LTX-2 is truly open-source, providing not just model weights but also full training code, benchmarks, and LoRAs, making it adaptable for developers and studios

The distilled model variant enables practical local video generation on consumer hardware, with significantly faster render times (53 seconds vs 2 minutes 27 seconds) while maintaining good quality

LTX-2 uses a two-stage rendering process: generating a lower-resolution base video first, then upscaling it to final resolution, which requires applying LoRAs to both stages

The model supports multiple workflows including text-to-video, image-to-video, video-to-video, and audio-conditioned generation, with resolutions up to 4K

Camera control LoRAs enable precise control over camera movements (dolly, pan, etc.) and require specific prompting techniques describing the destination of movement and hidden scene elements

LTX-2 UNLEASHED AI VIDEO

Name: LTX-2 UNLEASHED AI VIDEO
Uploaded: 2026-02-03T19:57:11.215123
Duration: 21 min 1 s
Description: This video introduces LTX-2, Lightricks' fully open-source AI video generation model that represents a major breakthrough in local video generation. Unlike typical "open" models that only provide weights, LTX-2 includes full training code, model weights, and optimization for consumer hardware. The h

By Wes Roth

Watch on YouTube (21:01)

Overview

This video introduces LTX-2, Lightricks' fully open-source AI video generation model that represents a major breakthrough in local video generation. Unlike typical "open" models that only provide weights, LTX-2 includes full training code, model weights, and optimization for consumer hardware. The host demonstrates how to install and use the model locally on ComfyUI, comparing full vs distilled versions, and showcasing various features including text-to-video, image-to-video, and camera control LoRAs.

Key Takeaways

LTX-2 is truly open-source, providing not just model weights but also full training code, benchmarks, and LoRAs, making it adaptable for developers and studios
The distilled model variant enables practical local video generation on consumer hardware, with significantly faster render times (53 seconds vs 2 minutes 27 seconds) while maintaining good quality
LTX-2 uses a two-stage rendering process: generating a lower-resolution base video first, then upscaling it to final resolution, which requires applying LoRAs to both stages
The model supports multiple workflows including text-to-video, image-to-video, video-to-video, and audio-conditioned generation, with resolutions up to 4K
Camera control LoRAs enable precise control over camera movements (dolly, pan, etc.) and require specific prompting techniques describing the destination of movement and hidden scene elements