Using API nodes in ComfyUI — LivePair Image and Video walkthrough
This is the hands-on version — from an empty ComfyUI canvas to a working hosted-generation chain. It assumes ComfyUI and ComfyUI Manager are already running; the LivePair side takes about two minutes.
Install
In ComfyUI Manager, search for LivePair and hit install — the package is on the official ComfyUI registry as livepair. Manual install works too: clone gitlab.com/livepairai/comfyui-livepair into ComfyUI/custom_nodes and reload.
There are no dependencies to resolve — the nodes talk to the API over the Python standard library and borrow torch/PIL from ComfyUI itself. If ComfyUI runs, the nodes run.
The API key
Generations bill an lp_ API key — create one on the billing page. Set it as the LIVEPAIR_API_KEY environment variable for the ComfyUI process, or paste it into the node's api_key widget. Same key, same prepaid balance as the Studio — usage shows up in one ledger.
Building the graph
Drop a LivePair Image node on the canvas. It takes a prompt, a model choice from the dropdown, and the usual size controls — and outputs a standard IMAGE tensor plus the result URL. Wire the tensor into whatever comes next in your pipeline: a preview, an upscale, a composite.
The dropdown mirrors the live catalog: LivePairRefreshModels pulls new entries from the API, and the custom option with model_override accepts any current model id — new models work without waiting for a node update.
Image inputs route themselves
Both nodes take an optional image input. Attach one to LivePair Image and the call becomes an instruction-based edit; attach one to LivePair Video and it becomes the i2v source frame. No separate node type, no mode toggle — the wire decides.
Jobs, failures, retention
Video jobs can take a few minutes — the node polls the job endpoint server-side until done and returns the clip as a VIDEO output. A failed job raises inside ComfyUI like any node error; prepaid debits only happen for completed results on the API's normal accounting.
One operational note: hosted result URLs live for 48 hours, then auto-delete. The tensor lands in your session either way — save outputs you want to keep.