ComfyUI

There are two things to count: how much VRAM your card has and whether you are ready to spend the first evening on installation rather than pictures.

Who it is for. For people who generate a lot and want control over every step: a designer maintaining a brand style through custom LoRAs, a video creator who needs Wan and MiniMax H3 without a subscription, a studio that builds one pipeline and runs it hundreds of times. And for those who already have a graphics card with 12 GB of memory or more: then the tool itself costs nothing.

Who else. For people who do not want to hand prompts and images over to someone else's server. Your own installation works without the internet, and desktop telemetry is enabled only with your consent and sees neither workflows, prompts, nor results.

Who it is not for. Those who want to write a phrase and get an image. The first screen greets you with a graph of a dozen nodes. And those without a graphics card: cloud at $20 per month or hourly server rental is all that remains.

Suitable: section 3 will take you from a graphics card to your first image. If you are choosing between ComfyUI and Automatic1111, see the choice in section 7.

What vibecoding isStarting from scratch: how to set tasks for agents with no coding experience

Three layers: open models on your graphics card, third-party nodes and templates from the registry, and partner nodes that call closed models directly into the workflow for credits.

Workflow. One node loads a model, one reads text, one generates, one saves, with lines between them. Only the parts of the workflow that changed since the previous run are calculated, so editing one word in a prompt does not run everything again. The finished image carries the workflow inside it: drag someone else's file into the window, and ComfyUI rebuilds its entire pipeline along with the seed. Workflows are stored in JSON, and a complex workflow can be hidden behind a simple form through App Mode.

If your team needs a shared cloud workflow for game art, compare the Layer workspace: references, style training and a team file library.

Models. As of Sep 7, 2026, the vendor's README lists the families it supports without third-party servers: for images Stable Diffusion, SDXL, Flux and Qwen Image, for video Wan 2.2, LTX-2 and HunyuanVideo, for audio the ACE-Step model and MiniMax Music, for 3D Hunyuan3D, plus the text models Gemma and Qwen. New models arrive on release day: Wan 2.2 on Jul 28, 2025, Flux.2 on Nov 25, 2025, MiniMax H3 on Aug 3, 2026. Weights repackaged for ComfyUI live in the Comfy-Org organization on Hugging Face, 84 repositories, downloaded 69.9 million times in the 30 days to Sep 7, 2026. The full list of families is in the FAQ.

Nodes and templates. The vendor counts more than 60 thousand community nodes across over five thousand extensions; as of Sep 7, 2026, the registry contains 5,496 packages, each scanned for malicious code. They are installed by ComfyUI-Manager, the extension store now built into the core and enabled with a flag. The template library has 617 ready-made workflows in nine sections: 177 for video, 160 for images, 37 for 3D, 35 for audio.

Partner nodes. Closed models are called from the same workflow without your own keys: Veo 3.1, Kling, Runway, Nano Banana, GPT-Image, 182 models from 25 vendors in total. You pay with Comfy credits: a Kling image costs 5.91 credits, one second of Veo 3.1 with audio costs 84.4. The nodes are optional; the --disable-api-nodes flag disables them and all internet connectivity.

Exit route. A workflow is exposed as an API to your own app, and ComfyUI connects to AI agents through MCP: cloud Comfy MCP launched on Jun 29, 2026, the vendor released a local MCP server on Aug 18, 2026, and an agent can build and run workflows itself. Graphics-card support is also broad: NVIDIA, AMD on Linux and Windows, Intel Arc, Apple Silicon through Metal, and without a graphics card it runs on the CPU, slowly.

What the vendor itself shows. The header video is clipped from a tutorial on the Comfy channel dated Jun 9, 2026: the camera moves across the instructor's large working graph, where Veo, Kling, and Seedance nodes sit side by side with run prices in credits, and the Comfy Cloud label is visible below. This is what a professional's working workflow looks like, not a beginner's first screen, which is shown in the image above. We did not run ComfyUI ourselves for this task; the figures come from documentation, pricing pages, the GitHub and Hugging Face APIs, and third-party tests.

The easiest way to install is the Comfy Desktop installer.

Step 1. A graphics card. The vendor names no memory minimum, by its tutorials Stable Diffusion 1.5 runs on 6 GB and the smaller Wan 2.2 on 8 GB, while practitioners advise 12 GB and up for Flux and 24 GB and up for video. The main build is built for NVIDIA 20-series and newer, AMD and Intel have separate archives, and a Mac on Apple Silicon is supported through Metal.

Step 2. Download Comfy Desktop from comfy.org/download: Windows 10 and later, macOS 13 and later on Apple Silicon, 4.85 GB of disk space, 8 GB of RAM. The installer will install Python and dependencies itself. On Linux and for those who want the latest commits, use the command line:

pip install comfy-cli

comfy install

Or manually: clone the repository, install PyTorch for your GPU, run python main.py, and open 127.0.0.1:8188. A portable Windows archive with Python included is also on GitHub, but the vendor recommends it only for experienced users.

Step 3. Models and nodes. Open the template library, choose a workflow, and ComfyUI will offer to download the required weights from Hugging Face. Put your own models and LoRA from Civitai in the models folder. Manager, the extension store, is included with the desktop version; in a manual installation, enable it with the --enable-manager flag after the pip install -r manager_requirements.txt command.

Without your own GPU. RunPod offers an RTX 4090 for $0.74 per hour, while RunComfy and ThinkDiffusion offer ComfyUI as a service from $0.99 per hour (Sep 7, 2026). Free Colab prohibits this kind of launch under its rules.

