Get legible text and typography in images

The best image models now spell short text right most of the time. In this lesson, GPT Image 2.6 Sunburst paints a perfumer's name on the side of a delivery cart from the prompt alone. Then, with a logo that already exists, it paints that exact logo onto a cart with no lettering, using reference images.

Before you start

• Approve the wording first. The model paints exactly what you type, typos included. Get the text signed off before you generate, not after.
• Keep it short. A name, a title, a short tagline. The more words, the more chances one comes out wrong. Long copy, small print and anything legal belongs in a layout tool (see Set the text in code).
• Know your two routes. If the text doesn't exist yet as artwork, write it in the prompt (Steps 1 to 3). If you already have the exact lettering, like a logo, give it to the model as a reference image (Steps 4 to 6).
• Where it lives. Both routes use the image generation panel, under Create. The cost of each run shows on the Generate button.

Step 1: Pick the model and describe the scene

Under Model, pick GPT Image 2.6 Sunburst. In Prompt, describe the scene first: the kind of image, the setting, the light, the subject. The film's prompt starts with the camera and the place (a tracking shot on a mountain road at golden hour), then the cart and its driver.

Step 2: Write the text exactly

The lettering goes at the end of the prompt. This is the part of the film's prompt that sets it, word for word:

On the side of the box, hand-painted lettering in a bold ivory brush script with a long swash, a burgundy shadow and a gold inline reads exactly "Liseron", and the small spaced capitals below it read exactly "PARFUMEUR", with no other text anywhere: no road signs, no number plate, no markings

Three rules make it work:

• Put each piece of text after "reads exactly", in quotes, typed just as it should appear. The model copies what's between the quotes. Capitals, and even a full stop, show up in the picture. "Liseron" comes out in upper and lower case; "PARFUMEUR" in capitals.
• Say how it looks and where it sits. Name the surface ("on the side of the box"), the style ("a bold ivory brush script with a long swash"), the details ("a burgundy shadow and a gold inline") and how the lines relate ("the small spaced capitals below it"). Keep the words themselves short.
• End with "no other text", and name the places it could appear. Models like to add text: road signs, number plates, labels. Listing them ("no road signs, no number plate, no markings") keeps the picture clean, so the only words in it are yours.

Step 3: Generate and read every letter

Click Generate. The result opens in the viewer.

Don't judge the text at the size the viewer opens at. Zoom in and read each word, letter by letter. Here both are right.

Step 4: Add your image and your logo as references

If you already have the exact text, like a logo, and want it on an existing image, don't retype it: add it as a reference.

Open Reference Images and add two images, in this order:

1. Image 1: the cart without lettering, the image you want to keep.
2. Image 2: the logo.

You can drag and drop or paste them, choose them with Select from Library, or Upload them. Each one also appears as a chip under the prompt (@im...), so you can point to it.

The order matters: the prompt calls them image 1 and image 2.

Step 5: Ask for the logo, keep the rest

Keep GPT Image 2.6 Sunburst. In Prompt, ask the model to paint the logo from image 2 on the cart, and to keep everything else the same. The film's prompt:

Paint the logo from image 2 on the side panel of the cart in image 1, as hand-painted sign-writing that follows the panel's angle. Keep every letter, the swash and the colours of the logo exactly as in image 2. Keep everything else in image 1 exactly the same. No other text.

The prompt does four things: it says where the logo goes and how it's applied ("follows the panel's angle"), what must not change in the logo (every letter, the swash, the colors), what must not change in the image, and that no other text should appear.

Click Generate. In the film, this ran with Image Count 1, Width 2,048, Height 1,152, Output Format PNG and Quality High.

Step 6: Check the logo letter for letter

The logo is on the cart, letter for letter, following the panel.

Compare it with the logo, not with your memory of it: the letter shapes, the swash, the colors, the spacing of the capitals. Then check that the rest of image 1 (the driver, the road, the light) didn't change.

Which route to use

Models for text

Whatever the model, read every letter. A text-first model can still add lettering you didn't ask for. On this poster, Ideogram 4.5 got the title and tagline right, and filled the neon sign in the background with invented letters:

That's why the prompt in Step 2 ends with "no other text" and names the places it could appear.

Set the text in code

When the wording must be exact, is long, or changes often, don't ask the model for it. Generate the image with no text and leave room for it, then set the type on top in your layout tool or in code. This is the right route for:

• legal lines, small print and prices;
• translations, where one image needs the same layout in several languages;
• text in video, which a video model can redraw or warp from frame to frame;
• anything that changes often, like a date or an offer.

To leave room, ask for it in the prompt (for example, an empty dark area at the top) and add "no text" so the model doesn't fill it.

Common problems

What's next

Text is one part of a good prompt: Prompt images covers the rest, from subject to constraints. To put lettering to work in a full campaign, see Key art and store assets. Next in this course, Live generation shows how to sketch on a canvas and watch a model render it as you draw.