An Interactive Installation Generating Brutalist Buildings with Pix2pix GAN

Description

This project is an interactive installation that utilizes machine learning to generate Brutalist buildings in real-time. The installation invites participants to capture images via webcam and witness their concrete and bricks collages transformed into towering, monolithic structures.

By allowing individuals to create these monumental edifices, the work discusses the typical power dynamic between humans and Brutalist architecture.

Medium

Materials: transparent mapping fabric, rebar, wood, webcam, projector, Mac mini
Technologies: pix2pix GAN, machine learning, OpenCV
Dimensions: 200 × 150 × 226 cm

Exhibition

Bargehouse Gallery, London, UK
April 2023

Process

Dataset & Training
A custom dataset of Brutalist architecture was compiled and structured for image-to-image translation. Architectural outlines were used as inputs, paired with photographic images as outputs, enabling the model to learn the transformation from sketch to built form.

Model
The system is based on a pix2pix GAN trained in TensorFlow, generating architectural imagery from real-time inputs. The model continuously interprets webcam captures and outputs evolving structures.

Installation System
A live camera feed captures participant input. The generated images are projected onto a semi-transparent architectural surface, creating a layered spatial experience between physical structure and synthetic imagery.

Gallery