• SootySootySoot [any]@hexbear.netEnglish
    22·
    2 天前

    As people don’t seem to realise. It’s a known trick that when you ask people to draw bicycle from memory, the vast majority won’t do it correctly. Even (to some degree) including cyclists.

    Though it was originally just a weird fun thing an artist discovered, it has been touted as a neurological trick that demonstrates the difference between the brain being able to recognise something and actually knowing what it looks like.

    • ToastedRavioli@midwest.socialEnglish
      15·
      2 天前

      Semi-related, but this was also my personal test for image generative AI. Unsurprisingly, despite having significant reference information for what a Zamboni looks like, it always fails to properly make an accurate image of one. And I dont mean “close enough”, I mean just way off.

      Here is output from like 2021 or so:

      Here is output from 2023, which is even worse (in terms of accuracy) than it was in 2021 despite being more visually impressive overall:

      Apparently by now it has nearly passed the test, kind of, as here is an image from today. It still cant make anything other than a Zam 552. I asked it to make an Olympia millennium and it just made a 552 with augers on the front for some reason:

      If you just want it to make a Zam 552 it does a pretty decent job, which Ill give it credit for:

      • aanes_appreciator [he/him, comrade/them]@hexbear.netEnglish
        7·
        2 天前

        Another good one is to check the perspective lines. While Generative models are quite good at making perspective “real enough” to pass at a glance, but they struggle immensely with perfect perspective. Plus, its a good trick to use if you dont know the subject - I dont know what a Zamboni looks like so I could be fooled by something you’d catch.

        • ToastedRavioli@midwest.socialEnglish
          4·
          2 天前

          It actually is pretty spot on with the machine itself, unless you ask it to have the dump tank up or basically anything other than “make an image of a Zamboni”. Here is a real photo of a 552 for reference:

          It still has some wonk. Like its missing the brush on the board brush. And there is some missing/extra stuff on the top towards the operator. But in terms of broad strokes it passes the test I would say

      • invalidusernamelol [he/him]@hexbear.netEnglish
        4·
        2 天前

        Those last 2 look better at first glance, but the more you pay attention the worse it gets.

        The jumbotron showing an active game and score (that is doubled!) is funny.

  • edinbruh@feddit.itEnglish
    16·
    2 天前

    How does this happen? How does someone who sees bicycles at least once a month fail to draw one?

    • KuroXppi [they/them]@hexbear.netEnglish
      13·
      2 天前

      I’m not surprised? It’s like when in drawing faces you’re taught to place the eyes at the middle point of the head, not three-quarters up the face. Conceptually we overestimate the size of interesting facial feaures, and shrink the others, so we don’t give as much space to the forehead and hair (which occupies half of the head) causing the proportions to be thrown off. We look at faces all day but can’t consistently place them (across people) without training, recall or good observation skills.

    • 7bicycles [he/him]@hexbear.netEnglish
      13·
      2 天前

      this is easily replicable at least to unsuspecting people, try it at a party sometime it’s fun. Most people, like here, get the outline correct - wheels, handlebars, saddle are generally positioned correctly - everything else is a mess. Check this out: https://www.gianlucagimini.it/portfolio-item/velocipedia/

      C is an odd case here because it seems to include a rear derailleur, the fork, the double rear tubes but the pedals are kangaroo mode and everything else is bananas

    • NephewAlphaBravo [he/him]@hexbear.netEnglish
      14·
      2 天前

      doubt it, dementia test results are so much less accurate than this, it’s probably just a random “hey try drawing this thing you never really think about” experiment

      • Multiplexer@discuss.tchncs.deEnglish
        18·
        2 天前

        What always gets me with this kind of tests:
        This is remarkably like the kind of errors generative AI typically does when told to generate images of stuff.