Translate text in any image

Why some text comes back untranslated

This is the most common surprise, and it is worth understanding because it is not a translation failure. Text is only translated if it was read first. When a line scores too low to trust, we leave it as it is rather than guessing — an obviously untranslated line tells you something went wrong and can be reshot, while a confident mistranslation reads as a fact. On a photographed card in our own testing, lowering that bar turned a stone's name into an unrelated word that looked perfectly plausible in Korean. We would rather show the gap.

Backgrounds: flat beats photographic

• Best case — text on a flat colour. Menus, signs, labels, packaging and UI screenshots. The background is measured from the image and reproduced exactly, and only the glyph pixels are repainted, so the label keeps its own grain and edges.
• Harder — text over a photograph or texture. The background behind the letters has to be invented by an inpainting model. It is usually convincing, but strong structure (tiles, patterned fabric, a face) can leave a visible seam.
• Hardest — text with outlines, shadows or gradients. The effect has to be removed along with the letters, or a ghost of the original remains.

Length changes, and what we do about it

Japanese and Chinese are compact; Korean, English and German usually expand. The layout was designed for the source, not the target, so the renderer searches for the largest font size at which the translation fits its box in both dimensions, wrapping at word boundaries for space-delimited scripts and per character for CJK. Very long translations in small boxes shrink; there is a floor, below which text would be unreadable anyway.

Vertical text

Vertical Japanese and Chinese columns are reassembled into whole sentences before translation — otherwise each column would be translated as an isolated fragment and drawn back in the wrong place. A column that stays vertical after that is rendered vertically: stacked character by character for Korean, Japanese and Chinese targets, or rotated for Latin-script targets. The reasoning is in this article.

Getting the best result

• Shoot square to the surface. Angled labels make the erase harder to keep inside its own edge.
• Fill the frame. Resolution per character bounds the reading stage, and everything after it inherits the result.
• Avoid your own shadow. A shadow across part of the text is the most common cause of a line scoring just below the bar.
• Pick the source language explicitly. Chinese and Japanese use different recognition models; the wrong choice costs accuracy.

Language pairs

Sources cover Chinese, Japanese and 40-plus Latin-script languages; targets add Korean. Source and target must differ. Translation uses Google Cloud Translation for the text itself, while all reading, erasing and redrawing happens in our own pipeline — so the layout work is ours, and the language work is not.

What you get back

A PNG with the translated text drawn in place, downloadable, plus the extracted source text and its translation as a list you can copy. Up to 10 images per batch. Uploaded images are deleted immediately after processing; see the Privacy Policy.