inb4 “it’s actually 14 kb 🤓” it’s the joy that is 1kb in size, here’s an image that’s 1kb:

  • isolatedscotch
    link
    fedilink
    arrow-up
    94
    ·
    edit-2
    1 day ago

    since the 1kb example was actually 1.1kb I re-made it while also improving the quality ten-fold

    (check it out - it really is 1kb!)

    • Valmond@lemmy.world
      link
      fedilink
      arrow-up
      10
      ·
      20 hours ago

      I bet a hand crafted png can be smaller still, we used png for video games back in the day and there were some heavy tweaking bringing them down to close to nothing (removing all metadata, reducing the number of colors, …).

      Maybe they were bigger than 1kb, gotta dig up some examples…

      • MHLoppy@fedia.io
        link
        fedilink
        arrow-up
        6
        ·
        18 hours ago

        The maximum color quantization you can do on this image without huge information errors is something like:

        • 1x yellow-brown for stars / streamers / shorts / socks
        • 1x brown for tree base (though it might be better to remove the base to save a color)
        • 1x red for baubles / hat / sweater / shoes
        • 1x green for tree
        • 1x blue for hair (potentially you could merge the tree color if you want to really push it)
        • 1x skin tone for skin
        • 1x pink for mouth
        • 1x black for lines
        • 1x light yellow for background

        Which is 9 total colors. This would also require living with aliased text ( c r u n c h y ), since it would be data-expensive to add extra shades of gray. At that point you’re no longer making a low-quality copy of the original - you’d basically be making a pixel art version of it since you can’t afford any colors for anti-aliasing and gradients.

        Here’s an example PNG with 9 unique colors and some pretty simple patterns without huge information density: https://files.catbox.moe/bj0acl.png

        Even that’s 1,847 bytes! (i.e., basically 2KB)


        Edit: I made a big (in hindsight, obvious) mistake by forgetting I can literally just change the bit depth of the image when saving, so the example I’ve provided is actually very inefficient by comparison. Valmond has set me straight.

        • Hoimo@ani.social
          link
          fedilink
          arrow-up
          1
          ·
          8 hours ago

          Finally somewhere my knowledge of anime can improve the world! The character is Yotsuba and her hair is green, not blue, so one color for both the tree and her hair is absolutely fine!

        • Valmond@lemmy.world
          link
          fedilink
          arrow-up
          2
          ·
          10 hours ago

          Nice!

          Did you use some specific soft to compress the png & get rid of the meta data? Because if you don’t then it will be way bigger.

          Also, you could anti alias the text with colours, that’s how it’s done on screens toaday, you just don’t see it from afar. And lastly, you could reduce the colorspace even more by reusing similarish colors.

          • Valmond@lemmy.world
            link
            fedilink
            arrow-up
            4
            ·
            10 hours ago

            I dug up some examples, I don’t have the talent needed to remake that, nor the executables for it but this is what people could compress back in the day:

            699 bytes:

            923 bytes:

            1.4kb:

            The same but smaller so 770 bytes:

            and some eye candy:

            2.3kb:

            2.1kb:

            10kb !

            So yeah, hard to push that original under 1kb I guess, but who knows ^^ !

            Merry Christmas !

            • MHLoppy@fedia.io
              link
              fedilink
              arrow-up
              1
              ·
              9 hours ago

              Shit, you’re absolutely right, I missed an (in hindsight very obvious) optimization - bit depth. It’s been so long since I’ve actually needed to worry about it that I forgot that the setting existed! What makes it even worse is that I did already play with quantizing the colors dwon to a more limited space, I just never baked that in as the bit depth haha.

          • MHLoppy@fedia.io
            link
            fedilink
            arrow-up
            1
            ·
            9 hours ago

            To be honest I’m not sure if the metadata actually matters much or not (I’ve never had to ultra-optimize like this before), but I just ran it through a PNG size optimizer and let it figure it out haha.

      • MHLoppy@fedia.io
        link
        fedilink
        arrow-up
        15
        ·
        22 hours ago

        It’s quite challenging to keep the text legible within a 1KB limit. Here I manually removed a few details that more-or-less weren’t visible post-compression anyway, then cut the color palette a little. You have to use such a low resolution with such high compression that almost everything gets amputated to keep the text kinda-readable (and even AVIF and JPEG XL (which are usually better than WebP) struggled, at least in my editor): https://files.catbox.moe/eyp2w7.webp

        If you can live with 2KB, you don’t have to amputate nearly as much: https://files.catbox.moe/g5htfo.webp

        In both cases I manually reconstructed the top of the star, but that’s a bit “extra” lol.


        And just for comparison, no text and 10KB at “full” res: https://files.catbox.moe/9bkn21.webp

        The same thing but half res (more optimal at this file size): https://files.catbox.moe/cac65u.webp


        Valmond’s implicit suggestion of not just quantizing as a pre-processing step (which is what I foolishly did), but actually reducing the saved bit depth of the image might give you something that looks much better overall than what any of the WebP versions we’ve been playing with here do - if you put in more effort!

        Here’s an example of a not-fully-optimized implementation that gets down to ~2.5KB as a PNG, or ~2KB as a lossless WebP (i.e., the two images are identical in quality):

        With some judicious manual optimization (which I haven’t done here), it’s plausible you could get this down to 1KB with better overall fidelity than the lossy WebP versions we’ve been playing with. Not 100% sure, as optimizing images for file sizes this small is not really my wheelhouse!

        My main concern with this approach is that you’re bottlenecked by resolution - large areas of plain color have a hard limit on their compression with PNG, but lossy compression can go wild with stuff like that.