Thumbnail Generator

Thumbnail Generator — Face Swap + Style Variations

Goal

Take inspiration thumbnails from creators you like, generate new thumbnails in the same style but with your face, then apply stochastic modifications (background colors, text changes, styling tweaks). Produces 3 variations by default, each intentionally different.

Inputs

• Source thumbnails: YouTube URLs, direct image URLs, or local image paths
• Face reference: face.png in .tmp/thumbnails/face/ — this is the only reference photo. Always use --refs 1
• Modifications (optional): Additional instructions like "replace the logo with X" or "change background to dark blue"

Process

Step 1: Download & Inspect Source Thumbnail

Download the source thumbnail and visually inspect it yourself using the Read tool before generating anything.

# From YouTube URL — use --source with the direct image URL (faster, avoids video ID parsing)
python3 execution/recreate_thumbnails.py --source "https://i.ytimg.com/vi/VIDEO_ID/maxresdefault.jpg" --no-match --refs 1 --variations 1

# From local file
python3 execution/recreate_thumbnails.py --source "path/to/thumbnail.jpg" --no-match --refs 1 --variations 1

Step 2: Visual Analysis (ALWAYS DO THIS)

After inspecting the source thumbnail, identify 3-4 minor visual elements that can be subtly varied without changing the thumbnail's overall look. Examples:

• Background color/tint (white → warm cream, cool gray, light blue)
• Arrow or accent color (red → orange, gold, deeper red)
• Text shadow weight or outline thickness
• Logo color shift (subtle hue change)
• Lighting warmth (cool vs warm white)

Keep variations minor — the thumbnail should still look like the same thumbnail at a glance. The goal is subtle differentiation, not a redesign.

Step 3: Generate 3 Variations with Micro-Tweaks

Run 3 separate generations in parallel (1 variation each), each with a different micro-tweak from your analysis passed via --prompt. This ensures each output is meaningfully different rather than relying on stochastic model variance alone.

# Variation 1: tweak element A
python3 execution/recreate_thumbnails.py --source "URL" --no-match --refs 1 --variations 1 \
  --prompt "Make the background wall a very subtle warm cream instead of pure white"

# Variation 2: tweak element B
python3 execution/recreate_thumbnails.py --source "URL" --no-match --refs 1 --variations 1 \
  --prompt "Change the red arrow to a gold/yellow arrow"

# Variation 3: tweak element C
python3 execution/recreate_thumbnails.py --source "URL" --no-match --refs 1 --variations 1 \
  --prompt "Add a subtle cool blue tint to the lighting"

Always append the user's own instructions (if any) to each prompt alongside the micro-tweak.

Step 4: Edit Pass (Optional)

Pick your favorite, then refine:

python3 execution/recreate_thumbnails.py --edit ".tmp/thumbnails/20260228/115127_1.png" \
  --prompt "Change background to teal. Replace 'AI GOLD RUSH' with 'AGENT SKILLS'. Make text bigger."

Edit passes support:

• Background color changes
• Text replacement and resizing
• Logo swaps
• Graph/chart modifications
• Any visual tweak you can describe

Key Behaviors

• Face + body adaptation: The script adapts skin tone, hair, build, and neck to match Nick — not just a face paste
• Forced 1280x720 output: All outputs are resized to exact YouTube thumbnail dimensions
• Single reference photo: Always use --refs 1. The reference is face.png (clean headshot). Using noisy references (thumbnails with text overlays, etc.) degrades quality
• Always use --no-match: Face direction matching requires specially named files (nick_yawL30_pitchU10.jpg). We don't have those, so always skip with --no-match

File Locations

Environment

Requires in .env:

NANO_BANANA_API_KEY=your_gemini_api_key

Cost

• ~$0.14-0.24 per generation/edit
• 3 parallel variations = ~$0.50 total
• Edit passes are one image at a time = ~$0.14-0.24

Tips

• 1 clean reference photo beats 2 where one is noisy
• Run all 3 variations in parallel (separate Bash calls) for speed
• Keep micro-tweak prompts simple and direct
• Reference exact text strings when changing copy

First-Run Setup

Before executing, check if the workspace has a .gitignore file. If it doesn't, assume the user is new to this skill. In that case:

1. Ask the user if this is their first time running this skill
2. If yes, walk them through how it works and what they need to configure/set up (API keys, env vars, dependencies, etc.)
3. Let them know that Nick wishes them the best!