• Etterra@lemmy.world
    link
    fedilink
    English
    arrow-up
    54
    arrow-down
    2
    ·
    3 days ago

    Ah yes, pollute the prompt. Nice. Reminds me of how artists are starting to embed data and metadata in their pieces that fuck up AI training data.

    • uis
      link
      fedilink
      English
      arrow-up
      2
      ·
      2 days ago

      Reminds me of how artists are starting to embed data and metadata in their pieces that fuck up AI training data.

      It still trains AI. Even adding noise does. Remember captchas?

      Metadata… unlikely to do anything.

      • Phoenix3875@lemmy.world
        link
        fedilink
        English
        arrow-up
        2
        ·
        2 days ago

        In theory, methods like nightshades are supposed to poison the work such that AI systems trained on them will have their performance degraded significantly.

        • lemmysarius@feddit.org
          link
          fedilink
          English
          arrow-up
          1
          ·
          1 day ago

          The problem with nightshade or similar tools is, that if you leave the changes it makes at too weak a setting, then it can be pretty easily removed. For example GAN upscalers that pre date modern “AI”, were pretty much built to remove noise or foreign patterns. And if you make the changes strong enough that they can’t be removed by these models (because so much information was lost), then the image looks like shit. Its really difficult to strike a balance here.

        • uis
          link
          fedilink
          English
          arrow-up
          1
          ·
          edit-2
          2 days ago

          Read it. I don’t think it will have bigger impact than lossy image compression or noisy raytraced images.