APEX: New AI Metric for Image Quality Assessment

A new evaluation framework for assessing image quality, named APEX (Assumption-free Projection-based Embedding eXamination), has been developed by researchers. This innovative framework utilizes the Sliced Wasserstein Distance, providing a mathematically sound and assumption-free measure of similarity, thus addressing the shortcomings of conventional metrics such as FID. APEX is agnostic to embeddings and utilizes the open-vocabulary foundation models CLIP and DINOv2 for feature extraction. It demonstrates effective scalability in high-dimensional spaces, backed by both theoretical and empirical findings. This advancement tackles the limitations posed by outdated features and the bias inherent in rigid parametric models used in existing metrics.