Anyone can create AI imagery but right now it takes time, practice and judgement to create good AI images. And that's without editing the images to make them even better.
I guess your argument could be changed slightly to "Why buy stock photos when I can take a photo with my phone?"
the difference is, its not even remotely easy to snap a photo even halfway decent stock photo of an office space with random people, especially with so many people not working from home.
it's pretty easy to tell some generative software to "Female giving a presentation in a well lit office with windows open" and get something pretty decent.
This goes for anything that needs props that a normal person doesn't have.
> it's pretty easy to tell some generative software to "Female giving a presentation in a well lit office with windows open" and get something pretty decent.
Is it? I tried that exact prompt with dall-e and got 4 images with nightmare fuel faces: https://imgur.com/a/yyt5Tax
>it's pretty easy to tell some generative software to "Female giving a presentation in a well lit office with windows open" and get something pretty decent.
As someone who actually tried it, with current models it's possible but neither easy nor straightforward. Getting decent and coherent images is still hard and requires finetuning, so there is certain value in this. And for very specific and series-consistent images I'd still rather set up a photo session if I had a studio.
It will change for stock photos, but probably not very soon.
> it's pretty easy to tell some generative software to "Female giving a presentation in a well lit office with windows open" and get something pretty decent.
Tell me you haven't tried Stable Diffusion without telling me you haven't tried Stable Diffusion.
There is some skill involved but once you know how to use the tools it does not take long to generate a good image and upscale it. If a good stock photo exists I think I'd prefer that, but the real power in AI is to avoid hiring someone to create something original.
I agree that anatomy and hands are a problem. Give it 6-12 months and they won't be. That's how fast things are moving at the moment!
I think what is needed is specialized models. Say that you're in a business in which you just need pictures of residential interior designs and nothing else. A model which is just trained on billions of images of interiors is going to work better than one trained on billions of images of everything.
Agreed, but strictly speaking the model would probably still be based on a general model such as Stable Diffusion, then "fine-tuned" to render only interior designs. For example the NovelAI anime image generator was based on SD then fine tuned on anime art, and it works incredibly well - it can generate anime art of basically anything you ask, because the model has the wide knowledge of the Stable Diffusion model but is fine tuned to the anime art style.
"female with two hands that have five fingers, and normal looking face, giving a presentation in a well-lit office with correct shadows, and a clock on a wall that is perfectly round with legible roman numerals correctly placed around the dial ..."
Because if you're looking for something specific, or just a great level of quality (in situations where details matter), you will have to generate thousands of AI images and review them manually to get to a shortlist of a handful of good ones to choose from.
Or possibly you would have to have your own model which is trained on that space of images that is relevant to what you are looking for.
Look at what virtually all the hobbyists and enthusiasts out there are doing with this stuff. It's almost all exlusively these: surrealism, fantasy and dumb memes. In all these fun genres, bad details, like hands with six fingers, or laughably wrong shadows, do not matter.
> "The data we licensed from Shutterstock was critical to the training of DALL-E,” said Sam Altman
So there was legitimate Shutterstock data used to train DALL-E. I don't think I've read this before. The images with Shutterstock copyright watermarks were presumably scraped accidentally from the web.
The contributor fund is also setting a precedent that data used to train a model has a claim for payment.
> Shutterstock is launching a “Contributor Fund” that will reimburse creators when the company sells work to train text-to-image AI models. This follows widespread criticism from artists whose output has been scraped from the web without their consent to create these systems. Notably, Shutterstock is also banning the sale of AI-generated art on its site that is not made using its DALL-E integration.
It's also possible that they licensed the images after the fact so they wouldn't get sued, and what Altman is describing _are_ the watermarked images. It's possible Shutterstock got very upset with them and they worked this out quietly, outside of court.
> When asked about these issues, a spokesperson for Shutterstock told The Verge that there were “lot of questions and uncertainty around this new technology, specifically when it comes to the concept of ownership,” but that the company’s stance is that “because AI content generation models leverage the IP of many artists and their content, AI-generated content ownership cannot be assigned to an individual and must instead compensate the many artists who were involved in the creation of each new piece of content.”
As far as I know, this is Shutterstock opinion and there is no judge rules or new laws addressing the situation.
Shutterstock don't mind stealing others work and selling it as their own. These AI ownership questions are trivially resolved with their way of doing things.
The human mind has been doing this forever. Taking inspiration from others and creating new art.
As far as I know it’s even legal to study the style of one specific artist and create imitation art as long as it’s clear that this is a new artwork from a new artist and about ~50% different from any existing works.
