What is Real Anymore? A Solution to Detect Hyper Realistic AI-Generated Imagery

Building on the WIRA image dataset, this work introduces the Symbiote Particle, an enhanced dual stream CNN model architecture designed to detect AI-generated patterns when used to train on distinct image categories. Additionally, a methodology is proposed to find a single hyperparameter configuration that helps the Symbiote Particle generalize well not only on one WIRA image category, but multiple image categories. Across all 62 WIRA image categories, the modelstrategy configurations collectively achieved a mean accuracy of 65%, mean precision of 68%, mean recall of 77%, and mean F1 score of 66%. These configurations collectively achieved a mean accuracy of 73%, mean precision of 73%, mean recall of 88%, and a mean$F 1$of 77% on more than half of the 62 WIRA image categories. Finally, these configurations collectively achieved perfect performance with 100% accuracy, precision, recall, and F1 in 3 WIRA image categories. These results establish the Symbiote Particle as an effective solution to detect hyper realistic AI-generated content.