ChatGPT Image Generation: A Practical Prompting Guide

ChatGPT makes images inside the chat window using OpenAI's GPT Image model family, which you may see labelled as 4o Image Generation or GPT-Image-1 depending on where you look. It behaves differently from most image tools: there is no parameter syntax to learn, long descriptive prompts are a strength rather than a liability, and you refine the picture by talking to it. This guide covers how to write for it, which choices actually matter and how to work efficiently within the limits of your plan.

How image generation in ChatGPT works

You type a request such as "Create an image of..." and ChatGPT generates a picture in the conversation. Because the model is built around language understanding, it reads your whole prompt rather than latching on to a few keywords, and it keeps the context of the chat. That means a second message like "make the sky overcast" edits the previous image rather than starting again. Outputs can be square, portrait or landscape, and you can ask for a transparent background, which is delivered as a PNG.

Two practical notes. First, the exact pixel dimensions, how many images you can make per day and how fast they arrive depend on your plan and change over time; OpenAI's help pages have the current figures. Second, the model is strong at rendering readable text in images, which older systems struggled with, so signage, labels and headlines are fair game (see text and typography in ChatGPT images).

Writing a prompt that works

Because long prompts are followed well, the useful habit is to write a brief rather than a tag list. A structure that reliably gives good results:

1. Subject. Who or what, with two or three concrete details. "A ceramic teapot with a cracked blue glaze and a bamboo handle."
2. Setting. Where it is and what surrounds it. "On a weathered oak table by a window, a folded linen napkin beside it."
3. Light. Direction and quality. "Soft morning light from the left, gentle shadows."
4. Style and medium. Photograph, watercolour, flat vector illustration, 3D render, and any era or technique. Describe qualities rather than naming living artists.
5. Composition. Framing, angle, how much of the frame the subject fills, and the orientation. "Close-up, slightly above eye level, portrait orientation, subject centred with space above."
6. Constraints. Things to avoid or include: "no text", "plain background", "leave the bottom third empty".

Write it as sentences. The model does not need commas-and-keywords, and full sentences make your intent unambiguous. Avoid stacking contradictory style words ("photorealistic watercolour") unless you mean exactly that. If you want several ideas, ask for them in separate messages rather than one message with five requests; each image is generated on its own.

Settings at a glance

ChatGPT has no settings panel for images; the "settings" are things you say in the prompt. These are the ones worth saying every time.

Iterating in the same conversation

The chat is your editing tool. After the first image arrives, describe the change you want in a sentence: "keep everything but make the teapot green", "move the camera lower", "add a second cup". The model applies the change to the existing image, which is far more efficient than rewriting the whole prompt. A few habits keep this productive:

• Make one change per message so you can see what each request did.
• Say what to keep as well as what to change; it reduces unintended drift elsewhere in the picture.
• If an edit chain goes wrong, go back to the message with the last good image and branch from there rather than piling fixes on fixes.
• When you get an image you like, paste its full prompt and the sequence of edits into your notes. ChatGPT conversations are easy to lose, and reconstructing a look from memory is painful.

Uploading your own images to edit or to use as references is covered separately in image edits and references.

Try it

Create a landscape-orientation photograph of a small independent bookshop at dusk, seen from across a narrow cobbled street. Warm light spills from the shop window onto wet stones; a bicycle leans against the wall. Overcast sky, soft reflections, muted blues outside and amber inside. Documentary style, 35mm look, slight grain. No people, no readable text on the signage.

Change the subject sentence to your own scene, keep the light and palette sentences as a template, and swap "documentary style, 35mm look" for "flat vector illustration" or "loose ink and watercolour" to see how far the same brief can travel. If the result is too busy, add "simple composition, few objects".

Where results go wrong

• Generic images. The prompt was short. Add the concrete details a photographer would need to recreate the shot.
• Wrong orientation. Orientation was not stated. Say "portrait orientation" or "square" every time.
• Unwanted text or labels. Add "no text" and "no logos". The model is good at text, so it sometimes adds some uninvited.
• Hitting a limit. Plans cap how many images you can make in a period. Plan your drafts, use edits rather than regenerations, and check your plan's quota on OpenAI's help pages.