Realism Assessment of AI Generated Visual Content

With the unprecedented level of photo-realism attained by modern generative AI models such as Stable Diffusion, it has become possible to use AI generated visual content for entertainment, education, e-commerce, etc. However, generating photorealistic images still involves considerable amount of prompt engineering and hyperparameter tuning, and even the most successful generative models for producing visual content do not guarantee photo-realism under any prompt and model configuration. Thus, a reliable means for eveluating the photo-realism of AI generated content is necessary to determine their suitability for practical use. This research aims to investigate methodologies and metrics for subjectively and objectively evaluating the perceptual realism of AI-generated visual content. By leveraging insights from computer vision, psychology, and human perception studies, we seek to develop robust frameworks that can rank AI generated images in the order of their perceptual realism.

Through the insights developed from systematic experimentation and comparative analysis against human perception benchmarks, this research endeavors to develop a feedback mechanism through which generative AI models can progressively improve their performance in terms of photo-realism by developing an implicit understanding of the factors that contribute to the perceptual realism of visual content. By bridging the gap between technical advancements in AI and the qualitative assessment of visual realism, we aim to establish a foundation for future research and development of AI systems that produce visually compelling and ethically sound synthetic content.