Steps verified on Sep 7, 2026 against the repository README, Comfy documentation, the Citilink storefront, immers.cloud and RunPod pricing, and a check-host test with control nodes

There are no rental figures for models; ComfyUI is not a model, so these metrics cover scale and demand rather than image quality.

Code and community. 131,829 stars and 15,547 forks on GitHub (API, Sep 7, 2026), with only 74 repositories across all of GitHub ranking higher. The vendor’s Discord has 60,730 members.

Company funding. On Apr 24, 2026, Comfy Org raised $30 million at a $500 million valuation, $47 million total across two rounds. In the same post, the company claims 4 million users and 150 thousand downloads a day; these figures have not been independently verified.

Demand. The place among image tools is shown under the card title, and the whole tool ranking sits on the shared page.

ComfyUI itself costs $0 forever; you pay for two separate scarcities, a GPU and closed models, and those two budgets are independent of each other.

Currency. The cloud and partner nodes charge in credits. The vendor does not publish an exchange rate, but its resource table lists an RTX 6000 Pro hour at both $3.49 and 736.39 credits, or 211 credits per dollar; the same rate matches across four plans and Google's Veo pricing. A five-second Wan 2.2 video in the cloud uses 11 credits, about $0.05.

Cloud. 5 free runs before purchase, no card required. Standard is $20 per month for 4,200 credits, Creator $35 for 7,400, Pro $100 for 21,100, Team from $200 to $2,500 for a shared pool, and Enterprise on request. The storefront defaults to annual billing and shows $16, $28, and $80. Everything runs on RTX 6000 Pro with 96 GB of memory; a workflow can run for 30 minutes, or an hour on Pro, and you pay only for the time while the workflow is actually running.

Credits. Monthly ones burn out at the end of the period, topped-up ones live a year, and the vendor describes refunds two ways: the terms call purchases final, while the pricing page FAQ invites you to write to support and handles cases one by one. Closed models go at their vendor's price or slightly higher. Veo 3.1 costs exactly what Google charges, $0.40 per second with sound, Nano Banana Pro is 20 % more expensive, Runway's models are 43 % more expensive than their own API. A picture through Kling costs 5.91 credits.

Quiet price cut. The prices of the three plans have not changed since Dec 18, 2025; what they provide has changed. In December 2025, Standard promised 164 five-second videos; since February 2026 it has promised 380, because video became cheaper from 25.6 to 11 credits. The annual discount and Team plan appeared between 03.07 and 24.08.2026.

Four questions: what the license permits, who owns the output images, how dangerous third-party nodes are, and whether an update will break your workflow.

License. The GPL-3.0 core can be installed, modified, and run as your own service, but if you distribute a modified copy, you must provide the source code. Comfy Desktop is separately under AGPL-3.0, and anyone who wants to embed it in a closed product buys a license from Comfy Org. The vendor distributes workflow templates under MIT.

Third-party nodes. The main risk is not the core but extensions. In June 2024, the ComfyUI_LLMVISION node stole passwords and crypto wallets; in December 2024, an infected ultralytics package installed a miner on Mac and Linux; in October 2025, two registry packages were downloaded 790 times before a scanner and maintainer banned them. Since January 2025, the registry has checked code, prohibited eval and on-the-fly package installation, and given approved nodes a badge next to their name in Manager.

Core vulnerabilities. Five vulnerability records were published on Jul 31, 2026; four had already been closed by version 0.28.0 on 15.07, while the most serious one, CVE-2026-68771, critical at 9.3–9.8 on different scales, allowed a malicious data file to run foreign code on a computer with ComfyUI 0.23.0 from 01.06.2026. The fix was merged on Jun 18, 2026 and released in 0.26.0 on Jun 23, 2026, a month before the record was published. Keep your version current and do not expose ComfyUI to the internet without a password.

Data. A self-hosted installation sends nothing, and desktop telemetry is enabled only with consent and, under the vendor's rules, does not see workflows, prompts, images, or weights. In the cloud and partner nodes, prompts go to Comfy servers in the US and model vendors, but without your name or account. The agreement is with Comfy Organization of Delaware under California law, with disputes resolved through arbitration in San Francisco.

Updates break nodes. This is the community's loudest complaint. In the issue «What is happening in ComfyUI country» from Dec 16, 2025, the core author replied that compatibility with custom nodes had never been a priority and that he runs his tests without a single third-party node. The repository has 4,207 open issues and another 613 unreviewed pull requests as of 07.09.2026. Practical takeaway: update the core after, not before, delivering a project.

The choice is not image quality, since everyone uses the same models, but two questions: workflow or form, and is the project alive?

Automatic1111. The classic Stable Diffusion web interface with tabs and sliders, where most people started. It has more GitHub stars, 164,844 versus 131,829 for ComfyUI, but the main branch has not changed since Jul 27, 2024, the development branch since Mar 2, 2026, and the latest working release, 1.10.1, came out in July 2024. New models such as Wan and Flux.2 do not arrive there. Its Forge fork has stood still since summer 2025, while the simplified Fooocus by the same author has not been developed since August 2024.

InvokeAI. The only live neighbor from the old guard: release 6.14.1 came out Sep 6, 2026, Apache-2.0 license, 28 thousand GitHub stars versus 132 thousand for the main subject.

Neighbors inside. SwarmUI, from a Comfy Org co-founder, and a Krita plugin use ComfyUI as their engine and hide the graph behind a simple form. And anyone who wants no hardware at all goes to or to Midjourney : there are no nodes or custom models, but no installation either.

In three and a half years, a Quebec programmer's hobby project grew into a company valued at $500 million, with its own cloud and node registry, while surviving the author's departure from Stability AI, malicious extensions, and a move to a new interface.