mlx-vision

Image preprocessing and Apple Vision framework interop for MLX on Apple Silicon.

mlx-vision provides two layers:

• MLX-native image loading and preprocessing for model pipelines
• optional Apple Vision framework integration for OCR, face detection, classification, and saliency

GitHub: https://github.com/AmiraniLabs/mlx-vision

Install

Base install:

pip install mlx-vision

With Apple Vision support:

pip install mlx-vision[vision]

Development install:

pip install mlx-vision[dev]

Full install:

pip install mlx-vision[all]

Quickstart

import mlx_vision as mv

image = mv.load("photo.jpg")
image = mv.resize(image, (256, 256))
image = mv.center_crop(image, (224, 224))
image = mv.normalize(image)
image = mv.to_chw(image)

Batch Loading

import mlx_vision as mv

batch = mv.load_batch(["a.jpg", "b.jpg"], size=(224, 224))
batch = mv.normalize(batch)
batch = mv.to_chw(batch)

Apple Vision Integration

import mlx_vision as mv

if mv.is_vision_available():
    faces = mv.detect_faces("portrait.jpg")
    text = mv.detect_text("document.jpg")
    classes = mv.classify_image("photo.jpg")
    saliency = mv.attention_saliency_map("photo.jpg")

API

Core image preprocessing:

• load(path, color="rgb")
• load_batch(paths, color="rgb", size=None)
• resize(image, size, interpolation="linear")
• center_crop(image, size)
• pad(image, padding, fill=0, mode="constant")
• horizontal_flip(image)
• to_float(image, scale=True)
• normalize(image, mean=..., std=...)
• to_chw(image)
• to_hwc(image)

Apple Vision framework:

• is_vision_available()
• detect_faces(image)
• detect_text(image, recognition_level="accurate", languages=None)
• classify_image(image)
• saliency_map(image, kind="attention")
• attention_saliency_map(image)

Positioning

What mlx-vision is:

mlx-vision is the image and computer vision entry point for MLX pipelines on Apple Silicon. It provides model-ready image transforms and optional access to Apple's native Vision framework with outputs compatible with MLX workflows.

What it is not:

• not a full replacement for OpenCV
• not a training framework
• not a visionOS SDK
• not a general multimedia library

Publishing

This package is configured for PyPI Trusted Publishing from the AmiraniLabs/mlx-vision GitHub repository using .github/workflows/publish.yml.