acceptodds
Under review as a conference paper at ICLR 2027

From Physical Defects to Discriminative Residuals: Generalizable Synthetic Image Detection

Abstract

Existing synthetic image detectors often exhibit limited generalization across heterogeneous generative architectures due to their reliance on generator-specific artifacts. To address this limitation, we explore physical discrepancies shared across different generative paradigms to identify forensic cues that remain stable across architectures. Through physically motivated probe experiments, we further identify three shared physical defects across GANs, diffusion models, and DiTs, reflected in systematic discrepancies from natural images under resampling operations, spatial smoothing, and high frequency characteristics. Motivated by these observations, we propose a learnable physical prior decoupling framework that transforms such physical defects into discriminative residual representations. Specifically, we introduce three physical prior simulators, Resample, Blur, and Edge, together with learnable residual modulation to adaptively extract complementary residual cues. These residuals are further enhanced in the high-frequency domain and dynamically fused for classification without additional reconstruction supervision. Extensive experiments using only ProGAN generated images for training show that our method consistently outperforms state-of-the-art methods across multiple benchmarks covering different generative architectures and paradigms, demonstrating strong cross architecture generalization.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Synthetic Image Detection, Cross-Architecture Generalization, Physical Priors, Discriminative Residuals

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Q8kQOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a physical-prior-based residual learning framework for AIGC image detection. The authors design three physical simulators, including Resampling, Blur, and Edge, to extract residual differences between real and generated images, and combine frequency-domain features for classification.

Critical strengths
  • The framework is clearly structured, combining physical operations, residual extraction, and feature fusion into a unified detection pipeline.
  • The introduction of physical priors provides some interpretability for analyzing the differences between real and generated images.
  • The paper provides a relatively detailed description of the proposed method and the roles of different components.
Critical weaknesses
  1. Limited novelty The proposed method mainly relies on residual information, high-frequency features, and physical artifacts for AIGC image detection, which has significant overlap with existing artifact-based detection approaches. The physical operations used in this work, such as Resampling, Blur, and Edge, are based on well-established image processing techniques. The paper mainly combines existing components and does not demonstrate a clearly novel detection mechanism.

  2. Insufficient experimental settings and validation The experiments are mainly conducted on limited generators and standard benchmarks, where the reported performance is close to saturation. However, such high performance makes it difficult to distinguish the proposed method from existing approaches and does not sufficiently demonstrate its generalization ability in realistic scenarios, such as unknown generators, image compression, and post-processing operations.

  3. Insufficient justification of physical prior selection The authors select Resampling, Blur, and Edge as the main physical priors, but do not sufficiently explain why these operations can represent the key defects shared by different generation models. With the rapid development of GAN, Diffusion, and DiT-based generators, whether these physical priors remain generally applicable requires further investigation.

Questions for authors
  • Compared with existing residual-, frequency-, and artifact-based AIGC detection methods, what is the key novel contribution of this work?
  • Can the current experimental settings sufficiently demonstrate the effectiveness of the method in realistic scenarios?
  • Are the selected physical priors applicable to newer generation models, such as Diffusion and DiT?
Reviewer_vXo1OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a learnable physical prior decoupling framework for generalizable AI-generated image detection. The motivation is that synthetic images respond differently from natural images to resampling, spatial smoothing and high-frequency operations. The authors describe these differences as physical defects shared across GANs and diffusion models. The input passes through three physical prior simulators (Resample, Blur, Edge). Each combines a fixed operator with a zero-initialized learnable residual modulation. The residuals are concatenated and reduced to the DWT HH subband. Then, they are re-weighted by channel attention and classified by a ResNet. The model is trained only on ProGAN images and their real counterparts. The model achieves comparable or SOTA performance across several benchmarks covering both GAN-based and diffusion-based generators. The authors claim that this cross-architecture generalization comes from exploiting shared physical defects rather than generator-specific artifacts.

