This weekly roundup thread is intended for all culture war posts. 'Culture war' is vaguely defined, but it basically means controversial issues that fall along set tribal lines. Arguments over culture war issues generate a lot of heat and little light, and few deeply entrenched people ever change their minds. This thread is for voicing opinions and analyzing the state of the discussion while trying to optimize for light over heat.
Optimistically, we think that engaging with people you disagree with is worth your time, and so is being nice! Pessimistically, there are many dynamics that can lead discussions on Culture War topics to become unproductive. There's a human tendency to divide along tribal lines, praising your ingroup and vilifying your outgroup - and if you think you find it easy to criticize your ingroup, then it may be that your outgroup is not who you think it is. Extremists with opposing positions can feed off each other, highlighting each other's worst points to justify their own angry rhetoric, which becomes in turn a new example of bad behavior for the other side to highlight.
We would like to avoid these negative dynamics. Accordingly, we ask that you do not use this thread for waging the Culture War. Examples of waging the Culture War:
-
Shaming.
-
Attempting to 'build consensus' or enforce ideological conformity.
-
Making sweeping generalizations to vilify a group you dislike.
-
Recruiting for a cause.
-
Posting links that could be summarized as 'Boo outgroup!' Basically, if your content is 'Can you believe what Those People did this week?' then you should either refrain from posting, or do some very patient work to contextualize and/or steel-man the relevant viewpoint.
In general, you should argue to understand, not to win. This thread is not territory to be claimed by one group or another; indeed, the aim is to have many different viewpoints represented here. Thus, we also ask that you follow some guidelines:
-
Speak plainly. Avoid sarcasm and mockery. When disagreeing with someone, state your objections explicitly.
-
Be as precise and charitable as you can. Don't paraphrase unflatteringly.
-
Don't imply that someone said something they did not say, even if you think it follows from what they said.
-
Write like everyone is reading and you want them to be included in the discussion.
On an ad hoc basis, the mods will try to compile a list of the best posts/comments from the previous week, posted in Quality Contribution threads and archived at /r/TheThread. You may nominate a comment for this list by clicking on 'report' at the bottom of the post and typing 'Actually a quality contribution' as the report reason.
Jump in the discussion.
No email address required.
Notes -
@CeePlusPlusCanFightMe
Shutterstock will start selling AI-generated stock imagery with help from OpenAI
This strikes me as fantastically stupid. Why would I buy AI-generated imagery from Shutterstock when I could just make it myself? In the near future, people who don't have high-end PCs won't even need to pay Stability or Midjourney for a subscription. Getting the open source version of SD to run smoothly on your phone is a mere engineering problem that will eventually be solved.
Maybe they just understand this market better than me? Never underestimate just how little work people are willing to put into things. Even playing around with prompts and inpainting for a few hours may be too much for most people, when they could just hand over $10 for a pretty picture on Shutterstock instead.
The "Contributor Fund" also makes me slightly more bearish on the prospect of there being any serious legal challenges to AI art. If there was any sector of the art market that I thought would have been most eager to launch a legal challenge, it would have been the stock photo industry. They seem like they're in the most obvious danger of being replaced. Undoubtedly, copyrighted Disney and Nintendo art was used to train the models, and those companies are notoriously protective of their IP, but they would also like to use the technology themselves and replace workers with automation if they can, so, they have conflicting incentives.
According to the article though, Shutterstock was already working with OpenAI last year to help train DALL-E, so apparently they made the calculation a while back to embrace AI rather than fight it. The "Contributor Fund" is pretty much a white flag. But maybe Getty will feel differently.
Edit to clarify a bit: What this seems to come down to is that they're adding a "DALL-E plugin" to their website. Why I would use Shutterstock as a middleman for DALL-E instead of just using DALL-E myself, I'm not sure. Their announcement makes it clear that they're not accepting AI submissions from sources besides their own plugin, due to outstanding legal concerns:
There's been some talk here about corporations using AI art and then simply lying about its origin in order to retain copyright. If I use Megacorp X's art without their permission and they try to claim a copyright violation, and I claim they made it with AI so I can do whatever I want with it, I wonder where the burden of proof would be in that case?
