How to Install LTX 2.5 in ComfyUI: Step-by-Step Guide

The easiest way to install LTX 2.5 is ComfyUI's ready-made templates. When you open a template and click "Download all," it downloads the model files it needs by itself. If you want to install manually, work offline, or write code in Python, there are two more routes. This post is the second part of the series; I explained what the model is in What Is LTX 2.5, and for hardware details see requirements and speed.

This post is based on the official documentation; I have not tried the installation on my own machine. Commands and file names come from LTX's ComfyUI page and the Hugging Face model card. Sources are at the end.

Before you start

The prerequisites on LTX's ComfyUI page:

• ComfyUI installed
• A CUDA-compatible graphics card with 32 GB or more of VRAM (for lower-end cards, see the requirements post)
• More than 100 GB of free disk space for models and cache
• Python 3.12 or later

The weights on Hugging Face are also gated: you need to accept the license and sign in. For the license's revenue cap ($10 million per year), see the what-is post.

Route 1: With ComfyUI templates (recommended)

1. Open ComfyUI and click the Templates button.
2. Search for "LTX-2.5." To start, pick the Text to Video or Image to Video template.
3. Click Download all to download the required models.
4. When the download finishes, restart ComfyUI.

The nodes appear in the menu under the LTXVideo category. ComfyUI provides three ready-made LTX-2.5 templates:

Template	What it does
Text-to-Video (video_ltx2_5_t2v)	Two-stage, video from text
Image-to-Video (video_ltx2_5_i2v)	Two-stage, video from a starting image
First-Frame / Last-Frame (video_ltx2_5_flf2v)	Single-stage, generates the motion between the first and last frames

Advanced workflows (IC-LoRA control, inpainting and outpainting) are in the example_workflows/2.5/ folder.

Route 2: Installing the nodes manually

If the templates don't work or you don't use ComfyUI Manager:

cd ComfyUI/custom_nodes
git clone https://github.com/Lightricks/ComfyUI-LTXVideo.git
cd ComfyUI-LTXVideo
pip install -r requirements.txt

If you use portable ComfyUI:

.\python_embeded\python.exe -m pip install -r .\ComfyUI\custom_nodes\ComfyUI-LTXVideo\requirements.txt

Then fully close and reopen ComfyUI. To verify the installation, right-click the canvas and follow Add Node → LTXVideo; you should see categories such as loaders, samplers, conditioning, and utils.

Model files and folders

The templates' "Download all" step does all of this for you. If you are downloading manually, take these files from the LTX-2.5 repository on Hugging Face and put them in the matching folders:

File	Purpose	Folder
ltx-2.5-22b-distilled-transformer-bf16.safetensors	Distilled model	ComfyUI/models/diffusion_models/
gemma4-12b-with-proj-ltx-2.5-bf16.safetensors	Gemma 4 text encoder	ComfyUI/models/text_encoders/
gemma4_e2b_it_bf16.safetensors	Prompt enhancer	ComfyUI/models/text_encoders/
ltx-2.5-video-vae-bf16.safetensors	Video VAE	ComfyUI/models/vae/
ltx-2.5-audio-vae-bf16.safetensors	Audio VAE	ComfyUI/models/vae/
ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors	Spatial upscaler (two-stage templates only)	ComfyUI/models/latent_upscale_models/

Two details to watch:

• The prompt enhancer is not in the LTX-2.5 repository. The gemma4_e2b_it_bf16.safetensors file lives in the Comfy-Org/gemma-4 repository. The templates download it automatically; with a manual install, you have to download it yourself.
• The upscaler is only needed for the two-stage templates. The single-stage First/Last Frame template doesn't use it.

Files for low VRAM

If you have a card with less memory, use these three files together, in place of the BF16 transformer, Gemma encoder, and video VAE above:

• ltx-2.5-22b-distilled-transformer-comfy-int8-convrot.safetensors → diffusion_models/
• gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot.safetensors → text_encoders/
• ltx-2.5-video-vae-conv-bf16.safetensors → vae/ (a faster, lighter Conv VAE)

For cards with FP4 support, such as NVIDIA Blackwell, there is an even smaller NVFP4 distilled transformer. The *-comfy-int8-convrot files are for ComfyUI only and do not work on the Python command-line route.

