How do you make images with GPT-4o's new image generation?

Since March 25, OpenAI has been rolling out GPT-4o itself as ChatGPT's default image maker, in place of the separate DALL·E 3 model. OpenAI calls this native image generation: the same model that reads your messages also produces the picture. In practice that means two things for you. It follows long, specific descriptions more closely, and it remembers the conversation, so you can adjust an image by talking to it instead of starting over.

You don't need a special prompt language. Write the way you would brief a graphic designer: what the image is for, what must appear, what the words say, and the style. Expect to wait; detailed images often take up to a minute.

A worked example: a library event poster

Say a public library wants a poster for a free workshop. Here is a first prompt. The event details are illustrative.

Design a portrait poster (2:3) for a public library event.
Style: friendly flat illustration, warm colours, plenty of white space.
Main colour: #1F6FB2 (blue). Accent: #F2A33A (orange).
Illustration: three adults of different ages around a table with laptops.
Text, exactly as written, nothing else:
Top, large: "AI for Beginners"
Middle: "A free hands-on workshop"
Bottom: "Saturday, April 12 · 10 a.m. · Main Branch"
Use a clean sans-serif font. No logos.

A few habits make the difference:

• Quote every word that must appear, and add "nothing else" so the model doesn't invent extra slogans.
• Give the shape (2:3, square, wide) and exact colours as hex codes, the six-character codes designers use. OpenAI says both are understood.
• Keep text short. Three or four lines at readable sizes work far better than a paragraph.
• Describe placement ("top," "bottom right") rather than hoping for a good layout.

Editing by conversation

When the first image arrives, check it like a proofreader: every letter, every date, every colour. Then reply in the same chat with one change per message.

Keep everything the same, but make the bottom line larger and add a little
more space between the illustration and the title.

Change the background to a very light cream. Do not change any text.

Because the model keeps the earlier image and your instructions in view, people and objects tend to stay consistent from one round to the next. That consistency is what makes small series possible, such as the same poster in three colour schemes, or one character across several illustrations.

OpenAI is candid that edits are its weak spot: asking it to fix a single typo may not work and can change other parts of the image or add new mistakes. If that happens, go back to your best version, and restate the full prompt with the correction instead of patching repeatedly. For a printed piece, many people use the image as the background and lay the final text over it in a design program, where they control every character.

Limits and good practice

OpenAI lists several known weaknesses at launch: long images such as posters can be cropped too tightly near the bottom; scenes with many separate objects get confused; text in non-Latin alphabets can come out wrong; and very small, dense text is unreliable. So do a final check of names, dates and spellings on every image you share.

Every image carries C2PA metadata, an industry standard label recording that the picture came from GPT-4o. It helps others check that an image was AI-made. OpenAI also says it blocks certain content and applies stricter rules to images of real people, and it lets artists opt out; its policy also prevents images that directly copy the work of a living artist.

Why is text in images finally readable?

Lettering has long been the giveaway of AI images: a shop sign or a book cover that looks like writing at a glance but spells nothing. The style was right; the letters were wrong.

That is OpenAI's explanation for the change. GPT-4o was trained on images and text jointly, so it learned how images relate to language and to each other, and then went through heavy additional training after that first stage, which OpenAI calls post-training. Because the image comes from the same model that understands your words, the text you quote in the prompt is something it can spell, not just imitate. OpenAI names accurate text rendering as one of the model's main strengths, alongside following detailed instructions and using what was said earlier in the chat.

"Readable" still has limits. Short, large text in English and other Latin-alphabet languages is now generally reliable; long passages, tiny print and other writing systems are not yet. The practical rule follows: use GPT-4o for the words that carry the message, keep them few and large, and proofread every letter before anything goes to print.