LTX-2 is Lightricks’ DiT-based audio–video foundation model — the first open model to combine synchronized sound and motion in a single architecture. The official repository provides the Python inference package and LoRA training code, with model weights (currently the LTX-2.5 generation) on Hugging Face. It has close to 10,000 stars on GitHub and is under active development through 2026.
Most open video models generate silent footage; LTX-2 generates audio and video together, so dialogue, ambience, and motion arrive in one pass. The repository ships multiple production-oriented pipelines: text/image-to-video, two-stage high-quality generation with a distilled model and spatial upscaler, IC-LoRA pipelines for video-to-video transformation, dubbing, keyframe animation, and even HDR output. Attention backends are selected automatically per GPU, and the codebase targets both datacenter and high-end consumer hardware, making it one of the most complete open video-generation stacks available. The package layout separates the core transformer, pipelines, and training utilities, so developers can adopt a single pipeline or build custom generation workflows on top of the model.
Freemium. The LTX-2 code and open weights are available for self-hosting under the LTX-2 community license. Lightricks also sells hosted access through the LTX platform and API, plus the LTX Studio creative product, for those who prefer not to run GPUs themselves.
Video-generation researchers, AI filmmakers, and product teams who want an open, self-hostable audio–video model with production pipelines — and the option to fine-tune it with LoRAs for a house style.