Translate PDF & Academic Papers

When you translate a PDF, the hard part is not the language — it is the page. Columns, footnotes, tables, formulas and captions each sit somewhere specific, and if any step reads the order wrong, the translation comes back scrambled. Geyi joins parsing, translation and layout rebuilding into one pass: upload a whole PDF or academic paper and get back a translation you can use as is, with the layout, math and figures kept in place when the PDF has a text layer.

Upload a PDF to start translating

There are two routes through a PDF, and which one fits depends on the original. Digital PDFs with an intact text layer can go through layout-preserving mode, which replaces text page by page without touching the design. Scans, photographs and messily typeset files go through AI reflow, which recognises the text first and lays it out again.

Geyi PDF translation joins parsing, translation and layout rebuilding into one pass over a whole PDF. Digital PDFs with a text layer can keep their original pages; scans, photos and messy layouts go through AI reflow, which recognises the text first and lays it out again.

Key capabilities

Layout-preserving mode: pagination, tables and images match the original PDF
AI reflow for scans and photos, exporting Word, PDF, Markdown or EPUB
Optional bilingual PDF and Word, interleaved paragraph by paragraph
Every translated paragraph tagged with its source page
High precision adds AI polishing and drift checking

Choosing between the two routes

There is one test: can you select the text with your mouse in a PDF reader? If you can, there is a text layer, layout-preserving works, the translation matches the original page for page, figures and formulas stay put, and it costs fewer credits. If you cannot, the page is essentially an image and has to be recognised first — that is AI reflow.

One more case: the text layer exists but the layout itself is a mess, which is common in old archives that were scanned and then OCR'd. AI reflow usually reads better there, because it rebuilds the reading order rather than following a layout that was already wrong.

What you get back

Layout-preserving: still a PDF, with pagination, table positions and image positions matching the original.
AI reflow: export Word, PDF, Markdown or EPUB, whichever suits what you do next.
Bilingual output (optional): an extra PDF and Word pair with original and translation interleaved paragraph by paragraph, so proofreading needs only one window.
Each translated paragraph is tagged with the page it came from, so citations trace straight back to the source.

Which precision to pick

High precision adds an AI polishing and drift-checking pass: after the first translation it reworks wording and sentence order paragraph by paragraph and verifies nothing has drifted from the original meaning. It suits publishing, academic work and formal delivery. Standard runs in one pass — faster and cheaper, and fine for skim-reading, drafts and bulk work.

Both tiers handle layout identically; the difference is only how far the prose itself is polished.

Before uploading

Sign in to upload and submit a translation. File size and concurrent-task limits follow your account allowance and the upload panel. Choose from the source languages, target languages and output formats available in the tool. Check estimated credits and your balance before submitting; translation is not an unlimited free service.

Layout and content limits

Original Layout and AI Reflow produce different outputs. OCR, complex tables, text inside images and formulas may need manual checking. Review the public examples and preview the result before using it for formal delivery.

The two PDF modes

	Layout-preserving	AI reflow
Suits which originals	Digital PDFs whose text can be selected; cleanly typeset papers and publications	Scans, photos, messy layouts, or PDFs that need a different output format
What the translation looks like	Original pagination, text layer replaced in place	Rebuilt layout, reading order restored
Export formats	PDF	Word, PDF, Markdown, EPUB
Credits per page, high precision	4 credits/page	8 credits/page
Credits per page, standard	2 credits/page	4 credits/page

Geyi compared with generic machine translation and a manual workflow

Aspect	Geyi	Generic machine translation	Manual workflow
Formulas and layout	Layout-preserving mode keeps the original pagination and replaces text in place, leaving formulas, tables and images where they are; AI reflow re-typesets in reading order and keeps formulas intact	Often outputs plain text or simplified formatting; formulas and tables can shift, and the layout often needs fixing by hand	Copied text often has broken lines and the wrong order; formulas, tables and captions are rebuilt one by one
Scanned documents (OCR)	Scanned PDFs without a text layer are recognised first, then translated; multi-column pages follow reading order, and pages that fail recognition are not billed	Whether scans and images can be read varies by tool; the text often has to be recognised separately first	Run a separate OCR tool, proofread each page, then translate
Long documents	Upload the whole file once; it is processed in segments in the background and keeps running if you close the page; the file size limit depends on your plan	Often limits text length or file size per request, so long documents tend to be split into batches	Copy and paste section by section; the longer the document, the more time it takes and the harder terms are to keep consistent
Zotero	Right-click an item in Zotero 7 through 9 to translate it; the translated PDF is attached to the same entry; the plugin needs a Research plan or above	Usually not connected to Zotero; files are exported, translated and imported back	Export the PDF, translate it, then attach the result to the item by hand
Right-to-left languages (RTL)	Arabic, Hebrew and Persian are available as target languages; right-to-left layout is applied to PDF output only, where the direction of digits and Latin text is still being improved, so preview before use; Word, EPUB and other formats do not yet set right-to-left paragraphs	The text can be translated, but right-to-left layout and embedded digits or Latin text often need manual adjustment	Re-typeset in software that supports right-to-left layout and check text direction throughout
Billing	Credits per page, with free credits on sign-up; the estimated charge is shown before you submit, and the final charge follows the parsed page count	Usually billed by characters or by subscription; rules vary by tool	The main cost is working time, which grows with length

The generic machine translation and manual workflow columns describe common cases; actual results vary by tool and method. The Geyi column lists only capabilities that are live in the product today.

