Ask for a shop sign, a product label or a poster headline and get lettering you can actually read. Compare the models the catalog describes as text-capable, learn how to word the prompt, and proofread the result before it goes anywhere near a customer.
Anyone who has asked an image model for a shop sign and received a storefront reading BAKRY knows the problem. Garbled lettering was the giveaway of AI images for years. Newer models spell short strings correctly far more often, and a few are described in the catalog as built for it. This page covers which models to try on Picasso IA, how to word the prompt, and where the letters still fall apart.
An image model does not spell. It learned what letters look like, the way it learned what leaves look like, and it paints shapes that resemble them. Most of the time the shapes are close enough that your eye fills the gap, but lettering is the one thing people read exactly. One swapped letter in the middle of a word and the whole picture looks wrong, even when the lighting and composition are perfect.
That is why text inside images mattered enough for model makers to work on it directly. A poster, a menu board, a product label, a book cover or a storefront mockup is only useful if the words are right. Before, you generated the picture and added the type afterwards in a layout tool. With text-capable models you can get the words into the image in one pass, which is faster for drafts, social graphics and concept art, although not always right for final artwork.
If your goal is a full poster, the AI poster design page covers the layout side, and AI book cover design covers titles on covers. This page stays focused on one question: how to get the letters themselves right.
Picasso IA lists 502 models, and 212 of them are text-to-image. Only some of those listings say anything about lettering, so here is what the catalog actually states for the six models worth trying first.
Ideogram v4 Balanced is the clearest case. Its listing says it renders words, labels and headlines inside generated images, and names posters, banners and quote cards among its uses. GPT Image 2.5 Flare is the other explicit one: the listing describes sharp on-image text for posters, labels and social graphics, and says it can also edit an existing photo from a reference image.
Ideogram v4 Quality sits in the same family. Its listing does not spell out a text claim, but the example prompts attached to it put quoted words into the picture, so it is a sensible second test. Recraft v4.1 Pro, Seedream 5 Pro and Nano Banana Pro are strong image models, with print-size output, long prompts and reference images respectively, but their catalog descriptions make no statement about lettering. Treat them as candidates you test yourself, not as promises.
The honest way to choose is to run the same short phrase through three of them and read the results. The full catalog lets you filter further, and the collection browses by category.
Not all text is equally hard. A single word on a sign and a menu with eight lines are different jobs, and the table below groups the common cases by how to word them and what usually goes wrong first.
| Kind of text | How to write it in the prompt | What tends to break first |
|---|---|---|
| One word on a sign or storefront | Quote the word and name the surface: a hand-painted wooden sign that reads "OPEN" | A swapped or doubled letter in the middle of the word |
| Short slogan on a poster | Quote it, say where it sits (top third, centered) and how large it is | The last word drifts into something invented |
| Product label | Use a made-up brand name in quotes and describe the label as flat, front-facing and well lit | Small secondary lines turn into lookalike squiggles |
| Headline plus subhead | Name the hierarchy: large bold headline, smaller subhead below it | The subhead breaks before the headline does |
| Menu, schedule or list | One item per line, few items, quoted exactly | Rows merge and digits swap places |
Read the right column as a warning order, not a verdict. The same model that nails a three-word slogan may fail the subhead under it, which is why the steps below end with proofreading.
This is the routine that works whichever model you pick. It assumes you already know the words you need, because the single best predictor of clean lettering is a short, settled string.
Short strings win. Three words are safer than six, and six are safer than a sentence. If a design needs more copy than that, generate the picture with a clear empty area and add the rest in a layout tool. Models also follow a quoted phrase much better than text that is merely implied, so write the sign that reads "FRESH BREAD" instead of a bakery sign that says fresh bread.
Capitals help on signs and labels because the letterforms are simpler and more uniform. Describe one style of lettering, such as bold sans-serif or hand-painted brush, rather than several. Give the text a calm background: lettering over a busy photo or a strong pattern breaks more often than lettering on a plain wall or a flat label.
