I tested 6 ways to translate a scanned PDF without losing the formatting

Disclosure up front: I build one of the tools in this list (Reglyph). I've
tried to be concrete enough about the others that the comparison is useful
anyway — every claim below is something you can reproduce in five minutes with
one scanned page.

"Translate this PDF" is a solved problem — until the PDF is a scan.

A digital PDF has text inside it. A tool can swap that text in place and the
layout survives. A scanned PDF has no text at all, just a photograph of a
page. Most translation tools handle this by OCR-ing the page, translating the
text, and giving you back… a plain text dump. The table is gone. The stamp is
gone. The two-column layout is one long column.

That matters because scans are exactly the documents where layout matters:
birth certificates, transcripts, contracts, invoices, technical manuals. A
reflowed wall of text is often unusable for the purpose you needed the
translation for.

I ran the same test document — a scanned certificate with a table, a seal, and
a signature line — through six approaches. Here's what each one actually does.

1. Google Translate (document upload)

What it does with a scan: accepts the upload, OCRs it, and returns the
translation as reflowed text. On image-only PDFs you often get an HTML-ish page
or a text-layer overlay; the visual layout of the original is not reconstructed.

• ✅ Free, instant, fine for "just tell me what this says"
• ❌ Tables collapse, images/stamps disappear, no output that looks like the original document
• Use it when: you only need to understand the content, not present it.

2. DocTranslator / online doc translators

What it does with a scan: these tools shine on digital Office/PDF files,
where they genuinely preserve formatting by editing the file's own text runs.
Feed them a scan and there are no text runs to edit — you get OCR-then-reflow,
same failure mode as above, sometimes with a "this file appears to be scanned"
warning, sometimes silently.

• ✅ Great for .docx and born-digital PDFs
• ❌ The formatting-preservation promise quietly stops applying at scans
• Use it when: your PDF was exported from Word, not from a scanner.

3. DeepL (PDF upload)

What it does with a scan: best-in-class translation quality, and its PDF
handling re-typesets digital PDFs impressively well. Scanned/image-only PDFs
are the weak spot — depending on plan and settings you either get a rejection
or an OCR-and-reflow result. Layout reconstruction of a bitmap page isn't what
it's built for.

• ✅ The translation itself is excellent
• ❌ Scans are either unsupported or come back reflowed; page limits on free tier
• Use it when: translation nuance matters more than the page looking like the original.

4. OCR + Word, by hand

The DIY route: run OCR (Acrobat, ABBYY, or ocrmypdf), open the result in
Word, translate chunk by chunk (or with Word's built-in translator), then fix
the layout manually.

• ✅ Full control; can reach genuinely faithful results
• ❌ It's work. A 10-page certificate packet with tables took me well over an hour, and OCR errors in table cells mean you're also proofreading a layout you didn't design
• Use it when: one high-stakes document, and you're billing someone for the time.

5. A vision LLM (GPT-4o / Gemini / Claude)

Screenshot a page, ask the model to translate it. The translations are good and
the model understands the layout — it can even output a Markdown table that
mirrors the original.

• ✅ Surprisingly strong comprehension of messy scans; great for Q&A about a document
• ❌ Output is text/Markdown, not a document. Nobody will accept a Markdown birth certificate. Multi-page = manual loop, and hallucinated digits are a real risk on numbers-heavy pages
• Use it when: you need to interrogate a document, not reproduce it.

6. Reglyph (the one I built)

Reglyph exists specifically for the scan case. It
OCRs the page, erases the original text from the image with an inpainting
model, translates the text, and re-typesets the translation back into the
original positions. The output is a PDF that still looks like your document:
tables, columns, seals, signature lines and numbers stay where they were.

• ✅ Layout survives on scans and even phone photos; 19 languages both directions, including CJK, Arabic (proper RTL + cursive joining) and Devanagari
• ❌ Machine translation quality, not human-translator quality — for certified translations you still need a human to review; it's a web app, no offline mode
• Pricing: 5 free pages every day, no credit card.

The actual decision table

The pattern behind the whole list: tools preserve formatting by editing text
that's already in the file. The moment the "text" is pixels, preservation has
to be rebuilt from scratch — erase ink, re-typeset, keep the geometry — and
almost nothing does that. That gap is why every roundup of "PDF translators
that keep formatting" quietly tests on digital PDFs only.

If you've got a scanned document workflow that beats any row of this table,
I'd genuinely like to hear it — especially for right-to-left targets, which is
where most pipelines (including ours, for a long time) fall over first.