Visual Art AI is Theft
December 15, 2022 3:21 PM Subscribe
Portfolio site Artstation is dealing with an ongoing protest over AI. Artists are not compensated for the use of their art for AI datasets, which charge money and are a commercial product.
Vice about this protest.
Kotaku on the same subject.
Artstation responded poorly, and made the situation worse.
Artist Lois Van Baarle on the protest.
Vice about this protest.
Kotaku on the same subject.
Artstation responded poorly, and made the situation worse.
Artist Lois Van Baarle on the protest.
I don't think we want to rely on the egregiously lawsuit happy arm of a global multinational to copyright-strike AI art into non-existence! Copyright law is a net bad for creativity and art and it is only useful as a defence for people who have the money and resources to pursue legal options.
I pointed this out recently but Mat Dryhurst and Holly Herndon have developed a tool for tracking whether your work has been used in a training set, and for opting out of training sets: https://haveibeentrained.com/
posted by Cpt. The Mango at 3:32 PM on December 15, 2022 [6 favorites]
I pointed this out recently but Mat Dryhurst and Holly Herndon have developed a tool for tracking whether your work has been used in a training set, and for opting out of training sets: https://haveibeentrained.com/
posted by Cpt. The Mango at 3:32 PM on December 15, 2022 [6 favorites]
The point is that it should be opt-in in the first place. Artists shouldn't have to take the time to chase down bad-actor AI companies.
Music AI like Stability is an example of opt-in AI.
posted by ishmael at 3:37 PM on December 15, 2022 [5 favorites]
Music AI like Stability is an example of opt-in AI.
posted by ishmael at 3:37 PM on December 15, 2022 [5 favorites]
I don't think we want to rely on the egregiously lawsuit happy arm of a global multinational to copyright-strike AI art into non-existence!
Perhaps you don't understand that any art created within the lifetime of an artist plus however many decades Disney has established as being "copyright" is under copyright, and all the art being used is stolen and in violate of copyright. Disney has the money to pursue the lawsuits. That furry artist doing art for fun or for $25 per convention badge doesn't have that, but their rights needs to be enforced too.
I don't see any governmental agencies stepping up and filing lawsuits on behalf of artists whose art is being stolen and used for AI feeding. Disney might be the only weapon we have.
posted by hippybear at 3:45 PM on December 15, 2022 [10 favorites]
Perhaps you don't understand that any art created within the lifetime of an artist plus however many decades Disney has established as being "copyright" is under copyright, and all the art being used is stolen and in violate of copyright. Disney has the money to pursue the lawsuits. That furry artist doing art for fun or for $25 per convention badge doesn't have that, but their rights needs to be enforced too.
I don't see any governmental agencies stepping up and filing lawsuits on behalf of artists whose art is being stolen and used for AI feeding. Disney might be the only weapon we have.
posted by hippybear at 3:45 PM on December 15, 2022 [10 favorites]
Moralize all you want, but violation of one's right to copy is district from theft.
posted by grokus at 3:50 PM on December 15, 2022 [2 favorites]
posted by grokus at 3:50 PM on December 15, 2022 [2 favorites]
Perhaps you didn't read the second sentence of my post? I have terrible news for furry artists and anyone who's ever doodled Micky Mouse making out with Captain America if you think that Disney enforcing copyright law will be at all equitable, or ever in favour of someone "doing art for fun or for $25 per convention badge."
posted by Cpt. The Mango at 3:52 PM on December 15, 2022 [1 favorite]
posted by Cpt. The Mango at 3:52 PM on December 15, 2022 [1 favorite]
Furry artists are generally creating original characters. They do things with existing characters (somehow especially Judy from Zootopia, amongst others), but most of the fandom is about making your OWN anthropomorphic animal and not trying to live as something someone else did for a series or whatever.
posted by hippybear at 3:56 PM on December 15, 2022 [4 favorites]
posted by hippybear at 3:56 PM on December 15, 2022 [4 favorites]
Moralize all you want, but violation of one's right to copy is district from theft.
From the comments in one of the Kotaku articles:
"if i paint something in the style of another artist, say i paint a Monet style version of the Mona Lisa, am i stealing their work or is it my creation inspired by theirs?"
That’s an invalid analogy. AI art is more like tracing a copy of somebody else’s art and changing a few details, then calling it your own art, which, by the way, is also incredibly frowned upon in the art community.
posted by ishmael at 3:58 PM on December 15, 2022 [4 favorites]
From the comments in one of the Kotaku articles:
"if i paint something in the style of another artist, say i paint a Monet style version of the Mona Lisa, am i stealing their work or is it my creation inspired by theirs?"
That’s an invalid analogy. AI art is more like tracing a copy of somebody else’s art and changing a few details, then calling it your own art, which, by the way, is also incredibly frowned upon in the art community.
posted by ishmael at 3:58 PM on December 15, 2022 [4 favorites]
Perhaps you didn't read the second sentence of my post? I have terrible news for furry artists and anyone who's ever doodled Micky Mouse making out with Captain America if you think that Disney enforcing copyright law will be at all equitable, or ever in favour of someone "doing art for fun or for $25 per convention badge."
I read that sentence.
I agree that it won't be equitable, but I also agree with hippybear that it will probably be the most effective tool against AI companies. Individual artists having to chase down each and every AI company will not get the job done.
posted by ishmael at 4:02 PM on December 15, 2022 [1 favorite]
I read that sentence.
I agree that it won't be equitable, but I also agree with hippybear that it will probably be the most effective tool against AI companies. Individual artists having to chase down each and every AI company will not get the job done.
posted by ishmael at 4:02 PM on December 15, 2022 [1 favorite]
Right but if it's not equitable... why do we want it to happen? It won't serve small struggling artists and it won't make AI go away. I personally don't have a problem with AI companies per se, I consider them technology companies like Ableton or whatever. I don't think Ableton should be using an artist's work for their product without permission either. But I certainly don't want Disney suing Ableton to protect their own IP and no one else's, thereby creating precedent for even further conglomeration and misuse of copyright law to strike down less powerful artists... And anyone who thinks that Disney is going to sue to make AI go away, as opposed to simply make it accessible only to them and no one else, and then start AI-generating 45 Marvel movies a year without paying human actors, writers, production crews, and so on, is just deluded as to what purpose copyright law serves multinational global brands.
Has anyone checked out the haveibeentrained.com link? If you have artists in your life who are concerned that AI is stealing their work, please encourage them to register there and opt out of Stable Diffusion v3 training.
posted by Cpt. The Mango at 4:17 PM on December 15, 2022 [2 favorites]
Has anyone checked out the haveibeentrained.com link? If you have artists in your life who are concerned that AI is stealing their work, please encourage them to register there and opt out of Stable Diffusion v3 training.
posted by Cpt. The Mango at 4:17 PM on December 15, 2022 [2 favorites]
There is a technique called "textual inversion" in which an image generating model can be augmented to associate a specific abstract "token" with a particular visual concept. That is, you provide the textual inversion training system with a small number of example images, say five to ten, and after an hour or so of training you have a new "word" that you can use to ask the model to make more of what you showed it. This can be a concrete thing (like you show it images of a specific person), or an abstract style, both will work.
So even when somebody produces a model trained only on vetted public domain images, it will be possible to augment that model to then reproduce whatever you want. It'll be possible to even do this without violating copyright: simply commission a human artist to make you some illustrations "in Thomas Kinkade's style", which we all agree is fine, and then feed those pieces into textual inversion. Now you have a model that can act like Thomas Kinkade without having ever seen a real Kinkade.
