acceptodds
Under review as a conference paper at ICLR 2027

Masked Quaternion Phase Modeling for AI-Generated Image Detection

Abstract

Since Oppenheim and Lim's classic experiments, the phase of the Fourier spectrum is known to carry the structural content of an image. The spectral representations currently used to detect AI-generated images nevertheless rely almost entirely on the amplitude spectrum. We build a detector on the phase itself. We hypothesize that a real image inherits phase consistency from its physical imaging pipeline, across frequency bands and color channels, and that a generated image fails to reproduce this consistency. To capture this consistency, we take the quaternion Fourier transform (QFT) of the RGB image. Unlike three separate transforms, the QFT places the three color channels in one shared phase frame, as a luminance and a chrominance component whose relative phase is well defined. We pretrain a vision transformer on real images only to restore the quaternion phase of masked spectral regions and freeze the model, with its decoder, as the backbone. On this backbone, we build two scores: a phase-deviation score that measures how far an image departs from real-image phase structure without any generated image, and the score of a lightweight attention specialist trained on a single generator. Their decisions are combined by a logical OR at thresholds set on real training images. In the GenImage benchmark, real and generated images differ in compression format and resolution, and these alone predict the label. Under a protocol that canonicalizes both properties for every method, the detector averages 88.2% accuracy over eight generators, 14.9 points above the best of six retrained baselines. QPP transfers zero-shot to the nine generators of Synthbuster and remains stable under JPEG recompression down to a quality factor of 50. Controlled ablations attribute the gain to supervising the luminance–chrominance relative phase, a quantity defined in the shared phase frame that per-channel objectives leave unsupervised.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
AI-generated image detection, quaternion Fourier transform, Fourier phase, self-supervised learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_9fNVOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a method for AI-generated image detection based on the quaternion Fourier phase and learns phase-related representations through masked quaternion modeling. The paper also attempts to address the compression and resolution shortcuts in GenImage, which is a relevant issue. However, the core motivation, problem setting, and main technical narrative of this paper overlap substantially with recent work. The paper does not demonstrate enough independent novelty. In addition, the experiments only compare against methods published in 2025 or earlier, which is not sufficient to support the paper’s conclusions.

Critical strengths

1.The paper attempts to address the compression and resolution shortcuts in GenImage, an issue that has often been overlooked in previous work.

2.The idea of using the quaternion Fourier transform (QFT) to capture properties of the physical imaging process is interesting and potentially useful.

Critical weaknesses
  1. Although the paper cites related works, it does not provide a direct experimental comparison or clearly explain the fundamental differences between these methods, such as Any resolution AI-generated image detection by spectral learning, CPTFormer.

2The experiments do not sufficiently compare against the latest relevant methods. Most of the baselines are from 2023–2025. This is not sufficient for an ICLR 2027 submission and makes it difficult to assess the actual state-of-the-art performance of the proposed method.

3.The experimental results suggest that the main performance gain comes largely from the input preprocessing rather than from the quaternion phase itself.

4.The diversity of the generators used for evaluation is insufficient to support broad generalization claims. The current set of generators is relatively limited, and it does not adequately cover newer generation models or substantially different generation mechanisms.

5.The method relies on many manually designed and data-dependent choices, and its generalization requires further validation. Authors should provide more extensive sensitivity analyses and evaluate whether the method remains effective under different datasets, real-image domains, and preprocessing settings.

Questions for authors

1.Why does the paper not compare against methods published after 2025? Since this is an ICLR 2027 submission, please clarify whether the authors evaluated the latest relevant methods, especially those published in 2026.

2.The authors have only provided empirical validation for the core hypothesis. Could you provide a more in-depth mechanistic analysis?

2 comments

We thank the reviewer for the constructive and detailed comments.

W2 / Q1 — Baselines from 2026

Most baselines are indeed from 2023–2025. Methods from 2025 are included (SPAI, AIDE, LOTA, and C2P-CLIP retrained under the same protocol in Table 1; Effort and AlignedForensics as literature values in Tables 1 and 2), but none from 2026. We will evaluate the 2026 methods with public implementations and GenImage-SD1.4 weights (OmniAID [ICML 2026], IAPL [CVPR 2026]) under the same protocol on GenImage, Synthbuster, and Chameleon, and add them to Tables 1 and 2 in the revision.

