Glassmorphism effects are trivial when the overlay is a rectangle, or a rectangle with border-radius. Irregular shapes — emoji or SVG icons — are a different problem entirely. What follows is the path to a working implementation for a nav bar whose icons get the frosted-glass treatment, along with the traps found along the way.

Starting point: emoji layers

The demo avoids SVG icons, using emojis with filters applied to match the theme instead. The first idea was to stack the two pseudos of each navigation link — both holding emoji content — offset and rotated via transform so they only partly overlap, then make the top one semitransparent and give it backdrop-filter: blur().

Animated gif. Shows a nav bar with four grey icons. On :hover/ :focus, a tinted icon slides and rotates, partly coming out from behind the grey one. In the area where they overlap, we have a glassmorphism effect, with the icon in the back seen as blurred through the semitransparent grey one in front.
The effect we’re after.

The Pug below generates the nav bar:

- let data = {
-   home: { ico: '🏠', hue: 200 }, 
-   notes: { ico: '🗒️', hue: 260 }, 
-   activity: { ico: '🔔', hue: 320 }, 
-   discovery: { ico: '🧭', hue: 30 }
- };
- let e = Object.entries(data);
- let n = e.length;

nav
  - for(let i = 0; i > n; i++)
    a(href='#' data-ico=e[i][1].ico style=`--hue: ${e[i][1].hue}deg`) #{e[i][0]}

and compiles to the HTML that follows:

<nav>
  <a href='#' data-ico='🏠' style='--hue: 200deg'>home</a>
  <a href='#' data-ico='🗒️' style='--hue: 260deg'>notes</a>
  <a href='#' data-ico='🔔' style='--hue: 320deg'>activity</a>
  <a href='#' data-ico='🧭' style='--hue: 30deg'>iscovery</a>
</nav>

Layout comes first: the elements become grid items, the nav sits in the middle, links get explicit widths, both pseudos of each link go in the top cell (pushing the link text into the bottom cell), and the link text and pseudos are middle-aligned.

body, nav, a { display: grid; }

body {
  margin: 0;
  height: 100vh;
}

nav {
  grid-auto-flow: column;
  place-self: center;
  padding: .75em 0 .375em;
}

a {
  width: 5em;
  text-align: center;
  
  &::before, &::after {
    grid-area: 1/ 1;
    content: attr(data-ico);
  }
}
Screenshot. Shows the four menu items lined up in a row in the middle of the page, each item occupying a column, all columns having the same width; with emojis above the link text, both middle-aligned horizontally.
Firefox screenshot of the result after we got layout basics sorted.

Emoji rendering varies by browser. A legible font is picked and its size bumped up, icons are made larger, backgrounds are set, and each link gets a color derived from its --hue custom property:

body {
  /* same as before */
  background: #333;
}

nav {
  /* same as before */
  background: #fff;
  font: clamp(.625em, 5vw, 1.25em)/ 1.25 ubuntu, sans-serif;
}

a {
  /* same as before */
  color: hsl(var(--hue), 100%, 50%);
  text-decoration: none;
  
  &::before, &::after {
    /* same as before */
    font-size: 2.5em;
  }
}
Screenshot. Shows the same layout as before, only with a prettier and bigger font and even bigger icons, backgrounds and each menu item having a different color value based on its --hue.
Chrome screenshot of the result (live demo) after prettifying things a bit.

The two emoji layers are then differentiated. The ::before pseudo is slightly moved and rotated, made monochrome with sepia(1), brought to the right hue and given extra contrast(). The ::after pseudo gets filter: grayscale(1) and is made semitransparent, otherwise the other pseudo would be invisible through it.

a {
  /* same as before */
  
  &::before {
    transform: 
      translate(.375em, -.25em) 
      rotate(22.5deg);
    filter: 
      sepia(1) 
      hue-rotate(calc(var(--hue) - 50deg)) 
      saturate(3);
  }
	
  &::after {
    opacity: .5;
    filter: grayscale(1);
  }
}
Screenshot. Same nav bar as before, only now the top icon layer is grey and semitransparent, while the bottom one is slightly offset and rotated, mono in the specified --hue.
Chrome screenshot of the result (live demo) after differentiating between the two icon layers.

