LangSplat: 3D Language Gaussian Splatting

LangSplat targets open-vocabulary querying in 3D by attaching language features to 3D Gaussians and rendering them with a tile-based splatter analogous to RGB splatting, avoiding ray marching through a NeRF. A scene-wise language autoencoder reduces memory versus storing full CLIP embeddings, and hierarchical semantics derived from segmentation priors sharpen object boundaries. Experiments report large speedups over prior NeRF-grounded language fields at high resolution with improved grounding quality.