Remote rendering and pixel streaming are both used to describe enterprise XR streaming infrastructure, including by vendors and in technical documentation. They are not two competing technologies. Remote rendering is the category, meaning the rendering happens somewhere other than the headset. Pixel streaming is one implementation inside that category.
Knowing which of the two labels a product carries tells you very little. The question that decides whether a product fits an enterprise requirement is simpler, and neither term answers it: does your model stay on the render host, or does it travel to the headset?
Three approaches compete for the same job of getting a heavy model into a headset. Two of them are remote rendering. The third, geometry streaming, is often classified as remote rendering because the model is stored and served remotely, but the headset is what renders it.
What you will find here
Three ways a model reaches a headset
| Factor | Pixel streaming | Hybrid rendering | Geometry streaming |
|---|---|---|---|
| Is this remote rendering? | Yes | Yes | No. The headset does the rendering. |
| Does the model stay on the render host? | Yes | Yes | No. It is copied to every device that opens it. |
| What crosses the network | Encoded pixels only | Encoded pixels, plus content the device draws itself | The 3D model |
| Where the frame is assembled | Entirely on the server | Server output and device output are combined on the device | Entirely on the device |
| What happens to the source model | Nothing. It is rendered as authored. | Nothing. The remote portion is rendered as authored. | Converted in advance into tiles at several levels of detail |
| What limits model size | The server you choose | The server, plus what the device can composite | The headset's onboard hardware |
| Worked example | Hololight's XR pixel streaming technology | Azure Remote Rendering, retired September 2025 | 3D Tiles |
Remote rendering describes where the work happens, not how the result arrives. The GPU doing the rendering sits somewhere other than the headset. That is the whole definition.
Pixel streaming is one way to satisfy it. The server renders the complete frame and sends encoded pixels. The headset decodes the image and displays it, while the underlying data stays on the server and never touches the headset.
So the two terms sit at different levels, because while all pixel streaming is remote rendering, not all remote rendering is pixel streaming. Asking which of the two a product uses is the wrong question, since one contains the other.
The term "remote rendering" does not tell you:
One question separates the options available to you, and it is not about where the rendering happens. It is about where the model ends up.
In some setups the model stays on the render host for the whole session and the headset receives only an image. In others the model is sent to the headset and drawn there. That difference decides whether design files leave your environment, and whether the model can be reviewed at the complexity it was authored at.
Two approaches keep the model on the host. Pixel streaming renders the complete frame on the server and sends encoded pixels. Hybrid rendering splits the job, so the server renders the heavy content, the headset renders its own layer such as the user interface, and the two are combined on the device before display. Both are remote rendering.
The third approach, geometry streaming, sends the model itself. It belongs in this comparison because it is the one most easily mistaken for the other two, and the name is the reason. A server holds the model and delivers it in pieces over the course of a session, which looks like streaming from the outside. Nothing is rendered remotely and no video stream is involved, so despite the label it is not remote rendering at all.
Where the model ends up
| Approach | Where the model ends up | What the headset receives |
|---|---|---|
| Pixel streaming | Stays on the render host | A finished image |
| Hybrid rendering | Stays on the render host | An image, which it completes with its own layer |
| Geometry streaming | Copied to the device | Geometry, which it renders itself |
With Hololight's XR pixel streaming technology, the application runs and is rendered on server hardware, and the encoded image is sent to the headset. The source model is never converted and never reaches the device, so it can be reviewed at full complexity and the data stays where it already sits. The render host can be a workstation, an on-premises server, or a cloud instance, which is a product choice rather than a property of the approach.
Azure Remote Rendering is the clearest documented example. Microsoft's documentation describes a service that rendered high-quality interactive 3D content in the cloud and streamed it in real time to devices such as HoloLens 2, while the headset drew interface elements locally and combined the two images before display. The problem it addressed was fidelity, and Microsoft illustrates the gap with one model at more than 18 million triangles against roughly 200,000 once decimated. Microsoft retired the service on September 30, 2025.1,2
3D Tiles is an open standard for streaming large 3D datasets, approved by the Open Geospatial Consortium in 2018 and covering BIM and CAD content alongside photogrammetry and point clouds. The model is cut up in advance into tiles at several levels of detail, and during a session the device requests only the tiles it currently needs, asking for finer versions of whatever the user is looking at. It stops refining once the detail is sufficient or once the hardware runs out of room, so the device rarely holds the whole model but always holds real geometry. Preparing CAD data for this kind of real-time display is a known and error-prone step in its own right.3,4
The distinction becomes concrete the moment you write requirements down.
Start with data. If design files have to stay inside your own environment, the approach decides whether that is possible at all. Server-side rendering sends images, so the model never moves. Geometry streaming sends the model, so it does.
Then fidelity. A model rendered on the device is limited by the device, which is why it is reduced beforehand. A model rendered on a server is limited by the server, and you choose the server.
Then connectivity. Server-side rendering needs a reachable render host for the whole session, because every frame comes across the network. Geometry streaming also pulls data during a session, but what it has already received stays on the device, so a poor connection degrades detail rather than stopping the session. Where the host sits, in a cloud or in your own server room, is a product question rather than a question of approach.
Finally, the fleet. A headset that only displays an image needs to do nothing except decode video, so one setup covers everything from an older device to the newest one. A headset that renders has to be powerful enough for your model, which means every new device class has to be tested and every older one may fall short.
Requirement to approach
| Requirement | What it rules in | What it rules out |
|---|---|---|
| Design data must not leave your infrastructure | Rendering the frame on hardware you control | Any approach that sends the model to devices |
| Review the model as authored, with no reduction | Rendering on hardware you size yourself | Rendering bound by the headset |
| No network between headset and render hardware at all | Rendering on the device | Both server-side approaches |
| Mixed headset fleet that will keep changing | Headsets that decode and display | Headsets that each have to render the model |
A product can describe itself accurately and still leave the approach unclear, since the same terms cover all three. Five questions settle it, and they are worth asking in exactly this order.
Questions to ask a vendor
The first answer tells you most of what you need. Geometry arriving at the headset means the headset is rendering it, with the limits that come with that.
With Hololight's XR pixel streaming technology, pixels cross the network. The source file is not converted or reduced. The render host can be a workstation, an on-premises server, or a cloud instance, including on an isolated network. No copy of the model exists on the headset during a session or after it, and the headset renders no part of the scene. Those answers are why teams in manufacturing, automotive, aerospace, and defense choose Hololight.
Remote rendering tells you that the application is not rendered on the headset, and little else. Inside the category, pixel streaming and hybrid rendering both keep the model on the render host, and they differ mainly in whether the headset draws anything itself. Outside it, geometry streaming sends the model to the device, which is the one approach where design files leave the machine they started on. Ask the five questions above and you will know which you are being offered.
See how enterprise XR streaming infrastructure is deployed
Last Updated: October 1, 2026
Sources