Can I combine SAM with text prompts?

The core SAM model uses points and boxes. To segment via text, you typically pair this skill with a detection model like GroundingDINO to generate the initial bounding box from text.

What makes the Segment Anything Model (SAM) unique?

SAM is a foundation model trained on over 1.1 billion masks, allowing it to segment objects in images it has never seen before (zero-shot) based on simple user prompts like clicks or boxes.

Can I use this for real-time video segmentation?

While SAM is optimized for images, it can be used for video frame-by-frame. For native video support with temporal consistency, SAM 2 is the recommended successor.

Which SAM model size should I choose?

ViT-B (375MB) is best for speed and edge deployment; ViT-L (1.2GB) offers a balance of speed and accuracy; ViT-H (2.4GB) provides the highest precision but requires more VRAM.

Does SAM require a GPU?

While SAM can run on a CPU, it is highly recommended to use a CUDA-enabled GPU for acceptable inference speeds, especially when using the larger ViT-H model.

Segment Anything Model (SAM)

Name: Segment Anything Model (SAM)
Author: Orchestra-Research

byOrchestra-Research

•

3,983

•

数据科学与机器学习

Implements zero-shot image segmentation using Meta AI's SAM to identify and extract objects via points, boxes, or automatic mask generation.

The Segment Anything Model (SAM) skill provides a comprehensive implementation framework for Meta AI's foundation model for computer vision. It enables developers to perform high-quality object segmentation on any image domain without the need for task-specific fine-tuning. This skill is particularly useful for building interactive annotation tools, automating image processing pipelines, and generating precise masks for complex visual data using flexible prompts like coordinate points or bounding boxes.

主要功能

01Zero-shot transfer to new image domains without retraining

02Multi-mode prompting via points, bounding boxes, and existing masks

033,983 GitHub stars

04Integration patterns for HuggingFace Transformers and ONNX deployment

05Automatic 'Everything' mode for generating all masks in an image

06Support for ViT-B, ViT-L, and ViT-H model architectures

使用场景

01Automating object extraction and background removal for creative tools

02Analyzing specialized visual data in medical, satellite, or industrial imaging

03Building AI-assisted image annotation and data labeling platforms

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add orchestra-research/ai-research-skills segment-anything

For use in Claude.ai and ChatGPT

Download Skill