S

Slimsam 50 Uniform

Developed by nielsr
SlimSAM is a compressed version of the Segment Anything (SAM) model, significantly reducing parameters and computational requirements through pruning and distillation techniques while maintaining high-quality object segmentation capabilities.
Downloads 430
Release Time : 1/7/2024

Model Overview

SlimSAM is a compressed version based on the SAM model, capable of generating high-quality object masks from input prompts like points or boxes. Through an innovative pruning-distillation framework, it achieves performance close to the original SAM with extremely low training costs.

Model Features

Efficient Compression
Parameter count is only 0.9% of original SAM, computation reduced to 0.8%, significantly lowering resource requirements
Low-cost Training
Training costs reduced by over 10x compared to existing methods, requiring only 10k training images (0.1% of original data)
Alternate Slimming Strategy
Innovative pruning-distillation framework that finely optimizes model structure through progressive steps
Label-free Pruning Criterion
Pruning objectives aligned with optimization goals, enhancing distillation effects post-pruning

Model Capabilities

Prompt-based Image Segmentation
Automatic Mask Generation
Object Recognition & Segmentation
Zero-shot Segmentation

Use Cases

Computer Vision
Interactive Image Editing
Quickly segment objects in images using point or box prompts
Generate high-quality object masks
Automatic Image Annotation
Automatically generate masks for all objects in images without manual intervention
Supports batch processing to improve annotation efficiency
Resource-constrained Environments
Mobile Image Processing
Achieve real-time object segmentation on mobile devices like smartphones
Low computational requirements enable deployment
Featured Recommended AI Models
AIbase
Empowering the Future, Your AI Solution Knowledge Base
© 2025AIbase