I had to put together a presentation that was somewhat graphic-heavy. I used Dall-e 2, and the ability to get things so tailored to what I needed, so quickly, was absolutely worth it.
I don't know if AI will replace stock photography, but I wouldn't bet against it.
The ability to charge a high price for stock photography is over though.
Stable Diffusion is still good but you need to fine tune it (not that expensive with Lora) and use controlnet to get good results. The model is just too small on its own.
What's behind MidJourney, and how has it not been matched/defeated? There's a lot of interest in the space and money to be made, and MidJourney proves that quality results are possible. Is it because of the sheer costs or a secret sauce?
They combine multiple custom algorithms that are designed to run on high-RAM (>24GB) GPUs. And, they incorporate a lot of user-provided feedback into the training of new models. If you have more time on your hands than cash, you can get some free generation credit by rating images.
That's mostly because they make a finished end-user product, which nobody else does (except OpenAI and DALL-E, but they are mostly focused on language processing right now).
- It runs in the cloud with more VRAM available than on consumer machines, so it's able to use more parameters and custom layers. This makes it more coherent.
- It's finetuned on a large RLHF dataset, which they have from the users voting for bot-generated images in their massively popular Discord server.
- It includes a decent share of custom processing and manual tuning (system prompts, custom layers trained on god knows what etc)
Their closest contender, Stability AI, is currently busy training SDXL, their new model, on a RLHF dataset they are collecting on their much smaller Discord channel. So Midjourney is SOTA for text-to-image.
There's DeepFloyd IF which is large enough and includes a 11B LLM for natural language understanding, so it has much less issues understanding complex prompts such as "koi fish doing a handstand on a longboard rolling down the slope of a moon crater". But it's not a finished product (no complex pipelines, no RLHF), and the license is restrictive.
That said, Midjourney's strengths are also its weaknesses. Midjourney is good for the text-to-image, but not much else, and image generation is much more than just text-to-image. MJ gives great results off the bat, but struggles with unusual requests, complex compositions with multiple objects, and can't use various elaborate trickery available in the much smaller Stable Diffusion (which is mediocre by default but has powerful higher-order guidance and training, and is also much better at complex things). Midjourney is also censored - you can't have Joe Biden eating a slice of porcupine.
Well you can gain a free GPU hour by rating images. I am pretty sure midjourney is doing some kind of reinforcement learning from human feedback.
The 5.1 raw model is simply unbelievable for photo realistic images. There is 4a,4b,4c, 5, 5.1, 5.1 raw, the anime model I never use. All are interesting but how often do you read anyone mention what midjourney model they are talking about? Basically never.
There are certain circumstances that I have seen stable diffusion match the output but overall I don't even think it is close.
Personally, it wouldn't even occur to me to look for stock photos now as opposed to just making variations in midjourney. There is a bit of a learning curve with so many options in midjourney so I take anything I read about it with a grain of salt. Most people are using these tools to make absolute shit.
Has been for a long time from what I have gathered. Stock photography has been a very hard way to make money as a photographer for awhile based on the forums I read.
When I looked at stock photography as a possible side income, licensing my photographs (those that Getty wanted - they didn't care for most of them because they were more artistic than stock) would have gotten pennies for each photograph.
I can certainly appreciate those who take stock photography (and I've seen some of them in my travels - taking 100s photos of slightly different poses of a model at a location because the stock buyer could say "I like this one but the arm is in the wrong spot for the page layout" and go with someone else's photograph).
Being a stock photographer has never been easy. In the days before digital it represented countless rolls of film shot (and processed at the photographer's expense) and now its even more digital images to ensure that you've got just the right one for some client a few years from now.
It's not an easy job and no one has gotten rich off of it.
And the sad thing is that these models are most likely trained upon most of these stock photos to begin with. I can understand the frustration of the industry.
Nope - a.i. is essentially just rearranging pixels because you can formally delineate (via mathematical expressions) how to arrive from input set to output image. To compare this in any respectable degree to human creativity is not reasonable.
That is an incredibly strong assertion that evidence does not seem to support currently. So, you may take it on faith if you wish, however there is growing interest in the notion that creativity itself is that which is not systemetizable.
You gotta back up claims like that with a source.
As far as I know, nobody can prove we are not in a simulation.
Meaning everything is made out of math.
Today's AI has taught us that everything people show us they think is unique, isn't really. Everything comes from somewhere, including the things that add up to what we call creativity.
That's actually untrue to me as a professional artist. A.I. art has shown me that cheap permutations (which have been done in art even before A.I.) can still capture the untrained eye. I don't think that's anything particularly new.
And absolutely, however creativity is the tip of the spear which pushes beyond constraints to develop novel AND meaningful style (like Beksinski, Kopera, Kostetsky, Moebius, etc).
If they couldn't tell it was AI, yes. Winning an art competition is by itself paradoxical in undermining the point of art, so I don't know how that's relevant, but it is certainly an impressively weak point.
So how would one define "trained eye" without having to defer to your personal opinion? You don't think there is any merit to the point that people chosen by the community of artists to judge art are fooled by it?
I never said it can't. And you're mistaken unfortunately - it is extremely difficult to prove a negative, as most people trained in philosophy intimately understand. For instance, "prove God is NOT real".
They're not arguing for the general case. In this case the negation, "just" requires showing a single computation done by a brain which provably can't be simulated by a universal turing machine.
I think my point flew over your head. The burden of proof is on you to provide the source that we can't simulate the whole universe, humans included.
This is something that cannot be proved to be true, even if it was the truth, it is possible that you can only prove it to be false, and nobody managed to do it yet.
That's why it's on you to prove it to be false. Especially since you claim that something cannot be computed with maths.
I never said it can't. And you're mistaken unfortunately - it is extremely difficult to prove a negative, as most people trained in philosophy intimately understand. For instance, "prove God is NOT real". You might be new to analytical thinking, and I would recommend diving a bit deeper into "burden of proof" type arguments before proceeding further.
> You might be new to analytical thinking, and I would recommend diving a bit deeper into "burden of proof" type arguments before proceeding further.
I guess my last response was too hard for you to understand since you respond with this. Perhaps you are new to things like "reading between the lines".
You are the one who made this claim:
> That is an incredibly strong assertion that evidence does not seem to support currently.
To avoid loosing more time, I'll write what you failed to understand:
"That evidence does not seem to support currently" is your claim, something you can prove, shows the evidence you are talking about.
No, you are so far up in your ivory tower, you started to use pseudo-science, and now accuse peoples of being a troll, it's easier to lie to yourself than face the reality.
You still didn't send any link to the evidence you were talking about.
Proving that we can't reliably simulate human brains only requires a single example of a signal emanating from anywhere in any brain that can not be traced to a cause via known physics.
Proving that we can reliably simulate human brains requires showing that nowhere in any human brains does such signals ever emanate. We can get close if we one day have the capacity to run a simulated brain for long enough to show it appears to function like a human, but when the counter hypothesises a violation of known physics that may well be intermittent and extremely limited, we can't realistically absolutely prove its absence.
As such, the former is tractable if such signals can occur, the latter is not, and so I think my comparison of it elsewhere to a claim of Russell's proverbial orbiting teapot is reasonable - people will always be able to claim that there is some so far unobserved difference, and given it postulates highly localised violations of known physics and this seems like an absolutely extraordinary claim, the burden must be theirs.
If we can infinitely reach things untraceable to known physics then I don't see how it could ever be proven that we could perfectly simulate human brains, but for the sake of the argument: what resolution do we care to simulate the brain at? Technically speaking, a 1990's chatbot simulates the human brain to some non-zero resolution.
My point is that to make the claim "human creativity" can be simulated surely would have the burden.
The extraordinary claim would be the claim that something that violates known physics is happening in the brain, and that claim is a necessary prerequisite for it to be impossible to replicate the processes of the brain. As such the burden is squarely on anyone suggesting they can't be replicated to show at the very minimum a plausible indication that there's unknown physics going on.
Just one measurement that doesn't fit would be enough to shift the burden. Absent that, any suggestion we can't is no more than a religious belief.
It's an assertion that rests only on the assumption that the brain adheres to known physics or if it relies on unknown physics that the computational effects of those differences can be simulated or the physical effect replicated.
It'd take a truly extraordinary find that'd upend both physics and areas of logic for that assertion to be false.
Let's try something else, because you seem to take it that "brain = physics" is a strong enough argument to believe that all of human creativity is inherently systematic.
The definition of creativity itself should include the ability to subvert a system through abstraction. AI is a system so it cannot subvert itself (or, it hasn't yet). As humans, we subvert systems in mathematics by creating higher dimensional abstraction and shattering previous understanding of the fields. Often, mathematicians achieve this with a "lightening bolt-like realization" that they do not attribute to systematic thinking.
Please reconsider whether systematizability should be so quickly assumed through napkin philosophy. If you want, prove with a topological map of the creative apparatus in the brain that it produces results in a systematic way. That would at least scratch the itch and get the conversation going somewhere interesting.
What you call "systematizability" is irrelevant. Only whether or not a function can be computed matter, and if a function can be computed by a brain restricted to known physics, there is no reasonable theory under which it can compute anything which can't be computed by any universal turing machine. Given the absence of even any real hypothesis for any category of functions which can't, it'd be an absolutely extraordinary claim to suggest such a function exist.
