• ZILtoid1991@kbin.social
    link
    fedilink
    arrow-up
    12
    ·
    1 year ago

    There’s some people that specializes in reverse engineering prompts. Sometimes it’s funny, as they often disprove posts claiming "this is what the AI gave to me to “average [political viewpoint haver]”, only to turn out their prompt never contained the words “liberal”, “conservative”, etc, but words describing the image.

    • Omega_Haxors@lemmy.ml
      link
      fedilink
      arrow-up
      5
      ·
      edit-2
      1 year ago

      Reversing prompts is kind of pseudoscience. It’s like using a C# decompiler on a JAR file. Yes it produces working code from the binary, but it’s nowhere near what the original writer intended. They are also rife with false positives and negatives and that’s ignoring the weird idiosyncrasies such as nonsense tokens based on random artists who it thinks is 0.001% the style of, because it can’t find a better token to use instead. They can’t really tell you what the AI was thinking, but rather just make an educated guess which is more often than not completely wrong. Anyone claiming to be able to reverse engineer these black boxes flawlessly is outright lying to you. Nobody knows how they work.

      EDIT: and prompt reversal also assumes that someone is using a certain model, not switching it out half way for another model (or multiple times even) using reference photos through img->img or adding in custom drawing through inpaint at any point of the process. Like, I can’t even begin to begin on how literally impossible it would be to untangle that absolute mess of chaos when all you have is the end result.