# ComfyUI-LTXVideo

> Lightricks' custom ComfyUI nodes and workflows for the LTX-2 video model

- Category: Multimodal Generation / Video Generation
- Language: Python
- Repository: https://github.com/Lightricks/ComfyUI-LTXVideo
- Canonical page: https://ai-tldr.dev/tools/comfyui-ltxvideo/

## In short

ComfyUI-LTXVideo is Lightricks' collection of custom nodes that extends ComfyUI for the LTX-2 video generation model. It sits in the Video Generation category of the Multimodal Generation tools and is open-source, written in Python. To get started: check the prerequisites.

## Overview

ComfyUI-LTXVideo is Lightricks' collection of custom nodes that extends ComfyUI for the LTX-2 video generation model. LTX-2 itself is already built into ComfyUI core, so every ComfyUI user can reach the basic pipeline without this repository; what the node pack adds is the advanced surface — the control adapters, the staged workflows and the example graphs that get more out of the model than the built-in nodes expose.

The headline additions are IC-LoRA control workflows: depth, human pose and edge conditioning, motion tracking on image-to-video, HDR, a Dub-It pass and pixel spatial upscaling, plus a text-to-audio distilled workflow. A Union IC-LoRA combines depth and edge (canny) conditioning into a single adapter, and it operates on a downsampled latent, which cuts memory use and speeds up inference while holding quality. Single-stage and two-stage (upsampling) variants of the text/image-to-video graphs are both shipped.

Everything arrives as example workflow JSON files installed alongside the nodes, so the practical entry point is loading a graph rather than wiring one. The pack is installable from ComfyUI Manager and required model files are fetched on first use. It is hardware-hungry: the project lists a CUDA-compatible GPU with 32GB or more of VRAM and 100GB or more of free disk space for models and cache as prerequisites.

## What it does

- IC-LoRA control workflows for depth, human pose and edge conditioning on LTX-2 generations
- A Union IC-LoRA that merges depth and canny control into one adapter and runs on a downsampled latent for lower memory use and faster inference
- Motion-tracking image-to-video, HDR, Dub-It and pixel spatial upscaling workflows
- Single-stage and two-stage (with upsampling) text/image-to-video graphs for both the full and distilled models
- A text-to-audio distilled workflow alongside the video graphs
- Example workflow JSON installed with the nodes, and models downloaded automatically on first use

## Getting started

Install from ComfyUI Manager rather than cloning by hand, then load one of the bundled example workflows. Check the hardware requirements first — this is a large-model pipeline, not a laptop node pack.

1. Check the prerequisites

   You need ComfyUI installed, a CUDA-compatible GPU with 32GB+ VRAM, and 100GB+ of free disk space for models and cache.

2. Install the nodes from ComfyUI Manager

   Open ComfyUI, press Ctrl+M for the Manager, choose Install Custom Nodes, search for “LTXVideo”, install, and restart ComfyUI. The nodes then appear in the node menu under the LTXVideo category.

```text
Manager (Ctrl+M) → Install Custom Nodes → search "LTXVideo" → Install → restart
```

3. Load an example workflow

   The installation ships several ready-made graphs; the recommended starting point is one of the LTX-2 workflows available through Comfy Manager. Required models download on first use.

```text
ComfyUI/custom_nodes/ComfyUI-LTXVideo/example_workflows/
```

4. Move up to the control workflows

   Once text-to-video works, switch to the IC-LoRA graphs for depth, pose and edge control, motion tracking, HDR or spatial upscaling — and use the two-stage distilled workflow when you want the upsampling pass.

## When to use it

- Drive an LTX-2 generation with depth, pose or edge control instead of prompt text alone
- Add an upsampling second stage to a distilled LTX-2 run when the single-stage output is not sharp enough
- Use one Union IC-LoRA for depth and canny control to keep VRAM use and inference time down
- Start from a maintained example graph rather than reverse-engineering the model's node wiring yourself