acceptodds
Withdrawn from ICLR 2027

Modulo Folding: Rethinking AI-Generated Image Detection Beyond Low-Bit Planes

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Reliable detection of AI-generated images is increasingly important for media forensics and trustworthy visual content. Recent studies often use image-based reconstruction errors as an important feature for determining whether an image is AI-generated. However, these approaches typically incur high computational costs and also fail to capture intrinsic noisy features present in the raw images. The recently proposed LOTA method assumes that low-bit planes provide the key discriminative cue and uses bit-plane guided noise generation for detection, which achieves good performance on AI-generated image detection. However, the underlying mechanism of LOTA remains unclear. In this paper, we reveal that LOTA's bit-plane guided noise generation is mathematically equivalent to taking the image modulo eight. This equivalence shows that the essential ingredient is not the bit-plane representation itself but periodic modulo folding. Indeed, we observe that a broad range of modulo periods achieve comparable performance. We provide a theoretical explanation for this phenomenon: modulo folding compresses the large-amplitude semantic component into a small remainder while preserving small-amplitude artifacts, thereby substantially increasing the relative variance and information of artifacts. Based on this finding, we propose Multi-Period Modulo Folding Detector (MP-MFD), which generalizes the single-period modulo eight design of LOTA to multiple periods and injects the resulting folded images into different stages of a backbone network to capture complementary multi-level artifact cues. MP-MFD requires no reconstruction or iterative sampling and treats LOTA as a special case. Extensive experiments show that MP-MFD achieves SOTA results. These results validate the effectiveness of modulo folding.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
AI-generated image detection

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_2BnqOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper revisits LOTA’s low-bit-plane representation and shows that extracting the three least significant bits is equivalent to applying a modulo-8 operation. It argues that modulo folding suppresses large-amplitude semantic content while preserving small-amplitude generation artifacts. Based on this interpretation, the authors propose the Multi-Period Modulo Folding Detector (MP-MFD), which injects representations computed with different modulo periods into multiple stages of a ResNet-50 backbone.

Critical strengths
  • Simple and computationally efficient design. MP-MFD operates directly on pixel values without requiring diffusion reconstruction, inversion, or iterative sampling, making it potentially suitable for efficient detection.
  • Useful reinterpretation of low-bit-plane extraction. The equivalence between extracting the lowest three bit-planes and modulo-8 provides a clear perspective on LOTA and motivates exploring alternative modulo periods.
  • Promising reported empirical results. The method achieves strong average accuracy across four benchmarks, including cross-dataset evaluation on Co-Spy-Bench. The ablations also provide evidence that the proposed multi-period, multi-stage configuration improves over a single-period input.
Critical weaknesses
  1. Reliability is insufficiently defined and evaluated. The paper primarily reports accuracy at a fixed threshold, without establishing whether “reliable” refers to calibrated confidence, low false-positive rates, robustness, or interpretability. The authors should specify the intended meaning and provide corresponding evaluation.

  2. The theoretical explanation relies on unverified assumptions. The analysis assumes small artifact amplitudes and approximately uniform semantic remainders without empirical validation. Its variance calculations omit covariance terms and the effects of modulo-boundary crossings. Consequently, the claimed increase in artifact variance contribution does not establish improved discriminative information.

  3. The ablations do not isolate the contribution of multiple periods. Adding periods also introduces adapters and additional feature-injection pathways. Matched-capacity controls with repeated periods are needed to distinguish the benefits of different modulo representations from architectural changes. Moreover, modulo-6 is deterministically recoverable from modulo-12, making the claim of complementary information insufficiently justified.

  4. The evaluation does not adequately exclude dataset shortcuts. Near-saturated GenImage accuracy provides limited evidence of practical reliability without controlling for compression history, resolution, and other acquisition differences. Although the paper includes cross-dataset evaluation, additional tests on challenging, bias-controlled and in-the-wild data are needed. The robustness experiments also lack comparisons against baselines trained with equivalent augmentation.

  5. Coverage of relevant recent approaches is incomplete. Despite including Effort and two 2026 baselines, the paper does not adequately position its contribution against AIDE, FakeVLM, FakeVLM-R1, and related explainable detection approaches such as SIDA and FakeShield. Task-aligned comparisons, with explicit accounting for training data and computational resources, are necessary to substantiate the broad state-of-the-art claims.

Questions for authors
  1. What specific aspect of “reliability” does MP-MFD improve beyond detection accuracy, and what evidence supports this improvement?