The backdrop-filter wall

The next step was to set backdrop-filter: blur(5px) on the top (::after) layer. Firefox needs the gfx.webrender.all and layout.css.backdrop-filter.enabled flags set to true in about:config before the property works at all.

Animated gif. Shows how to find the flags mentioned above (gfx.webrender.all and layout.css.backdrop-filter.enabled) in order to ensure they are set to true. Go to about:config, start typing their name in the search box and double click their value to change it if it's not set to true already.
The flags that are still required in Firefox for backdrop-filter to work.

The result is an overlay the size of the entire top icon bounding box, with no visible blur on the bottom icon.

Screenshot collage. Shows the not really blurred, but awkward result with an overlay the size of the top emoji box after applying the backdrop-filter property. This happens both in Chrome (top) and in Firefox (bottom).
Chrome (top) and Firefox (bottom) screenshots of the result (live demo) after applying backdrop-filter.

Reducing to a known-working case

The debugging method: take an example that does work, adapt it until it breaks. The simplified working demo is an article inside a section. The section gets an image background, the article a semitransparent one, plus backdrop-filter on the article.

section { background: url(cake.jpg) 50%/ cover; }

article {
  margin: 25vmin;
  height: 40vh;
  background: hsla(0, 0%, 97%, .25);
  backdrop-filter: blur(5px);
}
Screenshot. Shows a working glassmorphism effect, where we have a semitransparent box on top of its parent one, having an image background.
Working glassmorphism effect (live demo) in a simplified test.

That works, but the two layers need to be siblings, not nested. Making both layers sibling article elements and partially overlapping them keeps the effect in Chrome and, mostly, in Firefox — except that Firefox's blur() edge handling looks awkward. Based on the images in the spec, the Firefox result also appears incorrect.

<article class='base'></article>
<article class='grey'></article>
article { width: 66%; height: 40vh; }

.base { background: url(cake.jpg) 50%/ cover; }

.grey {
  margin: -50% 0 0 33%;
  background: hsla(0, 0%, 97%, .25);
  backdrop-filter: blur(5px);
}
Screenshot collage. Shows the case where we have a semitransparent box on top of its sibling having an image background. The top panel screenshot was taken in Chrome, where the glassmorphism effect works as expected. The bottom panel screenshot was taken in Firefox, where things are mostly fine, but the blur handling around the edges is really weird.
Chrome (top) and Firefox (bottom) screenshots of the result (live demo) when the two layers are siblings.

For layers sitting on a solid background (white here), one fix is a box-shadow on the bottom layer (.base) with no offsets, no blur and a spread radius twice the blur radius of the top layer's backdrop-filter. It works in this case. If the parent has an image background that isn't fixed, a layered backgrounds approach could handle Firefox — out of scope here.

The real icons can't be square boxes, so semitransparency can't come from an hsla() background; it has to use opacity.

.grey {
  /* same as before */
  opacity: .25;
  background: hsl(0, 0%, 97%);
}
Screenshot. Shows the case where we have a subunitary opacity on the top layer in order to make it semitransparent, instead of a subunitary alpha value for the semitransparent background.
The result (live demo) when the top layer is made semitransparent using opacity instead of a hsla() background.

That is the break: making the top layer semitransparent via opacity kills the backdrop-filter effect in both Chrome and Firefox.

Spec and documentation

MDN's backdrop-filter page states:

Because it applies to everything behind the element, to see the effect you must make the element or its background at least partially transparent.

Read plainly, that suggests opacity shouldn't break the effect — but it does, in both browsers. The spec offers this, which may or may not be relevant:

The effect of the backdrop-filter will not be visible unless some portion of element B is semi-transparent. Also note that any opacity applied to element B will be applied to the filtered backdrop image as well.

An attempt with filter: opacity() instead of the opacity property produced inconsistent results across identical browsers, and a Twitter query returned mixed answers, so no reliable conclusion emerged.

Text, and what the blur respects

Swapping rectangles for text: the article elements get text content, explicit sizing is dropped, font-size is bumped up, the margin producing partial overlap is adjusted, and the background declarations are replaced with color. The bottom one gets aria-hidden='true'.

