acceptodds
Under review as a conference paper at ICLR 2027

COMPASS: Contextual Moment Prior Alignment for Generalizable AIGI Detection

Abstract

Generalizable AI-generated image (AIGI) and deepfake detection remains challenging due to the asymmetry phenomenon, where adaptation to limited training distributions can degrade the transferable representations of Vision Foundation Models (VFMs). Recent subspace adaptation methods mitigate this issue by freezing high-magnitude principal components while adapting the residual subspace. However, such magnitude-based decomposition is data-agnostic and does not explicitly account for target-relevant semantics, potentially leading to semantic misalignment. We propose COMPASS (COntextual Moment Prior Alignment for Subspace Separation), a data-aware subspace adaptation framework that uses a small unlabeled Semantic Proxy Set to estimate target activation statistics. COMPASS incorporates activation covariance into the decomposition of pre-trained weights, preserving target-relevant representations as a stable semantic reference while the residual adapter learns forgery-related variations. The proxy calibration requires no gradient-based optimization. Extensive experiments demonstrate strong generalization across cross-dataset, cross-method, joint cross-dataset-method, and open-world cross-generator evaluations, achieving state-of-the-art performance on both Deepfake and AIGI detection benchmarks. Source code and trained models will be publicly released.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
AI-Generated Image Detection, Domain Generalization, Orthogonal Subspace Decomposition, Vision Foundation Models, Parameter-Efficient Fine-Tuning

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_34bdOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes COMPASS, a data-aware subspace adaptation framework for generalizable AI-generated image and deepfake detection. Building upon EFFORT, the method introduces an unlabeled Semantic Proxy Set to estimate activation covariance, which guides the decomposition of pretrained VFM weights into frozen principal components and trainable residual components. Experiments on cross-dataset, cross-method, joint cross-dataset-method, and cross-generator detection benchmarks demonstrate promising generalization performance. Additional ablations investigate proxy-set selection, covariance guidance, and different VFM backbones.

Critical strengths
  1. Strong empirical performance across diverse evaluation settings. The method demonstrates improvements over competitive baselines, particularly under joint dataset-method shifts and cross-generator evaluation. These results suggest that the proposed calibration strategy can improve the generalization of forgery detectors.

  2. Simple and parameter-efficient adaptation framework. Incorporating activation statistics into pretrained weight decomposition is an intuitive extension of existing subspace adaptation methods. The calibration requires no gradient-based optimization, and the method retains the parameter efficiency of low-rank residual adaptation.

Critical weaknesses
  1. Insufficient theoretical justification The response-preservation objective in Eq. (3) naturally leads to covariance-root weighting, yet the proposed method empirically adopts the full covariance matrix without a corresponding theoretical justification. More importantly, Eqs. (8)–(10) require inverting the covariance matrix, which may be rank-deficient when estimated from a small proxy set. The paper does not clearly explain how this issue is handled or whether regularization affects the resulting decomposition. These concerns weaken the theoretical foundation and reproducibility of the central method.

  2. Unclear generalization protocol and potential dependence on target-domain information. COMPASS relies on a Semantic Proxy Set sampled from the target-category distribution, whereas the compared baselines do not appear to receive equivalent target-distribution information. Although the authors state that test samples are excluded, the precise relationship between the proxy sources and evaluation domains remains unclear. Consequently, it is difficult to disentangle the benefits of covariance-guided decomposition from those of additional target-domain exposure, making the fairness of the generalization comparison uncertain.

Questions for authors
  1. How is the inverse covariance matrix computed when the estimated covariance is rank-deficient? Is regularization or a pseudoinverse employed, and how sensitive are the results to this choice?

  2. Can the authors clarify the exact construction of the Semantic Proxy Set for each evaluation protocol? In particular, does it contain samples from the target evaluation domains or manipulation categories? How does COMPASS perform when the proxy set is restricted to the source training distribution, and how does EFFORT perform when given equivalent target-domain information?

