Convert raster images into real vector geometry
Every border between two colours is fitted once, so no edge doubles.
Convert and preview without an account. Downloading is free too — it just needs a Google sign-in. No card, no subscription.
01 · 5 colours · 193 nodes
How it works
- Upload one raster image — PNG, JPEG, WEBP, BMP, GIF or AVIF, up to 25 MB and 40 megapixels, one image per conversion.
- Set the colour count between 2 and 64. It decides how many flat regions the segmentation produces, and every later stage inherits that map.
- Choose a draw style and shape stacking: filled closed shapes with cut-outs for cutting and printing, or stroked edges for the border network on its own.
- Select the output formats — SVG, PDF, EPS, DXF, PNG — and a size in millimetres if the geometry has to land at a physical scale.
- Check the 800 px preview, then sign in with Google (free) to download. Pick more than one format and a ZIP is produced as well.
"Raster to vector" is a resampling problem in a drawing problem's clothes: the input is a lattice of colour samples, the output has to be closed regions bounded by curves. This converter combines geometric tracing with machine learning (AI) — same input, same geometry — working in four explicit stages. Colour segmentation comes first, quantising in LAB space and regularising the label field, so the stage emits a map of flat regions rather than a per-pixel palette.
An anti-aliasing model follows: between two flat colours a rendered edge is a two- or three-pixel ramp, not a step, and the boundary belongs where that ramp is half of each side. Each border is sampled along its own normal and slid onto that crossing, landing at sub-pixel position.
The third stage decides whether the result behaves inside a CAD, GIS, signage or prepress tool. A planar subdivision is built on the lattice of cracks between pixels, so each border between two regions is traced once and both regions are assembled from that chain, one of them reversed. Contouring each colour on its own — what per-region contour routines do — fits every shared border twice, and the two answers disagree by up to a pixel: that is the double edge a cutter runs twice and the hairline that shows on plate. Here there is no second copy to drift, and a region's polygon area equals its pixel count by construction.
Curve fitting runs last, on chains already in the right place: simplified, corners kept as corners, the rest fitted with least-squares cubic Béziers and falling back to a line or a circular arc where one covers the run within tolerance. Every subpath closes, and the DXF and PDF writers promote every curve to a cubic, so the SVG curve and the curve AutoCAD opens are the same.
What actually happens
- Raster 6×6 · 8 px · PNG 77 B
- 6 nodes · 1 path
- SVG 91 B · vector, not pixels
What a raster trace is designed to hold downstream, and what it cannot
A two-pixel ramp is a raster border, not a third colour
Colour quantisation has no notion of coverage. Where two flat colours meet, the ramp of blended pixels is a third colour to the clustering, and left alone it comes back as its own thin region with an inner and an outer edge. The engine models that ramp and dissolves it into the border between its two neighbours, which is why a two-colour mark yields one contour rather than three.
An arc stays an arc in SVG and becomes a cubic elsewhere
The fitter covers each run with the simplest primitive that stays within tolerance: a line, a circular arc, or a cubic. The SVG keeps all three in their own path commands, which is what keeps it small. PDF, EPS and DXF have no arc in the vocabulary the writers use, so an arc is promoted to cubics there, up to four of them, with the same geometry.
A scanned line is a region with two sides, not a centreline
A scanned ink line is a thin dark region: filled shapes give you its outline, a closed path running up one side and back down the other, and stroked edges give the same two sides as the border network. Neither is a centreline. A CAD user who needs a single stroke to offset or dimension from will draw it over the outline; there is no skeleton pass here.
Choosing the colour count
This page assumes you own the vocabulary already: a draughtsman who needs a scanned detail as polylines and splines to snap to, a GIS technician lifting region boundaries off a raster plan sheet, a sign shop turning artwork into cut paths, a prepress operator separating flat artwork by colour.
What those jobs share is geometry that has to survive a second program: an outline only nearly closed will not offset, will not fill and will not cut, and two contours that nearly coincide give a router a double pass and a press a visible ridge. Both come from borders fitted twice.
