Research map / Image generation / Autoregressive and latent generation
Training-free Detection of AI-generated images via Cropping Robustness
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes WaRPAD, a training-free AI-generated image detector that scores DINOv2 embedding sensitivity to Haar-wavelet high-frequency perturbations over rescaled-and-patchified image crops, averaging across patches.
Problem
Detecting AI-generated images without training on any real or fake data, with generalization across generators, resolutions, and domains.
Approach
Self-supervised models trained with RandomResizedCrop should be invariant to high-frequency perturbations on real images but not on generated ones; WaRPAD measures this sensitivity with Haar-wavelet perturbations and deterministically simulates RRC by rescaling images to a multiple of the input size and averaging the score over 224x224 patches.
Key results
- WaRPAD reaches mean AUROC 0.834 on Synthbuster, versus 0.703 for the best baseline AEROBLADE, and 0.587 for RIGID.
- On GenImage WaRPAD averages 0.946 AUROC, above AEROBLADE (0.928) and AIDE (0.969 is higher, so WaRPAD is second on mean only versus AIDE's reported 0.969 in the table, see notes).
- On Deepfake-LSUN-Bedroom WaRPAD averages 0.934 AUROC, versus 0.920 for Manifold Bias.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use WaRPAD instead of AEROBLADE | training-free detection on Synthbuster · Synthbuster | AUROC (Mean): 0.703 -> 0.834 | Helped |
| Use WaRPAD instead of AEROBLADE | training-free detection on GenImage · GenImage | AUROC (Mean): 0.928 -> 0.946 | Helped |
| Use WaRPAD instead of Manifold Bias | training-free detection on Deepfake-LSUN-Bedroom · Deepfake-LSUN-Bedroom | AUROC (Mean): 0.920 -> 0.934 | Helped |
| Use WaRPAD instead of RIGID | detection in Art domain · Art domain | AUROC: 0.725 -> 0.765 | Helped |
| Add HFwav to the patch-based detection method (ablation) | WaRPAD on Synthbuster · Synthbuster | AUROC: 0.656 -> 0.834 | Helped |
| Add HFwav to the patch-based detection method (ablation) | WaRPAD on GenImage · GenImage | AUROC: 0.800 -> 0.946 | Helped |
| Add HFwav to the patch-based detection method (ablation) | WaRPAD on Deepfake-LSUN-Bedroom · Deepfake-LSUN-Bedroom | AUROC: 0.861 -> 0.934 | Helped |
| Add HFwav without patch augmentation (ablation) | RIGID on Synthbuster · Synthbuster | AUROC: 0.587 -> 0.636 | Helped |
| Add HFwav without patch augmentation (ablation) | RIGID on GenImage · GenImage | AUROC: 0.820 -> 0.809 | Hurt |
| Add HFwav without patch augmentation (ablation) | RIGID on Deepfake-LSUN-Bedroom · Deepfake-LSUN-Bedroom | AUROC: 0.861 -> 0.890 | Helped |
| Use more RIGID patch views at equal computational cost (ablation) | RIGID · Synthbuster; GenImage; Deepfake-LSUN-Bedroom | AUROC: 0.587 -> 0.589; 0.820 -> 0.823; 0.861 -> 0.872 | Helped |
| Replace mean aggregation with median aggregation (ablation) | WaRPAD with ViT-L14 · Synthbuster; GenImage; Deepfake-LSUN-Bedroom | AUROC: 0.83/0.95/0.93 -> 0.84/0.94/0.92 | Mixed |
Surprising findings
- Other self-supervised backbones (BEiT) with the same method underperform RIGID, and WaRPAD gains are weaker for masked-image-modeling models.
- Wavelets with more vanishing moments make DINOv2 lose robustness, so the Haar wavelet is the best choice.
Limitations
- Extra computation cost from computing scores over many patches after rescaling.
- Dependence on the backbone self-supervised model; may not generalize to high-resolution images beyond the backbone's scope.
- May become less effective as generators produce realistic high-frequency content that fools the foundation model.
Experimental setup
Datasets: Synthbuster (RAISE-1k real) GenImage (ImageNet real) Deepfake-LSUN-Bedroom (LSUN real) WikiArt-based Kaggle art dataset
Benchmarks: Synthbuster GenImage Deepfake-LSUN-Bedroom Art domain (Kaggle WikiArt-based)
Metrics: AUROC
Baselines: AIDE FatFormer RIGID MINDER AEROBLADE Manifold Bias
Builds on: RIGID DINOv2 AEROBLADE RandomResizedCrop augmentation
Closest papers by meaning
- Any-Resolution AI-Generated Image Detection by Spectral Learning (CVPR 2025 · similarity 0.87)
- A Sanity Check for AI-generated Image Detection (ICLR 2025 · similarity 0.86)
- A Bias-Free Training Paradigm for More General AI-generated Image Detection (CVPR 2025 · similarity 0.85)
- Few-Shot Learner Generalizes Across AI-Generated Image Detection (ICML 2025 · similarity 0.85)
- Breaking Semantic Artifacts for Generalized AI-generated Image Detection (NeurIPS 2024 · similarity 0.84)
- WildFake: A Large-Scale and Hierarchical Dataset for AI-Generated Images Detection (AAAI 2025 · similarity 0.84)
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.