acceptodds
Under review as a conference paper at ICLR 2027

Beyond Natural Images: Rethinking AI-Generated Image Detection in Documents

Abstract

AI-generated image detection has attracted increasing attention, but existing evaluations mainly focus on natural images, leaving AI-generated document images largely underexplored. This omission is concerning because documents often appear in sensitive real-world scenarios, such as invoices, expense reports, certificates, and medical records. In this paper, we first construct a controlled diagnostic benchmark, AIGDoc-Pilot, and reveal that existing detectors suffer substantial performance degradation on AI-generated document images, with the mean AUC dropping by more than 7%. Based on this, we further reveal two document-specific properties behind this gap: generation artifacts exhibit strong spatial inconsistency across local regions, and text density significantly affects real-synthetic separability, where text-dense regions offer stronger discriminative evidence. Motivated by these findings, we construct AIGDoc, a larger document-centric dataset containing diverse real-world documents and AI-generated counterparts produced by multiple advanced generation and editing models. Extensive experiments on AIGDoc demonstrate that existing detectors still struggle to reliably identify AI-generated documents, while document-based training partially narrows the gap. Together, these results offer valuable insights for developing dependable and generalizable detectors in document-centric scenarios. The code and datasets will be made publicly available upon acceptance of the paper.

Primary area:
datasets and benchmarks
Keywords:
AI-Generated Image Detection, Document Images, AI Security

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_zbVUOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper first introduces AIGDoc-Pilot to investigate the potential limitations of existing AI-generated image detectors when applied to AI-generated document images. The authors further propose AIGDoc, a large-scale benchmark designed to comprehensively evaluate the performance of existing detectors in this setting.

Critical strengths
  • The identified problems faced by existing detectors on AI-generated document images are potentially valuable and could provide useful insights for future research.
Critical weaknesses
  • The experimental analysis is not sufficiently comprehensive to support the reported findings. Most of the analyses are conducted on AIGDoc-Pilot, which is relatively small and contains only images generated through image-to-image generation. Furthermore, the number of evaluated detectors is quite limited, with several important families of detection methods being overlooked. These include simple classifiers based on pretrained models such as CLIP, which has been shown to exhibit strong robustness in AIGIBench, as well as methods that leverage frequency-domain information, such as FreqNet, AIDE, and SAFE. In addition, all generated images are saved in JPEG format, which may introduce format-specific biases. Since PNG is also a widely used image format, it remains unclear whether the reported findings generalize to images stored in other formats or are partially influenced by the specific compression settings.
Questions for authors

I have concerns about the overall organization of the benchmark and the experimental analysis. For a benchmark-oriented study, I would generally expect the authors to first establish a comprehensive benchmark and then derive findings and conclusions from experiments conducted on that benchmark. However, in the current paper, most of the findings are based on experiments conducted on a relatively small dataset with a limited selection of detectors. In contrast, the proposed large-scale AIGDoc benchmark appears to be primarily used for evaluating overall detection performance, with relatively limited additional analysis. This organization makes it difficult to understand why the main findings are derived from AIGDoc-Pilot rather than the more comprehensive AIGDoc benchmark. More importantly, it raises concerns about the reliability and generalizability of the conclusions presented in Section 3.

I would therefore suggest conducting the analyses in Section 3 on the full AIGDoc benchmark, incorporating a broader range of representative detectors. This would not only provide stronger empirical support for the reported findings but also improve the clarity and logical flow of the paper. In particular, introducing the benchmark before presenting the main experimental findings would make the paper easier to follow and help readers better understand how the proposed benchmark supports the authors' conclusions.

4 comments

We thank the reviewer for these closely related and constructive comments. We agree that the original presentation did not sufficiently distinguish the roles of AIGDoc-Pilot and AIGDoc, nor did it make the validation chain from controlled analysis to large-scale benchmarking sufficiently explicit.

1. Expanded baseline evaluation.

We have expanded the evaluation on the full AIGDoc benchmark by adding representative detectors from complementary families, including the CLIP-based UnivFD and frequency-oriented methods FreqNet, SAFE, AIDE, and SPAI. The corresponding Acc. and AUC results are included in the revised Tables 5 and 9. As shown in Table I & II below, despite these stronger and more diverse baselines, PE-LoRA$^\ddagger$ remains the best method in terms of mean Acc., supporting the effectiveness of document-aware training and region selection on the full benchmark.

Table I. Detection Accuracy across AIGDoc Settings.

