Lara-LTX-2.5-MLX-BF16

Native MLX/Metal BF16 runtime for local LTX-2.5 video and synchronized-audio generation on Apple Silicon. No CUDA or server process is required on Mac.

Release 0.0.5 matches the GitHub source tag v0.0.5.

This repository is a release descriptor. The exact official gated BF16 files remain in Lightricks/LTX-2.5; accepting upstream access is required. The Lara loader follows lara_ltx_model.toml to that pinned source revision and does not publish an unnecessary second copy of the official checkpoint.

Quick Start

Requirements: Apple Silicon, macOS, Python 3.12, ffmpeg, 128 GB unified memory, and accepted access to the gated upstream model. The packaged profile defaults to the measured 512x320/17-frame workload and rejects unverified memory or Stage-2 token grids before checkpoint download or loading.

git clone https://github.com/azossy/Lara-LTX-2.5-MLX-BF16.git
cd Lara-LTX-2.5-MLX-BF16
uv sync
export HF_TOKEN="your_read_token"

from lara_ltx import LTXPipeline

pipe = LTXPipeline.from_pretrained("LaraAI-Labs/Lara-LTX-2.5-MLX-BF16")
video = pipe(prompt="A cinematic aerial shot of Seoul at night.", seed=42)
video.save("output.mp4")

The CLI and optional thin ComfyUI adapter invoke this same in-process Python pipeline. See the GitHub README for measured hardware results, installation, configuration, known limitations and verified parity evidence.

CUDA vs MLX: matched 10-second cinematic demo

Both native outputs use the same prompt, negative prompt, seed 25082110, 512×320 source grid, 241 frames, 24 fps and 15 inference steps. The scene is a photorealistic close-up of an adult East Asian dancer with wind-driven hair, facial micro-expressions and eye acting. Cross-backend stochastic generation is not expected to produce byte-identical or composition-identical videos.

4K side-by-side presentation

The side-by-side file has CUDA on the left and MLX/Metal on the right. It carries separate CUDA and MLX audio tracks. All 4K files are deterministic Lanczos presentation upscales to 3840×2160, not native 4K model generations, and were never used as evaluator inputs.

Native output	Video	4K presentation
CUDA	Play/download	Play/download
MLX/Metal	Play/download	Play/download

Measured generation time

Backend run	Elapsed	Generated frames/s
CUDA	74.05 s	3.2547
MLX/Metal	2,221.30 s	0.1085

CUDA was 30.00× faster in these two observed runs. This is a deployment result, not a controlled hardware benchmark: the runs used different systems and the table must not be read as an intrinsic CUDA-versus-Metal hardware ratio.

Objective evaluation on native outputs

VBench metric	CUDA	MLX/Metal
Subject consistency	0.8786	0.8799
Background consistency	0.9314	0.9151
Motion smoothness	0.9866	0.9877
Dynamic degree	1.0000	1.0000
Aesthetic quality	0.5204	0.5641
Imaging quality	0.6663	0.6288

VideoScore2 (1–5)	CUDA	MLX/Metal
Visual quality	4	4
Text alignment	4	5
Physical/common-sense consistency	4	4

Audio metric	CUDA	MLX/Metal
Audiobox CE	3.0175	7.6048
Audiobox CU	5.5600	8.3217
Audiobox PC	2.3645	5.6478
Audiobox PQ	6.1328	8.3602
LAION CLAP cosine	0.2976	-0.0937

The metrics measure different properties and are not combined into one winner. Inspect the videos directly alongside the raw, revision-pinned reports: VideoScore2, CUDA VBench, MLX VBench, CUDA Audiobox, MLX Audiobox, CUDA CLAP, and MLX CLAP.

Status

This is an engineering preview, not a final CUDA-quality-parity claim. Exact CUDA inputs verify the checkpoint-metadata-driven transformer input path and final transformer/output-head component boundaries. The v6 CUDA trace contains 408 named block-0 Attention boundaries across all six modules and three guidance passes; payload deduplication stores 118 unique tensors plus 290 aliases. All 176 block-0 MLX exact-input sub-operation comparisons pass the frozen component gate. The v7 sequence trace replays blocks 24-39 from the verified block-23 boundary, reproduces all 12 CUDA source outputs exactly, and adds passing 176-operation MLX reports at both blocks 31 and 39. Their worst NRMSE values are 0.00396 and 0.00395.

CUDA-aligned FP32 RMSNorm and RoPE accumulation reproduce the captured AdaLN boundary exactly. The complete stochastic replay ends at video/audio latent NRMSE 0.1916/0.1110; the deep exact-input results rule out a discrete attention implementation or mapping failure, while cross-backend BF16 rounding differences still accumulate. The frozen final-latent gate remains open without weakening its thresholds. All four MLX quality-corpus cases and paired CUDA diagnostics are complete. A non-blind frame-sequence audit accepts candidate usability but confirms that same-seed composition differs materially, so independent blind parity is not claimed. The reviewed decoder profile completes 512x512/33 frames in 92.0 seconds at a 40.25 GB MLX peak instead of selecting the old pathological 51,200-tile plan. Resource-policy failures use localized LARA-RUNTIME-010 guidance and never silently quantize or select a fallback.

License

This derivative is subject to the complete LTX-2.x Community License Agreement, including Section 4 and Attachment A. Recipients must comply with all use restrictions. Commercial Entities, as defined by the agreement, must obtain the required paid license before commercial use. The full license and modification notice are distributed with the source release.