This Neural Network Dreams In Cities

Inspired by Italo Calvino, a team of engineers teaches neural networks to imagine entirely new cities.

From above, each city has its own distinct character: something beyond the sum of its parts, a unique fingerprint of streets and houses that makes it quintessentially itself. Even if you didn’t know the landmarks, you’d be unlikely to mistake a satellite image of, say, Milan, for one of New York or Hong Kong. They just have different styles.


Now imagine if an aerial image of New York could camouflage itself as one from Hong Kong, or Milan. That’s just one of the things Invisible Cities, a new project from Italy’s Opendot (which includes Gene Kogan, Gabriell Gambotto, Ambika Jib Samsen, Michele Ferretti, Andrej Boleslavsky, Damiano Gui, and Fabian Frei), can do. It’s a neural network that can not only transfer the style of one city onto a map of another city . . . it can dream up realistic satellite imagery of entirely new, imaginary cities.

The backbone of Invisible Cities is pix2pix, machine learning code that radically changes the style of an image–making a sketch of a handbag into a realistic photo of a handbag, for example, or translating a picture taken by day into one taken at night. In Invisible Cities‘ case, though, the bot has been trained on a MapBox database of aerial satellite images of cities, along with their corresponding map tiles of OpenStreetMapData, which color-code the images by roads, parks, buildings, and more.


By learning the relationship between the satellite images and the map tiles, the bot is able to take a map tile from one city and give it the “style” of another. For example, give a more Mediterranean city like Naples the industrial look of New York. Even more interestingly, Invisible Cities can generate entirely new realistic-looking satellite images based on hand-drawn sketches. In other words, creating an imaginary city is as easy as sketching down some lines on a piece of paper, then feeding it into the bot, which anyone can do through the publicly available code (technical knowledge allowing).

Given the name of the project, it’s no surprise that the team was inspired by Italo Calvino’s 1972 novel Invisible Cities, which is largely made up of 55 prose poems about mythical cities, which are in turn inspired by the fanciful (and mostly made-up) cities that Marco Polo allegedly visited on his trip across Asia in the 13th century. “We thought it was an appropriate analogy as we started making these generative city maps,” Kogan says.

Since the technology can be used to create realistic fictional cityscapes with minimal effort, it could be useful for video game developers or even Hollywood special effects artists. Shorter term, though,the team says they might bring Invisible Cities to the web, making it easier for people with less technical acumen to bring the fanciful metropolises of their minds to life.