When looking at the two technologies from the outside Gaussian Splatting and photogrammetry, the simple – or perhaps, let's say, a little naive – question almost inevitably arises:
What is actually better now?
Of course, we all know that such questions are rarely easy to answer in reality. Nevertheless, it is valid – because why should one engage with a new technology at all if it is not superior to the tried and tested, at least in certain aspects?
We started our small practical experience with precisely this question.
We had the opportunity to digitize a special architectural model, approximately three meters in size, which shows the never-realized design for a Munich opera house for Richard Wagner (currently on display at the Herrenchiemsee Museum).
A perfect opportunity to combine both methods – photogrammetry, with which we already have a lot of experience, and the still quite new Gaussian Splatting with the Lixel K1 LiDAR Scanner – to test them directly side by side and honestly compare them.
And yes, we also wondered at the beginning:
Does it even make sense to test this new workflow, or is photogrammetry ultimately the safer option anyway?
As is so often the case in such technological comparisons, the answer is:
It depends.
There are strengths and weaknesses on both sides—and that’s exactly what we’d like to balance out in detail and in a practical way in this blog post.
Our goal: Not a black-and-white answer, but a well-founded basis for decision-making on when which procedure is sensible for which application.
A Direct Comparison of Admission Processes
Photogrammetry – the safe, albeit elaborate, path
As usual, we opted for the tried-and-true photogrammetry approach using over 1,000 images. The shoot took about an hour. Thanks to our experience, we knew that this number of photos would yield a solid result.
However, this involves a considerable amount of effort—especially on-site at a museum, where time is often limited and special care must be taken with the exhibits.
Despite all care, a typical problem showed up during the later evaluation:
Especially in shaded areas, such as behind the rear arches of our model, slight distortions and artifacts occurred. Lines that should have been straight appeared partially warped – almost like in a Dali painting. No wonder, as these areas were in shadow, where very little image data was available. This reveals a known weakness of photogrammetry without additional LiDAR scan data or point clouds.
Photogrammetry model
Gaussian Splatting Model with the Lixel K1 Lidar Scanner
Lixel K1 – the quick solution with a surprise effect
The Lixel K1 LiDAR Scanner brought an interesting alternative into play. Although the scanner is actually designed for larger rooms and objects, our attempt showed: Even with a three-meter model, it delivered surprisingly good results.
The recording only took 15 minutes, including multiple laps and various perspectives from above and below. Especially positive: We were able to capture the model from a safe distance without endangering the exhibit.
We were aware that the scanner also captured the surrounding area—but the gaps that resulted in the room weren't a problem given our focus on the model.
Why not use a handheld scanner?
An important point:
Although theoretically we could have used a handheld scanner, we did not have one ourselves.
In addition: Even if the model was presented freely in space on a table and we had physical access to the table, working with a handheld scanner would have been very time-consuming and potentially risky.
Because:
Handheld scanners generally have to be used at a very close range (less than 20 cm). This would have made it particularly difficult to reach the top sections of the model. To scan these areas properly, we would have had to lean far over the model or hold the scanner directly over the delicate exhibit.
That would have been not only cumbersome but also dangerous—both for the exhibit and for the stability of the mount.
In addition, with handheld scanners, the question always arises as to how stable the movements can actually be in such positions without skewing the results.
So it was clear to us: LiDAR with Gaussian Splatting This seemed to be the most pragmatic, safest, and at the same time fastest solution, since we were able to fully capture the model from a distance without having to bend over.
Important Note: Gaussian Splatting is a workflow, not hardware
There is one point that is often misunderstood, and it should be clarified here:
The Lixel K1 LiDAR Scanner does not necessarily create proprietary Gaussian Splatting data. Rather, the result depends on, which processing workflow and which software to use.
The ones from Lixel provided software offers a direct export in formats that are very easy to use for Gaussian Splatting by default – for example, for real-time engine applications or for web-based visualizations.
However, it would be equally possible to Raw data from the Lixel K1 also for classic photogrammetry workflows to be used. For example, the LiDAR data would have to be exported as point clouds and combined with photogrammetric methods or converted into polygon models.
This is entirely achievable – but with a associated with significantly higher effort and corresponding expertise. Specialized software solutions such as Agisoft Metashape or similar can support such workflows, but they are not the standard application of the Lixel K1.
For us, in this project, the use of Gaussian Splatting: the most pragmatic way, for quick and high-performance results for a VR- and WebAR-compatible visualization to achieve.
Quality of Results – Photogrammetry vs. Gaussian Splatting
Both methods yielded valid results in comparison – each with its own strengths and weaknesses.
The Photogrammetry generated a polygonal model with fine textures, but showed weaknesses in difficult areas (e.g., in shadow zones).
The Lixel K1 delivered a Gaussian splatting model with a soft, flat look that reproduced the captured views in a particularly convincing way.
File size and performance in mobile use
An interesting difference was revealed in the amount of data and performance on mobile devices.
