Nougan Krea 2 · LoRA Stack 🌀

Same deal as the single NouganKrea2Loader - bundled uncensor LoRAs for Krea 2, applied from the pack's own lora/ folder - but this one applies the same LoRA stack to up to five models at once, and hands you all of them back. It's the node you reach for when you're testing a prompt across Krea 2 Raw, Krea 2 Turbo, and a couple of finetunes at the same time and don't want to wire the uncensor stage five times.

How it works

The mechanism is identical to the single loader: lora_data JSON from the themed UI, parsed into LoRA entries, each applied with comfy.sd.load_lora_for_models at strengths clamped to ±10, read from the pack-bundled lora/ folder. The difference is the fan-out - the node applies the stack to the main model and to whatever you plug into model_2 through model_5, and emits every patched model on its own output.

Useful detail: the extras are applied with no CLIP (the _apply call passes None for clip on the secondary models), which is fine - you're almost always just patching the diffusion weights on the alternates, and the main CLIP output covers the text encoder once.

The inputs that matter

• model - the primary model. The stack always applies here, and its patched result is output 1.
• lora_data - the shared stack, edited once in the UI.
• clip - optional, connects the text encoder for the primary model.
• model_2 … model_5 - up to four extra models, each gets the same stack and comes out its own MODEL 2 … MODEL 5 output.

Outputs: MODEL, CLIP, applied (the STRING summary of the stack), plus MODEL 2–MODEL 5.

Installing

Search Nougan in ComfyUI Manager, or:

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

Same zero-dependency story as the rest of the pack; the uncensor LoRAs ship inside the repo.

Where people get burned

Memory, mostly. Krea 2 is a 12B model and you're now holding up to five patched copies of it in memory during execution - the node itself does nothing to manage VRAM, so on a 24GB card you'll realistically wire two or three alternates, not five. The other recurring trap is assuming each alternate got its own stack tuned differently; it didn't - lora_data is shared, which is the feature. If you need per-model strengths, use five single loaders instead. For A/B/C sampling with the same uncensoring, this is the cleanest way to do it without turning your graph into spaghetti.