<article class='base' aria-hidden='true'>Lion 🧡</article>
<article class='grey'>Lion 🖤</article>
article { font: 900 21vw/ 1 cursive; }

.base { color: #ff7a18; }

.grey {
  margin: -.75em 0 0 .5em;
  color: hsla(0, 0%, 50%, .25);
  backdrop-filter: blur(5px);
}
Screenshot collage. Shows the case where we have a semitransparent text layer on top of its identical solid orange text sibling. The top panel screenshot was taken in Chrome, where we get proper blurring, but it's underneath the entire bounding box of the semitransparent top text, not limited to just the actual text. The bottom panel screenshot was taken in Firefox, where things are even worse, with the blur handling around the edges being really weird.
Chrome (top) and Firefox (bottom) screenshots of the result (live demo) when we have two text layers.

Two observations. First, setting color to a value with a subunitary alpha makes emojis semitransparent too, in both Chrome and Firefox — the other channels don't influence emojis. Second, both browsers blur the entire area of the orange text and emoji underneath the top layer's bounding box rather than only what lies under the actual glyphs. Firefox adds its sharp-edge artifact on top of that.

The box blur is consistent with the spec: "transparent" here means the full filtered backdrop image is visible, as with background-color: transparent. A test with non-text shapes — a background gradient, clip-path, mask — shows both browsers blurring the whole top-layer box for the gradient case, while Chrome honors the clip-path and mask shapes and Firefox does not.

Screenshot collage. Shows the case where we have semitransparent non-rectangular shaped layers (obtained with three various methods: gradient background, clip-path and mask) on top of a rectangular siblings. The top panel screenshot was taken in Chrome, where things seem to work fine in the clip-path and mask case, but not in the gradient background case. In this case, everything that's underneath the bounding box of the top element gets blurred, not just what's underneath the visible part. The bottom panel screenshot was taken in Firefox, where, regardless of the way we got the shape, everything underneath its bounding box gets blurred, not just what's underneath the actual shape. Furthermore, in all three cases we have the old awkward sharp edge issue we've had in Firefox before
Chrome (top) and Firefox (bottom) screenshots of the result (live demo) when the top layer is a non-rectangular shape.

Chrome: mask the backdrop to the glyphs

Applying a mask clipped to text on the top layer (.grey) makes the blur follow the text edges. This is Chrome-only: text is a non-standard mask-clip value supported only in WebKit browsers, and Firefox (above) doesn't confine the blur to a mask shape anyway.

/* same as before */

.grey {
  /* same as before */
  -webkit-mask: linear-gradient(red, red) text; /* only works in WebKit browsers */
}
Chrome screenshot. Shows two text and emoji layers partly overlapping. The top one is semitransparent, so through it, we can see the layer underneath blurred (by applying a backdrop-filter on the top one).
Chrome screenshot of the result (live demo) when the top layer has a mask restricted to the text area.

Chrome now shows the right result, but only because the demo used a heart emoji, which exists in orange and black. Other emojis don't come in black and white, so the next attempt made both layers identical, the top one semitransparent with filter: grayscale(1) — which also desaturated the blurred area underneath.

article { 
  color: hsla(25, 100%, 55%, var(--a, 1));
  font: 900 21vw/ 1.25 cursive;
}

.grey {
  --a: .25;
  margin: -1em 0 0 .5em;
  filter: grayscale(1);
  backdrop-filter: blur(5px);
  -webkit-mask: linear-gradient(red, red) text;
}
Chrome screenshot. Shows two text and emoji layers partly overlapping. The top one is semitransparent, so through it, we can see the layer underneath blurred (by applying a backdrop-filter on the top one). The problem is that applying the grayscale filter on the top semitransparent layer not only affects this layer, but also the blurred area of the layer underneath.
Chrome screenshot of the result (live demo) when the top layer gets a grayscale(1) filter.

The fix is a third layer: the base layer stays slightly offset, the middle layer is a transparent "ghost" carrying the backdrop-filter, and the top layer is semitransparent with grayscale(1).

body { display: grid; }

article {
  grid-area: 1/ 1;
  place-self: center;
  padding: .25em;
  color: hsla(25, 100%, 55%, var(--a, 1));
  font: 900 21vw/ 1.25 pacifico, z003, segoe script, comic sans ms, cursive;
}

.base { margin: -.5em 0 0 -.5em; }

.midl {
  --a: 0;
  backdrop-filter: blur(5px);
  -webkit-mask: linear-gradient(red, red) text;
}

.grey { filter: grayscale(1) opacity(.25) }
Chrome screenshot. Shows two text and emoji layers partly overlapping. The top one is semitransparent grey, so through it, we can see the layer underneath blurred (by applying a backdrop-filter on a middle, completely transparent one).
Chrome screenshot of the result (live demo) with three layers.

All that remains is making the base layer monochrome, and the Chrome effect is what was wanted.

/* same as before */

.base {
  margin: -.5em 0 0 -.5em;
  filter: sepia(1) hue-rotate(165deg) contrast(1.5);
}
Chrome screenshot. Shows two text and emoji layers partly overlapping. The bottom one is mono (bluish in this case) and blurred at the intersection with the semitransparent grey one on top.
Chrome screenshot of the result (live demo) we were after.

Firefox: element() as a background

Firefox alone supports element(), which uses an element as another element's background — the basis for a Firefox version. .base and .grey keep the Chrome styles; the middle layer's background is, via element(), a blurred copy of the layers.

<article id='blur' aria-hidden='true'>Lion 🦁</article>
<article class='midl'>Lion 🦁</article>

The blurred element is absolutely positioned (and kept in view initially), made monochrome and blurred, then used as the .midl background:

#blur {
  position: absolute;
  top: 2em; right: 0;
  margin: -.5em 0 0 -.5em;
  filter: sepia(1) hue-rotate(165deg) contrast(1.5) blur(5px);
}

.midl {
  --a: .5;
  background: -moz-element(#blur);
}
Firefox screenshot. Shows a blurred mono (bluish in this case) text and emoji element below everything else. 'Everything else' in this case is another text and emoji element that uses a semitransparent color so we can partly see through to the background which is set to the blurred element via the element() function.
Firefox screenshot of the result (live demo) when using the blurred element #blur as a background.

One immediate issue: margin offsets the actual #blur element but has no effect on its use as a background; a transform is needed for that, which also allows rotation and other transforms. A translation is used for now:

Firefox screenshot. Shows a slightly rotated blurred mono (bluish in this case) text and emoji element below everything else. 'Everything else' in this case is another text and emoji element that uses a semitransparent color so we can partly see through to the background which is set to the slightly rotated blurred element via the element() function.
Firefox screenshot of the result (live demo) when using transform: rotate() instead of margin on the #blur element.
#blur {
  /* same as before */
  transform: translate(-.25em, -.25em); /* replaced margin */
}
Firefox screenshot. Shows a slightly offset blurred mono (bluish in this case) text and emoji element below everything else. 'Everything else' in this case is another text and emoji element that uses a semitransparent color so we can partly see through to the background which is set to the slightly offset blurred element via the element() function. This slight offset means the actual text doesn't perfectly overlap with the background one anymore.
Firefox screenshot of the result (live demo) when using transform: translate() instead of margin on the #blur element.

Part of the blurred background gets cut off outside the middle layer's padding-box — harmless, since the next move clips the background to text, but useful space because .base will be translated the same distance. So padding is bumped up a little, alongside background-clip: text on .midl:

Firefox screenshot. Shows a slightly offset blurred mono (bluish in this case) text and emoji element below everything else. 'Everything else' in this case is another text and emoji element that uses a semitransparent color so we can partly see through to the background which is set to the slightly offset blurred element via the element() function. This slight offset means the actual text doesn't perfectly overlap with the background one anymore. It also means that the translated background text may not fully be within the limits of the padding-box anymore, as highlighted in this screenshot, which also shows the element boxes overlays.
Firefox screenshot highlighting how the translated #blur background exceeds the limits of the padding-box on the .midl element.
article {
  /* same as before */
  padding: .5em;
}

#blur {
  position: absolute;
  bottom: 100vh;
  transform: translate(-.25em, -.25em);
  filter: sepia(1) hue-rotate(165deg) contrast(1.5) blur(5px);
}

.midl {
  --a: .1;
  background: -moz-element(#blur);
  background-clip: text;
}
Firefox screenshot. Shows a text and emoji element that uses a semitransparent color so we can partly see through to the background which is set to a blurred element (now positioned out of sight) via the element() function. This slight offset means the actual text doesn't perfectly overlap with the background one anymore. We have also clipped the background of this element to the text, so that none of the background outside it is visible. Even so, there's enough padding room so that the blurred background is contained within the padding-box.
Firefox screenshot of the result (live demo) after clipping the .midl element’s background to text.

