Prompt tips for Qwen video/image analysis

General principles

• Be specific about what you want described. "Analyze this video" gives vague results. "Describe the character's body movement, arm positions, and weight shifts frame by frame" gives useful results.
• Ask for structured output (JSON, numbered lists) when you need to parse the response programmatically.
• Mention what to ignore if the video has irrelevant elements (e.g., "ignore the background, focus on the person's gestures").

Video analysis prompts

Motion/animation analysis

Describe the character's body movement in detail:
1. Starting pose
2. Transition movement
3. Main action/loop
4. Hand and arm positions
5. Weight distribution and shifts
6. Head movement
Ignore facial expressions and eye movement.

Scene composition

Analyze the visual composition:
- Camera angle and movement
- Lighting setup (key, fill, rim)
- Color palette and mood
- Depth of field and focus
- Subject placement in frame

Technical quality

Evaluate the technical quality of this generated video:
- Motion smoothness and consistency
- Artifact presence (flickering, morphing, distortion)
- Temporal coherence (do objects maintain shape between frames?)
- Hand/finger quality
- Face consistency
Rate overall quality 1-10 with explanation.

Comparison

When comparing multiple videos, analyze them sequentially and ask Qwen to compare:

I'll show you video clips of the same character performing similar actions.
For each, describe the motion quality, smoothness, and naturalness.
Then rank them from best to worst as animation references.

Image analysis prompts

UI/screenshot analysis

List all UI elements visible in this screenshot:
- Navigation elements
- Content areas
- Interactive controls
- Text content (transcribe accurately)
- Visual hierarchy

Reference image analysis

Describe this character reference image in detail:
- Body proportions and pose
- Clothing and accessories
- Art style and rendering technique
- Color palette
- Distinguishing features

Output format tips

For machine-readable results, always end your prompt with:

Respond in this exact JSON format:
{ ... your schema ... }
Return ONLY the JSON, no other text.

For human-readable results, ask for markdown with headers.