LTX models (LTX-2, LTX-2.3, LTX-2.5) are multimodal video generation models: AI models that create and edit video from text prompts, still images or audio. They are made by LTX, a lab founded in 2024 that began as a division of Lightricks.
LTX-2.5 is a diffusion transformer, a neural network that turns random noise into video step by step under the guidance of the prompt, with 22 billion parameters and two streams that generate audio and video together. The models do text-to-video, image-to-video and audio-to-video, extend clips, regenerate parts of a video (retake), and upscale to HDR, with output up to 4K. LTX-2.5 adds multi-shot scenes that keep characters and lighting consistent across shots. The API offers each of LTX-2.3 and LTX-2.5 in a fast and a pro variant.
There are two ways to use them. The weights and code are published on GitHub and Hugging Face; LTX-2.5 is free for companies under US$10 million in revenue, with a commercial license for larger ones; they also run in the free LTX Desktop app. The LTX API is a hosted pay-per-use service billed per second of generated video.