AI Background Remover

How background removal actually works

This isn't edge-detection or a colour-key like an old green screen — it's semantic segmentation. A neural network (here, the IS-Net dichotomous segmentation model) looks at the whole image and predicts, for every pixel, the probability that it belongs to the foreground subject rather than the background. The output is an alpha matte: a greyscale mask where white is 'keep', black is 'drop', and the grey values along the edges encode partial transparency. That matte is what lets the model handle a wispy hair edge or a soft shadow as a gradient rather than a hard, scissored line. The model is described in the research on highly accurate dichotomous image segmentation (Qin et al., 2022).

Because it reasons about the image semantically, it generalises well across subjects it was trained to recognise — people, products, animals, vehicles, isolated objects — without you drawing a single selection. The trade-off is that it's a learned model: it has no understanding of your specific intent, only its training, so it occasionally decides a held object isn't part of the person, or that a strong background reflection is foreground.

Where segmentation reliably struggles

• Fine hair and fur against a busy background — usually good, but flyaway strands over a cluttered scene can drop out, leaving a slightly chewed edge.
• Transparency and glass — a wine glass, spectacles, a plastic bottle: the model can't decide whether to keep what's visible through the object, and tends to make it opaque or punch a hole.
• Low subject/background contrast — a grey jacket against a grey wall has no clear boundary to find, so the matte gets mushy exactly where the colours converge.
• Motion blur and soft focus — blurred edges have no crisp transition for the model to lock onto, so the cutout inherits the fuzz.
• Very fine structures — chain-link, net, lace, thin wires: detail finer than the model's effective resolution gets simplified or lost.

Local versus server tools — and trusting the output

The headline difference from remove.bg and similar services is that the model runs in your browser via transformers.js and WebAssembly. The weights download once (around 25 MB) on your first cutout and are cached, after which it works offline and your images are never uploaded — which matters when the photo is an unreleased product, a person who hasn't consented to a cloud upload, or anything under NDA. The cost is the one-time download and that processing speed is bound by your own CPU or GPU rather than a server farm, so a very large image takes a few seconds locally instead of being near-instant.

Treat the result as a strong first pass, not a finished cut. The matte is AI-generated and can carry artefacts — a faint colour halo inherited from the old background, a stray pixel on an edge, a small hole inside a complex subject. For casual use it's ready as-is; for commercial product shots, zoom to 100% and check the edges. After the cutout, you'll usually want to resize it to a target canvas or compress the PNG, and if the source is small or soft, run it through the AI upscaler first so the model has more edge detail to segment.