Designers, take a deep breath: Really and truly, AI is not on the verge of stealing your clients. No matter how good an algorithm gets at generating beautiful renderings, your business is safe until a computer can deal with the messy complexity of real people in the real world. (The day AI can simultaneously settle a dispute between spouses, zhuzh a flower arrangement and field an angry call from a contractor, start to worry.)
However, we are going to see technology offering AI-powered design help to everyday consumers—especially a version deployed by retailers who want to sell their own merchandise, or by independent platforms that want to make a cut on the sale of a variety of products. Right now, a lot of the technology on the market is impressive, but not really all that useful. In other words, we lack what’s referred to in the tech industry as a “killer app”—a tool so powerful it becomes ubiquitous.
What would such an app look like? Imagine pulling out your phone, snapping a pic of a room and getting a near-instantaneous, perfect-to-the-half-inch AI-generated rendering of what the space would look like in, say, Memphis Milano style. You then add a few parameter changes, which the program executes perfectly—try an orange geometric-print sofa instead of teal tweed, you say—before delivering a shoppable list of product seconds later. That’s a killer design app. But in order for it to happen, three fundamental AI challenges need to be solved.
The Envelope Problem
The very first AI-powered interior design apps shared a similar problem: They didn’t just redesign a room, they tore it apart. The algorithm would move windows, demo walls and, in some cases, completely ignore physics and gravity. (Try giving that rendering to a contractor!) This is the envelope problem: AI doesn’t really know the dimensions of the room you’re asking it to re-imagine.
Of the three big hurdles, this one is probably the easiest to solve. In the past six months alone, AI has gotten a lot better at preserving the basic physical properties of the images it’s redesigning, and many of the current AI design tools have evolved to stick to the basic shell of a room. If developers can find a way to tie in data from a smartphone’s depth-sensing Lidar scanner, the envelope problem will be solved.
The Precision Problem
If you want AI to generate an image of a serene coastal living room with a cream boucle sofa, bleached wood and blue accents, you’re in luck. If you like the results, but decide to swap linen for boucle—well, that’s difficult. Instead of making your edits, the AI is likely to overcorrect at best, or come up with a new room entirely. (In one test we tried, the AI blanketed the entire room in the new sofa textile—not exactly what we had in mind.) This is the precision problem: The AI tools currently on the market are not good at making small, finely tuned changes.
This quirk is a byproduct of the way these systems are trained: By “looking” at millions of images and coming up with averages, they can produce approximations of a particular style. But the algorithm does not “know” that a sofa is a sofa, so asking it to swap out the fabric on one is a tall order.
There are ways around this challenge, and the developers making AI design tools are working on it. But getting AI to successfully make hyperprecise adjustments will be a significant challenge—especially the kind of nuanced requests humans often have (“I want this exact room, but make the pillows bolder and the curtains more sheer—but not too sheer”).
The Product Problem
AI tools are capable of creating surprisingly good interior design schemes, but even if you wanted to buy everything in an AI room, you couldn’t—none of it is real. This is the product problem: Artificial intelligence only generates approximations of real furniture and decor, not the real thing.
There are some exceptions to that. Because iconic pieces like the Barcelona chair are so ubiquitous in popular media, AI—which is trained by looking at millions of images—will often replicate them accurately. But for the most part, AI is generating something like the mathematical average of 150,000 chairs, not a specific piece you can buy at DWR.
Developers are working on this issue too. Sometimes, the attempted solution involves a bit of reverse engineering: Let the AI do its thing, then let another AI scan the image to try to identify a close-enough look-alike. In other cases, the solution involves training the AI exclusively on the inventory of a particular retailer. Neither works perfectly. However, of all the technological roadblocks, solving the product problem is probably the most lucrative, so expect to see a full-court press to find an answer. Where there’s a will—and billions of dollars at stake—there’s a way.
Homepage photo: ©Andrey Popov/Adobe Stock