acceptodds
Under review as a conference paper at ICLR 2027

FoRM: Preserving First-Order Real-Manifold Geometry for Generalizable AI-Generated Image Detection

Abstract

As generative models continue to evolve rapidly, generalization to unseen generators has become a central challenge in generalizable AI-generated image detection. Recent reconstruction-based, real-centric envelope learning methods construct near-real fake samples to characterize the outer boundary of the real manifold, offering a promising approach to cross-generator generalization. However, these methods do not sufficiently guarantee a key prerequisite of envelope learning—the stability of the real manifold enclosed by the envelope. Once the real manifold becomes distorted during training, even a well-characterized outer boundary cannot compensate for the resulting discriminative bias caused by the degradation of its internal structure. To address this limitation, we propose FoRM, a first-order real-manifold preservation framework. Specifically, FoRM performs pointwise feature anchoring against a frozen visual foundation model to establish zeroth-order consistency in real representations. Nevertheless, pointwise consistency alone is insufficient to determine the local shape and geometric directions of sample neighborhoods. We therefore introduce Local Directional Jacobian Alignment (LJ), which matches the local responses and their magnitudes between the teacher and detector along natural image variation directions. LJ extends pointwise consistency to local first-order geometric consistency, thereby mitigating local geometric contraction and distortion of the real manifold. Furthermore, FoRM introduces distributionally diverse real samples to prevent real-side optimization from persistently concentrating on homogeneous local regions, enhancing the structural stability of the real manifold from a data-centric perspective. It also employs heterogeneous reconstruction mechanisms to construct diverse near-real fake samples to strengthen outer-boundary supervision. By jointly modeling the internal geometry of the real manifold and its surrounding discriminative envelope, FoRM achieves competitive cross-generator generalization performance across 12 benchmarks.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
AI-Generated Image Detection, AI Content Safety, Multimedia Forensics

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_D3QhOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper argues that existing AI-generated image detection methods primarily focus on refining the discriminative boundary surrounding the real-image manifold, while overlooking potential distortions of the manifold itself during detector training. To address this issue, the authors propose FoRM, which leverages a frozen visual foundation model to preserve the feature representations and local geometric structure of real images during detector adaptation, thereby improving the accurate recognition of real images, including those subjected to post-processing.

Critical strengths
  1. This paper is well organized and easy to follow.

  2. Strong and consistent empirical performance across benchmarks.

Critical weaknesses
  1. This paper lacks a quantitative evaluation of the manifold consistency between clean real images and their post-processed counterparts. Therefore, it is difficult to determine whether the proposed alignment mechanism effectively preserves the geometry of the real-image manifold.

  2. Although the proposed method alleviates the performance degradation caused by image post-processing to some extent, the robustness evaluation remains insufficient. The limited variety of post-processing operations makes it difficult to fully assess the method's robustness under more challenging real-world conditions.

  3. The paper lacks a computational efficiency analysis. Given the dual-branch DINOv3 architecture, the authors should report training time, computational overhead, and resource consumption to better assess the practical cost of the proposed method.

  4. The paper motivates its approach by arguing that FakeAcc is nearly saturated in existing envelope-learning methods. However, a major challenge in cross-generator detection is that unseen generated images are often misclassified as real. The proposed manifold-preservation constraints are applied exclusively to real images, making it unclear how they improve discrimination against unseen generators rather than simply reducing false positives on real images. The authors should provide a clearer explanation of how preserving real-manifold geometry contributes to cross-generator generalization.

  5. The ablation studies and post-processing robustness experiments are essential for validating the main contributions of this paper. However, both are relegated to the appendix. I believe these key experiments and their analyses should be presented in the main paper rather than the supplementary material.

Questions for authors

Please refer to the Weaknesses section above for detailed questions and concerns, particularly regarding the effectiveness of manifold alignment, post-processing robustness, computational efficiency, and the presentation of key experimental results.

Reviewer_xs3MOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors observe that the errors of existing AI-generated image detectors mainly come from real images being misclassified as fake. Using real images, transformed versions of real images, and generated images reconstructed from real images, the authors train FoRM, an AI-generated image classifier with DINOv3-L as the backbone, which improves the feature distribution in the vicinity of real images so that it is closer to the feature distribution of the backbone. Experiments show that FoRM achieves the highest average accuracy on 12 datasets and markedly improves the classification accuracy on real images.