Isn't a benefit of Shutterstock that if an image you use somehow violates copyright, they'll take on the liability?
With AI art being a potential copyright minefield, most people probably don't want to end up in a situation where they are defending against a relatively novel suit. So having access to AI art that is produced based on Shutterstock's images, and presumably the protection against copyright violations, that's probably a great thing to have for many companies/users, big and small.
Isn't that in stark contrast to DALL-E, which says anybody who generates an image is the full legal owner of it, and can do whatever they want? If DALL-E were to spit out two identical images, I wonder who owns it? lol
The very last job to be automated will be "scapegoat."
More options
Context Copy link
More options
Context Copy link
Aside from the time savings that other people have mentioned, the other big advantage of Shutterstock is that it handles all of the relevant copyrights. Using AI art generation is probably OK from an IP rights perspective, but there's still a chance that a generated image will be close enough to the source material that the original artist could sue.
Using Shutterstock already insulates companies from the risk of traditional artists giving them a copied picture. I can see how people would see the same service as valuable for AI images too, at least until the legal issues get straightened out.
More options
Context Copy link
If this were Stability, I'd have suspected a finetune (or these days they seem to prefer using hypernetworks etc.) to produce that classical soulless stock image feel, with old men making the Harold grimace, smartly dressed black entrepreneurs crossing the finishing line, brain tissue looking like lava lamp clumps in a nondescript neon-lit emptiness, and so on. Not sure if OpenAI intends to offer customization of that kind.
Ad for your dashed hopes of regulatory action and the «Contributor Fund»: this looks very unsurprising to me, because as you say, relying on red tape when the models and techniques have proliferated and people run Stable Diffusion on their iPhones (literally) would be long-term disastrous.
I think corporations will try the embrace-extend-extinguish route with regards to content creation, and regulators will attempt to choke the next generation of AIs with preemptive X-risk framing.
More options
Context Copy link
Probably the same reason you could buy potatoes in the store though it's possible to grow them on your own back yard? Convenience, separation of labor, specialization, availability of the expertise, etc. I mean, I was eating tomatoes grown on my backyard this whole summer, and they were excellent, but that doesn't mean I think commercial tomato growers are stupid to expect to find any market.
More options
Context Copy link
as an aside i'm curious about how much Shutterstock got paid for the training data they sold to OpenAI.
More options
Context Copy link
OH HELL YES MY HOBBY HORSE, thank you for pinging me
I actually had a completely separate post that I was just going to throw in main which made essentially the same points as you made here. In particular, Shutterstock gives absolutely no clue whatsoever what the genuine Shutterstock value add would be. Like, as a customer why on earth would I ever go generate an AI image on Shutterstock using DALL-E when I could just use the DALL-E 2 API? Their editing tools? If that's what they're banking on, i'll just note that if Shutterstock wants to compete with Adobe in generative content creation, they... actually i don't know how to finish that sentence because it seems self-evidently like a terrible terrible idea.
Other notes:
Lol @ “crucially”. A ban on non-Shutterstock-sponsored AI art seems like transparently a non-functional fig leaf given that (1) there’s no method even in principal of checking whether a piece of art is AI-generated, and (2) adobe’s announced integration of AI with their products means that there will soon no longer be any kind of hard-and-fast distinction between “ai art” and “not ai art”. You know: “AI art? Oh, no, you misunderstand entirely, I made this myself using Adobe Illustrator.”
As an aside: this article gets a primo place in the Shutterstock blog. You will of course notice there is no corresponding article in the OpenAI blog, since OpenAI does not give a shit about this partnership except in the sense that it marginally pads their coffers if it works, and if it doesn’t, hey, it’s not their problem.whoops, missed the part where they were providing training data.I 100% don't get why this protects the Shutterstock business model as opposed to burning a whole bunch of money on developing an API integration that's strictly inferior to every other possible way of accessing that API.