Critical strengths
  • Despite training only on ProGAN, the method ranks first or second in mean accuracy on all four benchmarks, which cover both GAN and diffusion generators.
  • The design is simple and lightweight. It combines standard image processing operators with a small learnable modulation.
  • The ablation study demonstrates that combining resample, blur, and edge improves performance over individual simulators.
Critical weaknesses
  • The technical novelty is limited. Each branch targets a cue family already exploited in cited work: up-sampling traces (NPR), median-filter reconstruction residuals (FerretNet) and high-frequency cues (FreqNet, SAFE). The improvement over SAFE is small on most benchmarks, and SAFE performs better on Self-Synthesis.
  • The "physical defect" interpretation is not established. Responses to resampling, smoothing and high-frequency operations are signal-processing characteristics rather than established physical properties of image generation. GANs and diffusion models use different generation mechanisms, and cross-generator accuracy shows only that the cues are predictive, not that they come from a shared mechanism. The paper should define physical defects rigorously and rule out generic image statistics and dataset biases such as resolution, compression or file format.
  • Figure 2 does not establish why the selected operations should generalize. Figure 2(a) appears to show one image pair, and Figure 2(b) does not state which images or generators it uses. The low-pass probe also does not clearly correspond to the Laplacian-based Edge simulator.
  • Several comparisons are hard to interpret. Table 4 states all models are trained exclusively on ProGAN-generated images, yet F3Net, GramNet and DIRE score near-perfect on SD-family generators and near chance on most others, including BigGAN. Baseline provenance is not reported. Moreover, no variance across runs is reported despite several margins below one point.
Questions for authors
  • Figure 4 states that "None" uses no simulator. In that setting, is the DWT-HH step applied to the input image, or is the image fed to the ResNet directly?
  • Which images and generators are used for Figure 2? Does the gap in Figure 2(b) hold for each test generator, including the diffusion models?
  • What exact ResNet configuration is used to obtain the reported 1.44M parameters? Is it trained from scratch or initialized from pretrained weights?
  • For each baseline in Table 1-4, was the result reproduced, obtained from a released checkpoint or taken from the original paper? On what data was each one trained?
  • Images are resized to 256 x 256 (Section 4.1), which is itself a resampling operation. Since real and generated images may differ in native resolution and file format, how sensitive are the results to this preprocessing?
Reviewer_rzQ6OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a learnable physical-prior framework for generalizable synthetic image detection. Its main contributions are: 1) using resampling, smoothing, and spectral probes to motivate potentially shared forensic cues across generative models; 2) constructing three residual branches with fixed operators and learnable modulation, followed by DWT-based high-frequency extraction and attention-based fusion; and 3) evaluating a detector trained on ProGAN against multiple GAN and diffusion sources, with supplementary component ablations, plug-and-play experiments, and robustness tests under blur and resizing.

Critical strengths
  • The method combines explicit image-processing operators with learnable residual modulation and a compact classifier. This provides a practical way to incorporate several low-level cues within an end-to-end detection pipeline, with a reported parameter count of 1.44M.
  • The supplementary ablations report performance decreases when learnable modulation, the three simulators, or DWT is removed. The plug-and-play experiments also report average improvements for NPR, an FFT-based detector, and FerretNet. These results support the empirical usefulness of the components and their potential applicability beyond the authors’ own detector.
Critical weaknesses
  • Table 4 states that all methods are trained exclusively on ProGAN. However, its eight C2P-CLIP results exactly match a row in Table 3 of the original C2P-CLIP paper, whose caption specifies training on SD v1.4. This creates an unresolved discrepancy between the stated protocol and the apparent source of the results. Additionally, Table 1 compares averages over different evaluation subsets: PiD lacks StyleGAN2, whereas the proposed method’s average includes it. These issues make it difficult to interpret the reported advantages as comparisons under a common protocol.
  • The evaluation covers numerous classical GANs and earlier diffusion models, but provides no quantitative results for representative modern generators such as FLUX, PixArt, or SD3. Several evaluation settings also correspond to sampling variants of the same generator family rather than distinct architectures. Since the paper motivates its method through physical characteristics purportedly shared across heterogeneous generation mechanisms and explicitly discusses DiTs, evidence on these additional mechanisms would be particularly important. Existing benchmarks such as DiTFake make this a targeted validation request rather than a requirement to evaluate every newly released model.
  • The single-branch, pairwise, and component ablations support the usefulness of the proposed combination. However, they do not establish whether the fully trained model consistently exploits distinct cues from all three branches, relies predominantly on one branch, or changes its reliance across generators and image conditions. The Grad-CAM comparisons do not directly resolve these alternatives. Moreover, the observed residual and spectral differences have not been sufficiently separated from potential effects of image content, compression history, and resolution. Stronger evidence would help connect the claimed shared physical defects to the actual behavior of the learned detector.
Questions for authors
  • Could the authors clarify whether the results in Table 4 were reproduced using ProGAN-trained models or taken from published evaluations with other training sources? In particular, how should the matching C2P-CLIP results and the different training-source descriptions be reconciled? Please identify the training source and result provenance for each method, and report common-subset averages where evaluation entries are missing.
  • Could the authors characterize branch reliance in the fully trained model across generators, real/fake classes, and perturbation conditions? Per-sample gating distributions, together with branch interventions in a fixed model, would help distinguish complementary use from persistent dominance or sample-dependent specialization. A comparison between dynamic and static fusion would also clarify whether input-dependent weighting contributes materially to the gains. Attention weights alone should be interpreted alongside their downstream effects.
  • Could the authors provide the sampling and preprocessing protocol for the physical probes in Figure 2, including the generators, sample counts, and controls for content, compression, and resolution? Do the observed differences and detection gains persist under matched compression and resolution conditions? This would help distinguish generation-related cues from dataset-specific processing differences.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.