I did something similar in a big fashion company in the past
They had a non-utilized collection of tens of thousands of tagged clothing photos going back to the 80s. I Used DCGAN to generate images of new clothing, then model arithmetic to request specific ("sparkly winter dress") pieces. It was pretty amazing to see the strength with existing photos
The higher ups stared, blinked a few times, and just muttered they don't get it
I have no doubt this is in the future of fashion design in the next 20 years. I just hope those images don't get lost when the company inevitably shuts down
Fashion is so diverse that you can hardly ever create something that looks distinctively new. Me with these shoes. They either look weird or regular, and the regular probably exist in a very similar style.
I work with a shoemaker and I've commissioned a few bespoke sneakers (trainers as we call them here) in the past. He comes up with new designs from scratch, and I made suggestions and ask for changes based on my taste.
Usually I browse online shops for inspiration, but I'm curious about whether this service can produce original designs sensible enough to, at least, be the base of a new shoe.
Yeah the process is being used to generate sneakers well before this guy even started scraping the net. We had our first prototype in production last winter and just finished our second two. http://aire-gan.com
Yeah, it looks like the server was not able to handle the visitors. I'm completely new to webhosting, but I will try to get everything back online soon.
I get that the title is in the theme of this-x-does-not-exist but I don't think it's true in this case. Many of these look like existing models.
The amount of variation in sneakers is limited so when you have 50k training images you end up copying some of them with little to no change because if you deviate too much you end with something that doesn't pass as a sneaker at all.
In most cases, GAN produces a random mixture of features extracted from real photos. That's also why you can tempt GitHub Copilot to regurgitate Carmack's swearing comment.
Sufficiently advanced random mixture of features extracted from real photos is indistinguishable from genuine understanding.
That's why I don't understand this argument. It's not true - you can make the AI generate specific samples, but most samples can not be found in the training data - but even if it was true, it wouldn't prove anything anyways, because you can always say "well the features it copies are just smaller" until the features are just combinations of pixels. It's unfalsifiable even for human artists.
It depends on your definition of "useful." If you look at a single GAN-generated image in isolation it's almost certainly "useless" for most definitions. But if you need a bunch of faces, sneakers, buildings, whatever, and you want to be cognizant of people's privacy, and you don't mind the occasional weird image artifact or completely fake-looking image, these can be good alternatives to sourcing a large number of real images.
So, for example if i was creating art or designing something, say i could use one of these generated sneaker images instead of a stock photo (could be useful, right).
Now the question is, since the training material were images of copyrighted designs/shoes could this get me in trouble?
I feel like there are many cool uses for generated stuff such as generated music or generated code also, but ultimately the uses are held back by existing laws and regulations around copyright and so on. So it really doesn't solve anything that also wouldn't be solved by just getting rid of copyrights entirely.. (sorry, just thinking "out loud")
"Just" getting rid of copyrights entirely isn't going to happen so it's kind of pointless to speculate on that front...
I don't think it's a foregone conclusion that generated images are subject to copyright even if the model's source images are copyrighted. You can make an argument that if all the source materials are subject to the same copyright, the generated images could be. But even in that case the copyright applies to the source. It doesn't apply to the next step (at least not explicitly). It depends on the legal definition of derivative work, if these images even qualify as a derivative work, etc.
I found it frustrating, since most texts make no sense. You can force yourself to imagine a situation where they make sense, but then it becomes tiring.
I guess it's just lightly trained with the captions users have submitted for each image. It's a random meme generator that didn't even need AI if this was the intended end product.
Maybe it being crappy is actually the joke and I'm missing it.
A few years ago, I interviewed at StockX, and the team was working on exactly this – using GANs to generate pictures of shoes. Interesting to see essentially the same thing here a few years later.
Wow, I think this would be pretty useful as a shopping guide in an online shop: Let the user pick a few models they like and render more items based on the chosen ones. Present actual sneakers you have in stock at the last step.
I think the experience is more like an employee coming to you and asking you what you're looking for in a shoe, instead of entering an empty store and browsing all available shoes. I would certainly prefer some guidance, as I usually have no clue what I'm looking for.
Yeah, I love it when stores lure me in with products that I want, and then only have products that I don't want. I never rage-quit the store when that happens; I only ever buy dregs that they have in stock. I would love spending hours at a store browsing imaginary products that I can't buy; engagement is a perfect metric to predict profitability and the world needs a lot more than that.
All that said... if you change the contract a bit, it could work. Kinda like groupon, but people can browse AI-generated shoes and the most popular ones enter a tournament, and the winning design gets made for purchase. That could be pretty cool.
Too many of those shoes look like they have a munged Nike logo which means that any company trying to ship them would soon be receiving a kindly call from the Nike megacorp lawyers. You would have to improve the training to avoid generating trademark / logo infringing designs. Could this be trivially done?
