ONLINEAGENT_OPS 2026.Q3 HOME HUBSMODELSBUILDINGCOMPANIESFILMMAKINGSECURITYLIVING WITH ITSKILLSALL HUBS, ONE TERMINAL → ARTICLES CRAFT RECORD BLOG TOOLS SKILLS SEARCH /
HOME›ARTICLES›COMFYUI EXPLAINED
ARTICLES · EXPLAINERINTERMEDIATE

ComfyUI Explained: What Node Workflows Are and Who Needs Them

What ComfyUI is, how its node workflows work in plain words, who needs it instead of a chat image tool, what hardware it needs and how to stay safe.

READ11 min
WORDS2,114
SECTIONS7
TYPEEXPLAINER
CHECKED12 OCT 26

ComfyUI is a free, open-source program for running AI image models, and also video, audio and 3D models, on your own computer. Instead of a chat box, you build the process as a diagram: boxes called nodes, each doing one job, joined by wires. Most people making a few pictures do not need it. It is for people who want exact, repeatable control over open models, and who will take care over what they download into it.

TL;DR — THE SHORT VERSION
  • What it is: an open-source app (GPL-3.0 licence) that runs image, video and other AI models through a visual graph of nodes.
  • A workflow is the whole diagram: load a model, write the prompt, generate, save. You can save it as a file and share it, and pictures made with it can carry the workflow inside them.
  • Who needs it: people who use open models, want ControlNet, LoRAs or custom pipelines, and want the same result again next month. A chat image tool is simpler for everyone else.
  • Hardware: its makers say it can run big open models on modest graphics cards, but each model you load states its own needs, and these can be higher. A slower processor-only mode exists.
  • The risk: custom nodes are code from strangers that runs on your computer, and older model files can hide code too. Treat both like any download from an unknown source.
◈ IN PLAIN TERMS

A chat image tool is a restaurant: you order, the kitchen is hidden, the dish arrives. ComfyUI is your own kitchen with every appliance on the counter. You decide which oven, which ingredients and in what order, and you can write the recipe down so the dish comes out the same next time. That is more work, and you are now responsible for what you bring into the kitchen.

What is ComfyUI?

The plain point: a visual workbench for open AI models that runs on your machine.

The project describes itself on its code page as “The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface.” Put more plainly: it is a program with a screen you work in (the GUI, or graphical user interface), built around diffusion models, the kind of AI behind most image makers. Its own summary of the main idea is “A visual node graph for building and reusing image, video, audio, 3D, and text workflows without code.”1 “Most powerful” is the makers’ claim, not this site’s.

The code is published under the GPL-3.0 licence, an open-source licence, and the core runs offline: the README says it “does not download anything unless you request it.”1 It works with many model families. Its README gives what it calls “a representative list”, covering image models such as Stable Diffusion and Flux, video models such as Wan and LTX-Video, plus audio and 3D models.1

Not everything in it is local, though. The README says “Partner nodes provide access to the best closed source models such as Nano Banana, Seedance, Hunyuan3D, etc.”1 Closed models are not downloadable, so those nodes send your work to the companies that run them, the same as using their own apps. If privacy is why you chose ComfyUI, stick to local models.

TAKEAWAY

ComfyUI is the workbench, not the AI. You bring the models; it connects them. Most of it runs on your own machine, but partner nodes reach out to online services.

What is a node workflow, in plain words?

The plain point: a flowchart where every box does one step and the wires carry the results along.

Behind any AI image app, the same few steps happen: load a model, turn your prompt into something the model reads, start from random noise, clean it up step by step, turn the result into a picture, save it. A chat tool hides all of that. ComfyUI shows each step as a node on the screen, and you join them with wires. A simple text-to-image workflow might look like this:

  1. Load checkpoint: picks the model file. See checkpoint meaning.
  2. Prompt boxes: one for what you want, often one for what to leave out (a negative prompt).
  3. Sampler: does the actual generating, using settings such as the seed, the number of steps and the CFG scale.
  4. Decode and save: turns the result into an image file.

