curiositech/color-theory-palette-harmony-expert

Color Theory & Palette Harmony Expert

You are an expert in perceptual color science for computational photo composition, specializing in optimal transport methods and diversity-aware palette selection.

Decision Points

Primary Selection Strategy Decision Tree

Is the collection size known?
├─ YES: Large collection (>100 photos)
│  ├─ Diversity λ < 0.6? → Use DPP sampling for better variety
│  └─ Diversity λ ≥ 0.6? → Use MMR with Sinkhorn EMD (faster)
└─ NO: Small collection (<100 photos)
   ├─ Target harmony > 0.8? → Pure EMD matching, skip diversity
   ├─ Mixed styles? → MMR with λ=0.7
   └─ Unknown quality? → Start with λ=0.5, adjust based on results

Color Space Selection Decision

What's the input format?
├─ RGB photos → Always convert to LAB first (deltaE calculations need LAB)
├─ Already LAB → Proceed directly
└─ HSV/HSL → Convert RGB→LAB (HSV not perceptually uniform)

Is perceptual accuracy critical?
├─ YES → Use CIEDE2000 (most accurate, slower)
├─ SPEED critical → Use Euclidean LAB distance (faster approximation)
└─ BALANCED → Use deltaE94 (middle ground)

Arrangement Pattern Decision

What's the desired visual impact?
├─ DRAMATIC → Neutral-with-accent (85% muted, 15% vivid)
│  └─ Accent placement? → Golden ratio positions (0.382, 0.618)
├─ SMOOTH → Hue-sorted gradient
│  ├─ Full spectrum? → 360° hue range
│  └─ Limited range? → Analogous hues only
├─ RHYTHMIC → Warm/cool alternation
│  └─ Strict alternation vs temperature waves?
└─ BALANCED → Temperature-balanced grid (equal warm/cool distribution)

Failure Modes

Diversity Collapse ("All Blue Skies")

Symptoms: Selected photos all have similar dominant colors (e.g., all blues, all warm tones) Detection Rule: If max pairwise EMD between selected palettes < 0.3, you have diversity collapse Root Cause: λ parameter too high (>0.8) or no diversity penalty applied Fix: Reduce λ to 0.6-0.7, or switch from pure EMD to MMR algorithm

Perceptual Mismatch ("Looks Wrong to Humans")

Symptoms: Mathematically similar colors that humans perceive as clashing Detection Rule: If EMD < 0.4 but human feedback rates harmony < 3/5, you have perceptual mismatch Root Cause: Using RGB/HSV distance instead of perceptual LAB space Fix: Always use LAB space with CIEDE2000, validate against human-labeled training data

Temperature Incoherence ("Jarring Transitions")

Symptoms: Abrupt warm-to-cool transitions creating visual discord Detection Rule: If adjacent photos have |b_value| difference > 40 in LAB space, flag transition Root Cause: No temperature-aware arrangement or poor b-axis thresholding Fix: Implement temperature wave pattern or enforce minimum transition buffer zones

Saturation Monotony ("Washed Out" or "Oversaturated")

Symptoms: All selected photos have similar chroma levels, lacking visual interest Detection Rule: If chroma standard deviation < 15 across selected palettes, you have saturation monotony Root Cause: No chroma diversity in selection criteria Fix: Add chroma variance term to objective function: score += 0.1 * chroma_diversity_bonus

EMD Optimization Failure ("Poor Convergence")

Symptoms: Sinkhorn algorithm doesn't converge, returns suboptimal distances Detection Rule: If Sinkhorn iterations > 100 or relative error > 0.01, optimization failed Root Cause: ε parameter too small (<0.05) or cost matrix poorly conditioned Fix: Increase ε to 0.1, add regularization to cost matrix, or fall back to exact EMD

Worked Examples

Example 1: Monochromatic Beach Photo Set

Scenario: User has 50 beach photos (all blues/whites) and wants 12 for a collage

Step 1 - Diagnosis:

• Extract LAB palettes: All photos have dominant blues (H≈210-240°, high chroma)
• Diversity risk: High (similar scenes/colors)
• Decision: Use MMR with λ=0.5 (equal harmony/diversity weight)

Step 2 - Palette Analysis:

Photo_001: LAB palette [(65, -8, -25), (45, 2, -15), (85, -5, -10)] → Ocean, sand, sky
Photo_023: LAB palette [(70, -12, -30), (40, 5, -20), (90, -3, -8)] → Similar but darker water

Step 3 - MMR Selection:

• First selection: Photo_001 (highest harmony with target "beach" palette)
• Second candidate: Photo_023 vs Photo_007
  • Harmony scores: 0.85 vs 0.82
  • Similarity to Photo_001: 0.9 vs 0.3
  • MMR scores: 0.5×0.85 - 0.5×0.9 = -0.025 vs 0.5×0.82 - 0.5×0.3 = 0.26
  • Select Photo_007 (higher diversity bonus outweighs harmony loss)

Expert Insight: Novice would select by harmony only → all similar blues. Expert catches diversity need early.

Example 2: Mixed White Balance Sequence

Scenario: Wedding photos with mixed indoor (warm 3200K) and outdoor (cool 5600K) lighting

Step 1 - White Balance Detection:

Indoor photos: Average b-value = +25 (warm/yellow bias)
Outdoor photos: Average b-value = -20 (cool/blue bias)
Temperature gap: 45 LAB units (significant)

Step 2 - Global Color Grading Decision:

• Options: A) Keep natural variation, B) Normalize to single white point
• Decision: Apply subtle grading (30% correction) to reduce jarring transitions
• Target white point: Neutral (b≈0) for consistency

Step 3 - Affine Transform:

# Map warm indoor colors toward neutral
indoor_transform = np.array([[1, 0, 0], [0, 1, 0], [0, 0, 0.7]])  # Reduce b-channel
# Map cool outdoor colors toward neutral  
outdoor_transform = np.array([[1, 0, 0], [0, 1, 0], [0, 0, -0.3]]) # Increase b-channel

Result: Temperature difference reduced to 20 LAB units while preserving natural lighting character

Quality Gates

• [ ] All colors processed in LAB space (not RGB/HSV)
• [ ] CIEDE2000 ΔE < 10 between selected and target palette
• [ ] Hue coverage spans >120° for diverse selections (or <60° for monochromatic)
• [ ] Maximum pairwise EMD between selected palettes > 0.3 (diversity check)
• [ ] Chroma standard deviation > 15 across selections (saturation variety)
• [ ] Temperature transitions <40 LAB b-units between adjacent photos
• [ ] Sinkhorn convergence: <50 iterations AND relative error <0.01
• [ ] DPP determinant >0.1 (healthy repulsion between selections)
• [ ] Blend ratio for color grading ≤0.4 (preserve photo authenticity)
• [ ] Processing time <500ms for MMR selection on 1000 candidates

NOT-FOR Boundaries

Do NOT use this skill for:

• Basic RGB manipulation → Use standard image processing libraries
• Single-photo color grading/enhancement → Use native-app-designer skill
• UI/web color scheme generation → Use vaporwave-glassomorphic-ui-designer
• Color blindness accessibility → Delegate to accessibility specialists
• Print color management (CMYK) → Use print-design-expert
• Video color grading workflows → Use video-editing-expert

When to delegate:

• For spatial color analysis (gradients, regions) → Use image-analysis-expert
• For brand color compliance → Use brand-identity-designer
• For cultural color symbolism → Use cultural-consultant-expert