Watermark Remover / Content-Aware Fill

Box the watermark, logo or object you want gone and let content-aware fill rebuild it from the surrounding picture. Runs entirely in your browser — nothing is uploaded.

All processing happens in your browser. Nothing is uploaded.

How to use

  1. Drop an image in, or click to choose one.
  2. Drag a rectangle over the watermark you want gone.
  3. Pick a quality level and hit Remove.
  4. Hold Hold to compare to check it against the original, then Download.

To clear several watermarks, repeat steps 2–3 — each pass works on the result of the last.

How the fill works

This uses PatchMatch content-aware fill (Barnes et al., SIGGRAPH 2009) — the algorithm Photoshop’s Content-Aware Fill is built on.

For every pixel in the hole, it searches the rest of the image for the small patch whose surroundings look most like this pixel’s surroundings, then copies that patch’s centre colour across. Each pixel picks its own source independently, so gradients, sky-to-ground transitions and regular textures follow the surrounding image naturally instead of averaging into a blur.

Each iteration does three things:

After a few rounds it converges. Only the perimeter ring gets a light blur at the end, to hide any residual colour step at the seam; the interior stays sharp.

What it cannot do

The algorithm does not imagine content. It relocates it.

What it can do is continue texture that already exists nearby into the covered area. Flat backgrounds, sky, grass, wood, walls, smooth gradients — these usually come out clean enough that the repair is invisible.

What it cannot do is produce something that was never in the picture. A watermark across a face will not become that half of the face. One sitting in a line of text will not become those letters. Over a building’s distinctive structure, it will fetch some plausible-looking texture from elsewhere, and it will not survive close inspection.

When that happens, a smaller selection — or several passes over different parts — often beats one big box.

About image size

Images longer than 2048px on their longest edge are scaled down before processing.

Cost scales with area: on a 4000×3000 phone photo, a watermark box that looks small can be hundreds of thousands of pixels and take tens of seconds. At 2048 the watermark is still perfectly legible and the work drops by an order of magnitude.

The download is at the processed size. If you need the original resolution preserved, this is not the right tool — it is built for a quick fix, not for retouching work.

Frequently asked questions

Are my images uploaded?

No. Decoding, filling and exporting all happen in your own browser, with the algorithm running in a Web Worker. No request is made to any server, and the tool keeps working offline once the page has loaded.

Why does it look seamless in some places and terrible in others?

Content-aware fill borrows pixels from the surrounding picture — it cannot invent anything. Regular backgrounds (flat colour, sky, grass, wood grain, gradients) come out well. But if the watermark sits on a face, on text, or on some one-of-a-kind structure, there is nothing similar nearby to borrow from and the result will be mush. That is not a tuning problem; it is the boundary of this class of algorithm.

How tight should the selection be?

Just enough to cover the watermark plus a pixel or two. Many watermarks have a semi-transparent anti-aliased edge, and a too-tight box leaves a faint halo — the tool already expands your selection by 4px to catch that. But bigger is not better: the larger the hole, the less reference is left to fill it from.

What is the difference between Fast, Balanced and Best?

Patch size and iteration count. Larger patches preserve texture continuity better; more iterations converge on better matches. Both cost time. Fast is plenty for a small badge; Best is worth the wait only for a large watermark over complex texture.

Why does it refuse when my selection is large?

The tool declines when the selection covers more than half the image. At that point there is not enough surrounding content to fill from, and spending twenty seconds to produce a blurry mess helps nobody.

Articles about this tool