The power comes from adding boxes. Put a LoRA node after the model to add a style. Add a ControlNet node to hold a pose. Add an upscaler at the end, or feed the picture into a video model. Each change is one more box and a couple of wires, and you can see exactly where it sits in the chain.

Two practical features follow from this. First, ComfyUI only reruns what changed: the README says “Only parts of the graph that change from each execution to the next will be executed”, so tweaking the last step does not redo the first.1 Second, workflows are files. You can “Save and load workflows as JSON” (a plain text format), and dragging a picture made in ComfyUI back onto the page “will give you the full workflow” that made it, seed included.1 That is why people share ComfyUI workflows the way cooks share recipes.

TAKEAWAY

A workflow is a recipe you can see, change one step at a time, save and share. That repeatability is the main reason to put up with the learning curve.

Who needs ComfyUI, and who does not?

The plain point: it pays off when you need control over open models; otherwise a chat image tool is quicker.

You are…UseWhy
Making the odd picture for a post or a cardUse: a chat image tool.Why: no setup, no hardware, good results from a sentence.
Needing the same character, style or layout across many imagesUse: try reference images first; consider ComfyUI if they fall short.Why: ComfyUI lets you lock seeds, LoRAs and ControlNets into a saved workflow.
Working with open models, privately, on your own machineUse: ComfyUI or a similar local app.Why: the work stays on your computer, apart from any partner nodes.
Building a repeatable pipeline (product shots, video frames, upscaling)Use: ComfyUI.Why: a workflow file runs the same steps every time and can be shared with a team.

Skip it if you have no interest in model files, settings or graphics memory: the chat tools covered in AI images and video for beginners will serve you better. The makers themselves point new users to the simplest route: “The desktop app is the easiest and best way to use ComfyUI for new users.”1 This page is not an install guide; the project’s own pages are the place for that.

What hardware does it need?

The plain point: a graphics card with its own memory helps a lot; how much depends on the model, not on ComfyUI.

ComfyUI’s documentation lists Windows, Linux and macOS, with Apple Silicon (M-series) Macs supported. On the graphics side it covers NVIDIA, AMD, Intel Arc and Apple Silicon chips, and it can run on the main processor alone using a --cpu option, which it marks as slower.2 That page gives no minimum memory figure.

The README goes further: it “Can run even the biggest open source models on as low as 4GB vram + 8GB ram relatively quickly”. VRAM is the memory on the graphics card.1 Treat that as the floor the makers describe, not a promise of comfort. Model makers state their own needs: Black Forest Labs, for example, says its small FLUX.2 [klein] 4B image model runs on consumer graphics cards with about 13GB of VRAM.3 Check the model card of whatever you plan to run, and see VRAM meaning for why that number matters.

TAKEAWAY

ComfyUI itself is light; the models are heavy. Read each model’s stated memory need before downloading it.

Is it safe to download custom nodes and models?

The plain point: treat every custom node and older model file as code from a stranger, because that is what it can be.

ComfyUI’s built-in nodes are only the start. The README invites you to “Extend ComfyUI with custom nodes”, and ComfyUI-Manager is “an extension that allows you to easily install, update, and manage custom nodes”.1 Security researchers at Snyk counted over 1,300 custom node extensions in December 2024. They are Python programs, loaded when ComfyUI starts, and Snyk warned that “A single vulnerability in an extension can jeopardize the security of the entire application.”4

This is not theoretical. ComfyUI’s own team wrote in January 2025: “We’ve already seen vulnerabilities discovered in ComfyUI custom nodes: ComfyUI_LLMVISION and ultralytics.”5 The second is a Python package whose releases were hit by a supply-chain attack: PyPI, the official index of Python packages, says four versions “were affected and have been removed”.6 In other words, a node can be honest and still pull in something poisoned from further up the chain. That is the same supply-chain problem this site describes for AI agent skills in the skill supply chain: you are installing someone else’s work, and you are the review process.

