Research map / Image generation / Autoregressive and latent generation
Forensic-MoE: Exploring Comprehensive Synthetic Image Detection Traces with Mixture of Experts
Mingqi Fang, Ziguang Li, Lingyun Yu, Quanwei Yang, Hongtao Xie, Yongdong Zhang
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 multiple interacting experts instead of a single universal detector (ablation) | Forensic-MoE · Ablation mean over 16 test subsets | AccM: 93.78 -> 95.14 | Improved |
| add expert interaction finetuning versus independent frozen experts with AWG (ablation) | Forensic-MoE · Ablation mean over 16 test subsets | AccM: 94.03 -> 95.14 | Improved |
| add EKD loss (ablation) | Forensic-MoE · Ablation mean over 16 test subsets | AccM: 94.72 -> 95.14 | Improved |
| use EKD loss instead of single center loss (ablation) | Forensic-MoE · Ablation mean over 16 test subsets | AccM: 94.47 -> 95.14 | Improved |
| add adaptive weight generator AWG (ablation) | Forensic-MoE · Ablation mean over 16 test subsets | AccM: 94.50 -> 95.14 | Improved |
| add multi-stage adapter (ablation) | Forensic-MoE expert · Ablation mean over 16 test subsets | AccM: 93.30 -> 95.14 | Improved |
| add patch decentralization (ablation) | Forensic-MoE expert · Ablation mean over 16 test subsets | AccM: 90.14 -> 95.14 | Improved |
| add DWT frequency information (ablation) | Forensic-MoE expert · Ablation mean over 16 test subsets | AccM: 93.79 -> 95.14 | Improved |
| train experts on GAN + diffusion data rather than 2 GANs or 2 DMs (ablation) | Forensic-MoE · Ablation mean over 16 test subsets | Acc | Improved |
| increase number of experts from 1 to 2 (ablation) | Forensic-MoE · Ablation mean over 16 test subsets | Acc / AP | Improved |
| add diffusion images to training of FatFormer and NPR | FatFormer, NPR · ForenSynths+GenImage (trained on ProGAN+SDv1.4) overall mean | Acc | No clear effect |
| Forensic-MoE vs FatFormer on unseen generators (Kandinsky 2) | Forensic-MoE · FakeInversion unseen | Acc (%): 69.02 -> 88.14 | Improved |
Closest papers by meaning
- Community Forensics: Using Thousands of Generators to Train Fake Image Detectors (CVPR 2025 · similarity 0.86)
- Seeing What Matters: Generalizable AI-generated Video Detection with Forensic-Oriented Augmentation (NeurIPS 2025 · similarity 0.86)
- A Bias-Free Training Paradigm for More General AI-generated Image Detection (CVPR 2025 · similarity 0.85)
- ForensicHub: A Unified Benchmark & Codebase for All-Domain Fake Image Detection and Localization (NeurIPS 2025 · similarity 0.85)
- CO-SPY: Combining Semantic and Pixel Features to Detect Synthetic Images by AI (CVPR 2025 · similarity 0.85)
- Forensic Self-Descriptions Are All You Need for Zero-Shot Detection, Open-Set Source Attribution, and Clustering of AI-generated Images (CVPR 2025 · 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.
