The furry fandom is a subculture interested in anthropomorphic animal characters with human personalities and characteristics. Examples of anthropomorphic attributes include exhibiting human intelligence and facial expressions, speaking, walking on two legs, and wearing clothes. The term "furry fandom" is also used to refer to the community of people who gather on the internet and at furry conventions. [1]
One key differentiation for the furry fandom is that Furries usually have their own made up characters instead of other fandoms where roleplay/cosplay/avatars are typically of characters from media.
Should be careful with calling them "made up characters". I know a significant percentage of furries where their fursona is considered a real second personality.
> As the images are generated by an AI, they are non-copyrightable and are therefore public domain.
I find this claim on the "about page" quite interesting. Some of those images might be so close to the training data that the copyright protection for fictional characters becomes relevant, even if the image is not identical. This is visible in this topic as people recognize characters from popular-culture (video-games or movies), because the training data seems to also contain fanart.
This is the whole conundrum of creation and copyright. Every creative work is protected by copyright yet every creative work is the sum of unconscious derivations to varying degrees of something an author has perceived, creative works existing ex nihilo are at best vanishingly rare; personally I'm not even sure they exist, I'm leaning more towards we're just mistaking unusually big jumps of derivations/combinations of those for ex nihilo creative works. We readily recognise "influences" of great artists (whether it is music, literature, painting...).
Doesn't mean creative work should not be protected, but drawing the line of infringing vs not is by definition extremely blurry and subjective.
AIUI, copyright holds for a specific fixed expression. So if an artist drew a face that looks incredibly like yours, the artist would hold the copyright over that drawing, but have no claim over your face.
I'd like to propose a "does not exist" site that does not exist: This Family Does Not Exist. Generate faces and a family tree, where I can see familial resemblance and where that resemblance follows our understanding of genetics.
Related business: let me upload photos of my relatives and a family tree, and show me generated faces for other people on the family tree (e.g., common ancestors) based on these inputs and genetics. I wonder what kind of accuracy can be achieved in generating a person's face, based on how many descendants' faces we have photos of, over how many generations, with how much inbreeding, etc.
One of those interesting side-effects of furry avatars that I noticed is that accounts bearing those avatars were always real people, with the added bonus that you can generally authenticate the human behind the mask if you know how. The reality of being online today is that we have to understand whether we're interacting with real people or just a clever piece of software, and this is far more true for folks who are not technically savvy.
Oddly enough, furries were the last bastion of humanity. (And a welcoming one at that, but I digress.)
This complicates that heuristic somewhat. This brings furry avatars on the same level as human headshots. I now need to, e.g., read and process the full account bio and spend more time authenticating whom I interact with online.
Ok, thank you (and no offense meant to the OP) that makes sense in an odd way, and I’m clearly just not up on these things. I wasn’t aware that the corresponding fake-person generator was a thing, but of course it is. There’s something odd about the fact that a cartoon rendering would be more trustworthy than a photo. Then again, there’s always been the joke about people who keep the sample photo that comes in the picture frame to pretend they have a s.o. or family ...
• logotypes — actual logos, icons, arbitrary aesthetically-pleasing typographical art
• headshots of real people
• photographs of real-life things — places, nature, buildings, etc.
• crops from TV/film — headshots of actors playing characters, or whatever you'd call Baby Yoda
• famous works of art — crops of prints of paintings, crops of photos of sculptures
• commercially-marketable art — box art, movie posters, crops from cartoons/anime, professionally-commissioned CG paintings that fit the style of their source material
• unknown, non-marketable art — works that are clearly either self-made, or commissioned as a one-off, where the work has traits that make it specific-enough to someone's tastes that it obviously would never have been produced as spec work without an arranged buyer/planned use
The parent's point is that, for all the categories except the last, there's an obvious way to scrape or generate a million such images, that someone can include in their spambot/voting-ring registering algorithm.
The last category, though—custom competent-but-not-commercial-looking illustrations—were, in some sense, a Proof of Work token for the profile it was attached to: someone had to draw that (and even more, gather requirements to draw that, rather than it just being one keyframe following the same rules of thousands of others.) It cost a few dollars for that person to get that image; and therefore, it's less likely (though not impossible) that ten users with ten different such illustrations in a forum thread were all secretly the same person/bot.
There hasn't even been an AI that can do face-detection on funny-animal cartoons until now, AFAIK, so there was until now no way to even automate+scale scraping of "authentic" profile-pictures from some art-hosting website, let alone a way to automate+scale generating them. But now the cat's out of the bag. Bit of a shame.
"Automate and scale" means something other than what you think it means. The scraping of relevant raw images is the easy part. But you can't use scraped images as profile pictures directly, without first cropping them to be headshot-ish, because slapping random raw images with no sense of photographic composition into your profile picture is also a common signal of there not being a human in the loop.
And human labor—even the cheapest human labor—still costs way too much to have real people sitting there cropping pictures to use as profile pictures, if you need a basically-infinite stream of them for your spambot network.
So you'd need, at minimum, a face-detection algorithm that you can rely on to do auto-cropping, so that you can just throw it a whole scrape-dump of the art site, and get back that basically-infinite stream of headshots.
As I said, until this very project, there was no face-detection algorithm that worked on illustrations of, er, "demihuman" faces. (There was one for humans, that would fail horribly on this data; and then specific other ones for cats/dogs/etc in photos, that also would fail here.)
All is fair in war, love, and the name of science (with exceptions).
After spending a lot of time online as well, I've found just as many folks using content they stole that weren't real people, but were generative content.
Stolen content in the furry subculture is actually easily enough dealt with. Given a little bit of work, you can sniff out who drew the original work and piece things together. (This goes notwithstanding that artists actually enjoy knowing when their work is misused so that they can take the appropriate steps to report it, so there's an incentive at play to track this kind of thing down.)
Also, because the corpus of source material is comparatively pretty small, you can use image hashing and a distance metric to match something you see to its source. It takes more effort to draw something than it does to snap a photo, and the entirety of E621 can fit on a single SSD.
This is quite unlike headshot photographs of people. Authenticating the photographer is almost impossible, and because of the low-effort nature of photos, it's infeasible to have a control data set to match with observed data.
You could "just" draw a furry avatar if you have the art skills.
Or you could pay an artist to do that for you.
Neither tactic really scales, especially when your game is coordinated inauthentic behavior i.e. https://www.youtube.com/watch?v=V-1RhQ1uuQ4 and you need thousands of convincing fake accounts.
Evidence suggests that the Russians were dropping $400k a month on trolls. That's a lot of commissions on Fiverr or FurAffinity, and still plenty left over to pay east european IT contractor rates. Or hell, just steal enough existing avatars and assume that 80% will get taken down fast.
Furry art could be copied en masse if anyone had ever wanted to use them for trolls. It's no more difficult than, say, anime avatars. If Russian troll farms haven't been using furry avatars, that probably has more to do with the stigma against furries impeding their propaganda mission than about the difficulty of downloading images from e621.
Doing this is generally not advisable; you'll be found out pretty quickly if you do.
Furries are a pretty tightly-knit bunch of people, and the folks who occupy the artwork-having set are usually two or three degrees of separation from each other.
As an added bonus, in case the social network effects weren't enough, DMCA claims become easier because the provenance of a piece of art is often quite clear, what with the subculture's emphasis on always attributing the source of the art in question.
Russian propaganda agents actually infiltrated furry Tumblr pretty well; it seems to have been one of their more successful operations in terms of reaching a wide audience. It just didn't get so much media coverage, probably because (as you'd expect for an operation targeting Tumblr and furries) the political messages they used to weren't ones that fit the existing narrative about Russian interference.
Interesting. Would you happen to have any good sources for that? All I'm coming up with are pop culture rags that are scant on the details and don't offer up any real analysis of what went on there.
