Lightricks has released LTX-2, an open-source toolkit for generating synchronized audio and video, with its LTX-2.5 model now available. The project positions LTX-2.5 as an open-weights video, audio, and world-simulation model, with code on GitHub and weights on Hugging Face as of August 2026.
Unlike simple clip generators, LTX-2 is a comprehensive framework designed for developers. It provides a modular, high-performance base for building real-time graphics and visual effects applications. The toolkit is split into packages for core model functions, generation pipelines, and model training.
How is the LTX-2 architecture structured?
LTX-2 features a modular architecture based on a Diffusion Transformer (DiT). Developers can select specific components like transformers, VAEs, and text encoders to fit their project's needs. This allows for customized pipelines that balance performance and quality without being locked into a monolithic system.The latest version, LTX-2.5, breaks the model down into individual `safetensors` files. This granular approach means a developer only needs to download the specific weights required for a given task, such as the 66 GiB needed for the quick start guide.
Component Type | Description | Example File (LTX-2.5) |
|---|---|---|
Transformer | Core diffusion model. A full 'dev' version and a faster 'distilled' version are available. | ltx-2.5-22b-distilled-transformer-bf16.safetensors |
Text Encoder | A fine-tuned Gemma 4 12B model required by all pipelines to interpret prompts. | gemma4-12b-with-proj-ltx-2.5-bf16.safetensors |
Video VAE | Decodes latent representations into video frames. Options include a high-quality diffusion VAE or a lighter convolutional VAE. | ltx-2.5-video-vae-bf16.safetensors |
Audio VAE | Handles audio generation and decoding, enabling synchronized sound. | ltx-2.5-audio-vae-bf16.safetensors |
What generation pipelines are available?
The LTX-2 toolkit provides multiple, distinct pipelines tailored for different use cases, from rapid prototyping to production-quality rendering. These pre-built scripts allow developers to quickly implement complex generation tasks like text-to-video, audio-to-video, and video-to-video transformations.DistilledPipeline: The fastest option for text-to-video, designed as a quick starting point.
DFRPipeline: The 'Diffusion Fidelity Rendering' path for production-quality output. It uses more VRAM and time to add spatial detailing and extra keyframes.
ICLoraPipeline: Enables video-to-video and image-to-video editing and transformation.
A2VidPipelineTwoStage: Generates video conditioned on an input audio file, creating visuals that match a soundtrack.
DubItPipeline: Rephrases dialogue in a video while matching the original speaker's identity and lip movements.
How does it handle LoRA training?
LTX-2 simplifies fine-tuning through its integrated LoRA trainer, which treats LoRA weights as pipeline-agnostic. This means the same checkpoint can be used for both text-to-video (T2V) and image-to-video (I2V) inference. This flexibility streamlines the training process for custom styles or characters.The system uses different configurations depending on the intended use. For mixed-use cases, the `i2v_lora` config learns both I2V and T2V generation in a single run. It achieves this by applying first-frame conditioning with a probability of 0.5 during training. For text-only generation, the `t2v_lora` configuration is used instead.
