A suite of 3D reconstruction models released by Meta on November 19, 2025. Hand it a single photo and it infers the shape, pose, and texture of the object or person in the frame — including the back side the camera never saw — and exports the result as a textured 3D model. Conventional 3D capture assumed multi-view photography, dozens of shots taken all the way around the subject; SAM 3D works from a single snapshot instead. It is made up of purpose-specific sub-models — “SAM 3D Objects” for general objects and “SAM 3D Body” for human bodies — and the model weights and inference code are published under Meta’s own SAM License.
Key Features
- 3D reconstruction from a single image: No multi-view capture or depth sensor required; a single handheld snapshot is the input. Occluded regions (parts hidden behind foreground objects) and areas cut off by the frame are filled in by inference, and the output is a closed solid
- SAM 3D Objects for general objects: Given an image and a mask, it recovers shape geometry, texture, and layout within the scene together. Output can be a mesh, or Gaussian splats exported as
.plyfiles - SAM 3D Body for human bodies: Specify the target person with a bounding box or mask and it estimates a full-body mesh. An additional decoder for refining hand position and shape is also provided, aimed at reproducing pose down to the fingers
- Mask creation handled end to end via SAM 3: The intended flow is to cut out the target with SAM 3, which handles image and video segmentation, then pass that mask straight to SAM 3D. Even the work of specifying “what to turn into 3D” shifts to the AI side
- Segment Anything Playground, testable in the browser: An official demo environment lets you upload an image, select a target, and go all the way to 3D without any local setup
- Published weights and code: Checkpoints and inference code are distributed on GitHub (
facebookresearch/sam-3d-objects), bundled with Jupyter notebooks demonstrating single-object and multi-object inference
Pricing
| Plan | Price | Key points |
|---|---|---|
| Segment Anything Playground | $0 | Official demo for uploading an image and going from target selection to 3D in the browser |
| Model weights and code | $0 (SAM License) | Pull checkpoints and inference code from GitHub and run them in your own environment; compute resources such as GPUs are on you |
Pricing is as of August 2026. The SAM License broadly permits both research and commercial use, but comes with the Acceptable Use Policy and trade control restrictions attached as conditions. Always check the official site and the repository’s LICENSE before use.
Pros & Cons
✅ Pros
- 3D from a single photo, which sharply reduces the effort and cost of shooting or scanning
- Weights and inference code are published, so you can embed it in your own environment and build an in-house pipeline
- Output supports both meshes and Gaussian splats, widening your choice of downstream tools
- Separate models for objects and bodies make it easier to get accuracy tuned to each use case
- A browser demo exists, so you can verify quality on your own images before committing
⚠️ Cons
- Running it yourself requires a GPU-equipped environment; it is not as effortless as a cloud SaaS
- The SAM License is not a fully open source license, and its terms of use and trade control constraints have to be checked case by case
- The unseen back side is filled in by inference, not measured. It is not suited to uses that demand dimensional accuracy
- Output is not necessarily a production-quality asset as-is; retopology and texture adjustment tend to be assumed as post-processing
- The offering centers on distributing trained models, so Japanese-language documentation and support channels are not in place
Comparison with Similar Services
| Item | SAM 3D | Hunyuan3D | TRELLIS | Meshy |
|---|---|---|---|---|
| Provider | Meta | Tencent | Microsoft | Meshy |
| Main input | One image (+ mask) | Image, text | Image, text | Image, text |
| Body-specific model | Yes (SAM 3D Body) | No | No | No |
| Delivery | Published weights (SAM License) | Published weights | Published weights | Web service (SaaS) |
| Setup | Own GPU or demo environment | Own GPU | Own GPU | None (browser only) |
Who Is It For
- E-commerce and catalog operators who want to mass-produce 3D assets from product photos
- Developers prototyping AR/VR content who want to hold down the cost of scanning real objects
- Engineers building applications that handle human motion or pose and need body mesh estimation from a single image
- Researchers working on 3D reconstruction who are looking for published models to compare against
- Teams that want to build a 3D conversion feature into their own service in-house rather than through an external API
Summary
SAM 3D takes on the hard problem of “raising a solid from a single photo, unseen parts included,” and makes it approachable with published weights and code. Splitting the suite into object-facing and body-facing models is a practical strength that general-purpose image-to-3D tools do not have. On the other hand, running it takes a GPU environment, and the license is not fully free use. A realistic path is to throw your own images at the official browser demo first, see whether output quality clears your bar, and only then consider embedding it in your own environment.