Nougan Krea 2 · LoRA 🌀

Krea 2's open weights come with a problem baked in: the alignment pass that flattened expressions, dropped body/violence prompt clauses, and gated the explicit stuff. The community's settled fix is an uncensor LoRA - the KB's Krea 2 panel is explicit that the LoRA is preferred over the conditioning-rebalance node because it costs less quality and restores general prompt responsiveness, not just NSFW. NouganKrea2Loader is that answer as a ComfyUI node: it bakes in bundled uncensor LoRAs that travel inside the pack, with per-LoRA toggles and an uncapped strength slider.

So this is less a "loader" and more a "load your model, then lay down the uncensoring foundation with one node" step. The README calls it the "Krea 2 · Uncensored" stage, and it's designed to sit between your model loader and your regular LoRA stack: base model in, uncensored model out, then you stack style/character LoRAs on top.

How it works

Under the hood it's a normal LoRA apply - comfy.sd.load_lora_for_models, same as any LoRA loader - but with two twists. The LoRA files are read from the pack's own lora/ folder instead of models/loras, so they're guaranteed present, and the strength is clamped to ±10 rather than the usual 0–1 range. That's deliberate: Krea 2's refusal behavior sometimes needs a hard shove, and the tooltip note about "no upper limit" is the author leaning into the fact that 1.0 isn't always enough on this architecture.

The lora_data JSON widget is managed by the node's themed UI - you get toggles, per-LoRA strengths, and a live readout of what's applied, and the hidden widget is what actually serializes your choices into the workflow.

The inputs that matter

• model - your loaded Krea 2 model (typically from this pack's Diffusers Loader).
• lora_data - managed by the UI; don't hand-edit.
• clip - optional, but connect it if you want the LoRAs applied to the text encoder too.

Outputs: MODEL (uncensored), CLIP, and applied - a STRING summary of which LoRAs hit, e.g. [email protected], [email protected]. That last one is handy for logging or feeding into a text display node.

Installing

Manager search Nougan, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Winnougan/Nougan_Nodes
# restart ComfyUI

Zero pip dependencies - the LoRAs come with the repo, so no extra downloads beyond the clone. Each family loads in its own try/except; if the Krea 2 family fails you'll see a [Nougan] ⚠️ krea2_loader NOT loaded line in the console and the rest of the pack carries on.

Where people get burned

The big one is ordering: this has to run before your other LoRAs, because a global uncensor LoRA stacked on top of a character LoRA in the wrong order can fight it. Also, the "no upper limit" freedom cuts both ways - strength way past 1.0 can start melting quality and prompt-following, which defeats the point. And remember the Krea 2 caveat that applies to every bypass: both the LoRA route and the conditioning-rebalance route cost something, the LoRA just costs less. If you're seeing quality loss you didn't cause, back the strength off and let the model do more of the work.