Skip to content

DirectoryTools

SAM 3D logo

SAM 3D

Meta's models that rebuild objects and bodies in 3D from one image.

Visit SAM 3D

3D & GamingFreeWe checked the prices on ai.meta.com on 28 Sep 2026.

SAM 3D homepage screenshot

Recaply review

SAM 3D review: free 3D from messy photos, if you have the GPU

SAM 3D is Meta's pair of free, open-weight models that rebuild objects and people in 3D from a single photo, built for researchers, robotics teams and developers who run their own GPUs.

$0
To download both models under Meta's royalty-free SAM License
5:1
Minimum win rate for SAM 3D Objects over other models in Meta's human preference tests
32 GB
Of NVIDIA GPU memory Meta lists as the minimum for SAM 3D Objects, on Linux
840M
Parameters in the larger SAM 3D Body checkpoint; a 631M version scores almost the same

What is SAM 3D?

SAM 3D is two models. SAM 3D Objects takes one photo and a mask of the object you want, then predicts its shape, texture and position in the scene, and it copes with clutter and partly hidden objects that trip up tools trained on clean studio shots. SAM 3D Body turns each person in a photo into a posed body mesh, hands and feet included, and accepts masks or 2D keypoints to guide it.

It is Meta's current 3D release with no successor yet. There is no paid plan or API: you download the weights, or try both models free in the Segment Anything Playground.

What is the difference between SAM 3D Objects and SAM 3D Body?

Objects rebuilds things and scenes; Body rebuilds people. Both start from one photo and can share a scene.

SAM 3D Objects
Takes a photo and a mask per object and returns its shape, texture and pose, saved as a Gaussian splat.
SAM 3D Body
Takes a photo of people and returns a full-body mesh with hands and feet on Meta's Momentum Human Rig.
Together
A Meta notebook places objects and people in the same frame, so a whole scene can be rebuilt from one image.
Limits
Objects predicts one object at a time without reasoning about contact, and Body treats each person on their own.

Who is SAM 3D best for?

SAM 3D is best for robotics and 3D perception researchers, developers who need body pose and shape from single photos, and teams that want free, self-hosted 3D reconstruction.

  • Robotics and 3D perception research

    Objects come back posed in the scene rather than floating alone, and Meta publishes its SA-3DAO benchmark to measure progress on real photos.

  • Body pose and shape from one photo

    SAM 3D Body returns a full-body mesh on Meta's Momentum Human Rig, and you can steer it with masks or keypoints when a pose is unusual or partly hidden.

  • Teams that want free, self-hosted weights

    The SAM License is royalty-free with no non-commercial clause, so you pay only for your own GPU time.

  • Trying 3D reconstruction before installing anything

    The Segment Anything Playground runs both models in the browser after a login, free for personal, non-commercial use.

What's included with SAM 3D?

Object and scene reconstruction from a masked photo, a promptable full-body model on Meta's Momentum Human Rig, two Body checkpoints, and a public benchmark.

Object reconstruction from a mask
Mask any object in a photo and SAM 3D Objects predicts its geometry, texture and pose, even when other things block part of it.
Multi-object scenes
Rebuild several objects from one photo, and a Meta notebook lines them up with SAM 3D Body people in the same frame.
Promptable body model
SAM 3D Body takes optional masks and 2D keypoints, and a separate hand decoder can refine hand poses.
Momentum Human Rig
Body output uses MHR, which keeps the skeleton separate from body shape, and Meta released MHR under a permissive commercial licence.
Two Body checkpoints
Choose DINOv3-H+ at 840M parameters or ViT-H at 631M; they score within a few points of each other on Meta's benchmarks.
SA-3DAO benchmark
1,000 photos paired with meshes made by 3D artists; 100 are public and 900 are held back for a Hugging Face leaderboard.

How does SAM 3D work?

  1. Request access

    Ask for the checkpoints on Meta's Hugging Face pages; approval is manual, and requests from comprehensively sanctioned countries are refused.

  2. Set up a GPU machine

    SAM 3D Objects needs 64-bit Linux and an NVIDIA GPU with at least 32 GB of memory; install the Python environment from Meta's GitHub repo.

  3. Give it a photo

    Pass an image and a mask for each object to SAM 3D Objects, or just the image to SAM 3D Body, with optional keypoints.

  4. Save the result

    Meta's sample code saves each object as a Gaussian splat .ply file, and SAM 3D Body returns mesh parameters you can render or combine with the objects.

How much does SAM 3D cost?

SAM 3D is free: both models download at no cost under Meta's royalty-free SAM License, and the Playground is free for personal, non-commercial use, so your only cost is GPU time.