Route 3: Python command line (ltx-pipelines)

For when you want to write code or automate. Versions recommended on the model card: Python 3.12 or later, CUDA 12.7 or later, PyTorch around 2.7. (LTX's low-VRAM post says CUDA 13.2 or later; the two sources don't agree, so check the README in the repository.)

git clone https://github.com/Lightricks/LTX-2.git
cd LTX-2
uv sync
source .venv/bin/activate

Download the weights (about 66 GiB in total according to the README):

hf auth login
hf download Lightricks/LTX-2.5 \
  diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors \
  text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors \
  vae/ltx-2.5-video-vae-bf16.safetensors \
  vae/ltx-2.5-audio-vae-bf16.safetensors \
  model_patches/ltx-2.5-duration-head-bf16.safetensors \
  latent_upscale_models/ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors \
  --local-dir models/ltx-2.5

Run the distilled pipeline:

uv run python -m ltx_pipelines.distilled \
  --transformer-path models/ltx-2.5/diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors \
  --text-encoder-path models/ltx-2.5/text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors \
  --video-vae-path models/ltx-2.5/vae/ltx-2.5-video-vae-bf16.safetensors \
  --audio-vae-path models/ltx-2.5/vae/ltx-2.5-audio-vae-bf16.safetensors \
  --duration-head-path models/ltx-2.5/model_patches/ltx-2.5-duration-head-bf16.safetensors \
  --spatial-upsampler-path models/ltx-2.5/latent_upscale_models/ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors \
  --prompt "A golden retriever running through a sunny meadow, cinematic lighting" \
  --seed 42 \
  --output-path output_distilled.mp4

If you don't pass --num-frames, the duration predictor picks the length from the prompt. If you set it manually, the frame count must leave a remainder of 1 when divided by 8 (frames % 8 == 1, for example 121), and the width and height must be divisible by 32.

If you run into memory trouble, you can add --quantization fp8-cast --offload cpu.

If you want to use Diffusers, there is a separate package (Lightricks/LTX-2.5-Diffusers), but according to the model card, released versions of Diffusers don't support LTX-2.5 yet; you need to install from the main branch with pip install git+https://github.com/huggingface/diffusers.

Common errors

Problem	Fix
Nodes are missing from the menu	Verify the custom_nodes/ComfyUI-LTXVideo folder, repeat pip install -r requirements.txt, fully close and reopen ComfyUI, check the Python version
Red nodes in the workflow	Click "Install Missing Custom Nodes," update ComfyUI, update the nodes with git pull
Model won't load	Check the file sizes, verify the folders are correct, download again
Commonly missing dependencies	ComfyUI-VideoHelperSuite, ComfyUI-Manager, ComfyUI-Impact-Pack

To update the nodes:

cd ComfyUI/custom_nodes/ComfyUI-LTXVideo
git pull
pip install -r requirements.txt --upgrade

Once the installation is done, move on to the last post in the series to generate your first video and write prompts.

Frequently Asked Questions

What is the easiest way to install LTX 2.5?

Open the LTX-2.5 template from the Templates menu in ComfyUI and click Download all. The nodes and models come automatically.

Which files do I need to download?

The distilled transformer, the Gemma 4 text encoder, the video and audio VAEs, and, for the two-stage templates, the spatial upscaler. The prompt enhancer comes from a separate repository.

Does it work on Windows?

The ComfyUI route is documented for Windows with a portable install. The main supported environment for the Python command-line pipelines is Linux.

Why does Hugging Face ask me to sign in?

The weights are in a gated repository that requires license acceptance. You need to sign in and accept the license.

How much space does the installation take?

The README gives about 66 GiB of downloads in total, and the ComfyUI page recommends more than 100 GB of free disk space for models and cache.

Sources

• LTX, Using ComfyUI with LTX: prerequisites, installation steps, file table, templates, error fixes.
• Lightricks, LTX-2.5 model card (Hugging Face): Python command-line setup, file list, restrictions.
• Lightricks, LTX-2 repository (GitHub): pipelines, download command and size.
• Lightricks, ComfyUI-LTXVideo (GitHub): ComfyUI nodes and example workflows.
• LTX, How to Run a Video Generation Model Locally: CUDA and Python version notes.