W3 — Preprocessing or quaternion phase?

The point is partly correct numerically, so we separate the contributions (GenImage average, %):

ComparisonValuesGap
Imaging channel off → on (QPP)74.8 → 88.2+13.4
Pretraining objective (channel on, identical conditions)MCFM 83.3, no L_cross 82.2 → MQM 88.2+4.9 to +6.0
Best baseline, channel offNPR 72.2 → QPP 74.8+2.6
Best baseline, channel onC2P-CLIP 73.3 → QPP 88.2+14.9

The imaging channel is the largest contribution and part of the method; it adds 13.4 points to QPP while working against the baselines, so most of the 14.9-point gap comes from this asymmetry. With the preprocessing fixed, the pretraining objective accounts for 4.9–6.0 points, and without the channel QPP still leads the best baseline by 2.6 points. The revision will stop presenting the 14.9 points as a representational gain and report Table 1 with and without the channel.

The preprocessing does not create the sensitivity. The channel is applied identically to every baseline, and QPP leads without it (Table D.10). Keeping the preprocessing, readout, and mixture model fixed and swapping only the backbone for a public model lowers the deviation AUROC and removes the response to phase perturbations of real photographs:

BackboneDeviation AUROC (GenImage avg.)FPR under phase perturbation of real photographs
MQM (ours)0.84811.7% → 97.4% (monotonic)
MAE (public)0.79511–18% (no change)
MFM (public)0.78511–18% (no change)

The sensitivity to phase regularity therefore comes from pretraining that masks the spectrum and restores real images. (MAE and MFM use a decoder fitted post hoc, so the match is not perfect.)

W4 — Diversity of generators

Agreed. The evaluated generators are the eight of GenImage and the nine of Synthbuster, without latest-generation models. We will drop “broad generalization” and state the evaluated range as a limitation in Section 7.

W5 — Sensitivity analysis

Threshold sensitivity is in Table 4 and Tables D.7–D.9. Sensitivity to the number of clusters, the channel settings (JPEG quality, downscaling ratio), and the real-image reference set (ImageNet / COCO / RAISE) will be added to the appendix.

Q2 — Mechanistic analysis

We examined what the detector reads from both sides. On real photographs, the score ignores white balance, chrominance gain, and uniform rotation of the relative phase (FPR changes by at most 4.7 points) but rises monotonically from 11.7% to 97.4% under per-bin phase noise (response to Reviewer rFw5, W2). On generated images, matching the amount of high-frequency detail changes the AUROC only from 0.848 to 0.820 (response to rFw5, W1), the relative-phase measures alone carry most of the discrimination (0.842; response to Reviewer 3ML5, W1), and the public MAE and MFM backbones show no phase sensitivity (table above). Accordingly, we restrict the central hypothesis to measuring the deviation from the statistical regularity of the chrominance phase structure of imaged photographs, and reposition the illuminant-cast derivation (Appendix D.1.4) as a property of the representation.

  1. The authors only promise comparisons with methods published in 2026, without providing results. Despite having sufficient time during the rebuttal period, they have not provided additional experimental evidence sufficient to establish superiority over existing methods.
  2. The authors still do not adequately explain the substantive differences between their method and closely related work. This distinction is essential for assessing the paper’s independent novelty.

Overall, I maintain my original rating.

Reviewer_rFw5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This work proposes Quaternion Phase Probing (QPP), a method for detecting AI-generated images. A vision transformer is first pretrained on authentic photographs to restore spectral content that has been masked out, and an extra loss term directly supervises the relative phase between luminance and chrominance. Once pretrained, the reconstruction network is frozen and used as a feature extractor for two components: a deviation detector fit only on real images, and an attention-based specialist trained on images from one generator. Each component is given its own calibrated threshold, and the final prediction is the logical OR of the two. Experiments use a common preprocessing pipeline and are supplemented by studies of cross-dataset generalization, robustness to compression, and the quality of the learned representation.

