Download Models

ComfyUI needs AI model files to generate images and videos. This guide covers where to download them and where to put them.

Model Types

Type	Folder	What It Does	File Size
Checkpoint	models/checkpoints/	The main generation model	2-12 GB
LoRA	models/loras/	Fine-tuned style/concept add-on	10-200 MB
VAE	models/vae/	Image encoder/decoder	300-800 MB
ControlNet	models/controlnet/	Guides generation with reference images	700 MB-1.5 GB
Upscaler	models/upscale_models/	Increases image resolution	20-200 MB
CLIP	models/clip/	Text encoder	200 MB-2 GB
UNET	models/diffusion_models/	Core diffusion model (Flux, etc.)	5-24 GB

Where to Download

Hugging Face

The largest open-source model hub. Most official model releases happen here.

• Browse: huggingface.co/models
• Search for models by name and download .safetensors files

CivitAI

Community-driven platform with thousands of fine-tuned models, LoRAs, and embeddings.

• Browse: civitai.com
• Filter by Type (Checkpoint, LoRA, etc.) and Base Model (SD 1.5, SDXL, Flux)

Recommended Starter Models

For Image Generation

Model	Base	VRAM Needed	Best For
Stable Diffusion 1.5	SD 1.5	4 GB+	Fast generation, huge plugin ecosystem
RealVisXL	SDXL	8 GB+	Photorealistic images
Flux Schnell	Flux	12 GB+	Fast, high-quality generation
Flux Dev	Flux	12 GB+	Higher quality, slower

For Video Generation

Model	VRAM Needed	Notes
Wan 2.1	12 GB+ (quantized)	Text-to-video and image-to-video
HunyuanVideo	16 GB+	High-quality video generation
LTX-Video	8 GB+	Lightweight video model

How to Install

1. Download the .safetensors file (avoid .ckpt files when possible — .safetensors is safer)
2. Place it in the correct subfolder under models/:

ComfyUI/
└── models/
    ├── checkpoints/     ← Main models go here
    ├── loras/           ← LoRA files go here
    ├── vae/             ← VAE files go here
    ├── controlnet/      ← ControlNet models go here
    └── upscale_models/  ← Upscaler models go here

1. Restart ComfyUI (or click Refresh in the model dropdown)

Using GGUF Quantized Models

If you have limited VRAM, you can use GGUF quantized versions of large models (like Flux). These use less memory at a small quality cost.

Quantized models require the ComfyUI-GGUF custom node:

1. Install the ComfyUI-GGUF node (see Install Custom Nodes)
2. Download a GGUF version of your model (e.g., flux1-schnell-Q4_K_S.gguf)
3. Place it in models/diffusion_models/
4. Use the GGUF Loader node instead of the regular Load Checkpoint node

Tips

• .safetensors vs .ckpt — Always prefer .safetensors. The .ckpt format can contain executable code and is a security risk.
• Check the base model — A LoRA trained on SD 1.5 won't work with SDXL checkpoints. Always match the base model.
• Organize subfolders — You can create subfolders inside models/checkpoints/ (e.g., models/checkpoints/sdxl/). ComfyUI scans subdirectories automatically.
• Share models — If you also use A1111 or Forge, configure extra_model_paths.yaml to share model files. See Portable Package setup.
• Check your VRAM — Not sure if a model will run on your GPU? See System Requirements and GPU Compatibility for VRAM guidance.

Next Steps

• Generate your first image
• Install custom nodes
• Common issues — troubleshooting model loading errors