So I think the question of the provenance of the training data is ultimately an argument of convenience to avoid having to engage with the more difficult argument, which is that AI image generators are simply too easy, too good, and too fast, and that in our system which measures the value of everything by how competitive it is in the market, this directly threatens human artists.
posted by Pyry at 4:23 PM on December 15, 2022 [14 favorites]
So even when somebody produces a model trained only on vetted public domain images, it will be possible to augment that model to then reproduce whatever you want. It'll be possible to even do this without violating copyright: simply commission a human artist to make you some illustrations "in Thomas Kinkade's style", which we all agree is fine, and then feed those pieces into textual inversion. Now you have a model that can act like Thomas Kinkade without having ever seen a real Kinkade.
So I think the question of the provenance of the training data is ultimately an argument of convenience to avoid having to engage with the more difficult argument, which is that AI image generators are simply too easy, too good, and too fast, and that in our system which measures the value of everything by how competitive it is in the market, this directly threatens human artists.
posted by Pyry at 4:23 PM on December 15, 2022 [14 favorites]
I don't think Ableton should be using an artist's work for their product without permission either.
And yet you're opposed to the tool that forces them not to, calling it a net loss for creativity. And you argue against Disney while ignoring that the other side of the argument is backed by multinational corporations as well, who argue against copyright and other such protections for their own self interest. Disney might not be a friend to artists, but at least they argue for creators to own their work - which is more than I can say for Alphabet and Silicon Valley.
posted by NoxAeternum at 4:25 PM on December 15, 2022 [4 favorites]
And yet you're opposed to the tool that forces them not to, calling it a net loss for creativity. And you argue against Disney while ignoring that the other side of the argument is backed by multinational corporations as well, who argue against copyright and other such protections for their own self interest. Disney might not be a friend to artists, but at least they argue for creators to own their work - which is more than I can say for Alphabet and Silicon Valley.
posted by NoxAeternum at 4:25 PM on December 15, 2022 [4 favorites]
I'm definitely not FOR Alphabet and Silicon Valley, nor am I FOR whoever is backing these huge AI projects! I'm pro-artist haha. Neither Disney, Alphabet, or Silicon Valley have artist's best interest in mind, I can promise you that. If you could name me a single "creator" of a Disney IP who owns their work outright I will, I dunno, do something very drastic and improbable that demonstrates how extremely wrong I was about everything.
Also, if you think that copyright is a tool that forces large institutions to respect the rights of small artists, I have a bridge to sell you. Copyright law is, and always has been, a cudgel that only serves people rich enough to afford to use it.
I want all artists to be empowered and to own their work outright. I don't think the vast majority of current AI projects agree with me. And I know for a fact that Disney doesn't agree with me. It's possible for two things to be bad at once!
posted by Cpt. The Mango at 4:32 PM on December 15, 2022 [8 favorites]
Also, if you think that copyright is a tool that forces large institutions to respect the rights of small artists, I have a bridge to sell you. Copyright law is, and always has been, a cudgel that only serves people rich enough to afford to use it.
I want all artists to be empowered and to own their work outright. I don't think the vast majority of current AI projects agree with me. And I know for a fact that Disney doesn't agree with me. It's possible for two things to be bad at once!
posted by Cpt. The Mango at 4:32 PM on December 15, 2022 [8 favorites]
in the long run i feel like there will be a feedback loop where more and more training data for models was produced by a model in the first place. the results will be insipid and the range of output constrained. its just a shitty system to produce content when looked at as anything other than a fad. its more like a convoluted way to index and retreive unstructured data than anything creative.
posted by AlbertCalavicci at 4:37 PM on December 15, 2022 [1 favorite]
posted by AlbertCalavicci at 4:37 PM on December 15, 2022 [1 favorite]
It's a simple labor rights issue masquerading as some kind of technological conundrum Artists have the right to own the products of their own labor, and negotiate a fair value for it. To challenge those rights on the grounds that people who produce nothing of their own have a 'right to copy' is frankly appalling.
posted by Sing Or Swim at 4:37 PM on December 15, 2022 [10 favorites]
posted by Sing Or Swim at 4:37 PM on December 15, 2022 [10 favorites]
This is kind of like sampling in the '80s and '90s, but at least sampling was human beings mixing and matching snippets of copyrighted music to make new compositions. Some originality will, inevitably, enter the process when a human intelligence is involved; anything that looks like serendipity that a machine does is completely accidental. Anyway, sampling isn't legal unless you have the original artists' consent now. Without that, it's just stealing.
posted by kittens for breakfast at 4:42 PM on December 15, 2022 [2 favorites]
posted by kittens for breakfast at 4:42 PM on December 15, 2022 [2 favorites]
I think Disney is a good example: what happens when Disney trains a model on the vast amount of IP they own? Or when Google / Facebook / whoever simply buys 'AI training rights' to enough images to bootstrap their own free-and-clear model? Would we all be happy that after those foundational artists got their one-time payout, now Disney can mint an endless amount of new content without ever paying new artists again?
posted by Pyry at 4:42 PM on December 15, 2022
posted by Pyry at 4:42 PM on December 15, 2022
So much misinformation flying around about this technology. The large image generation projects (Midjourney, DALL-E, etc.) are all trained on licensed images. Nothing is being "stolen" or used without permission. Yes, if some random hobbyist wants to train a stable diffusion model at home, there's nothing stopping them from yoinking images from wherever but that's a different issue.
I don't believe that AI tools devalue art at all. It's an accessibility technology, for one thing, not a mechanization or replacement for anybody's current skills. And while it's a technical disruption so was photography and we consider that art, right? In fact, nearly every one of the complaints I read about AI assisted art would read exactly the same if you put the word "photography" in there instead of "AI."
posted by riotnrrd at 4:42 PM on December 15, 2022 [11 favorites]
I don't believe that AI tools devalue art at all. It's an accessibility technology, for one thing, not a mechanization or replacement for anybody's current skills. And while it's a technical disruption so was photography and we consider that art, right? In fact, nearly every one of the complaints I read about AI assisted art would read exactly the same if you put the word "photography" in there instead of "AI."
posted by riotnrrd at 4:42 PM on December 15, 2022 [11 favorites]
The large image generation projects (Midjourney, DALL-E, etc.) are all trained on licensed images.
*citation needed
posted by hippybear at 4:45 PM on December 15, 2022 [12 favorites]
*citation needed
posted by hippybear at 4:45 PM on December 15, 2022 [12 favorites]
So even when somebody produces a model trained only on vetted public domain images, it will be possible to augment that model to then reproduce whatever you want. It'll be possible to even do this without violating copyright: simply commission a human artist to make you some illustrations "in Thomas Kinkade's style", which we all agree is fine, and then feed those pieces into textual inversion. Now you have a model that can act like Thomas Kinkade without having ever seen a real Kinkade.
i’d say there’s something to be said for the differences in that example. namely that you paid a living, breathing artist an amount to rip someone off. and then you trained that model on that data, which means you have a model that produces work inspired by rip-offs, which is pretty different in my mind from using the model to actually extract values from the unpaid source. Still super tacky and gross, but hey, at least the model paid the artist for the data.
I’m just stunned that some toad out there thought “you know who we should fuck with? people just barely getting by on work they’ve struggled their whole life to produce.” Every day I regret computers more.
posted by gorestainedrunes at 4:46 PM on December 15, 2022 [2 favorites]
Gosh, I'm not sure what I said that makes everybody think that I believe "that people who produce nothing of their own have a 'right to copy'."
I'm simply saying that we don't want Disney sorting this out, because they won't make life better for any working artists. They will simply quite quickly replace anyone who works for Disney with whatever shitty AI model they take over with their lawsuits.
As a working artist for roughly 10 years I think I have a right to talk about whether copyright law is valuable or not. My work has been - literally - stolen. Ask me how useful copyright law was for getting me compensated! I have used samples in my work - cleared with the artist beforehand. Ask me how stressful it was to get a notice to cease & desist from the holding company that holds the publishing rights to samples that constitute over 51% of the artist who I sampled's work, which means that they hold ownership over the original artists entire song, unbeknownst to both myself and the artist! And then ask me about how delightful it was for a numbered company stationed on some sunny island to receive 100% of the profits of this song while both myself and the original artist received nothing!
