Cover the watermark and the area is rebuilt from the pixels around it using OpenCV inpainting. Most AI watermarks sit along the bottom, so there are one-click boxes for the bottom corners and a full-width bottom bar, each running flush to the image edge. Everything runs inside this browser tab: your image is never uploaded, never stored, and never seen by a server.
Most AI watermarks sit along the bottom. If a corner box misses part of the mark, use Bottom bar, or drag your own box — every preset runs flush to the image edge so nothing is left behind.
Telea rebuilds sharper, Navier-Stokes smoother. Blending the two hides the seam better than either alone.
Inpainting reconstructs the covered area from the pixels around it. It works best on flat or gently textured backgrounds; over fine detail such as text or faces, expect a soft patch.
Inpainting quality is decided almost entirely by the mask. A little care here is worth more than any slider.
No. The file is read by the browser, decoded to a canvas, and repaired by OpenCV compiled to WebAssembly, all on your device. Nothing is sent to a server, so nothing can be stored or seen by anyone else. You can confirm it by opening your network tab: after the library downloads, there are no further requests.
It is inpainting, not AI generation. The masked pixels are thrown away and rebuilt by propagating colour and texture inward from the surrounding edges. That is why it is flawless over sky, walls, and gradients, and imperfect over detailed texture: there is no information under the mask to recover, only a plausible continuation of what surrounds it.
Two different inpainting algorithms. Telea marches inward from the boundary and keeps edges sharper; Navier-Stokes treats the image like a fluid and produces smoother, more even fills. Each leaves its own kind of artefact, so the tool runs both and blends them, which hides the seam better than either does alone. The slider lets you favour one.
Because the pixels underneath are genuinely gone. Inpainting invents a continuation of the surroundings, so a mark sitting over flat background disappears cleanly while one over text or a face leaves a soft patch. Try a mask that hugs the mark more tightly, and a lower repair radius.
The mask almost certainly missed part of the watermark. Inpainting takes its colour from the edge of the masked area, so if even a few pixels of a bright mark sit just outside your box, that brightness becomes the source it copies inward and it smears across the whole patch. The result looks far worse than leaving the mark alone. The fix is to make the mask bigger, not smaller, and make sure it runs right off the edge of the image — watermarks sit flush in the corner, so a box that stops a few pixels short is the usual cause.
Only from images you have the right to edit — your own photos, your own generated output, or content you are licensed to modify. A watermark often marks ownership or an AI provenance signal, and stripping one from someone else's work is a copyright problem regardless of the tool used. Note too that invisible provenance marks such as SynthID are unaffected by this.
None imposed by the tool, since it is your own machine doing the work. Very large images simply take longer, and the repair time depends far more on the size of the masked area than on the image itself.