How painting away a watermark works
A watermark is just a cluster of pixels sitting on top of the picture underneath. If you can say which pixels belong to the mark, the rest is a fill-in-the-blank problem: look at everything around the hole and extend it inward. The tool on this page splits that job in two — you do the pointing, the browser does the filling.
Painting the mask is the step that matters most. Cover the mark fully but leave clean pixels alone: stray red over detailed areas forces the fill to guess where it did not need to. For semi-transparent tiled marks, pick the watermark colour once and let the tolerance slider grab the rest, then tidy up with the brush and eraser.
Fast is a classical onion-peel diffusion. Starting from the edge of the mask, it fills each unknown pixel from a weighted average of its already-known neighbours and peels inward until nothing is left, then softens the patch into its surroundings. It needs no model, no network, and no warm-up — it is instant on any device.
AI fill runs LaMa, a neural network trained specifically for inpainting, converted to ONNX and executed inside your browser by ONNX Runtime Web — on the GPU over WebGPU where available, otherwise on the CPU over WebAssembly. It crops a square of context around your mask, runs the network at 512×512, and composites the result back at full resolution. Large masks are tiled automatically with feather-blended seams.
Which engine for which mark
| Mark type | Suggested engine | Why |
|---|---|---|
| Small text on a flat background | Fast | Diffusion is exact, instant, and needs no download |
| Mark over texture, foliage, fabric | AI fill | The network reconstructs texture instead of smearing it |
| Large logo or tiled pattern | AI fill | Tiled inference covers areas diffusion cannot reach |
| No internet / weak device | Fast | Fully offline; runs anywhere a canvas runs |
Privacy: where your image actually goes
Your image is read from the file you choose into a canvas in this page, then into memory, and then processed. The application contains no upload endpoint, no form action, and no code path that transmits pixel data. The requests this page does make are ordinary downloads: the inference library, the model weights, and the site’s privacy-friendly page-view analytics.
- Open your browser’s network tab before you start and watch it during a removal.
- Every entry before an AI run is a GET of library or model files; during the run the list stays still.
- Once cached, the model does not re-download, so the page works with no outbound model request at all.
One honest limit: no fill can recover what was never captured. A fill is an estimate built from surrounding pixels — convincing on photos, but never the true original. And please only remove marks from images you own or have the rights to edit.
Frequently asked questions
Does my image leave my device?
No. It is decoded into a canvas in this page and processed in memory. There is no upload endpoint and no code path that transmits pixel data — the only downloads are the runtime, the model weights, and page-view analytics.
Which engine should I use — Fast or AI?
Fast diffuses surrounding colour inward: instant, offline, best for small marks on smooth backgrounds. AI runs the LaMa network and reconstructs texture and structure, winning on busy backgrounds and larger marks at the cost of a one-time ~200MB download.
Why is the first AI run slower than later ones?
The first run downloads the ONNX Runtime library and the LaMa weights (~200MB) from public CDNs. The weights are cached in IndexedDB afterwards, so later visits skip the download and go straight to compiling the session.
What do the brush, eraser and colour-pick tools do?
The brush paints the removal mask, the eraser takes mask back, and colour-pick samples the watermark colour and auto-selects similar pixels — ideal for semi-transparent tiled marks. Undo and Clear mask redo the selection.
Does the output keep the same dimensions?
Yes. The working canvas is created at exactly the uploaded image’s pixel size and is never resized; exports are pixel-identical in width and height.
Does it work on a phone?
Yes. Fast needs only a canvas. AI needs WebAssembly, which every modern browser ships; where WebGPU is also available the model runs on the GPU and is several times faster.
May I remove the watermark from any image?
Only on images you own or have the rights to edit. Watermarks often signal ownership or licensing terms, and removing them from someone else’s work can infringe those rights.