ComfyUI
ComfyUI is an open-source engine under GPL-3.0 for generating images, video, audio and 3D with AI, built as a visual node graph you wire together without code. It supports the newest open models, including Stable Diffusion, SDXL, Flux, Qwen Image, and Wan and LTX for video, with LoRAs, ControlNets and inpainting. It runs fully offline on NVIDIA, AMD, Intel and Apple hardware, with a paid cloud version available.
What does ComfyUI do?
ComfyUI calls itself "the most powerful and modular AI engine for content creation", its own claim. Its interface is "a visual node graph for building and reusing image, video, audio, 3D, and text workflows without code", and an App Mode turns a finished graph into a simple form. The README says it "natively supports the latest open-source state of the art models", listing Stable Diffusion 1.5, SDXL, SD3.5, Flux.1 and Flux.2, Qwen Image, HiDream and others for images, and Wan 2.1 and 2.2, LTX-Video, HunyuanVideo, CogVideoX and Mochi for video.
It loads LoRAs, ControlNets, adapters and upscalers, and has inpainting, outpainting, masks, model merging, frame interpolation and segmentation built in. Only the parts of a graph that change between runs are executed again. It "runs fully offline: core does not download anything unless you request it", and a flag turns off the paid API nodes that reach closed models. A generated PNG carries the full workflow and seeds that made it. A local API lets other applications run workflows.
Key facts
- Licence: GPL-3.0, with no added conditions. The models it runs carry their own licences, which the README does not discuss.
- A node-graph interface for image, video, audio, 3D and text workflows, plus App Mode for simple forms.
- Image models listed include Stable Diffusion 1.5, SDXL, SD3.5, Flux.1, Flux.2, Qwen Image, Z-Image, HiDream and Chroma; video models include Wan 2.1 and 2.2, LTX-Video, HunyuanVideo, CogVideoX and Mochi.
- LoRAs, ControlNets, adapters, upscalers, inpainting, outpainting, model merging and frame interpolation.
- Hardware: NVIDIA, AMD, Intel, Apple Silicon and Ascend; the standard NVIDIA portable build supports 20-series cards and newer.
- Runs fully offline; optional paid API nodes reach closed models and can be turned off with a flag.
- Install: a desktop app for Windows and macOS, a Windows portable build, comfy-cli with pip, or from source with Python and PyTorch 2.7 or later.
- Hosted: Comfy Cloud, the official paid version. The repository moved to Comfy-Org/ComfyUI. Latest tagged release when read: v0.36.0 on 2026-09-15; 133,702 stars on 2026-09-17.
What does it replace, and where does it fall short?
ComfyUI replaces Midjourney for someone who wants to generate images and video on their own GPU with open models they choose, and to control every step of the process as a graph they can save, share and rerun. It is the pick for control and the newest models on the open-source alternatives to Midjourney page.
Where it falls short: a node graph takes time to learn, where Midjourney takes a sentence. Quality depends on the model you download and the GPU you have, and each model carries its own licence. The README warns that commits between stable tags can break custom nodes. For a simpler, canvas-first interface, InvokeAI or SwarmUI.
How does ComfyUI run?
The README calls the desktop app for Windows and macOS the easiest way to start. There is also a Windows portable build for NVIDIA, AMD and Intel, installation through pip install comfy-cli and comfy install, or a manual install with Python and PyTorch. Models download separately. Comfy Cloud is the official paid hosted version for those without the hardware.
Who is ComfyUI for?
A creator or technical artist who wants the widest model support and full control over image and video generation on their own hardware. Someone who wants to type a prompt and get a picture should start with InvokeAI or SwarmUI's simpler tabs.
What limits does the README state?
From the README: commits outside stable release tags "may be very unstable and break many custom nodes"; Python 3.14 works but some custom nodes may have issues; the portable build is not recommended for regular users.
Questions people ask
Is ComfyUI open source?
Yes. ComfyUI is GPL-3.0 with no added conditions. The image and video models you run in it are separate downloads with their own licences, some of which restrict commercial use, so check each model's licence. Comfy Cloud and the paid API nodes are optional.
Can ComfyUI run offline?
Yes. The README says it runs fully offline and that the core does not download anything unless you ask it to. The paid API nodes that reach closed models can be disabled with a flag, so nothing leaves your machine.
What GPU does ComfyUI need?
It supports NVIDIA, AMD, Intel, Apple Silicon and Ascend. The standard NVIDIA portable build supports 20-series cards and newer, with a separate build for older cards. The README does not state a minimum amount of video memory; that depends on the model.
How does ComfyUI compare with Midjourney?
ComfyUI runs open models on your own hardware with full control over each step, for free, and handles video, audio and 3D as well as images. Midjourney is a hosted service driven by a prompt. The open-source alternatives to Midjourney page compares it with InvokeAI, SD.Next and SwarmUI.
Which models does ComfyUI support?
The README lists Stable Diffusion 1.5, SDXL, SD3.5, Flux.1 and Flux.2, Qwen Image, Z-Image, HiDream and Chroma for images, and Wan, LTX-Video, HunyuanVideo, CogVideoX and Mochi for video, plus audio, 3D and text models. It adds support for new open models as they appear.
Sources
- ComfyUI README and GPL-3.0 LICENSE: github.com/Comfy-Org/ComfyUI, read 2026-09-17.
Compared with the others
On the open-source alternatives to Midjourney page, ComfyUI is the pick for control, and the newest models. GPL-3.0 node graph for image, video, audio and 3D, with Stable Diffusion, SDXL, Flux, Wan and LTX, LoRAs and ControlNets, running fully offline.
Also on that page: InvokeAI for a canvas for inpainting and editing, SD.Next for modest or unusual hardware, SwarmUI for a simple tab now, a graph later.
More image and video generation
An Apache-2.0 creative image tool with a unified canvas for inpainting and outpainting, node workflows, and support for 28 model families from Stable Diffusion to Flux and Qwen Image, run as a local web app.
An Apache-2.0 all-in-one web interface for AI image and video generation, descended from AUTOMATIC1111's WebUI, with LoRA, ControlNet, quantisation that cuts video memory up to four times, and the widest hardware list here.
An MIT image and video generation web interface, formerly StableSwarmUI, with a simple Generate tab for beginners and a raw ComfyUI workflow tab for experts, that can spread work across several GPUs.
Added September 17, 2026. Every claim above comes from the project's README, LICENSE or model card, read on September 17, 2026, or from the GitHub API on the date shown in the panel. Found an error? Write to reveneau@licheo.com and it is fixed in the next weekly pass. Repository: github.com/Comfy-Org/ComfyUI.
Running one of these inside your own environment, with your own data and your own security rules, is the kind of work Reveneau does. Read how a forward deployed engagement works.