Trends play a pivotal role in shaping consumer behavior and influencing market demand in various industries: for example, in the realm of fashion, a particular hue or style may become the talk of the town, while in the world of art, music, or cinema, a particular genre becomes the new wave.
The allure of keeping up with the latest trends is undeniable: it signals that you are in the know and ahead of the curve, with a finger firmly on the pulse of the zeitgeist. However, while following trends can be advantageous, it can also have its drawbacks. By conforming to the prevailing trends, one risks losing their individuality and creativity, becoming just another follower in a sea of sameness.
This tendency towards conformity is a trap that many fall into, unwittingly surrendering their uniqueness and originality. Let us consider a simplistic yet illuminating instance: a ubiquitous phenomenon in the contemporary cafe culture is the widespread usage of blackboards to display menus.
This trend transcends geographical boundaries and appears to have established itself as the norm in various cities across the globe - be it the bustling streets of New York, the romantic alleys of Paris, the vibrant urbanity of London, the fashionable avenues of Milan, or the sunny beaches of Sydney. The prevalence of this trend is not limited to corporate franchises but also encompasses independent cafes and bars. The reason behind this phenomenon is rather simple: we stumble upon an endless stream of images on social media depicting the same trend, we internalize the visual narrative and we conform to what eventually becomes an established norm.
But this pervasive inclination to adhere to a distinct standard or aesthetic seems to have permeated the realm of Artificial Intelligence as well.
It is indeed not rare to spot something cool generated on one of the #general servers on Discord by the Midjourney Bot or an image created using another AI text-to-image generator that, posted on Reddit, quickly catches up, leading to the spawning of similar visuals, thereby creating a trend.
Let's make some examples: it is not so rare to spot on Instagram AI generated images of popular film or cartoon characters turned into white and blue Ming Dynasty porcelain pieces.
Darth Vader, Bart Simpson and Hello Kitty prevail on @anchorball's Instagram page, a very popular pose in which they are portrayed is with their legs crossed, like Buddha. A version of a Ming Dynasty Mandalorian in the same pose can also be spotted on @str4ngething's Instagram page.
The idea is not completely new: months ago, Paul Parsons created porcelain versions of pop characters' masks, from superheroes to villains of the Star Wars saga such as Darth Vader and Imperial stormtroopers.
But the origins of the blue porcelain Star Wars character craze should be traced back to a few years ago when the trend became so popular at pop culture events and fairs in Asia that, in 2016, Hot Toys decided to release a Star Wars Stormtrooper figure with a blue porcelain pattern. So, in a way, this is a real life trend that has been applied to the digital world and that Artificial Intelligence has learnt to replicate pretty quickly.
But let's go back to exploring similarities in AI generated images and consider further examples.
Atelier Blancmange's recent "Emboss/Deboss" collection, designed for the first Artificial Intelligence Fashion Week, is a mix of Classical, Renaissance, Mannerist, Baroque and Neoclassical sculptures, filtered through costume design, cinema and music inspirations.
The collection looks unusual, but @str4ngething (who like @blancmange_atelier must be using Midjourney) seems to have obtained almost the same effects not just for what regards AI generated garments, but also accessories.
Obviously, there are differences in the shapes of the bags and sunglasses, in the effects obtained on the garments and in the palettes they chose for their works - a colourful one for Atelier Blancmange and a neutral one for @Str4ngething. Yet the main inspiration - three-dimensional sculptures - remains the same.
Apart from that, another user, @rickdick__ produced men's corsets that have got something sculptural about them, even though in this case the Ming Dynasty reference eventually led the system to the creation of something different.
Elsewhere one candy / jelly inspired gown posted a while back on Atelier Blancmange's Instagram page seems echoed in a collection by Pierce Gibbon, another participant in Artificial Intelligence Fashion Week.
Actually, if you go on Midjourney and include in your prompt words such as "candy", "jelly", "edible", "plastic" and "melted" or "melting", associated with "dress", you will generate gowns with different colours and configurations, that are in essence reminiscent of this collection.
