In January 2026, Lightricks officially open‑sourced the LTX‑2 model, a diffusion‑transformer (DiT) based audio‑video foundation model capable of generating synchronized video and audio within a single model. As the first fully open‑sourced multimodal video generation tool, LTX‑2 supports high‑fidelity outputs, including native 4K resolution, 50 frames per second (FPS), and videos up to 20 seconds. It integrates core components of modern video generation — spatial and temporal upsamplers, LoRA adapters, and control modules — and supports tasks such as text‑to‑video, image‑to‑video, and video‑to‑video. This open‑source move marks a shift in AI video generation from closed ecosystems to community‑driven development; Lightricks CEO Zeev Farbman said in a Reddit AMA that the goal is to let developers "actually use" the model and drive evolution from simple prompt generation toward professional rendering engines. This article, based on authoritative sources including Hugging Face, GitHub, and arXiv, details LTX‑2's technical features, installation and usage steps, and provides three demonstration YouTube links to help readers get started.
Architecture and Capabilities of LTX‑2
At its core, LTX‑2 uses an asymmetric two‑stream transformer architecture: the video stream has 14B parameters and the audio stream 5B parameters, coupled by bidirectional audio‑video cross‑attention layers and supporting shared temporal conditioning. This design allows the model to generate audio synchronously while producing video, avoiding the inefficiencies of separately processed approaches. The model family includes five checkpoints: a base model for general generation, trainable and customizable; a distilled model that accelerates inference (only 8 steps); and various LoRAs for controls such as Canny edges, depth maps, and pose. LTX‑2 is optimized for efficiency, capable of running locally on consumer GPUs and supporting FP8 quantization to reduce memory usage.
In terms of performance, LTX‑2 supports multiple modes: fast prototyping (single‑stage pipeline), production quality (two‑stage upsampling), and keyframe interpolation. It can handle complex prompts, such as detailed action descriptions, environmental details, and camera motion, achieving cinematic effects. Compared to closed models like Sora and Veo, LTX‑2's open‑source nature makes it easier to customize and integrate, for example with ComfyUI for graphical workflows. Lightricks emphasizes the model's "production‑grade" orientation and has applied it in company products, demonstrating its reliability for long‑form creation.
Installation and Usage Steps
LTX‑2's installation is based on a Python environment; the official repository provides full code and training packages, supporting local inference and LoRA fine‑tuning. System requirements include Python 3.12+, sufficient VRAM (Hopper GPUs recommended to enable Flash Attention 3) and the uv package manager. Below are detailed steps:
- Clone the repository and set up the environment：
- Install uv (if not installed): pip install uv.
- Clone the repo: git clone https://github.com/Lightricks/LTX‑2.git.
- Enter the directory: cd LTX‑2.
- Sync dependencies: uv sync ‑‑frozen (this creates a virtual environment and installs core packages such as ltx‑core, ltx‑pipelines and ltx‑trainer).
- Activate the environment: source .venv/bin/activate.
- Download model weights：
- Download core checkpoints from Hugging Face, such as ltx‑2‑19b‑dev‑fp8.safetensors (FP8 version recommended to save memory).
- Download upsamplers: spatial ltx‑2‑spatial‑upscaler‑x2‑1.0.safetensors and temporal ltx‑2‑temporal‑upscaler‑x2‑1.0.safetensors.
- Download the Gemma text encoder assets from google/gemma‑3‑12b‑it‑qat‑q4_0‑unquantized.
- Optional LoRAs: e.g. ltx‑2‑19b‑distilled‑lora‑384.safetensors for the distilled pipeline. Place them in a local directory or use the Hugging Face cache.
- Optimization setup：
- Install attention optimizations: uv sync ‑‑extra xformers (or manually install Flash Attention 3).
- For high‑VRAM devices, disable memory cleanup to speed up runtime.
- Usage examples：
- Text‑to‑video：Use TI2VidTwoStagesPipeline (recommended for production). Example command：
from ltx_pipelines import TI2VidTwoStagesPipeline pipeline = TI2VidTwoStagesPipeline.from_pretrained("Lightricks/LTX-2", fp8transformer=True) video = pipeline("A cinematic scene of a mountain landscape at sunset", num_inference_steps=40, guidance_scale=7.5) video.save("output.mp4")- Parameters: num_inference_steps controls the number of steps (20–40 recommended), guidance_scale adjusts prompt adherence.
- Image‑to‑video: Similar, using an image path as input.
- Distilled mode: Use DistilledPipeline to accelerate (only 8 steps): pipeline = DistilledPipeline.from_pretrained(...).
- Prompt tips: Prompts should be detailed and describe actions and environment in temporal order, keep length under ~200 words, and use enhance_prompt=True to auto‑optimize.
- ComfyUI integration: Clone https://github.com/Lightricks/ComfyUI‑LTXVideo/ and install the nodes to generate via the graphical interface.
- Training and fine‑tuning：
- Use the ltx‑trainer package: after preparing a dataset, run ltx‑trainer lora ‑‑config config.yaml.
- Supports full fine‑tuning and IC‑LoRA; see the repository README for details.
Representative YouTube Demonstrations
LTX‑2's practical results have been demonstrated in multiple YouTube videos; below are three representative links covering installation, generation, and comparisons:
- How To Use LTX‑2 in ComfyUI | FREE AI Videos With Synced Audio (https://www.youtube.com/watch?v=IVs2uBNyHX0)：This video provides a detailed walkthrough of installing and using LTX‑2 in ComfyUI, demonstrating text‑to‑video and image‑to‑video examples and highlighting the advantage of free local execution.
- LTX‑2 Released opensource! (https://www.youtube.com/watch?v=wXv4qhS4Oc0)：Focuses on the open‑source release, showcases 4K 50FPS video generation examples, and compares capabilities with models like Sora, suitable for newcomers to understand overall performance.
- NEW Open‑Source Leap to Create ALL Camera Movements in AI Filmmaking (30 Cinematic Moves) (https://www.youtube.com/watch?v=OlsyJx5YRBs)：A tutorial‑style video demonstrating 30 cinematic camera moves such as the Dolly Zoom, showing how precise prompts can achieve professional effects.
Industry Impact and Outlook
LTX‑2's open‑source release fills a gap in multimodal models and accelerates AI applications in film, gaming, and VFX. Community feedback has been positive, noting ease of local use but also the need for high‑end GPUs. Looking ahead, Lightricks plans to expand the roadmap to include longer videos and advanced controls. Overall, LTX‑2 represents a leap for open‑source AI video generation, and developers can contribute via GitHub to help shape the multimodal future.
▶️ Enjoy this article? Creating content is hard!
Please follow our YouTube channel:
AI Rosetta
Your likes, subscriptions, and shares are our greatest support! 🙏
More AI news at: https://www.airosetta.com