The rapid rise of artificially intelligent text-to-image generators has captured the imagination. But the stormy development also leads to concerns about the future of art and artists.
Artificial intelligence (AI) will make great strides in a variety of areas by 2022. One of the biggest shocks has been the rise of AI models like DALL E 2which can convert a simple textual description into realistic images.
‘At the end of 2021 we had not expected this at all, I would say. It is mind blowing‘ says Thomas Wolf, co-founder of hugging facea website where people can share AI codes and datasets.
READ ALSO
Dossier: Know what, how, why and for what you eat During the holidays we often eat and drink a little more than necessary or sensible. Immediately afterwards follows the period in which we make our good intentions …
Before 2022, text-to-image AIs were a fairly nascent technology that delivered only rough results. This year, development has moved so fast that one of the winning entries at the Colorado State Fair art contest even created by an AI.
Transformers
According to AI researcher Mark Lee from the University of Birmingham in the UK, the technology developed so quickly thanks to a convergence of hardware and software improvements. First, researchers started so-called transformers apply to image generation. That’s a type of algorithm invented by Google engineers in 2017. Transformers examine a series of data and based on that, predict the next part of the series. Originally they were used for text generation models, or chatbots, such as GPT-3.
Second, the hardware became much more powerful in 2022. This allowed huge numbers of graphics cards to be bundled together to create efficient supercomputers. They are ideal for training AI models.
Read also: Artificially intelligent chatbot ChatGPT is tricking you
But according to Lee, the main reason for the progress is that big companies, with the money and resources to train these models, started giving away some of their findings. They even offered limited access to the general public, generating an outpouring of interest and new research.
“You would expect these big companies to keep all this work for themselves just to make money,” says Lee. “Releasing it to the wider community is part of a kind of long-term vision, because when you do that, more scientists get to work in this area.”
Wow effect
According to Wolf, transformer models indeed initially made gains in image generation. But in recent months, he argues, a new type of algorithm, called diffusion.
“Transformers work, but they tend to give rather contrived results,” he says. ‘Diffusion models are very different from transformers. They can create very fine-grained images. That, I think, is what sets these new models apart in terms of the ‘wow effect.’
Sewage
These AIs are already proving to be disruptive. Adrian Alexander Medina, editor of literary website and magazine Aphotic Realm and book cover designer, says he’s already losing business to AI. Customers would rather generate images for free than pay human designers.
‘I was talking to people a few times, after which they decided to go in a different direction and buy an AI-generated cover or make it themselves. It’s their money, their right. But it sure is discouraging and irritating,” he says.
In October, photo licensing company Shutterstock and research organization Open AI even closed a deal. Shutterstock customers can now purchase access to the latest AI model and have it generate images on demand. Medina compares this to ‘sewage leaking into the drinking water supply’.
Data sets
What makes the loss of assignments rather bitter is that the AI models are trained on huge datasets that include millions of images scraped from the internet – images that were produced by human creators. On the website Have I Been Trained people can search these datasets for evidence that their work has been used to feed the AI.
Ultimately, this AI breakthrough could have a dramatic effect on human creators. AI costs less time and less money, and still delivers customization. “One person can generate dozens of images in a matter of hours and push them to customers who don’t know any better, or who don’t care,” says Medina.