I tried something new this time and it seems like a winner. I chose 4 of the highest quality images taken during the same time period (2008 - 2010). I verified that two images had the subject smiling and two had the subject with a neutral expression. (This ensures that the resultant LoRA model has the ability to create images with smiles or not smiles!) I made sure each image I chose was easily identifiable as the subject (by myself as well as Bing image search.) I then trained using 10 repeat trains per image but only 8 train epochs. This ended up being 320 training steps. However, it came out perfect. Each image I generate is identified by myself as well as Bing (which is notoriously picky) as the subject. (As a tip, I found that the latest version of Photos (Windows 11) has an icon for Bing image search at the bottom which when clicked takes you to search for an immediate result.) Let me know your thoughts.
