| abstract |
| Background Generative AI technology is rapidly expanding its application in the field of image production. In particular, platforms such as DALL·E 3 and Midjourney are capable of generating high-resolution human portraits. However, there remains a lack of standardized criteria to quantitatively evaluate users’ perceived realism of these images. This study aims to analyze how visual attributes in AI-generated portraits contribute to users’ perception of realism.
Methods Eight portrait images were generated using identical prompts across two platforms. Based on the visual attribute framework proposed by Fan et al. (2018), a survey consisting of ten items (Q1-Q10) was developed. A total of 106 valid responses were collected. The analysis focused on image-wise mean scores, comparisons between platforms, and correlation analysis among visual attributes.
Result Midjourney images received overall higher evaluation scores. Visual attributes such as facial expression, photographic quality, and compositional layout exhibited strong positive correlations with perceptual realism. These findings suggest that realism is formed through a structural interplay among multiple visual attributes.
Conclusion This study provides foundational data for establishing a perceptual realism evaluation framework for AI-generated images and offers meaningful implications for assessing image quality and designing prompt-based generation strategies. |
|
|
| Key Words |
| 생성형 인공지능, 시각 속성, 실제감, Generative AI, Image Attributes, Perceptual Realism |
|
|
|
 |
|
Full Text