AI does one thing. It takes an equation like y=f(x) and a bunch of examples of x and y and gives you a guess as to what f might be. This f doesn't know the actual value of y for the value of x you're giving it, but it can guess because there were patterns to the data it was trained on and your x value should generally fall somewhere within the range of the input data set. If I train a neural network on the x and y values from something like y=x^2, it's going to give me something that resembles y=x^2. It won't be exactly that, but if I'm willing to accept a slightly wrong answer, it's good enough. There are infinite possible values for x and almost as many for y, so the neural network will eventually learn to give an answer that's good enough. For a real world example, if I train a neural network that x is what a raytraced render looks like after 100 samples and y is what it looks like after 1000 samples, the neural network can "learn" patterns about where light tends to fall in scenes built by humans, and where raytracing tends to need more or fewer samples to get it right. You can assemble a huge library of different kinds of scenes and there are an infinite number of possible camera positions for each, so again you've got a lot of data. And as raytracing works essentially the same way no matter what the scene is, the x value the neural network will be fed after training is done is going to resemble the x values it was trained with. Now let's look at what would be required to generate a 3D model. When you say "a baby wrapped in a blue blanket", that means something very specific. Of all the words in there, the only one that you could change without changing the entire image is "blue". So if a neural network "knows" how to approximate a blue object and "knows" how to approximate a baby wrapped in a [color] blanket, it can generate something reasonable. But if you change "baby", "wrapped", or "blanket" to a different word, the mesh is completely different. Knowing what a car wrapped in a blue blanket looks like doesn't help you here, nor does knowing what a baby holding a blue blanket looks like, nor does knowing what a baby wrapped in a blue garden hose looks like. Let's pretend the AI is actually intelligent and it can guess that the baby's head should visible somewhere in the model. The rest of the model has a very specific connection location (it has to be on the neck side of the baby's head - you wouldn't look at a decapitated Marge Simpson and say "that's the right image") and a very specific shape (knowing what pretty much any other object wrapped in a blue blanket looks like or a baby wrapped in any other blue object wouldn't help you at all here). So to circle back to the original point, AI can figure out an approximate answer to an objective question if you have a lot of examples of correct answers to similar questions. And unless the neural network was trained on a lot of models of babies wrapped in different colored blankets (and at that point, you'd just buy one of those models) it's going to have no idea what you're asking it for. AI can only help you if you're asking it a question it already basically knows the answer to. It's a fuzzy search engine, but for fuzzy results in addition to fuzzy queries. If it produces a model that looks remotely like the thing you're asking it to make a model of, that's because someone gave it that answer ahead of time.
1:43 - I can't believe these people listed PBR as an """Art Style""". I didn't need more proof that the venture capitalists/tech bros jumping on-board the AI hype train don't understand the medium they're trying to automate, but there's some more.
I don't believe that AI-generated 3D art will ever be a true 'one click solution.' It's the same problem with photogrammetry, in that it MAY eventually be able to actually produce a 'pretty' looking model that's accurate to the prompt, but will be essentially unusable in any real-world context without massive amounts of manual labor added in after the fact. Terrible topology, weird texture seams, weird artifacts, etc. At the end of the day, it will always be better to have a competent real human artist at the helm than AI.
Wow. That is so much worse than anything I could imagine. I thought it would at least give you a baby model with terrible topology but it just completely failed in every aspect.
The problem is that while AI is capable of very impressive things, it is also very naive, and it doesn't "know" anything. It has to be SO thoroughly trained in order to generate anything coherent.
Currently it cannot make anything new or independently assess the quality of its work. I don't know of any reason to believe this will ever change. There are likely to be solid improvements in its model creation capabilities but I seriously doubt you will get anything game ready or close to human artists this decade, and being able to reliably request changes from it like you would request changes from a freelancer is extremely unlikely.
Chet, I don't know what website you're using, and when you say "AI", there's so many different artificial intelligence systems. Whatever junk you found online isn't any good. All of these popular websites that you found is a cash grab, they cobbled together a bunch of trash and stole your money, you got conned. My local AI, my own personal system, that uses my own personally trained database, generates very high quality 3D models. Also, using "baby" is pretty frowned upon, because generating child abuse material. Most AI's have that data removed and erased. This reminds me of when you were working for Valve, being negative about Virtual Reality. It seems to be a copy of that same scenario.
If it gets better and usable in the next 12-24 months, would you use it in your workflow? Shortly enough you will be able to make a video generation of a camera rotating on anything and these 3d generations will get massively better. Saying it’s a scam is massively undervaluing what’s happening, and I don’t have a newsletter and just have a usual job. I understand your point of view and where it comes from, I just ask you to make another video in about 12 months to compare. Will be nice to see! Been following you for many, many years. Thank you for everything you’ve done.
What makes you think it's going to get "better" such that it's "usable" in 24 months? What is your definition of "usable"? It can generate a 4k video for more than 10 seconds? Or it can design a navy ship for you from scratch? One of those use-cases is basically not useful to anyone outside of some extremely low-hanging prototyping, the other is not possible without some radical breakthroughs that are nowhere on the horizon with our current understanding of AI.
2:52 "I don't know what 'refine' means here"
Amungus Baby
I was thinking the same thing
🎵My Baby Blue🎵
AI does one thing. It takes an equation like y=f(x) and a bunch of examples of x and y and gives you a guess as to what f might be.
This f doesn't know the actual value of y for the value of x you're giving it, but it can guess because there were patterns to the data it was trained on and your x value should generally fall somewhere within the range of the input data set. If I train a neural network on the x and y values from something like y=x^2, it's going to give me something that resembles y=x^2. It won't be exactly that, but if I'm willing to accept a slightly wrong answer, it's good enough. There are infinite possible values for x and almost as many for y, so the neural network will eventually learn to give an answer that's good enough.
