This might be a surprise to you, but real artists tend to actually enjoy the process of creating art.
Opinions on AI art.
- Thread starter Angel1984
- Start date
- Jun 10, 2017
- 11,357
- 17,258
Except that this acceleration only apply to hardware, not to software.
We can always make camera smaller, CPU more powerful, and all, but to have better AIs, we need better algorithms, and for this, it need someone to come with the right idea. It can happen tomorrow, like it can need two decades.
AIs that we have now, we could have had them 30 years ago. Not for personal use, because at this time the computing capabilities were more limited, but maxi calculators were powerful enough for searchers to obtain good result in a reasonable amount of time. Yet, it's only ~10 years ago that they appeared, just because we didn't had the right algorithms before this.
And it's why AIs will more than probably not improve the way people expect it, at least in a near future. They'll be more powerful, they'll give more realistic results, but they'll still face the same limitations, the ones due to nowadays algorithms. And I'm not really optimist on that subject, because even the best professionals don't really understand what happen in the black boxes; difficult to improve something, when you don't know what this "something" is.
And it's precisely the problem. Tools always come with a price to pay. They improve our life, but slowly lead to the disappearing of a knowledge or an expertise.
But this is just the old grumpy folk talking...
one of the things that bothers me about AI art is how it's always 'random portraits of girls looking into camera'. on that single aspect alone this POS movie with its terrible acting already beats the AI xena 'acting' a 100x.
- Jan 28, 2021
- 270
- 905
I can't really speak to this as I am far from an expert on AI. I wouldn't even consider myself an amateur hobbyist because I haven't bothered to mess with Chat GPT or any of the others for myself. But it would seem to me that the further we push the hardware the software will follow. Innovation is key sure but faster processing power will assist with that.
Yeah that's standard old people talk. There are tons of things that were important for people to know in the past that just don't matter now. I tried to learn how to drive a stick shift when I was a kid but then I said fuck it and never looked back. I'm not even sure how many manual transmission cars are left these days. Probably a few. But not nearly as many as there used to be.
See, I don't see that as an AI problem but rather a Directorial problem. The people using AI to do this stuff aren't film majors. And if an actual director tried their hand at utilizing AI, directing it would be different than directing a human actor. What we need is AI specialists that know how to train their AI models and direct them in such a way that would make for good cinema.
To recap: AI is a tool that is rapidly evolving and we don't quite have the specialist craftsmen trained to use it properly yet. I predict we'll have our first "decent" (read: not absolute garbage) AI feature length film in the next 10 years. It will probably go straight to a streaming service though. Not released to theaters.
Last edited:
- Nov 2, 2017
- 1,509
- 1,695
I have no idea how I got here but for now AI generated images used, especially in porn games are very easy to recognize for the time being imo.
They have a certain shading to it and facial expressions that need to be ironed out. I don't care if it has soul or not.
I just don't like to look at it.
Look at taffy tales and the nose dive of the artstyle. Sure Uberpie is training his own artstyle into the AI but he can't get rid of that shading I mentioned.
Anyway....
They have a certain shading to it and facial expressions that need to be ironed out. I don't care if it has soul or not.
I just don't like to look at it.
Look at taffy tales and the nose dive of the artstyle. Sure Uberpie is training his own artstyle into the AI but he can't get rid of that shading I mentioned.
Anyway....
The europeans would like to have a word with you.
- Jan 28, 2021
- 270
- 905
Pfft. Do those tiny European cars even count as cars? Just kidding. I wish the roads here weren't so overrun with excessively large vehicles. I swear most people buy them just so they get better visibility to see over other cars. But that just results in more and more oversized cars.The europeans would like to have a word with you.
At this point your typical American SUV could run over a toddler and not even notice.Which is why I think devs who want to use AI Art should be prepared to heavily edit the output with their own digital art editing skills. Get good enough at it and people won't notice nearly as much.
Can you give some examples of Taffy tale before and after AI? Never played the game and preview pics are all from his old art, I think.I have no idea how I got here but for now AI generated images used, especially in porn games are very easy to recognize for the time being imo.
