What background removal actually does
It produces a mask that separates subject from background, then composites the subject over a transparent layer. The output is a PNG (or sometimes WebP) with an alpha channel.
How browser-side AI works
An ONNX or TF.js model — usually a U-Net variant trained on millions of foreground/background pairs — runs in a Web Worker. Modern phones can do a 1MP image in 1–2 seconds; a desktop CPU does it in <1.
Trade-offs vs Photoshop
Photoshop's Object Selection is more precise on hard edges (glasses, hair). The browser model is faster, free, and never sees your image. For most product-shot use cases the gap is small.
Avoid the soft-edge halo
Most models leave a 1–2 pixel feather of background around hair. Two fixes: matte against a similar background colour, or set the model's threshold higher and accept slightly tighter cuts.