by Chloe Karnezi
Using machine learning and Arduino, “Digitised Dysmorphia” is an interactive triptych, and a subtle critique on the shortcomings of digital technologies as a means of capturing our physical selves accurately. It is also a comment on the effects of these shortcomings on our perception of our physical selves. The word “dysmorphia” in the title alludes to body dysmorphia, a condition where the way one views their body is vastly different from the way other people do, predominantly in a negative way.
The phrase “Do I actually look like that?” is one that has an all-too-familiar ring to it for countless women, or in fact anyone who has experienced seeing themselves figured in a digital way, usually in photographs or videos. Every smartphone brand, for example, has a different camera. Each of them skew us slightly differently, stretching, warping us in infinitesimal ways that leave us wondering which version of ourselves is real. These distortions are all down to small differences in manufacturing, and yet they have the power to impact our view of reality, our body image, and self-esteem.
If body dysmorphia makes us view our bodies as distorted in our mind’s eye, digital technologies are the real-life facilitators of those distortions.
On the Interaction
I like to think of this triptych as “self-conscious”: the bodies “hide” if you get too close. The way to interact with it is to slowly move towards the bodies. The lights dim as you get closer, and when you’re “dangerously” close, meaning at a distance where you can really see the details, the marks and blemishes, the lights dim completely, and you are confronted with your own reflection in the mirror.
Technical: I made the work interactive using Arduino: I used an HC-SR04 ultrasonic distance sensor, attached to the bottom of each frame of the triptych, to record the distance of the audience from the work, and LED strips inside the structure whose brightness depends on the distance data. When at first I mapped the brightness to the raw distance data of the sensor, I noticed the lights were too flickery. Due to the quickly-changing distance values that the sensor recorded, I calculated an average distance to smooth out the results. This allowed for smoother LED dimming, as the brightness of the LEDs is mapped to the average distance.
I used a two-way mirror, which means that when you are at a safe distance from the bodies and the LEDs inside the structure are on, the mirror becomes transparent, and you are able to look into the structure, at the bodies. When you get close and the lights dim, meaning there is more light on the outside of the structure than the inside, the glass becomes a mirror and you can only see your own reflection.
There are a few reasons why I wanted to use the lights and mirror interaction and play on distance:
First of all it’s about control: The female nude, classically, throughout the history of art, has been figured, painted, sculpted, by male artists and filtered through the prism of the male gaze. This female nude attempts to turn that trope on its head in that I took the photos myself, of myself, and the bodies of this work have agency through the interaction. They have the power to conceal or reveal themselves to you: you can only view them on their own terms.
The interaction also calls people out on our voyeuristic tendencies: we use platforms such as Instagram to scrutinise and "dissect" other people’s digital bodies on social media etc. In this case those tendencies are rudely shut down as when you come closer the lights turn off and you’re invited to reckon with those tendencies as you, quite literally, take a long hard look in the mirror.
The LED/distance interaction is also an observation on self-esteem and vulnerability: we are more comfortable being seen from a distance but often want to quite literally turn the lights off when someone is too close and able to see our imperfections.
I used the mirrors as a comment on digital vs analog technologies, where the analog technology is the mirror and the digital technology is the camera. It’s interesting to think about how the weight and significance of the mirror shifts as we see ourselves increasingly on screens or use our phone cameras as mirrors instead. Are the two really the same? Are they equally trustworthy?
Lastly, the interaction is about vanity: This work is, in a sense, a bi-directional selfie. The generated images come from selfies of my body that I took, and when people interacted with the work during the exhibition they would usually hold their phones up, recording my body selfies. As they slowly closed in on them the glass would become a mirror, in which they were recording a mirror selfie of themselves. It was fascinating to observe the audience interact with the work: some chose to spend more time close to it and enjoy looking at themselves in the mirror rather than take a step back where the lights would come back on and they could see the bodies through the glass.
Generating the Bodies (Machine Learning)
Technical: To generate the images of the bodies I used a program called Runway ML. More specifically, I used a class of machine learning called a generative adversarial network, or a GAN, as I will be referring to it.
First, I had to create a dataset. I created one by taking 600 photos (selfies) of my body, which I took over time. The GAN then studied these images and was tasked with creating new, non-existent images of my body, convincing enough to pass as original images of the dataset. The “adversarial” nature of the GAN comes down to the way it works: it is comprised of two neural networks. It’s helpful to think of these as a forger and a detective, contesting with each other. The forger attempts to generate the most convincing new image of my body, while the detective is tasked with working out why the image is not good enough to belong to the original dataset.
The longer you let the machine learning model train, the more similar to the original images these new generated images will look.
So it then becomes a question of: how long are you comfortable training a machine learning model that over time creates increasingly accurate and explicit images of your nude body, before these images become too accurate for your comfort and you become too self-conscious?
It would have taken about six hours for the generated images to be quite similar to the original photos of my body that I took for the training dataset. I only felt brave enough to train it for two.
The resulting bodies are abstract, yet simultaneously not unlike what you would see if you were to look at the details of a body through a microscope. They are abstract enough that anyone can recognise something of themselves in the folds and furrows of the flesh, in the cracks and the blemishes and the discolourations. The shapes of the bodies also look quite dynamic: the body looks soft like clay, and like it’s morphing from one shape to another.
The female form is always fluid, stuck in motion, with each woman trying to mould and sculpt her body to reach the transient ideals of female beauty.
Using a GAN allowed me to disrupt the automatic, instinctive process of feeling embarrassed, vulnerable, and critical at the sight of my naked body. The generated bodies are at once my own, in that they were generated using images of me, and not my own, in that they are entirely new and previously non-existent. Am I therefore critical of them and embarrassed by them? To what extent can I claim them as my own?
These bodies also raise the issue of evading online censorship of the female body which happens so often, with the famous example of instagram censoring female nipples. As the GAN continues to train, at some point the generated images would become too accurate to evade online censorship and they’d be taken down from a platform like instagram.
The painter Jenny Saville is a great inspiration of mine. Her Grotesque, raw representations of the female nude with all its textures and folds and blemishes forever shifted the way we see women’s bodies.
For my process I was influenced by that of Cindy Sherman, the American artist whose work consists primarily of photographic self-portraits. There is something very powerful about the way she points the lens at herself as opposed to giving someone else that authority and authorship of her own image. When I see her self-portraits I see her at once in front of the lens and behind it.
I was also inspired by the Theory of Disobedient Objects, which we studied in our theory class here at the university. Disobedient objects are ones which are repurposed or fulfill a different purpose from what you would expect them to. In the case of my work you would expect the lights to come on when you get close to the pictures as opposed to dim, because the reason you would come closer would be to see more. But that’s when the lights go dark, so it’s a disobedient object in that it refuses to obey.
Self-evaluation and Future Development
By using images of my own body, this work largely becomes a diary of my own struggles with body dysmorphia. Whilst it can be an important and powerful experience to peek into other people’s diaries, the experience would be different if the audience were reckoning with images of their own body. Whilst the audience were able to experience the “voyeurism” aspect I wanted them to, they would only be able to experience the vulnerability and embarrassment aspect if the bodies were more explicitly their own. I would like to look into creating a more personalised experience for the audience in the future. As a first step I would like to create a collection of more diverse bodies of varying shapes, sizes, and skin tones.
Jenny Saville, painter
Cindy Sherman, photographer
Hertz, Garnet. "Disobedient Electronics." (2016)