So a prompt like “a photo of Times Square in the style of Ansel Adams” should be ~50%+ unique.
The trickiest area for these models might be prompts that recreate existing work.
Given a prompt like “a black and white photo of Yosemite Valley in the style of Ansel Adams”, if the model is too good, then it seems like copyright infringement.
But as far as training data goes, as long as it’s acquired legally, I don’t think it can (or should) be regulated.
You claim "difference", but how can you say such things when we have no idea how human brains mix and match different sources of inspirations? I've definitely composed a succession of melodies only to turn around and realize I'd just copied an entire chunk of a commercial jingle I'd heard in the background a while back. What is the precise definition of inspiration?
Known and unknown are relative. I know my first name. I don't know yours. Your theory suggests that the two names will certainly be different. Is that always the case?
The mechanisms of human brain is indeed unknown to us. So how can we say with confidence that its underlying workings are this way or that way with certainty? We can only observe the inputs and outputs and do our best to compare the two. Which is the whole point of AI - to simulate the human brain (or what we can observe about it) to the best of our ability.
The work of art was 'created' by an anonymous author, the AI. Title 17 explicitly handles works by anonymous authors. I can't just claim them if I didn't create them. If I can claim I own works created by my code, and the code is the artist, then Adobe owns every work created with it's tools that does not have a seperate copyright claim. Which Adobe obviously doesn't. This is not new or novel, it's just like with Napster, the current law in inconvenient to what people WANT to do, so they claim the reality they want. If the work is unique enough to claim copyright, then the algorithm added the uniqueness/copyrightable characteristics, not your writing the algorithm, and therefore the AI is an anonymous Author. If it is your algorithm that is doing the creating, then Adobe owns the copyrights to the work generated using it's algorithm, which it obviously doesn't. I am missing the gray (grey) area here.
Are the resulting images copyrighted in the name of whiever generated them? What proves that this generation happened first - Shutterstock upload?
What is to prevent someone from generating that same image with a prompt?
What would prevent them from simplu resharing that image freely with everyone, or undercutting shutterstock? If Shutterstock tries to sue, the courts may say AI generated images do not enjoy copyright protection !
How is this not just a marketing stunt to try and keep them looking relevant in a world where the most bargain basement stock photography is entirely commoditised?
If I were Shutterstock, I'd turn this into a service to generate instant stock photography for any situation you can describe -- for a fee, of course.
Yes, "anybody" could just spin up a Stable Diffusion server and do the same, but plenty of companies would be perfectly happy to outsource this to Shutterstock and, quite importantly, also have them take all the heat for any legal complications behind reuse, copyright, etc etc.
> If I were Shutterstock, I'd turn this into a service to generate instant stock photography for any situation you can describe -- for a fee, of course.
The problem is that with the amount of prompt tweaking one needs to do, it doesn't scale
That may very well change in the next X years, but right now it doesn't really work as a business
DALL-E and such is often not good enough yet to generate conceptually specific stock stuff (just try to generate nice looking cartoonish illustrations that look natural); as someone who uses stock photos/illustrations pretty often, just looking up human creations is often still easier and faster (and not that much more expensive). And if they offer both AI and non-AI generated stuff, isn't the whole point for them to stay relevant as the main platform for getting stock photography.
Fair question. For something as simple as "women laughing with salad", totally, you can generate yourself. But, for anything less concrete/more conceptual, I think needing to explore various options will remain. Often when you are looking for a stock photo you don't have an exact idea of what you want in mind - just something that might fit with the article/concept. So you search a few different things on Shutterstock and see what kinds of stuff exists and looks good. The question is, is it going to be quicker and easier to explore the prompt space (try a prompt, see what happens, revise, see what happens...), or will it be easier explore the space of uploaded photos/generated media by other creators by just searching on a site such as Shutterstock. Basically I think there is still going to be a market for people who just want to find stuff that is generated instead of generating it themselves.
I was thinking the same about widespread trucking jobs 5 years ago.
There's a recurring theme in neural networks. You can go from nothing to 90% in a year. And it's amazing. That naturally leads you to believe that perfection is little more than a bit more compute and a bit more massaging the data.
Then over the next year you manage to get to 95%, further affirming your beliefs. But from then on every single fraction of a percent becomes exponentially more difficult. And it soon becomes clear that there's an apparently insurmountable asymptote long before you get anywhere near 100%.
Trucking is all or nothing - once self-driving cars are working and legal, those jobs are gone.
AI art is not - jobs are already being lost in real time, users can create stock photos, logos, bespoke art right now. And more and more jobs will be lost as the AI improves.