They have a certain shading to it and facial expressions that need to be ironed out. I don't care if it has soul or not.
I just don't like to look at it.
Look at taffy tales and the nose dive of the artstyle. Sure Uberpie is training his own artstyle into the AI but he can't get rid of that shading I mentioned.
Anyway....
The europeans would like to have a word with you.
Partly. The general idea of NNs and many of the algorithms that are nowadays used to train NNs like "gradient descent" and "backpropagation" were already known in the last century. And it was also known or at least anticipated, that NNs can be used for tasks like pattern recognition and language/image generation. But for letting them learn something worthwile and do it in decent time takes lots of data, memory, fast processors and if I understand it right, the processor architecture is also very important. So the current progress we see was more the result of the availability of the right hardware and digital data than about the availability of new algorithms. There are still algorithmic innovations for Neural Networks like using "GANs" for image processing and "Transformers" in natural language processing, but the importance of the hardware-power and structure is definitely something that is not to underestimate.
There might still be currently much space for developments. My own idea would be to adapt the image-generation technology to 3D-spaces to have a simple possiblity to create better 3D-models and move them in the space instead of creating new images each time, but I don't know, if this would actually work.
Last edited:
- Jan 28, 2021
- 270
- 905
This last bit here made me think about what an AI Art generator could do with 3D modeling and an integrated 3D printer. But then I remembered all my favorite dystopian AI Sci-fi movies and the existential dread kicked in.
"Open the pod bay doors, HAL."This last bit here made me think about what an AI Art generator could do with 3D modeling and an integrated 3D printer. But then I remembered all my favorite dystopian AI Sci-fi movies and the existential dread kicked in.
- Jul 25, 2020
- 214
- 345
Still... If everyone is capable of doing "art" in seconds with just one phrase is it really art?. Maybe it's me being old but i cannot enjoy AI "art" ... it just feels so shallow. It's sort of downgrading... I could not play AI generated games.. i could not watch AI generated movies.
Regarding the progress aspect.. I heard that AI is somewhat a poznie sheme... like a bubble ready to burst i i would make a prediction that we'll kind of forget about AI in few years. Guess we'll see.
Regarding the progress aspect.. I heard that AI is somewhat a poznie sheme... like a bubble ready to burst i i would make a prediction that we'll kind of forget about AI in few years. Guess we'll see.
- Jan 28, 2021
- 270
- 905
It's best not to put "Art" into a conceptual box. Anything can be art. And, from a certain point of view, Hollywood has already been doing something functionally similar to AI when it comes to certain genres.
We call tv shows like Law and Order "procedural cop dramas" because the structure of every episode is formulaic. The episodes practically write themselves. The main cast is a collection of characters that all adhere to certain tropes fundamental to the genre, each episode features a new "villain of the week", and the plot progresses from the crime, to the investigation, to the prosecution, with good triumphant over evil wrapping up everything with a big red bow on top.
Obviously shows like that have writers but long term shows that go through different writers commonly have a blueprint or "bible" for the writers to follow. A set of instructions to help ensure the show maintains its core vision. Not too dissimilar from how we give AI instructions to ensure it spits out the images we want.
Eh, are there people out there who are just trying to ride the AI Wave/Bubble for financial gain? Sure. There's gold in them hills and it's a rush to grab the market. Sort of like the whole bitcoin/nft craze there's a lot of stupid/scammy shit going on but the technology does have it's uses. But this whole "use AI to write better work e-mails" fad is dystopian as fuck.
- Jun 10, 2017
- 11,357
- 17,258
There's a really good reason for this.
To simplify, AIs learn from what is the best sources. So, there's movies for the way human bodies moves, photography for the global anatomy of a human body, and portraits for the details of a human face.
Portraits... That particular kind of photography where people will be face to the camera, looking straight in front of them.
And of course, even photography that aren't portraits still depict humans looking more or less in front on them. Your friend take a photo of you, you'll look at the camera. A group photo from a family dinner, people will look at the camera. Holidays photos, still looking at the camera. Browse the photos you have, at least half of them have the person(s) you wanted to photography look at the camera.
