Q1: How do negative prompts and modifiers influence image characteristics and quality?

Firstly, we provide an ablation study, comparing the average quality scores of SDXL images with and without negative prompts and positive modifiers. The table below suggests that prompt engineering improves the average quality of generated images.

Since the modifiers are randomly sampled for each image, we compute the average accuracy and quality scores with and without each modifier for SD 3 images. The table below shows that none of these modifiers alone has significant and consistent effects on image quality or accuracy. We argue that the specific effect of modifiers may depend on the prompt and the synergy between modifiers.

Q2: Test on closed-source models like MidJourney and DALL·E 3.

Please refer to Q4 of Reviewer rmyb.

Q3: Discussion on failure cases (e.g., when high-quality images are misclassified).

We take the high-quality images with top 30% ImageReward and HPS v2 scores, and regard the samples with at most 2 correct predictions for the 6 detectors (#Correct<=2) as failure cases, comparing them with the cases of #Correct>=5. We find that the failure cases consistently have lower average contrast and saturation, which is consistent with our conclusions in Sec. 3.

Q4: Consider testing adversarial modifications to verify detector robustness.

Table 1(a) of the paper indicates certain adversarial directions for image modification, such as lowering the lightness/contrast/saturation, or increasing the sharpness. We implement these manipulations with different factors (0.5 means decreasing by 50%; 1.5 means increasing by 50%). The results below suggest that the adversarial modifications are effective for DRCT and RINE, while CoDE and SuSy are less sensitive to such modifications. Interestingly, NPR shows stronger performance on any manipulated data.

Q5: Are there specific image transformations that can fool existing AIGI detectors while maintaining a high human preference score?

To the best of our knowledge, there may not be simple transformations with these properties. However, one may consider replacing the perceptual constraint in the perceptual adversarial attack [a] with the preference score constraint and implement an effective attack for this purpose.

[a] Perceptual Adversarial Robustness: Defense Against Unseen Threat Models. ICLR 2021.

Q6: Would a multi-task learning approach (joint training on quality scoring & detection) further improve detector performance?

We believe that this is possible as training on quality scoring could encourage the detector to learn on more quality-related features, especially those of higher levels, which could make the detector more robust to the distribution shift of low-level features in the cross-generator scenario.

Q1: It is recommended to discuss similar research focusing on resolution, such as [1,2].

Thank you for your valuable suggestion. While high resolution is usually an important aspect of "high-quality" images in a broad sense, this paper considers a narrower sense of image quality, which is evaluated by human preference models. The generated images studied in this paper generally have the same resolution for each generator, while [1,2] emphasize the discrepancy between low-resolution and high-resolution deepfake images, as well as the influence of super-resolution. The paper [3], as you mentioned, focuses on "forgery quality" (i.e., "whether the deepfake image is realistic or not"), instead of the blurriness or resolution of images, which is closer to our definition of quality. We will include more discussions on these related works and clarify our focus.

Q2: This paper does not provide a comprehensive discussion on potential broader applications.

As you kindly suggested, the methodology of this research could be extended to related tasks such as image manipulation detection, and the results in this paper may benefit the research on improving the generative models or building more robust detectors. We will add these discussions to our revision.

Q3: The paper is missing evaluations with more model architectures, such as CLIP, ViT (ImageNet), and CNNs trained from scratch.

We agree that evaluation with diverse model architectures is important for reaching a reliable conclusion. The experiments in Sec. 5 already included detectors with different architectures, such as CLIP (RINE, DRCT), ImageNet-pretrained ViT (CoDE), and ResNet trained from scratch (NPR). Besides, we add more evaluations of the quality-accuracy correlation on generators of different architectures (please refer to Q4 of Reviewer rmyb).

Q4: The proposed method to address the resolution issue is somewhat limited, focusing primarily on patch-based detectors.

We acknowledge that the proposed patch selection strategies are restricted to patch-based detectors. However, other kinds of detection methods could also benefit from the results of this paper. For example, our regression models obtained in Sec. 4.3 could be applied to identify the hard samples for detection, and emphasizing these samples in model training may enhance the generalization.

Q5: The paper could benefit from using more recent and comprehensive datasets.

Thank you for your advice. In our response to Q4 of Reviewer rmyb, we validate our main conclusions on recent datasets of real-world generated images from commercial models. As for the experiments in Sec. 5, DRCT-2M is the latest published comprehensive benchmark for AIGI detection to the best of our knowledge.

Q6: Since resolution is strongly related to the frequency domain, could analyzing from the frequency domain provide some new insights?

Certain fingerprints of fake images such as the up-sampling traces utilized by NPR (Tan et al., 2024) may be witnessed in the frequency domain. Nonetheless, our preliminary experiments suggest that frequency domain analysis may not bring meaningful insights into the image quality we focus on.

Q7: Would high-quality images, such as super-resolution images, be easier to detect?