The same thing often happens with rainbow coloured gowns: Midjourney seems indeed to tend to generate gowns with ample cape-like cloaks or panels that open up into rainbows splashing on the runway (this post features a collection between a design by @anchorball and one from Maritza Rubio, another AIFW contender).
So, why is this happening? How come there are such connections and in some cases almost the same image is generated? Sure, it is possible that these users were influenced by similar cultural or social trends or by each other, besides there are creators who spend time on the Midjourney #general servers on Discord or on Reddit, bouncing ideas off each other.
In the case of Atelier Blancmange, though, the collection was unveiled on 20th April for AIFW 23, the images were sent to the organisers at the beginning of April and weren't posted online before that, still there is an evident connection with the images generated by @str4ngething in mid-April. So, what caused the similarities?
Well, there is actually a tech explanation: the AI text-to-image system the two creators used learnt to generate specific images based on certain prompts or keywords. This is known as "mode collapse," where the model tends to generate similar or identical images for different prompts due to overfitting to a certain set of images or concepts.
Mode collapse is a common problem in Generative Adversarial Networks (GANs), which are a type of deep learning architecture used for image and video generation.
Several text-to-image systems use GANs as their core component: Midjourney employs StyleGAN2, which is an advanced version of the original StyleGAN model, while Stable Diffusion uses BigGAN, a large-scale GAN architecture that produces high-quality images.
Other popular text-to-image systems that use GANs include DALL-E from OpenAI and AttnGAN from Microsoft Research Asia.
In a GAN, there are two parts: a generator network and a discriminator network. The generator network tries to create new images, while the discriminator network tries to determine whether an image is real (from a training dataset) or fake (generated by the generator).
Mode collapse occurs when the generator network starts producing a limited range of images, even though the input data should allow for a much wider range of outputs. This happens when the generator network becomes too good at fooling the discriminator into thinking its generated images are real, but it does so by only generating a limited subset of images that are similar to each other. In other words, the generator network "collapses" many possible outputs into a few "modes" or clusters of similar images.
One reason for mode collapse is that the generator network has not learnt enough variation in the input data, and so it produces outputs that are similar to each other. Another reason is that the discriminator network has become too strong, making it difficult for the generator network to create new and diverse images.
Researchers have developed various techniques to mitigate mode collapse, such as modifying the GAN architecture, changing the loss functions, adding noise to the input data, and using regularization techniques. Despite these efforts, mode collapse remains a common problem in GANs and a topic of ongoing research in the field of deep learning.
So, the similarities between images you may see on Instagram feeds, may not be a question of plagiarism or of stealing ideas from each other, but of the system learning to react to a prompt in certain ways and in accordance with the input data it was trained on.
The final danger? Exactly what we were talking about at the beginning of this post - generating not just a trend, but a sea of sameness.
Does this mean that we should stop creating on these systems? After all, it must be frustrating for a creator to think they have come up with something unique only to discover that somebody else did more or less the same thing.
Maybe one way to avoid all this, rather than just creating something and posting it immediately on Instagram to prove you were there before everybody else, would be to generate images on AI text-to-image systems based on your own exclusive images fed to the system. That said, the system would then start learning from them and it would still be able to reproduce something similar.
In future things may change, but GAN architecture at the moment constitutes a problem for those fashion designers and houses interested in using Artificial Intelligence in their collections on a merely visual level: they may indeed generate an image of the most beautiful gown today thinking of producing the design IRL in six weeks' time at a fashion show. However, tomorrow, an anonymous user may use the same words in a prompt and generate a design that is more or less the same.
This suggests that we have to sit down and think about how to use Artificial Intelligence in fashion in more original ways.
Sure, AI can be employed for personalized styling, supply chain optimization, waste reduction and virtual try-ons, but, when it comes to product design, rather than using Artificial Intelligence to create entire fashion collections, we could maybe employ it to generate ideas that could then be developed into innovative solutions and cutting-edge applications and integrated into fashion collections, from fastening systems to textiles elaborations for surface treatments and texturing.
In a nutshell, the ideal remains to merge human creativity with AI optimization and reinterpret the outputs generated by AI through the insights gained from human knowledge and experimentation.
Comments