Peter Zhang
Nov 21, 2025 06:51
Meta introduces Segment Anything Model 3, enhancing image and video segmentation with new features. The launch includes the Segment Anything Playground for interactive experimentation.
Meta AI has announced the release of Segment Anything Model 3 (SAM 3), an advanced model designed to enhance the capabilities of image and video segmentation. This new model introduces innovative features such as text and exemplar prompts, enabling comprehensive detection, segmentation, and tracking across diverse visual concepts, according to Meta AI.
Enhanced Segmentation Features
SAM 3 is equipped with promptable concept segmentation, allowing users to define and segment visual concepts using text or exemplar prompts. Unlike traditional models limited by fixed label sets, SAM 3 supports open-vocabulary segmentation, addressing a wider array of user requests. This flexibility is expected to significantly improve the model’s performance in both image and video contexts.
Segment Anything Playground
Alongside SAM 3, Meta has introduced the Segment Anything Playground, a platform enabling users to experiment with SAM 3’s capabilities. This tool is designed to simplify the application of advanced AI models for media modification, making it accessible to a broader audience without the need for technical expertise.
Applications and Collaborations
SAM 3 and its counterpart, SAM 3D, are being integrated into various Meta platforms, including Facebook Marketplace and Instagram’s Edits app. These integrations aim to enhance user experience by enabling new creative effects and visualization tools, such as the ‘View in Room’ feature for home decor items.
In collaboration with Conservation X Labs and Osa Conservation, Meta is leveraging SAM 3 for wildlife monitoring, contributing to a publicly available video dataset for research purposes. Similarly, partnerships with organizations like FathomNet are advancing AI tools in ocean exploration.
Technical Innovations
The SAM 3 architecture builds on previous AI advancements, incorporating components like the Meta Perception Encoder and the DETR model. These enhancements have resulted in a significant leap in segmentation performance, with SAM 3 doubling performance metrics on the Segment Anything with Concepts (SA-Co) benchmark.
Meta has also developed a scalable data engine using SAM 3 and human annotators, dramatically improving the speed and efficiency of data annotation. This hybrid system has enabled the creation of a diverse training set with over 4 million unique concepts.
Future Exploration
While SAM 3 demonstrates strong performance, Meta is committed to further advancements, particularly in challenging visual domains. The open-source release includes fine-tuning approaches, encouraging the community to adapt SAM 3 for specific use cases.
The Segment Anything Playground provides a user-friendly interface for exploring SAM 3’s capabilities, offering practical templates for media modification. This platform exemplifies Meta’s vision of democratizing advanced AI tools for creative expression.
Image source: Shutterstock