• Mossy Feathers (She/They)@pawb.social
    link
    fedilink
    English
    arrow-up
    10
    ·
    4 months ago

    I had a similar thought about AI; that it’s more like imagining something than actually drawing it. When you ask a program like stable diffusion to draw something, you’re basically asking it to imagine something and then you reach inside its head to pull the image out. I think that if AI was forced to draw the “ol’ fashioned way” then it’d be both better and worse. The results would be more “correct” but the actual quality would probably be worse. It’d also take it longer to get to the same level as a professional artist.

    There are a ton of shortcuts you can take in the digital world to save time; you’re basically a god limited only by your computer’s specs. You can do extremely complex things near-instantly. This saves significantly on training time when it comes to AI. An AI forced to learn how to do art the ol’ fashioned way would take significantly longer because it can’t take the same shortcuts.

    • Redjard@lemmy.dbzer0.com
      link
      fedilink
      English
      arrow-up
      5
      ·
      4 months ago

      Yeah. You want to preserve the AI’s abilities. Hence adding the “paste imagination” feature for example. If you simply use that and finish “editing” that is current AI. Then you can quickly redo only sections from imagination until they look good, maybe with a specific prompt or other form of understanding about what needs to be done and changed there.

      We can invert our visual center, so basically we see an image, think about it, then can summon a mental version of that painting back as an image by converting the abstraction of it and change things about the abstraction until the mental image seems good. This abstraction can handle ideas like recognizing, moving, scaling, recoloring objects. It can do all we can imagine because it is literally how we interpret the world. Then we spend hours trying to paint that mental image we created using limited tools. If we could just project something the same way we see, that would probably match image-AI in the initial output but after tens, hundreds of passes you could likely within minutes create something completely impossible by any other means.