Thank you for raising this valuable question. Since generative super-resolution with diffusion models is popular in real-world applications, we collect 1000 fake images generated by an SDXL variant without or with generative super-resolution ($1024\to1536$). We test the detectors on these images, and the results below suggest that whether the super-resolution images are easier or harder to detect depends on the detector. We believe that this question is worthy of further study.

Q8: Why are high-quality images easier to detect, given that both high-quality and low-quality images contain traces of DNN generation models?

According to our analyses in Sec. 4, existing detectors are sensitive to certain features that correlate with high quality scores, such as high contrast and high texture richness, which means the traces of generation may be more prominent in such images for the detectors.

Q9: How can we alleviate the overfitting of detection models to the forgery traces in high-quality images and reduce the performance gap between different quality levels?

We may collect more diverse and balanced data in terms of image quality. Moreover, designing models that are more capable of detecting higher-level artifacts like the distortion of object structures could reduce the performance gap, as such artifacts are common in low-quality images.

Q1: The relationship between the curve and the histogram in the plots is not clearly explained.

Thank you for pointing out the potential difficulty for readers to understand the figures. In Figure 1-3, the red curve illustrates how variable $y$ (e.g., the accuracy in Fig. 1) changes with respect to $x$ (e.g., the quality score in Fig. 1), and the blue histogram depicts the distribution of $x$. The curve and the histogram share the same $x$-axis. We will revise the captions and related descriptions of Figure 1-3 to improve the clarity.

Q2: The underlying reasons for the phenomena illustrated by the figures are not well explained.

The main observation illustrated by Fig. 1 (i.e., generated images of higher quality tend to be easier to detect) can be explained by our analyses in Sec. 4. To clarify, we do not suppose that the relation between the detector accuracy and the quality scores is causal; instead, we aim to explore the potential confounders underlying the counterintuitive positive correlation between the two variables. Our results in Sec. 4.3 suggests that certain low-level image characteristics and high-level features may be the confounders for this correlation, and our experiments in Sec. 5 further validate that these low-level image characteristics can be utilized to predict the detectability of image patches on a broader range of data. We will make these points clearer in the revision.

Q3: This study lacks a deeper theoretical analysis or exploration of underlying causes.

We understand that theoretical analysis can provide deeper insights into the relationship between the detector accuracy and the image quality, although this paper is supposed to be an empirical study. We will provide a discussion from the causal perspective (as described in Q2) with a theoretical causal graph that explicitly depicts the relations among the variables we studied in this paper.

Q4: Provide comprehensive statistics on the dataset construction, such as the quantity and distribution of each category.

Thank you for your kind suggestion. The distributions of the image quality scores and the prompt lengths are depicted by the histograms in Fig. 1 and Fig. 3, respectively. However, different from some previous datasets for AIGI detection that are based on certain object categories (e.g., ForenSynths (Wang et al., 2020) and GenImage (Zhu et al., 2023)), the images in our dataset contain diverse content including complex scenes with various objects. Therefore, it is difficult to categorize the images and provide reliable statistics on the image content.

Q5: Will the dataset be publicly available in the near future?

Yes. As stated in the footnote on Page 3 of the paper, the dataset will be publicly available upon acceptance. More specifically, we will release the generated images, the corresponding prompts, and other metadata such as the denoising steps for diffusion models and the JPEG compression quality.

Q1: Discrepancy between existing datasets and real-world applications.

The discrepancy between existing datasets and real-world applications lies in many aspects, such as semantics, quality, and image compression. This paper focus on the quality (as explained in lines 110-117, left column), while some previous studies emphasize that the real-world performance of detectors can be affected by the image compression bias in existing datasets.

We acknowledge that some detectors [1,2,3] are reported to perform well on some OOD data. However, we find that they still have poor performance on other datasets, as suggested by the results in: https://anonymous.4open.science/r/ICML2025-rebuttal-4584/table.pdf. We did not test the method proposed in [1] as it is not open-sourced.

Q2: Evidence for lacking diversity.

Thank you for pointing this out. To clarify, prior studies (such as [1,2,3] as you mentioned) collect images generated by diverse generators to study the cross-generator generalization of detectors. In this paper, we instead focus on the data diversity corresponding to the same generator, especially the diversity of (1) quality-related image features and (2) prompt complexity. Specifically:

(1) The diversity of quality-related features is achieved by the independently sampled positive modifiers, which are not applied in most existing datasets. (2) In contrast to Synthbuster, which has the highest diversity in prompt complexity but 98.5% of its prompts are under 60 words, our dataset (Fig. 3) exhibits a significantly wider distribution of prompt lengths. We will revise the related descriptions in Sec. 2 and Sec. 3.1 to improve the clarification on dataset diversity.

Q3: The role of "advanced generators".

We agree that current benchmarks already contain diffusion-based models like SD (1/2/XL series). However, by "advanced generators", we refer to those more capable of producing high-quality images, especially DiTs. SD 3 and PixArt-α are selected as representatives of DiTs, and the average quality of their generated images is higher than SD 2.1/XL with U-Net architecture, as indicated by the histogram of Fig. 1. We will improve the related descriptions and provide more statistical comparisons.

