ComfyUI without a GPU — run hosted models inside the workflow UI
The reason most people want 'ComfyUI online' or 'ComfyUI cloud' isn't the node editor — it's the GPU underneath it. The editor runs on almost anything; the models are what need the hardware. LivePair's custom nodes split those two apart: ComfyUI stays the editor, generation happens on the hosted API.
What ComfyUI actually is
ComfyUI is a node-based editor for diffusion pipelines. The graph itself is lightweight — nodes, wires, queues — and runs happily on a laptop. What eats the hardware is the checkpoint: an SDXL or Wan model loaded into VRAM doing the sampling.
When people search for a ComfyUI cloud, they're really asking for the model side to live somewhere else. API nodes are the built-in way to do that — a node that sends the request to a hosted endpoint and returns the result as a normal tensor.
The LivePair nodes
Two nodes cover the whole thing: LivePair Image (text-to-image, instruction edit, upscale, background removal) and LivePair Video (text-to-video and image-to-video on the Wan and Seedance lines). Both return standard ComfyUI outputs — IMAGE/VIDEO tensors plus the result URL — so they slot into existing graphs like any sampler.
Install is the ComfyUI Manager route: search LivePair in the registry and click install. The package is stdlib-only — no pip dependencies, no version juggling. Details live on the ComfyUI integration page.
What it costs
There's no GPU instance billing by the hour. Generations debit prepaid credits per result — the same ledger as the Studio and the REST API, with the per-model price shown on the models page before anything runs. A $10 top-up covers real experimentation, not just a demo.
Where the hybrid setup wins
Keep the editor and post-processing local — masks, composites, upscalers you already like — and let API nodes do the model inference. You get the workflow control ComfyUI is good at without managing driver updates, VRAM ceilings, or a cloud GPU lease that bills while you edit.
If you do have the GPU already, the nodes still pull weight for the models you don't want to download: trying a new Wan or Seedance line is a dropdown change instead of a 20GB checkpoint hunt.