夜半-Krea2 六区LoRA后置调节器 V3

Once you've built a six-area matrix in the loader, you don't want to edit every cell every time you feel like the whole thing is half a step too strong. That's what Krea2LoraFunctionalMatrixPostAdjusterV3 (夜半-Krea2 六区LoRA后置调节器 V3) is for: it reads the loader's LoRA权重记录 and re-applies the entire matrix with one group multiplier plus six per-area multipliers, leaving the grid itself untouched. Think of it as a mixing-desk master fader bolted on after your routing.

How it works

The loader's output is a structured LORA_STACK - a JSON record of every LoRA, its top strength, and its six-area assignment. The post-adjuster consumes that record and applies a simple chain:

final effective weight = matrix value × top strength × group scale × area scale

No normalization anywhere, and every multiplier defaults to 1.0, so dropping the node into your graph changes nothing until you touch a knob. That's the behavior you want from a post-stage: it should be invisible by default.

The genuinely useful bit is that it merges up to eight stacks. The required LoRA权重记录 input plus seven optional ones (LoRA权重记录2 through LoRA权重记录8). If you're running two or three loaders on different model paths or groups, you can pipe all their records through one adjuster and trim the combined result with a single set of faders instead of editing three grids. The node emits a fresh LoRA权重记录 of its own, so you can chain a second adjuster after it if you're the kind of person who needs two master faders (no judgement).

Inputs and outputs

Required: model (the Krea 2 MODEL) and LoRA权重记录 (a stack straight from the loader V2 - or from a previous adjuster). The seven extra stack inputs are optional. Outputs are the same trio as the loader:

• MODEL - the re-patched model, onward to your sampler
• 作用说明 - a summary line showing which group settings were applied, including per-area scales
• LoRA权重记录 - the updated stack, for chaining or for the test controller

Install

Same as the rest of the pack - this isn't a standalone node:

cd ComfyUI/custom_nodes
git clone https://github.com/wsq194/yeban-krea2-lora-matrix.git Yeban_Krea2_Lora_Matrix

or install "Yeban Krea2 LoRA Matrix" from ComfyUI Manager and restart. It requires ComfyUI ≥ 0.27.0 and frontend ≥ 1.45.20 (the matrix UI is a JS widget), and there are no extra Python dependencies. It's a Krea 2 pack through and through - you need Krea 2 weights, the Qwen3-VL text encoder and Qwen-Image VAE in place, and it's only verified on Windows + CUDA so far.

The trap to avoid

The per-area multipliers multiply against whatever the loader already did. If you've zeroed an area in the loader and want to bring it back, the adjuster can't - zero times anything is still zero. The post-adjuster is a trim stage, not a rescue stage; if an area is dead in the matrix, fix it in the matrix. And remember the pack's own warning: those six area labels are empirically derived operating tendencies, not official Krea semantics, so adjust with a fixed seed and check what actually changed rather than trusting the name on the fader.