Reglyph is our product, so we can't claim to be neutral. That's why every verdict here comes from the same set of files run through every tool, and every tool's result on every page is shown below as it came out.
We scored every tool on five criteria. Reglyph scored 4.5 and came first, losing half a point on typography such as indents and bold. DeepL is close behind on 4, and typography is its strongest point. ChatGPT scored 3.5: good on old paper and tables, marked down on the photo background and on the paper's details. Doclingo and Reflo scored 3 each, with Doclingo translating tables more completely. Google Translate and Gemini scored 1.5, and Smallpdf 0.
| Reads scans | Laid-out output | Complex background | Equations & tables | Typography | Score | |
|---|---|---|---|---|---|---|
| ReglyphBest overall | Yes: Yes | Yes: PDF | Yes: Fully restored | Yes: Complete | Partly: Loses first-line indents and bold | 4.5 / 5 |
| DeepL | Yes: Yes | Yes: PDF | Partly: Smudges on icons | Partly: A label covers a figure; one paper ran to two pages | Yes: Indents, italics and bold kept | 4 / 5 |
| ChatGPT | Yes: Yes | Yes: A PDF, when asked | Partly: Old paper restored, photo replaced with grey | Partly: Tables complete, inline symbols turned into ? | Partly: Tidy plain pages; apostrophes became ? and a heading went missing | 3.5 / 5 |
| Doclingo | Yes: Yes | Yes: PDF | No: Solid backing instead | Partly: Tables complete; one equation lost a term | Partly: Indents kept, type size jumps | 3 / 5 |
| Reflo | Yes: Yes | Yes: Laid-out page | No: Solid backing instead | Partly: Often skips tables | Partly: Words split mid-word | 3 / 5 |
| Google Translate | No: Refused | Yes: Digital only | No: Refused | No: Equations broken, table flattened | Partly: Indents kept, type size uneven | 1.5 / 5 |
| Gemini | Yes: Yes | No: Text in the chat | No: Background and figures gone | Partly: Clean equations, figures gone | No: No page | 1.5 / 5 |
| Smallpdf | No: Refused | No: Plain text | No: Refused | No: Equation flattened to one line | No: No page | 0 / 5 |
How we scored: 1 point for each criterion met, 0.5 for partly met, 0 for not met, refused, or text only.
Below are the six test pages, each with every tool we ran on it. Tools that returned a page show that page; tools that didn't show what they gave us instead. Click any page to enlarge all the results together and compare the same spot.
Annual report page, text printed straight onto an aerial photo (scan, EN → ZH)
Swipe to compare, tap to enlarge.Click a page to enlarge.
The hardest background, and the widest spread of results. We restored the photo. DeepL kept the photo but left smudges on the icons behind the statistics. Doclingo translated everything and kept the big numbers, but each line of text over the photo sits on a grey band. ChatGPT and Reflo laid a solid backing under the translation, so most of the sea is hidden and some English is left uncovered. Gemini answered in the chat, without the photo. Google Translate and Smallpdf refused the scan.
Dietary guidelines page, green title bars and dark-blue tabs on white (scan, EN → ES)
Swipe to compare, tap to enlarge.Click a page to enlarge.
On a white page with colored panels, everyone did well. ChatGPT's page is the tidiest, but it dropped a word and flattened the title-bar gradient. DeepL and we tied. Doclingo was a little better than us: its titles and tab label didn't wrap, though its type size jumps between lines. Reflo put a pale backing under the titles and left the footer in English. Google Translate and Smallpdf refused the scan; Gemini answered in the chat and put two items under the wrong heading.
Twenty Thousand Leagues Under the Sea, 1871 edition, page 9: yellowed paper and an engraving (scan, FR → EN)
Swipe to compare, tap to enlarge.Click a page to enlarge.
DeepL has the best detail on the old book, keeping first-line indents and italics. DeepL, ChatGPT and we all kept the grain of the paper, though ChatGPT set its text in only the left part of the column. Doclingo kept the indents too, but each line sits on a beige band. Reflo set its English on a flat beige backing, so the paper texture doesn't come through. Gemini replaced the engraving with a line of text. Google Translate and Smallpdf refused the scan.
U.S. Department of the Interior financial report, net-cost table and bar chart (scan, EN → JA)
Swipe to compare, tap to enlarge.Click a page to enlarge.
We, DeepL and Doclingo all translated the whole table; Doclingo misplaced a few $ signs. ChatGPT rebuilt the table neatly, but squeezed text together in one header and set the chart labels far below the bars. Reflo translated the heading and body text only, leaving the table and chart labels in English. Gemini laid out the whole table neatly in the chat and turned the bar chart into a list. Google Translate and Smallpdf refused the scan.
Gravitational-wave physics paper, page 3, densely set with equations and a figure (digital, EN → FR)
Swipe to compare, tap to enlarge.Click a page to enlarge.
