Queue on remote (single)

If you've got two GPUs, or a second PC gathering dust, and you've wondered whether ComfyUI can actually use both at once - this is the node that does it. RemoteQueueSimple ("Queue on remote (single)") takes your workflow, mirrors it over to a second, already-running ComfyUI instance, and queues it there while your local graph keeps working its own copy. Pair it with FetchRemote and you get both results back into one graph.

It's from ComfyUI_NetDist, a pack by city96 - better known these days for ComfyUI-GGUF, the quantized-model loader half the community runs. NetDist is older and does something unrelated to model quantization: it treats a second ComfyUI process as a worker you talk to over HTTP, not a magic tensor-parallel split. Someone who actually paired a 4060 Ti with a spare 3060 using it reported a real 15-20% speedup over running the 4060 Ti alone - a genuine win, if a fiddly one to set up.

How it works

Under the hood this is a plain HTTP POST of your current queue to remote_url, using the requests library. That second instance has to already be running - you start it yourself, typically with its own --port and --cuda-device flag to land it on a different card, or --listen plus the machine's real IP if it's a separate PC. This node doesn't spin that instance up for you.

You can also chain several RemoteQueueSimple nodes together for more than two machines - the README calls this pattern "the simple multi-machine" setup - but it's honest that it's "a bit jank," and warns you can get duplicate images if you reuse the same seed across machines. Randomize the seed instead.

The inputs and outputs that matter

• remote_url (default http://127.0.0.1:8288/) - the address and port of the second instance. Change the port for a second local GPU; change the whole host for a second PC.
• batch_local / batch_remote (1-8 each) - how many images stay on this machine versus how many get shipped to the other one.
• trigger (on_change / always) - the README flags this directly: set it to always "if it isn't getting executed - i.e. you changed a sampler setting but not the seed." ComfyUI's caching can otherwise skip re-queuing the remote job when it thinks nothing relevant changed.
• enabled (true / false / remote) - a kill switch for the node.
• seed - the seed NetDist uses to keep multi-machine output from duplicating.

Outputs: seed and batch (the actual values used, useful for logging), and remote_info - the token FetchRemote needs to pull the finished image back.

How to install it

Via ComfyUI Manager: search ComfyUI_NetDist. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/city96/ComfyUI_NetDist
pip install requests

Restart. That single dependency - requests - is the whole footprint; no models to fetch.

Common issues & troubleshooting

The remote node just doesn't run when you tweak something. That's the trigger field. Flip it to always if a change that isn't the seed (a different sampler, a different step count) isn't reaching the remote instance.

Duplicate images across machines in a chained setup. Expected per the README when the seed isn't randomized - set it to random rather than fixed when you're running several RemoteQueueSimple nodes in sequence.

Nothing happens and there's no error. Confirm the second instance is actually up first - this node has no way to launch it, and a wrong remote_url or an instance that hasn't started yet just looks like silence.

Linux controlling a Windows remote (or the reverse). The README's own roadmap still lists this as unresolved: an os.sep mismatch between platforms. If you're bridging OSes, don't be surprised by path-related weirdness - it's a known, unfixed edge case, not something you're doing wrong.