夜半-Krea2 六区矩阵测试控制器 V1

The Krea 2 matrix loader tells you where a LoRA thinks it's working, but the only way to know what an area actually does visually is to isolate it and render it. Doing that by hand - duplicating sampler branches for every LoRA and every region - is exactly the kind of tedious, error-prone work that should be a node. Krea2LoraFunctionalMatrixTestControllerV1 (夜半-Krea2 六区矩阵测试控制器 V1) is that node: feed it a LoRA权重记录 from the loader and it generates the whole battery of test models in one shot.

What it generates

Give it the matrix record and it builds a sequence of patched models (plus labels and structured snapshots for every case):

• baseline - every LoRA off, so you have the control condition
• current - the loader's matrix exactly as configured
• controller combo - all LoRAs using the controller's own six-area weights
• then, for each candidate LoRA: a single-LoRA case, and six per-area isolation cases where only that one region of that one LoRA is active

You wire the models list into a batch sampler with a fixed seed and prompt, render, and compare. This is the workflow the pack's README keeps hammering: fix seed, prompt, reference image, latent size and sampler, change only the matrix variable, then read the result. If an area's isolation render changes nothing, you've found a dead region or a mislabeled one.

The inputs that matter

• test_plan - full (everything above), quick (baseline + current + controller + single-LoRA cases, skips the six per-area renders), or areas_only (just the isolation cases). quick is your friend when you're iterating on weights; full when you're actually auditing a LoRA.
• isolation_weight - total weight used for area isolation cases (default 1.0, 0–10).
• area_weight_source - where the per-area isolation weights come from: the controller's own per-LoRA total_weights, the loader's current matrix, or the fixed isolation_weight. Defaults to the controller weights.
• max_cases - a hard cap (default 64, up to 256). full on several LoRAs generates a lot of cases; if you blow past the cap the node refuses rather than silently dropping cases. Raise it or switch to quick.
• include_disabled_loras - whether LoRAs you've muted in the matrix still get test cases.
• disable_txtfusion_for_area - TextFusion layers are deliberately not area-controlled in this pack; this (default on) turns them off for isolation cases so they don't muddy the comparison.

One detail that surprises people: the controller applies its own weights directly, so it is not affected by whether the loader's areas are disabled, whether a cell is zeroed, or what the top multiplier says. If the controller's row says 20, the case runs at total weight 20 and the label says 20. There's also a per-LoRA "uniform total weight" toggle that greys out the six-area inputs and uses one total for every case.

Outputs and wiring

models, case_labels and matrix_snapshots are lists; case_count, columns and cells_per_grid are plain integers. The last three exist to feed the pack's YebanLoraTestXYZGrid node directly - batch-sample the models, then pass images, case_labels, columns and cells_per_grid into the grid and you get a labeled comparison sheet.

Install

Not a standalone node - same pack install as the loader:

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

or grab "Yeban Krea2 LoRA Matrix" via ComfyUI Manager and restart. ComfyUI ≥ 0.27.0, frontend ≥ 1.45.20, no extra Python deps, Krea 2 model files required (UNET + Qwen3-VL encoder + Qwen-Image VAE), verified on Windows + CUDA. If the node errors with "No LoRA test cases were generated", check that the LoRA权重记录 input is actually from the loader V2 - it won't work with a bare stack from anywhere else.