Skip to content

Models & Files

You’ve downloaded a bunch of .safetensors files and the last thing you want is to put them in the wrong place — ComfyUI won’t recognize files that aren’t in the right directory. Below is a quick reference by file type → default directory, which ties into Introduction and Workflow.

  • Checkpoint (base model): The full set of model weights bundled together, usually including CLIP + VAE (though some have a mediocre built-in VAE that people swap out).
    Directory: ComfyUI/models/checkpoints

  • VAE: The small network that “translates” latent space into the RGB image you can see; sometimes downloaded separately, sometimes embedded in a checkpoint.
    Directory: ComfyUI/models/vae (standalone files go here; use Load VAE or the node to select it in the UI)

  • LoRA / LyCORIS etc.: Small fine-tuning patches that attach to a base model; relatively small file size.
    Directory: ComfyUI/models/loras

  • ControlNet: A conditioning model used to “lock down” structural constraints.
    Directory: ComfyUI/models/controlnet

  • CLIP / text encoders etc.: A small number of workflows reference these separately; common bundles also place them under a subfolder in models. Don’t worry about memorizing this yet — deal with it when you encounter it.

Exact subpaths depend on your ComfyUI installation — portable builds, bundles, and custom installs may have different root names — but the subfolder names under models are widely agreed upon in the community.

  1. Check someone else’s workflow: The author usually lists required filenames in a Note node or README — download whatever is missing.
  2. Read the model page: The author will state compatibility with SD1.5 / SDXL or other architectures — base model architecture, LoRA, and ControlNet must match; a mismatch at best does nothing, at worst throws an error.
  3. Don’t judge by file size: Bigger isn’t always better for your use case — get it running first, then optimize.
  • CheckpointLoaderSimple → reads from checkpoints
  • Load LoRA → reads from loras
  • Load ControlNet → reads from controlnet
  • Standalone VAE → reads from vae (when the workflow is set up that way)

Once you understand the pipeline, these are simply: each node pulls files from its own drawer.