Your image is processed on this device and is not uploaded for vectorization. Tracing, clean-up, editing and export all run in your browser; optional AI models download to your browser and run there.
1. Reading the file
Your browser decodes the image with its own built-in decoders — the same code that displays images on any web page. The file's first bytes are checked to confirm its real type, whatever its name says. Very large images are measured before anything heavy happens, and you are offered a resize if the image would strain your device's memory.
2. Measuring the picture
A small copy of the image is analysed to answer one question: what kind of picture is this? The tool counts distinct flat colours, measures how sharp the edges are, checks whether the border is a solid background, looks at how much of the image is ink against paper, and checks for compression noise. Those numbers choose a mode — Logo, Illustration, Line art, Sketch, Signature, Photo, Portrait or Pixel art — and the numbers are shown next to the choice so you can see why.
3. Preparing the pixels
Depending on the mode and your settings, the image is cleaned: JPEG noise smoothed, uneven paper flattened, halos left by background removal trimmed, small sources enlarged so the tracer can find edges between the pixels. A solid background can be keyed out by colour; a complex one can be separated by an optional on-device AI model.
4. Finding colours
Colours are clustered in OKLab, a colour space built so that equal distances look like equal differences. For flat graphics, pixels on edges count for less than pixels inside shapes, because edge pixels are blends of two colours, not colours of their own. The palette is then snapped to the real colours that appear most in the image, so a red logo stays the red it was.
5. Finding shapes
Every pixel is assigned to its nearest palette colour, and touching pixels of the same colour become regions. Regions smaller than the minimum size are not deleted — that would leave holes — but merged into the neighbour they most resemble. In photos, region borders are settled by a neighbourhood vote so they are smooth rather than hairy.
6. Following the edges
Each region's boundary is walked along the cracks between pixels, giving an exact outline. The outline is then straightened into the polygon that best explains it — the staircase of a diagonal becomes one straight segment — and each vertex is classified: a real corner where the outline turns sharply, a curve where it bends gradually.
7. Fitting curves
Between corners, cubic Bézier curves are fitted to the outline within a tolerance set by your Detail and Smoothness settings. Long straight runs become single line segments. The result is a path with as few nodes as the shape allows.
8. Stacking the shapes
In layered mode, larger regions are painted first and allowed to extend slightly beneath the smaller ones laid on top, so no background can ever show through a hairline gap. In cut-out mode, shapes meet edge to edge without overlapping, which cutting machines need.
9. Writing the file
The SVG is written directly from the path data, with coordinates rounded to the precision you choose. Before download it is checked: it must contain only paths, no embedded bitmap, no scripts, and a valid viewBox. PDF, EPS and DXF are written by small exporters in the page; PNG is rendered by your browser from the SVG.
Where the work happens
The heavy steps run in a Web Worker — a background thread in your browser tab — so the page stays responsive and a trace can be cancelled at any time. The optional AI models download from a model file store the first time you ask for them and then run on your device, using WebGPU when available and WebAssembly otherwise. The privacy page lists every network request the site can make.