For a real world example, if I train a neural network that x is what a raytraced render looks like after 100 samples and y is what it looks like after 1000 samples, the neural network can "learn" patterns about where light tends to fall in scenes built by humans, and where raytracing tends to need more or fewer samples to get it right. You can assemble a huge library of different kinds of scenes and there are an infinite number of possible camera positions for each, so again you've got a lot of data. And as raytracing works essentially the same way no matter what the scene is, the x value the neural network will be fed after training is done is going to resemble the x values it was trained with.
Now let's look at what would be required to generate a 3D model.
When you say "a baby wrapped in a blue blanket", that means something very specific. Of all the words in there, the only one that you could change without changing the entire image is "blue". So if a neural network "knows" how to approximate a blue object and "knows" how to approximate a baby wrapped in a [color] blanket, it can generate something reasonable.
But if you change "baby", "wrapped", or "blanket" to a different word, the mesh is completely different. Knowing what a car wrapped in a blue blanket looks like doesn't help you here, nor does knowing what a baby holding a blue blanket looks like, nor does knowing what a baby wrapped in a blue garden hose looks like.
Let's pretend the AI is actually intelligent and it can guess that the baby's head should visible somewhere in the model. The rest of the model has a very specific connection location (it has to be on the neck side of the baby's head - you wouldn't look at a decapitated Marge Simpson and say "that's the right image") and a very specific shape (knowing what pretty much any other object wrapped in a blue blanket looks like or a baby wrapped in any other blue object wouldn't help you at all here).
So to circle back to the original point, AI can figure out an approximate answer to an objective question if you have a lot of examples of correct answers to similar questions. And unless the neural network was trained on a lot of models of babies wrapped in different colored blankets (and at that point, you'd just buy one of those models) it's going to have no idea what you're asking it for.
AI can only help you if you're asking it a question it already basically knows the answer to. It's a fuzzy search engine, but for fuzzy results in addition to fuzzy queries. If it produces a model that looks remotely like the thing you're asking it to make a model of, that's because someone gave it that answer ahead of time.
2:30 Dunno, that looks like a very realistic Kirby to me
congratulations on your baby lamp.
idk what you're talking about. these are beautiful baby trash fires
1:54 Hideo Kojima must have used this on Death Stranding
Seeing what LLMs are capable of, I thought maybe AI of any kind would be amazing. I stand corrected XD. Great video as always Chet!
1:43 - I can't believe these people listed PBR as an """Art Style""".
I didn't need more proof that the venture capitalists/tech bros jumping on-board the AI hype train don't understand the medium they're trying to automate, but there's some more.
I don't believe that AI-generated 3D art will ever be a true 'one click solution.' It's the same problem with photogrammetry, in that it MAY eventually be able to actually produce a 'pretty' looking model that's accurate to the prompt, but will be essentially unusable in any real-world context without massive amounts of manual labor added in after the fact. Terrible topology, weird texture seams, weird artifacts, etc. At the end of the day, it will always be better to have a competent real human artist at the helm than AI.
It's fine, We love Lamp
boutta ruin this,
2:21 among us baby
you beat me to it
you didn't specify which trimester, chet
Ai is really cool if you dont have to actually use it!
Wow. That is so much worse than anything I could imagine. I thought it would at least give you a baby model with terrible topology but it just completely failed in every aspect.
Lol
2:20 AMOGUS 😮
I wonder what exactly are the limits of AI, is it flawed in a way that will always produce slop, no matter what you feed it?
The problem is that while AI is capable of very impressive things, it is also very naive, and it doesn't "know" anything.
It has to be SO thoroughly trained in order to generate anything coherent.
Currently it cannot make anything new or independently assess the quality of its work. I don't know of any reason to believe this will ever change. There are likely to be solid improvements in its model creation capabilities but I seriously doubt you will get anything game ready or close to human artists this decade, and being able to reliably request changes from it like you would request changes from a freelancer is extremely unlikely.
Chet, I don't know what website you're using, and when you say "AI", there's so many different artificial intelligence systems. Whatever junk you found online isn't any good.
All of these popular websites that you found is a cash grab, they cobbled together a bunch of trash and stole your money, you got conned.
My local AI, my own personal system, that uses my own personally trained database, generates very high quality 3D models.
Also, using "baby" is pretty frowned upon, because generating child abuse material. Most AI's have that data removed and erased.
This reminds me of when you were working for Valve, being negative about Virtual Reality. It seems to be a copy of that same scenario.
If it gets better and usable in the next 12-24 months, would you use it in your workflow?
Shortly enough you will be able to make a video generation of a camera rotating on anything and these 3d generations will get massively better.
Saying it’s a scam is massively undervaluing what’s happening, and I don’t have a newsletter and just have a usual job. I understand your point of view and where it comes from, I just ask you to make another video in about 12 months to compare. Will be nice to see!
Been following you for many, many years. Thank you for everything you’ve done.
What makes you think it's going to get "better" such that it's "usable" in 24 months? What is your definition of "usable"? It can generate a 4k video for more than 10 seconds? Or it can design a navy ship for you from scratch?
One of those use-cases is basically not useful to anyone outside of some extremely low-hanging prototyping, the other is not possible without some radical breakthroughs that are nowhere on the horizon with our current understanding of AI.
This is quite unlikely to happen.
It won’t and no.
@@niamhleeson3522 let’s see! We can come back here in 24 months and see if something actually worked.
@@chetfaliszek I genuinely hope you’re right, because I’m pretty worried about the next decade. Thanks for replying!