GeneratorUnivFDDRCTFreqNetSAFEAIDEEffortSPAIB-FreeDDAMIRROROmniAIDPROBESICAPE-LoRAPE-LoRA$^\dagger$PE-LoRA$^\ddagger$
Text2Image
Z-Image-Turbo49.461.849.350.049.385.274.952.361.449.949.855.144.968.092.890.5
HiDream-O1-Image56.562.184.650.050.372.748.597.095.266.863.297.975.895.297.993.5
Seedream4.550.151.764.050.548.358.275.995.291.852.853.399.561.898.097.495.7
Nano-Banana-250.046.256.050.048.960.256.166.058.950.451.063.955.459.867.560.5
GPT-Image-250.049.552.749.948.574.259.051.347.049.850.351.247.449.459.783.9
Image-Edit
FLUX.1-Kontext49.851.251.150.045.054.562.549.653.750.051.250.751.951.558.993.5
FLUX.2-Klein49.543.752.650.344.348.346.250.752.150.050.551.152.050.253.383.9
Hunyuan-3-Instruct49.857.065.350.044.161.174.772.495.854.266.096.664.566.497.099.5
LongCat-Image-Edit49.847.957.750.047.753.361.771.968.050.453.169.156.875.784.299.2
OmniGen256.157.252.850.361.758.156.781.365.772.466.973.868.572.283.494.1
Qwen-Image-Edit51.049.856.150.048.457.252.582.278.953.756.481.860.473.389.897.3
Step1X-Edit49.655.558.449.944.959.075.374.376.250.653.778.555.575.182.384.5
Seedream4.551.462.154.849.944.254.882.797.294.750.253.099.253.597.199.299.5
Seedream5.0-Lite52.862.155.649.944.048.384.095.685.850.352.798.954.493.698.098.1
Nano-Banana-249.345.049.649.944.354.966.849.050.850.054.450.451.350.257.460.6
Digital-Image-Edit
Qwen-Image-Edit47.834.950.050.022.050.150.030.623.950.050.039.18.355.975.465.6
Nano-Banana-248.136.850.050.119.124.564.517.524.651.150.036.150.949.550.150.0
Average50.651.456.550.044.457.364.266.766.153.154.470.253.769.579.185.3

Table II. Detection AUC across AIGDoc Settings.

GeneratorUnivFDDRCTFreqNetSAFEAIDEEffortSPAIB-FreeDDAMIRROROmniAIDPROBESICAPE-LoRAPE-LoRA$^\dagger$PE-LoRA$^\ddagger$
Text2Image
Z-Image-Turbo48.069.579.479.678.393.281.160.978.456.431.886.142.984.297.797.2
HiDream-O1-Image82.666.796.788.282.883.546.799.699.193.587.199.978.999.399.998.7
Seedream4.567.851.489.270.266.864.091.499.097.867.241.0100.064.899.699.999.8
Nano-Banana-255.945.269.653.466.663.658.570.663.667.847.078.557.558.279.064.5
GPT-Image-267.751.076.952.156.283.361.047.346.964.438.275.848.335.872.794.3
Image-Edit
FLUX.1-Kontext46.853.246.963.550.558.566.847.672.454.354.571.458.961.368.699.4
FLUX.2-Klein43.542.456.665.648.436.940.544.954.950.354.051.358.441.741.291.8
Hunyuan-3-Instruct53.362.189.480.051.572.681.487.399.474.866.899.778.786.699.7100.0
LongCat-Image-Edit44.149.065.065.954.055.965.587.487.161.258.889.570.594.696.6100.0
OmniGen268.862.966.268.473.967.257.685.475.282.770.690.376.992.190.399.3
Qwen-Image-Edit55.951.274.563.362.664.849.292.194.068.872.096.470.689.797.099.8
Step1X-Edit42.158.066.875.148.666.782.980.084.958.461.388.067.785.983.097.0
Seedream4.556.369.178.352.650.068.592.099.499.459.552.4100.064.099.8100.0100.0
Seedream5.0-Lite62.168.176.046.349.258.196.298.797.457.752.299.963.799.299.899.8
Nano-Banana-232.944.236.035.748.467.071.448.163.460.757.774.856.456.286.685.5
Digital-Image-Edit
Qwen-Image-Edit36.929.599.934.48.748.198.619.414.26.310.22.02.577.091.299.2
Nano-Banana-23.830.027.326.10.614.999.815.613.361.354.00.057.011.228.913.3
Average51.153.170.360.052.862.873.069.673.061.553.576.759.974.884.290.6

2. AIGDoc-Pilot is an analytical dataset under controlled conditions.

AIGDoc-Pilot serves a different purpose from AIGDoc. It is a controlled diagnostic testbed rather than a substitute for a comprehensive benchmark. To study whether document images introduce a distinct detection domain gap, we must control for confounding factors such as the generation model, generation procedure, and semantic content. AIGDoc-Pilot therefore uses aligned generation settings, in which generated samples are produced through image-to-image generation or VAE reconstruction. This design enables paired comparisons that isolate the effect of the document domain. In contrast, AIGDoc intentionally covers diverse generators and three realistic generation scenarios; this diversity is essential for evaluating generalization, but makes causal attribution to a single factor much more difficult.

