When Zillow needs computer vision technology that doesn’t exist, its engineers invent it

These days, when consumers use Zillow’s tech to interactively browse media on a floor plan on a Zillow listing or fly-around a home's exterior as if they’re operating a drone from their computer, they may take for granted the state-of-the-art AI that enables these experiences. But when Zillow's scientists set out to build it, the technology they needed did not exist. So they invented it.

"Zillow works on the cutting edge of computer vision [a field within AI] more than people know," said Manju Narayana, a lead scientist on the team. "When there's no known solution for certain problems, which happen even with the best methods available, we invent the solution."

The fact is that Zillow is a technology company that operates in the real estate space. And it explains why a company most people associate with home listings keeps turning up at the field's top academic conferences. Zillow's research has appeared repeatedly at the Computer Vision and Pattern Recognition Conference (CVPR) and the European Conference on Computer Vision (ECCV), the two most competitive events in the field, where reviewers wave through only about one in five submissions. Its newest paper, titled "GaussFusion: Improving 3D Reconstruction in the Wild with Geometry-Informed Video Generator," was accepted to CVPR 2026. It solves a problem that has stumped the field: the visible glitches and distortions that even the most advanced methods leave when building a 3D model of a home.

But GaussFusion is far from a one-off. In 2021, Zillow’s Indoor Dataset generated a string of papers that each cracked another piece of the same puzzle: reading a room's shape from a single image, stitching panoramas together and working out where each photo was taken inside a home. Behind it sits a substantial operation, involving a cross-functional team of science, engineering, product and operations, along with a growing portfolio of computer vision patents. The team has also developed and published state of the art models for floorplans, localization of images for real estate tours and reconstruction of floorplans. "It tells the community that we are doing serious computer vision work," Narayana said.

Over the years, that work has come down to two kinds of problems. Sometimes the technology Zillow needs simply does not exist, and the team has to invent it from scratch. Other times a promising method already exists somewhere in the research world but is nowhere near ready for real use, and the job is to rebuild it until it holds up on millions of listings.

"The through-line is that we don't stop at what's already out there," Narayana said. "If a tool exists and it works for our customers, we use it. When it doesn't, we build the technology that should exist."

Sometimes the technology has to be built from nothing

Each day, Zillow users pull up an interactive floor plan of a home and, without a second thought, tap a listing photo to see exactly where in the house it was taken, grasping not just what a kitchen looks like but where it sits and how the rest of the home flows around it. To make that feature happen, the Zillow engineers needed AI that could take the everyday photos and 360-degree panoramas captured during a listing shoot, figure out the shape of each room, fit them together into an accurate floor plan, and then drop every photo onto the exact spot where it was taken. No product on the market could do it well.

"Typically, we will work with the latest and greatest that is acceptable to us," Narayana said. "But we often have to invent something beyond it."

Often the first thing missing is the data. So the team built its own, a large public collection of annotated home interiors called the Zillow Indoor Dataset.

"We really believe that modeling indoor spaces, building maps, building floor plans is a customer benefit," Narayana said. "But that technology didn't exist. So we needed to capture a dataset that truly represented the problem we're trying to solve."

Over many months, engineers then went to work building the technology that now runs quietly behind Zillow 3D Home tours and interactive floor plans on listings nationwide.

Sometimes technology exists, but needs help

Every so often a promising method already exists in the research world, and the job becomes making it work in the real one. SkyTour is the clearest example.

On a Zillow Showcase listing, SkyTour lets a shopper swoop around the outside of a house, circling the roofline or checking the distance to a neighbor's fence, fully user-controlled and entirely from real drone footage. The effect comes from a fairly new rendering method called Gaussian Splatting, which blends thousands of overlapping frames into one smooth 3D scene. Zillow was the first real estate platform to bring it to shoppers at scale.

"SkyTour is a great example of leveraging what's been contributed by the broader computer vision community and building on top of it,” said Will Hutchcroft, another scientist on the team. "We didn't substantially change the core technology there. The challenge is building everything around that technology so it actually works with the reliability a real product requires."

That surrounding work, he said, is the hard part. "Even if you can find this technology that's been released by somebody else, to actually turn that into something that works reliably, over and over, in a live product, is a task in itself."

The idea in every case is that the result is the same for the person on the other side of the screen: a home they can walk through and study without leaving the couch. And with a roadmap full of new features still to come, the scientists behind them expect to keep running into the same wall, and keep building their way through it. "Our roadmap is full of computer vision-based customer features and experiences," Narayana said. "Users should absolutely be expecting more."