The remarkable realism of images generated by diffusion models poses critical detection challenges. Current methods utilize reconstruction error as a discriminative feature, exploiting the observation that real images exhibit higher reconstruction errors when processed through diffusion models. However, these approaches require costly reconstruction computations and depend on specific diffusion models, making their performance highly model-dependent. We identify a fundamental difference: real images are more difficult to fit with Gaussian distributions compared to synthetic ones. In this paper, we propose Forgery Identification via Noise Disturbance (FIND), a novel method that requires only a simple binary classifier. It eliminates reconstruction by directly targeting the core distributional difference between real and synthetic images. Our key operation is to add Gaussian noise to real images during training and label these noisy versions as synthetic. This step allows the classifier to focus on the statistical patterns that distinguish real from synthetic images. We theoretically prove that the noise-augmented real images resemble diffusion-generated images in their ease of Gaussian fitting. Furthermore, simply by adding noise, they still retain visual similarity to the original images, highlighting the most discriminative distribution-related features. The proposed FIND improves performance by 11.7% on the GenImage benchmark while running 126x faster than existing methods. By removing the need for auxiliary diffusion models and reconstruction, it offers a practical, efficient, and generalizable way to detect diffusion-generated content.
Figures & tables
Figure 1: Comparison between FIND and previous noise-based methods. (a) DIRE utilizes complete reconstruction by adding noise and then denoising it with a Diffusion Model (DM) in multiple steps; (b) LaRE 2 compresses the noise addition and denoising steps into a single step within the latent space; (c) