This is really cool, I love seeing folks use CSS in clever ways. :)

My one feedback would be to avoid using attr selectors on the style attribute like [style*="--lqip:"]. Browsers normally lazy compute that string version of the style attribute [1], but if you use a selector like this then on every style recalc it'll force all new inline styles (ex. element.style.foo = bar) to compute the string version.

Instead if you use a separate boolean attribute (or even faster a class) it'll avoid that perf foot gun. So write <div lqip style="--lqip: ..."> and match on that.

[1] https://source.chromium.org/chromium/chromium/src/+/main:thi...

    <img src="…" lqip="192900">

It's obviously mostly an aesthetic nitpick for this blog post specifically and not the project itself, because few people are going to be exploring the encoded space outside of the blog post, but the sliders letting you explore the LQIP space would "flash" a lot less if the base color was encoded in the high bits instead of the low bits.

Since there're independent Lightness values set for each section (I'd say quadrant but there are 6 of them), I wonder if two bits can be shaved from the `L` value from the base color. It'd take some reshuffling and might not play well with color customization in mainly flat images, but I think it could work.

I'm also curious to see that they're doing solely grayscale radial gradients over the base color instead of tweaking the base color's `L` value and using that as the radial gradient's center, I'd imagine you'd be doing more math that way in the OKLab colorspace which might give prettier results(?).

Tempted to play around with this myself, it's a really creative idea with a lot of potential. Maybe even try moving the centers (picking from a list of pre-defined options with the two bits stolen from the base color's L channel), to account for varying patterns (person portraits, quadrant-based compositions, etc).

Was expecting the common "background-image: data url + filter: blur" that a lot of static site generators produce, not a binary encoding algorithm implemented in CSS! Very impressive.

I wonder what other things could be encoded this way. Those generic profile pictures, perhaps? (The ones where your email or account id is hashed to produce some unique geometric pattern.)