Deskew
Open the app: Straighten a scan → (products.documentize.ai)
In one sentence
Deskew looks at every page of a scanned PDF, measures how much the text lines are tilted, and turns the page back until they are level.
What makes it unique
- It decides page by page. Each page gets its own angle. A pack where every sheet went through the scanner a little differently is fixed in one go; pages that are already straight are left alone.
- Small angles are its business. It looks for tilt within roughly ±15°. It will not turn a page that is lying on its side — that is Rotation's job.
- It never over-corrects a straight page. Below about 0.3° it does nothing at all to the page (no warp, no re-encoding), because the correction would be invisible and only cost quality.
- It shows you what it did. For the most skewed pages (tilt above 0.5°) the result panel shows a small corrected preview, up to three of them.
- Fast and parallel. Several pages are worked on at once, with progress shown as "Deskewing page X of Y".
- Unlike Rotation (which turns a page by a fixed angle you choose, or auto-uprights it), Deskew measures a free angle and fills the corners that the turn uncovers with the colour of the paper.
When to use it / when not to
- Use it for flatbed or phone scans where the page is a few degrees off, before OCR or before archiving.
- Not for sideways or upside-down pages → Rotation first.
- Not for photographs, drawings or very dark/very empty pages. The detector needs a normal amount of dark text on light paper. A blank page, or a page that is more than half ink, is left uncorrected (the page keeps angle 0).
- Not for perspective problems (a phone photo taken at an angle, curved book pages). Deskew only rotates; it does not flatten trapezoids.
- Not if your PDF's pages are born-digital text (exported from Word). They are not crooked; there is nothing to do.
Before you start
- Fix pages that are on their side first.
- Know that the output is made of page images (see "Quality & limits"). If you need searchable text, plan Searchable afterwards.
- The site's FAQ says images (JPEG, PNG) can be uploaded as single-page PDFs, and that the free tier takes PDFs up to 100 MB (about 500 pages), with higher tiers allowing more and batch upload. Treat those as the site's statement.
- Password-protected PDFs need their password; otherwise unlock first.
Step by step
- Open the Deskew page and drop your scanned PDF.
- Press the upload/deskew button. There are no options to set.
- Wait while it reports progress page by page.
- Look at the preview of the most-corrected pages, then download the straightened PDF.
Options explained
| Option | What it does | Default | Recommended |
|---|---|---|---|
| (none) | There is nothing to configure. The angle is detected automatically for every page | – | Just upload |
What you get
One PDF named Deskew_<original name>.pdf (if you upload several files they come as a ZIP, Deskewed_documents). The page count and each page's size stay as they were. The original is never modified.
Quality & limits
- Pages are rebuilt as pictures. Each page is placed back into a fresh PDF as an image: straight pages as they were, corrected pages re-encoded (JPEG at quality 85, or lossless PNG for pure black-and-white scans). Expect no visible loss on a normal scan, but this is a re-encode, not a lossless edit.
- Text layer — read this carefully. The site's FAQ says the existing searchable text is preserved. The way the worker is built (it re-assembles each page from the page image only), do not count on a text layer surviving. Treat the output as an image PDF and run Searchable afterwards (verify on the live app with your own file).
- Accuracy: the site's FAQ claims sub-degree precision (about ±0.1°). In practice it will not fix what it cannot measure: pages with little text or no clear lines are left as they are.
- Very large page images (above about 40 million pixels) are scaled down before correction, to protect memory.
- If one page fails, the original image of that page is kept, so the page count never changes. If no page can be processed at all, you get an error.
- An empty PDF gives the message "No content found", not a crash.
- A free-plan time limit can stop a large batch early and leave a continuation to finish later.
- Results are kept for about two days.
- The marketing text mentions a searchable index of detected angles; the worker only returns the previews described above.
Works well with
- Before: Rotation, Merger (photos → one PDF), Recovery.
- After: Searchable (OCR), Resize, Compressor.
- Recipes: Photos → searchable PDF, Digitise an archive, and the lab in this chapter.
Common mistakes & fixes
- "Nothing changed on my page" → its tilt is below about 0.3°, or the page has too little text for a reliable measurement. That is intended.
- "The page is still sideways" → Deskew only handles small angles. Use Rotation.
- "The text is no longer selectable" → the output is image-based; run Searchable.
- "The file got bigger" → pages were re-encoded as images; finish with Compressor.
- "Pages look cut off at a corner" → the page was turned inside a frame of the same size; a large correction can push an edge outward. Check the preview and, if needed, rescan.
Same thing elsewhere
- CLI plugin: Documentize.Deskew; it takes no options beyond the task id.
- Web API / MCP: see Beyond the browser (no dedicated MCP tool for Deskew was verified).
Try it
Scan or photograph one printed page, deliberately a few degrees off, and make a 3-page PDF with Merger (one tilted, one straight, one tilted the other way). Run Deskew and compare the previews with the pages in your PDF viewer. Then try selecting text in the result and see whether it works.