Shot Quality Validator Skill

Purpose

Perform quality assurance on generated shot images, flagging issues for regeneration and tracking quality metrics.

Trigger

Shots generated by shot-image-generator.

Inputs Required

• •SHOTS_EP{{XX}}/*.png - Generated images
• •SHOT_LIST_EP{{XX}}.json - Shot specifications
• •CANON_DB.json - Reference data
• •CHARACTER_REFS/*/refs/*.png - Character references

Outputs Produced

• •SHOT_QA_REPORT_EP{{XX}}.md - Quality assessment report
• •Updated SHOT_LIST_EP{{XX}}.json (with QA status)
• •List of shots requiring regeneration

Quality Dimensions

1. Reference Consistency

Does the shot match the reference images?

2. Technical Quality

Is the image technically sound?

CRITICAL: Single Moment Check Frame prompts must produce a single clean moment, not composites. Watch for:

• •Multiple versions of same character in frame (before/after states)
• •Split-screen or triptych compositions (unless explicitly requested)
• •Motion blur suggesting multiple moments
• •Transitional language in prompt interpreted literally

Common cause: Prompts containing transitional language like "expression shifts from X to Y" will cause models to render both X and Y states in a single image.

Fix: Move transitional language to VIDEO prompts only. Frame prompts should describe a static moment (the START state).

3. Composition

Does the shot match specification?

4. Style Consistency

Does the shot match show aesthetic?

Process

Step 1: Load Shot and Specification

For each shot in SHOTS_EP{{XX}}:

1. •Load the generated image
2. •Load corresponding shot specification
3. •Load relevant character/location references

Step 2: Automated Checks

Technical Checks (automated):

• •File exists and loads correctly
• •Dimensions match specification
• •File size is reasonable
• •No obvious corruption

Basic Visual Checks (if using vision model):

• •Number of figures matches character count
• •General composition matches type
• •No obvious artifacts

Step 3: Reference Comparison

Compare generated shot against:

• •Character reference images
• •Location reference images

Score similarity:

• •High (0.85+): Very consistent
• •Medium (0.70-0.84): Acceptable with minor drift
• •Low (<0.70): Needs regeneration

Step 4: Manual Review Flags

Flag for human review if:

• •Multiple characters present
• •Complex action/pose
• •First shot of scene
• •First appearance of character
• •Similarity score < 0.80
• •Automated checks flag issues

Step 5: Generate QA Report

# Shot QA Report: EP{{XX}}

## Summary
- Total Shots: {{COUNT}}
- Passed: {{COUNT}} ({{%}})
- Flagged for Review: {{COUNT}} ({{%}})
- Failed (Regenerate): {{COUNT}} ({{%}})

## Failed Shots

### EP01_SC03_SH02
- **Issue**: Character face drift
- **Severity**: High
- **Action**: Regenerate with higher ref weight
- **Notes**: Eyes and hair differ from reference

## Flagged for Review

### EP01_SC05_SH01
- **Reason**: Complex two-shot
- **Checks Passed**: 8/10
- **Notes**: Verify character positions

## Quality Metrics

### By Scene
| Scene | Shots | Pass Rate | Notes |
|-------|-------|-----------|-------|
| SC01 | 12 | 83% | |

### By Character
| Character | Appearances | Consistency Score |
|-----------|-------------|-------------------|
| ALICE | 45 | 0.87 |

### Common Issues
1. {{ISSUE}} - {{COUNT}} occurrences

Step 6: Update Shot List

Add QA status to each shot:

{
  "qa": {
    "status": "passed|flagged|failed",
    "score": 0.85,
    "issues": [],
    "reviewed_by": "auto|human",
    "timestamp": "..."
  }
}

Step 7: Generate Regeneration List

Create list of shots needing regeneration:

{
  "regeneration_queue": [
    {
      "shot_id": "EP01_SC03_SH02",
      "reason": "character_drift",
      "suggestion": "increase_ref_weight",
      "priority": "high"
    }
  ]
}

Severity Levels

CRITICAL (Regenerate Required)

• •Character unrecognizable
• •Wrong number of figures
• •Major anatomical errors
• •Completely wrong location
• •Composite/multi-moment image (frame shows multiple states of same subject)

HIGH (Likely Regenerate)

• •Significant character drift
• •Wrong shot type
• •Major composition issues
• •Obvious artifacts

MEDIUM (Review Needed)

• •Minor character variations
• •Slight composition drift
• •Lighting inconsistency
• •Minor artifacts

LOW (Acceptable with Notes)

• •Very minor variations
• •Style slightly off
• •Could be improved but usable

Automation Levels

Fully Automated

• •Technical file checks
• •Dimension verification
• •Basic presence detection

Semi-Automated (Vision Model)

• •Figure counting
• •General composition check
• •Obvious error detection

Manual Required

• •Detailed character verification
• •Nuanced expression matching
• •Final approval decisions

Notes

• •Start with automated checks to filter obvious failures
• •Use sampling for large batches
• •Track patterns in failures to improve generation
• •Quality threshold can be adjusted per project
• •Some variance is expected and acceptable
• •Perfect consistency is not the goal