ComfyUI Model Folder Cheat Sheet: Where Every Model File Goes (2026)

The most common reason a ComfyUI workflow fails is not a bug. It is a model file sitting in the wrong folder. You download a 12 GB file, drop it into checkpoints, and the loader node shows nothing, or you get a red “value not in list” error. This cheat sheet fixes that for good. It lists every ComfyUI model folder, which loader node reads it, what kind of file goes there, and how to tell from a file name where it belongs. It also has a quick-reference table of exact paths for the popular 2026 models (Qwen Image 2.1, FLUX 3 Klein, Krea 2, MiniMax H3, MOSS-Image, Anima and more), plus GGUF files, custom node folders and how to share one model folder between several installs. Bookmark it; you will come back to it.

Where Is the ComfyUI Models Folder?

Every model lives under one models folder. Where that folder is depends on how you installed ComfyUI:

In this guide, every path starts with ComfyUI/models/. On the portable build, put ComfyUI_windows_portable\ in front of it.

Master Table: Every Model Folder

These are the folders a current ComfyUI install creates or understands. The core ones come with ComfyUI itself; a few newer ones appear when you update.

If a folder in a guide does not exist on your PC, you can simply create it. ComfyUI reads any of the folder names above.

Which Loader Node Reads Which Folder

When a dropdown in a node is empty, look at which folder that node reads:

A common mistake is putting a diffusion-only file in checkpoints and using Load Checkpoint. That fails, because the file has no text encoder or VAE inside. Use Load Diffusion Model instead.

Checkpoint vs Split Models: The Big Difference

Older Stable Diffusion models shipped as one file that held all three parts. One Load Checkpoint node gave you MODEL, CLIP and VAE outputs.

Almost every big model since FLUX ships split. The diffusion model is huge, and the text encoder is shared between many models, so they come as separate files:

ComfyUI/models/diffusion_models/   the main model (Load Diffusion Model)
ComfyUI/models/text_encoders/      the text encoder (Load CLIP)
ComfyUI/models/vae/                the VAE (Load VAE)

The upside is that you download a shared text encoder or VAE only once. For example, qwen_image_vae.safetensors is used by Krea 2, Anima and other models, and flux2-vae.safetensors is shared by several FLUX-family models.

How to Tell the Folder From the File Name

Model authors usually hint at the folder in the name. Use these rules when a download page does not say:

The precision suffix tells you nothing about the folder. bf16, fp16, fp8_scaled, int8_convrot and nvfp4 only describe how the weights are stored.

Quick Reference: Exact Paths for Popular 2026 Models

These are the paths we have used in our own model guides. Each row links to the full guide.

Image models

Video, audio and 3D

Tools

Where GGUF Files Go

GGUF is a compressed format (Q8_0, Q6_K, Q5_K_M, Q4_K_M and so on) that makes big models fit on smaller GPUs. ComfyUI does not load GGUF on its own. You need the ComfyUI-GGUF pack by city96:

cd ComfyUI/custom_nodes
git clone https://github.com/city96/ComfyUI-GGUF
pip install -r ComfyUI-GGUF/requirements.txt

On the Windows portable build, install the requirement with the embedded Python:

python_embeded\python.exe -m pip install -r ComfyUI\custom_nodes\ComfyUI-GGUF\requirements.txt

Then:

• GGUF diffusion models go in ComfyUI/models/unet/, for example ComfyUI/models/unet/flux1-fill-dev-Q4_0.gguf. Load them with Unet Loader (GGUF).
• GGUF text encoders go in ComfyUI/models/text_encoders/ (or clip), for example ComfyUI/models/clip/t5-v1_1-xxl-encoder-Q5_K_M.gguf. Load them with CLIPLoader (GGUF) or DualCLIPLoader (GGUF).
• The VAE stays a normal .safetensors file in ComfyUI/models/vae/.

Not sure which quant to download? Our GGUF quantization guide explains Q4_K_M vs Q5_K_M vs Q8_0 and how to match them to your VRAM, and the MiniMax H3 GGUF workflow shows a real example.

Folders Created by Custom Nodes

Many custom node packs add their own folder under ComfyUI/models/. Always check the pack’s README, but these are the usual ones:

The custom node code itself never goes in models. It goes in ComfyUI/custom_nodes/<pack-name>/, usually by git clone or through ComfyUI Manager.