EDIT: On reflection I should not have referred to the customer using the Shutterstock site to access DALL-E 2, since the plan seems clearly to sell the DALL-E 2 generated images as stock images (where the artist is the one using DALL-E 2). Which also seems... pointless, as a customer. Why would I want to buy limited rights to an image an AI generated when I could generate one myself for free? And why would Shutterstock have any advantage in vending out such AI-generated images as opposed to a random hypothetical AI startup?
Their plan seems clearly to exist in the very very narrow gap between "I want something complex and specific, I'll use Adobe Illustrator" and "I want something straightforward, I'll just use a generative image directly". This gap only narrows over time.
EDIT EDIT: My understanding right now of how art copyright works is that if you use an image you don't own the rights to, the enforcement mechanism for that is the artist coming out of the woodwork and demanding money, with proof of some kind that she created the image. I do not know what the plausible enforcement mechanism for AI art is even if it's theoretically problematic from a copyright perspective. Is a judge gonna grant you a subpoena to get the chain of custody for the image so you can verify you have the right to sue over it? What does that conversation sound like? "You can see it's AI! Just look at the hands!"
EDIT EDIT EDIT: On reflection right now the Shutterstock curation process (so you only get to see the good generations) does represent a concrete value add, but one that decreases in value over time as image generation products get better.
I was assuming this was the value-add. At least with existing tech, figuring out how to phrase the query is a skill. And I've seen some tutorials on getting good results out of Stable Diffusion talking about making multiple img2img passes. I could certainly believe some future tech will be able to give a perfectly acceptable endless list of good images to an unsophisticated human asking for "a stock image of [...]", but the tech isn't there yet. Shutterstock is betting on there continuing to be a gap for some human effort to fill for some time; given that these techniques seem to be progressing quite quickly, I guess we'll see how much time they're actually buying themselves. At some point the human effort is small enough that it gets pushed back onto the graphic designer who is currently doing the effort of querying and selecting an image from Shutterstock.
Well that and continuing their existing job of providing legal/copyright insurance on the images: you're paying for the issue of copyright of AI generated images to be Shutterstock's problem.
More options
Context Copy link
More options
Context Copy link
If one is an employee, playing around with something for a few hours probably costs their employer far more than an order of magnitude beyond $10.
More options
Context Copy link
Why does McDonald's exist when you can make a hamburger yourself? People pay for convenience .
Yeah, but worth considering the inconvenience involved in having to track which rights you have purchased to which media, especially if you're a small business using a bunch of them. AI art lacks this issue, since you know nobody has the rights to the image because it's unique.
And people using stock images are people who are, for the most part, running small businesses, not consumers who we might expect to be lazy.
Not necessarily. AI art often involves seed images which may be copyright. Also, the person who produced the AI art my still try to claim copyright.
Ah, an unstated but crucial assumption in the post was that you personally the one who created the image. it's true, AI images grabbed off of a stock website are basically similar to regular stock images in all relevant respects.
More options
Context Copy link
More options
Context Copy link
There is something of FUD campaign going on with AI art property rights based on the idea of the model and model produced works being derivative from the works they were trained on. You've probably seen some of the comments reason along those lines in earlier threads here. Of course with IP rights, buying a right from someone who may themselves not have that right does not fully protect the purchaser, but that aspect is less well advertised so it may still seem worthwhile to purchase a license from a known entity.
IANAL, but I don't see that as FUD, I see it as an open legal question.
Ofc judges can decide whatever. But there's no way they're going to side with the artists, destroying AI tech. It'd be just yielding to China.
Ignoring practicalities, it just doesn't make any sense. Why couldn't you train AI on copyrighted works while still able to train your own biological neural network on them?
Have you tried to reproduce a copyrighted photo using only the latent representations stored in your biological network?
While biological networks and computer models have some similarities in abstract, in practice there are crucial differences.
More options
Context Copy link
More options
Context Copy link
it does seem fair to not want to be the test legal case for AI art
More options
Context Copy link
More options
Context Copy link
seems fair. I do think there's also a thing where it's not clear yet how suing someone for using ai art works since the way it works for art now is that if you use an image you don't have rights to, the way that shakes out is the person who originally made the image sues you for damages (and they can prove they made the image because they presumably have some timestamped evidence indicating so).
