ComfyUI models folder: where every model type goes (and Forge too)

ComfyUI models folder structure

These are the folders ComfyUI reads, as listed in its own folder_paths.py:

ComfyUI/
└── models/
    ├── checkpoints/        full checkpoints (SDXL, SD 1.5 ...)
    ├── diffusion_models/   diffusion models on their own (older installs: unet/)
    ├── text_encoders/      text encoders (older installs: clip/)
    ├── vae/                VAEs
    ├── loras/              LoRAs, image and video
    ├── embeddings/         textual inversion embeddings
    ├── controlnet/         ControlNet models (also t2i_adapter/)
    ├── clip_vision/        CLIP vision models
    ├── upscale_models/     upscalers (ESRGAN-style)
    ├── style_models/       style models
    ├── hypernetworks/      hypernetworks
    ├── vae_approx/         preview decoders
    └── configs/            .yaml configs

In the Windows portable download the folder is inside the unpacked folder: ComfyUI_windows_portable\ComfyUI\models\. The ComfyUI Desktop app keeps a shared model library in %LOCALAPPDATA%\Comfy-Desktop\ComfyUI-Shared on Windows and ~/ComfyUI-Shared on macOS.

Where to put LoRAs in ComfyUI

Every LoRA, image or video, goes in ComfyUI/models/loras/. ComfyUI also reads subfolders, so keep one per family or base and the right file is always at hand. Our family pages list the subfolders and sizes for each:

• Moss LoRAs: loras/moss-base/, loras/moss-sft/, loras/moss-turbo/
• Flux LoRAs: loras/flux/
• Wan 2.2 LoRAs: loras/wan22-5b/, loras/wan22-14b-high/, loras/wan22-14b-low/, loras/wan21/
• Z-Image LoRAs: loras/zimage-turbo/
• Flux 2 Klein LoRAs: loras/klein4b/, loras/klein9b/, loras/flux2-dev/
• Qwen Image LoRAs: loras/qwen-image/, loras/qwen-image-21/
• Krea 2 LoRAs: loras/krea2/
• SDXL LoRAs: loras/sdxl/, checkpoints/sdxl/, vae/sdxl/
• LTX-2 LoRAs: loras/ltx2/, loras/ltx23/
• Video LoRAs: Wan and LTX-2 side by side

After you add a file, refresh the LoRA list or restart ComfyUI.

Where checkpoints, diffusion models and text encoders go

A full checkpoint (model, text encoder and VAE in one file, like most SDXL and SD 1.5 checkpoints) goes in models/checkpoints/.

Newer families often come in parts: the diffusion model goes in models/diffusion_models/, the text encoders in models/text_encoders/ and the VAE in models/vae/. Older guides say unet/ and clip/: those are the old names, and ComfyUI still reads both.

VAE, ControlNet, upscalers and embeddings

• VAEs: models/vae/
• ControlNet models: models/controlnet/ (T2I adapters can also go in models/t2i_adapter/)
• Upscalers: models/upscale_models/
• Textual inversion embeddings: models/embeddings/
• CLIP vision models: models/clip_vision/

Use one models folder for ComfyUI and Forge

ComfyUI ships a file called extra_model_paths.yaml.example. Rename it to extra_model_paths.yaml, then either set base_path in its a1111 section to your Forge folder, or point the comfyui section at a central models folder. ComfyUI then reads models from there too, so you don't keep two copies.

Forge models folder

Forge uses the same layout as the AUTOMATIC1111 web UI, inside its install folder (webui/ in the one-click package):

• Checkpoints: models/Stable-diffusion/
• LoRAs: models/Lora/
• VAEs: models/VAE/
• Text encoders: models/text_encoder/
• ControlNet: models/ControlNet/
• Upscalers: models/ESRGAN/
• Embeddings: embeddings/

Where to download models, and how to keep a copy

Add a Civitai model to your vault in one click: paste the model link and BigVault imports it in the background, with its details and trigger words. Later, download it straight into the right folder on any computer.

More on this: keep your own Civitai archive.

Back up your whole models folder

Upload your whole models folder to BigVault and it sits in our own encrypted storage, private to you. Download it again on any computer with a link or a script (Downloads).

Next on our roadmap: a Windows tray app that keeps a folder in sync with your vault, so you can point ComfyUI at it.

ComfyUI models folder questions

Where do I put LoRAs in ComfyUI?

In ComfyUI/models/loras/. Subfolders are fine, such as loras/flux/ or loras/wan22-14b-high/. Then refresh the LoRA list or restart ComfyUI.

Where do checkpoints go?

Full checkpoints go in ComfyUI/models/checkpoints/. A model that comes as a separate diffusion model goes in models/diffusion_models/, with its text encoders in models/text_encoders/ and its VAE in models/vae/.

Why can't ComfyUI see my model?

Usually it is in the wrong folder, ComfyUI needs a refresh or restart, or the file has an extension ComfyUI doesn't load. Check the folder for that model type and use .safetensors files where you can.

Can ComfyUI and Forge share one folder?

Yes. Rename extra_model_paths.yaml.example to extra_model_paths.yaml in your ComfyUI folder and set base_path in its a1111 section to your Forge folder. ComfyUI then also reads Forge's models.

How much space does a models folder need?

It adds up fast: in our samples a Flux LoRA had a median of 73 MB and an SDXL LoRA 218 MB, and the SDXL base checkpoint is about 6.9 GB. Estimate; real numbers vary.