Aesthetic Scorer Skill#

This skill provides comprehensive aesthetic evaluation and improvement suggestions for images and photographs through a dynamic weighted two-tier architecture.

Architecture Overview#

Two Evaluation Sources with Dynamic Weight:

1. Improved Aesthetic Predictor: CLIP ViT-L/14 + MLP for content-level aesthetic scoring
   • Understands image semantics and visual impact
   • Base Weight: 45% (adjustable 45%-70% based on NIMA consensus)
2. NIMA (Neural Image Assessment): MobileNet for technical quality scoring
   • Provides detailed quality distribution and standard deviation
   • Base Weight: 55% (adjustable 30%-55% based on NIMA consensus)

Dynamic Weight Logic:

• NIMA returns a standard deviation (std) indicating score distribution spread
• High std = controversial image = more weight to IAP (content-based)
• Low std = consensus = use balanced weights

Dynamic Weight Formula#

text

normalized_std = min(nima_std / 2.5, 1.0)
weight_iap = 0.45 + normalized_std * 0.25    // Range: 45% - 70%
weight_nima = 1.0 - weight_iap

// Penalty when scores diverge significantly (diff >= 2.0)
penalty = 0.05 * |IAP_score - NIMA_score|

weighted_score = weight_iap * IAP + weight_nima * NIMA - penalty

ParameterValueDescriptionBase IAP Weight45%Balanced starting pointBase NIMA Weight55%Balanced starting pointMax Adjustment±25%IAP weight increases with controversyIAP Weight Range45% - 70%Dynamic adjustmentDivergence Threshold2.0Score difference triggers penaltyDivergence Penalty0.05Per point of difference

Evaluation Text Generation:

• The detailed evaluation text (composition, color, lighting, technical quality, improvement suggestions) is generated by the AI (WorkBuddy) based on:
  • The weighted scores from both models
  • Visual understanding of the photo content
  • Professional photography knowledge and best practices
• This provides professional-grade analysis without requiring external API calls

Workflow#

Phase 1: Execute Both Evaluations#

Step 1: Improved Aesthetic Predictor (dynamic weight)

1. Execute scripts/score_improved_predictor.py <image_path>
2. Parse output score (0-10 scale)
3. Record as score_improved

Step 2: NIMA Model (dynamic weight)

1. Execute scripts/score_nima.py <image_path>
2. Parse mean score AND standard deviation
3. Record as score_nima and nima_std

Phase 2: Calculate Weighted Comprehensive Score (Dynamic)#

Step 2.1: Calculate Dynamic Weights

• Extract NIMA's standard deviation (std_score)
• Normalize: normalized_std = min(std_score / 2.5, 1.0)
• Calculate weights: weight_iap = 0.45 + normalized_std * 0.25

Step 2.2: Apply Penalty if Needed

• If |IAP - NIMA| >= 2.0, apply penalty: penalty = 0.05 * |IAP - NIMA|

Step 2.3: Calculate Final Score

text

weighted_score = weight_iap * IAP + (1-weight_iap) * NIMA - penalty

Example:

• IAP = 6.44, NIMA = 4.52, NIMA_std = 1.87
• normalized_std = 1.87/2.5 = 0.75
• weight_iap = 0.45 + 0.75×0.25 = 0.637 (63.7%)
• weight_nima = 0.363 (36.3%)
• Score_diff = 1.92 < 2.0, penalty = 0
• Final = 0.637×6.44 + 0.363×4.52 = 5.74

Security Note: All processing is 100% local - no data leaves your device

Phase 3: Generate Evaluation at Appropriate Detail Level#

CRITICAL: Three Detail Levels Available

Always generate the detailed evaluation (10分) in the background first and save it. Then present the evaluation at the requested detail level:

LevelNameWord Count per PhotoDescription1简要~200字Concise overview, key points only3中等 (默认)300字Balanced, covers all aspects10详细4000字Comprehensive, in-depth analysis

How User Requests Different Levels:

• 默认/未指定 → 使用 3分（中等），约300字
• "详细评价" / "详细版" / "完整评价" → 使用 10分（详细），约4000字
• "简要评价" / "简洁版" / "简要说明" → 使用 1分（简要），约200字

Important:

• ALWAYS generate detailed evaluation (10分) in background first
• Save detailed evaluation so it can be retrieved immediately if user requests it
• Present the appropriate level based on user request (default: 3分)
• If user requests detailed evaluation after seeing summary, retrieve the saved detailed version

Output Format by Detail Level#

Level 1: 简要 (~200字 per photo)#

markdown

## [Photo Name]

综合评分: X.XX/10 (等级: 夯/顶级/人上人/NPC/拉完了)
构图: [2-3句话]
色彩: [2-3句话]
光线: [2-3句话]
技术: [2-3句话]
建议: [3-4条关键建议]

