Image Translate

Image translate reads the text inside a picture and renders it in another language, keeping the layout intact. It is meant for screenshots, menus, packaging, posters and diagrams — cases where the words are part of the image and retyping them would mean rebuilding the design.

How to get a clean result

The model has to find the text before it can translate it, so legibility decides the outcome more than any setting does.

• Give it a straight, well-lit capture. Perspective distortion and glare are the two most common reasons a line gets skipped.
• Use enough resolution. Text that is only a few pixels tall is guessed at, not read.
• Send the original, not a re-compressed one. Screenshot tools and chat apps re-encode images and soften small type.

Telling it what kind of text this is

A domain hint matters more than it looks. A phrase that means one thing in a menu and another in a medical leaflet will be translated differently depending on the domain you name. Alongside it:

• Terminologies — terms that must survive translation unchanged, such as product names or units.
• Sensitive words — words to handle carefully or leave alone.
• Skip image segmentation — send the whole image to be translated as one unit instead of detecting text blocks first. Useful for stylised lettering that a detector would split awkwardly.

Settings worth understanding

• Source and target language — set both; the model does not always guess the source correctly from a short line of text.
• Model route — the routes offered in the form are the ones reachable right now; live controls are the source of truth, not this page.
• Resolution — affects the credit cost shown beside the form.

Where it fits, and where it does not

It fits translating signs, labels, interface screenshots, comic panels and slides. It is a poor fit for handwriting, for text baked into a photograph at an angle, and for layouts where the translation is much longer than the original — a language pair that expands the text will crowd the box it came from. Expect to check every character that carries meaning: numbers and units in particular are worth reading twice.

Questions people ask

Will the design change? The intent is to keep the layout and replace only the wording. Heavily stylised lettering is where that gets hard.

How much does it cost? The credit cost appears beside the form before you submit. Credit prices are on pricing.

Where do I start? Open the Wan workspace and choose the image-translate tab.