removebgpictureStart

Published test results

Edge test corpus: how the cutout edges score

Eight hard edges with a known right answer, cut by the same engine your browser runs, and scored against that answer. Scores were generated on 2026-11-01 from the engine as of commit 14ef8064, at 192 × 192 pixels. Lower is better; the failures stay in.

The scores

Alpha error per scene, lower is better
SceneMethodMean errorEdge errorRMSE
Hair strands against a blue wallColour solve0.5%10.7%2.91
Fur on a grey groundColour solve1.75%19.7%7.37
A feathered vignetteColour solve8.54%36.2%20.75
Anti-aliased lettering on whiteColour key0.01%0.5%0.17
Thin wire and a pendantColour solve0.14%9.5%1.33
A soft contact shadowColour solve1.65%11.9%6.03
A glass-like see-through shapeColour solve10.51%48.7%23.66
A hard product edgeColour solve0.23%19.1%2.14

Each scene: input, right answer, our answer

Hair strands against a blue wall: the test picture
Input
Hair strands against a blue wall: the reference alpha
Right answer
Hair strands against a blue wall: the alpha this site produced
Our alpha

Hair strands against a blue wall

Thirty one-pixel strands leaving a head; coverage is fractional along every strand.

Mean 0.5% · edge 10.7% · 1,534 edge pixels

Fur on a grey ground: the test picture
Input
Fur on a grey ground: the reference alpha
Right answer
Fur on a grey ground: the alpha this site produced
Our alpha

Fur on a grey ground

A soft, noisy boundary several pixels deep, like an animal's coat.

Mean 1.75% · edge 19.7% · 2,637 edge pixels

A feathered vignette: the test picture
Input
A feathered vignette: the reference alpha
Right answer
A feathered vignette: the alpha this site produced
Our alpha

A feathered vignette

An edge that fades over 24 pixels: the right answer is a long, even ramp.

Mean 8.54% · edge 36.2% · 8,696 edge pixels

Anti-aliased lettering on white: the test picture
Input
Anti-aliased lettering on white: the reference alpha
Right answer
Anti-aliased lettering on white: the alpha this site produced
Our alpha

Anti-aliased lettering on white

A logo-style mark on pure white, cut with the colour key and its unmixing.

Mean 0.01% · edge 0.5% · 616 edge pixels

Thin wire and a pendant: the test picture
Input
Thin wire and a pendant: the reference alpha
Right answer
Thin wire and a pendant: the alpha this site produced
Our alpha

Thin wire and a pendant

Lines one to two pixels wide crossing open background; easy to lose entirely.

Mean 0.14% · edge 9.5% · 523 edge pixels

A soft contact shadow: the test picture
Input
A soft contact shadow: the reference alpha
Right answer
A soft contact shadow: the alpha this site produced
Our alpha

A soft contact shadow

A dark, blurred shadow under a box. Whether it should stay is a choice; the reference keeps it at its real density.

Mean 1.65% · edge 11.9% · 5,109 edge pixels

A glass-like see-through shape: the test picture
Input
A glass-like see-through shape: the reference alpha
Right answer
A glass-like see-through shape: the alpha this site produced
Our alpha

A glass-like see-through shape

Half-transparent all the way across. Colour-based solves treat it as solid or as gone; it is a known failure.

Mean 10.51% · edge 48.7% · 7,925 edge pixels

A hard product edge: the test picture
Input
A hard product edge: the reference alpha
Right answer
A hard product edge: the alpha this site produced
Our alpha

A hard product edge

A crisp manufactured outline on near-white: the reference is binary apart from one pixel of anti-aliasing.

Mean 0.23% · edge 19.1% · 440 edge pixels

Questions about the corpus

What do the scores mean?

Mean error is the average difference between the engine's alpha and the reference alpha over the whole picture, as a percentage of full opacity: 0% would be a perfect match. Edge error is the same average taken only over the pixels the reference marks as part-way transparent, which is where cutouts actually go wrong. RMSE punishes large mistakes more than small ones.

Why are the scenes drawn rather than photographed?

Because a fair score needs the true answer, and only a drawn scene has one exactly: each picture is made by laying a foreground over a background through a known alpha, so the reference is not a guess or a hand-painted mask.

Why publish the bad results?

Because the glass and feathered scenes are real limits of a colour-based cutout, and hiding them would make the good scores worth less. When the engine changes, the corpus is run again and this page shows the new numbers with the commit they came from.

Were the optional AI models tested here?

No. These scores are for the colour solve and the colour key, which run with nothing downloaded. The optional models are fetched only when a visitor asks for one, so they are not part of this build-time test.

Related