It was dark, and the streets of Altadena were empty enough that a car's headlights were the only usable light source. Nonny de la Peña had deployed a camera operator on short notice, and together they aimed the headlights at the wreckage and started capturing. There were no birds. There were no other cars. There was, as de la Peña would later describe it, nothing.
That scene, shot within days of the January 7, 2025 fires, became part of the raw material for "Out of the Ashes," a documentary de la Peña co-directed with Rory Mitchell that premiered and won the Special Jury Prize at the Venice Film Festival’s Immersive Island this September. The VR film reconstructs seven families' destroyed homes in navigable 3D, built with Gaussian Splatting, a technique that only became practical for fieldwork in the last two to three years.
The Eaton and Palisades fires killed 31 people, burned more than 37,000 combined acres, and destroyed more than 16,000 structures, according to Los Angeles County's own accounting. De la Peña's team, working out of ASU's Narrative and Emerging Media program, decided within days that photos and video alone wouldn't convey what had happened. They needed a format people could walk into and through.
What Gaussian Splatting Actually Does
Most 3D capture built for the web has relied on one of two approaches. Photogrammetry converts overlapping photographs into a polygon mesh. Neural Radiance Fields, or NeRF, train a neural network to predict what a scene looks like from any angle. Both work. Both are slow. A NeRF scene can take tens of hours to train and still needs a network inference pass for every pixel it renders.
Gaussian Splatting skips the mesh and skips the network. It represents a scene as millions of tiny, semi-transparent 3D ellipsoids, or three-dimensional closed geometric shapes, almost like a 3D bell curve. The technique takes its name from the Gaussian distribution, the bell-shaped mathematical function associated with Carl Friedrich Gauss.
In the splat, each ellipsoid has its own position, size, orientation, and color, derived directly from a set of photographs and their calculated camera positions.
The scene is stored in a form a GPU can display right away, not one it has to run through an AI model first. That's why it plays at full speed on a laptop or inside a browser tab, the same as any other 3D graphics, while a comparable NeRF scene can choke trying to render every frame.
The original Gaussian Splatting paper is barely three years old. In that time, it has gone from an academic curiosity to something Zillow, Niantic, Esri, and Autodesk ship in production.
Into the Burn Zone
Rory Mitchell reached the fire zone by filming along the police lines at Santa Monica until he realized an e-bike could get him past the blockade down the beach path. Once ASU and the Walter Cronkite School of Journalism arranged press credentials, the team pushed into Altadena and the Palisades with a rig nicknamed the “Caronna stick,” after Mike Caronna, a colleague and a creator of the rig. The stick was equipped with four Sony RX0 action cameras mounted on a telescoping pole extended to between 8 and 12 feet. This managed to capture a property from every angle in a single pass, without a drone, in areas still closed to the public.
The work was not clean or easy. De la Peña recalled the difficulty of “trying not to step in glass or risk your own welfare while trying to use these new technologies” in burned-out lots still too hot and hazardous for normal foot traffic. Several team members, including ASU’s Jet Olaño, scanned the homes of friends and neighbors, which more than one described as the hardest part of the assignment.
De la Peña had her own reference point. A fire had destroyed her studio about 18 months before the LA fires, and when she filed an insurance claim for a couch lost in it, the insurer initially balked for lack of a receipt. A scan of the wreckage that captured a surviving metal armrest was, in her account, enough evidence that “they couldn't deny it.” That experience shaped how she talked about the project's second purpose. Beyond the documentary, the scans became spatial evidence survivors could use directly, at a moment when insurers were processing a wave of claims with no independent way to verify what had burned.
What began as an emergency documentation effort became, in de la Peña's words, an attempt to move “from archive journalism to actual major storytelling.”
Real Estate Moves Fastest
Zillow calls itself among the first companies to implement and scale Gaussian Splatting, and its SkyTour feature is one of the clearest large-scale proofs of the technology outside a research lab. Launched in July 2025, SkyTour turns drone footage of a listing into an interactive splat that shoppers can fly around from a phone or laptop, no app required. It runs only on Zillow Showcase listings where an agent has ordered drone footage through Zillow's premium media package.
