Research map / Image generation / Autoregressive and latent generation
Where Did I Come From? Origin Attribution of AI-Generated Images
Zhenting Wang, Chen Chen, Yi Zeng, Lingjuan Lyu, Shiqing Ma
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 MSE instead of SSIM as reconstruction distance (ablation) | RONAN · StyleGAN2-ADA CIFAR-10 | detection accuracy: 85.5% -> 97.0% | Improved |
| Use MSE instead of MAE (ablation) | RONAN · StyleGAN2-ADA CIFAR-10 | detection accuracy: 94.5% -> 97.0% | Improved |
| Use MSE instead of LPIPS (ablation) | RONAN · StyleGAN2-ADA CIFAR-10 | detection accuracy: 96.0% -> 97.0% | Improved |
| Add reconstruction loss calibration (ablation) | RONAN · Stable Diffusion v2 | detection accuracy: 75.0% -> 87.5% | Improved |
| Replace DDIM inversion with RONAN reverse-engineering (ablation) | Stable Diffusion v2 inversion · Stable Diffusion v2 | detection accuracy: 55.5% -> 87.5% | Improved |
| Replace Parmar et al. inversion with RONAN reverse-engineering (ablation) | Stable Diffusion v2 inversion · Stable Diffusion v2 | detection accuracy: 62.5% -> 87.5% | Improved |
| Optimize in W+ space instead of Z space (ablation) | RONAN on StyleGAN2-ADA · CIFAR-10 | detection accuracy: 97.0% -> 96.0% | No clear effect |
| Optimize in W+ space instead of Z space (ablation) | RONAN on StyleGAN XL · ImageNet | detection accuracy: 93.0% -> 93.5% | No clear effect |
| Change reference model for calibration (Consistency Model, StyleGAN XL, Stable Diffusion) (ablation) | RONAN · StyleGAN2-ADA CIFAR-10 | detection accuracy: 97.0%, 95.0%, 96.0% | No clear effect |
| Use 2D entropy instead of reference model as complexity measure (ablation) | RONAN · StyleGAN2-ADA CIFAR-10 | detection accuracy: 97.0% | No clear effect |
| Increase training data overlap of M2 with M1 from 50% to 90% | RONAN · DCGAN CIFAR-10 | detection accuracy: 98.5% -> 96.5% | No clear effect |
| Increase box blur size from 1 to 4 | RONAN under adaptive attack · DCGAN CIFAR-10 | detection accuracy: 92.5% -> 53.5% | Hurt |
Closest papers by meaning
- Single-Model Attribution of Generative Models Through Final-Layer Inversion (ICML 2024 · similarity 0.85)
- How to Trace Latent Generative Model Generated Images without Artificial Watermark? (ICML 2024 · similarity 0.85)
- CO-SPY: Combining Semantic and Pixel Features to Detect Synthetic Images by AI (CVPR 2025 · similarity 0.82)
- Where's the Liability in the Generative Era? Recovery-based Black-Box Detection of AI-Generated Content (CVPR 2025 · similarity 0.81)
- ManiFPT: Defining and Analyzing Fingerprints of Generative Models (CVPR 2024 · similarity 0.81)
- Dual Data Alignment Makes AI-Generated Image Detector Easier Generalizable (NeurIPS 2025 · similarity 0.81)
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.
