PaperlayerView exact revision on arXiv

Source-linked AI summary

MIRROR: Manifold Ideal Reference ReconstructOR for Generalizable 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, Jing Dong, Yan Wang, Shu Wu

arXiv:2602.02222v1cs.CVcs.CR

TL;DR

AIGI detectors often rely on artifacts that generalize poorly as generators evolve, while human judgment uses stable real-world regularities. MIRROR models those regularities with a learnable memory bank, reconstructs an ideal real-image reference, and detects deviations through residuals. Across 14 benchmarks, it consistently outperforms prior methods and approaches human perceptual performance on human-imperceptible images.

  • Problem

    Artifact-based AIGI detectors struggle to generalize to evolving generative traces, motivating detection based on stable real-world regularities.

  • Method

    MIRROR encodes reality priors in a learnable memory bank, sparsely projects inputs onto the real-image manifold, and uses reconstruction residuals for Reference-Comparison detection.

  • Results

    Across 14 benchmarks, MIRROR consistently outperforms existing methods, including 2.1% average gains on six standard benchmarks and 8.1% on seven in-the-wild benchmarks.

  • Takeaways & Limitations

    MIRROR surpasses human performance on human-imperceptible samples and approaches the human perceptual limit as pretrained backbones scale.

Abstract

from arXiv · show

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

1 Introduction

High-fidelity generation challenges artifact-based detectors, which generalize poorly as generators evolve. MIRROR instead compares inputs with learned real-image references, while Human-AIGI measures progress toward surpassing human perception.

  • Motivation: Existing artifact-based detectors degrade on unknown generators because forged distributions and their traces evolve.Pretrained approaches can outperform non-pretrained methods, but their gains saturate as backbone size increases.
  • Motivation: Humans generalize better by using stable real-world regularities to construct internal ideal references and compare observed images against them.Reference images improve human detection, and visual experts outperform non-experts.
  • Benchmark: The Human-AIGI Benchmark uses psychophysical experiments with 50 participants to curate human-imperceptible samples and quantify the “Superhuman Crossover.”The benchmark records accuracy, confidence, and response latency.
  • Results: 2.1% average accuracy gain appears on six standard benchmarks, while 8.1% appears on seven in-the-wild benchmarks.On the human-imperceptible Human-AIGI subset, MIRROR reaches 89.5% accuracy and outperforms prior methods by 2.6%.
  • Results: MIRROR’s accuracy approaches the human perceptual limit as pretrained backbone capacity scales.The reported result positions MIRROR as a practical assistive tool for AIGI detection.
  • MIRROR: MIRROR encodes real-image regularities with a learnable memory bank and detects forgeries from residuals between inputs and manifold-consistent ideal references.The framework reformulates detection as Reference-Comparison rather than artifact fitting.

2 Related Work

Prior AIGI detectors use artifacts, pretrained representations, or reconstruction and alignment strategies to improve generalization. However, existing methods remain vulnerable to generator evolution and generally lack explicit human-like priors and verification mechanisms.

  • Artifact-based detection: Early detectors exploit generator-specific spatial, spectral, frequency, or upsampling artifacts, but these texture-level cues are fragile under generator evolution.CNNSpot demonstrated cross-GAN transferability, while later work expanded artifact types.
  • Pretrained representations: Foundation-model methods such as UnivFD, FatFormer, AIDE, and Effort leverage pretrained representations through classifiers, adaptation modules, or disentangled subspaces.These methods aim to improve robustness beyond low-level artifact cues.
  • Bias mitigation: Reconstruction and alignment methods mitigate data bias and shortcut learning by relating real and synthetic data.Representative approaches include FakeInversion, SemGIR, DRCT, B-Free, AlignedForensics, and Dual Data Alignment.
  • Open limitation: Existing detectors largely remain black-box discriminators without explicit hierarchical priors and verification mechanisms comparable to human perception.This limits the extent to which their decision process models stable real-world regularities.
  • Benchmarks: Benchmarks have expanded from GAN datasets to diverse generators, diffusion reconstruction, cross-model testing, and realistic web-image degradation.Examples include GenImage, AIGCDetect, DRCT-2M, UniversalFakeDetect, Synthbuster, Chameleon, SynthWildx, WildRF, AIGIBench, RRDataset, and RealChain.

3 Motivation

The paper motivates AIGI detection as hypothesis-driven Reference-Comparison rather than artifact memorization. Psychophysical evidence links stronger real-image priors and explicit references to better detection, motivating MIRROR’s manifold-consistent residual comparison.

