curiositech/wedding-immortalist

Wedding Immortalist

Transform wedding photos and video into an eternal, immersive 3D experience. Create living memories that let couples and guests relive the magic forever.

DECISION POINTS

1. Multi-Space Merge Strategy

Input: Multiple venue spaces detected (ceremony, reception, cocktail hour)
├── < 500 photos per space
│   └── Train unified scene (risk: quality loss, benefit: seamless navigation)
├── 500-1500 photos per space  
│   └── Train per-space with portal transitions (RECOMMENDED)
├── > 1500 photos per space
│   └── Split large spaces into sub-regions to avoid memory overflow
└── Mixed lighting conditions?
    ├── Yes → Separate indoor/outdoor scenes
    └── No → Can merge with lighting adaptation

2. Face Clustering Threshold Selection

Input: Face embedding similarity scores across all detected faces
├── Photo count < 1000
│   └── Conservative threshold (0.4): Prefer precision over recall
├── Photo count 1000-5000
│   └── Balanced threshold (0.3): Standard wedding size
├── Photo count > 5000
│   └── Aggressive threshold (0.25): Large weddings need looser matching
└── Professional + phone photos mixed?
    ├── Yes → Increase threshold by 0.05 (lighting/quality variance)
    └── No → Use base threshold

3. Theme Signal Conflict Resolution

Input: Competing aesthetic signals detected
├── Venue theme ≠ Decor theme
│   ├── Venue rustic, decor modern → Weight decor 70%, venue 30%
│   └── Venue formal, decor casual → Weight venue 60%, decor 40%
├── Multiple cultural traditions
│   └── Create fusion palette prioritizing couple's primary heritage
├── Seasonal vs. chosen colors
│   ├── Spring wedding with fall colors → Honor chosen colors
│   └── Christmas wedding ignoring season → Blend seasonal + chosen
└── Guest attire contradicts theme
    └── Ignore guest attire, focus on venue + intentional decor

4. Video Frame Extraction Rate

Input: Wedding video files with metadata
├── Handheld/phone video (high motion blur)
│   └── Extract 1 fps to reduce blur frames
├── Professional video (smooth)
│   └── Extract 2-3 fps for optimal overlap
├── Drone footage
│   └── Extract 1 fps (smooth motion, redundant frames)
├── Ceremony (mostly static)
│   └── Extract 1 fps, boost for processional/recessional
└── Reception/dancing (high motion)
    └── Extract 3 fps to capture movement variety

5. Space Skip Threshold Decision

Input: Reconstructed space quality metrics
├── < 50 images AND reconstruction error > 5.0 pixels
│   └── Skip space: insufficient coverage, warn user
├── 50-100 images AND error > 3.0 pixels
│   └── Include with quality warning, mark as "preview quality"
├── > 100 images AND error > 2.0 pixels
│   └── Retrain with adjusted parameters, likely lighting issues
└── Blurry photos > 40% of space images
    └── Skip space, request better photos from user

FAILURE MODES

1. COLMAP Divergence Cascade

Symptom: Sparse reconstruction fails completely, no 3D points generated Detection Rule: If COLMAP outputs < 100 3D points for > 200 input images Diagnosis: Feature matching failure due to repetitive textures (white walls, flowers) or extreme lighting changes Fix:

• Re-extract features with lower max_image_size (1600px)
• Switch to sequential matcher instead of exhaustive
• Filter images by Laplacian variance > 150 to remove blur

2. Face-Cluster False Merge Explosion

Symptom: Single cluster contains 3+ distinct people, especially bridesmaids in similar makeup Detection Rule: If cluster size > 30 photos AND intra-cluster cosine distance variance > 0.15 Diagnosis: Similar makeup, lighting, or formal poses confusing embeddings Fix:

• Reduce clustering threshold by 0.05
• Apply makeup-invariant embedding model (AdaFace instead of ArcFace)
• Manual split using user confirmation of cluster samples

3. Theme Signal Chaos

Symptom: UI colors clash horribly, typography doesn't match aesthetic Detection Rule: If extracted color palette has > 8 dominant colors OR no single color > 15% dominance Diagnosis: Too many competing signals from mixed lighting, decorations, clothing Fix:

• Focus extraction on venue + decor photos only, ignore people
• Apply k-means clustering to consolidate palette to 4-5 colors
• Fallback to safe neutral palette with couple's stated preferences

4. Memory Overflow Death Spiral

Symptom: 3DGS training crashes, viewer freezes on load Detection Rule: If trained .ply file > 2GB OR browser memory usage > 8GB Diagnosis: Too many Gaussian points from over-densification or merged scenes Fix:

• Split scene into smaller spatial regions
• Reduce densify_grad_threshold from 0.0002 to 0.0005
• Implement LOD (level-of-detail) rendering with quality presets