Critical strengths

The authors look at how phase relates across color channels instead of modeling each channel's spectrum in isolation. The relative-phase loss is a sensible inductive bias, and the controlled ablations suggest that this supervision adds value on top of the reconstruction loss alone.

Keeping the backbone frozen after pretraining on real data neatly decouples representation learning from the supervised detection stage. Because of this, generated images seen during detector training have limited ability to reshape the restoration behavior, which makes the setup well suited to studying reconstruction cues that might generalize.

Critical weaknesses
  1. The deviation detector presumes that a model reconstructing only real images acts like a projection onto genuine photographic structure. Yet the backbone is optimized to fill in missing spectral content, not to map arbitrary inputs onto a manifold of real images. Its reconstruction error may therefore track how predictable the content is, how complex the texture is, or how far the input departs from the corruptions seen in training, rather than whether the image is authentic. This matters most for the unmasked-input view, since reconstructing a complete input is not the task the model was trained on. The authors should explain more clearly why these restoration signals would distinguish synthetic images from real photographs that are simply unfamiliar.

  2. The phase loss is selective in ways that muddy its physical meaning. Its weighting by amplitude, coupling strength, and frequency favors spectral regions where luminance and chrominance are both strong and where reconstruction is assumed to be reliable. That choice is defensible for stable training, but it means the learned features need not reflect a general physical consistency principle; they may instead favor color and texture regularities of the pretraining data. Likewise, better illuminant prediction shows that the features are sensitive to color, not that this color information is what drives real-versus-fake decisions. The proposed mechanism would be more persuasive if the authors tested how the detector responds to controlled manipulations that alter these statistics while keeping the image authentic.

  3. The imaging-channel step and the real-image reference aim to handle nuisance factors, but they do not explicitly separate those factors from the intended authenticity cue. Passing every image through the same JPEG compression and resizing cannot in general undo its prior processing, and may simply alter how that history shows up in the reconstruction features. The clustered Mahalanobis model then counts any distance from the resulting reference as a sign of generation, whether it stems from synthesis, the camera, post-editing, or scene content. The core problem is one of identifiability: the paper does not show which sources of deviation are actually measured, nor why the chosen preprocessing removes nuisance variation while preserving authenticity information.

Questions for authors

Please respond to each of the weaknesses listed above.

3 comments

Response to Reviewer rFw5

We thank the reviewer for the careful reading. The three concerns share one question: what does the phase-deviation score measure? We answer with three new experiments and narrow two claims accordingly.

W1 — Predictability rather than authenticity?

The intact view is not essential. The mid-band-masked view alone, the pretraining view, reaches 87.7% against 88.2% with all three views; removing the intact view costs 1.3 points. Section 4.2 will present the masked views as primary and the intact view as a complement.

Views usedDeviation score only, acc. (%)Deviation AUROCFull system, acc. (%)
All three (paper)75.80.84888.2
Intact only75.60.84084.9
Mid-band masked only75.30.84387.7
Complement masked only69.30.80583.2
Two masked views74.00.83386.9

Complexity does not explain the separation. Stratifying by a complexity index (fraction of spectral energy above 32 cycles; quartiles of the calibration reals), the deviation AUROC stays at 0.78–0.86 within every quartile (0.779 / 0.854 / 0.863 / 0.832 from low to high), the index alone gives 0.629, and regressing it out leaves 0.822 (0.848 before). On real images the rank correlation between score and index is −0.14, so complex real images are not the ones flagged.

The score does not fire on unfamiliar real photographs: with thresholds fixed on ImageNet, the false-positive rate is 12.8% on MS-COCO and 5.3% on RAISE (15.0% on GenImage validation reals; 21.9% on Chameleon).

W2 — Is the phase loss selective, and what does illuminant estimation show?

On the weighting: the weights are confidence weights that down-weight bins where the relative phase is undefined or unpredictable from context. As the reviewer notes, what is learned is a statistical regularity of imaged photographs, not a physical law; we will restate the claim at that scope.

