I Love When AI Go All-Conquering

Seriously, sometimes we humans and our AI friends completely fail to get on the same page.

As an experiment, I asked Midjourney to “clean up and re-render this logo $(link)” and instead it came up with, I guess, a messed up logo for a not very successful cleaning service:

Midjourney AI and mjr: “clean up and re-render this logo”

Midjourney AI and mjr: “clean up and re-render this logo”

The last few weeks I have been stuck into a home renovation project with some local folks – tearing out walls and old plumbing and hauling things to the dumpster. Did the AI somehow read that out of my subconscious?


  1. Tethys says

    Huh, maybe it’s using a different definition of re-render?
    It appears that the roof has fallen into both rooms, though the plant in #2 is unscathed. Perhaps the levitating sign has rendered some new skylight’s?

    Is there supposed to be an original image at the link?

  2. sonofrojblake says

    While the original (apparently now mostly solved) problem with hands was interesting, I’m more fascinated by AI’s blind spot when it comes to text.

    There’s an app on my phone that allows me to point a camera at a sign that’s written in German or Japanese, and the phone will render the sign in English text LIVE ON THE SCREEN WHILE I WATCH. On some level, this app “understands” that the patterns it’s seeing are words, formed perhaps into sentences, and is able to produce (close to) the equivalent sentence in a different language and display it appropriately on the live image of the sign/book/whatever. And it’s been able to do this for YEARS now.

    By contrast, our new AI overlords (at least, the image-generating variety), don’t seem to have the concept of sentences. They don’t even, apparently, have the concept of words, beyond the idea that letters usually happen in strings and often the first one is an uppercase variant. It seems weird to me that there’s not some subroutine or other that looks at the almost-complete image and goes “hang on, that bit of the image is letters – I’ll just make it say something vaguely coherent”, especially given that the other headline grabbing AI tech that’s conquered the world in the last year or so is Turing-test-passing text generation. It seems a strange disconnect.

  3. says

    @sonofrojblake: Apparently new models handle hands, and Midjourney is actually developing a text model specifically to make signs and dialogue. Apparently there is a very active AI comic scene.

  4. sonofrojblake says

    I didn’t think the gap would last long. It’s an interesting time to be alive, watching this stuff morph in front of us.

    Speaking of things morphing in front of you and AI art – have you SEEN the opening titles to the latest MCU TV show? Animated AI art specifically to give it an otherworldly feel, and it’s beautiful. Check it out.

Leave a Reply