Recreate Thumbnails

Recreate YouTube Thumbnails with Nick Saraev

Goal

Face-swap YouTube thumbnails to feature Nick Saraev using Nano Banana Pro (Gemini image model). The system:

• Analyzes face direction (yaw/pitch) in the source thumbnail
• Finds the best-matching reference photo of Nick by pose
• Generates 3 variations by default
• Supports iterative edit passes for refinements

Quick Start

# Generate 3 variations from a YouTube video
python ~/.claude/skills/_youtube-execution/recreate_thumbnails.py --youtube "https://youtube.com/watch?v=VIDEO_ID"

# Generate from a local thumbnail
python ~/.claude/skills/_youtube-execution/recreate_thumbnails.py --source ".tmp/thumbnails/source_yt_thumb.jpg"

# Edit pass on a generated thumbnail
python ~/.claude/skills/_youtube-execution/recreate_thumbnails.py --edit ".tmp/thumbnails/recreated_v3.png" \
  --prompt "Change colors to teal. Change 'AI GOLD RUSH' to 'AGENTIC FLOWS'."

Full Workflow

Step 2: Generate Thumbnails

# From YouTube URL (auto-analyzes face, finds best reference, generates 3 variations)
python ~/.claude/skills/_youtube-execution/recreate_thumbnails.py --youtube "https://youtube.com/watch?v=VIDEO_ID"

# From local image
python ~/.claude/skills/_youtube-execution/recreate_thumbnails.py --source "path/to/thumbnail.jpg"

# Custom variation count
python ~/.claude/skills/_youtube-execution/recreate_thumbnails.py --source "thumbnail.jpg" -n 5

# Skip direction matching (use default references)
python ~/.claude/skills/_youtube-execution/recreate_thumbnails.py --source "thumbnail.jpg" --no-match

What happens:

1. Downloads/loads source thumbnail
2. Analyzes face direction using MediaPipe (yaw/pitch in degrees)
3. Finds best-matching reference by Euclidean distance in pose space
4. Loads 2 reference photos (best match + one other)
5. Sends to Gemini with face swap prompt
6. Generates 3 variations (results vary between runs)

Step 3: Select Best & Edit

Pick your favorite variation, then refine with edit passes:

# Single edit
python ~/.claude/skills/_youtube-execution/recreate_thumbnails.py --edit ".tmp/thumbnails/recreated_v3.png" \
  --prompt "Change colors to teal brand colors. Change 'AI GOLD RUSH' to 'AGENTIC FLOWS'."

# Chain multiple edits (each builds on the previous)
python ~/.claude/skills/_youtube-execution/recreate_thumbnails.py --edit ".tmp/thumbnails/edited_1.png" \
  --prompt "Change background to pure white. Make the graph show two bounces instead of one."

python ~/.claude/skills/_youtube-execution/recreate_thumbnails.py --edit ".tmp/thumbnails/edited_2.png" \
  --prompt "Make 'AGENTIC FLOWS' text bigger. Change 'The Moment' to 'its happening'."

Edit pass tips:

• Be specific about what to change
• Reference exact text strings when changing copy
• Colors, text, graphs, backgrounds all work well
• Each edit is a separate API call (~$0.14-0.24)

File Locations

Output Organization

Thumbnails are stored in date-based folders:

.tmp/thumbnails/
├── 20251205/              # December 5, 2025
│   ├── 104016_1.png       # Generated at 10:40:16, variation 1
│   ├── 104016_2.png       # Generated at 10:40:16, variation 2
│   ├── 104016_3.png       # Generated at 10:40:16, variation 3
│   ├── 104532_edited.png  # Edit pass at 10:45:32
│   └── 110000_1.png       # Another run at 11:00:00
└── 20251206/              # December 6, 2025
    └── 090000_1.png

• Same-day runs go to the same folder
• New day creates a new folder
• Files named with HHMMSS timestamp + variation number (or _edited for edit passes)

recreate_thumbnails.py

analyze_face_directions.py

API Notes

• Model: gemini-3-pro-image-preview (Nano Banana Pro)
• Cost: ~$0.14-0.24 per generation/edit
• Latency: 10-60+ seconds per image
• Output resolution: ~1376x768 (close to 16:9, model rounds slightly)
• API key: NANO_BANANA_API_KEY in .env

Naming Convention

Reference photos use this format:

nick_yaw{L/R}{degrees}_pitch{U/D}{degrees}.jpg

Examples:

• nick_yawL30_pitchU10.jpg — looking 30° left, 10° up
• nick_yawR45_pitch0.jpg — looking 45° right, level
• nick_yaw0_pitch0.jpg — dead center

Learnings

• Face direction matching reduces uncanny valley — using a reference with similar pose to the target dramatically improves results
• Uses native 1280x720 input resolution for crisp output on large displays
• 2 reference photos is optimal — 1 loses likeness, 3+ can cause full image regeneration
• Must explicitly request 16:9 format in prompt or model outputs squares
• Label images explicitly in prompt ("IMAGE 1: Reference photo, IMAGE 2: Thumbnail") to prevent confusion
• Keep prompts simple and direct — complex instructions cause regeneration instead of editing
• Generate 3 variations by default since results vary between runs
• "100% exact duplicate except face" instruction works well for pure face swaps
• Edit passes work well for text, colors, graphs, backgrounds
• Multiple edit passes can be chained (each builds on previous output)

Troubleshooting

Full Specification

Complete details, decision trees, protocols, and implementation specs: references/full-details.md