Converter test
PDF to EPUB converters tested: Calibre (CloudConvert), Marker, Docling and WandPDF on 31 real PDFs
We took 31 public PDFs - novels, papers, manuals, laws and government reports in 8 languages - and ran each through Calibre, Marker, Docling and our own converter, WandPDF. Then the same automatic check counted the books a reader would call broken: paragraphs cut in half, page numbers in the middle of the text, lost words.
Results: books with no major defects
The short version: Calibre, the tool most guides point to and the engine behind CloudConvert, copies the printed page line by line. Novels keep their line-break hyphens and page numbers, and a two-column paper can come out with the two columns interleaved (see the BERT passage below). Marker and Docling read the layout with neural models and do much better. Marker's most common failure is a paragraph split at a page break; Docling's most common are running heads and garbage characters left in the text. WandPDF had the most clean books, but 6 of 31 still failed. If you are weighing Calibre against WandPDF, the Calibre alternative page covers the differences beyond this test.
The converters
WandPDF
this site
Rebuilds the book from the position of every glyph in the PDF, with rules and small layout models on our own CPU server.
The same pipeline as the converter on this site, EPUB output.
Calibre 8.5
free desktop ebook manager
Its ebook-convert command is the usual free way to turn a PDF into an EPUB. CloudConvert uses this same version.
ebook-convert book.pdf book.epub
Default settings. With --enable-heuristics, as CloudConvert runs it, also 2 clean.
Marker 2.1.0
open source, by Datalab
Converts PDFs to Markdown, HTML and JSON with neural layout and reading-order models; aimed at AI pipelines.
marker_single book.pdf --output_format html --mode fast --disable_ocr --disable_image_extraction
CPU, no LLM mode.
Docling 2.145.0
open source, by IBM Research
Converts documents to Markdown, HTML and JSON with a layout model and TableFormer for tables; aimed at AI pipelines.
docling book.pdf --no-ocr --to html
CPU.
Marker and Docling write Markdown, HTML or JSON rather than EPUB, so we scored their HTML: the check reads the text, not the packaging. Marker can also send pages to an LLM for better results; we left that off, and OCR was off for every converter. Every book in the set has a text layer.
Calibre has a heuristic processing option meant to repair this line-by-line output. With --enable-heuristics it was clean on 2 of the 31 books; 3 books did not finish within our 15-minute limit and count as failures. This is how CloudConvert runs Calibre (see limits). Its free tier stops a conversion after 5 minutes, so those 3 books would fail there too, and on the free tier the count stays at 2.
What goes wrong, and how often
Number of books, out of 31, where the check found each kind of defect. One book can have several; the definitions are in how the check works.
| Defect | WandPDF | Calibre | Marker | Docling |
|---|---|---|---|---|
| Paragraphs split mid-sentence | 2 | 20 | 9 | 4 |
| Page numbers in the text | 1 | 15 | 3 | 5 |
| Running heads and footers | 1 | 15 | 2 | 7 |
| Lost text | 2 | 5 | 3 | 2 |
| Line-break hyphens | 0 | 6 | 0 | 0 |
| Lines not joined | 0 | 8 | 1 | 0 |
| Garbage characters | 0 | 6 | 5 | 6 |
| Broken Arabic or CJK text | 1 | 1 | 2 | 2 |
Calibre's PDF to EPUB output vs WandPDF: three passages
Three passages as they come out of Calibre and WandPDF. They come from books we use to develop WandPDF, not from the 31 test books: we don't open the test books, so they stay a fair test.
A two-column paper: BERT (Devlin et al., 2018), page 1
2018). These include sentence-level tasks such as In this paper, we improve the fine-tuning based
natural language inference (Bowman et al., 2015; approaches by proposing BERT: Bidirectional
Williams et al., 2018) and paraphrasing (Dolan Encoder Representations from Transformers.
In this paper, we improve the fine-tuning based approaches by proposing BERT: Bidirectional Encoder Representations from Transformers.
Calibre reads straight across the page, so every line glues the left column to the right one.
A programming book: Think Python, pages 185-186
19.4 any and all
Python provides a built-in function, any, that takes a sequence of boolean values and re-turns True if any of the values are True. It works on lists:
186 Chapter 19. The Goodies
>>> any([False, False, True])
True
19.4 any and all
Python provides a built-in function, any, that takes a sequence of boolean values and returns True if any of the values are True. It works on lists:
>>> any([False, False, True]) True
The page number and running head land in the text, the line-break hyphen stays, and the code becomes two lines of ordinary text.
A novel: The Time Machine, H. G. Wells
6 The Time Machine
‘Not exactly,’ said the Medical Man. ‘There are bal-loons.’
‘But before the balloons, save for spasmodic jumping and the inequalities of the surface, man had no freedom of verti-cal movement.’
‘Not exactly,’ said the Medical Man. ‘There are balloons.’
‘But before the balloons, save for spasmodic jumping and the inequalities of the surface, man had no freedom of vertical movement.’
