Research map / Image generation / Autoregressive and latent generation
MIRROR: Manifold Ideal Reference ReconstructOR for AI-Generated Image Detection
Ruiqi Liu, Manni Cui, Ziheng Qin, Zhiyuan Yan, Ruoxin Chen, Yi Han, Zhiheng Li, Junkai Chen, ZhiJin Chen, Kaiqing Lin, Jialiang Shen, Lubin Weng and 2 more
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MIRROR, an AI-generated image detector that anchors detection on the stationary real-image distribution rather than on forgery artifacts, to generalise to unseen generators.
Problem
AI-generated image (AIGI) detectors typically rely on artifact classification and fail to generalise to new generators as generative models become more photorealistic.
Approach
Anchor detection on the stationary real-image distribution (modelled as a manifold with a reconstruction reference) instead of chasing the non-stationary forgery distribution, yielding a more generalisable signal.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| build memory bank from real images only instead of generated-only (ablation) | MIRROR · Standard benchmarks (avg) | B.Acc: 90.5 -> 94.0 | Improved |
| build memory bank from real images only instead of real+generated mix (ablation) | MIRROR · Standard benchmarks (avg) | B.Acc: 87.1 -> 94.0 | Improved |
| add reconstruction perplexity signal to classifier baseline (ablation) | MIRROR baseline · Standard benchmarks (avg) | B.Acc: 88.4 -> 90.1 | Improved |
| add comparison residual signal to classifier baseline (ablation) | MIRROR baseline · Standard benchmarks (avg) | B.Acc: 88.4 -> 92.5 | Improved |
| use full Reference-Comparison (both signals) instead of direct classification (ablation) | MIRROR · Human-AIGI (avg) | accuracy: 83.7 -> 89.8 | Improved |
| scale memory bank size K beyond 4096 or below it (ablation) | MIRROR · sensitivity analysis | accuracy | Mixed |
| vary top-k sparsity of prototype projection (ablation) | MIRROR · sensitivity analysis | accuracy | Mixed |
| switch backbone from DINOv2 to DINOv3 | MIRROR · Standard benchmarks (avg) | B.Acc: 92.4 -> 94.0 | Improved |
| switch backbone from DINOv2 to DINOv3 | MIRROR · In-the-wild (avg) | B.Acc: 87.4 -> 91.2 | Improved |
| MIRROR vs B-Free on DRCT-2M | MIRROR · DRCT-2M | B.Acc: 99.2 -> 93.0 | Hurt |
| scale pretrained backbone from Base to Huge for artifact-based detectors | NPR, UnivFD, DDA · Human-AIGI | accuracy | No clear effect |
Closest papers by meaning
- Dual-Branch Asymmetric Discrepancy Learning Based on Fake Image Pattern-Coexistence for AI-Generated Image Detection (AAAI 2026 · similarity 0.88)
- A Bias-Free Training Paradigm for More General AI-generated Image Detection (CVPR 2025 · similarity 0.87)
- Is Artificial Intelligence Generated Image Detection a Solved Problem? (NeurIPS 2025 · similarity 0.87)
- Community Forensics: Using Thousands of Generators to Train Fake Image Detectors (CVPR 2025 · similarity 0.87)
- Stay-Positive: A Case for Ignoring Real Image Features in Fake Image Detection (ICML 2025 · similarity 0.87)
- CO-SPY: Combining Semantic and Pixel Features to Detect Synthetic Images by AI (CVPR 2025 · similarity 0.86)
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.
