LTX-Video is the official video generation model repository released and maintained by Lightricks on GitHub. According to its official description, it is the first DiT-based video generation model, concentrating the core capabilities of modern video generation into a single model: synchronized audio-video, high fidelity, multiple performance modes, production-ready output, API access, and open access. It can produce videos at native 4K resolution, up to 50 FPS, and with synchronized audio in a single generation, and is trained on a large-scale dataset of diverse videos, capable of generating high-resolution videos with realistic and varied content.
In terms of main capabilities, the model supports image-to-video, multi-keyframe conditioning, keyframe-based animation, video extension (forward and backward), video-to-video, and any combination of these features. The repository provides examples of image-to-video and controlled video. Regarding parameters, the model is suitable for inputs with resolutions divisible by 32 and frame counts that are a multiple of 8 plus 1 (e.g., 257 frames); if they do not conform, the input is padded with -1 and then cropped to the target resolution and frame count. The official recommendation is resolutions below 720×1280 and frame counts below 257. The recommended Guidance Scale is 3–3.5; more steps (40+) benefit quality, while fewer (20–30) benefit speed; the seed can be saved to reproduce style or composition.
In terms of version progress, v0.9.0 was released on November 21, 2024 as the initial version, supporting text-to-video and image-to-video; v0.9.1 was released on December 20, 2024, improving quality and supporting STG / PAG, loading checkpoints in Diffusers format, offloading unused parts to CPU, and a new timestep-conditioned VAE decoder; the research paper was released on December 31, 2024; on February 20, 2025, more inference options were added, spatiotemporal guidance was improved, macOS MPS (PyTorch 2.3.0) was supported, along with the 8-bit model LTX-VideoQ8, TeaCache, and ComfyUI-LTXTricks and Diffusion-Pipe were added; v0.9.5 was released on March 5, 2025, adopting the OpenRail-M commercial license, supporting keyframes and video extension, higher resolution, improved prompt understanding, and launching an online web app and automatic prompt enhancement in LTX-Studio; on April 15, 2025, v0.9.6 checkpoints and distilled models were released, with distilled model inference speedIt is 15 times faster than non-distilled models, with a default resolution of 1216×704 at 30 FPS, and supports random inference; on May 5, 2025, the 13B v0.9.7 model, FP8 quantized model, and temporal/spatial upscalers were released; on May 14, 2025, the 13B v0.9.7 distilled model was released, capable of generating HD video within 10 seconds and low-resolution previews within 3 seconds (H100), without requiring classifier-free guidance and spatiotemporal guidance, supporting sampling with 8 diffusion steps or fewer, and providing a LoRA version (requiring only 1GB of VRAM) and an FP8 quantized distilled model; on July 8, 2025, three control models for depth, pose, and Canny were released; on July 16, 2025, the v0.9.8 distilled model was released, supporting videos up to 60 seconds long, including 13B and 2B distilled models, both distilled from the same base model, which can be used together in the same multi-scale pipeline, and the LTX-Video-ICLoRA-detailer-13B-0.9.8 detail model was released; on October 23, 2025, LTX-2 was announced as a DiT-based audio-video foundation model, featuring synchronized audio-video generation, native 4K and up to 50 fps, continuous clips up to 10 seconds long, up to 50% lower computational cost compared with competitors, multi-keyframe conditioning, 3D camera logic, and LoRA fine-tuning, with its model weights, code, and benchmarks to be released later in 2025.
In terms of usage, the models can be used online through LTX-Studio (13B-mix, 13B distilled image-to-video), Fal.ai (13B full, 13B distilled image-to-video), and Replicate. The codebase has been tested under Python 3.10.5, CUDA 12.2, and PyTorch >= 2.1.2; macOS MPS has been tested under PyTorch 2.3.0 and should support PyTorch == 2.3 or >= 2.6. To run locally, follow the inference code in inference.py, and it can be used as a library. Input video clips must contain a number of frames that is a multiple of 8 plus 1 (such as 9, 17, 25), and the target frame count should be a multiple of 8. Paths to images and/or short video clips and their target frame numbers can be provided for conditioned generation, and the conditioning strength for each item can be specified (default 1.0). For ComfyUI integration, see ComfyUI-LTXVideo; for Diffusers integration, see the official documentation, and Diffusers also supports the 8-bit version. When writing prompts, it is recommended to describe actions and scenes in detail in chronological order, including specific motion, appearance, camera angles, and environmental details, using a single flowing paragraph of text, starting directly with the action, with literal and precise descriptions, like a photographer writing a shot list, kept within 200 words, and a recommended structure is also given. When using LTXVideoPipeline directly, prompt enhancement can be enabled via enhance_prompt=True. For advanced parameters, see python inference.py --help.
In terms of community contributions, ComfyUI-LTXTricks provides nodes for RF-Inversion, RF-Edit, FlowEdit, I+V2V, enhancement, and interpolation/frame settings; LTX-VideoQ8 is an 8-bit optimized version, delivering up to 3x speedup on NVIDIA Ada GPUs with no loss of precision, and can generate a 720x480x121 video in under a minute on an RTX 4060 (8GB VRAM), and can fine-tune a 2B transformer model using precomputed latents; TeaCache for LTX-Video is a training-free caching method that uses timestep differences between model outputs to accelerate inference by up to 2x with no noticeable degradation in visual quality.
In terms of limitations, official materials state that input resolution and frame count must satisfy divisibility conditions, otherwise they will be padded and then cropped; the model performs best below 720×1280 resolution and below 257 frames. FP8 kernels provide performance improvements only on supported GPUs (Ada architecture and later). The specific model weights, code, and benchmarks for LTX-2 are not yet available and are planned for release later in 2025. Apart from the above, facts regarding pricing, fees, and specific usage steps are not yet provided.
