Research map / Image generation / Autoregressive and latent generation
Unveiling Perceptual Artifacts: A Fine-Grained Benchmark for Interpretable AI-Generated Image Detection
Yao Xiao, Weiyan Chen, Jiahao Chen, Zijie Cao, Weijian Deng, Binbin Yang, ZiYi Dong, Xiangyang Ji, Wei Ke, Pengxu Wei, Liang Lin
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 PAD pretraining with linear probing instead of AJ training (ablation) | Swin-B · X-AIGD test AJ | balanced accuracy (%): 89.3 -> 82.1 | Hurt |
| add PAD pretraining then full fine-tuning (ablation) | Swin-B AJ model · X-AIGD test AJ | F1 (%): 90.2 -> 92.5 | Improved |
| add PAD as joint auxiliary task (multi-task) (ablation) | Swin-B AJ model · X-AIGD test AJ | F1 (%): 90.2 -> 92.3 | Improved |
| add PAD as joint auxiliary task (multi-task) (ablation) | Swin-B AJ model · X-AIGD test AJ | balanced accuracy (%): 89.3 -> 89.1 | No clear effect |
| add multi-task training vs PAD-only for PAD (ablation) | Swin-B · X-AIGD test category-agnostic PAD | IoU (%): 27.2 -> 27.3 | No clear effect |
| train on artifact annotations (PAD-only) vs existing detector heatmaps | Swin-B · X-AIGD test category-agnostic PAD | IoU (%): 9.0 -> 27.2 | Improved |
| add attention alignment loss to DINOv2 fine-tuning (ablation) | DINOv2 ViT-B · Synthbuster, CommFor, Chameleon, X-AIGD | F1 | Improved |
| replace artifact masks with saliency masks in attention alignment (ablation) | DINOv2 ViT-B · cross-dataset AIGI detection | F1 | No clear effect |
| raise benign region weight lambda toward 1.0 (ablation) | DINOv2 ViT-B · cross-dataset AIGI detection | F1 | Mixed |
| use multi-task model for high-level and cognitive artifacts (ablation) | Swin-B · X-AIGD test PAD (commonsense, physics) | pixel recall (%): 2.2 -> 2.2 (commonsense); 0.3 -> 0.8 (physics) | No clear effect |
| use general MLLM for artifact presence detection | MLLMs · X-AIGD test artifact presence | image-level F1 (%) | No clear effect |
Closest papers by meaning
- A Sanity Check for AI-generated Image Detection (ICLR 2025 · similarity 0.88)
- Is Artificial Intelligence Generated Image Detection a Solved Problem? (NeurIPS 2025 · similarity 0.88)
- SALART-VQA: Diagnosing Whether VLMs Understand Salient Artifacts in Generated Images (NeurIPS 2026 · similarity 0.88)
- Breaking Semantic Artifacts for Generalized AI-generated Image Detection (NeurIPS 2024 · similarity 0.88)
- Are High-Quality AI-Generated Images More Difficult for Models to Detect? (ICML 2025 · similarity 0.87)
- BIAS-ID: A Framework for Analyzing Transformation Biases in AI-Generated Image Detectors (NeurIPS 2026 · 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.
