Proof

Background removal benchmark

How well FileGizmo's background remover cuts out nine test scenes whose true transparency is known exactly, and how long it takes, measured in the browser on 16 September 2026.

At a glance

Best edges BiRefNet Lite, 1024 px on the graphics card
Mean IoU
0.905
Edge error
0.089
Seconds, warm
7.5
9 of 9 scenes scored
People only MODNet, 512 px on the short side, on the processor
Mean IoU
0.143
Edge error
0.335
Seconds, warm
1.9
9 of 9 scenes scored
Flat colour No model; keys out a smooth background
Mean IoU
0.868
Edge error
0.076
Seconds, warm
0.1
6 of 9 scenes scored; it declines the rest

Higher IoU is better, lower edge error is better. Figures are for the 384 x 288 scenes; the tables below have every scene at both sizes.

Scenes at 384 x 288

SceneBest edgesPeople onlyFlat colour
MeasureIoUEdge errorIoUEdge errorIoUEdge error
Portrait0.9980.0110.6770.071DeclinedDeclined
Product0.9990.0080.0670.3991.0000.003
Animal with whiskers0.9920.0110.1490.3260.9680.016
Fine hair, similar colour0.8060.1790.3650.240DeclinedDeclined
Text logo0.9990.0080.0000.4861.0000.000
Signature on paper0.7200.0590.0030.1601.0000.000
Glass0.9410.2550.0000.3740.2400.438
White product, light backdrop0.9980.0100.0000.4851.0000.001
Busy background0.6910.2560.0220.471DeclinedDeclined

The same scenes at 1536 x 1152

SceneBest edgesPeople onlyFlat colour
MeasureIoUEdge errorIoUEdge errorIoUEdge error
Portrait1.0000.0120.4420.145DeclinedDeclined
Product0.8670.0840.0710.3521.0000.003
Animal with whiskers0.9940.0130.1470.3370.9680.048
Fine hair, similar colour0.8470.1780.2530.260DeclinedDeclined
Text logo1.0000.0160.0000.4961.0000.000
Signature on paper0.9910.0130.0980.3861.0000.000
Glass0.6560.3200.0000.4360.2390.510
White product, light backdrop0.9990.0500.0000.4981.0000.001
Busy background0.6950.2570.0140.473DeclinedDeclined

How to read it

IoU, intersection over union, compares the cutout with the true shape once both are reduced to in or out. 1.000 is a perfect match. It says whether the subject was found, not whether its edge is good.

Edge error is the mean difference in transparency, from 0 to 1, inside a band 6 pixels either side of the true edge. That is where hair, glass and soft outlines succeed or fail, and where a cutout that scores well on IoU can still look wrong.

Declined means Flat colour refused the scene because its background is not one smooth colour, which is the right answer for a photo it cannot separate. The tool then suggests Best edges.

Best edges and People only are different tools. People only is a portrait model and is expected to fail on products, logos and signatures; its numbers are here so that the difference is visible, not hidden.

How it was measured

Each scene is drawn twice from the same vector artwork: once whole, which is what the tool is given, and once as the subject alone on transparency, whose alpha channel is the exact answer. The scenes cover what cutouts are judged on: a portrait, a product on a gradient, an animal with whiskers, fine hair against a similar colour, a text logo, a signature on ruled paper, a glass, a white product on a light backdrop, and a busy pattern. The larger set draws the same scenes at four times the size, bigger than either model predicts at, where the mask has to be brought up to the photo.

Every image went through the page as the site serves it, on an AMD Ryzen 5 5500U laptop with integrated Radeon graphics and 16 GB of memory, running Chrome 152 headless, and the download was scored against its reference. The seconds run from adding the image to the result appearing. The first image of each model includes loading the model and is left out of the warm median. A faster graphics card or processor will be quicker.

The device also decides which Best edges model runs. With WebGPU it is the 1024 pixel model above. Without it, or on a graphics adapter the browser cannot use for it, Best edges runs a 512 pixel model on the processor, which on 16 September 2026 scored a mean IoU of 0.849 and an edge error of 0.126 on the same small scenes, taking about 9 seconds a scene on this laptop.

The measurement is scripted, so every change to the remover is scored the same way before it ships, and this page is updated from a named run rather than from whatever ran last.

What this does not show

These are synthetic scenes. They catch regressions and compare settings precisely, but they are cleaner than photographs, with no noise, blur or compression. Scores on real photographs will be lower, most of all for hair and glass. Public matting sets with hand-made references are the next addition.

Other background removers are not measured here. Testing them would mean uploading images to them, and a comparison built that way is not one we publish. The comparison of free background removers sets out what each vendor states about price, resolution and where your photo goes.