Does Meta still limit text in ad images?

No. Meta’s help center says there is no longer a limit on the amount of text in an ad image, and its text overlay tool is gone. It still advises that text should not block the visuals, in a clean font at a readable size with strong contrast. Keep text inside the Stories and Reels safe zone: see Facebook and Instagram ad sizes.

Can an AI model use my brand font?

Not reliably. You can describe a style, and Google’s Nano Banana guide even suggests naming a font, but the model draws letters that look like the description; it does not load the font file. Google’s Imagen guide says not to rely on precise font replication. For brand type, set the words in a design tool on the clean photo.

Why does AI put random letters on signs and packaging in the background?

Signs, labels and screens in photos usually carry writing, so models tend to draw letter-like marks on them. Keep backgrounds plain, write "no text anywhere else in the image" on models that follow rules, and check every surface at 100% zoom.

How to get text and headlines right in AI images

Checked September 24, 20267 min read

Short answer

Keep words out of the first render. Make a clean photo with room left for type, then add the headline in a second pass, either with an image model that handles text or in a design tool with your brand fonts. Keep model-made text to a few words in quotes, check every letter at 100% zoom, and repeat the message in the ad copy and alt text, because screen readers cannot read words inside a picture.

Which models handle text best (September 2026)

Google and OpenAI both market their newest image models on text, and both still list text as a limit. On the Arena text-rendering leaderboard, which ranks models by public blind votes, OpenAI’s GPT Image 2.5 Sunburst and Flare were first and second on September 24, 2026 (both marked preliminary), GPT Image 2 was third, and Google’s Nano Banana 2 was ninth. Rankings move month to month, so test with your own headline.

Model	What the maker says about text	The limit the maker states
GPT Image 2.5 Sunburst and Flare (OpenAI, September 2026)	Sunburst is for work where editing precision matters most; Flare is the faster everyday model	OpenAI’s docs: the models "can still struggle with precise text placement and clarity"
GPT Image 2 (OpenAI, April 2026)	The previous OpenAI model, third on the Arena text board	Same as above
Nano Banana Pro (Google, November 2025)	Google’s best model for legible text, "whether you’re looking for a short tagline, or a long paragraph", in several languages	Small text, fine detail and grammar in translated text can still come out wrong
Nano Banana 2 (Google, February 2026)	"Reliable text rendering", per Google’s developer docs	Its model card rates text "poor in small text (often blurry in 1k model), long paragraphs, page length"
Ideogram 4.0 (June 2026)	Ideogram says it has led on text rendering since launch; adds multilingual text and layout control	The open weights carry a non-commercial license; Ideogram says commercial deployments need its license
Recraft V4 (February 2026)	Clear, legible text for infographics, menus, signage and packaging, with a vector (SVG) option	The claim covers short and mid-length phrases
FLUX.2 (Black Forest Labs, November 2025)	Legible fine text for typography, infographics and UI mockups; the [flex] version is tuned for text and detail	Takes no negative prompts, so "no other text" has to be written as what you want

Claims and limits in the makers’ own words where quoted, checked September 24, 2026.

For the wider comparison of image models, see which AI image model is best for product photos.

Render clean, then set the type

On a gummy supplement campaign our agency made in 2026, the jar was rendered alone and the headline was set in a separate pass over the finished photo. It took 3.2 renders per keeper, the fewest of our four campaigns that year, and no render had to spell anything. The brownie job, with motion, took 12, and the surf job, with one model in three locations, took 13.6.

Plan the copy space in the shot list: where the headline goes and how much of the frame it needs. See the shot list template.
Render the photo with that area left plain, and with no words in the prompt that you do not want printed.
Pick the keeper on the photo alone.
Set the type: either as an edit on a model that handles text, with the clean photo as the input, or in a design tool.
Keep the clean file. Other sizes, languages and offers start from it.

Second-pass prompt for a model that handles text

Using INPUT IMAGE 1 as the finished photo, add the headline "[EXACT WORDS]" in [bold white sans-serif], [about one fifth of the frame height], in the [upper left, on the plain wall]. Spell [the brand name] letter by letter: [B-R-A-N-D]. Add no other text. Keep every other part of the image exactly the same.

