Concetti Chiave
The core message of this article is that the representation space around generated images exhibits distinct properties compared to real images, specifically in terms of complexity and vulnerability. The authors propose two novel metrics, anomaly score (AS) and anomaly score for individual images (AS-i), to effectively evaluate generative models and individual generated images based on these properties.
Sintesi
The article presents an extensive investigation into the relationship between the representation space and input space around generated images. The authors first propose two measures related to the presence of unnatural elements within images:
Complexity: Indicates how non-linear the representation space is with respect to linear input changes.
Vulnerability: Captures how easily the extracted feature changes by adversarial input changes.
Based on these measures, the authors introduce a new metric called anomaly score (AS) to evaluate image-generative models in terms of naturalness. AS is the difference of joint distributions of complexity and vulnerability between the sets of reference real images and generated images, quantified by 2D Kolmogorov-Smirnov (KS) statistics.
The authors also propose AS-i (anomaly score for individual images) to assess generated images individually. Through subjective tests, they demonstrate that AS-i outperforms existing methods for image evaluation, such as rarity score and realism score, in terms of alignment with human perception of naturalness.
The experimental results show that the representation space around generated images is less complex and more vulnerable compared to that of real images. The authors validate the effectiveness of their proposed metrics by demonstrating their strong correlation with human judgments on the naturalness of generated images, outperforming the conventional FID metric.
Statistiche
The article does not provide any specific numerical data or statistics to support the key arguments. The authors rely on qualitative observations and comparisons of the distributions of complexity and vulnerability between real and generated images.