This Anime Does Not Exist
thisanimedoesnotexist.ai
thisanimedoesnotexist.ai
https://thisanimedoesnotexist.ai/results/psi-1.1/seed0300.pn...
https://thisanimedoesnotexist.ai/results/psi-1.5/seed7458.pn...
https://thisanimedoesnotexist.ai/results/psi-1.2/seed1289.pn...
https://en.wikipedia.org/wiki/Hentaigana#Gallery
(That's hentai as in 変体 "variant form", not hentai as in 変態 "sexual perversion".)
Unfortunately their profile [1] is gone, and the Internet Archive doesn't seem to have saved the post [2] either. Nevertheless, we've looped back to neural networks again.
"I think that あ, お, and ま are the same thing, and every line needs to start with it!"
Some Fun Monsters:
https://thisanimedoesnotexist.ai/results/psi-1.8/seed3103.pn...
https://thisanimedoesnotexist.ai/results/psi-1.8/seed4437.pn...
https://thisanimedoesnotexist.ai/results/psi-1.8/seed8810.pn...
https://thisanimedoesnotexist.ai/results/psi-1.8/seed3309.pn...
https://thisanimedoesnotexist.ai/results/psi-1.0/seed9278.pn...
https://thisanimedoesnotexist.ai/results/psi-1.8/seed8300.pn...
The dataset is actually ~3x bigger than ImageNet, FWIW.
Oh that is very interesting. Thanks. Do you know anything else about the dataset? I was looking at the link for some quick stats and didn't see them. I'm curious if there are metrics to support or disprove my hypothesis about body features. If so ADA would still help because there isn't an even distribution of the features.
https://thisanimedoesnotexist.ai/results/psi-0.6/seed38513.p...
I specially like this one.
https://thisanimedoesnotexist.ai/results/psi-1.8/seed4437.pn...
https://thisanimedoesnotexist.ai/results/psi-1.0/seed7820.pn...
And a few more monsters:
https://thisanimedoesnotexist.ai/results/psi-1.0/seed18905.p...
https://thisanimedoesnotexist.ai/results/psi-1.0/seed47133.p...
https://thisanimedoesnotexist.ai/results/psi-1.0/seed2954.pn...
https://thisanimedoesnotexist.ai/results/psi-1.0/seed16924.p...
https://thisanimedoesnotexist.ai/results/psi-1.0/seed30742.p...
I get migraines with visual hallucinations, and while I don’t get kaleidoscopic hallucinations, this kind of reminds me of what I’ve read about them.
They seem to be frequent, but only because there are an ungodly number of submissions to danbooru. Think of it more like “for every male character, there are at least 10+ female characters.”
And it doesn’t help that a lot of the male characters are drawn in a female style...
This explains a lot about why the generated images are the way they are (having a lot of borderline NSFW pictures and even if not having poses which are sexualized and/or imply sexual content).
While Anime tend to be more open to such things and tend to have often quite a bit of fan service the generated images are way to much biased compared to the "normal" anime/manga content you find on platform like crunchyroll or similar.
I mean a non negligible part of the pictures they used have likely hentai variations if you disable safe mode (If Danbooru is anyway similar to a site I used a view years ago to get nice screen background, if used in safe mode. And this seems to be the case.)...
But then due to practical (and maybe copyright) reasons it's very viable to use Danbooru as source but not so much to get snapshots of Anime's or pages from "real" Mangas.
https://i.imgur.com/9eO6ZpU.png [standard warning that it might contain nudity and violence exceeding your personal threshold]
It seems to me that if it truly used Danbooru, it also filtered for specific tags, which is also indicated by the art-style that it uses. The ratio of ostensibly male to female characters there is more so 7:18
There are tags for male/female which will be a better representation.
1girl: 3038k
1boy: 559k
Your idea about they filtered images could still be true, to be clear.
https://thisanimedoesnotexist.ai/slider.html?seed=11761
https://thisanimedoesnotexist.ai/slider.html?seed=3817
https://thisanimedoesnotexist.ai/slider.html?seed=12386
Maybe the source dataset was too small, and it was not trained strictly enough, or it should have actually trained on anime stills instead of these Moe manga style drawings.
The state of the art, for example that site that generates images from sentences is much more impressive to me.
The fact that it looks overfitted is a sign that the model is legit. The interpolation videos are much more convincing.
https://images-ext-2.discordapp.net/external/4Om03hvtvjDhT4c...
Come on, DALL-E came out like last week - give us a break! That was a major research project contributed to by at least 11 fulltime OpenAI researchers using 400 million images; and this was like 3 hobbyists in a Discord occasionally.
Anyway, I did the most hated thing on HN, casually dismissing someone's hard work or careful thoughts. I'll go chew on some sand now.
