Project type: Art installation (Museum of Modern Art, 2022).
1. What type of machine learning models did the creator use?
The installation used Generative Adversarial Networks (GANs), specifically advanced forms such as StyleGAN2 with Adaptive Discriminator Augmentation (ADA). GANs are generative models where a “generator” creates images and a “discriminator” evaluates them, pushing the system to produce realistic new visuals.
The training data was drawn from MoMA’s digital collection of artworks — about 138,000 pieces including paintings, sculptures, and photographs. These digitized images provided the “visual memory” that the GAN learned from in order to generate new forms.
Refik Anadol wanted the AI to behave as if it were “dreaming” about MoMA’s collection — not just copying works, but hallucinating entirely new ones. GANs are ideal for generative creativity, they can blend features from many artworks and produce surreal, novel imagery. They support continuous transformation: GANs can generate outputs that flow smoothly into one another, creating the immersive, dreamlike animations seen in the installation.They align with the concept of machine memory: by training on MoMA’s archive, the GAN becomes a kind of artificial imagination of the museum’s collective history.
Refik Anadol’s Unsupervised is an art installation that uses GANs (StyleGAN2/ADA) trained on 138,000 MoMA artworks. The model was chosen because GANs can create new, dreamlike images that continuously evolve, allowing the artist to explore the idea of a machine hallucinating and reimagining the museum’s collection.
p5.js sketch
https://editor.p5js.org/Sitong_Zhou_Silvia/full/k96LeNXaD
his is the first version, but the preview is overexposed.
I changed the blendMode(ADD) to blendMode(SCREEN) and fill(0,15) to fill(0,30). To fix this problem.
Lastly, I changed the color of the particles.