How Well Does AI Watermark Removal Actually Work? We Measured It
We ran six hard cases through our own remover and compared every result with the original image the watermark was hiding
Every watermark remover, ours included, shows you its best before-and-after. That tells you very little. The useful question is how close the result gets to what was actually under the watermark, and where it falls apart.
Measuring that is normally impossible, because nobody has the original image. So we made our own. Every test image below was drawn by us from scratch, so we know every pixel the watermark covered. We stamped an invented watermark on each one, sent it through the same live tool you use, and compared the output with the clean original.
The short version: small marks on plain backgrounds are close to solved. Tiled marks across detail are not. Anything covering text produces something that looks plausible but is made up. And the numbers turned out to be kinder to the tool than our own eyes were.
How We Tested
- Six scenes, six hard cases. Each scene puts a white text watermark somewhere AI removal is known to find difficult: across a hard edge, over a subject, over a large smooth area, tiled across detail, and over text. A small corner mark on plain sky is the easy baseline.
- The live tool, unedited. Every “after” image is exactly what our production server returned. Nothing was retouched or cherry-picked; each case was run once.
- A perfect mask. Because we drew the watermark, we told the tool exactly which pixels it covered. That is the best case. A mask you paint by hand will be less precise.
For each case we measured two things, only inside the area the watermark touched:
- Damage undone: how much of the watermark’s effect on the image was reversed. 100% would mean the output matches the original exactly.
- Still visibly off: the share of watermark pixels whose color still differs from the original by more than about 9% on any channel. That is roughly where a difference becomes noticeable when you look closely.
The Results
| Case | Damage undone | Still visibly off | What we saw |
|---|---|---|---|
| Small corner mark on plain sky | 96.8% | 0.0% | Gone, no trace |
| Text across a horizon line | 95.9% | 5.9% | Horizon turned wavy, with a light streak |
| Text over a person | 98.5% | 0.8% | Clean, but the figure is a flat silhouette |
| Large mark on a sunset sky | 97.9% | 0.6% | Clean |
| Tiled marks over a forest | 86.4% | 9.4% | Smeared, broken branches; clear failure |
| Mark over text | 98.1% | 1.6% | Hidden lines invented, lengths changed |
Four of the six look like clean wins in the table. Two of those four are less impressive than they look, and the next sections explain why.
Where It Works: Small Marks on Plain Backgrounds
A small mark over sky, sand, a wall or any other plain area is the case AI inpainting was built for. The tool only has to continue a smooth color from the surrounding pixels, and it does. The same holds for the large “SAMPLE” over a sunset sky: big, but over soft gradients and blurry clouds that are easy to continue.
Where It Wobbles: Hard Edges
Put the same kind of mark across a sharp boundary, such as a horizon, a roofline or the edge of a product, and the tool has to guess where the line runs under the text. It got close, but zoomed in, the horizon is no longer straight. It bends slightly and picks up a pale streak where the letters were.
That is the first lesson about the numbers. Six percent sounds small, but here those pixels all sit on the edge your eye follows. If you are removing a mark from your own photo and it crosses a straight line, check that line at full size.
Where It Fails: Tiled Marks Over Detail
A watermark repeated across the whole frame, over branches and trunks, is the hardest case we tried. There is no clean background to borrow from, only more detail, which is itself partly covered. At a glance the result passes. Zoomed in, it does not:
The tool invented branch shapes that are not in the original and left pale smears where the letters crossed dark needles. This is why stock agencies and photographers proofing for clients tile their marks: it is the style that costs a remover the most quality. Our test of six watermark styles looks at this from the photographer’s side.
Where the Numbers Lie: Text and Faces
The mark over a screenshot scored 98.1%, one of the best results in the table. It is also the clearest example of why you should not trust a single score.
Our screenshot used simple placeholder lines instead of real words, and the tool filled the gap with more placeholder lines. They look right at a glance, but compare them with the original: one line got longer, and two separate lines became one blurred bar. The tool cannot know what was under the watermark, so it draws something that fits. With real text it would produce letter-like shapes that spell nothing. With a real face it would produce a plausible face, not the person’s.
The same applies to the “text over a person” case. It scored 98.5%, but our person is a solid dark silhouette, which is easy to continue. That result says nothing about a real face, where the tool has to invent skin, eyes and expression.
So treat anything that removes a watermark from text or faces as reconstruction, not recovery. If the words or the face under the mark matter, the only real fix is the original image without the mark.
Why These Numbers Flatter the Tool
We would rather overstate the limits than the results, so here is everything that makes this test easier than real life:
- Our scenes are simpler than photos. They are drawn, not photographed: smooth gradients, clean shapes, and no camera noise, compression artifacts or fine texture. Real photos are harder, especially the tiled-over-detail case.
- Our masks were perfect. We knew exactly which pixels each watermark touched. A hand-painted mask that misses edges leaves outlines behind; one that is too generous makes the tool rebuild more than it needs to.
- One run per case, 1200×800 pixels. This is a small, honest sample, not a benchmark.
- Averages hide the defects you notice. As the horizon and text cases show, a small share of wrong pixels in the wrong place is what your eye catches.
Even on these easy scenes, the two failure modes are clear: detail covered by repeated marks, and anything where the content under the mark is information (text, faces) rather than texture. They match what we already list under what this tool does badly.
What This Means If You’re Removing a Watermark
- Plain backgrounds: expect a clean result. Paint the mark slightly larger than the letters and you are done.
- Edges and lines: check them at full size afterwards, and use Fine Detail mode with a small brush near the edge.
- Repeated marks over detail: expect smearing. Removing the marks a few at a time and touching up often gives a better result than one large mask, but some damage is likely.
- Text or faces under the mark: the tool will invent what was there. Fine for decoration; not fine if the content matters.
And only remove watermarks from images you own or have the right to edit; our guide to the legal side covers when that applies.
Frequently Asked Questions
Can AI recover what was hidden under a watermark?
No. It reconstructs a plausible guess from the surrounding pixels. On plain backgrounds the guess is usually indistinguishable from the original. Over text, faces or other detail the guess is invented: it can look right and still be wrong, as our text test shows.
What kind of watermark is hardest for AI to remove?
In our tests, a watermark tiled across the whole image over detailed areas, such as the forest scene, was the clear worst case: 86.4% of the damage undone and visible smearing. Small marks on plain backgrounds were the easiest.
Are these results from real photos?
No. We drew the test images ourselves so we would know exactly what was under each watermark, which is the only way to measure accuracy. Drawn scenes are simpler than photos, so expect real results to be somewhat worse, especially on detailed images.
Why does my result look blurry or smeared?
Usually because the mask covers detail the AI has to rebuild from guesswork, or because it is much larger than the watermark. Tighten the mask to the mark itself and use Fine Detail mode near edges and faces.