SAGA (Segment Any 3D GAussians) enables open-vocabulary 3D semantic segmentation by applying SAM to 3D Gaussian Splatting scenes. The method extracts mask scales from pre-trained 3D-GS models and associates them with 2D SAM predictions to achieve 3D mask generation and editing. It works on standard 3D reconstruction datasets including MipNeRF360, LLFF, and LERF.