3 comments

  • eth_hack77 10 hours ago
    Sounds interesting! Do you train your image model / customized existing one or what is the source for the thumbnail images?
    • gits997 9 hours ago
      No, I use existing model(s) as a base along with LLMs to: 1. enhance the prompt and add technical details that are important for high quality thumbnails (e.g. high contrast image) 2. validate the generated images and choose the best one.
  • nbbaier 1 day ago
    The example you give is very clearly AI generated. Has there been any work done on how different such thumbnails behave vs. clearly non-AI thumbnails?
    • gits997 12 hours ago
      Thanks for the comment. Appreciate the interest.

      I did the experiment for my fantasy youtube channel (https://www.youtube.com/@AIBookBytes) and CTR increased from 1.2% to 2% which for me was a huge deal.

      This is probably not the tool for super high subscriber youtube channels but from amateurs to mid size youtube channels could be very helpful. Obviousy it is clear that it's AI generated by many channels use that and still are successful. As long as expressions are clear and text is bold and visible, should be good.