Cesium has added support for 3D Gaussian splats across CesiumJS, Cesium for Unreal and Cesium ion, with level-of-detail streaming built directly into the 3D Tiles standard. The update lets teams stream photorealistic 3D captures — from an entire city down to sub-centimetre detail — without loading a whole dataset at once.
From polygons to points of light
Gaussian splatting reconstructs a scene from ordinary photographs or video, but instead of building a mesh of triangles and stretching textures over it, the algorithm produces millions of small volumetric points, each carrying its own position, colour, opacity and orientation. Rendering blends these points together in real time, which is why the technique — first described by researchers at Inria in 2023, according to Utsubo — has moved quickly from academic paper to production tool.
The practical gain over photogrammetry shows up exactly where mesh-based capture struggles most: cables, foliage, glass, water, fire, hair. Cesium points out that spherical harmonics embedded in each Gaussian let the renderer reproduce view-dependent reflections and anisotropic surfaces that a static texture cannot fake. Utsubo frames the same point from the other end of the pipeline: splats combine the visual fidelity long associated with neural radiance fields with rendering speeds fast enough for a browser, something NeRF itself never managed.
What actually changed with this release
Splats have existed as research output and as isolated web viewers for a couple of years. What Cesium’s announcement adds is infrastructure: 3D Tiles now acts as a spatial index for splat data, and glTF — through two new extensions, KHR_gaussian_splatting and KHR_gaussian_splatting_compression_spz — acts as the payload format. In practice, a splat dataset can be broken into tiles of increasing detail and streamed only where and when a viewer needs it, the same logic already used for terrain and photogrammetry meshes in Cesium’s ecosystem.
File size, historically the main obstacle to using splats outside a demo, is addressed through SPZ, an open-source compression format from Niantic Spatial. According to Cesium, SPZ reduces splat data by up to 90 percent compared to standard PLY files while preserving the spherical harmonic information needed for realistic lighting. Its most recent release also improves how thin linear features — antennas, power lines, scaffolding — are encoded, which matters directly for infrastructure inspection work.
Interoperability, not just a plugin
The extensions were developed jointly by Cesium, Khronos Group, the Open Geospatial Consortium, Esri and Niantic Spatial, and splats are being folded into the proposed 3D Tiles 2.0 community standard. Khronos president Neil Trevett described the collaboration as a model for how “new rendering techniques should enter the industry: quickly, openly” — worth taking seriously, since it means splats are being standardised before they fragment across incompatible proprietary viewers, roughly what happened with early photogrammetry tools.
Cesium ion now offers an end-to-end pipeline too: source photos go in, and the platform automatically reconstructs a georeferenced 3D Tileset as mesh, point cloud or Gaussian splat, ready to sit alongside existing datasets such as Cesium World Terrain or Google’s Photorealistic 3D Tiles.
Where the technique is already working
Utsubo’s overview places real estate furthest along: platforms such as Zillow and Apartments.com have shipped splat-based property tours in 2026, replacing the “teleporting” feel of stitched 360° photography with continuous walkthrough movement, alongside adoption in e-commerce, film VFX and volumetric video. Cesium’s list points the other way, towards infrastructure monitoring, telecom towers, electrical substations and long-term situational awareness, where the value lies in tracking how a physical asset changes over its full life cycle, not in a walkthrough experience.
Museums and archaeological sites sit at the intersection of both lists: they need faithful capture of complex, often fragile surfaces, and they increasingly need to monitor how a site changes over years, not just present a single snapshot to visitors.
What this means for immersive projects in Italy
For a museum, an archaeological park or a municipal technical office, LOD-enabled splat streaming changes what is realistic to promise a client. Until now, a photorealistic digital twin of an extended site — a whole excavation, a historic centre, a stretch of coastline — meant either a heavily simplified mesh or a dataset too large to deliver over the web at usable frame rates. Tiled splat streaming removes that trade-off: a visitor’s device loads the detail relevant to where they are standing, whether that is a city-scale overview or the texture of a single carved surface, without downloading the whole site first.
This opens a category of project that was previously a hard sell: territorial digital twins that are actually navigable in a browser or headset, rather than pre-rendered flythroughs. An archaeological park with several hectares of standing structures, or a municipality that wants a living record of its historic centre for conservation and tourism purposes, can now be captured, tiled and published as one coherent, explorable dataset — and updated over time to track deterioration, restoration work or seasonal change, essentially the “situational awareness” use case Cesium describes for infrastructure, applied instead to heritage.
The opportunity comes with a genuine change in required skills, and it would be misleading to present this as simply plugging a new viewer into an existing pipeline. Capturing splat data well still depends on careful photo or video acquisition, and the reconstruction, tiling and compression steps introduce a data-management layer that most cultural or public-sector teams have not had to handle before: choosing compression settings, deciding how much detail is worth streaming at which distance, keeping datasets georeferenced correctly so they sit alongside terrain or existing GIS layers. None of that is exotic, but none of it is optional either — a splat dataset published without that discipline degrades quickly into an oversized, slow-loading curiosity rather than a usable tool.
There is also a standards question that public bodies, in particular, should not ignore. The fact that splats are being formalised inside glTF and 3D Tiles as open, royalty-free extensions rather than a proprietary format is what makes long-term investment defensible: a public administration commissioning a decade-long digital twin needs to know the data will still be readable in ten years, by more than one vendor’s viewer. That guarantee did not exist eighteen months ago. It exists now, at least on paper, and it is the detail that should shape how any public tender for this kind of work is written from here on.