← AI Terminology
Gaussian Splatting
3D Gaussian splatting represents scenes as clouds of anisotropic Gaussians rasterised quickly for real-time novel-view synthesis, often matching NeRF quality at higher speed.
It rapidly became a leading practical 3D representation (2023+).
It rapidly became a leading practical 3D representation (2023+).
Why It Matters in AI
NeRFs were beautiful but slow; Gaussian splatting delivered real-time rendering with trainability from images. It dominates many production novel-view and 3D content workflows now.
Key Points
| Aspect | Description |
|---|---|
| Pro | Real-time viewers; strong quality |
| Rep | Millions of 3D Gaussians with opacity/colour/covariance |
| Origin | Kerbl et al., 3D Gaussian Splatting (2023) |
| Render | Differentiable splatting/rasterisation |
| Related | NeRF, photogrammetry, radiance fields |
| Ecosystem | Open viewers, editing research exploding |
Simple Analogy
Painting a scene with countless tiny soft airbrush ellipses in 3D space you can rephotograph instantly from any angle.
Common Usage Examples
- Train 3DGS from multi-view photos
- Real-time web viewers
- Edit/remove Gaussian primitives
- Benchmark FPS vs NeRF
Summary
In short: Gaussian splatting builds real-time renderable 3D scenes from image sets — a practical successor path to slow classic NeRFs.