A Gaussian splat has no surface. That one fact is why splats, four years after they arrived, still sit in XR applications as pretty inert scenery: no shadow falls from them, nothing collides with them, and nothing can be occluded by them, because every technique for those things wants a polygon to ask about.
SplatXR: A Toolkit for Multi-Layered Extended Reality via On-Headset 3D Gaussian Splatting (arXiv:2610.26338) is an attempt to close that gap in Unity. It was posted on 19 October 2026 by Cyrus Vachha, Mohamed Kari and Parastoo Abtahi, all at Princeton University, and the comments field places it at ACM UIST 2026.
Multi-layered reality
The framing matters as much as the toolkit. Conventional XR treats the physical and the digital as two separate spaces that get composited. SplatXR proposes organising XR instead as a multi-layered reality of interacting primitives, in which pixels and objects can cross the physical/digital boundary in either direction.
That is not just rhetoric when 3D foundation models can turn a physical object into a splat on the fly. If a thing on your desk can become a splat during the session, “physical” and “digital” stop being properties of objects and become properties of the moment.
The building blocks
The project page describes a Unity editor panel from which these are dropped into a scene:
- Splat Renderer — on-device rendering, from a single object up to LoD-managed scenes
- Splat Physics — generates collision meshes from splats, enabling rigid-body dynamics and shadow casting
- Splat Effects — shader-based disintegration, radial reveal and distortion
- Splat Blending — portal-style effects showing splat spaces through virtual windows blended with the real world
- Offloaded splat object reconstruction via on-headset capture
Splat Physics is the one that earns the paper. Deriving a collision mesh from a splat cloud is the pragmatic answer to the no-surface problem: you do not make the splats physical, you grow an invisible proxy that is, and then shadows and collisions work because they are talking to geometry again. The portal blending is the obvious right move for splats’ real weakness, which is that they look superb from the capture envelope and terrible from outside it — a window crops the view to where the data is good.
The paper reports demonstration applications and a small developer study for validation.
The part to be clear-eyed about
The project page describes SplatXR as an open-source Unity toolkit under the MIT licence, says the code is available on GitHub, links documentation — and carries a form to be notified on release.
We went looking. There is no SplatXR code repository under Cyrus Vachha’s GitHub account; the only thing that exists is splatxr/splatxr.github.io, the project page itself, created on 8 July 2026 with zero stars. A GitHub search for “SplatXR” returns exactly two repositories, the other being an unrelated MSc group project from Trinity College Dublin — yash-jain221/SplatXR, created April 2026, an end-to-end splat capture pipeline across Unity and Unreal Engine 5 evaluated on Quest 3. Same name, different work, two stars.
So the MIT licence is a stated intention and the toolkit is not yet something you can download. That is normal for a UIST paper landing before camera-ready, and worth saying plainly rather than reporting a release that has not happened. Vachha’s track record is a reason to expect it to arrive: he is behind instruct-gs2gs, the Nerfstudio-ecosystem instruction-based splat editing project, which sits at 135 stars.
Neither the paper nor the project page names a headset model, states a frame rate, or gives any performance number — notable for a system whose entire premise is running on-device, and the first thing to ask for when the code lands. The page confirms only that it uses video-passthrough mixed reality.
Sources
- SplatXR: A Toolkit for Multi-Layered Extended Reality via On-Headset 3D Gaussian Splatting — arXiv:2610.26338
- SplatXR project page
- ACM UIST 2026
- instruct-gs2gs
- Cyrus Vachha