LTX-2.5 Fast

LTX-2.5 Fast is a speed-optimized, open-weights video generation model developed by Lightricks. Built on a 22-billion parameter Diffusion Transformer architecture, it is designed to generate high-fidelity video and synchronized audio in a single pass. The model is part of a broader family of "world models" that prioritize temporal consistency and physical simulation, making it suitable for professional filmmaking, rapid prototyping, and emerging applications in robotics.

Technical Architecture

The model utilizes a custom Gemma 4 12B text encoder, which enhances its ability to follow complex, multi-subject instructions and specific camera directions. A defining feature of the LTX-2.5 architecture is Diffusion Fidelity Rendering, a system that dynamically allocates compute resources based on scene complexity to ensure fine detail in demanding sequences. Furthermore, a dedicated Diffusion Video Decoder replaces traditional VAE reconstruction, resulting in sharper facial features, legible on-screen text, and reduced motion artifacts.

Key Capabilities

LTX-2.5 Fast supports native multi-shot generation, enabling the creation of connected scenes that maintain character identity, lighting, and environmental consistency across cuts within a single generation. It is capable of producing videos up to 20 seconds in length at resolutions ranging from 720p to 4K. The model's unified pipeline handles text-to-video, image-to-video, and frame-to-frame interpolation tasks while simultaneously generating a matching audio track.

Prompting Guide

Official documentation recommends a prose-based prompting approach rather than the use of weighted tags or keywords. Effective prompts are typically structured as a single paragraph covering six core elements: the shot type, scene environment, specific actions, character descriptions, camera movement, and audio details. Lightricks suggests writing in the present tense and using quotation marks for any intended spoken dialogue. For maximum control, users can provide both a starting and an ending frame to guide the model's interpolation process.