acceptodds
Under review as a conference paper at ICLR 2027

PhysCons: Physics-Aware Consistency Modeling for Robust AI-Generated Image Detection

Abstract

Detecting AI-generated images from unseen generators remains challenging, as existing methods often overfit to generator-specific artifacts that evolve rapidly. We observe that while generative models change over time, the physical laws governing real-world image formation—such as optics, sensor noise, and illumination—remain invariant. Based on this observation, we propose PhysCons, a physics-aware consistency modeling framework that detects synthetic images by identifying violations of intrinsic physical consistency, rather than modeling generator-dependent signatures. PhysCons jointly exploits three coupled physical dimensions—spectral–spatial structure, luminance–chrominance coupling, and noise–texture statistics—to expose the physical decoupling inherent in generative processes. Extensive experiments on GenImage and AIGCDetectBenchmark demonstrate state-of-the-art generalization of PhysCons to unseen generators. Notably, PhysCons achieves this performance with only 0.05M parameters and 722 FPS, offering an exceptional balance between robustness, accuracy, and efficiency.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
AI-Generated Image Detection, Image Forensics, Physics-Aware Consistency Modeling, Cross-Generator Generalization, Explainable Computer Vision

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_qsB8OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PhysCons, a lightweight AI-generated image detector that models three physics-inspired relationships. It's central hypothesis is that physical consistency in real image formation is more invariant across generators than generator-specific artifacts, enabling stronger generalization to unseen synthesis models. Although, the performance is higher than other methods, the experiments do not convincingly establish that it's improvements arise from learning physically grounded invariances rather than discriminative statistical correlations as the authors argue.

Critical strengths
  • Strong generalization on GenImage with additional evaluations on other benchmarks.
  • Highly efficient architecture
  • Three complementary feature branches and violation loss are supported by ablation studies in terms of the detection performance.
Critical weaknesses
  • The paper assumes that the selected feature relationships represent invariant physical properties but does not quantitatively demonstrate that these relationships are consistently preserved in real images and violated in generated images.
  • Equation 7 explicitly supervises the physical violation maps using real/fake labels, encouraging low activation for real images and high activation for generated images. A concern is that the learned maps may simply encode discriminative statistical differences rather than genuine physical violations.
  • The proposed framework combines established feature extraction methods. Although, their integration into a unified consistency framework is a contribution, the paper does not convincingly establish what fundamentally new modeling capability emerges beyond conventional multi-feature fusion.
  • The authors do not compare the proposed physically motivated feature pairs against alternative, non-physics inspired feature combinations under equivalent architectures and training conditions. If their motivation is correct, the non-physics inspired feature combinations should drop the performance?
Questions for authors

See Weaknesses

2 comments

Thank you for your response. However, I believe two of my main concerns remain insufficiently addressed.

Q1

Table R1 demonstrates statistical differences in spectral–spatial correlations between real and generated images, but does not establish that these differences represent violations of invariant physical relationships. Moreover, no corresponding analysis is provided for the luminance–chrominance relationship.

Q2

The authors argue that the violation maps capture physical inconsistencies, but the provided evidence does not directly support this interpretation. Since Eq. (7) supervises these maps using real/fake labels, it remains unclear whether their activations reflect actual violations of physical relationships rather than class-discriminative features. Neither Table R1 nor Table R2 establishes this connection.

Thank you for additional detailed response.

However, my central concern remains unresolved. While Table R3 demonstrates statistical differences between real and generated images, it does not establish that these differences correspond to violations of physical image-formation constraints.

Furthermore, the reported within-image correlations (0.157 for SS and 0.068 for LC) are relatively weak and do not rule out the possibility that the violation maps primarily capture class-discriminative statistical features.

Therefore, although I acknowledge the effectiveness of the proposed feature combinations, I still find insufficient evidence to support the paper's central claim that the model explicitly learns and detects physical consistency violations, rather than general statistical discrepancies.

Reviewer_qX8EOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

AI-generated content detection has become an active research topic with significant practical relevance. This paper focuses on the generalization capabilities of AI-generated image detectors, aiming to address the challenge of cross-generator generalization. The key motivation behind this work is that effective detection should be grounded in the invariant physical principles governing real-world image formation. To this end, the authors propose PhysCons, a framework designed to robustly detect images produced by advanced AI generative models. The authors argue that this physics-consistency-based detection paradigm, which identifies violations of physical constraints, avoids relying on generator-specific artifacts. Extensive experiments demonstrate promising results in terms of cross-generator generalization, inference efficiency, and interpretability.

