Legible text in AI images — the models, the prompts, the limits

Text used to be the giveaway of an AI image: shop signs in no known alphabet, covers with a title that almost spelled something. The models that joined eroq in October 2026 change that for short lettering: a headline, a name on a label, a word on a sign. Long copy is still a job for layout software. The one-line answer: quote the exact words, keep them short, give them a place and a size, render four takes on a text-strong model, and read every letter. The rest of this tutorial is the detail.

Which models to use for words

Four engines on eroq are positioned for in-image text. Pick by the kind of frame the words live in:

• GPT Image 2.5 Sunburst (OpenAI) is the pick for text: the sharpest in-image lettering of the lineup and a precise, controlled layout. Posters, packaging mockups, UI screens. 31 credits / image.
• GPT Image 2.5 Flare (OpenAI) keeps the same text handling with a bolder, more dramatic rendering. Key art and film posters. 31 credits / image.
• Flux 3 Image (Black Forest Labs) gives reliable typography inside natural, photographic light. Signs, labels and packs in real-looking scenes. 32 credits / image.
• Nano Banana Pro (Google) has the most accurate text of the Nano Banana line on top of its fidelity, and it is the priciest image model on eroq at 124 credits / image. A finisher, not a drafting engine.

The rest of the 13 models are tuned for other jobs (speed, texture, character, anime): use them for the picture and give the lettering to one of these four, or to your editor. When you want no text at all, Anima is the engine with a negative prompt, where "text, watermark" keeps lettering out of an illustration.

Write the words like a typesetter

Quoted text is something the model draws, so brief it the way you would brief a typesetter:

• Quote the exact string. "PAPER MOON BOOKS", in double quotes, spelled exactly as it should appear.
• Type the case you want. Capitals for signs and titles, sentence case for a tagline.
• Name the text's job. A headline, a fascia sign, a label, a button. The job implies a size and a style.
• Give it a place and a size. "Large across the top third", "small, centred at the bottom".
• Describe the look in plain words. Condensed sans-serif, elegant serif, brush script, plus a material if the lettering is physical: painted, engraved, neon, foil.
• Say it once. Repeating the words elsewhere in the prompt invites a second copy of them.

Keep each block to a handful of words, and keep the blocks to one or two per frame. A title plus a short tagline is a poster. Eight labels is a layout job.

What still fails, and what to do instead

• Paragraphs and fine print. Ingredient lists, legal lines, the credits block at the foot of a poster. Prompt for a clean panel and set that copy in a design tool.
• Many separate labels. Dense menus, maps full of names, infographics with a dozen callouts. Cut to a few short labels, or render placeholders and replace them.
• Tiny text in a wide shot. If the words are a few pixels tall, move the camera closer or put them on a bigger surface.
• Steep angles and tricky surfaces. Lettering on a curved edge, a reflective bottle or wet glass is harder to read. Turn the surface toward the camera.
• Rare words and invented names. These are where letters most often slip. Try more takes, or a simpler spelling of a name you are still choosing.
• Your exact logo or label artwork. A model draws its own version of a mark. Composite your real artwork afterwards, or, for a photo of your actual product, shoot it and use Eroq One, which builds the shot from up to ten reference pictures.
• Text in motion. Animate a still with lettering in it and the letters can warp as the camera moves. Keep the move small, or add titles in the edit; animating AI images to video covers the hand-off.

The proofing loop

1. Choose the format first. Where the words sit depends on the frame, and every format costs the same.
2. Render four takes on a text model. That is 124 credits on Sunburst, 128 credits on Flux 3, 496 credits on Nano Banana Pro.
3. Read every letter at full size. Double letters, apostrophes and the last letter of a long word are the usual suspects.
4. None right? Change one thing. Shorter words, capitals, a bigger surface or a plainer type style, then run the batch again.
5. Still stuck after two rounds? Spend one batch on Nano Banana Pro, or set the type yourself. Some jobs are simply faster in an editor.

Failed renders refund themselves and a refused prompt is never charged, but a take with a typo is a finished image. Proofing all four before re-rolling is the cheapest habit in this guide.

Example prompts

1. Poster, Sunburst, 3:4. "A screen-printed travel poster of a red lighthouse on a cliff at dusk, flat colour blocks of navy, coral and cream, the headline "NORTH CAPE" in large cream condensed capitals across the top third, generous margins, subtle paper grain." The headline has a size, a place and a colour, and the margins keep it off the edges.
2. Shopfront, Flux 3, 4:3. "A corner bookshop at dusk with warm light in the windows, a painted fascia sign reading "PAPER MOON BOOKS" in dark green serif capitals on a cream board, a bicycle leaning on the railing, 35mm street photograph." A flat board facing the camera is the best seat in the house for lettering.
3. App screen, Sunburst, 9:16. "A clean smartphone weather app screen, a large "21°" in thin white numerals over a soft blue gradient, the word "SUNNY" below it, a simple sun icon, minimal flat UI mockup." Two text elements with an obvious hierarchy, and nothing else to read.
4. Book cover, Nano Banana Pro, 3:4. "A paperback cover showing a misty salt marsh at dawn with a lone heron, muted greens and pearl grey, the title "THE SALT ROAD" in a tall elegant serif centred in the upper half, the author name "MARA VENN" in small capitals at the bottom." Title and author are the only two blocks a cover needs; everything else stays image.
5. Key art, Flare, 16:9. "Cinematic key art of a red-lit submarine corridor, a lone engineer silhouetted against a glowing hatch, haze and sparks in the air, the title "DEEP SHIFT" in heavy brushed-metal capitals across the lower third." Giving the title a material makes it part of the scene instead of a sticker on top.

Next: the head-to-head in GPT Image 2.5 vs Flux 3 for text, a sign and label library in Flux 3 typography prompts, and the full workflow in the AI movie poster generator.

FAQ

Which AI image model is best for text?

On eroq, GPT Image 2.5 Sunburst is positioned for the sharpest in-image text of the lineup. Nano Banana Pro gives the most accurate text of the Nano Banana line at top fidelity, and Flux 3 Image handles typography in natural-light, photographic scenes.

Why does AI misspell words in images?

The model draws lettering as part of the picture rather than typing it, so every extra letter is another chance to slip. Short, quoted words on a large surface fail least; long strings, rare words and tiny text fail most.

How do I put a lot of text on an AI image?

Do not ask the model for it. Render the image with a clean panel where the copy will go, then set the text in a design tool, where it stays editable.

Do I get credits back for a misspelled image?

No. Refunds cover renders that fail, and a refused prompt is never charged, but a misspelled take is a delivered image. That is why it pays to read every take in a batch before you run another one.

Quote your words, pick a text model and open the image studio.