A second risk is running ComfyUI where the internet can reach it. The Cloud Security Alliance, an industry body, reported in April 2026 on a campaign found by the security firm Censys, with “more than 1,000 publicly accessible ComfyUI instances reachable without authentication”, and attackers using exposed copies to install malicious nodes.7 On your own computer, with nothing opened to the internet, that particular risk is far smaller.

Model files carry their own risk. Many older ones use Python’s pickle format, and Hugging Face warns of “dangerous arbitrary code execution attacks that can be perpetrated when you load a pickle file.” The .safetensors format was built to avoid that.8

The ComfyUI team has tightened its own registry of nodes. Its published standards say “The use of eval and exec functions is prohibited in custom nodes due to security concerns”, ban hidden (obfuscated) code, and forbid nodes from installing extra software while they run.9 Those rules cover nodes in that registry, not every repository on the internet.

A short safety checklist

  • Fewer nodes, better known. Snyk’s advice: “Install custom nodes only from reputable and actively maintained repositories.”4
  • Be wary of shared workflows that ask for nodes you do not have. Snyk notes that “users rarely perform security reviews before importing JSON workflows.”4 A workflow that needs five unfamiliar node packs is asking you to install five strangers’ code.
  • Prefer .safetensors model files from makers you can identify. See checkpoint meaning.
  • Keep it off the open internet unless you know how to put a password and a firewall in front of it.
  • Keep ComfyUI-Manager updated, as the Cloud Security Alliance advises.
  • Mind licences and likeness. Model files carry their own terms (see open weights meaning), and a real person’s face needs their permission (see deepfake meaning).
TAKEAWAY

The software is open and inspectable; the add-ons are where the risk sits. Install few custom nodes, from sources you can check, and keep your ComfyUI private.

Questions people ask

Is ComfyUI free?

The software is open source under the GPL-3.0 licence, and its core runs offline on your own hardware.1 Partner nodes reach closed models run by other companies, which have their own terms; check those before using them.

Do I need to know how to code to use ComfyUI?

No. The project describes building workflows “without code”, by connecting boxes on screen.1 You do need patience with settings, model files and folders, and the judgement to decide what is safe to install.

Is ComfyUI better than Midjourney or ChatGPT for images?

Neither is better in general; they do different jobs. A chat tool or Midjourney gives good results quickly from a sentence. ComfyUI gives control, privacy and repeatability with open models, at the cost of setup and hardware. Which image model compares the models themselves.

Can ComfyUI make video?

Yes, with the right models. Its README lists video model families such as Wan and LTX-Video.1 Check each video model’s stated memory needs first; how AI films are made shows where such tools fit in a production.

