An open-weight world model for video generation, physical AI and real-time applications, with support for local deployment and domain-specific fine-tuning.
LTX has released LTX-2.5, an open-weight world model designed for applications including film, robotics and real-time world generation. The model provides publicly accessible weights, allowing developers to customise and run visual simulations locally. LTX says its models have already recorded more than 33 million downloads.
The introduction of the Diffusion Video Decoder in LTX-2.5 aims to minimise visual artefacts during high-motion shots without compromising the compression capabilities of previous LTX models. LTX-2.5 also supports native multishot generation, creating sequences while maintaining consistency in characters, setting and voice across cuts. A customised Gemma 4 language backbone and prompt enhancer were utilised for the processing of complex, multi-character prompts.
A pretrained checkpoint is included in LTX-2.5 to enable teams to fine-tune the model using datasets from their domains. A distilled version is also available, designed to deliver similar output quality at lower cost and with reduced inference times.
LTX-2.5 is available in ComfyUI through a launch partnership, allowing developers to deploy the model using its node-based workflow. Local deployment gives users greater control over the hardware, generated data and IP, while also allowing the model to be customised for specific requirements.
The model has open weights on the platform HuggingFace and is accessible via ComfyUI and LTX API for Managed Generation services. The model can run on a wide range of hardware, from cloud data-centre GPUs to Mac computers. The model is intended for application including media and entertainment, real-time generation, robotics and other physical AI workloads.