I found this [0], which is pretty scant on details. It is also scant on details but amusingly it lists the usernames they used. A few seemed somewhat political or related to race issues, including:
1) voteforwest2020
2) weproudto2black
3) aaddictedtoblackk
4) best-usa-today
5) black-to-the-bones
6) blackness-by-your-side
7) blacknproud
8) blacktolive
9) bleepthepolice
10) guns4l1fe
11) sumchckn previously known as: blondeinpolitics, blvckcommunity, classylgbthomie, hwuudoin, politixblondie
Yeah, I stumbled upon that as well, but it just looked like the connections to furry merely involved political issues that coincided with the causes that furries tend to care about. Even furries, still being human at the end of the day, contribute to the mainstream political discourse.
My definition of "infiltrating furry Tumblr" is perhaps a fair bit stricter: Infiltration implies being accepted as a member of the community, which also implies some level of gossip around someone new or cool whom folks are excited to introduce or meet.
The problem is that Tumblr killed these accounts, so there isn't any good record today of what their profiles looked like, what avatars they had, etc. I also can't seem to find anything particularly furry-centric beyond the fact that a malicious actor saw fit to target a subculture for some political trolling.
I'm chalking this up to the interface of a subculture with mainstream society being compromised and not necessarily the subculture itself.
Er, that's irrelevant. Obviously people with avatars of dragons and wolves are not actually dragons and wolves, nor are they generally claiming to be the original artist, any more than the millions of people with anime avatars are usually claiming to actually be Japanese schoolgirls or the mangaka who drew the avatar. There's nothing to 'find out' in either case.
I don't feel it's irrelevant. The fact is that the characters depicted in the artwork you're proposing could be ripped off belong to individuals. People like you and me. People with whom I'm personally good friends, with whom I have shared beers with at conferences, and whom I have visited en route on my trips to far-flung places.
These are characters that are recognizable as being digital representations of those individuals. They aren't the trademarks or personality rights of some faceless commercial entity.
My point is that trying to use artwork depicting a non-commercial character that isn't yours—that doesn't accurately reflect the representation of your digital self but instead someone else's—is as taboo as impersonation. It isn't tolerated, and that kind of behavior is ripe for gossip.
So, yes, while there's a wealth of source material that's ripe for cropping into a profile picture, there's the added hurdle of overcoming the bullshit detectors of furries who will test you to see if you're actually their friend or not.
You're changing arguments in mid-stream. What does any of that have to do with making Twitter more vulnerable to mass-influence operations? No one, Russian trolls least of all, is trying to pretend to be your specific friend (and if someone were doing such a spearphishing attack on you personally, they wouldn't be using new artwork of any kind whatsoever, human or AI-generated, in the first place, as that would defeat the point of trying to pretend to be an existing acquaintance!).
No, I'm making the same argument. I'll restate my proposition: Furry, as a subculture, is mostly secure against the kinds of deception employed by, e.g., Russian trolls because of social network effects.
There are a few facts I feel are worth emphasizing here. Sorry, it's a lot of words, but I want to make sure the required context is here for drawing the logical connectives that are missing.
Notably, furries tend to be acquainted with each other, in the sense that if two people within the community don't personally know each other, the chance is very good that they at least know of each other or have mutual friends who can make introductions. For individuals who exhibit problematic behaviors, interventions happen, and they happen reasonably swiftly.
This goes hand-in-hand with the fact that there are artifacts that act as shibboleths within the community and help to signal "hey, you're one of us." The subculture has a very well-defined market, literally and proverbially, for these kinds of artifacts, and the market is amazingly insular with known players who are also themselves part of the community.
One common example is the artwork I've been talking about, given its relevance here, but I've left a thing or two unstated in my conversation about it so far. There's a reverence of sorts for the artists who do this work, and I'd feel safe in asserting without evidence that artists don't tend to become popular or get work at all unless they've had a good inculcation of what furry as a subculture is all about, up to and including the social connections. Ultimately—and forgive the hand waving—this means that the artwork is readily identifiable as "furry" (as opposed to "just a cartoon") and that there's an understanding as to where it came from. It's able to be authenticated as having originated from within the furry subculture. The implication that there's widespread interest to know the identity of the individual who did a piece of artwork is hopefully clear.
("Ooh, are they open for commissions?" is a common refrain.)
That said, these artifacts extend far beyond just raster images on the Internet. There are T-shirts, stickers, enamel pins, and what seems like myriad other things that are innocuous (or just plain weird, let's admit) to folks outside the community but will start a conversation between folks within. As an anecdote: A homemade sticker on my ThinkPad given to me by an artist friend was seen by a passenger sitting next to me on a flight, and it wound up leading to us talking for the whole flight and getting margaritas while on layover to our separate final destinations. I still talk to that individual today, and knowing them has actually strengthened my connections to other folks.
Those artifacts wind up being a very good way of establishing common ground with someone else within the community very quickly, further cementing the "hey, you're one of us" effect.
… phew.
Let's recap. The furry subculture is particularly tight-knit and egalitarian with strong social connections and community cohesion that are reinforced by cultural artifacts and shared experiences that come with the ability to be verified as genuine.
Or to repeat something I said earlier: "[Copying furry artwork en masse] is generally not advisable; you'll be found out pretty quickly if you do," given a definition of copying probably being closer to co-opting.
Even if that's the argument being made, nothing changes. Co-opting furry artwork without actually being part of the community will be uncovered soon enough as being fake.
I won't assert that the furry subculture is always better than any intelligence agency at this game, but it has a lot of things going for it that certainly make this kind of deceptive, divisive activity harder than it would ordinarily be.
That's true. Especially all the people who've had to pass serious background checks. I didn't expect to meet rocket scientists and defense contractors when I became a furry. Someone is going to figure it out, and furry breached containment a while ago. I see furry tweets going viral all the time now.
> (for people not familiar, furries are a vibrant online community and one of the last outposts of collaborative, creative Internet culture. They're also heavily LGBTQ and have a deep commitment to inclusiveness and social justice. Pretty great people to have on your side!)
It's easier to tell whether or not an account is a bot if it uses a furry avatar.
There are a couple of important facts to consider: The furry subculture is very tightly knit, and there's a reasonable chance that you're only two or three degrees of separation from someone else. There's also an emphasis on attributing furry artwork back to its source, not to mention tools for doing just this.
Those two things make authenticating the identity behind an account with a furry avatar comparatively very easy when contrasted to a human headshot photograph.
I'm convinced that for 90% of these "Does not exist" generators, you can identify two or three hugely influential images, and almost all the rest is noise. There's an image where the hair comes from, an image for the facial structure, and then some perturbations are made to eye color or jawline.
Not to say that isn't impressive! Generating convincing fakes, even 1% of the time, even if they're not super unique (There's plenty of people who "look exactly like" in the real world, too!) is a big deal.
You would not. Many GAN papers do nearest-neighbor lookups, and typically, generated samples are clearly different. Since Arfa's model is high-quality and the interps look fine, I would not expect it to be any different. (It would be hard to check because there are no pretrained classifiers to provide an embedding to do the search in, but one could do reverse-encoding.)
The fact that nearest-neighbor lookups do not find exact overlaps between the training dataset and a large number of random samples has always been one of the arguments I use against the widespread misconception that GANs 'just memorize' data: https://www.gwern.net/Faces#faq
That so many people are convinced that a given datapoint must be an exact copy of a training datapoint - "it looks exactly like a Zootopia character I recognize!" - is really quite a compliment to the GAN...
This has far fewer oddities than the people based ones. Maybe that's because furry art all have similar qualities already, so it's easier? A few of them were just straight up Judy from Zootopia, so that says something gross about the furry community, I'm sure.
Also, I hate this. I especially hate the loading messages.
In my experience, any activity or community enjoyed by humans has a subgroup for “that thing, but also with sex”. It’s more of a reflection of what it means to be human than anything to do with the means of that expression.
It doesn't need to be gross to dislike the weird baby talk "uwu" stuff. Although seeing it in much more technical contexts will make me laugh - https://github.com/mpaland/printf/issues/15 - "Scawy big no functionality at all UwU #15"
Wow, that person seems to do this in all of the issues they file. They're also really good at filing issues despite the baby talk. All of their issues are well-researched and immediately actionable with a surprising amount of debugging completed before the initial submission.
Devs tend to have a penchant for the absurd. The details may change along with the zeitgeist, but the personality quirk itself seems to be ubiquitous across the decades.
not really sure what you mean, can you elaborate on why you would say that?
