Which mode to choose
Product or object uses U²-Netp, a small salient-object model (4.6 MB) that finds the main thing in the picture: a cup, a shoe, a car, an animal or a person. It is quick and works well on clear subjects, but it draws hair and fur with a soft, slightly see-through edge. Person uses MODNet (26 MB), a portrait matting model trained on people, which keeps strands of hair and curly hair much better. Person mode only looks for people: on a product photo it can cut away parts of the object, so switch back to Product or object for anything that is not a person.
What we measured, and the honest limits
We tried both modes on a test set: a studio portrait, a person with a large afro, a cup on a plain background, a car in a street, a dog running on grass and a man sitting on a car. Object mode cut the cup, the car and the dog cleanly, but the afro came out blurred at the edge. Person mode gave clean hair on both portraits, but on the cup photo it left dark patches and removed half the saucer. With a man sitting on a car, both modes kept only the man and dropped most of the car. Expect trouble with glass, smoke, thin cables, a subject the same colour as the background, and busy scenes with several subjects, where the model keeps what stands out most.
Edge sharpness and feather
The model gives every pixel a value between fully kept and fully removed. Edge sharpness pushes the half-transparent pixels towards one or the other, which removes a faint halo around products. Feather softens the edge by a few pixels, which helps a cut-out sit naturally on a new background. Changing these settings does not run the model again, so the preview updates at once.
New background: transparent, colour or picture
Transparent keeps the cut-out as a PNG with an alpha channel, ready for slides, designs and online shops that accept PNG. Choose a colour, such as white for a product listing, or upload a second picture: it is scaled to cover the photo without stretching and centred. With a colour or picture you can also save a JPG, which is smaller and opens everywhere. Photos larger than 16.8 megapixels are reduced to that size, the limit of Safari on iPhone and iPad.
Private by design
The tool uses onnxruntime-web to run the models with WebAssembly in your browser. The only downloads are the engine (about 4 MB, from this site) and the model you pick, from a fixed address on Hugging Face whose fingerprint is checked before use; both stay in your browser cache, so the next visit starts at once. Your photos are read from your disk, processed in memory and saved back to your device; no request carries them. Both models are open source under the Apache 2.0 licence. On our test laptop each photo took about three seconds; phones are slower.