Make a Transparent PNG

How it works

Three steps, and only one of them is interesting.

A segmentation model — BiRefNet — looks at the image and produces a mask: a grayscale image the same size as yours where bright means subject and dark means background. That mask is then written into the PNG as its alpha channel, and the color pixels are passed through untouched. Your subject is not redrawn, recolored, or regenerated. The only thing that changes is which pixels are declared visible.

That is worth knowing because it tells you what can and cannot go wrong. The subject cannot come back distorted, because nothing generated it. What can go wrong is the boundary: too much cut away, too little, or a soft edge resolved into a hard one.

Reading a mask

There is one habit that catches almost everything. Look at the result twice, once on a dark background and once on a light one.

Masks fail asymmetrically. Leftover background pixels are pale in a photo shot on white, so they vanish against a white page and appear as a gray fringe on a dark one. Over-cutting does the opposite: the missing sliver of shoulder is invisible against a matching backdrop and obvious against a contrasting one. If you check on one background only, you will ship whichever error that background happens to hide.

What the edge pixels actually contain

The fringe question comes up often enough to be worth stating plainly.

In the original photograph, a pixel sitting exactly on the subject's outline is not subject-colored. It is a mixture — part subject, part whatever was behind it — because the sensor averaged both across that pixel's area. When the background is removed, that mixture stays. A dark jacket photographed against a white wall has an outline of mid-gray pixels, and those pixels remain mid-gray when you place the cutout on a black page.

No mask can fix this, because the information needed to fix it was never recorded. The practical response is to accept a one-pixel erosion of the edge, which trades a sliver of the subject for a clean boundary.

Where this is the wrong tool

Glass, smoke, water, veils, and anything in motion blur are all cases where the correct alpha value is fractional and varies per pixel. A segmentation model is answering a two-way question and will give you a firm boundary through a region that genuinely was not firm. The output usually looks confident, which is worse than looking wrong.

Fine hair is the borderline case. You will get a believable silhouette and lose the individual strands. For a profile picture that is often fine; for a composite against a busy scene it usually is not.

Publication gate

This page ships when the workflow has been run against a product photo, a portrait with loose hair, and an image with a low-contrast subject, with the output inspected at full size on both a dark and a light backdrop.