AgentSkillsCN

Visual Continuity Validator

视觉连续性验证器

SKILL.md

Visual Continuity Validator Skill

Purpose

Check visual consistency across all shots in an episode, identifying drift patterns and enforcing visual canon.

Trigger

Shot quality validation complete for episode.

Inputs Required

  • SHOTS_EP{{XX}}/*.png - All episode shots
  • SHOT_LIST_EP{{XX}}.json - Shot specifications
  • SHOT_QA_REPORT_EP{{XX}}.md - Individual shot QA
  • CANON_DB.json - Visual canon
  • CHARACTER_REFS/*/refs/*.png - Reference images

Outputs Produced

  • VISUAL_CONTINUITY_REPORT_EP{{XX}}.md
  • Drift alerts and correction recommendations

Continuity Dimensions

1. Character Continuity

  • Same character looks consistent across all appearances
  • Outfit consistency within scenes
  • Expression/pose progression makes sense

2. Location Continuity

  • Same location looks consistent across all shots
  • Time-of-day matches within scenes
  • Props/furniture don't move unexpectedly

3. Scene Continuity

  • Shots within a scene are visually cohesive
  • Lighting consistent within scene
  • Color grading consistent

4. Style Continuity

  • Overall aesthetic maintained
  • Color palette consistent
  • Mood appropriate throughout

Process

Step 1: Group Shots

Organize shots by:

  • Scene (same location, continuous time)
  • Character (all appearances of each character)
  • Location (all shots at each location)

Step 2: Character Thread Analysis

For each character:

  1. Collect all shots featuring character
  2. Compare character appearance across shots
  3. Identify drift patterns
  4. Flag significant inconsistencies

Check Points:

  • Face consistency
  • Hair consistency
  • Outfit consistency (within same scene)
  • Body proportions
  • Age appearance

Step 3: Scene Thread Analysis

For each scene:

  1. Collect all shots in scene
  2. Verify temporal continuity
  3. Check spatial consistency
  4. Validate lighting continuity

Check Points:

  • Lighting angle doesn't jump
  • Props don't move unexpectedly
  • Character positions make sense
  • Background consistent

Step 4: Location Thread Analysis

For each location:

  1. Collect all shots at location
  2. Compare architectural elements
  3. Verify key features present
  4. Check time-of-day consistency

Check Points:

  • Key architectural features
  • Furniture/prop placement
  • Window positions
  • Color of walls/surfaces

Step 5: Drift Detection

Calculate drift metrics:

Progressive Drift:

  • Character slowly changes over episode
  • Detected by comparing first vs. last appearance

Sudden Drift:

  • Abrupt change between adjacent shots
  • Detected by sequential comparison

Pattern Drift:

  • Consistent error across multiple shots
  • Detected by comparing against reference

Step 6: Generate Report

markdown
# Visual Continuity Report: EP{{XX}}

## Summary
- Overall Continuity Score: {{0-100}}
- Characters: {{SCORE}}
- Locations: {{SCORE}}
- Scenes: {{SCORE}}
- Style: {{SCORE}}

## Critical Issues

### Character Drift: {{CHARACTER}}
- First appearance: SC01_SH03
- Drift detected: SC05_SH02 onward
- Issue: Hair color shifted darker
- Action: Regenerate SC05+ with corrected refs

## Scene Continuity Issues

### Scene 03
- Shots affected: SH04, SH06
- Issue: Lighting angle inconsistent
- Action: Regenerate with matched lighting

## Drift Patterns

### {{CHARACTER}} Over Episode
[Visual timeline or scores showing consistency]

### {{LOCATION}} Across Scenes
[Visual timeline or scores]

## Recommendations

### High Priority
1. Regenerate shots: [list]
2. Update references: [if refs are causing drift]

### Medium Priority
1. Review and potentially regenerate: [list]

### Acceptable Variance
[List of minor issues that don't require action]

Step 7: Prioritize Corrections

Rank issues by:

  1. Critical: Breaks story clarity
  2. High: Noticeable distraction
  3. Medium: Visible but not distracting
  4. Low: Minor, acceptable variance

Step 8: Update Canon if Needed

If generated shots are consistently better than references:

  • Flag for reference-library-updater
  • Document what works better

Continuity Rules

Within Scene (STRICT)

  • Characters must look identical
  • Outfits cannot change
  • Props cannot move
  • Lighting must match

Across Scenes, Same Day (MODERATE)

  • Characters should look consistent
  • Outfits can change if justified
  • Minor lighting variation acceptable

Across Episode (FLEXIBLE)

  • Characters recognizable
  • Style consistent
  • Overall aesthetic maintained

Common Continuity Errors

Character Issues

  • Face features drifting
  • Hair changing
  • Age appearing different
  • Outfit changing mid-scene

Location Issues

  • Architecture changing
  • Props moving
  • Lighting direction flipping
  • Color of surfaces changing

Scene Issues

  • Shot-to-shot jumps
  • Lighting inconsistency
  • Color grading shifts

End-of-Clip Continuity Review (Video Production)

When generating video clips, continuity must be validated BETWEEN clips, not just within shots.

Workflow

code
Generate Clip N → Extract Last Frame → Claude Reviews → Decision:
    ├─ Continue to Clip N+1 (if continuity is good)
    ├─ Generate Bridge Clip (if gap needs bridging)
    └─ Generate New Frame(s) → New Bridge Clip (for exceptional continuity)

Review Process

After each clip is generated:

  1. Extract last frame using ffmpeg:

    bash
    ffmpeg -sseof -1 -i clip.mp4 -update 1 -q:v 2 last_frame.png
    
  2. Claude reviews the extracted frame:

    • What is the character's position/state/expression?
    • Does this flow naturally into the next clip's requirements?
    • Are there any continuity breaks?
  3. Decision:

    • Continue: Last frame flows naturally into next clip's start frame
    • Bridge needed: Gap exists but can be bridged with additional video
    • New frames needed: Generate supporting frame(s) using Nano Banana Pro

Bridge Clip Strategy

When a bridge clip is needed:

  1. Use extracted last frame as start frame (strategy: last_frame)
  2. Write prompts that transition FROM the current state TO the next clip's expected state
  3. Keep prompts START FRAME AWARE - describe continuation, not contradiction

Continuity Decision Criteria

ScenarioDecision
Character in same position, same framingContinue
Character in similar position, different framingContinue (cut handles it)
Character needs to move to new positionBridge clip
Significant time/action gapBridge clip
New locationNew generated start frame

Notes

  • Perfect continuity is impossible with current tech
  • Focus on maintaining character recognition
  • Location minor drift is less noticeable
  • Scene boundaries can hide more variance
  • Some drift is acceptable if not distracting
  • Progressive drift is harder to catch than sudden
  • End-of-clip review is essential for video production