CogVideo Prompting Guide
Write prompts that get you the video you pictured. General, practical guidance for crafting effective prompts across all CogVideo tools.
A good prompt is the single biggest lever on your result in CogVideo. The model can only act on what you tell it, so a clear, specific description consistently beats a vague one. This guide covers general principles that apply across Text to Video, Image to Video, and Video to Video.
Principles
Use descriptive language
Include specific details about the subject, the scene, the action, and the mood. Concrete nouns and verbs give the model something to render.
- Weaker: "A dog on a roof."
- Stronger: "A golden retriever wearing sunglasses runs playfully across a sunny rooftop."
Add context
Describe the background, lighting, time of day, or atmosphere. Context shapes the look as much as the subject does.
- Example: "In a futuristic city, a robot chef prepares a gourmet dish under neon lights."
Set a style or tone
Style words steer the overall feel. Try terms like cinematic, surreal, realistic, or documentary to push the output in a direction.
- Example: "A surreal, dreamlike shot of a floating castle drifting through the clouds."
Avoid ambiguity
Be specific. Generic or vague descriptions leave too much open, and the result is harder to predict. If something matters, say it explicitly.
From simple to detailed
The same idea improves quickly as you add detail. These examples mirror the kinds of prompts each tool handles well.
| Tool | Simple prompt | Detailed prompt |
|---|---|---|
| Text to Video | A person walking in a park. | A young woman walks through a serene park in autumn. Golden leaves fall gently as a soft breeze rustles the trees. |
| Image to Video | An astronaut. | An astronaut plants a flag on Mars, with a red dusty sky and distant mountains in the background. |
| Video to Video | A bike rider. | A motorbike rider speeds through a forest road at sunset, golden light streaming through the trees. |
Iterate, do not perfect
You rarely get the exact result on the first try, and that is fine. Change one thing at a time, regenerate, and compare. Small wording changes often make a large difference.
Practical tips
- Lead with the subject and action, then layer in setting, lighting, and style.
- Keep it within the limit. Text to Video prompts are capped at 500 characters, so spend them on the details that matter most.
- Change one variable at a time when iterating, so you can tell what each change did.
- Use a seed to reproduce a result in Text to Video, then vary the prompt slightly to explore around a clip you like.
- Match the prompt to the tool. With Image to Video, describe the motion you want rather than re-describing what is already in your image.