Quality-preserving per-layer conditioning control for Krea 2 — rebalance the 12 Qwen3-VL taps without the artifacts and likeness drift the ×4 default introduces.

A single, fast ComfyUI node that gives you direct control over the multi-layer text conditioning Krea 2 feeds to its denoiser — and does it without the quality collapse the original "rebalance" approach can cause when pushed.

Forked from — and crediting — nova452/ComfyUI-ConditioningKrea2Rebalance (Apache-2.0), which introduced the per-layer-weighting idea. This fork fixes its main flaw (below), adds presets, and hardens the engineering.

The original rebalance node reweights Krea 2's 12 Qwen3-VL conditioning taps — then hits the entire tensor with a global multiplier (default 4.0). The two compound: with the default weights the conditioning magnitude is inflated ~8.7× (4× multiplier × ~2.2× from the per-layer gains). That doesn't just "boost" — it destabilises the conditioning. That magnitude blowup is the setup-independent part; how it presents depends on your setup. Some report it as oversaturation. In our tests (Krea 2 Turbo) it showed up as skin artifacts (scarring + a birthmark-style spot) and likeness drift (a younger face, head and body tilt), with no visible saturation shift. We're not claiming oversaturation doesn't occur — only that in our setup the dominant symptom was different. Either way the root cause is the same inflated magnitude, and the rebalance works against itself.

This node flips the default: RMS-renormalised per-layer rebalancing. It shifts the ratios between taps (boost the deep detail layers relative to the shallow ones) while holding the overall conditioning magnitude constant — so the denoiser sees a rebalanced signal rather than an inflated one (in our tests: no artifacts, no likeness drift).

Same prompt, same seed (123), 8-step Turbo. The only difference is the conditioning node.

In this test there was no visible saturation shift between the three — but that's just our test (Turbo, 8 steps), not a claim that oversaturation never occurs. It's a commonly reported symptom on other setups (Raw, different weights, quantisation); here the magnitude blowup presented as artifacts and likeness drift instead. Renormalising holds the magnitude — which is the part that actually matters.

Reproduced on a second, independent render (a different prompt at 1024² / 8-step Turbo): 10.2 / 255 — 96.0% similar, still clean. The quality-preserving behaviour is prompt-stable, not a one-off.

Krea 2 doesn't condition on a single text embedding. Its DiT runs a small txtfusion module that turns 12 selected Qwen3-VL hidden-state layers (packed as (B, seq, 12·2560)) into one refined embedding, in three stages (mapped by fblissjr, verified against the code): (1) layerwise blocks — cross-layer attention where the 12 taps attend across the layer axis and L20 is the hub every layer attends to; (2) projector — a learned Linear(12 → 1) collapse whose weights are contrastive (positive on the mid layers, peak L14; negative on the deep ones L23/26/29/32) — roughly "mid minus deep"; (3) refiner blocks — cross-token self-attention that refines the now-single embedding per token before the DiT sees it. The deep detail is actively subtracted at stage 2 — the mechanism behind the under-representation.

This node intervenes upstream of all three stages: it reshapes the packed tensor to expose the layer axis and reweights each tap before txtfusion runs, boosting the deep taps to recover what the projector will subtract. Because that boost is then re-mixed through the L20-centric layerwise attention, its effect is indirect and gentle — a 5× deep-tap boost shifts the image only ~10/255 — which is why this is the forgiving lever. The weight-space alternative edits the projector (stage 2) directly and reaches a comparable magnitude at ~⅛ the strength; different insertion point relative to the L20 re-mix, so the two read as complementary (see Related work).

 conditioning   (B, seq, 12·2560)
      │   reshape to expose the layer axis
      ▼
               (B, seq, 12, 2560)
      │   × per-layer gain   (boost deep taps 7–10)
      ▼
      │   RMS renormalise  → ratios change, magnitude held   ← on by default
      ▼
      │   × global multiplier                                ← 1.0 by default
      ▼
 conditioning   (B, seq, 12·2560)    ← masks & pooled output untouched

• ⚖️ Quality-preserving by default — RMS renormalise is ON and the global multiplier defaults to 1.0. Shift tap ratios, hold magnitude. Reproduce the original node's behavior anytime with renormalize = false, multiplier = 4.0.
• 🎛️ Presets — balanced (the classic profile), detail, subtle, uniform, or custom.
• 🧱 Structure-safe — recurses the ComfyUI CONDITIONING structure; masks, pooled outputs, and arbitrary payloads pass through unchanged.
• 🔢 Tap-agnostic — defaults target Krea 2's 12 taps, but the math generalises to any multi-layer-tap conditioning.
• ⚡ Zero dependencies — pure PyTorch, nothing extra to install.

ComfyUI Manager — search Krea 2 Conditioning → Install.

Manual

cd ComfyUI/custom_nodes
git clone https://github.com/huwhitememes/comfyui-krea2-conditioning.git
# restart ComfyUI

• Start with the default (balanced, renormalize = true, multiplier = 1.0) — that's pure ratio rebalance with magnitude held. This is where the quality lives.
• Want it louder? Raise multiplier — but expect artifacts/likeliness drift past ~2–3×.
• Cleanest A/B against the un-rebalanced conditioning: multiplier = 1.0, renormalize = true.
• Reproduce the original node exactly: renormalize = false, multiplier = 4.0.
• Drop the node between your CLIPTextEncode (or Krea 2 text encode) and the sampler.

Full credit to nova452 for the per-layer-weighting technique.

• Krea 2 (Raw / Turbo) — primary target.
• Generalises to any diffusion model that conditions on a flattened multi-layer hidden-state stack; just match the weight count to the tap count.

• nova452 — ComfyUI-ConditioningKrea2Rebalance: introduced per-layer conditioning reweighting for Krea 2 — the activation-space technique this node refines with magnitude-preserving defaults.
• fblissjr — krea-explorations: interpretability + a weight-space complement. Rather than scaling the conditioning activations, it edits the learned txtfusion.projector weight (the [1,12] combiner over the taps), and shows that projector is contrastive (mid-minus-deep) and that L20 is a universal attention hub. The two approaches sit at different stages of the pipeline — ours pre-scales the input taps; theirs re-weights the combination — and are combinable.

This node is the activation-space, magnitude-preserving entry in that lineage.

Cross-tested (same seed/prompt, Krea 2 Turbo): the activation-space lever is the more forgiving of the two — a 5× deep-tap boost shifts the image ~10/255 and stays coherent, while a 4-band projector-weight edit only reaches a comparable magnitude around strength ~0.1 and breaks by 1.0 (the projector coefficients are the combination, so they're ~8× more sensitive). The two read as complementary, not redundant.

Per-layer conditioning rebalancing technique by nova452 — ComfyUI-ConditioningKrea2Rebalance. This fork extends it with quality-preserving defaults, RMS renormalisation, presets, and hardened parsing.

Apache-2.0. Krea 2 weights are released under the Krea 2 Community License — verify it covers your use case.