---
title: "SAM 3D Objects: image-to-3D technology & benchmarks — Segmentit"
description: "Understand SAM 3D Objects and Body, single-image 3D reconstruction and SA-3DAO benchmarks against TRELLIS and Hunyuan3D. Explore the Segmentit workflow."
lang: "en"
canonical: "https://segmentit.com/sam-3d/"
---

# SAM 3D. Give your photo volume.

Explore Meta’s single-image reconstruction model, how it builds objects, and its published results against TRELLIS, Hunyuan3D and other image-to-3D methods.

[Try the studio](https://segmentit.com/app?lang=en) [Explore benchmarks](https://segmentit.com/sam-3d/#benchmarks)

[![Object workflow in the Segmentit studio](https://segmentit.com/images/blog/studio-monument.png)](https://segmentit.com/app?lang=en) [Open the studio](https://segmentit.com/app?lang=en)

An independent guide by Segmentit. SAM 3 and SAM 3D are developed by Meta; Segmentit is not affiliated with Meta.

## What is SAM 3D?

SAM 3D, also searched as SAM3D, is a family of Meta models for reconstructing 3D from an image. SAM 3D Objects estimates object geometry, texture and layout. SAM 3D Body is a separate model for human body shape and pose. Segmentit’s object workflow focuses on Objects.

[Meta · SAM 3D ↗](https://ai.meta.com/research/sam3d/)

### SAM 3

Image → 2D mask

### SAM 3D Objects

Image + mask → 3D object

A precise selection and a reconstruction in volume solve different needs. Choosing the right subject comes before shape, texture and scene composition. [Explore SAM 3](https://segmentit.com/sam-3/).

### The models compared

-   [TRELLISMicrosoft · Tsinghua · USTC](https://microsoft.github.io/TRELLIS/)
-   [Hunyuan3D-2.1Tencent](https://github.com/Tencent/Hunyuan3D-2)
-   [Hunyuan3D-2.0Tencent](https://github.com/Tencent/Hunyuan3D-2)
-   [Direct3D-S2Nanjing · DreamTech · Fudan · Oxford](https://arxiv.org/html/2505.17412v1)
-   [TripoSGVAST / Tripo](https://github.com/VAST-AI-Research/TripoSG)
-   [Hi3DGenByteDance · CUHK Shenzhen · Tsinghua](https://github.com/bytedance/Hi3DGen)
-   [SAM 3DMeta](https://github.com/facebookresearch/sam-3d-objects)

Countries of the organizations and collaborations. Each model links to its official project.

## SAM 3D Objects or Body?

### Objects & scenes

**SAM 3D Objects**

Reconstruct the geometry, appearance and placement of an object from an image.

### People & poses

**SAM 3D Body**

Estimate human body shape and pose. A separate model for a different task.

### Selection before reconstruction

**SAM 3 → SAM 3D**

Identify the subject in 2D, then use its mask to guide object reconstruction.

## Choose your creative workflow

Start with a prop for a game, a decorative object for a scene, or a product to explore from another angle. A clear reference image makes it easier to assess the generated shape.

[Select an object](https://segmentit.com/docs/select-an-object/) [Create in 3D](https://segmentit.com/docs/generate-3d/) [Download & export](https://segmentit.com/docs/download-and-export/)

## How does SAM 3D work?

### Image + mask

The source image provides context; the mask identifies the object to reconstruct.

### Geometry + pose

A first generative stage estimates shape and placement. Hidden surfaces are inferred from learned visual patterns.

### Texture + refinement

A second stage refines appearance and details. The research pipeline supports mesh and Gaussian representations.

Research model architecture. Video features, representations and tools exposed in an application can differ. [Official code ↗](https://github.com/facebookresearch/sam-3d-objects)

