Powder Clear Conditioning Cache

Powder Clear Conditioning Cache is the escape hatch for the one failure mode that caching always introduces: stale results. If you've been batch-testing LoRAs and styles through Powder Conditioner and suddenly the output isn't changing when you know the inputs did, this is the node you're looking for.

It does one thing. It wipes the conditioner's internal encoding cache. That's the whole job, and honestly, that's why you want it - it's a utility node, not a feature.

How it works

Powder Conditioner stores the encoded conditioning for every (CLIP model + prompt) pair it has seen, so repeated runs with the same prompt are instant instead of re-encoding. That cache is keyed on the CLIP object - the pack detects when a different LoRA produces a different CLIP and re-encodes automatically. But the detection has limits. If you switch to a completely different base model, or you've been testing with disable_clip toggled, the cache can hold onto encodings that no longer match what you're sampling, and you get images that quietly ignore your new setup.

ClearConditioningCache clears both the conditioning cache and the CLIP dimension cache. It's an output node - nothing comes out of it. Its one input is the trigger that makes it run.

The clever part is how it fires. The node only executes when the trigger value changes. ComfyUI otherwise re-runs every node in a queued graph, which would mean the cache gets nuked on every queue if you left the node sitting in your workflow. Instead, IS_CHANGED returns the trigger value, so the clear only happens when you bump it. To bump it: edit the number, or click the Clear now button on the node, which auto-increments for you.

What you actually set

• trigger (INT, default 0) - change this value to fire the clear. That's it. Clicking "Clear now" does the increment for you, so you may never touch the field directly.

When to use it

The pack's own usage guide calls out three cases: after loading a completely different base model, if you suspect stale results, and to free GPU/RAM when the conditioning cache is bloated. The practical advice: don't leave it in the graph permanently. Drop it in, run it once after a big change, then remove it. That's what the trigger mechanism is designed around - it's meant for occasional manual use, not to sit in your pipeline.

Installation

This pack installs once and provides all eight Powder nodes:

cd ComfyUI/custom_nodes
git clone https://github.com/E2GO/e2go-comfyui-nodes.git e2go_nodes

Restart ComfyUI. Or search e2go_nodes / E2GO in ComfyUI Manager's "Install Custom Nodes" and install from there. There are no external Python dependencies to worry about - the pack is pure Python against ComfyUI's own runtime, so installation is genuinely painless. The README targets ComfyUI 0.17+.

Common issues

If you clear the cache and results are still "stuck," check two things. First, that you're not actually getting cache hits from somewhere else - this node only touches this pack's caches. Second, if you're on FLUX or a quantized CLIP, remember that the conditioner's auto cache mode already disables caching for those models, so there may be nothing to clear - your stale result is coming from somewhere else entirely. And if you ever wonder whether the clear actually happened, the node logs the number of entries it removed to the console.

It's a small node, and it doesn't pretend to be more. But when you're ten LoRA comparisons deep and the results stop changing, it's the difference between five seconds of debugging and a confused half hour.