Lightricks/LTX-2

Project overview

An open-source Python inference and LoRA trainer package for the LTX-2 DiT-based audio-video generative model. It enables developers to generate production-ready, synchronized audio and video from text, images, or existing video clips.

Project type

Model Runtime · Video · Audio & Speech

Use cases

Video Creation

Deployment

Refer to project documentation

License

License pending

Best for

• Developers and AI engineers seeking an open Python library for DiT-based audio-video generation.
• Creators needing localized inference and fine-tuning tools for generating video content with synchronized audio.

Key capabilities

• Provides tools for training and fine-tuning models using LoRA, full fine-tuning, and IC-LoRA techniques.
• Inference pipelines include an enhance_prompt parameter to automatically process and improve input prompts before generation.

Limitations and risks

• Inference is constrained by GPU memory capacity, requiring users to implement FP8 quantization or CPU/Disk offloading on hardware with limited VRAM.
• The system requires specialized hardware (GPUs) and manual installation of attention optimizations, such as FlashAttention, to achieve documented performance levels.

Getting started

• Setup involves cloning the repository, downloading multi-gigabyte LTX-2.3 models and the Gemma 3 12b text encoder using the HF CLI, authenticating with a Hugging Face token, and running specific pipeline commands.

Evidence and sources

• GitHub project description: Official Python inference and LoRA trainer package for the LTX-2 audio–video generative model.
• README: LTX-2 is the first DiT-based audio-video foundation model that contains all core capabilities of modern video generation in one model: synchronized audio and video, high fidelity,…
• README: hf download Lightricks/LTX-2.3 \ ltx-2.3-22b-distilled-1.1.safetensors ltx-2.3-spatial-upscaler-x2-1.1.safetensors --local-dir models/ltx-2.3 hf download google/gemma-3-12b-it-qat…
• README: uv run python -m ltx_pipelines.distilled
• README: In cases of GPU memory constraints, consider `--quantization fp8-cast --offload {cpu, disk}`.

Reviewed at: 2026-07-16T10:52:38Z