MLEP: Multi-granularity Local Entropy Patterns for Generalized AI-generated Image Detection

Thank you for your positive feedback. In response to your concern about the generalization ability of the proposed approach against more recent generators, we conducted additional experiments on several other released generative image datasets, including high-quality images produced by GPT-4o, Flux, and Stable Diffusion. Specifically, we evaluated our model—originally trained on ProGAN—against these datasets in a fully out-of-distribution (i.e., cross-generator) setting. The results, summarized in the table below, report mean detection accuracy (Acc.) and average precision (A.P.), with balanced numbers of real and fake samples in each case. As shown, our method consistently delivers strong performance across most tested datasets, demonstrating its robust generalization ability in detecting AI-generated content from previously unseen generators.

The two values in each cell of the above table denote Acc./A.P. respectively.

The rebuttal has addressed all the concerns and I will keep the score to accept the paper.

Response to Weakness 1 - Impact of prompts of text-to-image generation models

Thank you for your thoughtful comment. We agree that in text-to-image diffusion models, prompt specificity can greatly affect the visual quality and details of generated images. We acknowledge that our manuscript did not provide enough details about the prompts, as the test images were sourced from public datasets that cover diverse content but do not specify the prompt details for each image.

To address this, we conducted an additional experiment using DiffusionDB (ACL 2023), a large dataset with 14 million Stable Diffusion images generated from 1.8 million unique prompts. We randomly selected two subsets of 3,000 images each: one from complex prompts (over 200 characters with keywords like “high quality,” “detailed,” and “realistic”) and one from simple prompts (under 100 characters and without those keywords). We evaluated our trained detector on both subsets and found almost no difference in performance (see Table below). This shows that MLEP remains effective even with high-detail prompts, demonstrating its robustness to variations in prompt complexity and image detail.

Response to Weakness 2 - Potential real/fake performance imbalance

First, we would like to clarify that all of our evaluations were conducted under balanced settings—each set of AI-generated images was paired with a comparable number of real image samples. Therefore, we opted to report standard metrics such as accuracy and average precision, which are commonly used by baseline methods, including NPR. In response to your suggestion, we conducted additional analyses to examine detection performance for real and fake images separately, and further included false positive rate (FPR) and false negative rate (FNR). The results indicate that misclassifications are generally well-balanced, with the FPR being slightly higher than the FNR, suggesting that the model maintains relatively even sensitivity across both classes.

Response to Weakness 3 - Lack of robustness tests against image post-processing

Thank you for raising this important point. We acknowledge that the current manuscript does not explicitly address the robustness of our method under common image post-processing operations. Indeed, robustness is not the primary strength of the proposed MLEP approach. For example, when applying different levels of JPEG compression, blurring, or noise, we observed a 17% to 45% drop in detection accuracy—performance that is less satisfactory compared to methods explicitly optimized for robustness. This limitation stems from the fact that MLEP was not specifically designed to handle such transformations, and no data augmentation techniques were employed during training. Instead, our focus was on exploring the potential of information entropy as a discriminative signal for AIGI detection and on revealing the intrinsic differences in local entropy patterns between real and AI-generated images. Interestingly, our method shows strong robustness to image rescaling: when images are downsampled to half their original resolution, the mean detection accuracy remains above 92.4% (only 4.7% drop). We attribute this to the multi-scale resampling strategy used during training, which effectively introduced resolution variability as a form of implicit data augmentation.

We fully agree that robustness to post-processing is critical for practical deployment. Promising directions for future work include incorporating a broader range of image transformations during training and developing more advanced entropy-based features that are resilient to such distortions. We plan to explore these aspects in greater depth in our subsequent research. Thank you for your understanding.

Response to Weakness 4 - Limited comparison and outdated baselines

Thank you for your comments. To address your concern, we conducted additional experiments on several other released generative image datasets, including high-quality images produced by GPT-4o and Stable Diffusion. Specifically, we evaluated our model—originally trained on ProGAN—against these datasets in a fully out-of-distribution (i.e., cross-generator) setting. The results, summarized in the table below, report mean detection accuracy (Acc.) and average precision (AP), with balanced numbers of real and fake samples in each case. As shown, our method consistently delivers strong performance across most tested datasets, demonstrating its robust generalization ability in detecting AI-generated content from previously unseen generators.

