COMPASS: Contextual Moment Prior Alignment for Generalizable AIGI Detection

Zhan Yang; Tiancheng Shen; Weijie Lyu; Shuo Chen; Bernard Ghanem; Yibo Yang; Ming-Hsuan Yang

cs.CVOct 13, 2026

COMPASS: Contextual Moment Prior Alignment for Generalizable AIGI Detection

Authors: Zhan Yang, Tiancheng Shen, Weijie Lyu, Shuo Chen, Bernard Ghanem, Yibo Yang, Ming-Hsuan Yang

Organizations: Zhejiang University · University of California, Merced · KAUST · Shanghai Jiao Tong University

Abstract

Generalizable AI-generated image (AIGI) and deepfake detection remain 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.

Figures & tables

Figure 1: Pilot study on feature retention prior to fine-tuning. (a) PCA visualizations of intermediate features across 1K images. Magnitude-based truncation in EFFORT increasingly alters the original feature geometry in deeper layers, with the average cosine similarity to the original CLIP representation dropping to 0.26 in Layer 24, compared with 0.61 for COMPASS. (b) Pairwise similarity density plots using only the principal subspace ( $\mathbf{W}_{prin}$ , before fine-tuning). Each point compares the cosine similarity of the same image pair in the original CLIP and principal-only spaces. COMPASS better preserves the relative similarity structure, achieving a higher Spearman rank correlation ( $\rho_{s}=0.5533$ ) than EFFORT ( $\rho_{s}=0.1345$ ). Since truncation and feature re-normalization may shift absolute cosine similarities, we focus on rank preservation rather than proximity to the diagonal; further analysis is provided in section A.1 .

Figure 2: Overall pipeline of COMPASS. Left: A small unlabeled Semantic Proxy Set is used to estimate the activation covariance $\mathbf{C}$ . Middle: The pre-trained weights are factorized via covariance-guided SVD into principal weights $\mathbf{W}_{prin}$ and residual weights $\mathbf{W}_{res}$ . Right: During fine-tuning, $\mathbf{W}_{prin}$ is frozen as a stable semantic reference, while the residual adapter $\Delta\mathbf{W}$ , initialized from $\mathbf{W}_{res}$ , is optimized to learn forgery-related variations.

Panel of Figure 3 — Figure 3: Sensitivity to Proxy Set Size ( $N_{p}$ ). COMPASS improves with additional proxy set samples and gradually stabilizes on both (a) Deepfake Detection and (b) AIGI Detection. The facial domain requires substantially fewer samples than the more semantically diverse open-world AIGI setting.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 4: Grad-CAM Visualization on Deepfake Detection. Left: Input fake image. Middle: EFFORT exhibits relatively dispersed attention across facial and contextual regions. Right: COMPASS places more attention on the central facial region.

Generator	CNN-Spot	UnivFD	FreqNet	NPR	Ladeda	AIDE	C2P-CLIP	DFfreq	SAFE	Effort	NS-Net	FerretNet	LOTA	BSF	Ours
ProGAN	99.9	99.9	100.0	100.0	100.0	99.6	100.0	100.0	100.0	100.0	100.0	100.0	99.9	100.0	100.0
R3GAN	52.4	91.2	55.8	61.1	72.6	97.1	91.3	90.6	99.2	98.2	91.3	99.0	94.7	93.9	99.7
StyleGAN3	85.2	84.5	92.4	91.7	96.9	91.4	99.5	98.1	94.7	98.9	99.4	92.2	96.5	95.8	99.5
StyleGAN-XL	50.2	93.3	83.7	75.3	93.1	93.2	98.9	69.5	90.8	97.0	96.1	95.6	97.4	97.2	98.7
StyleSwim	76.2	95.2	91.6	94.9	98.5	89.3	99.6	99.5	96.7	99.1	97.8	99.8	96.4	98.6	99.9
WFIR	65.5	83.0	57.0	65.5	86.9	90.8	99.6	89.0	60.7	100.0	99.4	98.1	78.8	51.5	98.6

Appendix

Table 6: AP Results on AIGIBench. All detectors are trained on the Setting-II of AIGIBench.

