Lightricks/ComfyUI-LTXVideo

What it is

ComfyUI‑LTXVideo is a set of custom nodes for the visual‑programming UI ComfyUI that let you run Lightricks’ LTX‑2 video‑generation model inside ComfyUI. The repository bundles extra nodes, example workflows and helper scripts so users can do text‑to‑video, image‑to‑video, video‑to‑video, audio‑only generation, HDR conversion, dubbing, up‑scaling and many LoRA‑based control tricks.

Who it’s for

Artists, creators, and developers who already use ComfyUI and want to experiment with state‑of‑the‑art video generation.
Users with a high‑end GPU (32 GB VRAM) and enough disk space (≈100 GB) who can host the large LTX‑2 checkpoints.
Anyone interested in applying LoRA‑based controls (depth, pose, HDR, dubbing, etc.) to generated video.

How to get started

Install ComfyUI (download from the official site).
Open ComfyUI → press Ctrl + M → Manager → search for “LTXVideo” and install the custom node package.
Restart ComfyUI; a new LTXVideo category appears in the node menu.
On first use the required LTX‑2 model files are downloaded automatically (or you can pre‑download them from the HuggingFace links listed in the README).
Open any of the example workflow JSON files under custom_nodes/ComfyUI‑LTXVideo/example_workflows/ to see a ready‑made pipeline, then edit prompts or parameters as desired.

What you can do with it

Feature	What it does	Typical use
Text‑to‑Video / Image‑to‑Video	Generates full video clips from a text prompt or a single image.	Quick storyboards, visual effects mock‑ups.
IC‑LoRA controls	LoRA adapters that add depth, pose, edge, HDR, dubbing, water‑simulation, etc., to the base model.	Fine‑grained artistic control, style transfer, visual effects.
HDR conversion	Turns SDR video into ACEScct HDR frames; can export 16/32‑bit EXR sequences.	High‑dynamic‑range grading pipelines.
Dub‑It	Re‑generates lip‑sync and audio to match a new dialogue (multilingual or same‑language).	Localisation, voice‑over creation.
Pixel Spatial Upscaler	Generative up‑scaling (2× or 4×) that adds new detail rather than simple interpolation.	Turning low‑res drafts into high‑resolution final renders.
Audio‑only generation (T2A)	Uses the joint audio/video transformer in audio‑only mode, outputting a wav/FLAC file.	Sound design, music generation.
Tiled Fusion Sampler	Samples large canvases (FullHD‑8K) by tiling while keeping a single latent canvas, enabling high‑resolution video‑to‑video work.	Upscaling or detail‑refining 4K/8K video.
Low‑VRAM loaders	Helper nodes that off‑load parts of the model to fit within 32 GB VRAM.	Users with slightly less memory.

How it works (high‑level)

LTX‑2 is a large transformer that jointly predicts video frames and audio. The core model lives inside the official ComfyUI repository, so this package only adds wrapper nodes that expose extra functionality (LoRA injection, control‑signal handling, up‑scaling, HDR post‑processing, etc.).
Workflows are built by connecting nodes: a model loader node brings the checkpoint into memory, optional LoRA nodes attach control adapters, sampler nodes run the diffusion process, and decode nodes turn latent tensors into image frames or audio waveforms.
For high‑resolution work the Tiled Fusion Sampler splits the canvas into overlapping tiles, runs diffusion on each tile while sharing a single noise field, then stitches the results back together.

When it might not be right for you

If you lack a GPU with ≥32 GB VRAM, the base model (22 B parameters) will not fit; you would need to rely on the low‑VRAM loaders and accept slower generation.
The repository is tightly coupled to ComfyUI; it cannot be used as a stand‑alone script or with other UI frameworks.
It requires manual download of many large checkpoint files (model, up‑scalers, LoRAs), so setup can be storage‑heavy.

All information above is taken directly from the repository’s README; no external assumptions have been added.