The two values in each cell of the above table denote Acc./A.P. respectively.

Moreover, in response to your suggestion, we compared our method with several competitive approaches recently published in 2025. Using their publicly available pretrained models, we evaluated all methods—including ours—on the same set of GAN- and diffusion-based datasets. The comparison results, reported in terms of accuracy (Acc.) and average precision (AP), are summarized in the table below (with balanced real and fake samples). As shown, MLEP consistently outperforms these state-of-the-art methods, demonstrating its effectiveness and robustness. We hope these results offer a more comprehensive and convincing evaluation of our approach.

Two values in the above table are Acc./A.P. respectively. References:

[1] D3: Scaling Up Deepfake Detection by Learning from Discrepancy, CVPR 2025
[2] Towards Universal AI-Generated Image Detection by Variational Information Bottleneck Network, CVPR 2025
[3] ForgeLens: Data-Efficient Forgery Focus for Generalizable Forgery Image Detection, ICCV 2025

Response to Question 5 - Differences between MLEP and ZED

The key difference between our method MLEP and ZED (Zero-Shot Detection of AI-Generated Images) lies in how they leverage entropy for detecting AI-generated images. The specific differences are given below:

MLEP detects AI-generated images by explicitly computing Shannon entropy over small pixel patches (typically 2×2) to capture local statistical irregularities. It treats entropy as a spatial signal and uses patch scrambling to eliminate semantic bias. A CNN is then trained on both real and synthetic images to learn discriminative entropy patterns across multiple scales. The core insight is that real images contain more structured high-frequency details, reflected in distinct entropy distributions.

In contrast, ZED is a zero-shot method trained only on real images. It uses a pretrained lossless encoder to estimate pixel-level probabilities and measures the difference between model-based entropy and negative log-likelihood (NLL)—known as the coding cost gap—as the detection signal. This gap highlights inconsistencies commonly found in AI-generated images.

It is hard to say MLEP is much more advanced, but the additional contributions of MLEP compared to ZED include:

• Explicit entropy: Computes Shannon entropy over small patches without relying on learned distributions.
• Multi-scale analysis: Captures entropy patterns across resolutions for better robustness.
• Semantic-agnostic: Uses patch scrambling to avoid reliance on content semantics.
• Efficient and lightweight: Uses vectorized ops and standard CNNs for fast, scalable detection.
• Trainable end-to-end: Optimized directly for detection, unlike ZED’s fixed encoder.

Nevertheless, both methods tackle AIGI detection using entropy from different angles, offering complementary insights and demonstrating entropy’s potential as a strong cue for identifying AI-generated images.

Thank you for your valuable comments. Indeed, our work is based on a strong assumption that AI-generated images (AIGIs) may exhibit different levels of high-frequency components or randomness compared to real photographic images. This stems from the observation that generative models primarily focus on producing semantically coherent content, but often fail to replicate the low-level noise characteristics inherent in real imaging processes. This motivated our curiosity to investigate the differences between real and AI-generated images through the lens of information entropy—a widely used metric for quantifying image randomness. To this end, we conducted a preliminary study comparing the local entropy distributions (based on 2×2 pixel patches) of real and AI-generated images. As shown in Fig. 1 of the submitted manuscript, real images consistently display a higher likelihood of containing blocks with an entropy value of 2.0, indicating the presence of finer local details. This finding supports our hypothesis that AIGIs, while semantically plausible, tend to exhibit greater structural randomness and lack the detailed local variability seen in real images. Building on this insight, we proposed MLEP (Multi-granularity Local Entropy Patterns), a method that leverages local entropy as a discriminative signal for AIGI detection. The core idea of MLEP is to suppress high-level semantic cues by scrambling small image patches, thereby forcing the detector to rely on low-level entropy patterns rather than content semantics, which may introduce bias. By analyzing the distribution of information randomness both within local patches and across randomly sampled regions, MLEP operates in a content-agnostic manner. Furthermore, we introduce a multi-scale resampling strategy to generalize the entropy estimation across different resolutions, enhancing the method’s robustness in diverse scenarios.

