LoRC: Detecting AI-Generated Images via Low-Rank Collapse in Semantic Residuals

Haozhen Yan; Ruoxin Chen; Jiahui Zhan; Bo Wang; Youchang Xiao; Shouhong Ding; Liqing Zhang; Taiping Yao; Jianfu Zhang

Source-linked AI summary

LoRC: Detecting AI-Generated Images via Low-Rank Collapse in Semantic Residuals

Haozhen Yan, Ruoxin Chen, Jiahui Zhan, Bo Wang, Youchang Xiao, Shouhong Ding, Liqing Zhang, Taiping Yao, Jianfu Zhang

arXiv:2608.20882v1Aug 21, 2026cs.CV

TL;DR

AIGI detection needs cues beyond increasingly faithful macroscopic semantics, especially for generalization to unseen generators. LoRC geometrically separates the dominant semantic direction from the orthogonal residual space and models its low-rank collapse, achieving a 7.0% absolute average improvement and 97.0% average accuracy on unseen generators.

Problem
AIGI detectors need reliable cues beyond macroscopic semantics because modern generators produce semantically faithful images and unseen architectures make zero-shot generalization difficult.
Method
LoRC uses orthogonal decomposition with a frozen vision foundation model to suppress dominant semantics and models the collapsed semantic-residual geometry with low-rank attention and subspace separation.
Results
7.0% absolute average improvement over prior methods and 97.0% average accuracy on T2I-CoReBench demonstrate LoRC's reported performance across benchmarks and unseen generators.
Takeaways & Limitations
Low-rank collapse provides an architecture-agnostic and resilient forensic signature for cross-model AIGI detection.

Abstract

from arXiv · show

Modern generators faithfully model macroscopic semantics, producing synthetic images that appear highly realistic. Consequently, decisive forensic cues reside in subtle non-semantic visual discrepancies. To reveal these cues, we revisit AIGI detection from a geometric perspective and identify an architecture-agnostic signature. Specifically, modern generators exhibit low-rank collapse (\textit{i.e.}, rank degeneracy) in the semantic-residual orthogonal subspace while largely preserving the dominant semantic direction. This structural flattening consistently emerges during the final decoding stage, forming a shared bottleneck across diverse generator architectures. Motivated by this signature, we propose \textbf{LoRC}, a framework that decouples semantic dominance to capture the collapsed residual geometry induced by the generative decoding bottleneck. Our method improves accuracy by an average of 7.0\% across multiple benchmarks and achieves 97.0\% accuracy on 39 unseen generators. These results demonstrate strong cross-model generalization and robustness, making LoRC a reliable approach for AIGI detection in complex real-world environments.

1 Introduction

Modern generators increasingly match real images in dominant semantics, shifting detection toward a shared low-rank collapse in the orthogonal semantic-residual space. LoRC targets this geometry by decoupling semantic alignment and capturing residual discrepancies.

AIGI detection is increasingly difficult because generators match real images in object identity and global layout, while unseen-generator shifts challenge cross-model generalization.Generated images show stronger semantic alignment but weaker semantic-residual energy than real images, reducing the discriminative value of semantic features.
Controlled reconstructions reveal feature shifts that reduce residual magnitude while largely preserving the dominant semantic direction.The shifts are highly directional in the semantic-residual dimensions and introduce no significant semantic-direction variation.
Low-Rank Collapse concentrates real-to-generated discrepancies on only a few dominant residual dimensions, flattening the residual geometry.This systematic collapse is identified as a generalizable generative signature.
The collapse appears across diffusion and autoregressive architectures at the final decoding stage, whose generative mapping acts as a shared information bottleneck.Its geometric origin also makes the signal largely insensitive to superficial perturbations such as JPEG compression.
LoRC isolates the semantic-residual subspace by using a frozen vision foundation model and projecting patch tokens onto the orthogonal complement of the global [CLS] semantic anchor.This pipeline is motivated by the need to neutralize dominant semantic alignment before capturing residual geometry.

2 Related Works

Prior AIGI detectors use low-level artifacts, frequency traces, semantic-null representations, information bottlenecks, and generator-component traces. These cues can weaken across generator families or reflect spurious dataset biases, motivating more transferable detection signals.

Early detectors primarily exploit low-level generation artifacts such as spectral duplication and upsampling traces.Frequency patterns are reported to be substantially weaker for diffusion-generated images, although later methods use them as auxiliary signals.
Frequency- and artifact-based cues can be spurious, with file format, resolution, compression, and acquisition pipelines harming cross-domain generalization.Prior work identifies non-causal dataset biases as a detection risk and seeks alignment of real and synthetic data distributions.
Frozen generic vision representations improve generalization to unseen generators and may already encode forensic cues useful for in-the-wild detection.Related approaches pair frozen features with simple classifiers or suppress semantic interference through semantic null-spaces.
Other methods impose information bottlenecks or use final-generator components and parameter-space SVD to obtain transferable detection traces.These approaches target semantic compression or architectural structure from different perspectives.

