ZML_XY_LoRA加载器V3

The V3 loader is the family's answer to a specific question the README names outright: "use the same multiple prompts to test LoRA effects." Where V2 pulls each LoRA's own txt trigger, V3 inverts the setup - you supply one fixed prompt plus a multiline list of variable prompts, and the grid cross-products LoRAs × your prompts at a single weight. Same prompt set, many LoRAs, straight comparison.

The inputs tell the story. LoRA文件夹路径, LoRA数量, and a single LoRA权重 (no weight sweep - V3 drops the start/end interpolation in favor of one strength, which is the right call when the variable axis is the prompt). Then 固定提示词 (a non-multiline base prompt) and 多行变量提示词 - one line per cell. Each cell's final prompt is the fixed part plus one variable line, encoded to conditioning. 无LoRA对比 adds the unpatched baseline, XY互换 swaps which axis is LoRAs and which is prompts.

Outputs match V2: 模型 (MODEL list), 条件 (CONDITIONING list, pre-encoded per cell), 图表信息 for ZML_XY_图表拼接. CLIP is required, same as V2, because it encodes. The wiring pattern is identical: loader → sampler → grid drawer, and the grid drawer labels rows with LoRA names and columns with your variable prompt lines (or flipped).

Why would you pick this over V2? V2 is for "each LoRA has its own style to test" - character LoRAs with different trigger words. V3 is for "I have a style-neutral prompt, does this LoRA make it better or not" - style/effect LoRAs, checkpoints-style sweeps, or when you want every cell on identical ground. It's also simpler to reason about: one weight, prompts as the variable, no txt-file scavenging required. The V3 note in the README's changelog (2026.01) frames it exactly as "same prompts, test multiple LoRAs."

Installing it

Ships in ComfyUI-ZML-Image. ComfyUI Manager → ComfyUI-ZML-Image, or:

cd ComfyUI/custom_nodes
git clone https://github.com/zml-w/ComfyUI-ZML-Image

Restart. No extra downloads. Chinese UI; English patch at https://github.com/zml-w/ZZZ_ZML_English_Patch.

Common issues

The usual ZML loader gotchas apply - the E:\Models\Loras default path is meaningless on your machine, and LoRA数量 grabs the first N files in sorted order, so name files deliberately. 无LoRA对比 here adds one baseline cell exactly like V1/V2 (single baseline even with many prompts, reused via image copy), which is the pack's way of keeping the grid cheap.

Prompt-wise, the trap is empty variable lines: if 多行变量提示词 has a blank line or is entirely empty, you get one cell with just the fixed prompt - harmless alone, confusing when you expected N cells. The fixed+variable split is literal, so put punctuation between them yourself (the fixed prompt's trailing comma matters - the node doesn't add one). And remember there's no per-prompt weight here: every variable line rides at LoRA权重. If you need per-cell weight control, you want ZML_XY_提示词 (Prompt Loader) instead, which sweeps weights around a single variable prompt.