万相提示词

Wan's video models want long, cinematic, natural-language prompts - "a Hanfu woman in a classical corridor, camera tilting up as she slowly raises her head" - and most people are bad at writing those from a blank box. Wan_Prompt is a prompt builder: instead of running an LLM, it gives you a stack of dropdowns - light source, time of day, shot size, composition, lens, camera move, color tone, emotion, and more - and stitches your picks together into exactly that kind of Wan-flavored Chinese prompt. No model download, no inference, no VRAM. It's instant, offline, and deterministic until you turn on the random button.

It's part of AIGCZero/ComfyUI-tools_zero, the same pack that ships the Qwen prompt nodes. Where those run a real language model to write for you, this one is a pure template assembler - think of it as the "structured natural language" option for people who want control over every shot element without typing a sentence of film jargon.

What it does

Type what you want into the 用户输入 (user input) box, then fill in as many of the fourteen category dropdowns as you like. They cover the full shot vocabulary in Chinese: 光源类型 (light source - daylight, neon, candlelight), 光线类型 (light quality - soft, hard, rim), 时间段 (time of day), 景别 (shot size - close-up to wide), 构图类型 (composition - rule of thirds, symmetry), 镜头描述 and 镜头角度 (lens and angle), 色调 (color tone), 运动描述 (motion), 人物情绪 (character emotion), 基础运镜 (basic camera moves), 运镜速度 (camera speed), 视觉风格 (visual style - 19 of them, from pixel art to cyberpunk), and 特效镜头 (effect shots like dolly zoom and timelapse). Any dropdown left on none gets skipped.

The node does one piece of smart assembly: if you pick both a camera move and a speed, they merge into one phrase like "快速地镜头推进" (fast camera push-in) instead of two awkward clauses. Everything gets joined with Chinese commas, your user input lands at the end, and out comes a single prompt STRING ready to feed Wan's text encoder.

There's also 开启随机提示词 (enable random prompt) plus a seed. Flip it on and the node randomly fills the seven "atmosphere" categories - light, time, shot size, camera move, tone, speed - while keeping your explicit picks, so you can brute-force variations and use the seed to make any one of them reproducible. It restores the global random state afterwards, so it won't disturb other random nodes in your graph.

How it works

There's no mechanism beyond dictionaries and string joining. The PROMPT_DATA table defines every option per category; the node walks your selections, applies the random overlay if enabled, merges speed+camera-move pairs, and joins the parts. That's the whole thing - which is why it's the safest node in the pack to trust: nothing runs except string formatting.

Installing it

ComfyUI Manager → search "ComfyUI-tools_zero", or:

cd ComfyUI/custom_nodes
git clone https://github.com/AIGCZero/ComfyUI-tools_zero

Restart after cloning. Despite the pack pulling in transformers and bitsandbytes for its other nodes, this one works with zero model files.

Where people get burned

The honest catch: it's a Chinese-first node. All the labels, options, and the output itself are Chinese, and Wan genuinely does like Chinese cinematic prompts - many Wan workflows are built around them, so that's a feature. If your workflow is English-only, you can still use it: pick values from the dropdowns and type your own English subject into 用户输入; you'll get a mixed prompt that mostly works. And remember it builds, it doesn't write - for full sentences and story you want the pack's Qwen2_prompt video modes, which run an actual model. For controlled, repeatable shot selection, this is the faster tool.