AI tech generates entire bodies of people who don't exist
ctvnews.ca
ctvnews.ca
Something genuinely useful might, perhaps, be achieved by applying GANs to parameters of those 3D models to allow easier customization, but I haven't seen any work in that direction.
Looks better to me than a lot of what is on your video. The video is very nice and it's really good, however it's often not really photorealistic when it's supposed to be and I also don't believe it'll be possible to do raytracing at such performance it'll be usable for most people (mobile phones), for VR. This approach is a new way that sidesteps the issues of raytracing.
Your comment has value, but your tone is unnecessarily harsh. Would you talk to a coworker like this?
https://deepmind.com/blog/neural-scene-representation-and-re...
Fashion isn’t about looking good in of itself. It’s about social proof-of-work. Knowing what to wear and pulling it off isn’t easy. It’s the raison d’être of the multi-billion dollar fashion and fitness industries. Women eat grass all day everyday and go on ridiculous juice cleanses while contorting their body into shapes that would put a pretzel to shame. If the world endures so much pain to look effortlessly good, Why should manufacturers be allowed to forego the pain? If the root of aspiration is making the impossible look like a breeze, then why would anyone want to buy from a computer generated perfect model that never had to feel the pull of a difficult stretch and stare in the mirror with guilty eyes?
The suspension of disbelief necessary for the social economy to function crumbled as soon as something is too perfect. Like Instagram photos. It’s not cool to slap on a filter anymore and stage a perfect photo. It’s too easy now. Just like Prisma was. What’s cool is to put in the work and make something imperfect so that everyone knows that you put in the effort and looked unflappably blasé while doing it.
Most models that are hired today are hired partly on their social presence. Having an audience, even a micro-audience, is essential for getting high paid gigs, because brands aren’t just selling the look; they’re selling the whole package associated with the look. Women like to buy from other women (and men like James Charles) whom they can aspire to. Tweens and teens want to look like the models on tik-tok. Young-adults and grown women want to look like their favorite YouTuber and Instagram celeb. They all want to participate in making something cool happen with their friends.
All of this and more is intrinsically social and human.
All of this and more means that as long as humans buy from other humans, brands will hire models to appeal to what’s most human; sexy, sexy crazy-making envy.
> Because what people (women and men) are buying is aspiration wrapped in dead plant fibers
In this case I agree. If someone buys the watch that Elon Musk wears, it's not just about wanting to look like Elon, it's about want to BE him. This effect works even if rationally people know that they are not him. Emotionally, they just want the feeling, however fleeting, that they are him.
> The suspension of disbelief necessary for the social economy to function crumbled as soon as something is too perfect.
Here I disagree and the example of fake Instagram profiles are my counter point. Some of these accounts used real people but the entire "lifestyle" was faked and yet they still had thousands and thousands of followers. What's the difference between creating a fake person to drop into that fake profile and having it be a real person?
One side effect of that is that while modeling has always been very hard to pierce into, at least there are modeling opportunities for non celebrities. As the practice of using AI to generate models for all but higher end stuff grows, it will get harder and harder to pierce into. It’ll go from an industry with tens of thousands of workers - some paid a lot, some much less - to just a few hundreds of highly paid models who have every incentive to prevent newcomers from taking their spot.
It’s not a phenomenon unique to fashion, and it’s entirely problematic when tech kills the long tail of this sort of occupation. How will those people pay their bills now? Not all of them can become AI engineers, that’s for sure.
The same way the luddites paid theirs I suppose
What feels "human" can be manufactured. Imperfections can be artificial, too.
We like to believe we're flawless arbiters of what's human, but we can be fooled in almost all cases. Most of all in photos. The soul remains an abstract concept.
You buy the dress because you want to convince yourself you look like old school Crawford or you want others to be convinced that you look like old school Crawford.
The point of models 99% of the time is to show the clothes. This isn't going to get rid of super models, but it's going to step on the turf of all but the most popular models.
There are some people who are famous, but the vast majority of working models are unknown people who will usually stay that way.
This aspect of fashion gives you narratives you can use to weave the story you tell about yourself and present to others. AI generated characters can tap into this with created narrative while constantly also posessing the supernormal stimuli currently achieved by using photoshop on real models. That coupled with the switch in the focus of mass media from films to games and I think that while real models will not die away, digital models will not just be used in catalogue settings. I wouldn't be surprised if 2-D from the Gorillaz became the model for a clothing brand for instance, so it isn't even a question of photorealism.
