SAM 3D is Meta's family for reconstructing 3D information from images. For a creator preparing a prop, start with SAM 3D Objects: choose the object in a source image, prepare its mask, run the supported reconstruction workflow and inspect the output from more than the original camera view.
This independent guide focuses on preparation and acceptance. We inspected the official repository and provide original input and asset-review worksheets. We did not run model inference, download gated weights or produce a reconstructed mesh for this article. The cover is independently generated editorial art, not a SAM 3D output.
Choose SAM 3D Objects or SAM 3D Body
Meta's introduction distinguishes SAM 3D Objects for object and scene reconstruction from SAM 3D Body for human pose and shape estimation. A product prop and a human body are different tasks. SAM 3, without the D, is a segmentation model rather than the same reconstruction operation.
| Your task | Relevant starting point | First acceptance question |
|---|---|---|
| Reconstruct a selected prop | SAM 3D Objects | Does the selected object reconstruct usefully beyond the source view |
| Estimate a human body | SAM 3D Body | Does the recovered body suit the intended downstream task |
| Select an object in pixels | Segmentation and mask preparation | Does the mask select the intended subject |
| Make a finished campaign image | Image generation or editing | Does the exported picture meet the brief |
Do not treat these operations as interchangeable. A clean mask is useful input, but it is not a reconstructed object. A beautiful rendered picture also does not prove that an editable 3D asset exists.
Start with the official implementation
Use the official SAM 3D Objects repository for the implementation and notebooks. A similarly named hosted site is not automatically Meta's official service. Record the repository revision and checkpoint used with your own result so a later test has a clear starting point.
The documented setup specifies Linux 64-bit and an NVIDIA GPU with at least 32 GB of VRAM. It also describes requesting checkpoint access and authenticating for the download. That is not a promise of a native Mac installation or instant access. We did not execute the environment on this workstation.
Read the linked SAM License and applicable model-access conditions for your intended use. Repository visibility does not establish unrestricted reuse. Keep real access tokens out of shared prompts, worksheets and published notebooks.
Prepare one object before a whole scene
For a first creator exercise, choose a single owned prop with a recognizable outline. A watering can is useful because its body, handle opening and spout provide distinct inspection points. Photograph it clearly enough that the selected object can be separated from the background, and keep the untouched image.
A single view leaves hidden surfaces uncertain. Decide whether that uncertainty is acceptable for a distant cinema prop, a rough scene blockout or the intended asset. Exact product geometry and readable branding require stronger review than a background shape. Do not assume a plausible hidden side is the actual manufactured object.
Download the original single-prop input brief. It defines the reference role, mask check and acceptance views without pretending to contain a completed reconstruction. Fill in your own source, checkpoint, execution environment and results.
Check what the mask loader actually reads
At the inspected repository revision, the notebook mask loader turns values above zero into selected pixels and uses the last channel when the image has multiple channels. For an opaque RGBA file, that can make the alpha channel select the entire image even if its RGB picture looks like a black-and-white mask.
Prepare the intended binary selection in a format compatible with the exact loader you use. A simple single-channel black-background and white-subject mask avoids the ambiguity of an unrelated opaque alpha channel. Check that its pixel dimensions and orientation match the source. Inspect openings such as a handle hole rather than just the outside outline.
This observation comes from reading the input code, not from running a reconstruction benchmark. A correctly shaped input mask still does not guarantee an accurate 3D result. If you use a wrapper or another notebook, inspect that implementation's input handling as well.
Keep the first reconstruction traceable
Follow the current official notebook for the operation you need. Keep the image, mask, repository revision, checkpoint, settings and returned files together. First confirm the existing job completed and its outputs can be opened before launching another variation.
Our blank run record leaves execution and timing fields empty. Record actual observations rather than substituting estimates from another machine. No speed, memory-use or success-rate result is prefilled.
Distinguish a Gaussian splat from a mesh
The official quickstart demonstrates saving a Gaussian splat as PLY. A file extension alone does not establish a polygon mesh with faces, UVs or a rig. Check which output representation your notebook returned and whether the downstream application supports it.
For a mesh-based scene, inspect the actual mesh export path and open the resulting file in the intended application. Do not promise that renaming a PLY to GLB or FBX performs a conversion. Keep the original output alongside any converted or cleaned asset so changes remain inspectable.
Review the prop from views the source did not show
Turn the asset around. For a watering can, inspect the handle opening, spout attachment, underside and the transition between the body and lid area. Check whether thin structures disappear, merge or become implausibly thick. A matching front silhouette can conceal a poor side view.
Download the blank asset acceptance worksheet. Record front, side, rear and underside observations, representation, import status and the intended use. Add scale, topology or material checks when those are requirements of the actual downstream workflow. The sheet is a review aid, not a certification of geometric accuracy.
For cinema, test the asset in the intended camera and lighting before calling it ready. For close product shots, compare visible shape and lettering with the owned source. When a defect matters, correct or replace the asset before expanding the sequence.
Choose the next workflow by the deliverable
For another image-to-3D route, compare the preparation and export decisions in our TRELLIS 2 workflow and Hunyuan 3D prop guide. Those are separate model workflows, not interchangeable checkpoints or independent comparative benchmarks.
If you need a finished visual rather than an editable 3D prop, explore Clout image and video workflows or plan a scene in AI Cinema. This guide does not claim Clout hosts SAM 3D inference. Choose the tool that returns the actual deliverable, then review that file against your brief.
