ltx-mlx

ltx-mlx

LTX-Video generation optimized for Apple Silicon. Pure MLX — no PyTorch, diffusers, or transformers at runtime.

Supports LTX-Video 2B (fast) and LTX-Video 13B (high quality) with automatic model detection.

Performance

Benchmarked on M3 Ultra (76 GPU cores, 256 GB). PyTorch numbers use optimized MPS pipeline with bf16 + patched SDPA.

Install

pip install -e .

Dependencies: mlx, numpy, pillow, sentencepiece, imageio[ffmpeg]. No PyTorch needed.

Model Setup

Download the LTX-Video checkpoints from Hugging Face:

models/LTX/
├── ltxv-2b-0.9.8-distilled.safetensors          # 2B model (5.9 GB)
├── LTX 13B/
│   └── ltxv-13b-0.9.8-distilled.safetensors      # 13B model (27 GB)
├── text_encoder/                                   # T5v1.1-XXL (shared)
│   ├── config.json
│   ├── model-00001-of-00004.safetensors
│   ├── model-00002-of-00004.safetensors
│   ├── model-00003-of-00004.safetensors
│   ├── model-00004-of-00004.safetensors
│   └── model.safetensors.index.json
└── tokenizer/                                      # T5 tokenizer (shared)
    ├── spiece.model
    └── tokenizer_config.json

Both models share the same T5 text encoder and tokenizer.

Usage

CLI

# 2B model — fast (~7s)
ltx-mlx --prompt "a cat sitting on a windowsill, golden hour"

# 13B model — sharp (~60s)
ltx-mlx --prompt "a portrait, cinematic" --model 13b

# 720p for maximum quality
ltx-mlx --prompt "ocean waves crashing" --model 13b --resolution 720p

# Custom resolution + settings
ltx-mlx --prompt "a dancer" --height 544 --width 960 --steps 8 --seed 7 -o dance.mp4

Python API

from ltx_mlx import LTXPipeline

# 2B (fast)
pipe = LTXPipeline("models/LTX")
frames = pipe.generate("a cat on a windowsill", seed=42)

# 13B (high quality)
pipe = LTXPipeline("models/LTX", checkpoint="LTX 13B/ltxv-13b-0.9.8-distilled.safetensors")
frames = pipe.generate(
    "a woman smiling, soft studio lighting, portrait",
    height=720, width=1280,
    num_frames=41,
    seed=42,
)

# Save video
import imageio
imageio.mimwrite("output.mp4", frames, fps=24, codec="libx264")

Resolution Guide

Higher resolutions need fewer frames to fit in memory. Recommended:

• 480p: up to 97 frames (~4s)
• 720p: up to 41 frames (~1.7s)
• 1080p: up to 17 frames (~0.7s)

Architecture

Prompt → T5v1.1-XXL Encoder → Text Embeddings
                                    ↓
Noise → DiT Transformer (28 or 48 blocks) → Denoised Latents
                                    ↓
         VAE Decoder (timestep-conditioned) → Video Frames

• T5 Encoder: 4.8B params, sentencepiece tokenizer (~0.1s)
• DiT (2B): 28 layers, 2048 hidden, 32×64 heads (~1.1s/step)
• DiT (13B): 48 layers, 4096 hidden, 32×128 heads (~7.0s/step)
• VAE Decoder: 553M params, timestep-conditioned 3D convolutions (~2.5s)
• Scheduler: FlowMatch Euler with dynamic sigma shifting

Zero-copy weight loading via mx.load (memory-mapped safetensors).

License

Apache-2.0. Model weights are subject to Lightricks' license.