Install ComfyUI

ComfyUI overview

What is ComfyUI

ComfyUI is a powerful and modular graphical interface for running and creating Stable Diffusion workflows for generating images and videos from text or images.

It allows you to build complex diffusion pipelines visually by connecting nodes, without needing to write code. It is designed for flexibility, performance, and transparency, letting you see exactly how each part of a diffusion model (VAE, CLIP, sampler, etc.) is connected and processed.

Installation

Installing on linux

Make sure you have Python 3.10 or higher and git installed:

sudo apt install python3 python3-venv python3-pip git

Clone the ComfyUI repository:

git clone https://github.com/comfyanonymous/ComfyUI.git
cd ComfyUI

Create a virtual environment and install dependencies:

python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

Install NVIDIA dependencies

pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu130
pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu130

Run ComfyUI:

python main.py

The interface will open in your browser at

http://127.0.0.1:8188

Installing and running on windows

1. Download comfyUI
2. Unzip the downloaded file
3. Double click either run nvidia_gpu.bat or run cpu.bat depending on your computer’s configuration to launch ComfyUI.
4. You can access to http://127.0.0.1:8188 when you see it in the opened window.

---

Understanding workflows and nodes

What workflows are

A workflow in ComfyUI is a graph that defines how data (text prompts, models, images, etc.) flows between nodes to generate an image.

Example workflow on image to video:

Each workflow is saved as a .json file that can be shared or loaded into ComfyUI. They are stored in your ComfyUI/user/default/workflows folder.

You can import a workflow by dragging directly into the interface:

• a JSON file
• media (image or video) generated with comfyUI

What nodes are

Nodes are the building blocks of a workflow. Each node performs a specific task, such as:

• loading a model
• encoding a text prompt
• generating noise
• decoding the latent image

Nodes are connected through inputs and outputs, creating a chain of operations that define the diffusion process.

Common node types include:

Checkpoint Loader - loads the main diffusion model

VAE Loader - loads a Variational Autoencoder

CLIP Text Encoder - processes text prompts

Sampler - controls the diffusion steps

Save Image - exports the final result

You can double-click on the canvas to add new nodes or search for them by name.

---

ComfyUI folder structure

ComfyUI organizes its files in a clear directory structure:

/ComfyUI/models/checkpoints - main diffusion models

/ComfyUI/models/vae - Variational Autoencoders (VAEs)

/ComfyUI/models/clip - CLIP encoders for text and image embeddings

/ComfyUI/models/upscale - upscaling models

/ComfyUI/custom_nodes - external or user-made node scripts

/ComfyUI/output - generated media

ComfyUI/user/default/workflows - saved workflow files in JSON format

Loading assets on your workflow

Depending on how much VRAM does your graphics card has and how much RAM you have, you will be able to use resources up to the total amount of memory you have.

Example

• Computer available memory 56GB, distributed in:
  • graphics card VRAM: 24GB
  • computer RAM: 32GB

• Resources on nodes 38,5GB:
  • Model wan2.2_i2v_high_noise_14B_Q8_0.gguf : 15GB
  • Model wan2.2_i2v_low_noise_14B_Q8_0.gguf : 15GB
  • LoRA wan2.2_i2v_A14b_high_noise_lora_rank64_lightx2v_4step_1022.safetensors: 0,6GB
  • LoRA wan2.2_i2v_A14b_low_noise_lora_rank64_lightx2v_4step_1022.safetensors: 0,7GB
  • VAE wan_2.1_vae.safetensors: 0,3GB
  • Text Encoder umt5_xxl_fp8_e4m3fn_scaled.safetensors: 6,7GB
• ComfyUI processing and temporary data: 10-20% of assets

With this configuration, we would fit the requirements, as we have 56GB of memory and the process requires less than 50GB.

If comfyUI stops abruptly, it's probably due to an excessive usage of memory. You will have to reduce the size of the assets you are using.

Assets resources

Diffusion models

Wan GGUFs (Models)

https://huggingface.co/bullerwins/Wan2.2-I2V-A14B-GGUF/tree/main

Diffusion models

https://civitai.com - community hub for diffusion models (SD1.5, SDXL, LoRA, etc.)

https://huggingface.co/models - official and community models

https://modelscope.cn/models - AI models for various domains

GGUF models

GGUF format is used for quantized models optimized for inference.

You can find them at:

https://huggingface.co - TheBloke profile (repository of GGUF models)

https://github.com/ggerganov/llama.cpp - technical details on the GGUF format

VAE (Variational Autoencoders)

VAEs are used to improve image color, clarity, and detail.

Common sources:

• https://civitai.com
• https://huggingface.co

CLIP models

CLIP encoders transform text or images into embeddings.

They are usually included with diffusion models, but standalone versions can be found on:

https://huggingface.co/openai/clip-vit-large-patch14

LoRAs and ControlNets

These extend model behavior for styles, poses, or compositions. Available on:

https://civitai.com/models (LoRA and ControlNet section)

https://huggingface.co/models?q=controlnet index.php?title=Category:AI: Generate media locally