Using Subfolders to Stay Organised

You can make subfolders inside any model folder, and ComfyUI shows them in the dropdown as subfolder/file.safetensors:

ComfyUI/models/loras/
  krea2/
    krea2_turbo_3step_rank_64_lora_comfyui.safetensors
  qwen-image/
    p_q_image_2.1_6_steps_v0.2_comfyui_bf16.safetensors
  minimax-h3/
    minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors

One catch: a shared workflow saves the full name, including the subfolder. If you load someone else’s workflow and the file is in a different subfolder on your PC, just pick it again in the dropdown.

Share One Model Folder Between Installs

If you run the portable build and Comfy Desktop, or ComfyUI and another app, do not keep two copies of 100 GB of models. ComfyUI can read extra folders listed in extra_model_paths.yaml.

1. In your ComfyUI folder, find extra_model_paths.yaml.example.
2. Copy it and rename the copy to extra_model_paths.yaml.
3. Add a section that points at your shared folder:

shared:
    base_path: D:/AI/models/
    checkpoints: checkpoints/
    diffusion_models: |
        diffusion_models
        unet
    text_encoders: |
        text_encoders
        clip
    vae: vae/
    loras: loras/
    upscale_models: upscale_models/
    controlnet: controlnet/
    clip_vision: clip_vision/

4. Restart ComfyUI. Files from the shared folder now appear next to the local ones.

Use forward slashes in the YAML even on Windows, and keep the indentation exact. Another option on Windows is a folder link, made in an administrator Command Prompt:

mklink /D "C:\ComfyUI_windows_portable\ComfyUI\models\loras" "D:\AI\models\loras"

Delete or rename the original loras folder first, because the link needs the name to be free.

Model Not Showing Up? Checklist

1. Right folder? Check the loader table above. Diffusion-only files do not work in Load Checkpoint.
2. Right install? Many people have two ComfyUI folders and copy files into the one they are not running. The console shows the base path at startup.
3. Refreshed? Press R in the ComfyUI window to refresh the model lists, or restart.
4. Download finished? A partial download shows as a smaller file or ends in .part or .crdownload. Compare the size with the download page.
5. Right extension? Files must end in .safetensors, .ckpt, .pt, .pth, .bin or .gguf. Some browsers add .txt or strip the extension.
6. GGUF without the pack? GGUF files only show in the GGUF loader nodes after ComfyUI-GGUF is installed.
7. “Value not in list” on a shared workflow? The workflow expects an exact file name. Click the dropdown and pick your file.
8. Error while deserializing header? The file is corrupt or not really a safetensors file (often an HTML error page saved with the model’s name). Download it again.

For the full list of error messages and fixes, see our ComfyUI errors and fixes guide.

FAQ

Is diffusion_models the same as unet?

Yes. unet is the older name. Current ComfyUI reads both for Load Diffusion Model. Use diffusion_models for safetensors files and unet for GGUF files to match most guides.

Is text_encoders the same as clip?

Yes, clip is the older name and both are read. New guides use text_encoders.

Where do I put a .ckpt file?

If it is a full SD 1.5 or SDXL model, in ComfyUI/models/checkpoints/. Prefer the .safetensors version when one exists, because .ckpt files can contain code.

Where do embeddings go and how do I use them?

In ComfyUI/models/embeddings/. Write embedding:filename in the prompt, without the extension.

Can I rename model files?

Yes. ComfyUI does not care about the name. But shared workflows look for the original name, so keeping it saves you re-selecting files.

Do I need to restart after adding a model?

No. Pressing R refreshes the lists. New custom node packs, however, do need a full restart.

Final Thoughts

Once you know the pattern, ComfyUI’s folders are simple: one-file models go in checkpoints, split models use diffusion_models, text_encoders and vae, add-ons go in loras, and GGUF files go in unet with the GGUF pack. Use the quick-reference tables when you set up a new model, the file-name rules when a download page is vague, and extra_model_paths.yaml once your collection outgrows one drive.

More ComfyUI guides:

• ComfyUI Errors and Fixes: The Complete List
• GGUF Quantization Explained: Which File to Download
• How Much VRAM Do You Need for ComfyUI?
• Best Open-Source Text-to-Image Models

Sources

• GitHub: ComfyUI (folder layout and extra_model_paths.yaml.example)
• GitHub: city96/ComfyUI-GGUF
• comfyart.org: Models guide (tool model folders)
• Model paths from the official model cards linked in each Freakyjolly guide above