Load workflow (image upload)

If you've ever dragged a random ComfyUI output PNG onto your canvas and watched the whole graph that made it snap into place, you already understand this node - it's that same trick, just wired as a node instead of a drag-and-drop. ComfyUI embeds the full workflow graph into any image it saves, as metadata sitting quietly inside the PNG. Load Workflow JSON reads that embedded graph back out and hands it to you as a plain Workflow JSON output you can feed into something else.

Why it exists

This node belongs to the "advanced" side of NetDist_Plus, a fork of city96's ComfyUI_NetDist built by nux1111 for splitting a render across two GPUs or two networked machines. The simple version of that (see RemoteQueueSimple-Nux) just re-sends whatever's currently loaded in your local graph. The advanced version lets you queue an entire saved workflow on a remote worker instead - useful once you've got more than two GPUs and want to fire a known-good pipeline at a farm of remote instances without rebuilding it fresh every session. To do that, something has to load the saved workflow back into graph form first. That's this node.

Its one real job downstream: feed its Workflow JSON output into RemoteChainStart-Nux's workflow input, and that node dispatches the whole thing to a remote instance.

Inputs and outputs

There's exactly one input, image - a standard upload picker, same as the core Load Image node. Pick any PNG that ComfyUI itself produced (or one you explicitly saved with its embedded workflow intact). The single output is Workflow JSON, a JSON type carrying the reconstructed graph.

That's the whole node. No dials, no options - it's a metadata reader, not an editor.

Installing it

Search ComfyUI_NetDist_Plus in ComfyUI Manager and install, or do it by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/nux1111/ComfyUI_NetDist_Plus

Restart ComfyUI. The pack's only external Python dependency is the requests library - if it's not already on your system, pip install requests before you launch. There are no model downloads for this node or the pack in general; it's pure orchestration, no weights involved.

Where people get burned

The most common failure is feeding it a PNG that never actually carried a workflow. Any image editor, Discord, or messaging app that re-compresses or re-saves a PNG will strip the metadata clean off - so if you downloaded an image from somewhere other than directly off your own ComfyUI output folder, there's a decent chance there's nothing left to extract. Grab the original file, not a re-shared copy.

Second: loading the graph is only half the battle. The remote instance you're about to dispatch it to still needs every custom node that graph uses actually installed - this is the same "why does no workflow I download work" problem that plagues the whole ComfyUI ecosystem, just now happening on a machine you can't directly see the console of. If a dispatched job silently fails on the remote side, check that instance's own startup log for missing-node errors before assuming this node did something wrong.

Third, a scope thing worth knowing going in: this node only loads. The pack's author is upfront in the README that editing a loaded workflow - search-and-replace on a seed, swapping a checkpoint name - isn't built yet ("I have nodes to save/load the workflows, but ideally there would be some nodes to also edit them... PRs welcome"). If you need to change a value inside the graph you just loaded, that's what RemoteApplyValues-Nux and RemoteApplyValuesMulti-Nux are for - they patch individual node parameters on top of whatever this node hands you, rather than editing the JSON itself.

One more honest note on the pack as a whole: it's a niche, lightly-maintained fork built by a solo hobbyist for a very specific multi-GPU setup, not a widely-adopted production tool. It works for what it was built for, but don't expect much beyond that.