The other 50% of the time it goes wrong, and the result is either a human that is not anatomically correct, or something that doesn't even seem to be a two-dimensional projection of three-dimensional Euclidean space.
From the detailed writeup:
> Broadly, his StyleGAN2-ext increases the model size and disables regularizations, which are useful for restricted domains like faces but fail badly on more complex and multimodal domains.
https://thisanimedoesnotexist.ai/results/psi-1.0/seed9022.pn...
I have no idea if this comment is off limits for HN. I'm surprised to see no one else voicing their discomfort with the results of this StyleGAN output.
More reference images with what looks convincingly (to a nonspeaker's eyes) like text:
https://thisanimedoesnotexist.ai/results/psi-1.5/seed9501.pn...
https://thisanimedoesnotexist.ai/results/psi-1.0/seed6369.pn...
https://thisanimedoesnotexist.ai/results/psi-1.0/seed0527.pn...
https://thisanimedoesnotexist.ai/results/psi-1.0/seed8073.pn...
https://thisanimedoesnotexist.ai/results/psi-1.0/seed8119.pn...
apparently it's a very fine line between "dream girl" and "nightmare girl"...
EDIT: It also a very reasonable source as it's very hard to get large amount of anime/manga images which mostly are in color, which more importantly are tagged and you know you will most likely not get in trouble for using them.
edit: well it's translating a div but still pretty neat
This an uninformed thought but if you're extrapolating something. If we took the base knowledge math/physics can AI/ML stumble onto the next undiscovered stuff?
Somewhat tangent, like how you can conceptualize something in your brain, you could pre-load your neurons randomly to figure out some new idea. I imagine most of it would just be noise.
I feel really excited to experiment on such projects in my spare time but reading how it costs researchers hundreds and thousands of dollar don't even try.
The good news is that there's been huge progress in GAN efficiency over the past few years, and we are rapidly approaching the point where ImageNet-scale datasets can be reasonably trained with 1 GPU-month or less.
Worth noting in that vein are VQ-GAN https://compvis.github.io/taming-transformers/ and not-so-BigGAN https://arxiv.org/abs/2009.04433 . There's also a resurgence of VAE work which is relatively efficient: VD-VAE https://arxiv.org/abs/2011.10650#openai and NVAE https://arxiv.org/abs/2007.03898#nvidia ; DALL-E https://openai.com/blog/dall-e/ also appears to be surprisingly lightweight on the VAE side.
So, if you feel left out, don't worry. In the near future, between the RTX GPUs and optimization and new VAE/GANs, I expect things like Danbooru2020 https://www.gwern.net/Danbooru2020 to be totally doable on hobbyist hardware.
Somebody recommend Andrew NG courses but I wasn't able to complete it and got bored. Then I tried watching sentdex and it was amazing and I'm able to fully comprehend what he does since he teaches more practical aspects like running gta v on autopilot etc.
So just wondering what I could learn next.
This AI creates tumor-like deformities, where limbs are missing or melded together. As an example, here's an image that couldn't decide if it wanted the girl to have a large chest or a dislocated shoulder: https://thisanimedoesnotexist.ai/results/psi-0.8/seed9182.pn...
With everything else and how it mashes together being an afterthought.
We can patch our behavior and pretend we are looking at other things, but the GAN aggregates what all of the artists did and has an aesthetically pleasing result even when they don't make sense upon closer inspection, because it is including the features we look at.
https://thisanimedoesnotexist.ai/results/psi-2.0/seed37303.p...
https://i2.wp.com/overmental.com/wp-content/uploads/2015/03/...
Now with photorealism:
https://dvdnewsflashthereviews.files.wordpress.com/2018/05/6...
Hopefully someday those top quality Anime would be cheaper to produce with the help of Machine learning and AI. And hopefully those Japanese companies [2] hopefully take the world wide market more seriously.
[1] https://anime.stackexchange.com/questions/39689/why-is-anime...
[2] https://hyperjapan.co.uk/anime/anime-studios-unite-to-create...
> As animated television goes, anime is actually made on the cheap.
Regardless, outside of videogames autogenerated content turns me off (and sometimes even in videogames, you can tell if something was handcrafted or autogenerated). If the way to make animé "survive" means removing meaningful human authorial input, then maybe it's best if it dies?
Well respected sci-fi authors of all times have seen this scenario as a dystopia. I remember one short story -- was it by PKD -- where an author inputs a novel "high concept" into a computer, a second later says "wait, I have a minor correction" and the computer replies "too late, I've already written the novel and its sequels based on your concept, it's already being distributed".
Surely the answer to "this requires top talent and human artistry, and this stuff costs money" cannot be "replace them all with Machine Learning", can it?
what if it simple generated frames between the ones drawn by hand? so the artist doesn't have to draw every frame just key frames say on ever second or half second and the ai filled in the in between frames while maintaining consistancy
They will be bought by the same sort of people who buy Amish made furniture, listen to vinyl records, or who dabble in vintage computers.