Steve Anderson, a Zillow product manager, has described the goal as pushing splat rendering to run smoothly even on ordinary consumer hardware inside a browser.
“When people try SkyTour, their reaction is visceral – ‘I feel like I’ve been there.’ That’s incredibly satisfying from both a product and technology standpoint.”
That browser-native pipeline is also why the technology undercuts older 3D tour equipment on cost. A Matterport scan requires a dedicated device and a tripod setup at multiple points through a property. A splat can be captured with a smartphone or drone quickly, at a fraction of the equipment cost. This is why competing platforms built around Matterport-style scanning have started adding their own splat features rather than cede the format entirely.
Construction Sites and Digital Twins
Construction is where Gaussian Splatting is furthest past the demo stage. Skender, a Chicago commercial contractor with more than 70 years in the business, has used it on a 33-story West Loop tower, capturing an entire floor's rebar and post-tension cabling the afternoon before a concrete pour so the work is on record before it disappears behind concrete. Its construction technologist, Ben Stocker, has reported cutting the time to resolve site disputes by roughly 40 percent once a photorealistic, walkable model replaces a stack of photos and a point cloud only specialists can open. He says a full commercial building can now be captured in about 25 minutes, against the hours a traditional laser scan takes.
“We can share this out with all of our bidding subcontractors... Usually, there's a site walk where everyone has to go around and document what they can and then make their bids off that. But if we share this with everyone, they can more accurately create those bids for us."This can also be seen in Bentley's 2024 acquisition of the Philadelphia-area geospatial company Cesium, which already has a flagship project to show for it. Cesium built a Gaussian splat reconstruction of Microsoft's Redmond campus in Washington, stitching together 20,169 aerial photographs into a 3.7-square-kilometer tileset of 110 million splats at three-centimeter ground resolution.
Cesium says it can now monitor its assets consistently rather than relying on a one-time visualization. The idea is to rescan the same telecom tower, substation, or crane every few months and compare the new splat against the old one, catching corrosion, sagging cables, or vegetation encroachment before it turns into a failure. This is the kind of detail a mesh-based model tends to smooth away since it struggles to hold fine geometry on cables, power lines, and reflective surfaces the way a splat does. Because each tileset is georeferenced, it drops into the same coordinate system as a utility's existing GIS data, so an inspector is comparing the same physical point in space each time instead of reconciling two separately aligned models.
Museums Add Speed to Preservation
Cultural institutions are also utilizing the technology. At the LWL Open-Air Museum in Detmold, Germany, one of Europe's largest open-air museums, students from a local media institute used a drone and a handheld scanner to capture the site's courtyards and interiors, including areas normally closed to the public. They built an interactive exhibit from the results for a four-meter screen for the museum's new entrance building.
Heritage researchers are also testing the technique on damaged and fragile objects specifically. Cairo University researchers recently used Gaussian Splatting to digitize a Ptolemaic-era Egyptian statue, comparing it against laser scanning and photogrammetry, and found it to be the fastest of the three to capture while producing a textured, web-deployable model good enough for an interactive online museum viewer, at a moderate cost in geometric precision. That tradeoff, speed and accessibility over laser-scan-grade accuracy, is exactly what made the format usable inside a burn zone with a fire crew still working nearby.
Where This Leaves AV, XR, and Spatial Computing
For the AV and XR industry, the significant shift is not visual quality. Photogrammetry could already approximate that for static objects. What's different is the deployment path. A splat scene can be compressed, streamed into a browser tab, and rendered on commodity graphics hardware without a dedicated app, a game engine build, or specialized inference silicon. That is a materially lighter integration lift than NeRF or traditional volumetric capture.
Standards bodies are converging on compression formats, and headset makers are building native splat rendering into their runtimes. That combination will likely turn the technique into a standard capture layer, alongside 360 video and photogrammetry, for any spatial computing deployment that needs to render a real place in real time.
For the families in “Out of the Ashes,” the technology's real value has nothing to do with formats or frame rates. It is that a home nobody can walk back into now exists somewhere a person can. As de la Peña put it, describing what it means for a family to see their home again inside a reconstruction built from their own photos and memories, “It's really quite powerful.”
