Gaussian Splatting and Virtual Tours: What It Is, What It Replaces, and What It Doesn't

A prospective guest exploring a country estate wants to stand in the courtyard, turn around, and see the cottage fifty metres behind them. They don't want to click through fixed scan positions. They want to move. Gaussian splatting is how you give them that — and understanding what it is, what it can't do, and when to pair it with a traditional virtual tour is increasingly valuable knowledge for anyone commissioning property and venue content in 2026.

---

What Gaussian splatting actually is

Gaussian splatting is a method for building a navigable three-dimensional scene from a collection of overlapping photographs. The name refers to the mathematical primitives used to represent the scene: "Gaussians" — millions of small, semi-transparent ellipsoids, each with a position in 3D space, a colour, and an opacity. Together, calibrated to match the appearance of the scene from every captured angle, they form a photorealistic environment a viewer can move through in real time.

The foundational technique was published in 2023 by researchers at INRIA and the Max Planck Institute for Informatics (Kerbl, Kopanas, Leimkühler, and Drettakis, "3D Gaussian Splatting for Real-Time Radiance Field Rendering," ACM Transactions on Graphics, SIGGRAPH 2023. DOI: 10.1145/3592433). In the two years since, it has moved with unusual speed from academic paper to commercial deployment.

To understand why it matters, it helps to know its predecessor. Neural Radiance Fields (NeRF), introduced in 2020 (Mildenhall et al., ECCV 2020), showed that overlapping photographs could reconstruct photorealistic 3D scenes — but required significant compute time to render each new viewpoint. Gaussian splatting achieves comparable or better quality while rendering in real time, in a web browser, on consumer hardware. That shift from minutes per frame to real-time is what made commercial applications practical.

---

How it differs from a 360° virtual tour

A traditional 360° virtual tour — built from panoramas captured by a camera like the Realsee Galois M2 — is a structured sequence of fixed viewpoints. From each scan position, you can look in any direction. Between positions, you click to move. The experience is intentional, room-by-room, suited to interiors.

A Gaussian splat is structurally different. It is not a collection of viewpoints — it is a full three-dimensional model of a real space. The viewer does not jump between defined positions; they move through the scene continuously, in any direction, at any speed, from any viewpoint. There is no predefined path and no "next scan point". The space exists as a whole.

The practical consequence: a Gaussian splat communicates spatial relationships — how far apart buildings sit, how the garden wraps around the main house, what the grounds feel like to walk through — in a way that fixed-position scanning cannot replicate.

---

What a traditional virtual tour still does better

Gaussian splatting is a complement to the traditional virtual tour, not a replacement. A camera like the Realsee Galois M2 — capturing 134-megapixel panoramas at 16K resolution — delivers several things a Gaussian splat cannot match at present:

Detail at any zoom. A 16K panoramic scan renders cleanly when a viewer zooms into artwork, signage, or material texture. Gaussian splat models at current quality levels degrade under close inspection of fine detail.

Structured, floorplan-linked navigation. Interactive hotspot markers on a floorplan — click a room, jump to that scan position — are only possible with a positionally anchored tour. Gaussian splats do not map to room-by-room structured navigation.

Measurement data. LiDAR-equipped virtual tour cameras produce point clouds suitable for as-built surveys, space planning, and CAD output. A Gaussian splat is a visual representation, not a measured environment.

Indoor legibility. For interiors, most viewers benefit from a structured walkthrough. A guest inspecting a hotel suite wants to see it clearly from a defined position, not navigate through it as a free-floating 3D cloud. The panoramic tour is more intuitive for enclosed spaces.

Broad device compatibility. Traditional virtual tours run reliably across all devices. Gaussian splatting is computationally heavier and, while increasingly well-optimised, performs best on modern hardware.

Floorplan-linked navigation — click any room marker to jump directly to that scan position

---

When to use each — and when to use both

---

The White Heron Estate example

White Heron Estate in Lyonshall, Herefordshire — a multi-building luxury estate with multiple distinct properties across a single site — illustrates why the hybrid approach matters. Photography could show each space. It could not show prospective guests how the buildings related to one another across the grounds, or help a group booking understand which accommodation sat where.

See3D produced a virtual tour for the estate combining both methods: a Gaussian splat of the external grounds, freely navigable from any viewpoint, with hotspot markers at each building entrance linking through to the full interior 360° tour for that building — complete with interactive floorplans. A guest can stand in the courtyard, explore the grounds spatially, then click into The Forge, The Colloquy, or Gardeners Cottage for the detailed interior walkthrough and room-by-room navigation.

