How the model decides what to keep
A segmentation model predicts, for every pixel, how likely it is to belong to the foreground subject. That produces a mask rather than a hard outline, which is what lets soft edges — hair, fur, motion blur — come out semi-transparent instead of being cut off with scissors.
It is a learned judgement, not a measurement, so it works best on the kinds of images it saw most during training: a clear subject, reasonable separation from the background, ordinary lighting. It has no understanding of what you consider the subject, only a statistical sense of what usually is one.
Where it struggles
Low contrast between subject and background is the main failure mode — a grey coat against a grey wall gives the model very little to work with. Fine detail like flyaway hair, chain-link, lace or foliage is the second, because a single pixel genuinely contains both subject and background.
Transparent and reflective objects are the hardest case of all: glass, water and glossy surfaces show the background through the subject, and no per-pixel mask can resolve that cleanly. Where the result is close but imperfect, the usual approach is to use the mask as a starting point and clean up the edge in an editor.
Why this one runs in your browser
Every other background remover uploads your photograph to a server. This one downloads the model instead — around 55 MB the first time — and runs it on your own device, so the image never leaves the machine.
The trade is that first download and a slower first result, since your laptop is doing work a datacentre GPU would do faster. After the initial download the model is cached, and subsequent images are processed without touching the network at all. For photographs of people, documents or products that are not public yet, that difference is the entire point.
Save as PNG or WebP, never JPG
The output has a transparent background, and JPEG has no alpha channel — saving as JPG would flatten the transparency straight back onto white and undo the work. PNG is the safe choice; WebP is smaller if the destination is a web page.
Bear in mind that a cut-out's edges are semi-transparent by design, so the result looks correct only when composited over something. Viewed in a file manager against a white preview it can appear to have a fringe that is not really there.