Getty made an AI generator that only trained on its licensed image
theverge.com
theverge.com
I'm pretty firmly on the side of "large-scale analysis of content that's accessible to the public should be entirely outside the scope of copyright, regardless of what one does with the output", even though I also sympathize with artists that see generative AI as a potentially existential threat to their careers. But I'm curious if Getty is actually going to make meaningful payments to the creators of the images that went into the training set, or if this is more of a "we could continue to pay our existing royalties to content creators, or we could keep most of the money for ourselves by selling access to a model trained on that content" type of situation.
[1] "We’ve created a model that compensates our world‑class content creators for the use of their work in our AI model, allowing them to continue to create more of the high‑quality pre‑shot imagery you depend on." (https://www.gettyimages.com/ai/generation/about)
Adobe is also dealing with this question with Firefly. So far they plan to pay creators annually, but how much each artists gets and how the amount is calculated is yet to be determined. Since a model is only as good as its training data, hopefully this can be a new revenue stream for artists?
https://techcrunch.com/2023/09/13/adobe-starts-paying-bonuse...
the model is made from X images, from Y artists, each artist has a percent contribution to the total. between 0 and 100 percent.
so each time money is generated by the model, the contributors get their share.
as a bonus, it is easy to understand how that share was derived. so it is implicitly fair.
(please keep in mind that this “artist’s share” will be a percentage of the net revenue generated, and that this percentage will probably be a trade secret or some such.)
(any works from artists unable to be paid, meaning public domain or otherwise, will be handled by the corporation for the continuous and never-ending improvement of the platform that makes it all possible.)
I don't think I understand this. How does something being simple to understand make it fair?
The model would be almost entirely, if not entirely, on your images.
a company could always train a model on what constitutes a low value or malicious image. submissions get screened by the bouncer before going on to integration with the commercial model.
throw in a submission fee, and enough of the malicious submissions will be stopped to plausibly call this solved.
plus, new cash stream.
If this isn't already a scam it's about to be. Paying copyright holders money for this will cause people to produce mountains of low effort images, which then get the same payment as things people put actual work into, ensuring that the payments are negligible because they're diluted by such volume. And will reduce the quality of the model so nobody wants to use it.
And we know what the cheapest way to generate a large volume of low effort images is, which is probably not going to bode well for the purity of the system either.
At the same time though, Getty and Adobe have a vested interest in not allowing that content into their training sets. They are obviously curating here, so that their generator outputs better stuff than the generators of competitors. I'm sure they are wargaming out the possibility of MS/OpenAI and Google coming in. Or maybe even worse, Disney or the video game makers, entering this market.
So you would need to generate a mountain of low effort content, and then hope that some of it is good enough to make it through the curation process. It's not a guaranteed thing.
The threshold for getting accepted has to be much lower than the threshold for making much money, because the company needs the long tail (and so does the AI model), but a photographer previously only made money if their photos were popular.
Now as long as you can meet the minimum bar for inclusion, you get the same payment as anyone else, so the quantity of photos at that minimum level of quality goes way up.
They will totally train on every image they can get their hands on, and then finetune on a curated set of the best images.
AI. I’m ok with closing the loop and AI folding on itself.
More obfuscation will mean more shaving.
Authors, artists, etc. are suing projects whose models just inhaled their work and generate derivative works and I'm surprised the studios aren't getting more involved with going after AI models that included their properties in them.
I can see this going the route of us having a Disney AI, Sony AI, Discovery AI, Amazon AI, etc. where you can generate stuff using models owned by the studios, but only those studios and any public domain stuff they suck in too.
Exactly. These companies are probably hoping the artists win as many legal battles as they can, since the result will be that only big companies will be able to create useful AI models.
> He told me the training process took about 2.5 hours on a GPU at Vast.ai, and cost less than $2.
This is a SaaS waiting to happen.
https://waxy.org/2022/11/invasive-diffusion-how-one-unwillin... This was on the front page last year.
Presumably when photographers/artists submit their images to Getty they hand over full ownership, or otherwise pretty broad licensing agreement. If Getty's the rights holders for these images they can use that by training their own models.
In my mind, OpenAI/Midjourney/Stable Diffusion are the Napsters of generative AI. Adobe Firefly, and now Getty, are coming up with the iTunes Store/Spotify. For better or worse.
The alternative to Spotify and the traditional record labels are places like Bandcamp, where artists get a far more significant portion of the earned money. Any industry where 99% of the money goes to the middle man or service provider rather than the creator is an industry in dire need of disruption.