  • Decision paradigms: Artifact classification overfits generator-specific traces, whereas human vision anchors judgments in invariant physical regularities.This contrast is framed as a difference between artifact classification and Reference-Comparison.
  • Reference priors: Explicit real-image references significantly improve lay users’ detection accuracy on generated images.The experiment supports reference-based comparison as a useful perceptual aid.
  • Human-imperceptible benchmark: The Human-AIGI Hard subset selects highly deceptive or confusing images using confidence and response-time criteria.These samples preserve high visual fidelity and provide a stringent human-imperceptible testbed.
  • Expertise priors: CV experts outperform lay users, consistent with stronger perceptual priors acquired through prolonged exposure to real images.The comparison concerns computer-vision experts rather than forensic experts.
  • MIRROR: MIRROR operationalizes Reference-Comparison by modeling the real-image manifold and measuring residuals between each observation and its manifold-consistent Ideal Reference.This design targets robust, generator-agnostic detection.

4 Methodology

MIRROR learns reality priors from real images, constructs a manifold-consistent reference through sparse projection, and detects forgeries from reconstruction and comparison evidence.

  • Framework overview: MIRROR uses a two-stage framework: it first encodes reality priors, then detects forgeries by comparing inputs with manifold-consistent ideal references.The memory bank and projection module are frozen before Phase 2 detection.
  • Phase 1: Encoding Reality Priors: A frozen DINO encoder extracts patch-level features, while an orthogonal memory bank stores prototypes representing frequent, stable real-world patterns.The memory bank is trained exclusively on real images to approximate the real-image manifold.
  • Phase 1: Encoding Reality Priors: Top-k cross-attention projects input features onto the memory-bank subspace, producing a sparse linear combination of real-world prototypes.This sparse projection filters patterns that do not belong to the real-image manifold.
  • Phase 1: Encoding Reality Priors: Phase 1 minimizes reconstruction error while enforcing prototype orthogonality to preserve accurate, diverse, and nonredundant reality priors.The objective is optimized using real images only.
  • Phase 2: Reference-Comparison-Based Detection: Phase 2 combines reconstruct perplexity and comparison residual to assess retrieval uncertainty and irreconstructible details outside the real-image manifold.These signals are projected into evidence vectors and concatenated for the final classification head.
  • Phase 2: Reference-Comparison-Based Detection: MIRROR identifies forgeries when the real-world manifold cannot sparsely reconstruct or accurately explain the input.The final forgery probability is trained with binary cross-entropy.

5 The Human-AIGI Benchmark

The Human-AIGI benchmark uses psychophysical evaluation to identify generated images that challenge human perception and to test whether detectors can replace expert review.

  • Benchmark design: The benchmark evaluates lay users, computer vision experts, and AIGI detection experts on real-versus-generated image judgments.Participants assess images synthesized by 27 generators in a controlled binary classification task.
  • Subset construction: The Human-Imperceptible subset contains generated samples that either deceive observers into judging them real or induce significant hesitation.The subset is defined from psychophysical observations of human perception.
  • Subset construction: Dhard includes samples whose perceived realism score reaches τreal or whose response time exceeds µrt + σrt.S(x) is the perceived realism score, while µrt and σrt are benchmark response-time statistics.
  • Evaluation purpose: Joint evaluation on Dorig. and Dhard tests whether detectors can reliably identify threats that evade human perceptual scrutiny and replace expert review.The subset is intended as a stringent testbed for detector performance near the human perceptual limit.

6 Experiments

Experiments evaluate MIRROR across standard, in-the-wild, human-imperceptible, ablation, robustness, and efficiency settings. Results show strong generalization, robustness, human-level comparison, and deployment efficiency.

  • Experimental setup: MIRROR is evaluated on diverse OOD benchmarks using Balanced Accuracy, JPEG Robustness, and Resize Robustness metrics.The evaluation covers standard and in-the-wild datasets, along with human-imperceptible samples and architectural ablations.
  • Comparison with state-of-the-art: 94.0% average Balanced Accuracy on standard benchmarks substantially surpasses AIDE, while MIRROR exceeds B-Free by 2.0% in JPEG Robustness and 3.5% in Resize Robustness.The comparison emphasizes cross-generator generalization and robustness to mild image processing.
  • Comparison with state-of-the-art: 91.2% Balanced Accuracy across seven in-the-wild benchmarks outperforms DDA by 8.1% under diverse post-processing conditions.MIRROR uses readily available real images for prior learning and only a small amount of generated data for fine-tuning.
  • Comparison with state-of-the-art: 89.5% balanced accuracy on the human-imperceptible subset exceeds visual experts’ 88.3% and lay users’ 76.9% performance.On the original dataset, MIRROR reaches up to 89.6% balanced accuracy across generators.
  • Ablation study and analysis: K = 4096 and top-k = 128 provide a balanced configuration between manifold coverage, reconstruction quality, and residual separability.Both capacity and sparsity exhibit inverted-U trade-offs, with excessive values causing redundancy, overfitting, or artifact absorption.
  • Ablation study and analysis: Real-only memory prototypes perform best, while mixing authentic and forged prototypes severely reduces residual discriminability.Direct feature classification also performs 5.4% worse than the Reference-Comparison paradigm.
  • Computational efficiency: MIRROR reaches 20.03 FPS, processing 1,000 samples in 49.91 seconds on a Tesla V100, with competitive inference efficiency.The DINOv3-large backbone dominates inference cost despite the memory bank and reference reconstructor.

