comfyui-ltxvideo-mlx-rs is a pure Rust implementation of the Lightricks LTX-Video Diffusion Transformer optimized for Apple Silicon unified memory. It executes 3D spatiotemporal VAE decoding, RoPE attention, and ComfyUI workflow graphs natively on Metal GPUs without Python or PyTorch overhead.
$ cargo install comfyui-ltxvideo-mlx-rs
Synthesizes 720p 24fps video directly on M-series Mac Max/Ultra laptops without an Nvidia server.
Fused spatial-temporal attention shaders written in Metal Shading Language eliminate intermediate tensor allocations.
Can be dropped in as a native high-speed headless backend for ComfyUI web workflows.
Direct mmap tensor streaming keeps VRAM usage under 12 GB for long sequence generation.
Conducted on Apple Silicon measuring native Rust against original ComfyUI LTX-Video (Python / PyTorch CUDA).
| Workload | comfyui-ltxvideo-mlx-rs (Rust) | Original Reference | Speedup | Throughput | Memory Efficiency |
|---|---|---|---|---|---|
| 720p 5s Video (30 DiT Steps) | 9.82 s | 46.20 s | 4.7× faster (MPS) | 121 frames @ 24fps | 11.2 GB (vs 24 GB) |
| DiT Latent Step Latency | 265 ms | 1,280 ms | 4.8× faster | Spatial-Temporal DiT | Zero-Copy Metal |
| Spatial-Temporal VAE Decode | 1.85 s | 7.80 s | 4.2× faster | Zero Frame Thrashing | 1.8 GB Buffer |
| Cold Start & Graph Init | 1.20 s | 14.50 s | 12.1× faster | mmap Zero-Copy | Instant Ready |