LTX released LTX-2.5, the latest version of its open-weight world model, on August 11, 2026. The model is designed for video and audio generation, real-time generation, and robotics use cases, with native multi-shot generation, 4K HDR support, and stronger prompt understanding. Model weights were published on Hugging Face, and ComfyUI support was available from the first day of release.

LTX-2.5 rebuilds much of the generation pipeline and introduces a new Diffusion Video Decoder, a generative component for video reconstruction. The change is intended to reduce visual breakdowns in scenes with large motion and improve details such as faces and textures. Earlier models generally generated one continuous shot, while LTX-2.5 can generate a scene made of multiple cuts at once and aims to keep characters, backgrounds, lighting, voices, and visual style consistent across cut changes.

The text encoder uses Gemma 4 12B tuned for LTX. LTX said the model improves adherence to complex prompts involving multiple people, camera work, lighting, and actions, and can expand short instructions into more detailed prompts for video generation. It also supports Diffusion Fidelity Rendering, which allocates compute according to scene complexity, automatic duration prediction from prompts, native 4K HDR, and RAW workflows. The official model page says LTX also provides distilled models that aim for quality close to the full model while using less compute.

Hugging Face hosts both the 22B full model and distilled models. Users can run the models in their own environments and fine-tune them with their own data. The official LTX-2 GitHub repository also provides a Python inference environment and a LoRA training environment.

LTX-2.5 is distributed under LTX’s own “LTX-2.x Community License.” Organizations with annual revenue under $10 million can use it for commercial and production purposes at no cost under certain conditions, while companies above that threshold need a separate license. Open-weight means model weights are available, but it does not necessarily mean all training data or the full training process is public, or that the model can be reproduced from scratch.

LTX describes LTX-2.5 as a world model that can be used not only for video generation but also for “Physical AI,” meaning AI that operates in the real world, such as robots. The Hugging Face Model Card states that LTX-2.5’s established use cases are video and audio generation, while robotics and Physical AI applications are “developing.”

LTX released LTX-2.5, the latest version of its open-weight world model, on August 11, 2026. The model is designed for video and audio generation, real-time generation, and robotics use cases, with native multi-shot generation, 4K HDR support, and stronger prompt understanding. Model weights were published on Hugging Face, and ComfyUI support was available from the first day of release.

LTX-2.5 rebuilds much of the generation pipeline and introduces a new Diffusion Video Decoder, a generative component for video reconstruction. The change is intended to reduce visual breakdowns in scenes with large motion and improve details such as faces and textures. Earlier models generally generated one continuous shot, while LTX-2.5 can generate a scene made of multiple cuts at once and aims to keep characters, backgrounds, lighting, voices, and visual style consistent across cut changes.

The text encoder uses Gemma 4 12B tuned for LTX. LTX said the model improves adherence to complex prompts involving multiple people, camera work, lighting, and actions, and can expand short instructions into more detailed prompts for video generation. It also supports Diffusion Fidelity Rendering, which allocates compute according to scene complexity, automatic duration prediction from prompts, native 4K HDR, and RAW workflows. The official model page says LTX also provides distilled models that aim for quality close to the full model while using less compute.

Hugging Face hosts both the 22B full model and distilled models. Users can run the models in their own environments and fine-tune them with their own data. The official LTX-2 GitHub repository also provides a Python inference environment and a LoRA training environment.

LTX-2.5 is distributed under LTX’s own “LTX-2.x Community License.” Organizations with annual revenue under $10 million can use it for commercial and production purposes at no cost under certain conditions, while companies above that threshold need a separate license. Open-weight means model weights are available, but it does not necessarily mean all training data or the full training process is public, or that the model can be reproduced from scratch.

LTX describes LTX-2.5 as a world model that can be used not only for video generation but also for “Physical AI,” meaning AI that operates in the real world, such as robots. The Hugging Face Model Card states that LTX-2.5’s established use cases are video and audio generation, while robotics and Physical AI applications are “developing.”