LTX-2 is here. And completely open.
The first fully open-source text-to-video model with real production-grade quality. Download. Customize. Use.
Finally. A video AI model that isn't just called "open source," but truly is. LTX-2 by Lightricks gives you everything: model weights, code, training framework. Free. Runnable locally. Customizable.
Lightricks has released the complete LTX-2 stack. You get the main model and the distilled version. You can run both locally, fine-tune them, or deploy them into production. No cloud dependency, no API limits.
What makes it special: It's the first time a company has truly put everything out there. Model weights, LoRA adapters for reference, a modular training framework with multimodal pipelines, multi-GPU support, and evaluation tools. The full stack.
What you actually get
LTX-2 has been optimized for NVIDIA hardware. It runs exceptionally well on an RTX 5090. You can generate videos up to 20 seconds, in 4K, at up to 50 frames per second. That is cinema level.
Full Control
Define the first and last frame of your 20-second clips. Connect them into videos of any length. Storyboard control just like in a film studio.
Lip Sync & Audio
Precise lip synchronization and audio generation included. No post-processing required.
Multi-Keyframe Conditioning
Set multiple keyframes and let the model interpolate the intermediate frames. 3D camera logic for cinematic shots.
LoRA Customization
Train your own LoRA adapters for specific styles. Lightricks provides reference adapters for inspiration.
The training framework is modular. You receive multimodal data pipelines, reference configurations, and multi-GPU support. Everything you need to adapt the model to your specific purposes.
Local Power, Cloud Option
Download from GitHub or Hugging Face. Install locally. Or use a cloud provider if you lack the hardware. Both ways work.
LTX-2 has already claimed the number one spot on the Artificial Analysis Open Weights Leaderboard. The best video model with open weights. Period.
Lightricks has even published a research paper documenting all the techniques they used. Complete transparency across the board.
Why this matters
For the first time, you get a production-ready video model with no ifs or buts. No half measures, no hidden restrictions. You can download it, customize it, and use it commercially.
The generated videos show impressive quality. Realistic physics, diverse styles, compelling details. You can integrate it directly into ComfyUI.
Conclusion
LTX-2 democratizes professional video generation. No expensive cloud APIs needed anymore. No more black box. You are in control. Download the model, experiment, build with it. This is open source the way it should be.