Local Watermark RemoverLocal Watermark RemoverPick a photo and the subject is cut out on the spot — or erased entirely, with the background painted back in. The image never leaves your device.
Three steps, no sign-up in between.
Two different models, both running on your own machine.
A segmentation model (IS-Net) predicts which pixels belong to the subject and writes that into the alpha channel. The output is a PNG with a genuinely transparent background, not a white rectangle.
Erase mode inverts that same mask and hands it to MI-GAN, the inpainting model behind the watermark tool. The subject is deleted and the background is regenerated from the surrounding pixels.
Both models are downloaded to your machine and run through WebAssembly. No request carrying image data is ever made — open your devtools Network tab and check.
Edges are where cutouts fail, so you can zoom to 32× and drag a compare divider across the result instead of judging it from a thumbnail.
No watermark on the result, no resolution locked behind a paid tier, no daily quota. The alpha channel is preserved as computed.
First use pulls about 95 MB of model and runtime. The CDN caches it for four hours, so repeat use in the same sitting starts immediately. Erase mode adds a further 28 MB the first time you use it.
Background removal is the step people most often reach for with photos they would rather not hand over: product shots before launch, headshots, ID photos, screenshots of internal tools. Every online cutout service starts by asking you to upload exactly those files.
This tool removes that step. The models are downloaded once and the computation happens where the image already is. There is no question of whether to trust the server, because there is no server in the path.
The trade-off is honest and worth stating: you pay for that with a large one-time download instead of a fast round trip to someone else GPU.
It is strongest on a clearly separated subject: a person, a product, an animal, anything with a definite outline against a background that is not the same colour. That covers most of what people want cutouts for.
It struggles exactly where cutout models generally struggle — fine hair, fur, motion blur, glass, smoke, and anything semi-transparent. The alpha ramp across those edges is only a few pixels wide, so you get a hard edge where reality is soft.
Erase mode has a separate limit: it fills the hole from the surrounding pixels, so it is convincing on continuous backgrounds like a wall, sky or grass, and much less so when structure ran behind the subject — a doorway, a bookshelf, a face. The bigger the subject, the more the model is inventing.
The specifics, with numbers rather than reassurance.
No. The image is decoded to a canvas, the models run through WebAssembly on your own CPU, and the result is written back locally. The only network requests are the one-time model downloads.
The first cutout downloads about 95 MB — an 84 MB segmentation model plus 11 MB of ONNX Runtime WebAssembly. After that the CDN caches it for four hours, so a second image in the same sitting starts right away.
Erase mode downloads a further 28 MB for MI-GAN the first time you use it. That one is stored with the browser Cache API and stays available offline.
They share one pass of the segmentation model and use its result in opposite directions. Keep writes the mask into the alpha channel so the background disappears. Erase inverts the mask and inpaints, so the subject disappears and the background is filled in.
Switching between them does not re-run the segmentation — the second mode is computed once and then cached for that image.
Yes, in keep mode. The download is a PNG with a real alpha channel, which is why the preview shows a checkerboard behind it. Erase mode output is fully opaque, since the background has been filled in rather than removed.
The cutout works at whatever resolution you give it. Erase mode runs inpainting at up to 4096 px on the long edge; anything larger is scaled down for the model pass and scaled back afterwards, so very large images lose some detail inside the filled area only.
Yes, with no account and no watermark on the output. There is no server cost to cover, because your own machine does the work.
Back to the watermark remover · All questions about the watermark tool