Copyright law is a tool for rich companies to protect their product. It does not, has not, and will not ever stick up for, or work in service of, small artists. This should be obvious! Disney does not want to pay artists! AI companies do not want to pay artists! Spotify does not want to pay artists!
posted by Cpt. The Mango at 4:48 PM on December 15, 2022 [13 favorites]
I'm simply saying that we don't want Disney sorting this out, because they won't make life better for any working artists. They will simply quite quickly replace anyone who works for Disney with whatever shitty AI model they take over with their lawsuits.
As a working artist for roughly 10 years I think I have a right to talk about whether copyright law is valuable or not. My work has been - literally - stolen. Ask me how useful copyright law was for getting me compensated! I have used samples in my work - cleared with the artist beforehand. Ask me how stressful it was to get a notice to cease & desist from the holding company that holds the publishing rights to samples that constitute over 51% of the artist who I sampled's work, which means that they hold ownership over the original artists entire song, unbeknownst to both myself and the artist! And then ask me about how delightful it was for a numbered company stationed on some sunny island to receive 100% of the profits of this song while both myself and the original artist received nothing!
Copyright law is a tool for rich companies to protect their product. It does not, has not, and will not ever stick up for, or work in service of, small artists. This should be obvious! Disney does not want to pay artists! AI companies do not want to pay artists! Spotify does not want to pay artists!
posted by Cpt. The Mango at 4:48 PM on December 15, 2022 [13 favorites]
*citation needed
If this is anything more than lazy snark, you can read the papers where they detail the source of all their training data (the literal citations!). They're all available on arxiv.
But think about it: would Google, Meta or Apple take the enormous financial and PR risk of releasing a model trained on unlicensed, copyrighted imagery?
posted by riotnrrd at 4:50 PM on December 15, 2022 [1 favorite]
If this is anything more than lazy snark, you can read the papers where they detail the source of all their training data (the literal citations!). They're all available on arxiv.
But think about it: would Google, Meta or Apple take the enormous financial and PR risk of releasing a model trained on unlicensed, copyrighted imagery?
posted by riotnrrd at 4:50 PM on December 15, 2022 [1 favorite]
Gosh, I'm not sure what I said that makes everybody think that I believe "that people who produce nothing of their own have a 'right to copy'."
I believe this was directed at grokus's comment above.
posted by ishmael at 4:51 PM on December 15, 2022
I believe this was directed at grokus's comment above.
posted by ishmael at 4:51 PM on December 15, 2022
Phewf!
posted by Cpt. The Mango at 4:51 PM on December 15, 2022
posted by Cpt. The Mango at 4:51 PM on December 15, 2022
But think about it: would Google, Meta or Apple take the enormous financial and PR risk of releasing a model trained on unlicensed, copyrighted imagery?
If those artists are already basically nobodies who are doing art for their friends and fans and don't have any financial backing to fight for their rights? If you're a megacorp, you'd do the math, know you'd never be sued because those you're ripping off have no ability to fight you in court, and not give a shit.
posted by hippybear at 4:52 PM on December 15, 2022 [1 favorite]
If those artists are already basically nobodies who are doing art for their friends and fans and don't have any financial backing to fight for their rights? If you're a megacorp, you'd do the math, know you'd never be sued because those you're ripping off have no ability to fight you in court, and not give a shit.
posted by hippybear at 4:52 PM on December 15, 2022 [1 favorite]
Still super tacky and gross, but hey, at least the model paid the artist for the data.
I mean, is it really fair that I can pay one artist once, produce a textual inversion which I give away to everyone, and now the whole world can create more Kinkade's on a whim? It's OK as long as one person gets paid? It seems to me that this creates a race to the bottom, where the value of art is increasingly devalued, which makes it cheaper and cheaper to commission human artists for training data, and copyright does little to prevent this vicious cycle.
I think the problems with AI art are far deeper and more complicated than simply "the training data is stolen".
posted by Pyry at 4:54 PM on December 15, 2022 [7 favorites]
I mean, is it really fair that I can pay one artist once, produce a textual inversion which I give away to everyone, and now the whole world can create more Kinkade's on a whim? It's OK as long as one person gets paid? It seems to me that this creates a race to the bottom, where the value of art is increasingly devalued, which makes it cheaper and cheaper to commission human artists for training data, and copyright does little to prevent this vicious cycle.
I think the problems with AI art are far deeper and more complicated than simply "the training data is stolen".
posted by Pyry at 4:54 PM on December 15, 2022 [7 favorites]
So much misinformation flying around about this technology. The large image generation projects (Midjourney, DALL-E, etc.) are all trained on licensed images.
One of the reasons why this became an issue is because Stable Diffusion was sourcing the work of artists without their permission.
posted by ishmael at 4:57 PM on December 15, 2022 [6 favorites]
One of the reasons why this became an issue is because Stable Diffusion was sourcing the work of artists without their permission.
posted by ishmael at 4:57 PM on December 15, 2022 [6 favorites]
I mean, is it really fair that I can pay one artist once, produce a textual inversion which I give away to everyone, and now the whole world can create more Kinkade's on a whim? It's OK as long as one person gets paid?
oh i do not see fairness anywhere in this land. it is far from here. but seeing as how poorly we already treat artists, do you (the hypothetical artist) want to fight against a model fed off your rip-offs or fed directly from your table? do you believe in your art vs. a model that is like you or a model that IS you? .. with the promise that every time you publish something it will be used against you, without your permission?
in my point at least the jerk trainer has to hire somebody to continue trying to mimic you.
posted by gorestainedrunes at 5:02 PM on December 15, 2022 [1 favorite]
oh i do not see fairness anywhere in this land. it is far from here. but seeing as how poorly we already treat artists, do you (the hypothetical artist) want to fight against a model fed off your rip-offs or fed directly from your table? do you believe in your art vs. a model that is like you or a model that IS you? .. with the promise that every time you publish something it will be used against you, without your permission?
in my point at least the jerk trainer has to hire somebody to continue trying to mimic you.
posted by gorestainedrunes at 5:02 PM on December 15, 2022 [1 favorite]
Yes, but the data was licensed in good faith by Stability.AI. As that article says, they used the LAION-5B data set "which was compiled by the German nonprofit LAION." If this artist believes their art was incorrectly licensed, it is an indictment of LAION and their legal process, not of the underlying technology.
This is all very similar to the concerns around ImageNet, an early and important image dataset that was the gold standard for a while. Unfortunately, some of the component datasets (ImageNet is a collection of many datasets) were badly sourced and could cause legal problems. Thus, ImageNet has now been largely abandoned by industry. I'm sure similar things will happen with these (much larger) image datasets that are being used for stable diffusion, etc.
posted by riotnrrd at 5:02 PM on December 15, 2022 [5 favorites]
This is all very similar to the concerns around ImageNet, an early and important image dataset that was the gold standard for a while. Unfortunately, some of the component datasets (ImageNet is a collection of many datasets) were badly sourced and could cause legal problems. Thus, ImageNet has now been largely abandoned by industry. I'm sure similar things will happen with these (much larger) image datasets that are being used for stable diffusion, etc.
posted by riotnrrd at 5:02 PM on December 15, 2022 [5 favorites]
That's why all these images are gathered by arms-length non-profits, I reckon. So profit-making users of the datasets can just shrug and say "not me!"
posted by seanmpuckett at 5:07 PM on December 15, 2022 [11 favorites]
posted by seanmpuckett at 5:07 PM on December 15, 2022 [11 favorites]
Most of the modern batch of text to image ML models were trained on a data set derived from one of of the LAION image sets.
LAION Has this to say about copyright:
posted by thedward at 5:12 PM on December 15, 2022 [8 favorites]
LAION Has this to say about copyright:
Does LAION datasets respect copyright laws?AFAICT, they don't do any sort of licensing.
