Meta's SAM 3D turns flat images into three-dimensional reconstructions. Upload a photo and it generates full 3D models with accurate geometry, texture, and positioning — both for objects and people.
The technical approach splits into two systems. One handles objects using a two-stage DiT architecture. The other reconstructs human bodies with a transformer-based encoder-decoder that captures pose and shape, even when parts of the body are hidden. Both work from single images, which is tougher than it sounds.
Does it actually work? Meta claims beyond state-of-the-art performance across benchmarks. They built the SAM 3D Artist Object Dataset specifically to test real-world scenarios with diverse objects that challenge existing 3D systems. The models handle occlusion well and can place multiple reconstructed elements into shared scenes with proper spatial relationships.
You can prompt the system for joint reconstructions and visualize results through 3D AR overlays. The outputs aren't just static meshes — they're independent, posed models you can manipulate and interact with.
The Facebook Marketplace integration makes practical sense. Sellers could generate 3D product views without special equipment. Beyond e-commerce, this fits use cases in physical therapy assessment, robotics training, and AR development.
What's missing from the facts, and any mention of public access. No API details. No information about whether this is a research preview or production tool. The technical performance sounds solid but availability determines whether you can actually use it.
Meta tested this against challenging real-world images. That's more valuable than perfect results on clean datasets. For researchers and developers building 3D reconstruction pipelines, the architecture details suggest serious engineering. Whether it's accessible to your project remains the open question.