The estate's tour is live at tours.see3d.ai/WhiteHeron.

---

Structure from Motion — and why capture method matters

Gaussian splatting relies on Structure from Motion (SfM): a photogrammetric process that determines the 3D position of points in a scene by analysing how the same features appear across many photographs taken from meaningfully different angles. The more varied the camera positions, the stronger the 3D reconstruction.

This is where traditional virtual tour cameras present a challenge. A camera like the Realsee Galois M2 is designed to capture complete 360° panoramas from fixed tripod positions. Each scan position captures the scene fully from a single point in space — but SfM needs parallax: the same subject seen from different positions to triangulate depth. A static panorama at a single height gives one data point in 3D space; it does not provide the varied angular coverage SfM requires.

Gaussian splatting with a traditional virtual tour camera is achievable — but requires a substantially higher number of scan positions, at varied heights and locations, to generate the parallax needed for a robust reconstruction. For large open outdoor spaces, dedicated capture methods work considerably better:

• 360° action cameras — cameras like the Insta360 X5 capture full-sphere imagery continuously as the operator walks through a space, generating overlapping frames with natural parallax from constant motion. The X5's dual 1/1.28-inch sensors and 8K/30fps capture make it a practical ground-level option for Gaussian splatting data collection.
• 360° drones — aerial platforms dramatically expand coverage for exterior spaces. The Antigravity A1 (launched December 2025) captures 8K 360° footage from a 249g drone invisible in the final output, suited to estate grounds and architectural exteriors. The DJI Avata 360 (launched March 2026) adds 8K/60fps 360° capability to DJI's FPV platform with a 20km transmission range — a significant development for aerial Gaussian splatting capture. Both produce the varied viewpoints from multiple angles and heights that SfM requires.

The proliferation of accessible 360° capture hardware is one of the primary factors making Gaussian splatting practical at commercial scale.

---

Where the technology is going

The 2023 SIGGRAPH paper has since generated a substantial body of follow-on research — improving rendering quality, reducing model file sizes, enabling dynamic scenes, and integrating Gaussian splatting into commercial software pipelines.

Real-world deployment is accelerating. In July 2025, Zillow launched SkyTour — a feature on its Showcase listing tier that uses Gaussian splatting to give property browsers a navigable 3D exterior view generated from drone footage, describing itself as "among the first in the industry to implement and scale Gaussian splatting technology" for real estate. NVIDIA's Omniverse NuRec, announced at SIGGRAPH 2025, integrates Gaussian splatting into professional simulation and robotics workflows. Luma AI — a commercial platform for Gaussian splatting capture and delivery, backed by over $1 billion in total funding including a $900M Series C in November 2025 — has built browser-embeddable tools that run on standard hardware. In heritage preservation, peer-reviewed research published in Frontiers in Computer Science and npj Heritage Science in 2025 has documented the use of Gaussian splatting to digitise damaged artefacts and historic structures.

The near-term direction for property applications is higher quality at smaller file sizes — making the technology practical across a wider range of devices and connections. The longer-term possibility is the integration of interactive elements — floorplan overlays, measurement tools, data labels — directly within the splat model, which would close much of the functional gap with traditional tour platforms.

Platforms at the frontier

Some practitioners are already building toward this convergence. Olli Huttunen — a Finnish developer who has become one of the more visible pioneers in applied Gaussian splatting — is developing Dioramix, a narrative editor for 3D Gaussian splat models. The description is precise: "like PowerPoint in a 3D world." Dioramix allows creators to define camera viewpoints, build button-based navigation between views, place annotations on specific positions in the scene, and construct structured interactive presentations around real-world 3D scans — all running in a browser via Babylon.js and WebGPU. The platform supports multiple formats including Gaussian PLY and the compressed SOG format, and the community interest it has generated points toward database integration, measurement tools, and AI-rendered outputs as the next logical steps.

This is the kind of platform-level thinking that indicates the industry's direction of travel: not just photorealistic spatial capture, but structured, navigable, annotated experiences built on top of Gaussian splat models. 3DVista — whose Virtual Tour Pro platform is one of the more capable tools in the traditional 360° tour space — shipped native Gaussian splatting support in their 2025.0 update: users can now import splat models directly into a Virtual Tour Pro project and navigate them in Fly Over, Orbital, or First Person modes, with interactive hotspots and precise measurement tools embedded within the splat. The models render at up to ten times smaller than equivalent GLB files. That is a commercial tour platform that has absorbed Gaussian splatting as a first-class feature, not an experimental add-on. The gap between "a Gaussian splat" and "a virtual tour built on a Gaussian splat" is narrowing fast.