LAION datasets are simply indexes to the internet, i.e. lists of URLs to the original images together with the ALT texts found linked to those images. While we downloaded and calculated CLIP embeddings of the pictures to compute similarity scores between pictures and texts, we subsequently discarded all the photos. Any researcher using the datasets must reconstruct the images data by downloading the subset they are interested in. For this purpose, we suggest the img2dataset tool.
posted by thedward at 5:12 PM on December 15, 2022 [8 favorites]
riotnrrd, the LAION don't contain images, only links, nor any licenses for those images they link to.
posted by thedward at 5:15 PM on December 15, 2022 [2 favorites]
posted by thedward at 5:15 PM on December 15, 2022 [2 favorites]
My question is, if you are using say Midjourney and you ask it to produce an image of Mickey Mouse, and it does, whose ass gets beat by the Disney paddle?
posted by seanmpuckett at 5:15 PM on December 15, 2022 [1 favorite]
posted by seanmpuckett at 5:15 PM on December 15, 2022 [1 favorite]
the LAION don't contain images, only links, nor any licenses for those images they link to.
Well, THAT'S pretty fucking stupid. I spend a day filling out legal paperwork every time I publish a model. I'm gobsmacked that someone with Stability.AI's money and talent would just yolo it. Baffling.
posted by riotnrrd at 5:24 PM on December 15, 2022 [1 favorite]
Well, THAT'S pretty fucking stupid. I spend a day filling out legal paperwork every time I publish a model. I'm gobsmacked that someone with Stability.AI's money and talent would just yolo it. Baffling.
posted by riotnrrd at 5:24 PM on December 15, 2022 [1 favorite]
I think most artists (including myself) would be okay with the ML-derived generators if the dataset were opt-in or royalty-free art or just somehow not leeching off of personal data (amongst other things found in LAION... private medical photos, non-consensual pornography). Maybe there is some existential threat of "what if you could truly make the pic in your head zap into existence?" that maybe, eventually, we will reach, but now is not that time. Especially when I see img2img stuff like this that is just enabling IP laundering and banking on the average person's inability to recognize random art. If I were to lose my arms tomorrow, I would not be satisfied with this as a replacement. Nor the text2img, knowing where it comes from. I don't find this "accessible" either as there are already disabled artists with shaky hands and missing fingers who hate AI as well. There are accessibility tools that already exist. Lazy Nezumi, perspective rulers, and it's entirely possible to make 3D models entirely with a mouse (I do this often in blender and sketchup). And accessibility should not be riding on unpaid labor. Is sweatshop labor for textiles and plastics okay because you find it convenient?
And, I personally do not care about Letter of the Law debates about what is technically copyright violation or "what a human does" or what a judge rules in court. I want satisfactory outcomes for independent artists who have already struggled in this hell industry as it is.
For years I've watched people disrespect the creators of entertainment they consume--racists listening to hip hop, kpop and watching anime, people dissing pornographic art and sex work while jacking off to it, the entirety of exposure.txt entitlement to wanting free work and boasting that art isn't "a real job". And, this is like a little cherry on top of all that.
I will eat my shoe if the end result favors artists even a smidge. I expect that even if Disney or Nintendo were to be involved, the outcome would be something that suits them but still screws artists over. probably by putting the onus on artists to sue every individual person doing an Image Remix with no credit. Or even, they may not take action at all. If I make or sample Pikachu in an image generator, everyone will know it's Pikachu and Nintendo can sue as they like. If the sampling is undetectable and unrecognizable? Doesn't actually affect a big studio and their already-popular franchise. Someone makes a bootleg of a random commission artist? Oh well.
As an aside, I have a double career as a laboratory RA. Of all the things I would want automated, it's lab work. Where's my AI troubleshooting software for experiments not panning out as expected? There are already automated liquid handlers, samplers, etc.... but my workplace vetoed it because it's too expensive. It's cheaper for me to do everything manually.
So basically my creative and intellectual pursuits are being automated but my physical body remains as fodder for capitalist greed. Great outlook over here, really.
posted by picklenickle at 5:27 PM on December 15, 2022 [10 favorites]
And, I personally do not care about Letter of the Law debates about what is technically copyright violation or "what a human does" or what a judge rules in court. I want satisfactory outcomes for independent artists who have already struggled in this hell industry as it is.
For years I've watched people disrespect the creators of entertainment they consume--racists listening to hip hop, kpop and watching anime, people dissing pornographic art and sex work while jacking off to it, the entirety of exposure.txt entitlement to wanting free work and boasting that art isn't "a real job". And, this is like a little cherry on top of all that.
I will eat my shoe if the end result favors artists even a smidge. I expect that even if Disney or Nintendo were to be involved, the outcome would be something that suits them but still screws artists over. probably by putting the onus on artists to sue every individual person doing an Image Remix with no credit. Or even, they may not take action at all. If I make or sample Pikachu in an image generator, everyone will know it's Pikachu and Nintendo can sue as they like. If the sampling is undetectable and unrecognizable? Doesn't actually affect a big studio and their already-popular franchise. Someone makes a bootleg of a random commission artist? Oh well.
As an aside, I have a double career as a laboratory RA. Of all the things I would want automated, it's lab work. Where's my AI troubleshooting software for experiments not panning out as expected? There are already automated liquid handlers, samplers, etc.... but my workplace vetoed it because it's too expensive. It's cheaper for me to do everything manually.
So basically my creative and intellectual pursuits are being automated but my physical body remains as fodder for capitalist greed. Great outlook over here, really.
posted by picklenickle at 5:27 PM on December 15, 2022 [10 favorites]
Copyright law is a tool for rich companies to protect their product. It does not, has not, and will not ever stick up for, or work in service of, small artists.
This is such a ridiculous take it's hard for me to take anything you have written in this thread seriously. Copyright belongs to the creator, and that artificial monopoly on your own artistic output is the primary means you have to turn your work into money as a working artist. This is how every artist, from small to evil corporation, is able to charge money for their work through licensing. Copyright is the only thing that allows small artists to negotiate with larger commercial entities. If you've made it 10 years as a professional artist and don't understand this (or misunderstand it, as it seems) you are in desperate need of a mentor to sit you down and explain it to you for real.
The large image generation projects (Midjourney, DALL-E, etc.) are all trained on licensed images.
If that's the case, how come when I type my professional name into one of these databases of images they're using it brings up dozens of my works that I retain full rights to? Because they've structured this in a sneaky way where the the datasets and the model itself are generated by open source and educational organizations (license not needed), and then the commercial entities who fund those same organizations turn around and use the model produced and claim "the model doesn't contain any actual images so we don't need a license."
Licensing my art has been a daily part of my one-man business for over 2 decades and IMO that is some fucking bullshit.
I think a big signal that this is going to face serious legal challenges is Getty Images (no hero of mine!) has removed all works created with any AI tools or assets. I'm also of the opinion all this AI stuff will just become incorporated into the tools artists use and we'll on go on with our lives, just like when Photoshop was introduced.
posted by bradbane at 5:44 PM on December 15, 2022 [9 favorites]
This is such a ridiculous take it's hard for me to take anything you have written in this thread seriously. Copyright belongs to the creator, and that artificial monopoly on your own artistic output is the primary means you have to turn your work into money as a working artist. This is how every artist, from small to evil corporation, is able to charge money for their work through licensing. Copyright is the only thing that allows small artists to negotiate with larger commercial entities. If you've made it 10 years as a professional artist and don't understand this (or misunderstand it, as it seems) you are in desperate need of a mentor to sit you down and explain it to you for real.
The large image generation projects (Midjourney, DALL-E, etc.) are all trained on licensed images.
If that's the case, how come when I type my professional name into one of these databases of images they're using it brings up dozens of my works that I retain full rights to? Because they've structured this in a sneaky way where the the datasets and the model itself are generated by open source and educational organizations (license not needed), and then the commercial entities who fund those same organizations turn around and use the model produced and claim "the model doesn't contain any actual images so we don't need a license."
