• MacN'Cheezus@lemmy.todayOP
      link
      fedilink
      English
      arrow-up
      31
      arrow-down
      1
      ·
      edit-2
      11 months ago

      To be fair, it did spit out a couple of completely nonsensical images afterwards:

      I think the AI might be biased against dogs.

      • lseif@sopuli.xyz
        link
        fedilink
        arrow-up
        8
        ·
        11 months ago

        did it give you the images in base64 from an llm, or from an image generation model ?

        • ChaoticNeutralCzech@lemmy.ml
          link
          fedilink
          arrow-up
          7
          ·
          edit-2
          11 months ago

          I think you can guess that part. I doubt a current LLM can create a valid PNG, even if it’s just a 1x1px one that has been created before. This is partially because PNGs have a checksum and the LLM has definitely not seen enough PNGs in base64 to figure out the algorithm, and is not optimized to calculate checksums. In fact, I analyzed the image and the image header checksum is wrong even though the header makes sense (was likely stolen). Also, it gets penalized for repetition, which occurs a lot in image headers.

          AFAIK, the smallest valid image you see mentioned on the web is a 35-byte transparent pixel GIF, and the smallest PNG is a black pixel with 67 bytes:

          data:image/gif;base64,R0lGODlhAQABAAAAACH5BAEAAAAALAAAAAABAAEAAAIBAAA=
          data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABAQAAAAA3bvkkAAAACklEQVR4AWNgAAAAAgABc3UBGAAAAABJRU5ErkJggg==
          

          Testing rendering: Alt text for the GIF; if you see it, it failed, Alt text for the PNG; if you see it, it failed, another 67-byte PNG but 8 px wide: , or 1 gray pixel: , or a green one:

          The article + the generator