Translate image text without losing the layout

To translate text inside an image while keeping its layout, separate the task into approved wording, image editing and final typesetting. First decide exactly what the new text should say. Then replace the original words. Finally, check the exported image against both the original design and the approved translation.

For a poster that will be reused in several languages, an editable master is usually the more controllable path. Reference-based AI editing can produce a candidate quickly, but it can also redraw a logo, change a digit or move an object. Asking it to “translate everything and preserve the layout” does not lock the non-text pixels.

This guide includes a complete prompt and a downloadable, deliberately simple example. The example is an authored SVG layout, not an Ofox generation result or a customer campaign. We checked Ofox’s image interface on September 29, 2026; we did not run a paid image-localization generation for this article. The generative branch below is a workflow to try and inspect, not a measured success claim.

Choose the path from your source file

Look for the original design file before uploading a flattened image. In a layered editor, select the headline: can you put a text cursor between its letters? If so, you may be able to replace the wording without reconstructing the image.

A .psd extension alone does not establish that the words remain editable: a flattened image can be stored inside a PSD. Likewise, some SVG lettering is converted to paths. You need text objects, not only the right filename.

Photopea documents the distinction between editable type layers and raster exports such as PNG and JPG. Turning a PNG into a PSD does not recover its original font, layers or background behind the letters.

Make a text inventory before translating

Use one row per meaningful block: headline, supporting line, date, price, button and disclaimer. Copy the text yourself or use OCR and verify its transcription against the image. OCR can confuse 0 with O, omit punctuation or join two columns.

Here is the complete source copy for our fictional workshop poster:

These translations are editorial examples, not observed model output. For a real campaign, have a fluent reviewer approve the wording before putting it into artwork. The sample deliberately has no ticket price, discount or brand: those create extra facts to validate.

Create a separate “do not translate” list for brand names, product codes, URLs and packaging text. If the image says Model AB-12, preserve that identifier rather than replacing it with a plausible local model name. If a date is ambiguous, leave a question for the owner; do not silently choose a regional interpretation.

Use AI for constrained copy, then approve it

Paste this into a text model along with the transcription. You can use the prompt independently of the image editor; there is no need to ask an image generator to invent the translation and render it in the same request.

Task: localize the following image copy from Chinese into English.
Audience: adults considering a creative workshop.
Tone: clear, friendly, no exaggerated claims.

Source blocks:
H1: 周六创作工坊
BODY: 把灵感做成作品
DATE: 10月17日 · 14:00–16:00
CTA: 查看活动详情
NOTE: 示例活动 · 非真实报名页

Constraints:
- Preserve every date and time. Do not invent a year or timezone.
- Do not add prices, discounts, booking availability or guarantees.
- Keep the CTA informational; it does not register the reader.
- Keep NOTE explicitly about a fictional example.
- H1 may occupy two lines; BODY may occupy two lines.
- Provide a natural standard version and a shorter alternative.
- Never drop facts merely to fit the shorter version.

Return a table with ID, source, standard translation, short alternative,
meaning changes (if any), and questions requiring human confirmation.
Flag ambiguity instead of guessing. Do not render an image yet.

Check each row, not just the fluent-looking headline. A translation can sound excellent while changing “view details” into “book now.” Back-translation can reveal a mismatch, but is not proof that the translation is correct. Resolve the questions, select one version per block and save that approved copy separately.

Do not impose a universal character limit to solve fit. Ten Japanese characters, ten Cyrillic characters and ten Latin characters occupy different widths depending on the font. Test the actual text at the final size. If it does not fit, shorten wording without dropping its meaning, allow another line or enlarge the text area before reducing the font.

Try a reference-based edit in Ofox

Open the Ofox image tool and sign in if required. At the time of our interface check, the page exposed a prompt field, reference-image upload, model selection and model-dependent image parameters. It also displayed generation history with downloads. Those controls establish an editing entry point; they do not establish perfect localization or a text-layer export feature.

Use an image you own or have permission to edit. Add the poster as a reference image, select one image model and check the current upload requirements and available sizes in the interface. Start with one model so that you can associate each result with its settings. Match the original aspect ratio where supported; a different output canvas makes layout comparison harder.

