ComfyUI is a free, open-source app for running generative AI models, where each job is a graph of connected nodes instead of a form. It runs open image, video, audio and 3D models on your own computer, and saves every workflow as a JSON file or inside the files it produces. The latest release on 14 October 2026 is v0.39.0. Beginners should install Comfy Desktop and open a template. The paid parts, Comfy Cloud and credit-based Partner Nodes for closed models, are optional.
ComfyUI is a free, open-source app for running generative AI models, where you build each job as a graph of connected nodes instead of filling in a form. One node loads a model, another encodes the prompt, another samples the image, another saves it. That graph is the workflow, and you can save it, share it and run it again with the same settings.
Checked on 14 October 2026 against Comfy Org's own pages: the ComfyUI README and releases on GitHub, the docs on workflows, templates, Comfy Desktop, system requirements and Partner Nodes, and the Comfy Cloud pricing page. Prices are US dollars. An earlier version of this page carried speed and adoption claims with no source; they are gone.
ComfyUI in one minute
Four facts place it. It is open source under GPL-3.0. It is local first: the README says it "runs fully offline: core does not download anything unless you request it". It is maintained by a company, Comfy Org. And it moves fast, which is the main reason guides about it go stale.
- 9 Sep 2026v0.35.0
- 15 Sep 2026v0.36.0
- 21 Sep 2026v0.37.0
- 29 Sep 2026v0.38.0
Added Qwen-Image 2.1 support, among others.
- 5 Oct 2026v0.39.0
Added the --offline and --disable-partner-nodes flags. --disable-api-nodes is deprecated.
Source: ComfyUI releases on GitHub, read 14 October 2026.
The README describes the cycle as a new stable version roughly every two weeks, with releases aimed at Mondays, and adds that this changes often. Treat any version number in a tutorial, including this one, as a date stamp. For a plain-language definition, see ComfyUI in the glossary.
How ComfyUI works
A workflow is a recipe, and each node is one instruction. Comfy's docs put it this way: "A node performs one specific job", "A link connects an output from one node to an input on another node", and "Technically, a workflow is a node graph." The default text-to-image workflow shows the whole pattern in six nodes.
- 01Load Checkpoint
Loads the model, its text encoder and its VAE.
- 02CLIP Text Encode
Turns the positive and negative prompts into conditioning.
- 03Empty Latent Image
Sets the size and the batch.
- 04KSampler
Runs the diffusion steps.
- 05VAE Decode
Converts the result into pixels.
- 06Save Image
Writes the file, with the workflow inside it.
Add a node and you add a capability: a loader for a LoRA, a ControlNet for pose, an upscaler, or a video model that animates the still. That is the whole idea. Everything else is knowing which nodes a job needs.
The features you use every day
- Workflows travel with your files. Save a workflow as JSON, or let the built-in save nodes embed it in the output. The metadata docs list PNG, animated WebP, MP4 and WebM among the formats. Drag the file back onto the canvas to recover the graph. Start with
--disable-metadataif you share originals and do not want to share the workflow. - Templates. Click Templates in the sidebar, pick a workflow, and ComfyUI checks for the model files it needs and prompts you to download any that are missing (docs).
- Subgraphs. A subgraph packs part of a workflow into one reusable node, which the docs compare to a folder. It needs frontend 1.24.3 or later.
- App Mode. App Mode, officially supported from frontend 1.41.13, puts a simple input and output screen on top of a finished workflow so someone else can run it without touching nodes.
- Low-memory running. The README says ComfyUI can run large open models on "as low as 4GB vram + 8GB ram" by streaming weights. It works. It is not quick.
What can you run in ComfyUI?
Most open image, video and audio models, usually within days of release. The README lists the families below and calls its list representative, not complete.
