# How well does browser watermark removal work? Test results

Source: https://kwr.ytools.in/blog/watermark-removal-test-results
Last updated: September 30, 2026

By The Kling Watermark Remover team. We measured our in-browser watermark removal on controlled test videos: how close the result gets to the original, what affects quality, and what didn't work.

We tested this tool’s watermark removal on videos where we also had the clean original, so every result could be measured rather than eyeballed. On moving footage, a semi-transparent text watermark went from 18.3 dB to 35.9 dB PSNR in brightness, which in practice means going from obvious to very hard to see. Here is how we tested, what we found, and the limits of these numbers.

## How we tested

- Test clips were generated with FFmpeg: a detailed fractal image with a panning camera, a slow zoom, and a completely static shot. Each is 1280×720, 30 frames per second, with added grain so compression behaves like real footage.
- We rendered a clean version of each clip, then the same clip with watermarks burned in: white text at 55% opacity and a solid-looking square at 50% opacity, fixed in one corner. One clip carried three marks, including one visible only between 2 and 4 seconds.
- Each watermarked clip went through the same pipeline the website uses, in Chrome. We compared the result with the clean original using PSNR (peak signal-to-noise ratio), measured separately for brightness (Y) and colour (U, V). Higher is closer to the original; as a rough guide, differences above about 40 dB are very hard to see.

## Results

| Test clip | Watermarked | After removal |
| --- | --- | --- |
| Panning camera, text and square | 18.3 dB | 35.9 dB |
| Slow zoom, text and square | 22.1 dB | 34.5 dB |
| Three marks, top-left text | 25.3 dB | 39.6 dB |
| Three marks, bottom-right text | 20.2 dB | 41.2 dB |

Values are brightness (Y) PSNR inside the watermark area, measured against the clean original. The largest gains come from see-through text on moving backgrounds, which is the most common kind of watermark on AI-generated video.

## What made the biggest difference

Working in the video’s own colour format. Our first version converted every frame to RGB for processing and back again. That shifted colours across the whole picture, not just the watermark, and left faint coloured fringes around letters. Editing the decoder’s original frames instead, and writing back only the pixels that belong to the watermark, raised colour accuracy in areas the tool never touched from about 34 dB to about 42 dB, so only normal re-compression loss remains.

Finding the watermark’s own pixels. Our first version filled the whole selected box from its edges, which always leaves a visible smudge, because a box is much larger than the letters inside it. Detecting exactly which pixels belong to the mark, and reversing only those, is what made the results above possible.

## Automatic detection

Auto-detect found every whole-clip watermark in our test clips. On the slow-zoom clip it first also proposed two false boxes over scenery that simply held still near the centre of the zoom. Adding a check that a candidate’s edges are neutral in colour, as white, grey and black overlays are, removed both false positives without losing any real marks. The trade-off is that strongly coloured logos aren’t detected automatically and need a box drawn by hand. A mark visible for only 40% of the clip was detected in some runs and not others, so the editor lets you give any box its own time range.

## What didn’t work

We also tried estimating the watermark’s opacity from how much each pixel’s brightness varies over time compared with its neighbours, and forcing each shape to a single opacity. Both scored worse on every test clip (for example 35.9 dB fell to 26.7 dB and 28.3 dB), so neither ships. Smoothing the recovered pixels and learning from more than 80 frames made no meaningful difference.

## Limits of these results

- The clips are synthetic. Real watermarks vary in shape, opacity and animation, and real scenes vary in how much they move.
- PSNR is one objective measure. It doesn’t capture everything the eye notices, such as a faint shape in flat areas.
- Solid parts of a watermark hide the picture completely, so they can only be filled in from their surroundings.
- Static shots give the tool nothing to learn from, so it falls back to filling the area in.

[How it works](https://kwr.ytools.in/how-it-works) explains the method, and you can try it on your own video in the [editor](https://kwr.ytools.in/video-watermark-remover).