Proofread every letter, always. A word can look right at a glance and still carry a missing letter or a doubled one, so read it backwards, slowly, before the image is used anywhere.
For wording that must stay consistent across a set, such as a series of posters, keep the quoted text and the style description identical between runs and change only the scene. If you are weighing a dedicated text-first tool against a catalog, the Picasso IA vs Ideogram comparison walks through the trade.
Better does not mean solved. These are the places where it still stops working, and where you should plan a different route from the start:
None of that makes the tool less useful for what it does well: short, bold, quoted lettering in a picture. It just marks the point where a designer's layout software takes over. If the image shows something you are selling, keep the lettering truthful to the actual product and check the platform's current requirements before you publish.
None spell perfectly, but some are described as built for it. On Picasso IA the catalog listings for Ideogram v4 Balanced and GPT Image 2.5 Flare both say they render legible text inside images, so start there. Ideogram v4 Quality, Recraft v4.1 Pro, Seedream 5 Pro and Nano Banana Pro are worth a test on the same prompt, although their listings make no text claim. Run the same short phrase through a few and keep whichever spells it right.
Yes, for short strings. Write the words in quotation marks inside your prompt, say what surface they appear on, and generate in the toolkit with a text-capable model. Check the result letter by letter, because even the better models occasionally swap or add a letter. For anything long, or anything with legal or safety weight, set the words in a layout tool and use the generated picture only as the background.
Because the model paints what letters look like instead of spelling them from a dictionary. It learned letterforms from images, so a word is a pattern of shapes, and a small slip in one shape changes the word without looking wrong to the model. Models built with text in mind handle this much better on short strings, but the weakness shows again whenever the text gets long, small or crowded.
There is no fixed number, and it depends on the model, the font size and the background. As a working rule, one to three words is the safe zone, a short slogan is a fair bet, and a full sentence starts to fail. Several lines of body copy almost never survive. If the design needs more, generate the image with space left for the copy and add the text afterwards.
Put the exact words in quotation marks and name where they go: a hand-painted sign above the door that reads "FRESH BREAD". Describe the style once, such as bold capitals or brush lettering, and mention strong contrast with the background. Avoid paraphrasing, because a model asked for a sign about bread will invent its own words. Quoted text in capitals is usually the most reliable form to start with.
Treat it as unproven until you test it. The catalog listings for these models say nothing about non-Latin scripts, and joined or stacked letterforms are harder to draw than separate Latin capitals. Try a very short sample, have someone who reads the language check every character, and plan to add the real text in a layout tool if the sample fails. Never publish non-Latin lettering that nobody fluent has read.
Sometimes. You can attach the image to a model that accepts reference images, such as GPT Image 2.5 Flare or Nano Banana Pro, and ask for only the lettering to change. Results vary, and the model may redraw more than the word you named, so keep the original. If the edit fails, shortening the phrase and generating again, or fixing the word in an ordinary image editor, is often quicker.
Use it for concepts and mockups, not as final packaging. Real labels carry legal, safety and ingredient information that must be exact, and an image model can silently change a letter or a digit. If the picture shows a product you actually sell, the label in it must match the real item. Set final label text in a layout tool, have a person proofread it, and check the platform's current requirements.
Not as a separate charge that this page can quote. Cost depends on the model and the settings you choose, and text-capable models are priced like any other model in the catalog. Several attempts on a stubborn phrase will add up, so keep the string short to save retries. New accounts get free credits to start, and the pricing page has the current plans.
Use a design tool when the layout is mostly text, such as a flyer with a schedule, a menu or a certificate. Use an image model when the picture is the star and the lettering is a short line inside it: a sign in a scene, a headline on a poster, a label on a bottle. Many projects use both, with the model making the artwork and the tool setting the long copy.
Pick a phrase of three words, quote it, and try it in the Picasso IA toolkit on two different models to see which one spells it right.
Choose your language
