this post was submitted on 15 May 2026
15 points (89.5% liked)
Funny
322 readers
71 users here now
Funniest content on all Lemmygrad
founded 2 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
I think you might be misunderstanding how image generation models like Stable Diffusion actually work here. They encode a compressed statistical model of the patterns that exist across billions of images the training data was broken down into. When you give Stable Diffusion a text prompt it doesn't go find a relevant image and regurgitate it. What actually happens is that it generates an image pixel by pixel based on the probability distribution it learned during training through a process of denoising in a latent space.
And that process is not actually all that different from the way humans imagine or combine visual ideas. We absorb visual patterns from countless examples over our lifetimes and internalize them. When you draw a picture of a cat you do not pull up a specific cat image you memorized from a photo to copy. Drawing is a process of generating novel combinations based on general patterns we associate with the subject.
Similarly the weights of a model encode statistical regularities about pixel co occurrence, shapes, textures, object relationships, and higher level visual concepts. When it generates an image it is combining those patterns in ways that were never present in the training data. That is how Stable Diffusion produces entirely novel images that have never been seen before and why it can be prompted to combine concepts from different domains in creative ways.
This doesn't really change the argument, and for your information I do know how so called "ai" works.
In fact I'd argue that's the crux of my point.
What do these probability functions do? They create emphasis. This, of course, is not all of what art is. There is no one of what art is, by its nature. However, emphasis is a large part of what makes a style unique.
Let's say I were to ask an ai for a soviet style propaganda poster. What would I get? I would, most likely, a poster with harsh contrasting colours with an emphasis on deep or bright reds and yellows. Ignore the content, it can be anything. I could make one with Joe Biden or Donald Trump or Augusto Pinochet or Netanahyu, but that's simple enough. It is probabilistically determining that this figure would be in the center or near center of the image, doing some action.
Objectively it is making something never seen before. That student from before did too. Didn't they? I find it highly unlikely that, even copying Monet's style, they somehow made something someone exactly made before.
But that's not the point. Just like how when looking at modern society you need to term labour in "socially necessary labour time," here too you need to look at it socially. Yes, objectively, these models are making new things, and can even combine them to make new things. But I'm struck by a quote Dan Olsen had about Doug Walker, saying, "He is a fundamentally incurious person who thinks a groundbreaking idea is 'what if Batman met Mario.'"
To put it simply, ai art can't experiment. Obviously you can regenerate a prompt. Let's say you only had training data of socialist realist paintings. Would you be able to make "defeat the whites a red wedge"? Or the visa versa. Perhaps that's a strawman. What of you gave training data of all the non-cubist paintings to that point and training data of real world images, would it be able to make a cubist painting?
Theres also a consideration to be made, could it make an inverse? Many movements were made to be the inverse to a movement at the time. Realism itself was an inverse to the romantic movement. Could I ask an ai to generate the inverse of a romantic painting? Or the inverse of an anime style?
"Ah-hah" you say "well you can ask it to generate it with specific charecteristics." Well, what would those characteristics be? Maybe you could ask it to generate the inverse of anime charecters, small eyes, textured skin, etc., but what's the inverse of the style?
Does ai understand humans? That's the key question. I don't mean this metaphyscially, I do mean it literally.
Can someone who hasn't seen the world, hasn't seen other people, hasn't lived, produce new art? If you put a homonculous in a box and showed it all the art of humanity, would it be able to make something new?
I'd say no.
Lets go back to your sunset example. When someone draws a sunset, do they simply copy it? Ignoring the difficulty that actually entails, no. They stylyze it. They emphasize, they warp it, not to objective reality but to what they feel.
If you strapped a camera onto a person and recorded every frame of their life from birth to death, could you get it to make an oil painting of a sunset?
