Krea 2 Conditioning Rebalance

Here's the honest version: the Krea 2 Conditioning Rebalance is the most experimental node in the Sigmax pack, and even its author is careful not to oversell it. It's a CONDITIONING transform that rebalances Krea 2's text conditioning - "conditioning tap rebalance with fixed RMS preservation," if you want the full phrase - and it makes no prompt-adherence claims and patches nothing. You use it when you've got a Krea 2 RAW or Turbo workflow whose conditioning feel you want to nudge, and you want a scalpel rather than a sledgehammer.

The mechanism matters more than the marketing (there is no marketing; it's called Experimental in the node itself). Krea 2 uses a Qwen3-VL text encoder with multilayer feature aggregation, and its conditioning path carries a lot of signal. This node rebalances the primary conditioning tensor before it fuses into the model, preserving overall RMS - meaning it changes the balance of the conditioning signal without blowing up its scale. At strength 0 it's an identity pass: in, out, nothing changed. That's your safety rail.

Inputs

• conditioning - the CONDITIONING from your CLIP text encoder, wired straight in.
• variant - RAW or Turbo. Pick the one matching the checkpoint you're running. This isn't cosmetic; the rebalance is tuned per variant.
• profile - Disabled, Subtle Experimental, or Classic Experimental. Disabled is the off switch, Subtle is the sane starting point, Classic is the stronger tweak. The pack says nothing about which is "better" because it deliberately doesn't claim either is.
• strength - 0.0 to 1.0, default 0.5. Zero is identity; start low and work up. This is the knob you actually tune.

Outputs: the transformed conditioning (wire it where the original conditioning went) and modifier_info, a string describing what the node did.

What it is not

It is not a prompt-adherence booster, and the node says so in its own description. It does not patch the model, it does not touch the scheduler, and it does not rescue a workflow whose real problem is CFG or steps. If you're coming from the "add a conditioning zero-out / negative prompt tweak" school of Krea 2 debugging - and the Krea ecosystem is full of those - think of this as a similar lever with the volume control made explicit.

Installing

It's part of the one-pack install, same as everything else here:

cd ComfyUI/custom_nodes
git clone https://github.com/rookiestar28/ComfyUI-Sigmax comfyui-sigmax

Restart ComfyUI, find it under Sigmax. ComfyUI Manager has it as "ComfyUI-Sigmax". Python 3.10+, ComfyUI 0.29.0+, no extra dependencies.

Where people get burned

The trap is treating an experimental tuning node as a fix. If your Krea 2 output is mushy or off-prompt, rebalance is not the first dial - the variant/scheduler/step choices are, and Sigmax's own Krea2SigmaScheduler covers those. Reach for this only once the schedule is right and you're chasing a specific conditioning feel. And when you do: run it at strength 0 first to confirm the identity path works, then move up in 0.1-ish steps. modifier_info will tell you what profile actually applied.