Licensing my art has been a daily part of my one-man business for over 2 decades and IMO that is some fucking bullshit.
I think a big signal that this is going to face serious legal challenges is Getty Images (no hero of mine!) has removed all works created with any AI tools or assets. I'm also of the opinion all this AI stuff will just become incorporated into the tools artists use and we'll on go on with our lives, just like when Photoshop was introduced.
posted by bradbane at 5:44 PM on December 15, 2022 [9 favorites]
Also, a bit more on the original topic--
Artstation, which is so popularly used as a prompt tag along with its artists, is basically a LinkedIn for artists to find work. So even if AI were to become the new norm of art, it doesn't make sense to flood the place with AI images or use it as a scraping hub. Who's gonna be able to find and hire Greg Rutkowski amongst all the "epic, dragon, fantasy, fire, trending on artstation, Greg Rutkowski" images?
posted by picklenickle at 6:03 PM on December 15, 2022 [1 favorite]
Artstation, which is so popularly used as a prompt tag along with its artists, is basically a LinkedIn for artists to find work. So even if AI were to become the new norm of art, it doesn't make sense to flood the place with AI images or use it as a scraping hub. Who's gonna be able to find and hire Greg Rutkowski amongst all the "epic, dragon, fantasy, fire, trending on artstation, Greg Rutkowski" images?
posted by picklenickle at 6:03 PM on December 15, 2022 [1 favorite]
I'm also of the opinion all this AI stuff will just become incorporated into the tools artists use and we'll on go on with our lives, just like when Photoshop was introduced.
Speaking for myself, I'm not gonna use it if it's still trained off of copyrighted images. Royalty-free or volunteered, maybe. Clip Studio already had some AI tools like an auto-coloring feature, though it wasn't very good and artists deemed it as only useful for testing color palettes out at best.
There are things that *could* be useful--not text2img or img2img but things like an auto-flatter that does a layer per color, something to get rid of the lingering transparent pixels after flatting, some kind of face-realignment or other smart realignment so that images look normal when flipped, auto retopology for 3D, smarter tweening for animation... I would love all those but I'm not riding on other people's stolen art and I'm *especially* not gonna pay for stolen art. It would be like chopping off my hair without consent and trying to sell it back to me as a wig.
However I actually fear that because of this, the well has been poisoned for AI and they'd have to at least package it under a different term, and be transparent about how the tool was made.
posted by picklenickle at 7:10 PM on December 15, 2022 [2 favorites]
Speaking for myself, I'm not gonna use it if it's still trained off of copyrighted images. Royalty-free or volunteered, maybe. Clip Studio already had some AI tools like an auto-coloring feature, though it wasn't very good and artists deemed it as only useful for testing color palettes out at best.
There are things that *could* be useful--not text2img or img2img but things like an auto-flatter that does a layer per color, something to get rid of the lingering transparent pixels after flatting, some kind of face-realignment or other smart realignment so that images look normal when flipped, auto retopology for 3D, smarter tweening for animation... I would love all those but I'm not riding on other people's stolen art and I'm *especially* not gonna pay for stolen art. It would be like chopping off my hair without consent and trying to sell it back to me as a wig.
However I actually fear that because of this, the well has been poisoned for AI and they'd have to at least package it under a different term, and be transparent about how the tool was made.
posted by picklenickle at 7:10 PM on December 15, 2022 [2 favorites]
Copyright law is a tool for rich companies to protect their product. It does not, has not, and will not ever stick up for, or work in service of, small artists.
This is such a ridiculous take it's hard for me to take anything you have written in this thread seriously.
There's a significant difference between what copyright was intended to be and what it has become. Originally, a copyright term lasted for 14 years, with one renewable 14-year term. Nowadays, specifically due to lobbying by companies like Disney, it is 95 years from publication or 120 years from creation. That is definitely because of rich companies protecting profits.
posted by CheeseDigestsAll at 7:18 PM on December 15, 2022 [7 favorites]
This is such a ridiculous take it's hard for me to take anything you have written in this thread seriously.
There's a significant difference between what copyright was intended to be and what it has become. Originally, a copyright term lasted for 14 years, with one renewable 14-year term. Nowadays, specifically due to lobbying by companies like Disney, it is 95 years from publication or 120 years from creation. That is definitely because of rich companies protecting profits.
posted by CheeseDigestsAll at 7:18 PM on December 15, 2022 [7 favorites]
Speaking for myself, I'm not gonna use it if it's still trained off of copyrighted images.
Sure, but it seems like the most obvious single-artist use of this type of technology would be training it on your own work and portfolio. Maybe I input my life's archive and that colorize thing is more useful when it's trained on my own color palette, or Adobe Creative Suite's tools get better at anticipating the changes I'd make or speed up tedious parts of the workflow. Just like any other digital tool.
Having the algo spit out a mish mash of other people's stuff isn't interesting or useful to me either. However there are some other AI tools that looks extremely promising for the kinds of stuff I produce - AI sharpening, interpolation, etc.
Most of the hand wringing and moral panic I see about AI tools seems highly concentrated in hobbyist type spaces. I don't mean that in a pejorative sense, just that professionals provide a service and in order for something like Midjourney to be a service, it would require better tooling and a highly skilled operator to deliver on time/on budget and then you're right back to where you started... you need a professional artist (and what tools they use are irrelevant).
posted by bradbane at 7:45 PM on December 15, 2022 [3 favorites]
Sure, but it seems like the most obvious single-artist use of this type of technology would be training it on your own work and portfolio. Maybe I input my life's archive and that colorize thing is more useful when it's trained on my own color palette, or Adobe Creative Suite's tools get better at anticipating the changes I'd make or speed up tedious parts of the workflow. Just like any other digital tool.
Having the algo spit out a mish mash of other people's stuff isn't interesting or useful to me either. However there are some other AI tools that looks extremely promising for the kinds of stuff I produce - AI sharpening, interpolation, etc.
Most of the hand wringing and moral panic I see about AI tools seems highly concentrated in hobbyist type spaces. I don't mean that in a pejorative sense, just that professionals provide a service and in order for something like Midjourney to be a service, it would require better tooling and a highly skilled operator to deliver on time/on budget and then you're right back to where you started... you need a professional artist (and what tools they use are irrelevant).
posted by bradbane at 7:45 PM on December 15, 2022 [3 favorites]
Stable Diffusion can be run locally on your own PC and a 6 year old graphics card.
The model is 4GB in size, condensed from training the neural network on 2 billion images. It's not actively scraping the entire 2 billion images on the web each time you give it a prompt.
It would be hard to argue that the 4GB model itself is infringing. It would be like arguing that, after reading a book, and I wrote 10 pages on my blog analyzing the character motivations, the use of themes, summarizing the major plot points, and describing the general writing style, that those 10 page are infringing / stealing from the author because it's a summary that could allow someone, in theory, to create something similar to the book I read.
posted by xdvesper at 7:56 PM on December 15, 2022 [9 favorites]
The model is 4GB in size, condensed from training the neural network on 2 billion images. It's not actively scraping the entire 2 billion images on the web each time you give it a prompt.
It would be hard to argue that the 4GB model itself is infringing. It would be like arguing that, after reading a book, and I wrote 10 pages on my blog analyzing the character motivations, the use of themes, summarizing the major plot points, and describing the general writing style, that those 10 page are infringing / stealing from the author because it's a summary that could allow someone, in theory, to create something similar to the book I read.
posted by xdvesper at 7:56 PM on December 15, 2022 [9 favorites]
I think it's a fool's errand to try and stop it. I see no difference between an AI that has been 'trained' on a lot of copyrighted art and an Artist who has 'looked at/learned from' a lot of copyrighted art.
