Why does the paper cutout fail on a busy background?

The cutout is a colour flood-fill, not object recognition

Almost every tool in this niche advertises 'AI background removal' or 'object selection'. This one does not, and it is worth knowing exactly what it does instead, because that single fact is the whole reason a busy background fails.

The cutout reads the image pixel by pixel. It samples the four borders — every third pixel along the top, bottom, left and right edges — and averages them into one reference colour. Then it starts a breadth-first flood-fill from every border pixel, spreading inward to any neighbouring pixel whose colour is close enough to that reference. Whatever the fill cannot reach stays; whatever it reaches becomes transparent.

That is a colour operation. There is no model, no segmentation step and no understanding of the picture. It has no idea that the thing you care about is a person, a bottle or a dog. It only knows which pixels look like the edge colour — which is exactly why it separates a wall cleanly and falls apart on a forest.

The tolerance slider, in real numbers

The single control for this is the tolerance slider, and its range is not arbitrary. It maps 0 to 1 onto a colour distance of 26 to 156 in RGB. The default is 0.35, which is a distance of roughly 72.

Low tolerance, near 0, means a pixel must be almost identical to the border average to be cut. High tolerance, near 1, means pixels that are only vaguely similar get cut too. The comparison is done on a squared distance, so the slider is not linear in effect — most of the visible change happens in the upper half of the range, which is why small nudges near the top move the result far more than small nudges near the bottom.

Two mechanical details matter when you are chasing a stubborn edge. First, the reference colour is a single average of the whole border, so a background that is dark at the top and bright at the bottom has no one colour to match, and the fill stops wherever the tone drifts out of range. Second, the resulting mask is feathered across a five-by-five pixel neighbourhood, about a two-pixel soft edge, so the boundary is not aliased — but it is also not a razor cut.

What the percentage tells you, and what to do about it

After each pass the tool counts the pixels it actually cut and prints their share of the whole frame. That number is the diagnostic, and there are only three useful readings.

• Under about 1.5%: the background is not even enough for a flood-fill. Raising the tolerance rarely saves this, because the colour is varying too much for one reference value to describe.
• Between roughly 2% and 92%: the cut worked. The number tells you how aggressive it was, so you can move the slider and watch the figure change instead of guessing from the preview.
• Over about 92%: the fill has crossed onto the subject and is eating it. Lower the tolerance.

The fix for a busy background is not this slider

Here is the part that the 'one-click AI' wording skips. When the background is busy, textured or a gradient, no tolerance setting will separate your subject cleanly, because the operation has nothing consistent to grab. The right move is to stop tuning and change the input.

Cut the subject out in a tool that uses a real segmentation model, or do it by hand, save it as a PNG with a transparent background, and bring that in. This tool animates a transparent PNG exactly as well as a photo; the tear, the paper shadow and the motion do not care what sits behind the cutout.

There is one hard case worth naming. If the image comes from a cross-origin source, the browser will not let the page read its pixels at all, and the cutout cannot run. The status line says so rather than pretending the cut succeeded.

Choosing a photo that cuts cleanly in the first place

Most cutout failures are decided before the image is ever loaded. These conditions let the flood-fill do its job:

• A subject against one fairly even tone — a wall, a studio sweep, a plain table, a clear sky.
• Soft, even light. A hard shadow falling across the background adds a second colour the fill will stop at.
• Clear separation between subject and background, with no large area of the background sharing the subject's colour.
• If you already hold a transparent PNG of the subject, skip the cutout step entirely — the tool reads it as an ordinary image.