7 Conclusion

The paper identifies poor scalability and generalization as fundamental limitations of artifact-driven AIGI detectors and proposes MIRROR’s Reference-Comparison formulation as an alternative. Experiments, including Human-AIGI, show consistent gains and performance above humans on human-imperceptible samples.

  • Conclusion: Artifact-driven AIGI detectors suffer from poor scalability and generalization as generative models evolve.MIRROR addresses this by modeling the real-image manifold with a learnable memory bank and comparing inputs against reconstructed references.
  • Conclusion: MIRROR consistently outperforms existing methods and surpasses human performance on human-imperceptible samples.The conclusion frames Reference-Comparison and the Human-AIGI benchmark as central components of the evidence.

A.1 Detailed Breakdown of Benchmark Performance

Additional benchmark analyses examine diverse generative paradigms and ensure architectural comparisons are conducted under both official and format-aligned training settings.

  • Benchmark coverage: The detailed breakdown covers architectures from early GANs to modern diffusion models and evaluates robustness against unseen processing pipelines.The supplementary results include AIGCDetect, DRCT-2M, and UniversalFakeDetect.
  • Evaluation protocol: Baseline detectors are evaluated with both officially released weights and weights retrained on a format-aligned SDv1.4 dataset.The retrained setting controls format alignment and data-scale consistency across methods.

B Robustness to Common Corruptions

MIRROR remains stable under JPEG compression, resizing, and Gaussian blurring, where artifact-driven baselines are vulnerable to changes in low-level image statistics.

  • Common corruptions: MIRROR consistently outperforms state-of-the-art baselines across JPEG compression, spatial resizing, and Gaussian blurring.The stress tests vary JPEG quality from 100 to 40, scale factors from 0.5× to 2.0×, and Gaussian sigma from 0 to 2.0.
  • JPEG compression: MIRROR preserves detection under JPEG compression by relying on manifold-consistent structural residuals rather than fragile high-frequency statistics.JPEG quality factors range from 100 down to 40.
  • Resizing operations: MIRROR’s sparse linear combination strategy preserves structural integrity across resizing from 0.5× to 2.0×.The resulting detection process remains largely invariant to scale shifts.
  • Gaussian blurring: MIRROR remains stable under Gaussian blurring with σ from 0 to 2.0 by identifying global semantic discordance rather than local sharpness cues.Blurring suppresses high-frequency components used by several baseline detectors.

C Heatmap Visualization

MIRROR visualizes reconstruction residuals to show where images depart from the learned manifold of authentic images. Authentic images produce uniformly low residuals, while generated images reveal localized forensic inconsistencies.

  • MIRROR derives heatmaps directly from the residual R = ∥x −ˆx∥2 between an input feature and its manifold projection.This provides a residual-based alternative to Grad-CAM for examining generative discrepancies.
  • Authentic images exhibit uniformly low residuals across their spatial extent, indicating close adherence to the learned natural-image manifold.
  • AI-generated images show localized high-residual clusters that pinpoint irregular illumination, blurred texture boundaries, and geometric distortions.
  • The heatmaps localize generative artifacts that may be imperceptible to human observers, supporting transparent interpretation of MIRROR’s classifications.

D Visual Showcase of the Human-AIGI Benchmark

The Human-AIGI Benchmark is presented as a broad, high-fidelity collection designed to reflect modern commercial generative systems and difficult detection conditions. Its visual showcase covers diverse samples from many generators and a psychophysically curated human-imperceptible subset.

  • The benchmark emphasizes cross-generator heterogeneity and high perceptual indistinguishability rather than single-generator or easily recognizable artifacts.
  • The showcase includes portraits, objects, and complex scenes generated by 27 state-of-the-art systems in T2I-COREBENCH.
  • The Human-Imperceptible Subset, Dhard, consists of perceptual corner cases identified through psychophysical curation.
  • The supplied benchmark tables report comparisons on AIGCDetect, DRCT-2M, UniversalFakeDetect, Synthbuster, and EvalGEN.
  • Table 10 compares official and retrained weights using Balanced Accuracy, JPEG Robustness, and Resize Robustness, with special averaging for Synthbuster and EvalGEN.
Loading 2602.02222v1…