Take a product shot, ad or poster with words in one language and get the same design back in another. An image editor rewrites the text in place while the layout, photo and colors stay put, and a language model drafts the translation first. You still proofread every line before it goes live.
Most marketing images carry words baked into the pixels: a headline on an ad, a flavor name on a package, a sale tag on a poster. When the audience speaks another language, that text is the part nobody can edit, and rebuilding the design from scratch for every market is slow. Image editors can now rewrite those words in place, so the photo, the layout and the colors stay as they were.
A translated caption under an untouched image is a half measure. If the packaging, the banner or the poster still shows the original language, the reader sees a foreign product with a note attached. Translating the text that sits inside the picture makes the asset look made for that market, which is the whole point of localizing.
The traditional route is a designer with the source file, a font match and an hour per language. That works well when the file exists. It falls apart when you only have the exported JPG, which is the usual situation for a supplier photo, an old campaign or a screenshot of your own ad. An image editor needs only the picture and an instruction.
Picasso IA runs 502 models in one workspace, so the two jobs live side by side: a language model drafts and checks the translation, and an image model puts it into the picture. Both run in the toolkit, and new accounts get free credits to see whether a given image works before you spend anything.
Two different skills are involved, and they should not be mixed up. Translating is a language task, so give it to a language model such as claude-sonnet-5 or gemini-3.5-flash, both in the large language model category of the catalog. Rewriting pixels is an image task, so give it to an editor: google/nano-banana-pro accepts up to 14 reference images with a prompt, openai/gpt-image-2.5-flare edits an existing photo from a reference and is described as rendering on-image text cleanly, and picassoia/picassoia-image-editor-pro edits from a plain-language instruction with up to three reference images.
| Approach | How it works | Best for | Main risk |
|---|---|---|---|
| Edit in place | Upload the original and ask the editor to replace the words with a translation you supply | Product shots, packaging, ad banners | Fonts drift slightly from the original |
| Regenerate from a brief | Describe the design and the new text, generate fresh | Posters you no longer have a source for | The layout will not match the original |
| Edit, then upscale | Run the in-place edit, then upscale the result | Small text on high-resolution print files | Upscaling can soften thin letters |
| Cover and retype | Remove the old text, then set the new words in a design tool | Legal lines, price tags, exact brand fonts | Needs a design tool and a few minutes |
| Batch per language | Reuse one edit prompt, swap only the translation | A catalog of images across several markets | Errors repeat when one mistake gets copied |
The first row is what most people want, and the page walks through it below. The fourth row is not a failure of the tool: for text that must be exact, a human typing it is still the safest path.
This is the loop that ends with an image you can actually publish. It assumes you have the original picture and a target language.
The prompt decides whether you get a clean swap or a redesign. Name what changes and what must stay: replace the headline text with a new line, keep the font style, color, position, background and product exactly as they are. Image editors respond well to that split, because the second half tells them not to get creative.
Give the replacement text in quotation marks and one line at a time, for example: replace the top line with "Livraison gratuite" and the bottom line with "Offre limitée". Quoted strings are copied far more reliably than words buried in a sentence. If an edit changes something you did not ask for, rerun it with the unwanted change named and the instruction to keep it unchanged.
Translations are usually longer or shorter than the original. German and Portuguese often run long, so a tight banner may need a shorter wording. Ask the language model for two or three versions at different lengths before you touch the image.
Run each language as its own edit from the original file, not from the previous translation. Editing an edit stacks small drifts, and after two rounds a font that looked fine starts to look off. The static ad creatives page covers building the original in the first place.
The most common slip is trusting a literal translation. A slogan that rhymes in one language is a plain sentence in another, and a pun usually dies on the way. Ask the language model for the meaning and the tone of the line, and for natural options in the target market, not a word-for-word version.
The second is skipping the character-level check. Image models can swap a letter, drop an accent or turn a word into something that looks right at thumbnail size and reads wrong at full size. Zoom to 100 percent and read every line, including the small ones.
The third is forgetting what else is on the image. A price, a unit, a date format or a size chart may need conversion, not translation, and that is a business decision. If the picture is a product photo, it must still show the item as it really is: changing the words is fine, changing the product is not. The product photography page covers that line.
An honest list of limits saves you a bad launch. Here is where this approach needs a human, or a different tool.
If any of those apply, use the AI edit as a draft for the layout and bring in a person for the finish. That still saves most of the work.
Yes, in two steps. A language model translates the words, and an image editor rewrites them in the picture while keeping the layout and photo. Picasso IA has both kinds of model in the toolkit, so you can do the translation and the edit in one place. The result still needs a human check, especially for small text or complex scripts, because image models can misspell a letter without any warning.
Image editors that take a reference image and an instruction can do it. In the catalog, google/nano-banana-pro, openai/gpt-image-2.5-flare and picassoia/picassoia-image-editor-pro all accept reference images alongside a prompt. Flare is described as rendering on-image text cleanly. Try the same edit on two of them, because each one handles fonts and spacing a little differently, and keep the one whose letters match the original best.
Use a language model, not the image editor. Both anthropic/claude-sonnet-5 and google/gemini-3.5-flash are in the large language model category of the catalog. Give the model the context, such as headline versus flavor name, the tone and a character limit, and ask for a couple of options. The language model is better at wording, and the image model is better at pixels, so each should do only its own job.
Mostly, but not perfectly. Editors try to match the typeface, color, size and position of the old text, and on clean, flat lettering they often land close. On decorative fonts, curved text or tight layouts you may see small differences. Always compare the result against the original at full size, and if the match matters for a brand, retype the text in a design tool instead.
It can try, but treat the output as a draft. Right-to-left direction, joined Arabic letters and Hindi conjuncts are easy for an image model to get subtly wrong, in ways a non-reader cannot spot. Have a native reader check every line before publishing. If the text is short, such as a single headline, asking the editor for it and then proofreading is realistic; for dense text, set it in a design tool.
Keep those out of the AI workflow. Ingredient lists, safety warnings, certifications and disclaimers are regulated and differ by country, so a translation error is a real problem and not a typo. Have a qualified person produce and approve the correct wording, then place it on the image yourself. This page offers general information, not legal advice, so ask a professional when the text carries legal weight.
You can repeat the same edit across many images, which is the practical way to localize a catalog. Keep one edit prompt, swap only the translated text, and always start from the original file. The risk is that one mistake gets copied into every image, so check a few results closely before you run the rest, and spot-check the batch afterward.
Only on images you own the rights to, or on fictional characters. The edit should change the words and leave the people as they are, so tell the editor to keep faces and the rest of the scene unchanged. Do not use it to alter pictures of real, identifiable people in ways they have not agreed to, and do not use it to make a product photo show something untrue.
Each generation costs credits, and the cost depends on the model and the settings you choose. A language model call is usually small next to an image edit, and more resolution costs more. New accounts get free credits to test a few images. Current plans and credit packs are on the pricing page, which is the only place to trust for numbers.
Read it twice: first against the translation you approved, letter by letter, then as a customer would, at the size it will be seen. Check accents, line breaks and whether anything else in the picture shifted. Then have a fluent speaker read it, because a native eye catches tone problems a spelling check misses. Last, check the rules of the platform where it will run.
Your next image is closer than a new design file. Open the Picasso IA toolkit, draft the translation with a language model and let an image editor put it where the old words were.
Choose your language
