Getting text right in an image, and knowing when to stop trying

Recent models produce readable words fairly reliably, which is a genuine change and has convinced a lot of people that generated typography is a solved problem. It is not.

Spelling a word correctly and setting type are different skills. Generated text has inconsistent letter spacing, drifting baselines, mismatched weights within a single line, and no relationship to any real typeface. It reads as text and looks, to anyone who works with type, obviously wrong.

For a background sign in a busy frame, that is fine. For anything a client will look at directly, it is not — and the fix is almost never a better prompt.

Generated text succeeds when it is:

• Short — one or two words. Reliability falls off sharply past that.
• Large — occupying a real portion of the frame. Small text becomes noise.
• Isolated — on a clean surface, not wrapped around an object or in perspective.
• Unimportant — nobody will read it closely or compare it to a brand standard.

Ambient environmental text — a shop sign in the distance, a word on a wall — meets all four and works well. A product label meets none of them.

If your case fails any condition, skip to the fallback rather than spending an afternoon confirming it.

Put the exact string in quotation marks and say nothing else about its content. Models handle a quoted literal better than a described one, and adding adjectives about the text invites the model to reinterpret it.

A matte black sign on a bone wall reading "OPEN",
large sans-serif letters, evenly spaced, flat to camera.

Then describe the physical surface the text sits on and the angle it is viewed from. Text flat to camera on a plain surface is dramatically more reliable than text in perspective or on a curve.

Asking for a specific typeface by name rarely works — the model has no font library — but asking for a category does. 'Heavy grotesk', 'condensed sans', 'serif with high contrast' shift the letterforms in the right direction.

Generate the image with a deliberate empty area where the text belongs — a plain wall, a blank sign, a clear lower third — and set the type in a design tool.

Everything improves at once: the typeface is real and correct, the spacing is controlled, the text is editable, and it can be produced in five languages without regenerating anything.

A blank matte sign mounted on a bone wall, no text,
no markings, flat to camera, evenly lit.

Ask explicitly for no text. Models will helpfully invent some on any surface that looks like it should have writing, and removing it afterwards is more work than preventing it.

A client's wordmark has a defined typeface, defined spacing and a defined clear space, and they have a document that specifies all three. A generated approximation violates every line of it.

It will also be immediately obvious to them, in a way that undermines confidence in everything else in the frame. A slightly wrong logo reads as carelessness even when the rest of the work is excellent.

Generate the surface, composite the supplied asset. Ask for the vector file at the start of the job; it takes one email and removes the problem entirely.

Text that carries meaning — a headline, a product name, a price, a brand — gets set properly in a design tool over a generated plate.

Text that is texture — a distant sign, a word on a passing van, writing on a background object — can be generated and usually should be, because setting it would cost more than it is worth.

Deciding which category you are in takes a second and saves the afternoon. Most of the frustration with generated typography comes from treating case one as case two and hoping the model closes the gap.