The ability of the wealthy to transmit this to their heirs, the tendency of children to enter the same professions as their parents. Social proof-of-work.
How much of our life is just social proof-of-work...
Take for example Japan - the land of anime girls on every piece of advertising. How much would the demand for human models expand if illustrated characters were not an alternative?
There's a scarcity in the type of human beauty that's sought after by the fashion industry. What if that scarcity didn't exist?
If people grow up with this new economic model of beauty and scarcity, maybe it would change the landscape of fashion entirely.
That said, one recent interesting development I saw is VRoid Wear, a collaboration between fashion brands and an anime-style 3D avatar creation tool. The brand Chloma is releasing the same apparel for avatars and real people simultaneously. The avatar-wear is not a mere promotional giveaway for the physical-wear: both are treated equally as products for sale, as they are selling the model data and textures for the former at around $50. (The physical version goes at something closer to $500, but still.) https://hub.vroid.com/wear/chloma
While i know few instances where they exist, that’s just not the reality in japan. Akihabara is as much of a japan as Silicon Valley is Usa. At certain level yes but on the other hand not at all.
There might be an collaboration with some anime and product once in a while but from my experience (and this is an experience of an foreigner living in japan) it’s far from norm.
You’re assuming that brands won’t create their own personas that drive that same result.
In fact, this has already happened - the best known one isn’t even a human - he’s a mouse, and millions of people buy t-shirts and other things with his picture on them. The company that owns him is rather large, and very good at appealing to humans with non-human characters: Disney.
This is the beginning of the end for models. Actors are next.
It's honestly exciting to imagine where we'll be in twenty years. I don't think we'll resent it - it's got fascinating artistic potential and will probably be quite liberating to finally realize that the things we strive for are artificially distilled.
You don't have anything to backup your assertion (or even have the possibility of evidence, before what you are assertion is what's in the mind of people).
When I browse through a fashion magazine, most models are unknown to me, they might as well be AI-gen.
The people that you know that tries to sell you cloth comprise a small fraction of people-images that you are exposed to.
Much to my surprise, I remember reading some of them. I don't know who you are, but you're a good writer.
However, if I were just an above-average model, where the main function is to display clothes rather than weave a narrative, I might be worried about this technology because many use-cases of models may be in jeopardy from this technology.
Virtual influencers may still work well for some brands/purposes. Here is the digitally generated Colonel from KFC:
https://www.adweek.com/creativity/kfc-created-a-virtual-infl...
I suppose it's possible that there's a man out there somewhere who thinks like this, but I've never met him and this is the first time I've heard of him.
If anything, most men probably prefer the opposite -- clothing that hasn't been worn by some obnoxious socialite.
It would be interesting to take some of these GAN demos and write a utility where for each output, have it find the closest-matching training input, and see how similar it is.
I don't think GAN work that way (I haven't followed it last year though). My understanding is that a GAN simply tries to determine if an image is generated or authentic. I suppose they used one there.
One way to do what you propose would be to plot all the training pictures in the kernel space of the autoencoder and match the closest one to the generated picture.
To answer this one would need to invest time into researching something that might potentially deflate the AI hype, so no one is doing it. Training dataset similarity analysis I usually see in these papers is downright laughable.
> Child pornography under federal law is defined as any visual depiction of sexually explicit conduct involving a minor (someone under 18 years of age). Visual depictions include photographs, videos, digital or computer generated images indistinguishable from an actual minor, and images created, adapted, or modified, but appear to depict an identifiable, actual minor.
But that's not so in Japan. So one ought take care when torrenting Japanese films in the US.
0) https://en.wikipedia.org/wiki/Child_pornography_laws_in_the_...
Obscenity itself requires an average expectation of what obscenity is, otherwise the exception dissipates entirely.
It is therefore easy for me to see that the technological availability of this would change it. there would be attempts by executive branch and legislature that are obliged to simply say something “should” be illegal, and a court wouldnt be able to agree with that on constitutional grounds
I think Denmark briefly allowed “all” pornography. Zoo was legal there until a few years ago, even had commercial production.
3D body scanning to get a representation of your body. Than use software to automatically generate and cut clothing patterns to get the exact fit you're looking for. TNow all that's left is stiching, which would be fairly difficult to automate when patterns are variable so would still be a significant cost point for custom clothing.
