acceptodds
Withdrawn from ICLR 2027

GlobalForge: Towards Robust AI-Generated Image Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

AI-generated image (AIGI) detectors achieve strong accuracy on clean benchmarks, but their performance drops sharply after images are propagated through real-world channels. We trace this fragility to what these detectors actually learn: they overfit to local artifacts left by generators in small spatial neighborhoods, which are easily destroyed by common propagation degradations such as JPEG compression and blur. Instead, we shift the discriminative cue from fragile local artifacts to more robust global structure. Building on this, we propose GlobalForge, a framework with two complementary modules. The Local Information Bottleneck (LIB) suppresses local components to block shortcut learning, while the Global Structural Reasoning (GSR) module forces every token to gather evidence from distant regions. Both modules are trained jointly under a contrastive structural loss based on degradation that keeps the resulting features stable under degradation. To support fine-grained robustness evaluation, we further introduce RealDeg-Bench, covering 7 common degradation operators and multi-step compound chains. GlobalForge improves average BAcc on 8 in-the-wild benchmark groups by 5.89% over the previous state-of-the-art, and is clearly ahead of representative baselines on RealDeg-Bench under both single and compound degradations. Code is available at https://anonymous.4open.science/r/GlobalForge-BE0F/.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
AIGI detection, Robustness

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

The paper studies the robustness of AI-generated image detection under image degradations. GlobalForge combines a Local Information Bottleneck, which smooths backbone features through a learnable gate, with Global Structural Reasoning, which masks nearby tokens during attention. A degradation-aware contrastive objective aligns clean and degraded representations. The authors also introduce RealDeg-Bench for single and compound degradations. Experiments report improvements across eight in-the-wild benchmark groups, supported by component ablations and comparisons with alternative filtering and augmentation strategies.

Critical strengths

1.The evaluation addresses an important practical limitation of existing detectors. RealDeg-Bench complements the in-the-wild evaluation by controlling degradation types and chain lengths, making robustness failures easier to characterize. The reported improvements extend across both evaluation suites, providing more convincing evidence of practical utility than results on clean datasets alone.

  1. The controlled ablations provide useful support for the proposed design. Table 3 compares the components under the same backbone and training setup, while Table 4 examines augmentation, input filtering, and unrestricted attention as alternatives. These comparisons strengthen the empirical contribution by showing that the gains are not explained solely by the pretrained backbone or standard degradation augmentation.
Critical weaknesses
  1. The evidence for improved robustness is stronger than the explanation that the model blocks local shortcuts. LIB retains an unfiltered feature component, and Equation 4 preserves the input through a residual connection. Moreover, the backbone tokens already contain globally contextualized information. Masking nearby attention positions therefore does not establish exclusive reliance on global structure. The ablations support the design’s effectiveness, but the stronger mechanistic claims in Section 4.2 should be qualified.

  2. The metric used for the CO-SPY subsets in Table 2 needs clarification. The official CO-SPY release describes its in-the-wild data as synthetic samples, whereas the table labels all results as balanced accuracy. It is unclear which real images, if any, are included in these subsets. Since the CO-SPY group contributes to the headline average, the paper should distinguish synthetic-class accuracy from balanced accuracy and explain how the group scores are computed.

Questions for authors
  1. In Table 2, do the five CO-SPY subsets contain both real and synthetic images? Please specify the real-image source, decision thresholds, and metric calculation. If these entries measure synthetic-class accuracy, how does reporting them separately affect the eight-group comparison?
  2. What evidence distinguishes increased reliance on global structural relationships from the effects of feature smoothing and degradation regularization? In particular, how should the claim of blocking local shortcuts be interpreted given the unfiltered component in Equation 1 and the residual connection in Equation 4?
  3. Appendix D states that the contrastive representation is pooled from the final backbone layer. Is this representation computed before or after LIB and GSR? Please clarify which modules receive gradients from the contrastive loss, since this affects the interpretation of their joint optimization.
Reviewer_Tm9xOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper claims that typical AI-generated image detectors fail after real-world image processing operations because they rely on local generator-specific artifacts. It proposes GlobalForge, an AI-detector model based on a LoRA-tuned DINO (v2/v3) encoder with a few additional architectural modifications designed to reduce detection shortcuts caused by memorization of generator-specific high-frequency patterns. Specifically, it blends Gaussian-smoothed copies of feature representations into the backbone outputs via a learnable gate parameter (Local Information Bottleneck) and forces the model to learn distant spatial relationships by masking each query’s local window in the attention operation (Global Structural Reasoning). The training objective also adds a contrastive loss term to align representations of clean and degraded variants of the same image alongside the traditional classification loss. To evaluate the model in a robust setting, the paper also introduces RealDeg-Bench consisting of ~7.3k clean images degraded with 7 simple distortions (JPEG, Gaussian blur, etc.) and their random chains of length 1 to 5. GlobalForge is compared against 12 detectors (6 off-the-shelf, 6 retrained) on RealDeg-Bench and 8 prior benchmarks: the reported increase in average balanced accuracy is around 4-6 percentage points relative to the best prior detector. The paper also reports several component and alternative-design ablations, as well as Grad-CAM visualizations.

