Dev.to AI πŸ€– Ai πŸ‘ 2 πŸ“– 3 min read

Run AI video with sound locally on a Mac (Apple Silicon, MLX, no cloud)

A year ago, "AI video on a Mac" meant a cloud tab and a credit counter. Today an Apple Silicon Mac can render a short clip, picture and sound together, without a network connection. This is a practical walkthrough: what

Run AI video with sound locally on a Mac (Apple Silicon, MLX, no cloud)

A year ago, "AI video on a Mac" meant a cloud tab and a credit counter. Today an Apple Silicon Mac can render a short clip, picture and sound together, without a network connection. This is a practical walkthrough: what your Mac needs, what to expect, and how to get your first clip.

All the examples use Phosphene, a free, open-source (MIT) Mac app that wraps the open video models in a simple panel. The same ideas apply if you prefer to script the models yourself.

Why this works on a Mac at all

Video models are big. The base download for LTX-2.5 is about 27.5 GB. On a PC that has to fit in the graphics card's own memory. On Apple Silicon, the CPU and GPU share one pool of unified memory, so a 64 GB Mac can hold weights that would need a very expensive GPU elsewhere.

Apple's MLX framework is built around that shared memory. Arrays live in one place, the GPU reads them directly, and nothing gets copied back and forth. That is why MLX ports of these models run well on a Mac, and why memory, not chip speed, is the first thing to check.

What your Mac needs

  • Apple Silicon (M1 or newer). Intel Macs are not supported.
  • 16 GB of memory to start. That runs LTX-2.5 at smaller sizes.
  • More memory opens more settings: 48 to 64 GB is where video gets comfortable.
  • Disk space: about 27.5 GB for the base models, about 75 GB more if you add Hailuo H3, about 11 GB for the song model.

What to expect, roughly

Rounded typical times for a 5-second LTX-2.5 clip with sound at Balanced settings:

Mac Typical time
M5 Max about 1.5 minutes
M3 Ultra about 2 minutes
M4 Max about 3 minutes
M4 Pro, 48 GB about 6 minutes
M5, 16 GB (480p) about 5 minutes
M4, 16 GB (480p) about 11 minutes

The first render after launch is always slower because the engine loads and warms up. Time the second one.

Two engines, both with sound

The Video tab in Phosphene: modes on the left, the take on the right

LTX-2.5 (from Lightricks) turns text or a picture into video with sound in one pass. You describe the shot and what should be heard, and the audio comes out matched to the picture: footsteps, rain, a door, room tone. It also does keyframes, extending a take, lip-sync to your own audio, upscale and face fix.

Hailuo H3 (from MiniMax) is heavier and slower, and it can do audible dialogue. Write a line in quotes and the character says it. On an M4 Max, a 5-second H3 draft takes about 5 minutes. Its model licence has territory limits, so read it for your country before you download it.

Your first clip, step by step

  1. Download the Mac app from phosphene.video (or install through Pinokio if you already use it).
  2. Let the first download finish. Start appears once the base models are in.
  3. Open Video, choose Text, then Balanced and 5 s.
  4. Write the scene, the camera and the sound. For example:
A lantern boat drifts through a still fjord at dusk. Slow push in from the shore.
Water laps at the hull, a distant gull, wind over the water.
  1. Press Generate. Make a second take with the same prompt and compare.

Prompt tips that help on small Macs

  • Describe the sound explicitly. If you don't, the model guesses, and the guess is often generic hum.
  • Keep one action per shot. Five seconds is short; two actions usually blur into neither.
  • Start from a picture when you can. Image to video holds the look better than text alone.
  • Use Draft or Balanced while exploring, then re-render the keeper at High.

When local is the wrong choice

Be honest with yourself about the trade-offs. Local means big downloads, minutes per clip and a Mac that runs warm while it works. If you have 8 GB, or you need a minute-long shot at 4K, a hosted service is the better tool. If you want privacy, no credits and the freedom to make fifty takes of the same idea, local is hard to beat.

The code is on GitHub: mrbizarro/phosphene. If you try it, I'd love to hear which Mac you used and how long your second take took.

Mr. Bizarro

πŸ“° Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes β€” full credit and traffic to the original publisher.