Remove a Watermark From a Scanned Document: Yes — but Watch the Text Underneath
Yes, you can remove a watermark from a scanned document in a browser tool — if the scan is a JPG or PNG image and the mark sits on plain paper. We built three test pages and ran the tool's actual inpainting engine (Telea fast-marching, the same one the watermark remover uses) over a faint rotated watermark. On blank paper the mark came out with 0% of its ink remaining (RMSE 1.5 against a clean copy — essentially pixel-exact). But there is a scan-specific trap: when the watermark crosses printed text, 100% of the text pixels underneath it were erased, because inpainting rebuilds the paper, not the letters. And on a scanned photo the mark was only half gone (44% ink remaining). One hard limit: the tool only opens image files (PNG, JPEG, WebP, BMP, GIF) — a PDF scan is not supported, so rasterize it to an image first.
How we tested it (a real run, not a guess)
A scanned document is mostly black text on white, often with a faint diagonal 'SAMPLE' mark or a stamp. We generated a scanned test page — white paper, printed body text, plus a grayscale photo patch — stamped it with a semi-transparent rotated SAMPLE watermark, then ran the inpainting engine over the mark footprint and compared the result to a clean, unwatermarked copy. The engine is the same Telea fast-marching fill (brush radius 24) the browser tool applies; we dilated the mask by the brush radius to mirror how a painted brush actually covers pixels. Every number below comes from that run.
Result 1: on blank paper, the mark just disappears
Where the watermark landed on empty margin, the fill reconstructed the paper cleanly. The mark was gone and the surrounding paper matched the untouched original almost exactly.
| Region | Watermark ink remaining | RMSE vs clean | Mean residue |
|---|---|---|---|
| Blank paper (no text under mark) | 0.00% | 1.5 | 1.14% |
This is the easy case: a diagonal watermark over a clean white area is exactly what inpainting is good at. Treat any mark on a blank edge of a scan as a near-free cleanup.
Result 2: when the watermark crosses text, the text is the casualty
This is the part every scan guide skips. A real watermark rarely stays in the margin — it usually runs straight through the lines you want to keep. We measured what happened to the text under the mark, not just the mark itself.
| Region | Watermark ink remaining | RMSE vs clean | Text pixels under mark | Text destroyed |
|---|---|---|---|---|
| Watermark over printed body text | 7.53% | 20.78 | 2,743 | 100% |
Read that carefully: the watermark band looks clean because the SAMPLE mark is gone, but every one of the 2,743 text pixels that sat under it was replaced by paper fill. Inpainting does not know there was text there — it only fills the hole with neighboring pixels. So the mark leaves, and the sentence under it leaves with it. The 7.53% 'ink remaining' figure is the mark band measured against the original; it is low precisely because the band is now blank paper, not because the text survived.
What to do about it: paint only the watermark, never the text, and zoom in so your brush stays on the mark. If a line of text is partly under the mark, accept that you will lose those few letters and re-type them — no pixel tool recovers text it has painted over. For general text-over-image cleanup outside scans, our remove-text test covers the same engine on captions and overlays.
Result 3: a scanned photo keeps about half the mark
Scans often include a grayscale photo (a figure, a signature photo, an attached snapshot). On that patch the engine could not reconstruct the lost detail.
| Region | Watermark ink remaining | RMSE vs clean | Mean residue |
|---|---|---|---|
| Grayscale scanned photo patch | 44.38% | 30.43 | 19.44% |
Almost half the watermark ink stayed visible and the error against the original was high — the fill smeared the photo detail instead of recovering it. This is the same behavior our screenshot test and photo tests found on busy backgrounds: inpainting handles flat paper and simple graphics, not rich imagery. If the mark is on a photo inside the scan, expect a visible patch and plan to crop or accept it.
The PDF problem: a scanned-document tool that won't open a PDF
Most scans arrive as PDF, not images. Here is the hard limit, confirmed directly in the tool's code: the file picker accepts only image/png, image/jpeg, image/webp, image/bmp, image/gif (the page's input accept attribute). There is no PDF reader in the code, so a PDF scan will not open — and that is by design, not a bug. Before you can clean a mark, turn the PDF page into an image: print the PDF to PNG, or run it through any PDF-to-image step, then load the resulting JPG/PNG into the tool. A photographed scan (a phone picture of the page) is already an image and drops straight in.
Step-by-step on a scanned page
- Get the page as an image. JPG or PNG from a scan app, a phone photo, or a PDF-to-image export. Keep it at the scan's full resolution — the tool works at native pixels and does not downscale.
- Load it. Nothing leaves your device; the image is drawn onto a local canvas (see our no-upload test).
- Paint the mark, not the text. Zoom in so the brush covers only the watermark. This is the single most important step on a scan — painting over text is what destroys it.
- Remove, then download. The fill runs on the painted region. Export as PNG to avoid re-compressing a JPEG and adding new artifacts (our JPEG test explains why).
If you need the text back
The tool is a pixel editor, not an OCR. If the watermark sat over text you care about, run OCR on a clean copy of the page first (a PDF or a scan app's text layer) and save that text separately, then clean the image for display. Don't expect the watermark step to return letters it painted over. For marks that are logos or stamps rather than diagonal text, our logo test and signature test cover those shapes on flat and textured backgrounds.
FAQ
Can I remove a watermark from a scanned PDF directly? No. The tool opens PNG, JPEG, WebP, BMP, and GIF only — confirmed in its file picker. Convert the PDF page to an image first.
Will removing a watermark from a scanned document keep my metadata? No. Like every in-browser edit, the export re-encodes through a canvas, which drops EXIF and other metadata — see our EXIF test and our metadata-clearing guide.
What if the watermark is over text? Paint only the mark and zoom in; any text you paint over will be erased, not recovered. Re-type the lost characters afterward.
Why is half the mark still there on a photo? Inpainting reconstructs flat or simple regions well but cannot recreate photo detail; the residual mark on a scanned photo is expected, not a failure of the run.