FWIW, I've met more normies in my 10years of software development than absurd people. Absurd people are just memorable, it's a bias.
If you're actually curious, this username was stolen from an absurd Atlanta based rapper named blunt fang and I get more weed culture association than furry association with it.
The dataset grabbed has both fursonas and fan art in it.
There are not many furries with fursonas that look like Judy, though I'm not sure why that would be gross.
Fursonas tend to mirror the style the content one was watching when they were a kid, so once kids who grow up watching Zootopia become old enough to enter the fandom, then those individuals might bring in a 3D style, but it has yet to happen.
I was about to say I recognised Toriel [0] and was suspicious it might have been memorising a bit too much.
As for the idea that furry are has too many similarities… I don’t think so. The variance in fursona body morphology is huge. For example, one thing I didn’t notice in this set was Sergals [1].
Arfa excluded a few categories like ponies and scalies to keep it from being too broad for StyleGAN. Sergals would probably get excluded as part of that.
I saw a lot of verbatim copies of both the fox and the rabbit from Zootopia (possibly with shades of Disney's Robin Hood) and distinctive elements of Sonic the Hedgehog and Pokemon characters. I think it was trained on popular anthropomorphic media, not just furry art, and there's some overfitting. I don't think stylistic inspiration can quite explain its fondness for cheeky foxes wearing green.
It was trained on images from e621[1]. Considering the most popular characters[2] on that site, you can see where the neural net got it from.
[1]: The most popular furry booru. A booru is an image board where everyone can edit the tags of images. Strict tagging rules and advanced search operators make it easy to find specific images. A safe-rated only mirror is available at e926.net.
Oh no, I don't mean to diminish the achievement on display here. It's very impressive how many of these could pass for original, hand-drawn artwork, and in such a wide variety of styles too.
What’s most interesting to me vis-a-vis other GAN portrait generators is the range of illustration styles employed here. It’s like a GAN that generates both manga and western comic-art faces without any incongruous hybrids.
Has anyone ever made some variant of these "* does not exist" sites with some control sliders/options, that controls various aspects of what's generated?
So things like: gender, hair color, face shape, etc?
Basically, an AI-driven character generator that isn't completely random?
Absolutely. Waifu Labs https://waifulabs.com/ implements a grid-based choice system for evolving anime portraits, and Artbreeder https://artbreeder.com/ implements controls plus crossbreeding and other features for a variety of StyleGAN/BigGAN models. There's plenty of scripts and Colab notebooks as well for various kinds of editing or control if Artbreeder doesn't do it for you. (I think Runway may also do editing but I haven't used them in ages.)
GAN models do not need to be specifically architected to enable control, because you can reverse them to get the latents/seed and manipulate that to 'edit' images: https://www.gwern.net/Faces#reversing-stylegan-to-control-mo... So if someone wanted, they could use Arfa's model to edit images.
Truly impressive results, though I find it unsettling that there's a decent chance that the future of all Internet aggregators is going to be "We threw data into computer and it threw out magic thing!"
As far as the Elsagate stuff goes, it's not that the videos themselves were algorithmically generated, but that their creators were driven by YouTube's algorithms to create the content that would generate the most views.
The funny thing about the AI meme generator is per the technical write up (https://towardsdatascience.com/meme-text-generation-with-a-d...), that network is even less advanced than the old recurrent neural network approaches. (it uses convolutional neural networks which are more used for images than text nowadays)
However, this is the one case where more incoherence works better.
Yes, it's a testament to crowdsourcing and filtering. Many people have argued that 'generation + filtering = creativity', and it's impressive what NNs+crowds can do. The NNs are unhindered by any merely human considerations and are indefatigable, while the crowd brings the filtering & evaluation to select just the best candidates.
The heterochromia is probably not 'incidental'. Heterochromia, quite aside from lighting/shading differences in eye color, is present in many real images: Danbooru has a ton, and e621 has 11k images tagged 'heterochromia'. Just a common trope in illustrations.
The thumbnails have human-level quality. The full images have oddities, but they disappear when downsized.
DL is going to put animators, voice actors and even actors out of work sooner than I ever expected. It's going to be feasible for one person to create a full anime episode in three days. The amount of content is going to explode by orders of magnitude and shared culture is basically going to die, with nearly zero chance another random person consumes the same content. What a weird time to live in.
That's like saying snowclones and memes are going to kill book authors out of work... if anything, this is taking us to the next level of original work.
> The amount of content is going to explode by orders of magnitude and shared culture is basically going to die, with nearly zero chance another random person consumes the same content.
Why? I believe there's already so much content out there that this could easily happen today if the only question would be the raw amount. However if you ask people around the water cooler, instead of one guy watching joesmith34's Source Film Maker video on YouTube and another watching janedoe157's flash animation yesterday evening, everyone watched the latest episode of a popular Netflix series.
> As the images are generated by an AI, they are non-copyrightable and are therefore public domain. Feel free to use them any way you see fit. Just don't try to pass them off as your own art or sell them or anything.
The furry fandom is extremely touchy about "art theft." Considering how many of these images clearly resemble individual artists' styles, I'd tread a little more carefully.
> As the images are generated by an AI, they are non-copyrightable and are therefore public domain.
This part seems false. If I take copyrighted photos and run it through something like an "AI" filter to change the colors slightly, the photos aren't suddenly public domain.
If you steal from enough people at once is it then legal?
> If I take copyrighted photos and run it through something like an "AI" filter to change the colors slightly, the photos aren't suddenly public domain.
> If you steal from enough people at once is it then legal?
How many people does a great human artist steal from to become an artist? If you steal solely from one other artist, you're a plagiarist; if you steal from a thousand other artists, you're an innovator and pioneer of a new style...
This isn't transformative enough to pass a sniff test. They took multiple images and put them in a computer to make images that look like those images. What's transformative about that?
That's extremely transformative: it's learning from a large corpus what faces are and how to create brand new ones which cannot even be traced back to an original. (If you think even that is not enough to count as 'transformative', I suggest you look at the examples in the WP article of what has been considered transformative, like 'thumbnail screenshots of web pages' and 'putting a mustache on the Mona Lisa'.)
When a significant number look like zootopia characters I don't think you can claim they "can't be traced back to the original", lots of these are overfit
They are not 'overfit' unless they are exactly the same and they cannot generalize. But you can see the wide variety of high quality faces which are not copied from the dataset and that the interpolations are fine: https://twitter.com/arfafax/status/1258344026706599937
Generating a recognizable character is no more 'overfit' than a painter being able to paint George Washington means their brain is 'overfit'. If those characters are part of the data distribution, as they are, then the GAN can and should generate them and countless variants on them.
This demonstrates, at the very least, that it isn't simply memorizing the datapoints. You can see that it is able to smoothly transition between images of Zootopia characters and other characters, which indicates that it has learned a lot more about the actual features.
I believe the prevalence of certain characters (Zootopia, Sonic characters, Pokemon) showing up is because a large portion of the input space maps to those regions of the latent space. So I'd expect there to be a roughly equal proportion of images that look like Nick Wilde in the random samples as there are in the training data.
No, because a fursona is something that's deeply personal to the person, and typically the process for creating one looks like this:
1. Decide what you want to look like
2. Create a reference sheet (or commission an artist to create one based on a description from step 1)
3. Commission a headshot for an avatar
4. Commission some art of your fursona, possibly with your friends' fursonas
5. GOTO 4
6. At some point, your friends may ask you for your reference sheet so they can gift you an art piece of your fursona and theirs (see optional stage of 4)
Even if you can churn out AI-generated headshots, getting the colors/markings just right for your character is nontrivial for non-generic fursonas (see https://soatok.com/static/soatok-johis-responsive.jpg for example).
And besides, most art commissions are conducted after you have a reference sheet and/or headshot of your character.
If anything, this will give artists something they can point the "steal other people's character art for their roleplay accounts" types of (especially younger) furries towards. "Can't afford to commission an artist? Just use the AI thing and stop the misbehavior!"
> I assume that this would only work reliably for headshots/avatars?
