Technique · AI content creation

Text and typography in AI images.

Image models render short words far better than they used to, but generated text is still the wrong place for anything that must be exact: brand typefaces, prices, legal lines and translations. The reliable method is to generate the image with room for copy, then set the words as real type in layout, where they can be proofread, edited, localised and read by screen readers.

Why models struggle with text

Image models learn what letters look like, not how words are spelled, so they reproduce the shape of text more reliably than its content. Errors grow with length, small sizes, unusual words and scripts the model saw less of, such as Arabic, Thai or Devanagari. Brand typefaces come out as near misses: the right mood, the wrong letterforms, spacing and weights. The most dangerous errors are subtle, one wrong letter in a long word, because they survive a quick review and get noticed by customers instead.

When generated text is acceptable

Incidental text inside a scene, such as a shop sign in the background or the spine of a book, can be generated, provided it is proofread letter by letter and does not spell out a real company's name. Hand-lettered words that are part of an illustration, and placeholder text in concepts and storyboards, are also reasonable. Anything a customer is meant to read and rely on, such as an offer, a price, an ingredient or a date, belongs in live type instead.

Live type for everything that matters

Compose images with deliberate empty space for the headline and supers, then set the copy in a layout tool with the brand's licensed fonts. Prices, legal lines and claims stay exact and can be updated without touching the image. Localisation becomes one image with many language layers instead of one image per language. Accessibility improves too: screen readers cannot reliably read text inside an image unless it is repeated in alternative text, and WCAG 2.2 success criterion 1.4.5, Images of Text (Level AA), asks for real text wherever the presentation allows it, with meaningful alternative text for the rest.

Labels, packaging and lookalike brands

Product labels must be exactly right, so they should come from the product photograph, not be redrawn by a model. Synthetic White's product lock keeps the real product, label included, pixel-accurate while the scene around it changes. For packaging concepts, treat any generated label text as placeholder and set the real copy in design. Check backgrounds too: models sometimes invent signage or logos that resemble real brands, which is a trademark problem waiting to happen.

Updated 26 October 2026

Questions

Text and typography in AI images, answered.

Can AI write text inside images correctly?

Short, common words often come out right now, but accuracy drops with longer phrases, small type, unusual words and non-Latin scripts, and brand typefaces are never reproduced exactly. Use generated text for incidental details at most, and set real copy as type.

Why does AI misspell words in images?

Image models learn the visual shapes of letters and words rather than spelling, so they produce text that looks plausible at a glance. Errors are often a single letter in a longer word, which is why every piece of generated text needs proofreading.

How do we localise images that contain text?

Generate a text-free master image with space reserved for copy, then add the words as separate layers per language. One approved image then serves every market, and translations can be corrected without generating anything again.

Get started

See your AI studio generating this week.

Thirty minutes. A live studio in your colours, your brand hub loaded, a real campaign brief.

See plans