  2. Can you empirically validate the assumptions underlying the theoretical analysis, particularly small artifact amplitudes and approximately uniform semantic remainders? How would accounting for covariance and modulo-boundary crossings affect your conclusions?

  3. Do multiple periods outperform repeated-period inputs, such as {8,8,8}\{8,8,8\} or {12,12,12}\{12,12,12\}, under identical architectures and training settings? Since modulo-6 is recoverable from modulo-12, what explains its additional benefit?

  4. Why does the single-period modulo-8 configuration achieve approximately 96.0% on GenImage, whereas LOTA achieves 98.6% in Table 1? Which implementation or training differences account for this gap?

  5. How does MP-MFD perform when compression history and resolution are controlled across real and generated images, and on challenging benchmarks such as Chameleon? Does its robustness advantage persist when competing methods receive identical JPEG augmentation?

  6. Can you provide task-aligned comparisons with AIDE, FakeVLM, and FakeVLM-R1, while discussing how your contribution relates to SIDA and FakeShield?

  7. Were modulo periods, injection stages, and adapter configurations selected exclusively using an independent validation set? Are the reported improvements consistent across multiple random seeds?

Reviewer_9AZwOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper focuses on fake image detection, especially the low-bit-plane detection methods. Low-bit plane detection methods operate on the fact that generator artifacts are often microscopic and contained in the bits of lesser significance (of the pixel intensity). The baseline method (LOTA), only uses the last 3 bits, the authors make the argument that taking the last-3 bits is akin to taking modulo 8. Based on this, they present an argument where assuming artifacts predominantly reside in the last m-bits (where m is a number like 3), the modulo operation removes the semantic info, which provides no discriminative signal, and increases the amount of artifact content present in the input. Based on this, they design a method which feeds multiple images which have each been processed with a modulo m operation (in their case 6, 8, 12) and train a detector on that. The authors report results measured by accuracy on GenImage, DRCT and Co-Spy-Bench which show improvement of the method over its baselines. They conduct tests ablating design choices such as bit-plane selection, and also measure robustness to certain post-processing artifacts.

Critical strengths
  1. Method is evaluated on 3 prominent benchmarks, some of which use modern generative models such as FLUX.
Critical weaknesses
  1. This is a concern that I have with the whole paradigm of low-bit plane detection methods. But the method strips away anything except the last n-bits (in this case modulo 12, modulo 8 and 6 info). If so an adversary could almost trivially bypass the detector by removing the last 3 bits from an image. From the tests that I conducted, the images almost, always perceptually look the same (even if i take modulo 12 and remove the information). I would therefore like to see the authors report robustness to quantization, essentially I would like to see a graph where x axis shows the number of bits removed (1 would imply the least significant bit is removed, 2 would imply that the first and second least significant bit is removed, so on and so forth) and y axis shows the performance metric (accuracy and average precision i.e AP). For control, I would like to see till 5, and also request the authors to report for the LOTA method as well.
  2. The JPEG compression range assumed in the paper is very small, for instance at JPEG 30, for lot of images do look perceptually the same. The paper currently only reports until JPEG 80, I would like to see the results from a similar sweep on JPEG (100 till 0). I would prefer if the authors would report accuracy and AP, on both this method and LOTA for control purposes. Additionally, I would also like to see the results on a unseen compression technique (WEBP)
  3. The paper does not currently have experiments which justify the requirement of the adapter, for instance I would like to know what happens if the other bit planes are simply concatenated with the input (channel wise) and fed through only the CNN backbone in Fig 2.
  4. The paper draws strong conclusions based on numerical differences which I feel are not significant, for instance ablation study 1, figure 3, argues that the module 6 performs the best and later on 6,8,10 and 12 are taken to be the best bit planes. Line 415-431 also has a similar issue to me where the differences are rather small in order to make such conclusions with respect to stage 3 and 4. But these are based differences under a single accuracy point and the experiments also have a single seed value, hence I am not convinced that these are significant so far.
  5. I am eager to see Table 1,2,3 but with the thresholdless average precision reported as the metric instead of accuracy. While a threshold of 0.5 is reasonable, I feel that the gap between LOTA and MP-MFD is rather small in terms of accuracy, and am curious to know how a threshold-less metric, commonly used in fake image detection research such as AP measures the differences.

