Research map / Image generation / Autoregressive and latent generation
Prior-Conditioned Gaussian Discriminants for Generalizable AI-generated Image Detection
Shashank Kotyan, Makoto Shing, Yuki Imajuku, Rujikorn Charakorn, Tarin Clanuwat
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 |
|---|---|---|---|
| replace released head with Mah-NCM on same features | UnivFD · Percept-Lens, CNNSpot prior | mean class accuracy (%): 54.63 -> 57.62 | Improved |
| replace released head with Mah-NCM on same features | AIDE (Full GenImage) · Percept-Lens, GenImage prior | mean class accuracy (%): 54.98 -> 64.64 | Improved |
| replace released head with Mah-NCM on same features | DRCT-UnivFD (Full GenImage) · Percept-Lens, GenImage prior | mean class accuracy (%): 65.77 -> 73.01 | Improved |
| replace released head with QDA on same features | DRCT-UnivFD (DRCT-SDv2) · Percept-Lens, DRCT prior | mean class accuracy (%): 62.37 -> 70.31 | Improved |
| replace released head with Mah-NCM on same features | CF-384 · Percept-Lens, CF prior | mean class accuracy (%): 87.54 -> 84.55 | Hurt |
| use shared full covariance (Mah-NCM) instead of isotropic Euc-NCM (ablation) | PE-Core-bigG-14-448 head · Percept-Lens, CF prior | mean class accuracy (%): 86.38 -> 94.46 | Improved |
| use shared covariance (Mah-NCM) instead of diagonal GNB (ablation) | DINOv3-ViT-7b head · Percept-Lens, CF prior | mean class accuracy (%): 83.18 -> 88.91 | Improved |
| use class-specific covariance (QDA) instead of Mah-NCM (ablation) | PE-Core-bigG-14-448 head · Percept-Lens, CF prior | mean class accuracy (%): 94.46 -> 86.15 | Hurt |
| use class-specific covariance (QDA) instead of Mah-NCM (ablation) | PE-Core-bigG-14-448 head · Percept-Lens, DRCT prior | mean class accuracy (%): 83.25 -> 91.09 | Improved |
| change training prior from CNNSpot to CF (ablation) | PE-Core-bigG-14-448 Gaussian head · Percept-Lens | mean class accuracy (%): 77.92 -> 94.46 | Improved |
| scale encoder from CLIP ResNet-50 to ViT-L/14 (ablation) | Mah-NCM head · Percept-Lens, CF prior | mean class accuracy (%): 72.00 -> 76.96 | Improved |
| change pretraining objective from MAE-Huge to DINOv3-ViT-7b (ablation) | Mah-NCM head · Percept-Lens, CF prior | mean class accuracy (%): 69.37 -> 88.91 | Improved |
Closest papers by meaning
- Orthogonal Subspace Decomposition for Generalizable AI-Generated Image Detection (ICML 2025 · similarity 0.86)
- Breaking Latent Prior Bias in Detectors for Generalizable AIGC Image Detection (NeurIPS 2025 · similarity 0.86)
- A Bias-Free Training Paradigm for More General AI-generated Image Detection (CVPR 2025 · similarity 0.86)
- A Sanity Check for AI-generated Image Detection (ICLR 2025 · similarity 0.86)
- What Would Gauss Say About Representations? Probing Pretrained Image Models using Synthetic Gaussian Benchmarks (ICML 2024 · similarity 0.86)
- Discriminative Score Function: Turning Pretrained Models into Functional Generative Priors (NeurIPS 2026 · 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.