I don't think anyone ever suggests that weird extreme outside ai-as-art-itself. Usually the idea with ai creation is "would your work be easier if I provided something 90% there that you fix up as needed". People will not get replaced soon, but rather provided with tools like clippy "hey, you're drawing a person, would you like 100 auto-generated running posses that you don't have to draw manually?"
In the general case I'm still skeptical though. Consider the case of comics, with the jobs of penciller and inker. Is the inker an automatable job? Do they not add important artistic decisions to what at first sight would seem like a merely mechanical job, to the point some inkers have their own fan base? (And, it is argued, a good inker can "make or break" the pencil art). And if this is true, could it be also the case that even the lowly in-betweener brings his/her small grain of artistic input to the final piece?
In the context of anime, I'm talking about a hypothetical future (hinted at by the person I was replying to) where characters are designed by ML/computer automation, with minimal human input -- or in a nightmarish scenario, even entire plot lines and episodes! -- to drive costs down.
In the context of videogames, I mean the battle-tested technique of sandbox and roguelikes (and the games they inspired) of generating missions, scenarios and levels programmatically and randomly, so that you don't get the same game twice. It's cute, and for roguelikes it works, but it's also very different to levels created by hand.
Of course, all salaries and in general production of almost everything is cheaper in most nations than in the U.S.A. simply because very few taxes are paid in the U.S.A., so employees receive more salary, but every product also costs more.
For instance, in 2019, the average net compensation for a salaried worker in the U.S.A. was 51 916 U.S.D., with median beng 34 248 U.S.D.
I found those numbers for Japan to be 39 851 U.S.D. converted, and median 62 9475 U.S.D., actually making median higher due to the top-heavy nature of U.S.A. incomes.
This top-heavy nature might also be why it's not far to compare absolute star shows with each other.
I, frankly, do not really understand how The Simpsons could ever take more resources to produce than anything which involves choreographed, three-dimensional fighting scenes that involves characters dazzling and jumping through buildings:
These Lyrics Do Not Exist https://theselyricsdonotexist.com
This Eye Does Not Exist https://thiseyedoesnotexist.com
This Cat Does Not Exist https://thiscatdoesnotexist.com
https://thisanimedoesnotexist.ai/results/psi-1.0/seed3756.pn...
I had someone walk in as I clicked on it. "What that?" I closed the browser and replied. "This anime does not exist".
It uses not a GAN but a different deep learning mechanism (the concrete language model is called GPT-2). See also https://datascience.stackexchange.com/questions/48005/which-...
The model has no idea about what arms are supposed to look like [1] but has no problem making sure that the resulting abominations are wearing underwear only. Also missing: men.
[1] https://thisanimedoesnotexist.ai/results/psi-1.0/seed3578.pn...
Worldwide anime and related media have a mostly female audience and certainly don't lack male characters. In Japan women write more fanfiction relative to men who draw more fanart, I think, but I've never counted.
There are other sites with more female users (eshuushuu and zerochan iirc) but danbooru's data is better because of a combination of 1. not allowing horrible poorly drawn fetish art or selfposts 2. ignoring the fact that most of the good artists don't want to be posted in the first place.
The a.i. was clearly trained using moe only.
Moe and fan-service tend to walk as one. You will also notice that the overwhelming majority of the characters look female. A high to absolute number of female characters is also common to moe.
I'd actually be interested what turns out if an a.i. actually be trained from a completely random selection of last year's releases in animation from Japan and whether it would manage to keep different art styles apart, or somehow blend them.
I don't really believe in the common claim that it is easy to recognize animation from Japan from stylistic elements, it just so happens that it seems that moe is often what is being talked about:
- https://i.ytimg.com/vi/dR-648aYD6s/maxresdefault.jpg
- https://pbs.twimg.com/media/ENxzRVVVAAIzbOC.jpg
- https://encrypted-tbn0.gstatic.com/images?q=tbn:ANd9GcQHB0hi...
- https://i.pinimg.com/474x/57/71/bb/5771bbd550c96e1799659e2a0...
- https://assets.mycast.io/characters/monkey-d-luffy-12086-nor...
- https://www.indiewire.com/wp-content/uploads/2016/07/blame.p...
- https://miro.medium.com/max/3840/1*opC8ndIEWzVartwBG16Chg.pn...
I think it would be quite interesting to see what were to happen if the a.i. would be trained across a library of very different art styles.
Something I noticed is that AIs appear to have trouble with differentiating perspective vs art style. What I mean by that often characters are in 3/4 view, which naturally means one eye will be smaller than the other due to perspective. But some art styles have bigger eyes than others. What I see sometimes is that the AI then produces an image where each eye comes from a different art style.
