Scored against the truth.
Have the watermark file? Open it in the app and remove the mark from any photo. No file? Learn it from 5–8 photos and save it. In a hurry? Two photos with the same mark. Every example here is synthetic: a picture with a generated mark stamped on it, so the true clean image exists and each result is scored against it. No number on this page comes from a real photo.
Three ways to remove a mark
Have the watermark file? Open it and remove the mark from any photo.
In the Workbench (Advanced mode), step 2, choose Open a watermark file and pick the mark's file; the app finds the mark on each photo and removes it. On 42 synthetic pictures (21 kinds of mark, two pictures each): 41 ended closer to the true picture, 1 further, 0 declined; mean change +14.47 dB.
It works when the file matches how the mark was applied: a mark the app learned and saved, or your own exported mark file. A plain logo at a guessed opacity is not that: in a separate test with the logo at 50 % opacity, the pictures ended further from the truth on average (−2.01 dB over 84 pictures).
No file? Learn it from 5–8 photos.
Give the app 5–8 photos with the same mark; it learns the mark, you save it, and it becomes a watermark file for new photos. Learned from 8 photos and applied to the same pictures: 33 of 42 closer, 1 further, 8 declined; mean +4.77 dB.
Quick: two photos.
Two photos with the same mark, cleaned in one go: 34 of 42 closer, 0 further, 8 declined; mean +3.07 dB. Here the learn sees the very photos it cleans.
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.
Every example is synthetic, so the true clean picture exists and each result is scored against it; higher is closer.
Tiled embossed text
Watermark file (full strength) 
Photo 
Result Change against the true picture: picture 1: +18.80 dB; picture 2: +17.68 dB.
Faint tiled text, many small copies
Watermark file (full strength) 
Photo 
Result Change against the true picture: picture 1: +11.43 dB; picture 2: +10.07 dB.
Repeating two-tone signature: pale wash under dark script
Watermark file (full strength) 
Photo 
Result Change against the true picture: picture 1: +7.58 dB; picture 2: +7.75 dB.
Dark tiled banner text
Watermark file (full strength) 
Photo 
Result Change against the true picture: picture 1: +11.25 dB; picture 2: +11.15 dB.
Faint tiled text, blended in linear light
Watermark file (full strength) 
Photo 
Result Change against the true picture: picture 1: +11.55 dB; picture 2: +10.95 dB.
Tiled handwritten script
Watermark file (full strength) 
Photo 
Result Change against the true picture: picture 1: +9.56 dB; picture 2: +8.66 dB.
Dense, stronger tiled banner text
Watermark file (full strength) 
Photo 
Result Change against the true picture: picture 1: +16.13 dB; picture 2: +15.42 dB.
Tiled text in a saturated brand colour
Watermark file (full strength) 
Photo 
Result Change against the true picture: picture 1: +13.24 dB; picture 2: +12.71 dB.
Repeating soft pale patches with dark script
Watermark file (full strength) 
Photo 
Result Change against the true picture: picture 1: +14.38 dB; picture 2: +14.21 dB.
Clustered pale patches with dark script, repeating
Watermark file (full strength) 
Photo 
Result Change against the true picture: picture 1: +12.89 dB; picture 2: +13.49 dB.
One bold wordmark across the middle
Watermark file (full strength) 
Photo 
Result Change against the true picture: picture 1: +20.05 dB; picture 2: +20.21 dB.
A single script signature
Watermark file (full strength) 
Photo 
Result Change against the true picture: picture 1: +16.36 dB; picture 2: +15.49 dB.
A single corner logo with a handle
Watermark file (full strength) 
Photo 
Result Change against the true picture: picture 1: +17.71 dB; picture 2: +14.30 dB.
A single handle in plain text
Watermark file (full strength) 
Photo 
Result Change against the true picture: picture 1: +21.36 dB; picture 2: +17.83 dB.
A small credit line near the edge
Watermark file (full strength) 
Photo 
Result Change against the true picture: picture 1: +23.97 dB; picture 2: +18.47 dB.
One large diagonal stamp
Watermark file (full strength) 
Photo 
Result Change against the true picture: picture 1: +15.64 dB; picture 2: +14.84 dB.
A faint single corner logo
Watermark file (full strength) 
Photo 
Result Change against the true picture: picture 1: +25.34 dB; picture 2: +20.03 dB.
A single strong “made with” label
Watermark file (full strength) 
Photo 
Result Change against the true picture: picture 1: +21.26 dB; picture 2: +17.12 dB.
A small credit line, blended in linear light
Watermark file (full strength) 
Photo 
Result Change against the true picture: picture 1: +16.66 dB; picture 2: +10.98 dB.
A single embossed shadow signature
Watermark file (full strength) 
Photo 
Result Change against the true picture: picture 1: −0.15 dB; picture 2: +14.83 dB.
A single two-tone signature
Watermark file (full strength) 
Photo 
Result Change against the true picture: picture 1: +6.10 dB; picture 2: +10.53 dB.