Sure, a human wouldn't be able to either, but they could make their own "training data," per se. They can make it, change it, improve it. Conversely, ai's can't train of themselves or else the recursion gets even worse
I get what you are saying and I think you are making a good point about the limits of using AI as a tool to produce art as it exists right now. But here is the thing, human art does not come from a vacuum either. It comes from material conditions, from culture, from all the art we have seen before and our experience of the world we live in. That is also a form of training data, just more varied and embodied than what the LLM gets.
You say a human can make an inverse of a style like realism being an inverse of romanticism. But that inverse was not random inspiration that came from some separate realm of Platonic objects. It came from a specific social and political context the artist is embedded in. Realism emerged because artists were responding to industrialization, to photography, to new ideas about truth and observation. They did not just think what if I do the opposite of romanticism in a vacuum. New styles are a reaction to the material conditions people experience. These are the same conditions that shaped their training data of the world.
Now imagine you take that homunculus in a box and show it all the art of humanity. You say it could not make something new, but could a human raised in complete sensory isolation with only a library of art make something truly new? That's also highly unlikely because novelty in human art comes from bumping up against reality, from making mistakes, and having a body that feels cold and hunger. On top of that, it is ultimately a human user that generates the prompts. The idea to create something doesn't originate within the LLM, it's just a medium somebody uses because they had a though in their head and they wanted to visualize it in some way.
The issue is not that AI cannot experiment either, you could create a system where an agent produces images and people rate them on how much they like them, and it learns the style people appreciate most. This type of stuff has already been done by the way. The key gap is that training data that LLMs are fed is narrow and static while ours is dynamic and embodied. But, say, you gave an AI a robot body and let it wander the world for twenty years, let it have a lot of embodies experiences, its outputs would end up having the same material basis as human art. And it wouldn't be because it has a soul, it would simply be a product of having a richer set of statistical patterns to draw from and new ways to contextualize them.
You say AI cannot train on itself without recursion getting worse. That is true for current methods, but humans can also get stuck in recursive loops. Think of artists who just copy their own style forever which is a form of recursion too. The ones who do break out do so by encountering new experiences.
So I do not think the difference is of kind, but rather one of breadth and embodiment. The human advantage is that we are embedded in the messy physical and social world. But we should not treat that as some sort of a metaphysical miracle.
The homonculus thing was simply to avoid the logical problems of somehow having someone both completely sensory deprived, and able to draw and such.
No, a human completely sensory deprived human with all the knowledge of art would not be able to make something new. That's exactly the point I'm trying to make. If an ai cannot experience, then it cannot make anything new, just as how a human who cannot do those things cannot.
If you want to talk about the future, then that may be a different matter. But it would have to be a big leap in technology
But my point is that the AI already has a lot of existing human art baked into it, and the human driving it provides the genuine novel experience you talk about. The idea for a prompt comes from the human user of the system, then the system combines all its training to produce an image, and the human reviews it and decides whether it matches what they intended, then iterates on that. So, that's precisely the element where novelty comes from with the technology as it works today. What the model does is automate the mechanical process of producing the image. The person doesn't need to learn how to paint or use photoshop, they just bring their creativity to the table.
Art is a lot more than just the concept of what you want. I'm not saying something has intrinsic value by being worked on nk matter its content, but just fundamentally that by its very nature you cannot mathematically/statistically create art that appeals at any higher value level. Maybe now it's good but if this technology doesn't improve rapidly, I'm sure in 5, 10 years everyone is gonna be sick of it. Edit: actually people are already sick of it, on social media it'll exist because the bar is through the floor as it's free. But otherwise, I doubt anything more is going to come of it
Also, saying humans can get in a recursion loop is fundamentally different from ai. You're a marxist, you should know the difference. Corruption can occur in socialist countries, corruption will occur in capitalist ones. It's the same here. Humans, individually, can get stuck in loops, but humanity rarely will get stuck in one for a long time. Ai fundamentally will get stuck, not being able to cannibalize it's own creations.
At the very least, it is not you or me making this art. If we asked someone else to make a concept for us, that's not us doing work. It's the programmers who make these ai's that are, effectively, making the art. They at the very least make the machine that makes the art.
