Ümüt Yildiz - Grace-2020-13

Grace was my initial self-trained StyleGAN model. When I delved into AI in generative art, I used BigGAN initially. Although it was an interesting model, it didn't allow me to utilize my own pictures. In 2019, StyleGAN was released, marking an incredibly captivating moment for me. Towards the end of 2019, I extensively worked on my local PC while pursuing my Bachelor of Art, aiming to complete it by early 2020. During this period, I pushed my computer to handle challenging tasks for my thesis. AI back then differed significantly from today's capabilities. It required collecting a substantial dataset, around 1000-2000 pictures, which I curated manually. Tools like Midjourney and Runway emerged around that time, but being a student with limited income, I couldn't afford virtual machines. I did my best with my local machine. StyleGAN allowed the machine to learn from the data I provided, enabling me to infuse a part of myself into it. I explored how AI perceives humans and the beauty of our species, delving into a study that fascinated me. Additionally, I was intrigued by social media as a phenomenon, collecting over 1000 smartphone pictures of various people. Over a span of more than two weeks, I collected and curated the pictures. Despite the challenges posed by my thesis and freelancing for a living, I resumed experimenting with StyleGAN in December 2019. The training process ran continuously for about three weeks until the machine produced its first evolved picture. StyleGAN initially replicates the dataset before attempting to create new variants. The images were initially pixelated and blurry, prompting adjustments to parameters and another month of training. Unfortunately, the computer crashed during this period, coinciding with the onset of the COVID-19 pandemic. Despite social media's prominence in Germany and the support among artists, the StyleGAN results still didn't resemble the dataset. It needed more training time, but an unfortunate turn of events occurred—the computer crashed. However, the StyleGAN model persisted on the hard drive. After rescuing the model to a notebook, the rendering process began. Witnessing how AI thought and interpreted was exhilarating; the results were abstract, focusing primarily on eyes and hair. It seemed like the machine was on the brink of understanding how a human face is constructed. AI thinks and sees in patterns, so it's natural that eyes and hair, prevalent in the dataset, were learned first. While some considered my experiment a failure in the context of beauty images, it was not so for me. Throughout the output creation process, I discovered various things, such as the initial aspects we remember about a person, like their eyes and hair. The beauty of abstract morphing shapes forming new creatures and the eyes providing a living, organic feeling to the work became evident after multiple renderings from this model. After three years of working with AI, I'm thrilled to finally present this piece, showcasing the hard work invested in its creation and the progress we've made in the AI community. No Sound