Sam3D

sam3d is Meta’s single-image 3D reconstruction model that fuses SAM 3 open-vocabulary segmentation with geometry, texture, and layout predictions. It turns ordinary photos into ready-to-use 3D assets. As a research-grade model, it reconstructs 3D geometry, texture, and layout from just one image, extending the Segment Anything family to enable text or visual prompts to isolate objects and quickly convert them into 3D assets.

How to Use

To use sam3d, first capture a single, well-lit RGB image. Optionally, apply SAM 3 with a text or box prompt to isolate the target object. Next, run inference using the released checkpoints and code; sam3d will predict geometry, texture, and layout directly. Finally, export the mesh/texture for deployment into AR viewers, 3D engines, robotics simulators, or marketing experiences.

Key Features

• Single-image 3D reconstruction (geometry, texture, layout)
• Open-vocabulary segmentation using SAM 3 prompts (text, points, boxes)
• Open ecosystem with checkpoints, inference code, and benchmarks
• XR (AR/VR) ready for immersive applications
• Efficient inputs, compatible with legacy photos and user-generated content

Use Cases

• Creative production (games, CGI, social content) by scanning props from photos
• E-commerce & AR shopping for 'view in room' experiences with product shots
• Robotics & autonomy to provide 3D priors and infer shape from cameras
• Medical & scientific visualization to turn 2D scans into 3D forms