Research map / Image generation / Autoregressive and latent generation
All Patches Matter, More Patches Better: Enhance AI-Generated Image Detection via Panoptic Patch Learning
Zheng Yang, Ruoxin Chen, Zhiyuan Yan, Ke-Yue Zhang, Xinghe Fu, Shuang Wu, Xiujun Shu, Taiping Yao, Shouhong Ding, Zequn Qin, Xi Li
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 margin-based Euclidean contrastive loss instead of cosine distance (ablation) | PPL · GenImage | mean accuracy: 92.8 -> 97.2 | Improved |
| use margin-based contrastive loss instead of InfoNCE (ablation) | PPL · GenImage | mean accuracy: 95.8 -> 97.2 | Improved |
| randomly select reconstructed patches instead of fixed half of the image (ablation) | PPL · GenImage | mean accuracy: 95.0 -> 97.2 | Improved |
| use diffusion patch replacement instead of random patch dropout (ablation) | PPL · GenImage | mean accuracy: 94.2 -> 97.2 | Improved |
| change reconstruction patch size from 14 to 28 (ablation) | PPL · GenImage | mean accuracy: 97.2 -> 97.6 | No clear effect |
| increase reconstruction patch size from 14 to 112 (ablation) | PPL · GenImage | mean accuracy: 97.2 -> 96.6 | Hurt |
| use diffusion reconstruction instead of stitching synthetic patches (ablation) | PPL · GenImage | accuracy | Improved |
| combine RPR and PCL rather than either alone (ablation) | PPL · GenImage | accuracy | Improved |
| set contrastive weight lambda to 0.3 (ablation) | PPL · GenImage | accuracy | Improved |
| raise reconstruction ratio to excessively high values (ablation) | PPL · GenImage | accuracy | Hurt |
| vary probability of replacing fake images with RPR images (ablation) | PPL · GenImage | accuracy | No clear effect |
| add PCL contrastive loss to CLIP-Large LoRA training | PPL · training cost | GPU memory and epoch time: 16 GB -> 19 GB; 3 -> 4 hours | Hurt |
Closest papers by meaning
- Breaking Semantic Artifacts for Generalized AI-generated Image Detection (NeurIPS 2024 · similarity 0.87)
- Training-free Detection of AI-generated images via Cropping Robustness (NeurIPS 2025 · similarity 0.87)
- Is Artificial Intelligence Generated Image Detection a Solved Problem? (NeurIPS 2025 · similarity 0.86)
- A Sanity Check for AI-generated Image Detection (ICLR 2025 · similarity 0.86)
- A Unified, Resilient, and Explainable Adversarial Patch Detector (CVPR 2025 · similarity 0.86)
- Rethinking the Readout: Unlocking Video Backbones for AI-Generated Video Detection (NeurIPS 2026 · similarity 0.85)
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.