Next, the .base element gets essentially the Chrome styles, with transform replacing margin:

<article id='blur' aria-hidden='true'>Lion 🦁</article>
<article class='base' aria-hidden='true'>Lion 🦁</article>
<article class='midl'>Lion 🦁</article>
#blur {
  position: absolute;
  bottom: 100vh;
  transform: translate(-.25em, -.25em);
  filter: sepia(1) hue-rotate(165deg) contrast(1.5) blur(5px);
}

.base {
  transform: translate(-.25em, -.25em);
  filter: sepia(1) hue-rotate(165deg) contrast(1.5);
}

Shared styles are factored out by adding the .base class to #blur to avoid duplication:

<article id='blur' class='base' aria-hidden='true'>Lion 🦁</article>
<article class='base' aria-hidden='true'>Lion 🦁</article>
<article class='midl'>Lion 🦁</article>
#blur {
  --r: 5px;
  position: absolute;
  bottom: 100vh;
}

.base {
  transform: translate(-.25em, -.25em);
  filter: sepia(1) hue-rotate(165deg) contrast(1.5) blur(var(--r, 0));
}
Firefox screenshot. Shows two text and emoji layers slightly offset from one another. The .base one, first in the DOM order, is made mono with a filter and slightly offset to the top left with a transform. The .midl one, following it in DOM order, has semitransparent text so that we can see through to the text clipped background, which uses as a background image the blurred version of the mono, slightly offset .base layer. In spite of DOM order, the .base layer still shows up on top.
Firefox screenshot of the result (live demo) after adding the .base layer.

Because .base has a transform, it paints above .midl despite DOM order. The simplest fix is z-index: 2 on .midl:

Firefox screenshot. Shows two text and emoji layers slightly offset from one another. The .base one, first in the DOM order, is made mono with a filter and slightly offset to the top left with a transform. The .midl one, following it in DOM order, has semitransparent text so that we can see through to the text clipped background, which uses as a background image the blurred version of the mono, slightly offset .base layer. Having explicitly set a z-index on the .midl layer, it now shows up on top of the .base one.
Firefox screenshot of the result (live demo) after fixing the layer order such that .base is underneath .midl.

A subtler issue follows: .base remains visible under the semitransparent parts of the blurred .midl background, because blurring makes edge-adjacent pixels semitransparent. On a solid parent background — the case here — setting background-color on .midl to that same value removes the sharp edges showing through.

Screenshot. Shows two lines of blue text with a red outline to highlight the boundaries of the actual text. The text on the second line is blurred and it can be seen how this causes us to have semitransparent blue pixels on both sides of the red outline - both outside and inside.
The blur effect around the edges.
.midl {
  /* same as before */
  background: -moz-element(#blur) #fff;
  background-clip: text;
}
Firefox screenshot. Shows two text and emoji layers slightly offset from one another. The .base one, first in the DOM order, is made mono with a filter and slightly offset to the top left with a transform. The .midl one, following it in DOM order, has semitransparent text so that we can see through to the text clipped background, which uses as a background image the blurred version of the mono, slightly offset .base layer. Having explicitly set a z-index on the .midl layer and having set a fully opaque background-color on it, the .base layer now lies underneath it and it isn't visible through any semitransparent parts in the text area because there aren't any more such parts.
Firefox screenshot of the result (live demo) after ensuring the .base layer isn’t visible through the background of the .midl one.

Adding the top .grey layer with the Chrome styles fails: .midl's z-index: 2 puts it above what should be the top layer, despite DOM order. The fix is z-index: 3 on .grey. It's layer after layer of z-index, but low effort and functional.

.grey { filter: grayscale(1) opacity(.25); }
Firefox screenshot. Shows two text and emoji layers slightly offset from one another. The .base one, first in the DOM order, is made mono with a filter and slightly offset to the top left with a transform. The .midl one, following it in DOM order, has transparent text so that we can see through to the text clipped background, which uses as a background image the blurred version of the mono, slightly offset .base layer. Since the background-color of this layer coincides to that of their parent, it is hard to see. We also have a third .grey layer, the last in DOM order. This should be right on top of the .midl one, but, due to having set a z-index on the .midl layer, the .grey layer is underneath it and not visible, in spite of the DOM order.
Firefox screenshot of the result (live demo) after adding the top .grey layer.
.grey {
  z-index: 3;
  filter: grayscale(1) opacity(.25);
}
Firefox screenshot. Shows two text and emoji layers partly overlapping. The bottom one is mono (bluish in this case) and blurred at the intersection with the semitransparent grey one on top.
Firefox screenshot of the result (live demo) we were after.