3 Methodology

LoRC detects AI-generated images by separating dominant semantic alignment from semantic-residual geometry, where generative decoding produces consistent low-rank collapse. The framework models these structured residual cues across architectures and under JPEG compression.

Motivation: Modern generators closely match high-level structure and object identity, shifting decisive forensic evidence toward process-dependent residuals rather than global semantics.The shared final decoding stage may imprint transferable synthetic structure in feature space.
Geometric Decoupling: Orthogonal decomposition uses the normalized frozen [CLS] token as a semantic direction and projects patch features into semantic and residual components.X_sem captures global semantic alignment, while X_res isolates the remaining real–synthetic discrepancy.
Systematic Semantic–Residual Shift: Generated images show higher semantic alignment and smaller residual norms, whereas real images extend toward larger residual norms.Across eight architectures, this pattern indicates semantic over-alignment and systematically attenuated residual energy in generated images.
Low-Rank Collapse: 29.1% and 16.1% of variance lie in the first two residual-subspace principal components, compared with 10.7% and 7.4% in the semantic subspace.The spectra indicate that structural discrepancy concentrates in a few residual directions while semantic representations remain more dispersed.
Universality and Robustness: Every tested model exhibits a similarly pronounced low-rank tendency in the semantic-residual subspace across Stable Diffusion, DiT, FLUX, PixelFlow, and JanusPro.The comparison covers fundamentally different generative paradigms.
Universality and Robustness: JPEG compression at quality factor 96 preserves the pronounced residual-subspace low-rank tendency, indicating robustness beyond high-frequency reconstruction details.The analysis compares uncompressed PNG and compressed JPEG reconstructions across diverse pipelines.
LoRC Architecture: LoRC applies low-rank attention to residual features in a rank-r bottleneck, then combines pooled residual features with the frozen [CLS] token for classification.An auxiliary subspace-separation objective encourages decorrelation between real and fake residual subspaces.

4 Experiments

LoRC is evaluated across standard, in-the-wild, and recent-generator benchmarks, with ablations and robustness tests examining its components and stability under image degradations.

Evaluation Protocol: Evaluation uses official baseline implementations, JPEG recompression at quality 96, and Balanced Accuracy as the primary metric.Benchmarks are grouped into standard, in-the-wild, and recent SOTA generator settings.
Overall Results: 97.2% overall accuracy across seven datasets surpasses the second-best method by 7.0%.LoRC improves consistently across standard and in-the-wild benchmarks.
Zero-Shot Comparison: 97.0% average accuracy across unseen generators on T2I-CoReBench improves over existing approaches by 5.9%.The evaluation targets zero-shot generalization to rapidly evolving generator architectures.
Ablation Study: Semantic Decomposition, Low-Rank Attention, and Subspace Separation Loss progressively improve performance in the cumulative ablation.Semantic Decomposition benefits T2I-CoReBench, Low-Rank Attention improves in-the-wild performance, and Subspace Separation Loss yields the best overall results.
Hyperparameter Sensitivity: λSS = 0.1, feature rank 32, and LoRA rank 16 provide the best reported hyperparameter settings.These settings balance subspace separation, feature compression, information preservation, and adaptation capacity.
Robustness Analysis: LoRC maintains consistently high balanced accuracy under JPEG compression, image resizing, and Gaussian blurring.The robustness analysis tests three common image perturbations.

5 Conclusion

The paper identifies low-rank collapse in the semantic-residual orthogonal subspace as an architecture-agnostic signature and introduces LoRC to model it for AIGI detection.

Conclusion: LoRC detects a pronounced low-rank collapse in semantic-residual dimensions while the dominant semantic direction remains largely preserved.The collapse is described as a stable degeneracy introduced by the final decoding bottleneck.
Conclusion: LoRC achieves a 7.0% absolute average improvement across seven benchmarks and 97.2% overall accuracy.The conclusion reports strong performance across the evaluated benchmarks.
Conclusion: LoRC reaches 97.0% accuracy across 39 unseen generators, supporting low-rank collapse as a reliable signature for real-world AIGI detection.The result concerns zero-shot generalization to rapidly evolving generator suites.

Supplementary Material

The supplementary material adds method comparisons, detailed benchmark results, further ablations, backbone analysis, and computational-cost analysis for LoRC.

