Wan2.2提示词预设生成器

Wan 2.2 is the last open-weight Wan, and it has a prompting problem: the community's own advice for good clips is basically "describe the image like a robot, then narrate the motion like George R.R. Martin." That's a lot of writing, and it's the exact spot where people stall. Wan22PromptGenerator (display name "Wan2.2提示词预设生成器") is a shortcut through it - five dropdowns in, a complete descriptive motion prompt out. No API, no API key, no model download. It's just a very large curated template library stapled to a string-joiner.

Here's the honest framing, though: this is a tool for Chinese-language prompting, and a training-wheels one at that. The whole pack is a single commit from August 2025 by a hobbyist author, sits at zero community traction, and its UI and output are entirely in Chinese. Wan 2.2 is Alibaba's model, so it understands Chinese prompts natively and this works fine - but if you prompt in English, you'll be translating the output, and at that point the convenience mostly evaporates.

How it works

The node loads two JSON files from its own directory (wan22_templates.json plus character_actions.json) and assembles a prompt in a fixed order: subject description → emotion → action → camera move → lighting. It does a bit of text surgery on the way, swapping placeholder words like "女孩" (girl), "她" (she), and "主体" (subject) for whatever you typed in the 自定义主体 field - which is why that field wants a plain noun, not a sentence.

The output is a single STRING named 生成的提示词 ("generated prompt"). Wire it straight into the text/prompt input of your Wan 2.2 T2V or I2V sampler - the native ComfyUI Wan nodes or Kijai's WanVideoWrapper both accept it. That's the entire integration; there's nothing else to connect.

The inputs that matter

Only two are required, and the one you'll actually touch is 自定义主体 (custom subject) - default is "女孩", so change it or every clip stars the same girl. 主体类型 (subject type) toggles between 人物 (person) and 物体 (object), which flips which camera preset applies.

The optional dropdowns are where the library lives:

• 人物运镜类型 / 物体运镜类型 - 19 person camera moves and 18 object moves ("slow push-in", "orbit + zoom", "diagonal rise"...). Only the one matching your subject type is used; the other you can ignore.
• 光线类型 - 21 lighting looks, from daylight to neon to aurora.
• 人物动作 - 31 action presets (walk, run, turn, react...).
• 情绪表情 - 10 expressions, from grinning to screaming.

Installing

cd ComfyUI/custom_nodes && git clone https://github.com/ZB678232321/ComfyUI-wan-prompt-generator, then restart ComfyUI. You'll find the node under Wan2.2/提示词生成. ComfyUI Manager can also grab it if you search the pack title. There are no model files and nothing GPU-heavy.

One quirk worth knowing: requirements.txt lists openai and anthropic alongside the usual pillow/requests. The node never calls an API - those are leftover deps from a template project, harmless dead weight. Don't let them make you think it phones home.

Where people get tripped up

• The output is Chinese. Fine for Wan 2.2, but keep your whole prompt in one language or you'll get odd adherence.
• It's a preset, not a collaborator. Every prompt is assembled from the same skeleton, so batch-testing camera moves across subjects is where it shines; don't expect it to surprise you.
• Zero support risk. One commit, no issues, no roadmap. Treat the template library as the product - if it works at install, it'll keep working, but don't hold your breath for updates.