Keep model-made text short

A headline of a few words is within reach of the models above. A paragraph, a price grid or a legal line is where letters go wrong. Google’s prompt guide for its Imagen models suggests keeping text to 25 characters or fewer, in no more than three phrases. Use that as a starting budget on any model.

Words	Who sets them	Why
Headline of a few words	A text-capable model in a second pass, or a design tool	Short, large, easy to check
Button or call to action	Either	Short, but must match the landing page
Price, offer, dates	Design tool	Must be exact and changes often
Legal lines and disclaimers	Design tool	Must be exact and readable at small sizes
Words on the product’s own label	Neither: they come from the reference photo	Retyping a label invites misspellings. If the label comes out wrong, fix it from the real packshot.
Several languages	Design tool, with native-speaker review	One layered file per market is easier to check

When to set type in a design tool instead

Brand fonts. An image model imitates the look of a typeface from its description. It does not load your font files, so letter shapes and spacing drift.
Exact copy: prices, claims, legal lines, product names with unusual spelling.
Many markets. Translations are easier to check and change on a text layer.
Many sizes. Type usually needs resetting for each placement, and a text layer moves in seconds.
Live text. On a web page or in an email, real text can be read by screen readers and search engines, and it scales.

How to prompt for words

Put the exact words in quotes. Google, OpenAI and Black Forest Labs all give this advice.
Describe the type and where it goes: "bold, white, sans-serif" is Google’s own example, and Black Forest Labs asks for the placement relative to other elements.
Spell tricky words, such as brand names, letter by letter. OpenAI suggests this in its prompting guide.
Use a higher quality setting for small text. OpenAI recommends medium or high quality for small or dense text.
Say "no other text". OpenAI’s own billboard example asks for the text "EXACT, verbatim, no extra characters".
For another language, write the prompt in yours and name the target language for the text. Google documents this for Nano Banana.

Check every letter

Read every word at 100% zoom against the approved copy, letter by letter.
Look for letters added, dropped, doubled or swapped, and for melted letter shapes.
Look for fake writing elsewhere in the frame: on signs, packaging, fabric and props.
Check the spacing and baseline: letters sit on one line and do not collide.
View the ad at the size people will see it, such as a phone screen, and make sure the words still read.
Check the contrast between the words and the photo behind them.
Confirm the type sits inside the placement’s safe zone.
Have someone who did not write the copy read it once more.

Words in pictures and accessibility

Screen readers cannot read words inside an image, and people who need larger text or different colors cannot restyle them. WCAG 2.2, the W3C’s accessibility standard, asks for real text instead of images of text wherever the design allows (success criterion 1.4.5, level AA), with logos counted as an exception. When words must be in the picture, the W3C’s tutorial says the alt text must contain the same words.

Put the offer and the key claim in the ad’s own text fields as well as in the picture.
Write alt text that repeats the headline word for word. See how to write alt text for product images.
In email, set the headline as live text. Mailchimp notes that many email clients turn images off by default, and readers then see the alt text instead.

Questions people also ask

Does Meta still limit text in ad images?: No. Meta’s help center says there is no longer a limit on the amount of text in an ad image, and its text overlay tool is gone. It still advises that text should not block the visuals, in a clean font at a readable size with strong contrast. Keep text inside the Stories and Reels safe zone: see Facebook and Instagram ad sizes.
Can an AI model use my brand font?: Not reliably. You can describe a style, and Google’s Nano Banana guide even suggests naming a font, but the model draws letters that look like the description; it does not load the font file. Google’s Imagen guide says not to rely on precise font replication. For brand type, set the words in a design tool on the clean photo.
Why does AI put random letters on signs and packaging in the background?: Signs, labels and screens in photos usually carry writing, so models tend to draw letter-like marks on them. Keep backgrounds plain, write "no text anywhere else in the image" on models that follow rules, and check every surface at 100% zoom.

Where Overs fits

Overs can set a headline, subline and button onto a finished photo in a separate typography step, and it keeps the clean version, so you can also set the type in your own design tool.

Free for 40 photos a month. The AI that makes the photos is billed separately, on your own key, with no markup from Overs.

Start free How it works

Sources

Checked on September 24, 2026. Prices, specs and rules change; follow the links for the current versions.