IAMCCS Redhead Node (Krea2)

Krea 2 was the best base model the open ecosystem got in the first half of 2026, and it shipped with a self-inflicted wound: the open weights went through an alignment pass the hosted model never got, and the damage isn't confined to explicit prompts. It flattens facial expression, drops clauses about bodies and violence, and generally behaves like a model that has decided not to hear part of what you said. The community's main fixes are an uncensor LoRA and conditioning rebalance nodes.

IAMCCS_RedheadNode is a third, stranger option. It doesn't touch your prompt or your conditioning. It patches the model so the text-fusion path inside Krea 2's diffusion model runs with a stronger internal signal, on the theory that if the model has the knowledge and is merely gating it, pushing the fusion pathway harder gets some of it back.

Inputs and output

Four widgets and one socket:

• model - the diffusion model, required.
• enabled - default true. This is the on/off switch.
• strength - 0 to 2, default 1.0, step 0.05. 0 is a true bypass.
• debug - prints what it measured.

Output: model, the patched model. Wire it between your Krea 2 loader and the sampler, the same place you'd put a LoRA loader.

How it works, as far as the code will tell you

It attaches a diffusion-model wrapper under a fixed key, storing enabled, strength and debug in the model's transformer options. On every step, the wrapper checks that the model actually is Krea 2's text-fusion layout - twelve tap layers of dimension 2560, twenty-four chunks of 1280 - and if it isn't, it silently skips (printing a note only when debug is on). That check is the reason this node is safe to leave in a graph you sometimes run on other models.

When it does match, it runs the text-fusion forward twice: once as normal to get a reference output, then again on an input scaled by a hand-authored 24-chunk gain profile, with a global multiplier applied on top. At strength = 1 that global multiplier is 15×, so yes, the default setting is the loud one. The two results are compared and the enhanced one is blended back toward the reference, with a per-token cap: the relative change introduced at any token is clamped to 0.75 of its base RMS. So the node cannot run away with the conditioning. strength scales both the chunk gains and the global multiplier, so it fades the whole intervention in rather than switching the profile off.

Then it restores the original forward pass, and it's careful about it - the original is stashed, the enhanced version is swapped in for one call, and the finally block puts it back. The wrapper also carries a re-entry guard, so if the model is invoked again inside that call it goes down the unpatched path instead of recursing forever. The wrapper is removed and re-added on each apply, so chaining two of these doesn't stack them.

debug is not decoration

With debug on you get up to eight printed lines per run:

[IAMCCS_RedheadNode] strength=1.000 progress=0.412 sigma=0.744 global=15 proj_ratio=... out_rel=... out_cos=... clamp=...

out_cos is the cosine similarity between the normal and enhanced outputs - near 1.0 means the intervention is doing almost nothing, and it's worth knowing that before you attribute a good render to it. out_rel is how much the output moved relative to its own RMS. clamp_mean tells you how often the per-token cap is actually engaging; if it's clamped nearly everywhere, you're at the ceiling of what this can do and turning strength up won't help.

Should you use it?

Honest answer: it's a lever, not a fix, and it's closer in spirit to a conditioning hack than to the uncensor-LoRA route people broadly settled on. If your problem is flat expressions or prompts that get ignored, try it with debug on, look at out_cos and clamp_mean, and judge for yourself whether the change is worth the loss of predictability. It costs almost nothing to test - no model downloads, no graph surgery, one node.

What it isn't: a licence workaround. Krea 2's acceptable-use policy forbids circumventing safety measures, and a node that pushes the fusion pathway harder sits in the same grey zone as the rebalance nodes and uncensor LoRAs. Local, personal use has never been enforced against; commercial and public use is where that exposure starts to matter.

Install

Pack install as usual:

cd ComfyUI/custom_nodes
git clone https://github.com/IAMCCS/IAMCCS-nodes.git

Or Manager → search IAMCCS, restart. You need a working Krea 2 setup to use it - the model, the Qwen3-VL text encoder and the Qwen-Image VAE, loaded through stock Krea 2 nodes - and remember to save your favourite: because the node registers under conditioning/krea2, it appears next to ComfyUI's own Krea 2 conditioning nodes.