RemRun Input Graph(s) Output

Read the article on RemRun Input Graph(s) first, because this node is literally the same class with one flag flipped. Same inputs, same outputs, same remote-execution machinery. The only difference: OUTPUT_NODE is set to True.

What that changes, in ComfyUI terms, is that this node counts as a terminal for execution. When the graph runs, execution stops and waits here instead of just passing through. For this pack that's a meaningful distinction: it marks the remote run as the point of the workflow - "run everything upstream of me on the other machine, and this is where that work lands."

When you'd use it over the plain Input node

Reach for this when the remote run is the end of the branch and there's nothing interesting left to do locally. The README's offload pattern is a good example: if the whole point is to decode the VAE on the big machine and just preview the result, the Input Graph(s) Output node is the natural sink. It makes ComfyUI treat the remote execution as the "output node" of the graph, which matters for how prompts are constructed and when the queue considers the job done.

The plain Input node is the right call when you're splitting a workflow in the middle - run the sampler locally, then ship the decode upstream of the node and continue with upscaling or previews locally afterward. The Output variant is for when the remote does the heavy lifting and the local side is mostly the front door.

The inputs

Identical to RemRun Input Graph(s): remote_url, total_timeout, request_timeout, serialization, then the tuning set - max_size_mb, gzip_compression, gzip_level, lazily_transfer, skip_lazy_transfer_under_mb, response, ws_compression, forward_progress_messages, run_outputs_connected_to_inputs, inputs_when_local, is_changed, run_sync. Ten wildcard inputs (input_0 through input_9), ten wildcard outputs (out_0 through out_9).

The only input you'll actually set on a first run is remote_url. Everything else inherits the same sensible defaults as the Input node: fancy_safetensors serialization, 90-second total timeout, binary response.

Gotchas worth knowing

Because it's an output node, a remote run that hangs will hold the whole queue for as long as total_timeout allows - a hung VAE decode on the far side looks like a frozen ComfyUI to the local machine. Raise the timeout for big video jobs, and remember both instances need the pack installed, because the subgraph this node ships over contains the pack's internal Serializer/Deserializer nodes. No models to download, install is the standard Manager or git clone:

cd ComfyUI/custom_nodes
git clone https://github.com/LatentRat/comfy_remote_run

Restart, and it shows up as "RemRun Input Graph(s) Output" in the Remote Run menu.