Remove background

Cut out a subject on a plain background, transparent PNG. Free, no upload. 1 GB max · Up to 1 GB · Processed locally, never sent to a server.

Lifting an object away from its background is what a shop listing needs for a product shot, what a logo needs before it sits on a coloured panel, what an article thumbnail needs. This tool does it by colour analysis, and it is worth knowing exactly how, because that is what decides which images come out clean and which ones fail.

No machine learning model is involved. There is no neural network, no subject recognition, no call to a cut-out service. The algorithm reads the four corners of the image, averages the colour over a block of three pixels by three, and keeps those four values as its references for the background. It then starts from the pixels along the edge and moves neighbour to neighbour: any pixel close enough to a reference becomes transparent, and the spread continues from there.

Two consequences follow. Only background that is connected to the edge of the image disappears, which protects a central subject even when it contains the same shades as the backdrop. And the tool does not know what it is looking at: it tells one colour from another colour, not a person from a wall. The result always comes out as PNG, the only common format that carries an alpha channel.

How to use it

  1. Drop one image One file at a time, in any format the browser can read, up to about 24 megapixels. Beyond that the cut-out is refused, with a message inviting you to shrink the image first.
  2. Set the tolerance The slider runs from 5 to 90 and starts at 32. That value defines the colour distance allowed from the corner references: the lower it is, the stricter the selection.
  3. Decide on edge softening Set to “Yes” by default. It lowers the opacity of the one-pixel ring that borders the removed area, which takes the edge off the staircase effect on diagonal outlines.
  4. Click “Run” Three phases are announced: analysing the image, cutting out, applying transparency. The whole computation runs pixel by pixel on your device’s processor.
  5. Check the percentage removed The note gives the share of the image made transparent. If it drops below 2%, the tool reports that it found little background and suggests raising the tolerance.

How the selection is computed

The tolerance shown on screen is multiplied by three, then compared with the sum of the deviations across the red, green and blue channels. At the default value of 32, a pixel is counted as background if its components differ on average by less than 32 levels out of 255 from one of the four references. The test is binary: a pixel either belongs to the background or it does not.

The spread from the edges has one consequence worth knowing. An area the colour of the background but enclosed inside the subject — the gap between a mug and its handle, for instance — stays opaque, since nothing connects it to the edge. It will not go, however high you push the tolerance.

Images that will come out well

The ideal case is a subject with crisp outlines on a plain background that touches all four corners and whose colour is clearly different from the subject’s own: a product photo on a studio backdrop, a logo on flat colour, a scanned stamp, a screenshot. The threshold then separates the two without ambiguity.

Leave a margin of background all around the subject. The references are taken at the corners: if the subject reaches the edge and occupies a corner, its own colour becomes the reference for the background, and the tool erases the subject while leaving the backdrop in place. Cropping the image beforehand settles that.

Images it will fail on

Hair and fur are the most common case. Along a strand, every pixel mixes hair with backdrop; a binary test has to pick a side, and it produces either a jagged silhouette or holes in the hair. What that needs is a matte with continuous values, which this algorithm does not compute.

Transparent and reflective objects raise the same problem from another angle. A glass, a bottle, a veil, a pair of spectacles all let the background show through them: those pixels resemble it because they genuinely contain it, and they are removed along with it.

Then there is plain lack of contrast. A white shirt against a white wall, a grey car on asphalt, a gradient backdrop: raising the tolerance far enough to take in the whole background ends up taking in the subject too, and no value separates the two properly. JPG artefacts make it worse by scattering intermediate pixels around the subject, which survive as a fringe.

Setting the tolerance without guessing blind

Run it once at 32 and read the percentage removed. If patches of background are left, raise the value in steps of ten; if the subject starts developing holes, come back down. An image that gives nothing usable above 60 is not a good candidate for this method.

Turn softening off when you need a hard matte, for a die-cut shape or for printing: it caps the alpha of the pixels next to the removed background at 130 out of 255, which smooths the staircase but leaves the subject’s border slightly translucent.

Frequently asked questions

Is this an AI cut-out?

No. No model is loaded and no request is sent: the tool compares colours against those of the four corners and spreads the selection from the edges. Hence its speed, and hence its limits.

My background is a gradient and only part of it disappears.

The threshold keeps only shades close to the corners. A gradient drifts away from them steadily until it falls outside the tolerance. Raising that widens the selection, but it reaches the subject before it has covered the whole background.

An area of background stayed opaque in the middle of my subject.

That is the expected behaviour: only pixels connected to the edge of the image are removed. An enclosed area, such as the inside of a handle or of a letter, is kept on purpose.

Can I get a transparent JPG?

No, and no tool can: the JPG format has no alpha channel. The output is therefore always a PNG.

The tab freezes for a few seconds on a large image.

The cut-out walks every pixel and then spreads the selection, which comes to several million operations on a twenty-megapixel photo. That time is local computation, not a transfer.

I get a message saying the image is too large to cut out.

The limit is about 24 megapixels. Reduce the resolution with the Resize image tool, then run the cut-out on the result.

Similar tools

Discover other tools

Categories