Asset 20 8 2
Does AI recommend your business? Run the free check →

Join 15,000 business owners, marketers and entrepreneurs. The Sunday newsletter you'll be annoyed only arrives once a week.

Article

How Do You Write Effective ChatGPT Picture Prompts?

If you are skim reading
To write an effective ChatGPT picture prompt, describe up to six things in this order: the exact subject, the style or medium, the composition, the light and colour, one mood word, and the format.

To write an effective ChatGPT picture prompt, describe up to six things in this order: the exact subject, the style or medium, the composition, the light and colour, one mood word, and the format. Be specific about each one, cut words like "beautiful" that tell the model nothing, and ask for changes one at a time, because details can drift with each edit.

Below you'll find that structure, a worked candle example with the exact steps to run it in ChatGPT, how to get short text onto an image, a fix list for the usual failures, and five fill-in patterns you can copy.

Why most picture prompts come out flat

Most people type something like "create a picture of a cosy coffee shop" and then wonder why the result looks like every other AI coffee shop image ever made. The problem isn't the tool. It's that the prompt gave the model almost nothing to work with. ChatGPT's image generation fills in every gap you leave with the most statistically average version of that thing. Vague in, generic out: "a dog" gets you a stock golden retriever on a lawn, while "a scruffy grey whippet asleep on a green velvet sofa" gets you a picture nobody could have stumbled into by accident.

Writing a good picture prompt isn't about being clever. It's about removing the guesswork the model would otherwise have to do for you. Every detail you leave out gets replaced with a default, and the defaults are almost always the least interesting version of your idea.

The six-part structure that gets consistent results

I use the same order every time, because it matches how a photographer or illustrator would think about a shot before taking it. Ignore any part that doesn't apply to your image, but keep the order.

  • Subject: name the exact thing, not the category. "A single amber glass jar candle" beats "a candle."
  • Style or medium: photograph, watercolour illustration, 3D render, flat vector, oil painting.
  • Composition: camera angle, what's in the foreground, what's blurred in the background.
  • Light and colour: golden hour, overcast, neon, monochrome, warm tones versus cool tones.
  • Mood or texture detail: one or two words maximum, not a paragraph.
  • Format: aspect ratio and the platform it's for, square for Instagram, 16:9 for a banner.

That's it. Six lines, not a wall of adjectives. If you've read anything on this site about the broader ChatGPT prompts and daily limits, you'll know the same principle applies across text prompts too: structure beats volume every time.

A worked example: turning a vague brief into a usable image

Say you run a small online candle business and you need a hero shot for an autumn launch on Instagram. Here's the prompt most people would type first:

"Make me a nice candle photo for my shop, autumn theme, make it look professional."

That prompt has a subject and nothing else. No style, no composition, no light, no format. ChatGPT will produce something plausible looking but generic, probably a candle floating in the middle of a white void with vague orange lighting, because "professional" and "nice" don't mean anything specific to a model that has to translate words into pixels.

Here's the same brief run through the six-part structure:

"A photorealistic product photograph of a single amber glass jar candle with a cream label, lit and burning, sitting on a rustic wooden table scattered with dried autumn leaves and two cinnamon sticks. Shot from a slightly high angle, soft natural window light from the left, warm orange and brown tones throughout. Shallow depth of field with a cosy, softly blurred living room in the background. Square 1:1 format for an Instagram feed post."

That's around 65 words, and you can label every part of it. Subject: "a single amber glass jar candle with a cream label, lit and burning". Style: "photorealistic product photograph". Composition: "slightly high angle", the table, leaves and cinnamon in the foreground, "shallow depth of field". Light and colour: "soft natural window light from the left, warm orange and brown tones". Mood: "cosy". Format: "square 1:1 format for an Instagram feed post". You'll get a candle sat in an actual scene, with light behaving the way real light behaves, instead of a candle that looks like it was cut out and pasted onto a background.

Running it in ChatGPT, step by step:

  1. Open a new chat, so an earlier conversation doesn't colour the result.
  2. Paste the full prompt as one message. Asking for the image in plain words is enough.
  3. When the image arrives, ask for one change per message: "Same image, make the window light slightly cooler."
  4. After each edit, check the things you wanted kept: the label text, the leaves, the camera angle.
  5. If something you liked has moved, say what must stay: "Keep the label, leaves and angle exactly as they are, only change the light."

