cross-posted from: https://programming.dev/post/3974080

Hey everyone. I made a casual survey to see if people can tell the difference between human-made and AI generated art. Any responses would be appreciated, I’m curious to see how accurately people can tell the difference (especially those familiar with AI image generation)

  • Gork@lemm.ee
    link
    fedilink
    English
    arrow-up
    15
    arrow-down
    4
    ·
    1 year ago

    14 / 20 here. I dunno why there are so many people, particularly on Reddit, who absolutely hate AI art. Yeah some of it can look janky, uncanny valley, or such but a lot of it looks really damn cool.

    And not all of us have talents to create visual art of our own so text creation is much more accessible for us to explore our imaginations. Or lack the money to commission pieces from human artists.

    • Pregnenolone@lemmy.world
      link
      fedilink
      English
      arrow-up
      1
      ·
      1 year ago

      A lot of Redditors don’t even know why they think a certain way, they think that way because everyone else around them thinks that way. There are some legit criticisms of AI art but most of the time it’s just bullshit lip service to artists when they don’t actually care

      • Gork@lemm.ee
        link
        fedilink
        English
        arrow-up
        1
        ·
        1 year ago

        Yeah I’ve had posts deleted on Reddit before because “ew AI art”. Like, I’m just trying to share interesting images. I’m not profiting off them in any way. But they take it so personally.

    • Sekoia@lemmy.blahaj.zone
      link
      fedilink
      English
      arrow-up
      3
      arrow-down
      3
      ·
      1 year ago

      Personally, I have no issue with models made from stuff obtained with explicit consent. Otherwise you’re just exploiting labor without consent.

      (Also if you’re just making random images for yourself, w/e)

      ((Also also, text models are a separate debate and imo much worse considering they’re literally misinformation generators))

      Note: if anybody wants to reply with “actually AI models learn like people so it’s fine”, please don’t. No they don’t. Bugger off. https://arxiv.org/pdf/2212.03860.pdf here have a source.

      • Even_Adder@lemmy.dbzer0.com
        link
        fedilink
        English
        arrow-up
        3
        ·
        1 year ago

        This paper is just about stock photos or video game art with enough dupes or variations that they didn’t get cut from the training set. The repeated images were included frequently enough to overfit. Which is something we already knew. That doesn’t really go to proving if diffusion models learn like humans or not. Not that I think they do.

        • Sekoia@lemmy.blahaj.zone
          link
          fedilink
          English
          arrow-up
          1
          ·
          1 year ago

          Sure, it’s not proof, but it gives a good starting point. Non-overfitted images would still have this effect (to a lesser extent), and this would never happen to a human. And it’s not like the prompts were the image labels, the model just decided to use the stock image as a template (obvious in the case with the painting).

          • Even_Adder@lemmy.dbzer0.com
            link
            fedilink
            English
            arrow-up
            1
            ·
            1 year ago

            Non-overfitted images would still have this effect (to a lesser extent),

            This is a bold claim to make with no evidence. When every trained image accounts for less than one byte of data in the model. Even the tiniest images file contain many thousands of bytes. One byte isn’t even enough to store a single character of text, most Latin-based alphabets and some symbols, use two bytes.

            and this would never happen to a human.

            There are plenty of artists that get stuck with same-face. Like Sam Yang for instance. Then there are the others who can’t draw disabled people or people of color. If it isn’t a beautiful white female character, they can’t do it. It can take a lot of additional training for people to break out of their rut, some don’t.

            I’m not going to tell you that latent diffusion models learn like humans, but they are still learning. https://arxiv.org/pdf/2306.05720.pdf Have a source.

            I recommend reading this article by Kit Walsh, a senior staff attorney at the EFF if you haven’t already. The EFF is a digital rights group who most recently won a historic case: border guards in the US now need a warrant to search your phone.

            This guy also does a pretty good job of explaining how latent diffusion models work, You should give this a watch too.