the image is for. Google recommends you provide context and intent, and OpenAI suggests stating the intended use (an ad, a UI mock-up, an infographic) to set the level of polish.

Example to adapt: ...in a small, warm bakery at dawn. Soft morning light comes through a side window, with steam rising from fresh bread on the shelves behind him. For a bakery website header.

Step 4: Frame the shot

Tell the model where the "camera" is. Google's guide suggests photographic terms like wide-angle shot, macro shot, low-angle perspective. OpenAI adds placement instructions such as "subject centered with negative space on left" when you need room for a headline.

Then choose an aspect ratio that suits the use: 16:9 for a banner, 9:16 for a phone story, 1:1 for a social post. BFL warns that the ratio and the prompt should agree: if the frame is ultra-wide, give the model something to fill the width.

Example to adapt: Shot at eye level from a slight side angle, shallow depth of field, the baker on the right third of the frame with empty space on the left. 16:9.

Watch out: don't overdo camera jargon. Honestly, I'd skip exact lens numbers. OpenAI says detailed camera specs may be interpreted loosely, so use them for the general look rather than exact settings.

Step 5: Generate, then refine one change at a time

Run the prompt and look at the results. Google is blunt about it: "Don't expect a perfect image on the first try." Instead, reply with small follow-ups, such as "make the lighting a bit warmer".

OpenAI gives the same advice under the heading "Iterate instead of overloading": start with a clean base prompt and make single-change follow-ups. When something keeps drifting, re-state the details that matter.

Using an AI photo editor prompt

Editing your own photo works the same way, but you say what to change and what to keep. Google's template reads: "Using the provided image, change only the [specific element]", followed by an instruction to keep everything else the same. OpenAI recommends you repeat the preserve list on each iteration to stop the image drifting.

Example to adapt (on your own photo): Using the provided image of my living room, change only the grey sofa to a dark green velvet sofa. Keep the walls, rug, lighting and camera angle exactly the same.

Only edit photos of people with their consent. Creating intimate or deceptive images of real people without consent is harmful, and it's illegal in many places.

Image to prompt: let AI write the first draft

Stuck for words? Use an image to AI prompt tool; it's a handy shortcut. Midjourney's Describe feature analyses an image you upload and writes four suggested prompts. Midjourney notes the suggestions won't precisely copy your image, so treat them as vocabulary, not a recipe. You can also upload your own photo to Gemini or ChatGPT and ask it to describe the image as a prompt. Our guide to using an AI prompt generator covers this in more depth.

Example: a complete worked example

Here's one AI image prompt built from start to finish with the five steps. The goal: a header image for a small café's autumn menu page.

1. Type and medium: a photorealistic overhead photograph.
2. Subject: a ceramic mug of spiced latte with a leaf pattern in the foam, next to two cinnamon sticks and a slice of pumpkin bread.
3. Scene and light: a worn oak table, scattered dry maple leaves, soft late-afternoon window light from the left. For a café menu header.
4. Framing: top-down view, the mug on the right, open space on the left for text, 16:9.
5. Refine: after the first result, one change per follow-up.

Put together, the first prompt reads:

A photorealistic overhead photograph of a ceramic mug of spiced latte with a leaf pattern in the foam, next to two cinnamon sticks and a slice of pumpkin bread on a worn oak table scattered with dry maple leaves. Soft late-afternoon window light from the left. Top-down view, the mug on the right third of the frame, with clean empty table space on the left for a menu title. Header image for a café's autumn menu page. 16:9.

Then refine with short follow-ups, one at a time:

Make the light a little warmer and the shadows softer.

Keep everything the same, but replace the pumpkin bread with a cinnamon roll.

If you want text on the image, put the exact words in quotes. OpenAI advises putting literal text in quotes or ALL CAPS, and Google says Gemini handles text best if you first generate the text and then ask for an image with the text.

Want to turn the finished still into a short clip? See our picks of the best AI image animator tools.

Common mistakes and how to fix them

• Writing what you don't want. Midjourney warns that if you ask for a party with "no cake," a cake might still appear. Fix: describe what fills the space. BFL's example swaps "a street with no cars" for "a quiet pedestrian walkway with cobblestones".
• Stuffing in quality tags. BFL says to omit quality tags such as "masterpiece", and OpenAI says camera and composition terms steer realism better than generic "8K" labels. Fix: describe the actual light, lens and texture.
• Writing a shopping list. Midjourney says to avoid making long lists or detailed instructions. Fix: write full sentences that say how things relate, such as who is looking at what.
• Changing five things at once. You can't tell which change helped. Fix: one edit per follow-up.
• Copying a living artist or a famous character. Prompts like "in the style of [living artist]" or a well-known cartoon character raise copyright and fairness problems, and many tools block them. Fix: describe the qualities you like instead, such as "loose ink linework, muted earth tones, heavy paper texture".

Tips to get better results

• Put what matters first. BFL calls its order "a habit, not a rule": lead with whatever is most important.
• Break busy scenes into steps. Google suggests describing the background first, then the foreground, then the final detail, for complex scenes with many elements.
• Use full sentences for people. OpenAI suggests details like "full body visible, feet included" or where someone is looking, which helps with proportions and gaze.
• Match the prompt to the model. Midjourney says short and simple prompts typically generate the best images on its platform, while Gemini and GPT Image handle longer, conversational descriptions.
• Draft cheap, finish big. BFL recommends you iterate at a low resolution, then switch to 2K or 4K for the final image.

For more AI prompts for photos in action, this video from Tao Prompts walks through 14 tips for Nano Banana images.

Next steps

Save your best prompts as templates and swap out one part at a time. That's likely the quickest way to get good at this.

• Moving from stills to motion? Our AI video prompt guide uses the same idea, with camera movement added.
• Planning to sell what you make? Read can I sell AI-generated art? first.
• Comparing tools for editing real photos? See our best AI picture tools and our Leonardo AI review.

FAQ

How to write AI prompts for art?

Describe the finished artwork in sentences: the medium (watercolour, oil painting, pencil sketch), the subject and what it is doing, the setting, the light and colours, and the framing. BFL suggests naming both the art form and the style. Describe the qualities you want rather than copying a living artist's name, then refine one change at a time.

What is the best AI image prompt structure?

Most official guides use a similar order. OpenAI recommends background/scene, then subject, then key details, then constraints, plus the intended use. BFL suggests image type, subject, setting, light, then framing. Pick one order, stick to it, and put the most important element first.

Can AI turn an image into a prompt?

Yes, to a point. Midjourney's Describe tool writes four prompts based on an image you upload, and chat assistants like Gemini or ChatGPT can describe an uploaded photo as a prompt. The result is a draft: it won't recreate the original exactly, so edit it before you generate.

Do long prompts work better than short ones?

It depends on the model. Midjourney favours short, clear prompts, while BFL says to start with a sentence and add detail as you review results. Longer prompts seem to help when you have several subjects, a specific layout or text in the image.

Is it OK to use photos of real people in AI prompts?

Only your own photos or photos of people who have given consent. Google's terms tell users not to generate images that deceive, harass, or harm. Making intimate or misleading images of real people without consent is harmful and illegal in many places.