Updated on August 11, 2026
Meta SAM 3D is a pair of 3D reconstruction models from Meta: it rebuilds complete objects, scenes and human bodies from a single photograph. Everything is free, from the browser demo on the Segment Anything Playground to the weights available on GitHub and Hugging Face. In human preference tests, its reconstructions won by at least 5 to 1 against earlier single-image methods. Cluttered rooms and half-hidden objects are, oddly enough, where it feels most at home.
SAM 3D bundles two models published under the Segment Anything collection: SAM 3D Objects, which reconstructs the geometry, texture and layout of objects and scenes, and SAM 3D Body, which estimates human pose and shape.
In practice, you upload a photo of your living room, click the armchair, and out comes a full 3D object, hidden side included. SAM 3D Body relies on Momentum Human Rig (MHR), a mesh format that separates the skeleton from the body shape, and can be steered with masks or 2D keypoints.
SAM 3D was trained with a human-and-model-in-the-loop annotation pipeline that mixes synthetic pretraining with real-world alignment; SAM 3D Body started from a dataset of billions of images.
Where photogrammetry calls for dozens of shots from every angle, one smartphone picture is enough here, even in a messy room. Meta already runs it in production: the View in Room feature on Facebook Marketplace places a lamp or a table in your home before you buy. Next to Tencent's Hunyuan3D, which leans toward polygon-efficient meshes for real-time rendering, SAM 3D bets on human reconstruction and whole-scene composition.
| Profile | What SAM 3D brings | Example use |
|---|---|---|
| 3D artist, game dev | A textured starter mesh from a photo | Props and set pieces to refine |
| E-commerce | Products viewable in volume | View in Room style previews |
| Robotics, research | Spatial understanding of real scenes | Simulation and object handling |
| VFX, AR studios | Posed, rigged bodies via MHR | Digital doubles, virtual try-on |
The Segment Anything Playground is the shortest route: you drop your images in the browser, select people or objects, and collect the reconstruction with no install and no credit card.
For deeper work, checkpoints and inference code ship under the SAM License, with commercial use permitted within a few guardrails. Fans of open-source AI projects on GitHub will also find MHR there, the parametric human body model (plan on a solid GPU, your fans will remember the session).
Meta also worked with 3D artists on SA-3DAO, an evaluation benchmark built to measure progress on real-world images rather than clean studio renders.
Yes. The Segment Anything Playground demo costs nothing in the browser, and both models' weights download free from GitHub and Hugging Face. Third-party APIs also host the models and bill per generation, but nothing forces you down that road.
Commercial use is permitted under the SAM License, with restrictions such as a ban on military applications. You keep ownership of your derivative works, and the MHR rig ships under a permissive commercial license. Read the official terms before releasing a product.
GLB export makes the generated meshes compatible with Blender out of the box, and step-by-step tutorials cover the whole photo-to-Blender pipeline. Expect a cleanup pass on the mesh before production use.
SAM 3D Body handles each person separately, without modelling contact between people or objects, and Meta says hand pose estimation still has room to grow. On the object side, output fidelity closely tracks the quality of the source photo.
Verdict: Game studios, robotics labs and Blender regulars get a serious free source of 3D raw material here; for lean assets headed to real-time engines, a retopology pass rounds out the pipeline.
Leonardo AI 
Genie by Lumalabs 
Get3D Nvidia 
