Classical CV · deterministic · self-hostable
Drop an image below. You get the subject back at full resolution with a soft alpha channel, no watermark and no account — plus the engine’s own confidence in the cut, so a bad result announces itself instead of quietly shipping.
Or paste from your clipboard with ⌘V / Ctrl V.
PNG · JPEG · WEBP — up to 25.0 MB — full resolution out
Every cutout carries a confidence score built from border uniformity, alpha certainty and colour-mass statistics. When that score falls below the engine's gate the result is returned flagged — shown to you as an unfinished cutout, never dressed up as a finished one.
The engine is classical computer vision: colour-difference keying, guided filtering and a matting solver. There are no trained weights in the image and no training data behind it, so there is no IP-provenance question to answer and the same input keeps producing the same output.
The usual free background remover caps you at a preview-sized image and sells you the pixels back. Here the free web tool returns the original dimensions with a soft alpha channel. The paid product is the API and the self-hostable image, not your resolution.
What it is bad at
Classical matting has known, mechanical failure modes, and pretending otherwise would make the confidence score worthless. Hair and fur break the colour-line assumption the solver rests on. Semi-transparency is mathematically underdetermined. Low-contrast subject-against-background collapses the segmentation. A cluttered natural scene is far harder than a studio backdrop, and the gap is not small.
So the engine measures the cut it just made and flags the ones it cannot stand behind. On the API those jobs are never billed.
How the routing works