So, while they know how to represent human heads and faces from all the angles when it's necessary, what AIs learned is that humans look at the camera. And like they learned how to represent details from portraits, the eyes will also looks straight in front of them. There's what, something like ~75% of the photos they learned from that represent humans that way, what are they, simple AIs, to think that it's not how humans should be represented ?
It's one of the limitations I was pointing previously. Like they know, but don't understand, they aren't in a position to get rid of that posture for the faces. They can, if prompted otherwise, but when it's them that have to decide, it's what they'll use.
It don't need to be an expert. Having some knowledge about codding is enough to know that it don't works that way.
Take FPS by example.
Late 1991, id Software released
You must be registered to see the links
. The concept was basic, take pixels, and give them a third dimension as well as a texture, you then get a cube that make it looks like you were immersed in a labyrinth. You can see it if you play the game, or its few brothers, all face of the cube have the exact same texture. Before this, all games in 3D, even from id Software, used wire frame, therefore lines representing the borders of any surface. Then, two years later, the same guys released the first
You must be registered to see the links
, for the same computers...Why did they pass by that intermediary hybrid, while it was possible to already have something way better, that use raytrace-like mechanism to represent the world in third dimension ? They knew how to represent more realistic 3D environment since there were the 3D wire frame games, and raytracing was already something a those times. But the fact is that, to represent what was just an instant in DOOM, and to do it on domestic computers, raytracing algorithms needed something like a full hour.
And it's the reason for that intermediary step. Not that it was needed to learn how to do DOOM, but it was all they could do at this time, by lack of algorithms that permit them to do better.
It even goes further, because what Wolfenstein 3D was doing don't needed all the power of those times domestic computers. Therefore it would have been possible to do it before. The visual would have been more limited, but computers like the Amstrad CPC or the Commodore C64 could have done the same, just with plain colored faces due to their RAM limitations. And of course, arcade cabinet could have done it too, this time with fully textured faces.
Both could, but haven't. Not because it wasn't technically possible, or because the public wasn't there. Imagine an arcade game with a 3D labyrinth... teenager me would have lost a fortune playing this. But like no one came with the algorithms for this, no one did it, because no one knew how to do it.
And it's the same for AIs. Especially since they rely on black boxes, meaning that people who develop them know "more or less" what is happening;
You must be registered to see the links
. On this, AIs don't really differ from human brains, we have a good knowledge regarding how our brain works technically (neurons, synapse, etc), and we more or less know what part of the brain is responsible for what. But how our brain pass to "I want to express my thoughts regarding AIs" to typing lots of words that have a meaning and are actually the said thoughts, it's a big mystery.And, yes, it's the same for AIs. Why an AI think that "this image" is an orange ? Because it's what it think... There's no switch telling it that it's one, there's no decision tree starting by "it's round shaped" and ending by "it's an orange". It's a magic process that works but no one can really explain why exempt than by saying "it's what the AI learned".
What, for a coder is both marvelous and terrifying. We can design algorithms that will give the correct result, but we have globally no idea how they really works...
... Wait... It's an estimate, but in Europe they probably still represent between 70% and 80% of the cars...
You must be registered to see the links
. So, if just five years ago you had to travel to Europe and rent a car, your chance for one with an automatic transmission to be available would have been low, and you would have been doomed.It's the problem with lost knowledge, they doesn't matters... as long as things goes well. But things don't always goes well. Take flooding, big fires, and all natural disasters. Your region is hit by one, there's power outage for two days... And if you only know how to put meal in the microwave, for two days you'll eat what ? Cereals ?
And it goes like this for everything and every fields. The day you can't anymore count on the technology, whatever the reason, you need to go see one of the elder, because he learned before that technology appeared and know how to do without it. But when all the elders will be retired, or dead, there will be no one to remember how to do without technology, and you'll just tell everyone that "it can't be done". Not because it really can't be done, but because no one know how to do it.