---

The ecosystem: tools, resources, and practitioners

Behind the academic papers and commercial deployments is a practitioner and developer community that has contributed tools, resources, and open infrastructure at a speed that has kept pace with the research itself. The following are worth knowing.

Staying informed

Radiance Fields, run by Michael Rubloff, is the resource to bookmark if you want to keep pace with developments across the whole field — new papers, software releases, platform updates, hardware reviews. As Founder and Managing Editor, Rubloff has built the closest thing the 3DGS and NeRF community has to a dedicated trade publication: consistently covering new research and practitioner workflows with enough technical depth to be genuinely useful. He has also developed Gaussian King — an Apple app for iPhone and iPad that brings Gaussian splatting to anyone with a modern iOS device, removing the GPU compute barrier that has historically made the technology the preserve of desktop hardware.

Free training tools

Brush, built by Arthur Brussee, is a real-time Gaussian splatting training tool that runs across Windows, macOS, Linux, Android, and in a browser — freely available, actively maintained, and designed from the ground up to make training approachable rather than the preserve of high-end compute infrastructure. Lichtfeld Studio, built by Janusch Patas, takes a similar ethos: a free, evolving desktop application for training and working with Gaussian splat models, developed with genuine care for the practitioner community. Both are volunteer-led projects. In a landscape where commercial Gaussian splatting platforms have attracted substantial venture investment, the continued existence and improvement of capable free tools is largely down to individual contributors giving their time. If you are a business, agency, or developer who has benefited from either — or who can see the value in backing industry-leading open-source tooling before it becomes infrastructure everyone depends on — 2026 is a good time to lend your support.

Extending the tools

What distinguishes Lichtfeld further is its Python plugin architecture, which allows third-party developers to extend the platform — and that ecosystem is already producing meaningful results. Shady Gmira has developed a densification plugin for Lichtfeld that directly addresses one of the most persistent practical challenges in Gaussian splatting: the sparse dataset. When a capture session yields limited or unevenly distributed photographs, the initial SfM reconstruction is thin — not enough correspondences to triangulate dense geometry, and the training that follows begins from an impoverished foundation. Gmira's plugin runs a densification pass on that sparse reconstruction using RoMa v2 (Edstedt et al., arXiv:2511.15706), a dense feature matching algorithm that identifies correspondences where conventional methods fail. Those matches generate additional 3D points before training begins — effectively filling in the geometric gaps. Measured results: +0.63% PSNR and +1.58% SSIM over non-densified training, converging 13% faster. For practitioners working with constrained capture conditions — incomplete coverage, difficult access, fast-moving subjects — this is the difference between a workable reconstruction and an abandoned one.

Automating the pipeline

ronskiuk — a UK-based developer with real depth of practical knowledge in the field — has built a 360 Video – Stills Prep Tool that addresses the step most practitioners find unglamorous but indispensable: preparing source material before training begins. Converting 360° video footage into stills suitable for SfM sounds straightforward, but it is easy to get wrong — frame selection, spatial distribution, coverage density, and image quality all materially affect the reconstruction that follows. Ronskiuk's tool automates this preparation stage, handling the frame extraction and optimisation workflow that would otherwise demand both technical understanding and considerable manual effort. At $20, it is priced to be accessible rather than profitable — a 4.8/5 rating across 19 reviews suggests it is doing what it claims. Going considerably further than the typical "vibe coder" approach, it reflects the kind of applied expertise the field needs at every level of the stack.

Practitioners at the front

Gradeeterna — a London-based creator — is among the most compelling examples of what applied Gaussian splatting looks like in the UK right now. Capturing extraordinary locations with the Insta360 X5, training models in Brush, and navigating the results in real time, their output consistently demonstrates the full pipeline from capture to experience — and the visual ambition on display is genuinely striking. In this author's view, they are the leading practitioner in the UK producing work at this level. Their YouTube channel is essential viewing for anyone serious about understanding where hands-on Gaussian splatting is heading.

If you work in or adjacent to this space, 2026 is a good time to pay attention to the people building it from the ground up. The tools above are not built on salary; they are built on time. In a field moving this fast, that is not something to take for granted.

---

Gaussian splatting will not replace the panoramic virtual tour. The two serve different purposes, and for properties where the spatial relationship between spaces is part of what you're selling, they are more valuable together than either is alone. See3D offers Gaussian splatting as part of our virtual tour service — for estates, venues, and any project where spatial context is as important as interior detail. Get in touch to discuss whether a hybrid approach is right for your project.