Approximately 5% to 10% of children globally are affected by developmental delays, resulting in deficits in language comprehension, motor skills, and daily self-care abilities. Visual perception deficiencies are particularly common and significantly impact children's learning and daily activities. Traditional visual perception training materials often lack personalization and diversity, failing to engage children and meet their unique needs. This study applies generative AI technology to create personalized visual training materials and evaluates their effectiveness in enhancing visual perception abilities in children aged 3 to 5 years. The study involved 12 children, utilizing GazePoint 3 eye-tracking equipment to record data and the Children Participation Questionnaire (CPQ) to gather feedback. Results indicated that AI-generated materials effectively captured children's attention and improved information processing efficiency. Although 4 of the 12 participants could not complete the tasks, the remaining 8 participants had an average task completion time of 10.11 s, an average fixation count of 5 times, an average fixation duration of 2.53 s, a time to first fixation of 1.2 s, and an average fixation duration percentage of 75%. The CPQ survey showed high approval from both guardians and trainers for this new training method. Future research should explore generative AI's application for children with various developmental disorders, aiding in creating more complex therapeutic plans and developing specialized training programs and toolkits for therapists. This study highlights the substantial potential of generative AI in developing personalized, interactive therapeutic materials, promising more effective visual perception training programs to support the comprehensive development of children with developmental delays.