LTX-2.5-MLX

Unofficial Apple Silicon MLX conversion of the official Lightricks/LTX-2.5 distilled checkpoint.

This repository contains model weights only. It is intended for local inference on Apple Silicon with the MLX port of LTX. The checkpoint was converted from the official BF16 release; the distilled transformer and Gemma 4 text encoder are stored in 8-bit MLX format, while the remaining components retain BF16 precision.

Included components

• LTX 2.5 distilled video/audio transformer
• Gemma 4 LTX text encoder and modality connectors
• Video encoder and decoder VAE
• Audio VAE and vocoder
• Spatial and temporal latent upscalers
• LTX 2.5 model configuration

The optional prompt rewriter is intentionally not included. Prompt enhancement uses a separate instruction-tuned multimodal model such as FrankDase/Gemma-4-E2B-it-MLX.

Capabilities

• Text-to-video
• Image-to-video with an optional start frame
• Image-to-video with an optional end frame
• Interpolation between start and end frames
• Audio-to-video
• Combined image and audio conditioning
• Synchronized video and audio generation
• Distilled two-stage generation with latent upscaling

Support for individual modes depends on the runner and pipeline version used.

Requirements

• Apple Silicon Mac
• Native ARM64 Python environment
• MLX 0.31 or newer
• A compatible LTX 2.5 MLX runtime

The model was tested locally on a Mac with an M4 processor and 64 GB unified memory. Lower-memory systems may require reduced resolution, frame count, and a low-RAM execution mode.

Basic example

With a compatible ltx-2-mlx installation:

caffeinate -dimsu /usr/bin/arch -arm64 \
  ltx-2-mlx generate \
  --model "/path/to/LTX-2.5-MLX" \
  --distilled \
  --low-ram \
  --prompt "A red fox walking through fresh snow, cinematic natural light" \
  --height 320 \
  --width 448 \
  --frames 97 \
  --frame-rate 24 \
  --seed 42 \
  --output "output.mp4"

LTX 2.5 expects frame counts in the form 8n + 1, for example 9, 97, or 121 frames. Distilled inference uses its dedicated few-step schedule; guidance settings intended for the full development checkpoint should not be applied.

Prompting

LTX 2.5 works best with a detailed, continuous description covering the scene, subject motion, lighting, camera behavior, and sound. Short prompts can be expanded before generation with the separate Gemma 4 prompt enhancer linked above. For image-to-video, the reference image and text prompt should both be provided to the multimodal prompt enhancer.

Model provenance

• Original model: Lightricks/LTX-2.5
• MLX runtime: MrMoferFRAN/ltx-2-mlx
• Conversion format: MLX Safetensors
• Transformer: distilled, 8-bit
• Text encoder: Gemma 4 LTX encoder, 8-bit

This is a format conversion, not a fine-tune. No training data was added and the underlying model behavior remains attributable to the original LTX 2.5 release.

License

The converted weights remain subject to the LTX-2.x Community License. Review the original license before using, redistributing, or deploying this model. In particular, eligibility and commercial-use conditions depend on the terms defined by Lightricks.

This repository is not affiliated with or endorsed by Lightricks.