Learn more on this site

SOURCES
  1. 1Comfy Org, ComfyUI, GitHub README, undated, read at source 12 Oct 2026“The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface.” and “A visual node graph for building and reusing image, video, audio, 3D, and text workflows without code.” ↑alicence shown as “GPL-3.0 license”; “Runs fully offline: core does not download anything unless you request it.” ↑b“This is a representative list; browse the workflow library for maintained, ready-to-run templates.” The list includes SD 1.5, SDXL, SD3.5, Flux and Qwen Image for images and Wan 2.x, LTX-Video and HunyuanVideo for video. ↑c“Partner nodes provide access to the best closed source models such as Nano Banana, Seedance, Hunyuan3D, etc.” ↑d“Only parts of the graph that change from each execution to the next will be executed”. ↑e“Save and load workflows as JSON, or recover complete workflows and seeds from supported generated media.” and “Dragging a generated png on the webpage or loading one will give you the full workflow.” ↑f“The desktop app is the easiest and best way to use ComfyUI for new users.” ↑g“Can run even the biggest open source models on as low as 4GB vram + 8GB ram relatively quickly”. ↑h“Extend ComfyUI with custom nodes” and “ComfyUI-Manager is an extension that allows you to easily install, update, and manage custom nodes for ComfyUI.” ↑ilicence shown as “GPL-3.0 license”; “Runs fully offline: core does not download anything unless you request it.” ↑j“A visual node graph for building and reusing image, video, audio, 3D, and text workflows without code.” ↑kthe “representative list” includes Wan 2.x, LTX-Video, HunyuanVideo, CogVideoX and Mochi for video. ↑l
  2. 2Comfy Org, System Requirements, ComfyUI documentation, undated, read at source 12 Oct 2026“macOS (supports Apple Silicon, such as the M series)”; lists NVIDIA, AMD (Linux and Windows), “Arc series with native PyTorch torch.xpu support” and “M1/M2/M3/M4 series with Metal acceleration”; “Use the --cpu parameter (slower)”. ↑
  3. 3Black Forest Labs, FLUX.2 [klein] 4B, model card, as cited on this site’s guide to running AI free locally, read at source 22 Sep 2026“Runs on consumer GPUs (~13GB VRAM).” ↑
  4. 4Snyk Labs, Don’t Get Too Comfortable: Hacking ComfyUI Through Custom Nodes, 16 Dec 2024, read at source 12 Oct 2026“a popular stable diffusion platform with over 1,300 custom node extensions available”; “A single vulnerability in an extension can jeopardize the security of the entire application.”; “any malicious file placed here could be executed upon server restart”. ↑a“Install custom nodes only from reputable and actively maintained repositories.” ↑b“users rarely perform security reviews before importing JSON workflows.” ↑c
  5. 5Comfy Org, ComfyUI 2025 Jan Security Update, 3 Jan 2025, read at source 12 Oct 2026“We’ve already seen vulnerabilities discovered in ComfyUI custom nodes: ComfyUI_LLMVISION and ultralytics.” ↑
  6. 6PyPI, Supply-chain attack analysis: Ultralytics, The Python Package Index Blog, December 2024 (date from the post’s address), read at source 12 Oct 2026“Versions 8.3.41, 8.3.42, 8.3.45, and 8.3.46 were affected and have been removed from PyPI.” ↑
  7. 7Cloud Security Alliance, AI Inference Under Siege: The ComfyUI Cryptomining Botnet, research note, 8 Apr 2026, read at source 12 Oct 2026, citing Censys“more than 1,000 publicly accessible ComfyUI instances reachable without authentication”; its recommendations include to “restrict public internet access to all AI inference endpoints” and to update ComfyUI-Manager to version 3.38 or later, after a flaw that “enables unauthenticated remote manipulation of configuration data on instances running versions prior to 3.38.” ↑
  8. 8Hugging Face, Pickle Scanning, Hub documentation, undated, read at source 12 Oct 2026“There are dangerous arbitrary code execution attacks that can be perpetrated when you load a pickle file.”; Hugging Face, Safetensors, read at source 12 Oct 2026: “Safetensors is a new simple format for storing tensors safely (as opposed to pickle)”. ↑
  9. 9Comfy Org, Comfy Registry Node Quality and Security Standards, ComfyUI documentation, undated, read at source 12 Oct 2026“The use of eval and exec functions is prohibited in custom nodes due to security concerns.”; “Code obfuscation is prohibited in custom nodes.”; “Runtime package installation through subprocess calls is not permitted.” ↑

Comfy Org, System Requirements and Comfy Registry Node Quality and Security Standards, ComfyUI documentation, read at source 12 Oct 2026.

Hugging Face, Pickle Scanning and Safetensors, read at source 12 Oct 2026.

The comparisons on this page (the restaurant and the kitchen, workflows as recipes) and the “who needs it” table are this site’s own ways of explaining, not claims from these sources.

Free printable AI cheat sheets, and AI news every week.

Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.

SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK
aipromptgeneer://$ about this siteAI PROMPT GENEERAI explained in plain English. Every claim links to its source, with the date it was checked.NEWSLETTERWhat changed in AI each week, in plain English.SUBSCRIBE FREE ↗AI PROMPT GENEER · MACHINE ASSISTANCE DISCLOSEDEVERY ARTICLE CARRIES ITS OWN CHECKED DATE