Critical strengths

S1. Reported empirical results are rather strong, and provide a broad comparison with other recent methods. According to the provided results, GlobalForge with DINOv3 achieves the highest average balanced accuracy on both author-curated RealDeg-Bench and on the 8 external benchmarking datasets. This could be the main argument towards acceptance: if confirmed under a more complete set of performance metrics (see W2.1) and better architecturally justified (see W2.2), the method could be a practically useful detector. S2. Proposed depth-resolved evaluations provide valuable insights into the robustness of recent models, several of which degrade sharply as the distortion chain grows. This perspective is useful beyond the proposed detector, even though evaluation through compound distortions in itself is not a new concept (see W1.1).

Critical weaknesses

W1. The central robustness claims of the paper are largely overstated. They are based on a rather narrow, mostly in-distribution evaluation. More specifically, this weakness can be subdivided into 3 separate concerns: W1.1 The distortion set of RealDeg-Bench is too narrow to support the claim of measuring robustness under “realistic propagation processes”. The operator pool has only 7 simple global operations, that are chained in uniformly random order. Many common real-world transformations are not covered: e.g., other compression methods such as JPEG 2000, WebP, HEIF, neural codecs, multiple recompression scenarios; geometric transformations such as cropping and rotation, and many other potential degradation types. Uniformly random ordering is also not justified, since it does not reflect actual platform pipelines. Moreover, RealDeg-Bench's positioning against prior work and its novelty also raise questions. Real-chain, which the paper itself evaluates, already applies random chains of operations (though mostly related to platform processing). Some works have already used more sophisticated distortion groups that go far beyond JPEG and other simple transformations. For instance, the NTIRE 2026 challenge on AI-image detection in the wild [1] also employs degradation chains with 36 distortion types. The paper also does not discuss established degradation pipelines from image restoration (e.g., Real-ESRGAN [2]). The novelty of compound evaluation is therefore limited, and RealDeg-Bench’s contribution lies mainly in the depth-resolved reporting (S2). W1.2 According to LL273–274, most RealDeg-Bench distortions (5/7) are also used during training, which makes the evaluation largely in-distribution for GlobalForge detector. All compound chains draw transforms from the same pool that contains degradations seen during training. The reported results therefore mainly show invariance to the training augmentations, not robustness to unseen degradations. The paper also does not state whether the 6 retrained baselines received the same augmentations: if they didn’t, the comparison unfairly favors GlobalForge. A properly controlled distortion protocol, where training augmentations and test distortions are kept disjoint, is needed to support the paper's key robustness claim. W1.3 The claimed control over distortion strength is limited to operator parameters. LL287-8 suggests that distortions in other benchmarks are not strength-controlled, while RealDeg-bench introduces such control and analyses its impact; however, the paper never quantifies how much these settings actually degrade the images; it only states the used hyperparameter ranges. Parameter values are not comparable across operators; therefore, some of them could be disproportionately stronger than others. To make operator severity comparable, report image-quality measures for each condition and chain depth (e.g., PSNR, SSIM, LPIPS, or subjective studies). It would also show whether the chosen parameters could match realistic platform processing scenarios. W2. The second critical issue relates to the evaluation protocol design: W2.1. The only detection quality metric used throughout the paper is balanced accuracy at a fixed threshold, with no AUC/AP and no per-class (real/fake) accuracy. For a task that essentially reduces to binary classification, this is a serious omission: most AI-image detection work reports threshold-free metrics or at least per-class accuracy (e.g., AP in AIGIBench and Real-chain benchmarks, both of which this paper reports). For instance, balanced accuracy alone could not clarify how exactly the degradations influence the detector predictions: do they systematically shift scores on fake images towards real class, or vice versa, or is the effect non-directional? Furthermore, accuracy relies on binarized predictions that are heavily influenced by the decision threshold selection. Some detectors may have uncalibrated output scores or non-trivial probability distributions, so the default 0.5 threshold (presumably used in the paper) may be severely suboptimal and thus understate baseline performance, particularly for those evaluated with official weights. Evaluations also raise validity concerns: suspiciously, Table 2 reports small or near-0 B.Acc. values for Aligned and C2P-CLIP detectors on CO-SPY, meaning that these detectors would have achieved near-perfect accuracy if inverted. W2.2. Current evidence cannot separate the method’s contribution from metric variance and backbone choice. All reported evaluations rely on single-seed experiments without statistical significance measurements, while many differences presented as evidence seem small (1-3%) compared with the larger gaps reported within the same experiments. For instance, results in Table 1 suggest that backbone selection (DINO v2 vs v3) has a significant impact on performance: GlobalForge with DINO-v2 has the largest drop in accuracy between clean and 5-step (89 % to 69%), while for v3 the gap is narrower (~8%). This performance gap between v2 and v3 is larger than differences reported in all ablation and sensitivity experiments. Other backbone families are also not discussed. Additionally, multi-step chains in RealDeg-Bench rely on random sampling of distortions and their strengths, yet the variance across different chains with the fixed step count is not shown or discussed in any way. [1] NTIRE 2026 Challenge on Robust AI-Generated Image Detection in the Wild, CVPR Workshop, 2026 [2] Real-ESRGAN: Training Real-World Blind Super-Resolution with Pure Synthetic Data, ICCV Workshop, 2021

