ScaledCFGGuider

ScaledCFGGuider lets you add a second positive conditioning into the mix the way model merging adds trained differences - scaled, not averaged. It's one of the quieter nodes in the ComfyUI Extra Samplers pack by Clybius, and it's for when you want prompt B's influence layered on top of prompt A's, rather than blended 50/50.

The mechanism, straight from the source: each step it runs the model on both positives (cond1, cond2) and the negative, then computes the difference between the two positive predictions and adds it to cond1 scaled by cond2_alpha. That is the "Add Trained Difference" trick from model merging, transplanted into conditioning space: instead of cond1 + cond2 (which blows up or muddies), you get cond1 + cond2_alpha * (cond2 - cond1). At cond2_alpha = 0 it's just cond1; at 1.0 you've added the full difference, effectively standing on cond1 and pushing toward cond2. The README says it outright - this is the "Add Trained Difference" method, applied to conditionings.

The practical shape of that: prompt A ("a red car, studio photo") is your anchor, prompt B ("neon cyberpunk lighting, rain") is the flavor you want layered in. cond2_alpha controls how much of B's character you inject without letting B take over composition. You can push it negative or past 1.0 for exaggerated or inverted influence, though both are experiments more than production settings.

The inputs that matter:

• cond1 / cond2 - the two positive conditionings. cond1 is the base, cond2 is the difference source.
• cond2_alpha - how much of the difference to add. Default 1.0; 0.2–0.6 is the sensible starting range for "a hint of B."
• negative - the negative conditioning for the CFG pass.
• cfg - the CFG scale applied on top. Default 8.

Output: a GUIDER for a custom sampler node's guider input.

Where people get burned: cranking cond2_alpha to 2 or 3 "for more effect" and getting artifacts that look like a broken sampler. Difference-additions scale fast; small values are where the useful work happens. And as with its geometric sibling, this rewards a fixed-seed comparison: run cond2_alpha 0, 0.3, 0.6 and pick, rather than guessing.

Install is the shared pack routine - ComfyUI Manager → "ComfyUI Extra Samplers", or:

cd ComfyUI/custom_nodes
git clone https://github.com/Clybius/ComfyUI-Extra-Samplers

then restart ComfyUI. No model downloads; dependency kornia. Want the geometric (multiplicative) flavor of the same idea instead? That's GeometricCFGGuider in the same pack.