S

Sam Hq Vit Large

Developed by syscv-community
SAM-HQ is an enhanced version of the Segment Anything Model (SAM), capable of generating higher-quality object masks from input prompts such as points or boxes.
Downloads 60
Release Time : 5/5/2025

Model Overview

By introducing high-quality output tokens and global-local feature fusion components, SAM-HQ significantly improves the quality of segmentation masks, especially for objects with complex boundaries and fine structures.

Model Features

High-Quality Output Tokens
Introduces learnable HQ output tokens specifically designed for predicting high-quality masks, significantly improving segmentation accuracy.
Global-Local Feature Fusion
Combines early and final ViT features to fuse high-level semantic context with low-level boundary information, enhancing mask details.
Efficient Training
Requires only 4 hours of training on 8 GPUs, with less than 0.5% additional parameters compared to the original SAM.
Zero-Shot Generalization
Retains SAM's original zero-shot generalization capability while performing better on 10 datasets.

Model Capabilities

High-quality image segmentation
Prompt-based mask generation
Automatic mask generation
Complex boundary handling
Fine structure recognition

Use Cases

Image editing
Precise object segmentation
Used for accurately separating objects in image editing software
Generates finer mask boundaries compared to the original SAM
Automated annotation
Data annotation assistance
Automatically generates segmentation annotations for training data
Reduces manual annotation workload and improves annotation quality
Featured Recommended AI Models
AIbase
Empowering the Future, Your AI Solution Knowledge Base
Š 2025AIbase