As a materialist, I absolutely do think that art is something you can mathematically/statistically create. Thinking otherwise involves believing in magic basically. As a Marxist, I think humans are product of the material world, and that our brains evolved to do computation through the selection pressures in the environment. In my view, the brain is a biological computer that evolved to allow us to interact with our environment more predictably, which created obvious survival advantages for us. I don't think what the brain does is exclusive to biology, and I see no reason why artificial systems would not be able to do similar types of computation.
I don't really see any evidence to support the argument that what the brain does is fundamentally different from AI. There's certainly a huge difference in complexity, and I'm not suggesting that LLMs are qualitatively comparable to what human brains do in any way here. But on a fundamental level, I think that both are inference engines which build an internal model through reinforcement training. I'm also not sure what you're basing the argument that AI must fundamentally get stuck in loops.
And the programmers making these AI are not making this art any more than Adobe makes art by creating photoshop. The art is the product of human intention and the tool. If we asked someone to draw an idea that we had to a specification that we came up with, then we would be collaborating on art.
I think that art requires suffering. If there was a way that could make an art piece without suffering and spending any time on it, it would be worthless
I'm not even talking about studying it, although it does also requires a lot of effort, but the process of making art, process of trying and trying again and again, making errors, fixing them, making more errors, until it's something you at least fine with it. It applies not only for drawing, but for every artistic expression I think that if we remove that part from the process of creating art, we strip it of any worth. That's why hand-made stuff is still relevant, hand-made mug might be worth more for you because you made it, or someone you love made it, or you just bought it from some weird funny looking guy who makes mugs and sell it for overprice, even though there's factory made objectively better mugs for objectively better price. So, maybe AI does the same thing as human brain but faster, just like factory machines can do things better than humans. But the time, effort, the suffering of making something is what gives art any value in my opinion. Just like, well, any kind of work requires effort, but if in production faster and cheaper is desired thing, art has different goals, I think that human effort to just make pretty picture or depict your idea is more desiarable than whatever AI might provide with art. Although I do agree that if your goal is to depict your idea and you lack the skills required but you really want to, using AI is a way, especially if you lack time and/or will to commit to learning art, and you lack finances to commision artist. But if I see a painting I can be impressed with artist learning all the skills required, all the details or even lack of them, it is cool. If I see ai generated image the best I can say that it might be neat idea, I could not possible comment on execution, just like I couldn't comment on a factory made mug. (Well I could say that factory is cool thing too, and that AI can generate pictures is pretty cool too, I just can't agree that it can be considered art) In art, as I see it, the skill of the author is equally, if not more important than idea. Because whatever idea you might put in it, people will intrepret it however they see it unless you just spell it out and if you do what even the point
Sorry if it's a bit incoherent I'm also thinking while writing..
I mean that's where the whole term artisanal comes in. People place more value on things that were crafted by hand end to end as opposed to manufactured using an industrial process. That's why I don't really see image generation replacing traditional art. People still paint on canvas even though photoshop exists, and we tend to place more value on a physical painting than a digital one.
I'd argue AI generated content serves a different purpose from high art though, same way a mug made in a factory coexists with artisanal mugs you'd attach sentimental value to. For example, take the whole Iranian Lego videos that turned out to be extremely effective propaganda. Nobody would argue they're high art you'd display in a gallery, but they are effective, they evoke an emotional response, and they convey the message in a clever and creative way.
Before AI tools were available, the US effectively had a monopoly on such messaging because it was a highly labor intensive process that could only be accomplished by skilled 3d artists. It took a company the size of Pixar to produce something equivalent. But now, this process has been democratized, and a small studio in Iran can punch way above their weight to challenge imperial propaganda.
I think that's where the real value for this stuff comes in. Iran proved that we can use it for effective messaging. We can make memes and catchy short videos that promote our ideas and do it in a way that connects with people.