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.
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.
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.
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.
• 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.
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.
> (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.
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...
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.
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.
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.
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.
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"
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.
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.
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.
Readit News
[1] https://en.m.wikipedia.org/wiki/Furry_fandom
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.
As you say, some of the output images are clearly of specific characters, which turns this from "legally grey" into "definitely not public domain".
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.
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.
This is great work, but did you really have to?
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.
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.
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.
• 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.
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.
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.
What do you mean by this?
> (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!)
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.
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.jpghttps://thisfursonadoesnotexist.com/v2/jpgs-2x/seed13901.jpghttps://thisfursonadoesnotexist.com/v2/jpgs-2x/seed31957.jpghttps://thisfursonadoesnotexist.com/v2/jpgs-2x/seed25075.jpg
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...
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.
I would expect some over prevalence of certain popular characters. For 25075, I would assume https://sonic.fandom.com/wiki/Amy_Rose#Heroes
Also, I hate this. I especially hate the loading messages.
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
https://www.youtube.com/watch?v=DiSs-sGclcU
https://www.youtube.com/watch?v=3HwdsjXEbOQ
https://www.youtube.com/watch?v=gGOFU62wVco
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.
Zootopia: https://thisfursonadoesnotexist.com/v2/jpgs-2x/seed18100.jpg https://thisfursonadoesnotexist.com/v2/jpgs-2x/seed73222.jpghttps://thisfursonadoesnotexist.com/v2/jpgs-2x/seed46870.jpghttps://thisfursonadoesnotexist.com/v2/jpgs-2x/seed88525.jpghttps://thisfursonadoesnotexist.com/v2/jpgs-2x/seed16983.jpghttps://thisfursonadoesnotexist.com/v2/jpgs-2x/seed38798.jpghttps://thisfursonadoesnotexist.com/v2/jpgs-2x/seed34557.jpghttps://thisfursonadoesnotexist.com/v2/jpgs-2x/seed20308.jpghttps://thisfursonadoesnotexist.com/v2/jpgs-2x/seed65770.jpghttps://thisfursonadoesnotexist.com/v2/jpgs-2x/seed13269.jpghttps://thisfursonadoesnotexist.com/v2/jpgs-2x/seed31499.jpghttps://thisfursonadoesnotexist.com/v2/jpgs-2x/seed78829.jpghttps://thisfursonadoesnotexist.com/v2/jpgs-2x/seed02032.jpghttps://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
[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
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.
* original character https://knowyourmeme.com/memes/original-character-do-not-ste...
http://studiohunty.com/punkemon/
http://studiohunty.com/punkemon/ptown/punkachu.html
So things like: gender, hair color, face shape, etc?
Basically, an AI-driven character generator that isn't completely random?
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.