To prompt an AI image generator, describe one picture the way you would brief a photographer or illustrator: who or what is in it, what they are doing, where, in what light, from what angle and in what style. That brief is most of how to make AI images that look like what you pictured, and it works for AI art, product shots and birthday cards alike. Below are six steps, the six parts of a prompt, how ChatGPT, Gemini, Midjourney and Ideogram each want a prompt written, how to get words into an image, a worked example, and how to edit a result instead of starting over.
How to prompt an AI image generator in 6 steps
These six steps cover how to create AI images from a blank text box, and they work the same way in ChatGPT, Gemini, Midjourney and Image Generator in Toybox AI. If you are working out how to use an AI image generator for the first time, start here.
- Say what the picture is for. Google's Gemini docs recommend explaining the purpose, with the example that "Create a logo for a high-end, minimalist skincare brand" works better than "Create a logo." "A header image for a bakery's autumn menu" gives the model a use and a mood to work from.
- Say who or what is in it, and the action. Use numbers and specifics: "two golden retriever puppies tugging on a red rope toy" instead of "dogs playing". Midjourney's guide makes the same point with "three cats" over "cats".
- Add the setting, the light, the framing and the style, once each. "On a sunny back porch, late afternoon light from the left, low angle close-up, photo."
- Put exact words in double quotes. Anything that should appear as text, such as "Open Late", goes in quotes and stays short.
- Pick the shape in the tool. Choose square, portrait or wide in the size setting before you generate. ChatGPT has an aspect ratio picker for this.
- Generate, compare, and change one thing. Read the result against your prompt, find the biggest miss, and fix only that in the next round. OpenAI's prompting guide recommends small single-change follow-ups, and Google's docs recommend iterating in conversation.
The parts of an image prompt
Prompt engineering for image generation is mostly filling in the same six slots and leaving out what doesn't matter. OpenAI's prompting guide suggests a steady order: the scene, then the subject, then the key details, then the constraints. The table starts with the subject instead; to follow OpenAI's order, move the setting to the front.
| Part | What to write | Example |
|---|---|---|
| Subject and action | The main figure and its action, with counts | "An older man repairing a bicycle wheel" |
| Setting | A specific place and what surrounds the subject | "In a cluttered garage with tools on a pegboard" |
| Light | One source, its direction and the time of day | "Morning sun through the open garage door" |
| Framing | Distance, angle, and a lens if it's a photo | "Eye level, medium shot, 35mm lens" |
| Style or medium | One medium, with its texture | "Documentary photo, natural color" |
| Words and limits | Exact text in quotes, plus what must stay or go | "No text anywhere in the image" |
For the limits slot, Google's Gemini docs recommend what they call "semantic negative prompts": describe the scene you want ("an empty, quiet street") instead of listing what to remove ("no cars"). The article on negative prompts covers when a "no" works and when it backfires. When a photo should set the subject or the style, add it as a reference and name in the prompt what to take from it ("the dog from my photo"); the page on generating from a reference image has five worked prompts. Thirty ready-to-copy prompts built from these six parts are in the AI image prompts library.
How ChatGPT, Gemini, Midjourney and Ideogram read prompts
How to generate images with AI depends partly on the tool: the same idea needs a different shape in each one, because each maker's documentation asks for something different. As of September 2026:
| Tool | Prompt style its maker recommends | Words in the image | Shape |
|---|---|---|---|
| ChatGPT Images | Describe the image in a chat and keep asking for changes there | Give the exact wording; OpenAI says ChatGPT Images adds text on request | Aspect ratio picker, or name the ratio in the prompt |
| Gemini (Nano Banana 2) | A narrative scene description, refined over several turns | Settle the text first, then ask for the image with it | Named in the prompt, as Google's examples do |
| Midjourney | Short, simple phrases; instructions work too with the Edit model (V8.1 and later) | Double quotes only; short words in the Latin alphabet | The --ar parameter at the end |
| Ideogram | Complete sentences with punctuation | Quotes, with the text named early in the prompt | Presets between 1:3 and 3:1 |
The main split is length. Midjourney's Prompt Basics favors short, simple prompts and cautions against long lists of instructions, while Google's Nano Banana guide asks for the scene written out as a narrative, not a keyword list. So a prompt copied from a Midjourney gallery usually needs to be written out in full for ChatGPT or Gemini, and a ChatGPT paragraph needs trimming for Midjourney. For step-by-step help with one tool, see how to use Gemini's image generator or how to make images with ChatGPT. For Gemini-shaped prompts to copy, the Nano Banana prompts follow Google's own structure.
