Research map / Image generation / Autoregressive and latent generation
Breaking Semantic Artifacts for Generalized AI-generated Image Detection
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.
Shows that AI-image detectors overfit "semantic artifacts" of training scenes and proposes shuffling image patches before training an end-to-end patch-based ResNet-style classifier, improving cross-scene and open-world generalization.
Problem
AI-generated image detectors trained on one scene/dataset fail to generalize to images with different semantics (cross-scene), with high AP but low accuracy due to score-threshold shift.
Approach
Attribute failures to "semantic artifacts" inherited from dataset content and preprocessing; apply random patch shuffle to break global semantics and train a patch-based convolutional classifier with small receptive fields and a linear head over patch features.
Key results
- Zero-padding variant reaches 92.52% average cross-scene accuracy versus 90.44% for the best baseline NPR, a gain of about 2.08 points.
- Open-world average accuracy of 85.97% (zero padding) versus 75.38% for NPR, a 10.59-point gain.
- Patch shuffle sharply improves cross-scene thresholded accuracy (e.g., Church to Bedroom accuracy at 50% threshold rises from 0% to 94.15%) in the preliminary ResNet-50 study.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the patch-based classifier with zero padding instead of resizing (ablation) | Ours · Open-world | Avg. Acc.: 82.12 -> 85.97 | Helped |
| Use the patch-based classifier instead of NPR | AI-generated image detection · Cross-scene | Avg. Acc.: 90.44 -> 92.52 | Helped |
| Use the patch-based classifier instead of NPR | AI-generated image detection · Open-world | Avg. Acc.: 75.38 -> 85.97 | Helped |
| Increase the number of same-convolution blocks from 0 to 21 at patch size 64 (ablation) | Patch-based classifier · Cross-Diffusion | Avg. Acc.: 81.53 -> 88.91 | Helped |
| Increase same-convolution blocks to 24 rather than 21 at patch size 64 (ablation) | Patch-based classifier · Cross-Diffusion | Avg. Acc.: 88.91 -> 87.10 | Hurt |
| Use patch size 32 rather than patch size 16 (ablation) | Patch-based classifier · Open-world | Avg. Acc.: 77.13 -> 82.12 | Helped |
| Increase training images from 10k to 80k (ablation) | Patch-based classifier · Open-world | Avg. Acc.: 74.65 -> 81.69 | Helped |
| Add MSE denoising loss to BCE loss (ablation) | Patch-based classifier · Open-world | Avg. Acc.: 82.12 -> 78.18 | Hurt |
| Fuse global ResNet18 features with patch features (ablation) | Patch-based classifier · Open-world | Avg. Acc.: 82.12 -> 69.51 | Hurt |
| Add self-attention before the linear classifier (ablation) | Patch-based classifier · Open-world | Avg. Acc.: 82.12 -> 79.95 | Hurt |
| Use zero padding rather than resizing (ablation) | Patch-based classifier · Cross-GAN/CNN | Avg. Acc.: 72.58 -> 80.13 | Helped |
Surprising findings
- Detectors trained on one scene keep high AP but collapse in thresholded accuracy under cross-scene shift, implying score calibration rather than ranking is the failure mode.
- Shallower networks (N=6) generalize better cross-GAN/CNN while deeper ones (N=21) work better cross-diffusion, indicating a tradeoff between in-domain artifact fitting and cross-generator transfer.
Limitations
- Semantic artifacts are only qualitatively analyzed via spectra and CAM visualizations; no quantitative metric or theoretical proof.
- Training data is proprietary per the checklist, so data cannot be shared publicly.
- No error bars reported; all methods use a single seed due to compute cost.
Experimental setup
Datasets: DiffusionDB LAION-5B ForenSynths LSUN-bedroom LSUN-church CelebA FFHQ ImageNet
Benchmarks: 31 open-world test sets (18 diffusion variants, 7 GANs, 6 CNN-based generators) cross-scene Latent Diffusion variants (6 datasets)
Metrics: accuracy average precision (mAP)
Baselines: CNNDetection PatchFor F3Net Durall DIRE Dogoulis Ojha LGrad NPR
Builds on: CNNDetection (ForenSynths) NPR Ojha et al. (CLIP-based detection) LGrad
Closest papers by meaning
- A Sanity Check for AI-generated Image Detection (ICLR 2025 · similarity 0.89)
- CO-SPY: Combining Semantic and Pixel Features to Detect Synthetic Images by AI (CVPR 2025 · similarity 0.88)
- Community Forensics: Using Thousands of Generators to Train Fake Image Detectors (CVPR 2025 · similarity 0.87)
- A Bias-Free Training Paradigm for More General AI-generated Image Detection (CVPR 2025 · similarity 0.87)
- Beyond Generation: A Diffusion-based Low-level Feature Extractor for Detecting AI-generated Images (CVPR 2025 · similarity 0.86)
- Is Artificial Intelligence Generated Image Detection a Solved Problem? (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.