The output stays honest about what it is. The SVG is paths only, with no embedded bitmap. The PDF is PDF-1.4 with genuine cubic curve operators. The DXF holds outlines, one layer per colour, no filled areas. The PNG output is the vectors re-rendered on transparency — a cleaned raster, not vector data.
Settings we suggest
Colour count is the real control, since every later stage inherits the region map it produces. Two colours give one silhouette with a single clean contour, the stencil and cut-file case; four to eight covers most signage and flat logo artwork; scanned line work sits around six to twelve; tonal artwork needs twenty-four or more.
Draw style: filled shapes gives each region as a filled closed path, which is what a press or a cutter expects; stroked shapes draws those same closed paths as outlines; stroked edges draws only the shared border network, each border exactly once, for a single-line drawing rather than areas. Shape stacking: cut-outs keeps holes as reverse-wound subpaths so the nonzero fill rule punches them out, which is what you want when a hole is a real hole in a cut; stacked drops the holes and lets paint order cover them, easier when shapes get pulled apart in an editor.
Group by colour puts each colour's regions into one SVG group — spot separations, or a colour grabbed as a single cut layer; the DXF gets a layer per colour regardless. The millimetre size field applies to every vector output: the SVG width and height, the DXF coordinates, and the PDF and EPS page box, which is set to that physical size rather than to the pixel count. Above 1.5 megapixels the image is downscaled before tracing, so the output's pixel dimensions are the downscaled ones.
Questions
Will neighbouring shapes have duplicated borders?
No — each border is traced once and both regions reuse that chain, one reversed, so no second copy exists to drift: no overlapping slivers, no white hairlines. Filled SVG does add a 0.7 px seam stroke in each pair's mean colour.
Are all paths closed, and how are holes handled?
Every subpath closes — a region's rings leave the subdivision as complete cycles. Holes follow shape stacking: cut-outs keeps them as reverse-wound subpaths for the nonzero rule to punch out, stacked drops them and lets whatever sits inside paint over the top.
What exactly does the DXF contain?
DXF R2010 (AC1024), so it opens in AutoCAD 2010 and later, Fusion 360, LightBurn, Inkscape and FreeCAD. LWPOLYLINE for straight runs, degree-3 SPLINE for curves from the fitted Bézier control points with clamped knots — control points, not fit points, so the curve matches the SVG. One layer per colour, named FILL_RRGGBB. Coordinates are at millimetre scale, where 96 pixels correspond to 25.4 mm unless you ask for a physical size, and the header declares that unit ($INSUNITS = 4), so an importer that reads it places the part at size without being told which unit to assume.
Does the tracer use a learned model?
Yes, alongside the geometry: colour quantisation and label regularisation, an explicit anti-aliasing model, a planar arrangement and least-squares curve fitting do the tracing, and a machine-learning (AI) model running on our own servers is part of the pipeline. The failure modes are still ones you can reason about: too few colours, or an input with no flat regions.
How long does a conversion take, and what are the limits?
Simple logos usually finish quickly; detailed drawings and photos can take several minutes. Each conversion has a 300-second limit; if it is not going to finish in time, the converter tells you and suggests a smaller image. Inputs go up to 25 MB and 40 megapixels. Free and watermark-free: study the preview first, then a free Google sign-in unlocks downloads — 10 conversions per hour before signing in, 60 per hour and 300 per day signed in.
Can I feed a vector file or a scanned drawing back in?
The tracer reads raster only: PNG, JPEG, WEBP, BMP, GIF and AVIF. TIFF is not accepted — save the sheet as a PNG first — and there is no round-trip path for an existing vector file. Download links expire after 1 hour; uploads and generated files are removed by an automatic 1-day deletion rule.
Related pages
- Convert an image into vector shapes you can scale, edit and cut
- Turn a bitmap into real vector paths
- Convert an image to a DXF your CAD program will open
- Convert an image into a vector EPS
- Two colours, one clean silhouette contour
- DXF, explained: what is inside the file and what goes wrong
- A free alternative to Vectorizer.AI, trade-offs included