Drop in a photo and get a clean transparent PNG. A small segmentation network called U²-Net runs directly in this browser tab through WebAssembly, so your image is never uploaded. It keeps just the main subject by default, and everything else — background colour, edge tuning, and optional text behind the subject — is tucked away until you want it.
Tighten to cut away background haze, loosen to keep wispy edges like hair. These re-composite instantly — the model only looks at the image once.
Leave the field empty and nothing is drawn. Text behind the subject works best on a solid or blurred background.
The model is downloaded once on first use (about 4.5 MB) and cached by your browser after that. It handles people, products, animals and objects; very fine hair and glass are where it struggles most.
No. The only thing downloaded is the model file itself, once. After that the image is decoded, segmented, and composited entirely inside this browser tab, with no request carrying your picture anywhere. Nothing is stored and nothing is seen by a server.
U²-Net, in its small u2netp form — the same architecture behind a lot of open-source background removal. It is a salient-object segmentation network: it predicts which pixels belong to the thing the picture is actually about, then that prediction becomes the alpha channel.
The model file has to download once, roughly 4.5 MB, and WebAssembly has to compile it. Your browser caches both, so every run after the first starts immediately. Changing the background or edge settings never re-runs the model at all, it just re-composites the mask it already has.
The model outputs a confidence value per pixel rather than a hard yes or no. Tightness moves the line where confidence becomes opaque. Tighter cuts away the low-confidence haze that can ring a subject; looser keeps soft edges like hair and fur, at the cost of a little background coming with them.
Yes, and it's on by default. "Keep only the main subject" measures the connected regions of the cutout and keeps the single largest one, so a person standing near a chair or a plant comes out on their own. It runs on the mask rather than through a second model, so it costs nothing and is instant. Turn it off when you genuinely want every object the model found, such as two people side by side who aren't touching.
Yes. Open Add text, type a word, and leave "Place it behind the subject" ticked. Because the cutout is drawn after the text, the subject overlaps the word and you get the poster look. Leave the field empty and no text is drawn at all. It reads best on a solid or blurred background, since a transparent one has nothing for the text to sit against.
A clear subject against a background that differs from it: people, products, animals, furniture, food. It struggles where the boundary itself is ambiguous — flyaway hair, glass, smoke, chain-link, motion blur — and with busy images where nothing is obviously the subject.
Yes, subject to your rights in the original image. The tool adds no restriction of its own; U²-Net is Apache-2.0 licensed.