Translating a PDF in four steps

Upload Drop the whole PDF onto the file translation page — no splitting pages, no pasting paragraphs. The per-file size cap depends on your plan.
Pick the modes Choose a layout mode (preserving or reflow) and a precision tier (high or standard), and tick bilingual output if you want a side-by-side version.
Check the cost The estimated credits for your file appear before you submit. Billing is per page, with a 10-credit minimum per file.
Collect the result The job runs in the background, so you can close the page. When it finishes, preview it online or download your chosen format from History.

See it in action

The first three pages of NASA’s magnetic-gear paper are processed through layout-preserving and reflowed bilingual pipelines, showing two columns, a specification table, captions and alternating English and Simplified Chinese paragraphs.

Page 2 of the English original: two-column body text, the E-DRIVES RIG section and Table 1 (test-rig specifications) — Layout-preserving translation, page 2: the two-column text and the caption and row labels of Table 1 (specifications) are translated into Chinese in their original positions; the table rules are kept.

Chinese page 2: two-column body text, the E-Drives Rig section and Table 1 (test-rig specifications) in the right column, in Chinese — Layout-preserving translation, page 2: the two-column text and the caption and row labels of Table 1 (specifications) are translated into Chinese in their original positions; the table rules are kept.

Page 3 of the English original: labelled photo of the test rig (Fig. 1), torque-uncertainty contour plot (Fig. 2) and two-column text — Layout-preserving translation, page 3: the Fig. 1 and Fig. 2 captions and the two-column text are translated into Chinese in their original positions; English labels inside the images are not translated.

Chinese page 3: the labelled test-rig photo (Fig. 1) and the torque-uncertainty contour plot (Fig. 2), with Chinese captions and two-column Chinese text — Layout-preserving translation, page 3: the Fig. 1 and Fig. 2 captions and the two-column text are translated into Chinese in their original positions; English labels inside the images are not translated.

Before

The prototype’s response was measured at output speeds between 124 rpm and 744 rpm for a controlled output torque of 10 Nm (8% of the prototype’s maximum torque). After correcting for tare losses, the prototype’s efficiency was found to decrease from 90.0% to 83.0% as speed increased.

After

在受控输出转矩为10 Nm（原型机最大转矩的8%）的条件下，测量了该原型机在124 rpm至744 rpm输出转速范围内的响应。校正空载损耗后发现，随着转速升高，原型机的效率从90.0%降至83.0%。

Abstract excerpt (layout-preserving output, not hand-edited): units, figures and the glossary terms for “tare loss” and “torque” are applied consistently.

Generated through Geyi’s pipeline, then checked against the source and edited for display. These are excerpts; translations are in Simplified Chinese. · Source: Dynamic Testing of a High-Specific-Torque Concentric Magnetic Gear, Justin J. Scheidler, Zachary A. Cameron, Thomas F. Tallerico (NASA Glenn Research Center), 2019 (U.S. Government work, not subject to copyright: the NTRS record's copyright determination is GOV_PUBLIC_USE_PERMITTED, and the footnote on page 1 states “This is a work of the U.S. Government and is not subject to copyright protection in the U.S.” All three authors are NASA Glenn Research Center research engineers.)

Frequently asked questions

What does a 100-page PDF cost?

It depends on mode and precision. Layout-preserving on high precision is 4 credits a page, so about 400 for 100 pages; AI reflow on high precision is 8 a page, about 800. Standard is half of each. The exact estimate for your file is shown before you submit.

Can scanned PDFs be translated?

Yes. Scans have no text layer, so they go through AI reflow, which recognises the characters before translating. Clean print scanned straight gives the best results; where the image is blurry or largely handwritten, check against the original before exporting.

Is the result still a PDF?

In layout-preserving mode, yes, and it matches the original layout. In AI reflow mode you can export Word, PDF, Markdown or EPUB, whichever fits what you do next.

Do formulas and tables survive?

In layout-preserving mode formulas, figures, headers and footers stay exactly where they were, because only the text layer is replaced. AI reflow rebuilds the layout and places figures according to the restored reading order.

Why is layout-preserving PDF-only?

It replaces text inside the original PDF and reuses its pagination, so both input and output have to be PDF. Use AI reflow when the original is another format or you need a different output.

Can I translate only some pages?

Jobs are submitted per file and billed on the actual page count. If you need only part of a document, export those pages into a new PDF first and upload that — the charge drops accordingly.

Do I have to wait for a large file?

No. Jobs run in the background once submitted, so you can close the page and collect the result from History. How many files run at once depends on your plan.

Can I pin specialist terms?

Yes. The personal glossary in your account is applied automatically and takes effect on the next file as soon as you edit it.

What if a file exceeds a limit or fails?

For oversized files, split the document or check your account allowance before retrying. For parsing failures, check for damage, encryption or DRM. Read the task failure reason, then use the offered retry action or contact support with the task ID. A processing task does not yet have a downloadable result.

简体中文 · Français · Deutsch · Русский · 日本語 · 한국어 · العربية · 繁體中文

Geyi is a proprietary AI document platform owned and operated by Nayuta Technology and Culture Company Limited.