Who is Hidreley Diao and why should you care?
“The medium is the message.”
Marshall McLuhan
Hidreley Diao is the coolest artist you’ve never heard of.
He uses artificial intelligence to create photorealistic images of historical figures whose lives predated the invention of photography. He’s also brought beloved animated characters to life, like Moana and Bart Simpson, to show what they would look like as real people.
Source: Hidreley Diao
Hidreley’s work has been featured on Yahoo, Buzzfeed, and Upworthy, and he’s always posting new stuff on his Instagram. I first came across it via a tweet from mega-blogger Tim Urban.
I find Hidreley’s work to be incredibly cool. As a history nerd, I can’t get enough.
As I ogled an appropriately handsome rendering of the man who inspired the statue of David, I found myself deeply curious about the technology behind it.
How exactly did it work? If AI can create photographs, can it paint a landscape? Write a novel? How will this change the way we think about art, and on a deeper level, what it means to be human?
This might sound dramatic, but art – going back to the time of cave paintings – has always been a defining quality of our species. What if we’re not the only ones who can do it anymore? What if the machines can do it better?
These questions led me down the rabbit hole of artificial intelligence and computer-generated art.
In this post, I’ll explain how Hidreley and other digital artists are using AI to push the boundaries of art and explore the implications.
How does he do it?
Hidreley’s work is made possible by an artificial intelligence model called GAN, which stands for generative adversarial network. If that sounds like a mouthful, it’s because it is. Don’t worry I’m going to break it down.
However, to understand GANs, you first need to understand deep learning.
Deep Learning
Though you may not know it, deep learning plays an important role in your life. It powers the tech behind spam filters, facial recognition systems, the voice commands you give your phone, and which ads and news articles you’re served on social media platforms.
Deep learning is a subfield of artificial intelligence. It was first conceived in the 1960s when scientists, inspired by the vast networks of neurons that send information through the human brain, proposed creating artificial “neural networks” that would work much the same way for computers.
This idea remained purely theoretical for nearly 50 years. In 2016, DeepMind, owned by Google’s parent company Alphabet, achieved a major breakthrough when its deep-learning powered software AlphaGo beat a human player at Go, one of Asia’s oldest and most intellectual board games. This represented a major leap forward for the technology.
So, how does it work?
Many people assume that artificial intelligence is programmed, but it’s more accurate to say that it is trained by humans.
To apply deep learning, engineers create layers of artificial neural networks. There is an input layer and an output layer.
Data is fed into the input layer of the network, and a result emerges from the output layer. There may be thousands of additional layers between these two layers. This is where the “deep” in deep learning comes from.
Say we want to train AI to recognize an elephant.
We’ll first give the AI millions of pictures. Some will be of elephants, and some won’t. We’ll correctly label each photo as “elephant” or “not elephant”
Then, we’ll provide the AI with an objective: correctly identify pictures of elephants. Or, to be more specific, maximize the probability of correctly recognizing “elephant” vs. “not elephant”
It’s up to the AI to independently analyze the the labeled photos to determine what features distinguish “elephant” from “not an elephant”.
Gradually, AI will build a highly precise mathematical model that distinguishes elephants from non-elephants and apply that model to novel photos.
Generative Adversarial Networks
Ok, you’ve got deep learning and neural networks down! Let’s move on to GANs or generative adversarial networks. This is the technology that Hidreley uses to create his photos.
A GAN consists of a pair of neural networks with opposing objectives. They are going to be pitted against one another. This is where the “adversarial” comes from.
The first network is known as the forger. It tries to generate something that looks real.
The other network is called the detective. Its job is to determine if the forger’s creation is real or fake. It compares the forger’s creation with actual photos and determines if it passes the “test” or not.
Based on the detective’s feedback, the forger re-trains itself. Its objective is to dupe the detective, so it will use “learnings” from its first attempt to improve its second effort.
These two processes repeat, sometimes millions of times. With each loop, the forger’s creation gets closer to the “real” thing.
The generator’s images become progressively “better” as the networks cycle through millions of loops 
The generator’s images become progressively “better” as the networks cycle through millions of loops 
The generator’s images become progressively “better” as the networks cycle through millions of loops 
The generator’s images become progressively “better” as the networks cycle through millions of loops 
The generator’s images become progressively “better” as the networks cycle through millions of loops
Eventually, the forger wins, the detective is fooled, and equilibrium is reached. At this point the process is complete.
Say you want to create a real-life version of Aladdin.
The forger will be directed to create a fake image using a database of human faces. The detective will be given a “real” picture of Aladdin, which is up to the artist to select.
Let the games begin. The forger will try and fool the detective, gradually creating an artificial image of Aladdin. Once the detective is fooled, the process is complete.
Nvidia created something called StyleGan in 2018, which gives the artist more control over the images being generated.
The future of GANs
GANs, in general, are very powerful and have lots of potential applications. They will play a significant role in a number of areas in the coming decades. Here’s a non-exhaustive list of the things GANs can do:
- Create deepfakes: artificial videos
- Photography applicationsAge or de-age photos
- Colorize black and white photos
- Make animated paintings
- Detect glaucoma
- Discover new drugs
- Predict the effects of climate change
For AI-generated art, GANs are only the tip of the iceberg. As Hidreley’s work shows, GANs can produce some awesome results. However, it’s early days, and this technology continues to evolve.
Last year, OpenAI released DALL-E 2. Like a GAN, DALL-E uses multiple neural networks to produce novel images but uses an algorithm called a diffusion model. Some claim diffusion is more effective at creating artificial phots than a GAN. And unlike a GAN, it can generate images directly from text. It’ll be interesting to see where this technology goes in the coming years.
A double edged sword
This stuff is a lot of fun. Who doesn’t want to know what Elvis would look like if he was alive today, or what the Harry Potter characters would look like solely based on the way they’re described in the books.
And, remagining a historical figure like Cleopatra as someone you could walk by on the street is, I think, healthy. We tend to construct mythical narratives around historical figures, to the extent that they don’t seem real. They seem more like Gods. Seeing these photos allows us to take them off the pedestal and gives us a different perspective on their lives and the times they lived in. It makes them more relatable, more human.
And, in a culture that runs on soundbites, spectacle, and airbrushed snippets of our lives tailored for Instagram and Tik-Tok, it’s a reminder that perhaps we would be wise to take ourselves and one another off the pedestal as well.
The Artificial Artist
An artist in Colorado made headlines when his AI-generated artwork won the top prize at a state art competition. You can use AI to write emails or articles with tools such as Jasper or even have AI create music for you.
How much of this can be called art?
The creation of AI-generated images is a collaborative process. Humans are guiding the blind hand of the machine, drawing on imagination and intuition. But the machine is the one doing the creation. They both need each other.
Looking at the range of products that AI can create today when the technology is still in its relative infancy, can invite existential pondering.
The handwringing over new art-making technology is an old story. As Kevin Roose writes:
“Many painters recoiled at the invention of the camera, which they saw as a debasement of human artistry. (Charles Baudelaire, the 19th-century French poet and art critic, called photography “art’s most mortal enemy.”) In the 20th century, digital editing tools and computer-assisted design programs were similarly dismissed by purists for requiring too little skill of their human collaborators.”
But AI raises new questions. Why will anyone pay for art if they can simply download a program on their computer and make their own? What happens when AI’s art is better than our own?
The question we may need to answer is whether the purpose of art resides more in the effect it evokes on those who consume it, or in the act of creating it.
