Research map / Image generation / Autoregressive and latent generation
Beyond Artifacts: Real-Centric Envelope Modeling for Reliable AI-Generated Image Detection
Ruiqi Liu, Yi Han, Zhengbo Zhang, Liwei Yao, Zhiyuan Yan, Jialiang Shen, ZhiJin Chen, Manni Cui, Boyi Sun, Lubin Weng, Jing Dong, Yan Wang and 1 more
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 feature-level perturbation in MBR instead of plain VAE reconstruction (ablation) | REM · RealChain (Chain Degradations) | B.Acc: 3.4% drop without it | Improved |
| add tangency regularization to the envelope estimator (ablation) | REM · RealChain (Chain Degradations) | B.Acc: 2.7% drop without it | Improved |
| add cross-domain consistency (anchor and residual) with degradation augmentation (ablation) | REM · RealChain (Chain Degradations) | B.Acc: 16.4% drop without it | Improved |
| train on DDA-aligned data instead of GenImage (ablation) | AIDE · Seven-benchmark average | Avg B.Acc: 68.0 -> 72.7 | Improved |
| train on DDA-aligned data instead of GenImage (ablation) | DINOv3 · Seven-benchmark average | Avg B.Acc: 87.4 -> 90.3 | Improved |
| train with REM instead of DDA data (ablation) | DINOv3 · Seven-benchmark average | Avg B.Acc: 90.3 -> 93.5 | Improved |
| scale up DINO backbone (v3-L/16 to v3-7B/16) (ablation) | REM · RealChain (Chain Degradations) | B.Acc: 83.3 -> 86.1 | Improved |
| use DINOv3 instead of DINOv2 backbone (L vs L) (ablation) | REM · RealChain (Chain Degradations) | B.Acc: 78.5 -> 83.3 | Improved |
| increase number of real training samples up to 100k (ablation) | REM · all benchmarks | accuracy | Improved |
| replace artifact-driven detectors with REM | REM vs DDA · RealChain (Chain Degradations) | B.Acc: 65.8 -> 84.2 | Improved |
| REM vs DDA on in-the-wild benchmarks | REM · In-the-wild (4 datasets) | Avg B.Acc: 88.3 -> 94.8 | Improved |
| REM vs DDA on source attribution | REM · AIGCDetect source attribution | accuracy: 68.0 -> 75.7 | Improved |
Closest papers by meaning
- Stay-Positive: A Case for Ignoring Real Image Features in Fake Image Detection (ICML 2025 · similarity 0.88)
- Community Forensics: Using Thousands of Generators to Train Fake Image Detectors (CVPR 2025 · similarity 0.86)
- Aligned Datasets Improve Detection of Latent Diffusion-Generated Images (ICLR 2025 · similarity 0.86)
- Where's the Liability in the Generative Era? Recovery-based Black-Box Detection of AI-Generated Content (CVPR 2025 · similarity 0.86)
- A Bias-Free Training Paradigm for More General AI-generated Image Detection (CVPR 2025 · similarity 0.86)
- Is Artificial Intelligence Generated Image Detection a Solved Problem? (NeurIPS 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.