## SAM 3D against earlier image-to-3D methods

SA-3DAO · Shape reconstruction · F1@0.01 · Higher is better

F1@0.01 · 0 → 0.25

TRELLIS

**0.1475**

Hunyuan3D-2.1

**0.1399**

Hunyuan3D-2.0

**0.1574**

Direct3D-S2

**0.1513**

TripoSG

**0.1533**

Hi3DGen

**0.1629**

SAM 3D

**0.2344**

**Data and metrics**

**SA-3DAO · Table 2**

| Model | F1@0.01 ↑ | vIoU ↑ | Chamfer ↓ | EMD ↓ |
| --- | --- | --- | --- | --- |
| TRELLIS | 0.1475 | 0.1392 | 0.0902 | 0.2131 |
| Hunyuan3D-2.1 | 0.1399 | 0.1266 | 0.1126 | 0.2432 |
| Hunyuan3D-2.0 | 0.1574 | 0.1504 | 0.0866 | 0.2049 |
| Direct3D-S2 | 0.1513 | 0.1465 | 0.0962 | 0.2160 |
| TripoSG | 0.1533 | 0.1445 | 0.0844 | 0.2057 |
| Hi3DGen | 0.1629 | 0.1531 | 0.0937 | 0.2134 |
| SAM 3D | 0.2344 | 0.2311 | 0.0400 | 0.1211 |

F1: surface precision and recall at threshold 0.01. vIoU: volumetric overlap. Chamfer and EMD: geometric distances (lower is better). Raw paper values, not success percentages.

Source: [SAM 3D Team et al. · Table 2 · arXiv v1 (2025) ↗](https://arxiv.org/html/2511.16624v1#S4.T2)

SAM 3D scores 0.2344 on F1@0.01 in this comparison. These are paper-reported results on SA-3DAO, not a ranking of today’s commercial products or a benchmark run by Segmentit.

### Volume overlap

How closely the predicted volume matches the reference. Higher is better.

vIoU ↑0 → 0.25

**TRELLIS**

**0.1392**

**Hunyuan3D-2.1**

**0.1266**

**Hunyuan3D-2.0**

**0.1504**

**Direct3D-S2**

**0.1465**

**TripoSG**

**0.1445**

**Hi3DGen**

**0.1531**

**SAM 3D**

**0.2311**

[Published data · 2025 ↗](https://arxiv.org/html/2511.16624v1#S4.T2)

### Surface distance

Geometric distance to the reference surface. Lower is better.

Chamfer ↓0 → 0.12

**TRELLIS**

**0.0902**

**Hunyuan3D-2.1**

**0.1126**

**Hunyuan3D-2.0**

**0.0866**

**Direct3D-S2**

**0.0962**

**TripoSG**

**0.0844**

**Hi3DGen**

**0.0937**

**SAM 3D**

**0.0400**

[Published data · 2025 ↗](https://arxiv.org/html/2511.16624v1#S4.T2)

### Overall shape alignment

Distance between point distributions. Lower is better.

EMD ↓0 → 0.25

**TRELLIS**

**0.2131**

**Hunyuan3D-2.1**

**0.2432**

**Hunyuan3D-2.0**

**0.2049**

**Direct3D-S2**

**0.2160**

**TripoSG**

**0.2057**

**Hi3DGen**

**0.2134**

**SAM 3D**

**0.1211**

[Published data · 2025 ↗](https://arxiv.org/html/2511.16624v1#S4.T2)

### Reading the results

A historical comparison of the versions evaluated in the 2025 publications, reviewed on October 7, 2026. Datasets, prompts and metrics determine what is measured. These scores predict neither generation time nor the quality of every photo in Segmentit.

## From research to your next creation.

Bring a photo into Segmentit, isolate the object and launch its 3D generation. Inspect the result in the studio, then download a GLB for your next scene. Your selection and your model stay together in the project.

[Start creating](https://segmentit.com/app?lang=en) [Explore the workflow](https://segmentit.com/services/photo-to-3d/)

## Questions about SAM 3D

**SAM 3 or SAM 3D: which one do I need?**

Use segmentation to choose the subject in an image, then reconstruction to create its volume. SAM 3 and SAM 3D are complementary technologies, not two versions of the same task.

**What is SA-3DAO?**

Meta’s benchmark pairs real-world images with shapes made by professional 3D artists. Its release describes 100 public image/3D pairs and 900 held-out pairs for benchmark competition.

**How should I use a reconstructed object?**

Treat it as a creative asset: inspect every side, set the scale for your scene and adjust materials in your 3D tool. A single image gives a useful starting point, while unseen details remain inferred.

## Sources and documentation

-   [Meta · SAM 3D](https://ai.meta.com/research/sam3d/)
-   [Research paper · SAM 3D](https://arxiv.org/html/2511.16624v1)
-   [Official code · SAM 3D Objects](https://github.com/facebookresearch/sam-3d-objects)
-   [Meta · SA-3DAO dataset](https://ai.meta.com/datasets/sa-3dao-sam-3d-artist-objects-downloads/)

Reviewed October 7, 2026 · Segmentit editorial