Exploration of Reproducible Generated Image Detection

In recent years, Generative Artificial Intelligence (AIGC) technology has developed rapidly and been widely applied, significantly lowering the threshold for generating high-fidelity images using AI tools. With the continuous iteration of model architectures-from early Convolutional Neural Networks (CNNs) and Generative Adversarial Networks (GANs), to current mainstream Diffusion models, and further to the latest autoregressive and FlowMatch architectures-the fidelity of generated images has reached a level of being "indistinguishable to the naked eye." While this technological advancement has injected efficiency into fields such as creative design and content production, it also provides opportunities for malicious activities: for instance, using generated images to spread fake news, creating infringing works that mimic the style of specific artists, and even forging identity information to commit fraud. Consequently, the necessity and urgency of "AIGC image detection" technology have become increasingly prominent, and related detection algorithms have since emerged as a research focus in both academia and industry. The core challenge in current AIGC image detection stems from the "method lag" caused by the rapid iteration of generative models. Early detection methods targeting GANgenerated images have been proven ineffective for content generated by Diffusion models; and as Diffusion models have become the primary source of current AIGC images due to the popularity of commercial products such as Stable Diffusion (Stability AI (2025) ) and MidJourney(MidJourney Inc. ( 2025) ), a large number of detection methods specifically designed for Diffusion-generated images have emerged in academia. Most of these methods claim to possess "cross-generator generalization ability" in their papers-meaning they can effectively detect images produced by new generative models not seen during training-but they have exposed numerous issues in practical applications, making it difficult to achieve real-world deployment. From a technical perspective, AIGC image detection is essentially a binary classification task between "real images" and "generated images," and existing methods can be categorized into three types: The first type continues the traditional GAN detection approach, focusing on local image differences (e.g., high-frequency noise, texture deviations), and enhances the extraction of such differential features by designing dedicated modules (e.g., attention mechanisms, multi-scale feature fusion modules) to distinguish between real and generated images; The second type designs training-free detection schemes based on the inherent characteristics of generative models, with the most representative being "reconstruction error-based methods"-which achieve classification by leveraging differences in reconstruction errors between real and generated images (Chu et al. (2024); Ricker et al. (2024); Wang et al. (2023) )-but such methods generally have interpretability flaws, and no existing research has elaborated on the underlying principles of reconstruction errors in detail; The third type explores cross-modal fusion approaches, such as introducing Large Language Models (LLMs) to analyze image semantics and generation logic for improving the interpretability of detection (Zhang et al. (2025) ), but they have not yet been applied on a large scale due to limitations in detection efficiency.

Despite the widespread attention on the claimed performance of various detection methods, their “reproducibility” and “practical applicability” have significant shortcomings, which severely hinder technical deployment. We investigated 7 AIGC image detection papers published in top conferences in the computer science/AI field, and after organizing their code repositories and community feedback, we found that the problems in the reproduction process are specifically manifested in three aspects: Poor code availability, as some papers do not make their code public or only provide fragmented scripts; Discrepancies between reproduced results and those reported in the papers, mainly due to insufficient specific reproduction details; Significant performance degradation when methods are transferred to new datasets-a phenomenon that raises great doubts about the effectiveness of detection methods (see Figure 1). In-depth analysis shows that the core cause of this problem is likely that such methods are only trained on “images generated by a single generator,” leading to the features learned by the model lacking generality and being unable to adapt to the practical needs of multi-scenario and multigenerator applications. Previous research has pointed out this possibility, and this paper further verifies the validity of this view in some scenarios by reproducing that work and expanding experiments. 2025) ) show that some methods perform poorly on datasets in real-world application scenarios (Yan et al. (2024) ). The data on the left and right side

…(Full text truncated)…

This content is AI-processed based on ArXiv data.