Conditioning (Multiply)

Conditioning Multiply scales your prompt's embedding by a number. It's the volume knob of the conditioning world - a single conditioning in, a single CONDITIONING out, and one float that decides how loudly that prompt shouts at the sampler. Set it to 1 and nothing changes. Crank it to 2 and the prompt's influence roughly doubles. Drop it to 0.5 and it goes quiet.

It sounds like a gimmick until you're staring at a generation where one of two combined prompts is drowning the other. That's the real use case: you've got a regional or multi-prompt setup, one subject is muscling out the other, and you need per-prompt gain without re-typing anything. Multiply is the mixer fader.

How it works

Under the hood it multiplies every embedding tensor in the conditioning by multiplier. It also scales the pooled_output - the compact summary used by class-conditioned models - by the same factor. So it's not just the cross-attention signal that changes; the pooled conditioning scales too, which means the effect is felt by the parts of the model that read the pooled vector.

The interface is two fields:

• conditioning - whatever you want to amplify or damp.
• multiplier - -100 to 100, default 1. Below 1 softens; above 1 amplifies; negative values invert the signal.

One output, CONDITIONING, which replaces the original in the chain.

What matters in practice

• It's a scaler, not a new prompt. Multiply doesn't change what the text means, only how hard it pushes. If a prompt is fighting another one, reducing its multiplier is often cleaner than rewriting it.
• Negative multipliers are weird but real. Inverting an embedding isn't the same as a negative prompt, but in combined setups it can cancel a competing prompt's influence. Think of it as phase cancellation rather than "do the opposite."
• Watch the metadata. The node preserves the conditioning's metadata - masks, areas, strengths - so you can scale a regional prompt without losing its spatial constraints. Useful: crank the background prompt down without touching its mask.

Gotchas

It doesn't always behave linearly. Embeddings aren't a simple "twice the number, twice the adherence" - the cross-attention math is more complex than that, and at high multipliers you can get artifacts or outright garbage. Treat 0.5–1.5 as the sane working range and test before you go wild.

You can't "undo" a model's training with it. Multiplying to extreme values won't fix a prompt the checkpoint simply doesn't understand; it just amplifies the misunderstanding. If a concept is absent from the model's vocabulary, volume won't conjure it.

On guidance-distilled models, remember the negative side is inert - multiplying a negative that does nothing still does nothing. This node shines on the positive side and in regional combos.

It ships with ComfyUI core, search aliases "scale conditioning" and "scale prompt." It's one of those quiet nodes that repays knowing about - five minutes to understand, and it shows up in every serious multi-prompt graph.