SAM 3D Objects is a foundation model developed by Meta's Superintelligence Labs that reconstructs complete 3D shape geometry, texture, and layout from single images. The model is specifically designed to handle real-world scenarios involving occlusion and clutter, employing progressive training and a data engine incorporating human feedback to achieve robust performance. According to the repository documentation, it outperforms prior 3D generation models in human preference tests on real-world objects and scenes.
The repository is written in Python and addresses the intersection of 3D object segmentation, computer vision, and point cloud processing. It encompasses multiple related domains including image annotation, deep learning, object detection, semantic segmentation, multi-view analysis, 3D reconstruction, and machine learning. The codebase provides functionality for both single and multi-object 3D generation, allowing users to convert masked objects in images into complete 3D models with pose, shape, texture, and layout information.
The project includes substantial supporting materials and resources. The repository provides access to code, model weights, an online demonstration, a research paper, and a new challenging benchmark for evaluating 3D reconstruction performance. Recent updates tracked in the repository show that encoder weights were released on June 1, 2026, followed by the release of a 3D Artist Object Set and Hugging Face Leaderboard on June 2, 2026. The initial checkpoint launch, web demo, and paper publication occurred on November 19, 2025. Installation instructions and setup documentation are provided in the repository's doc directory, with quick-start functionality available through a demo.py script.
The project maintains active development and community engagement. GitGenius tracking shows a median issue and pull request response latency of 12.3 hours across 125 tracked items, with a mean response time of 141.7 hours. The most active contributors tracked by GitGenius are gleize with 80 events, alexsax with 40 events, and luffy-yu with 19 events. The repository connects to other major open-source projects through overlapping contributors, including nuxt/nuxt, huggingface/transformers, and huggingface/huggingface_hub.
SAM 3D Objects is part of a larger SAM 3D initiative that includes a companion model called SAM 3D Body, which focuses on single-image 3D human mesh recovery. The repository provides example notebooks demonstrating how to combine results from both models within a unified frame of reference. The project includes comprehensive contributor acknowledgments and maintains clear licensing under the SAM License, along with established contributing guidelines and code of conduct documentation. The model is designed to be robust in challenging natural images, handling small objects, unusual poses, and occlusions commonly found in uncurated natural scenes.
