Plugging ComfyUI into Claude: Free AI Images and Video on Your Home PC

The video opens with a disarmingly simple promise: high-quality AI images and video, free, on your own computer. The instrument is ComfyUI, the well-known node-based tool that lays out every step of image synthesis like stations on a production line. That power gave it a fearsome reputation — beginners saw a maze of boxes and wires and walked away. The video's whole argument is that two recent changes, built-in templates plus an AI assistant at your side, have flattened that learning curve.

Installation is now consumer-grade. You download the installer from the official site, double-click, and face one decisive question: cloud or local. Cloud means renting GPU power by subscription; local means your own machine does the work. The video chooses local and is honest about the price of that choice: modest hardware caps which models you can run and how fast frames appear. A couple of agreement checkboxes, an install path, a skip-and-install click — and the components download themselves.

The blank-canvas era is presented as history. Where early users wired every node by hand, today's ComfyUI ships with ready-made templates you can browse by model. A small crown badge marks templates that only run in the cloud — either the hardware appetite is too big or the model is not open — and a checkbox hides those paid items so you only see what runs on your desk. The demo picks a Z-Image-Turbo image template, and the red error badge in the corner is treated as normal: the model weights simply have not been downloaded yet.

Then comes a patient tour of node literacy: drag boxes to rearrange, follow the wires that carry data left to right, tweak each node's parameters, unfold the ones that hide an entire sub-workflow inside. Canvas navigation gets its shortcuts — scroll to zoom, hold space to pan, select and press period to focus — before the video returns to that error badge and shows the kindest feature in the program: one button downloads every missing model automatically, no hunting across mirror sites. Large files take a while, a progress button watches the queue, and a refresh shortcut clears the error when the weights land.

Where did several gigabytes of weights actually go? The answer sits on a card at the canvas edge listing each model's folder, and a Storage tab link opens the real directory so you can verify — or delete what you no longer need to reclaim disk. The first real test generates a still: a woman in traditional dress brewing tea in a courtyard, 1280 by 720, batch count set, Run pressed, the result blooming in the final node. Right-click saves it; an Assets panel collects everything you have ever generated; an Output folder mirrors it on disk. One more habit closes the loop: any edit dots the tab to mark unsaved work, the save shortcut names the workflow, and the Workflows panel reopens it any evening you want to generate again.

The second demo switches media entirely: ACE-Step, a text-to-music model, under the audio category. Same ritual — missing-model error, download-all, refresh — then a completely different form: a style field set to energetic Funk, a lyrics box supporting more than fifty languages including Chinese and English, tempo at 120, a bright major key, and a deliberately short sixty-second test length. Run, listen, and a small menu downloads the song. The point lands quietly: the same canvas that painted the tea scene now composes.

Video is where the free story meets its limit, and the video is unusually candid about it. The newest toy, MiniMax H3, can run locally but wants serious NVIDIA hardware, so on the presenter's Mac it would struggle — this workflow goes to the cloud instead. A filter narrows the gallery to external API workflows, the crown-only list appears, and an image-to-video template opens. This time the corner error means something different: no model is missing, but no input image has been supplied, because cloud compute needs nothing downloaded. Using it requires a Comfy account — a mail login, an avatar in the corner — and a credits wallet topped up with, in the demo, ten dollars by card.

With the wallet full, the earlier tea painting is dragged in as the first frame and the shot is described in words: the woman brews gracefully while the camera pulls slowly back. Two special toggles appear — AI prompt enhancement and 2K upscaling — left at defaults, before the yellow output node sets resolution and length up to fifteen seconds. A live detail sells the economics: dragging the duration slider moves the credit cost in the corner, longer video burning more points. Five seconds is enough for the test, Run is pressed, and a short clip lands in the neighboring node.

Then the video names its biggest change of the last two years: ComfyUI speaks MCP, the open protocol through which assistants like ChatGPT and Claude can operate software on your machine. The payoff is stated plainly — even if you never learn what the nodes do, you can manage workflows by talking to an assistant. Setup takes minutes: an Install MCP button on the site, a local-install choice, a copied setup prompt, then in Claude you point a project folder at the ComfyUI install directory, click trust workspace, paste the prompt, hit enter, and restart Claude once the installer finishes.

What the assistant can do starts with the simplest job: running workflows you already built. With three workflows saved, a fresh chat asks Claude to generate a Chinese song in ComfyUI with a named style and theme — and before touching anything, Claude interviews you: male or female vocalist, preview the lyrics first? You answer in plain language, Claude calls ComfyUI, the computation happens offstage, and the finished song waits in the Assets panel for playback. The demo plays it, and the loop feels complete: you never opened a node.

The finale goes one level deeper: the assistant rewrites the workflows themselves. A Z-Image text-to-image flow is ordered to swap its model for GPT Image 2 and to emit three styles per run — photoreal photo, vector illustration, 3D animation — and the test subject is a police penguin, which duly appears in three versions inside the chat window while the rebuilt canvas waits back in ComfyUI. The same flow then takes a new theme typed directly — a uniformed Shiba chef tossing a giant pan of fried rice — and Run produces three dog portraits without further help. Finally the assistant turns story editor: a rough plot about the chef dog becomes a shot list with distinct framing and camera moves per shot, and moments later a small animation is ready to watch.