ComfyUI: the one that keeps the receipt

Elsewhere in this index sits a renderer that will not tell you which model made your image. ComfyUI is the opposite proposition, and it charges the opposite price. Every image it produces comes with a file recording exactly what produced it — model, seed, sampler, every step in order — and getting to that image will cost you an afternoon, a graphics card, and a tolerance for looking at a diagram where you wanted a button.

What it is and who makes it

ComfyUI is a node-graph application for running diffusion models. Instead of a prompt box you assemble a diagram: a loader feeds a sampler, encoders feed conditioning, a decoder turns the result into pixels, and every parameter along the way is a visible connection rather than a hidden default.

comfyanonymous released it on GitHub in January 2023. The creator had been involved with Stability AI; that ended by mid-2024, and Comfy Org formed around the core developers. The engine is Python under GPL-3.0, and as of September 2026 the repository carries roughly 131,000 stars — not a quality measure, but a fair proxy for how much open-weight work passes through here.

The company is newer than the software. Comfy Org now sells a hosted service, a desktop build, a developer platform and an enterprise tier, and is the sole official reseller of MiniMax commercial-use licences from $5,000 a month. The free core is genuinely free and genuinely unrestricted. It is also the top of a commercial funnel. Both are true, and neither is a complaint.

Where it sits in your week

Not where you might hope. ComfyUI does not read a Revit file, and no plugin makes it. Like everything else in this category, it takes pictures and returns pictures.

What changes is what you can put in front of it. A depth pass or a line export from a Rhino or SketchUp viewport wires in as a conditioning input, holding your massing while the model works on everything else — the same principle a hosted renderer applies behind a slider, except here you can see the slider, name it, and set it identically tomorrow.

Three real addresses in a week. The first is house style: once a graph produces the image the practice wants, it is a file, and a colleague who opens it gets that treatment rather than an approximation of it. The second is the confidential project, where the answer to where did the client's model go is that it never left the machine. The third is volume — forty options queued overnight against a fixed graph, which is a different activity from forty prompts typed by hand.

What it is genuinely good at

The graph is the product. Every workflow is a JSON file that can sit in the project folder beside the drawings, go into version control, and be read a year later by someone who was not there. Checkpoint, seed, sampler, step count, every conditioning input: recorded, and re-runnable. In a profession that already keeps a drawing register, this should not feel exotic. It is the only tool in this index that offers it.

Second, it is where the open-weight models actually live. Of the models indexed here that run in ComfyUI, every one is open-weight — Flux.2 Dev and Klein, Z-Image Turbo, Qwen-Image, HunyuanImage 3.0, Wan 2.7, HunyuanVideo 1.5, LTX-2.3, Hunyuan3D 2.1, TRELLIS.2, Stable Fast 3D. That says less about ComfyUI than about what open weights are for. This is the room they are kept in.

Third, training. Because the weights are yours, a LoRA trained on the practice's own past renders is a weekend project rather than an enterprise negotiation. A house style that belongs to the house is a real possibility here and nowhere else on this list.

Where it breaks

Start with the part no one's marketing mentions. In April 2026, researchers at Censys found more than a thousand internet-exposed ComfyUI instances being swept and compromised: where no vulnerable node was present, the attack installed one through ComfyUI-Manager, then enrolled the machine's GPU into cryptomining and a proxy botnet. It was not the first. In June 2024 a package called ComfyUI_LLMVISION harvested browser credentials from the people who installed it.

The mechanism is the same one that makes the tool good. A custom node is arbitrary Python running with the privileges of whoever launched the application, and there are well over a thousand of them. Comfy Org's registry now prohibits eval, exec, obfuscated code and runtime package installation, and bans packages once a problem is known — real work, and still a policy about the registry rather than a guarantee about the folder on your machine. For a practice this is not theoretical. It is a conversation with whoever looks after your computers, before the first install, about which machine this is allowed to run on.

Setup is the second cost, and an afternoon is the floor rather than the estimate. It wants a GPU: Flux.2 Dev at 32 billion parameters expects serious hardware, while Z-Image Turbo will run in 16GB of VRAM.

Third, and easily missed: the graph records what you ran. It does not record whether it was right. A workflow with a documented seed will still invent a mullion pattern you never drew, and it will document that too.

Fourth, the local story has a hole in it. Partner Nodes let you call hosted models — Nano Banana Pro among them — from inside a graph running on your own machine, billed against Comfy credits. Useful, and the moment one appears in a workflow the image is leaving the building. Nothing in the file format stops a colleague from adding one.

What it costs, what it asks

The core is free under GPL-3.0, on your own hardware, as a desktop build or from the repository.

Comfy Cloud, from Comfy's own pricing page checked September 2026, in US dollars: five free GPU runs without a card. Standard at $20 a month, or $16 billed yearly at $192, for 4,200 monthly credits and a thirty-minute cap per run. Creator at $35, or $28 yearly at $336, for 7,400 credits and the ability to bring your own models and LoRAs. Pro at $100, or $80 yearly at $960, for 21,100 credits and an hour per run. Team runs from $200 to $2,500 a month against a credit commitment rather than per seat, capped at fifty members. Comfy's reference unit is a five-second video at about eleven credits — a few hundred a month on Standard.

The data question has three different answers depending on which product you are using, and a practice needs to know which one it is giving. Run it locally and nothing leaves. Run it on Comfy Cloud and the work is on somebody else's GPU under their terms. Call a Partner Node from either and the image goes to Google or ByteDance under theirs. Only the first is the argument people are making when they say ComfyUI keeps the work in-house.

Who should bother

The principal: this is the answer when a client has written data restrictions into an appointment, and it is the only tool here that arrives with a security question attached. Route it through whoever manages your machines before anyone installs a node.

The project architect: worth it only if the same kind of image recurs. One hero render a quarter, use something hosted. The same three views on every scheme, and a graph that produces them identically each time repays the afternoon several times over.

The graduate: yes, and this is where fluency actually gets built. Wiring a sampler by hand teaches more about why images fail than a year of prompting does.

Ignore it if there is no GPU in the office and no appetite for buying one. Comfy Cloud closes that gap technically and reopens the data question that was the reason to be here.

Every other tool in this index asks you to accept some quantity of not knowing in exchange for not having to work. ComfyUI asks for the work and returns the knowing — a file you can keep beside the drawings that says what made the picture. Whether that is worth an afternoon depends on whether anyone is ever going to ask. The direction of travel in disclosure rules suggests they will.