LTX-2.5: Multi-shot video generation that maintains character consistency across cuts

LTX-2.5 is an open-world model that accepts text, image, and video inputs to generate synchronized video and audio. It runs directly on local infrastructure, offering complete control and customization without per-generation billing or API dependencies. Companies with annual revenue under $10 million can use it for commercial and production purposes for free under the Community License.

Unlike previous versions, it supports native multi-shot generation that connects multiple shots in a single pass. It maintains character identity, environment, lighting, voice, and visual style across cuts. It also applies diffusion fidelity rendering, which dynamically adjusts compute allocation based on scene complexity, applying detailed processing to important areas and efficient processing to the rest.

The new diffusion video decoder replaces the VAE reconstruction step, enhancing the clarity of faces, textures, and on-screen text while reducing motion artifacts. It includes a custom Gemma 4 12B text encoder designed to handle complex prompts, and a prompt enhancer that expands short prompts into rich cinematic instructions. An optional duration predictor estimates clip length based on the prompt and automatically sets the frame count.

The model is distributed as a split pack consisting of component-specific .safetensors files rather than a single file. Each component, including the distilled DiT, full DiT, text encoder, video/audio VAE, LoRA, and upscaler, can be loaded individually. It is available in three ways: Python-based ltx-pipelines, ComfyUI, and Diffusers. The distilled model offers faster inference speeds with smaller checkpoints.