We ran the proposed controls: one manipulation at a time in the quaternion Fourier domain on 2,000 real photographs (GenImage validation split), false-positive rate at the declared thresholds.

ManipulationStrengthPSNR (dB)FPR (%)
None——11.7
White balance±30%19.214.4
Chrominance amplitude×227.816.4
Relative phase, same rotation in every binπ/224.215.1
Chrominance shifted spatially (structure kept)8–32 px27–3123
Relative phase, per-bin noiseσ = π/835.415.9
〃σ = π/430.046.7
〃σ = π/225.893.6
〃random24.397.4
Common phase, per-bin noise (relative phase kept)σ = π/821.840.2
〃σ = π/416.369.5

σ: standard deviation (rad) of Gaussian noise added per frequency bin.

  1. No response to color statistics: white balance, amplitude, and uniform rotation change the image down to 19–28 dB yet move the FPR by at most 4.7 points; a structure-preserving chrominance shift gives 23%.
  2. The score fires when the chrominance phase structure is broken: per-bin noise raises the FPR monotonically from 11.7% to 97.4%.
  3. The response is not specific to the relative phase: common-phase noise that preserves it also raises the FPR (40.2% at σ = π/8).

On illuminant estimation, the reviewer is right: since the detector ignores uniform rotation and white balance, a global color cast is not used for the decision. We will reposition illuminant estimation as a property of the representation and describe the decision signal as the deviation from the regularity of the chrominance phase structure.

On generated images too, the relative-phase group carries the detection: reading out only r_ψ and r_χ gives AUROC 0.842 against 0.848 for all four measures (0.785 for the others alone); per-generator breakdown in our response to Reviewer 3ML5 (W1).

Matching the amount of fine detail preserves the result. After matching the distribution of chrominance energy above 32 cycles between real and generated images, the average deviation AUROC moves from 0.848 to 0.834 (0.820 when luminance detail is matched as well).

GeneratorBeforeAfter matchingDetail amount alone
SD1.40.7200.7060.374
Midjourney0.8020.7900.398
ADM0.9410.9340.704
GLIDE0.9590.9540.682
BigGAN0.9610.9350.837
Average (8)0.8480.834—

For the SD models and Midjourney the generated images contain more chrominance detail than the reals (right column below 0.5) and are still detected, so “detected because detail is scarce” does not hold.

W3 — Identifiability and the imaging channel

We agree with the core of this point. A one-class distance cannot tell whether a deviation comes from synthesis, the camera, post-editing, or the scene; real photographs with heavy processing histories can be flagged (on Chameleon the score inverts to AUROC 0.382). We will state this scope in Sections 4.2 and 7. The imaging channel does not undo past processing; it brings images with different histories to a common one and tightens the real-image reference (74.8% without it vs. 88.2% with it, Table D.10; 87.5–89.4% under upstream JPEG recompression down to quality 50, Table D.11).

The preprocessing does not create the discriminability. Without the channel, QPP at 74.8% still exceeds NPR (72.2%), Swin-T (72.0%), and C2P-CLIP (71.6%) (Table D.10), and swapping only the backbone for the public MAE or MFM lowers the deviation AUROC to 0.795 / 0.785 and removes the response to the manipulations above (details in our response to Reviewer 9fNV, W3).

Reviewer_rFw5OpenReview ↗

RE: Rebuttal

Thank you for the detailed response and the additional controlled experiments. The masked-view ablations and complexity analyses substantially alleviate my first concern, and I appreciate the more cautious interpretation of the phase signal and illuminant estimation results. While the distinction between synthesis-related deviations and those caused by other distribution shifts remains unresolved, the additional evidence helps clarify what the detector measures and the limitations of its interpretation. Please incorporate these analyses and the revised scope of the claims into the manuscript.

Reviewer_3ML5OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Quaternion Phase Probing (QPP), an AI-generated image detection framework based on the hypothesis that AI-generated images may fail to reproduce the cross-channel phase regularities that arise from the physical imaging process of real photographs.

