Research map / Image generation / Autoregressive and latent generation
Your AI-Generated Image Detector Can Secretly Achieve SOTA Accuracy, If Calibrated
Muli Yang, Gabriel James Goenawan, Henan Wang, Huaiyuan Qin, Chenghao Xu, Yanhua Yang, Fen Fang, Ying Sun, Joo Hwee Lim, Hongyuan Zhu
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 supervised scalar logit calibration (100 target images) | CNNSpot · AIGCDetectBenchmark | average accuracy: 70.83 -> 78.22 | Improved |
| add supervised scalar logit calibration | RINE (CLIP-based) · GenImage | average accuracy: 81.78 -> 97.94 | Improved |
| add supervised scalar logit calibration | Effort · AIGCDetectBenchmark | average accuracy: 79.35 -> 89.48 | Improved |
| add unsupervised scalar logit calibration (no labels) | RINE · AIGCDetectBenchmark | average accuracy: 86.77 -> 93.79 | Improved |
| add calibration to a detector already strong in-distribution | AIDE · AIGCDetectBenchmark | average accuracy: 92.80 -> 94.42 (Sup.), 93.69 (Unsup.) | Improved |
| add unsupervised calibration instead of supervised | FreDect · AIGCDetectBenchmark | average accuracy: 64.03 -> 65.09 | No clear effect |
| add calibration under JPEG compression (QF=90) | AIDE · AIGCDetectBenchmark with JPEG QF=90 | average accuracy: 60.22 -> 75.61 | Improved |
| add calibration under Gaussian blur sigma=2.0 | LGrad · AIGCDetectBenchmark with blur sigma=2.0 | average accuracy: 50.03 -> 49.98 | No clear effect |
| use supervised KDE-based calibration vs unsupervised (ablation) | Fusing · GenImage | average accuracy: 80.94 -> 90.20 | Improved |
| reduce validation set to 10 images (ablation) | supervised and unsupervised calibration · AIGCDetectBenchmark and GenImage | average accuracy | No clear effect |
| estimate alpha with KDE instead of binary search or BCE training (ablation) | supervised calibration of CNNSpot · AIGCDetectBenchmark | average accuracy | Improved |
| add calibration to CNNSpot on GenImage (Unsup.) | CNNSpot · GenImage | average accuracy: 69.37 -> 70.86 | Improved |
Closest papers by meaning
- BIAS-ID: A Framework for Analyzing Transformation Biases in AI-Generated Image Detectors (NeurIPS 2026 · similarity 0.89)
- A Bias-Free Training Paradigm for More General AI-generated Image Detection (CVPR 2025 · similarity 0.88)
- Stay-Positive: A Case for Ignoring Real Image Features in Fake Image Detection (ICML 2025 · similarity 0.87)
- Epistemic Uncertainty for Generated Image Detection (NeurIPS 2025 · similarity 0.87)
- Is Artificial Intelligence Generated Image Detection a Solved Problem? (NeurIPS 2025 · similarity 0.86)
- Dual Data Alignment Makes AI-Generated Image Detector Easier Generalizable (NeurIPS 2025 · similarity 0.86)
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.
