aiqbench analysis
This research exposes a critical vulnerability in current AI-generated image detectors, revealing that they rely on global artifacts from the VAE encoding process rather than identifying synthesized content itself. The proposed "Inpainting Exchange" method neutralizes these detectors by restoring original background pixels while preserving edits, causing accuracy to drop to chance levels across commercial APIs and bias-aware models. This finding fundamentally challenges the reliability of existing detection pipelines for verifying AI-generated media.
AI-Generated Image Detectors Exploit Global Artifacts: Evidence from Inpainting Exchange
arXiv:2602.00192v2 Announce Type: replace-cross Abstract: Modern diffusion-based inpainting enables realistic, localized image manipulation, raising critical challenges for reliable detection. We show that state-of-the-art AI-generated image detectors largely fail to identify \emph{what} was synthesized and instead exploit a \emph{global} artifact introduced as a side effect of the generation pipeline: latent-diffusion inpainting passes the entire image through a VAE encoder--decoder, inducing a subtle but pervasive spectral shift even in unedited regions. We turn this observation into a simple, model-free diagnostic, \emph{Inpainting Exchange} (INP-X), that restores the original pixels outside the edited region while preserving all synthesized content. To study the effect at scale, we build a benchmark of $30$K matched real, standard-inpainted, and exchanged triplets ($90$K images) spanning four datasets and three inpainters. INP-X requires no model access and is a single compositing step, yet it collapses pretrained detectors, commercial APIs, and bias-aware methods toward chance (commercial APIs: $>$$90\%\!\to\!\sim$$55\%$; the bias-aware DDA: $97.6\%\!\to\!58.4\%$). We give formal propositions linking the effect to high-frequency attenuation from the VAE information bottleneck, and we causally isolate it with a factorial battery and a variance decomposition that attributes $97\%$ of the background spectral shift to the autoencoder round-trip alone. A perceptual study ($n{=}153$) shows the shift is also perceptible to humans, whose accuracy falls from $63.8\%$ to $56.2\%$ ($p{=}0.002$): significantly reduced, though not eliminated. Finally, we show that mechanism-driven mitigations (mixed-distribution and localization-supervised training) improve robustness, but detecting INP-X edits remains an open problem. Our benchmark and code are available at: https://github.com/emirhanbilgic/INP-X
Abstract:Modern diffusion-based inpainting enables realistic, localized image manipulation, raising critical challenges for reliable detection. We show that state-of-the-art AI-generated image detectors largely fail to identify \emph{what} was synthesized and instead exploit a \emph{global} artifact introduced as a side effect of the generation pipeline: latent-diffusion inpainting passes the entire image through a VAE encoder--decoder, inducing a subtle but pervasive spectral shift even in unedited regions. We turn this observation into a simple, model-free diagnostic, \emph{Inpainting Exchange} (INP-X), that restores the original pixels outside the edited region while preserving all synthesized content. To study the effect at scale, we build a benchmark of $30$K matched real, standard-inpainted, and exchanged triplets ($90$K images) spanning four datasets and three inpainters. INP-X requires no model access and is a single compositing step, yet it collapses pretrained detectors, commercial APIs, and bias-aware methods toward chance (commercial APIs: $>$$90\%\!\to\!\sim$$55\%$; the bias-aware DDA: $97.6\%\!\to\!58.4\%$). We give formal propositions linking the effect to high-frequency attenuation from the VAE information bottleneck, and we causally isolate it with a factorial battery and a variance decomposition that attributes $97\%$ of the background spectral shift to the autoencoder round-trip alone. A perceptual study ($n{=}153$) shows the shift is also perceptible to humans, whose accuracy falls from $63.8\%$ to $56.2\%$ ($p{=}0.002$): significantly reduced, though not eliminated. Finally, we show that mechanism-driven mitigations (mixed-distribution and localization-supervised training) improve robustness, but detecting INP-X edits remains an open problem. Our benchmark and code are available at: this https URL
Accepted to WACV 2027. 23 pages, 16 figures, 20 tables
Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
arXiv:2602.00192 [cs.CV]
(or arXiv:2602.00192v2 [cs.CV] for this version)
https://doi.org/10.48550/arXiv.2602.00192 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 350 words