But who would be responsible for noticing and suing somebody who made an image with an AI trained on copyrighted images? How would they know it was AI generated? Sure, subpoena a chain of custody for the image, fine. How are you going to get a judge to agree with you that this piece of art looks like it was generated on copyrighted images if the image itself does not contain those images? Gotta get the judge onboard to get the subpoena.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
It’s about getting in the corporate market quickly.
More options
Context Copy link
This makes sense to me. Shutterstock already has relationships with all of the companies that need stock images. Relationships like that are hard to build from scratch and are a kind of "social technology" that is easy to dismiss but critical to succeed as a B2B business. So even if all they can offer from a technical perspective is a slightly worse version of Midjourney, that's still going to add a lot of value for their customers. Maybe at some point in the future this tech will be ubiquitous, but my guess is even at that point there's going to be a compatibility layer to embed it in the workflows that users of stock images are accustomed to and Shutterstock would be well positioned to offer that.
I expect this to work right up to the point where there's an economic downturn and customers look around for line items they can cut from their budget.
EDIT: Ahh, that was probably not actually right given that shutterstock's subscription plan is actually fairly reasonably priced.
More options
Context Copy link
More options
Context Copy link
I work for an organization that uses shutterstock. That is absolutely a tradeoff worth taking. We use a dozen or so images a day, so spending 1 minute on an image instead of 10 minutes changes it from a part time job, to just a small additional task. The person's salary that gathers these images is in a mediumish salary range. But it would still be worth it to us if we were paying this person minimum wage.
Shutterstock only needs to save an hour of time of a minimum wage employee once a month to be worth it. It saves us thousands of dollars, its easily worth it. Until someone creates a giant library of free AI art with clear image rights then shutterstock will continue to be worth it.
Currently you're totally right. But I'll point out that the reason it takes ten minutes is because right now AI art kinda sucks (so it takes a while to get a prompt that looks okay), and the tech only gets better from here on out.
I don't know if the tech matters too much. There is only so much mind-reading that a computer can do. Any image generator has to be met halfway by someone that has played around with the generator enough to understand how to get useful images out of it.
The value-add of shutterstock is to be able to quickly search through a bunch of generic pictures. Even a curated list of AI generated images would work fine for this value-add.
The guy who currently does this at my company basically searches the tags of an article "inflation, money", and then gets a list of images that match those tags. He quickly visually scans a gallery of images, and then picks out the one he wants. While looking at the images he might spot one that has a building in it that looks like the federal reserve, and he thinks 'oh even better match' and he picks that one.
The images aren't supposed to be special. They exist mostly just to break up what would otherwise be an ugly wall of text. We might all be fine reading sites like reddit, but apparently a bunch of people like more variety in their visual space.
There is a very real sense in which Stable Diffusion and its ilk do represent a search process, it's just one over the latent space of images that could be created. The Shutterstock search process is distinct primarily in that it's a much much much more restricted search process that encompasses only a curated set of images.
This isn't (just) a "well technically" kind of language quibble, I'm pointing this out because generative prompt engineering and search prompt engineering are the same kind of activity, distinguished in large part by generative prompts yielding useful results far less frequently, with the search process being far slower as a result.
But this is a temporary (maybe) fact about the current state of the AI tool, not a permanent fact about reality.
More options
Context Copy link
But this is equally true of search engines and tagged galleries. When you venture out to find a picture, you have a description, at least in the form of a search query. Text to image models can minimize the difference between this text, treated as a caption, and the output; the ability to minimize FID is how they are evaluated. The best that Shutterstock can do is give you a set of similar images out of a necessarily finite pool of discrete samples. If we account for the time you waste on generations and shuffling through offered options, Shutterstock may have an edge for now due to curation and human common sense and the learned «style» of genericness that has become standard in corporate illustration, partially thanks to such platforms... but that can be learned as well. In the limit, generators will be strictly better at providing the same type of content.
Incidentally I prefer imageboards and would appreciate a reddit-esque forum with the AIB style of media attachments (also would be nice to add documents). But beggars can't be choosers. Wonder what's that «intelligent internet» Emad talks about.