Where you'd get a Nobel in physics if you could show unknown physics in the brain, you'd secure a Nobel in mathematics if you could come up with a class of functions the brain can compute which can't be computed by a computer, because it'd fundamentally break basic assumptions in any number of subfields of both maths and computer science.
What you describe as "subverting itself" does not require anything as drastic. It merely requires thought processes unobservable by the person in question, and processes to complex to casually understand and/or randomness. Given a large proportion of people don't even have an inner voice they can "listen to" in order to observe their thought processes other than via outputs, and we know e.g. from split-brain experiments that the conscious part of human reasoning often outright invents knowledge and observation of thought processes, we have no reliable way of analysing the extent to which thinking is "systematic", and plenty of evidence to suggest we can not trust any such notion.
Slight jitter + strong and complex selection functions, like "is it useful?", "is it more useful than what came before?", "is it appreciated by the audience?", etc.
The creative generation part - adding some randomness to the process - is trivial with machines. Much easier than with humans adults, who often have to learn to allow more randomness and in their thinking/feeling, in order to make some creative work. But the selection functions are deeply social, which is something we're instinctively good at, and had little luck teaching the machines so far.
Well, with the possible exception of GPT-4, which may have learned just enough about what humans like, that it could do a passing job at discriminating art/innovation from noise. I say may - it feels possible, but we won't know until someone properly tests for it.
> Well, with the possible exception of GPT-4, which may have learned just enough about what humans like, that it could do a passing job at discriminating art/innovation from noise. I say may - it feels possible, but we won't know until someone properly tests for it.
I've tried giving GPT4 some axes to rank writing on, and feed it some paragraphs. It's not perfect at it, but it's a far better critic than writer.
E.g. a while back I had it write something "in the style of Douglas Adams" and then had it critique it in another chat, and it recognised the imitation and proceeded to quite harshly judge its own writing for being flat and lacking direction and purpose. It was quite entertaining (more so than its superficially amusing Douglas Adams imitation).
Any output from a brain can also be represented mathematically-- it's a physical object, after all. Coming up with the parameters for those expressions is the much harder part, for both brain and AI.
Once again, that is an incredibly strong assertion and if you wish to be reasonable then the burden of proof is on you to prove those parameters before making such claims (e.g. that creativity = math).
The notion that creativity lies beyond what is systemetizable (or otoh, the value in creativity is to defy that which can be computed), seems more in line with my sense of it.
LLM and text-to-image models already show empirically that creativity is computable, of course, as the AI effect[1] says, if a machine did it, then it's just predicting the next word or rearranging pixels (which doesn't really make sense), if a human did it, then it's intelligence or creativity.
The reverse is also true which makes your entire point hypocritically paradoxical. If a human makes it we immediately assume it can be computed (creativity). That is the strong claim which I think requires the burden of proof.
I'm pointing out that your "AI Effect" is one-sided in the wrong direction.
Surely you agree that we care more to prove the affirmative of whether a machine is conscious rather than prove that it isn't. If not, I don't think we can proceed.
The issue is that the AI Effect you linked is often actually used in the reverse, including by yourself, where the burden of proof should be to prove a machine is NOT conscious. However, like we've agreed, we care for proof in the affirmative.
Thus, the paradox lies in that the "AI Effect" concept is critical of those demanding the proof which is more valuable.
I don't know where the concept of "consciousness" came from in the discussion but as I said in the first comment these models already empirically show intelligence and creativity, proving otherwise would be much more meaningful than aligning with the evidence.
Yeah I think my point flew over your head, which is fine.
And you're incorrect - you cannot "empirically prove creativity". As a professional artist I don't think you quite understand that the type of creativity we are interested in is not just "permutation", there is clearly a human component.
I'm trying to show you that you are falling into the reverse of the AI effect by assuming that permutation is creativity.
It doesn't seem like you can even define creativity in the first place for any of us to even attempt to answer your supposed point that flew over everyone's heads.
>you're incorrect - you cannot "empirically prove creativity"
Do humans even have creativity at this point or do you require a formal proof even for that? Ahah
>As a professional artist
Now I understand where the friction against accepting that even machines show signs of creativity comes from, I imagine you're an illustrator, so the feeling of becoming obsolete is the bias against machines.
First of all, we're discussing stock photography, not art. Second of all, Turing's arguments for why the Turing machine is a decent model of the brain seem pretty hard to bet against, and the Turing machine underlies all computation.
No - we're discussing whether someone producing stock photography can be trained in the same sense as a.i. However, I will give you that I'm potentially falsely assuming creativity is required to a comparable degree to art.
Also, gun-to-my-head I would not bet on creativity being systemetizable, so I completely disagree.
You know, I actually think you're right about the process, but totally wrong about the results. Yes, AI art is different in terms of creativity. It is only grounded in reality to a lesser extent than human creativity.
Where you're going wrong is what society will go for. Given the choice between reality and fantasy ... society will choose fantasy. AI will "win" fantasy, hands down.
AI will always lose because the result can only be a permutation of what already exists and the next artist will grow up and push the boundary - hence, AI will always lag behind the tip of the spear which is human creativity.
Not to mention, you are already wrong because art communities around human art are thriving now more than ever for the simple fact that the result of AI itself is non-communal (does not produce thriving community) and cannot compete at the bleeding edge of style.
Again, this is already happening regardless of your armchair hypothesis.
If the brain is a physical object that computes, and we have creativity, it stands to reason that creativity is computable. Why would you assume that creativity is not systematizable when we already have evidence on the physicality of the brain influencing behavior (ie cut out one part, personality changes, take drugs like LSD that change the chemical components of the brain, notice that creativity improves)?
Your assumption is IMO unfounded, it actually seems like you have the burden of proof given what I said above. Just because it seems more in line with your sense of it doesn't make it true, seems like you think creativity is some mystical nonphysical force (because if it were physical, it would be able to be modeled and computed, as the parent says).
Edit, I see, you're an artist so it's likely why you believe in creativity as a nonphysical force. I've noticed that many artists say this, with no evidence, while most engineers say the opposite, that it's systematized, likely that's why we're in our respective fields.
Once again since you missed the entire point: I don't believe creativity is non-physical, I believe it is non-systematizable, and my brother achieving a PhD in computational neuroscience from a top university actually initially shared this with me.
How are you defining systematizable? As in, creativity can be made via a systematic process? How do you think a brain works, physically speaking, otherwise? As others have mentioned, it is you who holds the burden of proof, not everyone else who already knows that the brain produces creativity.
I gave you the benefit of the doubt and re-read the thread.
Creativity: producing what is novel + valuable
Systematizable: can be enumerated via algorithmic operations
"As others have mentioned": that is not an argument
"not everyone else who already knows": also not an argument
Now, my point is that creativity as defined above requires a burden of proof to be defined as systemetizable. For instance, mathematical leaps are often achieved by a spark of "creativity" which breaks existing structures in a way which is difficult to grasp, and to call this systemetizable is a leap that requires a complex argument calling upon deep notions of computational theory and computational neuroscience (like producing a topological map of the creative structure of the brain and proving that it can be mimicked by a listable process).
Now, to apply this to art in a convincing way would surely be even harder. Do you at least understand my point? Or is "brain = physics = creative" satisfactory enough for you still? Thanks for reading in good faith this time. :)
I'd argue that the presence of anything in the brain which violates the laws of physics (necessary for the above claim to be false) is equivalent to Russells teapot, and absent extraordinary evidence it's a reasonable assumption to make that the brain adheres to known physics.
Yes, it does, because if the brain adheres to known physics it means the brain has the same limits on computability as any other universal turing machine, and while replicating the process can still be difficult, it can't be impossible.
That's a very fancy way with lots of fancy words to say "I have no idea how NN work, but if I sound smart maybe ppl will not figure out how stupid I sound". Well, you sound stupid and nothing what you said makes sense.
Delineate? Diffusion models can't be further away for delineating. They literally work by throwing random shit at the wall.
Input set? There is no input set once training is complete.
Human creativity? In stock photo industry? What next? "How I write while loop instead of for loop and I have achieved the nirvana?"
Incorrect - I'm clearly using delineate to mean providing the exact set of computational instructions which produce the output (image), which by the definition of "computable" must be true.
You seem emotional, "you sound stupid and nothing what you said makes sense", I wonder, why are you so triggered by a challenging idea?
The very point of computation is to evaluate clear, distinct, mathematical expressions (like reading a list of algorithmic steps to produce an NN).
Hmmm. Did I threaten your religious belief in the power of AI? :)
Seriously though, I used to think this was mostly a done deal that's being occasionally brought up by people fascinated with socialism or something. But the longer I live, the more I realize that the issue has not meaningfully improved at all - we've only been distracted from it. And it's much, much older than Marx and co. too. This, and not some general fear of technology, was the whole point of the Luddite movement (very relevant in context of AI). And you can probably find examples stretching back way before industrial revolution too.
What workers own the means of production in the case of an AI? Democratization via open source works, but it actually extends beyond workers to end users. If all the workers at OpenAI were equal shareholders and owned the means of their production, this would not change anything about them profiting from the company and all of their users. Stallman was and continues to be right, open source and freedom for end users is really the only way.
Yeeaasss, but so is a lathe or a kiln or a chainsaw.
I don't know if OpenAI has actually confirmed what hardware they use or if the press and bloggers did citeogenesis on each other, but the claim is they're using A100 GPUs, and the bigger models have to use several in parallel. I've seen detached houses in Eastern Europe going for the price of two of those cards.
