>>1925419
You're severely over-estimating what AI actually is. Shit's cool I agree with that, but the level of complexity of something like just text or matching text with images is MAGNITUDES easier than trying to re-create the creative process of taking even a handful of drawn 2D images and having it spit out a halfway decent 3D model.
The AI dungeon example, only text, an extremely limited and strict set of parameters (A linear language) that you have literally 1000's of years worth of training data to use. Again, shit's impressive, but it's not complicated in comparison to wanting a 3D model. Take the word you're on, compare against the words that came before, maybe keep certain key nouns in memory as reference for what to type, type new words that frequently come after the word that comes before. Not to dismiss the complexity, but as the current running theme, FAR, FAR easier than creating a model from scratch.
The text to image example, again, not that mind bending when you break it down. Feed it a shit ton of images that all have a variety of key words attached to each image, let it figure out the elements in the 2D array that are consistent between any given set of key words, then when it's asked for a certain key word it spits out an amalgamation of what it observed between the photos with that key word. This isn't too bad because we're sticking to a 2D pixel array and comparing to a linear text string. Your
>3d and even light simulation without even training for it
doesn't hold water, it's not dealing with a 3D model or doing RTX shit, it's still purely some form of 2D image it's dealing with. Which if it gets fed data that has consistent lighting and dimensionality parameters, it's going to fart out generated images with the same parameters.
Even the example you linked with the video is even more limited in that you can't just type out a text string and have the AI figure out the trillions upon trillions of ways you could ask it for a new 2D pixel array, it has to be in a mad-libs style limited text prompt. That much I could easily see being an extension of what is there now in the semi-near future. 3D models? Not so much, especially with how current systems work.
Why? A combination of literally adding another axis makes it a magnitude more complex to compute (The example of spitting out human 3D models you could argue is probably only 2.5D + 2.5D as it's figuring out a 2D array, figuring out approx. what Z height any given point in the image has, then takes what it knows the back of humans look like, generating a new back from what it gets fed from the front, then runs the Z depth process again, combining the two results on top of each other.), and there's no real significant training data. Hell, trying to AI generate voice, which given the extremely limited nature of JUST a waveform would in theory be easy as shit, is still very much in the "If you want to generate a voice, you need hours upon hours of recorded audio to let it figure out how to do it." stage. Hence why most examples are of famous characters from long running series and/or video game characters that you can easily get a shit ton of data for. (Outside of an example currently that doesn't require AS much training data by way of adapting an existing training model to a smaller snippet of voice, but is a shitty pay-for thing? Been a little bit since I last looked into it.)
How are you planning on training this AI model to create 3D models? 2D art CONSTANTLY breaks rules of how something should be constructed in 3D, in ways that not even humans could reliably predict. Which is why 3D artists, when asked to translate 2D characters, will never have the exact same results between them the farther away from perfect a design adheres to 3D space (and all 2D art by nature will at least a tiny bit), it's a creative interpretation. In the truest sense of the word. You won't get predictable results an AI could latch onto, because how one person interprets the linework won't be exactly the same as how someone else will. Hell, the same 3D artist will interpret linework differently between different sessions. Hell, even on top of THAT, even if you have a perfectly proportioned series of drawn front/back/side turntable views, you're STILL going to run into issues where the exact 3D shape of an object is going to be up to artistic interpretation.
I don't want you to read the above block of text and go
>Well he's just a nay-sayer, AI can do anything!
No, that's not what's happening. What AI can do as I've already said is cool as shit, but in purely practical terms, it's just taking shit loads of input with parameters, you ask for a specific parameter, and it generates something based on what it 'knows' for that parameter. You can rightly argue that's literally what the human consciousness is in a nutshell, but the difference is, AI as it is now is in an ameba state comparatively speaking. Will it EVENTUALLY get to the point where you can feed an AI 2D character art and spit out SOMETHING? Yes. But the level of complexity of where we are now compared to the level of complexity that would be is, say it with me, MAGNITUDES higher. To the point we're scraping some form of silicone consciousness instead of the braindead "Compare pixels/character strings" we have now. So not for quite some time. Might be sooner than later who knows, but AI as a whole would have to make metric leaps before we get "2D art, make 3D model."
|
