Understanding AI-Generated Food Images Through Aesthetic Cues and Cross-Dataset Evaluation

The increasing realism of AI-generated food images raises important concerns regarding authenticity, transparency, and trust in food-related visual content. Although modern vision models can detect synthetic food images with very high accuracy, considerably less is known about the visual properties that distinguish AI-generated food imagery from real photographs and whether these properties generalize across different datasets. In this work, we investigate handcrafted aesthetic and photographic descriptors as interpretable cues for understanding real-vs-AI differences in food images. Experiments are conducted on FoodFake-30K using descriptors related to color, texture, sharpness, frequency content, and composition, and their discriminative ability is compared with foundation-model baselines. To assess robustness beyond the training distribution, we further evaluate the identified cues on the external Gourmet Photography Dataset, which contains only real food photographs. Handcrafted descriptors alone provide a strong discriminative signal, achieving up to 88.63% classification accuracy and 95.34% AUROC. The largest differences are associated with edge density, local contrast, sharpness, grayscale entropy, and high-frequency energy, indicating that AI-generated food images tend to be visually denser and more locally enhanced than real photographs. However, these descriptors exhibit substantial distribution shifts on the external dataset, causing many real photographs to resemble synthetic images in the handcrafted feature space. In contrast, a fine-tuned vision transformer generalizes robustly and exhibits a negligible false-positive rate on the external dataset. These findings suggest that handcrafted aesthetic and photographic descriptors provide an interpretable perspective on the visual signatures of AI-generated food images but should primarily be regarded as an explainability layer rather than as portable authenticity detectors.