Meta has just released SAM 3D — extending their famous Segment Anything Model into 3D generation. This is open source and completely free, making it a significant moment for the 3D AI community.
I deployed the inference on Azure A100s and tested all capacities of this model. Here's my honest take after extensive testing.
How SAM 3D Works
SAM 3D is part of the Segment Anything family. It takes images, segments objects, and then generates a 3D representation. But here's the key thing to understand:
If you use the Meta playground, you'll only get Gaussian Splats output. To get meshes with textures, you need to run the model locally.
Speed is Insane
One thing that immediately stands out: the generation speed is incredibly fast. This is a major advantage over many other 3D generation tools that can take minutes per generation.
System Requirements
Meta's documentation states 32GB VRAM minimum, but I managed to run geometry generation under 12GB VRAM — that includes Gaussian Splats and mesh output.
However, for the full pipeline with textures, you'll need more VRAM and Linux.
Quality: SAM 3D vs Hunyuan 2.1
Meta claims an 80% preference rate over other open-source models. I compared it directly against Hunyuan 2.1.
My honest assessment? It's not quite 80%. I'd say more like 60-70% preference — but it's definitely better than Hunyuan 2.1 in many cases.
What It's Good (and Bad) At
Works Well For
- • Furniture and props
- • Environmental objects
- • Simple geometric shapes
- • Quick prototyping
Struggles With
- • Characters and faces
- • Organic shapes
- • Text and fine details
- • High-res textures (max 2K, still blurry)
The textures tend to be blurry even at maximum 2K resolution. This is because the Gaussian Splats are converted to mesh with baked textures — the conversion process loses some detail.
The mesh itself is relatively simple, but the overall shape accuracy is quite good for an open-source model.
Ideal Use Cases
The use case for SAM 3D is somewhat narrow but specific:
- •VR/AR development — Quick environmental props
- •Unreal Engine developers — Rapid prototyping of scene objects
- •Gaussian Splat workflows — If you already work with splats
For traditional 3D workflows (Blender, Maya, etc.), the Gaussian Splat output can be a hurdle. Opening and converting splats isn't as straightforward as importing a GLB or OBJ file.
The Verdict
SAM 3D is a genuinely exciting release for the open-source 3D AI community. The speed is impressive, and for certain use cases — especially furniture, props, and environmental objects — it delivers solid results.
The limitations are real: Linux-only, Gaussian Splat output, blurry textures, and struggles with organic shapes. But the fact that this model exists, is open source, and is making progress is absolutely amazing for the field.
Bottom line: If you're an Unreal developer doing VR/AR props, or you already work with Gaussian Splats, SAM 3D is worth trying. For traditional 3D workflows, the other tools on our leaderboard might still be more practical.