StyleGAN works best on single centered objects which are not too diverse in shape. So heads work great if you align them to the middle, and you can get OK results with single centered figures like someone standing, but anything beyond that and it falls apart. We have a theory about that (https://github.com/tensorfork/tensorfork/issues/21) but we've long since moved onto BigGAN, which can model much more diverse datasets like ImageNet successfully.
As far as fursonas go, see my other comment: https://news.ycombinator.com/item?id=23096442 I'd say that GANs are better at letting someone explore fursonas than any human artist is or ever will be. You cannot click a button and have a human artist generate 32 variants in a second for you, toggle on or off attributes, nor can they slide a slider from 'human' to 'neko' to 'dangerously cheesy!'. You might pay for a commission of your final fursona for the highest-quality image with no artifacts, but for exploring...? And as NNs get better over time, they'll creep up the value chain, as it were. (I'm still a little shocked how good the human voices in OpenAI's Jukebox https://openai.com/blog/jukebox/ are, even if the highest level of abstraction is still very shaky and they need to tack on another layer or two to get choruses etc.)
I can't wait for the first "your AI stole my original character!!!" furry thread on Twitter.
Impressive achievement, I'd love to see the outcome of re-training it with fan-art removed from the training set. There's a lot of Nick Wilde and Judy Hopps in there...
> They all seem to trend towards feminine cat/fox 'sonas.
The gender bias is actually really interesting. I'm curious how much of it is due to:
* Inherent bias in the source dataset. (For example, nekomimi art is overwhelmingly female -- there's a reason they're referred to as "catgirls". The GAN does a great job of distinguishing this from Western anthropomorphic art styles, incidentally.)
* Bias introduced through filtering of the dataset, e.g. by excluding NSFW source material or certain tags, or a bias in what types of faces were recognized.
* Androgynous faces being interpreted as feminine by default.
* GAN-specific characteristics of the output (like smooth features) being interpreted as feminine.
I think it's bais towards catgirls and other anthropomorphized anime characters which is skewing the results. If you go to FA, you would have to look hard to find feminine like this. It's either using a catgirls, as you say, heavy dataset or mixing two unrelated datasets.
I don't like the word feminine, because it's not really what furry feminine is. It's a human-applied-to-furry concept.
ACTUALLY. As I refresh the page, I get the feeling that it's the result of weird mixing between flat shaded art and mixing anthromorphic and "cargirls" (for lack of a better term- and I don't want to say east vs western, because catgirls dont' fall under furry)- anime inspired furry. It's weird, and makes me ponder. If the underlying datasets are a more similar art style and didn't have anime inspired art mixed in, I think the result would be less confusing and cooler.
The furry fandom is a subculture interested in anthropomorphic animal characters with human personalities and characteristics. Examples of anthropomorphic attributes include exhibiting human intelligence and facial expressions, speaking, walking on two legs, and wearing clothes. The term "furry fandom" is also used to refer to the community of people who gather on the internet and at furry conventions. [1]
[1] https://en.m.wikipedia.org/wiki/Furry_fandom
One key differentiation for the furry fandom is that Furries usually have their own made up characters instead of other fandoms where roleplay/cosplay/avatars are typically of characters from media.
Should be careful with calling them "made up characters". I know a significant percentage of furries where their fursona is considered a real second personality.
That doesn't make them not made up. Even a primary personality is made up.
Fredrik Knudsen has done an amazing video on Furries: https://youtu.be/8aF2GxWi7Ag
> As the images are generated by an AI, they are non-copyrightable and are therefore public domain.
I find this claim on the "about page" quite interesting. Some of those images might be so close to the training data that the copyright protection for fictional characters becomes relevant, even if the image is not identical. This is visible in this topic as people recognize characters from popular-culture (video-games or movies), because the training data seems to also contain fanart.
I suspect in legal terms, if you feed copyrighted images into a computer program, the output is a derivative work, AI or no.
As you say, some of the output images are clearly of specific characters, which turns this from "legally grey" into "definitely not public domain".
Depends how creative the AI is. Human artists are trained on copyrighted images too.
Good artists copy, great artists steal?
This is the whole conundrum of creation and copyright. Every creative work is protected by copyright yet every creative work is the sum of unconscious derivations to varying degrees of something an author has perceived, creative works existing ex nihilo are at best vanishingly rare; personally I'm not even sure they exist, I'm leaning more towards we're just mistaking unusually big jumps of derivations/combinations of those for ex nihilo creative works. We readily recognise "influences" of great artists (whether it is music, literature, painting...).
Doesn't mean creative work should not be protected, but drawing the line of infringing vs not is by definition extremely blurry and subjective.
That talk is as relevant as ever:
https://www.ted.com/talks/lawrence_lessig_laws_that_choke_cr...
Man what happens when thispersondoesnotexist.com generates something that is exceptionally close to my face ... ?
Am I now in the public domain?
Not that I'm looking to sue for such things but this "AI did it so it's public" thing ... easily could cross over into real life.
AIUI, copyright holds for a specific fixed expression. So if an artist drew a face that looks incredibly like yours, the artist would hold the copyright over that drawing, but have no claim over your face.
I'd like to propose a "does not exist" site that does not exist: This Family Does Not Exist. Generate faces and a family tree, where I can see familial resemblance and where that resemblance follows our understanding of genetics.
Related business: let me upload photos of my relatives and a family tree, and show me generated faces for other people on the family tree (e.g., common ancestors) based on these inputs and genetics. I wonder what kind of accuracy can be achieved in generating a person's face, based on how many descendants' faces we have photos of, over how many generations, with how much inbreeding, etc.
Thanks, I hate it.
One of those interesting side-effects of furry avatars that I noticed is that accounts bearing those avatars were always real people, with the added bonus that you can generally authenticate the human behind the mask if you know how. The reality of being online today is that we have to understand whether we're interacting with real people or just a clever piece of software, and this is far more true for folks who are not technically savvy.
Oddly enough, furries were the last bastion of humanity. (And a welcoming one at that, but I digress.)
This complicates that heuristic somewhat. This brings furry avatars on the same level as human headshots. I now need to, e.g., read and process the full account bio and spend more time authenticating whom I interact with online.
This is great work, but did you really have to?
What?
> What?
What nrr is saying is simply:
If someone replies to a comment and they have a human avatar, there was always a chance it was the output of https://www.thispersondoesnotexist.com
Their heuristic for "if it's a furry avatar, it's probably not from a Russian troll farm" is now invalid.
Ok, thank you (and no offense meant to the OP) that makes sense in an odd way, and I’m clearly just not up on these things. I wasn’t aware that the corresponding fake-person generator was a thing, but of course it is. There’s something odd about the fact that a cartoon rendering would be more trustworthy than a photo. Then again, there’s always been the joke about people who keep the sample photo that comes in the picture frame to pretend they have a s.o. or family ...
If you look at the profile pictures on Twitter, Tumblr, WordPress, etc., they fall into a few categories:
• fallback — e.g. the egg image on twitter
• explicitly algorithmic — e.g. http://identicon.net/
• logotypes — actual logos, icons, arbitrary aesthetically-pleasing typographical art
• headshots of real people
• photographs of real-life things — places, nature, buildings, etc.
• crops from TV/film — headshots of actors playing characters, or whatever you'd call Baby Yoda
• famous works of art — crops of prints of paintings, crops of photos of sculptures
• commercially-marketable art — box art, movie posters, crops from cartoons/anime, professionally-commissioned CG paintings that fit the style of their source material
• unknown, non-marketable art — works that are clearly either self-made, or commissioned as a one-off, where the work has traits that make it specific-enough to someone's tastes that it obviously would never have been produced as spec work without an arranged buyer/planned use
The parent's point is that, for all the categories except the last, there's an obvious way to scrape or generate a million such images, that someone can include in their spambot/voting-ring registering algorithm.
The last category, though—custom competent-but-not-commercial-looking illustrations—were, in some sense, a Proof of Work token for the profile it was attached to: someone had to draw that (and even more, gather requirements to draw that, rather than it just being one keyframe following the same rules of thousands of others.) It cost a few dollars for that person to get that image; and therefore, it's less likely (though not impossible) that ten users with ten different such illustrations in a forum thread were all secretly the same person/bot.
