ChatGPT imagesPrompting basics

ChatGPT Images: Getting Text and Typography Right

By Marco Cavazzana · Updated · 5 min read
ShareEmail
Short answer

To get accurate text in a ChatGPT image, put the exact words in quotation marks, keep them short, describe the type style and where the text sits, and say that nothing else should be written. The GPT Image model renders legible text far better than older image models, and a follow-up message can correct any remaining spelling error.

For years the reliable way to spot an AI image was the writing: signs full of invented letters and menus in no known language. The GPT Image model behind ChatGPT changed that. It renders short, legible text well enough that posters, labels, packaging and social graphics are realistic jobs, provided you prompt for the text deliberately. This guide shows how to specify wording, typographic style and layout, how to fix errors with edits, and when you should still add text in a design tool instead.

The model is built on a language system first, so it understands that the characters in your quotation marks are a string to be reproduced rather than a vague visual texture. It also reads long prompts in full, which means you can describe the words, the typeface feel, the colour and the position in separate sentences and expect all of them to be honoured. It is not perfect: long passages, very small type and unusual spellings still go wrong sometimes. The habits below keep you on the side where it works.

Specifying the words

  1. Quote the exact text. Write: the headline reads "Open Late". The quotation marks signal that this is literal.
  2. Keep it short. One to five words is where accuracy is high. A short sentence is usually fine. A paragraph is a gamble.
  3. Spell unusual words out. For a brand name or a made-up word, add "spelled exactly as written" and, if needed, spell it letter by letter.
  4. Say how many text elements there are. "Two lines of text only: a headline and a smaller subline." This stops the model adding filler text.
  5. Forbid extra writing. "No other text, no watermarks, no small print." Because the model is good at text, it sometimes volunteers some.
  6. Mind the case. If you want capitals, write them in capitals inside the quotes and say "all capitals".

Describing the typography

You cannot pick a font file, but you can describe type in the way a designer would brief a colleague, and the model follows that surprisingly well. Useful vocabulary:

  • Classification. Serif, sans-serif, slab serif, script, monospaced, blackletter, hand-lettered.
  • Weight and width. Bold, light, condensed, extended.
  • Feel. Geometric, humanist, high-contrast editorial, rounded and friendly, industrial stencil.
  • Treatment. Letterpress impression, neon tube, embossed metal, painted on brick, cut paper, chalk on a board.
  • Colour and contrast. "Cream letters on deep green" is clearer than "nice colours".

Describe one typographic idea per text element. "Headline in a bold geometric sans-serif, subline in a light serif" is fine; piling five adjectives on one line produces a muddle. Avoid naming commercial typefaces; describe the qualities instead, which is both safer and more reliable.

Placing text in the layout

Position is the part people forget, and it is why text ends up slapped across a face. Give the text a home: "the headline sits in the empty sky in the top third, centred", "the label wraps around the front of the jar", "the caption runs along the bottom edge in a thin band". Mention the relationship between text and image: "leave clear space around the text", "text does not overlap the figure". If the image is a poster or cover, say so, since the model then composes with text in mind from the start rather than squeezing it in afterwards.

For designs where the text will change later, such as a template for a series of social posts, you can ask for the image with the text area left empty and add the words in your design tool. The ChatGPT image generation guide covers composition cues in more depth.

Settings at a glance

SettingWhat it doesStart with
Quoted textThe literal words to renderOne to five words per element, in quotation marks
Element countHow many pieces of text appearState it, plus "no other text"
Type descriptionClassification, weight, feel, treatmentOne clear idea per element
PlacementWhere the text sits and what it must not overlapA region of the frame plus "clear space around it"
Aspect ratioAny ratio: use the aspect ratio picker or state it in the promptPortrait for posters, square for social
Follow-up editFixes a misspelling or moves text without redoing the image"Correct the headline to read … keep everything else"

Fixing mistakes and knowing when to add text later

When a letter is wrong, do not regenerate from scratch. Reply in the same chat: "The headline currently misspells the word. Correct it to read exactly "Harbour Market" and keep the layout, colours and everything else identical." The model will usually fix the text and preserve the rest. The same works for moving text ("shift the subline down so it does not touch the illustration") and for changing style ("make the headline a condensed sans-serif"). Check the whole image after each fix, because regeneration can nudge other details.

Try it

Create a portrait-orientation poster for a weekend food market. The headline reads "Harbour Market" in bold condensed sans-serif capitals, cream on deep teal, centred in the top third with clear space around it. A smaller subline underneath reads "Saturdays from 9" in a light serif. Below, a simple flat illustration of market stalls with striped awnings by a quayside, limited palette of teal, cream and tomato red. Exactly two lines of text, no other writing.

Change the two quoted strings and the illustration sentence. Keep the structure: one sentence per text element, a placement for each, and the closing line that forbids extra text. If the headline comes out wrong, fix it with a follow-up edit rather than starting again.

When to add text afterwards

Generate the text in ChatGPT when it must look part of the scene: painted on a wall, baked into packaging, lit as a neon sign. Add it in a design tool when it is a clean flat layer that must be pixel-perfect, use a licensed brand typeface, be edited by a client, or run to more than a sentence or two. Many working designers do both: generate the scene with the integrated text, then overlay the small print afterwards.

Frequently asked questions

Can ChatGPT generate images with readable text?
Yes. The GPT Image model renders short text legibly far more reliably than earlier image models. Put the exact words in quotation marks, keep them brief, describe the type style and placement, and state that no other text should appear.
How do I fix a spelling mistake in a ChatGPT image?
Reply in the same conversation and ask it to correct the text to the exact wording while keeping everything else identical. This usually fixes the letters without changing the layout, though it is worth checking the rest of the image afterwards.
Can I choose a specific font in ChatGPT image generation?
You cannot load a font file, but you can describe the type in detail: serif or sans-serif, weight, width, feel and treatment such as neon or letterpress. Describing qualities works better than naming commercial typefaces and avoids licensing questions.

Nothing you makeis lost.

Free · No card · Read-only · 2-minute setup

DreamdriveFree plan · Chrome
Add to Chrome — it’s free