Ghosts Beneath the Texture: DTS-Det Uses Texture Relationship Cues to Detect AI-Generated Images Across Paradigms

DTS-Det achieves cross-paradigm AI-generated image detection by modeling texture relationships; its approach comes from the semantics-independent texture patterns observed after suppressing semantic interference. On the newly constructed ConImageGen benchmark, it reaches 99.6% ACC, 10.5% higher than the best baseline; in the PicoBanana/RAID cross-dataset evaluations it achieves 93.2%/94.1% ACC, respectively, and under reconstruction attacks and black-box adversarial attacks the detection rates remain 95.2%/88.1%.