Paste your approved replacements explicitly:

Edit the attached workshop poster for an English-language audience.
Replace only these text blocks with the exact wording below:

周六创作工坊 -> Saturday Creative Workshop
把灵感做成作品 -> Bring your ideas to life
10月17日 · 14:00–16:00 -> 17 October · 14:00–16:00
查看活动详情 -> View event details
示例活动 · 非真实报名页 -> Sample event · Not a live booking page

Keep the same canvas composition, illustration, background colors,
text hierarchy and block alignment. Keep the date and hours exact.
The headline and supporting line may each use two lines.
Do not add a logo, price, slogan, booking claim or extra text.
Do not translate any other product or brand lettering.
Return the edited image, not a written explanation.

Review the displayed cost information before submitting. If the interface shows no estimate, that does not mean generation is free; confirm the model’s billing information first. After a successful run, download the result rather than judging only its preview, and record the selected model, settings and prompt.

Inspect the new wording letter by letter. Then compare the entire non-text area: object shapes, number of items, shadows, logos and product labels. An attractive result with a wrong date is a failed localization. If the model keeps changing details that must stay exact, move to the editable-layer method instead of expanding the prompt indefinitely.

Keep the layout with an editable master

The following example is useful when you need a repeatable process rather than a one-off generated candidate. Download both files:

• Chinese source poster, editable SVG
• English localized poster, editable SVG

Both use a 1200 × 1200 canvas, the same background and illustration, and the same text-block positions. The English headline uses two lines already reserved in the template. We checked the SVG structure: after text content is removed, the element structure and attributes are identical. This verifies that the authored files change only wording; it does not prove that a generative edit will behave the same way. Font substitution can still change the rendered width on another machine.

For this sample, open the SVG in a text editor. The editable blocks are named headline, support, date, cta and note. Replace only the words inside the <text> and <tspan> elements. Preserve the tags, coordinates and quotation marks around attributes. The headline and supporting line each have two tspan lines; replace both deliberately rather than pasting an entire paragraph into the first line. Save a separate locale copy and open it in a browser to inspect the result. In XML source, write an ampersand as &.

For a ready-to-use raster example, download our rendered English PNG or Chinese PNG. These are 1200 × 1200 renders of the SVG files, not model outputs. After changing your own SVG, open it in Photopea with File → Open and export through File → Export As → PNG. Keep the original SVG separately; do not assume imported text remains editable.

This source-edit method needs no account and makes a small change easy to audit. It is not convenient for rebuilding a complex flattened poster. For that, use your original design application or a layered image editor:

1. Open the original PSD in Photopea using File → Open. Duplicate the file for the new language. If the source is only a PNG, you will need to rebuild the text areas; it will not become editable merely by opening it.
2. Select the relevant type layer and use the Type tool to replace its text. For a new paragraph block, drag a text rectangle rather than placing a single unbounded line. These are Photopea’s documented text operations.
3. Keep the image layers untouched. If you have only a flat image with words on a solid panel, place a matching opaque panel over those words and add a separate type layer. When lettering overlaps detailed imagery, use a clean original or retouch the background first; a rectangular patch is not an invisible repair.
4. Use a font with the required script and inspect every line. Adjust the text box and line breaks; do not stretch the lettering to force it into the old dimensions. Keep the CTA label within its button.
5. Save the editable document with File → Save as PSD, then produce the delivery image with File → Export As → PNG. Those are the documented save and export paths. Reopen the exported file to verify it.

You can keep the exact original photograph while changing the typography this way. It still takes design judgment: a longer headline may require a different break, and two-language artwork may need a different composition. Preserving visual hierarchy matters more than mechanically reproducing every line break.

Fix the specific failure you see

A detected failure should change the next action. Repeating an identical generation is not a systematic fix for a font problem or an unapproved translation.

Deliver the image and the information needed to reuse it

Before publishing, compare the approved copy with the final image in reading order. Check numbers, punctuation, units, spelling, missing lines and the CTA. View it both at 100% and at its intended display size. Then compare non-text elements with the original. If exact preservation is required, use a layered original and inspect an aligned overlay; do not substitute a subjective “looks similar” judgment.

Keep these together: the original, approved copy table, editable master and final export. For generated candidates, also keep the prompt and model settings. A practical naming convention is workshop-en-v1.svg for the master and workshop-en-v1-1200.png for the delivery file. Mark the approved version so a future campaign does not start from a rejected draft.

For your next image, start with one headline and one supporting line in the Ofox image tool, then compare the result with this checklist. If the real task is a new cover rather than localization, use the AI thumbnail workflow. If the dimensions also need to change, decide whether to crop, pad or regenerate before translating the finished layout.