mlx-serve/Deep dive · Video generation

Text to video. Photo to video.
Soundtrack to video.

LTX-Video 2.5, 2.3 and MiniMax-H3 run natively on Apple Silicon — clips with synchronized audio, muxed straight to an mp4 on your desk. Start from a prompt, from your own photo, or from a voice clip your characters perform to.

Download MLX Core Talking characters
Audio included — synced, in the mp4 Validated tensor-by-tensor vs the reference Nothing uploaded, ever
Three ways in

Prompt, photo, or soundtrack

Text → video

Describe the shot

Type a prompt, pick a quality preset, get a clip with synchronized audio. The one-stage path runs guidance-free by default — ~2× faster per step with a more natural look.

Photo → video

Animate your own picture

Drop an image into the First-frame slot and the clip begins exactly from it — VAE-encoded on-device and locked as the clean opening frame, then animated forward.

Audio → video

Perform to a soundtrack

Attach speech or music, and the video is generated against that clip — timing, performance, and lip sync follow the audio, and the original recording lands in the mp4, not a re-synthesis.

MLX Core Video generation pane with prompt, quality presets, first-frame slot, and Speech & sound section
The Video pane: prompt, presets, First-frame slot, and the Speech & sound section for soundtrack-driven clips.
Talking characters

Put the lines in quotes. They speak.

Write spoken words in quotes in your prompt — short phrases with acting directions between them — and LTX generates the voice, timed to the picture. Want a specific voice? Attach a real recording, or type a line and have the local Qwen3-TTS speak it — including a voice you cloned from a few seconds of reference audio.

Any WAV/MP3/M4A works, the frame count auto-fits the clip length, and audio guidance on the Quality presets steers toward clean speech — clearer voices, less stray background noise.

Video pane with a personal photo loaded in the First-frame slot and a speech clip attached
A photo in the First-frame slot plus a voice clip: your picture, performing your audio.
Quality presets

Fast when you're iterating, two-stage when it matters

The Fast preset runs the distilled one-stage pipeline. The Quality and Super-Quality presets run the full reference two-stage pipeline natively: a guided half-resolution pass on the dev model (CFG + modality guidance — Super adds a second-order sampler), a learned 2× latent upscale, then a distilled refine at full resolution.

Style LoRAs apply here too — the same runtime "filters on steroids" that restyle image generations restyle your clips, zero quality loss on the base weights.

  • Fast — distilled one-stage, guidance-free, ~2× faster per step.
  • Quality — guided dev pass → 2× latent upscale → distilled refine.
  • Super — same, with a second-order sampler on the guided pass.
  • LoRA restyle — attach a .safetensors, dial the strength.
Model lineup

LTX-Video 2.5 & 2.3, and MiniMax-H3

LTX-Video 2.5

Lightricks' newest, text encoder included

LTX-Video 2.5 ships in a 4-bit pack (36 GB) and an 8-bit quality pack (59 GB) with its own text encoder bundled, so there is no separate download on first use. The 8-bit pack keeps faces, legs and fur that 4-bit loses, and adds a Diffusion decoder toggle: the decoder Lightricks' own published clips use, which denoises the frames for sharper texture and edges. The default canvas and frame ladder are sized to your Mac's memory, and 2.3 keeps working right beside it.

  • Two packs — 4-bit 36 GB, 8-bit 59 GB, encoder bundled.
  • Diffusion decoder on the 8-bit pack ("decoder": "diffusion" over the API).
  • Canvas sized to your Mac — big machines get a big picture by default.
  • Everything works — text, photo, audio input, and the two-stage pipelines.
MiniMax-H3

Hailuo 3.0: the soundtrack is generated with the picture

MiniMax-H3 denoises the clip and a stereo soundtrack together in one pass, so the sound is produced with the video rather than dubbed on after. Describe the scene, then what you want to hear after overall_soundscape:. The REF2VA build composes the clip around pictures, clips or audio you attach (<Picture 1>, <Video 1>, <Audio 1> in the prompt), and Turbo renders in 4 steps instead of 30.

  • Video + stereo audio in one denoise, never dubbed after.
  • Reference pictures, clips and audio steer the shot (REF2VA).
  • Turbo — 4 steps instead of 30, about twice as fast end to end.
  • Long clips chain window by window (chain_windows), LoRAs stack on top.
Honest requirements

What it takes to run

Local video generation is the heaviest thing a consumer Mac can do. Here's the real bill — once.

24 GB
recommended unified memory
36–50 GB
one-time download — LTX 2.5 with encoder included, or 2.3 with every preset
1 click
download lives right in the Video pane, resumable
$0
per clip, forever after

The app checks free RAM before starting and offers to stop the chat server if it's competing for memory. Outputs land in ~/.mlx-serve/generations/videos/, by date. API users: POST /v1/video/generations with prompt, optional first_frame_image, audio, pipeline, lora_paths (up to 8, stacked), and on LTX 2.5 "decoder": "diffusion".

FAQ

Local video generation, answered

Does it really run on-device — no cloud render farm?

Really. The full LTX-Video pipelines (2.5 and 2.3) and MiniMax-H3 — diffusion transformers, 3D VAEs, audio decoders — were ported natively to MLX and validated tensor-by-tensor against the reference implementation. Your prompts, photos, and clips never leave the Mac.

How does the soundtrack feature work?

The audio clip is encoded on-device and held fixed while the video denoises around it, so motion and lip sync follow the sound. The mp4 gets your original recording at native quality — trimmed to the clip length, never re-synthesized.

Can I use a cloned voice for the character?

Yes — the Speech & sound section chains into the local TTS: type the line, pick (or clone) a voice with Qwen3-TTS, and the generated speech drives the video. All three models run through the same local server.

Why is the download so big?

The ~50 GB snapshot ships both transformer variants (one-stage distilled + two-stage dev, ~11 GB each), the upscaler, and the Gemma text encoder — so every quality preset works offline without re-downloading. The downloader pulls only the files the engine actually reads, not the repo's full ~70 GB. LTX 2.5 travels lighter: 36 GB in 4-bit or 59 GB in 8-bit, text encoder bundled.

Go deeper

More deep dives

Start here

New to local AI? →

The 5-minute tour — free, private, offline.

 Media

Image generation & editing →

FLUX.2 + Krea-2, instruction edits, style LoRAs.

 Media

Voice cloning & TTS →

Zero-shot cloning from seconds of audio.

 App

Always-on assistant →

⌃Space launcher, voice, Telegram, scheduled tasks.

 Coding agents

Claude Code, fully local →

Free, offline, private — point it at your own Mac.

 Comparison

LM Studio alternative →

Same weights, faster engine — every cell we benchmark.

 Safety

Agent Sandbox →

Agent shell commands in an isolated Linux VM.

A film set on your desk.

Download MLX Core, grab LTX-Video with one click, and turn a prompt, a photo, or a voice memo into a clip — audio and all.

Download for Mac View on GitHub