Generative design tools powered by artificial intelligence stand to transform how architects and planners do their jobs, and how residents see urban problems.
When designers laboring away on a virtual cityscape began observing and tweaking their creation, one of the first things that jumped out at them, when they scanned down boulevards or looked across main streets, was the trees. More aptly, it was the unsettling lack of tree cover.
The streets of InfiniCity, a model 3D city synthesis built using AI that was released earlier this year, can appear a bit like a crude crayon-on-cardboard metropolis. The project renders an unending artificial city via a multi-step process; a series of software programs work together, pulling in satellite data and using other design tools and algorithms to create a streetscape users can endlessly cruise. Co-developer Chieh Hubert Lin says the goal is to “automatically recreate lifelike real-world cities,” generated entirely from the imagination of a synthetic mind.
As mesmerizing as InfiniCity looks, you wouldn’t want to live in this urban ouroboros. Early versions lacked greenery — the geometry of foliage was too fine-grained for the computer to render, Li says — and the built environment is a mish-mash of wonky-looking buildings on streets that randomly dead-end or empty unexpectedly into lakes. It’s sort of a 3D vision of a ChatGPT-authored term paper that descends into made-up quotations and awkward grammar. The organic logic of real-world cities can’t quite be reduced to a SimCity-esque platform just yet.
InfiniCity’s vision of a metropolis-without-end comes amid a wave of interest in the use of artificial intelligence to perform feats of human-like creativity. The latest generation of AI-powered image-making and planning tools can utilize large datasets and neural models to render buildings and cityscapes that range from the uncannily plausible to the purely fantastical; numerous fictional examples created via programs like Midjourney can be found online. Meanwhile, city officials are beginning to wrestle with the relevance — and risk — of AI in urban governance (also the topic of a recent United Nations report).
The goal of automating aspects of urban design and planning in order to reduce costs and improve the final product predates this current crop of generative design tools, notes Imdat As, an architect and co-editor of The Routledge Companion to Artificial Intelligence in Architecture. Computer-aided design, or CAD, software used in the design process emerged in the 1960s and has been in wide use since the 1980s, when Peter Eisenman used computer modeling to create his unrealized Biocenter project for the J.W. Goethe University in Frankfurt and Frank Gehry, whose firm would famously develop their own digital design interface, experimented with CAD. Today, technology such as Building Information Modeling (BIM), digital twins and generative design are used across the field.
Programs such as Delve, a generative real estate design tool from Alphabet-affiliated Sidewalk Labs, and Forma (formerly the Swedish company Spacemaker that was acquired by Autodesk for $240 million) are being used to plan neighborhood-scale projects. Forma can help developers adjust for sunlight and wind on large housing projects, for instance. Spacemaker CEO Håvard Haukeland has said that these tools’ ability to quickly devise and incorporate things like plumbing and power lines into proposed plans avoids what he calls “oh s--t” moments later on.
But issues of quality and data bias can skew the usefulness of these applications, and underscore that the dream (or nightmare) of an entirely AI-planned city remains out of reach — for now.
How to Train Your AI
Even some of the more well-known and pioneering users of this technology have blunt impressions of its current potential. Brooklyn-based musician and cycling advocate Zach Katz, whose Better Streets AI project uses the image-generation system DALL-E 2 to give virtual makeovers to car-centric streetscapes, views the AI image generation as more about messaging.
“AI is useful to create an impressionistic image and artwork to facilitate someone’s imagination, not a blueprint,” he says. He pointed to Brooklyn-based startup Laneform, currently developing a visioning tool for community engagement for public projects, as an example of using digital renderings to build the consensus and political will to turn renderings into reality.
Like ChatGPT and other generative AI chatbots, programs that leverage artificial intelligence for urban design devour vast amounts of data to create their output, gathering imagery of streets and buildings to recreate and imagine new urban forms. The variable quality of this data is part of what’s holding back the ability of AI to have a more concrete impact on cities, says Kory Bieg, a professor of architecture and technology at the University of Texas at Austin who has experimented with AI-generated buildings.
Screening for lung cancer, for instance, requires processing billions of images of lungs, all very similar compared to, say, photos of street corners in large US cities. The sheer diversity of urban imagery can lead to odd hallucinations, like upside-down signage, because variety makes it harder to predict what’s next. “With cities, it’s just such a diverse dataset,” Bieg says. “It’s like everything you can imagine, from a traffic light to the street grid, with many different resolutions.”
Architect and preservationist Mark Hewitt, in the design publication Common Edge, points out another critical problem with the AI-ification of design. The data sources that “train” these architecture programs lean toward the modern and middling, he argues, encouraging these tools to simply replicate unfortunate ideas and trends. He makes the comparison to contemporary Chinese urbanism, which he says mass produces the worst examples of Modernism and car-centric postwar sprawl.
“The good modern city doesn’t exist in any dataset that will be sampled by AI,” he says. “What AI will sample is Chinese cities and cities built with massive highway interchanges. It’ll take a bad city and try to make it less bad.”