SAM 3D Objects$0Open weights under the SAM License
  • Checkpoints on Hugging Face after an access request
  • Needs Linux and an NVIDIA GPU with 32 GB or more
  • Code and notebooks on GitHub
SAM 3D Body$0Open weights under the SAM License
  • 840M and 631M parameter checkpoints
  • Checkpoints on Hugging Face after an access request
  • Training dataset also released
Segment Anything Playground$0In the browser, with a login
  • Upload your own photos and pick objects or people
  • Personal, non-commercial use only
  • Meta may use uploads to improve and train its models

Meta sells no paid plan or API for SAM 3D. The SAM License is royalty-free and has no non-commercial clause, but it bans military, weapons, nuclear and other trade-controlled uses. The MHR body model has its own permissive commercial licence, and the SA-3DAO benchmark data is for non-commercial use. Running the weights costs only your own hardware or rented GPU time. Full pricing on ai.meta.com. We checked these prices on .

How do readers rate SAM 3D?

No one has rated SAM 3D yet. Be the first.

No ratings yetThe score and its stars show once 3 readers have rated SAM 3D.
Your rating

Sign in to rate. One rating per account.

Your rating of SAM 3D

SAM 3D pros and cons

It is free, fast and strong on cluttered real photos, but it gives Gaussian splats rather than clean meshes, the Playground has no file download, and self-hosting needs a 32 GB GPU.

Pros

  • Free to download with a royalty-free licence, and users on Reddit and Hacker News point to it as the free, run-it-yourself option next to paid 3D generators.
  • Early testers on Hacker News found it much faster than rival 3D generators and good at separating overlapping things, such as a person and the chair they sit on; Meta reports at least a 5:1 preference win rate.
  • The body model draws praise for its pose estimates, with one user calling it better than competing models, and Meta measures 54.8 MPJPE on the 3DPW benchmark, where lower is better.

Cons

  • The Playground shows results but gives no 3D file to download; several users could only save a video or pulled the .ply file out through the browser's developer tools.
  • Objects come out as Gaussian splats rather than clean meshes, and users who convert them in Blender say the materials get lost; Meta itself says its moderate output resolution limits detail.
  • Setup takes work: the checkpoints need a manual access request, SAM 3D Objects needs a 32 GB NVIDIA GPU on Linux, and users report download token errors and install problems.
  • Results still need cleanup: 3D artists on Reddit say shapes and poses are not ready for direct use, and Meta notes that SAM 3D Body handles each person separately and trails hand-only models on hands.

What are the best SAM 3D alternatives?

Rodin for export-ready textured meshes, Seed3D 2.0 for PBR assets from an API, and Quick Magic AI Mocap for body motion from video.

Questions about SAM 3D

What is the difference between SAM 3 and SAM 3D?

SAM 3 finds, cuts out and tracks objects in images and video from text or clicks. SAM 3D turns objects and people from a single photo into 3D. They launched together, and SAM 3 can act as the person detector for SAM 3D Body.

Who is SAM 3D for?

SAM 3D is best for robotics and 3D perception researchers, developers who need body pose and shape from single photos, and teams that want free, self-hosted 3D reconstruction.

Is SAM 3D free?

Yes. SAM 3D has a free SAM 3D Objects plan: checkpoints on Hugging Face after an access request.

What are the best SAM 3D alternatives?

The closest alternatives on Recaply are Rodin, Seed3D 2.0 and Quick Magic AI Mocap.

Where can I try SAM 3D?

On its official website, ai.meta.com.

Can I use SAM 3D for commercial projects?

Yes, when you run the models yourself. The SAM License is royalty-free with no non-commercial clause, though it bans military, weapons and other trade-controlled uses. The browser Playground is different: its terms allow personal, non-commercial use only.

What hardware does SAM 3D need?

SAM 3D Objects needs 64-bit Linux and an NVIDIA GPU with at least 32 GB of memory, according to Meta's setup guide. SAM 3D Body lists no minimum and comes in 840M and 631M parameter versions.

Can I export a 3D file from SAM 3D?

Yes, from the code: Meta's SAM 3D Objects demo saves each reconstruction as a Gaussian splat .ply file, and SAM 3D Body returns a body mesh. The Playground is built for viewing, rearranging and adding effects in the browser, so run the models yourself when you need files.

Does Meta use photos I upload to the SAM 3D Playground?

It can. The Playground's terms let Meta use your inputs and outputs to improve the Playground and its SAM models and to train other research and AI work. Running the open weights on your own machine sends nothing to Meta.

Our verdict

SAM 3D is the free model to try first for 3D from real, cluttered photos, as long as you can run it yourself.

Meta reports at least a 5:1 preference win rate over other models, the licence is royalty-free, and users praise its speed and body poses.

Skip it for
Game artists who need clean, textured meshes to drop into an engine, and anyone without a 32 GB NVIDIA GPU.
Check first
Request Hugging Face access early and test whether Gaussian splat output fits your pipeline.