We also clarify why Sec. 3 focuses on a limited set of detectors. The purpose of AIGDoc-Pilot is diagnostic analysis. We therefore selected representative detectors with relatively strong robustness under JPEG95 compression, including B-Free, DDA, MIRROR, PROBE, and OmniAID. These methods provide sufficiently non-degenerate responses for examining systematic changes across domains, spatial locations, and region densities.

In contrast, many detectors exhibit near-chance accuracy after JPEG compression, as also reported in AIGIBench [1]. When a detector produces nearly constant or highly unstable predictions, its apparent variation across regions or domains is dominated by threshold instability and does not provide reliable evidence for identifying systematic trends. Nevertheless, we agree that this controlled analysis alone is insufficient for evaluating overall performance. We have therefore expanded the full AIGDoc benchmark comparison to include a broader range of detectors, including UnivFD, FreqNet, SAFE, AIDE, and SPAI, as reported in Tables 5 and 9 of the manuscript.

3. The experimental results for AIGDoc validate the discovery mechanism.

The full AIGDoc results provide large-scale validation of the controlled findings. Across its 17 generation settings, text-density-guided region selection improves PE-LoRA from 69.5% to 79.1% mean Acc. (+9.6 percentage points). Training with document-domain data further improves the mean Acc. to 85.3% (+6.2 percentage points). These results show that the observations from AIGDoc-Pilot are not used in isolation: they motivate the design choices, which are then validated under the substantially broader and more realistic AIGDoc benchmark.

4. The essence of JPEG compression is to avoid the detection shortcuts introduced by PNG.

JPEG compression is used to reduce format-induced shortcut bias. In the document-image workflows considered in this work, authentic images are commonly camera-captured or distributed after compression, whereas generated images are frequently available as lossless PNG files. Directly comparing compressed real images with uncompressed PNG-generated images can allow a detector to exploit file-format and compression differences rather than generation artifacts. We therefore evaluate generated PNG images after JPEG95 compression, following the practical evaluation protocol also adopted by MIRROR [2].

Reference

[1] Li Z, Yan J, He Z, et al. Is artificial intelligence generated image detection a solved problem?[J]. Advances in neural information processing systems, 2025, 38.

[2] Liu R, Cui M, Qin Z, et al. Mirror: Manifold ideal reference reconstructor for generalizable ai-generated image detection.[J]. Advances in neural information processing systems, 2026.

I appreciate the authors' detailed responses. However, my major concerns remain unaddressed.

My primary concern is that this work is fundamentally a benchmark- and analysis-oriented study. However, the main analytical conclusions are drawn from a relatively small subset with limited evaluation, while the larger benchmark is used primarily for overall performance evaluation. This makes the analytical findings, and consequently the claimed analytical contribution of the work, less convincing.

The authors mention in the rebuttal that their findings on text density motivated the design of the improved cropping strategy, which they demonstrate to be effective. However, two important questions remain unanswered.

First, the authors report three key findings, yet the proposed method only builds upon one of them. How are the other two findings further validated, investigated, or translated into practical implications? Without further exploration, it remains unclear how much of the analytical contribution is actually substantiated by the subsequent method design.

Second, the proposed method lacks sufficient ablation and comparison with alternative cropping strategies. The current experiments only compare it against strict center cropping. What about random cropping, or resizing the image instead of cropping it? These are important comparisons for establishing whether the proposed text-density-based strategy itself is effective, rather than merely outperforming a potentially suboptimal baseline.

More concerningly, according to the results in Table I of the response to Reviewer XvYE, incorporating text-density-based cropping also improves performance on general AI-image detection, where text is less prevalent. This raises another possibility: the observed improvement may stem primarily from the limitations of strict center cropping, rather than from the specific effectiveness of text-density-based cropping. Without comparisons against alternative cropping and resizing strategies, it is difficult to distinguish between these explanations.

