No image loaded yet
Load an image to see the traced result — updates live as you adjust settings
No image loaded yet
Original
Result
100%
Settings — defaults = byte-exact to source
Point — faithful
Regions — invents edges
Illustrator — few shapes
Triangles — mesh instead of shapes
Triangle mesh
Classic — Regions pipeline (comparison only)
⚠ Classic is not a second triangle method — it is the Regions pipeline, the same one as the method Regions above. On the photo it costs five times the time and file size (measured: 22.6 s instead of 4.6 s, 11.9 MB instead of 5.0 MB, 108,637 paths instead of 12,000). It is here for comparison, not for daily use. — Verified 2026-09-07: the output is pfadgleich with the method Regions (108,637 paths, 1.81 million curve segments in both). So this switch cannot deliver anything here that Regions does not already deliver. Kept on purpose, as long as the triangle mesh is still being worked on.
Foliage, fur, grain and fabric have no shape a vector can describe — only a statistic. It is measured here per image cell (grain and contrast, separated by direction) and generated in the renderer. Not your leaf, but a leaf of the same grain in the same place. Zero extra triangles.
The mesh refines itself where it matches the image worst. More triangles = closer to the original, but at some point a triangle is smaller than a pixel — then it is a raster, not a vector any more. The status line says how many those are.
Foliage, fur, grain and fabric carry no shapea vector could describe — only a statistic. It is measured and generated in the renderer: zero extra path segments, a few kilobytes, no pixel grid. Measured per image cell and additionally separated by direction. Only in the Triangles method. In the others the texture was removed on 2026-09-05 — it did not convince there. ⛔ This deliberately measures worse. It is not your leaf but one with the same grain in the same place. Judge it by eye, not by the number.
Simplify — leave out
Contrast — dig out
Two opposite ways of working — not two strengths of the same thing.
Simplify takes away: coarser outlines, fewer colours, a calmer image. Both sliders at 0 compute literally whatthis tool has always computed; 60/60 is the painterly illustration look, 15/0 a light drawing touch. You can run any mix.
Contrast goes in the opposite direction: it digs out differences that used to get lost. A finer region map, and the colours go where something stands out from its neighbours instead of where there is a lot of area. In practice the better choice for logos, banners and graphics and for anything with small important spots — eyes, lettering, highlights.
Simplify takes away: coarser outlines, fewer colours, a calmer image. Both sliders at 0 compute literally whatthis tool has always computed; 60/60 is the painterly illustration look, 15/0 a light drawing touch. You can run any mix.
Contrast goes in the opposite direction: it digs out differences that used to get lost. A finer region map, and the colours go where something stands out from its neighbours instead of where there is a lot of area. In practice the better choice for logos, banners and graphics and for anything with small important spots — eyes, lettering, highlights.
How freely the outlines are traced. At 0 the tool places its own small arc for every pixel — measured: average arc length 1.00 pixel. Formally that is a curve, but it reproduces exactly the pixel staircase, which is why shapes look punched out when you zoom in a lot. To the right an arc may span the whole shape and sit slightly off — the way a hand draws. Colours are left untouched.
⚠ From about 40 % it tips over: leaves turn into angular shards. Judge it while zooming in.
⚠ From about 40 % it tips over: leaves turn into angular shards. Judge it while zooming in.
How strongly the number of colours is pulled together, and how generously small slivers are swallowed by their surroundings. The palette is chosen by visual impact — strong colours get weight, pale ones less.
⚠ This scales down the colour count you set: at 60 %, 400 colours become eleven. That is exactly when your five fine controls above barely act any more. That is why this slider was separated from the shape. Shapes are left untouched.
⚠ This scales down the colour count you set: at 60 %, 400 colours become eleven. That is exactly when your five fine controls above barely act any more. That is why this slider was separated from the shape. Shapes are left untouched.
How far this method zeichnen statt abschreiben should go. To the right: small slivers are swallowed by their surroundings, outlines are simplified as shape instead of traced pixel by pixel, and the palette is chosen by visual impact (strong colours get weight, pale areas less) instead of by pure colour distance. 0 computes exactly the old result — the transcript. The deviation number grows to the right, and that is not a mistake: a drawing is not the photo.
