2022 is going to be a massive year for the fashiontech industry. From brands adapting fashion to a brave new world of non-fungible tokens (NFTs) and blockchain technology, to embracing the metaverse and capitalising on fashion in the gaming space. This era also holds the possibility to push the door open even wider for a particular kind of artificial intelligence (AI) – Generative Adversarial Networks (GANs).
A good way to understand GANs is to consider where they fit in the AI ecosystem. Within AI there is machine learning and within machine learning there is deep learning. One of the most straightforward things that deep learning can do is image classification. For example, given enough labelled images of shirts and skirts, we can train a deep neural network (DNN) classifier to tell the difference between them. GANs are a subfield within deep learning and consist of one such DNN classifier that aims to predict whether images are real or fake. This DNN classifier has a nemesis – which we can think of as a ‘DNN faker’ – hence ‘adversarial’ in the name.
The DNN faker is a neural network that generates fake images. Every time the DNN faker produces an image and obtains judgement of the veracity of the image, it improves. Similarly, the DNN classifier is fed real images alongside fake images created by the DNN faker. Both networks get better with training over many cycles and eventually it can be difficult for us humans to discern between real and fake images.
AI and generative art
Due to high costs, time constraints and limited resources, luxury fashion brands inevitably have a keen interest in creating new pieces using AI. Enter Robbie Barrat. Barrat is a self-taught artist and coder who has been exploring the intersection of GANs and fashion, notably creating a new Balenciaga collection using a set of images from their previous collections.
After seeing his work, Swedish luxury fashion house Acne collaborated with Barratt to create a collection for Autumn/Winter 2020, a collection that would not just exist digitally. Barrat trained a neural network on the past four seasons of Acne’s collections to generate new designs and craft new imagery, which was then printed onto the label’s clothing. In addition to this, he created a tool for the Acne designers that allowed them to select an area of a particular piece of clothing or outfit and modify it.
The collection is arguably one of Acne’s most interesting, as the clothes include reimagining of the traditional waist, collar and hemline. Barrat expressed that although satisfied with the outcome, he would have liked to see the collection be more eccentric and nonsensical. This also highlights an advantage for the digital world and high fashion: the creation of the pieces from Barrat’s Balenciaga collection would be impossible to produce in the physical world due to the textural component that appears in the images. However, in the virtual world of the metaverse or in gaming, these issues fall away.
Interestingly, Barrat’s interest lies in what happens not when an outfit or collection is perfectly created, but rather when the GANs misinterpreted the information. For example, the AI would recognise that certain images contained a handbag on the arm or over the shoulder, but would generate an image where the bag was attached to the model’s shin or foot. Herein lies a critical factor that Barrat highlights throughout his work: that AI is simply a tool. It ought not be important that the collection was designed using AI, what is important is the artist and their creative process.
Virtual celebrities and influencers: navigating a new, distorted reality
In the new and frustrating Covid era, frequent and unnecessary travel has been cut to a minimum. This is, in part, due to the ever-changing restrictions per country, along with testing and vaccination requirements (as Novak Djokavic has so spectacularlydenounced of late).
Brands will inevitably still want to work with the best talent in the fashion industry. Just as Naomi Campbell and Kendall Jenner were made into digital versions of themselves for Burberry’s TB Monogram 2021 campaign, such appearances could become more prominent as time goes on. It has been suggested that models, actresses and athletes could be digitally scanned and copied into a CGI version of themselves, allowing brands to use their likeness for future campaigns. This would inevitably leave travel to a specific destination for the production unnecessary and would mean a star could lend their faces and bodies to more than one brand at a time. Responsible for some of the best deepfakes to date – GANs are a strong candidate for improving CGI.
Similarly, consider influencers such as Lil Miquela and Lu do Magalu. They have followings of 5 million and close to 3 million apiece. Miquela has worked with brands such as Prada, Dior and Calvin Klein. Noonoouri, a wide-eyed, dark-haired teenager (with her 369,000 followers) has collaborated with the likes of Marc Jacobs and Versace. There’s something else interesting about all of these influencers – they are virtual humans. Social entertainment firm Fullscreen surveyed 500 people aged 13 to 34 in 2020 and 42% of them didn’t know an influencer they were following was virtual. With FaceTune and the multitude of filters available on each social media platform nowadays, this figure is not as shocking as it might have been 10 years ago. Even more disturbing is how indistinguishable deepfakes have become from actual human beings.
Here are examples of some of the best deepfakes to date:
All four individuals pictured here are computer generated using StyleGAN2, software created by Nvidia research labs. What is perhaps most concerning about this is that highly influential figures can be deepfaked, along with realistic sounding voices. Given the expanding reach of the metaverse, it is not difficult to imagine that we will soon be at virtual parties with digital people. In that setting, distinguishing who exists in the physical world and who does not will be challenging. A research group from the University of Buffalo suggested that one way to tell the difference between real and fake, would be to focus on the person’s eyes. Physics dictates that reflections of light in both eyes tend to be similar. GANs, good as they are at discerning patterns, cannot appreciate such idiosyncrasies of our reality. With this technique, the researchers were able to discern 94% of fake images. Whether this becomes a silver bullet against deepfakes remains to be seen, however. If deepfakers keep finding ways to improve, our last hope may be old fashioned technologies – trust and common sense.
The new efficient norm: virtual photoshoots and personalised fashion
Recently, Japanese tech company DataGrid used GANs to create virtual fashion models, complete with authentic human movements and numerous outfit changes.
The cost of producing such a catalogue would usually include models’ time, photographer’s time, hair, makeup, styling and post-production. GANs technology would save on all of these fronts. A virtual set is one with far fewer delays, meltdowns and madness. With a few lines of code, a tech-savvy employee is able to alter poses, and – more drastically – change the garment as well as the model’s appearance.
Now to a pivotal player: the consumer. According to experts at UNCTAD (United Nations Conference on Trade and Development), the e-commerce sector saw a rise in retail sales from 16% to 19% in 2020, in part as a result of the Covid pandemic. This equates to approximately $26 trillion worth of products being sold online. Clearly, online images have never been more important than now. Noting too that images of clothing alone do not sell as well as when displayed in a human body, it follows that personalised shopping experiences should inevitably sell more products.
According to research done at Cambridge University, when Dove launched an advertising campaign with women of all skin tones and body types in the United States, sales increased by 600% in two months. Age and ethnicity play a major role in influencing consumers to buy products – the study concluded that when shoppers saw models that were close to them in age, their purchase intent went up by more than 175%. Given this, the opportunity for the fashion industry has never been more perfect than now: with the help of GANs, shoppers could see themselves in adverts online and be able to try on various outfits after giving their dimensions.
Tip of the iceberg, start of a revolution
The fashion business is bound to undergo significant change with increased synthesised content, and one can hardly speak of such reality distorting technology without considering its broader societal ramifications. AI, and specifically GANs, are neutral tools and therein lies their beauty and curse: they could be used in the creative process, driving down costs and saving time, however, they also could be used for more nefarious purposes. Given the breadth of their potential applications, I believe that we are only scratching the surface and we will in future look back and see what was the genesis of a revolution.