One way this can be accomplished is by removing all shoes with such logos from the training data. But Nike and Adidas are disproportionally large parts of my training set, so this would not be feasible.
The other option would be to train a machine learning model to recognize said logo's and to use this model to remove sneakers with logo's from my generated images. This could however greatly reduce the variety of images on the website.
EDIT: To clarify - don't remove the entire image of a sneaker, just impute away the logo filling in the space with the 'context fill' algorithm found in photo editing tools
I don't think that is an issue: The process should just help the user find the right shoe. If the user repeatedly picks shoes with a logo, he will be presented shoes of that brand, that the vendor has in stock. This is actually beneficial to the trademark holder.
Thank you for all your support, guys! I was not expecting this much traffic. I am currently in the process of moving my site to a better server.
While this is in progress, only the images cached by Cloudflare can be loaded. This means that the editor for most shoes will not work. I apologize for the inconvenience!
Site is dead as of now, but looking at the renders [0] they all look like viable designs, and pretty close to actual sneakers.
I don't know if it's a testament to makers' creativity, as tbh there are way wilder sneaker designs released as actual products by Nike or Puma, or the limits set on the generating algorithms to stay within mainstream designs.
I have seen your project. However, I think claiming "first" is disingenuous. There are many projects like ours that date back at least 3 years. We just happen to undertake these projects at a time when algorithms are better and results are palatable.
This is actually very prescient. In many ways it is much easier to replace artists than labor, contrary to the current narrative.
A simple example in visual effects are "artist guided tools". Instead of having a team of artists place and guide every virtual hair, you have a single artist draw a curve for the hair to follow, and the computer figures out the rest. There are dozens of examples like this in the world right now.
They are "just composites" in the sense that the NN learns to decompose an image into multidimensional concepts both local and global (the latent space) and then assemble new images from anywhere in that space even at points that weren't in the training set. Of course if the training set is small then it'll overfit to mostly reproduce the inputs.
it sounds like jargon to me as idk much about neural networks. i guess its kind of a highly complex and nuanced composition. but essentially its still just a sophisticated combination of existing things.
i guess this is the same as human creativity in the sense it depends on prior experience. its more the "does not exist" that bugs me a bit.
tx for the detailed response tho, wish i was smart enough to understand.
> i guess this is the same as human creativity in the sense it depends on prior experience.
That's roughly the idea. Although I didn't want to make that analogy because someone would try to overextend it and point out that an artist can do X and a neural network can't and therefore it's wrong.
> its more the "does not exist" that bugs me a bit.
Well, if someone learns what a sneaker generally is and then draws a fantasy-sneaker that is still recognizable as a sneaker because it stays in the general confines what is considered "a sneaker" then does that sneaker suddenly exist?
You’re somewhat right. The machine learning algorithm tries to replicate the general patterns on the images it’s trained on (so in this case, images of a lot of sneakers). If the algorithm is trained on enough images, it should, to an extent, learn to generalize and “understand” what a sneaker should look like and generate new ones rather than copying images. In my case, there is definitely some memorization going on, based on some shoes looking suspiciously similar to existing shoes, but there is also definitely some design going on not copied from other sneakers.
But it builds the images “pixel by pixel” rather than e.g. taking the sole of one shoe with the upper of another if that’s what you’re asking.
As a bit of a sneaker head, it's been quite fun to see how the algorithm changes the shapes when playing with the sliders on the Sneaker Editor.
Impressive that it isn't copying/pasting parts of shoes but actually building them pixel-by-pixel, I was editing a very AF1-looking sneaker and playing with the sliders made actual sensible changes to the shape and design.
Pretty entertaining :)
Quick edit: as a side question, the training dataset included all kinds of sneakers or was it biased towards more popular ones? I ask because from what I scrolled I missed seeing more avant-garde designs and wasn't sure if the model was trained on these or not.
The training set included pretty much all sneakers I could find online. However, the algorithm tends to bias towards patterns that are more common e.g. hundreds of colorways of Air Jordans.
BUT I can fine tune the model by training on a small subset of sneakers. So if you have any types of brands or sneakers that you would like to see generated, please feel free to mail me some examples at stan@thissneakerdoesnotexist.com and I will see what I can do.
It could also be interesting to define multiple subsets of sneakers - "2021 avant garde sneakers", "skate-inspired shoes", "2010s most popular shoes in the NBA" (based on https://ballershoesdb.com/ or a similar database). It could be cool to start with a model trained on a subset of sneakers, and tweak/combine things from there.
They had a non-utilized collection of tens of thousands of tagged clothing photos going back to the 80s. I Used DCGAN to generate images of new clothing, then model arithmetic to request specific ("sparkly winter dress") pieces. It was pretty amazing to see the strength with existing photos
The higher ups stared, blinked a few times, and just muttered they don't get it
I have no doubt this is in the future of fashion design in the next 20 years. I just hope those images don't get lost when the company inevitably shuts down