Rebalance Guider

Rebalance Guider takes a positive and a negative conditioning and mixes the negative into the positive at a chosen strength, handing back a single conditioning that already carries the contrast. The interesting part is what it's for: getting some negative-prompt effect on a model that runs at CFG 1, where the sampler normally never computes a negative pass at all.

Here's the background that makes this click. Krea 2 Turbo - and basically every guidance-distilled model of 2026 - runs at CFG 1, which means guidance is baked into the weights and ComfyUI skips the second, negative-prompt forward pass entirely. So your negative prompt does nothing by default. That's a known pain point, and the community's answers range from stating constraints positively to attention-level tricks like NAG. Rebalance Guider is this pack's take: pre-mix the negative into the positive conditioning so the sampler carries the push-away in the one pass it does run.

Inputs and outputs

Three required inputs:

• positive - your positive conditioning.
• negative - the conditioning you want to push away from.
• guidance_strength - default 0.5, range 0 to 2. How hard the negative pushes. 0 leaves the positive untouched; raise it to make the negative bite harder. Past 1 it gets aggressive fast.

Output is a single conditioning. And here's the thing to get right: despite the "Guider" in the name, the output type is CONDITIONING, not a GUIDER - so wire it into your sampler's positive input, not into a SamplerCustomAdvanced guider socket. That mismatch is the single most likely way to misuse this node.

Installing it

ComfyUI Manager: search ComfyUI-Conditioning-Rebalance, install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/nova452/ComfyUI-ConditioningKrea2Rebalance.git

Restart ComfyUI. No pip deps, no model files.

Common issues

The name genuinely misleads - read the last paragraph of inputs again. It outputs conditioning; it plugs into the positive slot. If your graph errors on a type mismatch at the sampler, this is why.

guidance_strength is touchy at the top of its range. Start at the 0.5 default and move up slowly on a fixed seed - high values can over-correct and leave the image looking washed or off-color, the same way an overcooked CFG does on a normal model.

Keep the positive and negative conditionings in the same model's space, as always.

And be clear-eyed about what this is: a lightweight approximation of negative guidance for CFG-1 models, not a full CFG restoration. If you need real negative-prompt power on a guidance-distilled model, the KB's own advice still holds - state your constraints positively where you can ("clean background, sharp focus" beats a negative), and consider a dedicated tool like NAG for the heavier cases. This node is the quick, in-pack option when you just want to nudge the model away from something without leaving your Krea 2 / Ideogram 4 conditioning graph.