Putting translations back into scans in place
Short answer: Translation backfill writes the translation into the scan at the position of the original text, instead of laying out a new document. EnjoyPDF reads the full text and builds a "source → target" keyword table; you review the translations and run a batch replacement, while layout, table lines and untranslated parts stay as they are.
Which files it suits
Finalized scanned material that only needs a language change: product manuals, technical handbooks, labels and packaging artwork, internal training material, and archive copies of correspondence. The layout usually needs to match the original because the client compares page by page, or because the file must be archived in its original format.
Signed or issued files are detected by the risk levels and need confirmation, and their output pages carry an “EDITED COPY” stamp. Official or certified documents are never changed in place. For those, use a formal certified translation process.
Workflow
- Build the keyword table: use “AI-translate whole document…” to read all text, translate it in batches in reading order, and build a “source → target” keyword table. Or import an existing two-column translation table.
- Review the translations: check and edit each entry in the keyword table editor. Entries from your glossary are used as they are.
- Find and preview: after searching, check each position, then switch to “Preview replacement” to see the layout.
- Replace and save as: when saving, each replacement is checked by read-back verification in the new file. Amounts, dates and reference numbers are compared strictly.
- Deliver: give the output file and the operation record (manifest.json) to the reviewer or client.
Why not convert to Word first
Converting to Word and back to PDF re-lays out every page: fonts, line spacing and table widths all change, and the printed texture of the scan is lost. Backfill changes only the text that needs changing; all other pixels come from the original image.
Multiple languages
For scripts that need complex shaping, such as Thai and Devanagari, EnjoyPDF shapes the text with HarfBuzz before writing it, and adds invisible text so it can be searched. The current version does not yet support right-to-left scripts such as Arabic (including Persian and Urdu) and Hebrew, either as the source text or as the translation written into the file. For measured results by language, see the language support matrix.
FAQ
Do I have to use AI translation?
No. You can use your own translation tool or a glossary exported from CAT software, arrange it as a two-column table, and import it for batch replacement. AI translation is only one optional way to build the keyword table.
Does text leave my computer during AI translation?
If you choose a cloud platform, the text to be translated is sent to that platform. You are asked to confirm before the first send, and numbers (including ID and phone numbers), emails and URLs are replaced with placeholders and not sent; names and other ordinary words are sent. If you choose a local model such as Ollama or LM Studio, the text stays on your computer.
What if the translation is longer than the original?
After the keyword table is built, "When image text is too long" is switched automatically to "Stay within the original area", and the font size is reduced where necessary. Check the result with "Preview replacement".