Translate Image Text and Edit CJK Lettering: Four Real Tests
See actual English-to-Spanish translation and Chinese, Japanese and Korean replacements, with exact characters, output dimensions and review limitations.
Image translation has two separate tasks: choose the right words, then render them in a layout. A phrase can be translated correctly in meaning and still overflow a sign, lose an accent or acquire a malformed character. Review language and pixels separately.
Our Translate mode targets one phrase supplied by the user. It does not automatically find and translate every word in a picture. Our CJK mode is different: it replaces an explicitly supplied Chinese, Japanese or Korean phrase and asks the model to preserve those exact characters.
The following examples are live Nano Banana Pro requests on our own test artwork. They cover simple horizontal headings, not dense documents, natural handwriting or a representative language evaluation.
English to Spanish: one short sign
We entered OPEN TODAY, selected Spanish, and requested a translation in place. The observed output was ABIERTO HOY, which is an appropriate rendering of this simple phrase. Longer text, idioms and specialist terminology need their own review.
Owned test artwork, before and after. Actual Nano Banana Pro 2K request: 44 seconds observed, 4% pixel-difference estimate. OPEN TODAY became ABIERTO HOY. This is one simple English-to-Spanish phrase, not a general translation accuracy benchmark.
The request took 44 seconds in this run. The phrase grew in length, but the simple sign had ample space. That does not prove a narrow button or tightly designed product label would preserve its layout.
If a brand name must remain unchanged, or a translation must follow a glossary, validate the wording outside the image first. For an exact approved translation, use Replace text mode and supply the approved target phrase yourself.
Chinese: supply the intended script
For the Chinese test, the source was 春日咖啡 and the supplied replacement was 夏日咖啡. The requested change is visible in the output. The model was not asked to translate, proofread or convert between simplified and traditional Chinese.
Owned test artwork, before and after. Actual Nano Banana Pro 2K request: 44.7 seconds observed, 5% pixel-difference estimate. The supplied Chinese replacement is visible. Only a short horizontal phrase was tested; no font identity is claimed.
Read every stroke with a fluent reviewer. Visually similar Chinese characters can carry different meanings. A pixel-difference check cannot establish whether a character is correct, and a similar-looking typeface is not the same as the original font file.
Japanese: check kanji and kana together
The Japanese request replaced 春のカフェ with 夏のカフェ. The target phrase is visible in this controlled horizontal-heading test. It does not test vertical composition, ruby text, tightly spaced body copy or a mixture of several writing directions.
Owned test artwork, before and after. Actual Nano Banana Pro 2K request: 38 seconds observed, 4.1% pixel-difference estimate. The supplied Japanese replacement is visible. Vertical writing, dense body text and mixed-script paragraphs were not tested.
Punctuation and long vowel marks deserve the same attention as the larger kanji. If a replacement is longer, check whether the model has changed line breaks, spacing or the placement of neighboring elements.
Korean: review the syllable blocks
The Korean request changed 봄 카페 to 여름 카페. The supplied phrase is visible in the result. This is a literal replacement test, not a broad Korean translation or typography benchmark.
Owned test artwork, before and after. Actual Nano Banana Pro 2K request: 61.3 seconds observed, 5.6% pixel-difference estimate. The supplied Korean phrase is visible. Fine syllable-block detail and other Korean typefaces need separate review.
Korean syllable blocks group multiple components into one visual unit. Inspect the components, word spacing and stroke weight at full size. OCR can assist a review but should not be the only acceptance check.
What the modes do not automate
Translate mode does not perform whole-image OCR or certified localization. CJK mode does not decide the correct spelling or silently change the script. Neither mode guarantees exact fonts, correct semantics in every context or a suitable final design.
Images used as records, payment proof or chat screenshots remain outside our permitted editing uses. For your own creative assets, keep the untouched original and avoid uploading confidential content because this version uses public-by-link storage.
Use the translation editor for a one-phrase translation, or the CJK editor for an explicit replacement. See all controlled examples and the test methodology before extrapolating these four cases to other material.