Diffusion-Driven Forgery Detection: Distilling Latent Features for Generalized Image Forensics

Generalized detection of diffusion-generated forgeries across unseen generators is essential for maintaining visual authenticity in the era of generative media. Recent studies incorporate pretrained diffusion models into detection frameworks, primarily capturing forgery cues via image-level reconstruction through the complete diffusion-denoising process. However, such schemes lack an explicit separation between the semantic priors encoded in diffusion representations and the generator-specific noise trajectories introduced during synthesis, causing detectors to overfit to artifacts from specific training generators. To address this limitation, we propose a two-stage diffusion-guided framework that exploits latent-space priors within a pretrained denoising backbone and decouples the detection process into authentic manifold modeling and forgery discrimination stages. In the authentic manifold modeling phase, stable real-image representations are distilled from diffusion features using only authentic data to establish a generator-agnostic semantic space. During forgery discrimination, forged samples are incorporated and multi-layer residual modeling is designed to capture representation inconsistencies relative to the distilled real-image manifold. Extensive experiments conducted on multiple diffusion forgery benchmarks demonstrate that the proposed approach achieves consistent cross-generator generalization performance.