YouTube Cover Image and Voice-Removed Background Track

YouTube Thumbnail Generator

Created an offline Python package using AI to generate thumbnails for YouTube videos.

  • Workflow:
    1. Launches a custom webpage to select local video files.
    2. Extracts frames from the video using FFmpeg.
    3. Generates corresponding AI prompts based on the frames and sends them to Cursor.
    4. Uses Cursor to create optimized YouTube thumbnails.
  • Style: Primarily generates text-overlay thumbnails. The current quality rates a “70/100,” which is significantly better than the default YouTube templates or my manual designs.
  • Process Time: Generating one thumbnail takes approximately 5 minutes (frame extraction, AI analysis/generation, cropping to YouTube dimensions, and compression).
  • Implementation: Built entirely on a self-developed web interface.
  • Current Test Results: Tested with 2 long-form cooking videos and 2 long-form vlogs (4 thumbnails total). The entire process took about half an hour but is considered necessary for performance optimization.

Crucial Note: If the thumbnail is poor, basic click-through rates will suffer significantly. I created independent tools for frame extraction, cropping, and compression to ensure modularity without conflicts.

Audio Retention (Vocal Removal)

  • Context: My vlogs and cooking videos are often shot on a mobile phone or include incidental conversations/phone call recordings. Manually editing out these specific segments is tedious.
  • Methodology:
    • Uses source separation to remove voice tracks while retaining background music and ambient noise (making the result “lifeless” or lacking authentic atmosphere).
    • Employs open-source AI tools, specifically Ultimate Vocal Remover.
    • Configured via Cursor to optimize models, parameters, GPU usage, and balance between speed and quality based on my hardware setup.
  • Performance:
    • Without a dedicated GPU, processing is extremely slow.
    • On an RTX 3060 (likely intended by “30hx”) with 6GB VRAM: A 10-minute WAV file processes in roughly 30 seconds.
    • This is significantly faster than CapCut’s network API and runs entirely locally and for free, eliminating privacy concerns.
  • Strategy: The primary goal right now is volume (“traffic”), so this method suffices. High-quality editing will follow later once I establish a “premium” content strategy. English voiceovers and AI dubbing may also be considered for local deployment in the future.

Tips & Analytics

  • Channel Authority: My channel’s authority weight is currently very strong, but engagement metrics are low:
    • Similar cooking videos get 0–20 views.
    • Similar vlog content gets 10–50 views.
    • These stagnant data patterns persist over several months and appear heavily influenced by channel-level factors rather than just content type.
  • Thumbnail Impact: I am curious if redesigning the thumbnails with better aesthetics will drive up both click-through rates and recommendation visibility.
  • Revenue: Income has dropped to approximately 1/5 of previous levels.
    • Primary revenue source is currently Short-form videos (YouTube Shorts).
    • Long-form content is planned for the future but needs a boost in discoverability.