Research map / Image generation / Autoregressive and latent generation
Reverse Engineering of Generative Models: Inferring Model Hyperparameters From Generated Images
Vishal Asnani, Xi Peng Yin, Tal Hassner, Xiaoming Liu
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 fingerprint estimation constraints (all four losses) (ablation) | model parsing framework · network architecture continuous | L1 error: 0.170 -> 0.149 | Improved |
| add magnitude loss to fingerprint estimation (ablation) | model parsing framework · network architecture discrete | F1 score: 0.674 -> 0.718 | Improved |
| add spectrum loss to fingerprint estimation (ablation) | model parsing framework · network architecture | L1 error and F1: L1 0.149 -> 0.149; F1 0.676 -> 0.718 | Mixed |
| add repetitive loss to fingerprint estimation (ablation) | model parsing framework · loss function prediction | F1 score: 0.794 -> 0.813 | Improved |
| add energy loss to fingerprint estimation (ablation) | model parsing framework · network architecture continuous | L1 error: 0.162 -> 0.149 | Improved |
| use cluster parser plus instance parser instead of one parser (ablation) | model parsing framework · network architecture continuous | L1 error: 0.161 -> 0.149 | Improved |
| use weighted cross-entropy instead of standard cross-entropy for discrete parameters (ablation) | model parsing framework · discrete parameter confusion matrices | class prediction balance | Improved |
| average predictions over 10 test images instead of 1 (ablation) | model parsing framework · network architecture continuous | L1 error: 0.215 -> 0.151 | Improved |
| increase test images per GM from 100 to 500 (ablation) | model parsing framework · network architecture continuous | L1 error: 0.145 -> 0.146 | No clear effect |
| train on full 110 GMs instead of face GMs only (ablation) | model parsing framework · face test GMs, network architecture continuous | L1 error: 0.139 -> 0.118 | Improved |
| train on non-face GMs instead of face GMs when testing on face GMs (ablation) | model parsing framework · face test GMs, network architecture continuous | L1 error: 0.139 -> 0.213 | Hurt |
| use FEN plus parsing network instead of FEN alone (ablation) | coordinated attack detector · coordinated misinformation attack | AUC (%): 83.5 -> 87.3 | Improved |
Closest papers by meaning
- ManiFPT: Defining and Analyzing Fingerprints of Generative Models (CVPR 2024 · similarity 0.87)
- Community Forensics: Using Thousands of Generators to Train Fake Image Detectors (CVPR 2025 · similarity 0.86)
- WOUAF: Weight Modulation for User Attribution and Fingerprinting in Text-to-Image Diffusion Models (CVPR 2024 · similarity 0.86)
- CO-SPY: Combining Semantic and Pixel Features to Detect Synthetic Images by AI (CVPR 2025 · similarity 0.85)
- WildFake: A Large-Scale and Hierarchical Dataset for AI-Generated Images Detection (AAAI 2025 · similarity 0.85)
- Single-Model Attribution of Generative Models Through Final-Layer Inversion (ICML 2024 · similarity 0.84)
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.
