The implementation actually uses a single ffmpeg process with pipes for efficient frame extraction, not separate processes per frame as you suggested.