To summarize, the additional experimental results provided in the rebuttal do not directly address the questions I raised. Specifically, these experiments continue to use the entire AIGDoc benchmark for overall performance evaluation across more detectors, whereas my primary concern is the validity and robustness of the fine-grained analysis conducted on AIGDoc-Pilot in Section 3. These are fundamentally different aspects of the work, and the additional overall evaluation does not resolve my concerns regarding the latter. Therefore, I maintain my original score.

Reviewer_XvYEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates AI-generated image detection in document-centric scenarios, a relatively underexplored problem compared to natural-image detection. The authors first construct AIGDoc-Pilot to systematically evaluate existing detectors, revealing a substantial performance gap between natural and document images. Further diagnostic experiments identify two key observations: generation artifacts exhibit stronger spatial inconsistency in document images, and text-dense regions generally provide more discriminative evidence for distinguishing real and AI-generated documents. The authors then introduce AIGDoc, a larger benchmark covering diverse document types, generation models, and editing scenarios, and demonstrate the effectiveness of text-density-guided region selection in document-centric scenarios.

Critical strengths
  1. Meaningful problem and comprehensive benchmark. The paper investigates an important yet relatively underexplored scenario in AI-generated image detection. Through controlled comparisons and extensive evaluations of existing detectors, the authors reveal a clear performance gap between natural and document images. The proposed AIGDoc benchmark further provides a valuable resource for future research.

2.Well-supported empirical findings. The fine-grained analysis of spatial inconsistency and text density provides valuable insights into the characteristics of AI-generated document images. In particular, the authors demonstrate that text-dense regions generally offer stronger discriminative evidence, and further validate the practical value of this observation through text-density-guided region selection.

Critical weaknesses

Limited evaluation of generalization beyond document images. The proposed text-density-guided region selection strategy achieves promising improvements on document images, but its effectiveness on other image domains remains unclear. Since the strategy is specifically designed around the characteristics of document images, it may not provide similar benefits for natural images and could potentially introduce performance trade-offs. Although the paper primarily focuses on document-centric detection, evaluating the proposed strategy on conventional AI-generated image benchmarks would help clarify its broader applicability and limitations.

Questions for authors
  1. Consistency across generation scenarios. The paper demonstrates that text-dense regions generally provide stronger discriminative evidence for AI-generated document detection. Does this observation hold consistently across different generation scenarios, including text-to-image generation, camera-captured document editing, and born-digital document editing? It would be helpful to clarify whether the effectiveness of text-density-guided region selection varies across these scenarios.

  2. Challenges in born-digital document editing. The experimental results show that born-digital document editing remains particularly challenging, even for the document-trained detector. Could the authors further discuss the potential reasons for this performance degradation? In particular, is it primarily related to differences in image acquisition characteristics or to the artifacts introduced by specific editing models?

3 comments

Q1. Limited evaluation of generalization beyond document images

Thank you for this valuable suggestion. We agree that assessing generalization beyond document images is important for clarifying both the broader applicability and limitations of our text-density-guided region selection strategy. We therefore added an evaluation on the conventional AIGIBench benchmark under JPEG95 compression; the results are reported in Table 11 of the revised manuscript and summarized in Table I below.

On AIGIBench, applying text-density-guided region selection to PE-LoRA improves the mAcc. from 85.3% to 87.7%, and the mAUC from 95.3% to 97.9%. This result suggests that the strategy can also provide useful discriminative regions beyond document images, rather than being beneficial only for document data.

However, when PE-LoRA is further trained with our document-specific training data, its performance on AIGIBench decreases substantially (72.7% mean Acc.). This trade-off highlights the considerable domain gap between natural-image and document-image distributions, and indicates that document-centric training does not automatically generalize to conventional natural-image benchmarks. We will clarify this limitation in the revised manuscript. In future work, we will investigate domain-balanced training strategies that better accommodate both natural and document images.

Table I. Generalization on AIGIBench under JPEG95 Compression.