Q4: Results with more generators.

Thank you for your valuable suggestion. We supplemented our data with images generated by the commercial DiT model FLUX.1 [dev], and Infinity, an autoregressive model. We reproduce Fig. 1/2 on the extended data and the results are presented in Fig. I/II in [a]. In addition, we validate our findings on closed-sourced models Midjourney v6 and DALL·E 3 based on images sampled from existing datasets in Fig. V/VI in [a]. The phenomena in these figures are consistent with Fig. 1/2 of the paper.

[a] https://anonymous.4open.science/r/ICML2025-rebuttal-4584/quality_and_accuracy.pdf

Q5: The biases towards short prompts.

We agree that these pre-trained detectors might be biased towards images generated from short prompts if their training data only comprises such images. Fig. 3 does suggest that these detectors tend to have higher performance if the prompt length is below 40.

However, the effects of prompt length on the detector accuracy are indirect. Specifically, longer prompts may include more objects and depict a more complex scene, or they could be more likely to mention certain attributes of objects that affect their visual appearance. Therefore, the emphasis of our analyses (Sec. 4.2/4.3/5) is on the visual characteristics of generated images instead of the prompts, and our results in Sec. 5 suggest that our conclusions can be utilized to improve the performance of detectors on existing datasets based on short prompts. Please refer to Q2 of Reviewer 581z for further explanations.

Q6: Images generated from the shortest prompt exhibit opposite characteristics.

Thank you for underlining the importance of this intriguing phenomenon. As explained in lines 210-215 (right column) and Appendix B (lines 694-729), we notice that the a significantly lower ratio of the shortest prompts contain color-related descriptions, while such descriptions tend to increase the saturation of image. Hence, compared with medium-length prompts (21-40 words), the shortest prompts induce lower saturation, which indicates lower accuracy according to Sec. 4.3. Therefore, this phenomenon does not contradict with our further analyses concerning image characteristics.

Q7: Explain the classification and provide qualitative examples of short, medium, and long prompts, along with their generated images and corresponding results.

Thank you for your advice. We will revise the descriptions related to prompt lengths according to the following classification standard: "short" means 1-20 words; "medium" means 21-40 words; "long" means more than 40 words.

We provide randomly sampled qualitative examples accordingly in https://anonymous.4open.science/r/ICML2025-rebuttal-4584/qualitative.pdf.

Q1: It is questionable that the quality scores given by the pre-trained preference models can reflect actual human preference/image quality.

We agree that the preference models could not replace humans in assessing image quality, and it is difficult to predict the human preference on image quality due to its subjective nature. Therefore, we show the validity of our conclusions by the consistent phenomenon among different generators and different preference models. We acknowledge that two preference models may be insufficient to support the consistency, hence, we complement the main results with an additional preference model, MPS [b], and other generators (please refer to Q4 of Reviewer rmyb). The consistent results in Figure I/II in [a] further validate our conclusions.

[a] https://anonymous.4open.science/r/ICML2025-rebuttal-4584/quality_and_accuracy.pdf
[b] Learning Multi-dimensional Human Preference for Text-to-Image Generation.

Q2: Why is the quality score determined by both text prompt and image?

Existing preference models commonly take the text prompt for generation as input because a high-quality image in practice should be not only visually appealing but also aligned with the text prompt (i.e., satisfying the intention of users). However, as you kindly suggest, this paper should focus on the visual quality of the generated image alone.

To this end, we try to minimize the influence of the image-text alignment in the comparison of image quality by replacing the text input for the preference models: instead of the original prompt for generation, we use the BLIP-2 caption of the generated image itself, which is expected to be well-aligned with the image as evaluated by the preference models. The corresponding results are presented in Figure III/IV in [a].

Q3: The correlation between the quality of real images and the detection accuracy.

Thank you for your suggestion. We provide the results in Figure VII/VIII in [a], which suggests that there is no significant and consistent correlation between the quality and detection accuracy for real images, indicating that existing detectors may learn different features on real and fake images.

Q4: Why perform regression analyses at the cluster level but not image level in Sec. 4.3?

At the image level, the accuracy is discrete (0/6, 1/6, ..., 6/6) as only 6 detectors are evaluated, which could induce overly high variance of the data and hinder the linear regression analyses. Therefore, we use the cluster-level data, where each sample is representative of a group of images with similar characteristics, to ease the regression analyses.

Q5: Sec. 5 shows that selecting high-quality patches is not always helpful for AI-generated image detection, which is in conflict with the main finding.

We acknowledge that the proposed patch selection strategy may not always bring a performance gain for different data and different detectors. This is expected as our main observation (i.e., generated images with higher quality scores tend to be easier to detect) is statistically valid, and its application to the improvement of detectors may depend on the characteristics of the detector and certain implementations such as the image patchification strategy. We believe that the findings of this paper could motivate future studies to propose more effective detectors with improved algorithmic designs.

Q6: The ImageReward paper was published in NeurIPS 2023, not 2024.

We are grateful for your careful reading! We will correct this error and proofread the references in the paper.