Remove image background

The model looks at a 320-pixel thumbnail of your photo

U²-Netp works at a fixed 320 × 320 input. Your photo is squashed down to that size, the network returns a grayscale confidence map at the same resolution, and that map is smoothly scaled back up to full size and used as the alpha channel.

Everything about the result follows from that. Big, obvious separations — a person against a wall, a shoe on a table — are decided at thumbnail scale and come back clean. Detail that does not exist at 320 px cannot be recovered: individual hair strands, fur, chain-link, thin cables, a wire earring. You get a plausible soft edge there, not a strand-by-strand cutout.

It also means the edge is inherently soft. This is a mask, not a pen-tool path — do not expect a pixel-exact border.

Photos it handles, photos that fight back

The model picks the subject for you. There is no brush, no lasso and no way to click "keep this one" — so if a crowded frame makes it choose the wrong object, cropping tighter around what you actually want is your only lever.

• Works well: one clear subject, a background that differs in brightness, even light, the whole subject inside the frame
• Struggles: a white product on a white surface, glass, bottles and anything transparent, heavy motion blur
• Struggles: busy shelves behind a product, two people when you only want one, a subject cropped by the frame edge
• Unpredictable: cast shadows — sometimes kept as part of the subject, sometimes cut away with the background

What the Cutout level slider is really doing

It is a threshold sliding across that grayscale confidence map, not an eraser. The slider sets where the cut falls — 50 puts it in the middle — and a fixed soft band on either side of that point is what gives the edge its feather. Drag right and pixels the model was unsure about get dropped; drag left and they get kept.

Read the failure and pick a direction. A faint halo of the old background clinging to the edges means slide right. Chunks bitten out of dark hair or a black jacket means slide left. Because the cached mask is simply redrawn, the preview keeps up with the slider even on a phone.

The catch is that the threshold is global. You cannot fix one corner without moving the edge everywhere else in the image.

Transparency means PNG, and PNG is heavy

The download is a PNG because that is the format with an alpha channel. A cutout of a detailed photo can easily land at several megabytes — normal, and worth planning for if it is going on a web page.

To slim it down, send the PNG through Compress image and select WebP. WebP keeps the transparency and usually lands far smaller. Do not select JPG: JPG has no alpha channel, and your cutout will come back sitting on a solid rectangle.

If you picked the white or solid-color background instead of transparent, none of that applies. The download is still a PNG — that is the only output here — but a flat-background cutout has no transparency to protect, so running it through Compress image as JPG is usually the smaller, better call.

What this page will not do

HEIC is not on the accepted list. The drop zone takes JPG, PNG, WebP, BMP and GIF, so a photo straight off an iPhone needs a trip through HEIC to JPG first — bring the JPG back here.

There is no batch mode, no manual refinement, no shadow generation and no way to composite the cutout onto another photo. It produces one subject on a transparent or flat background, and stops there.

Your image is never uploaded, and that is the trade-off: the model runs on your own CPU, so a large photo on an old phone takes a few seconds — and a very old browser may not manage it at all.