The tool uses the Apache-2.0 MODNet portrait-matting model in a browser worker. It prefers WebGPU with 16-bit model weights and falls back to WebAssembly on the CPU when WebGPU is missing or fails. The first run downloads roughly 13 MB of model weights with WebGPU or 26 MB for the CPU path, plus runtime files. Browser caching can make later runs faster.
Removal works best when one person or a clear foreground subject is visually distinct from the background. Fine hair can produce a soft alpha edge, while glass, motion blur, deep shadows, groups, products, and similar foreground/background colours are harder. Preview the cutout on light and dark surfaces before downloading the transparent PNG.