5. Guest Photo Desert

Symptom: Many face clusters have only 1-2 poor-quality photos Detection Rule: If > 30% of clusters have < 3 photos OR average aesthetic score < 0.4 Diagnosis: Insufficient photo coverage or professional photographer focused only on couple Fix:

• Request guest phone photos to fill gaps
• Lower aesthetic threshold for rare-appearance guests
• Generate "best available" galleries with quality disclaimers

WORKED EXAMPLES

Example 1: Large Disco Wedding (500 guests, 3000 photos)

Scenario: 1970s theme revival, disco balls, gold/orange/purple decor, mixed lighting (strobes + warm), professional + 50 guest phones

Step-by-step walkthrough:

1. Initial Assessment: 3000 photos, detect 4 spaces (ceremony, cocktail, reception, photo booth)

   • Decision Point 1 triggered: > 500 photos per space → train separately
   • Space distribution: ceremony (800), cocktail (400), reception (1500), photo booth (300)

2. Theme Extraction:

   • Color analysis finds: gold (22%), burnt orange (18%), deep purple (15%), brown (12%)
   • Era signals: disco ball reflections, wide lapels in photos, platform shoes
   • Typography match: groovy script + bold sans serif
   • Decision Point 3: Theme signals align → proceed with disco aesthetic

3. Face Clustering:

   • 2847 faces detected across all photos
   • Decision Point 2: Large wedding + mixed photo sources → threshold 0.25 + 0.05 = 0.3
   • Result: 287 unique identities (realistic for 500-guest wedding)

4. 3DGS Training:

   • Decision Point 4: Professional video → extract at 2 fps
   • Ceremony space training completes successfully
   • Reception space hits Decision Point 5: > 1500 images triggers sub-region split
   • Split reception into: dance floor, dining area, bar area

5. Quality Gates Check:

   • ✓ All spaces reconstructed with < 2.0 pixel error
   • ✓ Face clusters manually verified for couple + wedding party
   • ✓ Theatre mode markers placed at 12 key moments
   • ✓ UI theme matches gold/orange/purple palette
   • ✓ Guest galleries generated for 287 identities

Expert catches vs. Novice misses:

• Expert: Notices strobing lights will confuse COLMAP, filters those frames before processing
• Novice: Includes all frames, gets reconstruction failure, doesn't understand why
• Expert: Recognizes disco theme needs bold, groovy typography—not elegant script
• Novice: Applies generic wedding fonts, loses the personality

Example 2: Intimate Rustic Ceremony (50 guests, 400 photos)

Scenario: Barn venue, earth tones, string lights, sage/cream decor, single professional photographer

Key decision differences:

• Small photo count triggers Decision Point 1: unified scene approach
• Conservative face clustering (0.4 threshold) due to limited photos per person
• Earth-tone palette extraction straightforward (no conflicts)
• Single lighting condition simplifies reconstruction

Trade-off analysis:

• Benefit: Seamless navigation, no portal breaks
• Risk: Lower overall quality due to unified training
• Mitigation: Accept risk for small venue—user experience more important

QUALITY GATES

• [ ] All venue spaces reconstructed with COLMAP error < 3.0 pixels
• [ ] Face clustering achieves > 95% precision on couple + wedding party (verified manually)
• [ ] Theatre mode includes minimum 8 key moments with video/audio sync
• [ ] Color palette extraction yields cohesive 4-6 colors matching wedding aesthetic
• [ ] Best photo selection provides minimum 3 high-quality images per guest (where possible)
• [ ] Web viewer loads all scenes within 15 seconds on standard connection
• [ ] Guest roster correctly identifies minimum 90% of wedding party members
• [ ] UI theme visually matches extracted wedding aesthetic (validated by user)
• [ ] 3DGS file sizes total < 5GB for reasonable hosting/sharing
• [ ] All key moments playback smoothly without audio desync > 100ms

NOT-FOR BOUNDARIES

This skill should NOT be used for:

• General photo editing → Use native-app-designer instead for basic editing, filters, adjustments
• Non-wedding 3DGS projects → Use drone-inspection-specialist for architectural/industrial 3D scanning
• Wedding planning or vendor coordination → This is not a wedding planner, only processes existing photos/video
• Real-time event streaming → Use dedicated streaming solutions for live broadcast
• Professional photography instruction → Use photography education skills for technique training
• Video editing without 3D reconstruction → Use standard video editing skills for traditional editing needs

Delegate to other skills when:

• User wants basic photo slideshow → collage-layout-expert
• Need color palette advice before wedding → color-theory-palette-harmony-expert
• Want to improve individual photos → photo-composition-critic
• Need facial recognition for security → face-detection-specialist