Gaussian Splatting vs Photogrammetry: Same Photos, Different Outputs

Both 3DGS and photogrammetry start from the same input: a set of photographs taken from multiple viewpoints. Both use COLMAP (or equivalent SfM) to estimate camera positions. But they produce fundamentally different outputs.

Photogrammetry produces a triangle mesh with texture maps — the same kind of geometry used in games, movies, 3D printing, and CAD. You can select faces, move vertices, boolean-cut, UV-unwrap, and 3D print the result. The output is a tangible, editable 3D object.

Gaussian Splatting produces a cloud of colored ellipsoids — not a surface. You cannot select edges, cut holes, or 3D print it. But it looks dramatically more photorealistic, especially on complex materials (vegetation, hair, fabric, glass). And it processes 5-20x faster.

This is not a "which is better" question — they serve different purposes. The right choice depends on what you need to do with the output.

Side-by-side, 3DGS scenes look more real. The reason: spherical harmonics. Each Gaussian encodes how its color changes with viewing angle, capturing the subtle material appearance that makes real objects look real — the way a wooden table shifts from warm brown to cool gray as you view it at a shallow angle, or the way fabric changes hue in different lighting directions.

Photogrammetry textures are view-independent: a pixel on the texture map shows the same color regardless of camera angle. This is fine for matte surfaces (concrete, painted walls) but looks artificial on anything with specular properties (wood, metal, skin, leaves). Professional photogrammetry compensates with PBR material maps, but generating accurate PBR from photos is a separate, difficult problem.

Where photogrammetry looks better: large flat surfaces (walls, floors, roads) where mesh geometry is clean and textures tile well. 3DGS sometimes shows "cloudy" artifacts on large flat areas because the Gaussians are optimized for visual accuracy, not geometric flatness.

Photogrammetry wins outright. 3DGS output cannot be 3D printed — it is not a surface, so slicers cannot process it. Research tools like SuGaR extract approximate meshes from Gaussian Splatting scenes, but results are noisy, low-detail, and far from print-ready.

If your goal is 3D printing a real-world object: use photogrammetry. Capture photos → process with RealityCapture or Meshroom → export STL/OBJ → repair with polyvia3d.com/repair/stl → slice and print. See our guide on converting 3D scans to printable files.

If your goal is visual documentation with the option to print later: capture photos and run both pipelines. Use the 3DGS output for interactive web viewing and the photogrammetry output for printing. Both use the same input photos and COLMAP step.

3DGS has a growing advantage for web display. A compressed SPZ file (15-20 MB) loads faster than a textured mesh + textures (often 50-200 MB even after Draco compression). The visual quality is better for organic scenes. And the rendering pipeline is simpler — no texture loading, no material setup, just sort-and-splat.

Photogrammetry meshes work well on the web via GLB/glTF + Draco compression + Google Model Viewer. For products, architecture, and anything that benefits from clean geometry, this is the established pipeline. But for natural scenes (gardens, forests, historical sites), 3DGS produces more convincing results.

Our recommendation for web: use 3DGS for photorealistic scene tours (real estate, heritage, landscapes). Use photogrammetry meshes for product visualization and anything that needs to look "clean" (architecture, furniture, manufactured objects).

3DGS is dramatically faster. Training a room-scale scene: 20-40 minutes on an RTX 3060. Processing the same scene with photogrammetry: 2-8 hours for a high-quality textured mesh (RealityCapture), or 30-90 minutes for a quick preview mesh (Meshroom with reduced settings). Cloud photogrammetry services (RealityCapture cloud, Capturing Reality) charge per scene or per month.

For rapid documentation (construction sites, insurance claims, archaeological digs), 3DGS's speed advantage is significant. Capture photos in the morning, have an interactive 3D scene by lunch. Photogrammetry would not finish processing until the evening.

Cost comparison: 3DGS is free (open-source tools + your own GPU) or cheap (cloud services from $0-10/month). Professional photogrammetry software: RealityCapture $15-30K/year, Metashape $179-549, Meshroom free. Cloud photogrammetry: $30-200/month depending on volume.