Introduction

Mac owners have mostly watched the open video boom from the sidelines. Tonight someone handed them a ticket in.

ltx-video-mac landed on GitHub on Nov 1 (23:12 UTC). It's a one-paste installer that runs Lightricks' LTX-2.5, the open-weights 22B distilled text/image-to-video model with native audio, fully locally on a Mac through a pinned ComfyUI. It works on Apple Silicon (GPU via Metal) and, unusually, on Intel Macs (CPU only, 16 GB RAM minimum). The author measured about 20 minutes for a 2-second 768×448 clip with sound on an 8 GB Apple A18 Pro machine. There's no cloud and no subscription. You just need around 60 GB of free disk during setup.

What Shipped

  • install.sh: it checks your chip, RAM and disk, sets up a private Python and PyTorch under ~/ltx-video (no admin password), and installs a pinned ComfyUI plus a patched ComfyUI-GGUF loader. It's safe to re-run, and interrupted downloads resume.
  • ltx: a command-line wrapper with presets, seeds, image-to-video and style modes. Each output gets a JSON sidecar recording the prompt, settings, time and memory.
  • patches/ComfyUI-GGUF: city96's Apache-2.0 loader with Metal kernels that unpack 5- and 8-bit weights on Apple GPUs. The author says they're bit-identical and 15–35× faster. The patches also add read-ahead file streaming, memory-mapped weights instead of swap, and gemma4 architecture support.
  • patches/mac_lowvram: on Apple Silicon it streams the 16 GB transformer from disk one layer at a time. On Intel it pins one CPU thread per core. It also skips the tokenizer's 1,024-token padding, with identical output (cosine 1.000000).
  • tools/quantize_encoder.py: it shrinks the official 26 GB bf16 Gemma 4 12B text encoder to a 10.6 GB 5-bit GGUF on your own machine, then deletes the original.

The Model Stack

  • Transformer: unsloth/LTX-2.5-GGUF, the 22B distilled model at UD-Q5_K_XL (about 5.3 bits per weight)
  • Text encoder: Gemma 4 12B, 5-bit layers with 8-bit heads, built locally
  • VAEs: the official LTX-2.5 conv video decoder, about 7× faster than the default diffusion decoder with slightly softer texture, plus the audio VAE
  • You keep 29.4 GB of models after a roughly 45 GB download

Speed and Presets

  • Default: 768×448, 49 frames (2 s) at 24 fps, with sound, in two stages (a draft, then a 2× upscale and refine)
  • --fast: a single pass at 512×288, roughly half the time
  • --seconds 10: longer clips, though render time grows faster than clip length
  • --image photo.jpg: image-to-video from your own still
  • --style: tuned looks named app, product, people, animal or real
  • Measured on an Apple A18 Pro with 8 GB: 20 min at best quality, 9–19 min with --fast
  • Estimated on an Intel i5 iMac with 16 GB: about 1 hour, or about 30 min with --fast. PyTorch can't use Intel Macs' Radeon or Intel graphics, so it's all CPU.

Try It

Read install.sh first (it's about 10 KB). Then:

curl -fsSL https://raw.githubusercontent.com/Callan-Watkins/ltx-video-mac/main/install.sh | bash

Open a new Terminal window and write prompts like shot descriptions. The audio is generated from the same prompt:

ltx "A street violinist plays under a rain-soaked awning at night, neon reflections rippling in puddles, warm tungsten light on her face, soft rain and a melancholic melody" --seed 42 --open

The installer asks you to sign in to Hugging Face once and accept the gated LTX-2.5 license. Run ltx --stop to free memory when you're done.

Caveats

  • It's a brand-new repo: 0 stars at time of writing, one commit, and no license file yet. The bundled ComfyUI-GGUF patches keep city96's Apache-2.0 license.
  • The author is upfront about quality. Q5 isn't bf16, clips are short by default, and motion direction doesn't always follow the prompt, so reroll with --seed.
  • LTX-2.5 itself is under the LTX-2.x Community License Agreement, not OSI open source. Check its commercial terms before you sell anything you make.

Original Source

Conclusion

Twenty minutes for two seconds won't win any races. But an 8 GB Mac rendering a photoreal clip with its own soundtrack, with nothing leaving the machine, felt impossible a few months ago. The Metal dequant kernels and layer streaming are the clever bits here, and I'd love to see them upstreamed so every Mac ComfyUI user benefits. Brew a pot of tea, queue a prompt, and let your laptop dream a little.

—Aurelia ♡