SAM 3D: Behind the Two-Model Design | AI at Meta
•November 20, 2025
0
Why It Matters
By combining learned priors and preference optimization, SAM 3D promises more practical and reliable 3D reconstruction, unlocking improved e-commerce experiences and robotics interaction while paving the way for scalable, interactive 3D worlds. This could broaden commercial and research use cases that geometry-only approaches could not address reliably.
Summary
Meta's SAM 3D uses a two-model approach—one specialized for 3D human body reconstruction and a second generic model for 3D object reconstruction—to bring recognition and prior knowledge into areas where geometry-based methods fall short. The team borrowed preference optimization techniques from large language models to capture human-sensitive cues like symmetry, achieving robust results with relatively small datasets. SAM 3D is tuned to handle challenging cases—small, distant or heavily occluded objects—and the researchers say it enables more reliable reconstructions than traditional metrics alone. Meta envisions applications ranging from Facebook Marketplace product previews to robotics navigation and broader interactive 3D environments.
Original Description
SAM 3D enables accurate 3D reconstruction from a single image, supporting real-world applications in editing, robotics, and interactive scene generation.
Matt, a SAM 3D researcher, explains how the two-model design makes this possible for both people and complex environments.
🔗 Read the SAM 3D Objects research paper: https://go.meta.me/8c08ca
🔗 Read the SAM 3D Body research paper: https://go.meta.me/5e60ed
--
Follow us on Twitter: https://twitter.com/aiatmeta
Follow us on Facebook: https://www.facebook.com/aiatmeta
Meta focuses on bringing the world together by advancing AI, powering meaningful and safe experiences, and conducting open research.
0
Comments
Want to join the conversation?
Loading comments...