LTX-2.5 Operational

LTX-2.5 is LTX's open-source foundational AI video generation model with 22B parameters and open weights available from launch. The model supports native multishot generation, 4K HDR and RAW/EXR professional workflows, synchronized audio generation, and precise video editing capabilities. As closed-source models become increasingly expensive, LTX-2.5 provides a professional-grade open-source alternative for the AI video field through its open ecosystem strategy.

Native multishot generation: A single generation can output multiple continuous shots with different camera positions and shot sizes, while maintaining a high degree of consistency in character appearance, environmental atmosphere, lighting conditions, and sound across shot transitions, solving the core challenge of shot continuity in AI storytelling.
Diffusion Fidelity Rendering: Rendering resources are allocated adaptively according to scene complexity. Complex dynamic shots receive more computing resources, while static or simple scenes automatically save resources, improving overall efficiency while maintaining image quality.
Precise video editing: While preserving the motion trajectories and spatial structure of original live-action footage, the model enables highly precise style rewriting, scene redrawing, and character replacement, realizing controllable creation through a “live-action skeleton + AI stylized overlay.”
Automatic duration prediction: The built-in Duration Head module reads action descriptions in prompts and automatically determines the most suitable clip length, eliminating the need for creators to repeatedly experiment.
Professional-grade output formats: Supports native 4K HDR, RAW, and EXR lossless output, allowing direct integration with professional film and television color grading, visual effects compositing, and post-production editing pipelines.
Synchronized audio generation: An independent Audio VAE module generates a matching audio track synchronized with the video.
Precise IC-LoRA control: Provides three reference-video control modes—Canny edge, Depth, and Pose—so generated results strictly follow the composition, spatial structure, or character movements of the reference material.

Follow on WeChat and reply “开源” to join the AI open-source project discussion group
Gemma 4 12B text encoder: A large language model encoder customized by Lightricks for LTX. It fully preserves multi-subject relationships, action details, lighting descriptions, and camera directions in complex long prompts, without losing information from subordinate clauses as the prompt becomes longer.
Prompt Enhancer: A lightweight auxiliary model that automatically expands short user inputs into detailed, cinematic instructions, lowering the barrier to prompt engineering with extremely low computational overhead.
Diffusion Video Decoder: A new architecture that replaces traditional VAE decoders and is specifically optimized for fast-motion scenes. It significantly reduces image smearing and artifacts, making faces clearer and text more readable.
Audio VAE: An independent audio latent-space encoder responsible for mapping video content to matching audio representations, enabling synchronized generation of visuals and sound.
Diffusion Fidelity Rendering: Uses a two-stage rendering strategy of “structure first, details second.” In the first stage, motion, composition, and shot structure are built in a latent space with 8× temporal compression, while high-fidelity keyframes are generated adaptively. In the second stage, the final video is rendered through joint diffusion from the structure and keyframes, analyzing textures, materials, and facial details with pixel-level precision.
Duration Head: A prediction module deployed before the diffusion process. By understanding the action semantics in the prompt, it automatically infers and outputs the most suitable video duration, giving the model an initial awareness of narrative pacing.

Choose a path: Decide whether to use the API cloud service or deploy locally with ComfyUI.
API integration: Apply for an API Key on the LTX official website and select the Fast or Pro tier to call the interface.
Install nodes: For local deployment, update ComfyUI to the latest version and install the ComfyUI-LTXVideo node package through Manager.
Download the model: Search for “LTX-2.5″ in the ComfyUI template browser, select the corresponding workflow, and download all model files with one click.
Configure parameters: Set the width and height to multiples of 32, the frame count to 1+ multiples of 8, and the frame rate to 24/25/48/50fps. Set the base resolution to half the target resolution to enable 2× upscaling.
Write prompts: Use a coherent paragraph to describe the shot size, scene lighting, character actions, camera movement, and audio. Avoid stacking tags or using weight syntax.
Run generation: Click run. Text-to-video generates the video directly; image-to-video requires uploading the first frame; first-and-last-frame mode requires uploading both the first and last frames.
Precise control: Advanced users can load IC-LoRA in Canny, Depth, or Pose modes and connect reference videos to lock the composition or movements.
Model selection: Use the distilled model for rapid iteration and the full dev model for final output to obtain finer details.

Narrative coherence: Native Multishot generation can maintain a high degree of consistency in character appearance, environmental atmosphere, lighting conditions, and sound in a single output, fundamentally solving the biggest challenge of shot continuity in AI storytelling.
Major leap in controllability: Precise video editing preserves the motion trajectories and spatial structure of original live-action footage. Combined with three IC-LoRA control modes—Canny/Depth/Pose—it enables a shift from “blindly rolling the dice” to controllable production based on a “live-action skeleton + AI stylized overlay.”
Industrial-grade workflow: Native RAW/HDR/EXR output can be directly integrated into professional color grading, visual effects compositing, and post-production editing pipelines without compression or transcoding, meeting the film and television industry's strict need for lossless materials.
Adaptive efficiency: Diffusion Fidelity Rendering dynamically allocates rendering resources according to scene complexity. Complex dynamic shots are rendered in detail, while static simple scenes automatically save resources, improving overall generation efficiency while maintaining pixel-level image quality.

Official website: https://ltx.io/model/ltx-2-5
Hugging Face model repository: https://huggingface.co/Lightricks/LTX-2.5

Professional film and television post-production: Native 4K HDR, RAW, and EXR lossless output can be directly integrated into color grading, visual effects compositing, and editing pipelines, meeting the post-production needs of theatrical and streaming content without the color-grading nightmare of compressed MP4.
AI short dramas and animated dramas: Multishot generation outputs coherent narrative storyboards in a single pass, while precise video editing supports “live-action movement skeleton + AI character replacement and stylized overlay,” fundamentally solving the two major pain points of shot continuity and action controllability.
Advertising and branded content: Rapidly iterate on high-quality commercial videos. Prompt Enhancer lowers the barrier to prompt engineering, the Distilled model supports fast previews, and the full model outputs final productions, balancing efficiency and quality.
Physical AI and robotics: Physical AI Checkpoint provides an underlying world model for embodied intelligence teams, helping robotic systems perceive the physical world and laws of motion, extending beyond the scope of pure content creation.