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.
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.)
Something to note is that Color Theif (Quantize) is using median cut on RGB, it would be interesting to try and extract dominant color in OKLab instead.
I also love the idea of a genetic algorithm to find an ideal match for a given image; it should be possible to simulate radial gradients server & client side with webgpu, but probably overkill for such a simple task.
EDIT: Although it works for me in Chrome, it doesn't seem to work in Safari v16.1.
Just in case anyone also misses it like I did, dark reader (at least on firefox) appears to apply itself to the final colors causing them to look quite bad and not match the input image at all. I would have discounted this entirely if it wasn't for all the praise I was seeing in the comments here.
Forgive my ignorance, feel like it's embarrassing to ask here to be honest, but can someone explain how this helps/works? I've never actually used these placeholders, but I always imagined that they work by processing the image beforehand on the server and using something like a super low quality image or gradient or such as the placeholder. If this is done in pure CSS, does the browser not need to download the image first to figure out what's in it, before then doing the placeholder effect? Perhaps it doesn't help that I've not had my morning coffee yet, but I don't understand.
Years ago before you could do anything this fancy with CSS I experimented with generating 3x2 pixel images server side and then presenting them as base64 encoded pngs in a "scoped" block of CSS to ensure they loaded before the src images. Coincidentally this was the same 3x2 layout as OP did here with CSS. I abandoned it because a 3x2 image scaled up looked terrible, and went with average color instead. This solution looks a lot better visually.
I still do the average color thing today since it's easy to calculate and store server side (I resize the uploaded image to 1x1 px and just record the result as a hex code in the DB).
We need to embrace WebP v2 for this kind of stuff. I took one of their images, resized it to 24x16px, and compressed it with Squoosh at 65% quality. It compresses to just 144 bytes. And it looks way way way better than these CSS gradients.
Definitely very low resolution, but compared to sites that use a solid color this seems much better. And only requiring one variable is really nice.
The article seems very well thought through. Though for both the algorithm and the benchmark algorithm the half blue / half green image with the lake shows the limitations of this technique. Still pretty good considering how light weight it is.
This works significantly better than I would have expected. I was just exploring extremely simple png strings as an alternative to the hash libraries requiring decoding. I had also explored two color css gradient, based on pregenerated major/minor colors, but too course to be useful (for a fast scrolling gallery). I’ll give this a test drive!
cool, but the fact that you can now do this with CSS is part of the reason that a new browser engine is so unlikely - one of 100000 things that css can do now and need to be supported :(
Maybe we should have kept CSS simple and JS optional. Maybe we took a few wrong turns...
I suppose the existence of bad uses does not invalidate the good but it feels like 99% of blurry image placeholder behavior is actually just preventing people from seeing anything unless they also run the ad and spying javascript that monetizes the site.
So a CSS-only way is neat and indisputably better but I think it's missing the point? The point of blurry placeholders isn't to make things easier or display better. The point is to make things worse. This write up is definitely making things better.
I know very little of css and to me it seems like a configuration file for rendering text, similar to changing default fonts ornsizes for matplotlib plots using plt.rcParams. How does this do inage blurring then?
Also a bit of nitpicking. While it provides a visual placeholder for an image that's being fetched, it does not reflect its content. So, when it's loaded we can see a completely different color palette and shapes.
Minimal CSS-only blurry image placeholders
(leanrada.com)439 points by ChiptuneIsCool 30 March 2025 | 69 comments
Comments
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...
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).
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.)
Something to note is that Color Theif (Quantize) is using median cut on RGB, it would be interesting to try and extract dominant color in OKLab instead.
I also love the idea of a genetic algorithm to find an ideal match for a given image; it should be possible to simulate radial gradients server & client side with webgpu, but probably overkill for such a simple task.
EDIT: Although it works for me in Chrome, it doesn't seem to work in Safari v16.1.
I still do the average color thing today since it's easy to calculate and store server side (I resize the uploaded image to 1x1 px and just record the result as a hex code in the DB).
https://squoosh.app/
Definitely very low resolution, but compared to sites that use a solid color this seems much better. And only requiring one variable is really nice.
The article seems very well thought through. Though for both the algorithm and the benchmark algorithm the half blue / half green image with the lake shows the limitations of this technique. Still pretty good considering how light weight it is.
https://0x0.st/820Q.html
Maybe we should have kept CSS simple and JS optional. Maybe we took a few wrong turns...
So a CSS-only way is neat and indisputably better but I think it's missing the point? The point of blurry placeholders isn't to make things easier or display better. The point is to make things worse. This write up is definitely making things better.
Also a bit of nitpicking. While it provides a visual placeholder for an image that's being fetched, it does not reflect its content. So, when it's loaded we can see a completely different color palette and shapes.