LTX Open Source Documentation
LTX-2.5 is an open world model with open weights, built for local execution and fine-tuning. It is a single DiT-based model that runs on consumer GPUs and generates synchronized video and audio from text, image, and video inputs. It also supports LoRA and IC-LoRA customization for fine-grained control over style, motion, and structure.
New in 2.5
LTX-2.5 brings updates to nearly every stage of the pipeline, on top of a larger, more filtered training dataset and reinforcement-learning post-training.
In this version:
-
Native multishot — a single generation can produce multiple connected shots while holding character, scene, lighting, visual style, and voice consistent across cuts. The Prompting Guide gives updated guidance.
-
Improved distilled model — the distilled model keeps more of the full model’s visual quality, prompt adherence, and motion performance, while requiring substantially less compute.
-
Diffusion fidelity rendering — instead of locking every scene to one compression rate, the model allocates compute by scene complexity and budget, dynamically allocating more compute to visually demanding moments and less where it is not needed.
-
New diffusion video decoder — a timestep-conditioned decode that reconstructs detail more deliberately from the latent representation, improving fine texture, facial detail, typography, product elements, and stability during fast motion. Video and audio are decoded by separate VAEs.
-
Stronger prompt understanding — a custom Gemma 4 12B text encoder converts your prompt into the conditioning the model generates from, and follows complex prompts more faithfully than before. A separate, optional prompt enhancer can first expand a short prompt with visual, cinematic, temporal, and contextual detail before it’s encoded. Use it for terse prompts, or turn it off to keep your exact wording.
-
Auto duration — the model interprets the event or motion described in the prompt and predicts the appropriate clip duration before diffusion starts. Duration is connected to what needs to happen in the scene rather than a fixed parameter.
-
Native EXR support — LTX-2.5 includes native support for EXR input and output. It can apply generative changes and return the result as EXR or color-matched video without forcing footage through a lossy 8-bit round-trip.
-
More permissive licensing — the model licensing now has fewer restrictive third-party dependencies and a clearer path to fine-tune, deploy, and commercialize.
To learn more about LTX’s architecture and the research behind it, read our research paper.