Translate a coffee bag label to Japanese with Ideogram 4.5

Translate a packaging label into Japanese with ideogram/ideogram-v4.5: give the exact string, one language per call, check characters. $0.075 per bag at medium.

5 min readSume
All posts

To translate the text on a coffee bag photo into Japanese, send the photo to ideogram/ideogram-v4.5 as the first input_references entry and give the exact Japanese strings in the prompt, one language per call. A medium edit is $0.075 on Sume, so twelve bags are $0.90. Ideogram's 4.5 page shows exactly this kind of work, with a coffee brand system translated into Japanese, a bakery marquee into French and a chalkboard into English.

The hard part is not the call. It is making sure the characters are right, because a plausible-looking kana string with one wrong character is still wrong.

Provide the translation, do not ask for it

Do the translation yourself, or have a native reader do it, and put the final text in the prompt in quotes: replace the English text DARK ROAST with 深煎り, replace 250 g with 250g. Asking the model to translate and render in one step makes it responsible for two things, and you cannot tell which went wrong when the output is off.

Check that your strings contain the characters you think they do. Fonts, copy-paste and keyboards produce look-alike characters. The snippet below prints each character with its Unicode name so you can see kana, kanji and half-width forms before you send the request.

import unicodedata

text = "深煎りコーヒー豆 250g"
for ch in text:
    print(ch, unicodedata.name(ch, "UNKNOWN"))

Prompt and call

Name each piece of text and where it sits, and tell the model to keep the bag shape, fold lines, colour, lighting and every other element. Do not send mask_url or output_format; Ideogram 4.5 on Sume does not accept them, and the edit keeps the source geometry when aspect_ratio is absent.

  • List every text element to replace and its Japanese string in quotes.
  • Say that fonts should match the existing style and that nothing else changes.
  • Run one language per call; a second language is a second edit of the original.
  • Keep the product photo, not the earlier edit, as the input for each language.

Cost

Prices are Sume's per-image amounts. Japanese text has more detail per character than Latin, so start at medium and move to high only if strokes look broken at full size.

Sume cost of Ideogram 4.5 label translations (read 2026-10-07)
QualityPer bag12 bags, one language12 bags, three languages
low$0.0375$0.45$1.35
medium$0.075$0.90$2.70
high$0.275$3.30$9.90

Text sizes and layout

Japanese text is denser than English. A short English word like ROAST becomes two or three characters in kana or kanji, which can leave empty space, while a long description becomes a longer line that needs more room. Say whether the text should keep its box size or fill it. Also say whether the layout stays horizontal; Japanese packaging often uses vertical lines, but a model asked to replace horizontal text will normally keep it horizontal.

Weights and fonts matter for readability. A decorative English script has no direct Japanese equivalent, so ask for a clean sans or a brush style explicitly rather than hoping for a match. If the bag has small print such as ingredients or a roast date, replace it only if you have a verified translation, and consider leaving regulatory text to a proper packaging file, since a photograph is not a compliant label.

Review checklist

Zoom into every character. Kana have small marks, and a missing dakuten changes the word. Compare the line breaks, the vertical or horizontal direction and the size of each text block against the source. Have a Japanese reader check the packaging copy for natural phrasing, since a literal translation can read oddly. A failed job is not billed, but a completed edit with a wrong character is, so a pilot of three bags at medium is $0.225 and worth it.

If a character keeps coming out wrong, fall back to a deterministic route: leave the artwork text-free and set the Japanese text in your own design file.

Sources

Related posts

More in Use cases

All Use cases posts

Written by Sume