SegAnyGAussians

Why we included this project

Teams already invested in 3D Gaussian Splatting often need to pick objects out of a scene without rebuilding the whole representation. This repo is the official implementation of SAGA, a method that moves Segment Anything's text or click prompts onto the 3D Gaussians domain, so a prompt resolves to a segmented object in milliseconds rather than minutes. That speed makes it useful for editing workflows and for AR or robotics work where you want to isolate something on the fly. The README walks through the full pipeline: pre-train the Gaussians, run SAM over the images to extract candidate masks, compute per-mask scales, then do promptable segmentation on the 3D representation. It's research code built for reproducibility, so expect a conda environment with CUDA and your own pre-trained 3D-GS model, but if you're evaluating segmentation methods for Gaussian scenes, this is much easier to build on than reimplementing the paper from scratch.