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

The paper argues that prior data-agnostic SVD decomposition of VFM weights may inadequately preserve task-relevant representations, impairing transfer to AIGI and Deepfake detection. It proposes COMPASS, which uses activation covariance from an unlabeled proxy set to guide decomposition, freezing principal components while adapting residuals. The paper validates the method’s generalization and robustness through evaluations under multiple settings and ablation studies on the proxy set, guidance matrix, and backbone.

Critical strengths
  1. Insightful limitation and simple improvement. The paper shows that EFFORT’s[1] magnitude-based decomposition may inadequately preserve target-relevant representations and introduces lightweight covariance-guided calibration to address this issue.
  2. Validation across two detection tasks. Evaluations on both deepfake and AIGI detection show gains under joint distribution shifts and cross-generator settings. This breadth strengthens confidence in the framework’s applicability beyond a single task.

[1] Orthogonal Subspace Decomposition for Generalizable AI-Generated Image Detection. (ICML 2025)

Critical weaknesses
  1. Limited methodological novelty. Prior studies, such as CorDA [1], have already proposed data-aware decomposition and principal-component preservation methods that are conceptually similar to those used in COMPASS. In this sense, COMPASS can be viewed primarily as adapting this paradigm to VFM-based Deepfake detection and AIGI detection tasks.
  2. Insufficient explanation of generalization gains. Figure 1 measures principal-subspace feature retention before fine-tuning, but does not establish that better retention drives generalization rather than changes in residual optimization.
  3. Unclear scope of proxy-robustness conclusions. Section 4.4.1 evaluates proxy-set robustness only for facial deepfake detection. Although the effect of AIGI proxy-set size is examined, corresponding analyses of proxy-set composition and data source are absent, limiting confidence in whether these conclusions generalize to semantically diverse AIGI detection settings.

[1] Context-Oriented Decomposition Adaptation of Large Language Models for Task-Aware Parameter-Efficient Fine-tuning. (NIPS 2024)

Questions for authors
  1. Beyond the feature-retention analysis in Figure 1 and the qualitative Grad-CAM analysis in the appendix, can you provide quantitative analysis to further explain the mechanism behind COMPASS’s improved generalization in forgery detection?
  2. Does proxy-set selection substantially affect COMPASS’s performance on AIGI detection? The proxy set is sampled from the AIGIBench training set, which covers only car, cat, chair, and horse. How does COMPASS perform with this proxy set on a more diverse benchmark such as Chameleon[1]?
  3. Have you compared COMPASS with more recent state-of-the-art AIGI detectors, such as OmniAID[2], PROBE[3], and LoRC[4]?
  4. Could you clarify whether COMPASS and the compared methods, particularly EFFORT, use the same training data in the AIGI detection experiments?

[1] A Sanity Check for AI-generated Image Detection. (ICLR 2025)

[2] OmniAID: Decoupling Semantic and Artifacts for Universal AI-Generated Image Detection in the Wild. (ICML 2026)

[3] Where Detectors Fail: Probing Generative Space for Generalizable AI-Generated Image Detection. (ICML 2026)

[4] LoRC: Detecting AI-Generated Images via Low-Rank Collapse in Semantic Residuals. (ECCV 2026)

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

This paper proposes COMPASS, a proxy-guided subspace adaptation method for generalizable AI-generated image and deepfake detection. It estimates activation covariance from a small unlabeled proxy set, uses it to guide weight decomposition, freezes the principal component, and adapts a low-rank residual. Experiments cover cross-dataset, cross-method, joint domain/method shifts, and AIGIBench. The paper reports 0.829 average AUC versus 0.710 for EFFORT under the joint-shift protocol, and 90.9% average AP across 25 AIGIBench generators. Additional analyses examine proxy selection, guidance statistics, backbone choice, and image corruptions.

Critical strengths
  1. The paper addresses a relevant limitation of weight-only subspace selection: directions important for reconstructing weights need not preserve responses on the data of interest. The small-proxy calibration strategy offers a practical way to incorporate distribution information.

