How to Run an AI Image Localization Pass on 20 Assets

9 min read

Make this article actionable

Send the article context into Vife Agent and turn it into a plan, checklist, or draft you can keep working on.

Open in Agent

A campaign goes live in five markets next month. The copy is approved. The images are not: a launch poster, eight marketplace banners, four app-store screenshots, three packaging mockups, two event cards, a PDF cover and a menu — twenty assets, each with words baked into the pixels.

The old answer is a designer rebuilding every layout five times. The newer answer is a model that retypes the words in place. Microsoft's November 17 announcement of MAI Image 2.7 shows that demo directly: one poster kept with "Same layout, same art, translated in place" across English, Spanish, German, Japanese, Korean and Arabic, and a second demo that swaps a headline typeface while holding "Same words, same photo, same layout."

A demo is a capability, not a workflow. Between "the model can retype a headline" and "twenty assets ship in five locales without a broken layout" sit a text inventory, a length budget, a font decision, a right-to-left plan and a QA pass. This guide is that workflow, sized for a small team with no localization vendor.

Quick answer

Inventory every text block first, including the small print. Freeze one master file per aspect ratio. Put the approved target strings in a table with a character budget per block, and translate them in context before you generate anything. Then run one locale per pass and one text block per instruction, always starting from the master rather than from the previous edit. Finish with a six-check QA pass that compares each locale against the master, not against the last version you liked.

Mid-read shortcut

Turn the useful parts into next steps

Vife Agent can convert this guide into a prioritized workflow with tasks, risks, and reusable prompts.

Create a brief

Why in-image text breaks when you translate it

  • Growth. A short English headline becomes a longer German or Spanish one, and text that fit a fixed-width band now crowds an edge or drops a line into a face.
  • Substitution. A typeface chosen for English may have no glyphs for Japanese, Arabic or Hindi, so the model falls back to another font or draws something that looks like letters and is not.
  • Direction. Right-to-left scripts change alignment, punctuation placement and sometimes the whole composition.

There is a fourth failure unrelated to translation: the model redraws something you did not ask it to redraw. Microsoft describes the editing behaviour as changing "an object's color, material, or position without disrupting the rest of an image" and says the model "preserves visual consistency across successive edits" (MAI Image 2.7 announcement). That is the design goal, not a guarantee. The same page publishes no text-fidelity benchmark and no stated limit on how much text one edit can replace, so treat both as open questions in your QA.

Step 1: Inventory every text block before you touch a pixel

Build one row per text block, not per asset:

AssetBlockSource stringCharsPriority
launch-poster-4x5
H1
Save the date
13
may shorten
launch-poster-4x5
LEGAL
Terms apply. See site.
22
exact
banner-marketplace-1
CTA
Shop the drop
13
may shorten
packaging-mockup-a
ING
Ingredients list
180
exact

Two rules make the table useful. Mark every block exact or may shorten before translation starts, so nobody negotiates a legal line under deadline pressure. And count characters, not words: overflow is a character problem.

Reading text off the image can be partly automated. Google's Gemini 4 Neon announcement says the model "reads uploads like spreadsheets, PDFs, and photos more precisely," and reports 73.5% on Chartography for reading dense charts — a vendor-reported score, not an independent test. Useful for a first pass, not sufficient alone: transcription errors are silent, and a missing comma in a legal line survives every later check. Read the inventory against the images once, by eye.

Step 2: Freeze a master file and a text table

Pick one master per aspect ratio and stop editing it. Every locale is generated from that master, never from the previous locale. This one rule prevents the most expensive failure in the workflow: drift that compounds across five sequential edits until nobody can tell what changed.

Then write the text table. For each block and locale, record the approved string and a character budget. If you have no measured data from your own campaigns, plan for the target to run up to about a third longer than English for German, and verify against real assets.

Decide now what happens on overflow: shorten, drop the line, or change the layout. Do not let the model solve it by shrinking the type. That is how a poster ends up with a 6-point legal line.

Step 3: Translate the strings, not the picture

Translate in a text tool, with context, before generating images. Give the translator each block's role (headline, button, legal, ingredient list), the character budget, the audience and a glossary of product names that must not be translated. Approve the strings in the table; only approved strings go into image prompts.

Text review is cheap and fast; image iteration is neither. A prompt carrying an unreviewed translation also hides the review: once the words are painted into a poster, a wrong word looks like a design decision.

