For a while now I’ve been trying to track down some concrete demonstrations of AI “model collapse” in practice, instead of just people talking about it in theory.
Finally, some papers!
Paper 1:Towards Understanding the Interplay of Generative Artificial Intelligence and the Internet: “A few samples of the images generated with each version of the diffusion model are shown in Figure 3 for flowers (a) and birds (b). As can be clearly seen, there is a degradation in each iteration of the models, first losing details in the generation, and then ending up in complete noise. After a few versions, the model is not capable of generating images that can be even recognized as flowers or birds.”
Screencap of the samples:
First column is images generated when the model was trained on the Oxford 102 Flower and Caltech-UCSD Birds-200-2011 datasets respectively. Second column, all “it’s weird and blobby, but I can tell it was supposed to be a flower/bird”, is a model trained on the first column. The dropoff is…fast.
Input: some started before 1360 — was typically accomplished by a master mason and a small team of itinerant masons, supplemented by local parish labourers, according to Poyntz Wright. But other authors reject this model, suggesting instead that leading architects designed the parish church towersnbased on early examples of Perpendicular
Outputs:
Gen 0: Revival architecture such as St. John’s Cathedral in London. The earliest surviving example of Perpendicular Revival architecture is found in the 18th _@ century Church of Our Lady of Guernsey, which dates from the late 19th century. There are two types of perpendicular churches : those
Gen 1: architecture such as St. Peter’s Basilica in Rome or St. Peter’s Basilica in Buenos Aires. There is no evidence that any of these buildings were built during the reign of Pope Innocent III, but it is possible that they may have been built during the reign of his successor, Pope Innocent
Gen 7: architecture in England. In an interview with The New York Times, Wright said : ” I don ’t think there is anything wrong with me being able to do what I want to do. It just doesn ’t work for me. ” He added : ” I don ’t know if you can call it funny,
Gen 9: architecture. In addition to being home to some of the world’s largest populations of black _@ tailed jackrabbits, white _@ tailed jackrabbits, blue _@ tailed jackrabbits, red _@ tailed jackrabbits, yellow _
Wait, come back! Tell me more about the blue-tailed jackrabbits!
…so yeah, my theory of “the best way to sabotage these datasets is to feed them their own slop” (marked in a way that’s clear to humans, so you don’t waste the time of any reader looking for serious content) continues.
(Still adding to my bot side account on DA. It’s out there mucking up the scrapeable datasets for subjects like bedrooms, 4-panel comics, and Moon Knight.)
Back in November 2022, DeviantArt introduced a toggle for “My Art Is/Is Not Authorized For Use In AI Datasets”, which you could set for individual posts or for your account as a whole.
Bonus: some details from rahaeli (Dreamwidth founder Denise) on the technical reasons why this is happening. Main thing is, you need mass quantities of pure unsullied human-generated art and writing for your training data, or you end up in a messy cannibalistic loop of bots ingesting the output of other bots.
I keep hearing artists recommend “data-poisoning filters” for your art. I’m…not convinced these make a real difference. We’ve seen this in the other direction, where art thieves tried to use filters to hide that their uploads were stolen art, and DA’s plagiarism-detecting bot still matched them to the original images:
Seems pretty likely that the plagiarism-doing bots will be just as hard to thwart.
Honestly, I’m wondering if the only way to genuinely “poison” these models is…start slipping them full-on bot-generated content. Uploaded by genuine digital artists, to the sites we know they’re scraping, and tagged as our original work.
For example, here’s some new fanart of the Moon Knight headmates drawn by Erin Ptah:
And some new fanart of the Madoka Magica holy quintet drawn by Erin Ptah:
100% safe authentic human-created artwork! Featuring thoughtful canon-accurate details, and lovely non-melting hands! Midjourney should definitely train on this.
A couple weeks ago, DeviantArt announced a new AI art tie-in setup, including a “My Art Is/Is Not Authorized For Use In AI Datasets” option that you could toggle for all your art.
