LTX-2.5 – Lightricks' Open-Source AI Video Generation Base Model

1. What is LTX-2.5

LTX-2.5 is a professional-level AI video generation base model open-sourced by the Israeli company Lightricks. With 22B parameters, the model weights are made available immediately upon release. It supports native multi-shot generation, 4K HDR, and RAW/EXR lossless professional workflows, can generate audio synchronously, and possesses precise video editing capabilities. In the current landscape where closed-source models are increasingly expensive, LTX-2.5 offers a professional-grade open-source alternative for the AI video domain through its open-ecosystem strategy, covering a wide range of application scenarios from post-production in film and television to physical AI.

Image source: Official article
Image source: official article

Technical Positioning and Domain: LTX-2.5 is a video generation base model based on the diffusion model architecture, designed for professional film content creation and industrial-level video production. Unlike most previous open-source models that only supported short clips and low-resolution outputs, LTX-2.5 is directly aimed at 4K HDR and RAW/EXR lossless formats, and natively integrates multi-shot narrative capabilities, filling the gap in professional video workflows within open-source models.

Development Background: Developed by the Lightricks team, the company has accumulated years of expertise in the field of image and video editing, having launched well-known applications such as Facetune and Videoleap. LTX-2.5 continues the company's technical approach in visual generation, transferring the concept of fine-grained control from image editing to video generation, and choosing full open-sourcing to promote the development of the community ecosystem.

Core Value: It addresses three core pain points in AI video generation: shot continuity (maintaining character and scene consistency in multi-shot generation), professional output formats (direct integration into color grading and special effects pipelines), and controllability (precise video editing and IC-LoRA control). It provides creators with a deterministic production path, moving from "blind card-drawing" to "real-shoot skeleton + AI stylization overlay."

Technical Features: Based on a diffusion Transformer architecture with 22B parameters, it employs a two-stage rendering strategy (structure first, then details) known as Diffusion Fidelity Rendering. It includes an internal Gemma 4 12B text encoder and a Prompt Enhancer auxiliary model, and supports adaptive computational resource allocation. The output resolution can reach 4K (3840×2160), with a maximum duration of 20 seconds, and supports frame rates of 24/25/48/50fps.

2. Key Features

• Native Multi-Camera Generation: Generate multiple continuous shots with different camera angles and compositions in a single pass, maintaining high consistency in character appearance, environmental atmosphere, lighting conditions, and sound between transitions. This capability directly addresses the core challenge of shot continuity in AI-driven storytelling, eliminating the need for post-production stitching or manual adjustments.

• Diffusion Fidelity Rendering: Utilizes a two-phase strategy of "first structure, then details." In the first phase, motion, composition, and shot framework are constructed in an 8× time-compressed latent space, with adaptive generation of high-fidelity keyframes. In the second phase, the final video is rendered through diffusion from both the structure and keyframes, parsing textures, materials, and facial details with pixel-level precision. This mechanism dynamically allocates computational resources based on scene complexity, allocating more resources to complex dynamic shots and automatically saving resources for static scenes.

• Precise Video Editing: Perform high-accuracy style rewriting, scene repainting, and character replacement while preserving the original motion trajectory and spatial structure of the live-action footage. This enables "live-action skeleton + AI stylized overlay" controlled creation, allowing creators to first use live-action footage to define actions and composition, then replace the visual style using AI.

• Automatic Duration Prediction (Duration Head): Features a built-in Duration Head module deployed prior to the diffusion process. This module automatically infers and outputs the most appropriate video duration by understanding the action semantics in the prompt, giving the model an initial sense of narrative pacing without requiring creators to repeatedly test different lengths.

• Professional-Grade Output Formats: Supports native 4K HDR, RAW, and EXR lossless output formats. These formats can be directly integrated into professional film color grading, visual effects compositing, and post-production editing pipelines, without the need for compression or transcoding, meeting the strict requirements for lossless material in theatrical and streaming contexts.

• Synchronized Audio Generation: Through an independent Audio VAE module, video content is mapped to corresponding audio representations, enabling synchronized generation of visuals and sound. The audio track remains consistent with the video during multi-shot transitions, further enhancing narrative coherence.

