hswq_AnythingEverywherePrompts

hswq_AnythingEverywherePrompts is the positive/negative pair in one node: connect your positive conditioning to positive (labelled +ve) and your negative to negative (-ve), and every sampler downstream gets the right one without a single wire running across the canvas.

The category says everywhere/deprecated and the description says "Deprecated - should automatically be replaced". Believe it, but don't panic - deprecated here means the node is an alias, not a broken node.

What "deprecated" actually means here

The Python class exists so old workflow JSON still loads. The moment the node touches a graph - loaded from a file or dropped from the menu - the frontend's type converter rewrites it into a plain hswq_AnythingEverywhere, keeps exactly two inputs, relabels them positive and negative, and hands the real work to the normal broadcast machinery. I read the conversion function to check, and that's literally what it does: sets keep_inputs = [0, 1], flips fixed_inputs off, relabels both slots.

So functionally, this node is an Anything Everywhere with two inputs. Which raises the obvious question: why does the two-input version exist at all?

Because CONDITIONING is a type where "send it everywhere" is wrong about half the time. A broadcast with no restrictions hits every unconnected conditioning input in the graph - including the negative one on the same sampler. The prompts node is the pack's answer to that: it declares which input is which, and the wiring rules then match by input name. The matching logic in use_everywhere_classes.js notices when a controller is broadcasting a type that's already being broadcast (duplicate types) and falls back to comparing the broadcasting input's name against the target input's name - exact match by default, with start/end/regex variants available in the properties dialog. That's how positive finds the positive input on your sampler's conditioner and stops there.

The pack also ships built-in i18n regexes for prompt-ish input names, so untouched setups work out of the box.

The inputs

Both are * (untyped ANY) and both are optional, so the node will happily sit there empty. There are no outputs - nothing to connect on the right-hand side. If you're staring at the node wondering where the conditioning comes out, that's the whole point: the wiring appears only for the duration of the queue.

Practically: wire +ve and -ve from your CLIP Text Encodes, then go look at your samplers. Their conditioning slots will be filled at run time, drawn as virtual links on the canvas, and the widgets they feed will grey out to tell you they're spoken for. If your graph is small enough to read already, this node buys you nothing - it earns its place in the 40-node ones.

Install

Same pack, same command:

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

Restart, then look under everywhere/deprecated. ComfyUI Manager's search finds it by the pack title (ComfyUI-NunchakuFluxLoraStacker), which is unhelpful if you were searching for "everywhere" - that's why you're here.

Heads up on the requirements: this family of nodes needs none of the pack's heavy dependencies (nunchaku, Florence-2, the TensorRT-RTX CCSR stack), but the pack imports all of its submodules at startup, so a failed pip install takes the entire pack - and therefore these nodes - down with it. On a machine without nunchaku the FLUX LoRA nodes are skipped with a log warning and the everything-else still works.

When it misbehaves

If both your positive and negative conditioning end up on the same input, you've probably also got a plain Anything Everywhere broadcasting conditioning somewhere - two controllers, one target, and the more specific one losing. Check with the extension's logging turned on; it prints a match line per input.

If the node loads and then nothing downstream gets conditioning at all, check whether the conversion happened: after the rewrite it should present as hswq_AnythingEverywhere with two inputs named positive and negative. If you're on a ComfyUI build so old that the extension's link conversion doesn't run at all - the pre-fork breakage presents as a queue error like Required input is missing: clip - that's a pack/version problem, not a wiring one.

And the standing caveat for this whole family: the links aren't in your workflow file. Send someone the JSON and they'll see the samplers with empty conditioning slots unless they have the pack too.