Pro
EN
Sign inAccount Download for Mac
Download for Mac
Journal / 2026-11-12

How to run LTX on a Mac: LTX‑2.5 step by step, M1 to M5

Run LTX-2.5 locally on your Mac with Phosphene. Measured render times for every chip from M1 to M5, what your memory allows, and the steps to a clip with sound.

A diver with a lamp in the ballroom of a sunken ocean liner, a frame from a clip made with LTX-2.5 on a Mac
by Mr Bizarro
tutorial · ltx-2.5 · mac · apple-silicon

Give LTX-2.5 a sentence or a picture and it makes a short clip with sound. Ask for footsteps in the rain and it makes those alongside the picture. It can make a voice, too. This open video model from Lightricks runs well on a Mac, too.

We make Phosphene, a free and open-source Mac app that runs LTX-2.5. The render times below are measured.

What your Mac needs

Chip. Any Apple Silicon M1 or newer. Intel Macs are not supported.

Memory. The model needs room to work in unified memory. Check how much you have under Apple menu → About This Mac, on the Memory line.

MemoryWhat you can do with LTX-2.5
16–47 GBText to video and picture to video at smaller sizes. Macs with 24 GB or less use a low-memory path automatically.
48–79 GBEvery mode, including Keyframes, Extend and the High quality tier, up to 768 px on the long side for Keyframes and Extend.
80 GB and upMost modes at full size. From 120 GB, no size limits.

A standard 1280×704 render peaks at about 22 GB of memory in use. More free memory lets you ask for a larger frame. Every memory size and chip, with disk sizes and render times, is on Will my Mac run it?.

Disk. The base set is about 28 GB: the LTX-2.5 engine, its text encoder and a preview decoder. At 100 Mbps, allow roughly 37 minutes to download it. At 300 Mbps, about 12.

Time. A Mac takes minutes to render a video. There is time to put the kettle on.

How long a clip takes, by chip (M1 to M5)

These are median times for one 5-second clip at Balanced quality and 480p, with sound, from anonymous render counts by Phosphene users between 10 August and 22 October 2026. Model loading is included. Your first clip after opening the app takes longer because the model has to load.

ChipMemory5-second clip
M116 GB18 min 30 s
M216 GB16 min 40 s
M316 GB9 min 30 s
M416 GB10 min 40 s
M424 GB7 min 50 s
M516 GB4 min 55 s
M1 Pro16 GB6 min 40 s
M2 Pro16 GB4 min 30 s
M3 Pro18 GB6 min 30 s
M4 Pro24 GB3 min 50 s
M5 Pro24 GB1 min 20 s
M1 Max32 GB3 min 40 s
M2 Max32 GB3 min
M3 Max36 GB2 min 40 s
M4 Max36 GB2 min 15 s
M5 Max36 GB1 min

Some rows come from only a handful of Macs, so treat them as a guide. Every chip, size and length is on the render times page. Larger frames and longer clips take more time. Before you press Generate, the app shows an estimate for your Mac next to each length.

A Pro or Max chip matters more than the year. A 16 GB M1 Pro beat a 16 GB M4 by four minutes. Pro and Max chips have more graphics cores and more memory bandwidth, and video models need both.

Memory sets what you can make, not just how fast. A 16 GB M4 and a 24 GB M4 make the same clips. The 24 GB one was about three minutes quicker.

MacBook Air and Mac mini count. They use the base chips, like the M4 above. Ten minutes for a 5-second clip with sound calls for some patience, but it works. On a smaller Mac, start with Quick and 3 seconds to find the shot: that had a median of 6 min 10 s on a 16 GB M4. Then render the keeper at Balanced.