There hasn't even been an AI that can do face-detection on funny-animal cartoons until now, AFAIK, so there was until now no way to even automate+scale scraping of "authentic" profile-pictures from some art-hosting website, let alone a way to automate+scale generating them. But now the cat's out of the bag. Bit of a shame.
> so there was until now no way to even automate+scale scraping of "authentic" profile-pictures from some art-hosting website
There's a few furry-oriented art websites, so this part has actually been really easy for like two decades.
"Automate and scale" means something other than what you think it means. The scraping of relevant raw images is the easy part. But you can't use scraped images as profile pictures directly, without first cropping them to be headshot-ish, because slapping random raw images with no sense of photographic composition into your profile picture is also a common signal of there not being a human in the loop.
And human labor—even the cheapest human labor—still costs way too much to have real people sitting there cropping pictures to use as profile pictures, if you need a basically-infinite stream of them for your spambot network.
So you'd need, at minimum, a face-detection algorithm that you can rely on to do auto-cropping, so that you can just throw it a whole scrape-dump of the art site, and get back that basically-infinite stream of headshots.
As I said, until this very project, there was no face-detection algorithm that worked on illustrations of, er, "demihuman" faces. (There was one for humans, that would fail horribly on this data; and then specific other ones for cats/dogs/etc in photos, that also would fail here.)
> did you really have to?
All is fair in war, love, and the name of science (with exceptions).
After spending a lot of time online as well, I've found just as many folks using content they stole that weren't real people, but were generative content.
Stolen content in the furry subculture is actually easily enough dealt with. Given a little bit of work, you can sniff out who drew the original work and piece things together. (This goes notwithstanding that artists actually enjoy knowing when their work is misused so that they can take the appropriate steps to report it, so there's an incentive at play to track this kind of thing down.)
Also, because the corpus of source material is comparatively pretty small, you can use image hashing and a distance metric to match something you see to its source. It takes more effort to draw something than it does to snap a photo, and the entirety of E621 can fit on a single SSD.
This is quite unlike headshot photographs of people. Authenticating the photographer is almost impossible, and because of the low-effort nature of photos, it's infeasible to have a control data set to match with observed data.
I mean you could just draw a furry avatar before, it's not that different.
You could "just" draw a furry avatar if you have the art skills.
Or you could pay an artist to do that for you.
Neither tactic really scales, especially when your game is coordinated inauthentic behavior i.e. https://www.youtube.com/watch?v=V-1RhQ1uuQ4 and you need thousands of convincing fake accounts.
Evidence suggests that the Russians were dropping $400k a month on trolls. That's a lot of commissions on Fiverr or FurAffinity, and still plenty left over to pay east european IT contractor rates. Or hell, just steal enough existing avatars and assume that 80% will get taken down fast.
Furry art could be copied en masse if anyone had ever wanted to use them for trolls. It's no more difficult than, say, anime avatars. If Russian troll farms haven't been using furry avatars, that probably has more to do with the stigma against furries impeding their propaganda mission than about the difficulty of downloading images from e621.
Doing this is generally not advisable; you'll be found out pretty quickly if you do.
Furries are a pretty tightly-knit bunch of people, and the folks who occupy the artwork-having set are usually two or three degrees of separation from each other.
As an added bonus, in case the social network effects weren't enough, DMCA claims become easier because the provenance of a piece of art is often quite clear, what with the subculture's emphasis on always attributing the source of the art in question.
Russian propaganda agents actually infiltrated furry Tumblr pretty well; it seems to have been one of their more successful operations in terms of reaching a wide audience. It just didn't get so much media coverage, probably because (as you'd expect for an operation targeting Tumblr and furries) the political messages they used to weren't ones that fit the existing narrative about Russian interference.
Interesting. Would you happen to have any good sources for that? All I'm coming up with are pop culture rags that are scant on the details and don't offer up any real analysis of what went on there.
I found this [0], which is pretty scant on details. It is also scant on details but amusingly it lists the usernames they used. A few seemed somewhat political or related to race issues, including:
1) voteforwest2020
2) weproudto2black
3) aaddictedtoblackk
4) best-usa-today
5) black-to-the-bones
6) blackness-by-your-side
7) blacknproud
8) blacktolive
9) bleepthepolice
10) guns4l1fe
11) sumchckn previously known as: blondeinpolitics, blvckcommunity, classylgbthomie, hwuudoin, politixblondie
12) superblygun
[0]: https://techcrunch.com/2018/03/23/tumblr-confirms-84-account...
Yeah, I stumbled upon that as well, but it just looked like the connections to furry merely involved political issues that coincided with the causes that furries tend to care about. Even furries, still being human at the end of the day, contribute to the mainstream political discourse.
My definition of "infiltrating furry Tumblr" is perhaps a fair bit stricter: Infiltration implies being accepted as a member of the community, which also implies some level of gossip around someone new or cool whom folks are excited to introduce or meet.
The problem is that Tumblr killed these accounts, so there isn't any good record today of what their profiles looked like, what avatars they had, etc. I also can't seem to find anything particularly furry-centric beyond the fact that a malicious actor saw fit to target a subculture for some political trolling.
I'm chalking this up to the interface of a subculture with mainstream society being compromised and not necessarily the subculture itself.
Er, that's irrelevant. Obviously people with avatars of dragons and wolves are not actually dragons and wolves, nor are they generally claiming to be the original artist, any more than the millions of people with anime avatars are usually claiming to actually be Japanese schoolgirls or the mangaka who drew the avatar. There's nothing to 'find out' in either case.
I don't feel it's irrelevant. The fact is that the characters depicted in the artwork you're proposing could be ripped off belong to individuals. People like you and me. People with whom I'm personally good friends, with whom I have shared beers with at conferences, and whom I have visited en route on my trips to far-flung places.
These are characters that are recognizable as being digital representations of those individuals. They aren't the trademarks or personality rights of some faceless commercial entity.
My point is that trying to use artwork depicting a non-commercial character that isn't yours—that doesn't accurately reflect the representation of your digital self but instead someone else's—is as taboo as impersonation. It isn't tolerated, and that kind of behavior is ripe for gossip.
So, yes, while there's a wealth of source material that's ripe for cropping into a profile picture, there's the added hurdle of overcoming the bullshit detectors of furries who will test you to see if you're actually their friend or not.
You're changing arguments in mid-stream. What does any of that have to do with making Twitter more vulnerable to mass-influence operations? No one, Russian trolls least of all, is trying to pretend to be your specific friend (and if someone were doing such a spearphishing attack on you personally, they wouldn't be using new artwork of any kind whatsoever, human or AI-generated, in the first place, as that would defeat the point of trying to pretend to be an existing acquaintance!).
No, I'm making the same argument. I'll restate my proposition: Furry, as a subculture, is mostly secure against the kinds of deception employed by, e.g., Russian trolls because of social network effects.
There are a few facts I feel are worth emphasizing here. Sorry, it's a lot of words, but I want to make sure the required context is here for drawing the logical connectives that are missing.
Notably, furries tend to be acquainted with each other, in the sense that if two people within the community don't personally know each other, the chance is very good that they at least know of each other or have mutual friends who can make introductions. For individuals who exhibit problematic behaviors, interventions happen, and they happen reasonably swiftly.
This goes hand-in-hand with the fact that there are artifacts that act as shibboleths within the community and help to signal "hey, you're one of us." The subculture has a very well-defined market, literally and proverbially, for these kinds of artifacts, and the market is amazingly insular with known players who are also themselves part of the community.
One common example is the artwork I've been talking about, given its relevance here, but I've left a thing or two unstated in my conversation about it so far. There's a reverence of sorts for the artists who do this work, and I'd feel safe in asserting without evidence that artists don't tend to become popular or get work at all unless they've had a good inculcation of what furry as a subculture is all about, up to and including the social connections. Ultimately—and forgive the hand waving—this means that the artwork is readily identifiable as "furry" (as opposed to "just a cartoon") and that there's an understanding as to where it came from. It's able to be authenticated as having originated from within the furry subculture. The implication that there's widespread interest to know the identity of the individual who did a piece of artwork is hopefully clear.
