People today, not products yet
The current model separates a person from the background. A product photo, a logo or a pet is refused with a message rather than returned as a bad cut-out.
Free background remover, in your browser
Drop a photo and get the same image back with the background gone: a transparent PNG, or a flat white or black, that keeps the original pixel dimensions. Everything happens in this tab, nothing is uploaded and there is no account.
Add up to 10 photosFull resolution · transparent PNG · up to 10 at once · nothing uploaded
On this device
Local separation progress
Your cut-out
Model passes over your photo
Drag across the image to compare the cut-out with the original.
Finished cut-outs
Add photos of people to fill this shelf.
The current model separates a person from the background. A product photo, a logo or a pet is refused with a message rather than returned as a bad cut-out.
The saved file keeps the source pixel dimensions. There is no watermark, no preview tier and no resolution held back for a paid plan.
The mask is computed at 256 by 256 and refined against your own pixels. Strands of hair and thin gaps stay soft; compare at full size before you use the result.
No. The photo, the mask and the finished cut-out stay in this browser tab. The page downloads its own model file once and sends property-free audience measurement; it has no image endpoint to upload to.
The mask is computed at 256 by 256 and then refined against your own pixels, so single strands of hair, thin straps and small gaps stay approximate. Compare at full size before you use the result. A flat white or black background hides a soft edge better than transparency does.
Choose before you run the queue. Transparent keeps an alpha channel and is saved as PNG, which is what a design tool, a sticker or a product listing needs. White or black is flattened and saved as the smaller JPEG, which is what a document or an ID-style photo needs. A flat colour has no transparency left to lose, so JPEG costs it nothing.
Not yet. The current model is a person-versus-background segmenter, so a portrait, a team photo or a full-length shot works and a product, a logo or a pet does not. It says so rather than returning a bad cut-out. The models usually recommended for objects were refused here on licence grounds; the licenses page says which and why.
Yes. The output keeps the source pixel dimensions exactly. Nothing is downscaled on the way out, there is no watermark and no resolution is held back for a paid plan.
Up to ten in one local queue, processed one at a time. A photo that fails never stops the rest, and Stop ends the run after the image being worked on.
No. It produces one foreground mask for the whole person. No facial landmark, face geometry or face template is computed, stored or sent, on purpose.
About 3.9 MB compressed the first time: a 447 KB model plus the local inference engine this site already uses. The browser caches both, so a second visit starts immediately.