SAM 3D Objects

Given a single cluttered real-world photo, SAM 3D Objects reconstructs full 3D geometry, texture, and scene layout for the objects in it, a setting where occlusion usually breaks single-view methods. Part of Meta's SAM 3D family announced in November 2025, the model was trained with a progressive scheme and a data engine that folds in human feedback, and it handles single or multi-object generation, producing object pose, shape, texture, and layout with Gaussian splat output in PLY format. The repository provides checkpoints, a demo script, Jupyter notebooks, and a web demo, with June 2026 additions including released encoder weights, a 3D artist object evaluation set, and a public leaderboard. Setup is Python-based with separate requirements files and a PyTorch3D dependency, which makes installation the fiddliest part, and inference realistically wants a CUDA GPU. Code and checkpoints are distributed under Meta's SAM License, a custom license that permits commercial use but bars military and trade-control-restricted applications and reverse engineering. The project sits around 7.2k GitHub stars.