Artificial intelligence continued to make great strides in a variety of fields in 2022, but perhaps one of the biggest shocks was the emergence of AI models that can generate photorealistic images from a simple text description.
鈥淚t was totally unexpected, I would say, at the end of 2021. Unexpectedly mind-blowing,鈥� says Thomas Wolf, co-founder of , a website that allows people to share AI code and data sets.
Before 2022, these text-to-image AIs were a fairly immature technology producing crude results. But this year things have evolved rapidly, to the point where .
Advertisement
at the University of Birmingham, UK, says the technology saw rapid progress from a collision of hardware and software improvements. Firstly, researchers started applying to image generation a type of algorithm called a transformer, invented by Google engineers in 2017. Originally this type of AI model, which examines a string of data and predicts the next likely piece, was used for text-generation models like GPT-3.
Secondly, hardware got really powerful, with cumulative advances in the ability to turn huge numbers of graphics cards into a unified and efficient supercomputer ideal for tasks like training AI models, making them the obvious choice in terms of the ratio of power to cost.
But perhaps most importantly, says Lee, large companies with the money and resources to train these models began to give away part of their findings and even offered limited access to the wider public, sparking a flurry of interest and fresh research.
鈥淵ou鈥檇 expect these big companies to just do all this work and then just keep it to themselves and make money,鈥� says Lee. 鈥淏ut releasing it to the wider community is a kind of very long-term view, because if you do that then more scientists start working in this area.鈥�
Wolf says that although transformer models made initial gains in image generation, such as the original DALL-E, a new type of algorithm called diffusion has been leading the way in recent months.
鈥淭ransformers work, but they tend to give some weird artefacts,鈥� he says. 鈥淒iffusion models are very different from transformers and they鈥檙e able to create this type of very fine-grain texture. Which is, I think, what鈥檚 made these new models stand out in terms of 鈥榳ow鈥� effect.鈥�
These AIs are already proving disruptive. Adrian Alexander Medina, editor of literary website and magazine Aphotic Realm and a creator of book covers, says that he is already losing commissions to AI, with clients opting to generate images for free rather than pay human creators.
鈥淚 was in discussions鈥� only for them to go another direction and purchase, or create their own, AI-generated cover. Their money, their prerogative. Disheartening and irritating for sure though,鈥� he says.
Photo licensing firm Shutterstock and research organisation Open AI have even signed a deal where customers can buy access to the latest model and generate images on request 鈥� something which Medina likens to 鈥渟ewer water leaking into the drinking supply鈥�.
What makes that loss of work all the more galling is that the AI models are necessarily trained on vast data sets that include millions of images scraped from the internet. A site called Have I Been Trained has been launched to allow people to search these data sets for evidence that their work has been pulled into these services. A search for 鈥�鈥� reveals that hundreds of previous covers have been incorporated, for example.
Ultimately the dramatic reduction in cost and time required to create custom artwork with AI may end up having a dramatic effect on human creators. 鈥淎 person can generate dozens of pieces in a couple of hours and flip them to authors who either don鈥檛 know better or don鈥檛 care,鈥� says Medina.
