夜半-Krea2 六区LoRA加载器 V2

This is the node that makes the whole Yeban Krea2 LoRA Matrix pack make sense. Every LoRA you've ever applied in ComfyUI has one strength slider and that's it - the whole delta gets dumped on the model. The Krea2LoraFunctionalMatrixLoaderV2 (夜半-Krea2 六区LoRA加载器 V2) replaces that with a grid: each LoRA gets a top multiplier plus six per-area weights, and those areas map onto real Krea 2 DiT blocks. Two LoRAs both at strength 1.0 are rarely doing the same job - one rewrites composition, one is mostly brushwork. This node lets you keep only the part you actually want.

What it does under the hood

Krea 2 is a 12B diffusion transformer with 28 blocks. The loader buckets blocks 0–27 into six functional areas:

• 构图形体 - composition, blocks 0–3
• 主体特征 - subject and pose, blocks 4–7
• 画风笔触 - brushwork and line, blocks 8–11
• 色彩光影 - color and light, blocks 12–15
• 材质纹理 - material and surface, blocks 16–20
• 细节质感 - detail and sharpness, blocks 21–27

Effective weight for an area is just matrix value × top LoRA strength. No normalization, so several LoRAs can pile onto the same area, and a new LoRA's six cells start at zero on purpose - the author wants you to assign areas deliberately rather than inherit defaults. The grid itself lives in the frontend widget: up to four LoRAs per row with auto-wrap, plus the usual matrix toolbox (solo/mute, row copy-paste, clear, lock, undo, and in-workflow presets).

The two things worth knowing before you trust it:

• The 分析 button is a diagnostic, not a tuning tool. It reads the LoRA's actual safetensors, computes the Frobenius energy of its Delta = B × A delta, and shows each cell as 分配值 | 能量占比 - your assigned value, then what share of the LoRA's real parameter energy lands in that area (like B 1.00 | 33.8%). Read-only: it never rewrites your weights or connections. Results cache by file size and mtime, so re-analyzing the same file is fast.
• The six area names are empirical tendencies, not official Krea semantics. Krea never published a "composition block" mapping. Use the analysis to find where a LoRA actually moves the model, then confirm with fixed-seed tests.

Inputs and outputs

Just one required input: model (a Krea 2 MODEL). Everything else happens in the matrix widget. Three outputs:

• MODEL - the patched model, straight into your existing Krea 2 sampling chain
• 作用说明 - a Chinese text summary of what was applied
• LoRA权重记录 (LORA_STACK) - the hand-off: feed this to the pack's post-adjuster or test controller instead of the MODEL

Install

ComfyUI Manager: search "Yeban Krea2 LoRA Matrix" and install, or:

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

Restart ComfyUI. It needs ComfyUI ≥ 0.27.0 and frontend ≥ 1.45.20 - the matrix is a JS widget, so this is a real constraint, not a formality. There are no extra Python dependencies beyond what ComfyUI already ships. The pack itself won't download anything: you need a Krea 2 UNET (Raw or Turbo), the Qwen3-VL text encoder and the Qwen-Image VAE on your own, and your Krea 2 LoRAs in the usual models/loras folder. The example workflow references one specific demo LoRA that isn't bundled - swap in your own.

Where people get burned

First, tuning dead knobs. A lot of Krea 2 LoRAs only train a subset of blocks, and cells mapped to untrained blocks look identical to live ones. Run the analysis before you start: if an area shows 0% energy, that cell does nothing no matter what you type in it.

Second, treating the analysis as a promise. A high 能量占比 in a region tells you where the delta lives, not what it'll do visually. The README is explicit about this, and it's right. Area names are operating tendencies - verify with fixed seed, prompt, reference image and sampler, changing one area at a time.

Finally, this pack is verified on Windows + NVIDIA CUDA with ComfyUI 0.27.0, frontend 1.45.20, Python 3.12. Other setups are unverified - if you're on Linux and hit something odd, that's likely the first place to look.