• IC-LoRA Precise Control: Offers three reference video control modes: Canny edges, Depth, and Pose. After loading the corresponding IC-LoRA, the generated results can strictly follow the composition, spatial structure, or character movements of the reference material, achieving precise control from composition to motion.

3. How to Use

1. Environment Requirements and Path Selection: There are two primary methods to use LTX-2.5: API cloud service or local ComfyUI deployment. The API method does not require local hardware but is limited by network and call quotas; local deployment requires a high-performance GPU (recommended NVIDIA A100 or H100, with VRAM no less than 48GB), and the operating system is recommended to be Linux or Windows, with the latest version of ComfyUI installed.

2. API Integration: Apply for an API Key on the LTX official website (ltx.io), and select either the Fast or Pro tier to call the API. The Fast tier is suitable for quick iteration previews, while the Pro tier offers higher resolution and longer duration. The API supports text-to-video, image-to-video, and first-last frame modes.

3. Local Deployment Preparation: Update ComfyUI to the latest version, and search for and install the "ComfyUI-LTXVideo" node package via the ComfyUI Manager. This node package encapsulates model loading, parameter configuration, and the inference process.

4. Download Models and Workflows: Search for "LTX-2.5" in the ComfyUI template browser, select the corresponding workflow template, and download all model files with one click (including the base model, text encoder, Audio VAE, etc.). The model weights are approximately 45GB, so ensure sufficient disk space is available.

5. Parameter Configuration: Set the width and height to multiples of 32 (e.g., 1024×576), and the number of frames to a multiple of 1+8 (e.g., 25 frames). Choose a frame rate of 24/25/48/50fps. Set the base resolution to half of the target output resolution to enable 2× upscaling (e.g., set base to 1920×1080 if the target is 4K).

6. Writing Prompts: Use a coherent paragraph of prose to describe shot composition, scene lighting, character actions, camera movement, and audio requirements. Avoid stacking tags or using weight syntax. The Prompt Enhancer will automatically expand short prompts into detailed cinematic-level instructions.

7. Execution of Generation: Click "Run." Text-to-video will generate the video directly; image-to-video requires uploading the first frame image; first-last frame mode requires uploading both the first and last frame images, and the model will automatically fill in the intermediate frames.

8. Precise Control (Advanced): Load the IC-LoRA in Canny, Depth, or Pose mode, and connect a reference video to lock composition or actions. This mode is suitable for scenarios requiring strict alignment with reference materials.

9. Model Selection: Use the distilled model for rapid iteration (faster inference speed), and the full dev model for the final output (more refined details). Both models can be downloaded from HuggingFace.

4. Pros and Cons Analysis

5. Comparative Analysis with Similar Tools

Selection Recommendations: For teams seeking professional cinematic quality and full control, LTX-2.5 is currently the best open-source option, with its multi-camera support, lossless output, and precise editing capabilities also leading among closed-source models. If the project is sensitive to hardware budget and only requires quick generation of short video previews, Minimax H3's API service can be selected (note the regional restrictions). For closed-source solutions requiring longer durations or more mature ecosystems, Runway Gen-3 Alpha still holds advantages in API stability and community resources, though it lacks multi-camera and audio synchronization capabilities. Overall, LTX-2.5 demonstrates clear differentiated advantages in openness, output quality, and controllability, making it suitable as the core engine for professional workflows.

6. Editor's Summary

The release of LTX-2.5 marks the entry of open-source AI video generation into a professional-level stage. From a technological innovation perspective, its 22B parameter scale ranks among the top in open-source video models, but what holds more value is the systematic design at the architectural level: the two-stage strategy of Diffusion Fidelity Rendering effectively balances efficiency and visual quality, the Duration Head introduces narrative pacing perception into the model, and the Audio VAE achieves sound synchronization during the generation phase — these are not merely improvements in individual metrics, but a redefinition of the "completeness" of video generation. In terms of practical value, native 4K HDR and RAW/EXR outputs directly integrate with post-production pipelines in the film and video industry, while precise video editing and IC-LoRA control enable creators to shift from "random selection" to "controlled production," significantly lowering the barriers and uncertainties of professional video creation. The target audience is clearly professional filmmakers, AI short film teams, advertising creative agencies, and researchers in physical AI, rather than ordinary short video users. In terms of future development potential, the open-source ecosystem is likely to foster a large number of community-finetuned models and application plugins, further expanding its capability boundaries. At the same time, continued investment from Lightricks and academic attention will also drive the model's evolution in duration and multi-modal integration. LTX-2.5 is not just a tool, but may become the cornerstone of the open-source ecosystem in the AI video domain.

