LTX-Video 2

Lightricks designed LTX-2 as a diffusion transformer that creates video and its soundtrack in one pass, so speech, effects, and ambience line up with the picture. The current checkpoint family, LTX-2.5, centers on a 22 billion parameter transformer in full and distilled versions, paired with a Gemma 4 based text encoder, separate video and audio VAEs, and latent upscalers that raise resolution or frame rate. The Python monorepo ships a core model library, a trainer for LoRA, IC-LoRA, and full fine-tuning, and ready pipelines covering text or image to video, keyframe interpolation, audio-driven video, regenerating a chosen time range, HDR output, and lip-matched dubbing. Lightricks also maintains ComfyUI nodes. Expect a heavy download, roughly 66 GB for the quick-start files, and a large NVIDIA GPU, though FP8 casting and CPU offload lower the bar. Code and weights both fall under the LTX-2.x Community License: free for people and companies under 10 million dollars in yearly revenue, while bigger firms must buy a commercial license. It supersedes the original LTX-Video repo and suits studios and researchers.