At best they'll be able to train the AI to not produce final works that are copyright infringing and then people will stop using it because ArtistIntellyArter refuses to create images in their favourite style.
posted by gible at 8:09 PM on December 15, 2022 [5 favorites]
At best they'll be able to train the AI to not produce final works that are copyright infringing and then people will stop using it because ArtistIntellyArter refuses to create images in their favourite style.
posted by gible at 8:09 PM on December 15, 2022 [5 favorites]
I think it's a fool's errand to try and stop it. I see no difference between an AI that has been 'trained' on a lot of copyrighted art and an Artist who has 'looked at/learned from' a lot of copyrighted art.
This has been my feeling on the matter as well. I realize there are some confounding factors that do make it different. But I really, I can’t see it as much different than the process a human brain has; albeit more linear and with less feeling. But overall, if AI is infringing because of its source material, then we all are.
Heck, as a once web designer, I had on more than a single occasion sat down with the goal of replicating a design trend to match a certain experience for the website; and gathered a bunch of examples, studied what made them what they were, and made something entirely original and yet replicating the style I had been studying. Some was a conscious process, some was going on outside of my awareness. No artist was consulted or consented to my use of their training data.
And what are inspiration boards if not more training data?
I mean, I’m sure AI isn’t to the point where it gets excited because it’s making glass buttons so shiny that they look tasty and it tickled my brain so much to have something I was making look delicious visually. So it IS a different thing. But by how much?
posted by [insert clever name here] at 1:23 AM on December 16, 2022 [1 favorite]
This has been my feeling on the matter as well. I realize there are some confounding factors that do make it different. But I really, I can’t see it as much different than the process a human brain has; albeit more linear and with less feeling. But overall, if AI is infringing because of its source material, then we all are.
Heck, as a once web designer, I had on more than a single occasion sat down with the goal of replicating a design trend to match a certain experience for the website; and gathered a bunch of examples, studied what made them what they were, and made something entirely original and yet replicating the style I had been studying. Some was a conscious process, some was going on outside of my awareness. No artist was consulted or consented to my use of their training data.
And what are inspiration boards if not more training data?
I mean, I’m sure AI isn’t to the point where it gets excited because it’s making glass buttons so shiny that they look tasty and it tickled my brain so much to have something I was making look delicious visually. So it IS a different thing. But by how much?
posted by [insert clever name here] at 1:23 AM on December 16, 2022 [1 favorite]
I see no difference between an AI that has been 'trained' on a lot of copyrighted art and an Artist who has 'looked at/learned from' a lot of copyrighted art.
The difference is there, but you're limiting your comparison to some theoretical plane of process, so you miss it: deep learning algorithms are running on computers, while human artists are human beings.
This means, among other things, that there are vastly different cost structures and thus economic consequences to these different methods of visual image production. We may think the benefits to society of AI image generation outweigh the costs, but if we do we should say that instead of ignoring the obvious. We try not to let kids make our clothes because we think it's bad for the kids and for society, even though a child making a shirt and an adult making a shirt are pretty indistinguishable on an abstract process level. Copyright itself is a completely made-up rule based on (heavily lobbied for) outcomes; process concerns only enter in as ways to affect those outcomes.
In other words, "I don't understand the difference between a synthetic neutral network and a human brain on some abstract level" is a bad reason to throw up or hands and ignore any potential real world consequences.
(Apologies if this comes off as harsh, I found myself having a visceral reaction to the comment.)
posted by ropeladder at 4:51 AM on December 16, 2022 [12 favorites]
The difference is there, but you're limiting your comparison to some theoretical plane of process, so you miss it: deep learning algorithms are running on computers, while human artists are human beings.
This means, among other things, that there are vastly different cost structures and thus economic consequences to these different methods of visual image production. We may think the benefits to society of AI image generation outweigh the costs, but if we do we should say that instead of ignoring the obvious. We try not to let kids make our clothes because we think it's bad for the kids and for society, even though a child making a shirt and an adult making a shirt are pretty indistinguishable on an abstract process level. Copyright itself is a completely made-up rule based on (heavily lobbied for) outcomes; process concerns only enter in as ways to affect those outcomes.
In other words, "I don't understand the difference between a synthetic neutral network and a human brain on some abstract level" is a bad reason to throw up or hands and ignore any potential real world consequences.
(Apologies if this comes off as harsh, I found myself having a visceral reaction to the comment.)
posted by ropeladder at 4:51 AM on December 16, 2022 [12 favorites]
I reactivated my account just to comment on this thread!
As a working artist for roughly 10 years I think I have a right to talk about whether copyright law is valuable or not. My work has been - literally - stolen. Ask me how useful copyright law was for getting me compensated! I have used samples in my work - cleared with the artist beforehand. Ask me how stressful it was to get a notice to cease & desist from the holding company that holds the publishing rights to samples that constitute over 51% of the artist who I sampled's work, which means that they hold ownership over the original artists entire song, unbeknownst to both myself and the artist! And then ask me about how delightful it was for a numbered company stationed on some sunny island to receive 100% of the profits of this song while both myself and the original artist received nothing!
That sucks, but it is clear that neither of you did your due diligence. The PROs all let you look up that information, and it is a big fail on the artist's part to not know their own stake in the publishing.
Copyright law is a tool for rich companies to protect their product. It does not, has not, and will not ever stick up for, or work in service of, small artists. This should be obvious! Disney does not want to pay artists! AI companies do not want to pay artists! Spotify does not want to pay artists!
As others have said, this is 100% incorrect. Copyright is how thousands upon thousands of indie artists get paid for having their music on TV shows, films, etc. I've been getting paid for my music for years and years, and I'm not a big rich artist, I have zero leverage. ASCAP does the legwork for me and gets me my money from cue sheets that are submitted by licensors. Why? Because copyright.
posted by grumpybear69 at 6:41 AM on December 16, 2022 [4 favorites]
As a working artist for roughly 10 years I think I have a right to talk about whether copyright law is valuable or not. My work has been - literally - stolen. Ask me how useful copyright law was for getting me compensated! I have used samples in my work - cleared with the artist beforehand. Ask me how stressful it was to get a notice to cease & desist from the holding company that holds the publishing rights to samples that constitute over 51% of the artist who I sampled's work, which means that they hold ownership over the original artists entire song, unbeknownst to both myself and the artist! And then ask me about how delightful it was for a numbered company stationed on some sunny island to receive 100% of the profits of this song while both myself and the original artist received nothing!
That sucks, but it is clear that neither of you did your due diligence. The PROs all let you look up that information, and it is a big fail on the artist's part to not know their own stake in the publishing.
Copyright law is a tool for rich companies to protect their product. It does not, has not, and will not ever stick up for, or work in service of, small artists. This should be obvious! Disney does not want to pay artists! AI companies do not want to pay artists! Spotify does not want to pay artists!
As others have said, this is 100% incorrect. Copyright is how thousands upon thousands of indie artists get paid for having their music on TV shows, films, etc. I've been getting paid for my music for years and years, and I'm not a big rich artist, I have zero leverage. ASCAP does the legwork for me and gets me my money from cue sheets that are submitted by licensors. Why? Because copyright.
posted by grumpybear69 at 6:41 AM on December 16, 2022 [4 favorites]
And what are inspiration boards if not more training data?
As I pointed out upthread, this is not the case.
An artist taking inspiration from another artist create marked differences from the source material.
The reason why AI is an issue is because artists have been finding AI-art that is clearly copying their work with tiny changes. Sometimes, slightly warped and erased traces of signatures are still in the AI images.
In any case, the point of all this is that the artists are protesting against the portfolio platform that is enabling the scraping. There has been zero effort on the platform's part to address this, no meta-data or watermarks or opt-in or anything.
posted by ishmael at 7:24 AM on December 16, 2022 [1 favorite]
As I pointed out upthread, this is not the case.
An artist taking inspiration from another artist create marked differences from the source material.
