MIRROR: Manifold Ideal Reference ReconstructOR for Generalizable AI-Generated Image Detection

TL;DR

This work reframes AI-generated image detection as a Reference-Comparison problem grounded in real-world priors, addressing the generalization shortcomings of artifact-based detectors. MIRROR encodes the real-image manifold into a discrete memory bank of orthogonal prototypes and constructs an ideal reference via sparse projection; detection hinges on reconstruction residuals and perplexity signals that reveal deviations from reality. The authors introduce the Human-AIGI Benchmark with a Human-Imperceptible Subset to quantify the boundary where detectors surpass human perception, showing MIRROR achieves strong cross-generator generalization and near-human performance as backbone capacity scales. Across 14 benchmarks, MIRROR delivers consistent gains and demonstrates robustness to common degradations, making it a practical tool for reliable AIGI detection in real-world, multi-generator scenarios.

Abstract

High-fidelity generative models have narrowed the perceptual gap between synthetic and real images, posing serious threats to media security. Most existing AI-generated image (AIGI) detectors rely on artifact-based classification and struggle to generalize to evolving generative traces. In contrast, human judgment relies on stable real-world regularities, with deviations from the human cognitive manifold serving as a more generalizable signal of forgery. Motivated by this insight, we reformulate AIGI detection as a Reference-Comparison problem that verifies consistency with the real-image manifold rather than fitting specific forgery cues. We propose MIRROR (Manifold Ideal Reference ReconstructOR), a framework that explicitly encodes reality priors using a learnable discrete memory bank. MIRROR projects an input into a manifold-consistent ideal reference via sparse linear combination, and uses the resulting residuals as robust detection signals. To evaluate whether detectors reach the "superhuman crossover" required to replace human experts, we introduce the Human-AIGI benchmark, featuring a psychophysically curated human-imperceptible subset. Across 14 benchmarks, MIRROR consistently outperforms prior methods, achieving gains of 2.1% on six standard benchmarks and 8.1% on seven in-the-wild benchmarks. On Human-AIGI, MIRROR reaches 89.6% accuracy across 27 generators, surpassing both lay users and visual experts, and further approaching the human perceptual limit as pretrained backbones scale. The code is publicly available at: https://github.com/349793927/MIRROR

Figures (10)

• Figure 1: Comparison of decision paradigms for AI-generated image detection. (Left) Traditional detectors treat detection as a Artifacts-Driven classification task, often overfitting to generator-specific artifacts. (Middle) Humans employ a Reference-Comparison strategy, anchoring judgment on the invariant priors of the physical world. (Right) Our MIRROR framework explicitly encodes these priors to act as a Reconstructor: it projects the input into a manifold-consistent Ideal Reference. The resulting reconstruction residual exposes forgeries as significant deviations from reality.
• Figure 2: Scalability comparison of AIGI detection between detectors and humans on Human-AIGI. While existing methods (NPR tan2024rethinking, UnivFD ojha2023towards, and DDA chen2025dual) exhibit saturated performance as model size increases, humans demonstrate strong prior-driven scalability. In contrast, MIRROR exhibits similar human-like behavior, achieving sustained performance gains as model capacity scales from Base to Huge.
• Figure 3: Psychophysical Analysis on the Human-AIGI Benchmark. Results are obtained from controlled psychophysical experiments on AI-generated images from 27 generators. (a) Reference Priors. Providing explicit reference images shifts hard samples (red) from high uncertainty with long response times to a higher-confidence regime (blue). (b) Expertise Priors. CV experts (green) consistently outperform lay users (red), reflecting the benefit of stronger internal perceptual priors. (c) Human-Imperceptible Subset. The Hard subset (red) includes samples with high deception or confusion, measured by confidence and response time; unlike Easy Fakes (gray), these images preserve high visual fidelity and define a rigorous benchmark.
• Figure 4: Architecture of MIRROR. The framework operates in two distinct phases. In Phase 1, a frozen DINO encoder extracts patch-level features from real images to encode Reality Priors within an orthogonal memory bank that approximates the real-image manifold. In Phase 2, the learned priors are frozen. Given an input image, MIRROR constructs a manifold-consistent Ideal Reference through sparse projection and identifies forgeries via Reference-Comparison. This process utilizes both the reconstruct perplexity and the comparison residual to evaluate whether the input deviates from the learned reality manifold.
• Figure 5: Sensitivity analysis of $K$ and $top$-$k$. Insufficient $K$ limits manifold coverage, while excessive $K$ causes overfitting via prototype redundancy. Sparsity $top$-$k$ exhibits a similar trade-off: strict $k$ hinders ideal reference reconstruction , while excessive $k$ allows generative anomalies to leak through prototype combinations, narrowing the residual gap.