Conditioning Scale (CCN)

CFG controls how hard the sampler steers toward your prompt. But there's a second, less famous volume knob: the magnitude of the conditioning tensor itself. Conditioning Scale (CCN) is exactly that - a node that multiplies your conditioning by a constant. Higher scale, stronger prompt influence; lower scale, weaker. One input, one output, one slider. It's the conditioning equivalent of turning a dial, and it's as close to foolproof as this pack gets.

How it works

It clones each conditioning entry's tensor and multiplies it by scale (default 1.0, range 0–10, step 0.01). The pooled output rides along untouched. scale = 1.0 is a true pass-through - the node returns your conditioning unchanged rather than rebuilding it, so at the default it costs you nothing. It even prints a line to the console when it actually scales, which makes it obvious when it's live.

Unlike the pack's Conditioning Normalizer, this is pure amplification or attenuation - no reshaping of the distribution, no math on the mean or std. That's the appeal: it's the blunt, predictable tool. If your prompt is under-driving the image (everything comes out generic and floaty), nudge the scale up a bit; if it's overpowering the scene (a single dominant concept crushing composition), back it off.

Inputs and outputs

conditioning in, conditioning out - inline-safe, slot it between your encoder and the sampler. scale is the only dial you'll touch. Because it's a straight multiply, it composes nicely: put it after a lerp or subtract to tune the result's overall strength without redoing the mix.

Install

Standard for the pack. ComfyUI Manager → search ComfyCollectorNodes → Install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/valkymaera/ComfyCollectorNodes

Restart ComfyUI; no pip step (no extra Python deps). It's under the (CCN) suffix.

The caveat that actually matters

Scaling the tensor is not the same as scaling CFG, and the difference will bite you if you expect it to substitute. CFG scales the difference between conditional and unconditional predictions at every step; this scales the whole embedding up front. So raising it can intensify a concept but also amplify whatever noise and cross-talk is in the embedding - it's a lever, not a precision tool. On distilled models running at CFG 1, remember the unconditional pass isn't even computed, so this is one of the few "guidance-ish" knobs that still does something there. Start with ±0.2 moves around 1.0 and A/B against plain; it's gentle enough to be safe and strong enough to matter.