Traces the outlines as a line. It runs on the paths that are there anyway — not a single extra path segment. The line is not a uniform black but the darkened colour of its own area: a black frame around every area looks like a comic, a darkened own colour like a drawing. 0 = off.
Foliage, fur, grain and fabric carry no shape a vector can describe —
only a statistic. This measures that statistic per region (grain size and
contrast) and hands it to an SVG noise filter, which generates it in the
renderer. Not your leaf, but a leaf of the same grain in the same place.
Zero extra path segments, zero bytes per grain, no pixel grid.
The colours you set are shared out across the image — and until now that happened by area: whatever takes up a lot of space gets a lot of colour. A pupil, however, is 0.03 % of the image and has practically no vote in such a ballot; its colour is swallowed by the nearest big blob. That is why eyes came out "roughly, but not reliably".
Switched on, every small area is asked how strongly it stands out from its neighbours — and whatever stands out gets more of a vote. niemandem etwas verloren: weight is only added, never taken away. A soft skin gradient barely differs from its neighbours, so it keeps exactly what it had and cannot fall apart into bands.
This is the only place in the whole engine that relates two areas to each other — contrast is not a property of one area but of a pair.
Switched on, every small area is asked how strongly it stands out from its neighbours — and whatever stands out gets more of a vote. niemandem etwas verloren: weight is only added, never taken away. A soft skin gradient barely differs from its neighbours, so it keeps exactly what it had and cannot fall apart into bands.
This is the only place in the whole engine that relates two areas to each other — contrast is not a property of one area but of a pair.
Before drawing, every pixel has to be assigned to an area. Until now most pixels do not get their own assignment but that of their neighbourhood block of about 6×6 pixels — all of them share one decision. By construction that is a mosaic, and it is the reason the region map looks coarser than it needs to. Only what was recognised as hair or edge is already handled individually.
Switched on, every pixel decides for itself. That gives even something smaller than a block — a pupil, a catchlight in an eye, a nostril — its own vote for the first time.
On by default since 2026-09-09. Checked on the cat's eyes: iris smoother, eyelid rim more continuous, fur finer — and the soft skin gradients do not break into patches. Cost: about a fifth more shapes, 7 % more curve segments, one second. The old worry that this would shatter into confetti was measured and does not hold.
Switched on, every pixel decides for itself. That gives even something smaller than a block — a pupil, a catchlight in an eye, a nostril — its own vote for the first time.
On by default since 2026-09-09. Checked on the cat's eyes: iris smoother, eyelid rim more continuous, fur finer — and the soft skin gradients do not break into patches. Cost: about a fifth more shapes, 7 % more curve segments, one second. The old worry that this would shatter into confetti was measured and does not hold.
When tracing an outline it has to be decided what is a real corner and what is only a step of the pixel grid. Until now the immediate neighbour was consulted — and that is one or two pixels away. At that distance jede grid step looks like a corner, and at every corner a gerades piece begins. That is why leaves become polygons with chopped-off tips, as if punched out. Switched on, the look is taken from a greater distance: a real corner is still sharp from there, a grid step averages out to the smooth direction it really means. Tiny shapes are exempt — for a thing of three pixels the corner split is the cheaper description.
⚠ Checked on the photo and decided OFF (2026-09-08): the difference is barely visible at eightfold magnification, because the foliage shapes are mostly smaller than the exemption limit. Costs 3 % more curve segments. On large-area subjects — logos, graphics, portraits — it can pay off, so the switch stays.
⚠ Checked on the photo and decided OFF (2026-09-08): the difference is barely visible at eightfold magnification, because the foliage shapes are mostly smaller than the exemption limit. Costs 3 % more curve segments. On large-area subjects — logos, graphics, portraits — it can pay off, so the switch stays.