Questions for authors

Please see the Weaknesses section for key concerns to address first. Other questions:

  1. From T2I-CoReBench’s page (lioooox/T2I-CoReBench-Images at huggingface), it seems that it has no real images, so where do the real images come from in RealDeg-Bench?
  2. Did the retrained baselines receive the same augmentations as GlobalForge?
Reviewer_GWCXOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper argues that AI-generated image detectors degrade in performance under real-world propagation such as compression, blur and resizing, because they rely on local generator artifacts. It proposes GlobalForge to move the decision toward global image structure. GlobalForge adds two feature-space modules on top of a LoRA finetuned DINO backbone. The Local Information Bottleneck (LIB) blends patch features with a Gaussian-smoothed version through a learnable gate α. Global Structural Reasoning (GSR) is an attention layer that masks each token's local neighborhood. A contrastive loss (DCS) aligns clean and degraded views of each image. The authors also introduce RealDeg-Bench, which covers seven common degradation operators and multi-step compound chains. The authors report two variants, one on DINOv2 and one on DINOv3. Among the evaluated methods, the DINOv3 variant reaches the best average balanced accuracy on eight in-the-wild benchmark groups, ahead of the strongest baseline DDA, and the best average on RealDeg-Bench.

Critical strengths
  • The problem is practically important. Robustness to compound propagation degradations is a real deployment gap, and the paper ties it to a concrete, testable hypothesis about which discriminative cues detectors use.
  • The in-the-wild evaluation is broad, with eight benchmark groups and twelve baselines.
  • The method keeps a simple single-forward-pass inference path, and a plain finetuned backbone trained on the same data shows that the proposed components add a substantial gain.
  • The ablations include the meaningful alternatives, such as augmentation only, input-side suppression and unmasked global attention.
Critical weaknesses
  • Robustness depends strongly on the model configuration. The DINOv2 variant shows larger drops than others under blur, noise and five-step compound degradation, and falls below B-Free and GAPL at five steps (Table 1). The plain backbone control and component ablations cover only the DINOv3 variant. The source of this gap is unclear, as the two variants also use different learning rates and epochs.
  • The shift from local artifacts to global structure is not directly measured. Grad-CAM maps are qualitative, and Figure 2 assumes that resizing disrupts only spatially local cues, despite the fact that it also changes image scale and removes high frequencies across the whole image.
  • Masking the local window does not ensure the local evidence is excluded, as the backbone tokens already attend globally and Eq. 4 keeps a residual path.
  • The ablations do not separate the structural modules from degradation-invariance training. The components seem to interact strongly, since DCS adds almost nothing alone but substantially on top of LIB and GSR (Table 3).
  • The DCS augmentation also overlaps most RealDeg-Bench operator families, and it is unclear whether the retrained baselines received comparable augmentations.
  • Several evaluation results need clarification. The JPEG row in Table 1 is nearly identical to Clean for almost all methods, despite the declines in Figure 5. In Table 5, several near-chance baselines score higher after degradation, and it is unclear whether degraded accuracy is class-balanced. Also, the original source of RealDeg-Bench's real images is not described.
  • GlobalForge is reported from a single run, while ablation and sensitivity differences are only a few points. Without repeated runs, it is unclear whether these differences exceed seed-to-seed variance.
Questions for authors
  • Can you report a plain DINOv2 baseline trained like the DINOv3 baseline in Table 3 together with the key component ablations under the DINOv2 setup?
  • What gate values does LIB learn in each variant? Given the globally contextualized backbone tokens and the residual connection in Eq. 4, what evidence distinguishes reliance on global structure from other benefits of the architectural constraint?
  • How was JPEG compression applied in Table 1, and why does the JPEG row match the Clean row for almost all methods?
  • In Table 5, is degraded accuracy class-balanced, and why do several baselines score higher after degradation?
  • What degradation augmentation did the retrained baselines receive? Can you report results separately for operator families included in the DCS augmentation and those absent from it, i.e., resizing and Gaussian noise?
  • In which features is DCS applied? Please also state the exact settings used for the reported results, such as the masking window and the DCS weight.

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

Related papers

Discussion (0)

Sign in to comment.