MethodProGANR3GANStyleGAN3StyleGAN-XLStyleSwimWFIRBlendFaceE4SFaceSwapInSwapSimSwapFLUX1-devMidjourneyGLIDEDALL-E3Imagen3SD3SDXLBLIPInfinite-IDInstantIDIP-AdapterPhotoMakerSocialRFCommunityAIMean
UnivFD99.2/100.060.5/73.457.8/68.557.0/63.571.6/90.887.2/97.251.9/55.476.6/89.074.2/87.953.3/64.356.7/70.148.7/26.451.5/32.459.3/71.454.0/51.149.3/31.253.0/53.857.0/63.552.5/63.160.7/72.254.6/69.252.0/55.953.7/62.553.8/51.451.0/57.560.0/65.7
DRCT73.2/87.245.9/44.266.1/72.165.3/73.566.0/84.463.1/92.460.3/70.664.7/90.667.3/88.267.0/83.466.9/82.652.7/52.448.6/50.664.9/73.562.8/76.059.6/65.461.3/65.865.3/73.567.4/76.657.1/56.265.8/81.262.6/68.964.8/80.451.1/56.059.0/71.261.8/72.3
FreqNet50.3/59.044.6/51.657.9/67.574.5/86.060.4/78.149.1/56.142.4/44.042.4/45.047.3/54.646.9/58.646.5/60.652.6/71.349.7/53.745.1/30.845.0/54.359.1/75.160.7/78.374.5/86.050.0/72.248.4/46.454.0/65.057.1/71.344.1/33.450.5/52.151.6/54.151.6/59.8
SAFE50.0/47.949.5/44.050.1/42.949.4/24.649.6/62.351.9/44.049.6/63.349.6/74.750.1/72.350.1/75.549.3/65.350.0/39.250.5/26.154.1/61.149.8/35.350.0/29.749.5/30.649.4/24.649.5/44.449.6/45.749.6/48.249.9/52.749.5/25.050.5/51.750.0/62.050.1/48.8
AIDE85.4/98.857.8/85.053.7/71.554.6/82.269.0/91.194.7/99.651.5/65.852.3/74.550.1/64.750.6/64.351.5/68.748.2/28.049.7/28.456.8/79.848.8/42.848.2/23.947.9/38.347.9/45.847.3/40.948.5/56.949.5/65.548.5/54.748.3/31.850.8/54.949.9/35.854.5/59.7
Effort70.2/85.659.1/66.059.7/75.551.5/63.045.8/51.782.0/97.054.9/67.278.4/87.866.3/75.759.5/70.058.1/68.259.1/64.971.2/79.860.6/67.269.4/78.068.1/77.157.2/65.851.5/63.075.8/83.646.4/50.950.7/59.159.3/68.451.6/61.656.1/59.251.8/51.460.2/68.8
SPAI76.0/87.070.3/87.765.6/60.573.7/78.765.8/58.059.5/58.133.1/27.542.4/42.733.5/28.933.8/30.528.2/23.054.4/50.961.9/73.163.4/66.065.9/68.653.0/50.474.3/82.873.7/78.768.2/79.872.6/81.766.7/65.968.1/74.559.9/53.471.8/80.056.5/58.759.4/61.2
B-Free95.2/98.793.5/99.270.8/93.499.3/100.097.9/99.690.5/99.962.1/94.784.1/98.464.9/91.566.3/92.865.7/92.269.7/88.193.6/99.187.7/96.996.9/99.476.7/94.499.2/100.099.3/100.098.7/100.099.1/100.099.2/100.099.1/100.097.2/99.785.4/92.681.4/89.186.5/96.7
DDA86.8/98.795.8/99.064.0/79.197.6/100.056.5/73.851.5/87.479.1/97.088.6/97.776.7/94.484.9/96.684.6/96.791.1/96.795.5/99.288.7/96.395.8/98.983.1/92.197.5/100.097.6/100.097.9/99.997.8/100.097.8/99.997.7/99.987.6/96.781.9/89.786.2/95.484.6/93.3
MIRROR85.9/95.565.3/80.674.1/91.397.1/99.972.4/90.260.8/83.676.5/91.177.0/92.482.4/94.985.7/96.181.0/94.696.8/99.589.5/95.496.4/99.496.3/99.276.7/87.696.0/99.397.1/99.996.5/99.396.8/99.597.6/99.786.5/95.165.9/83.386.5/94.387.4/94.683.6/93.7
OmniAID53.3/53.862.7/90.959.6/73.961.2/84.561.1/70.374.4/93.167.8/92.866.5/90.960.0/82.460.2/87.260.4/87.561.6/83.269.4/84.358.8/73.062.1/94.548.4/46.861.3/89.661.2/84.556.5/86.959.9/72.161.0/73.658.1/66.158.5/72.373.2/92.267.0/82.261.7/79.4
PROBE96.9/99.798.3/99.970.2/90.698.8/100.087.1/98.150.9/93.960.5/94.370.2/96.355.3/85.459.5/89.759.0/88.965.7/91.093.7/99.093.9/98.996.6/99.596.1/99.598.8/100.098.8/100.097.8/100.098.5/99.999.0/100.098.5/100.098.6/99.989.6/95.790.2/96.483.7/96.4
SICA89.7/98.282.5/96.989.0/99.085.9/94.183.2/99.682.1/99.661.8/64.481.0/87.580.4/87.869.3/72.469.6/73.475.6/83.264.2/68.779.6/84.679.2/86.356.2/62.380.5/86.785.9/94.183.5/93.283.5/93.983.6/96.777.3/82.280.3/86.578.0/86.049.1/37.777.1/84.8
PE-LoRA98.9/100.088.7/96.258.3/90.696.4/100.086.5/95.499.4/100.067.3/94.582.3/94.567.0/86.272.2/89.568.8/87.987.5/95.394.8/99.491.4/97.594.6/99.392.8/98.496.3/99.996.4/100.094.1/100.095.4/99.495.3/99.895.1/99.774.7/89.489.4/95.891.8/97.285.3/95.3
PE-LoRA$^\dagger$99.0/100.097.8/99.866.7/94.298.9/100.095.6/99.598.7/100.055.5/95.877.0/97.061.1/93.361.4/93.662.3/94.095.5/99.098.4/99.892.4/98.398.5/99.898.6/99.998.9/100.098.9/100.098.7/100.098.9/100.098.7/100.098.7/99.995.1/99.391.7/97.088.9/95.887.7/97.9
PE-LoRA$^\ddagger$68.2/98.061.0/98.351.1/84.799.3/99.963.4/93.061.6/90.850.2/94.950.2/95.650.6/91.650.7/93.550.2/92.875.5/98.495.0/99.064.1/98.282.0/98.998.4/99.999.7/100.099.3/99.996.9/99.992.8/99.893.8/99.889.8/99.684.5/99.077.8/94.959.6/77.272.7/93.6