At risk of exposing my rather shallow knowledge of the topic, one aspect of PoMo is "everything is an inauthentic quote of something else and everything has already been done to death - so embrace irony and self-reference and achieve sincerety by self-awareness of your own insincerity".
To some extent this already happens right? Popular fiction has its own laws for pacing, style of dialogue, tropes, the way characters behave etc. that is highly influenced from the canon of fiction before it and totally different to real life. Readers expect and anticipate this.
Good question. Had exactly this conversation yesterday. There were two
groups of opinions. Positive feedback usually does one of these two
things:
1 - It will deflate
2 - It will explode
Deflation is where the output of system tends toward a fixed point (AI
art gets blander and blander). As it feeds on it's own output (which
becomes the dominant training material because its cheap and
plentiful) each vector is compressed.
Chaos (explosion) is where the output gets less predictable and more
crazy on each iteration. Who knows what emerges before the system
breaks.
Positive feedback systems are generally bad. For example, in
biology/food; we started feeding cattle on the rendered down remains
of other dead cattle. The result was Bovine Spongiform Encephalopathy,
a lethal prion that thrived within the feedback loop.
Sell creators content to train DALL-E without consent. This is worse than just scraping without consent. After earning money to help train DALL-E, you earn again by selling the output of it.
But no worries, an unknown part is shared with creators.
Two revenue streams without consent but they still present themselves as the ethical AI company.
And it's not even clear what a customer is buying. As far as I know, AI art can't be copyrighted.
There is a common article that gets passed around where a person tried to get copyright assigned to a machine and that keeps getting denied. That is very different from "AI generated images cannot be copyrighted".
If I write a program to randomly generate an image of squares of different colors:
- I, a human, can have copyright over the program text itself as a literary work
- I, a human, can have copyright over the image output of the program
I can't claim copyright on one of your works, because I'm not the author.
I can't claim copyright on a work by an AI, because I'm not the author.
AI created works, like animal-created works [2], lack authorship, and are thus not eligible for copyright at all.
It's being litigated, and might change, but this is the current position of the US Copyright Office [1].
The only AI generated copyrighted work at the moment is a graphic novel, Zarya of the Dawn, and was granted a copyright as 'visual material', not 'visual arts work'. The most likely legal interpretation of which is that the author has added copyrightable material to non-copyrightable material, and obtained a copyright on the addition [3].
Go read Title 17 of the United States Code. AI would be considered an anonymous author. You can't claim title to another authors anonymous work. If you claim I can own copyright for works created using my software algorithms, you would then be claiming that Adobe owns the works to all works created within their tools as long as those works did not have copyright claims made by the original author, which obviously Adobe doesn't own.
Well, I'm fairly certain one can claim ownership of works they create via algorithms, as the entire digital media space wouldn't exist. I don't think authoring or hosting an algorithm gives one carte blanche ownership over things other people create with that algorithm, but cynically I could see that happening if enough money is involved unfortunately, but hopefully not. It's not clear at all why AI would ever be considered the author and in fact all the sources I've found assert the opposite, that AI cannot be considered an author. So I still haven't seen anything preventing people from claiming copyright protections on works they create via AI.
There is also this filing [1] from February regarding the "human authorship" requirement. Thaler was requesting that he be granted the copyright to this work as the AI was "work-for-hire". The Copyright Office found that Thaler “provided no evidence on sufficient creative input or intervention by a human author in the
Work.”
So, at least in this instance, the AI created the work with insufficient creative input from a person, and thus has no human author, and is not eligible for copyright.
You are misreading it and it's easy to do so. This is the money quote:
"Thaler does not assert that the Work was created with contribution from a human author"
All it takes is for the copyright submission for this Artist to say "I used a computer to generate this image through x y z steps selecting for q r s features until I was satisfied." I'm confident they would accept it.
This case is an attempt to set new precedent for copyright by granting it to a computer. And rightfully so, the Copyright office refuses to do so.
> I can't claim copyright on a work by an AI, because I'm not the author.
You can lie about authorship, and it will (probably) be undetectable/irrefutable.
Also, currently AI's are not given personhood under law, hence they are not the author of the work they create. Given that, it isn't much of a stretch to say that if you wrote the software that runs the AI, you are the author. Which is a problem for users of tools like DALL-E.
I don't believe this is the case, because the AI is what is known as an 'anonymous author' under copyright, and authorship does not just transfer because you're the first to claim it.
Are you claiming that whoever wrote Photoshop automatically owns digital art created in Photoshop if the artist of the digital art doesn't claim copyright?
