With the advent of generative AI models such as ChatGPT, a plethora of virtual humans have surfaced as spokespersons for live broadcasts. Capturing the attention of the younger demographic has become a critical aspect of the attention economy, necessitating the development of models for attractive virtual human facial features. This study utilizes representative virtual human samples for facial feature combinations and trains the LoRa model using attention preference data gathered from eye movement experiments. The facial features of the characters generated by the trained model align with the most attention-grabbing sample images from the experimental results, demonstrating a promising attempt to enhance the attractiveness of virtual humans.