Q2. Consistency across generation scenarios.

We sincerely thank the reviewer for this valuable feedback. Table 5 evaluates text-to-image generation, camera-captured document editing, and born-digital document editing separately.

Compared with the center-crop baseline PE-LoRA, PE-LoRA$^\dagger$, which uses text-density-guided region selection during both training and inference while keeping the same training data and optimization settings, consistently improves performance across all three scenario categories: from 74.1% to 83.1% (+9.0%) for text-to-image generation, from 70.5% to 80.4% (+9.8%) for camera-captured editing, and from 52.7% to 62.8% (+10.1%) for born-digital editing. At the individual-generator level, PE-LoRA$^\dagger$ improves over PE-LoRA on 16 of 17 subsets. The only exception is Seedream 4.5 in the text-to-image setting, where the accuracy changes marginally from 98.0% to 97.4%. Overall, the mAcc improves from 69.5% to 79.1% (+9.6%).

These consistent gains support that text-dense regions generally provide stronger discriminative evidence for AI-generated document detection across diverse document generation scenarios, rather than being limited to a particular generation type.

Q3. Challenges in born-digital document editing.

We thank the reviewer for raising this important point. We agree that born-digital document editing remains a major unresolved challenge. As stated in Sec. 4.4, our detector, including its document-trained variant, is trained primarily on camera-captured documents. In contrast, real born-digital documents are rendered by conventional software and lack camera-acquisition characteristics. Although authentic, their visual statistics may thus differ substantially from the real-image distribution learned by existing detectors, making reliable separation difficult.

Table 5 highlights this limitation: in the born-digital setting, many detectors operate at or below chance level, and even document-trained PE-LoRA$^\ddagger$ obtains only 50.0% on Nano-Banana-2. These results show that reliable detection of localized AI editing in born-digital documents remains far from solved.

We have clarified this point in Sec. 4.4 and strengthened the limitations discussion. Going forward, we will further investigate AI-generated document image detection and explore solutions tailored to born-digital documents.

Reviewer_1bSBOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigates AI-generated image detection for documents, an underexplored but security-critical setting. It introduces AIGDoc-Pilot and AIGDoc to show that existing detectors perform worse on document images than on natural images. The paper further finds that detection cues are spatially localized and more evident in text-dense regions, and shows that text-density-guided cropping and document-domain training improve performance.

