AI Background Remover

How removing a background inside your browser works

Removing a background is a labelling job, not a drawing job. The model looks at every pixel and answers one question about it: subject, or not the subject? The answer comes back as an alpha matte, a greyscale picture where white means keep and black means drop. A network small enough for that job runs on your own hardware, so this page needs no server and nobody has to pay per image.

The model is BiRefNet, an open source one, and it runs in a background thread of your browser so the page never freezes while it works. Your browser downloads it the first time and keeps it. After that every image takes the same path: the model returns the matte, and the matte is scaled up and applied to your untouched original on a canvas. The pixels you download are the pixels you uploaded, minus the ones the model marked as background.

One consequence of running locally: nothing is metered here, so there is no reason to hand you a small preview and keep the real file behind a paywall. The full-resolution PNG is the only thing this tool produces.

Where it does well, and where it does not

A clear photo of one subject is what this model was built for. People, a product on a table, pets, plants, a shoe on a floor. The edge comes out clean, and hair and fur fade out softly rather than ending in a hard staircase. Reach for those photos first.

It gets weaker when the frame is crowded. If several things could plausibly be the subject, the model has to choose, and it sometimes chooses wrong, so crop closer around what you actually want. Flat illustrations and screenshots give it nothing to read, and it tends to trim the outer margin and leave the rest opaque. A very dark or low-contrast scene can lose part of the subject along with the background. A large hosted model on a datacentre GPU handles all of that better, and it also charges you and reads your file. Pick whichever side suits the photo in front of you.

Getting a cleaner cutout on the first try

• Crop before you cut out. The tighter the subject sits in the frame, the less there is for the model to mistake for it.
• Check the edge against the checkerboard rather than on white. A pale halo is invisible on a white page and obvious against the transparency pattern.
• Preview the cutout on the colour you are going to post it on. Switching the background here repaints the same matte, so it happens instantly.
• Keep the PNG rather than a screenshot of it. PNG is the format that carries the alpha channel, and anything you re-save as JPG loses the transparency and flattens onto white.