## TL;DR
3D Gaussian Splatting (3DGS) has revolutionized 3D scene reconstruction, achieving real-time photorealistic novel view synthesis — 100x faster than NeRF — by representing scenes as millions of optimized anisotropic Gaussian primitives.

## Core Explanation
NeRF works by training an MLP on multi-view images: for each pixel ray, sample points in 3D, query the network for color and density at each point, and composite via volumetric rendering. 3DGS replaces the MLP with explicit Gaussians: each Gaussian has position (mean), covariance matrix (anisotropic spread), opacity, and spherical harmonic color coefficients. The entire scene renders via differentiable tile-based rasterization.

## Detailed Analysis
The key 3DGS innovation: (1) initialize Gaussians from SfM point cloud; (2) optimize position, covariance, opacity, and color via gradient descent; (3) adaptive density control — clone small Gaussians in under-reconstructed regions, split large ones in over-reconstructed regions. DreamGaussian (2023) extended this to text-to-3D generation. 2025 frontier: 4D Gaussian Splatting for dynamic scenes, SLAM integration.

## Further Reading
- 3D Gaussian Splatting Papers GitHub
- NVIDIA Instant NGP
- DreamFusion: Text-to-3D