It doesn't appear to account for the concept of a three-dimensionally spherical head and how perspective works with respect with such shapes.
It most certainly would not produce good results unless you specify that stuff like shirokuma cafe[0] characters are not human.
Was the This Person Does Not Exist-a.i. simply trained with pictures of human heads? As it seems to be quite capable of keeping many things apart that are commonly kept apart.
T Waifu DNE uses similarly centered and cropped portraits, which is why its output is consistently of higher quality than T Anime DNE.
There is a pretty large fanservice/borderline NSFW/smut and similar Anime industry.
But there are also quite many Animes which don't have the problem or at most have a very small amount of fan service "sprinkl in". But in this generator the borderline NSFW stuff seems to be the main topic.
A while it indeed get's better after some time (interestingly, why does that happen?) it doesn't change that most characters are in implicit poses.
Maybe clothed yes, but you can also put clothed characters in poses which imply sexual content.
I also realized that the source of the images show is largely from Danbooru (with safe mode), which pretty much explains everything ;=)
It seems to be in flux.
Boobs, the system models very well. This probably reflects the training data.
I don't know what I expected.
A storefront generator would be useful. There are procedural city generators that generate blank buildings in a reasonable way, but can't fill in the details.
I dare say that site is a mild threat to my productivity.
Sports is a different topic though: https://www.fakefake.news/article/17nwpk-dolphins-agree-to-r...
Most Japanese animation is entertainment targeting young teenagers and younger. Japan does however have a considerable adult animation market. The a.i. is not trained with any “anime”, however, but with still promotional art which is of far higher quality and clearly trained with the adult market where sexualization is, as one might expect, more common.
The paradox is that the unrealistically cute “moe” designs are typically targeted at adults and adolescents and most of the children's antertainment does not look like that.
Often in Japan, one speaks of four principal demographics: young males, young females, older males, and older females. — the a.i. is quite clearly entirely trained with the “older males” demographic, which should be undertood to target around 18-30 years of age, and even within that demographic it is trained almost exclusively with moe art.
Most “anime” would not even be instantly recognized as such, which is why I think the term is fairly useless. — it seems when people speak of “anime” more often than not what they mean is “moe art” whether it be animated or not.
This is what the average “anime” looks like: https://assets.moomin.com/uploads/2016/02/Moomin-animation-s...
So girls draw sexualized girls, some men do as well. Some says the high composition of female artists comes from the fact that moe arts are extension of Shoujo Manga market where predominantly female characters are portrayed by predominantly female artists for predominantly younger female audiences.
Whatever the actual reason might be, it’s probably more complicated than “person with polar opposite sexuality of the subject consume porn”.
Anime is far from a male-only thing in japan and honestly most of the time if I see anime/CGI boobs on my twitter timeline it's women retweeting them.
Best to just let people like what they like, and if you don't like it, let the people who do enjoy it in peace.
Well, this is not representative of all anime, but admittedly of a large subsegment.
> but why are so many anime women depicted in an overly sexualised manner?
Because it sells very well to young males, especially the stereotypical basement-dwelling sexually frustrated nerd population. And Japan is the land of the Hikkikomori, after all.
But it's not like this kind of thing doesn't exist in the USA either. Just take a look at some superhero comics.
I do not believe that many of these “western taboos”, invariably dealing with sexual themes, would be so taboo in most of continental Europe. — it is mostly the Anglo-Saxon who is known for his staunch moral control on sexuality, swearwords, and nudity.
I think it's a bit of a stretch to say that's what Oreimo is about, even though the theme and idea underlies the series :-)
There have been far more explicit works. — Aki-Sora and Yosugano Sora feature explicit sibling sex scenes that show everything but the genitals.
There's a difference between sexism and sexualization in entertainment. Say one could enjoy some kinky porn and at the same time not treat all women as objects in day to day life. The same way you could enjoy a violent movie and not go on a killing spree afterwards.
There's clearly a demand for this kind of stuff and I don't see any point in trying to shame it out of view - it won't change the fact that people enjoy it (they'll just enjoy it in secret).
_edit_
Not sure why op deleted the above comment, he did have some valid (altho a bit dramatic) points. Looking back the line that I quoted might have been sarcasm.
I find this to be a big difference that sets the U.S.A. apart from the other two media giants: India and Japan.
India and Japan are far better it seems at tapping the market of male sexualization and U.S.A. media has never quite done that.
I would in fact argue that this non-sexualization of males seems to be a very Anglo-Saxon cultural property, but many other cultures do not have the large media market to show it.
I've seen quite a few comments coming from the Anglo-Saxon world about the Internationally released Swedish series Love & Anarchy, about how surprisingly attractive and sexualized the males were.