What is ComfyUI? Node-based interface for creating images and videos with AI running on your machine: installation, requirements, and risks

What is ComfyUI?

ComfyUI is an open-source software with a node-based interface (a graph of connected blocks) for creating images, videos, audio, and 3D models using diffusion models like Stable Diffusion, SDXL, and Flux, running directly on your machine or server. The project, according to GitHub (as of 24/10/2026), uses the GPL-3.0 license, runs on Windows, Linux, and macOS, supports NVIDIA, AMD, Intel, and Apple Silicon GPUs, and saves workflows as JSON. Its strengths are step-by-step control and workflow reusability; its weaknesses are greater complexity than Automatic1111 and security risks from community nodes if installed carelessly.

The article goes through five parts: 1) how it works; 2) installation methods; 3) hardware requirements; 4) security risks; 5) OPCAIOS's installation experience on a GPU server. Finally comes a practical trial section. For example, a designer can build a workflow once in ComfyUI and then run it again for multiple images in the same style.

ComfyUI connects nodes into a workflow to create images and save them for running again or sharing. Diagram: OPCAIOS.

What is ComfyUI and who created it?

ComfyUI is a node-based image creation software released by an anonymous author with the nickname comfyanonymous on GitHub (16/1/2023 according to Wikipedia's infobox). The main milestones, according to Wikipedia (as of 24/10/2026), include three steps: 1) the author previously worked with Stability AI, the collaboration ending before 3/6/2024, after which Comfy Org was founded along with core developers; 2) in August 2024 ComfyUI added support for Flux and Comfy Org joined the Linux Foundation's Open Model Initiative; 3) today the codebase sits under the Comfy-Org organization, with around 142.3 thousand stars at the time of reading.

Wikipedia recorded 122 thousand stars in July 2026, so the star count is growing fast. Wikipedia describes ComfyUI as more complex than Automatic1111 but more flexible. For example, people used to a form-based interface will find the node graph harder at first, but in return can save and share the entire process as JSON or a PNG image. The star counts and timestamps we have not yet cross-checked with other sources.

How does ComfyUI work?

ComfyUI works by connecting nodes into a workflow, each node doing one thing like loading a model, encoding text prompts with CLIP, sampling, or decoding images. According to the GitHub repository (as of 24/10/2026), ComfyUI has five main points.

• Visual node graph to build and reuse workflows for images, video, audio, 3D, and text; has subgraph, workflow templates, App Mode, and local API.
• Asynchronous queue and partial graph rerun: only reruns the part that changed.
• VRAM and RAM management, pushing models out of memory, supporting quantized models.
• Workflow as JSON, and drag created PNG images into the page to restore the entire workflow, including random seed.
• Offline run: the --offline flag disables paid API nodes; --disable-partner-nodes disables only partner nodes.

The listed models (partial list): images (SD 1.5, SDXL, SD3.5, Flux.1, Flux.2, Qwen Image), video (Wan 2.1 and 2.2, LTX-Video, HunyuanVideo, CogVideoX, Mochi), audio (ACE-Step, Stable Audio), 3D and vision (Hunyuan3D, SAM 3, BiRefNet), and text (Gemma, Qwen3). To understand the language model foundation that comes with it, see What is an LLM?.

What are the ways to install ComfyUI?

ComfyUI has five installation methods, according to the GitHub repository (read on 24/10/2026).

• Method 1) Desktop application for Windows and macOS, recommended for beginners.
• Method 2) Windows portable version, standalone, unzip and run, supports NVIDIA, AMD, Intel GPU or CPU only.
• Method 3) Manual installation for all supported operating systems using git clone then pip install -r requirements.txt.
• Method 4) comfy-cli, installed via pip for those comfortable with the command line.
• Method 5) Comfy Cloud, paid hosted version for those without a GPU.

For manual installation on NVIDIA cards series 20 and above, the repository requires PyTorch cu130 or later, Python, PyTorch with the correct CUDA or ROCm version, and model files (checkpoint) placed in the models folder. For example, beginners on Windows should start with Desktop instead of manual installation. Read the README first because version requirements change. If you need a server with a GPU, see What is VPS?.

Table of installation methods

According to the GitHub repository:

Method	For	Notes
Desktop application	Windows and macOS	Recommended for beginners
Windows portable	NVIDIA, AMD, Intel GPU or CPU only	Standalone, unzip and run
Manual installation	All supported operating systems	git clone then pip install -r requirements.txt
comfy-cli	Command line	Installed via pip
Comfy Cloud	No GPU	Paid hosted version

For manual installation on NVIDIA cards series 20 and above, the repository requires PyTorch cu130 or later. Manual installation requires Python, PyTorch with the correct CUDA or ROCm version, and model files (checkpoint) placed in the models folder; read the README first because version requirements change. If you need a server with a GPU, see What is VPS? about how to rent and manage a server.