On an e-reader the running head and the hyphens turn up mid-line every few pages once the text reflows.
All 31 books
The test set is fixed: public PDFs picked for variety, never used to tune WandPDF. When we had to look into one of them to understand a failure, it moved to our development set and a fresh PDF replaced it (this has happened twice).
| Book | WandPDF | Calibre | Marker | Docling |
|---|---|---|---|---|
| Bayesian Inference for Dynamic Spatial Quantile Models with Interactive Effects (arXiv)Statistics paper · English · 30 pages | Clean | Failgarbage characters, page numbers in the text, running heads and footers, line-break hyphens, paragraphs split mid-sentence, lost text | Clean | Clean |
| How Do You Sleep? (Book Dash)Picture book · English · 18 pages | Clean | Failrunning heads and footers, paragraphs split mid-sentence | Clean | Clean |
| CC Certificate Course HandbookHandbook · English · 22 pages | Clean | Failparagraphs split mid-sentence | Clean | Failpage numbers in the text |
| Les licences Creative CommonsPoster · French · 1 page | Clean | Clean | Clean | Clean |
| A Christmas Carol, Charles DickensNovel · English · 106 pages | Clean | Failline-break hyphens, paragraphs split mid-sentence | Failparagraphs split mid-sentence | Failgarbage characters |
| Constitución Española (BOE)Law · Spanish · 39 pages | Clean | Failrunning heads and footers | Clean | Clean |
| Full Circle Magazine #200Magazine · English · 67 pages | Faillost text | Failparagraphs split mid-sentence, lost text | Failgarbage characters, running heads and footers, lines not joined, lost text | Faillost text |
| GNU grep manualSoftware manual · English · 44 pages | Clean | Failparagraphs split mid-sentence | Clean | Failrunning heads and footers |
| Le Grand Meaulnes, Alain-FournierNovel · French · 423 pages | Clean | Failpage numbers in the text | Failparagraphs split mid-sentence | Failpage numbers in the text |
| The Great Gatsby, F. Scott FitzgeraldNovel · English · 193 pages | Clean | Failrunning heads and footers, line-break hyphens | Failparagraphs split mid-sentence | Failgarbage characters |
| Heart of Darkness, Joseph ConradNovel · English · 131 pages | Clean | Failrunning heads and footers, line-break hyphens, paragraphs split mid-sentence | Failparagraphs split mid-sentence | Clean |
| IRS Publication 501Government guide · English · 31 pages | Clean | Failrunning heads and footers, paragraphs split mid-sentence | Failpage numbers in the text, paragraphs split mid-sentence | Failrunning heads and footers |
| LibreOffice: Using Styles and TemplatesSoftware manual · English · 21 pages | Clean | Failrunning heads and footers, lines not joined | Clean | Clean |
| LLaMA: Open and Efficient Foundation Language ModelsPaper · English · 27 pages | Failpage numbers in the text | Failpage numbers in the text, paragraphs split mid-sentence, lines not joined | Clean | Clean |
| The Not So Short Introduction to LaTeX (Chinese)Software manual · Chinese · 139 pages | Failparagraphs split mid-sentence | Failgarbage characters, page numbers in the text, running heads and footers, lines not joined, lost text | Failgarbage characters, broken Arabic or CJK text, page numbers in the text | Failgarbage characters, broken Arabic or CJK text, paragraphs split mid-sentence |
| GNU tar: an archiver toolSoftware manual · English · 270 pages | Clean | Failpage numbers in the text, running heads and footers, line-break hyphens, paragraphs split mid-sentence | Failgarbage characters | Failpage numbers in the text |
| The Merchant of Venice, Arabic translation by Elias FayyadPlay · Arabic · 60 pages | Failbroken Arabic or CJK text, lost text | Failbroken Arabic or CJK text, page numbers in the text, running heads and footers, lines not joined, lost text | Failbroken Arabic or CJK text, lost text | Failbroken Arabic or CJK text, running heads and footers, lost text |
| The Restaurant of Many Orders, Kenji MiyazawaShort story · Japanese · 6 pages | Clean | Failparagraphs split mid-sentence | Clean | Clean |
| NASA: Propulsion With the Space Launch SystemEducator guide · English · 48 pages | Failrunning heads and footers | Failgarbage characters, page numbers in the text, running heads and footers, paragraphs split mid-sentence | Failgarbage characters | Failgarbage characters, running heads and footers |
| NIST SP 800-63BStandard · English · 80 pages | Clean | Failgarbage characters, page numbers in the text, running heads and footers, paragraphs split mid-sentence | Clean | Failgarbage characters |
| Open Research (OER Hub)Book · English · 79 pages | Clean | Failrunning heads and footers, paragraphs split mid-sentence | Failparagraphs split mid-sentence, lost text | Clean |
| Domain Adaptation in NMT using a Qualia-Enriched FrameNet (LREC 2022)Paper · English · 12 pages | Clean | Failpage numbers in the text, paragraphs split mid-sentence | Failparagraphs split mid-sentence | Clean |
| The Queen of Spades, Alexander PushkinShort story · Russian · 16 pages | Clean | Failpage numbers in the text, paragraphs split mid-sentence | Clean | Failparagraphs split mid-sentence |
| 303 Creative LLC v. Elenis (U.S. Supreme Court)Court opinion · English · 70 pages | Clean | Failgarbage characters, page numbers in the text, running heads and footers, line-break hyphens, paragraphs split mid-sentence | Clean | Failpage numbers in the text, running heads and footers |
