Lightricks/LTX-2

What it solves

LTX-2 is a foundation model designed to unify the core capabilities of modern video generation into a single system. It addresses the challenge of creating high-fidelity, production-ready videos that feature synchronized audio and video, while providing flexible performance modes for different speed and quality needs.

How it works

Built on a Diffusion Transformer (DiT) architecture, LTX-2 uses a multi-stage pipeline approach to generate content. It supports various generation modes including text-to-video, image-to-video, and audio-to-video. The system utilizes a text encoder (Gemma 3) and offers multiple specialized pipelines—such as a distilled pipeline for fast inference and a two-stage pipeline for higher resolution via spatial upscaling. It also supports LoRAs for fine-grained control over camera movement, pose, and detail, as well as IC-LoRAs for tasks like lip dubbing and HDR output.

Who it’s for

This project is aimed at creators and developers who need production-quality video generation, ranging from those requiring rapid prototyping (via single-stage or distilled pipelines) to those needing high-fidelity outputs with precise cinematic control over camera angles and character movements.

Highlights

• Unified Capabilities: Combines synchronized audio and video generation in one model.
• Diverse Pipelines: Offers specialized workflows for text/image-to-video, audio-to-video, keyframe interpolation, and video-to-video transformations.
• Advanced Control: Includes LoRAs for specific camera movements (dolly, jib) and IC-LoRAs for lip dubbing and HDR.
• Optimization Options: Supports FP8 quantization and FlashAttention 3/4 to reduce memory footprint and increase speed on compatible GPUs.
• Training Tools: Includes a dedicated trainer package for LoRA and full fine-tuning.