DALL-E prompt tips: getting better images out of ChatGPT

Image generation in ChatGPT started with DALL-E and has since moved to newer OpenAI image models, but the way you prompt it has stayed broadly the same: you describe what you want in a chat, and a language model interprets the request before the image is made. That middle step is the key to prompting it well. ChatGPT reads your message as a conversation, not as a list of tags, so the habits that work in Midjourney or Stable Diffusion often don’t carry over.

The tips below apply whichever OpenAI image model your ChatGPT plan currently uses.

Describe it like you’d brief a person

Write in plain, complete sentences. Tag lists, weights such as (word:1.3) and Midjourney-style flags like --ar aren’t part of ChatGPT’s prompt syntax; at best they are read as ordinary words and at worst they confuse the request.

Before:

mountain cabin, snow, night, aurora, cinematic, 8k, ultra detailed, --ar 16:9

After:

A small wooden cabin on a snowy mountainside at night, warm yellow light glowing from its two windows. Green and violet aurora ripples across the sky above the ridge. Wide landscape format, cabin in the lower third, mostly sky. Photographic, quiet and cold.

The rewrite keeps the same idea but settles where things are, where the light comes from and what format the image should be. Because ChatGPT follows grammar, “warm yellow light glowing from its two windows” stays attached to the cabin rather than leaking across the whole scene.

Lead with the subject and medium

Start with what the image is: “A photo of…”, “A watercolour illustration of…”, “A 3D render of…”, “A flat icon of…”. That single choice changes how everything after it is read. Then add setting, composition, light, palette and mood, in roughly that order.

Be explicit about format

ChatGPT will choose a format if you don’t, and it usually picks square. If you need something else, say so in words and give the reason:

• “Wide landscape format, for a website banner.”
• “Tall portrait format, for a phone wallpaper.”
• “Square, with the subject centred and empty space around it, for a profile picture.”

Mentioning the purpose helps with composition as well as shape. A banner request tends to leave room for text; a wallpaper request tends to keep the important part away from where the clock sits.

If you need a transparent background, ask for it directly (“a sticker of a cartoon cactus on a transparent background”). Whether it is supported depends on the model and interface you’re using, so check the downloaded file.

Getting text right

OpenAI’s newer image models handle text in images far better than early DALL-E did, but clear instructions still matter:

• Put the exact text in quotation marks.
• Say where it goes and how it looks.
• Keep it short: titles, labels and short phrases, not paragraphs.

A vintage travel poster for a fictional island called "Isla Calma". The name in large cream serif letters across the top. Below, a turquoise bay with two sailboats and white cliffs, painted in flat gouache with visible brush texture. Small text at the bottom reading "Ferries daily".

If one word comes out wrong, ask for that fix specifically in the next message: “Keep everything the same, but correct the bottom line so it reads exactly: Ferries daily.”

Editing in conversation

Because the image lives inside a chat, you can refine it with follow-up messages instead of rewriting the whole prompt. This is the biggest practical difference from tools where every generation starts fresh.

Good follow-ups name one change and what to keep:

1. “Make the sky darker, closer to midnight. Keep the cabin and aurora as they are.”
2. “Add a thin column of smoke rising from the chimney.”
3. “Move the cabin slightly further left.”

Vague feedback (“make it pop”, “more epic”) forces ChatGPT to guess, and it may change elements you were happy with. If an image drifts after several rounds, it’s often faster to go back: “Return to the second version and only change the sky colour.” Some ChatGPT interfaces also let you select an area of the image and describe a change just for that region, which is the most precise way to make small fixes.

Ask to see the prompt

You can ask ChatGPT what it actually sent to the image model, or tell it to use your wording exactly. That is useful when results keep missing a detail: you can see whether the detail was dropped before generation, and add it back.

Wording that avoids refusals

OpenAI’s usage policies restrict some kinds of images, and requests that look like they cross a line are declined even when your intent is harmless. A few habits reduce false refusals:

• Describe style qualities, not living artists. Instead of naming a contemporary illustrator, describe what you like about the look: “bold black ink outlines, flat muted colours, lots of negative space”. This is more precise anyway, because it says which aspects of a style you want.
• Use fictional people and brands. Requests for images of real, identifiable people, or for trademarked logos and characters, are often refused or restricted. Describe an original character instead: “a woman in her sixties with short silver hair and round glasses”.
• Make benign context clear. A request that mentions injuries, weapons or medical scenes is easier to fulfil when the purpose is obvious: “an illustration for a first-aid training leaflet showing how to apply a bandage to a forearm”.
• Don’t try to rephrase around a refusal. If a request is declined because of what it depicts, rewording it to sneak past the rule isn’t a workaround the policy allows. Change what you’re asking for instead.

Examples

Product mockup:

A studio photo of a ceramic coffee mug in matte terracotta on a light oak table, soft daylight from the left, a sprig of eucalyptus beside it. Square format, mug centred, neutral warm-grey background, minimal and clean.

Follow-up edit:

Same image, but change the mug to a glossy deep blue glaze. Keep the lighting, table and eucalyptus unchanged.

Illustration for an article:

An editorial illustration of a person standing at a fork in a forest path, one path lit by warm sunlight, the other in cool blue shadow. Seen from behind, small in the frame. Flat shapes with subtle paper grain, palette of mustard, teal and charcoal. Wide landscape format for a blog header.

Icon set:

A set of four simple line icons on a white background: a house, a bell, a gear and a user silhouette. Uniform 2-pixel black strokes, rounded line ends, no fill, arranged in a row with equal spacing.

Moving prompts between tools

A prompt written for ChatGPT usually works in Gemini and Flux with little change, since both also prefer full sentences; our Nano Banana prompt guide covers Gemini’s editing workflow in detail. Moving to Midjourney or Stable Diffusion takes more reformatting, which the Midjourney prompt generator and Stable Diffusion prompt generator handle for you.

If you’re starting from an image rather than an idea, reverse-engineering it into a prompt gives you a plain-language description to paste into ChatGPT. If you only have a few words, the text to prompt tool expands them into a full description first.

Checklist

• Plain sentences; no tags, weights or -- flags.
• Open with the medium and subject.
• State the format and what the image is for.
• Exact text in quotation marks, short, with placement and lettering described.
• Edit with one change per message, and say what to keep.
• Describe style qualities instead of naming living artists; use fictional people and brands.
• When a detail keeps going missing, ask ChatGPT to show the prompt it used.