ASSESSMENT OF AI-GENERATED IMAGES USING COMPUTATIONAL METRICS AND HUMAN CENTRIC ANALYSIS
Abstract
The rapid advancements in AI-generated models for image generation have transformed image creation from entertainment to e-commerce, making robust evaluation essential. We designed and statistically validated a human study for assessing the photorealism and image quality of AI-generated images. The study indicated that camera images show more realism, while AI images, such as those generated by DALL-E2, excel in quality. Further analysis showed that existing metrics deviate from human judgments. To address this, we developed the Global-Local Image Perceptual Score (GLIPS), which evaluates photorealistic quality using Vision Transformer-based attention for local similarity and Maximum Mean Discrepancy (MMD) for global distributional similarity. Our evaluation showed that GLIPS aligns more closely with human perception compared to traditional metrics like FID and SSIM for photorealism, although MS-SSIM outperformed in image quality. Additionally, we introduced the Interpolative Binning Scale (IBS) for human-assisted metric scaling with human scores.