The DGX-2 is (was?) priced more like a detached house in Western Europe; the GH200 they just announced appears to be "if you have to ask, you can't afford it".
This is similar to how, despite that lathes, kilns, and chainsaws can be bought in supermarkets as impulse purchases, industrial versions of each are rather more expensive.
> The trajectory of ML appears to point towards massive transfer of wealth to capital owners and impoverishment of the working classes.
That’s the natural end of capitalism, as has been noted literally since the word “capitalism” waa coined for it (largely, for that reason.)
It is end to which every tool and technology is put under that system, and has nothing to do with the particular technology and everything to do with the system.
The capitalist system doesn't preclude workers from acquiring capital or capitalist from working. It also doesn't dictate the relative return on capital investment versus worker labor.
The rate of return on capital is what dictates the concentration of wealth, or conversely, the dispersal of wealth.
The conceptual systems introduced by Marx and others is that only workers can own capital and are based on a set of assumptions about Capital returns versus labor power.
> The capitalist system doesn't preclude workers from acquiring capital or capitalist from working.
In the same sense that the US electoral system doesn’t preclude more than two viable parties. The system of property rights that defines capitalism doesn’t specify any limit, but the structure it creates promotes it in function.
> The conceptual systems introduced by Marx and others is that only workers can own capital
No, the conceptual system introduced by Marx rejects the concept of ownership as applied to capital [0], though it favors workers exercising control of the capital to which their labor is applied. Different variants derived from this and related ideas have different specific mechanism for this exercise of control.
[0] When Marxists talk about rejecting/abolishing “private property”, it refers to this, because in the Marxists terminology “private property” refers not to property of non-state individuals and entities in general as it does in current common use, but specifically to property rights in the means of production, i.e., capital.
I understand the carve out in Marx's Theory for non-productive personal property in contrast to the means of production.
That said, I don't see you the distinction between controlling the means of production use for one's labor and owning it, when controlling the means of production also includes allocation of the additional value created using them.
It also is also worth noting how far current thoughts on socialism and communism have come from Marxist theories. contemporary sentiment seem to advocate that individuals (working or not) are entitled share of the labor value created by others, even if they are in another workplace.
All this is really an aside from what I thought my greater point was about the underpinning assumptions or conditions that lead to the "expected" systemic outcomes. The US federal electoral system currently favors two parties, but this could change with different conditions. For example, ranked choice voting, and proportion state electoral allocation could impact this without overturning the entire governmental system.
Similarly, the concentration of capital is based on an number of conditional assumptions. For example, if the real rate of return on capital is less than the rate of capital creation, you would not expect the concentration of wealth to increase. under this condition, capital investment could only serve to slow the losses due to inflation.
> That said, I don't see you the distinction between controlling the means of production use for one's labor and owning it
Conditional control while you are a laborer whose labor is applied to particular capital (which, to be fair, is one of many patterns advocated within Marxist-influenced socialism, not the only operationalization of the concept) is distinct from ownership which persists until the owner chooses to alienate the property.
> It also is also worth noting how far current thoughts on socialism and communism have come from Marxist theories. contemporary sentiment seem to advocate that individuals (working or not) are entitled share of the labor value created by others, even if they are in another workplace.
Yeah, its not like Marx said anything like: “In a higher phase of communist society, after the enslaving subordination of the individual to the division of labor, and therewith also the antithesis between mental and physical labor, has vanished; after labor has become not only a means of life but life's prime want; after the productive forces have also increased with the all-around development of the individual, and all the springs of co-operative wealth flow more abundantly—only then can the narrow horizon of bourgeois right be crossed in its entirety and society inscribe on its banners: From each according to his ability, to each according to his needs!” [0]
[0] Except, of course, that he did, in Critique of the Gotha Programme, 1875.
Im not sure if you are plugging the quote in support or disagreement. The internet is funny that way. Is a very modern take to utterly ignore the lengthy preconditions laid out and jump straight to the conclusion if it means the reader is entitled to free stuff. Why wait for a society where the products of human labor are given freely and consensually when they can simply be taken today.
> We need to find a way to distribute the benefits of such developments.
It’s called free software and making models at various ‘weight classes’ to support many GPUs. Basically don’t let companies like OpenAI/Google/Microsoft run the show and, ironically, support companies like Facebook who open source their models.
And yes, I know OpenAI has open sourced models in the past, but the fact that they won’t open source ChatGPT, GPT-4, and select other models like DALL-E 1/2 make me wary of them.
I don't know how to easily quantify these things but your analysis ignores the other part of the transaction. The consumer of the photos (or ML generated text) benefits through the lower cost, easier access, increased project velocity, etc.
Market disruption due to technology/innovation/progress is nothing new. There have always been winners and losers but the aggregate wealth/value continues to climb. Devising ways to mitigate the loses seems like a more feasible approach than pretending that we can prevent these sorts of changes through regulation or legislation.
You can run Stable Diffusion on any PC. A $50 desktop PC which is ten years old will generate an image in about 6 minutes. A more modern PC will do it in 2 minutes. That's the CPU. A $300 GPU is dramatically faster. And in a few years the same GPU will be $50.
You can rent the hardware from someone else, but that's a matter of convenience, not affordability.
> support what the industrial revolution did to cobblers and weavers
Which made cheap shoes and clothes available to everybody.
If you want to buy a tailor made shirt, it'll cost you at least $500. And that's still with machine made cloth. What do you think it would cost if the cloth was made by hand by a weaver?
> Which made cheap shoes and clothes available to everybody.
I'd argue the efficiency in production led to lower prices - while the concentration in wealth mostly kept buying power/living standards lower for the workers (or; capitalist class reaped an unfair share of productivity gains).
I believe commodisation of "knowledge" work will lead to lower prices (and unfair distribution) too.
The standard of living of the poor in America went up all through the 19th century. Things like average height and longevity are strong markers of that. Also, average clothing sizes increased dramatically.
Go take a look at Fort Williams, and at the clothes in it they preserved. They look like kids' clothes. The same for Civil War uniforms. Larger than the FW clothes, but still looking like kid clothes.
I flipped through a high school history book a few years ago, and it insisted that the middle class in America didn't emerge until FDR. This is laughably wrong. It was there from the beginning.
What? Do photographers not need a camera, at times costing thousands or tens of thousands of dollars when including lenses…to make that stock photography?
Painters are trained by copying the Masters. Shredding guitar players are trained on violin caprices. Martha Stewart was trained by watching Julia Child.
Real life is what good illustrators actually train from: figure studies, value studies etc. - from life. (Edit: you study like the old masters, not necessarily the paintings of them - though it is also done ofc, but not the most important part)
And the Masters developed their style over decades pushing beyond the previously systematized constraints of their peers, something which can definitionally not be achieved by AI which works within the dimensionality of the training set and thus will always lag behind the next artistic stylistic advancements.
Then how come everybody thinks AI has a more or less distinct style? Incredible amounts of detail everywhere, whereas human artists tend to focus details in service of the "object" or purpose of the artwork.
AI art is more like what would happen if you took Baroque art and applied it to modern styles. For a human artist the effort is not worth it.
Most artists actually can tell which style each AI piece is stealing from. If you want the perfect example google "Beksinski". The overall goopy feeling of AI art is actually a turn off for most people, btw.
Yes, and they own their creative output, instead of the firm who trained them.
The proper analogy to ML is that the owners of the latter train a slave, and then use that training process as justification for owning everything[1] the slave ever produces.
How can freemen compete with that?
[1] Which also happens to be a perversion of copyright law, since AI can't enter into a contract that can assign rights to creative output to it's owner.
I understand why people are making this argument about drawing/painting – I don't really agree with it, but I understand how it coheres – but it doesn't follow for photography.
This is incorrect. I was a photographer and trained very little looking at other pictures. There are a few guidelines about framing I got from both written rules and a few examples, but those examples were never of things I took pictures of.
I took pictures of a much wider variety of things than I looked at pictures of. Instagram and other sources of millions of pictures did not exist and I was perfectly capable of taking a picture without that “training”.
It may not have been conscious, proactive “training”, but before you ever picked up your camera you’d been exposed (pardon the pun) to thousands (at least) of photos others had taken. That unconscious exposure still counts, even if you don’t think of it as training.
No I hadn’t. I don’t think you understood how infrequently you would come across photos of things like rare wildlife and remote wilderness locations in the 80s.
You’re missing the forest for the trees here. The point is that a photographer doesn’t learn how to produce pictures of a screwdriver by looking at pictures of a screwdriver. A photographer learns how to take pictures generally and then can instantly be capable of producing pictures of far more objects than he/she has ever observed.
Generative AI is not capable of that because it’s not operating a camera in reality.
I don’t think AI will replace stock photos because AIs need some photos to build the AI photos.
Probably AI will help photographers to cut costs, I mean by taking less “real” photos and using the AI to generate other similar variants, example: “Photos Used to Generate AI Images for Client So Photographer Never Shoots Again Photos Used to Generate AI Images for Client So Photographer Never Shoots Again” https://petapixel.com/2022/12/21/photos-used-to-generate-ai-...
It’s a very cool project but totally different from taking only one real photo, edit it with an AI and sell the created AI photo variants to other clients.
There are so many photographers out there putting so many images on platforms that you get paid almost nothing for them. Unless you get a great image and have a proper agent or contact passing them along you don't get much. Payout for some of my images ranged from under 5 USD to at most 40 USD for one image one time.