In response to your concerns and two questions about the generalization capability of the proposed method on newly emerging, higher-quality generative images, we conducted additional experiments using the suggested datasets, including images generated by GPT-4o, Flux, Midjourney V5, and Stable Diffusion. Specifically, we evaluated the model originally trained on ProGAN against these newly introduced datasets. Each set of generated images were paired with comparable number of real images selected at random. These experiments were thus conducted entirely in out-of-distribution (i.e., cross-generator) settings. The mean detection accuracy (Acc.) and average precision (A.P.) are summarized in the table below. The SOTA method, i.e. NPR (CVPR 2024), are included for comparison as a reference. As shown, our method maintains consistently high performance across all these novel datasets, demonstrating its strong generalization ability in out-of-distribution scenarios.

The two values in each cell of the above table denote Acc./A.P. respectively.

Unfortunately, our method is less effective on generated video frames, with detection accuracy dropping by more than 30% in most cases. This is primarily because MLEP was not designed to handle image or video compression, and most available video clips have undergone some level of compression. In contrast, our study focused on exploring information entropy as a discriminative signal for AIGI detection, highlighting the intrinsic differences in local entropy patterns between real and AI-generated images, while placing less emphasis on robustness optimization. We fully agree that a robust image-based detector should generalize well to AI-generated video content. However, common video compression techniques can significantly degrade performance by suppressing subtle traces of the generation process—an issue that affects many image-based detectors. Additionally, such detectors are inherently limited in that they cannot capture temporal cues or generation artifacts related to object or scene motion in video. Promising directions for improving video detection include incorporating a wider range of compression operations during training and developing more advanced entropy-based features that also account for temporal information. We plan to investigate these aspects more thoroughly in future work. Thank you for your understanding.

In response to Question 1, we would like to clarify that the entropy comparison shown in Fig. 1 of the manuscript between real and AI-generated images (AIGI) represents an initial observation in our study. It reveals that real images consistently exhibit a higher likelihood of containing 2.0-entropy blocks, suggesting they preserve finer local details compared to AIGI. This aligns with our hypothesis that AI-generated images tend to exhibit greater randomness and less structured detail as quantified by entropy. Motivated by this finding, we further explored the potential of local entropy as a discriminative cue for AIGI detection.

In response to Weakness 1, we acknowledge the scalability limitation of the 2×2 window size. However, this configuration provides maximal computational efficiency, as a 2×2 patch has only five possible entropy states based on pixel occurrence patterns. This allows us to implement a fast, fully vectorized entropy computation method using PyTorch. Specifically, we extract all overlapping 2×2 patches via the unfold function, reshape them into flat vectors, and apply tensor operations to count unique pixel values and assign entropy values through broadcasting or indexing—without any explicit for loops. This approach enables parallel processing across all patches and is highly efficient, particularly when deployed on GPUs. The Python code used for local entropy computation is included in the supplementary materials of our initial submission.

Regarding the concern about using 2×2 local entropy (Weakness 2), we emphasize that the core idea of our method is to eliminate high-level image semantics by scrambling small image patches. This design choice intentionally avoids reliance on semantic or content-level information, which may introduce bias, and instead encourages the model to focus on lower-level entropy distributions—capturing information randomness both within local patches and across randomly positioned pixel groups. To complement the 2×2 setting, we further introduce a multi-scale resampling mechanism, which effectively extends the entropy estimation to smaller image scales and enhances detection robustness.

Finally, concerning the out-of-distribution evaluation (Question 2), we would like to confirm that our experiments were indeed conducted in a cross-generator setting. Our model was trained exclusively on ProGAN-generated images, and tested on images from a range of previously unseen GAN-based and Diffusion model-based generators, thereby validating its generalization across diverse generation methods. To further address your concern, we conducted additional evaluations using image samples generated by previously unseen models, including those produced by GPT-4o, Flux, Stable Diffusion, etc. Specifically, each set of generated images were paired with comparable number of real image samples selected at random, and we tested the model originally trained on ProGAN against these newly introduced datasets. The mean detection accuracy (Acc.)/average precision (A.P.) are reported in the following table, in comparison with the previous SOTA method, i.e. NPR (CVPR 2024). As shown, our method maintains consistently high performance across all these novel datasets, demonstrating its strong generalization ability in out-of-distribution scenarios.

The two values in each cell of the above table denote Acc./A.P. respectively.