The reason why AI is an issue is because artists have been finding AI-art that is clearly copying their work with tiny changes. Sometimes, slightly warped and erased traces of signatures are still in the AI images.
In any case, the point of all this is that the artists are protesting against the portfolio platform that is enabling the scraping. There has been zero effort on the platform's part to address this, no meta-data or watermarks or opt-in or anything.
posted by ishmael at 7:24 AM on December 16, 2022 [1 favorite]
"I don't understand the difference between a synthetic neutral network and a human brain on some abstract level" is a bad reason to throw up or hands and ignore any potential real world consequences.
Yes, thank you for putting this into words. To me it's like the difference between someone seeing your outfit and being inspired vs someone going around taking pics of everyone's bodies without permission and putting the photos in a fashion magazine. Isn't it basically the same because in both instances someone is looking at your publicly viewable clothes with their eyeballs? Even if I dressed up nicely to be admired in public, I think it would be creepy and violating. It's creepy and violating if a drone does it. It's creepy and violating if it were taken from security camera footage. In fact there are cosplayers and makeup artists who are annoyed that their pics are in the LAION database as well.
I feel like we have certain public etiquette when interacting in person but for some reason when new technology is involved, all etiquette goes flying out the window. Particularly where privacy and data ownership is concerned. If photocopiers were invented today and marketed as a "New Art Robot! Perfect at ink drawings of exactly what it sees!" I imagine people would eat that up as well. Isn't my Epson Inkjet ArtRobot printout of this Disney Book Cover basically just a really accurate still life painting?
The ramifications of this reach further than just art inspiration as well--this same technology can be used for new forms of deepfakes and I don't think "Actually this is no different than actors in makeup on TV!" is going to cut it. I feel like the eventual legal outcomes of this will set a precedent for quite a lot of other AI regulations as well as data usage and scraping in general.
posted by picklenickle at 8:20 AM on December 16, 2022 [2 favorites]
Yes, thank you for putting this into words. To me it's like the difference between someone seeing your outfit and being inspired vs someone going around taking pics of everyone's bodies without permission and putting the photos in a fashion magazine. Isn't it basically the same because in both instances someone is looking at your publicly viewable clothes with their eyeballs? Even if I dressed up nicely to be admired in public, I think it would be creepy and violating. It's creepy and violating if a drone does it. It's creepy and violating if it were taken from security camera footage. In fact there are cosplayers and makeup artists who are annoyed that their pics are in the LAION database as well.
I feel like we have certain public etiquette when interacting in person but for some reason when new technology is involved, all etiquette goes flying out the window. Particularly where privacy and data ownership is concerned. If photocopiers were invented today and marketed as a "New Art Robot! Perfect at ink drawings of exactly what it sees!" I imagine people would eat that up as well. Isn't my Epson Inkjet ArtRobot printout of this Disney Book Cover basically just a really accurate still life painting?
The ramifications of this reach further than just art inspiration as well--this same technology can be used for new forms of deepfakes and I don't think "Actually this is no different than actors in makeup on TV!" is going to cut it. I feel like the eventual legal outcomes of this will set a precedent for quite a lot of other AI regulations as well as data usage and scraping in general.
posted by picklenickle at 8:20 AM on December 16, 2022 [2 favorites]
I've been experimenting with SD and Dreambooth and I'm honestly not sure what can be done because those things are already out there even if future models are locked down (which Stability AI has been trying to do without much success so far). With Dreambooth someone can take 10-20 pictures and produce pictures "close" to the original even if the model didn't have them.
I've seen some artists suggest that every picture from now on should be accompanied with a timelapse of how it was made, or a file with layers, to prove they weren't AI generated and both solutions seem ludicrous to me.
posted by simmering octagon at 8:56 AM on December 16, 2022
I've seen some artists suggest that every picture from now on should be accompanied with a timelapse of how it was made, or a file with layers, to prove they weren't AI generated and both solutions seem ludicrous to me.
posted by simmering octagon at 8:56 AM on December 16, 2022
I think the eventual legal outcome is that the big companies will build training databases of known provenance, likely by forcing TOS terms onto their social networks of the form "you grant a perpetual, royalty-free license to any content you upload to be used for the purpose of training ML models". Like, if Twitter or Instagram put that in then people will complain but let's be honest only a tiny fraction will actually leave.
posted by Pyry at 9:12 AM on December 16, 2022
posted by Pyry at 9:12 AM on December 16, 2022
Man, this thread is depressing. The rules-lawyering takes about why this is okay just totally miss the point. It's especially tough hearing them from another artist!
Sharing art online was hard enough when people would literally just repost all your work on a different account and say it was theirs to get followers. It was hard enough with uncredited reposts on Pinterest and Instagram and people printing your art on Redbubble stuff to make money. It was hard enough with clients messaging you about commissioned work, only to berate you and tear you and your work down when you gave them a quote because they think they should get commerically usable custom artwork from someone for poverty wages. That's the kind of person I think of when I think about the driving force behind AI art--someone who wants something for nothing and doesn't care about the humans whose shoulders they're standing on to reach it. I had a friend tell me I was right to feel upset but that eventually I would have to get back to the business of sharing myself with the world and you know, I don't think I actually will. I'm tired, y'all. I think I'm shuttering my art accounts and taking my toys away.
posted by the liquid oxygen at 9:27 AM on December 16, 2022 [7 favorites]
Sharing art online was hard enough when people would literally just repost all your work on a different account and say it was theirs to get followers. It was hard enough with uncredited reposts on Pinterest and Instagram and people printing your art on Redbubble stuff to make money. It was hard enough with clients messaging you about commissioned work, only to berate you and tear you and your work down when you gave them a quote because they think they should get commerically usable custom artwork from someone for poverty wages. That's the kind of person I think of when I think about the driving force behind AI art--someone who wants something for nothing and doesn't care about the humans whose shoulders they're standing on to reach it. I had a friend tell me I was right to feel upset but that eventually I would have to get back to the business of sharing myself with the world and you know, I don't think I actually will. I'm tired, y'all. I think I'm shuttering my art accounts and taking my toys away.
posted by the liquid oxygen at 9:27 AM on December 16, 2022 [7 favorites]
I've been thinking about making a guide about how to defend your art from future models and dreambooth, because I've seen a lot of misinformation about the topic.
The disney thing and filling artstation won't do much, both things can be automatically filtered against, other than annoy artstation (which may be the real objective, but a ton of people think they are directly harming the models which they aren't)
I saw an artist recommending resizing your images and that also won't work unless you make things so small that the model can't pick up details at 512x512 (or 768x768 for SD 2) and that means people also won't be able to.
Some solutions also only work for the crawling for the initial training, like using javascript's canvas instead of image files. But dreambooth (and embeddings ) is really hard to defend against since manually picking, downloading/screnshotting, and editing 20-30 images from an artist and tagging them takes maybe an hour or two, including whatever edits you need to make.
Saving mediocre-quality JPGs also doesn't work because there are AI upscalers. They are not perfect but they just need to be good enough to be able to upscale an image enough that when resized back to 512x512 it looks sharp.
The only thing I've seen that can work is huge watermarks. And they need to be watermarks that are hard to edit or content-aware fill. But again, it all depends on how determined the person and how much time they have to edit your stuff.
posted by simmering octagon at 9:35 AM on December 16, 2022
The disney thing and filling artstation won't do much, both things can be automatically filtered against, other than annoy artstation (which may be the real objective, but a ton of people think they are directly harming the models which they aren't)
I saw an artist recommending resizing your images and that also won't work unless you make things so small that the model can't pick up details at 512x512 (or 768x768 for SD 2) and that means people also won't be able to.
Some solutions also only work for the crawling for the initial training, like using javascript's canvas instead of image files. But dreambooth (and embeddings ) is really hard to defend against since manually picking, downloading/screnshotting, and editing 20-30 images from an artist and tagging them takes maybe an hour or two, including whatever edits you need to make.
