Krea 2 Conditioning Rebalance

A small node that does something surprisingly structural. Krea 2 doesn't condition on the last layer of its text encoder like older models - it conditions on a 12-layer Qwen3-VL stack that its DiT mixes linearly through a projection. Shallow taps carry broad syntax and composition; deep taps carry fine detail - identity, texture, precise attributes. Krea2Rebalance lets you reweight those 12 taps, then insert the result between any conditioning node and the sampler.

Why you'd bother

Because "the model ignores my fine details" is a very Krea 2 problem. When the deep taps are underweighted relative to the shallow ones, you get the composition right and the identity or texture wrong. This node is the lever for that - you're not scaling the prompt, you're changing which layers of the text encoder actually drive the image. The pack's own examples run their positive through this node with the balanced preset before sampling.

The inputs

• conditioning - any Krea 2 conditioning (moodboard, identity edit, or plain encode). In, weighted, out.
• preset - balanced (the classic community profile: deep taps at 2.5 / 5.0 / 1.1 / 4.0), detail (stronger fine-detail adherence), subtle, uniform (no per-layer change), or custom.
• per_layer_weights - comma-separated gains, one per tap, shallow → deep. Only used when preset is custom. Must be exactly 12 for Krea 2; the node errors if you hand it 11, rather than silently doing something meaningless.
• multiplier - a global gain applied after per-layer weighting. Leave it at 1.0 when renormalize is on; cranking it >1 amplifies the whole tensor and mostly just oversaturates.
• renormalize - the default ON, and the thing that makes this node worth having. It holds the conditioning's overall RMS so only the tap ratios change. That's the quality-preserving mode; global amplification degrades likeness and color.

The honest history

You should know where this mechanic came from. The original community "Conditioning Krea2 Rebalance" node appeared on Krea 2's release day as a safety-filter workaround - the open weights shipped with an alignment pass, and rebalancing the conditioning was one of the first ways people found to get around its refusals. It worked, but the community quickly found it degraded quality and prompt adherence. That's exactly why this reimplementation defaults renormalize ON: it's the RMS-renormalized variant (credited to huwhitememes' comfyui-krea2-conditioning), which changes only the tap ratios instead of blasting the whole tensor. If you want the original node's exact behavior, set renormalize OFF and multiplier to 4.0 - it's a documented drop-in for workflows built around that node.

And a fair warning: if what you actually want is the filter gone without the quality tax, the community settled on an uncensor LoRA as the better tool. Rebalance is a per-layer dial, not a magic unfilter button. Treated as a dial, it's genuinely useful - the balanced profile is a real improvement for fine detail on default prompts.

Installing

Whole pack, one install - no third-party deps, GPL-3.0:

git clone https://github.com/RedNodeAI/ComfyUI-Krea2Moodboard ComfyUI/custom_nodes/ComfyUI-Krea2Moodboard

or search "Krea2Moodboard" in ComfyUI Manager and restart. This node doesn't even need the edit LoRA - it just needs ComfyUI's native Krea 2 support (0.25.0+), the qwen3vl_4b text encoder, and the VAE. If you're on a stock Krea 2 workflow and fine detail feels flat, insert it and leave balanced on. It's the cheapest quality lever in the pack.