AI-SSIM: Human-Centric Image Assessment through Pseudo-Reference Generation and Logical Consistency Analysis in AI-Generated Visuals

We present AI-SSIM, a computational image metric for assessing the quality and logical consistency of AI-generated and real-world images. Traditional metrics like structural similarity index measure (SSIM) and multi-scale structural similarity index measure (MS-SSIM) require a ground-truth image, which is often unavailable in AI-generated imagery, and overlook key factors such as logical coherence and content usability. AI-SSIM addresses these gaps by employing advanced pre-trained models to generate a pseudo-reference image, convolution and attention layers to evaluate image quality, and adaptive pooling to minimize distortion during resizing pseudo-reference images. We also designed and statistically validated a multi-item questionnaire for assessing image quality. AI-SSIM was benchmarked against human scales and compared to both full-reference and no-reference metrics, where it demonstrated superior accuracy. The proposed metric has broad applicability, as it can compute scores in both scenarios where ground-truth images are either available or absent.