| Sentence-BERTPaper · English · 11 pages | Clean | Failparagraphs split mid-sentence | Clean | Clean |
| Sugar the Robot and the Race to Save the EarthChildren's novel · English · 102 pages | Clean | Failpage numbers in the text | Failparagraphs split mid-sentence | Failparagraphs split mid-sentence |
| Sendov's Conjecture for Sufficiently High Degree Polynomials, Terence TaoMath paper · English · 38 pages | Clean | Failpage numbers in the text, running heads and footers, lines not joined, lost text | Clean | Clean |
| UK DfT TAG unit A4.2: Distributional Impact AppraisalGovernment guidance · English · 85 pages | Failparagraphs split mid-sentence | Failgarbage characters, page numbers in the text, paragraphs split mid-sentence, lines not joined | Failgarbage characters, paragraphs split mid-sentence | Failgarbage characters |
| La Lupa, Giovanni VergaShort story · Italian · 46 pages | Clean | Failparagraphs split mid-sentence, lines not joined | Clean | Failrunning heads and footers, paragraphs split mid-sentence |
| Later Poems, W. B. YeatsPoetry · English · 244 pages | Clean | Failpage numbers in the text, lines not joined | Failpage numbers in the text, running heads and footers | Failpage numbers in the text, running heads and footers |
| The Yellow Wallpaper, Charlotte Perkins GilmanShort story · English · 15 pages | Clean | Clean | Clean | Clean |
How the check works
For each book the check takes the text layer of the PDF itself as the reference, then reads the converter's output and looks for what makes a reflowed book hard to read. It applies the same rules to every converter's output, and it does not judge how the book looks.
- Paragraphs split mid-sentence. A paragraph stops in the middle of a sentence and the rest starts a new one, usually at a page break. Fails when more than about one paragraph in 50 is cut.
- Page numbers in the text. Printed page numbers left standing between paragraphs. Fails at two or more.
- Running heads and footers. The book title, chapter name or footer repeated through the text. Fails at one or more.
- Lost text. Stretches of 12 or more consecutive words from the PDF's text layer that never reach the output. Fails when they add up to more than 0.3% of the book's words, and more than 20.
- Line-break hyphens. Words still split where the printed line ended, like "re-turns". Fails at more than 5, plus 1 per 1,000 words.
- Lines not joined. Every printed line becomes its own paragraph. Fails when paragraphs average under 12 words, in books with 200+ paragraphs.
- Garbage characters. (cid:12), replacement characters, private-use glyphs or unexpanded ligatures that break search. Fails at one or more.
- Broken Arabic or CJK text. Arabic words cut apart by spaces, or spaces inserted between Chinese or Japanese characters. Fails above 1% of the Arabic words (and more than 5) or of the Chinese and Japanese characters.
A book fails if any one of these is found. The thresholds leave room for what authors do on purpose: a year on a title page is not a page number, and "down-stream" next to "downstream" is often the author's own spelling.
What this test does not show
- We make WandPDF and we wrote the check. It counts the defects we think matter most in an ebook. Another check would rank the converters differently.
- Tables, formulas and heading levels are not scored. Marker and Docling are better than WandPDF at borderless tables and at formulas, and this test does not show that.
- Marker and Docling are built for AI pipelines, not ebooks. They aim at Markdown and JSON for language models, where a stray page number costs less than in a novel.
- We did not upload the test books to online converters. For CloudConvert there was no need: its PDF to EPUB converter runs Calibre 8.5, the only engine its public API lists for this conversion (checked November 2026). A PDF of our own that we converted there came out word for word the same as from
ebook-convert --enable-heuristics, and with that option Calibre was clean on 2 of the 31 books, as above. Its separate "Convert with AI" option was not tested. Convertio and Amazon's Send to Kindle were not tested either, because that would mean uploading the books to them. MinerU is left out because its current version sends documents to its cloud service by default, and an earlier version took about 20 minutes per book on our CPU. - 31 books is a small sample. One book is about 3% of the set. Scanned PDFs are left out: all four converters ran without OCR.
How to repeat it
All the books are public PDFs, and the commands for Calibre, Marker and Docling are in the table above. We ran every converter on the same 8-core CPU server with no GPU, one book at a time. The check is part of WandPDF's code, which is not open source, so you can repeat the conversions but not our scoring of them.
Where your file goes
WandPDF converts on its own server, with its own code and models. Your PDF is not sent to an AI service or to any other third party. It is deleted as soon as it is converted, and the converted book within an hour (a week if you sign in). Calibre, Marker and Docling all run on your own machine.