RUST ยท APPLE SILICON VIDEO AI

mlx-video-rs โ€” Apple Silicon Diffusion Transformer (DiT) Video Engine in Rust

mlx-video-rs is a pure Rust port of mlx-video for generating high-definition video clips from text and image prompts on Apple Silicon Macs. It accelerates Wan2.1, LTX-Video, and CogVideoX diffusion transformers using Metal unified memory, 4-bit quantization, and tiled 3D-VAE decoding.

$ cargo install mlx-video-rs
GitHub โ†—

Built from the ground up for Metal

Combining the safety and speed of Rust with Apple MLX unified memory architecture.

โšก

Apple Silicon Unified Memory

Zero-copy checkpoint streaming from SafeTensors directly to Metal GPU unified RAM without duplicate CPU memory allocations.

๐ŸŒ€

3-Way Factorized 3D RoPE

Rotary Position Embeddings decomposed across temporal (F), height (H), and width (W) dimensions for fast spatial-temporal attention.

โฑ๏ธ

Flow Matching Schedulers

High-precision Euler and 2nd-order DPM++ 2M continuous vector-field integration with dynamic sample shifting.

๐ŸŽž๏ธ

Causal 3D VAE & Tiling

CausalConv3d decoding preserving temporal continuity, combined with spatial/temporal tiling to eliminate VRAM OOM errors on long generations.

๐Ÿงฉ

LoRA Dynamic Weight Fusion

Automatic rank/alpha extraction and instant on-the-fly matrix injection (ΔW = s • α/r • B × A) with multi-LoRA stacking support.

๐Ÿ–ผ๏ธ

Native Image-to-Video

Lanczos downsampling and 36-channel latent concatenation (16 noisy + 16 image + 4 mask) for Wan2.2 I2V animation.

Model Matrix & Presets

Comprehensive support across both Wan and LTX diffusion video transformer families.

Model Name Category Parameters DiT Dim Layers Heads Default Steps
Wan2.1 T2V (1.3B) Text-to-Video 1.3 Billion 1536 30 12 40 steps
Wan2.1 T2V (14B) Text-to-Video 14.0 Billion 5120 40 40 40 steps
Wan2.2 T2V (14B) Text-to-Video (Dual) 14.0 Billion 5120 40 40 40 steps
Wan2.2 I2V (14B) Image-to-Video 14.0 Billion 5120 (in=36) 40 40 40 steps
Wan2.2 TI2V (5B) Text+Image-to-Video 5.0 Billion 3072 32 24 40 steps
LTX-2 Distilled Multi-Stage Flow Variable 2048 28 32 8 steps
LTX-2 Dev Multi-Stage Flow Variable 2048 28 32 40 steps

Interactive CLI Command Generator

Configure generation parameters and generate ready-to-run mlx-video terminal commands.

Generated Terminal Command
โšก Runs natively on Apple Metal unified memory GPU.

Hardware Recommendations

Memory allocations scaled for Apple Silicon unified RAM.

๐Ÿ’ป

16GB โ€“ 24GB RAM

Recommended: Wan2.1-1.3B & LTX-2 Distilled at 768x512 with 49โ€“65 frames.

โšก

36GB โ€“ 48GB RAM

Recommended: Wan2.2-TI2V-5B and Wan2.1-14B (4-bit quantized / tiled) at 832x480 with 81 frames.

๐Ÿš€

64GB โ€“ 128GB+ RAM

Recommended: Full uncompressed 14B Dual models, Wan2.2 I2V, and 720p HD long-sequence video generation.