Scale Prediction

Scale Prediction is the simplest node in the pack: prediction * scale. One input, one float, one multiply per step. If you've used the pack's Combine Predictions, this is its boring cousin - the node you grab when a formula needs a weight applied to a single prediction rather than a whole arithmetic expression.

What it's for

Guidance formulas are full of lone coefficients. "CFG but at 9," "30% of this style vector," "half the negative's pull" - all of those are a single prediction times a constant. Scale Prediction is the dedicated knob for them. It's the Scale in the README's claim that Conditioned Prediction + Combine + Scale can express every other predictor: subtraction gives you a difference, Scale gives it a weight, and addition puts it back.

Two concrete uses:

• Weighting a guidance vector before recombining with Scaled Guidance Prediction. Build positive - negative with Combine, run it through Scale, and you've hand-rolled CFG at your chosen strength.
• Amplifying or damping a strategy before it enters an Interpolate or a Switch, when you don't want to touch the inner predictions.

The scale range is −100 to 100, default 1, step 0.01. Negative scales flip the direction - useful when you want to steer away from a difference instead of toward it. Zero zeroes it out entirely, which is a legit way to temporarily silence a branch of your graph without unplugging it.

Inputs and outputs

• prediction (PREDICTION) - any prediction from this pack's graph.
• scale (FLOAT, default 1, range −100–100).

Output: one prediction.

How it works

Nothing clever, and that's the appeal. Each step it asks the inner prediction for its output and multiplies the tensor by scale. It doesn't run the model itself, doesn't change which conditionings are evaluated, and doesn't cache anything of its own. The cost of a Scale node is whatever its child costs - it's free overhead. The only way it can surprise you is if you forget the inner prediction is still being fully evaluated on its own.

Installing it

Ships in the ComfyUI-Prediction pack by @RedHotTensors (Project RedRocket). ComfyUI Manager → search "ComfyUI-Prediction", or:

cd ComfyUI/custom_nodes
git clone https://github.com/redhottensors/ComfyUI-Prediction

Restart ComfyUI. No extra dependencies or model files. It's under Add Node > sampling > prediction.

Common issues

• "Why is my branch still slow?" Scale doesn't skip the child - it multiplies it. If the child is expensive, you still pay for it.
• Negative scale surprise. It works, but the result can be unstable in downstream arithmetic; test on a fixed seed before trusting it.
• ControlNet unsupported, as everywhere in this pack.