Why this was unsolvable for years
The image model does not know letters as signs. It learns pixel patterns, and a "g" is not a character to it but a shape that turned up in many forms and sizes across the training images. When you ask it for a caption, what you are really asking is that it draw a pattern resembling a row of letters — not that it write.
Two things changed this. One is that a full language model was put in front of the image models as the text encoder, so the characters of the word in your prompt reach the image model at all. The other is that the vendors deliberately trained the models on images containing captions. That is what made the text legible — but it did not turn it into typesetting.
Where it stands today
There is a measurement for this: in a comparison examining text rendering, the best closed model scores around 77 points and the best open-weight model around 70 on a scale of a hundred. In practice that means the following.
- One or two words, at a large size: reliably good by now. A product box caption, the text on a sign, a shop front will be fine.
- One short sentence: probably fine, but check it. This is where missing or doubled letters start to appear.
- A paragraph's worth of text: not happening. Models typically start to come apart around the two hundred character mark, and it deteriorates quickly from there.
- Lots of small captions at small size: hopeless. A dozen product labels on a shelf stay scrambled on every model.
Other languages are a separate story
Here is the part the vendor guides stay quiet about. According to a multi-language measurement from August 2026, languages written in accented Latin script come off 7 to 11 points worse than English — with the same model, on text of the same length. On a model built specifically for text the difference was English 0.87 against Spanish 0.80; on a general image model, English 0.68 against Spanish 0.57.
English is by some margin the best-supported case, and everything else falls off from there. What we know for certain comes from the vendor documentation: one provider specialising in text states that text rendering is most accurate in English, and that support for non-Latin punctuation and for accented Latin characters is limited. Accented Latin script loses those 7 to 11 points; non-Latin writing systems — Cyrillic, Greek, Arabic, the CJK scripts — fare worse still, because they are rarer in the training material and their character shapes are further from what the model has seen most of. If you are generating a caption in anything other than English, count on every third or fourth image failing on it.
How to ask for it
If you do have the model write it after all, these four tricks improve your odds the most.
- Put it in quotation marks. Every large provider's documentation says the same thing: text that is to be rendered exactly belongs in quotes. This is the single strongest trick.
- Say where it goes and how big. "On the front of the box, in large capitals" — specifying size and position reduces the distortion, because you are not leaving the model to try its luck at small sizes.
- Give a character of typeface, not a typeface. "Heavy grotesque", "serif, classical" work. The name of a specific font does not: the model is not using a font file, it remembers which images that name occurred alongside.
- One caption per image. Two different pieces of text in one picture still gives an order of magnitude worse result than one.
The length at which it falls apart
The two hundred character limit is not a rule but an empirical threshold — and one worth taking seriously. Below it the errors can usually be fixed with one or two regenerations. Above it the failure is of a different kind: it is not one letter slipping, it is the structure of the text breaking down, lines repeating, words getting mixed up.
From which follows the practical rule: the image should not carry the information. If the caption is functional — price, date, contact details, legal text — that is not the model's job. If it is decorative — a shop sign in the background, a word on a packet — then it can go in.
When it is better not to generate
There is a quiet industry admission in this area: the one provider that built its product on text rendering recommends, in its own documentation, that you put the typography onto the image afterwards. That is the most honest sentence in the whole field.
So the practice is two-step. The model makes the image — the composition, the light, the mood, and if needed the empty space for the text — and you put the type on in a proper editor. That is how it comes out sharply legible at every size, how you get to use your own typeface, and how you avoid twenty generations spent hoping it spells your company name right once. For brand elements this is not advice but an obligation: logos and brand assets explains exactly why.
Common misconceptions
- "Just regenerate enough times." For one or two words, yes. For a paragraph, no: there it does not come down to luck but to the model's capability.
- "Every language is the same as English, just different words." No. The measurements put even Spanish noticeably behind English, and languages with rarer characters come off worse still.
- "If I ask for a higher resolution, it will be more legible." Resolution does not help with spelling. You get the same wrong string of letters, only sharper.
- "Fixing it in an editor solves it." Rewriting one letter with inpainting sometimes works, but the model regenerates the surroundings as it goes — after a few attempts the area around the caption is visibly degraded.
What to take away
Captions are the one area of image generation where the right answer is often not to do it with the model. Let the picture be a picture, let the text be text, and let the two meet in an editor. This matters most when you are producing a lot of images with the same brand look — for that workflow, social visuals in series gives you a ready-made recipe.