Critical strengths

a) Important and timely problem. The paper addresses AI-generated and AI-edited document images, which are highly relevant to real-world applications such as invoices, receipts, certificates, and medical records. Compared with natural-image detection, this setting is less explored but has clear security and societal implications.

b) Carefully designed benchmark and analysis. The controlled AIGDoc-Pilot benchmark is a strong aspect of the paper because it compares natural and document images under matched acquisition and generation settings. This design helps support the claim that the observed performance gap is related to the document domain itself rather than solely to differences in data sources or generators.

c) Useful empirical findings with practical value. The paper goes beyond simply reporting lower performance on documents. Its findings on spatially inconsistent artifacts and the usefulness of text-dense regions are intuitive and potentially valuable for future detector design. The large-scale AIGDoc benchmark, spanning multiple generation and editing models, further increases the practical relevance of the study.

Critical weaknesses

a) ’Text density’ may be confounded with high-frequency structure rather than text itself. The density proxy is based on dark, high-gradient pixels, which can also select table borders, stamps, logos, handwriting, or textured backgrounds. Thus, the conclusion should be framed more cautiously as evidence concentrated in text-rich/high-frequency regions.

b) Further experimental verification. The experimental comparisons primarily focus on general-purpose AI-generated image detectors. Incorporating frequency-domain baselines, such as SPAI [1], would provide a more comprehensive assessment of whether existing document-forensics techniques can address emerging generative tampering, and would further substantiate the technical gap identified in this work.

c) Some minor inconsistencies. Figure 2 shows residual plots for "Qwen-Image 0.6," whereas Table 6 only lists quantitative results for Qwen-Image at noise levels of 0.2 and 0.4.

d) ** An open-ended question.** Why is born-digital document editing substantially harder to detect than camera-captured document editing? Is this mainly due to the lack of camera-acquisition traces or the domain gap in detector training data?

[1] Karageorgiou D, Papadopoulos S, Kompatsiaris I, et al. Any-resolution ai-generated image detection by spectral learning[C]//2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2025: 18706-18717.

Questions for authors

See Weaknesses

3 comments

Q1. ’Text density’ may be confounded with high-frequency structure rather than text itself.

Thank you for this important observation. In our work, “text density” is intended as an operational measure of dense document foreground structures rather than the amount of OCR-recognized text alone. In document images, printed characters, table borders, handwriting, stamps, and logos commonly form the text- and layout-related high-frequency structures that distinguish dense content regions from homogeneous backgrounds.

Our goal is to efficiently identify such densely structured regions for crop selection. Using an OCR-based localization pipeline would introduce substantial computational overhead and may miss handwritten text, borders, stamps, or other non-standard document elements. We therefore adopt the lightweight dark, high-gradient-pixel proxy, which effectively captures these dense text-and-layout regions without requiring language-specific OCR.

Q2. Further experimental verification.

We thank the reviewer for this helpful suggestion. We agree that including frequency-oriented detectors provides a more comprehensive evaluation of the gap between general AI-generated image detection and document-centric scenarios. We have therefore added FreqNet, SAFE, AIDE, and SPAI to the experimental comparison. Their results are reported in Tables 5 and 9 of the revised manuscript, with detailed results summarized in Tables I and II below.

These additional baselines cover complementary frequency-aware and robust detection paradigms. Despite these additions, PE-LoRA$^\ddagger$ remains the best-performing method overall in terms of mAcc. and mAUC, further supporting the challenges of AI-generated document image detection and the benefit of document-aware training and region selection.

Table I. Detection Accuracy across AIGDoc Settings.

GeneratorUnivFDDRCTFreqNetSAFEAIDEEffortSPAIB-FreeDDAMIRROROmniAIDPROBESICAPE-LoRAPE-LoRA$^\dagger$PE-LoRA$^\ddagger$
Text2Image
Z-Image-Turbo49.461.849.350.049.385.274.952.361.449.949.855.144.968.092.890.5
HiDream-O1-Image56.562.184.650.050.372.748.597.095.266.863.297.975.895.297.993.5
Seedream4.550.151.764.050.548.358.275.995.291.852.853.399.561.898.097.495.7
Nano-Banana-250.046.256.050.048.960.256.166.058.950.451.063.955.459.867.560.5
GPT-Image-250.049.552.749.948.574.259.051.347.049.850.351.247.449.459.783.9
Image-Edit
FLUX.1-Kontext49.851.251.150.045.054.562.549.653.750.051.250.751.951.558.993.5
FLUX.2-Klein49.543.752.650.344.348.346.250.752.150.050.551.152.050.253.383.9
Hunyuan-3-Instruct49.857.065.350.044.161.174.772.495.854.266.096.664.566.497.099.5
LongCat-Image-Edit49.847.957.750.047.753.361.771.968.050.453.169.156.875.784.299.2
OmniGen256.157.252.850.361.758.156.781.365.772.466.973.868.572.283.494.1
Qwen-Image-Edit51.049.856.150.048.457.252.582.278.953.756.481.860.473.389.897.3
Step1X-Edit49.655.558.449.944.959.075.374.376.250.653.778.555.575.182.384.5
Seedream4.551.462.154.849.944.254.882.797.294.750.253.099.253.597.199.299.5
Seedream5.0-Lite52.862.155.649.944.048.384.095.685.850.352.798.954.493.698.098.1
Nano-Banana-249.345.049.649.944.354.966.849.050.850.054.450.451.350.257.460.6
Digital-Image-Edit
Qwen-Image-Edit47.834.950.050.022.050.150.030.623.950.050.039.18.355.975.465.6
Nano-Banana-248.136.850.050.119.124.564.517.524.651.150.036.150.949.550.150.0
Average50.651.456.550.044.457.364.266.766.153.154.470.253.769.579.185.3

