Research map / Image generation / Autoregressive and latent generation
Leveraging Representations from Intermediate Encoder-blocks for Synthetic Image Detection
Christos Koutlis, Symeon Papadopoulos
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 intermediate CLIP block CLS tokens instead of last-layer features only (ablation) | RINE · 20 test datasets, all generators | average accuracy: 80.8 -> 89.8 | Improved |
| add Trainable Importance Estimator (TIE) (ablation) | RINE · 20 test datasets, all generators | average accuracy: 87.3 -> 89.8 | Improved |
| add supervised contrastive loss (ablation) | RINE · 20 test datasets, all generators | average accuracy: 88.4 -> 89.8 | Improved |
| add contrastive loss in 2-class setting (ablation) | RINE 2-class · 20 test datasets | accuracy: 87.9 -> 87.7 | Hurt |
| use CLIP ViT-L/14 instead of ViT-B/32 (ablation) | RINE · 20 test datasets | ACC and AP | Improved |
| vary contrastive loss factor xi (ablation) | RINE · 20 test datasets | ACC and AP | No clear effect |
| train for more epochs (3 to 15) instead of 1 (ablation) | RINE · 20 test datasets | ACC and AP | Hurt |
| increase training set size from 20% to 100% (ablation) | RINE · 20 test datasets | ACC and AP | No clear effect |
| replace projections and TIE with Feature Pyramid Network fusion (ablation) | RINE · 20 test datasets (4-class) | average accuracy: 91.5 -> 69.6 | Hurt |
| use ImageNet ViT or Wang detector backbone instead of CLIP (ablation) | RINE · 20 test datasets (4-class) | average accuracy: 91.5 -> 60.1 (ViT); 91.5 -> 64.7 (Wang) | Hurt |
| train on 20 classes instead of 4 (ablation) | RINE · 20 test datasets | average accuracy: 91.5 -> 90.7 | No clear effect |
| apply combined perturbations at test time | RINE 1-class · 20 test datasets | average accuracy: 90.3 -> 80.6 | Hurt |
Closest papers by meaning
- LEGION: Learning to Ground and Explain for Synthetic Image Detection (ICCV 2025 · similarity 0.89)
- FiSeR: Fine-Grained Source Representations for Cross-Domain AI Image Detection (ICML 2026 · similarity 0.88)
- Forgery-aware Adaptive Transformer for Generalizable Synthetic Image Detection (CVPR 2024 · similarity 0.88)
- CO-SPY: Combining Semantic and Pixel Features to Detect Synthetic Images by AI (CVPR 2025 · similarity 0.88)
- Preserving Forgery Artifacts: AI-Generated Video Detection at Native Scale (ICLR 2026 · similarity 0.88)
- Community Forensics: Using Thousands of Generators to Train Fake Image Detectors (CVPR 2025 · 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.