| Category | Examples named in the README |
|---|---|
| Image | Stable Diffusion 1.5, SDXL, SD3.5, Flux.1, Flux.2, Qwen Image, Z-Image, Hunyuan Image 2.1, HiDream, Chroma, Lumina Image 2.0 |
| Image editing | Flux Kontext, Flux.2 Klein, Qwen Image Edit, HiDream E1.1, OmniGen2 |
| Video | Wan 2.1 and 2.2, LTX-Video 2 and 2.3, HunyuanVideo 1.5, Kandinsky 5 Video, CogVideoX, Cosmos Predict2, Mochi |
| Video with audio | MiniMax H3, LTX-AV |
| Audio | ACE-Step 1.5, Stable Audio 3, MiniMax Music 3 |
| 3D and vision | Hunyuan3D 2.1, SeedVR2, SUPIR, Depth Anything 3, SAM 3, BiRefNet |
| Text | Gemma 3 and 4, Qwen3, Qwen3-VL |
Each open model carries its own licence, and the software being free does not make every model free for commercial work. Our best ComfyUI models guide compares licences and memory needs, and there are separate pages for SDXL and Z-Image. Closed models are reached through Partner Nodes, covered under costs below.
How is ComfyUI different from other tools?
It shows you the pipeline. A hosted generator or a form-based interface hides the steps and gives you a prompt box and a few settings. ComfyUI puts every step on the canvas, which costs you an afternoon of learning and pays you back every time you need to change one step without disturbing the rest.
- Every step is visible and editable
- A workflow is a file you can save, share and version
- New open models tend to arrive here first
- Runs on your own hardware, offline
- Slower to learn
- One box and a few settings
- Results are quick on day one
- The pipeline is fixed by the vendor
- Usually hosted and paid by subscription or credits
- Harder to repeat a result exactly
For the detail, read ComfyUI vs Stable Diffusion, which clears up the model-versus-interface confusion, and ComfyUI vs Automatic1111. If the worry is the learning time, how hard is ComfyUI lays out the stages.
Is ComfyUI free? What costs money
The software is free. What you can end up paying for is hardware, closed models and hosting, and all three are optional. Prices below were read on 14 October 2026.
| Item | Price | Notes |
|---|---|---|
| ComfyUI core | Free | Open source under GPL-3.0, run on your own hardware |
| Comfy Desktop | Free | Dual-licensed: AGPL-3.0-or-later or a commercial licence from Comfy Org |
| Open model files | Free to download | Each model has its own licence. Some do not allow commercial use |
| Partner Nodes | Prepaid credits | Optional. Calls closed models hosted by outside services. Top-up credits expire one year after purchase |
| Comfy Agent | Comfy Credits after a free trial amount | Optional and in beta. Generally available in Comfy Cloud; local Desktop support is listed as coming soon |
| Comfy Cloud | Standard $20, Creator $35, Pro $100 a month | Optional. $16, $28 and $80 a month billed yearly. A free tier gives 5 runs |
| Your own hardware | A GPU and disk space | Model files are the main storage need |
Two notes on the paid rows. Comfy's docs call Partner Nodes "completely optional"; they need a Comfy account with a positive credit balance, and you can remove them with --disable-partner-nodes. Comfy Cloud runs on RTX 6000 Pro GPUs with 96 GB of VRAM, and its plan credits reset each billing cycle without rolling over (pricing). To weigh it against renting a GPU yourself, see our cloud GPU comparison.
Comfy Agent is the newest paid piece: an in-app assistant that plans, builds and runs workflows from a plain-language request. Comfy describes it as in beta, generally available in Comfy Cloud, with local Desktop support still to come.
How do you get started?
Install Comfy Desktop, open a template, and change one setting at a time. The download page offers three routes, and the README adds a fourth.
| Route | Best for | What it is |
|---|---|---|
| Comfy Desktop | Most people starting out | A one-click installer with auto-updates. It sets up Python and dependencies for you. Windows 10+, macOS 13+ on Apple Silicon, Linux (.deb and AppImage) |
| Windows portable | Windows users who want a folder they can move | A self-contained build that runs from any folder, with builds for Nvidia, AMD, Intel or CPU |
| Manual install | Any operating system and GPU type | Clone the repository and install with Python 3.13, which the README recommends. Enable the Manager with --enable-manager |
| comfy-cli | People who script their setup | pip install comfy-cli, then comfy install |
| Comfy Cloud | No suitable GPU, or a team | The hosted version in a browser. Nothing to install, paid by plan |
- 1Install Comfy Desktop
At least 8 GB of RAM (16 GB recommended) and 4.85 GB of disk before any models.
- 2Open a template
Templates in the sidebar. Let it download the files it asks for.
- 3Run it, then change one thing
Seed, steps, prompt. One at a time, so you learn what each node does.