The usual mistake here is asking for three changes in one message, say cooler light, a second candle and a darker table. That tends to move something you wanted kept, and you can't tell which instruction caused it. Three messages, three checks.

What happens when you ask for a second edit

Editing matters as much as the first prompt. ChatGPT will happily edit an image you've already generated, softening the background, changing the candle colour, adding a second candle, and OpenAI's image documentation describes multi-turn editing as a way to iteratively make high fidelity edits. What OpenAI doesn't publish is how many rounds stay faithful, and in everyday use small details can drift as edits pile up: the label text shifts slightly, the leaf pattern changes, the angle tilts a degree or two. So check what you wanted kept after every round rather than assuming it survived.

The practical fix is to decide your full brief before you start, get as close as you can in one or two attempts, and if two edits in a row move something you wanted kept, start a fresh chat with an updated full prompt instead of patching again. If you need pixel-perfect brand consistency, for example the exact label text on every product shot, don't rely on any image generator for it: composite the real label onto the AI-generated background afterwards in a design tool.

Working from your own photo. Upload the image in the chat, say what must not change, then ask for one change: "This is my product shot. Keep the jar, label and table exactly as they are. Replace the background with a softly blurred kitchen in morning light." OpenAI's documentation describes edits that modify existing images using a new prompt, either partially or entirely, which is exactly why you spell out which part is which.

Wondering can ChatGPT enhance my photos? I've shared what it does well.

Words that help and words that waste your time

Some words carry real information for the model. Others are filler that feels descriptive but does nothing.

  • Useful: photorealistic, flat illustration, isometric, macro shot, wide angle, golden hour, overcast, muted palette, high contrast, film grain, 35mm.
  • Mostly useless: beautiful, amazing, stunning, professional, high quality, best. These describe your opinion of the outcome, not an instruction the model can act on.

Swap every useless word for a concrete one and your prompts get shorter and more effective at the same time. When unwanted things keep showing up in the image, try these negative prompts for AI images. That's the opposite of what most people expect: adding detail usually means adding length, but the good kind of detail replaces vague words rather than piling on top of them.

Work with me

Want AI doing the heavy lifting in your marketing?

I build the systems that handle the boring 80 percent, so you get your week back. Done properly, with the human kept in.

Leaving things out. There's no separate field for exclusions in ChatGPT, so put them in the prompt in plain words, at the end: "no people, no text, no logos".

Text inside the image. Keep any words to a short phrase in quotes and say how they should look: write the words "AUTUMN GLOW" on the label in cream serif capitals. Then check the spelling letter by letter, because a swapped letter is easy to miss at thumbnail size. Longer text, such as a full sentence or a price list, is where mistakes creep in, so if a short phrase fails twice, generate the image without text and add the words in a design tool afterwards.

Length has a ceiling, and it's lower than people think

OpenAI doesn't publish an ideal length for image prompts, so treat any word count you read, including the 65 words in the candle example, as a rule of thumb rather than a limit. What reliably goes wrong in long prompts is contradiction: "warm golden light" fighting "cool blue evening tones" three sentences later, or a composition note that clashes with the aspect ratio you asked for. Conflicting instructions tend to come back as a muddier, more generic image, which defeats the point of writing a detailed prompt. To find your own sweet spot, take one idea written at about 60 words and again at about 140, run each twice in fresh chats and compare. Whatever the length, put the must-have elements first. That's a practical habit rather than documented behaviour, but it means anything that gets lost is the least important part.

When the image comes back wrong

  • Wrong shape: state the format in words and numbers at the end, "landscape, 16:9". If it still comes back square, reply "Same image, landscape 16:9" and nothing else.
  • Extra fingers or stray objects: name the exact spot and ask for that fix alone, "Same image, the left hand has five fingers, change nothing else", or "remove the second spoon".
  • Brand colours off: give the colour in words and the hex code, "deep teal (#0F5E63)", then check it against your brand file. If it's close but not exact, correct it in a design tool rather than regenerating the whole image.
  • People, text or logos you didn't ask for: add the exclusion line, "no people, no text, no logos".
  • Misspelt words: shorten the phrase, put it in quotes, and after two failures add the text yourself.