Table II. Detection AUC across AIGDoc Settings.

GeneratorUnivFDDRCTFreqNetSAFEAIDEEffortSPAIB-FreeDDAMIRROROmniAIDPROBESICAPE-LoRAPE-LoRA$^\dagger$PE-LoRA$^\ddagger$
Text2Image
Z-Image-Turbo48.069.579.479.678.393.281.160.978.456.431.886.142.984.297.797.2
HiDream-O1-Image82.666.796.788.282.883.546.799.699.193.587.199.978.999.399.998.7
Seedream4.567.851.489.270.266.864.091.499.097.867.241.0100.064.899.699.999.8
Nano-Banana-255.945.269.653.466.663.658.570.663.667.847.078.557.558.279.064.5
GPT-Image-267.751.076.952.156.283.361.047.346.964.438.275.848.335.872.794.3
Image-Edit
FLUX.1-Kontext46.853.246.963.550.558.566.847.672.454.354.571.458.961.368.699.4
FLUX.2-Klein43.542.456.665.648.436.940.544.954.950.354.051.358.441.741.291.8
Hunyuan-3-Instruct53.362.189.480.051.572.681.487.399.474.866.899.778.786.699.7100.0
LongCat-Image-Edit44.149.065.065.954.055.965.587.487.161.258.889.570.594.696.6100.0
OmniGen268.862.966.268.473.967.257.685.475.282.770.690.376.992.190.399.3
Qwen-Image-Edit55.951.274.563.362.664.849.292.194.068.872.096.470.689.797.099.8
Step1X-Edit42.158.066.875.148.666.782.980.084.958.461.388.067.785.983.097.0
Seedream4.556.369.178.352.650.068.592.099.499.459.552.4100.064.099.8100.0100.0
Seedream5.0-Lite62.168.176.046.349.258.196.298.797.457.752.299.963.799.299.899.8
Nano-Banana-232.944.236.035.748.467.071.448.163.460.757.774.856.456.286.685.5
Digital-Image-Edit
Qwen-Image-Edit36.929.599.934.48.748.198.619.414.26.310.22.02.577.091.299.2
Nano-Banana-23.830.027.326.10.614.999.815.613.361.354.00.057.011.228.913.3
Average51.153.170.360.052.862.873.069.673.061.553.576.759.974.884.290.6

Q3. Some minor inconsistencies.

We thank the reviewer for pointing out this inconsistency. The label in Figure 2 was incorrect and should read “Qwen-Image 0.4” rather than “Qwen-Image 0.6.” We have corrected this error in the manuscript.

Q4. An open-ended question.

We thank the reviewer for raising this important point. We believe that the lack of camera-acquisition traces and the training-data domain gap are two closely related aspects of the same acquisition-domain mismatch. Our detector, including its document-trained variant, is trained primarily on camera-captured documents. In contrast, authentic born-digital documents are rendered by conventional software and lack camera-related characteristics. Their visual statistics may therefore differ substantially from the learned real-image distribution, making reliable real--fake separation difficult.

Table 5 highlights this limitation: in the born-digital setting, many detectors operate at or below chance level, and even document-trained PE-LoRA$^\ddagger$ obtains only 50.0% on Nano-Banana-2. These results show that reliable detection of localized AI editing in born-digital documents remains far from solved.

We have clarified this point in Sec. 4.4 and strengthened the limitations discussion. Going forward, we will further investigate AI-generated document image detection and explore solutions tailored to born-digital documents.

Comments on the paper

1 comment
Reviewer_1bSBOpenReview ↗

Rebuttal has addressed my concerns

Thanks for the rebuttal and the extra experimental results. My concerns are addressed. I will keep my positive score.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.