Photogrammetry vs Gaussian Splatting

Output Representation: Meshes vs. Splats

The most fundamental difference is what you get at the end of the pipeline. Photogrammetry produces triangle meshes — the universal currency of 3D graphics for decades. A photogrammetric mesh can be imported into Blender, retopologized in ZBrush (which added dedicated photogrammetry retopo tools in late 2025), rigged for animation, or dropped into Unreal Engine or Unity with full material and lighting support. This interoperability is photogrammetry's deepest advantage.

Gaussian Splatting produces something entirely different: a cloud of millions of semi-transparent ellipsoids that collectively reproduce the appearance of a scene from any viewpoint. There is no mesh, no UV map, no texture file. The representation is optimized for how the scene looks, not what it is geometrically. This makes splats remarkably efficient for view synthesis but means they don't slot neatly into traditional DCC workflows. The 2024 advances in SuGaR and 2DGS have begun to bridge this gap by extracting mesh-like surfaces from Gaussian representations, but these are still approximations rather than production-quality topology.

For teams whose deliverable is a rendered image or an immersive walkthrough, splats may be all they need. For teams that need to modify, animate, or measure the 3D data, photogrammetry's explicit geometry remains essential.

Speed: Capture to Render

Gaussian Splatting's speed advantage is dramatic at every stage. Training a splat scene from a set of photographs typically takes 5–30 minutes on a single consumer GPU, compared to hours or days for high-resolution photogrammetric reconstruction. At render time, splats achieve 30–200+ FPS by mapping directly to GPU rasterization, while photogrammetric meshes require careful LOD management — though Nanite in Unreal Engine 5 has largely solved this for game engines.

Photogrammetry has been closing the gap. AI-powered feature matching has reduced processing times by 30–60% on large datasets since 2024, and videogrammetry workflows introduced in 2025–2026 compress traditional multi-hour capture-to-model pipelines down to 2–10 minutes for time-sensitive applications like accident scene documentation. Cloud-based processing services have also eliminated the need for local GPU workstations.

Still, for rapid on-site visualization — showing a client a photorealistic walkthrough of a captured space within minutes — Gaussian Splatting's end-to-end speed is currently unmatched.

Visual Fidelity and Challenging Materials

Gaussian Splatting excels at rendering scenes that have historically been photogrammetry's weak points: reflective surfaces, translucent materials, fine geometry like hair and foliage, and areas with poor texture overlap. Because each Gaussian stores view-dependent color via spherical harmonics, the representation naturally handles specular highlights and subtle appearance changes across viewing angles.

Photogrammetry, by contrast, bakes appearance into static textures mapped onto mesh surfaces. This works beautifully for diffuse, textured surfaces — rocks, bark, fabric, concrete — which is why photoscanned asset libraries remain the gold standard for environment art. But reflective or transparent objects produce artifacts, and thin structures often result in missing or malformed geometry.

The visual quality gap is most apparent in novel view synthesis: rendering the scene from a viewpoint not in the original capture set. Splats produce consistently sharp, photorealistic results, while photogrammetric meshes may reveal texture seams, geometric holes, or lighting bake artifacts from unfamiliar angles.

Measurement, Accuracy, and Surveying

When the goal is measurement rather than visualization, photogrammetry remains the clear choice. Its reconstruction pipeline is grounded in geometric triangulation with well-understood error models. Sub-centimeter accuracy is achievable with proper capture protocols, and the output integrates directly with GIS, CAD, and BIM systems. Industries like construction, mining, agriculture, and forensics depend on this geometric fidelity.

Gaussian Splatting optimizes for visual appearance, not positional accuracy. While the Gaussian centers roughly correspond to surface positions, there is no guarantee of metric precision, and the representation lacks the topological structure needed for area, volume, or distance calculations. Hybrid workflows — using photogrammetry's SfM stage for georeferencing and then Gaussian Splatting for visualization — are emerging as a practical compromise, but for regulatory or engineering applications requiring certified measurements, photogrammetry (often augmented with LiDAR) is the only defensible choice.

Ecosystem Maturity and Standardization

Photogrammetry benefits from decades of ecosystem development. Software ranges from open-source (Meshroom, COLMAP) to enterprise (RealityCapture, Pix4D, Agisoft Metashape). Output formats (OBJ, FBX, glTF) are universally supported. Capture best practices are well documented. The supply chain — from drone manufacturers to scanning apps to cloud processing — is mature and competitive.

Gaussian Splatting's ecosystem is younger but advancing at remarkable speed. The August 2025 Khronos/OGC announcement adding splats to glTF via the SPZ container format was a watershed moment, signaling the transition from research artifact to interchange standard. The Foundry's Nuke 17.0 beta added native splat support for VFX pipelines. Esri, DJI Terra, and 3DVista have integrated splat workflows. WebGPU enables browser-based splat rendering, making captured 3D as shareable as a photograph.

The standardization trajectory suggests that by late 2026, splat support will be as routine as mesh support in major platforms. But today, teams choosing Gaussian Splatting must still navigate a less mature toolchain with fewer established best practices.

The Convergence: Hybrid Workflows

Increasingly, the real-world answer is not one or the other but both. Both technologies share the Structure from Motion stage for camera estimation. A single capture session can feed both a photogrammetric mesh (for CAD integration, measurement, and editing) and a Gaussian splat scene (for real-time visualization, client presentations, and web delivery). Tools like KIRI Engine and Polycam already support dual output from the same image set.

In virtual production, this hybrid approach is becoming standard: splats provide rapid on-set previsualization and lighting reference, while photogrammetric meshes feed the final CG pipeline. The Superman production in 2025 demonstrated this at scale, using dynamic Gaussian splats for real-time creative decisions alongside traditional geometry for final compositing. As the conversion tools between representations improve — particularly mesh extraction from splats via methods like SuGaR — the boundary between these technologies will continue to blur.