Matching the tool to what you're trying to make

ChatGPT is a strong general choice for product shots, illustrations, social graphics and quick concept images, especially if you're already using it for copywriting and want to stay in one tab. But it isn't the only option, and it isn't always the best one. If you're producing a high volume of images or need finer control over style consistency across a whole campaign, it's worth comparing how it stacks up against Gemini's image tools, which some people find stronger for photorealistic scenes. I've laid out where each tool wins in the ChatGPT versus Claude versus Gemini comparison if you want the fuller picture before you commit hours to one workflow.

Worth knowing too: free ChatGPT accounts can run into a limit on image generations, and OpenAI doesn't publish a fixed number for it, so plan a batch of images in one sitting and keep a spare tool ready for when the limit hits. What's known about the caps and how they reset is in the guide to free ChatGPT limits, and it's worth reading before you plan a whole content calendar around free-tier image generation.

Ready-to-copy prompt patterns

If you want a starting template rather than building one from scratch, these fill-in-the-blank patterns cover most small business needs:

  • Product shot: "A photorealistic product photograph of [exact object], [angle], [light source and direction], [colour palette], [background description], [aspect ratio and platform]."
  • Social graphic: "A flat vector illustration of [subject], minimal line style, [two colours named], white background, centred composition, square format for Instagram."
  • Blog header: "A wide photograph of [scene], shot at eye level, natural light, muted colour palette, empty space on the left third for text overlay, 16:9 format."
  • Illustration with no people: "A [watercolour / flat vector / paper-cut] illustration of [object or scene], [two or three colours named], [background], no people, no text, [aspect ratio and platform]."
  • Portrait-style social post: "A photograph of [person described by role, age range and clothing], [what they're doing], [setting], [light and direction], waist-up framing, vertical format for an Instagram feed post."

Here's the blog header pattern filled in: "A wide photograph of a tidy home office desk with a laptop, a mug of tea and an open notebook, shot at eye level, natural morning light, muted sage and oatmeal palette, empty space on the left third for text overlay, 16:9 format." Put that next to "a blog header about productivity" and it's obvious which one leaves the model guessing.

These patterns work for seasonal content too. If you've seen the specific, copy-paste style used in the Thanksgiving photo prompts for family edits, that's the same principle applied to a personal use case rather than a business one: name the exact people, setting and mood instead of describing the occasion in general terms.

If the picture is going on a card, there's a separate guide to writing a chatgpt happy birthday message that doesn't read like a template.

I run AI training and workshops for teams, and help businesses with AI enablement and workflow automation, so the busywork runs itself and people get time back. Most of that work is with SaaS marketing teams and small businesses. Here's how we can work together.

Frequently asked questions

Can ChatGPT put exact text on an image?

It can add a short phrase if you put the words in quotes and describe the lettering, but check the spelling every time. For anything longer than a few words, add the text in a design tool afterwards.

Why does ChatGPT ignore part of my prompt?

Most often because two instructions contradict each other, and the model averages them rather than picking one. OpenAI doesn't document how it weighs the order of your words, but putting the most important elements first and removing anything that clashes is the quickest fix.

How do I get the same style across several ChatGPT images?

Write your style lines once (medium, light, colours) and paste them word for word at the end of every prompt in the set, in the same chat. Then line the results up side by side and regenerate any that wander.

Can I tell ChatGPT what to leave out of an image?

Yes, say it plainly at the end of the prompt, for example "no people, no text, no logos". The negative prompts guide linked above has the patterns that hold up best.

Where to check the details

Run a business, a brand or a blog? I feature tools, products and expert guest posts for an audience of 15,000. See how it works

Published and maintained by the Lilach Bullock team, covering marketing, AI and business growth.
Your buyers are asking AI who to use. Does it say you?

See for free whether ChatGPT, Claude, Perplexity, Gemini and Google name you, and get the plan to become the answer.

Check my AI visibility →
Sundays only

Get the Sunday newsletter.

One email a week. AI experiments, marketing tactics, and the workflows Lilach is building right now in her own business.

Subscribe free

Let’s get your marketing running on AI.

Book a free 30-minute call

We figure out what you need, where AI fits in, and what working together would look like.

Book the call →

Or take the 30-second calculator

You’ll see the hours and the money quietly leaking out of your week, and the three workflows worth building first.

Take the calculator →

Or grab the free AI resource library

Prompt packs, templates, checklists, and swipe files. The exact tools I build for paying clients. Yours, free.

Get the library →
Keep reading

More from the blog.