Critical strengths
  1. The paper includes an adequate amount of experiments and follows the experimental setups of recent literature, which validates the effectiveness of the method.
  2. The paper clearly points out the shortcomings of current AI-generated image detection work in classifying real images, which is instructive.
Critical weaknesses

According to Eqs. (9), (10), and (11), LJ constrains Δθ(k)\Delta_\theta^{(k)} (the change in the classifier's features of a real image caused by a perturbation) so that its direction and magnitude are close to those of ΔT(k)\Delta_T^{(k)} (the change in the teacher model's features under the same perturbation). However, this idea has already been used in existing work, e.g., REM, which this paper cites and compares against: the LresL_\text{res} in that paper is a direct constraint on Δθ(k)−ΔT(k)\Delta_\theta^{(k)} - \Delta_T^{(k)}. In addition, many designs in the proposed method are quite similar to those of REM (e.g., hard fake samples, perturbations of real images, and the two-model training structure), and the authors should analyze in detail which specific modifications in FoRM lead to the performance improvement.

Questions for authors
  1. Architecture clarification: What are the architectures of WW and the MLP in Figure 4?
  2. Backbone ablation: Why DINO was chosen as the backbone, and how would the method perform with other self-supervised backbones (e.g., CLIP)?
  3. Hyperparameters: The paper does not report the weighting coefficients of the different terms in the loss. Is the performance of FoRM sensitive to them?
Reviewer_rjxXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FoRM, a framework for generalizable AI-generated image detection that focuses on preserving real-image representations during detector fine-tuning. The method combines pointwise feature anchoring with Local Directional Jacobian Alignment, which matches the feature responses of a frozen teacher and the fine-tuned detector under local image transformations. The authors also introduce heterogeneous real-image sampling and multiple reconstruction mechanisms to improve the coverage of real and near-real fake samples. Experiments across 12 benchmarks show strong cross-generator generalization, particularly in real-image classification accuracy.

Critical strengths
  1. The paper investigates an interesting aspect of real-centric detection. Rather than focusing only on constructing better fake samples, it considers how the representation of real images changes during fine-tuning.

  2. The authors provide separate real and fake accuracy results, which help explain where the performance improvements come from. The ablation study also shows that LJ provides additional gains beyond pointwise feature anchoring.

  3. The experimental evaluation covers different generator architectures and real-world datasets.

Critical weaknesses
  1. The connection between the proposed objective and manifold geometry preservation is not fully convincing. The main contribution is motivated by preserving the first-order geometry of the real-image manifold. However, LJ approximates local Jacobian responses using a small number of image transformations, including color adjustments and blurring. It is unclear whether these transformations represent tangent directions of the real-image manifold, rather than simply predefined perturbation directions. Matching feature displacements along these directions does not necessarily imply that the broader local manifold geometry is preserved.

  2. Stronger baselines are needed to establish the benefit of the proposed LJ objective. Table 15 shows that LJ improves performance when combined with pointwise feature anchoring. However, it remains unclear whether this improvement comes specifically from the proposed first-order alignment or from more general consistency regularization.

  3. The empirical evidence does not directly establish that manifold distortion is reduced. The motivation relies heavily on the claim that fine-tuning distorts the real manifold and that preserving its local geometry improves generalization. The reported improvements in RealAcc are consistent with this explanation, but classification accuracy alone does not establish that the underlying representation geometry is better preserved.

  4. The contribution of heterogeneous real sampling is not completely isolated. The two experimental settings are useful, especially the setting that selects samples only from MS-COCO. However, selecting images based on low-level statistics may also alter other properties of the training distribution, including semantic content and image quality. A comparison with randomly selected samples under matched data budgets would help clarify whether the improvement comes from the proposed selection criteria rather than simply changing the real training distribution.

Questions for authors
  1. How sensitive is LJ to the choice and magnitude of the image transformations? Does the method still provide similar improvements with a different set of perturbation directions?

  2. Can the authors provide evidence that local geometric distortion is reduced, independently of the final classification accuracy?

  3. Does the learnable projection affect the geometry of the original student feature space, and how sensitive are the results to this projection?

  4. How does heterogeneous real sampling compare with random sampling using an identical number of real images?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.