Method	Clean	JPEG (80)	Blur ( $\sigma$ =0.5)	Noise (0.01)	Resize (0.75 $\times$ )	Avg.
EFFORT	0.940	0.927	0.936	0.913	0.920	0.927
COMPASS	0.966	0.947	0.949	0.930	0.946	0.948

Appendix

Table 8: Robustness evaluation under real-world perturbations.

AI-Generated Image Detection

AI-generated image detection identifies synthetic or AI-edited images and develops methods that generalize across generators, domains, and image transformations.

See all 194 papers

PE-Mamba: Bidirectional Selective Layer Aggregation for AI-Generated Image Detection
Aug 8, 2026Kutub Uddin, Nusrat Tasnim, Khalid MalikVision Transformer Multi-Scale Feature Fusion
College of Innovation & Technology, University of Michigan-Flint, Michigan, US · School of Electronics and Information Engineering, Korea Aerospace University, Goyang, South Korea
Align, Don't Blend: Orthogonal Supervision Disentanglement As the Key to Generalizable AI-Generated Image Detection
Dec 7, 2025Ruoxin Chen, Jiahui Gao, Kaiqing Lin +5AI-Generated Image Detection
Tencent Youtu Lab · East China University of Science and Technology · Shenzhen University +1
FAIR: Feature-Augmented Implicit Regularization for AI-generated Fake Image Detection
Jul 24, 2026Md Redwanul Haque, Manzur Murshed, Manoranjan Paul +1Domain Generalization AI-Generated Image Detection
Deakin University, Burwood, VIC, Australia · Charles Sturt University, Bathurst, NSW, Australia

Aug 8, 2026J/K move · Enter open · S save

PE-Mamba: Bidirectional Selective Layer Aggregation for AI-Generated Image Detection

Kutub Uddin, Nusrat Tasnim, Khalid Malik

College of Innovation & Technology, University of Michigan-Flint, Michigan, US · School of Electronics and Information Engineering, Korea Aerospace University, Goyang, South Korea

Vision Transformer Multi-Scale Feature Fusion AI-Generated Image Detection

Open paper AskPDF arXiv

AI-generated image (AIGI) detection has become increasingly challenging due to the rapid advancement of generative models and the diminishing gap between synthetic and authentic content. Existing vision transformer-based detectors commonly rely on weighted-sum strategies to aggregate intermediate representations across transformer layers, often overlooking the inherently ordered semantic progression of hierarchical features from shallow texture cues to deep semantic representations. In this work, we propose PE-Mamba, a novel framework built upon a pre-trained PE-Core vision transformer with lightweight LoRA adaptation that introduces three complementary components for cross-layer feature aggregation and fusion. First, a bidirectional selective aggregator (BSA) processes layer-wise classification tokens through forward and backward selective scans, where the forward scan progressively accumulates shallow-to-deep forensic evidence, and the backward scan performs deep-to-shallow contextual refinement to reinterpret low-level cues in light of high-level semantic context. Second, a softmax-weighted aggregator (SWA) computes a learned global summary of all layer tokens as a complementary aggregation path. Third, a sigmoid-gated blend (SGA) adaptively fuses the BSA and SWA outputs via a learnable scalar gate, allowing the model to dynamically balance directional sequential evidence and global layer-wise aggregation. Extensive experiments on UniversalFakeDetect (96.6% mACC, 99.5% mAP) and AIGCDetect (95.3% mACC, 98.1% mAP) demonstrate that \methodname{} outperforms 18 detectors with superior generalization across diverse generative models, while training only 1.3% of total parameters (0.13% for LoRA alone).

Abstract

AI-Generated Image Detection

Explore similar work

PE-Mamba: Bidirectional Selective Layer Aggregation for AI-Generated Image Detection

Align, Don't Blend: Orthogonal Supervision Disentanglement As the Key to Generalizable AI-Generated Image Detection

FAIR: Feature-Augmented Implicit Regularization for AI-generated Fake Image Detection