{"id":5774,"date":"2026-03-24T19:46:27","date_gmt":"2026-03-24T11:46:27","guid":{"rendered":"https:\/\/crepal.ai\/blog\/?p=5774"},"modified":"2026-03-24T19:46:29","modified_gmt":"2026-03-24T11:46:29","slug":"how-to-install-ltx-2-3-comfyui","status":"publish","type":"post","link":"https:\/\/crepal.ai\/blog\/aivideo\/how-to-install-ltx-2-3-comfyui\/","title":{"rendered":"How to Install LTX 2.3 in ComfyUI: Step-by-Step Guide"},"content":{"rendered":"\n

Hello fellows! This is Dora \u2014 and recently I finally cracked the clean install sequence for LTX 2.3. I’d spent two hours on a broken node graph. Right package version, wrong folder. Right folder, missing VAE. Right VAE, wrong ComfyUI build. When it finally rendered a clean 9-frame clip, I immediately wrote down every exact step so I’d never have to do that again.<\/p>\n\n\n\n

The blog below explains no vague tips but every command, every folder path, every error I hit and how I fixed it. Let’s get into it.<\/p>\n\n\n\n

Prerequisites (ComfyUI version, GPU VRAM, disk space)<\/h2>\n\n\n\n

Before downloading anything, map your system against this table. I tested across three machines; this reflects what actually works.<\/p>\n\n\n\n

Component<\/td>Minimum<\/td>Recommended<\/td>Notes<\/td><\/tr>
ComfyUI build<\/td>Nightly 20260101+<\/td>Nightly 20260310+<\/td>Older builds missing LTXVModelLoader node<\/td><\/tr>
GPU VRAM<\/td>12 GB<\/td>24 GB<\/td>12 GB requires distilled + fp8 offload<\/td><\/tr>
Disk space<\/td>16 GB (distilled only)<\/td>35 GB (both models)<\/td>Include VAE + T5-XXL<\/td><\/tr>
Python<\/td>3.1<\/td>3.11<\/td>3.12 untested, likely broken<\/td><\/tr>
CUDA<\/td>11.8<\/td>12.3<\/td>ROCm: experimental, Linux only<\/td><\/tr>
OS<\/td>Windows 10, Ubuntu 22.04<\/td>Same<\/td>macOS: no CUDA, not viable<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

Check your ComfyUI build version<\/strong> before proceeding. Open ComfyUI’s terminal and run:<\/p>\n\n\n\n

cd ComfyUI\ngit log --oneline -1<\/code><\/pre>\n\n\n\n
If the commit date is before January 2026, update first:<\/p>\n\n\n\n
git pull origin master\npip install -r requirements.txt<\/code><\/pre>\n\n\n\n
Step 1 \u2014 InstallRequired Python Packages<\/a><\/strong><\/p>\n\n\n\n
This must happen before you download any models. The packages tell ComfyUI how to interpret LTX 2.3’s architecture.<\/p>\n\n\n\n
Identify which Python your ComfyUI uses.<\/strong> This trips people up constantly because they install packages into the wrong environment.<\/p>\n\n\n\n
\n
\"\"<\/figure>\n<\/figure>\n\n\n\n
If you use the ComfyUI portable package (Windows):<\/em><\/p>\n\n\n\n
# Navigate to your ComfyUI portable folder first\ncd C:\\ComfyUI_windows_portable\npython_embeded\\python.exe -m pip install ltx-core ltx-pipelines<\/code><\/pre>\n\n\n\n
If you use a venv or <\/em>conda<\/em> environment:<\/em><\/p>\n\n\n\n
# Activate your environment first\nconda activate comfyui   # or: source venv\/bin\/activate\npip install ltx-core ltx-pipelines<\/code><\/pre>\n\n\n\n
Verify the <\/strong>install<\/strong> worked:<\/strong><\/p>\n\n\n\n
python -c \"import ltx_core; print(ltx_core.__version__)\"\n# Should print: 0.9.2 or higher\npython -c \"import ltx_pipelines; print('OK')\"<\/code><\/pre>\n\n\n\n
If you get ModuleNotFoundError<\/code> on the verification step, you installed the wrong Python. Re-read the step above and confirm which python<\/code> executable is actually running ComfyUI.<\/p>\n\n\n\n
Step 2 \u2014 Download LTX 2.3 Weights<\/strong><\/p>\n\n\n\n
Dev (bf16) vs Distilled model \u2014 which to get<\/h3>\n\n\n\n
I’ve run both through hundreds of generations. Here’s the real tradeoff:<\/p>\n\n\n\n
<\/td>Dev (bf16)<\/td>Distilled<\/td><\/tr>
File size<\/td>~27 GB<\/td>~14 GB<\/td><\/tr>
Inference steps<\/td>40\u201350<\/td>6\u20138<\/td><\/tr>
Generation time (4090)<\/td>~4 min\/clip<\/td>~45 sec\/clip<\/td><\/tr>
Motion quality<\/td>High<\/td>Moderate<\/td><\/tr>
Complex scenes<\/td>Excellent<\/td>Struggles<\/td><\/tr>
Prompt iteration<\/td>Slow<\/td>Fast<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
My actual <\/strong>workflow<\/strong>: I use distilled for prompt exploration (testing 10\u201315 prompt variations quickly), then switch to dev for the final output I’ll publish. If you have a 12 GB GPU, start with distilled only.<\/p>\n\n\n\n
Where to download (HuggingFace, GitHub monorepo)<\/h3>\n\n\n\n
All official weights are hosted on HuggingFace<\/strong> by Lightricks:<\/p>\n\n\n\n
\n
\"\"<\/figure>\n<\/figure>\n\n\n\n