How to get words into an AI image
Each maker's docs include advice for text in images, and most of it lines up:
- Quote the text exactly. Midjourney needs double quotation marks; its docs say single quotes won't work. OpenAI's guide suggests quotes or ALL CAPS for the literal text.
- Keep it short. Ideogram's docs warn that each extra word raises the odds of a typo or a dropped word, and Midjourney's say shorter words and phrases render more reliably.
- Name the text early. Ideogram's docs advise putting the text in the first part of the prompt and saying where it goes.
- Describe the font rather than naming it. Google's Nano Banana guide says you can name a font, but Ideogram's docs say a typeface can't be specified by name there. A description such as "bold rounded sans-serif" or "thin formal script" is the safer choice in any tool.
- Spell hard words out. OpenAI suggests spelling uncommon names letter by letter, such as "M-A-R-I-S-O-L".
- Expect English to work best. Ideogram's docs name English as its most accurate language for text, and Midjourney's recommend the standard Latin alphabet.
A flyer carries more text than any of these tips assume: a date, a time, an address and a price. For that job, Flyer Maker lays out a whole flyer from the details you type, and you still check each date and price.
A worked example, from vague to usable
Three rounds on one idea show the method. The first attempt, "A cozy coffee shop", is a topic, not a picture: it has no subject, no moment and no light. The second round adds a subject, an action, the light, the framing and a style, and picks one moment:
A photo of a barista pouring steamed milk into a cup at the counter of a small coffee shop on a rainy morning. Warm light from pendant lamps, rain on the front window behind her, a chalkboard menu blurred in the background. Eye level, medium shot, 35mm lens, natural color.Say the result comes back with the barista facing away from the camera. The third round changes only that and restates what stays:
A photo of a barista pouring steamed milk into a cup at the counter of a small coffee shop on a rainy morning, facing the camera and smiling down at the cup. Keep the warm pendant lamps, the rain on the window, the blurred chalkboard menu, the eye-level medium shot and the natural color.Each round fixes the largest miss and nothing else, so you can tell which change did what.
How to edit an AI image instead of starting over
When a result is close, edit it rather than generating a new one from scratch. An edit instruction names the change and lists what stays: "Change only the mug to dark green. Keep her face, her hands, the table and the light the same." OpenAI's guide recommends repeating that keep-list every round, because details drift between edits.
Smart Editor in Toybox takes that kind of instruction in its "What should change?" box and runs on OpenAI's GPT Image 2.5. Smart Editor is a Pro-plan tool, currently 85 credits for HD or 110 for 4K. Each edit redraws the whole image, since there is no brush or mask, so end every instruction with "keep everything else the same". After a result, the "Editing" switch sets whether the next edit starts from "Latest result" or "Original photo". On Image Generator results, a Pro plan also adds "Want to change something?" under the image for quick fixes.
Troubleshooting
| Problem | Likely cause | Fix |
|---|---|---|
| The image ignores half the prompt | Too many ideas in one prompt | Split it: one subject, one setting, one style |
| Every result looks the same | The prompt names a topic, not a scene | Add an action, a place and a light source |
| A "no" item appears anyway | The prompt names the thing to avoid | Describe the scene without it, or use the tool's own exclusion setting |
| Words are misspelled | Long or unusual text | Shorten it, quote it, spell hard words letter by letter |
| The shape is wrong | The ratio was typed but not set | Choose the size in the tool before generating |
| A fix changed other things | The edit didn't say what to keep | Add "keep everything else the same" and list the key details |
| The subject is tiny in the frame | The prompt gave no distance | Add the framing: "close-up", "medium shot" or "full body" |
When the result looks generated rather than real, the fixes are different: see realism fixes for AI images.
Make your first AI image in Image Generator
- Open Image Generator. You need a free Toybox account to create anything.
- Type your description into Describe your image using the six parts, or tap "Try this prompt" under the round-three prompt in the worked example.
- To guide the subject or style with your own photo, use Add reference photos (three files at most, JPG or PNG).
- Choose a shape under Size: Square, Portrait, Landscape, Story or Wide. For designs with words, switch on Enhanced quality under More options (currently 45 extra credits).
- Tap Create image and wait about 30 seconds, the usual time for an image.
- Compare the result with your prompt and change one thing. On a Pro plan, you can also send the result to Smart Editor with "Open in Smart Editor" for a targeted edit.
Image Generator currently costs 50 credits per image, or 95 with Enhanced quality. New accounts get 20 free credits, which doesn't cover an image; see current prices for plans and top-ups.