cs.CVAug 21, 2026
SaveLoRC: Detecting AI-Generated Images via Low-Rank Collapse in Semantic Residuals
Organizations: Shanghai Jiao Tong University, China · Tencent Youtu Lab, China
Abstract
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 (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 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.
Figures & tables
Figure 1: Geometric perspective of AI-Generated Image (AIGI) detection. We illustrate the structural relationship between real images, AIGIs, and the semantic space of vision foundation models (VFMs). While real images reside on an intricate, high-dimensional manifold (bottom), modern generative models are increasingly aligning with them in the dominant semantic space (top), causing the Semantic Variation to shrink significantly. However, in the orthogonal semantic-residual dimension (vertical direction), the discrepancy between AIGIs and real images becomes correspondingly more pronounced. Our empirical analysis reveals that the final decoding stage of generative models acts as a severe information bottleneck, inducing a systematic low-rank collapse that forces the complex real manifold to flatten into a constrained AIGI space (middle). The resulting low-rank effect can serve as a universal forensic signature.
Figure 2: Decoupled semantic–residual feature space on GenImage (a) and SD v1.5 VAE reconstructions (b), showing a consistent generative bottleneck: semantic over-alignment with attenuated residual energy.
Figure 3: PCA variance spectra of the reconstructed transition deltas ( ). The semantic-residual subspace exhibits a stronger low-rank tendency.
Figure 4: PCA variance spectra of counterpart deltas ( ). Comparison of uncompressed PNG ( top ) and JPEG (Q=96) ( bottom ) reconstructions across diverse pipelines. The consistent eigenvalue distributions verify the universality and robustness of the semantic-residual structural deviation.
Figure 5: LoRC Architecture. DINOv3 patch features are decomposed using the normalized frozen [CLS] token as a semantic anchor, projecting tokens onto its orthogonal complement to obtain semantic-residuals. A Low-Rank Attention Block performs self-attention in a rank- bottleneck to amplify structured low-rank forensic cues. The pooled residual feature, concatenated with the frozen [CLS] token, is fed to a linear classifier. An auxiliary Subspace Separation Loss further encourages decorrelation between real and fake residual subspaces, reinforcing their geometric distinction.
| Setting | Standard | In-the-wild | T2I-CoReBench | Avg. |
| Baseline | 97.9 | 89.4 | 94.0 | 93.8 |
| + Semantic Decomposition | 97.2 | 86.9 | 98.1 | 94.1 |
| + Low-Rank Attention | 96.9 | 90.2 | 96.0 | 94.4 |
| + Subspace Separation Loss | 98.1 | 95.2 | 97.0 | 96.8 |
Table 3: Ablation study of the LoRC where modules are added cumulatively.
| Generator | NPR | UnivFD | C2P. | Fat. | SAFE | AIDE | DRCT | Align. | DDA | LoRC |
| Diffusion Models | ||||||||||
| SD-3-Medium | 51.6 | 5.0 | 71.0 | 3.8 | 1.6 | 19.2 | 59.4 | 60.5 | 96.8 | 99.6 |
| SD-3.5-Medium | 61.6 | 9.9 | 66.7 | 12.9 | 1.6 | 7.2 | 42.4 | 68.5 | 97.0 | 99.7 |
| SD-3.5-Large | 66.1 | 12.7 | 53.0 | 7.6 | 1.4 | 4.2 | 57.0 | 58.9 | 97.0 | 99.8 |
| FLUX.1-schnell | 51.6 | 4.5 | 59.9 | 2.5 | 0.7 | 16.8 | 36.0 | 27.1 | 94.8 | 98.0 |
| FLUX.1-dev | 58.7 | 3.5 | 73.5 | 0.8 | 0.4 | 18.7 | 33.5 | 21.0 | 90.6 | 96.2 |
Table 4: Comparison of fake accuracy on T2I-CoReBench. C2P., Fat., and Align. are abbreviations for C2P-CLIP, FatFormer, and AlignedForensics, respectively. All test images are compressed using JPEG with a quality factor of 96.
Figure 6: Hyperparameter sensitivity. (a) The impact of the weight parameter on Subspace Separation Loss. (b) The impact of the rank in the Low-Rank Attention Block. (c) The effect of varying the LoRA rank during fine-tuning.
Figure 7: Robustness analysis on GenImage.
| Strategy | Standard | In-the-wild | T2I-CoReBench | Avg. |
| w/o Semantic Decomposition | ||||
| Baseline ( original patch ) | 94.5 | 90.0 | 95.4 | 93.3 |
| + Low-Rank Attention | 91.6 | 83.7 | 94.9 | 90.0 |
| + Subspace Separation | 91.8 | 82.5 | 95.9 | 90.1 |
| w/ Semantic Decomposition | ||||
| Baseline ( subtraction ) | 94.5 | 88.7 | 97.9 | 93.7 |
Table 9: Ablation study on semantic decomposition. We evaluate the impact of low-rank attention and subspace separation with and without semantic decomposition. Modules are added cumulatively. Applying low-rank attention to entangled features (w/o decomposition) degrades performance. Compared to simple subtraction, our orthogonal projection ensures strict geometric decoupling, preventing semantic leakage and achieving the optimal average accuracy.
| CLS Token Strategy | Standard | In-the-wild | T2I-CoReBench | Avg. |
| Only Trainable [CLS] Token | 97.9 | 89.4 | 94.0 | 93.8 |
| w/o [CLS] Token | 97.1 | 94.2 | 97.3 | 96.2 |
| w/ [CLS] Token (LoRC) | 98.1 | 95.2 | 97.0 | 96.8 |
Table 10: Ablation study on the contribution of semantic [CLS] token concatenation. Using only the trainable semantic [CLS] token yields a lower average accuracy of 93.8%, indicating that global semantics alone are insufficient for robust detection. By comparison, residuals alone already achieve 96.2% accuracy, and adding the frozen [CLS] token brings only a marginal 0.6% gain. This confirms that the residual subspace drives detection, while global semantics merely provide auxiliary context.
| Setting | Standard | In-the-wild | T2I-CoReBench | Avg. |
| w/o SSL | 96.9 | 90.2 | 96.0 | 94.4 |
| SSL Before Low-Rank | 95.6 | 93.2 | 96.3 | 95.0 |
| SSL After Low-Rank (LoRC) | 98.1 | 95.2 | 97.0 | 96.8 |
Table 11: Ablation study on SSL placement. Applying SSL after the low-rank bottleneck achieves the best average accuracy.
| Backbone | Variant | Standard | In-the-wild | T2I-CoReBench | Avg. | Latency (ms) |
| DINOv3-L | Original Patch | 92.8 | 82.7 | 94.3 | 89.9 | 5.4 |
| LoRC | 96.5 | 88.1 | 96.3 | 93.6 | ||
| DINOv3-H+ | Original Patch | 94.5 | 90.0 | 95.4 | 93.3 | 9.5 |
| LoRC | 98.1 | 95.2 | 97.0 | 96.8 | ||
| DINOv3-7B | Original Patch | 96.1 | 92.9 | 96.5 | 95.2 | 85.4 |
| LoRC | 97.6 | 96.4 | 98.3 | 97.4 |
Table 12: Ablation study of LoRC across different DINOv3 backbones.
Figure 8: Efficiency and accuracy comparison on the Chameleon dataset. LoRC achieves a strong balance between accuracy and efficiency, attaining 92.6% balanced accuracy with a throughput of 106 images/s (9.5 ms latency).
More in this topic
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 198 papers