Gaussian Splatting in Shared VR
TUM-XR Student Initiative

Heavy scans. Standalone headsets. Full-detail Gaussian splatting explored together in VR, powered by Hololight's XR pixel streaming technology.

The TUM-XR student initiative, a student-run organization at the Technical University of Munich, set out to bring real, laser-scanned environments into virtual reality as explorable Gaussian splatting scenes. Working with Hololight's XR pixel streaming technology, the team moved detailed 3D captures off the limits of standalone headset hardware and into shared VR sessions, where several people can explore the same space at once.

• Full-detail Gaussian splatting scenes rendered on a workstation and streamed to standalone headsets, with scene size bounded by the workstation rather than the device
• Up to 3 people in one shared session, exploring the same scene together with no crashes or disconnects
• Around 90 FPS at roughly 7 million splats, where interaction felt natural and latency was not noticeable
• A longest uninterrupted run of about 20 minutes with 3 simultaneous users
• Well over 50 test sessions across single-user and shared setups

The Challenge

Gaussian splatting turns real spaces, captured here as 3D scans with Xgrids, into dense, photorealistic scenes made of millions of small 3D elements called splats. That density is what makes the scenes look convincing, and it is also what makes them heavy. A standalone headset has to store and render everything locally, and past a certain scene size it runs out of room to do that smoothly.

"We wanted to bring real spaces into VR, but the models were simply too heavy for the headset. It just wasn't technically possible."
– Joel Christlein, Founder & Head of Projects Department, TUM-XR Student Initiative

For the team, the goal was never a stripped-down preview. They wanted to stand inside the actual captured environment, a real space of roughly 30 by 30 meters, at full detail. The headset alone could not deliver that.

The Solution

Hololight's XR pixel streaming technology moves the rendering work off the headset and onto a workstation. The application runs on the workstation's GPU, and only the finished image stream is sent to the headset wirelessly. The headset no longer needs to hold or process the full scene, so scene size is bounded by the workstation, not the device.

The workflow connects directly to the OpenXR output of standard 3D tools such as Blender. Pressing play sends the scene straight into the headset, with no separate export step.

"Suddenly we were standing in the room we had only seen as a point cloud on the screen."
– Jakub Majewski, Senior Student Associate, TUM-XR Student Initiative

Setup and Testing

The team streamed from a single workstation to Meta Quest headsets over Wi-Fi 6, in both single-user and shared sessions.

They ran the same larger scene at several optimization levels. That let them compare performance cleanly across scene complexity, rather than across unrelated scenes.

Results & Performance

Across testing, the workflow handled medium-sized Gaussian splatting scenes comfortably. Smaller scenes ran smoothly with no extra optimization. The larger scene, the one shown in the initiative's video, needed some optimization first, and its behavior across detail levels mapped out a clear practical range.

The pattern was consistent: the practical ceiling came from the workstation's rendering capacity and scene complexity, not from the headset. Around 7 to 9 million splats marked the comfortable working range for smooth, shared exploration.

In shared sessions, up to three people explored the same scene at once with no crashes or disconnects, including one uninterrupted run of about 20 minutes with all three users active.

"You forget it's a model. You just discuss inside it like it's real."
– Abhishek Dubasi, Junior Student Associate, TUM-XR Student Initiative

Key Takeaways

• Streaming moves the rendering ceiling off the headset. The device only displays a stream, so scene size is limited by the workstation rather than the headset's onboard hardware.
• Medium Gaussian splatting scenes run smoothly in shared VR. Around 7 to 9 million splats delivered natural interaction and low perceived latency in the initiative's tests.
• Several people can explore one scene together. Up to three simultaneous users shared the same environment reliably across dozens of sessions.
• The practical limit is workstation-side. Very large scenes were bounded by rendering capacity and complexity, which leaves room to grow as workstation hardware scales.

Conclusion

For a student initiative preparing for careers in spatial computing, the value is not only the finished demo but the hands-on workflow behind it: capture a real space, turn it into a Gaussian splatting scene, and step inside it together at full detail. Hololight's XR pixel streaming technology made that possible by treating scene complexity as a workstation question rather than a headset limit. As scans grow larger and more detailed, the same approach leaves room to grow with them.