Step 4: Run one locale per pass, one block per instruction

Generate in this order: master, then one locale, then one text block at a time. A prompt that names a single block, quotes the exact approved string and states what must not change is far easier to check than "translate the poster into German."

text
Edit the master image only inside the headline block. Replace the headline text with: "Sichern Sie sich Ihren Platz" Keep: same typeface, same weight, same size, same position, same color, same background, same product, same lighting. Do not change any other text block.

If the output drifts — the product moves, the background gains texture, a second block changes — revert to the master and retry with a narrower instruction. Editing the edited file is how one small error becomes five inconsistent assets.

Step 5: Check fonts, scripts and direction before judging layout

Name two or three typefaces that cover every script in your locale set, and put them in the brief. A decorative Latin display face will not render Japanese or Arabic convincingly, and the model will not tell you it improvised.

  • Diacritics. Vietnamese, Polish, Czech and Turkish strings need stacked or dotted marks. Zoom to 200%; a flattened mark is invisible at thumbnail size and obvious on a poster.
  • Shaping. Arabic and Indic scripts connect and reorder glyphs. Look for broken joins, isolated letters mid-word, missing dots.
  • Line breaking. Japanese, Chinese and Korean break between characters rather than at spaces, so a line set for word-wrapping may break inside a term.
  • Direction. For Arabic and Hebrew, check alignment, call-to-action position and whether the composition needs mirroring. Never mirror photographs of people or products.

Step 6: QA each locale against the master

  1. Crop test. View at 25%. Anything that reads as a smudge there reads as a mistake on a phone.
  2. String match. Compare each block character by character with the approved table, including punctuation.
  3. Boundary check. Confirm no text touches its band edge and no line was silently deleted to make room.
  4. Non-text diff. Compare against the master outside the text regions: product position, shadow direction, background texture should be identical.
  5. Script check. Re-read Step 5 for this locale only.
  6. Legal check. If a regulated string exists, a second person reads it against the source.

Log the result as a row: asset, locale, model and version, attempt, reviewer, date, pass or fail. When a price changes next quarter, the log tells you which twelve files to regenerate.

What it costs

Microsoft lists about $34.8 per 1,000 images for MAI Image 2.7 and $10.24 per 1,000 for MAI Image 2.7 Flash, priced at 1024×1024 pixels, both in public preview at announcement (MAI Image 2.7). Prices and availability change, so treat these as list figures for planning.

Twenty assets in five locales is 100 localized files. At two draft attempts each, that is 200 drafts on the cheaper model (about $2.05 at the list rate), then 100 finals on the higher-quality model (about $3.48). Roughly $5.50 of image generation for work that would otherwise take days of layout. The budget that actually overruns is review time.

Pre-publish checklist

text
IMAGE LOCALIZATION CHECK [ ] Text inventory exists, exact vs. may-shorten marked [ ] Master frozen; every locale generated from the master [ ] Strings approved in a table with character budgets [ ] Glossary applied; product names untranslated [ ] Fonts named in the brief and present in every script [ ] Diacritics, shaping and line breaks checked at 200% [ ] Direction and alignment checked for RTL locales [ ] Non-text regions identical to the master [ ] Legal strings read by a second person [ ] Version log updated with model, attempt and reviewer

Common mistakes

  • Translating inside the image prompt. You lose review, glossary and character counts in one step.
  • Editing the previous locale. Drift compounds; restart from the master.
  • Asking for the whole poster at once. One block per instruction is slower to prompt and faster to fix.
  • Letting the model resize type to fit. Shrinking text hides overflow instead of solving it.
  • Trusting OCR for the inventory. A silent transcription error propagates into every locale.

FAQ

Can AI translate text inside an image without changing the design? Models can retype text in place, and vendors show demos of exactly that. Whether your asset survives depends on string length, script and how much the model decides to "improve." Verify per asset rather than assuming.

Do I still need a designer? For a small campaign on a fixed master, often not for production. You still need someone to set length budgets, choose fonts and sign off on regulated strings.

What about languages I cannot read? Use a native reviewer for the final pass and give them the checklist, not just the image. A reviewer without the source string can catch a broken layout but not a mistranslation.

Where this fits in a larger workflow

Run this pass after copy approval and before channel resizing, so localized assets go through the same export and accessibility checks as the originals. Keep the text table and version log with the campaign folder: next quarter's price change or new market then starts from a table instead of from twenty images.