Merging into one cross-browser version

The Firefox code is the more complete of the two, so it's the base. Its extra z-index declarations and the out-of-sight #blur element don't affect Chrome. What Chrome still needs is backdrop-filter and mask on .midl:

<article id='blur' class='base' aria-hidden='true'>Lion 🦁</article>
<article class='base' aria-hidden='true'>Lion 🦁</article>
<article class='midl' aria-hidden='true'>Lion 🦁</article>
<article class='grey'>Lion 🦁</article>
body { display: grid; }

article {
  grid-area: 1/ 1;
  place-self: center;
  padding: .5em;
  color: hsla(25, 100%, 55%, var(--a, 1));
  font: 900 21vw/ 1.25 cursive;
}

#blur {
  --r: 5px;
  position: absolute;
  bottom: 100vh;
}

.base {
  transform: translate(-.25em, -.25em);
  filter: sepia(1) hue-rotate(165deg) contrast(1.5) blur(var(--r, 0));
}

.midl {
  --a: 0;
  z-index: 2;
  background: -moz-element(#blur) #fff;
  background-clip: text;
}

.grey {
  z-index: 3;
  filter: grayscale(1) opacity(.25);
}
backdrop-filter: blur(5px);
-webkit-mask: linear-gradient(red, red) text;

Since neither declaration should apply in the other browser, @supports separates them:

$r: 5px;

/* same as before */

#blur {
  /* same as before */
  --r: #{$r};
}

.midl {
  --a: 0;
  z-index: 2;
  /* need to reset inside @supports so it doesn't get applied in Firefox */
  backdrop-filter: blur($r);
  /* invalid value in Firefox, not applied anyway, no need to reset */
  -webkit-mask: linear-gradient(red, red) text;
  
  @supports (background: -moz-element(#blur)) { /* for Firefox */
    background: -moz-element(#blur) #fff;
    background-clip: text;
    backdrop-filter: none;
  }
}
Chrome (top) and Firefox (bottom) screenshot collage of the text and emoji glassmorphism effect for comparison. The blurred backdrop seems thicker in Chrome and the emojis are obviously different, but the result is otherwise pretty similar.
Chrome (top) and Firefox (bottom) screenshots of the result (live demo) we were after.

The two browsers don't render identically, but the results are similar.

Why one element isn't enough

Collapsing the solution into a single element isn't possible. The Firefox version needs at least two elements, since one supplies the background for another via its id. For the remaining three layers (all the Chrome version needs), each carries at least one property that irreversibly propagates to children and pseudos: filter on .base and .grey, mask on the middle layer.

Plain text without emojis is a lighter case: two elements suffice, only one of which is needed by Chrome, the other being #blur for Firefox. The two pseudos of .text build the base layer (::before) and a combination of the other two layers (::after). Without emojis, no grayscale(1) is needed — saturation of the color value is controlled directly. The pseudos are stacked, the lower one offset by the same amount and colored like #blur; that color depends on a flag, --f, which drives saturation and alpha: 100% saturation and alpha 1 for #blur and ::before (--f: 1), 0% saturation and alpha .25 for ::after (--f: 0).

<article id='blur'>Blood</article>
<article class='text' aria-hidden='true' data-text='Blood'></article>
$r: 5px;

%text { // used by #blur and both .text pseudos
  --f: 1;
  grid-area: 1/ 1; // stack pseudos, ignored for absolutely positioned #base
  padding: .5em;
  color: hsla(345, calc(var(--f)*100%), 55%, calc(.25 + .75*var(--f)));
  content: attr(data-text);
}

article { font: 900 21vw/ 1.25 cursive }

#blur {
  position: absolute;
  bottom: 100vh;
  filter: blur($r);
}

#blur, .text::before {
  transform: translate(-.125em, -.125em);
  @extend %text;
}

.text {
  display: grid;
	
