Back to Gap Map Model components / Image, video, 3D & music generation

SAM 3D Objects

Meta
open weights / Overall score: 2.2

SAM 3D Objects is Meta's model for turning an object in a single photograph into a full 3D shape with texture and layout. It was built for real-world pictures where objects are cluttered or partly hidden, and was released in November 2025 together with SAM 3D Body, a separate model for human body meshes.

SAM 3D Body recovers human body meshes, which is reconstruction rather than generation, so it is not part of this entry; the Segment Anything segmentation models are a separate product. Encoder weights were added in June 2026.

Openness

3 high confidence
3.0
weights
open (checkpoints on the Hub behind a manual access request)
data
described (the README describes progressive training with a human-feedback data engine; no training set is released)
code
partial (inference code and a demo only; no training code)
license
SAM-License (the checkpoints and code, per the README and the LICENSE dated November 19, 2025)

SAM 3D Objects' code and weights are under Meta's SAM License, which allows commercial use and limits conduct and trade-controlled uses rather than the number of users or revenue. The weights need an approved access request on Hugging Face. Only inference code is published, and the training data is described but not released.

Adoption

1 high confidence
1.0

Adoption is measured as Hugging Face downloads of facebook/sam-3d-objects, the only first-party copy; the separate SAM 3D Body checkpoints are left out.

Capability

3 medium confidence
3.0

SAM 3D Objects places in the upper half of the image-to-3D arena, level with InstantMesh and just behind Hunyuan3D-2 and 2.1. It is ranked only as a Gaussian-splat output, so the topology board does not list it.

Verified 2026-11-12