" The finding that AI cannot be an author for copyright purposes does not mean that AI-assisted outputs are void of any copyright protection. Works created by a human using software on a computer (e.g., Microsoft Word or Adobe Photoshop) are arguably protectable under copyright law."
Now Stable Diffusion and friends exist, anyone can make images with a free Google Colab account or their own GPU. It's been democratised and waitlists are a thing of the past.
I am waiting for the next brilliant stage of this plan... where we set up a fund to compensate all the artists whose work was used to train the artists that generated the art to train the AI. Every artist that registers for the Shutterstock compensation plan will submit a list of every artist whose worked they studied, talked about, obtained inspiration from, or even viewed. When they get compensated from the fund, a portion of their compensation will be transferred to the artists or estates of the artists on their list. For artists no longer with us with no estates the funds will be donated to the EFF.
I haven't figured out how to compensate the artists that trained the artists that trained the artists that trained the AI. This gets 'meta' pretty fast.
76 comments
[ 3.2 ms ] story [ 154 ms ] threadI guess your argument could be changed slightly to "Why buy stock photos when I can take a photo with my phone?"
it's pretty easy to tell some generative software to "Female giving a presentation in a well lit office with windows open" and get something pretty decent.
This goes for anything that needs props that a normal person doesn't have.
Is it? I tried that exact prompt with dall-e and got 4 images with nightmare fuel faces: https://imgur.com/a/yyt5Tax
As someone who actually tried it, with current models it's possible but neither easy nor straightforward. Getting decent and coherent images is still hard and requires finetuning, so there is certain value in this. And for very specific and series-consistent images I'd still rather set up a photo session if I had a studio.
It will change for stock photos, but probably not very soon.
Tell me you haven't tried Stable Diffusion without telling me you haven't tried Stable Diffusion.
I agree that anatomy and hands are a problem. Give it 6-12 months and they won't be. That's how fast things are moving at the moment!
"female with two hands that have five fingers, and normal looking face, giving a presentation in a well-lit office with correct shadows, and a clock on a wall that is perfectly round with legible roman numerals correctly placed around the dial ..."
Honestly not a lot though, it's so easy. It's not really comparable to taking them with a phone because you don't have to have access to the subject!
Or possibly you would have to have your own model which is trained on that space of images that is relevant to what you are looking for.
Look at what virtually all the hobbyists and enthusiasts out there are doing with this stuff. It's almost all exlusively these: surrealism, fantasy and dumb memes. In all these fun genres, bad details, like hands with six fingers, or laughably wrong shadows, do not matter.
> "The data we licensed from Shutterstock was critical to the training of DALL-E,” said Sam Altman
So there was legitimate Shutterstock data used to train DALL-E. I don't think I've read this before. The images with Shutterstock copyright watermarks were presumably scraped accidentally from the web.
The contributor fund is also setting a precedent that data used to train a model has a claim for payment.
> Shutterstock is launching a “Contributor Fund” that will reimburse creators when the company sells work to train text-to-image AI models. This follows widespread criticism from artists whose output has been scraped from the web without their consent to create these systems. Notably, Shutterstock is also banning the sale of AI-generated art on its site that is not made using its DALL-E integration.
"DALL·E 2 is trained on hundreds of millions of captioned images from the internet."
As far as I know, this is Shutterstock opinion and there is no judge rules or new laws addressing the situation.
As far as I know it’s even legal to study the style of one specific artist and create imitation art as long as it’s clear that this is a new artwork from a new artist and about ~50% different from any existing works.
So a prompt like “a photo of Times Square in the style of Ansel Adams” should be ~50%+ unique.
The trickiest area for these models might be prompts that recreate existing work.
Given a prompt like “a black and white photo of Yosemite Valley in the style of Ansel Adams”, if the model is too good, then it seems like copyright infringement.
But as far as training data goes, as long as it’s acquired legally, I don’t think it can (or should) be regulated.
Difference being that the AI doesn't take inspiration, it takes raw data.
The mechanisms of human brain is indeed unknown to us. So how can we say with confidence that its underlying workings are this way or that way with certainty? We can only observe the inputs and outputs and do our best to compare the two. Which is the whole point of AI - to simulate the human brain (or what we can observe about it) to the best of our ability.
I guess the main difference for me is the extent of human involvement.
For example, with Adobe there is a huge amount of human interaction required to get from idea to end result.
With DALL-E, you basically go from idea to end result instantly.
What is to prevent someone from generating that same image with a prompt?
What would prevent them from simplu resharing that image freely with everyone, or undercutting shutterstock? If Shutterstock tries to sue, the courts may say AI generated images do not enjoy copyright protection !
I hope so.