Level 3: 中等 (默认, ~300字 per photo)#

markdown

## [Photo Name]

### 综合评分: X.XX/10 (夯/顶级/人上人/NPC/拉完了)

### 综合分析

#### 构图评价
[3-4句话]

#### 色彩评价
[3-4句话]

#### 光线评价
[3-4句话]

#### 技术质量评价
[3-4句话]

### 改进建议

#### 拍摄技巧
[3条建议]

#### 后期处理
[3条建议]

#### 构图优化
[3条建议]

### 总体评价
[2-3句话]

Level 10: 详细 (~4000字 per photo)#

markdown

## [Photo Name]

### 综合评分: X.XX/10 (夯/顶级/人上人/NPC/拉完了)

### 评分解读
[3-4句话]

### 综合分析

#### 构图评价
[6-10详细句话]

#### 色彩评价
[6-10详细句话]

#### 光线评价
[6-10详细句话]

#### 技术质量评价
[6-10详细句话]

### 改进建议

#### 拍摄技巧
[5-7详细条建议]

#### 后期处理
[5-7详细条建议]

#### 构图优化
[5-7详细条建议]

### 总体评价
[3-4段，每段6-8句]

Multiple Photos Comparison (Level 3, ~600字 total)#

markdown

## 照片对比分析

### 照片 1: [Name]
[Level 3 evaluation as above, ~300字]

### 照片 2: [Name]
[Level 3 evaluation as above, ~300字]

## 对比总结

| 对比项 | 照片1 | 照片2 | 胜出 |
|--------|-------|-------|------|
| 综合评分 | X.XX/10 | X.XX/10 | 照片X |
| 构图 | [评级] | [评级] | 照片X |
| 色彩 | [评级] | [评级] | 照片X |
| 光线 | [评级] | [评级] | 照片X |
| 技术质量 | [评级] | [评级] | 照片X |

## 综合建议
[3-4句话]

Score Interpretation Guide#

"从夯到拉" Rating System / "从夯到拉" 评分系统#

Score Range等级 / LevelDescription / 描述9.0-10.0夯 (Hāng)好到没话说,顶级水平 / Exceptional, top-tier, perfect8.0-8.9顶级极好,专业水准 / Excellent, professional level7.0-7.9人上人很好,超越常人 / Very good, above average, outstanding6.0-6.9NPC不起眼,普普通通 / Average, unremarkable, plain0.0-5.9拉完了差到没法再差 / Terrible, needs major improvement

Traditional Rating / 传统评分#

Score RangeLevelDescription9.0-10.0优秀Exceptional quality, professional level8.0-8.9很好High quality with minor improvements needed7.0-7.9良好Solid quality, above average6.0-6.9一般Average quality, noticeable room for improvement4.0-5.9较差Below average, significant improvements needed0.0-3.9很差Poor quality, substantial improvements needed

重要说明: 综合评分格式示例:

text

综合评分: X.XX/10 (夯)
综合评分: X.XX/10 (顶级)
综合评分: X.XX/10 (人上人)
综合评分: X.XX/10 (NPC)
综合评分: X.XX/10 (拉完了)

Error Handling#

If any evaluation source fails:

1. Improved Predictor fails: Use NIMA only
   • Score: NIMA score only
   • Note in report: "Improved Predictor 不可用，仅使用 NIMA 评分"
2. NIMA fails: Use Improved Predictor only
   • Score: Improved Predictor score only
   • Note in report: "NIMA 不可用，仅使用 Improved Predictor 评分"
3. Both sources fail: Inform user and suggest trying again later

Script Dependencies#

All scripts in scripts/ directory must be executable:

• score_improved_predictor.py: Fast aesthetic scoring
• score_nima.py: Detailed quality analysis
• comprehensive_score.py: Integrated weighted scoring

Model Paths (Local Installation)#

Default Paths (Windows)#

ModelLocationImproved Aesthetic Predictor (.pth)F:\software\skill\aesthetic-scorer\models\improved-aesthetic-predictor\sac+logos+ava1-l14-linearMSE.pthNIMA MobileNet weights (.h5)F:\software\skill\aesthetic-scorer\models\neural-image-assessment\weights\mobilenet_weights.h5

Environment Variables (Override Default Paths)#

You can override model paths by setting environment variables:

VariableDescriptionDefaultAESTHETIC_SCORER_MODEL_DIROverride IAP model directoryF:\software\skill\aesthetic-scorer\models\improved-aesthetic-predictor``AESTHETIC_SCORER_NIMA_DIROverride NIMA model directoryF:\software\skill\aesthetic-scorer\models\neural-image-assessment

Python runtime: F:\software\python\python.exe (Python 3.12)

Required packages (install via pip install -r requirements.txt): torch>=2.0.0, torchvision>=0.15.0, transformers>=4.30.0, tensorflow>=2.12.0, tf_keras>=2.12.0, pillow>=10.0.0, numpy>=1.23.0

Usage Examples#

User: "请评价这张照片" Action: Execute both evaluations, calculate weighted score, generate Level 3 evaluation (default, ~300字)

User: "详细评价这张照片" Action: Execute both evaluations, calculate weighted score, generate Level 10 evaluation (~4000字)

User: "简要评价这张照片" Action: Execute both evaluations, calculate weighted score, generate Level 1 evaluation (~200字)

User: "对比这两张照片" Action: Evaluate both photos, generate Level 3 comparison (~600字 total)

User: [评价后] "给我看详细版" Action: Retrieve the saved Level 10 evaluation and present it immediately

Notes#

• Always execute both evaluation sources when available
• Present evaluation as unified expert opinion
• Default to Level 3 (medium detail) unless user specifies otherwise
• Always generate and save Level 10 (detailed) evaluation in background
• Retrieve saved detailed evaluation when requested, don't regenerate
• Avoid repetitive references to evaluation sources
• Use natural, flowing language
• Adjust detail level based on user request
• Support both Chinese and English
• Always display "从夯到拉" rating level in the score