When this tool decides which two colours are "similar enough" to become one, until now it has gerechnet statt gesehen. In the dark the eye notices tiny differences, in the light only large ones — the pure number crunching does not know that difference and therefore distributes the colours wrongly: dark parts run together into mush, pale passages like haze and distant slopes lose their gradation. Switched on, the comparison is made in a colour space in which
distance roughly equals perception. Same number of colours, same computing time. Off is the default — viewed side by side on 2026-09-08 and decided that way. Off computes literally the old result: colour space and starting point of the colour search both go back.
The cross-border fit takes its slope from outside the region. This gives
the region its own average colour back, so a fit from outside can shape
the ramp without dragging the region's brightness towards its neighbours.
Off = pure vectors. On, an opaque photo carries a ~2 MB bitmap of itself
as the bottom layer. Measured 2026-09-02: it covers 8 pixels in the whole
interior, so it is off unless something proves otherwise.
Fill
Circles
Fill is the mode that counts. Circles is the historic first version and stays in for edge comparison.
Zigzag resolves the horizontal streaks and is the default. The other three are there only for comparison by eye.
Only an upper limit. Smaller images stay at their original size — sampling finer adds nothing, the elements are re-evaluated at every zoom level anyway.
Line art
Drawing
Line art sets black & white and fine strokes — for line drawings and technical images. Drawing sets few colours and large areas. That is the only way this method gets away from the pixel grid when enlarging : at the factory setting its regions are about two pixels in size, and a region of two pixels is a block, however smoothly its outline is drawn. Measured at 4× size: the factory setting sits practically on a blocky enlargement, Drawing breaks away from it — and pays with fidelity (deviation 0.30 → 5.93). Both are right, depending on what you want.
Black & White
Grayscale
Color
Also applies to the method Point: there the pixels are desaturated before handing over to the engine (luminance per Rec. 709); the engine itself is unchanged. Note: Black & white in the Point method becomes weich — this engine never guesses edges, so the hard threshold turns back into a gradient when drawing. If you need a hard, cuttable black & white drawing, use Regions.
Normal
Hand-drawn
Vector-only
Exact copy (raster)
Vector-only (default) is real vector output — pure cuttable, plottable, scalable paths, not a pixel image wrapped in an SVG. "Exact copy" pastes the source's own pixels back on top for byte-exact reproduction — useful when you specifically need a pixel-perfect copy, but it's not a vector-quality claim.
Standard 900
High 1400
Max 1800
Every upload is traced at this size on the long side. Standard is fast and small. Higher settings visibly reduce the pixel-grid look when you zoom in — a photograph's single-pixel regions drop from about a fifth of the image to a seventh — but the file and the trace time grow roughly with the pixel count. Only useful if your image is actually bigger than this.
Off
On (photos)
Removes the flat facets that show up in skin, faces and soft gradients when you zoom in. On a portrait it cuts the flat-patch share by a third and the worst pixel error by 3.5x, for about 18% more file. Leave it off for line art and logos — there is nothing there for it to smooth, and it only costs size.
Off
On
Places each boundary where the colour really crosses in the source instead of snapping it to the pixel grid, so edges come out as true curves rather than staircases. This is the one aimed at the “it still looks pixelly when I zoom in” problem. On flat art it is also slightly more accurate than leaving it off. Newer and less proven than the other controls — judge it by zooming in, not by the numbers.
Off
Gentle
Medium
Decides each pixel's region together with its neighbours instead of on its own, so a smooth area stops being split into thousands of one-pixel specks. This is the one aimed at the pixelly look when you zoom in — on a portrait it roughly halves the pixel-grid outline and improves the worst pixel. Leave it off for foliage or dense texture, where the pixel-scale variation is real detail. Judge it by zooming in, not by the numbers.
Off
On (photos)
Makes the result look like the photograph when you zoom in, instead of breaking into flat patches. It is a sub-pixel reconstruction, the same thing your image viewer does when it enlarges a photo — at normal size it changes nothing measurable (0.0004/255). It adds no detail that is not already there and does not touch the geometry; PDF, EPS and DXF export stay perfectly crisp either way.