hswq_AnythingEverywhere

You plug one wire into it. Nothing comes out. And somehow every sampler downstream now has a model, a clip and a VAE.

That's hswq_AnythingEverywhere, a broadcast node - you connect a source to its anything input and it silently feeds that same output into every unconnected input of the same type anywhere in the graph. No wires. This is the "spaghetti cutter" the plumbing layer of ComfyUI has been arguing about for two years, and it's genuinely the fastest way to keep a 60-node workflow readable.

One naming note before anything else. This lives in a pack called ComfyUI-NunchakuFluxLoraStacker, which has nothing to do with what the node does - that pack is a thirteen-node grab bag, and the Use Everywhere family in it is a port of chrisgoringe/cg-use-everywhere (Apache-2.0). The README's own account of why: the upstream repo stopped working on ComfyUI frontend 1.52.7 and later (the frontend dropped node.comfyClass, changed graphToPrompt's receiver binding, and started leaving dangling links behind), so the fork got merged here. The hswq_ prefix exists so its class_types don't collide with an upstream install.

How it actually works

This is the part that matters, because it explains every weird thing the node does.

The Python class is a stub. Its execute() returns an empty NodeOutput - it does nothing on the server. The work happens in the browser: the extension wraps app.graphToPrompt, and when you queue, it analyses the graph, works out which unconnected inputs your broadcast matches, and creates real links for the duration of the prompt. Then it tears them back down so your canvas looks the same afterwards. The pack's use_everywhere_apply.js even tags each temporary link with the controller node id so the cleanup can find them again.

Two consequences you'll trip over:

Your saved workflow has no wires in it. Someone who opens your graph without the pack gets a workflow with silently missing connections, not an error. (Exporting API-format is handled - the export command is wrapped too, so the JSON you get does contain the links.)

And your graph stops being self-explanatory. There's a whole thread titled "Please Stop using the Anything Anywhere extension" that argued exactly this - convenient, and impossible to debug later. It's a fair complaint. Reach for this when wire count is your actual problem, not reflexively.

Inputs and outputs: there's only one input

anything is the entire input surface - typed *, optional, so you can leave it empty and the node just sits there doing nothing. There are no outputs at all. If you're looking for something to wire into your sampler, you're looking at the wrong side of the node.

The controls that matter aren't sockets, they're properties: right-click the node and you get the Use Everywhere properties editor, where you can restrict a broadcast by node title (title regex), input name (input regex), group name, group colour, or node colour. A bare Anything Everywhere with no restrictions broadcasts to everything of that type, which is usually too much the moment you have two CLIP encoders in one graph.

When two controllers match the same input at the same priority, you get an ambiguity warning and nothing is sent - it won't guess. Priority is computed rather than set by hand: base 10, plus 20 if the node carries regex restrictions or is one of several broadcasting the same type. In practice the more specific node wins.

Wire two different types in and you get two independent broadcasts from one node - that's how people feed model and clip from a single loader.

Installing it

Manager → search the pack title, or:

cd ComfyUI/custom_nodes
git clone https://github.com/ussoewwin/ComfyUI-NunchakuFluxLoraStacker.git
cd ComfyUI-NunchakuFluxLoraStacker
pip install -r requirements.txt

Restart ComfyUI, then look under the everywhere category.

Here's the irritating bit: none of these nodes need any of that. They're a few frontend JS files plus no-op Python stubs - no GPU, no nunchaku, no models. But the pack's requirements.txt also pulls nunchaku>=1.0.0, the Florence-2 stack, and the full TensorRT-RTX CCSR stack (tensorrt-rtx, onnx, polygraphy, taming-transformers, …), and the file's own comment says the CCSR deps are required at pack import time - miss them and the whole pack fails to load, which means you get zero nodes, including these. So install the requirements even for the free stuff, and expect a chunky pip run. On AMD/ROCm, or anywhere nunchaku won't import, the FLUX LoRA nodes are gated off with a startup warning and the rest still loads.

If you already have chrisgoringe's original installed, pick one. Two packs both broadcasting the same type is the fastest route to an ambiguity warning you can't explain.

When it seems broken

The classic symptom of the pre-fork breakage is a queue failure reading Required input is missing: clip - the UE links were silently skipped so the sampler looked unconnected to the server. If you see that on a current frontend, you're on an old UE build, not a bad workflow.

Otherwise, check in this order: is the receiving input actually unconnected (UE won't override a real wire)? Does another controller already match it? Turn on the extension's logging setting if you want it to tell you, input by input, why each one was or wasn't matched.