The rapid progress of text-to-image models has made AI-generated images increasingly realistic, posing significant challenges for accurate detection of generated content. While training-based detectors often suffer from limited generalization to unseen images, training-free approaches offer better robustness, yet struggle to capture subtle discrepancies between real and synthetic images. In this work, we propose a training-free AI-generated image detection method that measures representation sensitivity to structured frequency perturbations, enabling detection of minute manipulations. The proposed method is computationally lightweight, as perturbation generation requires only a single Fourier transform for an input image. As a result, it achieves one to two orders of magnitude faster inference than most training-free detectors.Extensive experiments on challenging benchmarks demonstrate the efficacy of our method over state-of-the-art (SoTA). In particular, on OpenFake benchmark, our method improves AUC by nearly 10% compared to SoTA, while maintaining substantially lower computational cost.
Figures & tables
Figure 1: (a) Our method first generates a perturbed input by adding frequency-band-limited noise to each patch of the input image. These inputs are fed to Vision Transformer (ViT), then a Real/Fake classification is performed based on the distance between feature representations. (b) UMAP [ 16 ] projection of CLIP embeddings for original (blue, circle) and perturbed (orange, cross) images. The top panel corresponds to real images, and the bottom corresponds to fake (generated) images. As can be observed, real data exhibits distinct regions of original and perturbed images as opposed to fake data. For visualization purposes, we randomly sub-sampled 1,000 samples from OpenFake dataset
Dataset
# of Generators
# of Real
# of Generated
Openfake
34
29829
29829
GenImage
8
50000
50000
Semi-Truth
5
6967
26786
Table 1: Statistics of datasets in our experiments.
Generator
RIGID
MINDER
AERO BLADE
MIBD
ZERO FAKE
DTAD
WAR PAD
HFI
OURS
Aurora-20-1-25
0.473
0.471
0.343
0.788
0.684
0.228
0.420
0.487
0.850
Chroma
0.633
0.633
0.452
0.503
0.571
0.924
0.728
0.432
0.879
DALLE-3
0.469
0.468
0.311
0.572
0.560
0.386
0.525
0.354
0.858
Flux-1.1-pro
0.434
0.434
0.531
0.445
0.605
0.485
0.567
0.527
0.808
Flux-amateur snapshotphotos
0.506
0.505
0.559
0.460
0.502
0.934
0.627
0.498
0.797
Flux-mvc5000
0.371
0.371
0.560
0.501
0.543
0.772
0.381
0.476
0.857
Table 2: AUC scores of all methods evaluated on the OpenFake dataset. Rows correspond to image generators, columns to detection methods, and the final row reports the average AUC across generators. The best results are highlighted in bold.
Generator
RIGID
MINDER
AERO BLADE
MIBD
ZERO FAKE
DTAD
WAR PAD
HFI
OURS
Kandinsky22
0.450
0.451
0.723
0.548
0.507
0.829
0.852
0.606
0.844
OpenJourney
0.498
0.498
0.707
0.499
0.562
0.649
0.680
0.700
0.801
SDXL
0.352
0.352
0.531
0.482
0.525
0.686
0.805
0.659
0.834
SD1.4
0.350
0.351
0.686
0.525
0.679
0.730
0.750
0.975
0.872
SD1.5
0.447
0.448
0.724
0.530
0.668
0.780
0.785
0.967
0.854
Average
0.418
0.419
0.671
0.517
0.583
0.736
0.779
0.771
0.841
Table 3: AUC scores of all methods evaluated on the Semi-Truth dataset.
Generator
RIGID
MINDER
AERO BLADE
MIBD
ZERO FAKE
DTAD
WAR PAD
HFI
OURS
ADM
0.845
0.850
0.690
0.754
0.918
0.935
0.989
0.759
0.974
BigGAN
0.961
0.961
0.839
0.942
0.901
0.973
0.999
0.950
0.886
Glide
0.901
0.900
0.760
0.876
0.924
0.968
0.992
0.863
0.979
Midjourney
0.684
0.681
0.595
0.565
0.781
0.923
0.840
0.576
0.972
SD1.4
0.603
0.606
0.835
0.717
0.904
0.950
0.955
0.937
0.991
SD1.5
0.596
0.600
0.833
0.720
0.907
0.949
0.950
0.943
0.991
Table 4: AUC scores of all methods evaluated on the Genimage dataset.
Method
Openfake
Semi-Truth
Genimage
RIGID
1519.485
860.492
2550.385
MINDER
3019.285
1707.962
5055.123
AEROBLADE
11042.416
6246.216
18476.045
MIBD
75610.938
42668.291
126113.634
ZEROFAKE
2095768.396
1184961.851
3510376.209
DTAD
276861.686
156680.426
463905.786
Table 5: Total inference runtime (in seconds) with a batch size of 8.
Figure 2: ROC curves for the Openfake, Semi-Truth, and Genimage dataset.
Figure 3: Average AUC versus total inference runtime (batch size = 8). Higher AUC and lower runtime are preferred; points closer to the upper-left corner indicate better accuracy-speed performance.
Figure 4: Robustness evaluation across common perturbations, measured in AUC.
Figure 5: Hyperparameter analysis of our method w.r.t. AUC result on Openfake dataset.
Model
AUC
Runtime
CLIP-B/32 (Layer 8)
0.7652
245.3
DINOv2 (Layer 14)
0.6795
1437.6
DINOv3 (Layer 6)
0.5019
1172.4
CLIP-L/14 (Layer 13)
0.881
436.8
Table 6: Proposed method’s performance (AUC) and runtime in the Openfake dataset under different backbones. For each model, we extract the feature representations from the layer with the highest AUC.
AI-generated image detection identifies synthetic or AI-edited images and develops methods that generalize across generators, domains, and image transformations.