ChatGPT Image Prompts: How to Get Better Pictures From GPT
ChatGPT is not a slot machine where you pull the lever once. It is closer to briefing an illustrator who never gets tired of revisions. The people who get the best pictures write a clear brief and then talk the image into shape.
In this guide
- Why ChatGPT needs a different style of prompt
- The brief formula
- Copy-ready ChatGPT image prompts
- The conversation workflow
- Getting text right
- Keeping a consistent style across images
- Mistakes to avoid
- Rewriting keyword prompts for ChatGPT
- When to use ChatGPT versus other generators
- Example: one idea refined in four messages
- FAQ
Why ChatGPT needs a different style of prompt
Generators such as Midjourney respond well to comma-separated keyword lists. ChatGPT's image generation is built into a chat model, so it understands natural language, context and follow-up instructions very well. That changes the best approach:
- You can write full sentences and explain relationships ("the cat is sitting on the suitcase, not next to it").
- You can mention the purpose ("for a bakery's Instagram post") and the model uses it to make sensible choices.
- You can edit with plain requests ("make it evening", "remove the plant on the left").
Features, limits and usage rules change over time and vary by plan, so check OpenAI's help pages if something behaves unexpectedly.
The brief formula
Write your prompt in this order:
- Purpose and format: "Create a wide banner image for a travel blog header."
- Subject and action: "A woman in a yellow raincoat walks across a stone bridge."
- Setting and details: "Old town of Prague, early morning, almost empty streets."
- Style: "Realistic photograph" or "flat vector illustration with bold shapes."
- Lighting and colour: "Soft fog, cool blue tones with the yellow coat as the only bright colour."
- Composition: "She is small in the frame on the right third; lots of sky."
- Text, if any: exact words in quotes and where they go.
Put together:
Create a wide banner image for a travel blog header. A woman in a yellow raincoat walks across a stone bridge in the old town of Prague in the early morning, with almost empty streets. Realistic photograph, soft fog, cool blue tones with the yellow coat as the only bright colour. She is small in the frame on the right third, with plenty of misty sky on the left for a headline.
Copy-ready ChatGPT image prompts
Blog illustration
Make a friendly flat illustration for a blog post about saving money on groceries. Show a person comparing two price tags in a supermarket aisle. Use a limited palette of teal, cream and coral, simple shapes, no text, square format.
Recipe photo
A realistic overhead photo of a bowl of dal tadka with a swirl of ghee and fresh coriander, a small bowl of rice and a folded roti on the side, rustic wooden table, soft daylight from a window on the left, square format.
Kids' storybook page
A gentle watercolour illustration for a children's bedtime story: a small elephant reading a book by lantern light inside a tent, stars visible through the opening, warm and cosy colours, vertical format.
Simple infographic-style image
A clean diagram-style illustration showing the water cycle with four labelled stages: "Evaporation", "Condensation", "Precipitation", "Collection". Arrows connect them in a circle. White background, blue and green palette, wide format.
Event poster background
A vertical poster background for a summer music night: a crowd silhouetted against a warm sunset stage glow, string lights overhead, leave the top third clear for the title, no text.
The conversation workflow
The real advantage of ChatGPT is iteration. Use it like this:
- Generate from your brief.
- Name the biggest problem in one message: "The bridge looks too modern; make it an old stone arch bridge."
- Keep what works: "Keep everything else the same."
- Adjust style after the content is right: "Now make the colours more muted, like a film photo."
- Ask for variations once you are close: "Show a version from a lower angle."
Asking for six changes at once tends to produce a new, different image rather than an edit. One or two changes per message is more predictable.
Getting text right
Image models have improved at lettering, and ChatGPT's generator is often good with short text, but mistakes still happen. To improve your chances:
- Put the exact words in quotes.
- Keep text short: a title and maybe a subtitle.
- Say where it goes and what it looks like: "bold white sans-serif at the top centre".
- Proofread every letter before you publish.
Our detailed guide to readable text in AI images has more fixes.
Keeping a consistent style across images
If you need a set of images, for example for a presentation or a social feed:
- Write a style paragraph once and paste it into every request: "Flat vector illustration, thick rounded outlines, palette of navy, mustard and off-white, soft paper texture."
- In the same chat, refer back: "Use exactly the same style as the previous image."
- Describe characters with the same fixed traits every time.
For people and mascots, see consistent character design prompts. For feeds, read Instagram post prompts for consistent branding.
Mistakes to avoid
- Keyword soup. "dog, cute, 8k, masterpiece, trending" adds little in ChatGPT. Describe the scene instead.
- Vague size. Say square, wide or vertical so the layout fits your use.
- Contradictions. "Minimal" and "lots of detail" fight each other.
- Expecting exact real people or brands. Policies limit images of real people and trademarks. Describe fictional people and generic products.
- Not saving the prompt. When an image works, save the final wording along with the follow-ups that fixed it.
Rewriting keyword prompts for ChatGPT
Many prompts online are written as keyword lists. Convert them by turning each cluster into a sentence.
Keyword prompt: cyberpunk alley, neon, rain, lone figure, umbrella, cinematic, teal and pink
ChatGPT version: "A lone figure with a transparent umbrella stands in a narrow alley at night in a futuristic city. Neon signs in teal and pink reflect on the wet ground during light rain. Cinematic wide shot, moody atmosphere."
The AI Image Prompts Android app is handy here: it gives you a searchable, categorised library of prompts with preview images, one-tap copy and saved favourites, so you can grab a starting point, paste it into ChatGPT and expand it into a brief. The app is a prompt collection, not an image generator; the picture is made by whichever tool you paste into.
When to use ChatGPT versus other generators
ChatGPT is a strong choice when the image needs to follow instructions closely, includes simple text, or benefits from back-and-forth editing. Midjourney is popular for distinctive artistic styles, and Gemini also supports conversational creation and editing. Our guides to Midjourney prompt structure and Gemini prompts show how to adapt one idea across tools.
Example: one idea refined in four messages
Here is how a real refinement session might go for a café's menu board image:
- First message: "Create a square image of a chalkboard menu on a café wall, with a small plant and a cup of cappuccino on the counter in front. Warm morning light, realistic photo, no readable text on the board yet."
- Second message: "Good. Make the counter light oak instead of dark wood, and keep everything else the same."
- Third message: "Now add the words "Today's Special" in neat white chalk lettering at the top of the board."
- Fourth message: "Make the light a little softer and the photo slightly warmer, like early sunrise."
Each step changes one thing, so you can see exactly what each instruction did and roll back if a change makes the image worse.
FAQ
How long should a ChatGPT image prompt be?
Two to five sentences is a good range. Include purpose, subject, setting, style, lighting and format. Add more only if it describes something specific.
Can ChatGPT edit an image it already made?
Yes, you can ask for changes in follow-up messages. Describe one or two changes at a time and say what should stay the same.
Why did ChatGPT change my whole image when I asked for a small fix?
Big or multiple requests can trigger a fresh generation. Ask for a single change and add "keep everything else the same".
Does ChatGPT understand Midjourney parameters like --ar?
No, those are Midjourney-specific. Simply say "wide format", "square" or "vertical 9:16" in words.
Can I use ChatGPT images commercially?
Check OpenAI's current terms and your local law. Our article on copyright and AI images explains the general issues.