Machine-Augmented Creativity
To correct for AI’s faults, Katz uses DALL-E 2 in his BetterStreets AI project, because it allows for images to be edited; he normally takes about an hour to touch up the scenes that the program makes. More advanced AI models, if asked to create, say, a specific street in Los Angeles, but with more bike lanes and pedestrian infrastructure, would be likely to toss in palm trees and the Hollywood sign for good measure. Such visual cliches carry altogether too much weight in these models; efforts to recreate street scenes from specific years in cities like Paris typically deliver a pastiche of architectural styles and building types from many eras.
More detailed data would help overcome this tendency: Bieg says expert annotation would make billions of images of cities and buildings, which often contain more generic labels and metadata, usable for urbanists, architects and planners. He’s found that existing datasets often tag things like doors and windows in very simplistic ways, which reinforce biases. “Every time I’m in a conversation with other academics, that’s the topic that comes up,” says Bieg. “How do we start to gain control over this data?”
Other popular datasets, such as HoliCity, which is based on street images of downtown London, carry their own biases, since they draw from one metropolis with a specific climate and resulting building stock. Having architectural experts go through these datasets to identify specific design elements or features they do or don’t want — somewhat akin to new services offering to train writing bots on the existing work of a human writer — would “loosen the grip” technology has on defining design elements and make them infinitely more useful. He sees this eventually feeding ideation sessions, where designers and planners can mix styles and structures in Santa Fe, Manhattan and Detroit, and come up with new hybrids.
As is utilizing AI to help design the NAR Innovation District in Istanbul. The Spacemaker generative design program was utilized to help create the master plan for the district, which is envisioned as a testing ground for various “smart city” technologies like autonomous vehicles, drones and data sensors. (“Nar” mean pomegranate in Turkish, a reference to the project’s goal of seeding technological development throughout the city.) Artificial intelligence helped planners devise the most efficient use of the development site, located on a curved, uneven area, and test the environmental efficiency of different layouts.
One strength of AI design tools is their ability to mash up different aspects of buildings. One of As’s first experiences with this technology was in 2011 as the head of the startup Arcbazar, which hoped to develop technology that could mix and match different parts of homes to create new hybrid designs. The concept attracted funding from DARPA, the US defense innovation program, in 2017, because leaders there felt whatever tool could correctly fuse parts of a house might be able to fuse different parts of vehicles to create a Frankenstein-type battlefield device.
Other recent architectural applications of AI are a little more prosaic. British home developer Quintain used Delve to optimize a rental home project — the tool helped the designers find room for nearly 200 additional units while improving daylighting and open space access. A project by the Institute for Advanced Architecture in Catalonia, located in Barcelona, used AI and street view images to gauge the material inventory of an entire city, aiming to create a way to leverage this data to recycle old material and create a more circular building economy. Bostolena, a project from Sasaki Urban Planners, trained on images of Barcelona and transposed those lessons to Boston to suggest how to begin updating the New England city to accommodate more pedestrian-friendly superblocks.
AI’s Darker Side
While AI often gets employed to refine or improve human-generated designs, some fear it can also be tapped to more nefarious ends — feeding our worst biases and predispositions about urban life.
A recent campaign platform for conservative Toronto mayoral candidate Anthony Furey embodied several of the leading concerns about the use of AI-generated images in political ads. In addition to a quizzical three-armed woman, the candidate’s tough-on-crime messaging was illustrated with what appeared to be photos of Toronto streets and parks overrun with homeless encampments, all conjured via AI.
Such scenes may soon become a staple of local politics, as AI’s ability to create misleading but convincing images gets deployed by candidates or advocates in order to frame policy debates, especially around crime or urban dysfunction. By allowing just about anyone to generate photos or videos, AI could “empower a broader range of voices at the local level, so it’s not just going to be people with money who can advertise,” says Darrell West, a Brookings Institution fellow focused on technology. “It could be your mom-and-pop store in your neighborhood who has a preferred candidate. But it also could lead to a lot more chaos. Pandora’s box is going to be disinformation.”
Others warn that AI could further embed social inequality, especially around housing. A recent paper from MIT urban planning professor Carlo Ratti’s Senseable City Lab showed how a computer model fed with millions of street view images could predict neighborhood characteristics like home values and crime rates — a potentially powerful tool for urban policymaking, or a means of locking-in biases. “Imagine a dystopian future where everyone paints their walls a certain color to impress the bot, or local leaders focus on improving a neighborhood’s score on an AI metric rather than curing its real problems,” Ratti and co-author Antoine Picon write in the Boston Globe.
But Katz, who has deployed the technology to create more utopian versions of cities, is less concerned about the potential misuse of AI; to him, the explosion of AI-aided urban imagery reflects something healthier — an appetite for better places, and an avid desire to see cities remade with a little more intelligence.
“Good urbanism is now popular,” he says. “Even if someone created a vision of bike lanes causing chaos, people wouldn’t buy it.”
---
Autor(en)/Author(s): Patrick Sisson
Quelle/Source: Bloomberg, 1.09.2023
Bitte besuchen Sie/Please visit:
