Title: Finding Order in Chaos: Localizing AI-Generated Image Forgeries by Detecting Intrinsic Energy Anomalies

Abstract:

The rapid progress in generative artificial intelligence has enabled image editing models to create highly realistic forgeries. These synthetic manipulations often bypass traditional forgery localization techniques, which rely on detecting physical noise that is typically absent in computer-generated data. To overcome this limitation, our study provides a theoretical proof that the diffusion process naturally dampens local high-frequency variance. This suppression generates a distinct statistical energy gap, setting it apart from the natural entropy found in optical imaging.

Building on this finding, we introduce FLAME, a comprehensive framework designed for accurate, pixel-level forgery localization. FLAME leverages a LAD map to identify these inherent anomalies and integrates a parameter-efficient adapter for the Segment Anything Model (SAM). Additionally, to address the gap between forensic evaluation benchmarks and the swift evolution of generative models, we present EditStream. This is an automated pipeline that synthesizes training data continuously based on specific instructions.

Comprehensive experiments confirm that FLAME sets a new state-of-the-art standard. It significantly surpasses existing methods on datasets containing AI-generated forgeries and demonstrates strong generalization capabilities across unseen generative architectures. The source code for this work is accessible at https://github.com/phoenixnir/FLAME.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC