Cutting a subject out of its background used to be an afternoon with a pen tool. It is now a single click, and the reason is a class of model trained to answer one narrow question: for every pixel in this picture, how likely is it to belong to the thing the picture is of?
That framing explains both why it works so well and exactly where it fails.
What the model is actually doing
The technique is called salient object detection. The model has seen a very large number of images paired with a mask marking the main subject, and has learned the visual cues that separate a subject from its surroundings — depth of field, contrast at the boundary, the shapes objects tend to have, the way backgrounds tend to be more uniform.
It produces a probability map rather than a yes-or-no answer, which is then turned into a mask. Where the model is confident, the boundary is crisp. Where it is not, the edge is soft — and that softness is a feature, because a hard-edged cutout pasted onto a new background looks pasted.
Filum uses U²-Net, in its compact u2netp form — an open salient-object-detection model published under the Apache-2.0 licence, converted and served from this site rather than a third-party network.
Where it works and where it struggles
It works well when there is one clear subject with a distinguishable background: a person, a product on a table, a pet, an object shot for a listing. That covers the overwhelming majority of real requests, and the result is usually good enough to use directly.
It struggles in three specific places, and they are hard for every automatic tool, not just this one. Fine hair and fur, where the true boundary is thinner than a pixel and partially transparent. Semi-transparent material — glass, smoke, a veil — where the honest answer is a partial alpha the model was never trained to produce precisely. And busy scenes with no single subject, where the question the model was trained to answer does not have an answer.
In that last case the right behaviour is to say so. A tool that returns a blank image, or a cutout of an arbitrary region, has failed in the way that costs the most time — because you have to notice it yourself. Filum tells you when no confident subject was found rather than handing you something and staying quiet.
Why the result is a PNG
The output has to carry transparency, and PNG is the universal format for that. The subject keeps its exact original pixels; the background becomes fully transparent; and the file can be dropped onto a slide, a store listing, or any new background without a white box appearing behind it.
JPG cannot do this — it has no alpha channel at all, which is why a background removal exported as JPG comes back with the background replaced by white rather than removed. WebP and AVIF can carry transparency, but PNG is what every application accepts.
Running the model on your own device
Background removal is normally a server product. The model is large, it wants a GPU, and the business model of the well-known services is to charge per image or per resolution tier. Your photo is uploaded, processed on their hardware, and returned.
There is now another option. The model and the runtime that executes it can be downloaded into the browser — around 18 MB, once, from this site, cached after that — and run on your own machine. It takes a few seconds rather than being instant, because a laptop is not a datacentre GPU. In exchange, the photo never leaves your computer, nothing is metered, and there is no account.
That trade is worth naming honestly rather than pretending the local version is faster. It is not faster. It is private, and it is free, and for photographs of people, products before launch, and anything you would rather not hand to a company you have not read the terms of, that is usually the trade you want.