Writing Level Weaknesses

  1. Line 15-16: The first time a methods name is mentioned the full name should be used, for instance LOw-biT pAtch (LOTA), so that the reader is not confused. Line 53: First time the method is introduced a citation should be provided so that the reader can look it up.
  2. Line 36-37: while common, GAN (full form should be mentioned) and Diffusion, and both should cite the original papers for completeness.
  3. Line 38-40: Please provide citations for these early detectors.
  4. Whole Paper: Please use parenthetical citations (e.g., “DIRE (Wang et al., 2023)”) rather than placing author names directly after method names. This would improve readability and consistency throughout the paper.
  5. The Related Work section contains numerous duplicated author names (e.g., “Tan et al. Tan et al. (2024b),” “Qian et al. Qian et al. (2020a)”). These errors occur repeatedly throughout the section.
Questions for authors
  1. For the robustness to post-processing experiments, are the real images also being post-processed or is it only the generated images?
Reviewer_DsZ9OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper concentrates on the intrinsic evidence for distinguishing real and AI-generated images. Considering that LOTA has not provided the underlying mechanism of low-bit planes, this paper proposes MP-MFD to generalize the single-period modulo eight design to multiple periods. Extensive experiments exmaine the effectiveness and robustness of MP-MFD.

Critical strengths
  1. The author introduces a novel perspective of modulo folding to further explore AI-generated image detection.
  2. The author claims that MP-MFD generalizes the single-period modulo wight design to multiple periods.
Critical weaknesses
  1. The innovation of this paper is quite limited, as there is no essential difference between modulo folding (MP-MFD) and low-bit planes (LOTA):
  1. Modulo folding and bit planes are simply different perspectives from which to view artifacts of images, and bit planes additionally have theoretical support from digital image processing.
  2. The generalization claimed by the author from bit planes to modulo folding is essentially worthless, for the improvements of MP-MFD compared to LOTA and other methods (in Table 1, 2 and 3) are very small (e.g., from 97.6% to 98.6%). As the accuracy of the comparison methods is already very high, this small increase is more like adding something randomly to obtain a gain.
  3. Theoretical proof and analysis are unreliable, such as strong hypotheses, a lack of assumptions in some proofs, a rigorous mathematical proof, some errors lie in the proof.
  1. The datasets for evaluation are outdated, such as ForenSynths (2020), GenImage (2023) and DRCT-2M (2024). More recently proposed generators should be compared, such as Imagen 4 (2025), FLUX.1 Kontext (2025) and so on. The evaluation on these outdated datasets does not fully demonstrate the superiority of the proposed method, as many methods have already performed well on these datasets.
Questions for authors
  1. Could you provide any substantial gain from the generalization from bit planes to modulo folding?
  2. More evaluation on recently proposed datasets is required.
Reviewer_7NZSOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper focuses on AI-generated image detection. The authors find that the operation used in LOTA is mathematically equivalent to taking an image modulo 8, and that different modulo periods can achieve comparable performance. Based on this observation, the authors propose MP-MFD, which injects modulo-folded images with different periods into different stages of a backbone network to extract complementary artifact features at multiple levels. Experimental results demonstrate the effectiveness of the proposed method, achieving higher average detection accuracy than existing methods.

Critical strengths

The proposed method is simple and straightforward, with high computational efficiency. By extracting artifact features through modulo folding, it avoids complex and computationally expensive processes such as image reconstruction in AI-generated image detection. The method also achieves strong performance across multiple datasets, demonstrating its effectiveness.

Critical weaknesses

1.The theoretical analysis lacks sufficient persuasiveness. The authors claim that modulo folding suppresses semantic content while retaining generation‑induced artifacts. Nevertheless, such a conclusion builds upon assumptions regarding artifact amplitude and boundary‑crossing events, for which empirical validations are absent. Furthermore, elevated relative variance of artifacts does not equate to enhanced discriminative power for detection. 2.In addition, the rationale behind the multi‑period design is not well‑grounded. For instance, the feature yielded by modulo‑6 transformation can be directly derived from modulo‑12 outputs, which implies potential redundancy across the chosen set of moduli. The paper omits comparisons against baselines including single‑period modulo‑24 configurations, repeated‑period setups, and parameter‑matched controls. Consequently, it remains ambiguous whether performance gains stem from complementary artifact cues across different moduli or merely arise from extra network capacity introduced by auxiliary modules.

Questions for authors

Could the authors provide comparisons with single-period modulo 24, repeated modulo periods, and parameter-matched baselines to clarify whether the performance gains come from complementary artifact features or increased network capacity?

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

Related papers

Discussion (0)

Sign in to comment.