Contents

ComfyUI Model Download and Folders Tutorial: Where to Put Checkpoints, LoRAs, and VAEs

ComfyUI does not include any models. Before you generate images, download models and put them in the right folders.

Model folders

All models go in the models folder inside the ComfyUI directory:

ComfyUI/models/
├── checkpoints/        Complete models (SD 1.5, SDXL, etc.)
├── diffusion_models/   Diffusion model only (common for newer models such as Flux, Qwen-Image, and Wan)
├── text_encoders/      Text encoders (CLIP, T5, Qwen, etc.)
├── vae/                VAE
├── loras/              LoRA
├── controlnet/         ControlNet
├── upscale_models/     Image upscaling models (ESRGAN, etc.)
├── clip_vision/        Image encoder
├── embeddings/         Textual Inversion
└── vae_approx/         Small models for previews during generation

The desktop app's model folder is in the location you chose during installation. After adding a model, press R in the interface to refresh; the model will appear in the drop-down menu.

Two ways models are packaged

All-in-one checkpoints

Earlier models such as SD 1.5 and SDXL usually bundle the diffusion model, text encoder, and VAE into one file. Put it in checkpoints and load it with the Load Checkpoint node.

Separate files

Newer large models such as Flux, Qwen-Image, and Wan often split each component into separate files:

File	Folder	Load node
Diffusion model	`diffusion_models`	Load Diffusion Model
Text encoder	`text_encoders`	Load CLIP (or DualCLIPLoader, etc.)
VAE	`vae`	Load VAE

This lets different models share the same text encoder, and lets you choose versions at different precision levels. When using a new model, the easiest way to get started is with a built-in template. The template shows which files you need and where to put them. See Next-generation models such as Flux.

Where to download models

Site	Content
Hugging Face	Original models released by their authors and versions repackaged by the ComfyUI team
Civitai	Community fine-tuned models and a large selection of LoRAs

When downloading:

Prefer the .safetensors format. Older .ckpt / .pt formats may contain malicious code.
Check which base model the model is based on (SD 1.5, SDXL, Flux, etc.). LoRAs and ControlNets must match the base model to work.
Read the license terms. Some models restrict commercial use.

Recommended first model

If you are not sure where to start:

VRAM	Recommendation
4–6 GB	Models from the SD 1.5 family (about 2 GB), at 512×512 resolution
8 GB	SDXL models (about 6.5 GB), at 1024×1024 resolution
12 GB or more	Try quantized / FP8 versions of newer models such as Flux and Qwen-Image
24 GB or more	Full versions of newer models and video generation

Start with an SDXL model to get familiar with ComfyUI. The examples in the next section also focus on SDXL.

Model precision and file size

The same model often comes in several versions:

Label	Description
`fp16` / `bf16`	Standard precision
`fp8`	About half the file size with little quality loss; useful when VRAM is limited
`GGUF` (Q8, Q5, Q4…)	Quantized versions that are smaller and require the corresponding custom nodes to be installed
`pruned` / `emaonly`	Training data has been removed; this is all you need for generation

If you already have models for Stable Diffusion WebUI, you do not need to copy them. Rename extra_model_paths.yaml.example in the ComfyUI folder to extra_model_paths.yaml, then change base_path to point to the WebUI folder:

a111:
    base_path: D:/stable-diffusion-webui/

    checkpoints: models/Stable-diffusion
    vae: models/VAE
    loras: |
         models/Lora
         models/LyCORIS
    upscale_models: |
                  models/ESRGAN
                  models/RealESRGAN
    embeddings: embeddings
    controlnet: models/ControlNet

Restart ComfyUI to use the models from WebUI. In the desktop app, you can also add extra model paths in settings.

Tips for organizing models

Create subfolders to organize models (for example, checkpoints/SDXL/ and loras/characters/). ComfyUI will read them too.
Keep the URL of each model's download page so you can look up its trigger words and recommended parameters later.
Periodically remove models you no longer use; they take up a lot of space.

Updated October 12, 2026