Research map / Label-efficient and robust learning / Domain adaptation and generalization
FiSeR: Fine-Grained Source Representations for Cross-Domain AI Image Detection
Shan Zhang, Yongxin He, Mingming Zhang, Huiwen Tian, Lei Ma
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 |
|---|---|---|---|
| train with hierarchical (coarse + fine) contrastive learning on DINOv3 instead of DIRE | FiSeR vs DIRE · Average over WildFake/Community/AIGIBench/Chameleon/GenImage (trained on WildFake) | AUROC: 88.01 -> 98.23 | Improved |
| add fine-grained generator-identity contrastive loss (ablation) | FiSeR · Average over WildFake/Community/AIGIBench/Chameleon/GenImage (trained on WildFake) | AUROC: 94.75 -> 98.23 | Improved |
| add coarse-grained natural-vs-synthetic contrastive loss (ablation) | FiSeR · Average over WildFake/Community/AIGIBench/Chameleon/GenImage (trained on WildFake) | AUROC: 93.64 -> 98.23 | Improved |
| add fine-grained loss (ablation) | FiSeR · AIGIBench (trained on WildFake) | AUROC: 88.83 -> 98.00 | Improved |
| add coarse-grained loss (ablation) | FiSeR · AIGIBench (trained on WildFake) | AUROC: 93.15 -> 98.00 | Improved |
| fine-tune DINOv3 ViT-L/16 on WildFake with contrastive loss instead of pretrained (ablation) | FiSeR · AIGIBench (trained on WildFake) | AUROC: 89.87 -> 98.00 | Improved |
| use larger DINOv3 ViT-H/16+ instead of ViT-L/16 after fine-tuning (ablation) | FiSeR · AIGIBench (trained on WildFake) | AUROC: 98.00 -> 97.24 | Hurt |
| use DINOv3 backbone instead of CLIP ViT-L/14 (fine-tuned) (ablation) | FiSeR · AIGIBench (trained on WildFake) | AUROC: 93.78 -> 98.00 | Improved |
| refit an SVM head with 10 target samples per class on frozen features | ResNet50-Detection · Chameleon | AUROC: 63.65 -> 92.38 | Improved |
| refit an SVM head with 10 target samples per class on frozen features | CLIP-Detection · Chameleon | AUROC: 60.81 -> 85.39 | Improved |
| use SVM instead of k-NN head for few-shot adaptation (ablation) | all methods average · AIGIBench 10-shot | AUROC: 84.12 -> 87.82 | Improved |
| train on Community instead of WildFake | FiSeR · Average over 5 datasets | AUROC: 98.23 -> 91.12 | Hurt |
Closest papers by meaning
- A Bias-Free Training Paradigm for More General AI-generated Image Detection (CVPR 2025 · similarity 0.87)
- Community Forensics: Using Thousands of Generators to Train Fake Image Detectors (CVPR 2025 · similarity 0.87)
- CO-SPY: Combining Semantic and Pixel Features to Detect Synthetic Images by AI (CVPR 2025 · similarity 0.87)
- The Unmet Promise of Synthetic Training Images: Using Retrieved Real Images Performs Better (NeurIPS 2024 · similarity 0.86)
- Training-free Detection of AI-generated images via Cropping Robustness (NeurIPS 2025 · similarity 0.86)
- WildFake: A Large-Scale and Hierarchical Dataset for AI-Generated Images Detection (AAAI 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.
