Gaussian Splatting (3D Gaussian Splatting, or 3DGS) is a technique for reconstructing photorealistic 3D scenes from a collection of photographs or video frames. Unlike traditional photogrammetry, which produces dense point clouds requiring extensive cleanup, or NeRF (Neural Radiance Fields), which is computationally expensive to render, Gaussian Splatting produces real-time-renderable 3D scenes that can be directly integrated into design workflows.
How Gaussian Splatting Works
The technique represents a 3D scene as a collection of millions of tiny 3D Gaussian distributions — each one an ellipsoid with a position, orientation, size, opacity, and color. By optimizing these Gaussians to reproduce the input photographs from their respective camera positions, the system learns a compact and accurate representation of the scene that can be rendered from any viewpoint at real-time frame rates.
For practical use in AEC: capture 100–500 overlapping photographs of a space or building, run the optimization (minutes to hours depending on scene complexity and hardware), and receive a fully navigable 3D model with photorealistic appearance — no manual point cloud cleaning, no mesh reconstruction, no UV unwrapping.
Applications for Architects and Engineers
The primary AEC application is as-built capture: documenting existing conditions of a site, building, or construction phase without traditional photogrammetry workflows. An architect can capture a heritage building interior, a construction site at a specific phase, or an urban context, and integrate the result directly into their design software as a spatial reference.
Second application: the Gaussian Splatting representation preserves photometric accuracy — it records how light actually behaves in the captured space. This makes it valuable as a reference for material and lighting decisions in later design phases.
Third: for client communication. A Gaussian Splatting capture of a site, embedded in a real-time viewer or exported as video, provides spatial intuition about a place that no plan or section can convey.
Gaussian Splatting vs. Traditional Photogrammetry
Traditional photogrammetry pipelines (Reality Capture, Metashape, Pix4D) produce triangle meshes and texture maps — formats that design software already understands, but at the cost of hours of post-processing to clean artifacts, fill holes, and correct geometry. Gaussian Splatting skips the mesh entirely and produces a representation that is more photorealistic but less geometrically clean.
For AEC use cases that require precise measurements — structural surveys, BIM as-built documentation — photogrammetry or laser scanning remains the appropriate tool. For spatial context, client communication, and early design reference, Gaussian Splatting is significantly faster and more photorealistic.
MIAW Module F2 — Modeling with AI
Gaussian Splatting is covered in MIAW Module F2 alongside language-assisted geometric modeling. The deliverable: a 3D capture model of a real space integrated in the student's design environment, produced without intensive post-processing, alongside a natural language-assisted modeling exercise applied to an element of the student's actual project.