  2. The evaluation spans several distribution shifts and multiple foundation-model backbones. Proxy-source, composition, size, and guidance-matrix ablations provide useful evidence about the conditions under which calibration helps.

  3. Calibration is performed offline without backpropagation, and the residual adapter can be merged into the frozen weights for inference. The paper also distinguishes its empirical use of full covariance from the covariance-root weighting suggested by its derivation.

Critical weaknesses
  1. The proxy protocol makes the generalization claims difficult to interpret. Table 7 calibrates on training samples from manipulation methods also included in evaluation, whereas the baselines use source-trained models. Excluding test images avoids direct test-sample leakage but still permits target-method information. Please identify proxy sources for each main result, report a source-only proxy control, and compare against a target-aware baseline with the same unlabeled-data access.

  2. Covariance inversion is not adequately specified. Under Section 3.2's image-vector definition, 80 centered observations cannot produce an invertible covariance in a higher-dimensional layer; a single centered observation gives zero covariance, despite the reported one-shot result. If patch-token pooling or numerical regularization is used, please specify it and verify reconstruction of the original weights at initialization. This affects both reproducibility and the validity of the stated decomposition.

  3. Figure 1 evaluates the principal component before fine-tuning, while the proposed benefit concerns preserving useful representations during adaptation. It remains unclear how well this semantic reference is retained in the final detector. Please include a post-fine-tuning feature-retention comparison to better support the explanation for improved generalization.

The results are promising, but target-data access and the implementation of the core decomposition remain insufficiently specified. The evidence for semantic preservation during adaptation is also limited. These issues currently limit confidence in the interpretation of the reported generalization gains.

Questions for authors
  1. How much of the improvement remains when calibration uses only source-domain data, with all evaluated target generators excluded from the proxy set?

  2. Why do Table 3 and Appendix Table 6 report 90.9% average AP, while Figure 3 and Table 5(a) report 89.0% for the default AIGI proxy size? Please identify any differences in evaluation sets, checkpoints, or training configurations.

Reviewer_bmQLOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper targets AIGI/Deepfake detection methods such as EFFORT, which partition subspaces according to singular value magnitudes. It points out that such partitioning is data-agnostic and may retain directions that are unimportant for the target data. To address this issue, COMPASS is proposed: a small set of unlabeled target-class images (Semantic Proxy Set) is used to estimate the covariance matrix C of input activations at each layer. After performing SVD on WC, the principal components are frozen while rank‑1 residual components are trained. The entire calibration pipeline requires no backpropagation. Experiments cover three protocols for deepfake evaluation: cross-dataset, cross-method, and joint shift, together with the 25 generators in AIGIBench. Under the joint-shift setting, the AUC improves from 0.710 to 0.829, and the average AP on AIGIBench reaches 90.9%.

Critical strengths
  1. The method is lightweight with nearly no extra training overhead. Weights can be merged at inference time with no additional cost.
  2. It achieves substantial gains under challenging settings including cross-method and joint shift. Performance improvements grow with shift difficulty, forming a coherent story.
  3. Comprehensive ablation studies are conducted, covering the source and size of proxies, real/fake ratio, forms of the guiding matrix, and different backbone networks, yielding stable results.
Critical weaknesses
  1. Limited novelty. The core procedure (SVD on WC, restoration via right multiplication by C−1, freezing principal components, and adapter initialization using minor components) largely aligns with the knowledge-preserved paradigm of CorDA (NeurIPS 2024), which is not cited in the paper.
  2. The theoretically derived optimal weighting is C^1/2, yet C is adopted in practice without a theoretical justification.
Questions for authors
  1. Limited novelty. The core procedure (SVD on WC, restoration via right multiplication by C−1, freezing principal components, and adapter initialization using minor components) largely aligns with the knowledge-preserved paradigm of CorDA (NeurIPS 2024), which is not cited in the paper.
  2. The theoretically derived optimal weighting is C^1/2, yet C is adopted in practice without a theoretical justification.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.