Algorhythmic Collage: Representation in Latent Space
Can AI make art?
No, but artists make art using AI. According to Sofian Audry, the “triumph of machine learning” is based on several decades of computer science work in fields such as cybernetics that addressed the function of the brain (i.e., neurons) and how it could recognize patterns. In the contemporary art world of the 1960s, artists began conceiving works that could be created using computer-based, artificial systems. In this new world, there were few, if any Black or brown artists and no works representing Black women.
Three decades after Mark Dery published his essay on digital culture that launched Afrofuturism the unequal playing field of the digital divide still exists. Equity and representation in fields such as computer science is still a big issue. Some feel that AI, specifically machine learning or generative AI can change this.
Some assert that art (human creativity) has to play a role in this change. Last Saturday, John Pasmore (Latimer.ai) spoke to an audience at the Black Joy AI Summit about the lack of culturally relevant data in existing large language models or LLMs. Pasmore used ChatGPT, a chatbot that uses AI to generate responses to prompts, to demonstrate his point by querying who are “our most important artists.” He got a list of white, or European artists. Recently, I created AYA, which stands for Artificial Youth Algorhythm, as a ‘sociopolitical statement’ about AI. I prompted ChatGPT to generate a story about a young, multinational computer scientist involved with hip-hop culture who invents a virtual chatbot named AYA, that becomes a human being.
AYA exists in a place called NeoAccra, a culturally relevant representation of latent space. In simple terms, a latent space is a hidden world or layer of patterns and relationships that a machine learning model finds in data. Imagine training a model (machine learning) with a bunch of pictures of faces. The model can’t “see” the faces like humans, rather it breaks them down into numbers and patterns based on features such as nose shape, eye color, etc.
The patterns that are stored in latent space are like maps of how the different features relate to each other. In this unseen space, faces that look similar are closer together, and faces that are very different are farther apart. It’s a way for the ML model to understand and work with the data without directly using the original images.
So, latent space is like a hidden layer of meaning that ML models use to make sense of mountains of computer data.
Here is what latent space is NOT:
Latent space is not a human brain, nor does it exist on a plane or in a box. Rather, it is an abstract, multi-dimensional representation of data where similar items are mapped closer together. Artists can use latent space to remix or combine subject matter, compositions, and styles. We can use it to re-interpret existing works or generate new ideas based on those works. The AI-powered tools we use to generate images and other content can’t duplicate existing works. GenAI tools can only make predictions based on the information and data we give it. It takes a lot of knowledge, practice, and skill to use GenAI tools well and latent space opens doors to new possibilities (aesthetics, performances, experiences).
AYA exists in a new reality. She has one foot in culture and the other in machine learning. She appears as different representations of Black femininity in many of my AI-generated artworks.
The complexity of Black women’s experiences and perspectives have consistently challenged Western notions of femininity. At the intersections of race, gender, and class, the very existence of Black women pushes against narrow definitions of what it means to be women. — Black Future Co-op Fund
AYA represents me and many other Black women who are in spaces where they don’t see anyone like themselves. She is a creator who turns from the past to make art that is emerging, or disruptive. She is a maker, a self-taught engineer who uses machines to solve problems. AYA is reflected in my current process that melds the traditional with the technological. “Electric Kente” (see above) references Ghanaian Kente cloth, which is known for its bright colors, geometric patterns, and symbolic designs and inserts electronic circuitry into the Kente design. Hints of the Kente design can be seen in Black (African) American quilts: in a story that begins in Africa with the invention of the audiovisual style of polyrhythms.
The AYA project centers the African diaspora and other underrepresented, historically marginalized groups in mainstream technology development and use. This includes the mastery of prompt engineering and access to different tools such as Midjourney, Deep Dream Generator, and Adobe Photoshop. Three years ago, when text-to-image prompt-based AI tools were released I had already been using neural style transfer or NST to make AI art. I was using vintage metallic tapestries to add “shine” to enhance the faces of Black people. Text-to-image tools that generate highly stylized, culturally specific images are the result of artists’ knowledge and experiences, and not just the capabilities of the tools.
For “Loc-Hawk” I uploaded a personal photo to use as a reference. It was a photo of a Black boy with a painted face. I added text that helped the tool generate and predict an alternative outcome (ex. a Black woman instead of a boy). I upscaled and saved one. Next, I used DDG to apply neural style transfer or ‘Deep Style’. I repeated this process with other MJ-generated images, still using my photo as a source. I just changed the reference styles for each image, as seen in “Veiled” and “Helio.” I used Photoshop to extend the image and insert a different one of a star chart.
Skills such as coding, prompt engineering and digital imaging are very important but so is having a knowledge of culture, color theory, and visual language. Being a good storyteller is just as important as being a coder. The AYA series is inspired by Afrofuturism and addresses Dery’s questions:
Can a community whose past has been deliberately rubbed out, and whose energies have subsequently been consumed by the search for legible traces of its history, imagine possible futures? Furthermore, isn’t the unreal estate of the future already owned by the technocrats, futurologists, streamliners, and set designers — white to a man — who have engineered out collective fantasies?
Some of the ways we can answer this using generative and general AI is through the creation of alternative, foundational large language models (ex. Latimer.ai), entrepreneurship, and education/training. To this I add creativity and PRACTICE. Being in latent space, making AI-generated images for years has given me the ability to work around some of the data limitations, to create images that represent those of us who have been at the margins and are stepping to the center of the stage.