These locations may look eerily familiar, but none actually exist

Whether or not it’s Microsoft Paint, Adobe Photoshop, Snapchat’s Bitmoji, and even the (imaginary) Star Trek Holodeck, designers and engineers have been working for many years to determine learn how to flip our ethereal imaginations into tangible photographs—with as little technical experience as potential.

Strolling Cities—a brand new video undertaking out of the MIT-IBM Watson AI Lab—takes this custom to new heights. It’s a video that feeds poetry right into a machine, and permits synthetic intelligence to dream what the poetry appears like.

[Image: Mauro Martino/courtesy Strolling Cities]

And in an auteurial twist, this AI wasn’t skilled on all the pictures of the world to have the ability to dream something in any respect. As a substitute, consider the AI extra like a toddler that was raised on the streets of Italy. So its total level of reference are canals and porticoes, cobblestones and sea. It’s an AI mannequin that’s restricted by design, constructed to have a particularly Italian standpoint, to seize the nostalgic sensation of visiting a particular place.


The undertaking was born from COVID-19 lockdown, as Mauro Martino, the top of IBM’s Visible AI Lab, missed house. He was in Cambridge, Massachusetts, whereas the pandemic exploded in his house nation of Italy.

“I made a decision that the sweetness and sentiment, the social, historic, and psychological contents of my recollections of Italy might grow to be a creative undertaking, most likely a type of emotional comfort,” says Martino. “One thing stunning at all times comes out of nostalgia.”

To construct the video you see above, Martino’s crew enlisted college students from Politecnico di Milano. Throughout lockdown, they walked the streets of 9 totally different Italian cities, capturing 2 million images of those cityscapes, not from above or by automotive, but intimately on foot. The photographs had been then labeled (with phrases like “sky” or “window”) by means of automation, whereas an AI was skilled to think about cities from nothing but these photographs.

As Martino factors out, we’d already seen all types of technically proficient image-generating AI techniques, from imaginary Google Avenue View to tulips. But constructing these techniques requires piles of supply picture knowledge, so most AIs study what issues look like from publicly posted photographs on the web. Meaning you get an AI that may generate one thing that seems reasonable, but aesthetically, it’s not compelling. It’s a technically correct, boring common.

[Image: Mauro Martino/courtesy Strolling Cities]

“There isn’t a consciousness of the complexity of the cinematic language, there isn’t a authorship within the composition,” says Martino.

As a substitute, Strolling Cities wears blinders. It develops unmistakable, but additionally generally unplaceable, Italian landscapes—a psychedelic fever-dream mixture of Bologna, Venice, Rome, Como, and extra—all captured by means of the identical deliberate methodology and digicam system. The supply footage is curated, permitting the system to generate absolutely pretend photographs that also really feel like they’ve a standpoint.

[Image: Mauro Martino/courtesy Strolling Cities]

“There’s authenticity, in Strolling Cities you may see Rome as a Roman lives it,” says Martino. “One thing magical occurs, the landmarks disappear, but the cities are nonetheless recognizable.”


It’s straightforward to observe together with the best way the AI thinks. A point out of the ocean makes the ocean seem, and sidewalks makes walkways seem. The narrator saying “aerial verticality” makes the buildings stretch into the sky. And at occasions that there’s not a transparent sufficient Italian reference level, similar to a point out of “rice fields,” the system appears to do its greatest, providing a subject of one thing that appears not fairly like grass, but not fairly like rice or another plant both.

[Image: Mauro Martino/courtesy Strolling Cities]

As for the way forward for the undertaking, Martino is planning to debut real-time installations, which let you communicate and have the AI think about in actual time—whereas pushing the boundaries of the system’s creativeness. “Now we will generate full purple cities with blue streets, or be extra summary and generate a romantic location, or miserable place,” says Martino, teasing that quickly, we will communicate our minds to computer systems, and permit them to dream something we’d think about.

“It’s an exquisite time for ‘dreaming’ collectively!” he says.