Some photographers just spam TONS of images on the platforms to get some cash by betting on the long tail. Someone, somewhere with a shutterstock subscription might need your image for a local magazine/paper and you get 2 dollars for it.
That kind of stock photography might suffer even more with generated images. Actual premium stock or for hire photography will not.
The smarter producers will focus on building sets of images that feature common thematic elements. For example, using the same models in business situations. Or using the same style backgrounds with different subjects (people or vehicles). Sometimes it's a lot of shots of the same city with the same style. Other times it's different sets of illustrations (i.e. "corporate memphis meetings" or "ultra realistic athletes) that are the same style. This makes buyers more likely to select multiples of their work.
> Someone, somewhere with a shutterstock subscription might need your image for a local magazine/paper and you get 2 dollars for it.
Yep, that's how a lot of it works.
> Actual premium stock or for hire photography will not.
I kind of wonder about that. When I can upload a picture of my product and prompt, Photo of a scene with my product as the subject on 10 N Main street in Wherever, usa. I'm going to be a lot less likely to ship product (and get it lost or broken), commission a photographer in Wherever USA to go take the picture, wait 2 weeks for the right weather, only to find that some unreleased person's face in the reflection on the side of the product.
I totally agree that it’s going to be very hard for stock photography - but is that a fair comparison? Would you have used stock photography for that presentation to begin with? A lot of the time people tend to just use copyright stuff for presentations because it’s just internal / client presentation use.
For that kind of thing - just explaining an idea - midjourney is amazing. So much faster than making something in photoshop and so much more accurate than google searching.
I think the real issues for stock libraries is when midjourney starts replacing actual output. This is starting to happen but still very early. It’s gonna get a lot worse!
I haven't had to put together a new graphics-heavy presentation recently but in the future I could absolutely see using AI to augment Creative Commons and various graphs etc. from the Web. (I don't really use stock photos because I don't have a budget.)
As things stand, I would still probably use stock for "official" company marketing materials although I understand that some companies, especially smaller ones, are probably less risk adverse.
I am extremely skeptical when people say AI will replace all photographers, no one wants AI generated wedding photos. However, it has certainly replaced my use of stock photos.
> The ability to charge a high price for stock photography is over though.
Yes. Selling stock photos for crappy ads will be dead in a year. Unless you're at the level of Ansel Adams or Margeret Bourke-White, your content is not worth much.
Even Margaret Bourke-White's content isn't worth much. Her vintage silver gelatin prints, sure, but none of her LOC-housed imagery, and certainly none of her work available through Getty.
There is a certain class of images that will be largely unaffected - anything where authenticity matters. Such as a space walk photo from the ISS, a celebration moment in the Olympics, presidential speech, met gala fashion shots etc. and the authentic non generative photos could go up in value as photographers who do not have access to those moments either start to resort to AI or to fall out of the competition.
Hmm I wonder... Space walk for your average article might be more selling if generated. Olympics, maybe some posing setup would sell more too. Presidential speech surely they want more "representative" gallery or fix some defects. Met Gala fashion, again make things better.
Propaganda and need to sell things are powerful motives. And AI generation and modification are great tools for that. Authenticity sell to very few.
Like it or not, it's here like CGI in movies bestsellers. I don't envy artists, most of them will have to rethink their plans and switch to graphics or web designers, or something similar. Not sure who will be next. Probably GPT will take over virtual bookshelves on Amazon. Mandatory "AI generated" mark may help a bit. I will not waste my time reading it.
I think we should expect a flood of "masterpieces" in different channels. Including news reports, scientific papers. The solution could be closed humans only groups. But, like in chess today, it's hard to say who is cheating.
Dunno, 'brand name' writers of airport novels have been forever accused of having ghost-written their novels, I think AI will allow writers of lesser budgets to also get help in writing their novels. Or who knows, even the 'King'-s of the world will use AI to fill out their chapters.
yes, and we can expect a flood of those 'lesser budget' works.
One domain where I welcome AI is computer games. Having more natural flexible characters will make games more interesting. As opposite to current pre-written texts.
Also education, where kids can ask questions and give the answers. Not just read or listen passively.
Just need to figure out how to do AI inference without tanking rendering performance.
Offloading AI stuff to a cloud would make single player games unplayable offline and their servers won't stay online forever.
Another solution would be to use a second GPU just for AI stuff. But that would make most current gaming devices useless.
Real time might be difficult right now anyway. It takes too long to figure out what the AI would do with the more powerful LLMs.
On the bright side, compute power is probably going to continue increasing. Consumers might be able to run current LLMs on home hardware in a decade or two.
> Also education, where kids can ask questions and give the answers.
Not sure how parents feel about putting kids in front of a machine that can fabricate lies.
I mean a encyclopaedia getting information wrong is one thing - garbage in, garbage out - but one that actively invents lies is like added another layer of bad to things.
Probably better for kids learn to search for information and evaluate information sources.
> Not sure how parents feel about putting kids in front of a machine that can fabricate lies.
You mean like a human teacher?
Of course, I agree with your point, but worth remembering that teachers can and do (knowingly or unknowingly) give incorrect information to their students.
I think one big difference with ChatGPT etc. is, as far as I've seen, it will pretty much always give a confident sounding answer (unless it's on a banned topic) rather than saying something like "I don't know", which a (good) human teacher should do rather than just making something up.
This reminded me of one of my teachers who said that Cuba was a great place to live and the rest of Latin America should be more like them. She was essentially a communist who painted the old Soviet Union and communist countries in general as much better than the capitalist west.
I guess the task of finding something worth reading will get closer and closer to insurmountable under the flood of lesser and lesser productions. The more expensive ones will be deemed not worthy to produce and there, you'll be stuck between classics and lessers. I have no idea how to get out of this quickly approaching chaos. Stopping the lesser productions doesn't seem useful, also marking as "AI generated" will fail (is it still AI generated if I modified by hand two words on page 1?) and you'll have basically books generated on-demand wherever you click, hallucinations burying anything else I'm afraid.
Ghostwriting is actually a time honored way of breaking in. It’s similar to writing a great script that doesn’t get produced but getting offered work on another idea off the word of mouth.
This will never happen, promise you that a major CGI sequence will ship completely AI generated in a Marvel movie within 3 years. This tech will be so integrated into workflows of Disney and the like that the idea of an "AI Generated" watermark will be as ridiculous to them as a "3D Rendered" watermark and if any legislation comes about they'll be against it.
I could potentially see movies needing to credit AI in the end credits. Not sure how that would work out. Maybe list all the scenes and what AI was used? But none-the-less, not an easy clear cut solution.
In 5 years, the idea of AI being "cheating" will seem as silly as someone touching up an image a photoshop being "cheating" does to us today.
But I'm sure some people with expert darkroom skills had similar thoughts when photoshop was rising to prominence.
Consumers don't care what tools were used to make what they produce. "AI generated" mark is something only the people being made obsolete will want, and their power is rapidly declining.
The increase in income during a recession is because people are choosing stock instead of going and doing a photo shoot themselves.
Stock is still the easiest, safest, cheapest option - VS your own shoot.
So I don’t think that means it’s future proof in any way.
On the contrary it means that the same cost-saving decision will be made again and there’ll be a move to AI images when they are slightly more mature - ie when the legalities of copyright have been fully figured out.
(I work at a big ad agency and I know the copyright thing is currently a massive drag on potential usage)
You already know what you get with stock images. When you do a shoot, it's not a sure bet that you're even going to get the images you wanted or needed.
There is a lot of potential for aspiring photographers in creating (and licensing) photo libraries for the specific purpose of AI training.
Where previously you'd take 100 photos to choose the perfect one, now you'd include all of them, with labels that describe their quality or deficiences (which you do anyway as you curate the photos you took). This makes almost every shot the money shot.
Less glamorous for sure, a lot less money per photo, but more volume for virtually the same amount of work. And you can always treat the really unique and special photos as today and not include them in your AI dataset.
Combine that with the fact that a photo (and photoshop) pro can produce better results with AI image generation (because they can guide them better than people who don't know what they're doing), and the future doesn't look so bleak.
Where is the photographer's skill necessary in the process?
Eventually, all the "photographer" will need to do is to take a 8K video recording of the subject and AI will generate a set of great looking pictures (it already knows what people like).
Videos aren't a very good source for training data, because they usually contain compression artifacts and motion blur. I've experimented with AI training using video screencaps and, unfortunately - the resulting quality isn't as good as using high resolution still photos.
Another issue with using videos is overfitting, because video frames look too similar to each other. So you can get an inflexible model that always generates a similar background as that in the video.
Good AI training dataset is the opposite of that. It needs to contain different backgrounds and different lighting conditions. And as many different angles and poses as possible. Can't really get that from a single video.
Ok, but that's the least pertinent part of what they said. It seems you're busy arguing at different FPS -- a person takes endless stills ... that's still relatively low skilled?
I don't mean this in the context of training, but a "photographer" already using a trained model to apply on the video.
What the AI/model then needs to do is to extract the subject - person specific features and combine that with trained data. For that, artifacts/motion blur isn't a huge problem.
Good/great photographers have an eye for things like composition and color balance, which can help in refining training data to a place where AI can produce indistinguishable products from actual photographs.
Images generated by these systems look nice because the models were trained on photos that had already made it through curation filters before being published or distributed. Most people toss out their garbage photos and don't share them online.
