Research map / Image generation / Autoregressive and latent generation
LoRC: Detecting AI-Generated Images via Low-Rank Collapse in the Semantic-Residual Space
Haozhen Yan, Ruoxin Chen, Jiahui Zhan, Taiping Yao, Bo Wang, Youchang xiao, Shouhong Ding, Liqing Zhang, Jianfu Zhang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add orthogonal semantic decomposition to baseline (ablation) | LoRC · T2I-CoReBench | balanced accuracy (%): 94.0 -> 98.1 | Improved |
| add orthogonal semantic decomposition to baseline (ablation) | LoRC · In-the-wild | balanced accuracy (%): 89.4 -> 86.9 | Hurt |
| add low-rank attention after decomposition (ablation) | LoRC · In-the-wild | balanced accuracy (%): 86.9 -> 90.2 | Improved |
| add low-rank attention after decomposition (ablation) | LoRC · Standard benchmarks | balanced accuracy (%): 97.2 -> 96.9 | No clear effect |
| add Subspace Separation Loss (ablation) | LoRC · Average (Standard, In-the-wild, T2I) | balanced accuracy (%): 94.4 -> 96.8 | Improved |
| apply low-rank attention to original entangled patch features (ablation) | LoRC variant · Average | balanced accuracy (%): 93.3 -> 90.0 | Hurt |
| use orthogonal projection instead of subtraction for decomposition (ablation) | LoRC · Average | balanced accuracy (%): 96.0 -> 96.8 | Improved |
| add frozen [CLS] token to residual features (ablation) | LoRC · Average | balanced accuracy (%): 96.2 -> 96.8 | Improved |
| place SSL after low-rank bottleneck instead of before (ablation) | LoRC · Average | balanced accuracy (%): 95.0 -> 96.8 | Improved |
| scale backbone from DINOv3-L to DINOv3-7B (ablation) | LoRC · Average | balanced accuracy (%): 93.6 -> 97.4 | Improved |
| use LoRC (decomposition + low-rank + SSL) instead of original patch features (ablation) | DINOv3-H+ · Average | balanced accuracy (%): 93.3 -> 96.8 | Improved |
| LoRC vs DDA on unseen T2I generators | LoRC · T2I-CoReBench | balanced accuracy (%): 91.1 -> 97.0 | Improved |
Closest papers by meaning
- MIRROR: Manifold Ideal Reference ReconstructOR for AI-Generated Image Detection (NeurIPS 2026 · similarity 0.88)
- Scaling Up AI-Generated Image Detection with Generator-Aware Prototypes (CVPR 2026 · similarity 0.87)
- PiD: Generalized AI-Generated Images Detection with Pixelwise Decomposition Residuals (ICML 2025 · similarity 0.85)
- Locate-Then-Examine: Grounded Region Reasoning Improves Detection of AI-Generated Images (CVPR 2026 · similarity 0.85)
- A Bias-Free Training Paradigm for More General AI-generated Image Detection (CVPR 2025 · similarity 0.85)
- Community Forensics: Using Thousands of Generators to Train Fake Image Detectors (CVPR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
