Reglyph is our product, so we can't claim to be neutral. Both tools got the same files, and every result is shown as it came out.
| Page | Languages | Better result | What decided it |
|---|---|---|---|
| Colored panels on white (scan) | EN → ES | ChatGPT, slightly | Tidier layout; but it dropped a word and flattened the title-bar gradients |
| 1871 book page (scan) | FR → EN | Reglyph, slightly | Both kept the paper; ChatGPT's text fills only the left part of the column |
| Financial table (scan) | EN → JA | Reglyph, slightly | ChatGPT's table is tidier, but its chart labels sit far from the bars and one header is crowded |
| Text over a photo (scan) | EN → ZH | Reglyph | ChatGPT replaced the photo background with grey, some English left |
| Dense two-column paper (digital) | EN → FR | Reglyph | ChatGPT turned apostrophes and inline symbols into question marks and dropped a heading |
| Score | Reglyph 4 · ChatGPT 1 |
We asked in English: "Translate this PDF into Spanish and keep the original layout." ChatGPT worked on it for a minute or two and replied with a download link:
We compared the PDF it returned with the original pixel by pixel. The photos, icons and logos are identical; only the areas with text were changed. So it doesn't redraw the page: it replaces the areas where the original text was and draws the translation on top. That works well on flat backgrounds and on old paper; when the background is a photo, the replaced areas show.
ChatGPT's version of this page is tidier than ours. The typeface is close to the original and the "Salud intestinal" label fits its tab, where ours wraps onto two lines and leaves a smudge, and our first heading wraps too. Its problems only show up close. The green title bars were gradients in the original; ChatGPT painted them over with flat light boxes. And the gut-health paragraph lists sauerkraut, kimchi, kefir and miso in the original, but only three of them in ChatGPT's Spanish: miso is gone.
Reglyph is slightly ahead here, but ChatGPT did better on this page than we expected. The color and specks of the old paper are all there; at normal size you can't see where it worked, and only up close do a few faint horizontal streaks show where the original lines were. The difference is the layout. The original text runs the full width of the column; ChatGPT's English fills only the left part of it, leaving a strip of empty paper on the right. Paragraphs have no first-line indent, and the book title at the foot has lost its italics. Ours keeps the original column width; it also loses the indents, and the lines bunch up a little where paragraphs meet.
Reglyph is slightly ahead on this page. ChatGPT rebuilt the whole table, with neatly aligned rows and columns and every $ sign and bracketed negative in place; that part is cleaner than ours. The problems are outside the table. In the top-left header, the line giving the summary and units runs into the "2025年度" heading beside it. The five bar-chart labels sit far below the bars, with a few stray Japanese commas between them, and parts of the chart's frame are missing. In the right-hand column of body text, "DOI" is split across two lines. Our table and chart labels are where they belong; our flaws are a small smudge in the footnotes and the lost bold in the header.
This page prints its text straight onto an aerial photo of a coastline, which is where the solid-color approach runs into its limits. ChatGPT laid one big grey patch over the whole body-text area, so most of the sea on the left is no longer visible. A few spots aren't covered: half of "Our Mission" peeks out under the Chinese heading, the English captions under the three statistics are still there, and a line of English body text shows below the grey box. The original English photo caption is untouched, and ChatGPT put a white box with a Chinese caption above it, so the photo now has two captions. Ours puts the Chinese on the restored photo, with the three numbers still set large.
We then asked it to redo the page "as precisely as it can". It worked for over seven minutes and returned much the same result, using the same solid-color approach. On photos and gradients, asking for more care didn't noticeably change the result.
Reglyph wins this page. ChatGPT fitted the French onto one page and left the displayed equation and the figure untouched, with nothing overlapping. But every apostrophe in the text became a question mark, so "d’observation" reads "d?observation" and "qu’il" reads "qu?il", and inline math went the same way: "m₁ et m₂" became "m? et m?" and "ΔL" became "?L". The section heading "III. DÉTECTEURS" at the bottom of the left column is gone, the half-sentence "détecteur [33], un interféromètre de Michelson modifié…" appears both at the end of the left column and at the top of the right one, and the columns have large empty gaps. Our French has no garbled characters, and the heading and equations are all there; our weak spot is still the missing first-line indents.
Where ChatGPT ranks against the other tools we tested: the full PDF translation comparison.
More side-by-side tests: Reglyph vs Gemini, Reglyph vs DeepL, Reglyph vs Google Translate, Reglyph vs Reflo and Reglyph vs Smallpdf.
Upload a scanned PDF — even one made with your phone's scanner app. Reglyph OCRs it, translates it, and rebuilds the page so tables, stamps, and figures stay exactly where they were.
Translate 5 pages free, every dayIt can hand back a PDF. On a page of flat colored panels, an old book and a financial table the result was good; on a page with text printed over a photo, the grey patch it lays down hides most of the photo, and several pieces of English aren't covered.
On the page of colored panels ChatGPT was slightly ahead; on the old book and the financial table the two were close, with Reglyph slightly ahead; on the page with text over a photo, Reglyph was clearly better.
We can't be neutral, so we kept it checkable: the same files for both tools, ChatGPT's reply and output shown as they came out, and the page ChatGPT did better on included. The source pages are public, so you can run the test yourself.
Vision models genuinely read scans now. The question isn't whether ChatGPT can translate your PDF — it's whether a chat reply is actually what you need back.
Read guideCamera-translate apps are great for reading a page on the spot, but they don't give you a file — here's how to get an actual translated PDF from your phone.
Read guideRandom symbols, blank pages, or 'PDF tools are temporarily offline' — the cause is almost always the same one thing, and it's fixable.
Read guideSimple, scan-friendly pricing. Pages are pages — no multiplier for scanned files.
Translate a few pages. No commitment.
For individuals with scans to translate every month.
For high-volume users with scans to translate every week.