Without that, computer-generated images would be far uglier and weirder. And if the computers had to get their new material from, say, a bunch of webcam feeds, the composition of their output would be terrible.
The skill required to tell a good fake from a bad one is different from the skill of a photographer who doesn't need to worry about e.g. hands being messed up, but it's still a special skill just like prompt crafting. Stock photographers can move into that, or into strategy.
Since AI is mixing to create images and not being truly creative, I reckon it will be a worthy business to be able to detect where the pieces in an AI- generated image comes from.
> where the pieces in an AI- generated image comes from.
A single image "comes from" millions or even billions of pieces. It truly works like that (albeit counterintuitive). It doesn't just combine 4 or 5 images and apply a filter.
First of all, lawmakers and judges don't necessarily know how AI works.
But even an image is made of a lot of pieces*, it's still not 100% "legal-concern-clean". Consider this hypothetical scenario: you collect one million different photos of McDonald's logo, and one million other random photos. Then you train your diffusion model and tell it to generate a logo. Of course it'd be still very similar to McDonald's logo.
SD's training data is supposed to be much more diversed than that. But how could we be sure? That's where the fuss comes from.
Isn't it too early to evaluate that? AI image generation is seeing improvements on a monthly basis; whatever trend we observe today will likely be out of date later this year.
It won't kill it quite yet, but it will reduce demand. I expect the major players in the industry, who are quite litigious, will try to show that their IP was used to train the models so they'll try to get their cut stifle or kill them at birth.
Isn't the uncopyrightable nature of AI generated artwork a potential issue here?
If I produce ad content for a company whose public image is incredibly important (e.g. Apple) and that content includes AI art what is to stop a competitor, or anyone else, from reusing that same art in a counter ad or to produce other content to similarly harm the original company?
On the other hand, if I were to use images from a stock photo company I could simply purchase an exclusive license the artwork used (and likely the entirety of the similar work from the shoot, used or not) to avoid this type of imitation.
> Isn't the uncopyrightable nature of AI generated artwork a potential issue here?
For some uses of stock art, probably not (nonexclusive use is common and cheaper), for others yes, if AI art is not copyrightable (an opinion of the Registrar of Copyrights that it is not if the only input from the human claimant to the finished work is a text prompt is perhaps being grossly overread the way some people seem to think it is settled law that imagery generated through the use of an “AI” tool is categorically uncopyrightable.)
I really don't see this bright line between "AI" tools and the ton of other tasks that Photoshop, Lightroom, etc. can do for you to various degrees. Increasingly (finally) voice/text will become an important command tool for many types of graphical tasks. Unless you drastically raise the threshold for copyright protection (e.g. you can't use a computer), I'm not sure where you draw a line.
by that definition it is enough to pass through an AI generated image in an upscaling program and it is now not just only the prompt.
or open it in Photoshop and click once with a content-aware fill somewhere.
> by that definition it is enough to pass through an AI generated image in an upscaling program and it is now not just only the prompt. or open it in Photoshop and click once with a content-aware fill somewhere.
No, its not, becauase the statement at issue covers only what is (in the Registrar’s interpretation) definitely outside of copyright protection, it does not do much to clarify what it takes outside of that clear exclusion to qualify for protection. To assume any image where the human work is more thane exactly a text prompt is protected is as much reading too much into it as assuming all AI-generated images are completely excluded.
Copyright rests on nothing. As an issue, it only needs lobbyists to solve. It's easy to write Section 230 for AI, giving the owners of AI a ton of rights in return for no obligations..
The rapid increase in 'adobe stock' seems obviously to have been influenced by the rise of generative images in the first place (the big growth matches perfectly with when these things were released).
If you haven't tried to purchase a stock photo recently, it is incredibly difficult. I can order same day delivery of physical goods from Amazon much faster than trying to buy a single stock photo. All of the services are subscriptions now and they have dark UX patterns to convince you the only way to get the images is to sign up for an annual plan that will auto-renew.
In comparison, typing a prompt into a generative AI is a superior user experience. I'm willing to pay a photographer, they just don't make it easy.
I'm not sure that is the case... I've used Shutterstock and Adobe Stock in the past two weeks and both times, it took about 3-4 minutes to find something I liked, and 10 sec (subscription) to 1 minute (non subscription) go through the license/payment cycle.
> All of the services are subscriptions now and they have dark UX patterns
They do. I have a subscription to one of the services, and it's just super easy and I use it enough it's worth the monthly. I'm not sure there's a huge market for licensing an image occasionally... It would be interesting to see the breakdown there. Also, Adobe has a huge advantage because Adobe Stock is highly integrated into Adobe's design software.
>I'm not sure there's a huge market for licensing an image occasionally.
Maybe for an iconic newsworthy photo.
But, yeah. As an individual, paying $50 or whatever for a photo for a book cover or something like that would be a rather rare thing. And if I'm a company who needs a steady stream of stock for articles, corporate presentations, etc., I'm almost certainly going to have a subscription if only for the convenience of not having to get individual purchases approved all the time.
> 1 minute (non subscription) go through the license/payment cycle
No, it takes me several minutes to get my lazy ass off my chair, go find my wallet which is inside some pants inside the laundry basket, pull it out, find my credit card, enter the number in, enter the billing address, enter the zip code only to find that Google autofill has now decided to edit my address, then go edit the address, only to find that Google autofill has decided to edit my city, go edit the city, then just go "screw this" and use Stable Diffusion.
If you want to minimize friction for payments (a) don't display any GDPR or newsletter popups (b) don't ask me to register an account or verify an e-mail address or phone number (c) just let me hold my credit card to the webcam to pay. Do NOT display 17 text boxes to fill in.
> it took about 3-4 minutes to find something I liked, and 10 sec (subscription) to 1 minute (non subscription) go through the license/payment cycle
It took around 5 minutes for me to even be allowed to download the free low res version of an image. Between forced signup (by clicking the “free download” button, with no indication of needing an account), waiting to confirm my account via email, and having to return to the page after forced redirects begging me to pay for a subscription I’ll never use, all for a single image to play with in GIMP.
I genuinely hate how predatory UX has become in recent years.
I'm not sure what you're talking about, buying from Getty, iStock, Shutterstock, Adobe, etc is pretty seamless. Maybe you mean if you don't have an existing account? But in the case...why would you be buying a stock photo? I still also occasionally use Unsplash, but I've never had a single issue buying a stock photo.
There are many humans I'd turn off if it wasn't illegal.
And even more of them I'd turn off only if I knew with certainty that I could turn them back on whenever I want, and that their memories would simply continue where they stopped.
I, like most people here (I presume), do not murder. But can you truly say you have never met a person, or got to know them really well, and thought, “this person doesn’t deserve oxygen”? I think most can at least somewhat empathize with having that thought.
I disagree that this is vitriol, and would rather argue it’s natural to have dark thoughts that one does not actually act on. We’re all just a bunch of animals that developed strong moral codes and varying abilities to process complex thoughts.
Blatantly stating this on top of your other contributions to this thread. I hope you never wonder why "techbros" have the reviled reputation they do, as you're the epitome of it.
What did you do in your past that you have to virtue signal so hard to correct it in your head?
Every person is capable of murder, that's why murder happens. It's a natural thing, in nature, creatures die and kill every day. The fact that you consider yourself so outstanding and morally erect that you are completely sure that you'd never kill a person and have never wished to only tells me that its likely you had never had any real introspective thought in your life.
Of all evil I deem you capable: Therefore I want good from you. Verily, I have often laughed at the weaklings who thought themselves good because they had no claws.
Yeah, I wasn't referring to the part you think I was referring to.
>And even more of them I'd turn off only if I knew with certainty that I could turn them back on whenever I want, and that their memories would simply continue where they stopped.
Certainly a strain of solipsism I've never heard before now, but still pathological.
And art cannot be controlled once it is released to the public. If you don't want people to see it, if you're too cowardly to accept the all the consequences that come with it, then don't release your art. It's really that simple.
It doesn't matter if it's my eyes and brain, the computer working for me, or the eye of Sauron for that matter. Once the cat is out of the bag, any control you think you have over it is an illusion, propped up by laws that are easy to evade.
By way of comparison, look at the unreal marketplace or the unity asset store. It's easy to buy art on those sites. But almost nobody on there makes enough money to live off of. At the moment, I can only think of 1-2 creators on there who I think can live off their earnings, and they all live in parts of the world with low costs of living
IMO, the only value in selling assets on those stores is to attract the attention of companies who might want to hire you for real work
It's not viable to have an art company without some way to drive repeat transactions. Subscriptions are the only thing they can come up with that keeps the companies afloat and able to pay people's salaries
Photography is a valuable skill, but stock photography and footage have been too overvalued for a long time. I'm not surprised that they're having a rude awakening in the face of things like Stable Diffusion. For most purposes, stock photos and illustrations don't have to be perfect, but just good enough. Or good enough for an amateur to make finishing corrections to.
That said, it also seems like the industry needs to be disrupted. Not by AI per se, but by competition that wants to sell itself as stock photos without the bullshit.
I think it's too soon to see real data on this because the current leader in AI imagery is still quite limited in what it can produce. There are too many examples of stock photos where no combination of tokens is going to render a similar picture, there are still very prominent and noticeable limits on what AI can produce.
