leaxer-sam

CLI wrapper for Segment Anything Model (SAM), optimized for cross-platform deployment using ONNX Runtime.

Features

• Cross-platform GPU acceleration: CUDA (NVIDIA), DirectML (AMD/Intel on Windows), CoreML (Apple Silicon)
• Small binary size: ~50-100MB (vs ~5GB with PyTorch)
• Multiple prompt types: Box, point, and automatic segmentation
• Multiple model variants: MobileSAM for speed, SAM ViT-B/L/H for quality

Installation

Download the pre-built binary for your platform from the Releases page:

Platform	Binary	GPU Support
Windows	leaxer-sam-x86_64-pc-windows-msvc.exe	CUDA, DirectML
Linux	leaxer-sam-x86_64-unknown-linux-gnu	CUDA
macOS (Apple Silicon)	leaxer-sam-aarch64-apple-darwin	CoreML
macOS (Intel)	leaxer-sam-x86_64-apple-darwin	CPU

Usage

Box-prompted Segmentation

leaxer-sam segment \
  --image photo.jpg \
  --boxes "[[100,50,300,280],[500,400,620,550]]" \
  --output masks/ \
  --json result.json

Point-prompted Segmentation

leaxer-sam segment \
  --image photo.jpg \
  --points "[[200,150],[550,450]]" \
  --labels "[1,1]" \
  --output masks/ \
  --json result.json

Automatic Segmentation

leaxer-sam auto \
  --image photo.jpg \
  --points-per-side 32 \
  --min-area 100 \
  --output masks/ \
  --json result.json

Arguments

Argument	Required	Default	Description
--image	Yes	-	Path to input image
--boxes	No	-	Bounding boxes as JSON array [[x1,y1,x2,y2],...]
--points	No	-	Point prompts as JSON array [[x,y],...]
--labels	No	-	Point labels (1=foreground, 0=background)
--model	No	mobile_sam	Model variant
--output	Yes	-	Output directory for mask PNG files
--json	Yes	-	Output JSON path for metadata
--points-per-side	No	32	Grid density for auto mode
--min-area	No	100	Minimum mask area for auto mode

Output Format

{
  "image_size": [1920, 1080],
  "segments": [
    {
      "id": 0,
      "bbox": [100, 50, 300, 280],
      "confidence": 0.98,
      "mask_path": "masks/mask_0.png"
    },
    {
      "id": 1,
      "bbox": [500, 400, 620, 550],
      "confidence": 0.95,
      "mask_path": "masks/mask_1.png"
    }
  ]
}

Models

Models are downloaded separately by the Leaxer app and stored in:

• Windows/macOS: ~/Documents/Leaxer/models/sam/
• Linux: ~/.local/share/Leaxer/models/sam/

Model	Encoder	Decoder	Description
mobile_sam	~40MB	~15MB	Fast, lightweight
sam_vit_b	~350MB	~15MB	Good balance
sam_vit_l	~1.1GB	~15MB	High quality
sam_vit_h	~2.3GB	~15MB	Highest quality

GPU Acceleration

ONNX Runtime automatically selects the best available execution provider:

Platform	GPU Provider	Supported GPUs
Windows	CUDAExecutionProvider	NVIDIA
Windows	DirectMLExecutionProvider	NVIDIA, AMD, Intel
Linux	CUDAExecutionProvider	NVIDIA
macOS	CoreMLExecutionProvider	Apple Silicon (M1/M2/M3)
All	CPUExecutionProvider	Fallback

Building from Source

Requirements

• Python 3.10+
• PyInstaller

Build

# Install dependencies (choose one)
pip install -r requirements/cuda.txt  # NVIDIA GPU
pip install -r requirements/cpu.txt   # CPU only

pip install pyinstaller

# Build
pyinstaller pyinstaller/leaxer-sam.spec --distpath dist

Converting Models (CI only)

To convert PyTorch models to ONNX:

pip install -r requirements/convert.txt
python convert/export_onnx.py --model mobile_sam --output models/
python convert/export_onnx.py --model sam_vit_b --output models/

License

MIT License - see LICENSE for details.

This project wraps Segment Anything Model by Meta AI (Apache-2.0 License) and MobileSAM.