The manual ways

If you want to set LTX up yourself, two routes work on a Mac today:

  1. The reference Python code on PyTorch. Lightricks publishes the model and code. It runs through PyTorch’s Metal backend on a Mac, though it was written for NVIDIA cards. Expect some patching: Metal doesn’t support the 64-bit floats some calculations ask for, and FP8 checkpoints won’t load at all. You’ll need a recent Python and a development build of the libraries. Budget plenty of memory, too (64 GB is the comfortable starting point).
  2. A community MLX port. Several open-source LTX ports run from the command line on MLX, Apple’s machine-learning framework. They run faster and use less memory than the PyTorch route. Quantised (4-bit) weights fit a 16 GB Mac. You handle the Python environment and download the models yourself, then set the command-line flags. The clip becomes visible when it’s finished.

Both work if you like setting things up yourself. Phosphene is built on an MLX port (credited on our open source page) and handles that setup for you.

The easy way: Phosphene

1. Check your Mac

Apple menu → About This Mac. M1 or newer and 16 GB or more: you’re set.

2. Download Phosphene

Go to the download page and click Download for Mac. Open the disk image from your Downloads folder and drag Phosphene to Applications. It’s free. The page asks for an email for release news, then starts the download. There’s no account to create. The same disk image is attached to every GitHub release.

3. Download the video models

Open Phosphene from Applications. It lists every model and its size, alongside the disk space needed and available. Nothing downloads until you confirm. The base set is about 28 GB, and a progress bar shows how it’s going.

If the download stops, start it again. Finished files are kept. You can also set up the app now and download the models later if they won’t fit yet.

4. Write a prompt

On the Video tab, keep the model on LTX 2.5 and the mode on Text. Describe the action and where the camera is. Then give the sound a sentence of its own. LTX makes it from the same prompt:

The camera glides slowly forward through the ballroom of a sunken 1920s ocean liner. A diver in a brass helmet sweeps a warm hand lamp over a crooked crystal chandelier and a grand piano half buried in sand. Audio: muffled underwater ambience, slow breathing inside the helmet, rising bubbles, no music.

The Phosphene Video tab with LTX 2.5 and a text prompt. The finished clip plays on the right

That is a shorter version of the prompt used for this 10-second clip from our gallery. The same prompt made both picture and sound:

5. Pick a quality and length, then Generate

Start with Balanced and 5 seconds. Check the summary under the form for the size and estimated time on your Mac, then click Generate. You can watch the shot take shape in the Now panel. The finished clip lands in Outputs, sound included. Nothing is uploaded.

6. Try a picture instead

Switch the mode to Image, drop in a photo and describe the motion. Choose Anchor to use your picture as the first frame. The clip moves on from there. A phone photo works.

Picture to video in Phosphene: a still of a fjord with lanterns as the first frame, and the moving clip in the player

The other modes are a click away. Try Start & end frame or Keyframes, or use Extend to carry on a clip you like. Lip-sync matches an original character to your own audio. Each mode is covered in the Video guide.

Common problems and fixes

The render stops partway with a memory message. Your Mac ran out of memory. Close heavy apps, including that browser full of tabs, chat apps or the iOS Simulator, and try again. A smaller size or shorter length can help, too.

The first clip is much slower than the estimate. The first render after opening the app loads the model into memory. The second one is faster.

The download stalled or the disk filled up. Downloads keep 15 GB of your disk free and pause when space runs low. Free some space and click download again. Finished files are kept.

Some modes are missing. Phosphene hides what your Mac can’t run. Keyframes, Extend and High appear from 48 GB of memory.

The sound is not what you wanted. Be specific about what you want to hear. Add “no music” or “no speech” when you don’t want them.

You searched for LTX-2.3. LTX-2.5 is the newer version of the same family, and it’s what Phosphene uses to make clips. The app downloads LTX-2.3 only when you train your own character.

Supporting the project, with faster renders

Everything above is free. If you’d like to support Phosphene, Pro also speeds up LTX clips on Macs with 32 GB or more. A 5-second clip took 113 seconds instead of 152 on an M4 Max with 64 GB, using the same clip and seed. The speed-up is on by default, with a tick box beside Quality to render the usual way.