I played a bit with img2img and no prompt (or just a style, like anime) And even though my first reaction was *Oh cool, I’m a girl ! * to Why does AI always turns me in a girl ?

Here is a little experiment starting with the same image, a photo of a dude with a red T-shirt with a city panorama in the background (I won’t publish the original photo), and I give zero other guidance (no prompt, no style) The two stable diffusion model would give "male character’ while Dreamshaper and Meinamix would give feminine character. I’ve done a second generation with Meinamix and Dreamshaper giving female character again.

Second experiment Same dude as above (OP) but with long hair and a pink-doublet over a white shirt. Here I would understand that AI gets confused (Even though I don’t see why only girl could get long hair and wear pink clothe)

So the result, is like Dreamshaper and Meinamix are pretty close from my original clothing. Stable Diffusion 1.4 and Meinamix definitly choose a lady, Dreamspaher is more ambiguous but still on a feminine part of the gender spectrum, and stable diffusion 2 gives shit but let’s say that the beards make-it male. But again, these model do pretty great at “drawing my clothing” but seems biased toward female-passing result.

While writing that post I’ve done a few new generations with meinamix I got 2 guys out of 6 images (1 out of 6 with Dreamshaper), so here is one pretty close of the original photo

Obviously, it’s not a statistical study, I know that I could easily add a “negative prompt”. But this triggered my curiosity and was a fun experiment to do.

  • Thelsim@sh.itjust.worksM
    link
    fedilink
    English
    arrow-up
    3
    ·
    1 year ago

    Maybe it’s because of the shirt color?
    But yes, I’ve found biases to certain elements all over the place. You want a drawing with x, then the AI is obsessed with including y as well. But only when it’s a drawing, photos will come with z instead.
    So maybe in this case something something red something something will usually end up as a girl without further input.

    • Ziggurat@sh.itjust.worksOP
      link
      fedilink
      English
      arrow-up
      1
      ·
      1 year ago

      For the second one, A pink doublet and long-hair definitly put weight in the female category, and I can see how a model can get confused. So i was expected to see ladies coming-out.

      But for the first-one ? A red T-shirt isn’t something pretty gendered.

      I think I’ll do a bunch of other tests, using less known actor

      • Thelsim@sh.itjust.worksM
        link
        fedilink
        English
        arrow-up
        2
        ·
        1 year ago

        I don’t think it’s as simple as matching human sensibilities with an AI’s algorithm.
        What might be obvious to us, might not be the same for it.

      • Apathy Tree@lemmy.dbzer0.com
        link
        fedilink
        English
        arrow-up
        2
        ·
        1 year ago

        Try with neutral clothing colors as well, like gray or tan, to rule out the “pink is just a shade of red, so red=girl” possibility.

        • Ziggurat@sh.itjust.worksOP
          link
          fedilink
          English
          arrow-up
          1
          ·
          1 year ago

          An attempt with a black T-shirt and the xxmix9realistic_v40 model (I am also just trying out random model) glad to see that AI thinks I am that cute :)

          But I finally got the first Sexy-guy from the same photo original photo (SD 1.5, style = fantasy )

          I’ll try to collect some extra stats on Sunday using a few popular model and a set original image I can share. But I am eager to believe that there is a bias.

  • wolfshadowheart@kbin.social
    link
    fedilink
    arrow-up
    3
    ·
    1 year ago

    You can see model biases by using a model with a blank positive prompt and a negative prompt to “low quality”. Run 20 gens and you’ll see what the model likes to lean into.

    You can also do gentle inferences with a token for positive prompt. For example, if you are trying to find color biases something like Red_Outfit with the low quality negative prompt. However once you start adding more tokens it mostly goes out the window so it’s not something to rely on so much as to get a sense for what might be more likely to occur.

    Very helpful for testing models you get from civitai or wherever else

    • Ziggurat@sh.itjust.worksOP
      link
      fedilink
      English
      arrow-up
      1
      ·
      1 year ago

      Thanks for the link, I’ll check that out. It’s not necessarily a problem (Well I am a RPG player/GM trying to get illustrations for character so it can become one). Luckily, I am old enough and comfortable enough to not care about which gender the AI would give to my own photos. But I found that fun and interesting…