What hardware does ComfyUI need?

ComfyUI requires a supported GPU, and according to the GitHub repository (read on 24/10/2026) the project mentions GPUs from NVIDIA, AMD (ROCm on Linux and Windows), Intel (torch.xpu) and Apple Silicon, plus some NPUs. The repository claims the largest open models run with about 4 GB VRAM and 8 GB RAM thanks to memory management and model offloading. This is the project's claim; running is not the same as fast, and OPCAIOS has not verified the 4 GB figure. In reality, speed depends on three factors: model, resolution and number of steps, plus VRAM capacity. For example, in OPCAIOS testing on an RTX A4000 16 GB card, a 1024 px SDXL image at 25 steps took about 10 seconds, and VRAM used about 12 GB; the result only applies to one model and one card. For those without a GPU, Comfy Cloud or hourly GPU services are alternatives, but OPCAIOS has not read the pricing.

What security risks do ComfyUI custom nodes and models have?

Custom nodes are ComfyUI's extension system, both a great power and an attack surface. One incident, according to Wikipedia (read on 24/10/2026): in June 2024 the hacker group "Nullbulge" compromised the ComfyUI_LLMVISION extension, inserted malware and exposed login credentials of hundreds of users; vpnMentor said the extension could steal crypto wallets, take screenshots and steal files. General rules include five steps:

• Step 1: only install nodes from sources you trust, read the code or check community reviews first; limit the number of nodes.
• Step 2: run on your own machine or account, not shared with a browser logged into important accounts and crypto wallets.
• Step 3: do not open ComfyUI ports to the Internet; by default it listens locally, access remotely via SSH or authenticated VPN.
• Step 4: download models from trusted sources, prefer safe formats (safetensors).
• Step 5: back up workflows and record the versions of nodes used.

For example, those who installed ComfyUI_LLMVISION in 2024 on a machine with a crypto wallet may have had their wallet exposed, so Step 2 is the most important.

When should you choose ComfyUI?

Need	Suggestion
Control each step, repeatable workflows, automation via API	ComfyUI
Create images quickly with a simple interface	Simpler direct tools or hosted services
No GPU, need to try fast	Comfy Cloud or hourly GPU services
Sensitive data cannot leave the machine	Run locally, offline mode
Ad images need clear licenses	Read the license of each model before commercial use

About licensing: ComfyUI is GPL-3.0, but each model you download has its own license; read it before using for commercial purposes.

How OPCAIOS Installed and Tested ComfyUI

According to the operations log from October 22, 2026, OPCAIOS installed ComfyUI (version 0.42.0) on a GPU server with an RTX A4000 16 GB card, following three steps.

• Step 1: install ComfyUI together with the ComfyUI-Manager node manager and the SDXL base 1.0 model.
• Step 2: run as a background process, listening only to local address, access via SSH, not exposed to the Internet.
• Step 3: call the API to generate SDXL images at 1024 px, 25 steps.

In practice, we tested and measured around 10 seconds per image, the first time around 56 seconds because the model had to be loaded, and VRAM usage was around 12 GB, shared with another service on the same card. When checked again on October 24, 2026, the service was not responding so OPCAIOS cannot confirm the service was running at the time of writing. This was only a single experiment with one model and one card, not an official performance evaluation; OPCAIOS has not yet run video workflows or Flux.

Frequently Asked Questions

What is ComfyUI?

It is open-source software with a node interface for creating images, videos, audio and 3D using diffusion models, running on your machine.

Is ComfyUI free?

The software is open-source, GPL-3.0 license, runs free on your machine. Comfy Cloud is a paid hosted version; we haven't read its pricing.

What configuration does ComfyUI need?

According to the GitHub repository, it supports NVIDIA, AMD, Intel and Apple Silicon GPUs; the project claims large models run with around 4 GB VRAM thanks to memory management, but larger VRAM offers better speed and resolution.

How is ComfyUI different from Automatic1111?

ComfyUI uses node graphs, flexible and reusable workflows but more complex; Automatic1111 has a more familiar form interface according to general Wikipedia descriptions.

Is it safe to install custom nodes?

Not all nodes are safe; there was a case of malicious code injection in an extension in 2024. Only install from trusted sources and run in isolated environments. To learn how to safely use AI for image generation in your business, see OPCAIOS's AI application training course.

Sources: Comfy-Org/ComfyUI GitHub repository and Wikipedia "ComfyUI", read directly on October 24, 2026; docs.comfy.org page returned 404 so we didn't read it; star count changes quickly. The 4 GB VRAM claim is from the project, not yet verified. Performance measurements are from one test by OPCAIOS according to the October 22, 2026 log. Experience is from OPCAIOS.