How to Translate Text in an Image With AI Without Breaking the Layout

Choose the workflow from the source file

The right method depends on whether the text still exists as text. Check the source before uploading the image to another tool.

If the layered source is available, do not flatten it just to use an AI translator. Existing text frames already preserve the position, width, hierarchy, and styling information that OCR has to guess.

Why direct replacement breaks otherwise good artwork

Translations do not keep the same physical width or line count. W3C's guidance on text size in translation shows why a fixed box that fits one language can overflow in another. The visible symptom may be clipped letters, an extra line over an illustration, or a headline that is shrunk until it no longer carries the intended hierarchy.

OCR creates a separate risk. It can skip text, merge adjacent regions, or read decoration as characters. Google warns that small, unclear, or stylized image text can produce less accurate results. Canva's Grab Text documentation also says the feature may miss or invent text and can leave visual artifacts. Treat OCR as a draft inventory, never as proof that every word was captured.

Translation and layout are therefore two different acceptance decisions. A sentence may be linguistically sound and still fail its box. A perfectly fitted sentence may still be the wrong translation. Record both decisions instead of hiding them behind a single “looks good” status.

What happened in one retained poster example

We created a 1200 by 1500 event poster with six known text regions. Local Tesseract OCR recovered 17 of 19 expected tokens and missed the two-word headline, “OPEN STUDIO.” It also interpreted a decorative X as low-confidence punctuation. Because the source inventory was frozen before OCR, both errors were visible instead of silently entering the translation.

A direct Spanish replacement used the source font sizes and removed the original line breaks. The headline measured 1,488 pixels inside a 770-pixel box, and the body line measured 1,198 pixels inside a 720-pixel box. Both were rejected for overflow.

The layout candidate split the headline into two lines and reduced it from 96 to 78 pixels. It moved the body line break to a natural phrase boundary and reduced that text from 42 to 39 pixels. All six translated regions then fit their recorded bounds. The time, venue, and URL remained unchanged, and a pixel comparison found no change outside the six text regions.

That is a layout pass, not a publication pass. The Spanish copy is still marked NEEDS_NATIVE_REVIEW. This distinction matters: box measurements can verify clipping and background changes, but they cannot approve tone, idiom, or meaning.

Step 1: freeze the text before translating

List every visible string in reading order. Give each region a stable ID and record its source text, box position, approximate type size, alignment, color, and any text that must not change. Include small footer copy, dates, labels, disclaimers, and text inside buttons.

Compare the inventory with the image at full size. If OCR and the visible source disagree, the image wins. Never send a partial OCR transcript for translation and assume the missing lines will be noticed later.

Step 2: translate by region, not as one paragraph

Send each text box with its role and constraints. A headline, a button, and a legal disclaimer should not be optimized the same way. Ask first for a faithful translation, then for a shorter alternative only when the faithful version cannot fit.

Translate these image text regions from English to Spanish.

For each region, return:
1. a faithful translation;
2. a shorter fit-aware option only if needed;
3. a note if shortening changes meaning or tone.

Do not translate tokens marked LOCKED.
Do not merge regions or add new claims.
Keep dates, times, URLs, product names, and venue names exact.

Return a table with:
region_id | faithful_translation | fit_aware_text | warning

Do not ask the model to “make everything fit” without giving box constraints. That instruction encourages silent deletion. When space is tight, the reviewer should be able to see which wording was shortened and why.

Step 3: rebuild the text without repainting the artwork

Duplicate the source, hide or remove only the original lettering, and keep a clean background layer underneath. Add one editable text layer per inventory row. Preserve the original box coordinates before adjusting any typography.

Fit each region in this order:

1. Use natural line breaks at phrase boundaries.
2. Widen the text frame only into verified empty space.
3. Try a compatible condensed or language-appropriate font.
4. Reduce tracking and line spacing within readable limits.
5. Reduce font size while preserving the original hierarchy.
6. Use an approved shorter translation if the full version still cannot fit.

Do not begin by shrinking every translated string. A 10 percent reduction applied across the design may make body text unreadable while leaving the actual problem, such as one overlong headline, unsolved.

Use Canva when the text is straight and editable enough

For a flat poster with straight Latin text, Canva's documented path is to select the image, open Edit, and choose Grab Text in Magic Studio. The detected words become editable text boxes. Compare the result with the original before translating, because Canva documents limits for rotated, curved, arched, non-Latin, and multi-page text.

After correcting the extracted boxes, open Apps, choose Translate, set the target language, and translate the selected page or design. Canva says Translate can duplicate pages and resize text. Those options reduce manual work; they do not remove the need to check every box, especially headlines, buttons, prices, and small print. See Canva's current Translate documentation for availability and interface details.

If Grab Text leaves a halo, patched background, or invented lettering, return to the unedited source and rebuild that region manually. Continuing from a damaged intermediate file makes the artifact harder to isolate.

Use Google Translate for a reading copy, not a master design

On desktop, Google Translate lets you choose Images, upload a JPG, JPEG, or PNG, select languages, compare with the original, copy the translated text, or download the translated image. This is useful when the goal is to understand a sign, menu, screenshot, or notice quickly. The official image translation instructions also warn that small, unclear, or stylized text can reduce accuracy.

Keep the downloaded result as a reference unless it passes the same inventory, language, and layout checks as an editable design. If the image will be published, rebuild the approved translation in the source editor so future corrections do not require another round of OCR and repainting.

Handle right-to-left and vertical text as a layout change

Arabic and Hebrew change more than word length. The Unicode Bidirectional Algorithm governs how right-to-left text interacts with numbers and left-to-right strings such as URLs. Set the paragraph direction explicitly, then inspect punctuation, phone numbers, dates, arrows, and icon placement. Do not mirror a logo or product image just because the reading order changes.

Vertical Japanese or Chinese copy also needs an editor and font that support vertical layout. Unicode's bidirectional text FAQ distinguishes bidirectional behavior from vertical writing. Rotating a horizontal translation by 90 degrees is not equivalent to proper vertical typesetting; punctuation and glyph orientation can differ.

Run a release check at delivery size

Inspect the final export at 100 percent and at the size where readers will actually see it. Compare it with the source side by side, not from memory.

• Every inventory row appears once and in the correct region.
• Names, URLs, dates, times, prices, measurements, and legal references match their locked values.
• No line crosses its box, touches artwork, or disappears under a mask.
• Headline, body, caption, and button hierarchy remains clear.
• The chosen font contains every required glyph; no tofu boxes or fallback fragments appear.
• Text direction, punctuation, and mixed-script order are correct.
• No pixels outside the intended text regions changed unexpectedly.
• A fluent reviewer has approved meaning and tone in the target language.

If the image contains essential information on a web page, provide the same information as real text nearby or in an appropriate text alternative. W3C recommends actual text instead of images of text when the presentation can be achieved with text, because real text can reflow, resize, and work with assistive technology.

Release rule: export only when the text inventory is complete, all locked strings match, every box fits at delivery size, the artwork outside the edited regions is unchanged, and a qualified language reviewer has approved the translation. A tool's “translated” badge satisfies none of those checks by itself.