LTX-2.5 is here: open weights, native multishot, and quality that holds up on screen

LTX-2.5 improves multishot continuity, video rendering, and workflow control, with open weights and local execution.

LTX-2.5 advances our video model with a stronger focus on continuity across multiple shots. The system can generate a sequence made up of several shots while preserving characters, environments, lighting, voices, and visual direction from one cut to the next, whereas previous versions were more focused on a single continuous shot.

The main technical change concerns the final rendering stage. With Diffusion Fidelity Rendering, the model first builds the motion and structure of the sequence in a highly compressed space while generating more detailed reference frames alongside it. A second diffusion process then reconstructs the video, allocating more compute to visually complex sections. The new decoder is designed to better preserve faces, textures, on-screen text, and fine motion.

Prompt understanding now relies on a custom 12-billion-parameter Gemma 4 encoder adapted for LTX. It is designed to better retain information from complex prompts involving multiple characters, camera movements, actions, and lighting conditions. An optional system can also automatically determine the duration of a sequence based on the prompt.

For production workflows, LTX-2.5 supports HDR input and output within an ACES pipeline, as well as 16-bit linear data through EXR. Teams can also adapt the model through fine-tunes, LoRA, and IC-LoRA, run it on their own infrastructure, or integrate it with ComfyUI, Diffusers, and LTX’s Python pipelines.

The weights are available on Hugging Face in several variants, including a fully trainable version, a distilled version, and multiple quantized configurations. LTX.io also provides API access. The license allows free commercial use for organizations generating less than $10 million in annual revenue, while a commercial agreement is required above that threshold.