LTX-2 is a Python package for inference and training with an audio-video generative model that produces synchronized audio and video outputs from text prompts.
The tool addresses the challenge of generating high-fidelity video with matching audio by implementing a diffusion transformer architecture. It provides multiple performance modes through different model variants: a full 22-billion-parameter transformer for highest quality and a distilled version that runs in fewer steps for faster generation. The package includes several specialized pipelines for different use cases, such as guided two-stage generation and audio dubbing. For users with GPU memory constraints, the tool offers quantization and offloading options to reduce memory requirements during inference.
Developers should choose this tool if they need to generate video with synchronized audio in production settings. The package suits projects requiring high-fidelity outputs where audio-video synchronization is essential. The distilled model variant makes it accessible for faster iteration when quality can be traded for speed. Users will need approximately 66 gigabytes of storage for the model weights, which are distributed as separate component files so only necessary parts need downloading. The text encoder is a fine-tuned version of Gemma 4 bundled with the model, and the package automatically selects appropriate backends for different operating systems and hardware configurations.
The project maintains active development with regular updates to the model architecture and pipeline implementations. The codebase includes comprehensive documentation for optimization strategies and common configuration flags. The tool provides clear error handling for authentication and model version mismatches, ensuring users load compatible components. The package supports multiple inference backends with automatic fallback mechanisms across different platforms, indicating attention to broad hardware compatibility.