("Ooh, are they open for commissions?" is a common refrain.)
That said, these artifacts extend far beyond just raster images on the Internet. There are T-shirts, stickers, enamel pins, and what seems like myriad other things that are innocuous (or just plain weird, let's admit) to folks outside the community but will start a conversation between folks within. As an anecdote: A homemade sticker on my ThinkPad given to me by an artist friend was seen by a passenger sitting next to me on a flight, and it wound up leading to us talking for the whole flight and getting margaritas while on layover to our separate final destinations. I still talk to that individual today, and knowing them has actually strengthened my connections to other folks.
Those artifacts wind up being a very good way of establishing common ground with someone else within the community very quickly, further cementing the "hey, you're one of us" effect.
… phew.
Let's recap. The furry subculture is particularly tight-knit and egalitarian with strong social connections and community cohesion that are reinforced by cultural artifacts and shared experiences that come with the ability to be verified as genuine.
Or to repeat something I said earlier: "[Copying furry artwork en masse] is generally not advisable; you'll be found out pretty quickly if you do," given a definition of copying probably being closer to co-opting.
I think they mean a situation with these hypothetical state actors using furry avatars for propaganda in general, not just within the furry fandom.
Even if that's the argument being made, nothing changes. Co-opting furry artwork without actually being part of the community will be uncovered soon enough as being fake.
I won't assert that the furry subculture is always better than any intelligence agency at this game, but it has a lot of things going for it that certainly make this kind of deceptive, divisive activity harder than it would ordinarily be.
That's true. Especially all the people who've had to pass serious background checks. I didn't expect to meet rocket scientists and defense contractors when I became a furry. Someone is going to figure it out, and furry breached containment a while ago. I see furry tweets going viral all the time now.
>Oddly enough, furries were the last bastion of humanity.
What do you mean by this?
Quoth Pinboard: https://twitter.com/Pinboard/status/992819169593716737
> (for people not familiar, furries are a vibrant online community and one of the last outposts of collaborative, creative Internet culture. They're also heavily LGBTQ and have a deep commitment to inclusiveness and social justice. Pretty great people to have on your side!)
It's easier to tell whether or not an account is a bot if it uses a furry avatar.
There are a couple of important facts to consider: The furry subculture is very tightly knit, and there's a reasonable chance that you're only two or three degrees of separation from someone else. There's also an emphasis on attributing furry artwork back to its source, not to mention tools for doing just this.
Those two things make authenticating the identity behind an account with a furry avatar comparatively very easy when contrasted to a human headshot photograph.
Some of these have designs that are so... specific? holistic? Not sure what I'm shooting for.
But, I am wondering if you ran an image similarity search on them vs. the training set would you find matches or are they actually unique?
https://thisfursonadoesnotexist.com/v2/jpgs-2x/seed09985.jpg https://thisfursonadoesnotexist.com/v2/jpgs-2x/seed13901.jpg https://thisfursonadoesnotexist.com/v2/jpgs-2x/seed31957.jpg https://thisfursonadoesnotexist.com/v2/jpgs-2x/seed25075.jpg
I'm convinced that for 90% of these "Does not exist" generators, you can identify two or three hugely influential images, and almost all the rest is noise. There's an image where the hair comes from, an image for the facial structure, and then some perturbations are made to eye color or jawline.
Not to say that isn't impressive! Generating convincing fakes, even 1% of the time, even if they're not super unique (There's plenty of people who "look exactly like" in the real world, too!) is a big deal.
That is obviously not true if you watch any interpolation videos or do any nearest-neighbor lookups.
I know a guy who was developing a reverse image search system in this niche. I'll pass this on to him.
There are a couple of broad-based "fan art" reverse image search engines out there. One pretty comprehensive one is https://saucenao.com/.
It was trained on a tagged art gallery that also contains fan art.
I would expect some over prevalence of certain popular characters. For 25075, I would assume https://sonic.fandom.com/wiki/Amy_Rose#Heroes
For that same one, I see hints of https://sonic.fandom.com/wiki/Rouge_the_Bat (to the point where I thought of Rouge before seeing your comment)
You would not. Many GAN papers do nearest-neighbor lookups, and typically, generated samples are clearly different. Since Arfa's model is high-quality and the interps look fine, I would not expect it to be any different. (It would be hard to check because there are no pretrained classifiers to provide an embedding to do the search in, but one could do reverse-encoding.)
The fact that nearest-neighbor lookups do not find exact overlaps between the training dataset and a large number of random samples has always been one of the arguments I use against the widespread misconception that GANs 'just memorize' data: https://www.gwern.net/Faces#faq
That so many people are convinced that a given datapoint must be an exact copy of a training datapoint - "it looks exactly like a Zootopia character I recognize!" - is really quite a compliment to the GAN...
This has far fewer oddities than the people based ones. Maybe that's because furry art all have similar qualities already, so it's easier? A few of them were just straight up Judy from Zootopia, so that says something gross about the furry community, I'm sure.
Also, I hate this. I especially hate the loading messages.
What does that say that's gross?
I'm learning about some of my personal biases through this, apparently. I just assume furries/fursonas/furry fandom/whatever it's called is about sex.
> I just assume furries/fursonas/furry fandom/whatever it's called is about sex.
Nope! :3
https://faq.dhol.es/@Soatok/furry-fandom/is-furry-a-sex-thin...
It's a common misconception. We're more about community than anything.
(That's not saying that sex doesn't happen between adults, but that's not what it's about, for everyone.)
In my experience, any activity or community enjoyed by humans has a subgroup for “that thing, but also with sex”. It’s more of a reflection of what it means to be human than anything to do with the means of that expression.
It doesn't need to be gross to dislike the weird baby talk "uwu" stuff. Although seeing it in much more technical contexts will make me laugh - https://github.com/mpaland/printf/issues/15 - "Scawy big no functionality at all UwU #15"
Wow, that person seems to do this in all of the issues they file. They're also really good at filing issues despite the baby talk. All of their issues are well-researched and immediately actionable with a surprising amount of debugging completed before the initial submission.
Devs tend to have a penchant for the absurd. The details may change along with the zeitgeist, but the personality quirk itself seems to be ubiquitous across the decades.
>Devs tend to have a penchant for the absurd.
Ok this is just a stereotype that's not true.
:: notes reply comes from a commenter with a name that sounds like a fursona ::
...you're messing with me, aren't you. :D
not really sure what you mean, can you elaborate on why you would say that?
FWIW, I've met more normies in my 10years of software development than absurd people. Absurd people are just memorable, it's a bias.
If you're actually curious, this username was stolen from an absurd Atlanta based rapper named blunt fang and I get more weed culture association than furry association with it.
https://www.youtube.com/watch?v=bDal4GCyEJA
I don't agree with their framing but xxxxFang is a common naming scheme of wolf furries.
"uwu" and "owo" are emoticons people started vocalizing ironically, not baby talk.
The dataset grabbed has both fursonas and fan art in it.
There are not many furries with fursonas that look like Judy, though I'm not sure why that would be gross.
Fursonas tend to mirror the style the content one was watching when they were a kid, so once kids who grow up watching Zootopia become old enough to enter the fandom, then those individuals might bring in a 3D style, but it has yet to happen.
I was about to say I recognised Toriel [0] and was suspicious it might have been memorising a bit too much.
As for the idea that furry are has too many similarities… I don’t think so. The variance in fursona body morphology is huge. For example, one thing I didn’t notice in this set was Sergals [1].
[0] Undertail. And no, not like that.
[1] https://en.wikifur.com/wiki/Sergal
Undertale* though the typo makes sense given AI is on topic
Now I feel stupid. :)
Oh, I thought it was intentional, seeing as, uh, that spelling is used to refer to p*ographic fanworks of that work of fiction
Arfa excluded a few categories like ponies and scalies to keep it from being too broad for StyleGAN. Sergals would probably get excluded as part of that.
And protogens:
https://www.youtube.com/watch?v=DiSs-sGclcU
https://www.youtube.com/watch?v=3HwdsjXEbOQ
https://www.youtube.com/watch?v=gGOFU62wVco
Oh, so that’s what that critter is called. I had been wondering.
