A successful ComfyUI installation ends with one saved image from a known workflow. Seeing the interface is only the first checkpoint. You also need a compatible runtime, the right model files and a graph that connects them correctly.
Desktop versus Portable
The current ComfyUI getting-started documentation lists Desktop for Windows, Apple Silicon macOS and Linux, and Portable for Windows. Desktop follows stable releases by default. Portable bundles its own Python environment. Check the current operating-system and GPU guidance before downloading because packaging support changes.
| Option | Useful when | What you maintain |
|---|---|---|
| Desktop | You want a managed application setup | Selected instance, storage and compatible extensions |
| Windows Portable | You want a self-contained folder and explicit launch scripts | That folder’s Python, models and update procedure |
| Manual installation | You need direct control over the environment | Python environment, dependencies, runtime and update choices |
For Windows, use the official Desktop instructions or Portable instructions. Keep the Portable launch terminal open while working; closing the server process also stops generation. On another operating system, follow its matching installation page rather than adapting a Windows command blindly.
Install one model set, not a collection
Start with the model linked by an official example. Record the exact filename and where you put it. If a download contains a base model, encoder and VAE, all three may be required. A smaller file is not necessarily a drop-in replacement for a larger one.
File placement follows the component’s role. A checkpoint belongs in the checkpoint location, a LoRA in the LoRA location, and an upscaler in the upscaler location. The official model guide explains loader selection, model search paths and refreshing the available files. In particular, .safetensors is a storage format, not a model category.
How to add a safetensors file correctly
- Read the model card. Identify its architecture, intended loader and companion files.
- Find the active models directory. Desktop storage can differ from the folder used by a separate Portable installation.
- Place it in the documented subfolder. Do not move every file into
checkpoints. - Refresh or restart as required. Select the exact model in the matching loader.
- Run the matching example. Only then alter the creative brief.
If a model still does not appear, inspect the complete filename. A failed download, an extra extension or a file placed in a different installation can look like a loader problem. Confirm the downloaded file against the provider’s listed size before downloading another copy.
Your first-generation acceptance test
Use an ordinary scene: a ceramic mug on a wooden desk with soft window light. Keep the example’s image size and batch count. Check that the job finishes, the file is saved, and reopening the workflow preserves the selected model.
Then change the mug’s color while leaving the environment fixed. This is not a quality benchmark; it checks that you know which input controls the result. Save that baseline before adding a custom node, new model family or higher-resolution stage.
Keep the setup recoverable
Write down your installation method and location, the GPU and available memory, the ComfyUI version, and the date of the successful test. Keep your working graph and model list outside a folder you might replace during an update. If you use multiple installations, give them recognizable names so you know which one a launcher starts.
Next, learn how to save and import workflows or how to handle missing custom nodes. If your goal is an AI creator brand and you prefer a hosted production environment, start with a persona in Clout and create your first content batch there.
