Add support for LTX-2.5

[boxabirds]

LTX 2.5 was released last week, is richer and faster and has better memory management I believe. It'd be awesome to see an mlx-video to support this amazing update!

[kiarina]

I implemented LTX-2.5 distilled support and verified it on a Mac Studio M4 Max with 128 GB unified memory.

The implementation preserves the existing mlx_video.models.ltx_2 design and LTX-2/LTX-2.3 compatibility while adding the official split checkpoints, Gemma 4 Unified, the 22B transformer, the convolutional VAE, and LTX-2.5's ancestral Euler stage-1 sampler.

The current scope is T2V and I2V. Audio, the diffusion VAE, the duration predictor, and DFR remain follow-up work.

Measured at 768x512, 121 frames:

• T2V: 108.7 seconds, 37.81 GB peak
• I2V: 120.8 seconds, 39.54 GB peak

The related 47 tests pass, as does installation into a fresh Python 3.12 environment.

PR: #52

[kiarina]

Audio support is now implemented as a stacked follow-up to #52.

The added scope is:

• Synchronized audio-video generation
• Audio-to-video
• Combined audio and image conditioning
• The split LTX-2.5 audio VAE and BigVGAN + BWE vocoder

At 768x512 and 121 frames, generated audio took 118.1 seconds with a 37.81 GB peak, while A2V took 106.4 seconds with the same peak memory.

Stacked PR: #53

[kiarina]

Update: Audio support has been integrated into #52 so the implementation can be reviewed in one place. #53 is being closed to avoid duplicate review.

#52 is now the canonical PR for the core LTX-2.5 implementation, including audio generation and A2V.