When creatives are seeking out a stock photo, they're not looking for a piece of generic clip-art to fill a void on a page. They're looking to satisfy far more prescriptive set of requirements, the sort of thing that should really be shot anew, but budgetary and timing constraints typically mean that a pre-shot photo will have to do.
As AI improves it'll make a bigger impact, even right now in its limited form, it's not looking good for junk-stock peddlers, which will disappoint absolutely no one.
I sure hope so. I get that people like to take pictures, but paying for stock photos has always felt really bad. Thank god for AI.
Paying for something like wedding photos or art photos never felt bad for some reason. I think the work for photographers will really change the coming years
229 comments
[ 3.3 ms ] story [ 259 ms ] threadI don't know if AI will replace stock photography, but I wouldn't bet against it.
The ability to charge a high price for stock photography is over though.
Dall-E hasn’t improved since it was released, and the other good models like Google’s Imagen weren’t made public…
- It runs in the cloud with more VRAM available than on consumer machines, so it's able to use more parameters and custom layers. This makes it more coherent.
- It's finetuned on a large RLHF dataset, which they have from the users voting for bot-generated images in their massively popular Discord server.
- It includes a decent share of custom processing and manual tuning (system prompts, custom layers trained on god knows what etc)
Their closest contender, Stability AI, is currently busy training SDXL, their new model, on a RLHF dataset they are collecting on their much smaller Discord channel. So Midjourney is SOTA for text-to-image.
There's DeepFloyd IF which is large enough and includes a 11B LLM for natural language understanding, so it has much less issues understanding complex prompts such as "koi fish doing a handstand on a longboard rolling down the slope of a moon crater". But it's not a finished product (no complex pipelines, no RLHF), and the license is restrictive.
That said, Midjourney's strengths are also its weaknesses. Midjourney is good for the text-to-image, but not much else, and image generation is much more than just text-to-image. MJ gives great results off the bat, but struggles with unusual requests, complex compositions with multiple objects, and can't use various elaborate trickery available in the much smaller Stable Diffusion (which is mediocre by default but has powerful higher-order guidance and training, and is also much better at complex things). Midjourney is also censored - you can't have Joe Biden eating a slice of porcupine.
The 5.1 raw model is simply unbelievable for photo realistic images. There is 4a,4b,4c, 5, 5.1, 5.1 raw, the anime model I never use. All are interesting but how often do you read anyone mention what midjourney model they are talking about? Basically never.
There are certain circumstances that I have seen stable diffusion match the output but overall I don't even think it is close.
Personally, it wouldn't even occur to me to look for stock photos now as opposed to just making variations in midjourney. There is a bit of a learning curve with so many options in midjourney so I take anything I read about it with a grain of salt. Most people are using these tools to make absolute shit.
I can certainly appreciate those who take stock photography (and I've seen some of them in my travels - taking 100s photos of slightly different poses of a model at a location because the stock buyer could say "I like this one but the arm is in the wrong spot for the page layout" and go with someone else's photograph).
Being a stock photographer has never been easy. In the days before digital it represented countless rolls of film shot (and processed at the photographer's expense) and now its even more digital images to ensure that you've got just the right one for some client a few years from now.
It's not an easy job and no one has gotten rich off of it.
And absolutely, however creativity is the tip of the spear which pushes beyond constraints to develop novel AND meaningful style (like Beksinski, Kopera, Kostetsky, Moebius, etc).
Nice try.
Note that the same can be said about the negation: show me a proof that it can't.
I guess my last response was too hard for you to understand since you respond with this. Perhaps you are new to things like "reading between the lines".
You are the one who made this claim:
> That is an incredibly strong assertion that evidence does not seem to support currently.
To avoid loosing more time, I'll write what you failed to understand: "That evidence does not seem to support currently" is your claim, something you can prove, shows the evidence you are talking about.
You still didn't send any link to the evidence you were talking about.
Proving that we can reliably simulate human brains requires showing that nowhere in any human brains does such signals ever emanate. We can get close if we one day have the capacity to run a simulated brain for long enough to show it appears to function like a human, but when the counter hypothesises a violation of known physics that may well be intermittent and extremely limited, we can't realistically absolutely prove its absence.
As such, the former is tractable if such signals can occur, the latter is not, and so I think my comparison of it elsewhere to a claim of Russell's proverbial orbiting teapot is reasonable - people will always be able to claim that there is some so far unobserved difference, and given it postulates highly localised violations of known physics and this seems like an absolutely extraordinary claim, the burden must be theirs.
My point is that to make the claim "human creativity" can be simulated surely would have the burden.
Just one measurement that doesn't fit would be enough to shift the burden. Absent that, any suggestion we can't is no more than a religious belief.
It'd take a truly extraordinary find that'd upend both physics and areas of logic for that assertion to be false.
The definition of creativity itself should include the ability to subvert a system through abstraction. AI is a system so it cannot subvert itself (or, it hasn't yet). As humans, we subvert systems in mathematics by creating higher dimensional abstraction and shattering previous understanding of the fields. Often, mathematicians achieve this with a "lightening bolt-like realization" that they do not attribute to systematic thinking.
Please reconsider whether systematizability should be so quickly assumed through napkin philosophy. If you want, prove with a topological map of the creative apparatus in the brain that it produces results in a systematic way. That would at least scratch the itch and get the conversation going somewhere interesting.
What you call "systematizability" is irrelevant. Only whether or not a function can be computed matter, and if a function can be computed by a brain restricted to known physics, there is no reasonable theory under which it can compute anything which can't be computed by any universal turing machine. Given the absence of even any real hypothesis for any category of functions which can't, it'd be an absolutely extraordinary claim to suggest such a function exist.
Where you'd get a Nobel in physics if you could show unknown physics in the brain, you'd secure a Nobel in mathematics if you could come up with a class of functions the brain can compute which can't be computed by a computer, because it'd fundamentally break basic assumptions in any number of subfields of both maths and computer science.
What you describe as "subverting itself" does not require anything as drastic. It merely requires thought processes unobservable by the person in question, and processes to complex to casually understand and/or randomness. Given a large proportion of people don't even have an inner voice they can "listen to" in order to observe their thought processes other than via outputs, and we know e.g. from split-brain experiments that the conscious part of human reasoning often outright invents knowledge and observation of thought processes, we have no reliable way of analysing the extent to which thinking is "systematic", and plenty of evidence to suggest we can not trust any such notion.
The creative generation part - adding some randomness to the process - is trivial with machines. Much easier than with humans adults, who often have to learn to allow more randomness and in their thinking/feeling, in order to make some creative work. But the selection functions are deeply social, which is something we're instinctively good at, and had little luck teaching the machines so far.
Well, with the possible exception of GPT-4, which may have learned just enough about what humans like, that it could do a passing job at discriminating art/innovation from noise. I say may - it feels possible, but we won't know until someone properly tests for it.
I've tried giving GPT4 some axes to rank writing on, and feed it some paragraphs. It's not perfect at it, but it's a far better critic than writer.
E.g. a while back I had it write something "in the style of Douglas Adams" and then had it critique it in another chat, and it recognised the imitation and proceeded to quite harshly judge its own writing for being flat and lacking direction and purpose. It was quite entertaining (more so than its superficially amusing Douglas Adams imitation).
The notion that creativity lies beyond what is systemetizable (or otoh, the value in creativity is to defy that which can be computed), seems more in line with my sense of it.
[1]: https://en.m.wikipedia.org/wiki/AI_effect
Surely you agree that we care more to prove the affirmative of whether a machine is conscious rather than prove that it isn't. If not, I don't think we can proceed.
The issue is that the AI Effect you linked is often actually used in the reverse, including by yourself, where the burden of proof should be to prove a machine is NOT conscious. However, like we've agreed, we care for proof in the affirmative.
Thus, the paradox lies in that the "AI Effect" concept is critical of those demanding the proof which is more valuable.
And you're incorrect - you cannot "empirically prove creativity". As a professional artist I don't think you quite understand that the type of creativity we are interested in is not just "permutation", there is clearly a human component.
I'm trying to show you that you are falling into the reverse of the AI effect by assuming that permutation is creativity.
Do humans even have creativity at this point or do you require a formal proof even for that? Ahah
>As a professional artist
Now I understand where the friction against accepting that even machines show signs of creativity comes from, I imagine you're an illustrator, so the feeling of becoming obsolete is the bias against machines.
Also, gun-to-my-head I would not bet on creativity being systemetizable, so I completely disagree.
Where you're going wrong is what society will go for. Given the choice between reality and fantasy ... society will choose fantasy. AI will "win" fantasy, hands down.
Not to mention, you are already wrong because art communities around human art are thriving now more than ever for the simple fact that the result of AI itself is non-communal (does not produce thriving community) and cannot compete at the bleeding edge of style.
Again, this is already happening regardless of your armchair hypothesis.
Your assumption is IMO unfounded, it actually seems like you have the burden of proof given what I said above. Just because it seems more in line with your sense of it doesn't make it true, seems like you think creativity is some mystical nonphysical force (because if it were physical, it would be able to be modeled and computed, as the parent says).
Edit, I see, you're an artist so it's likely why you believe in creativity as a nonphysical force. I've noticed that many artists say this, with no evidence, while most engineers say the opposite, that it's systematized, likely that's why we're in our respective fields.