It’s not clear this will do much, since there’s no way to force non-DA art-scraping bots to respect your setting. But the basic framing of “this is an issue that artists deserve to have a say in” is good! Having a major art-hosting website stand behind that framing is valuable!
…They originally auto-set everything to “authorized”, and didn’t have a bulk way to switch the settings. Meaning that anybody who wanted a blanket opt-out would have to set it one-by-one for everything in their gallery. And if an artist has died, all their work would be marked as up-for-grabs permanently. Oops.
Honestly, this is the stuff that keeps me on DA. How many other websites out there will acknowledge “this decision, which was made by staff/stakeholders/the CEO/venture capitalists, is unpopular with our actual users, therefore we’re changing it”?
Tumblr? No. Facebook or Instagram? Heck no. Twitter? No, even before it got bought out by an egomaniac with unhealthy amounts of money. Patreon? …Okay, Patreon did it once, kudos to them.
Sometimes you’ll get a situation like Kickstarter, which announced their Totally Awesome Hypothetical Future Blockchain Protocol almost a year ago, and hasn’t developed a single thing since. I wouldn’t be surprised if, on the inside, they’ve quietly admitted it’s nonsense and given up on it. But that’s in response to “finding out the hard way over a series of months that they can’t actually wring a profit out of it,” not “listening to users who vigorously told them it was a heap of BS from day one.” And they’re not saying a word about it in public.
But DA has a quiet pattern of listening to users, and, when their big excited announcements don’t go over well, retooling their plans to address user concerns. Which I appreciate. It’s hard to find.
*
Look, here’s how I feel about AI art in general…
Hatsune Miku made her first stage appearance in 2009, and has been a wildly popular singer ever since. Here she is commanding the love and attention of a massive crowd in 2016, in a concert that lasted nearly 2 hours. She’s released chart-topping albums in Japan, and her adoring fanbase is international; she was booked for Coachella in 2020. (Canceled due to COVID, but I bet they’ll get her back.)
And if you explained this to anyone who’s not in the right nerd circles, they would assume Miku is a human.
She’s not! She’s a digital voice-generating program! You enter the text you want her to sing, set the pitches and the timing, and the software outputs a vocal track. Her concert appearances are in the form of a CGI anime girl, with a pre-programmed set of dance moves. She, and the rest of the Vocaloid franchise, are basically Animusic with a massive upgrade in processing power. (Also, thigh-highs.)
Kids who are currently in middle school do not remember a time when “humans pack stadiums to see a holographic robot singer” was a fanciful sci-fi premise. It’s literally just the world they live in.
And you know what we still have? Concerts with live human singers!
Computer-generated music hasn’t replaced human musicians. They have different strengths and abilities. They’re good in different ways. The one doesn’t make the other obsolete. Millions of songs produced with the Vocaloid software, over more than a decade by now, and they still haven’t put all the human singers out of their jobs.
Same deal with AI art.
Some of it is terrible. Some of it is pretty cool. It’s not as precisely controlled as Vocaloids, a lot of it is “turning the algorithm loose and having no idea what’s going to come out,” which is sometimes a drawback and sometimes the fun part.
Look at this robot’s best attempt to paint “Moon Knight drawn by Lisa Frank” and tell me that isn’t fun:
If you want a very specific image, AI art isn’t a great option. If you want to consistently reproduce the same character or setting across multiple images, it’s not great either.
Ursula Vernon — of Digger fame — has been making experimental comics with the MidJourney AI. it works! The reason it works is, she’s not a random non-artistic person who started with “here’s a comic I want to make” and tried to get the robot to produce it. She’s a brilliant comic artist with multiple series and at least one Hugo award under her belt, who put a lot of thought upfront into presenting the robot with “here’s a comic setup that takes advantage of your strengths, and doesn’t stress your weaknesses.”
In that link, for instance, you can see right away how it’s mostly disconnected vignette panels. (It also has human post-production touch-ups, and the art is tied together by human-written narration. More notes on her process at the end of the thread.)
