Tyler Perry Puts $800M Studio Expansion On Hold After Seeing OpenAI’s Sora: “Jobs Are Going to Be Lost”::Tyler Perry is raising the alarm about the impact of OpenAI’s Sora on Hollywood.

  • kescusay@lemmy.world
    link
    fedilink
    English
    arrow-up
    34
    arrow-down
    2
    ·
    11 months ago

    Yep. I watched their demo clips, and the “good” ones are full of errors, have lots of thematically incoherent content, and - this is the biggie - can’t be fixed.

    Say you’re a 3D animator and build an animation with thousands of different assets and individual, alterable elements. Your editor comes to you and says, “This furry guy over here is looking in the wrong direction, he should be looking at the kangaroo king over there, but it looks like he’s just glaring at his own hand.”

    So you just fix it. You go in, tweak the furry guy’s animation, and now he’s looking in the right direction.

    Now say you made that animation with Sora. You have no manipulatable assets, just a set of generated frames that made the furry guy look in the wrong direction.

    So you fire up Sora and try to fine-tune its instructions, and it generates a completely new animation that shares none of the elements of the previous one, and has all sorts of new, similarly unfixable errors.

    If I use an AI assistant while coding, I can correct its coding errors. But you can’t just “correct” frames of video it has created. If you try, you’re looking at painstakingly hand-painting every frame where there’s an error. You’ll spend more time trying to fix an AI-generated animation that’s 90% good and 10% wrong than you will just doing the animation with 3D assets from scratch.

    • Buelldozer@lemmy.today
      link
      fedilink
      English
      arrow-up
      10
      arrow-down
      1
      ·
      edit-2
      11 months ago

      Now say you made that animation with Sora. You have no manipulatable assets, just a set of generated frames that made the furry guy look in the wrong direction.

      “Sora, regenerate $Scene153 with $Character looking at $OtherCharacter. Same Style.”

      Or “Sora, regenerate $Scene153 from time mark X to time mark Y with $Character looking at $OtherCharcter. Same Style”.

      It’s a new model, you won’t work with frames anymore you’ll work with scenes and when the tools get a bit smarter you’ll be working with scene layers.

      “Sora, regenerate $Scene153 with $Character in Layer1 looking at $OtherCharacter in Layer2. Same Style, both layers.”

      I give it 36 months or less before that’s the norm.

      • Sprucie@feddit.uk
        link
        fedilink
        English
        arrow-up
        6
        ·
        11 months ago

        I agree, I don’t think people realise how early into this tech we are at the moment. There are going to be huge leaps over the next few years.

      • conciselyverbose
        link
        fedilink
        English
        arrow-up
        2
        ·
        11 months ago

        Or just “take the frame and replace the head with the same face pointed a different way”.

      • kescusay@lemmy.world
        link
        fedilink
        English
        arrow-up
        2
        ·
        11 months ago

        This seems like a fundamental misunderstanding of how generative AI works. To accomplish what you’re describing you’d need:

        • An instance of generative AI running for each asset.
        • An enclosing instance of generative AI running for each scene.
        • A means for each AI instance to discard its own model and recreate exactly the same asset, tweaked in precisely the manner requested, but immediately being able to reincorporate the model for subsequent generation.
        • A coordinating AI instance to keep it all working together, performing actions such as mediating asset collisions.

        The whole system would need to be able to rewind to specific trouble spots, correct them, and still generate everything that comes after unchanged. We’re talking orders of magnitude more complexity and difficulty.

        And in the meantime, artists creating 3D assets the regular way would suddenly look a lot less expensive and a lot less difficult.

        If all you have is a hammer, everything looks like a nail. Right now, generative AI is everyone’s really attractive hammer. But I don’t see it working here in 36 months. Or 48. Or even 60.

        The first 90% is easy. The last 10% is really fucking hard.

    • mindbleach
      link
      fedilink
      English
      arrow-up
      1
      ·
      edit-2
      11 months ago

      But you can’t just “correct” frames of video it has created.

      Yeah you can.

      Same way you can correct parts of a generated image, and have the generator go back and smooth it over again. Denoiser-based networks identify parts that don’t look right and nudge them toward the expectations of the model. Sora clearly has decent expectations for how things look and move. I would bet anything that pasting a static image of a guy’s head, facing the desired direction, will result in an equally-plausible shot with that guy facing the right way.

      There have been image-generator demos where elements can be moved in real-time. Yeah, it has wider effects on the whole image, because this technology is a pile of hacks - but it’s not gonna turn red wallpaper into a green forest, or shift the whole camera angle. You’re just negotiating with a model that expects, if this guy’s facing this way now, his hands must go over here. Goofy? Yes. Ruinous? Nope.

      And at the end of the day you can still have human artists modify the output, as surely as they can modify actual film of actual people. That process is not quick or cheap. But if your video was spat out by a robot, requiring no actors, sets, or animators, manual visual effects might be your entire budget.

      Really - the studios that do paint-overs for Hollywood could be the first to make this tech work. They’d only need a few extra people to start from the first 90% of a movie.