Generative AI in Artistic Style Transfer: Performance, Perception, and Evaluation

This thesis critically examines the intersection of generative AI and artistic style transfer, assessing its technological foundations, aesthetic implications, and avenues for further refinement. With rapid advancements in generative models, such as diffusion-based architectures and GANs, AI-generated artwork has reached remarkable levels of sophistication. However, significant challenges persist, including authenticity, compositional balance, and adherence to artistic conventions. A primary focus of this research is the evaluation of generative models in their ability to replicate artistic styles while maintaining structural integrity and stylistic fidelity. Through a systematic comparative analysis, this study explores how different AI models perform in generating artworks that align with human artistic expectations. The research also includes a large-scale user perception survey to assess whether AI-generated images can convincingly mimic human-created art. A major point of this work is the contribution to the AI-Pastiche Dataset, a structured collection of AI-generated images labeled across multiple criteria, including artistic style, subject matter, and historical period. This dataset serves as a benchmark for evaluating the strengths and weaknesses of various generative models. The study also highlights common distortions in AI-generated art, such as anatomical inaccuracies, hyperrealistic elements, and stylistic inconsistencies. Findings suggest that while AI can effectively replicate impressionistic and abstract styles, it struggles with intricate historical styles that require greater anatomical precision and nuanced brushwork. This research builds on the findings presented in the article “A Critical Assessment of Modern Generative Models’ Ability to Replicate Artistic Styles” which provides an in-depth evaluation of AI-driven style replication.