A watermark tiled across a picture is damage laid over something you already have the right to see. Watermark Remover learns that pattern from your photos and lifts it off, on the phone, without sending them anywhere.
Listed on Google Play as Watermark Remover (Offline).
Five steps
Add, remove, check, refine, save.
Screens from version 0.5.0 in Easy mode, on original drawn test pictures (no real photos). The ideas behind the learning step are on How it works.
1Add. Tap Add for pictures that carry the same mark — and clean originals or a watermark file if you have them. The app sorts them itself and asks only when unsure; Skip is always safe.
2Remove. Tap Remove the watermark. Learning and removal run on the phone, with progress on screen and a Cancel button. Time depends on the phone, the picture size and the mark.
3Check. Drag the wipe across each result to compare before and after. “Mark remaining” is an estimate relative to the input, not proof of a clean result.
4Refine if needed. Tap Not perfect to touch up the learned mark with Paint, Erase and Pick colour, then apply it to all the pictures again.
5Save. Save the results to Pictures or share them — after you have looked at each one.
Measured, not claimed
Before and after, with the score.
Examples use test images whose clean original is known, so every result can be scored over the whole picture. No score on this site comes from a real photo.
Have the watermark file? Open it and remove the mark from any photo. Measured on 21 kinds of mark, two pictures each: 41 of 42 ended closer to the true picture, 1 further, 0 declined. It works when the file matches how the mark was applied, such as a mark the app learned and saved. One of them:
Heat map: brighter = further from the true image; same scale on every example. Dark means
the pixel matches the true picture; full amber means it is off by 32 or more of 255 levels in
some colour channel.
Tiled embossed text
The watermark file, opened in the app.
Input
Restored
Change against the true picture: 24.58 → 43.38 dB (+18.80).
The app needs to see the mark more than once: repeating across your photos, on two or more of them. With one copy there is nothing to compare.
Handles
Repeating or tiled semi-transparent watermarks spread across the photo.
A mark learned once from several of your photos that carry it (5–8 work best), saved, and removed from new photos with the same mark — one at a time or many. A watermark file or a clean original helps when learning from marked photos alone is hard. With one photo and no usable watermark (0.6.0), a model on the phone estimates the mark and removes it; the median improvement on our test set was +6.5 dB (examples). You can also paint over the area instead.
Results that are not right yet: tap Not perfect and refine the learned mark with paint and erase tools, then apply it again.
Does not
Guarantee a clean result from one photo: the one-photo model (0.6.0) improved the picture in most of our tests but sometimes leaves a faint outline, and about 1 in 150 test pictures ended with a damaged patch (examples). Learned from 8 photos and saved, single marks did better.
Recover what a fully opaque mark hid: that area is filled in, and the fill is a reconstruction, not the original.
Promise results on other kinds of marks. Check each result before you save it.
Privacy
The photo stays on the phone.
Photos and results are processed on the device and never uploaded.
Learning and removal need no connection, and the app contains no advertising, analytics or crash-analytics SDK.
Reports the app can send — feedback you write, crash and memory reports, Hardware Test results, a self-test on built-in sample images once per install — carry no photos, and each automatic one has a switch in Settings. Opening the app makes no network connection of its own.