They took off when people started showing up at cons with protogen partials.
I saw a lot of verbatim copies of both the fox and the rabbit from Zootopia (possibly with shades of Disney's Robin Hood) and distinctive elements of Sonic the Hedgehog and Pokemon characters. I think it was trained on popular anthropomorphic media, not just furry art, and there's some overfitting. I don't think stylistic inspiration can quite explain its fondness for cheeky foxes wearing green.
Zootopia: https://thisfursonadoesnotexist.com/v2/jpgs-2x/seed18100.jpg https://thisfursonadoesnotexist.com/v2/jpgs-2x/seed73222.jpg https://thisfursonadoesnotexist.com/v2/jpgs-2x/seed46870.jpg https://thisfursonadoesnotexist.com/v2/jpgs-2x/seed88525.jpg https://thisfursonadoesnotexist.com/v2/jpgs-2x/seed16983.jpg https://thisfursonadoesnotexist.com/v2/jpgs-2x/seed38798.jpg https://thisfursonadoesnotexist.com/v2/jpgs-2x/seed34557.jpg https://thisfursonadoesnotexist.com/v2/jpgs-2x/seed20308.jpg https://thisfursonadoesnotexist.com/v2/jpgs-2x/seed65770.jpg https://thisfursonadoesnotexist.com/v2/jpgs-2x/seed13269.jpg https://thisfursonadoesnotexist.com/v2/jpgs-2x/seed31499.jpg https://thisfursonadoesnotexist.com/v2/jpgs-2x/seed78829.jpg https://thisfursonadoesnotexist.com/v2/jpgs-2x/seed02032.jpg https://thisfursonadoesnotexist.com/v2/jpgs-2x/seed24939.jpg
Pikachu: https://thisfursonadoesnotexist.com/v2/jpgs-2x/seed12878.jpg https://thisfursonadoesnotexist.com/v2/jpgs-2x/seed05116.jpg https://thisfursonadoesnotexist.com/v2/jpgs-2x/seed49143.jpg
Sonic: https://thisfursonadoesnotexist.com/v2/jpgs-2x/seed22797.jpg https://thisfursonadoesnotexist.com/v2/jpgs-2x/seed95651.jpg https://thisfursonadoesnotexist.com/v2/jpgs-2x/seed35704.jpg
It was trained on images from e621[1]. Considering the most popular characters[2] on that site, you can see where the neural net got it from.
[1]: The most popular furry booru. A booru is an image board where everyone can edit the tags of images. Strict tagging rules and advanced search operators make it easy to find specific images. A safe-rated only mirror is available at e926.net.
[2]: https://e926.net/tags?search[category]=4&search[order]=count
It’s still very realistic considering how many people’s actual OCs are just a popular character with a flood-fill of a different color.
* original character https://knowyourmeme.com/memes/original-character-do-not-ste...
Oh no, I don't mean to diminish the achievement on display here. It's very impressive how many of these could pass for original, hand-drawn artwork, and in such a wide variety of styles too.
There might be traces of Pikachu in that, but it's definitely Renamon. https://digimon.fandom.com/wiki/Renamon
More-comparable straight-on image here: https://neoencyclopedia.fandom.com/wiki/Renamon
This one looks almost straight up from the movie: https://thisfursonadoesnotexist.com/v2/jpgs/seed10592.jpg
I guess training on the source that has a massive crush on Zootopia wasn't the best choice for ‘doesn't exist’.
Also, those Sonics are apparently sourced from the gender-swap category.
Oi! Reminds me of Punkemon's Punkachu!
http://studiohunty.com/punkemon/
http://studiohunty.com/punkemon/ptown/punkachu.html
I mean it could just be that us non-furries aren't attuned to the subtle differences and conventions of fursonas.
What’s most interesting to me vis-a-vis other GAN portrait generators is the range of illustration styles employed here. It’s like a GAN that generates both manga and western comic-art faces without any incongruous hybrids.
Has anyone ever made some variant of these "* does not exist" sites with some control sliders/options, that controls various aspects of what's generated?
So things like: gender, hair color, face shape, etc?
Basically, an AI-driven character generator that isn't completely random?
Absolutely. Waifu Labs https://waifulabs.com/ implements a grid-based choice system for evolving anime portraits, and Artbreeder https://artbreeder.com/ implements controls plus crossbreeding and other features for a variety of StyleGAN/BigGAN models. There's plenty of scripts and Colab notebooks as well for various kinds of editing or control if Artbreeder doesn't do it for you. (I think Runway may also do editing but I haven't used them in ages.)
GAN models do not need to be specifically architected to enable control, because you can reverse them to get the latents/seed and manipulate that to 'edit' images: https://www.gwern.net/Faces#reversing-stylegan-to-control-mo... So if someone wanted, they could use Arfa's model to edit images.
This is incredible, thanks for sharing!
This looks like a Zoom meeting with 4chan.
Truly impressive results, though I find it unsettling that there's a decent chance that the future of all Internet aggregators is going to be "We threw data into computer and it threw out magic thing!"
Just wait until AI create random cartoon shows and comedy that end up being funny. That will be uncanny.
That's been going on for a while now with young kids television on youtube. It's usually more CGI than cartoon though.
Is there any evidence those are actually AI generated, or just quickly cranked out by humans using templates and scripts?
I tend to assume the latter, especially with any of the "Elsagate" style content.
That's a good point. I don't know if anyone has leaked any behind the scenes for those videos.
Is there a name for a Reverse Turing Test? Because those videos pass it.
As far as the Elsagate stuff goes, it's not that the videos themselves were algorithmically generated, but that their creators were driven by YouTube's algorithms to create the content that would generate the most views.
I saw an AI 'meme generator' the other day that seemed to make fairly insightful and funny memes. I think the future is close, and...very strange.
The funny thing about the AI meme generator is per the technical write up (https://towardsdatascience.com/meme-text-generation-with-a-d...), that network is even less advanced than the old recurrent neural network approaches. (it uses convolutional neural networks which are more used for images than text nowadays)
However, this is the one case where more incoherence works better.
Yes, it's a testament to crowdsourcing and filtering. Many people have argued that 'generation + filtering = creativity', and it's impressive what NNs+crowds can do. The NNs are unhindered by any merely human considerations and are indefatigable, while the crowd brings the filtering & evaluation to select just the best candidates.
https://www.reddit.com/r/SubSimulatorGPT2/ is already pretty amusing.
https://www.reddit.com/u/4chanGPT2Bot is pretty close to actual users on the site.
Or when AI starts producing improved AI algorithms.
Some of them are a bit weird, still.
This one is odd; are those horns or spikes? And what’s that sticking out the back (or is that from the ear?):
https://thisfursonadoesnotexist.com/v2/jpgs-2x/seed30101.jpg
It’s always fun when these kind of systems try to generate text:
https://thisfursonadoesnotexist.com/v2/jpgs-2x/seed25387.jpg
And then we have this one, which is pretty clearly Toriel from Undertale:
https://thisfursonadoesnotexist.com/v2/jpgs-2x/seed14017.jpg
The training set is (naturally) skewed towards a handful of characters. I wrote about it on twitter when I saw this: https://twitter.com/indrora/status/1258094389013831680
This AI passes the Furing Test.
In this case, the incidental heterochromia actually makes it feel even more authentic.
The heterochromia is probably not 'incidental'. Heterochromia, quite aside from lighting/shading differences in eye color, is present in many real images: Danbooru has a ton, and e621 has 11k images tagged 'heterochromia'. Just a common trope in illustrations.
This is the coolest use of AI I've seen in a while! :3
The thumbnails have human-level quality. The full images have oddities, but they disappear when downsized. DL is going to put animators, voice actors and even actors out of work sooner than I ever expected. It's going to be feasible for one person to create a full anime episode in three days. The amount of content is going to explode by orders of magnitude and shared culture is basically going to die, with nearly zero chance another random person consumes the same content. What a weird time to live in.
That's like saying snowclones and memes are going to kill book authors out of work... if anything, this is taking us to the next level of original work.
