中文提示词构建器

If you prompt Qwen-Image or Z-Image in Chinese - and honestly, if you're running those models you probably should be - ChinesePromptBuilder (中文提示词构建器) turns a giant wall of dropdowns into your final prompt. It's one node in the tiny prompt_builder pack, it needs no API key, downloads nothing, and takes up no VRAM. It's just JSON and Python glue, and that's kind of the point.

What it is

Open the node and you're greeted by roughly twenty dropdowns, grouped the way an art director would organize a brief: 主体/Subject (people, animals, objects), 场景/Scene (indoor and outdoor), 时间与天气/Time & Weather, 动作与姿态/Action & Pose, 服装与配饰/Clothing & Accessories, 风格/Style (art styles, named artists, render engines), 材质, 光照, 氛围, 构图, 参数, and an NSFW group that defaults to 无 (none). Every choice shows both Chinese and an English gloss - "单人 (Single Person)", "Van Gogh 梵高" - so you don't need to read Chinese to use it, even though the whole thing is clearly built for a Chinese-speaking audience.

Every dropdown has two special options beyond the real values: 无 (none, the default) and 随机 (random). Pick 无 and that category contributes nothing. Pick 随机 and the node rolls one value from that category's library each time the graph runs. Nothing forces you to fill all twenty; the default is deliberately empty.

How it works

The mechanism is refreshingly simple, and I checked the source so this isn't README hearsay. The pack ships prompt_categories.json, a nested dict of category → values. At import time INPUT_TYPES walks it recursively and turns every leaf into a combo widget. On execution, build_prompt walks your selections, swaps any 随机 for a random.choice out of that category's list, joins whatever survived the 无 filter with commas, then wraps the whole thing with your 前置提示词 (prefix) and 后置提示词 (suffix) and cleans up stray double commas.

The sneaky part is IS_CHANGED. ComfyUI caches node outputs, so a plain combo node would only re-roll your randoms once. This node detects that any category is set to 随机 and returns the current timestamp, forcing a fresh execution every run. That's what "每次执行自动刷新" in the description means - set two categories to 随机 and you get a new composition on every Queue, not a stuck one.

The single output is 最终提示词 (STRING). Wire it straight into your Qwen-Image or Z-Image text encoder - or a Save Text node if you just want to batch-generate prompt lists.

Why this shape works for Qwen and Z-Image

Here's the thing that makes this node smarter than it looks. Qwen-Image and Z-Image don't use CLIP; they encode prompts with a general-purpose LLM, so your prompt is an instruction, not a bag of tags. Attention weights like (word:1.3) get silently discarded, and the negative prompt box does nothing at all. What those models want is structured, comma-separated blocks - which is exactly what this node outputs. The named-artist and camera-angle dropdowns are especially worth reaching for, since naming real cameras and painters is the highest-leverage phrasing on LLM-encoded models.

One caveat from the community: Z-Image prompt testing shows shorter prompts often beat longer ones - the top reply to a popular enhancer comparison was "this just adds unprompted stuff." So treat the 20 dropdowns as a buffet, not a checklist. Fill the categories you care about, leave the rest on 无, and don't auto-random everything just because you can.

Installing it

Easiest path is ComfyUI Manager: search "prompt_builder" and install. Otherwise:

cd ComfyUI/custom_nodes
git clone https://github.com/tudoupaigu946/prompt_builder

then restart ComfyUI. No model files, no weights, no GPU setup.

The gotcha: PySide6

The pack's requirements.txt pins PySide6==6.7.2, and the README makes a fuss about it. Ignore it. The actual ComfyUI node imports only json, random, os, time, and typing - PySide6 belongs to the standalone GUI the author also describes, not to this node. If Manager chokes installing it (on Linux it often needs system Qt libraries like libxcb-cursor0), that's fine: the node will still load and run. It's one of those rare cases where a broken dependency in the install log is a non-issue.

If you edit prompt_categories.json to add your own categories or values - the format is just nested arrays - restart ComfyUI to pick it up, since it's loaded once at import. And if you somehow break the JSON, the node won't register, so keep a backup before you hand-edit.