Saving mediocre-quality JPGs also doesn't work because there are AI upscalers. They are not perfect but they just need to be good enough to be able to upscale an image enough that when resized back to 512x512 it looks sharp.
The only thing I've seen that can work is huge watermarks. And they need to be watermarks that are hard to edit or content-aware fill. But again, it all depends on how determined the person and how much time they have to edit your stuff.
posted by simmering octagon at 9:35 AM on December 16, 2022
Wait till the AIs are told to find more ways to profit through financial finagling. If they're not already...
posted by sneebler at 10:12 AM on December 16, 2022
posted by sneebler at 10:12 AM on December 16, 2022
It's ok everyone, ChatGPT has figured out a solution to all these problems:
Can you rewrite the poem 'This Is Just To Say', as if it were by WH Auden?posted by Gerald Bostock at 12:12 PM on December 16, 2022 [4 favorites]
Sure, I can try to rewrite the poem "This Is Just To Say" as if it were written by W.H. Auden. Here is one possible version:
I have eaten
the plums
that were in
the icebox
and which
you were probably
saving
for breakfast
Forgive me
they were delicious
so sweet
and so cold
Regards,
W.H. Auden
From Sarah Guzzardo on Twitter (via Paul F Tompkins):
We figured out a way to use technology to render the human creative impulse obsolete before figuring out a way to use technology to render any type of large-scale suffering obsolete, and that is actually… Peak Human. ☺️ A+ note to go out on, we did it guys
posted by ishmael at 1:25 PM on December 16, 2022 [1 favorite]
We figured out a way to use technology to render the human creative impulse obsolete before figuring out a way to use technology to render any type of large-scale suffering obsolete, and that is actually… Peak Human. ☺️ A+ note to go out on, we did it guys
posted by ishmael at 1:25 PM on December 16, 2022 [1 favorite]
I don't want to jump into the implications of AI art, because that would mean going on for way longer than I have any desire to do when typing on my phone... but I want to correct a couple of misconceptions that I saw in the thread.
First of all, someone made a reference to the "AI" pictures composed entirely of "NO AI" logos. Those weren't real, they were done as a prank! AI doesn't work that way, and even if it did it's generally using datasets of images that are at least a couple of years old. You can't make it malfunction by posting a bunch of mis tagged images on a website.
Second, the AI images that looked almost exactly like existing images? Those were mostly img2img generations where they plugged in the image and told the AI to make slight alterations. The users were doing that on purpose. AI has the potential for that kind of theft, but no more than pencils allow for tracing, or Photoshop allows you to apply a filter to someone else's image. And the vast majority of creations aren't doing that kind of direct copying. Even img2img is mostlu used to create images that are just vaguely reminiscent of the input image.
There is the potential for something called overfitting leading to copies. Overfitting is most common when a single image, or slight variations on it, occur dozens or hundreds of times in the dataset. For instance, typing in Mona Lisa will usually output slight variations on the painting. But that's generally not how AI generation works! You can only of that with very famous and frequently repeated images, not just any random image that was posted by some random artist. Also, AI software like Midjourney are taking steps to eliminate that issue with copyrighted art.
A lot of people are confusing the potentially problematic training of AI on copyrighted works with the idea that AI is mostly just making slight variations of those works. It's capable of making new images that don't even closely resemble any existing work!
posted by ThisIsAThrowaway at 4:38 PM on December 17, 2022 [1 favorite]
First of all, someone made a reference to the "AI" pictures composed entirely of "NO AI" logos. Those weren't real, they were done as a prank! AI doesn't work that way, and even if it did it's generally using datasets of images that are at least a couple of years old. You can't make it malfunction by posting a bunch of mis tagged images on a website.
Second, the AI images that looked almost exactly like existing images? Those were mostly img2img generations where they plugged in the image and told the AI to make slight alterations. The users were doing that on purpose. AI has the potential for that kind of theft, but no more than pencils allow for tracing, or Photoshop allows you to apply a filter to someone else's image. And the vast majority of creations aren't doing that kind of direct copying. Even img2img is mostlu used to create images that are just vaguely reminiscent of the input image.
There is the potential for something called overfitting leading to copies. Overfitting is most common when a single image, or slight variations on it, occur dozens or hundreds of times in the dataset. For instance, typing in Mona Lisa will usually output slight variations on the painting. But that's generally not how AI generation works! You can only of that with very famous and frequently repeated images, not just any random image that was posted by some random artist. Also, AI software like Midjourney are taking steps to eliminate that issue with copyrighted art.
A lot of people are confusing the potentially problematic training of AI on copyrighted works with the idea that AI is mostly just making slight variations of those works. It's capable of making new images that don't even closely resemble any existing work!
posted by ThisIsAThrowaway at 4:38 PM on December 17, 2022 [1 favorite]
img2img is actually, to me, way more interesting than text2image, in terms of recovering / enhancing old photos and images.
As I mentioned, I have Stable Diffusion running on my PC, with the 1.4 model based on the LAION dataset. I then use ERSGAN (Enhanced Super Resolution Generative Adversarial Networks) with the Remacri model (there are literally dozens of models available).
This is the result - on the left is the original low resolution lossy image, and Stable Diffusion creates a high resolution interpretation of the hair and tattoo. This is a zoomed in portion, the whole image is much larger. This is literally within a few hours of setting it up, I am sure a person proficient in this and familiar with the dozens of datasets and models available would be able to do a far better job.
This is one of those things where a person who is truly proficient with the available tools would be able to work literal magic, and do something in seconds that would have previously taken days or hours to do.
posted by xdvesper at 11:14 PM on December 17, 2022
As I mentioned, I have Stable Diffusion running on my PC, with the 1.4 model based on the LAION dataset. I then use ERSGAN (Enhanced Super Resolution Generative Adversarial Networks) with the Remacri model (there are literally dozens of models available).
This is the result - on the left is the original low resolution lossy image, and Stable Diffusion creates a high resolution interpretation of the hair and tattoo. This is a zoomed in portion, the whole image is much larger. This is literally within a few hours of setting it up, I am sure a person proficient in this and familiar with the dozens of datasets and models available would be able to do a far better job.
This is one of those things where a person who is truly proficient with the available tools would be able to work literal magic, and do something in seconds that would have previously taken days or hours to do.
posted by xdvesper at 11:14 PM on December 17, 2022
Petapixel: Midjourney Founder Admits to Using a ‘Hundred Million’ Images Without Consent
Twitter users have been sharing an interview that Holz did with Forbes back in September in which he readily confesses to using images that he didn’t have permission for.
When asked: “Did you seek consent from living artists or work still under copyright?”
Holz replies: “No. There isn’t really a way to get a hundred million images and know where they’re coming from.
“It would be cool if images had metadata embedded in them about the copyright owner or something. But that’s not a thing; there’s not a registry.
“There’s no way to find a picture on the internet, and then automatically trace it to an owner and then have any way of doing anything to authenticate it.”
posted by cendawanita at 9:47 PM on December 21, 2022 [1 favorite]
Twitter users have been sharing an interview that Holz did with Forbes back in September in which he readily confesses to using images that he didn’t have permission for.
When asked: “Did you seek consent from living artists or work still under copyright?”
Holz replies: “No. There isn’t really a way to get a hundred million images and know where they’re coming from.
“It would be cool if images had metadata embedded in them about the copyright owner or something. But that’s not a thing; there’s not a registry.
“There’s no way to find a picture on the internet, and then automatically trace it to an owner and then have any way of doing anything to authenticate it.”
posted by cendawanita at 9:47 PM on December 21, 2022 [1 favorite]
« Older Luddite’ Teens Don’t Want Your Likes | "The common good stands as a menace to the status... Newer »
This thread has been archived and is closed to new comments
posted by hippybear at 3:25 PM on December 15, 2022 [14 favorites]