The Photogrammetry Model The file size turned out to be significantly larger overall. In the quality we found acceptable, the file size was around 20 MB.
We've noticed that reducing the polygon count – for example, to optimize for mobile applications – leads to visible quality losses very quickly, especially with curves like the archways, which then appear polygonal and lose detail.
In contrast, the Gaussian Splatting Files are only about 4 MB in size with comparable visual quality.
However, testing on a Samsung S20 showed that despite the smaller file size, performance during Gaussian Splatting slightly behind Photogrammetry remained.
The framerate on the device was at Gaussian Splatting noticeably lower, although still completely acceptable and usable for the intended purpose.
This is mainly due to the rendering method of Gaussian Splatting, which generates a different GPU load than classic polygonal models due to the splats.
This point should be particularly considered on mobile devices, even if the amount of data appears smaller on paper.
Deployment in WebAR – data volume and performance are key
An important aspect of our project was the planned purpose of the 3D scan:
The model should later in a WebAR-application to be used, where users can place the object directly in the room via their smartphone or tablet – without an additional app, purely via the browser.
This means:
File size and performance are critical factors, as the model must be loaded over the internet. Here, one is forced to find a middle ground between high quality and the smallest possible data volume.
Our tests showed:
Both the optimized polygonal model from the Photogrammetry as well as the Gaussian Splatting Model could successfully into WebAR be integrated.
Both variations could be delivered via WebGL without any problems, and we paid particular attention to compactness in both versions.
Pleasing was that the Gaussian Splatting The model offered convincing display quality despite its smaller file size, and could also be delivered smoothly via WebGL – which is not a given for all modern 3D data formats.
In practical use in WebAR you had to look closely to even notice the differences between the two models.
Both versions were well-suited for the application and performed stably, even on normal smartphones like our Samsung S20.
Direct Comparison in WebAR (interactive links)
For those who want to see for themselves, we have both versions in WebAR-Applications provided.
Here you can try out the differences directly in the browser:
Display photogrammetry model in WebAR
Display Gaussian Splatting Model in WebAR
Challenges in Post-processing and Tool Support
Photogrammetry: Aids for Better Alignment – and Their Downsides
To be in the Photogrammetry To achieve the cleanest possible alignment of the photos, we used some tried-and-true tools in the project:
GPS module on the camera
Our photographer, Johannes Müller, used a GPS module mounted on the camera. This wrote position and direction information directly into the image metadata, which made later alignment in the Photogrammetry Software significantly eased.
Marker images (similar to QR codes):
Additionally, we have placed printed marker images on and around the exhibit. These markers help the software to correctly match images, especially for objects with uniform surfaces or few distinctive details.
These tools are in the Photogrammetry common practice – however, they have the disadvantage that the markers remain visible on the textures later and must be manually removed.
Fortunately, this is relatively easy to solve with polygonal 3D models, as there are established tools like Substance Painter that allow you to retouch textures directly on the model.
However, one must consider that the Photogrammetry Software generated UV atlases often appear very complex and cluttered, which makes texturing somewhat more challenging than with manually created models with clean UV unwrapping.
Gaussian Splatting: Retouching in the original test
Things look very different with Gaussian Splatting from
With Gaussian Splatting, colors and aspect ratio are determined in the splats rather than in classic UV textures. In our project at the time, we lacked a sophisticated retouching method to selectively remove markers or imperfections. The available tools are still evolving; this experience is described in the context of the project at that time.
In our project at the time, we were able to delete or move splats, but we weren’t able to specifically correct or retouch the colors.
At the time of our test, the tools we reviewed, PostShot and SuperSplat, did not offer a suitable function for this specific retouching case.
Only after extensive research did we come across a Blender addon called „3DGS Render Addon,“ which makes it possible to directly in Blender Gaussian Splats to edit and repaint.
This plugin ultimately allowed us to remove the marker positions from the splat model — even though the workflow was significantly more experimental and error-prone than in the classic Photogrammetry.
This experience clearly shows:
The workflow and tool support in thePhotogrammetry is mature and professionally standardized. In the original project,Gaussian Splatting When it comes to editing and retouching, it’s still significantly more experimental.
What has changed since our test
Our previous comparison described a specific practical test with the Lixel K1. Since then, the device and software landscape has evolved further. Above all, it has become clearer: Gaussian splatting and photogrammetry often start with very similar image data and a similar alignment. In both cases, overlapping images must be recognized and camera positions and viewing directions must be determined. This can be done using image features, an SLAM trajectory, LiDAR data, or a combination of these.
The decisive difference arises primarily after that. In photogrammetry, geometry and textures are reconstructed from the aligned images to create a polygonal model. In Gaussian Splatting, the images are used with their camera positions to optimize a perspective representation from 3D Gaussians. The methods therefore share important steps, but pursue different goals in reconstruction and result.
