• Tonuka@feddit.org
    link
    fedilink
    English
    arrow-up
    6
    ·
    17 hours ago

    ok but AI compression is legitimately a thing that has potential because of how extreme the size reduction could be. But yeah it’s definitely no lossless format lol

    • Tiresia@slrpnk.net
      link
      fedilink
      English
      arrow-up
      7
      ·
      12 hours ago

      Imagine replacing detail with the most common thing that mostly fits. Take a picture of a sculpture with a cool texture? Sorry, that’s incredibly hard to specify, but it’s marble so the AI knows to redraw it as a (smooth) marble statue.

      Take a picture of your partner’s eyes? Geez, do you know how complex eyes are? Look, I’m not unreasonable, the eyes are sea green in high detail with flecks of brown and a slight shimmer in the bottom right corner. Now you can have a nice compact image of your partner.

      Imagine a social media where images look weird and inconsistent unless they conform to conditions that don’t produce hallucinations. Expected textures, expected lighting conditions, expected locations, expected hairstyles and expected facial features.

      And yes, I do know AI sharpening is already deployed. Every time I zoom in on pictures I took with my phone I feel grossed out by the hallucinatory soup.

    • Ignotum@lemmy.world
      link
      fedilink
      English
      arrow-up
      1
      ·
      edit-2
      8 hours ago

      For text, if you have a decent probability distribution of the next piece of text based on the past context, then you can create a much more efficient encoding, so in that case you can create a good lossless compression
      decompression would be terribly slow though, and it doesn’t easily extend to images

      Are people downvoting just because LLM? It’s basic information theory, i’m not suggesting we start using it, it would be slow to the point of uselessness

      • Tonuka@feddit.org
        link
        fedilink
        English
        arrow-up
        2
        ·
        15 hours ago

        To be clear I’m not speculating on something that could potentially be cool, this already exists and is improving

        • Ignotum@lemmy.world
          link
          fedilink
          English
          arrow-up
          1
          ·
          14 hours ago

          Yeah i was just pointing out one case where it can be used for lossless compression, for lossy compression you can use autoencoders and stuff, but they are lossy and less efficient than jpeg and the like