Learn ComfyUI Foundations: From Zero to First Image Organizing Your ComfyUI Setup

Organizing Your ComfyUI Setup

Beginner 🕐 10 min Lesson 11 of 13
What you'll learn
  • Navigate the ComfyUI folder structure and identify which subfolder each model type belongs in
  • Set up a symlink so ComfyUI can read model files stored in a different location on disk
  • Describe a practical backup strategy for ComfyUI custom nodes and workflows

Organization Pays Off Quickly

A new ComfyUI installation is clean and simple. After a few weeks of downloading models, installing custom nodes, and building workflows, it can become a mess: gigabytes of model files scattered across your drive, duplicate downloads you forgot about, custom nodes you installed once and never used. Cleaning this up later is tedious. Setting up a good structure from the start takes 15 minutes and saves hours.

This lesson covers the folder structure ComfyUI expects, how to handle the fact that model files are huge and disk space is limited, and what to back up to avoid losing your work.

The ComfyUI Folder Structure

Inside your ComfyUI installation directory, the layout looks like this:

  • ComfyUI/models/ — all model files, organized by type in subfolders
  • ComfyUI/custom_nodes/ — installed custom node packages (each is a subfolder)
  • ComfyUI/output/ — generated images and videos (organized by date automatically)
  • ComfyUI/input/ — images you place here can be loaded by the Load Image node without specifying a path
  • ComfyUI/temp/ — temporary files created during generation (safe to clear)
  • ComfyUI/user/ — user settings, saved workflows, and preferences

The models/ subfolder is the critical one. Each model type has its own subfolder, and ComfyUI only looks in the correct subfolder for each node type. Placing a LoRA file in models/checkpoints/ will not cause an error, but it also will not appear in the LoRA loader dropdown — it will only appear in the checkpoint dropdown, where it does not belong.

Model Subfolder Reference

The standard model subfolders and what goes in each:

  • models/checkpoints/ — full checkpoint models (.safetensors, .ckpt). This is where your main generation models go.
  • models/vae/ — standalone VAE files.
  • models/loras/ — LoRA files. Can create sub-folders inside here for organization (e.g., loras/characters/, loras/styles/).
  • models/controlnet/ — ControlNet model files.
  • models/clip/ — standalone CLIP encoder models. Used in FLUX.1 workflows with the DualCLIPLoader node.
  • models/diffusion_models/ (also sometimes models/unet/) — standalone UNet files. Used in some FLUX.1 and other newer model setups.
  • models/upscale_models/ — upscaler files like 4x-UltraSharp or RealESRGAN.
  • models/embeddings/ — textual inversion (.pt, .bin) files.
  • models/ipadapter/ — IPAdapter model files.
  • models/animatediff_models/ — AnimateDiff motion modules.

Sub-folders within a model type folder are fine — ComfyUI searches recursively. Naming your LoRAs with descriptive prefixes (e.g., style-watercolor-v2.safetensors, char-product-photo-xl.safetensors) makes the dropdown lists much easier to navigate.

Symlinks: Solving the Disk Space Problem

Model files are large. A single FLUX.1 checkpoint is around 24 GB. SDXL checkpoints are 6–7 GB each. If you have multiple ComfyUI installations (e.g., one for stable work and one for testing) or if you also use A1111 and share models between them, downloading the same model twice wastes significant disk space.

The solution is symbolic links (symlinks). A symlink makes a folder in one location appear as if it exists in another location, without duplicating the files.

On Mac or Linux, to point ComfyUI's checkpoints folder at a shared model directory:

  • Remove (or rename) the existing empty folder: mv ComfyUI/models/checkpoints ComfyUI/models/checkpoints_backup
  • Create a symlink: ln -s /path/to/shared/models/checkpoints ComfyUI/models/checkpoints

On Windows, use mklink /J in an Administrator command prompt: mklink /J C:ComfyUImodelscheckpoints C:SharedModelscheckpoints

After creating the symlink, ComfyUI sees the models in the shared location as if they were in its own folder, with no duplication.

What to Back Up

Model files are large and re-downloadable — do not back them up unless you have unlimited backup storage and cannot re-download easily. Instead, back up the three things that represent your actual work:

  • ComfyUI/user/: Contains your saved workflows and settings. The most important thing to back up.
  • ComfyUI/custom_nodes/: Contains your installed custom node packages. Backing this up means you can restore your node setup without reinstalling everything through Manager.
  • A list of your model files: Not the files themselves — just a text file listing the names. If you ever need to rebuild from scratch, this list tells you exactly what to re-download.

A simple approach: keep your user/ folder synced to a cloud storage folder (Dropbox, Google Drive, etc.). Your workflows are then automatically backed up on every save. For custom nodes, running a git pull in each subfolder of custom_nodes/ is usually faster than restoring from backup — Manager can reinstall everything in 10 minutes.

Key takeaways
  • ComfyUI searches each model subfolder by type — checkpoints go in models/checkpoints/, LoRAs in models/loras/, ControlNet models in models/controlnet/ — placing a file in the wrong folder means it appears in the wrong dropdown.
  • Symlinks let multiple ComfyUI installations share the same physical model files on disk — no duplication needed; on Mac/Linux use ln -s, on Windows use mklink /J in an Administrator prompt.
  • Sub-folders within model type folders are supported and searched recursively — creating loras/characters/ and loras/styles/ sub-folders keeps the dropdown lists manageable as collections grow.
  • Back up ComfyUI/user/ (saved workflows and settings) and keep a list of your model filenames — the model files themselves are large and re-downloadable, so they are low priority for backups.
  • The output/ folder fills up quickly with generated images — it is organized by date automatically, but periodically moving finished work elsewhere and clearing old output keeps disk usage predictable.