GAN, Procedurally Generated. Collaboration with Steven Sych.
Featured in Towards Data Science.
Telephone Paintings is an infinite-3D art gallery. Inside it are paintings and texts created using artificial intelligences. The work uses three generative adversarial networks (GANs) working in tandem:
- The first GAN (StyleGan) creates an infinitely varied set of paintings in the style of abstract 20th century art.
- The paintings are then shown to a second GAN (Yolov3). Yolov uses object detection to name the objects in the paintings, which are then used to generate a descriptive phrase in fairly grammatical English.
- This descriptive phrase is then fed into a third GAN (GPT2) which creates a short description of the work’s history and (imaginary) artists involved. This network has been trained on the descriptions of various modern art museums.
The gallery is then procedurally created using Unity. Code controls the selection of artworks, the camera, and the 3D models. Nevertheless, the 3D gallery is site specific: its basis is high quality 3D scans of the Anteism Books Project Space in Montréal, Quebec, where the physical installation took place in 2020. The audio is an abstracted soundscape of server farms-- fans and rhythmic mechanical pulsing.
“I was not afraid of losing the “personal touch” so highly valued in previous painting. On the contrary, I even gave up signing my paintings. I put numbers and letters with the necessary data on the back of the canvas, as if they were cars, airplanes, or other industrial products.”
-László Moholy-Nagy (1944)