vanch007/LTX-2.5-mlx

LTX-2.5 MLX (Gemma4 Q4 + Transformer Q8)

This repository contains a local Apple-Silicon MLX conversion of the gated Lightricks/LTX-2.5 checkpoint. It is a derived conversion for MLX; it is not an official Lightricks release. Review and comply with the original model license and usage policy before redistributing or using the files.

Contents

• gemma4-q4/: Gemma4 Unified, affine 4-bit, group size 64, layer-sharded.
• transformer-q8/: 48-block LTX-2.5 AV Transformer, affine 8-bit, group size 64 on eligible Transformer Linear layers; non-block parameters remain BF16.
• Root media files: Conv video VAE encoder/decoder, audio VAE, vocoder/BWE, and the v1.0 latent x2 spatial upscaler.

The Transformer index and Gemma index are included. Conversion manifests record the source revision, tensor inventory, quantization policy, and completed shards. No original BF16 Transformer payload is included.

Local MLX usage

Clone the companion runtime from vanch007/ltx-2.5-mlx, download this model repository, and either place the files in the runtime's expected layout or point the local assembly/generation wrapper at the downloaded directories. The runtime supports one-stage and two-stage T2AV, repeatable I2V image anchors, and generated-keyframe conditioning.

The conversion was validated locally on an M3 Max/128 GB Mac:

• 48/48 deployed-Q8 Transformer blocks pass PyTorch-vs-MLX cumulative parity.
• One-stage, two-stage, I2V, multi-anchor, and generated-keyframe MP4 smoke paths pass.
• The largest 480x704x97, 8-step benchmark completed in 619.57 s with 17.05 GiB peak footprint and zero swap.

These checks do not claim full-pipeline pixel or waveform equivalence with the official PyTorch implementation. The full 42 GiB BF16 Transformer source sweep was intentionally skipped; only the converted Q8 artifacts are published.

Reproduce the conversion

See the runtime repository's conversion scripts and the pinned source manifest. The source model is gated and requires Hugging Face access approval. The local conversion used the LTX-2.5 source revision recorded in each conversion-manifest.json.