My limited understanding is that Getty is at least as bad as the music industry, if not much worse. No doubt Getty will make millions of dollars from their AI image generator, while the artists may get a few cents a year if they are lucky.
edit: also, re spotify, as far as I'm aware they do carry small and independent labels. Getty is comparable to the big music labels, not to spotify.
In this case: Getty is going to be paying less overall (otherwise, why do it at all?), and they will pocket the extra margin.
As entertaining as it would be to have a model where you type "photo of an astronaut riding a horse" and it defaults to Buzz Lightyear riding the horse from Tangled, that's not really what people use these models for in the main.
You don't actually want derivative works. The interesting training data isn't Hollywood movies, it's all the junk people post on social media. What you want is thousands of generic pictures of astronauts and thousands of generic pictures of horses. Pictures of Tom Hanks as Jim Lovell aren't any better (and may actually be worse) than actual public domain photos from NASA.
I think the vast majority of this is going to replace google image search, the way it replaced clip-art
If you're the first group to sue you have to spend millions extra on lawyers to establish the precedent. I don't see much reason for Disney and friends to rush to be first.
The current state of AI isn't really a threat to Disney and friends, just suggestive of a future threat. No need to rush on that account either. Especially since if they win the lawsuit - they're still going to benefit from all the R&D on neural networks that is happening on other peoples dime right now.
And all in, do studios stand to lose more or gain more to AI? Drastically cutting their costs might be worth some extra competition.
Most people seem to have a difficult time grasping that if the model is trained in such a way that Disney's IP is excluded, the model doesn't know Disney even exists. Consider if every Disney website was excluded from Google and every Disney related trademark was blacklisted.
As of mid 2023 it is very clear personal assistants are going to replace traditional web search. In my use case, they've replaced it 100%: excluding searching a single website or product name. Will these generative assistants -- which must be capable of both processing and generating images -- know the rights holder's IP even exists? The idea that a useful LLM is going to be trained on 100% public domain, copyright free data is absurd.
IP owners are suggesting that future personal assistants would need to pay them for the knowledge of even the mere existence of their IP. That would be like if every website indexed by Google and every trademarked keyword required Google to pay the copyright owner per search. That isn't possible. To the contrary, the reverse occurs.
What if the future is in fact the opposite of what IP owners are now suggesting? Nike pays the generative AI company for Nike shoes to appear in the one-off movie generated for a single viewer based on their personal preferences?
The world is drowning in IP: text, video, music. The future will be a deluge.
Admittedly, they're both nonsensically incompetent at times, and if you believe Disney's financials (do you believe any Hollywood financials?) then they somehow manage to lose money on $10,000,000,000 / year. I'm still not sure how you make the equivalent of 10's of 1,000's of tv episode costs ... and then lose money ... with Disney's back catalogue and vault.
There was this topic yesterday related to legal data https://news.ycombinator.com/item?id=37627129
I've been saying all year that LexisNexis absolutely HAS to be building a training corpus. It was a big deal when that dumb lawyer in TX filed a brief with bogus citations, but what if you had an LLM that was specifically trained on legal filings, understood rules like "If you cite a case, you MUST include a reference to that case and it must be valid in our system"?
It ended up not really mattering, because the product people preferred to use was YouTube, and the branding, media coverage, lawsuits, all that stuff took so long to resolve that it happened after YouTube was the winner of the space. Of course if the YouTube lawsuit had worked out like Napster then everything might have ended differently....
Not so long ago many predicted that stock image sites will eventually become obsolete or at least much less important due to AI image generators. But it might turn out that stock image sites are uniquely positioned to dominate commercial AI image generation because of the "ethically sourced" high quality training data they posses.
They might still have minor legal fights whether this is a legal way to use the data licenses to them, but the list of people who could possibly sue is much shorter, known to the stock image site, and consists of people who have already signed some form of contract about the use of these images. The legal risks are well scoped and predictable, unlike "we scraped the internet" models.
At the end of the day, we’re only talking about these models because someone published research completely tainted by the copyrighted training. So it’s still totally unknown, until a lawsuit, if even a “redo” of every part of the model on all licensed stuff, which may not even work because there isn’t enough variety, if it’s all that it promises legally.
the high-margin, very legal brand-name market. As others have pointed out, the means of production are already leaked, the goods easily copied and delivered, so there will be endless layers of gray market, not-legal, cheaper knock-offs and worse IMHO.
There could be an AI carveout going forward but that's not retroactive.
Derivative work and all that jazz.
If the copyright lawsuits do end up ruling against image AI statups, Getty may face problems when inevitably someone demands Getty takes down the images they had no right to buy and use, and retrain the AI to exclude those images from generated works as well.
However, until something like that happens, Getty has every right to generate images, of course.
End result: AI outperforms humans.