Gaussian Splatting vs Photogrammetry: Which to Use?

Gaussian splatting vs photogrammetry in one minute

Both methods start the same way. You take many overlapping photos or a video of a subject, and software works out where each picture was taken from. What they build next is different.

• Photogrammetry builds geometry: a polygon mesh, a surface made of triangles, with a texture image wrapped over it.
• Gaussian splatting builds appearance: a large number of small, soft, semi-transparent blobs, optimized so that from any angle the scene looks like your photos.

So the useful question is not which one is better, but what you need the result for. If you need a surface to measure, 3D print or drop into a game engine, photogrammetry is the safer choice. If you want a scene that looks like the real place as you move through it, Gaussian splatting usually looks more natural.

How photogrammetry works

Photogrammetry software looks for distinctive features in your photos, such as a corner, a stain or a pattern, matches them across images, and triangulates their 3D positions together with the camera positions. It then estimates depth for many more pixels to build a dense point cloud, turns that into a mesh, and projects the photos onto the mesh as a texture.

The output is ordinary 3D data. A textured mesh exported as OBJ, FBX or glTF opens in Blender and most other 3D tools. Professional software such as Agisoft Metashape goes further: its Professional Edition adds survey outputs like digital elevation models and georeferenced orthomosaics, plus tools to measure distances, areas and volumes. Because the result has a real surface, you can measure it once it is scaled correctly, 3D print it, use it for collisions in a game, or relight it.

Where photogrammetry struggles

Photogrammetry works best on still, matte, well-textured subjects. Some things make it harder.

• Shiny, transparent and plain surfaces. Reflections move as the camera moves, so the software cannot find stable points to match. Apple's Object Capture guidance specifically warns against highly reflective, transparent or translucent objects, and against single-color or very smooth surfaces.
• Thin and fuzzy detail. Hair, fur, foliage and wires are hard to turn into a clean surface and often come out lumpy, holey or missing.
• Baked lighting. The texture keeps the light and shadows that were present during capture. Agisoft even offers a separate free tool, Texture De-Lighter, to remove shadows from model textures. Without that kind of extra work, a mesh can look flat, and shine that changes with the viewing angle is lost.

How Gaussian splatting works

Gaussian splatting was introduced in a 2023 SIGGRAPH paper by Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler and George Drettakis. It reuses the camera positions and sparse points from structure-from-motion, often computed with COLMAP, places a 3D Gaussian at each point, and then adjusts the position, shape, opacity and color of every Gaussian until renders match the photos. During training it adds Gaussians where detail is missing and removes ones that become almost transparent. The paper's test scenes ended up with roughly 1 to 5 million Gaussians.

Color is stored with spherical harmonics, so each Gaussian can look slightly different from different directions. That is why splats can reproduce some view-dependent effects, such as highlights sliding across a glossy surface, that a single baked texture cannot. Fine structures like leaves and hair also tend to look softer and more natural than they do in a mesh. Splats still struggle with mirrors, glass and anything that moved during capture, so they do not remove the need for careful shooting.

Where NeRF fits

Neural Radiance Fields (NeRF), introduced in 2020, tackled a similar problem earlier. A NeRF stores the scene inside a neural network and renders it by sampling many points along every camera ray. Quality can be excellent, but rendering is slow.

The 3DGS paper put numbers on this. On the Mip-NeRF 360 dataset, the Mip-NeRF 360 method trained for about 48 hours and rendered at about 0.06 frames per second, while Gaussian splatting trained in about 41 minutes and rendered at around 134 frames per second at comparable quality, on the authors' test hardware.

For most creators today, the practical choice is between photogrammetry and Gaussian splatting. NeRF is best understood as the research step that made radiance fields popular.

Realism vs utility: an honest comparison

It is tempting to say splats are simply faster, but that is misleading. A finished photogrammetry mesh also renders very quickly; games draw meshes in real time all the time. Gaussian splatting was designed to render in real time too. The real difference is what each output is good at.

• Realism: splats usually win, especially for foliage, fine detail and view-dependent effects such as shine and soft reflections.
• Utility: meshes win for anything that needs a surface, including measuring, 3D printing, physics and collisions, retopology, and standard game or CAD pipelines.
• Compatibility: meshes open almost everywhere. Splats need a splat-aware viewer or editor.
• Editing: a mesh can be sculpted and remodeled. A splat can be cropped and cleaned, but not reshaped like a model.
• Processing time: for both, it depends heavily on the number of photos, their resolution and your hardware.

Converting a splat into a mesh is possible with some tools, but the mesh usually loses the view-dependent look and softness that made the splat attractive. If you need both outputs, capture once and process the same photos both ways.

Which should you use?

Match the method to the job.

• Choose photogrammetry if you need measurements, a 3D print, a game asset with collisions, a model you will keep editing in Blender, or a survey deliverable.
• Choose Gaussian splatting if you want a realistic walkthrough of a room or property, a product people can spin around in a browser, a record of a place, or an orbit video for social media.
• Capture the same way for both: plenty of overlap (Apple recommends 70% or more between neighboring shots), soft even light, and a subject that stays still.

Trying Gaussian splatting with SplatLoom

If you want to try Gaussian splatting without sending your photos anywhere, SplatLoom is a desktop app for Mac (Apple Silicon, macOS 15 or later) and Windows 10/11 x64. It turns overlapping JPG or PNG photos, or an MP4 or MOV video, into a 3D Gaussian splatting scene entirely on your own computer, with no upload, no account and no cloud processing.

You pick photos or a video and click Generate. Afterwards you can clean up the scene with box or sphere crop, rectangle select, one-click spike cleanup and undo/redo, and export a Gaussian PLY, a PNG still, a 12-second 1080p MP4 orbit video, or a self-contained offline HTML viewer. It bundles COLMAP for camera positions, Brush for splat training on your local GPU and PlayCanvas for the viewer, so there is no Python or command line to set up. Typical photo sets are 30 to 150 overlapping photos. In our test, a 13 to 20 second smartphone video was processed end to end in about 9 to 14 minutes on an Apple M5 Pro MacBook.

SplatLoom costs US$19.90 as a one-time purchase that includes the Mac and Windows installers, with no subscription and no credits, and a 7-day refund if it does not run on your computer. Since this article is about choosing: SplatLoom makes splats, not meshes. It is not a CAD or measurement tool, and its Windows GPU support is newer than its Mac support.