SAM 3D — Reconstruct Bodies & Objects from a Photo
3D Bodies and Objects from a Single Image
SAM 3D, from Meta, reconstructs 3D geometry from a single photo in two dedicated modes. In body mode it recovers one or more human bodies from an image, complete with pose; in object mode it reconstructs an object you point it at, guided by a short text prompt. Both modes output a GLB you can preview interactively right on this page and download — no multi-view capture or manual modeling required.
The two modes are tuned for different subjects. Body mode is built for people: upload a photo containing one or more humans and SAM 3D recovers their 3D body shape and pose, optionally including 3D keypoints in the model. Object mode is built for things: upload a photo and tell it what to reconstruct with a short prompt (for example 'car', 'chair', or 'backpack'), and it segments and reconstructs that object — with an optional detection threshold to tune how aggressively it picks the subject, and a textured-GLB option that bakes textures and UVs into the output.
Because SAM 3D is grounded in Meta's Segment Anything family, it's especially good at isolating the right subject from a busy image before reconstructing it. You don't need a clean studio shot or a turntable of photos — a single everyday image is enough to get a rotatable 3D result.
The workflow stays simple. Choose body or object, upload your image, set any options, then generate and inspect the model in the built-in interactive viewer — orbit, zoom, and check it from every angle. When it looks right, download the GLB and drop it into your engine, render, or DCC tool. Every generation is also saved to your dashboard gallery so you can return to it later.
How to Generate a 3D Model with SAM 3D
Pick a Mode & Upload
Choose Image to Body for people, or Image to Object for things, then upload a single photo of your subject.
Set Options
For body, choose whether to include 3D keypoints. For object, describe what to reconstruct and adjust the detection threshold or textured-GLB option.
Generate & Download
Click Generate, then orbit and zoom the result in the interactive viewer. Download the GLB — every model is also saved to your dashboard gallery.
SAM 3D Technical Specifications
| Provider | Meta |
| Platform | Cloud API (hosted) |
| Input Modes | Image-to-body and image-to-object |
| Input | A single image |
| Body Mode | Recovers one or more human bodies with pose |
| Object Mode | Prompt-guided object reconstruction |
| Output Format | GLB |
| Preview | Interactive in-browser 3D viewer |
| Commercial Use | Supported |
Why Choose SAM 3D
Bodies and Objects
One model, two dedicated modes — recover human bodies with pose, or reconstruct everyday objects, both from a single photo.
Smart Subject Selection
Built on Meta's Segment Anything family, SAM 3D isolates the right subject from a busy image — guided by a text prompt in object mode — before reconstructing it.
One Photo, No Setup
No turntable, multi-view rig, or studio lighting needed — a single everyday image is enough to get a rotatable GLB you can preview and download.
SAM 3D vs Other Image-to-3D Models
| Feature | SAM 3D | Hyper3D Rodin v2.5 | Tripo3D H3.1 |
|---|---|---|---|
| Input | Single image | Up to 5 images, or text | Image, multiview or text |
| Specialty | Human bodies + objects | General objects + characters | General objects |
| Subject Selection | Segment Anything (prompt/mask) | Whole image | Whole image |
| Output | GLB | GLB | GLB |
| Best For | People & in-the-wild objects | Detail tiers & rig-ready poses | Faithful multi-angle reconstruction |
What Can You Create with SAM 3D
Human Avatars
Recover a posed 3D body from a photo for avatars, previs, and character reference.
In-the-Wild Objects
Reconstruct an object from an everyday photo — no clean background or studio shot required.
AR & VR
Produce lightweight GLB assets that drop straight into AR experiences and VR scenes.
Product Capture
Turn a single product photo into a 3D model for e-commerce and configurators.
Pose & Motion Reference
Use recovered bodies with 3D keypoints as pose and motion reference for animation and rigging.
Game Assets
Reconstruct props and characters from reference photos and bring them into your engine via GLB.
Related AI Models
Frequently Asked Questions About SAM 3D
What does SAM 3D do?
SAM 3D reconstructs 3D geometry from a single image in two modes. Image-to-body recovers one or more human bodies with their pose; image-to-object reconstructs an object you specify with a short prompt. Both output a GLB you can preview and download.
What's the difference between body and object mode?
Body mode is built for people and recovers human body shape and pose from a photo, optionally including 3D keypoints. Object mode is built for things — you upload a photo, describe what to reconstruct (for example 'car'), and it segments and reconstructs that object.
How do I tell it which object to reconstruct?
In object mode you provide a short text prompt naming the object — like 'chair' or 'backpack'. SAM 3D uses Meta's Segment Anything to isolate that subject from the image before reconstructing it. You can also adjust the detection threshold to tune how it picks the subject.
Do I need multiple photos or a clean background?
No. SAM 3D works from a single everyday image and is designed to isolate the subject from a busy scene, so you don't need a turntable, a multi-view rig, or a studio backdrop.
What format do I get and can I preview it?
You download a GLB, which works across the web, game engines, and AR. The model is previewed interactively in your browser before you download, so you can orbit and zoom to check it from every angle.
How is this different from Hyper3D Rodin v2.5 and Tripo3D H3.1?
SAM 3D specializes in reconstructing a chosen subject — a human body or a prompted object — from a single in-the-wild photo, using Meta's Segment Anything for subject selection. Hyper3D Rodin v2.5 takes up to five images with quality tiers and rig-ready poses; Tripo3D H3.1 adds a multiview mode and standard/detailed quality. Pick the one that matches your input and subject.