More options
Context Copy link
Ugh, I hate when websites do that. I don't need a picture of a person smiling while looking at their laptop when I'm reading an article about filing taxes. What is it even doing there? I can't believe there are literally people paid money to find these pointless images.
More options
Context Copy link
My favorite variant of this is reading aggregator rehosted versions of news articles where only some of the inserted images make it in and not necessarily in the correct location or with the correct attribution/caption. The data mismatch being more interesting than the writing quality typically.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
Spending $10 instead of a few hours sounds like a good deal to me.
Yeah, 100% agree here. Also never underestimate how reluctant/incompetent people are at engaging with new tech. My DND group includes a couple of graphic designers and several people working in tech and aside from a couple of them mucking around on Dreamstudio all have claimed that they don’t have the time to figure out how to run a local instance of Stable Diffusion. Maybe that’s right, although I did it in a couple of hours (and that’s including training up several models on Dreamstudio to recognise pictures of me, my wife, a couple of friends, etc.).
Having to set up SD locally is an absolute hardstop for non-technical people right now. They’re going to see that you have to “clone a repo” and “run a script” and they’re just going to nope out. Someone will come up with a true idiot-proof GUI solution soon enough though.
There's already this creatively-named guide to get started.
More options
Context Copy link
In my case the power of horniness won through and helped galvanise my loins through the technical parts. I’ve now found a 0.3/0.7 merge of f111’s female anatomy model and SD 1.5 gives some very impressive results.
Same here.
I'm a programmer and was well aware of the marvel that is txt2img but the reason I immediately got it running the moment it came out was so that I can create the dankest porn.
I'm half curious, half terrified to see what kind of abominations motters are creating with this stuff. Fuck it, we must have nsfw tags for a reason, right?
Few random results. I only played with it for a few minutes after setting it up so far.
I like how #4 is practically just Projekt Melody.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
On first read I thought there was yet another mechamusume or moe anthropomorphism IP that I hadn't heard of featuring Aardvark-chan.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
it's possible that courts will start demanding a chain of custody for art, but I can't imagine that's terribly likely given the insane logistical challenges involved in enforcement.
More options
Context Copy link
Do you have a link for that RIAA story?
As someone who grew up steeped in seething hatred for the RIAA, I look forward to watching them try and fail to halt another inexorable advance in technology.
More options
Context Copy link
I’m happy that musicians have a powerful advocacy group who will defend their interests, but it’s unfortunate that artists don’t have something similar. The best I can hope for is that the RIAA does bring a lawsuit at some point and it sets a precedent that carries over to visual art.
The RIAA represents record labels, not musicians.
I’m aware. I was being intentionally cheeky. In this case their interests all happen to coincide, so they are representing the interests of musicians in this case, even if it’s just a byproduct.
Record companies care about profit and thus they stand to lose from unsigned artists creating better art, but if the calling of artists is art, they gain from having more tools to create it.
Let's try to keep things straight here.
AI is pretty much only going to harm visual artists. It's not going to "help them make better art", it's going to replace them, because people used to need artists to draw X and now they don't. One guy might get a productivity boost from using AI, but that's not going to do much for the other 10 people who got laid off.
Granted, the dynamics of the music market are quite different, because it's more dependent on live performances than commissions. But the concept of commission work still exists in music as well - lots of musicians make money by composing pieces for commercials, movies, and video games. I can't see any way that those musicians aren't going to be fully supportive of any legal action that the RIAA takes. They're at the highest risk of being replaced by AI - why hire someone to compose a jingle for my commercial when I can just type in a prompt instead?
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
My guess is that the average average marketer who wants to get stock images either:
has no clue that AI generated images even exist yet. Or if they do, they think it's always existed. Basically this person is completely out of the loop.
Thinks it's way more complicated than it is to get images from stable diffusion.
Both of these groups will diminish over time as the technology and the means to use it become more commonplace and common knowledge. But until then, there's a few bucks to be wrung out of them.
I think there's also currently the quality factor-- right now AI art honestly kinda sucks and I have to go through dozens of generations to get something half-decent. I expect this to get change very quickly over the next several months, since as Gwern says, attacks only get better.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link