An interesting experiment. It's noticeable that the neural network was able to recognize hairstyle, eyebrows, and partially eyes. Using SD 1.5 was the right decision, as it consumes fewer resources. Also, don't forget to use Gemini or another AI to help you configure Kohya and Stable Diffusion. My first attempts were on an RTX 2060s, and training LoRa took no more than 2 hours on SD 1.5 512px.