Getting Text Right Inside an Image

READ6 min

WORDS1,101

SECTIONS4

TYPEGUIDE

CHECKED31 AUG 26

Text in generated images used to be a punchline. It mostly works now — and the ways it still fails are specific enough to plan around.

TL;DR — THE SHORT VERSION

It got fixed deliberately. Text rendering is now a named feature vendors build and advertise, not an accident of scale.
Short is reliable, long is not. A few words on a sign usually work; a paragraph on a packet usually does not.
Quote the exact string. Put the words in quotation marks and say where they go — describing them is how you get near-misses.
The failures are predictable: extra letters, plausible-but-wrong spelling, and drift in the second and third lines.
For anything that must be correct, do not generate the text at all. Generate the image, set the type on top.
Always read it out loud before publishing. Your brain autocorrects text in images, which is exactly why these errors ship.

IN PLAIN ENGLISH

The model is not typing. It is drawing shapes that look like letters, the same way it draws a hand — and it is drawing them from having seen a great many pictures of words.

That is why it can get a familiar word perfect and an unusual one almost right. A brand name it has seen a million times is a shape it knows. Your client's name is not.

What works now

Text became a feature, not an accident

The improvement was engineered rather than emergent. Vendors now list text rendering as a headline capability — OpenAI's ChatGPT Images 2.0, released in April 2026, was announced with improved text rendering and multilingual support among its named improvements.OpenAI, "Introducing ChatGPT Images 2.0" (openai.com), released 21 Apr 2026; improved text rendering and multilingual text listed among the model's stated capabilities. Which model is best at text changes with every release, so this page names a documented example rather than ranking models. Checked 31 Aug 2026

The practical consequence: stop assuming it cannot do text. The assumption is a year out of date, and it makes people rule out approaches that now work.

Length is the variable that matters

Reliability falls off with the amount of text, sharply.

One to three words — a sign, a logo, a label. Usually correct.RELIABLE

A short line — a tagline, a headline. Often correct, worth checking.MOSTLY

Several lines — packaging, a poster with body copy. Expect errors in the later lines.RISKY

A paragraph — a page of a book, a form, a screen of code. Expect gibberish that looks convincing at a glance.NO

The rule that falls out of it: if the words matter more than the picture, this is the wrong tool for the words.

How to ask for it

Quote it, place it, describe it

Three parts, and most failures are a missing one.

Quote the exact string. Write the sign reads "OPEN LATE" rather than a sign advertising late opening. The second invites the model to invent wording, and it will.

Say where it sits. On the awning, across the top third, on the front of the packet. Unplaced text lands wherever there is room, which is often across your subject.

Describe the lettering as an object. Painted, embossed, neon, vinyl, chalk — plus a rough style. This anchors the rendering to a physical thing rather than a font idea.

Ask for less text than you want. Two words rendered perfectly beats six rendered nearly.

The failure modes, and what each means

An extra or doubled letter. The commonest failure. Usually fixed by a re-roll rather than a rewrite.RE-ROLL

A plausible wrong word. It rendered a word it knows instead of yours. Quote it more explicitly, or shorten it.QUOTE IT

Line two degrades. Reliability drops as it goes. Cut to one line, or set the rest as real type.CUT IT

Right word, wrong place. You did not say where. Placement is a separate instruction.PLACE IT

Invented extra text. The scene implies signage, so it fills it. Say what the other surfaces read, or say they are blank.BE EXPLICIT

When not to generate the text at all

For anything that has to be correct, generate a clean image and set the type on top. A price, a phone number, a legal line, a brand name, a URL.

This is not a workaround. It is better in every dimension: the text is selectable, searchable, translatable, readable by a screen reader, correctable without regenerating, and typographically controlled. It also cannot be subtly wrong.

Prompt for the space rather than the words — "clean unmarked awning with room for a sign", "plain packet front, no lettering" — then place real type in the gap. Is your AI output usable? covers why baked-in text is an accessibility problem as well as a production one.

THE CHECK EVERYONE SKIPS

Read the text out loud, letter by letter, before it ships.

Reading silently does not work. Your brain corrects near-miss words automatically — which is precisely why misspellings in generated images get published by the person who made them and spotted by everyone else. Spelling it out loud defeats the autocorrect.

Before you publish an image with words in it

1 — Did I put the exact string in quotation marks?□

2 — Did I say where it goes and what it is made of?□

3 — Is this the shortest version of the text that does the job?□

4 — Has anything correctness-critical been set as real type instead?□

5 — Have I read every word aloud, including background signage I did not ask for?□

6 — If it is a brand name, does it match the real spelling and casing exactly?□

SOURCES AND HONESTY ABOUT THEM

OpenAI, "Introducing ChatGPT Images 2.0", released 21 April 2026 — improved text rendering and multilingual support are the vendor's own stated capabilities. Checked 31 August 2026.

This page ranks no models and quotes no accuracy rate. Which model renders text best changes with each release, and a league table here would be stale within weeks. The failure modes and the fixes are the durable part — they follow from the fact that the model is drawing letterforms rather than setting type, and that has not changed. The prompting advice is craft, reasoned from those failure modes rather than measured in a controlled test.

The through-line: it can draw words now, but it is still drawing them. Anything that has to be right belongs in real type on top of the picture.