Refik Anadol — Machine Hallucinations: Nature Dreams
What type of machine learning models did the creator use?
Generative Adversarial Networks (GANs) to generate new, dream-like moving imagery from a huge visual archive. On the project pages and exhibition texts, the work is described as being trained with GAN algorithms, and the broader Machine Hallucinations series mentions specific GAN families like DCGAN, PGAN, and StyleGAN (and related “Generative AI” approaches).
In other Anadol works in the same “Machine Hallucinations” lineage, the studio explicitly mentions StyleGAN2-ADA for training on art-collection subsets (useful context for the technical direction of the practice).
What data might have been used to train the machine learning model?
For Nature Dreams, the studio states it used over 300 million publicly available photographs of nature, collected between 2018–2021, as the raw dataset to train the GAN. More broadly, Machine Hallucinations is framed as using vast visual datasets tied to “collective visual memories” of domains like nature, space, and urban environments.
Why did the creator of the project choose to use this machine learning model?
Because GANs are built for “dreaming up” new images: they learn a latent space of visual patterns (textures, shapes, compositions) and can generate novel outputs that feel related to the training archive but aren’t just direct copies—perfect for an artwork about “hallucination,” memory, and machine imagination. The project explicitly describes training a GAN and presenting the results as the machine’s “dreams” of nature, and also because GANs scale well to big image archives and high-impact visuals: immersive installations need large volumes of coherent, continuously generative imagery; GAN-based pipelines (e.g., StyleGAN-family approaches) are commonly used when artists want high-resolution, visually rich, continuously variable outputs for projection and spatial experiences—exactly how this work is presented (as a large-scale installation/data sculpture).
Week4_HW by ZhaoquanZheng -p5.js Web Editor
For this project, I explored how hand gestures could be used not as direct commands, but as subtle forces that influence a generative system. Instead of designing an interface where specific gestures trigger specific outcomes, I wanted to create an interaction that feels more atmospheric and less deterministic. The result is a flow-field–based visualization in p5.js that responds to hand movement captured through real-time hand tracking, producing an environment that behaves more like weather than a machine.
Visually, the system is composed of thousands of particles moving through an invisible vector field generated with Perlin noise. Each particle follows the local direction of the field and leaves a short trail behind it, allowing the structure of the flow to slowly emerge over time. As particles accelerate or slow down, their color shifts accordingly, moving from cooler blues and greens into warmer yellows, oranges, and reds. This gradual accumulation of trails creates an effect similar to wind maps or atmospheric simulations, where motion itself becomes the primary visual language.
Rather than controlling individual particles, the user’s hand acts as a disturbance within this environment. I used ml5.js handPose to track twenty-one hand landmarks from the webcam, then derived higher-level values such as hand position, movement speed, and openness. These values do not map to discrete actions. Instead, they continuously modulate the strength, radius, and rotation of a vortex embedded in the flow field. When the hand moves quickly, the system becomes turbulent and unstable. When the hand slows down, the flow begins to settle into smoother and more legible patterns. This creates an interaction that feels responsive but never fully predictable.