Yes, "anybody" could just spin up a Stable Diffusion server and do the same, but plenty of companies would be perfectly happy to outsource this to Shutterstock and, quite importantly, also have them take all the heat for any legal complications behind reuse, copyright, etc etc.
The problem is that with the amount of prompt tweaking one needs to do, it doesn't scale
That may very well change in the next X years, but right now it doesn't really work as a business
Once that changes why would you go to Shutterstock anymore?
Their whole thing is being a marketplace, when the supply side of the market gets eaten by someone else they have a major problem.
It seems like a pretty unlikely bet that we won't get something like SD which can do realistic "women laughing with salad" images within a few years.
There's a recurring theme in neural networks. You can go from nothing to 90% in a year. And it's amazing. That naturally leads you to believe that perfection is little more than a bit more compute and a bit more massaging the data.
Then over the next year you manage to get to 95%, further affirming your beliefs. But from then on every single fraction of a percent becomes exponentially more difficult. And it soon becomes clear that there's an apparently insurmountable asymptote long before you get anywhere near 100%.
There are more trucking jobs than ever today.
AI art is not - jobs are already being lost in real time, users can create stock photos, logos, bespoke art right now. And more and more jobs will be lost as the AI improves.
1 - It will deflate
2 - It will explode
Deflation is where the output of system tends toward a fixed point (AI art gets blander and blander). As it feeds on it's own output (which becomes the dominant training material because its cheap and plentiful) each vector is compressed.
Chaos (explosion) is where the output gets less predictable and more crazy on each iteration. Who knows what emerges before the system breaks.
Positive feedback systems are generally bad. For example, in biology/food; we started feeding cattle on the rendered down remains of other dead cattle. The result was Bovine Spongiform Encephalopathy, a lethal prion that thrived within the feedback loop.
But no worries, an unknown part is shared with creators.
Two revenue streams without consent but they still present themselves as the ethical AI company.
And it's not even clear what a customer is buying. As far as I know, AI art can't be copyrighted.
Why do you think that?
There is a common article that gets passed around where a person tried to get copyright assigned to a machine and that keeps getting denied. That is very different from "AI generated images cannot be copyrighted".
If I write a program to randomly generate an image of squares of different colors:
- I, a human, can have copyright over the program text itself as a literary work
- I, a human, can have copyright over the image output of the program
I can't claim copyright on a work by an AI, because I'm not the author.
AI created works, like animal-created works [2], lack authorship, and are thus not eligible for copyright at all.
It's being litigated, and might change, but this is the current position of the US Copyright Office [1].
The only AI generated copyrighted work at the moment is a graphic novel, Zarya of the Dawn, and was granted a copyright as 'visual material', not 'visual arts work'. The most likely legal interpretation of which is that the author has added copyrightable material to non-copyrightable material, and obtained a copyright on the addition [3].
[1] https://www.techdirt.com/2022/08/15/dude-who-keeps-suing-to-...
[2] https://en.wikipedia.org/wiki/Monkey_selfie_copyright_disput...
[3] https://en.wikipedia.org/wiki/L.H.O.O.Q.
Is there any source for this? The only things I can find are people trying to assert non humans as copyright holders and being told no.
So, at least in this instance, the AI created the work with insufficient creative input from a person, and thus has no human author, and is not eligible for copyright.
https://www.copyright.gov/rulings-filings/review-board/docs/...
"Thaler does not assert that the Work was created with contribution from a human author"
All it takes is for the copyright submission for this Artist to say "I used a computer to generate this image through x y z steps selecting for q r s features until I was satisfied." I'm confident they would accept it.
This case is an attempt to set new precedent for copyright by granting it to a computer. And rightfully so, the Copyright office refuses to do so.
You can lie about authorship, and it will (probably) be undetectable/irrefutable.
Also, currently AI's are not given personhood under law, hence they are not the author of the work they create. Given that, it isn't much of a stretch to say that if you wrote the software that runs the AI, you are the author. Which is a problem for users of tools like DALL-E.
Are you claiming that whoever wrote Photoshop automatically owns digital art created in Photoshop if the artist of the digital art doesn't claim copyright?
https://www.copyright.gov/rulings-filings/review-board/docs/...
https://www.jdsupra.com/legalnews/which-ai-components-are-co...
Current AI is "really fancy automatic photoshop collage".
https://openai.com/blog/dall-e-now-available-without-waitlis...
But I think a lot just moved on to Stable Diffusion
I haven't figured out how to compensate the artists that trained the artists that trained the artists that trained the AI. This gets 'meta' pretty fast.
I wonder if they or someone else wants to buy it?