Repo Image & Video

Comfy-Org/ComfyUI

Node-graph GUI, API and backend for diffusion and other generative models, with local-hardware optimization, a desktop app, portable builds and cloud options.

  • 137k GitHub stars
  • Python
  • ⚖️ GPL-3.0
  • 🎯 Intermediate
View on GitHub (opens in a new tab) Setup guide
pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu130
Comfy-Org/ComfyUI preview image

What it is

ComfyUI is a modular node-graph interface, API and backend for generating images, video, 3D models, audio and text. It focuses on efficient local inference on consumer hardware, using asynchronous queueing, smart VRAM/RAM management, model offloading and weight streaming. It runs on Windows, Linux and macOS via a desktop app, portable install, manual install or Comfy Cloud, and can be extended with custom nodes.

Who it's for

  • Visual professionals and creatives who want fine control over models, parameters and outputs
  • Developers integrating generative workflows into production pipelines via API endpoints
  • Users running large open-source models on limited local GPU hardware
  • Users on NVIDIA, AMD, Intel, Apple Silicon or Ascend hardware

Requirements

Requirements

  • Supported OS: Windows, Linux or macOS
  • Supported GPU types for manual install: NVIDIA, AMD, Intel, Apple Silicon, Ascend
  • Python 3.13 is very well supported; Python 3.14 works but some custom nodes may have issues
  • PyTorch: torch 2.7 is minimally supported and a newer version is strongly recommended; cu130 or above is required on Nvidia 20 series and above
  • Model checkpoints placed in models/checkpoints and VAEs in models/vae
  • Can run large models on as low as 4GB VRAM + 8GB RAM, per the README

Setup

  1. Desktop application (recommended)

    Download the desktop app from comfy.org/download; it is available on Windows and macOS and is described as the easiest way to get started.

  2. Windows portable

    Download the portable 7z build for your GPU (Nvidia, AMD or Intel) from the releases page and extract it with 7-Zip or Windows Explorer, then run. Put checkpoints in ComfyUI\models\checkpoints. Not recommended for regular users.

  3. Install with comfy-cli

    Install and start ComfyUI using comfy-cli.

    bash
    pip install comfy-cli
comfy install

  4. Manual install: NVIDIA PyTorch

    Git clone the repo, put checkpoints in models/checkpoints and VAEs in models/vae, then install stable PyTorch for NVIDIA.

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

  5. Manual install: dependencies

    Open a terminal inside the ComfyUI folder and install the dependencies.

    bash
    pip install -r requirements.txt

  6. Enable ComfyUI-Manager (optional)

    Install the manager dependencies, then run ComfyUI with the --enable-manager flag.

    bash
    pip install -r manager_requirements.txt
python main.py --enable-manager

Examples

Install and start via comfy-cli

bash
bash
pip install comfy-cli
comfy install

What it does: Installs comfy-cli and uses it to install and start ComfyUI.

Install stable PyTorch for AMD on Linux

bash
bash
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm7.2

What it does: Installs ROCm 7.2 PyTorch for AMD GPUs on Linux before installing ComfyUI dependencies.

Launch ComfyUI with the manager enabled

bash
bash
python main.py --enable-manager

What it does: Enables ComfyUI-Manager for installing, updating and managing custom nodes.

Run in offline mode

bash
bash
python main.py --offline

What it does: The README states --offline disables the optional paid Comfy API nodes and keeps all built-in functionality offline. The command form is minimally adapted from the flag description.

Pros & cons

Pros

  • Pro:Broad native model support across image, video, audio, 3D and text generation
  • Pro:Optimized local execution with smart VRAM/RAM management, model offloading and asynchronous weight streaming, enabling large models on as low as 4GB VRAM + 8GB RAM
  • Pro:Multiple install paths (desktop app, portable, comfy-cli, manual) covering NVIDIA, AMD, Intel, Apple Silicon and Ascend
  • Pro:Can run fully offline, and workflows can be saved as JSON and exposed through an API

Cons

  • Con:Commits outside stable release tags may be very unstable and break many custom nodes
  • Con:Manual install requires choosing and installing the right PyTorch build per GPU vendor
  • Con:Windows portable build is not recommended for regular users, and the cu126 build must not be used on 20 series and newer GPUs

Images

Comfy-Org/ComfyUI preview image
Comfy-Org/ComfyUI preview image · Comfy-Org/ComfyUI on GitHub · GPL-3.0
Comfy-Org/ComfyUI preview image