map-anything

Why we included this project

Most 3D reconstruction tools demand per-scene optimization before you get usable geometry. MapAnything skips that. A single feed-forward transformer takes monocular, stereo, or multi-view images and directly outputs metric-scale depth, point clouds, camera poses, and meshes in one pass. That makes it useful for robots that need to know real distances, AR apps that anchor content to physical space, and photogrammetry pipelines that currently assemble scenes offline. One model covers more than a dozen tasks, including structure-from-motion, multi-view stereo, monocular depth, registration, and depth completion, so teams can retire several separate tools. The modular interface lets other reconstruction models like DUSt3R, VGGT, and MASt3R run interchangeably, which makes it work as both a ready-to-use engine and a neutral base for comparing approaches.