Research map / Image generation / Autoregressive and latent generation
Zero-Shot Detection of AI-Generated Images
Davide Cozzolino, GIovanni Poggi, Matthias Niessner, Luisa Verdoliva
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 |
|---|---|---|---|
| use coding cost gap (NLL minus entropy) at full resolution instead of raw NLL or entropy (ablation) | ZED · Synthbuster | class separability | Improved |
| use level-0 gap instead of level-1 or level-2 gap (ablation) | ZED · Synthbuster | class separability | Improved |
| use absolute value of the gap statistic (ablation) | ZED · GLIDE (COCO real) | AUC: 47.8 -> 88.5 | Improved |
| use absolute value of the gap statistic (ablation) | ZED · StyleGAN2 (FFHQ real) | AUC: 43.1 -> 87.7 | Improved |
| use absolute value of the gap statistic (ablation) | ZED · DDPM (FFHQ real) | AUC: 76.6 -> 25.2 | Hurt |
| use slope Delta01 instead of D0 as decision statistic (ablation) | ZED · 30 test sets avg | AUC: 83.3 -> 88.8 | Improved |
| train lossless coder on COCO with JPEG augmentation instead of Open Images (ablation) | ZED · ImageNet and LAION real test sets | AUC | Improved |
| use zero-shot ZED instead of supervised detectors trained on ProGAN/latent diffusion | AI-generated image detection · 30 test sets avg | AUC: 90.0 vs 89.4 (Corvi2023), 88.4 (Ojha2023) | Mixed |
| single decision threshold across datasets for balanced accuracy | ZED · per-real-dataset balanced accuracy | balanced accuracy | Improved |
| apply ZED to highly compressed or resized images | ZED · web images | reliability | Hurt |
Closest papers by meaning
- Community Forensics: Using Thousands of Generators to Train Fake Image Detectors (CVPR 2025 · similarity 0.89)
- A Bias-Free Training Paradigm for More General AI-generated Image Detection (CVPR 2025 · similarity 0.89)
- Few-Shot Learner Generalizes Across AI-Generated Image Detection (ICML 2025 · similarity 0.88)
- Forensic Self-Descriptions Are All You Need for Zero-Shot Detection, Open-Set Source Attribution, and Clustering of AI-generated Images (CVPR 2025 · similarity 0.88)
- Seeing What Matters: Generalizable AI-generated Video Detection with Forensic-Oriented Augmentation (NeurIPS 2025 · similarity 0.88)
- Revisiting Tampered Scene Text Detection in the Era of Generative AI (AAAI 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.