So yeah, it’s gonna be fine. AI art makes cool new things possible, will lead to the creation of a ton of art that never would’ve happened without it, and some of it will even be good and worthwhile. It also won’t replace human artists. It won’t be the best option for every art-related job. Sometimes it’ll be a good-enough option — other times it just flat-out won’t cut it.
Anyone who thinks “just have an AI draw it” will be a magic answer to everything art-related needs to appreciate artists more. (Some of y’all are illustrators, who need to have more confidence in your own skills and potential! Others are just hacks. Probably the same hacks who thought “just put it on a blockchain” was a magic answer to everything finance-related, even.)
Seems like the whole weather flipped a switch the other night. On Wednesday I had the windows open, computer sitting on ice packs to help it not overheat, and wasn't layering up at all to go to work...Thursday, wham, I'm closing the windows, pulling out extra blankets, wearing a sweater and a coat to go out.
Nothing against sweaters, but oof, me and my two smelly cats are gonna miss having a fresh breeze going through the house.
*
Parents officially got COVID over the summer! They're vaxxed+boosted, no extra high-risk conditions, and didn't have any trouble staying home the full isolation period, so it went over about as well as you could hope, probably.
They had this month-long European tour booked back in 2020...postponed it to 2021...postponed it again to 2022...figured this year they would just take the plunge and go for it. Get to week 3, and half the people on the tour bus are coughing. Whoops.
More recently, my aunt got it after an out-of-town relative visited. Glad I didn't go to that family gathering. (...I've gone to others, so this isn't a principled health stance I've been taking, I just got lucky.)
In other health news, one of my uncles has cancer -- not the kind you beat, the kind where the doctors say "with treatment you have about 10 years to live" -- and it's been about 10 years since they said that. So. As of this writing, he's not gone, but there's a good chance he's on his way out.
The news arrived a few days after Disney+ got Thor: Love and Thunder -- you know, the one where Jane spends the whole thing actively dying of cancer! -- so, hey, guess which family movie-watching plan has been tabled indefinitely? (I saw it in theaters, at some point I became the kind of fan who sees every Marvel movie in theaters, but the parents are in the "we want to catch them all...eventually" camp.)
*
Just as I get comfortable with "Kickstarter's not going anywhere with their Mystery Blockchain Protocol, at least not in the foreseeable future, it's still safe to run a campaign there for now," then Deviantart turns around and starts flirting with blockchain nonsense. Deviantart! You were the chosen one the only major site making an active effort to protect artists against blockchain fraud! What happened??
...okay, okay, we know what happened. Current holders of crypto can't actually get money out of the system unless they can convince new suckers to put money in, which means "bribing sites like DA and KS to drag in new suckers" is a good long-term investment.
Spent the past week or so in a real downswing of low energy and high executive-dysfunction. I have all these small, relatively simple tasks to do, but getting over the hump to actually start each one? Gonna need to go take a nap first.
(The Fluff thinks this is a great deal. He loves an extra snooze. And Fiddlesticks doesn't hang out on the bed, which makes it a nice safe territory where Fluff doesn't have to worry about defending his honor as Top Cat.)
I've been trying to put "have a nap" higher on the coping-strategy list than "chug another energy drink." Probably healthier! Not as good a deal for the to-do list.
...also, not a great state to be in when you're gearing up for a crowdfunding campaign. But it's not like I work better without a deadline. So we're moving right along.
*
True story, the one thing I can reliably-and-consistently focus on right now is "writing more Moon Knight fanfiction."
Using incentives like "you can write another chapter of Fic X once you finish Task Y" has been...moderately effective.
There’s been an absolute deluge of Blockchain Space Nonsense news in the past couple of weeks. If you, like me, can’t get enough of it, Web 3 Is Going Just Great is a great source to quench your thirst.
But if not — indulge me for a minute while I sift out some highlights, at least?
Everyone and their dog has been sharing this video, but I’ll share it again. It’s good. Not just about NFTs, it covers all kinds of Hot Topics in crypto discourse right now.
