Gaussian Splatting vs. Photogrammetry

Key Takeaways

• Gaussian splatting and photogrammetry begin from the same overlapping photos but produce fundamentally different kinds of output.
• Splats render faster and look smoother because they blend through gaps rather than reconstructing an actual surface.
• A splat scene cannot be exported as a mesh or measured, since it has no reliable scale and no watertight surface.
• Moving a camera 10 cm in reality does not translate to 10 cm inside a splat, which rules the format out for survey work.
• Photogrammetry in Pixpro remains the route to measurable output, and the same roof scan was processed both ways for comparison.

Every few years, something new shows up online claiming to “revolutionize” 3D reconstruction. A while back it was NeRFs. Now it is Gaussian splatting — with endless YouTube clips of people walking through photorealistic 3D scenes that seem to render instantly.

It looks impressive, and the hype is high. However, if you already work with photogrammetry or scanning, you may be wondering: what exactly is this? Is it geometry? Can I measure it? Or is it just a visual rendering technique? Let me break it down without the math, just in plain language — what Gaussian splatting really is, how it works, and where it makes sense.

What Gaussian Splatting Actually Is

At its core, Gaussian splatting is a method of representing a 3D scene using “blobs” of color and transparency, rather than a solid mesh or a traditional point cloud. (Wikipedia refers to these as “primitives”; I would like to call them types of vectors.) Each blob or splat  is a small 3D Gaussian (a fuzzy sphere of color). Together, millions of these splats approximate how the scene looks from any direction, thus enabling you to navigate it in 3D.

We can think of it like painting space with semi-transparent colored dots that overlap and blend smoothly. When the system learns where to put those dots and how to color them, the scene appears 3D, continuous, and surprisingly realistic — even though it is not made of real geometry.

This approach is derived from Neural Radiance Fields (NeRFs), which also learn how light and color behave from different viewpoints. The key difference is that Gaussian splatting skips the slow “neural rendering” part. Instead, it turns that learned radiance field into discrete points that can be rendered extremely fast on a GPU. That’s why it looks so fluid in demos: it is optimized for display speed and visual fidelity, not for measurements.

Why It Looks Good

The look of splatting comes from how it handles light and color transitions. Instead of hard surfaces, it blends color and transparency smoothly across thousands of overlapping splats. This makes edges soft, lighting natural, and surfaces appear continuous, even when the data is incomplete.

It also avoids a big limitation of photogrammetry: texture seams and missing geometry. If an area has poor overlap or reflective surfaces, photogrammetry might fail to reconstruct it. Gaussian splatting simply “blends through” the gaps, which is visually forgiving even if it is inaccurate. This is why splatting demos often look so cinematic.

Gaussian Splatting Is Not Replacing Photogrammetry

You cannot export a Gaussian splat scene as a 3D polygon mesh file and measure distances. There is no reliable scale, no watertight mesh, and no consistent surface. Even if you somehow aligned it with GPS data, the reconstruction itself is based on appearance fitting, not triangulation.

That means accuracy is uncontrolled. If you move the camera 10 cm in real life, that does not necessarily translate to 10 cm in the splat world. There is no referencing by default.

This matters a great deal for practical work, such as architecture, mapping, inspection, or any task where measurements and consistency across sessions are crucial. In those cases, photogrammetry still reigns supreme because it is mathematically constrained by geometry.

Where It Actually Shines

Gaussian splatting is optimized for different goals:

• Virtual environments and games: instant, realistic environments for VR/AR, background sets, and walkthroughs.

• Cultural heritage visualization: quick immersive displays where accuracy is less important than presence.

• Preview tools: imagine capturing a site and instantly walking through it visually, before running a full photogrammetry process.

These are areas where the feel of the scene matters more than millimeter precision.

In fact, I can envision future tools utilizing splatting as a real-time preview layer, providing a live, smooth visualization immediately after capture, while photogrammetry runs in the background to produce the final, accurate model.

Our Example

Data for reconstructing splats is fundamentally the same - overlapping photos or video frames. I used our roof scan example to compare how a photogrammetric 3D reconstruction compares to a Gaussian splatting one.

This Gaussian splat example was created using Polycam's service, the most accessible web service available at the moment. For more tech-savvy users, the Gaussian splatting code is available free of charge and is open source. You can run it locally to try things for yourself.

This example is from Pixpro photogrammetry software, a standard 3D reconstruction process with all the necessary steps. You can see the camera positions and different layers as usual.

The differences are quite clear; the aesthetic of the Gaussian splatting environment is unique and interesting, in my opinion. While the photogrammetric 3D model is what we are quite used to.

The practical split is about what happens after capture. A splat is finished once it renders, whereas Pixpro photogrammetry software turns the same photos into a scaled point cloud, a mesh and an orthophoto you can take measurements from, which is what inspection, volume and solar work actually require. Flight planning for that capture runs through Pixpro Waypoints.

Conclusion

Gaussian splatting is an interesting piece of technology - fast,interesting, and full of potential. It is a rendering technique that prioritizes realism and visuals over precision.

Photogrammetry remains the standard for accurate, measurable 3D reconstruction, while splatting opens the door to smoother, more interactive visualization. The most exciting future is likely a blend of both: the reliability of geometry with the fluidity of AI-driven rendering. We will see where this takes us and take a second look in due time.

Frequently Asked Questions

What is Gaussian splatting?

A rendering technique that represents a scene as millions of overlapping three-dimensional blobs, each carrying position, colour and transparency. It grew out of neural radiance field research but skips the slow neural rendering step, which is why it displays in real time. The input is ordinary overlapping photos or video frames, exactly as photogrammetry uses.

Can you measure a Gaussian splat?

No, and this is the decisive limitation. There is no reliable scale in a splat scene, no watertight mesh and no consistent surface to measure against. Moving the camera 10 centimetres in reality does not correspond to 10 centimetres in the reconstruction, so distances, areas and volumes cannot be extracted with any confidence.

Is Gaussian splatting better than photogrammetry?

Neither replaces the other, because they answer different questions. Splatting wins on visual fidelity and speed of display, making it strong for virtual reality, cultural heritage and instant previews. Photogrammetry in Pixpro wins wherever the output must be measured, exported as a mesh, or defended as accurate to a client.

What is Gaussian splatting actually used for?

Applications where appearance matters more than measurement. Virtual and augmented reality scenes, cultural heritage documentation, and quick previews immediately after capture all suit it well. It also works as a real-time preview layer while a full photogrammetric reconstruction processes in the background, which is probably where the two techniques converge.

Do Gaussian splatting and photogrammetry use the same photos?

Yes, the capture requirement is essentially identical. Both need overlapping images or video frames covering the subject from multiple angles, so one dataset can be processed either way. That is why the roof scan shown here could be reconstructed as a splat and as a measurable model without reshooting anything.