Diffusion models (e.g., Stable Diffusion, DALL·E 3) can now generate images that are nearly indistinguishable from real ones, making synthetic image detection increasingly challenging. We propose DIFC-Net, a diffusion-intrinsic detection framework that identifies AI-generated images by analyzing their reconstruction behavior during diffusion inversion rather than relying on visual artifacts. DIFC-Net jointly captures spatial discrepancy signals and latent diffusion trajectory evolution, and adaptively fuses them into a unified forensic representation. Extensive cross-model evaluations show that DIFC-Net achieves 90.29% average AUC on multiple unseen diffusion generators, outperforming state-of-the-art detectors while maintaining strong generalization without relying on training-time knowledge of specific generative models.
Lu et al. (Mon,) studied this question.