Interpolate Predictions

Interpolate Predictions is the pack's lerp: a linear blend between two prediction strategies, applied at every denoising step. The formula is the shortest in the README -

prediction_A * (1.0 - scale_B) + prediction_B * scale_B

• and honestly, that's the whole node. Simple, transparent, and occasionally exactly what you need.

What it's for

A hard Switch Predictions between two strategies can be jarring - at the boundary you get a visible quality jump where one strategy hands off to another. Interpolate gives you a smoother alternative: instead of "CFG for these steps, something else for those," you can run a 50/50 blend of both everywhere, or a 30/70 lean.

Three genuinely useful patterns:

• Toning a heavy strategy down. "I want the quality of Characteristic Guidance but it costs a fortune - what if I blend it 30% with plain CFG?" That's scale_B 0.3, and you've bought most of the effect for less of the cost.
• A/B testing without rebuilding. Set scale_B to 0, render, set to 1, render, and you've compared two strategies on a fixed seed. The node is a controlled variable.
• Smoothing an early/middle boundary inside an Early/Middle/Late graph, so the phase transition isn't a cliff.

It's not the same as a per-step ramp - the blend ratio is fixed for the whole run - so don't reach for it expecting schedule-dependent interpolation. It's a global mix.

Inputs and outputs

• prediction_A (PREDICTION) - the "0.0" end of the blend.
• prediction_B (PREDICTION) - the "1.0" end.
• scale_B (FLOAT, default 0.5, range 0–1, step 0.01) - how much of B to mix in.

Output: one prediction.

The implementation is a straight torch.lerp; at exactly 0 or 1 it short-circuits and returns just the relevant child prediction, so there's zero overhead when you've effectively disabled it.

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

• Not doing what a schedule switch does. If you wanted "strategy B only in the middle steps," you want Switch Predictions or Early/Middle/Late, not a global blend.
• Both children evaluate every step. Unlike a switch, a blend can't skip a branch - both predictions run every step, so an expensive child still costs you. That's the price of smoothness.
• ControlNet unsupported, as with the rest of the pack.