- 4Add custom nodes only when a workflow needs them
Through ComfyUI-Manager, from authors you can identify.
- 5Build toward one goal
A consistent character, a product shot or a short video. Not everything at once.
On hardware, the system requirements cover NVIDIA, AMD (ROCm on Linux and on Windows for RDNA 2 to RDNA 4), Intel Arc and Apple Silicon from M1 to M4. A --cpu mode exists and the docs describe it as slower. The requirements page recommends Chrome 143 or later for the interface. Step-by-step instructions and error fixes are in our ComfyUI installation guide, and the ComfyUI Manager guide covers the node manager.
Where this leads for most readers here is a repeatable character: a base model, a trained face, a pose control and an upscaler saved as one workflow. The AI Influencers program teaches that pipeline, from Flux and SDXL setup to LoRA training, ControlNet and image-to-video.
Are custom nodes safe?
Custom nodes are ComfyUI's biggest strength and its main risk. The project's security policy is blunt about it: "Custom nodes are arbitrary Python code and are trusted as much as any other software the user chooses to install."
The Comfy Registry is the public collection behind ComfyUI-Manager. A published version cannot be changed, and the docs say nodes are scanned for malicious behaviour, with a verification flag for those that pass. That lowers the risk; it does not remove it. In April 2026, Censys reported a campaign that used vulnerable custom nodes on unauthenticated, internet-exposed ComfyUI servers to install mining malware, and counted over 1,000 such servers visible online.
- Install only the nodes a workflow needs
- Prefer authors you can identify and node packs with a history
- Keep ComfyUI-Manager at its default security level
- Never expose ComfyUI to the internet with --listen
- Update ComfyUI and your nodes regularly
- Read what a downloaded workflow will install before you accept it
What to read next
Pick by goal. For a persona that looks the same in every image, start with the consistent character workflow. For editing existing images, img2img explains denoise strength. For ready-made graphs, the workflow library guide lists safe sources. The AI Influencers library collects the rest.
ComfyUI: FAQ
What is ComfyUI used for?
ComfyUI runs generative AI models through workflows built from connected nodes, on your own computer or in Comfy Cloud. Creators use it for text-to-image, image editing, image-to-video, audio and 3D, and for repeatable pipelines that chain a base model with LoRAs, ControlNet, detailing and upscaling. The same graph can be saved, shared and run again with identical settings.
Is ComfyUI free?
Yes. The core is open source under GPL-3.0 and free to run on your own hardware, and Comfy Desktop is a free download. Two things are paid and optional: Partner Nodes, which spend prepaid credits to call closed models, and Comfy Cloud, the hosted version, which listed plans from $20 a month on 14 October 2026.
Is ComfyUI hard to learn?
It takes longer than a one-box prompt tool, because you see every step of the generation. You do not have to build graphs from nothing, though. The Templates browser opens ready-made workflows and offers to download the model files they need, and App Mode can turn a finished workflow into a simple form for someone else to run.
What hardware do I need for ComfyUI?
Comfy Desktop lists Windows 10 or later, macOS 13 or later on Apple Silicon, or Linux, with at least 8 GB of RAM (16 GB recommended) and 4.85 GB of disk per installation before models. A dedicated GPU is strongly recommended. The README says ComfyUI can run large open models on as little as 4 GB of VRAM and 8 GB of RAM by streaming weights, slowly.
Does ComfyUI work offline?
Yes. The README says the core runs fully offline and does not download anything unless you request it. Partner Nodes and Comfy Agent need a network connection and a Comfy account. Since v0.39.0 the launch flag --offline keeps everything local, and --disable-partner-nodes turns off only the paid nodes.
Is ComfyUI safe to use?
The core is maintained by Comfy Org, but custom nodes are a different matter. The security policy describes them as arbitrary Python code, trusted as much as any other software you choose to install. Install only the nodes a workflow needs, from authors you can identify, keep ComfyUI-Manager at its default security level, and never expose ComfyUI to the internet with --listen.
Can ComfyUI make videos?
Yes. The README lists open video models that run locally, including Wan 2.1 and 2.2, LTX-Video 2 and 2.3 and HunyuanVideo 1.5, plus MiniMax H3 and LTX-AV for video with sound. Closed video models are reachable through Partner Nodes, which need a Comfy account and prepaid credits.