I remember, it was something like 10 years ago, I was at a convention, and a child, around 10-12yo, asked what was probably an artist he was a real fan, a drawing with his autograph. The guy answered that he was sorry, but he only know how to draw on a computer. I don't remember how the kid reacted, but personally I was amazed, because the guy use a tactile tablet to draw... and not a second it crossed his mind that if he can use a fake pencil over a tactile surface, he can also use a real pencil over a piece of paper.
Like ray tracing exist since 1968, yet it needed around 15 years before it starts to be used for creation, 25 years before it starts to be used in games, and like nowadays algorithms have nothing in common with the first one.
Everything isn't just a question of raw computing power, but also a question of efficiency; something that I never cease to repeat to junior that interne at works. When, around mid 90's, John Carmack had the idea for his
You must be registered to see the links
, he didn't invented the wheel. The algorithm was known since (from memory) two decades. But no one before him found, or at least used/released an implementation that was efficient. And by "efficient" I don't mean that it need less computing power, but that it need less convolutions.It's like for drawing a line on a computer. Bresenham's algorithm date from 1962 and is still the base algorithm used by everyone for this. But nowadays implementation use tricks to make it more efficient, and therefore smaller, as well as less prone to bugs.
There's no interest in implementing what McCulloch and Pitts theorized in A logical calculus of the ideas immanent in nervous activity (1943) then What the frog’s eye tells the frog’s brain (late 50's), that are the base of AI researches, if you aren't sure that the result will be correct. And if one look at the draft they theorized, they would more than surely wonder how it led to nowadays algorithm. There's a logic of course, but it was just the base, and it's the lack of efficiency, more than the lack of power, that prevented to goes further right from those times.
this is one of the problems with AI as the whole point of it is that it's supposed to be used by complete noobs or with no people at all. experts using AI defeats its purpose.
not true. it's been shown already that increasing the generalization of LLMs takes exponentially more both computing power but more importantly training data, and training chatgpt already costs $700,000 a day. getting the AI work 1% better is taking up to 100x more resources. making AI much better than the current horse shit level of getting it wrong 40% of the time is getting prohibitively expensive.
but it's not even the worst problem. turns out that the training requires fresh untainted non-AI data, and as a million AIs are currently polluting internet with gargantuan amounts of AI slop, the amount of usable data is getting LESS not more.
then of course there's the copyright issue which will kill the current crop of AI companies the second legislation catches up with it. which is probably withing 2 years (based on how long it took with crypto).
AI images are okay as reference tools or as a starter for actual creations, but calling them "original images" is either Simone Biles-level mental gymnastics or just plain ignorance. I think by now we all understand how AI images are spawned and so forth: an aggregator vacuuming up original art and photographs from all over the Internet, melting them into a cauldon of pixel vomit, then spewing out whatever it considers to be whatever the prompt asked for.
All AI art looks like oil paintings to me. 95% of them have the EXACT same lighting (coming in from the top right). A majority of these AI image processors have no idea how may fingers or toes humans are supposed to have. And if the picture is supposed to have words or numbers, they are indecipherable...and I mean a combination of Gallifreyan, Cyrillic, English, and cursive WingDings.
Long story short, AI images are not "original" works of art. And all AI images should automatically be considered under public domain. So don't claim it's an original piece and get mad if someone uses is for their own purposes. After all, the art that was used to create it in the first place was stolen.
All AI art looks like oil paintings to me. 95% of them have the EXACT same lighting (coming in from the top right). A majority of these AI image processors have no idea how may fingers or toes humans are supposed to have. And if the picture is supposed to have words or numbers, they are indecipherable...and I mean a combination of Gallifreyan, Cyrillic, English, and cursive WingDings.
Long story short, AI images are not "original" works of art. And all AI images should automatically be considered under public domain. So don't claim it's an original piece and get mad if someone uses is for their own purposes. After all, the art that was used to create it in the first place was stolen.
I think AI images are nice for the games would generally use badly picked Real Porn videos and images from many sources with no consistency. At least this way they can have more consistent models mostly complete even if proofing the images is required.