Gaussian splatting 3DGS
A technique for building a photorealistic 3D scene from ordinary photographs and rendering it in real time. Presented at SIGGRAPH 2023 by a research group at France's Inria, it made NeRF look dated almost overnight.
The scene is represented as millions of small, translucent, view-dependent Gaussians in space — each "splat" has a position, shape, colour and opacity, and the parameters are optimized until the rendering matches the photographs. The decisive part is that splats can be rasterized like ordinary graphics instead of requiring rays to march through a neural network, which is the difference between seconds per frame and hundreds of frames per second. Training takes minutes rather than hours. Used in visual effects, real estate tours, heritage documentation and robotics. Drawbacks: large files, scenes that are hard to edit afterwards, and artifacts wherever few images provide coverage.