Beyond Artifacts: Real-Centric Envelope Modeling for Reliable AI-Generated Image Detection

TL;DR

This work reframes AI-generated image detection as real-distribution envelope modeling rather than artifact detection, addressing robustness to evolving generators and real-world degradations. It introduces Real-centric Envelope Modeling (REM) comprising Manifold Boundary Reconstruction, Envelope Estimator, and Cross-Domain Consistency to learn a boundary around the real image manifold and maintain stability across degradations. A new RealChain benchmark simulates chain degradations across open-source and commercial generators, demonstrating REM’s superior robustness and generalization, including forgery-source attribution. The results suggest that focusing on the real distribution yields more generator-agnostic and degradation-resilient detectors with practical impact for real-world content integrity. Future work will extend REM to video forensics and finer-grained real-distribution modeling.

Abstract

The rapid progress of generative models has intensified the need for reliable and robust detection under real-world conditions. However, existing detectors often overfit to generator-specific artifacts and remain highly sensitive to real-world degradations. As generative architectures evolve and images undergo multi-round cross-platform sharing and post-processing (chain degradations), these artifact cues become obsolete and harder to detect. To address this, we propose Real-centric Envelope Modeling (REM), a new paradigm that shifts detection from learning generator artifacts to modeling the robust distribution of real images. REM introduces feature-level perturbations in self-reconstruction to generate near-real samples, and employs an envelope estimator with cross-domain consistency to learn a boundary enclosing the real image manifold. We further build RealChain, a comprehensive benchmark covering both open-source and commercial generators with simulated real-world degradation. Across eight benchmark evaluations, REM achieves an average improvement of 7.5% over state-of-the-art methods, and notably maintains exceptional generalization on the severely degraded RealChain benchmark, establishing a solid foundation for synthetic image detection under real-world conditions. The code and the RealChain benchmark will be made publicly available upon acceptance of the paper.

Figures (8)

• Figure 1: Challenges of AIGI detection in real-world scenarios. (a) As generative models advance (from GAN models to Diffusion models to Autoregressive models), their outputs increasingly approach the real image manifold, rendering decision boundaries learned from older generators insufficient to distinguish emerging ones. (b) Images are often compressed (e.g., JPEG) during multi-round cross-platform/device propagation and post-processed by users (e.g., filters, digital stickers). These chain degradations severely impair the performance of artifact-based detectors.
• Figure 2: Challenges of Generator Evolutions. New generators continually emerge with evolving architectures and sampling strategies, causing the feature discrepancy between real and fake images to diminish over time. Detectors trained on obsolete generators overfit to seen artifacts and fail to generalize to new ones, highlighting the need for a real-centric modeling paradigm.
• Figure 3: Challenges of Chain Degradations. As the propagation chain length increases, multi-stage compression and processing progressively suppress generative artifacts, leading to reduced separability between real and fake samples in the frequency domain. This explains the sharp performance drop of artifact-driven detectors (e.g., high-frequency artifacts) in real-world conditions.
• Figure 4: Overview of the Real-centric Envelope Modeling (REM) framework. REM consists of three key components: (1) Manifold Boundary Reconstruction (MBR) generates diverse near-real samples by applying feature-level perturbations to self-reconstructed inputs; (2) Envelope Estimator (EE) then learns a compact and smooth boundary that tightly encloses the real distribution by separating real samples from the MBR generated near-real ones in feature space; and (3) Cross-domain consistency (CDC) further enforces stability of this boundary under various degradations, ensuring that the learned envelope remains consistent across domains.
• Figure 5: The RealChain dataset contains images with no degradation and images with chain degradations.