Frequently asked questions

Are my images uploaded to a server?

No. Every step runs inside your browser. The image is decoded to a canvas, the AI model runs through WebAssembly on your own CPU, and the result is written back locally. No request carrying image data is ever sent.

You can verify this yourself: open your browser devtools, switch to the Network tab, and remove a watermark. The only network request is the one-time model download.

Is it free? Does the output have a watermark?

It is free with no account, and the output carries no watermark of its own. There is no per-image limit, no daily quota, and no paid tier that unlocks a higher resolution.

This is possible because there is no server cost to cover — your own machine does the computation.

Which AI model does it use?

MI-GAN, an inpainting model, exported to ONNX and run with ONNX Runtime Web on the WebAssembly backend. The model file is about 28 MB.

Inpainting means the painted region is regenerated from the surrounding pixels rather than blurred or cloned, which is why text over a detailed background can disappear cleanly.

Does it work offline?

Yes, once the model has been downloaded. The model is stored with the browser Cache API on first use, so later visits — including offline ones — reuse the cached copy.

Why is the first removal slower than the rest?

The first run downloads the 28 MB model and compiles the WebAssembly inference engine. That happens once; the progress bar shows the download explicitly so it does not look frozen.

Every removal after that skips straight to inference.

Can it process several images at once?

Yes, and this is the main difference from most similar tools. All imported images are tiled on screen at the same time and each tile is independently paintable.

With "Sync strokes" on (the default), painting on one image applies the same stroke to every image — useful when a batch shares the same watermark position. "Remove on all" then processes the whole batch in sequence.

How large can the images be?

There is no hard limit, but inference is capped at a 4096 px long edge: larger images are scaled down for the model pass and the result is scaled back and composited onto the original. Anything above 8000 px triggers a confirmation because it will be slow.

Pixels outside the painted region are preserved byte for byte, including the alpha channel.

Which browsers are supported?

Any current Chrome, Edge, Firefox or Safari. The requirement is WebAssembly support. Inference runs single-threaded, because multi-threaded WebAssembly needs cross-origin isolation that most static hosting does not provide.

When is it appropriate to remove a watermark?

Use it on your own images — removing a stray logo from a photo you shot, cleaning up a screenshot you made, or taking out a template mark from a design you own.

A watermark on someone else's work marks authorship and licensing. Removing it does not grant you a licence, and using the result may infringe copyright regardless of the tool involved.