A Overview of Compared Methods: Section A summarizes the compared methods used in the main paper.It provides supplementary context for the baseline methods.
B Detailed Benchmark Results: Section B presents detailed results on GenImage, DRCT-2M, AIGCDetectionBenchmark, and WildRF.These results expand the benchmark reporting beyond the main summary.
C Additional Ablations: Section C provides additional ablations covering semantic decomposition, subspace separation loss placement, and backbone analysis.The supplementary ablations examine design choices beyond the main study.
D Computational Cost: Section D analyzes LoRC’s computational cost and accuracy-efficiency trade-off.The section addresses efficiency alongside detection performance.

A Overview of Compared Methods

The compared methods address AIGI detection through bias-reduced aligned training, low-level artifact modeling, pretrained visual features, multimodal adaptation, and reconstruction-based supervision.

Low-Level Artifact Modeling: NPR models neighboring-pixel relationships to capture generator-agnostic traces from upsampling artifacts.It is a CNN-based detector focused on low-level forensic cues.
Pretrained Visual Features: UnivFD uses frozen CLIP visual features with a linear classifier to improve transferability across generative models.This design reduces overfitting to generator-specific artifacts.
Pretrained Visual Features: FatFormer combines image- and frequency-domain forgery representations with text-image contrastive learning.Its adaptive module enhances forgery-sensitive features while textual prompts guide synthetic-artifact attention.
Robustness-Oriented Training: SAFE uses crop-based preprocessing, augmentations, and random masking to preserve forensic cues and strengthen local artifact awareness.The method approaches detection through image transformations.
Prompt and Contrastive Adaptation: C2P-CLIP injects category-common semantic prompts into caption-image contrastive training and uses an adapted image encoder at inference.The method is developed on top of CLIP.
Hybrid Representations: AIDE fuses CLIP semantic embeddings with extreme-frequency features and handcrafted noise cues.Its hybrid architecture combines high-level semantics with low-level forensic information.
Alignment and Bias Reduction: B-Free creates semantically aligned synthetic counterparts and uses content augmentation to reduce content bias and spurious correlations.Its training paradigm targets generation artifacts rather than mismatched content.
Alignment and Bias Reduction: DDA aligns real-fake pairs in pixel and frequency domains to reduce dataset bias and frequency-level mismatch.The method aims to limit reliance on spurious content or frequency cues.

B Detailed Benchmark Results

LoRC delivers the strongest average performance across the listed benchmarks, with substantial margins over the second-best method and robust gains on Chameleon and WildRF.

LoRC achieves the best average performance on GenImage, DRCT-2M, and AIGCDetectionBenchmark.It exceeds the second-best method by 6.2% on GenImage, 1.2% on DRCT-2M, and 8.5% on AIGCDetectionBenchmark.
10.2% is LoRC’s margin over the second-best method on Chameleon.
7.3% is LoRC’s advantage in average balanced accuracy on WildRF.The result is reported alongside the Chameleon comparison as evidence of robustness across diverse scenarios.

C.1 Effects of Semantic Decomposition

Semantic decomposition is essential to LoRC: orthogonal projection removes semantic leakage from residuals and exposes the features carrying most detection capability.

Semantic decomposition: 90.0% average accuracy results when Low-Rank Attention is applied directly to entangled features, down from 93.3%.
Semantic decomposition: 96.0% average accuracy follows simple subtraction, but semantic remnants remain in the residual space.
Semantic decomposition: 96.8% is achieved with orthogonal projection, which guarantees that isolated residuals are purged of semantic directions.
Residual contribution: 96.2% average accuracy is obtained using residual features alone, while adding the [CLS] token raises performance to 96.8%.Using only the trainable semantic [CLS] token yields 93.8%, indicating that global semantics are insufficient for reliable detection.

C.2 SSL for Residual Subspace Separation

SSL works best after the low-rank bottleneck, where it operates on cleaner residuals and better separates real and fake covariance supports.

SSL placement: Low-Rank Attention provides the rank bottleneck, while SSL further improves separation of real and fake residual distributions.
SSL placement: Applying SSL after the low-rank bottleneck achieves the best average accuracy.
SSL placement: Post-bottleneck SSL suppresses nuisance factors such as resizing and compression artifacts before subspace separation.
Backbone scaling: LoRC improves from 93.6% with DINOv3-L to 97.4% with DINOv3-7B across the DINOv3 family.DINOv3-H+ at 96.8% is adopted as the default extractor to balance performance and practicality.

D Inference Efficiency Analysis

On the full Chameleon dataset, LoRC combines high balanced accuracy with low latency and high throughput, indicating a favorable efficiency–accuracy balance.

Inference efficiency: 92.6% balanced accuracy is achieved with 9.5 ms latency and 106 images/s throughput.All methods are evaluated on the full Chameleon dataset with batch size 16 under their respective inference settings.