Conditioning Krea2 Rebalance

This is the node the whole pack is named after, and it does something narrow and specific: it takes the conditioning coming out of Krea 2's text encoder and re-weights it layer by layer before it hits the sampler. That's it. No API, no key, no extra model download despite the scary name - it's pure conditioning math on a tensor you already have.

Why you'd bother: Krea 2 reads your prompt (and, through the pack's encode nodes, your reference images) through the Qwen3-VL encoder, which does multilayer feature aggregation - it pulls features from several layers instead of just grabbing the last one. Different layers carry different stuff: some hold coarse structure, some hold fine detail and the reference-image signal. Conditioning Krea2 Rebalance lets you push some of those layers up and others down, which is how the pack fakes IP-Adapter-style control on a model that shipped with no native editor. Think of it as a graphic EQ for your conditioning.

How it works

You hand it a conditioning, and it multiplies each layer's contribution by a weight, then scales the whole thing. The two knobs are:

• per_layer_weights - a comma-separated list of 12 floats, one per layer. The default is 1.0,1.0,1.0,1.0,1.0,1.0,1.0,2.5,5.0,1.1,4.0,1.0. Notice the shape: most layers sit at 1.0 (untouched), but layers 8, 9, and 11 get boosted hard (2.5, 5.0, 4.0). Those are the layers the author found carry the punch. Leave the 1.0s alone and nudge the boosted ones if you want more or less of whatever effect you're chasing.
• multiplier - default 4. A global gain applied on top of the per-layer weights. Turn it up for a stronger, more aggressive rebalance; down toward 1 for a gentle touch.

The output is a single conditioning that wires straight into your sampler's positive input, exactly where your normal encoded prompt would go. Nothing else in your graph changes.

The honest read: the per-layer numbers are the author's, tuned by feel. You don't need to understand which layer means what to use this - start with the defaults, and if the effect is too strong, drop the multiplier or the big layer weights; too weak, raise them. It's a dial, not a theorem.

Installing it

Through ComfyUI Manager: open Manager, search for ComfyUI-Conditioning-Rebalance, install, restart. Or clone it:

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

Then restart ComfyUI. There are no pip dependencies and no bundled model files - the nodes are tiny. The real requirement is the thing they operate on: a working Krea 2 graph, which means the model plus its Qwen3-VL text encoder (~8GB) and the Qwen-Image VAE (~250MB). If you don't already have Krea 2 running, sort that first.

Common issues

The README is thin - screenshots, a couple of dated changelog notes, and linked workflow.json files, with zero parameter docs. Grab one of those example workflows from the repo and reverse-engineer it; that's faster than guessing.

This node is Krea 2 specific. It expects Krea 2's conditioning space. Feed it SDXL or Flux conditioning and you'll get either an error or garbage - there's an Ideogram 4 sibling (ConditioningIdeogram4Rebalance) for that model, with its own per-layer defaults.

One Krea-wide gotcha worth repeating: Krea 2 Turbo runs at CFG 1, and even then you want a negative-prompt slot wired in (an empty one is fine). The community found ConditioningZeroOut doesn't play nice with Krea's recommended settings, so don't reach for it here.

And keep expectations calibrated. This whole pack is a stopgap - Krea has said an official edit model is "in the works" but hasn't shipped one, so per-layer rebalancing is the community's workaround, not a polished feature. It's fiddly by nature. Save the seed, change one weight at a time, and compare.