The work decomposes Fréchet distance into contributions from individual feature-space directions. These contributions sum to the original distance, while leading eigenvectors identify where discrepancies between generated and reference distributions concentrate. Concept mappings then attach semantic interpretations to those directions.
Across images, videos, and proteins, a few directions explained much of the measured gap. Two-thirds of a COCO FID increase was associated with composition simplicity; video results exposed how the embedding determines sensitivity to appearance versus flicker; and Protein FID mainly reflected structural regularity rather than the previously emphasized alpha-versus-beta distinction. This makes aggregate Fréchet scores easier to diagnose when they conflict with perceived or task-specific quality.