Nowadays Artificial Intelligence (AI) is either a villain or a saviour, depending on who you are and what's your job or passion: while some are experimenting with the technology in different fields, governments are worried about users' privacy and critics, led by Elon Musk, are calling for a 6-month pause in the training of such systems.
Yesterday, US President Joe Biden met with his Council of Advisors on Science and Technology (PCAST) to discuss the "risks and opportunities" of artificial intelligence.
While the president highlighted that AI could help in addressing disease and climate change, he also stated that it is of vital importance to address potential risks to society, national security and the economy.
"Tech companies have a responsibility, in my view, to make sure their products are safe before making them public," Biden stated.
Asked if AI was dangerous, the President answered, "It remains to be seen. It could be."
Among the industries that are looking at the new possibilities offered by Artificial Intelligence to try and understand if AI is dangerous or helpful, there is fashion.
Will AI help brands saving money and time in product photoshoots and adverts, or will it generate an incredible amount of plagiarized collections (as if plagiarized collections didn't already exist…)?
Besides, will brands hire their own in-house AI departments or turn to specialized AI studios?
At the moment things are still at an experimental stage, but there is definitely an interest in the applications this technology may have in fashion as proved also by events such as the first AI Fashion Week (20-21 April) that will be held in New York at Soho's Spring Studios and will showcase collections from emerging AI designers.
AI models are developing at rapid speed: for what regards AI text to-image apps, the results that can be achieved with the latest version (v 5) of Midjourney are mind-blowing.
I've been experimenting with Midjourney since the end of last year, but I have used it more extensively since January.
In February I started developing ideas for a fashion collection that combined a sporty and casual mood with art inspirations from the Renaissance period.
At the end of February, Midjourney managed to produce coherent styles, but didn't respect the entire content of the prompt.
If you requested a classic fashion show picture with a model walking on a runway and people sitting on both sides, the system at times seemed to be struggling to recreate the set and focused mainly on the model.
Besides, quite often it didn't respect the given ratio and also tended to crop the figure that was usually slim and at times impossibly elongated.
The main distortions occurred around the nose, eyes and lips, so that models ended up having disquieting faces (in one case the face looked as if it had been sucked in by the hood...pure horror); hands at times presented awkward configurations with extra fingers, but that didn't happen too often to be honest, most times the hands weren't indeed a problem.
Hand distortions often occurred in the prints on the garments that featured portraits and paintings; in one instance, the hand of a woman portrayed on a top seemed to hold the hand of the model in an uncanny way.
The models produced were mainly white; when prompted to create an image of a Black model, the system tended to over-impose on a body a head that didn't match with the requested photographic rendition of a model, but that looked like a painting or a digital interpretation of a Black model.
Another mistake Midjourney was prone to, was creating a blend: the painting that was supposed to be a prop in the background became indeed a combination of a static figure and a human form.
So, from the head to the torso the image was flat and static, but from the waist down the image appeared like a pair of legs that came out of the painting and actually walked on the runway.
Style-wise Midjourney respected the prompt, coming up with a series of tracksuits or casual styles with a Renaissance twist about them.
At the beginning of March, prompted to create matching sports bags and sneakers, the system came up with oversized bags and rucksacks with prints that presented in some cases a few distortions.
Yet most times for what regarded the bags, the system respected the prompt that contained references to Renaissance costumes and three-dimensional elements: the latter were interpreted in the rucksacks as three-dimensional motifs applied to the printed portraits on the bags.
Midjourney made more confusion instead when it came to sneakers: the figures printed on them presented some anachronisms as they wore pleated tunics from Roman or Greek times, but also sported oversized sneakers.
The system was not capable of interpreting the words "three-dimensional" in these cases, but produced figures with clothes that came out of the shoes creating awkward reliefs.
A twist in the prompt in the days that followed pushed the system to refocus on costume details: Midjourney excelled at creating tracksuits with ample sleeves or that combined Renaissance costume details with tracksuits.
The adjective "Asian" in front of the word "model" contributed instead to generate a confusion with an intriguing result: Midjourney came up with an Asian model wearing an elegant tracksuit that incorporated a cropped jacket with elements vaguely reminiscent of Chinese warriors or samurai armours with floral Renaissance prints.
While sleeves became better, figures were still elongated, but the system improved slightly when generating Black models, even though close ups of Black models wearing jewelry weren't always convincing.
The images in these cases still looked like polished digital drawings with imperfect jewelry pieces (think chains that seemed to melt into a shirt), compared to portraits of white models that looked incredibly realistic.
In the meantime, the system also started coming up with better images of audiences, but requesting an image of a front row with multiple people still caused confusion.
Providing credible images of crowds proved confusing and Midjourney generated in these cases multiple hands, arms and legs, distortions in clothes, headgear and masks.
Proportions were rather bizarre with more elongated characters, and even those editors and influencers that Midjourney was prompted to generate and that looked perfect betrayed gross imperfections when it came to their hands that often integrated another hand or extra fingers.
Crowds represented a traumatic moment for the system that also generated hybrid humans, with more than two legs and heads that seemed over-imposed on random bodies.
In mid-March, Midjourney v5 brought a complete revolution: the prompt was the same, but the models looked definitely more polished and were absolutely more realistic.
The spaces and architectures of the show venue were also detailed, complete with front rows and there weren't any awkward blends between paintings in the backgrounds and models in the foreground.
The system responded well to the prompt: lights were perfect, models often looked at the camera and didn't show an alienated or absent stare.
There weren't any more elongated figures and Black models were represented in more credible ways (albeit the generic term "model" mainly prompted the system to create young, slim and white models).
The system also became smart enough to actually interpret the prompt more in depth, creating more precise prints and motifs: rather than generating garments with one badly rendered painting, it translated the "Renaissance jacquard motifs" prompt into tapestry-like interventions and variations.
Sneakers also looked more realistic: the same prompt generated more precise sneakers that combined sections of paintings; a variation in the prompt that referred to pearl embellishments generated instead platform sneakers covered in pearls, rhinestones and fabric flowers. The details were so precise that you could even see the glue around the appliqued pearls.
So what's the verdict after this experimental session on Midjourney? Well, the system is currently trained to create very polished images and is perfectly capable of generating pictures of an imaginary fashion collection.
Sure, it didn't look desperately innovative and was derivative, but the level of perfection the new version of Midjourney has reached is impressive.
As AI systems keep on progressing, they will definitely become part of the fashion industry and you can bet designers will start using Artificial Intelligence to create blends from their archives or generate quick ideas for their collections (well, those designers who aren't already doing so...).
Will this generate more "AIgiarisms" or the proliferation of blended collections? Probably, but for the time being it is fascinating and, yes, slightly scary as well, to see how Artificial Intelligence is developing and progressing. Ah, yes, one last note for those who appreciate those early and hilarious distortions: you can still generate them on Midjourney by adding "v 4" rather than "v 5" in the prompt.
Comments