  &::after {
    --f: 0;
    @extend %text;
    z-index: 2;
    backdrop-filter: blur($r);
    -webkit-mask: linear-gradient(red, red) text;

    @supports (background: -moz-element(#blur)) {
      background: -moz-element(#blur) #fff;
      background-clip: text;
      backdrop-filter: none;
    }
  }
}

Back to the nav bar

In this use case the glassmorphism applies only to the link icon, not the whole link with text, which simplifies things slightly. The Pug:

- let data = {
-   home: { ico: '🏠', hue: 200 }, 
-   notes: { ico: '🗒️', hue: 260 }, 
-   activity: { ico: '🔔', hue: 320 }, 
-   discovery: { ico: '🧭', hue: 30 }
- };
- let e = Object.entries(data);
- let n = e.length;

nav
  - for(let i = 0; i < n; i++)
    - let ico = e[i][1].ico;
    a.item(href='#' style=`--hue: ${e[i][1].hue}deg`)
      span.icon.tint(id=`blur${i}` aria-hidden='true') #{ico}
      span.icon.tint(aria-hidden='true') #{ico}
      span.icon.midl(aria-hidden='true' style=`background-image: -moz-element(#blur${i})`) #{ico}
      span.icon.grey(aria-hidden='true') #{ico}
      | #{e[i][0]}

and the resulting HTML:

<nav>
  <a class='item' href='#' style='--hue: 200deg'>
    <span class='icon tint' id='blur0' aria-hidden='true'>🏠</span>
    <span class='icon tint' aria-hidden='true'>🏠</span>
    <span class='icon midl' aria-hidden='true' style='background-image: -moz-element(#blur0)'>🏠</span>
    <span class='icon grey' aria-hidden='true'>🏠</span>
    home
  </a>
  <!-- the other nav items -->
</nav>

Part of these spans could be pseudos, but spans are more consistent and easier. Since every item has its own icon, each has a different blurred icon layer — the .midl element's background is set in the style attribute, so adding or removing entries in the data object (and thus changing the item count) requires no CSS changes.

Layout and prettifying styles stay close to the initial nav bar; the pseudos in each item's grid top cell are simply spans now:

span {
  grid-area: 1/ 1; /* stack all emojis on top of one another */
  font-size: 4em; /* bump up emoji size */
}

The icon layers keep the original transform and filter chains chosen when the pseudos were used. The color: hsla() declaration on the span layers is dropped: with emojis only the alpha channel matters, and while the default of 1 is preserved on .base and .grey, .midl gets color: transparent directly. background-color on .midl is only needed in the Firefox case, since background-image already comes from the style attribute:

.base { /* mono emoji version */
  transform: translate(.375em, -.25em) rotate(22.5deg);
  filter: sepia(1) hue-rotate(var(--hue)) saturate(3) blur(var(--r, 0));
}

.midl { /* middle, transparent emoji version */
  color: transparent; /* so it's not visible */
  backdrop-filter: blur(5px);
  -webkit-mask: linear-gradient(red 0 0) text;
  
  @supports (background: -moz-element(#b)) {
    background-color: #fff;
    background-clip: text;
    backdrop-filter: none;
  }
}
Chrome (top) and Firefox (bottom) screenshot collage of the emoji glassmorphism effect for comparison. The emojis are obviously different, but the result is otherwise pretty similar.
Chrome (top) and Firefox (bottom) screenshots of the desired emoji glassmorphism effect (live demo).

The effect shouldn't be permanent — only on :hover and :focus. A flag, --hl, is 0 normally and 1 in those states, controlling opacity and transform on the .base spans:

$t: .3s;

a {
  /* same as before */
  --hl: 0;
  color: hsl(var(--hue), calc(var(--hl)*100%), 65%);
  transition: color $t;
  
  &:hover, &:focus { --hl: 1; }
}

.base {
  transform: 
    translate(calc(var(--hl)*.375em), calc(var(--hl)*-.25em)) 
    rotate(calc(var(--hl)*22.5deg));
  opacity: var(--hl);
  transition: transform $t, opacity $t;
}

Open question: SVG

Would real SVG icons be simpler than a span sandwich? It would make more sense — not every icon exists as an emoji — but, having gone through all of the above for the emoji version, it is neither less code nor simpler.