Pipeline
Five stages, and one of them is admitting failure.
PureMatte is not a model you prompt. It is a fixed pipeline of measurable steps, which is exactly why it can report how well each step went.
- 00
Your browser uploads straight to object storage
The page asks our server for a presigned URL and nothing more. The image itself goes from your browser directly to object storage, and the finished cutout comes back to your browser the same way. The bytes never pass through the website or the API — that is an architectural rule, not an optimisation, and it is what keeps a free full-resolution tool affordable to run.
presigned PUT · exact-length signed
- 01
Ingestion normalises before anything else runs
CMYK JPEGs, non-sRGB ICC profiles, all eight EXIF orientations, pre-existing alpha channels and animated files are each either normalised or cleanly rejected before a single computer-vision operation touches the pixels. Anything over the pixel ceiling is refused rather than allowed to exhaust a worker. Every normalisation applied is reported back with the result.
libvips · header-first · 25 MP ceiling
- 02
A router classifies the image in under 50 ms
Stage 0 looks only at a thumbnail — never the full-resolution array — and measures border-ribbon uniformity in Lab space, edge density, saliency and border-histogram modality. From those it picks a route: uniform backdrop, natural scene, or neither. Because it reads a fixed-size thumbnail, the classification cost does not grow with your image.
thumbnail-only · Lab ΔE · Canny · saliency
- 03
Uniform backdrops get a colour-difference key
For a studio or uniform background the engine derives a soft alpha directly from colour distance, in the tradition of the Vlahos colour-difference matte. Producing alpha this way avoids a min-cut entirely, and therefore avoids the shrinking bias that makes graph-cut methods delete thin structures — so fine edges survive. Spill suppression and foreground decontamination run as separate stages afterwards.
route_a · soft alpha · never binarised
- 04
The result is measured, then gated
Border uniformity, alpha certainty and colour-mass statistics combine into a single confidence score. Below the gate, the job comes back flagged: you still get the image, clearly labelled as a cutout we do not stand behind, and on the API it is never billed. This is the part of the product that matters most — an automatic cutout that cannot admit failure is worse than no cutout.
confidence gate · flagged ⇒ unbilled
Try it on your own image.
Full resolution, no watermark, no account. A uniform or studio backdrop is where the engine is strongest today.