What we learnt when making AI images for the 2023 Power issue
For the 2023 Power issue our team taught artificial intelligence to make portraits. The results show how it is learning about 21st-century culture.
Parents know the importance of setting limits for their children. It turns out the same applies to teaching artificial intelligence how to create realistic images of famous people.
Our annual Power issue is an analysis of who are Australia’s most powerful. It’s also an exercise in portrait photography – our photographers are dispatched to get fresh shots of all the people deemed as Power listers by our panellists. Their output creates memorable images of prime ministers, corporate bosses and sports stars that are often republished and reused.
What’s with the hands? AI’s digit-al skills, as seen on this rendering of Margot Robbie, remain a work in progress.
For this year’s issue, it seemed apt to turn to artificial intelligence to generate the portrait images for us. We still sent our photographers out into the field to collect fresh photos. But we simultaneously set to work at our keyboards to create images through AI. We’ve put them in suitably outlandish scenarios, as is the style of much AI artwork. The AI output on these pages is a collaboration between our creative team, headed by Tim Beor, and Rodolfo Ocampo, a PhD candidate in AI at UNSW.
Anyone can have a stab at generating AI images of famous people through platforms such as Midjourney, but they’ll find that AI struggles to accurately replicate them. Because AI has been fed the contents of the entire internet, the more a person’s photo appears online, the better the AI version. It’s quite easy to create a fake image of Donald Trump – and from the reams of images of Trump being arrested, Trump dancing on the New York subway and Trump inventing penicillin, many already have.
Images of Robbie from the internet were used to create a profile ... then the AI file of Robbie’s face was created in Stable Diffusion.
To generate a likeness of someone with comparatively fewer images on the web – such as the Australian prime minister – takes more skill. Beor supplied Ocampo with a collection of dozens of photos of the PM as well as the other people who we’ve rendered in this issue. Ocampo then trained an AI called Stable Diffusion to generate an avatar of each person. The AI learnt the features of our subjects and got better through trial and error.
The fact that you can train Stable Diffusion – by limiting or controlling the inputs it draws upon – makes it better at rendering images of famous people. In effect, we trained it to become a specialist in how our Power listers look. Midjourney is more of a generalist; you can’t control its inputs, but it’s better at creating beautiful settings and scenes.
That is why Beor used Midjourney to create the various environments in which we wanted our Power listers to appear – negotiating with aliens, practising kung fu a la Bruce Lee. The final step involved Ocampo feeding that environment into Stable Diffusion as a reference image, and asking it to include the various avatars he’d already built. Sometimes real photos were used for the environment, such as the photo of Scott Morrison below.
You can see the results in these pages. They are both uncanny and yet slightly unreal. All have the distinctive fuzzy texture of AI images, as if they were drawn. Our prompts were very minimal and the output hints at the way AI is learning 21st-century human culture.
For the image of Margot Robbie depicted as the prime minister, the AI chose to dress her – unprompted by us – in a white suit. Is AI drawing upon the 2020s trend of women in power wearing white as a homage to the suffragettes? Look closely and you’ll notice odd numbers of fingers and weirdly deformed hands.
An AI-generated image of Margot Robbie at the prime minister’s desk that used a real photo of Scott Morrison as a reference. Dominic Lorrimer
There are ethical considerations in creating images like these, which, stripped of context, could be used as deepfakes. Such fakes are set to add a new dimension to misinformation campaigns, and we’re keen not to add to that problem. Our input images were all selected from what’s publicly available on the web; the fresh portraits that our photographers took for this issue were kept separate.
This is because we wanted to see how far we could push the AI to create magazine-quality portrait photos, without giving it a helping hand from our own shots. How quickly can you spot that they’re fakes?
Subscribe to gift this article
Gift 5 articles to anyone you choose each month when you subscribe.
Subscribe nowAlready a subscriber?
Introducing your Newsfeed
Follow the topics, people and companies that matter to you.
Find out moreRead More
Latest In Federal
Fetching latest articles