Move the mouse over the sheet, or focus the lens with Tab and use the arrow keys. The
type under the lens is still real text: select it, copy it, or read it with a screen
reader. Only its pixels are bent by an SVG displacement map.
Ridges, far to near: 1,240 m, 980 m, 610 m. Shoreline pines, eight. Birds in flight,
three. Water temperature at the jetty, 14 °C. Wind from the west, light.
Arrow keys move the lens, with Shift in bigger steps. Home centers it.
How it works
A loupe made of one displacement map
A magnifying glass shows the content from closer to its center than where you see it. That is a displacement: every pixel under the lens is fetched from a point nearer the middle. An SVG filter can do this to any HTML with a single generated map.
The pipeline
Generate the map once per lens shape: a PNG of the lens's square, whose red
and green channels hold, per pixel, how far to look towards the center.
Place it under the pointer with feImage's x and y.
Displace the content with feDisplacementMap, limited to the
lens square, and lay the result over the untouched content with feMerge.
Moving the lens only changes four attributes; the map is rebuilt only when the radius, zoom,
profile or rim change.
1. The zoom profile
For a pixel at distance t from the center (in units of the radius), a sampling factor f(t)
says where to look: f = 1/zoom magnifies, f = 1 shows the content where it is. The two
profiles differ in how they get from one to the other:
t = distance from the center / radius // 0 … 1
k = 1 − 1 / zoom
loupe: f(t) = 1 − k · (1 − smoothstep((t − (1 − rim)) / rim))
fisheye: f(t) = 1 − k · (1 − t²)²
a pixel at p shows the content at c + (p − c) · f(t)
so the offset stored in the map is (p − c) · (f(t) − 1)
A loupe magnifies evenly and squeezes everything that the magnified center covers
into a narrow rim, like a real magnifying glass. A fisheye is strongest in the center
and eases off all the way to the edge. Both reach f(1) = 1, so the lens meets its surroundings
without a seam, and t·f(t) only ever grows, so the picture never folds over; a unit test checks
that for every profile and setting.
This is what sets it apart from the liquid glass lens, which uses the same delivery path (a generated map in feImage) but
refracts only at its bezel and leaves the flat center alone. A loupe does the opposite: the
interior is the point.
2. Encoding the map
feDisplacementMap fetches each pixel from scale × (R − ½, G − ½) away. The offset (p − c)·(f − 1) is divided by the
largest offset in the lens, currently 0.355 × the radius, and stored around 128
in red and green, so the strongest pull reaches 1 or 255 and uses all 8 bits. The scale turns
that back into px: 68.52 for your settings. Outside the circle the map is 128, which moves nothing.
The PNG is encoded in plain TypeScript, with the same encoder as the CRT's barrel map, so it has no canvas color management in the
way.
<svg width="0" height="0" aria-hidden="true">
<!-- W × H: the element's size; X, Y: top-left of the lens square, in px -->
<filter id="loupe" x="0" y="0" width="W" height="H"
filterUnits="userSpaceOnUse" primitiveUnits="userSpaceOnUse"
color-interpolation-filters="sRGB">
<!-- 192 × 192 px PNG: red/green = offset towards the center, around 128 -->
<feImage href="data:image/png;base64,…" x="X" y="Y"
width="192" height="192" preserveAspectRatio="none" result="map" />
<feDisplacementMap in="SourceGraphic" in2="map" scale="68.52"
xChannelSelector="R" yChannelSelector="G"
x="X" y="Y" width="192" height="192" result="lens" />
<feMerge>
<feMergeNode in="SourceGraphic" />
<feMergeNode in="lens" />
</feMerge>
</filter>
</svg>
All coordinates are user-space px (filterUnits and primitiveUnits="userSpaceOnUse"): for an HTML element they start at the
top-left corner of its border box. Without them, the default filter region is the bounding
box plus 10%, and percentages would have to be recomputed on every resize.
3. Following the pointer, and the keyboard
A mouse anywhere over the element moves the lens; the lens follows with frame-rate
independent smoothing (the same helper as the holo card's
tilt), which is skipped when the user prefers reduced motion. The lens itself ignores mouse
clicks, because the pointer always sits in its middle and would otherwise block every link
underneath.
The lens is focusable and exposed as a two-dimensional slider, the pattern common color
pickers use: the arrow keys move it by 2% of the element (Shift: 10%), Home centers it, and
the value text reads both coordinates, for example "42% across, 38% down". On touch screens,
drag the lens itself; it sets touch-action: none only on itself, so the page still
scrolls everywhere else.
Readability
A filter magnifies the rendered pixels, it doesn't re-render the content larger. At 2× the
type under the lens is as sharp as a 2× bitmap zoom: fine on a high-DPI screen, visibly soft
on a 1× screen. It is a visual effect, not an accessibility zoom; for that the browser's own
zoom re-renders text properly. The loupe's rim does the opposite and squeezes content into a
narrow ring; displacement picks single pixels without averaging, so a very narrow rim looks
grainy. A wider rim or the fisheye profile keeps it smooth. The text never leaves the DOM,
so selection, find in page and screen readers work as usual.
Backdrop mode
By default the filter runs on the content, and every move of the lens reruns it over the
whole element. In Backdrop mode the same filter is applied as backdrop-filter: url(#id) on the lens instead, with the map at the lens's own
origin, so only the lens area is processed and anything behind the lens is magnified, not
just the component's children. Only Chromium renders SVG filters in backdrop-filter: Firefox drops the whole backdrop filter when it contains url(), and feature detection can't tell, since CSS.supports accepts
the syntax everywhere. The component therefore checks for a Chromium engine and otherwise falls
back to the content filter, which looks the same.
Browser support and performance
Content mode uses SVG filters on HTML, feImage with a data URL and feDisplacementMap, all Baseline widely available (web-features 3.44.0).
Backdrop mode is Chromium only, as above. Checked in Chromium; other engines are untested.
In content mode every lens move reruns the filter over the element. Only the lens square
is displaced, but the merge still covers the whole element, so keep the magnified area no
larger than needed. Backdrop mode only processes the lens.
Building a map is a loop over (2 × radius)² pixels plus an uncompressed PNG of the same
size, about 110 KB for the default radius. It happens when the lens shape changes, not
when it moves.