Students Involved:
Tran Dinh Khoi - 26511482 - Computer Science - Lead Author
Bui Viet Dat - 20521162 - Computer Science - Co-Author
Supervisor: Dr. Luong Ngoc Hoang
Paper Summary:
The ability to access models that generate images from text (Text-to-image generation) has been increasing in recent years. These demands finding appropriate prompts to create high-quality images that satisfy user criteria. A prompt is a specific text input created by users to instruct image generation systems. However, creating suitable prompts manually remains a significant challenge. Currently, there are automated methods using Evolutionary Algorithms to develop a population of prompts over generations until suitable prompts are found to generate desired images. The EvoGen framework, utilizing Genetic Algorithms, is an evolving framework to optimize highly anticipated prompts. However, the results obtained have lacked consistency. In this paper, we implement an alternative Genetic Algorithm setup for EvoGen. This implementation protects exceptional individuals, ensuring that valuable prompts are not accidentally eliminated due to the randomness inherent in genetic algorithms. Additionally, we introduce a new loss function, the cosine loss function, to achieve faster convergence and better guide image generation.
We would like to express our gratitude to Dr. Luong Ngoc Hoang, Ph.D. in Computer Science at University of Information Technology, for his dedicated guidance and for pointing out our limitations during our research and the publication of this international scientific paper.
The MAPR 2023 Conference (6-th International Conference on Multimedia Analysis and Pattern Recognition) is the 6th international conference on multimedia analysis and pattern recognition.
The conference serves as a scientific forum for academics, researchers, both national and international, to exchange experiences. It encourages Ph.D. students, postgraduate students, and young scientists to report and exchange their research results and learn from others, especially those with practical applications.
The conference organizers have received submissions from scientists. The presentations focus on areas such as recognition and machine learning, multimedia content analysis, biometrics and medical image analysis, computer vision and Robotics, text analysis and recognition, and other related applications.
To promote and develop research activities at University of Information Technology, the conference organizers invite experienced researchers, professors, and lecturers from prestigious universities, renowned research institutions, and powerful companies, both domestic and international, within related fields.
Detailed Information:
Hai Bang - Communications Collaborator, University of Information Technology
English version: Phan Huy Hoang