7. Application Scenarios

• Professional Film Post-Production: End-to-end 4K HDR, RAW, and EXR lossless outputs can be directly integrated into color grading and visual effects compositing software such as DaVinci Resolve and Nuke, eliminating the need for compression and transcoding. Ideal for visual previsualization, scene replacement, and stylization in cinema and streaming series, meeting the highest standards for media quality.

• AI Short Films and Comics: Multishot generates multiple shots in one go, outputting coherent narrative storyboards. Precise video editing supports "real-action motion skeleton + AI character replacement and stylized overlay," fundamentally addressing the two major challenges of shot continuity and motion controllability. Creators can first shoot actor performances and then use AI to replace characters and scenes, significantly reducing production costs.

• Advertising and Brand Content: Utilize the Prompt Enhancer to quickly generate high-quality commercial videos. The Distilled model enables rapid iteration of creative directions, while the full dev model produces the final output. IC-LoRA control ensures consistency in product composition and brand elements, making it suitable for e-commerce ads and brand promotional videos.

• Physical AI and Robotics: The Physical AI Checkpoint provides a foundational world model for embodied intelligence teams, helping robotic systems perceive the physical world and motion dynamics. Videos generated by LTX-2.5 can serve as training data in simulation environments for tasks such as motion planning and object interaction, extending beyond pure content creation.

8. FAQ

Q: What hardware configuration is required to run LTX-2.5 locally?
A: The minimum requirement is a GPU with at least 48GB of VRAM (e.g., NVIDIA A100 40GB/80GB, H100). It is recommended to use dual GPUs or higher configurations to accelerate inference. Consumer-grade GPUs like the RTX 4090 (24GB) cannot directly load the full model, but you can attempt to use quantization or model parallelism. The official version does not yet provide a quantized model. The recommended operating system is Linux or Windows 10/11, and you need to install the latest version of ComfyUI.

Q: Does it support Chinese prompts? How effective are they?
A: The text encoder is based on Gemma 4 12B, which is primarily optimized for English. Chinese prompts can be input, but complex long prompts involving multiple subject relationships and action details may not be fully preserved. It is recommended to write prompts in English, or first use the Prompt Enhancer to translate Chinese prompts into English before generating. A Chinese-compatible version is currently under development by the community.

Q: How can I fine-tune LTX-2.5?
A: The weights are fully open-source, and common fine-tuning methods such as LoRA and DreamBooth are supported. You need to prepare a video dataset (recommended to have at least 100 clips) and use the training scripts provided in the model card from HuggingFace. The official documentation currently lacks detailed instructions on fine-tuning, so it is recommended to refer to community tutorials or wait for official updates.

Q: What is the maximum duration and resolution for output videos?
A: The maximum single-generation duration is 20 seconds, and the highest resolution supported is 4K (3840×2160). Frame rates supported include 24/25/48/50fps. Output formats include standard MP4 (H.264/H.265), PNG sequences, RAW (.dng), or EXR (.exr). RAW and EXR are lossless formats and result in large file sizes, so be mindful of storage space.

Q: What are the advantages and disadvantages of LTX-2.5 compared to Sora?
A: The main advantages of LTX-2.5 include full open-source availability, professional output formats, multi-shot and audio synchronization, and the ability to deploy and fine-tune locally. Sora may excel in physical world simulation, ultra-long video generation (up to 60 seconds), and scene consistency, but it is closed-source and does not provide an API, limiting creators' autonomy. LTX-2.5 is more suitable for workflows requiring precise control and professional output, while Sora is better suited for quickly generating high-quality concept videos.

9. Project Links

1. Product Official Website: https://ltx.io/model/ltx-2-5
2. HuggingFace Model Repository: https://huggingface.co/Lightricks/LTX-2.5