Critical strengths

I consider cross-signal consistency modeling to be the most distinctive aspect of this paper. Existing AI-generated content detection methods often rely on statistical characteristics of synthetic images, which may limit their generalization capabilities. Rather than focusing on the absolute values of individual features, PhysCons emphasizes the consistency relationships among different physical signals. I believe this shift in perspective is particularly insightful and may offer a promising direction for future research on generalizable AI-generated image detection.

Critical weaknesses

The paper argues that the physical image formation process establishes structural relationships among different signals in real images, which AI-generated images fail to reproduce. While I find this underlying idea compelling, I still have some concerns about its current implementation. Specifically, I am not fully convinced that such physical consistency provides a fundamentally different basis for detection, rather than a more sophisticated integration of existing statistical cues. As generative models continue to improve, could AI-generated images also exhibit strong spectral–spatial consistency? Conversely, could real images exhibit weakened or unreliable consistency due to variations in camera viewpoints, imaging conditions, or post-processing operations?

Questions for authors

The proposed physical constraints may still be viewed as an extension of conventional detection features, rather than genuinely invariant principles that can reliably distinguish real images from AI-generated ones. Could the authors further discuss the long-term robustness of these physical constraints? In particular, how would the proposed method handle real images that no longer satisfy these constraints due to post-processing operations or other distortions?

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

PhysCons detects AI-generated images using three paired representations motivated by image formation: spectral–spatial structure, luminance–chrominance coupling, and noise–texture dependence. Each branch learns a spatial “physical violation map” and a classification score, with a label-conditioned violation loss and maximum-score aggregation across branches. The manuscript reports strong cross-generator results, including 92.23% accuracy on GenImage.

Critical strengths
  1. the architecture offers a lightweight alternative with potentially meaningful deployment advantages
  2. The proposed method got the promising transfer results across several evaluation settings.
Critical weaknesses
  1. The features are physically motivated, but the model is not explicitly constrained to measure physical inconsistency. The independently learned projections need not produce comparable representations, and Eq7 supervises violation maps through class labels rather than measurable physical relationships. The maps could therefore reflect ordinary discriminative artifacts. The current ablations show that the components help, but single-stream, duplicated-pair, or unrelated-pair controls are needed to substantiate the paper’s central claim.

  2. On AIGCDetectBenchmark, the ensemble achieves 85.44%, below both NT (92.84%) and AIDE (91.13%); on GenImage, LC+NT also outperforms the full model. Maximum aggregation preserves recall but needs evaluation at matched false-positive rates to establish its practical benefit. Moreover, JPEG compression substantially reduces accuracy, and the Sora evaluation does not demonstrate control over encoding and processing differences

Questions for authors
  1. What establishes that the model learns cross-signal consistency? Can you provide parameter-matched single-stream, duplicated-pair, or unrelated-pair controls? I would like to understand whether the physically motivated pairings are essential, or whether similar gains arise from combining filtered representations more generally.

  2. Is the full ensemble always the intended model, or is a branch selected using a predefined validation procedure? With thresholds chosen only on source-domain validation data, how do the full ensemble, individual branches, and LC+NT compare in recall at matched false-positive rates?

  3. How were processing differences controlled? Were real and synthetic samples matched for compression, resolution, resizing, and video encoding in the Sora evaluation? Please also reconcile the augmentation descriptions in Section 4.3 and A

  4. Please specify the values ​​for parameters λ\lambda and τ\tau, as well as the process used to determine them. Additionally, clarify whether the ℓ1\ell_1 penalty term for the real-valued image in Equation (7) is calculated as a summation or as an average over the spatial dimensions. Furthermore, to facilitate reproducibility, please specify the widths of each branch. Finally, please define the temporal scope covered by the 722 FPS result—specifically, whether this metric encompasses feature extraction, FFT/filtering operations, and the computations for all three branches.

Comments on the paper

1 comment
Area_Chair_JNjGOpenReview ↗

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.