视频指令词模板🐠meeeyo.com

Here's the situation: video models are picky about prompts. A sentence like "a cat on a roof" is fine for an image model but barely touches a video model's prompt encoder - video needs movement, camera, lighting, transitions, written in an order the model can follow. GenerateVideoPrompt (视频指令词模板 - "video instruction template") is a wrapper that does the prompt surgery for you: you drop in a plain description, pick a mode, and it returns a fully-structured cinematic prompt ready to feed Wan, Hunyuan, or whatever video pipeline you're on. No API, no key - it's pure string templating, right on your machine.

How it works

Two inputs: input_text (your raw description) and mode, a dropdown with five choices:

• 原文本 - passthrough. Returns your text untouched. Good for testing.
• 文生视频 (text-to-video) - wraps your text with a "you are a cinematic director" instruction that demands a chronological structure: main action, movement/gestures, character appearance, background, camera angles, lighting/colors, sudden changes - capped around 200 words.
• 图生视频 (image-to-video) - a similar director-style prefix aimed at expanding a single image or description into a full scene: action, environment, background details, camera movement, lighting, transitions.
• 首尾帧视频 (first-and-last-frame video) - the two-image mode. It instructs the model to narrate a smooth bridge from Image One to Image Two: scene continuity, emotional flow, output as one cohesive paragraph under 200 words.
• 视频负面词 (video negative words) - no input needed; it returns a fixed, prewritten negative-prompt block (overexposure, static artifacts, blurred details, extra fingers, distorted anatomy, and so on).

The output is a single STRING. And here's the thing to understand: the templates are written in English, aimed at English-prompted video models, even though the UI labels are Chinese. If your pipeline is Chinese-native (Wan2.x local tools often are), feed it accordingly.

Why it's useful

It bakes in the "prompt as structured instruction" approach the current models reward - see the KB's prompt-engineering notes on how modern video encoders want blocks and camera language, not tag soup. For beginners it's the fastest way to stop getting "static, near-image" video output and start getting actual motion, because the template forces the movement/camera/lighting details that make video video.

Install

Part of ComfyUI_StringOps:

cd ComfyUI/custom_nodes
git clone https://github.com/MeeeyoAI/ComfyUI_StringOps.git

Restart. Or ComfyUI Manager → "ComfyUI_StringOps". No models to download - the templates are hardcoded strings in the node source. Standard pack deps only. Chinese UI with the 🐠meeeyo.com branding.

Common issues

The biggest gotcha is mode leakage: the negative mode ignores your input entirely and returns the fixed block - don't type a description there expecting it to blend in. And the prompt-building modes prepend the template, so your text is the "main content" at the end; if you paste a whole pre-written cinematic prompt in, you'll get a doubled intro. Keep input plain and let the template do the dressing.

One honest caveat: the "200 words" limit is in the template text the LLM-ish prompt encoder sees, not enforced code - but since it's guiding an encoder, not an LLM, treating it as a soft target is fine. For the 首尾帧 mode, remember it only writes the bridge prompt - you still have to wire the two images into your video pipeline yourself. It's a prompt builder, not a video loader.

Handy, opinionated, and genuinely useful if you're doing text/image-to-video and keep writing prompts that come out static.