Research map / Image generation / Autoregressive and latent generation
Aggregating Diverse Cue Experts for AI-Generated Image Detection
Lei Tan, Shuwei Li, Mohan Kankanhalli, Robby T. Tan
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 position embedding shuffle to CI cue (ablation) | CI only · GenImage ablation (trained on SDV1.4, avg of 8 subsets) | avg accuracy: 86.3 -> 89.8 | Improved |
| add CI cue to HF-only model (ablation) | MCAN (HF) · GenImage ablation (trained on SDV1.4, avg of 8 subsets) | avg accuracy: 5.6% | Improved |
| add CI cue to Img-only model (ablation) | MCAN (Img) · GenImage ablation (trained on SDV1.4, avg of 8 subsets) | avg accuracy: 2.4% | Improved |
| add CI-Shuffled cue to Img+HF (ablation) | MCAN (Img, HF) · GenImage ablation (trained on SDV1.4, avg of 8 subsets) | avg accuracy: 95.4 -> 96.9 | Improved |
| add HF cue to Img+CI-Shuffled (ablation) | MCAN (Img, CI-Shuffled) · GenImage ablation (trained on SDV1.4, avg of 8 subsets) | avg accuracy: 92.6 -> 96.9 | Improved |
| add image cue to HF+CI-Shuffled (ablation) | MCAN (HF, CI-Shuffled) · GenImage ablation (trained on SDV1.4, avg of 8 subsets) | avg accuracy: 96.0 -> 96.9 | Improved |
| replace naive prediction averaging with unified MoEA aggregation (ablation) | three-cue model · GenImage ablation (trained on SDV1.4, avg of 8 subsets) | avg accuracy: 95.9 -> 96.9 | Improved |
| use HF cue instead of image cue alone (ablation) | single-cue model · GenImage ablation (trained on SDV1.4, avg of 8 subsets) | avg accuracy: 87.0 -> 93.6 | Improved |
| use fewer encoder experts than the number of cues (ablation) | MoEA · GenImage (trained on SDV1.4) | avg accuracy | Hurt |
| increase encoder experts beyond four (ablation) | MoEA · GenImage (trained on SDV1.4) | avg accuracy: 96.9 | No clear effect |
| insert MoEA in every layer instead of last four (ablation) | MoEA · GenImage (trained on SDV1.4) | avg accuracy | Hurt |
| MCAN vs DRCT on GenImage | MCAN · GenImage (avg of 8 subsets) | avg accuracy: 89.5 -> 96.9 | Improved |
Closest papers by meaning
- Deep Layer Aggregation (CVPR 2018 · similarity 0.85)
- All Patches Matter, More Patches Better: Enhance AI-Generated Image Detection via Panoptic Patch Learning (ICLR 2026 · similarity 0.85)
- A Sanity Check for AI-generated Image Detection (ICLR 2025 · similarity 0.84)
- Cue3D: Quantifying the Role of Image Cues in Single-Image 3D Generation (NeurIPS 2025 · similarity 0.84)
- ImageNet-trained CNNs are not biased towards texture: Revisiting feature reliance through controlled suppression (NeurIPS 2025 · similarity 0.83)
- Is Artificial Intelligence Generated Image Detection a Solved Problem? (NeurIPS 2025 · similarity 0.83)
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.