Creativity: producing what is novel + valuable
Systematizable: can be enumerated via algorithmic operations
"As others have mentioned": that is not an argument
"not everyone else who already knows": also not an argument
Now, my point is that creativity as defined above requires a burden of proof to be defined as systemetizable. For instance, mathematical leaps are often achieved by a spark of "creativity" which breaks existing structures in a way which is difficult to grasp, and to call this systemetizable is a leap that requires a complex argument calling upon deep notions of computational theory and computational neuroscience (like producing a topological map of the creative structure of the brain and proving that it can be mimicked by a listable process).
Now, to apply this to art in a convincing way would surely be even harder. Do you at least understand my point? Or is "brain = physics = creative" satisfactory enough for you still? Thanks for reading in good faith this time. :)
Delineate? Diffusion models can't be further away for delineating. They literally work by throwing random shit at the wall.
Input set? There is no input set once training is complete.
Human creativity? In stock photo industry? What next? "How I write while loop instead of for loop and I have achieved the nirvana?"
You seem emotional, "you sound stupid and nothing what you said makes sense", I wonder, why are you so triggered by a challenging idea?
The very point of computation is to evaluate clear, distinct, mathematical expressions (like reading a list of algorithmic steps to produce an NN).
Hmmm. Did I threaten your religious belief in the power of AI? :)
We need to find a way to distribute the benefits of such developments.
Until then, some sabotage may be in order.
Seriously though, I used to think this was mostly a done deal that's being occasionally brought up by people fascinated with socialism or something. But the longer I live, the more I realize that the issue has not meaningfully improved at all - we've only been distracted from it. And it's much, much older than Marx and co. too. This, and not some general fear of technology, was the whole point of the Luddite movement (very relevant in context of AI). And you can probably find examples stretching back way before industrial revolution too.
I don't know if OpenAI has actually confirmed what hardware they use or if the press and bloggers did citeogenesis on each other, but the claim is they're using A100 GPUs, and the bigger models have to use several in parallel. I've seen detached houses in Eastern Europe going for the price of two of those cards.
The DGX-2 is (was?) priced more like a detached house in Western Europe; the GH200 they just announced appears to be "if you have to ask, you can't afford it".
This is similar to how, despite that lathes, kilns, and chainsaws can be bought in supermarkets as impulse purchases, industrial versions of each are rather more expensive.
That’s the natural end of capitalism, as has been noted literally since the word “capitalism” waa coined for it (largely, for that reason.)
It is end to which every tool and technology is put under that system, and has nothing to do with the particular technology and everything to do with the system.
No evidence of that.
The rate of return on capital is what dictates the concentration of wealth, or conversely, the dispersal of wealth.
The conceptual systems introduced by Marx and others is that only workers can own capital and are based on a set of assumptions about Capital returns versus labor power.
In the same sense that the US electoral system doesn’t preclude more than two viable parties. The system of property rights that defines capitalism doesn’t specify any limit, but the structure it creates promotes it in function.
> The conceptual systems introduced by Marx and others is that only workers can own capital
No, the conceptual system introduced by Marx rejects the concept of ownership as applied to capital [0], though it favors workers exercising control of the capital to which their labor is applied. Different variants derived from this and related ideas have different specific mechanism for this exercise of control.
[0] When Marxists talk about rejecting/abolishing “private property”, it refers to this, because in the Marxists terminology “private property” refers not to property of non-state individuals and entities in general as it does in current common use, but specifically to property rights in the means of production, i.e., capital.
That said, I don't see you the distinction between controlling the means of production use for one's labor and owning it, when controlling the means of production also includes allocation of the additional value created using them.
It also is also worth noting how far current thoughts on socialism and communism have come from Marxist theories. contemporary sentiment seem to advocate that individuals (working or not) are entitled share of the labor value created by others, even if they are in another workplace.
All this is really an aside from what I thought my greater point was about the underpinning assumptions or conditions that lead to the "expected" systemic outcomes. The US federal electoral system currently favors two parties, but this could change with different conditions. For example, ranked choice voting, and proportion state electoral allocation could impact this without overturning the entire governmental system.
Similarly, the concentration of capital is based on an number of conditional assumptions. For example, if the real rate of return on capital is less than the rate of capital creation, you would not expect the concentration of wealth to increase. under this condition, capital investment could only serve to slow the losses due to inflation.
Conditional control while you are a laborer whose labor is applied to particular capital (which, to be fair, is one of many patterns advocated within Marxist-influenced socialism, not the only operationalization of the concept) is distinct from ownership which persists until the owner chooses to alienate the property.
> It also is also worth noting how far current thoughts on socialism and communism have come from Marxist theories. contemporary sentiment seem to advocate that individuals (working or not) are entitled share of the labor value created by others, even if they are in another workplace.
Yeah, its not like Marx said anything like: “In a higher phase of communist society, after the enslaving subordination of the individual to the division of labor, and therewith also the antithesis between mental and physical labor, has vanished; after labor has become not only a means of life but life's prime want; after the productive forces have also increased with the all-around development of the individual, and all the springs of co-operative wealth flow more abundantly—only then can the narrow horizon of bourgeois right be crossed in its entirety and society inscribe on its banners: From each according to his ability, to each according to his needs!” [0]
[0] Except, of course, that he did, in Critique of the Gotha Programme, 1875.
It’s called free software and making models at various ‘weight classes’ to support many GPUs. Basically don’t let companies like OpenAI/Google/Microsoft run the show and, ironically, support companies like Facebook who open source their models.
And yes, I know OpenAI has open sourced models in the past, but the fact that they won’t open source ChatGPT, GPT-4, and select other models like DALL-E 1/2 make me wary of them.
Market disruption due to technology/innovation/progress is nothing new. There have always been winners and losers but the aggregate wealth/value continues to climb. Devising ways to mitigate the loses seems like a more feasible approach than pretending that we can prevent these sorts of changes through regulation or legislation.
This is just an accentuation and further acceleration thereof.
So there is precedence for this scenario.
You can rent the hardware from someone else, but that's a matter of convenience, not affordability.
Hello post-industrial/ai/ml capitalism.
This development will do to designers, programmers, support what the industrial revolution did to cobblers and weavers ...
Which made cheap shoes and clothes available to everybody.
If you want to buy a tailor made shirt, it'll cost you at least $500. And that's still with machine made cloth. What do you think it would cost if the cloth was made by hand by a weaver?
I'd argue the efficiency in production led to lower prices - while the concentration in wealth mostly kept buying power/living standards lower for the workers (or; capitalist class reaped an unfair share of productivity gains).
I believe commodisation of "knowledge" work will lead to lower prices (and unfair distribution) too.
Go take a look at Fort Williams, and at the clothes in it they preserved. They look like kids' clothes. The same for Civil War uniforms. Larger than the FW clothes, but still looking like kid clothes.
I flipped through a high school history book a few years ago, and it insisted that the middle class in America didn't emerge until FDR. This is laughably wrong. It was there from the beginning.
Painters are trained by copying the Masters. Shredding guitar players are trained on violin caprices. Martha Stewart was trained by watching Julia Child.
AI art is more like what would happen if you took Baroque art and applied it to modern styles. For a human artist the effort is not worth it.
The proper analogy to ML is that the owners of the latter train a slave, and then use that training process as justification for owning everything[1] the slave ever produces.
How can freemen compete with that?
[1] Which also happens to be a perversion of copyright law, since AI can't enter into a contract that can assign rights to creative output to it's owner.
I took pictures of a much wider variety of things than I looked at pictures of. Instagram and other sources of millions of pictures did not exist and I was perfectly capable of taking a picture without that “training”.
You’re missing the forest for the trees here. The point is that a photographer doesn’t learn how to produce pictures of a screwdriver by looking at pictures of a screwdriver. A photographer learns how to take pictures generally and then can instantly be capable of producing pictures of far more objects than he/she has ever observed.
Generative AI is not capable of that because it’s not operating a camera in reality.
https://www.shutterstock.com/press/20435
Probably AI will help photographers to cut costs, I mean by taking less “real” photos and using the AI to generate other similar variants, example: “Photos Used to Generate AI Images for Client So Photographer Never Shoots Again Photos Used to Generate AI Images for Client So Photographer Never Shoots Again” https://petapixel.com/2022/12/21/photos-used-to-generate-ai-...
It's not exactly what you've said, but it itches a scratch from around there.
Some photographers just spam TONS of images on the platforms to get some cash by betting on the long tail. Someone, somewhere with a shutterstock subscription might need your image for a local magazine/paper and you get 2 dollars for it.
That kind of stock photography might suffer even more with generated images. Actual premium stock or for hire photography will not.
> Someone, somewhere with a shutterstock subscription might need your image for a local magazine/paper and you get 2 dollars for it.
Yep, that's how a lot of it works.
> Actual premium stock or for hire photography will not.
I kind of wonder about that. When I can upload a picture of my product and prompt, Photo of a scene with my product as the subject on 10 N Main street in Wherever, usa. I'm going to be a lot less likely to ship product (and get it lost or broken), commission a photographer in Wherever USA to go take the picture, wait 2 weeks for the right weather, only to find that some unreleased person's face in the reflection on the side of the product.
For that kind of thing - just explaining an idea - midjourney is amazing. So much faster than making something in photoshop and so much more accurate than google searching.
I think the real issues for stock libraries is when midjourney starts replacing actual output. This is starting to happen but still very early. It’s gonna get a lot worse!
As things stand, I would still probably use stock for "official" company marketing materials although I understand that some companies, especially smaller ones, are probably less risk adverse.