This affects the image capture. For a good Gaussian splatting result, the scene should be captured from many angles and preferably at the heights and positions from which it will be viewed later. If such views are missing, areas may appear soft, hollow, or semi-transparent. A photogrammetry pipeline can partially fill in missing areas during mesh generation or derive a coherent surface from the existing data. However, this does not automatically correct invisible areas: here too, sufficient overlap and appropriate coverage are required.
Since our K1 test, we have also tried out newer devices. In our opinion, the XGRIDS PortalCam provides visibly better image quality than the K1 used at the time and is especially designed for visual Gaussian-Splatting results. The Lixel K2 supplements the scan workflow with the ability to generate textured meshes from scan and panorama data. This is an important step towards polygonal post-processing. However, for high-quality media visualization, the mesh quality in our previous tests is still insufficient.
With the Lixel L3, a new scanner has now been added. XGRIDS cites a validated relative accuracy of 5 mm. This is interesting for spatial mapping and surveying, but by itself it does not say anything about whether textures, fine details, and surfaces are sufficient for high-quality media production. That is exactly what a practical test must show.
Our current situation: Photogrammetry remains particularly interesting when measurable, editable polygon models are required. Gaussian Splatting is suitable when a scene needs to be visually convincing and rendered from the actual angles captured. With SLAM scanners, the suitability for media production depends on whether point cloud, mesh, and texture resolution achieve the necessary level of detail.
Conclusion: What is better - Gaussian Splatting or Photogrammetry?
To conclude, the frequently asked, seemingly simple question:
What's actually better now? Photogrammetry or Gaussian Splatting?
Our honest feedback after this project is:
For exactly this test case, Gaussian Splatting the more pragmatic solution. No general ranking of the procedures can be derived from this.
Why?
Because in our application scenario – quickly capturing an exhibit in the museum for later use in a WebAR-application – mainly came down to two points:
- Efficiency and speed on site
- An acceptable balance between file size, quality, and ease of further processing
The following measurements were taken during our test of the architectural model at the time. They document this individual case and are not general performance values:
Gaussian Splatting:
- Recording time: approx. 15 minutes
- Processing time for preparation: approx. 4 hours
- Final file size: approx. 4 MB
- Ready to go immediately for WebAR without significant post-processing
Photogrammetry
- Recording time: approx. 1 hour
- Processing time for preparation: approx. 4 hours
- Initial file size after Reality Capture: multiple GB (with millions of polygons)
- Additional manual optimization and reduction for WebAR time-consuming
In our specific museum test, the recording with Gaussian Splatting much faster. However, the result depends on the recording plan, the scanner and the required coverage.
In this project, the GS workflow reduced our post-processing workload at the time. This does not translate into a general cost advantage: recording, reconstruction, cleaning, and the target platform must all be factored in.
However, to be fair, it must be added:
Gaussian Splatting Polygon workflows are not replacing everything everywhere: When CAD, editable geometry, UV textures, or classic material processing are needed, a mesh is still required.
The degree of flexibility with which a splat can be processed depends on the software, export format, and the desired manipulation. A polygon model is usually the better data type for direct geometry and texture changes.
So, anyone who needs data for classic CAD, offline renderings, or complex material processing will run into issues with Gaussian Splatting currently still at limits.
In our test at the time, there were major Gaussian Splatting-Files in Unreal and Unity Stability and performance issues. This is a project result from that time, not a general statement about current engine versions or plugins.
Curiously, our tests with WebGL and browser-based players the most stable and performant solution, which positively surprised us in practice.
In summary:
- For an aspect-dependent, visually convincing representation of the captured scene: Gaussian Splatting can have benefits.
- For measurable, editable polygon models and established 3D pipelines: Photogrammetry is often the more appropriate basis.
Our personal conclusion:
Today, we both view these two methods as important tools with different goals. What matters is what data the project needs: a visual-dependent representation, an editable mesh, or both.
Frequently Asked Questions (FAQs)
Both methods use overlapping images and need to align camera positions. Photogrammetry reconstructs a polygonal model with textures; Gaussian Splatting optimizes a perspective rendering from 3D Gaussians.
In our previous K1 comparison, photogrammetry took about an hour and the Gaussian Splatting scan took about 15 minutes. These values apply to this project and are not a general time comparison.
It depends on the capture, data quality, and target. Photogrammetry provides editable mesh geometry; Gaussian Splatting can reproduce the captured view very convincingly. For measurement accuracy, the respective capture and processing chain counts.
Yes. The same principle and a common alignment can serve as the basis for different outputs. Whether a mesh, a Gaussian splat, or both result from this depends on the reconstruction workflow.
Yes. Specialized tools support things like viewer, cleaning, or format conversion. What corrections are possible depends on the tool and file format. For direct geometry and UV texture changes, a polygonal mesh is usually the more suitable format.
This can vary depending on the device, model size, and viewer. The file size alone says little about the rendering performance; both formats must be tested and optimized for the respective target device.