To this end, the authors pretrain a backbone exclusively on real images using Masked Quaternion Modeling (MQM), explicitly learning the relative phase between luminance and chrominance components. The framework then detects AI-generated images by combining a phase-deviation score, which measures deviations from the reference distribution of real images based on discrepancies between input images and their reconstructions, with the prediction of an attention specialist trained on images from a specific generator, using a logical OR operation.

Experimental results show that the proposed method outperforms baseline detectors retrained under a controlled evaluation protocol that standardizes image compression formats and resolutions. The method also demonstrates generalization to unseen generators and robustness to JPEG recompression.

Critical strengths
  1. This paper explores phase information, which has been relatively underexplored in prior frequency-based AI-generated image detection research. In particular, the approach of explicitly learning the relative phase between luminance and chrominance components using quaternion representations is interesting. The ablation studies further demonstrate that the proposed relative-phase supervision contributes to improved detection performance.
  2. The authors carefully consider potential dataset biases arising from differences in compression formats and image resolutions, and retrain baseline detectors under consistent preprocessing conditions. This evaluation protocol is commendable, as it aims to mitigate the influence of dataset biases and enable a fairer comparison of detection performance.
  3. The paper provides extensive ablation studies examining the contributions of various components and design choices. Additionally, the evaluation across multiple benchmarks and the explicit discussion of the proposed method's limitations strengthen the empirical analysis.
Critical weaknesses
  1. The ablation studies and illuminant estimation experiments support the effectiveness of relative phase supervision, but it remains unclear whether the performance improvements arise from phase inconsistencies associated with the physical imaging process or from other frequency characteristics. Furthermore, the substantial variation in the discriminative ability of the phase deviation score across generators is not sufficiently explained in relation to the proposed physical hypothesis.
  2. On Chameleon, the phase deviation score exhibits an inverse discriminative trend, and the overall detection performance deteriorates significantly. The authors attribute this failure to differences in image processing histories, but there is insufficient empirical evidence to establish these differences as the primary cause. Furthermore, the OR decision rule cannot detect or mitigate the reduced reliability of individual detection scores under domain shifts.
  3. Although GenImage and Synthbuster are widely used benchmarks, many of the evaluated generators belong to earlier generations of image generation models. The limited evaluation on recent models makes it difficult to assess whether the proposed method can generalize effectively to more advanced generators.
Questions for authors
  1. Is there additional evidence that the detection improvements arise from phase regularities associated with the physical imaging process? How can the varying discriminative ability of the phase deviation score across different generators be explained?
  2. Is there empirical evidence supporting the claim that the performance degradation on Chameleon is primarily caused by differences in image processing histories? How could the reliability issues of the OR decision rule under such domain shifts be mitigated?
  3. Could the authors provide additional experimental results demonstrating whether the proposed method generalizes effectively to recent image generation models?
  4. Although the paper includes a Reproducibility Statement, code and checkpoints are not provided. Could the authors make these available to support reproducibility?
2 comments

We thank the reviewer for the detailed reading and for recognizing the evaluation protocol and the ablations.

W1 / Q1 — Phase inconsistencies of the imaging process, or other frequency characteristics?

What the detector reads is the chrominance phase structure; color statistics and the amount of fine detail do not explain it. Four pieces of evidence:

  1. Supervising the relative phase improves accuracy, reproducibly. Under identical conditions MQM reaches 88.2%, the per-channel MCFM control 83.3%, and MQM without L_cross 82.2% (Table 3); across four seeds the AUROC ranges (0.878–0.882 vs. 0.821–0.824) do not overlap.
  2. Interventions on real photographs. Breaking the chrominance phase structure of 2,000 real photographs raises the false-positive rate monotonically from 11.7% to 97.4%, whereas changing only color statistics moves it by at most 4.7 points, and a structure-preserving spatial shift of chrominance gives 23% (table in our response to Reviewer rFw5, W2).
  3. Matching the amount of fine detail preserves the separation. After matching the high-frequency energy distributions of real and generated images, the AUROC moves only from 0.848 to 0.820; for the SD models the generated images contain more chrominance detail than the reals and are still detected (table in our response to Reviewer rFw5, W1).
  4. The relative-phase group carries most of the discrimination (table below): r_ψ and r_χ alone give AUROC 0.842 against 0.848 for all four measures; the other measures alone give 0.785.

