Patois.

Docs · Features

Image localization

Upload an image and get it back with its text in another language. The pipeline detects the text regions, translates them, and re-renders the image, and the middle step is the ordinary document pipeline, so translation memory, glossary enforcement and workflows all apply to the text in an image for free. This feature is in beta.

What it does#

You have a screenshot, a banner, a diagram, or a UI mock with text baked into the pixels, and you need it in another language. Image localization detects the text, translates it through the same machinery that translates documents, and re-renders the picture with the translated text fitted back into place.

Each detected block of text becomes an ordinary translatable segment, which is what lets translation memory, glossary enforcement and workflow review apply to it with nothing extra to configure. Where the text will not fit its original space, that is flagged as overflow rather than clipped or silently shortened. Text location is always measured by a local detection component; a vision-capable model is used only to re-read a low-confidence region's text, never to guess at coordinates.

Getting started#

  1. Upload the image into a project, exactly as you would a document. Its text regions are detected immediately.
  2. Review what was detected: the regions, their confidence, and after translating, any overflow flags.
  3. Translate into a target locale. Each region goes through translation memory, glossary and any workflow, exactly as a document segment.
  4. Render to see the localized image and its fit, then export the file.
image upload 3 ./banner.png
image show 12 --locale es-ES
image translate 12 --locale es-ES
image render 12 --locale es-ES
image export 12 --locale es-ES

Works with#

  • Documents: an image's regions are ordinary segments, so the whole document pipeline, status, QA flags, the editor, the export guard, applies to the text in an image for free.
  • Workflows: a region is translated by the same engines and workflows a document is, so a multi-step workflow runs over image text unchanged.
  • Translation memory: region text reads from and writes to translation memory exactly as a document segment does.
  • Glossaries: glossary and do-not-translate checks fire on region text and surface as QA flags, the same as on a document segment.

Current limits#

Current limits

Text location is never inferred by a language model; it always comes from a local detector, so it is a measurement rather than a guess.

Text that does not fit its box is flagged as overflow, never truncated; the fix is a shorter translation or an accepted smaller fit.

If no text detector is available on the install, an image reports zero detected regions rather than presenting a false-empty result.

An exported image is always served as a download, and an exported SVG is restricted further, since an SVG file can carry executable script.

This feature is in beta: the pipeline is stable, but treat it as an extension of the document pipeline, which is where the broader quality tooling lives.