How it works
Fewer colors, on purpose
Posterizing throws away tonal detail: every color channel snaps to a handful of levels, so smooth shading turns into flat bands, like a screen print or a cel-shaded cartoon. One SVG primitive does it; ink lines along the edges complete the toon look.
1. Quantize: feComponentTransfer type="discrete"
<feComponentTransfer> runs each channel through its own transfer
function. With type="discrete" the function is a step function: n values in tableValues split the input range into n equal bins, and every input in bin k
comes out as the k-th value. Listing n evenly spaced values from 0 to 1 gives n levels per
channel with black and white both kept: 0 0.3333 0.6667 1 for
4 levels, up to 64 colors in total.
This is the same table the dither mode of the halftone page uses, but here it runs on R, G and B separately and keeps the hue, where halftone and duotone first reduce the image to luminance. That also means a channel can jump on its own: a gradient from orange to red crosses band borders in green only, which is where posterized images get their unexpected color steps.
2. Cleaner bands: blur and saturation first
- Smoothing blurs before quantizing. Without it, every bit of texture or noise near
a band border flips between two levels and the border frays. Blurred, the borders follow
the large shapes, which is what makes it read as cel shading rather than as a low color
depth. The blur also pulls transparent pixels into the border, so the transfer function
forces alpha to 1 and a final
<feComposite operator="in" in2="SourceGraphic">cuts the result back to the source's shape. - Saturation is an
<feColorMatrix type="saturate">before the steps. Quantizing tends to grey out mid-tones; a little extra saturation keeps the bands punchy. - Color space decides where the bins sit. In
sRGBthey are even on the gamma-encoded values, which is close to even for the eye. InlinearRGBthey are even in light, so on a perceptual ramp the darkest band is the widest, as the second ramp in the control panel shows.
3. Ink lines: edges from the edge filter
The ink comes from the same primitives as the edges and sketch page: two Sobel kernels in <feConvolveMatrix> with a bias of 0.5, folded
to a magnitude by a 1 0 1 table and added up. Here they run on R, G and B separately
and the three magnitudes are summed, because cartoon outlines sit between colors, and a red ball
in front of a blue wall can have almost the same luminance as the wall. They read the smoothed
source, not the bands, so lines follow shapes and not the shading steps.
An <feColorMatrix> then moves the edge strength into alpha (0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0), which turns it into the coverage of a flat <feFlood> in the ink color, merged over the bands with <feMerge>.
<svg width="0" height="0" aria-hidden="true">
<filter id="posterize" x="0" y="0" width="100%" height="100%"
color-interpolation-filters="sRGB">
<feGaussianBlur stdDeviation="1" result="soft" />
<feColorMatrix in="soft" type="saturate" values="1.15" />
<feComponentTransfer result="flat">
<feFuncR type="discrete" tableValues="0 0.3333 0.6667 1" />
<feFuncG type="discrete" tableValues="0 0.3333 0.6667 1" />
<feFuncB type="discrete" tableValues="0 0.3333 0.6667 1" />
<feFuncA type="linear" slope="0" intercept="1" />
</feComponentTransfer>
<!-- ink: Sobel on R, G and B (alpha forced to 1) -->
<feColorMatrix in="soft" type="matrix" result="opaque"
values="1 0 0 0 0 0 1 0 0 0 0 0 1 0 0 0 0 0 0 1" />
<feConvolveMatrix in="opaque" order="3" kernelMatrix="1 0 -1 2 0 -2 1 0 -1"
divisor="2.6667" bias="0.5" edgeMode="duplicate" preserveAlpha="true" />
<feComponentTransfer result="g0">
<feFuncR type="table" tableValues="1 0 1" />
<feFuncG type="table" tableValues="1 0 1" />
<feFuncB type="table" tableValues="1 0 1" />
</feComponentTransfer>
<feConvolveMatrix in="opaque" order="3" kernelMatrix="1 2 1 0 0 0 -1 -2 -1"
divisor="2.6667" bias="0.5" edgeMode="duplicate" preserveAlpha="true" />
<feComponentTransfer result="g1">
<feFuncR type="table" tableValues="1 0 1" />
<feFuncG type="table" tableValues="1 0 1" />
<feFuncB type="table" tableValues="1 0 1" />
</feComponentTransfer>
<feComposite in="g0" in2="g1" operator="arithmetic" k2="1" k3="1" />
<feColorMatrix type="matrix"
values="0.5 0.5 0.5 0 0 0.5 0.5 0.5 0 0 0.5 0.5 0.5 0 0 0 0 0 1 0" />
<feComponentTransfer>
<feFuncR type="linear" slope="1.5385" intercept="-0.5385" />
<feFuncG type="linear" slope="1.5385" intercept="-0.5385" />
<feFuncB type="linear" slope="1.5385" intercept="-0.5385" />
</feComponentTransfer>
<feMorphology operator="dilate" radius="1" />
<!-- edge strength → coverage of a flat ink layer -->
<feColorMatrix type="matrix" result="coverage"
values="0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0" />
<feFlood flood-color="#15111f" />
<feComposite in2="coverage" operator="in" result="ink" />
<feMerge>
<feMergeNode in="flat" />
<feMergeNode in="ink" />
</feMerge>
<!-- back to the source's shape -->
<feComposite in2="SourceGraphic" operator="in" />
</filter>
</svg> .poster {
filter: url(#posterize);
} Browser support and performance
SVG filters on HTML, including feComponentTransfer and feConvolveMatrix, have been Baseline widely available since July 2015
(web-features, svg-filters). Checked in Chromium and Firefox; Safari is
untested.
- Posterizing alone is one of the cheapest filters there is: a table lookup per channel. Smoothing adds a blur, and its cost grows with the radius.
- The ink lines add two 3×3 convolutions on three channels, a few cheap per-pixel steps and an optional dilate. Fine for static images, which are filtered once and cached; on animated, full-screen content it adds up.
- Kernels and dilation work on the browser's offscreen pixels, so on a high-DPI screen the lines are thinner in CSS pixels than at 1×.
- Text inside the filtered element is posterized and outlined too. Put the filter on images and illustrations, not on content that has to stay readable.