Yes. Selling stock photos for crappy ads will be dead in a year. Unless you're at the level of Ansel Adams or Margeret Bourke-White, your content is not worth much.
Propaganda and need to sell things are powerful motives. And AI generation and modification are great tools for that. Authenticity sell to very few.
I think we should expect a flood of "masterpieces" in different channels. Including news reports, scientific papers. The solution could be closed humans only groups. But, like in chess today, it's hard to say who is cheating.
One domain where I welcome AI is computer games. Having more natural flexible characters will make games more interesting. As opposite to current pre-written texts.
Also education, where kids can ask questions and give the answers. Not just read or listen passively.
Offloading AI stuff to a cloud would make single player games unplayable offline and their servers won't stay online forever. Another solution would be to use a second GPU just for AI stuff. But that would make most current gaming devices useless.
On the bright side, compute power is probably going to continue increasing. Consumers might be able to run current LLMs on home hardware in a decade or two.
GPUs will probably get a PhysX-like API for these kind of AI calculations in the future.
Not sure how parents feel about putting kids in front of a machine that can fabricate lies.
I mean a encyclopaedia getting information wrong is one thing - garbage in, garbage out - but one that actively invents lies is like added another layer of bad to things.
Probably better for kids learn to search for information and evaluate information sources.
You mean like a human teacher?
Of course, I agree with your point, but worth remembering that teachers can and do (knowingly or unknowingly) give incorrect information to their students.
I think one big difference with ChatGPT etc. is, as far as I've seen, it will pretty much always give a confident sounding answer (unless it's on a banned topic) rather than saying something like "I don't know", which a (good) human teacher should do rather than just making something up.
This reminded me of one of my teachers who said that Cuba was a great place to live and the rest of Latin America should be more like them. She was essentially a communist who painted the old Soviet Union and communist countries in general as much better than the capitalist west.
This will never happen, promise you that a major CGI sequence will ship completely AI generated in a Marvel movie within 3 years. This tech will be so integrated into workflows of Disney and the like that the idea of an "AI Generated" watermark will be as ridiculous to them as a "3D Rendered" watermark and if any legislation comes about they'll be against it.
But I'm sure some people with expert darkroom skills had similar thoughts when photoshop was rising to prominence.
Consumers don't care what tools were used to make what they produce. "AI generated" mark is something only the people being made obsolete will want, and their power is rapidly declining.
Stock is still the easiest, safest, cheapest option - VS your own shoot.
So I don’t think that means it’s future proof in any way.
On the contrary it means that the same cost-saving decision will be made again and there’ll be a move to AI images when they are slightly more mature - ie when the legalities of copyright have been fully figured out.
(I work at a big ad agency and I know the copyright thing is currently a massive drag on potential usage)
Can you elaborate on what you mean by "safest" vs shooting your own?
I’ve seen photo shoots costing hundreds of thousands of dollars not being used at all.
even multi million TV campaigns made and never run. Shooting is hard!
Where previously you'd take 100 photos to choose the perfect one, now you'd include all of them, with labels that describe their quality or deficiences (which you do anyway as you curate the photos you took). This makes almost every shot the money shot.
Less glamorous for sure, a lot less money per photo, but more volume for virtually the same amount of work. And you can always treat the really unique and special photos as today and not include them in your AI dataset.
Combine that with the fact that a photo (and photoshop) pro can produce better results with AI image generation (because they can guide them better than people who don't know what they're doing), and the future doesn't look so bleak.
Eventually, all the "photographer" will need to do is to take a 8K video recording of the subject and AI will generate a set of great looking pictures (it already knows what people like).
Another issue with using videos is overfitting, because video frames look too similar to each other. So you can get an inflexible model that always generates a similar background as that in the video.
Good AI training dataset is the opposite of that. It needs to contain different backgrounds and different lighting conditions. And as many different angles and poses as possible. Can't really get that from a single video.
What the AI/model then needs to do is to extract the subject - person specific features and combine that with trained data. For that, artifacts/motion blur isn't a huge problem.
Without that, computer-generated images would be far uglier and weirder. And if the computers had to get their new material from, say, a bunch of webcam feeds, the composition of their output would be terrible.
A single image "comes from" millions or even billions of pieces. It truly works like that (albeit counterintuitive). It doesn't just combine 4 or 5 images and apply a filter.
I saw mentions of companies being reluctant to use AI generated images because of potential copyright issues. What issues are they talking about?
But even an image is made of a lot of pieces*, it's still not 100% "legal-concern-clean". Consider this hypothetical scenario: you collect one million different photos of McDonald's logo, and one million other random photos. Then you train your diffusion model and tell it to generate a logo. Of course it'd be still very similar to McDonald's logo.
SD's training data is supposed to be much more diversed than that. But how could we be sure? That's where the fuss comes from.
* (pieces of information, not pixels)
If I produce ad content for a company whose public image is incredibly important (e.g. Apple) and that content includes AI art what is to stop a competitor, or anyone else, from reusing that same art in a counter ad or to produce other content to similarly harm the original company?
On the other hand, if I were to use images from a stock photo company I could simply purchase an exclusive license the artwork used (and likely the entirety of the similar work from the shoot, used or not) to avoid this type of imitation.
For some uses of stock art, probably not (nonexclusive use is common and cheaper), for others yes, if AI art is not copyrightable (an opinion of the Registrar of Copyrights that it is not if the only input from the human claimant to the finished work is a text prompt is perhaps being grossly overread the way some people seem to think it is settled law that imagery generated through the use of an “AI” tool is categorically uncopyrightable.)
No, its not, becauase the statement at issue covers only what is (in the Registrar’s interpretation) definitely outside of copyright protection, it does not do much to clarify what it takes outside of that clear exclusion to qualify for protection. To assume any image where the human work is more thane exactly a text prompt is protected is as much reading too much into it as assuming all AI-generated images are completely excluded.
Control-net is pretty much the best thing ever if the first shot didn't work.
In comparison, typing a prompt into a generative AI is a superior user experience. I'm willing to pay a photographer, they just don't make it easy.
> All of the services are subscriptions now and they have dark UX patterns
They do. I have a subscription to one of the services, and it's just super easy and I use it enough it's worth the monthly. I'm not sure there's a huge market for licensing an image occasionally... It would be interesting to see the breakdown there. Also, Adobe has a huge advantage because Adobe Stock is highly integrated into Adobe's design software.
Maybe for an iconic newsworthy photo.
But, yeah. As an individual, paying $50 or whatever for a photo for a book cover or something like that would be a rather rare thing. And if I'm a company who needs a steady stream of stock for articles, corporate presentations, etc., I'm almost certainly going to have a subscription if only for the convenience of not having to get individual purchases approved all the time.
No, it takes me several minutes to get my lazy ass off my chair, go find my wallet which is inside some pants inside the laundry basket, pull it out, find my credit card, enter the number in, enter the billing address, enter the zip code only to find that Google autofill has now decided to edit my address, then go edit the address, only to find that Google autofill has decided to edit my city, go edit the city, then just go "screw this" and use Stable Diffusion.
If you want to minimize friction for payments (a) don't display any GDPR or newsletter popups (b) don't ask me to register an account or verify an e-mail address or phone number (c) just let me hold my credit card to the webcam to pay. Do NOT display 17 text boxes to fill in.
It took around 5 minutes for me to even be allowed to download the free low res version of an image. Between forced signup (by clicking the “free download” button, with no indication of needing an account), waiting to confirm my account via email, and having to return to the page after forced redirects begging me to pay for a subscription I’ll never use, all for a single image to play with in GIMP.
I genuinely hate how predatory UX has become in recent years.
And even more of them I'd turn off only if I knew with certainty that I could turn them back on whenever I want, and that their memories would simply continue where they stopped.
Every person is capable of murder, that's why murder happens. It's a natural thing, in nature, creatures die and kill every day. The fact that you consider yourself so outstanding and morally erect that you are completely sure that you'd never kill a person and have never wished to only tells me that its likely you had never had any real introspective thought in your life.
>And even more of them I'd turn off only if I knew with certainty that I could turn them back on whenever I want, and that their memories would simply continue where they stopped.
Certainly a strain of solipsism I've never heard before now, but still pathological.
Though I’m not sure what place the US Supreme Court’s opinions have in rational discourse at this point.
It doesn't matter if it's my eyes and brain, the computer working for me, or the eye of Sauron for that matter. Once the cat is out of the bag, any control you think you have over it is an illusion, propped up by laws that are easy to evade.
IMO, the only value in selling assets on those stores is to attract the attention of companies who might want to hire you for real work
It's not viable to have an art company without some way to drive repeat transactions. Subscriptions are the only thing they can come up with that keeps the companies afloat and able to pay people's salaries
But yeah, I sure don't like them as a customer
That said, it also seems like the industry needs to be disrupted. Not by AI per se, but by competition that wants to sell itself as stock photos without the bullshit.
When creatives are seeking out a stock photo, they're not looking for a piece of generic clip-art to fill a void on a page. They're looking to satisfy far more prescriptive set of requirements, the sort of thing that should really be shot anew, but budgetary and timing constraints typically mean that a pre-shot photo will have to do.
As AI improves it'll make a bigger impact, even right now in its limited form, it's not looking good for junk-stock peddlers, which will disappoint absolutely no one.
https://twitter.com/drivenbyboredom/status/16618473784276090...
Paying for something like wedding photos or art photos never felt bad for some reason. I think the work for photographers will really change the coming years