Portrait Background Remover

Detailed guidance

What the engine decides about your photo

Behind the button is a small neural network that was trained on photographs of people, and it answers one question about each pixel: how likely is this to be the person rather than the room behind them? Your picture is first scaled down to a working size, then mapped onto a grid of roughly 256 by 256 positions, one soft probability per position. That grid is stretched back over your photo, smoothed while it scales so the boundary does not turn into staircases, and used as a stencil: whatever the model believed was you survives, and the rest is set to fully transparent. The output you download is a PNG, because PNG is the format that carries transparency without a licence argument attached to it.

Why the hairline is the whole difficulty

Everything else about a cut-out is easy. The edge is not. A probability grid is a fine instrument on a face and a coarse one on wisps of hair, so loose strands, the gap between an arm and a hip, and the tips of fingers are the first things to look wrong. The model was measured on its own test set at a little over ninety percent agreement with human tracings, and almost all of that disagreement lives on those edges rather than in the middle of the subject. Dark hair against a dark wall and fair hair against a bright window are the two inputs that cost the most, because there is no contrast for anything to lock onto. Check the result at full zoom, over both a dark and a light surface, before you assume the preview was kind to you.

What this page will not do for you

The honest boundaries are worth stating before you waste a run on them. It segments people, because the weights were only ever shown people, and it keeps one dominant subject, so a group shot leaves the other faces attached to the background. What it cannot do is tell you that it was handed the wrong picture. On our own test frames an aerial photograph of a coastline with nobody in it came back with 37 percent of the frame labelled as the subject, while a face cropped so tightly that it fills the picture came back with no subject at all — a non-portrait input earns a confident wrong answer rather than a clean refusal. That is why the result card prints the share of the frame it kept: compare that number with how much of the picture the person actually occupies, and treat a wide gap as a broken cut rather than a quirk. The training data also assumed a phone held at arm length, and accuracy drops off past roughly four metres from the camera, so a full-body shot taken from across a room is a worse input than a waist-up crop. And this is a boundary finder, not an editor: it cannot recover hair that was out of focus, see through a curtain, or reconstruct what the backdrop hid.

Getting a clean cut on the first pass

Four habits carry almost all of it. Put the subject far enough from the background that a soft shadow separates them from the wall. Choose a backdrop that differs from hair and clothing in brightness, not just in hue. Give the model a person rather than a face — head and shoulders, with the body still inside the frame, since a crop so tight that features fill the picture is the input it reads as nobody at all. And avoid motion blur and heavy grain, both of which smear the edge the model is trying to find. When a result has a halo, the fix is a different crop or a different source photo rather than another setting here, because there is only one setting on this page and it controls what goes behind the subject, not where the edge falls.

Where a portrait cut-out earns its keep

The obvious one is a formal headshot whose studio backdrop is the wrong colour for the destination: cut the person out once, then drop them onto the grey, white, or blue your authority specifies without reshooting. Team pages and staff directories are the second case — a set of photos taken on six different holidays becomes one coherent grid when every subject sits on the same flat colour, and this page has that exact option. Avatars, slide decks, newsletters, and profile pictures for a rebrand are the everyday version. For anything involving a face you did not photograph, the transparency is a tool rather than a toy, and the people in the picture deserve a say in the composite.

Privacy, weight, and what happens after you download

There is no upload step to trust or distrust: the engine and its model download into the tab, run against your file locally, and never post your photo anywhere, which is the entire reason a picture of your face can be used here without a second thought. The cost of that design is a first run that pauses while several megabytes arrive, and no support at all in browsers too old to run modern compiled code. Once you have the PNG, mind two things. Transparency flattens onto white or black the moment anyone saves it as a JPG, so hand off the PNG or set the flat colour here yourself. And an alpha channel is heavier than a plain photo, so if the file has to fit a form limit, run the cut-out through the compressor afterwards rather than hoping the shape alone came out cheap.