Keep the subject, lose the room

Every background remover answers one question per pixel: does this pixel belong to the subject or to the backdrop? Commercial services answer it with a segmentation network trained on hundreds of thousands of cut-outs. A browser tool answers it with colour, position and connectivity. The second answer is much simpler, much faster, and surprisingly good on the photographs people actually need cut out.

The method is called region growing, or flood fill with a tolerance. It starts from the border of the image, where the background is by definition present, and spreads inward. Each candidate pixel is compared with the pixel it came from, not with an average of the whole background. A local comparison is what makes a grey-to-white gradient disappear in a single pass: each step is a tiny colour change, well inside the tolerance, even though the two ends of the wall are far apart.

Why the walk stops where it does

The spread halts at a real discontinuity — the dark seam where a jacket meets the wall, the shadow line under a plate, the sudden red of a product label. That halt is the outline. The sensitivity slider sets how big a jump counts as a discontinuity. Too low and the walk crosses a soft shadow into the subject; too high and it stops early, leaving a rim of background.

Because the algorithm is connectivity-based, isolated patches of background that are not reachable from the border survive. A gap between an arm and a body, filled with wall, is enclosed by the subject and never touched. Raising the sensitivity frequently frees those patches, which is why the setting is best adjusted while watching the preview rather than guessed in advance.

The alpha channel is the real product

A cut-out is not a picture with the background painted white. It is a picture where every pixel carries an opacity value. The tool builds that value as a mask, blurs it slightly so the boundary fades over two to six pixels, and then composites the original photograph through it. That blur is the difference between a result that looks cut out with scissors and one that sits on a new colour naturally.

There is a second, subtler benefit: transparency lets the destination choose. The same PNG can sit on a white marketplace tile, a dark newsletter header or a coloured presentation slide without being re-exported three times.

Where this approach fails, and what to do instead

• Portraits in front of foliage or a crowd: the walk finds paths between leaves and limbs, and the result needs manual work in an editor.
• Glass, smoke and hair: the subject is partly transparent, so a binary mask destroys it.
• Subjects touching the frame edge: there is no background path around them, and the border pixels are treated as backdrop.

For those pictures the practical route is a reshoot against a plainer background. Ten minutes with a roll of paper and a window beats an hour of repairing a mask, and the result is repeatable for the whole catalogue.

A small quality checklist

Start with a sharp original, shoot slightly away from the wall so the shadow falls behind the subject, and keep the whole subject inside the frame. Run the removal at fine quality, set the edge softness to three or four, and check the outline at 100% on a dark background as well as a light one — halos hide on white and appear on black. If a rim survives, drop the sensitivity by four or five points and run it again; a second pass on the same file costs nothing.

A note for product catalogues

Marketplaces have their own rules, and most of them ask for a square image, a white or transparent background and the product filling roughly eighty-five percent of the frame. A cut-out produced here satisfies the background rule immediately; the framing usually needs a second step, because the subject in the original photo sits wherever it sat. Crop to the square after the removal, not before, so the mask still has background pixels to work with on all four sides.

Keep one master file per product at the highest resolution you have, transparent, and generate the marketplace sizes from it. When a second channel asks for 1500 pixels instead of 2000, you resize the master instead of photographing the item again — and every channel shows the same edge, which is what makes a catalogue look like a catalogue.