ZML_标签化提示词加载器

ZML_标签化提示词加载器 looks like a plain text box in the schema and is anything but in the canvas. The backend is a single-input, single-output pass-through - positive_prompt STRING in, 提示词 STRING out - and all the actual work lives in the JavaScript frontend: a visual prompt builder where you browse tags, search in Chinese or English, and assemble a prompt by clicking rather than typing.

This is the "tagged prompt loader" family the README mentions, borrowing the UI approach of WeiLin's prompt-all-in-one. It supports importing tag lists from txt files in 中文,英文 (Chinese, English) format, local word-bank search, and translation of tags - the author updates the search/translation logic and ships bundled txt files for it. The backend just echoes whatever the UI produced, so the entire value proposition is "a friendlier way to build a prompt than a raw text widget."

The interface

• positive_prompt - multiline STRING. What the UI assembles; you can also type directly.
• Output: 提示词 (STRING) → wire to your CLIP Text Encode positive input.

There's also an onResize handler in the source that pins the node to a 100px-tall body, so the node doesn't balloon - the UI is a compact panel, not a wall of text.

The honest read

The README marks this as "目前还是半成品" - a half-finished work. The author is transparent that it's a work-in-progress port of a tag-builder idea. So: it works for the tag-browse-and-search flow, it's genuinely nice if you think in booru-style tags and want a local searchable bank, and it is not the most polished thing in the pack. The same author has since put more effort into the tagged-image loader and the V3 text-select nodes, which are the mature members of this family - if the search/translate behavior feels flaky, those are the alternatives.

And a reminder that applies to every tag-based prompt tool: tag prompting is an SDXL-lineage habit (Illustrious, NoobAI, Pony, Wan). On LLM-encoded models your prompt wants to be an instruction, not a comma-separated tag list - the KB's prompt-engineering refresh makes that point hard. Use this where your model actually eats tags.

Install

Part of ComfyUI-ZML-Image - one install gets all 160+ nodes. ComfyUI Manager → search ComfyUI-ZML-Image, or:

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

restart. No model files, no dependencies beyond the base stack. The txt files for search/translation ship inside zml_w/txt. English users: the UI is Chinese-first, so the README's translation patch is worth it here more than most nodes.