Remove any background, right in your browser.

Remove a background with AI — and know when a simpler eraser wins

Cutting a subject out of its background used to mean tracing edges by hand with a pen tool. This does the tracing for you: a segmentation model looks at the whole image, decides which pixels belong to the subject, and hands back everything else as transparent. It is fast, it is free, and — because the model runs inside your browser — the photo never travels anywhere. Here is what the model is good at, where it slips, and how to get a clean transparent PNG out the other side.

How the cutout finds your subject

The tool runs a salient-object segmentation model called BiRefNet. Instead of matching colours, it predicts a matte — a soft, per-pixel map of how much each pixel belongs to the foreground — so a single strand of hair can be half-kept rather than crudely clipped. The image is analysed at 512×512, the matte is scaled back up to your full resolution, and the subject is painted onto a transparent canvas. Nothing here needs a server: the model weights download once (about 90 MB, then cached), and every cut-out after that happens on your own machine.

What it handles beautifully

Because it reasons about the subject rather than the colour, it copes with backgrounds a colour eraser would choke on — a person in front of a cluttered street, a dog on grass, a bottle on a wooden desk. It is at its best with:

• Hair and fur edges — the soft matte keeps flyaway strands and fuzzy outlines instead of leaving a hard, cut-with-scissors line.
• Product shots — a single item on a busy surface lifts cleanly, ready to drop onto pure white for a marketplace or catalogue.
• One clear subject — a portrait, a pet, an object with obvious edges: the model knows exactly what you meant to keep.

Where it struggles — and how to help it

No segmentation model is perfect. Three cases are genuinely hard: glass and other transparent objects, where the model cannot tell what should show through; subjects the same tone as their background, like a grey cat on a grey couch; and scenes so busy the subject itself is ambiguous. When the model is unsure it can overshoot and erase almost everything — so this tool watches for exactly that, and if the first pass keeps almost no pixels it quietly retries the same image on a larger, more careful model, showing "Refining…" while it works. When your background is instead a flat colour or a smooth gradient — game art, a logo, a product already on white — a segmentation model is overkill, and the deterministic colour eraser clears the background by colour without ever over-removing fine foreground detail.

Exporting a clean transparent PNG

By default you get a small transparent WebP, which keeps the alpha channel at a fraction of PNG's weight — handy when the cut-out is headed back onto the web. Tap PNG when you need the widest compatibility or a lossless master; both formats stay fully transparent. If you would rather not keep the transparency, flatten the subject onto white, studio grey or black before you download. Drop in a whole folder and every result zips together in one pass. The transparent PNG guide covers where each format pays off.

Everything runs on your device

There is no upload step here, and there is no account. The only thing that ever leaves your network is that one-time model download; your images are read, cut out and saved entirely inside the browser tab. Close it and nothing lingers on a server, because nothing was ever sent to one.

For a step-by-step walkthrough with example images, read how to remove a background, and for the export side — formats, edges, and when transparency matters — see make a transparent PNG.