Let the AI generate the color patterns (sticking with Gap as an example, it might retain their chosen /size as constraints) and boom, on demand styling
Basically start by changing colors around.
Until 3D printed textiles get there anyway.
Or perhaps there will be manufactured AIs that star in films and model clothing.
Funny enough, there was a case[1] of a real person thinking a model on a website was him but it turned out the image was that of an altogether different person that just looked like them, uncannily.
For illustration, let's say the AI generated a "body" that looked eerily similar to Kim Kardashian or the president of Burundi and used it to promote products or causes.
[1]https://www.npr.org/2019/03/10/702063209/man-inadvertently-p...
Or we could put it this way: a corporation might see this as a choice between A) creating a whole new branch of case law to test in the court system, fueled by highly paid lawyers, or B) just pay a part-time model to be able to say "this wasn't you nor inspired by you, because that's the the model and we entered this contract together"
This tech generates random images of models in random clothing. Maybe it could be instructed to do 'male' and 'brown' and 'jacket' - but never with the specificity required of a specific piece of clothing.
All clothes fit and drape in a unique way.
Also - models for regular clothing are a commodity and the process for getting generic imagery is industrialized and cheap.
Getting the 'photos of clothing on the model' is a small part of the cost.
Finally there is of course the other aesthetic choices, i.e. what's in the background that have to do with a particular design choice for whatever reason i.e. season choices. That won't be reproduced by AI either.
So it's all very cool, but still mostly novel.
https://www.cadcrowd.com/blog/why-ikea-uses-3d-renders-vs-ph...
Doing it for clothes on humans seems like a reasonable next-step.
- very accurate measurements of your body
- a very accurate representation of the materials used in that piece of clothing, including weight, tensile strength etc.
- a very accurate physics simulation matching the materials to your body measurements
The technology showcased here does nothing similar. It could perhaps generate pictures of what someone who looks sort of similar to your general body shape, as seen in your pictures, looks in clothes. It would be no better than what you currently have: look at the picture, and try to guess how well that fits your actual body.
Even worse, you would completely miss details like "the material is very stretchy, it looks good on the model's abs but it might look bad on my belly fat" because the AI would NOT be doing a simulation of how that clothing actually fits a human body - it's always producing an idealized simulation.
Edit: and to make this clear, I'm not saying that "the AI can't do that yet". I'm saying that the technique they are using can't achieve the goal you are talking about. There is nowhere near enough detail in pictures of clothing to get the kind of simulation you are looking for.
and with enough advances in brain/computer interfaces one could live a whole life without meeting a real human . In fact, in that case, one wouldn't even need accurate models since there would be no baseline to compare to.
Or maybe, we could implement surgical lenses that applies a filter on other people generating a fake person.
You can basically, eliminate _ism(s) in the workplace by making everyone look and sound the same.
People can pay to make their images in others' eyes different. Or one can pay to make their spouse look like a their true love in their eyes.
[1] https://www.youtube.com/watch?v=XOxxPcy5Gr4&t=56s
[2] http://www.arxiv-sanity.com/search?q=generative+adversarial+...
Uncompensated, or grossly under-compensated, use of personal data for AI systems is a serious issue that one no one seems to want to address.
Stock photos exist. It's not as if researchers are trawling facebook to build their AI models.
Just looking at the output, and the unnaturally specific body poses, I'd guess they just put in an order for 50,000 stock photos and went on from there.
If you negotiate, you can trivially get it down to 0.10 per image which is just 5000 dollars. And by trivially, I mean just email several sources saying your budget is in the low thousands.
So most likely, everyone got compensated. All rights were preserved.
More general: I'm not surprised to see some day, one AI bot could automatically register a Twitter account, and have "real" conversations there. One day, i might retweet their tweets, too.
BRB next startup idea.
Well my wife still respects me I really misused her I am having an affair With a random computer
I wonder if an easy hack would be to train a second separate neural network to clean up the kind of artifacting that you get out of the first neural network...
Shhhh, don't give the evil web overlords ideas for new nasty captchas :)
Here's the first 2 Google results for "GAN" meaning "Generative Adversarial Network"
https://skymind.ai/wiki/generative-adversarial-network-gan
https://en.m.wikipedia.org/wiki/Generative_adversarial_netwo...