A crowded paper. We fitted the whole page with the equations and figure intact, but lost the first-line indents. DeepL's running text is the most polished, but its translated label covers the main chart. Doclingo kept the indents but dropped part of the mass equation. ChatGPT kept the equation and figure, but apostrophes and inline symbols became question marks and a section heading went missing. Reflo split a few words mid-word. Google Translate broke the equations into fragments. Smallpdf and Gemini gave text only, with no figure.
Mistral 7B paper, page 2, with diagrams and a parameter table (digital, EN → KO)
Swipe to compare, tap to enlarge.Click a page to enlarge.
A paper with a parameter table. We and Reflo tied on layout. DeepL stretched one page into two and split the table. Doclingo kept one page and translated the table, but left the diagram labels in English. Google Translate flattened the table into a paragraph. Gemini and Smallpdf gave text only, without the diagrams, and Smallpdf flattened the table into one line.
On every scan we tested, Reglyph restored the background behind the translation, and it translated tables, captions, headers and footers almost without exception. On the digital papers the equations and figures stayed intact, with no overlapping text. Our weak spots, plainly: paragraphs lose their first-line indents, and on one page of colored panels a Spanish tab label ran long and spilled out of its box.
Best for: scanned PDFs (including ones saved from a phone scanner app), pages with photos or colored backgrounds, and files where every table cell has to be translated. 5 free pages a day; upload one here.
DeepL came closest of everything we tested. It read every scan and kept the backgrounds, and it's more careful than we are with typography: first-line indents, italics and bold. Where it slips is figures and tables. The page with the busiest background has smudges on the icons, a translated label on a physics paper is set far too large and covers the main chart, and a one-page paper came back as two pages with its table split. We won 6–4. Full comparison
Ask ChatGPT to keep the layout and it writes code on the spot and hands back a PDF. On the white page with colored panels the result is good, tidier than ours, though it dropped a word and flattened the title-bar gradient. It handled the 1871 book well too, with the paper grain intact, though the text sits in only the left part of the column. It rebuilt the financial table neatly and translated all of it, but set the chart labels far below the bars. Text printed over a photo is where it falls short: a grey patch hides most of the photo, a few bits of English aren't covered, and asking it to be more careful didn't noticeably change that. On the digital paper it fitted the page with the equation and figure intact, but every apostrophe and inline symbol became a question mark, and a section heading went missing. Full comparison
Doclingo read every scan and translated nearly all of it, tables and footers included. All six pages came back as single pages, with first-line indents kept. For backgrounds it works like Reflo, laying a solid-color backing under the translation: invisible on a white page, visible as bands of color where text sits on a photo or old paper. Its most noticeable layout issue is type size jumping between lines within a paragraph, and on the physics paper one equation lost part of a term. On the page of colored panels it did a little better than us. We won 5–2. Full comparison
Reflo reads image-only scans and translates the body text. Its approach to backgrounds is to lay a solid-color backing under the translation, which it did on three of the four scans. On white or simple backgrounds that reads fine; if the text sits on a photo or old paper and you want the translation to keep the look of the original, it's not the best fit. On five of the six pages it also left tables, chart labels, captions, or headers and footers untranslated. On the digital papers it kept first-line indents, but split a few words across two lines. Full comparison
Google Translate doesn't run OCR, so it refuses scans outright. It does translate digital papers, keeping the columns and figures, but it broke the physics paper's equations into fragments and flattened a parameter table into a paragraph. For a quick, free read of a text-only digital PDF, it's enough. Full comparison
Gemini reads scans and papers, translates well, and renders equations in the chat more cleanly than Google Translate. It laid out the financial table neatly in the chat. But even when we asked it to keep the layout, it answered with text in the chat: photos, colored panels and figures are gone, the 1871 engraving became a line of text, and on one scan two items ended up under the wrong heading. Full comparison
Smallpdf refused our scan. It translated a digital paper, but as plain text beside the original, without the columns, figures or equation layout. If you already use it to merge or compress PDFs and just want to read the content, it's convenient. Full comparison
You're welcome to cite these results; please credit Reglyph and link to this page.
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 dayIn our side-by-side tests, Reglyph. It restored the background on every scan, including the page with text printed over a photo. On that page Doclingo, Reflo and ChatGPT replaced the background with solid color, hiding most of the photo, and DeepL kept the photo but left smudges.
For digital PDFs, Google Translate is free and needs no sign-up, but equations and tables often come out wrong. It refuses scans; for those, Reglyph gives you 5 free pages a day and returns a PDF with the layout kept.
ChatGPT hands back a PDF when you ask it to keep the layout; plain pages, old paper and tables come out well, but on a page with text printed over a photo the background is replaced with grey. Gemini answers with text in the chat, without photos or figures.
Yes. It read all four image-only scans we tested and kept their backgrounds; on the page with the busiest background, it left smudges on the icons.
We can't be neutral, so we've tried to make the results checkable: every tool got the same files, every result on every page is shown in the post, and the pages where a rival did better are written up too. The source documents are public, so you can run them 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.
$0.30 a page, from 10 pages — you pay for exactly 10 pages.
For individuals with scans to translate every month.
For high-volume users with scans to translate every week.