Academic Papers in EPUB: What Survives and What Breaks

A PDF-to-EPUB conversion can look clean in the first chapter, then collapse at the first results table. Body text usually reflows because it is linear prose. Data tables and multi-panel figures do not, because they depend on a visual grid that reflow destroys. EPUB 3.3, a W3C Recommendation since 13 January 2026, packages HTML, CSS, SVG and other resources into a single-file container, but the spec does not tell you which table should become real markup and which should stay an image with a caption.

This tutorial from toolkit.bot moves past the reading-order question. It focuses on structural decisions: data tables, multi-panel figures, and captions. You will work through the sections Why Body Text Reflows and Grids Do Not; The Triage Decision: Table, Captioned Image, or Repair; Convert as a Table: Simple Grids With Real Data; Convert as a Captioned Image: Multi-Panel Figures and Dense Plots; Repair by Hand: The Cases No Converter Should Claim; Captions, Alt Text and the Accessibility Question; Verifying the Output on Real E-Readers; Worked Triage on a Two-Column Journal Paper; and Conclusion: Rules for the Next Paper.

The goal is not perfect conversion. It is a practical triage for your next paper.

Why Body Text Reflows and Grids Do Not

Reflowable EPUB3 has no fixed page. Its content is HTML styled with CSS, packaged with SVG and other resources in a single-file container, so paragraphs, headings and footnotes reflow to whatever screen size and font the reader chooses. EPUB 3.3 became a W3C Recommendation on 13 January 2026, and that packaging model, built on the same file formats the web runs on, is why running text travels so well to an e-reader.

A data table is a different problem. It is a grid whose meaning depends on column alignment. Once column width is negotiable, cells squash, wrap into unreadable stacks, or drift outside the viewport. The numbers survive; the relationship between a value and its column header does not.

Multi-panel figures are harder still. Four panels with axis labels, letters and legends make one visual argument. At phone width, the letters and tick marks become illegible before the figure itself becomes inconvenient.

Be precise about the mechanism: this is a practical observation about reflowable layout, not a rule written into the EPUB standard. The W3C pages describe packaging and accessibility guidance, not how any converter treats table markup.

The consequence is that you plan per element, not per document. Triage each element before you convert; the count of failures depends entirely on document structure and source PDF quality.

The Triage Decision: Table, Captioned Image, or Repair

For each table and figure, ask one of three questions: does this become real table markup, a captioned image, or a hand repair? Sort by grid dependence, not visual prominence. Two criteria dominate: grid size (rows and columns) and whether readers need numbers as selectable, searchable text or only need to see the shape of the data. A four-column results table with short cells is a strong markup candidate. A four-panel figure belongs as an image with a real text caption. A layout table that aligns definitions is ordinary text.

Triage before converting. Print the pages or place the PDF beside the output and mark each element while the original layout is visible. A converter sees boxes and coordinates; you see which table is a results grid and which is a layout device.

Call this a working heuristic from practice, not an industry standard. No published specification sorts tables this way, and converter behaviour varies by implementation and source PDF. Keep a repair list with page numbers attached. Items marked for hand work are exactly what goes missing silently after export; page references make the gap findable. For a wider account of what breaks and what gets fixed, compare your list against the output element by element.

Convert as a Table: Simple Grids With Real Data

Once an element is marked for table conversion, the criteria are mechanical: two to five columns, short cell contents, one clear header row, numbers or brief labels, no merged cells, no nested subheaders. Results tables, comparison tables and methods-section parameter lists usually qualify. Dense numeric grids, such as The hard part: model performance tables, are where the call stops being obvious.

Real HTML table markup beats a picture of a table whenever the structure holds. It lets readers move by row and column and lets assistive technology announce position, which a flat image does not support.

Header association is the usual failure point. Cells must be marked as headers with the correct scope so a screen reader announces the header and value together rather than a stream of bare numbers; W3C technique H63 covers scope values for headers outside the first row or column.

Merged cells and spanned headers degrade first. Multi-level headers, grouped columns and cells spanning two columns most often collapse into a flat grid or an image.

Verify by reflowing, not exporting. Change the font size, rotate the device, and confirm the table wraps or scrolls horizontally instead of collapsing. Data tables may scroll in two dimensions, but the cells must stay intact. If it reads correctly at one font size only, it has not survived.

Convert as a Captioned Image: Multi-Panel Figures and Dense Plots

Multi-panel figures, log-axis plots, annotated diagrams and schematics belong here. Their labels are part of the drawing, not a grid you can rebuild. Keeping them as images is a defensible structural choice, not a conversion failure; WCAG's Images of Text criterion explicitly excludes graphs, screenshots and diagrams from its definition of image-of-text.

Keep the caption as real text beside the image. A baked-in caption is not searchable, selectable or readable by assistive technology; a text caption is all three. Wrap both in figure and figcaption so the caption is programmatically attached. A loose caption paragraph drifts from its image during reflow, leaving a description with nothing to describe. For more on where this boundary sits, see what stays as an image and what to check first.

Prefer SVG for line art where the source allows it. EPUB 3.3 packages SVG alongside HTML and CSS; rasterised line art may lose quality when text is enlarged.

