Guide
Run an open video model in ComfyUI
ComfyUI is the free app most people use to run open image and video models on their own computer. This guide takes you from install to your first clip.
What ComfyUI is
ComfyUI is a free, open-source app where you build image and video generation as a flow of connected blocks. You do not need to code: you load a ready-made flow (a "workflow") and press run.
New to open source or to running things yourself? Read Start here first.
What you need
From the ComfyUI docs
- Operating system: Windows, Linux, or a Mac with an Apple Silicon chip (M-series).
- Graphics card: NVIDIA, AMD, Intel Arc or Apple Silicon are supported. A dedicated GPU is recommended. Without one it runs on the CPU, which is slower.
- Browser: a recent Google Chrome (the docs recommend version 143 or later).
- Disk space: at least 4.85 GB for the app itself. Model files are extra, and they are large.
Be realistic
- Video is demanding. The smaller Wan 2.2 model (5B) is said to fit on a card with 8 GB of video memory. The larger 14B models need more, and the docs do not give a minimum.
- Expect to wait. A short clip can take minutes, not seconds.
- No GPU? You can still try, but it will be slow. Comfy also offers Comfy Cloud, which needs no installation. Check its pricing first.
Step 1: install ComfyUI
There are three official ways. For most people, the first one is right.
Comfy Desktop
A normal installer for Windows, macOS (Apple Silicon) and Linux. Easiest, and it follows the stable release.
Recommended
ComfyUI Portable
Windows only. A folder you unzip and run, with its own Python included. It follows the latest changes.
Easy
Manual install
For all systems and GPU types. You install PyTorch and the requirements yourself.
Technical
Download from the official site (comfy.org), run the installer and follow the prompts. On first launch, the Welcome screen creates your first installation.
Step 2: your first clip
- Open a template. In ComfyUI, open the Template Library and search for "Wan 2.2". For a first test, pick the small one: Wan 2.2 5B Video Generation.
- Let it find the model files. If a model is missing, ComfyUI shows a warning with the file name and a download link. Click it and download. Downloads usually come from Hugging Face.
- Put each file in the right folder. Models live in
ComfyUI/models/, in subfolders by type. In Comfy Desktop, open the capsule menu at the top and choose the Storage tab to find the folders. - Refresh. Press R so ComfyUI sees the new files. If it does not, restart the app.
- Write a prompt and run. Describe the clip in the text block, then press Run or Ctrl + Enter.
- Save the result. The clip shows up in the save block at the end. Right-click it to save.
If the model list shows nothing or "null", check that the file is in the right folder, then refresh or restart.
Wan 2.2: which files go where
As listed in the ComfyUI docs on 12 October 2026. File names change between releases, so trust the missing-model warning in the template if it differs.
Folder in models/ | Wan 2.2 5B (text or image to video) | Wan 2.2 14B text to video |
|---|---|---|
diffusion_models | wan2.2_ti2v_5B_fp16.safetensors | the high-noise and low-noise 14B T2V files (fp8 scaled) |
text_encoders | umt5_xxl_fp8_e4m3fn_scaled.safetensors | the same file |
vae | wan2.2_vae.safetensors | wan_2.1_vae.safetensors |
The 14B image-to-video and first-last-frame workflows use their own pair of high-noise and low-noise files. The template names them for you.
A note on the licence: the code of Wan 2.2 is Apache-2.0, but check the model page for the licence of the model files themselves before you use clips commercially. Wan 2.2 on this site.
Other models
LTX-Video
The ComfyUI docs list "LTXV Text to Video" and "LTXV Image to Video" templates. Their example uses a checkpoint in models/checkpoints/ and a text encoder in models/text_encoders/. It works best with long, detailed prompts. LTX-Video on this site.
HunyuanVideo
Its license says it does not apply in the European Union, the United Kingdom and South Korea. If you live there, do not use it. Read the details.
Mochi, and voice models
We could not confirm a ComfyUI guide for Mochi, and voice models such as Chatterbox or Kokoro are normally run with their own tools. Follow each project's own install page rather than guessing.
Good to know
Custom nodes are other people's code. ComfyUI-Manager can install them for you. Only add ones from authors you trust, and look at how many people use them.
Model files are large. Some are many gigabytes each. Check your free disk space before you start downloading.
Licences differ. The code and the model files can have different rules. Read the model page before you sell or publish what you make.
Uninstalling Comfy Desktop does not delete your installations, models or outputs. They stay on your disk until you delete them yourself.