We nonetheless restrict the claim. The score also responds to common-phase noise that preserves the relative phase, so it is not specific to the relative phase. We will state the claim as measuring the deviation from the statistical regularity of the chrominance phase structure of imaged photographs, not as a test of a physical law.

The variation across generators has two sources:

  • Latent-diffusion models (SD family, Midjourney; 512 px): the relative-phase group carries the discrimination (SD1.4: 0.769 vs. 0.649). Newer generators come closer to real phase structure; on Synthbuster, Midjourney v5 and SDXL are the lowest columns (63.9 / 63.8, Table 2).
  • Pixel-space diffusion models (ADM, VQDM; 256 px): the amplitude and common-phase measures carry it (ADM: 0.926 vs. 0.848), and their high discriminability depends on the low native resolution.
ReadoutSD1.4MidjourneyADMGLIDEVQDMBigGANAvg. (8)
All four measures0.7200.8020.9410.9590.9580.9610.848
Relative-phase group only0.7690.8550.8480.9340.8780.8940.842
Other measures only0.6490.7570.9260.9040.9590.7750.786

We therefore withdraw the general statement that generative models fail to reproduce phase consistency; the degree of reproduction varies across generators.

W2 / Q2 — Chameleon and the OR rule

The reviewer is right that the cause has not been established. We can show only that the false-positive rate depends on the source of the real images (thresholds fixed: RAISE 5.3%, MS-COCO 12.8%, GenImage validation reals 15.0%, Chameleon 21.9%), and we acknowledge that the OR rule cannot detect a loss of reliability in one score (on Chameleon the specialist alone reaches AUROC 0.744 while the OR combination falls to 51.2%). The revision will state the dependence on processing history as the method’s scope (Sections 4.2 and 7), state the structural limitation of the OR rule, report AUROC and TPR at fixed FPR for each score, and list the cause of the Chameleon inversion and the estimation of score reliability under domain shift as important open problems in Section 7.

Q4 — Code and checkpoints

As stated in the Reproducibility Statement, upon acceptance we will release the code, the pretrained MQM backbone, the fitted phase-deviation reference, the trained specialist, the declared thresholds, and the scripts that reproduce every table. The pretraining and detector configurations are specified in Appendices B.1–B.3 and C, the evaluation protocol in Sections 4.2 and 5.2, and all thresholds are declared as quantiles of the real training images.

I thank the authors for their detailed responses and additional experiments. In particular, I found the analysis showing how the contributions of relative phase features and other frequency features vary across generators interesting. I also appreciate the authors' efforts to refine the interpretation of their central hypothesis and acknowledge the limitations of the proposed method. However, the following concerns remain unresolved.

  1. The reasons for the substantial variation in the discriminative ability of the phase-deviation score across generators, as well as the vulnerability of the OR decision rule when the reliability of individual detection scores deteriorates under domain shifts, remain insufficiently addressed. Although the additional analyses help clarify these issues, they do not provide sufficient evidence to establish the reliability of the proposed detection framework.

  2. The effectiveness of the proposed method on recent image generation models has not been adequately evaluated. The generators considered in this study were primarily released between 2018 and 2023, and thus do not sufficiently reflect recent advances in image generation technology. Although the authors acknowledge this limitation, the detection capability of the proposed method on more advanced, recent generators remains unverified.

Overall, I believe that the reliability of the proposed method has not been sufficiently established by the available evidence, and that the evaluation scope remains limited in its coverage of recent advances in image generation technology. Therefore, I will maintain my original rating.

Comments on the paper

1 comment
Area_Chair_58NUOpenReview ↗

A Gentle Reminder: About One Day Left in the Discussion Period

Dear Reviewers,

This is a gentle reminder that approximately one day remains in the discussion period. Please review the authors' responses and other reviewers' comments, and engage in the discussion to clarify which concerns have been addressed and which remain unresolved.

Thank you for your time and effort.

Best,
AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.