This page loads both models (~270 MB combined, cached after first load). WebGPU accelerates them where available; otherwise both fall back to WebAssembly and the page says so.
…or drop / choose your own
Original
Pop-out composite
backend –depth –matte –
For each pixel: the RMBG matte decides subject vs background; for the background, the depth value picks how much blur to blend in (further = blurrier) plus a slight dim. The subject stays fully sharp. Both models run once each in their own Web Worker; the compositing is canvas maths.