Then accept the trade-off: numbers inside an image chart are no longer searchable or copyable. Where readers must quote exact values, also convert the underlying data table or transcribe the key figures into the caption.

Repair by Hand: The Cases No Converter Should Claim

Some elements sit outside all three categories. Clear triggers for manual work: a table used as page layout rather than data, a table spanning two columns in a multi-column paper, a table broken across a page boundary, equation-heavy cells, and figures whose caption sits a page away from the image.

Rejoin split tables first. When a results table breaks across pages, merge the fragments into one table and restore the header row. Left alone, your reader meets two tables, the second with no headers and no context.

Rebuild column-spanning tables as either a single-column table with wrapped cells or a high-resolution image plus a short text table of the key values. Both are more usable than a half-converted grid.

Treat nested tables and sub-tables inside cells as unlikely to survive as nested structure. Flatten them into separate tables, or split the parent table at the boundaries between logical groups.

Budget the time realistically. If most tables in a document need hand work, reflowable EPUB may be the wrong output for that document, and a fixed-layout PDF workflow is the honest answer. PDF conversion to EPUB: a practical guide helps you judge which side of that line a paper sits on.

Captions, Alt Text and the Accessibility Question

Hand repair keeps the content usable; accessibility is a separate question that now sits inside the format itself. EPUB Accessibility 1.1 is a W3C Recommendation (17 October 2024) that specifies content conformance requirements and accessibility metadata for discoverability, it is part of the EPUB family of standards, not a separate bolt-on.

The EPUB Accessibility Techniques document lists requirements such as alternative text descriptions and a correct heading hierarchy. Treat it as guidance rather than settled law: Techniques 1.2 is a Group Note not endorsed by W3C itself or its Members, so verify its current status before citing it.

Autogenerated alt text is a starting point, not a finished description. A multi-panel figure needs its panel structure, overall trend, axes and units written out; a machine-generated one-liner about pixel content will not carry that.

Where a table became an image, push the essential numbers into the caption or an adjacent text table. In practice, image-only data is a frequent gap in converted papers -- worth checking before any accessibility review.

Finally, never claim compliance. A file can include alt text, accessibility metadata and a clean heading structure and still fall short of a full conformance review. Describe what the file includes, not what it guarantees. If the paper is heading to a distributor, plan for metadata and distributor submission alongside the content work.

Verifying the Output on Real E-Readers

A file that passes an accessibility review can still render badly. Test the reflow, not the export: open the EPUB on a phone-width screen, set the largest available font, and check three things. Do tables wrap or scroll? Do figures stay legible? Do captions sit with the images they describe?

Reading applications use different rendering engines, so the same file may display cleanly on one device and awkwardly on another -- Kindle, Kobo, BOOX, PocketBook, reMarkable and Supernote are each worth testing.

Use the free preview as your test harness. toolkit.bot is a web-based PDF to EPUB3 converter built for e-reader users, researchers, students and publishers. Its PDF to EPUB3 conversion service previews the first 50 pages without an account, enough to see how tables and figures behave before committing the whole document.

Converting the full document needs a free account; single-document purchase or a Pro subscription unlock clean unmarked copies. The service handles single- and multi-column PDFs, textbooks and scanned documents via automatic OCR into reflowable EPUB3, preserving reading order, headings, footnotes, tables and figures. Test your own file rather than assuming your worst table is covered.

Compare the output against your triage list element by element: tick, repair, or accept, and record which category each failure came from. That record is what makes the next paper faster.

Worked Triage on a Two-Column Journal Paper

Take a typical two-column research article: one results table, one demographic table, three figures, one wide cross-column table.

Step one: inventory. Walk the PDF page by page and list every table and figure with its page number and a one-word tag: grid, image or unknown.

Step two: apply the criteria. Grid dependence, not visual importance, sorts them. A four-by-six results table of numeric cells becomes table markup. A three-row table aligning a definition list is page layout, so it becomes ordinary text. A four-panel figure becomes a captioned image with a text caption and structured description. The cross-column table goes on the repair list.

Step three: convert and inspect. Run the PDF through toolkit.bot, whose notes explain why two-column academic papers and newsletters reflow correctly, then open the output beside your inventory and check each element. Triage attention belongs on grids and panels, since body text and headings typically need less intervention than structured elements.

Step four: repair the leftovers. Rebuild the wide table by hand, rejoin split fragments, and recheck three conditions: largest font size, phone-width screen, and a slow re-render to catch captions settling in the wrong place.

Step five: record the outcome. Note which element type failed and how you fixed it.

Conclusion: Rules for the Next Paper

Once you have run the triage sequence, the rules for the next paper reduce to four habits.

Apply the three-bucket triage before you convert. Keep captions as text, prefer SVG for line art, check header scope on converted tables, and never let essential numbers exist only inside a picture.

Treat the reflow-versus-grid explanation and the three-way triage as practical working rules, not published standards. Test on your own device at your own font size, because rendering engines vary across reading applications.

Use the free preview to test your specific file. Repair the short list; perfect conversion is not the goal.