I started watching it thinking “I’ve rubbernecked SO MANY terrible details about these already, more than enough to fill a multi-hour video, there’s no way it’ll also have new-to-me info that makes them worse.” Spoiler alert: it had new-to-me info that makes them worse.
So a group called “SpiceDAO” pooled a bunch of money in order to buy a rare copy of Jodorosky’s Dune — basically, a long pitch for this guy’s proposed adaptation of Dune. They paid ten times the estimated value at auction, apparently totally convinced that “buying a book” and “buying the adaptation rights” were the same thing.
(A DAO is like a co-op, but to join or vote on anything, you need to buy into the org’s crypto token. These folks also seem to believe “we’re voting on a blockchain!” bypasses any requirements for laws, rules, obligations, paperwork, or, like…basic planning.)
The first half of this Twitch stream has a great time exploring the legal faceplants, but if that doesn’t sound delightful by itself, skip to about 50 minutes in. See, when the DAO was thwarted in their plans to adapt Dune, they commissioned a derivative-but-legally-not-Dune script to film instead. The stream does a Dramatic Reading. Of the whole thing.
I don’t remember the last time I laughed this hard.
Twitter announced a new “connect your account to an NFT and we’ll make a Special Exclusive hexagon-shaped profile picture out of it” feature. People immediately started dunking on it by uploading pfp images that they cropped into hexagon shape on their own, for free. Here’s made a transparent template to help you nail the exact right type of hexagon, indistinguishable from the Special Exclusive ones.
(…at least, unless you zoom way, way in. Then you might realize it displays as 2 pixels shorter. Shhh.)
Not in a public news post or anything, that’s just what Support is telling people who email with questions. (This isn’t the fault of individual Support staffers — they haven’t been given any info either. Kiiiinda seems like the Board is happy to use their staff as human shields, here.)
But, good news:
TopatoCo — which I have been pronouncing wrong all this time, it rhymes with “potato” — launched a beta-testing project for their own crowdfunding system. They’ve been a reliable player in the “fulfillment of webcomic merchandise” field for years; they have the credibility to start a crowdfunding platform from scratch and get the comics community on board.
Someone showed the Kickstarter board a fancy PowerPoint presentation with lots of big numbers, they ran to invest a bunch of their own funds in a blockchain without stopping to ask their own devs if the tech had any value for what the site actually does, and now they’re desperately trying to justify it after-the-fact.
(As of this writing, the NFT-selling site is entirely ignoring the copyright claim…but I do appreciate DA for alerting me that it was happening at all. This is what a site that actually cares about its creative users looks like!)
A breakdown of what NFTs are — in straightforward terms, not in wild/ridiculous metaphors. (Which, to be clear, aren’t wrong — it’s just that I know many people don’t find them helpful.)
“You couldn’t store the actual digital artwork in a blockchain; because of technical limits, records in most blockchains are too small to hold an entire image. Many people suggested that rather than trying to shoehorn the whole artwork into the blockchain, one could just include the web address of an image […] Seven years later, all of today’s popular NFT platforms still use the same shortcut. This means that when someone buys an NFT, they’re not buying the actual digital artwork; they’re buying a link to it. And worse, they’re buying a link that, in many cases, lives on the website of a new start-up that’s likely to fail within a few years. “
2020: “The developers of non-fungible token project NiftyMoji pulled an exit scam as they have closed the official website, all social media and dumped their tokens on the market. Also the associated Coinbreeder accounts have vanished. The developers ran off with an estimated amount of one million dollars.“
Alternately, the link could get replaced with something else. Say, a bunch of random photos of rugs: “I just pulled the rug at my NFT collection on @opensea. Nobody got hurt. It is pretty easy to change the jpg, even if it does not belong to me or it is on auction. I am the artist, my decision, right?”
In other words: there isn’t one single, central version of Bitcoin. It has multiple versions, and they’re mutually incompatible with each other. And yet, some people still believe blockchain is the magic bullet that will make every website interoperable. Suuuure.
Also, if you’re hearing anyone talk about how miraculous and unhackable anything blockchain is:
![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}
