mlx-video 0.1.0

Video loading and preprocessing utilities for MLX on Apple Silicon

mlx-video

mlx-video is a small video preprocessing library for MLX on Apple Silicon.

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

It focuses on the first mile of video ML:

• load video files into mlx.core.array
• sample frames at a target FPS
• resize clips for model input
• convert uint8 frames to normalized float32 tensors

Install

pip install mlx-video

Publishing

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

Quickstart

import mlx_video as mv

frames = mv.load("video.mp4", fps=8)         # mx.array [T, H, W, C], uint8 RGB
frames = mv.resize(frames, (224, 224))       # spatial resize
frames = mv.normalize(frames)                # float32, ImageNet normalized

API

load(path, fps=None, start_time=None, end_time=None, max_frames=None)

Loads a video into a mx.array with shape [T, H, W, C] in RGB order.

resize(frames, size)

Resizes each frame to (height, width).

to_float(frames, scale=True)

Casts frames to float32. By default values are scaled to [0, 1].

normalize(frames, mean=..., std=...)

Applies channel-wise normalization using ImageNet defaults. Integer inputs are automatically converted to floating point and scaled before normalization.