Research map / Image generation / Autoregressive and latent generation
Where Detectors Fail: Probing Generative Space for Generalizable AI-Generated Image Detection
Zijie Cao, Weijie Tu, Yao Xiao, Weijian Deng, Weiyan Chen, Liang Lin, Pengxu Wei
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 PROBE boundary exploration and detector adaptation | ResNet50 detector · Avg over 7 benchmarks (trained on GenImage SD 1.4) | balanced accuracy (%): 63.5 -> 78.1 | Improved |
| add PROBE | DINOv2 detector · Avg over 7 benchmarks (trained on reconstruction set) | balanced accuracy (%): 87.4 -> 93.9 | Improved |
| add PROBE | DINOv2 detector · Avg over 7 benchmarks (trained on reconstruction set), average precision | average precision (%): 93.8 -> 97.8 | Improved |
| add PROBE | DINOv2 detector · GAN and NOVA generators (avg of 7) | balanced accuracy (%): 83.7 -> 93.8 | Improved |
| use PROBE samples instead of standard SD 1.4 samples (ablation) | ResNet50 detector · Avg over 5 benchmarks (augmentation comparison, ResNet50) | balanced accuracy (%): 67.5 -> 76.0 | Improved |
| use PROBE samples instead of pixel-level adversarial augmentation (ablation) | ResNet50 detector · Avg over 5 benchmarks (augmentation comparison, ResNet50) | balanced accuracy (%): 70.6 -> 76.0 | Improved |
| aggregate boundary samples from multiple detectors (ablation) | PROBE (ResNet50) · Avg over 5 benchmarks (augmentation comparison, ResNet50) | balanced accuracy (%): 76.0 -> 76.7 | No clear effect |
| add perceptual regularization (lambda=1.0) (ablation) | PROBE (ResNet50) · Avg over 7 benchmarks (ResNet50 ablation) | balanced accuracy (%): 74.7 -> 78.1 | Improved |
| use SD 3.5-Medium instead of SD 1.4 as probing generator (ablation) | PROBE (ResNet50) · Avg over 7 benchmarks (ResNet50 ablation) | balanced accuracy (%): 78.1 -> 76.2 | No clear effect |
| aggregate samples from three generators (ablation) | PROBE (ResNet50) · Avg over 7 benchmarks (ResNet50 ablation) | balanced accuracy (%): 78.1 -> 80.2 | Improved |
| increase PROBE samples from 20k to 40k (ablation) | PROBE (ResNet50) · Avg over 7 benchmarks (ResNet50 ablation) | balanced accuracy (%): 78.1 -> 78.9 | Improved |
| add a second and third probing iteration (ablation) | PROBE (ResNet50) · Avg over 7 benchmarks (ResNet50 ablation) | balanced accuracy (%): 78.1 -> 80.1 | Improved |
Closest papers by meaning
- Is Artificial Intelligence Generated Image Detection a Solved Problem? (NeurIPS 2025 · similarity 0.89)
- Breaking Semantic Artifacts for Generalized AI-generated Image Detection (NeurIPS 2024 · similarity 0.89)
- Are High-Quality AI-Generated Images More Difficult for Models to Detect? (ICML 2025 · similarity 0.88)
- Manifold Induced Biases for Zero-shot and Few-shot Detection of Generated Images (ICLR 2025 · similarity 0.87)
- A Sanity Check for AI-generated Image Detection (ICLR 2025 · similarity 0.87)
- CO-SPY: Combining Semantic and Pixel Features to Detect Synthetic Images by AI (CVPR 2025 · similarity 0.87)
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.
