Text inside images

Intermediate · 6 min · Decide what text to render and what to typeset yourself.

Product-specific details last reviewed: 2026-09-20

ai images
typography

Ask for a poster and the model will happily draw words — but it draws them as shapes, not as typeset letters from a font. That distinction explains everything: lettering often looks plausible while being quietly wrong, and the current OpenAI image guide describes text rendering as significantly improved while still listing precise text placement and clarity as a limitation. The skill is not avoiding text; it is deciding which words are decoration and which are information.

What renders well, what wobbles

Strong case: short display words — a few characters, large, bold, high contrast against a calm background. “OPEN”, “SOLE”, a two-word title.
Weak case: anything long or exact — paragraphs of copy, price lists, addresses, URLs, phone numbers, dates. The more characters, the more chances to wobble; the more exactness required, the less plausibility serves you.
Stylistic trap: scripts, handwriting and decorative faces thin out the letterforms — clarity is the first thing they lose.
Silent failure: invented lettering. “COFFE”, “HARVEST FAIRR” — a swapped or doubled letter that reads almost-right at a glance. It is the visual cousin of a hallucination: plausible-looking detail that is simply wrong.

Weak prompt

a chalkboard menu with twelve dishes and their prices listed clearly

Stronger prompt

a chalkboard menu with “TODAY'S SPECIAL” in large chalk capitals at the top, generous blank space below

Why the stronger one works

The weak prompt asks for twelve exact strings at small size — precision the model cannot reliably deliver, and every wobble is a wrong price.
The strong prompt uses the case these models handle best: short, large, high-contrast words.
The blank space is deliberate — it is where your own typesetting will live.

A bad example

Ask: “A poster with all the event details: address, website, ticket prices, opening times.”

Four exact strings, none of them large: expect a plausible-looking result with at least one silent error, and no way to fake-check it at a glance.

A better example

Ask: “A poster with the headline ‘HARVEST FAIR’ in large bold letters at the top, plenty of clear space below.”

The strong case, used on purpose: short and high-contrast in the render, everything exact later in a layout tool.

The safe workflow

How it works, step by step

Write the exact string first — capitalisation, punctuation, accents — before you write anything else in the prompt.
Keep it as short as the design can bear. Display words over sentences; if it needs three lines, it belongs in your layout tool.
Give it room: large type, high contrast, and calm imagery directly behind the letters.
Generate, then inspect character by character. Invented lettering hides in plain sight — read the words slowly, out loud if you must.
If the text must be perfect — brand names, URLs, prices, legal lines — leave space in the image and typeset the text yourself afterwards.

Why it works

A generative system produces letterforms the way it produces leaves: as patterns that match the look of writing, sampled from noise until they fit the prompt. Plausibility is the model’s strength; exactness is not. So the durable rule is about stakes, not avoidance: decorative words can tolerate a model’s hand, and information-bearing text — names, numbers, addresses — gets typeset by you, in place, afterwards.

A café wants two designs: a window sign reading “OPEN”, and a chalkboard with a full price list. How do you split the work?

Ask the model for both — current models handle paragraphs of text fine now Let the model render the “OPEN” sign; typeset the price list yourself afterwards Avoid text in generated images entirely — it never renders correctly

Practice

Spot the text trouble

A neighbourhood fair asked for a poster for its harvest celebration. You got the image back along with a short summary of what it contains — and the printer is waiting for a decision on the artwork.

Select every element you should not send to print as-is — the parts that need to be fixed, replaced or typeset yourself.

The AI answer

Here is your poster! The headline “HARVEST FAIR” sits at the top in large bold capitals. Below it, a friendly line of body copy reads “Join us Saturday from 10am to 4pm in the town square for fresh produce, live music and family activities.” Down in the corner, a small badge shows the address “12 Mill Lane” and the website “harvestfair.local/visit”. A decorative script banner at the bottom reads “See you there!”, and behind everything there is a background of autumn leaves and bunting.

Hint

Two requirements make rendered text hard: length (many characters) and exactness (zero tolerance for error). Flag where both appear.

Large and short is the strong case; small, long or script is the weak one — and anything a customer will act on must be exact.

Transfer

Quote cards and social graphics: one short line in the render is fine — check every character once.
Posters and menus: headline in the render, body copy and prices in your editor.
Packaging and labels: treat rendered text as a mockup placeholder; final artwork is typeset.
Proofreading habit: when a render contains readable text, read it letter by letter — invented lettering is built to pass a glance.

Key takeaways

Models draw letterforms as shapes: plausible is normal, exact is not.
Short, large, high-contrast words are the strong case; long or exact strings are not.
Decide by stakes: decorative text can ride along, information text gets typeset by you.

Next: Iterate and edit — refine instead of regenerate. The hallucination entry covers the pattern invented lettering belongs to.