> The amount of content is going to explode by orders of magnitude and shared culture is basically going to die, with nearly zero chance another random person consumes the same content.
Why? I believe there's already so much content out there that this could easily happen today if the only question would be the raw amount. However if you ask people around the water cooler, instead of one guy watching joesmith34's Source Film Maker video on YouTube and another watching janedoe157's flash animation yesterday evening, everyone watched the latest episode of a popular Netflix series.
You were so caught up on whether you could. You never stopped to ask the question whether you should.
> As the images are generated by an AI, they are non-copyrightable and are therefore public domain. Feel free to use them any way you see fit. Just don't try to pass them off as your own art or sell them or anything.
The furry fandom is extremely touchy about "art theft." Considering how many of these images clearly resemble individual artists' styles, I'd tread a little more carefully.
> As the images are generated by an AI, they are non-copyrightable and are therefore public domain.
This part seems false. If I take copyrighted photos and run it through something like an "AI" filter to change the colors slightly, the photos aren't suddenly public domain.
If you steal from enough people at once is it then legal?
> If I take copyrighted photos and run it through something like an "AI" filter to change the colors slightly, the photos aren't suddenly public domain.
That is not what is being done here. See also https://en.wikipedia.org/wiki/Transformativeness
> If you steal from enough people at once is it then legal?
How many people does a great human artist steal from to become an artist? If you steal solely from one other artist, you're a plagiarist; if you steal from a thousand other artists, you're an innovator and pioneer of a new style...
This isn't transformative enough to pass a sniff test. They took multiple images and put them in a computer to make images that look like those images. What's transformative about that?
That's extremely transformative: it's learning from a large corpus what faces are and how to create brand new ones which cannot even be traced back to an original. (If you think even that is not enough to count as 'transformative', I suggest you look at the examples in the WP article of what has been considered transformative, like 'thumbnail screenshots of web pages' and 'putting a mustache on the Mona Lisa'.)
When a significant number look like zootopia characters I don't think you can claim they "can't be traced back to the original", lots of these are overfit
They are not 'overfit' unless they are exactly the same and they cannot generalize. But you can see the wide variety of high quality faces which are not copied from the dataset and that the interpolations are fine: https://twitter.com/arfafax/status/1258344026706599937
Generating a recognizable character is no more 'overfit' than a painter being able to paint George Washington means their brain is 'overfit'. If those characters are part of the data distribution, as they are, then the GAN can and should generate them and countless variants on them.
My point is that just because you get the zootopia fox out of a GAN doesn't mean the zootopia fox is public domain.
Perhaps not, but that is completely different from what data it uses or whether it is 'overfit'.
My experience as a pro artist is that once you can swipe from eight people whose work is mostly unrelated to each other, you have a Unique Style.
Disney is kinda touchy with copyright, and the pirate bay still exists.
I don't think the internet, where 4chan or the dancing coffin meme were born, is really a medium that stops at the threat of an outrage.
All I see are an infinite number of unique user avatars for every Mastodon/Pleroma/Misskey instance of the Fediverse.
Regarding "overfitting": I posted an interpolation video here -https://twitter.com/arfafax/status/1258344026706599937?s=20
This demonstrates, at the very least, that it isn't simply memorizing the datapoints. You can see that it is able to smoothly transition between images of Zootopia characters and other characters, which indicates that it has learned a lot more about the actual features.
I believe the prevalence of certain characters (Zootopia, Sonic characters, Pokemon) showing up is because a large portion of the input space maps to those regions of the latent space. So I'd expect there to be a roughly equal proportion of images that look like Nick Wilde in the random samples as there are in the training data.
I'd love to see this for spaceships.
I assume that this would only work reliably for headshots/avatars? Still, if I were an artist whose income came from furry art, this would terrify me.
No, because a fursona is something that's deeply personal to the person, and typically the process for creating one looks like this:
1. Decide what you want to look like
2. Create a reference sheet (or commission an artist to create one based on a description from step 1)
3. Commission a headshot for an avatar
4. Commission some art of your fursona, possibly with your friends' fursonas
5. GOTO 4
6. At some point, your friends may ask you for your reference sheet so they can gift you an art piece of your fursona and theirs (see optional stage of 4)
Even if you can churn out AI-generated headshots, getting the colors/markings just right for your character is nontrivial for non-generic fursonas (see https://soatok.com/static/soatok-johis-responsive.jpg for example).
And besides, most art commissions are conducted after you have a reference sheet and/or headshot of your character.
If anything, this will give artists something they can point the "steal other people's character art for their roleplay accounts" types of (especially younger) furries towards. "Can't afford to commission an artist? Just use the AI thing and stop the misbehavior!"
there has been some progress in the anime category: https://twitter.com/theshawwn/status/1257683697777569792
This one generates a character and a plot: https://www.thiswaifudoesnotexist.net/
That's much older, this discussion is about extending beyond portraits.
> I assume that this would only work reliably for headshots/avatars?
StyleGAN works best on single centered objects which are not too diverse in shape. So heads work great if you align them to the middle, and you can get OK results with single centered figures like someone standing, but anything beyond that and it falls apart. We have a theory about that (https://github.com/tensorfork/tensorfork/issues/21) but we've long since moved onto BigGAN, which can model much more diverse datasets like ImageNet successfully.
As far as fursonas go, see my other comment: https://news.ycombinator.com/item?id=23096442 I'd say that GANs are better at letting someone explore fursonas than any human artist is or ever will be. You cannot click a button and have a human artist generate 32 variants in a second for you, toggle on or off attributes, nor can they slide a slider from 'human' to 'neko' to 'dangerously cheesy!'. You might pay for a commission of your final fursona for the highest-quality image with no artifacts, but for exploring...? And as NNs get better over time, they'll creep up the value chain, as it were. (I'm still a little shocked how good the human voices in OpenAI's Jukebox https://openai.com/blog/jukebox/ are, even if the highest level of abstraction is still very shaky and they need to tack on another layer or two to get choruses etc.)
I can't wait for the first "your AI stole my original character!!!" furry thread on Twitter.
Impressive achievement, I'd love to see the outcome of re-training it with fan-art removed from the training set. There's a lot of Nick Wilde and Judy Hopps in there...
Some furries were already having this discourse as soon as this hit HN.
This was expected from:
- It was already possible to create fursonas based on pictures of human faces.
- It was possible to generate pictures of faces of humans that do not exist.
Pipe these two together and you get this ...
Hah this is really neat and a fun use of AI.
No horned animals? Discrimination!
That would violate the "no horny on main" rule
Or scalies. They all seem to trend towards feminine cat/fox 'sonas. I guess it's a result of the data-set used.
> They all seem to trend towards feminine cat/fox 'sonas.
The gender bias is actually really interesting. I'm curious how much of it is due to:
* Inherent bias in the source dataset. (For example, nekomimi art is overwhelmingly female -- there's a reason they're referred to as "catgirls". The GAN does a great job of distinguishing this from Western anthropomorphic art styles, incidentally.)
* Bias introduced through filtering of the dataset, e.g. by excluding NSFW source material or certain tags, or a bias in what types of faces were recognized.
* Androgynous faces being interpreted as feminine by default.
* GAN-specific characteristics of the output (like smooth features) being interpreted as feminine.
I think it's bais towards catgirls and other anthropomorphized anime characters which is skewing the results. If you go to FA, you would have to look hard to find feminine like this. It's either using a catgirls, as you say, heavy dataset or mixing two unrelated datasets.
I don't like the word feminine, because it's not really what furry feminine is. It's a human-applied-to-furry concept.
ACTUALLY. As I refresh the page, I get the feeling that it's the result of weird mixing between flat shaded art and mixing anthromorphic and "cargirls" (for lack of a better term- and I don't want to say east vs western, because catgirls dont' fall under furry)- anime inspired furry. It's weird, and makes me ponder. If the underlying datasets are a more similar art style and didn't have anime inspired art mixed in, I think the result would be less confusing and cooler.
Still an awesome project.
According to a comment way up top, yes to the species: scalies and ponies were excluded to keep it relatively simple for StyleGAN to work with.
> Girls are fetching fursonas. Please wait warmly
Nice ref there.