Translating a Word doc is a five-minute job. Translating the words chalked onto a café menu in a photo is another story. For years you had two bad options. You could dig up the source files and rebuild the whole design, or drop a flat box over the old text and type on top. One costs a designer’s afternoon. The other looks like a ransom note.
That’s changed. Tools like nano-banana-pro.io can swap out text in a photo and keep the chalk dust, the curve of the bottle and the glare on the shop window. So a small team can now localize a campaign for four markets before lunch. The catch is that a menu can look flawless and still list the wrong price.
First, is this even a job for AI?
Sometimes not. AI text replacement shines when the words are part of the picture: a menu photographed on a table, a cereal box shot at an angle, a sign in the background of a street scene, or a thumbnail that only exists as a flat PNG.
It’s the wrong tool for anything a regulator might read. Nutrition panels, ingredient lists, allergen warnings and safety labels belong in the original design file, checked by someone whose name is on the line. To an image model, “milk, nuts” and “milk nuts” are just two slightly different shapes. A missing comma there can send someone to the hospital.
Meet Lucía and her chalkboard
Let’s make this concrete. Picture Lucía, who handles marketing for a small coffee chain about to open in Madrid. Her hero image is a chalkboard menu, and it needs to speak Spanish by Friday.
The worst thing she could do is open the editor straight away. Instead, she grabs a notepad and writes down every bit of text on that board. That means the headings, the drinks, the prices, the tiny “oat milk +50¢” squeezed into a corner, the Instagram handle, even the “Open 7 am” half-hidden behind a cup. She numbers each line and notes where it sits.
That boring list ends up doing double duty. Today it’s the translation brief. On Thursday it’s the checklist she reviews against.
Then she goes through it line by line. Translate, keep, or cut? The brand name stays as is, and so does the trademarked tagline. Prices and sizes don’t get translated so much as converted. Dollars become euros, and a 12 oz cup turns into roughly 355 ml.
Why a human picks the words
Machine translation will hand you a draft in seconds, and for a blog post that draft is often fine. Menus are different. They’re full of tiny phrases with no context around them, like “cold brew” or “limited edition,” and that’s exactly where automatic translation goes stiff or weird. The output reads smoothly, which makes the mistakes harder to spot. That trade-off is unpacked well in this look at AI translation’s hidden risks.
So Lucía gets her wording from a translator, or at least a colleague who actually speaks the language, before generating a single pixel. She also keeps a little glossary for repeat terms such as drink names, menu sections and “order here.” Across twenty images, sounding consistent beats finding the perfect phrase for one.
The text won’t fit, and that’s normal
Translations rarely fit in the original space. German and French usually run longer than English, sometimes by a lot. Japanese and Chinese use fewer characters but need bigger type to stay readable. Arabic and Hebrew flip the whole thing, reading right to left, which can change how a sign should be laid out at all.
Decide how to handle overflow before the model decides for you. Trim the copy with your translator, accept a smaller font, or let a line wrap. Leave it to the AI, and it’ll often squash the letters until they look like they’ve been through a trash compactor.
Talk to the editor like a production manager
“Translate this menu into Spanish” is a coin toss. Exact strings, exact positions, and a clear do-not-touch list get you something usable. Here’s roughly what Lucía types:
Replace the menu heading “Coffee” with “Café” and the line “Cold Brew $4.50” with “Café frío 4,50 €”. Keep the chalk lettering style, the board’s texture and lighting, and the position of every line. Do not change any other text or objects.
Spot the comma in “4,50”? That’s how Spain writes decimals, and it’s baked into the instruction rather than left to chance.
She also works in small bites, two or three lines per pass. If something goes sideways, she knows which edit did it. This is part of a bigger shift toward editing replacing one-shot generation, where you refine a real image instead of rerolling a fresh one.
And before committing to the full set, she does one board as a dry run. If that single test looks shaky, she’s saved herself nineteen more headaches.
The step everyone wants to skip
Proofreading. Specifically, reading the output one letter at a time against the approved translation. Skimming word by word misses things.
Image models trip up in fairly predictable spots. Accents vanish, so “café” becomes “cafe” and “año” turns into something far less polite. Non-Latin characters can look entirely believable and simply not exist. Decimal points and currency symbols wander. Every so often, a word you never mentioned gets quietly rewritten. Some letters that looked crisp at full size blur into the background once the image shrinks to phone width.
Then Lucía hands the finished image to someone who grew up speaking Spanish and asks them to look at it on a phone, at the size customers will see. A native reader catches a clunky phrase in two seconds. A non-speaker could stare for a minute and see nothing wrong. If the script is one you can’t read yourself, skipping this review isn’t an option.
Words are only half the localization
A Madrid storefront with a US license plate at the curb still looks American, however good the Spanish on the sign. Dates, phone formats, units, and prices shift from market to market. So can what a color or symbol means.
The trick is to stay deliberate about it. If the brief only says “translate the text,” leave the scene alone. If you do swap the plate or the plug socket, write it down so the reviewer knows to check those too.
Boring records save you next spring
For every image, Lucía keeps five things: the original file, the approved translation, the exact instruction she used, the final export, and the name of whoever signed it off. Files get names like menu-board_es-ES_v2.
It feels like paperwork until the cold brew goes up 30 cents. Then she can update Madrid, Lisbon and Lyon in one sitting instead of guessing which file is current.
One last look before you hit publish
Run through it once more. Does every visible string have an approved translation? Are the brand names untouched? Do prices, units, and dates match the local format? Has a fluent reader signed off on the accents and any non-Latin characters? Is the text you didn’t want changed still identical? Was all legal, safety, and ingredient copy typeset in the source file rather than generated? And have you actually looked at the final image at its real size?
If all of that checks out, ship it.
So, is it worth it?
Yes, with guardrails. AI editors really do make it practical to translate the text inside photos without rebuilding every design from scratch. The danger is an image that looks perfect and reads wrong. Let a person choose the words, give the model exact strings and firm limits, and put a native reader between the editor and the publish button. Done that way, one set of visuals can work in several markets without losing its polish or its accuracy.
For more, visit: AIInsightsNews