Every case, every method
| Mark | Kind | File, photo 1 | File, photo 2 | Saved, photo 1 | Saved, photo 2 | Two photos, photo 1 | Two photos, photo 2 |
|---|---|---|---|---|---|---|---|
| Tiled embossed text | repeating | +18.80 | +17.68 | +11.28 | +12.29 | +10.63 | +11.67 |
| Faint tiled text, many small copies | repeating | +11.43 | +10.07 | declined | declined | +6.63 | +6.81 |
| Repeating two-tone signature: pale wash under dark script | repeating | +7.58 | +7.75 | +0.07 | +0.17 | +0.02 | +0.01 |
| Dark tiled banner text | repeating | +11.25 | +11.15 | +0.78 | +0.51 | +0.61 | +0.37 |
| Faint tiled text, blended in linear light | repeating | +11.55 | +10.95 | +7.00 | +6.19 | +6.54 | +5.81 |
| Tiled handwritten script | repeating | +9.56 | +8.66 | declined | declined | +5.59 | +5.44 |
| Dense, stronger tiled banner text | repeating | +16.13 | +15.42 | +0.32 | +0.19 | +0.51 | +0.26 |
| Tiled text in a saturated brand colour | repeating | +13.24 | +12.71 | declined | declined | +0.19 | +0.27 |
| Repeating soft pale patches with dark script | repeating | +14.38 | +14.21 | +2.12 | +2.41 | +2.13 | +2.44 |
| Clustered pale patches with dark script, repeating | repeating | +12.89 | +13.49 | +1.65 | +2.50 | +3.68 | +3.93 |
| One bold wordmark across the middle | single | +20.05 | +20.21 | +3.72 | +4.31 | +0.60 | +0.65 |
| A single script signature | single | +16.36 | +15.49 | +5.18 | +5.12 | +5.36 | +5.50 |
| A single corner logo with a handle | single | +17.71 | +14.30 | +11.31 | +10.86 | +4.72 | +3.59 |
| A single handle in plain text | single | +21.36 | +17.83 | +10.84 | +10.54 | +6.65 | +6.12 |
| A small credit line near the edge | single | +23.97 | +18.47 | +8.12 | +8.33 | +4.76 | +1.63 |
| One large diagonal stamp | single | +15.64 | +14.84 | declined | declined | declined | declined |
| A faint single corner logo | single | +25.34 | +20.03 | +12.72 | +12.17 | declined | declined |
| A single strong “made with” label | single | +21.26 | +17.12 | +9.90 | +9.49 | +5.05 | +3.28 |
| A small credit line, blended in linear light | single | +16.66 | +10.98 | +6.31 | +7.26 | +4.45 | +3.12 |
| A single embossed shadow signature | single | −0.15 | +14.83 | +6.97 | +7.82 | declined | declined |
| A single two-tone signature | single | +6.10 | +10.53 | −0.08 | +2.06 | declined | declined |
One photo
No second photo, no watermark file? A neural model on the phone estimates the mark's opacity and colour from the single photo and undoes it. It also works for two photos when the learner above cannot find a shared mark. The cases below were picked to cover kinds of mark, not for their scores, and one of them is modest on purpose.
- +6.5 dB median improvement of the whole picture over the input, on 300 marked test pictures (100 photographs, 3 marks each; the photographs were not used to build the model). Measured on our test set.
- 0 of 100 clean photographs altered: the same 100 photographs without a mark came back unchanged. Measured on our test set.
- 2 of 300 marked pictures ended with a visibly damaged patch (0.7 %), against 9 of 300 without the self-check, which leaves pixels untouched where two passes disagree. Measured on our test set.
- About 20 s per 12-megapixel photo on a 2025 flagship phone, about 100 s on a 2020 mid-range phone (timed on one 12-megapixel test picture). Needs about 1.2 GB of free memory and the full, non-Lite build.
One large diagonal text stamp over sky and sea
Photo with the mark 
Result, one photo Change against the true picture: +17.15 dB (whole picture 27.02 dB before, 44.17 dB after). Stamp covering a third of the picture. The faint outline that is left is visible only when you look for it.
Tiled text over a night street
Photo with the mark 
Result, one photo Change against the true picture: +10.49 dB (whole picture 32.21 dB before, 42.70 dB after).
A single script signature
Photo with the mark 
Result, one photo Change against the true picture: +10.85 dB (whole picture 32.12 dB before, 42.97 dB after). A faint ghost of the signature remains in the result.
A circular corner logo
Photo with the mark 
Result, one photo Change against the true picture: +13.44 dB (whole picture 35.36 dB before, 48.80 dB after).
A boxed diagonal stamp
Photo with the mark 
Result, one photo Change against the true picture: +12.69 dB (whole picture 33.78 dB before, 46.48 dB after).
Tiled script text, many copies
Photo with the mark 
Result, one photo Change against the true picture: +6.82 dB (whole picture 32.00 dB before, 38.82 dB after). A middling result: the mark is much fainter, but a thin outline of every copy is still visible when you look closely.
A modest case: faint tiled outline text over snow and trees
Photo with the mark 
Result, one photo Change against the true picture: +1.05 dB (whole picture 27.14 dB before, 28.19 dB after). Modest on purpose: the mark is thin outline text over fine detail, and much of it is still visible in the result.
Eight screens from the app.
Screenshots of version 0.5.0 on a phone, working on three original drawn test pictures that carry the same DEMO mark. These are real screens, not scored examples: the percentages shown are the app’s own estimates.
Easy mode: one watermark per session and one Add button for pictures, clean originals and watermark files. 
The app sorts the pictures itself and asks only when unsure — is this the clean original, is this PNG a watermark file? Skip is always safe. 
Learning and removal run on the phone. The job can be cancelled. 
Compare with the wipe. Mark remaining is an estimate relative to the input. 
Not perfect? Refine the learned mark, then apply it to all pictures. 
Save or share the results once you have checked them. 
The Hardware Test measures this phone and picks the engine. It is first in the ⋮ menu on every screen. 
Settings: each kind of automatic report has its own switch.
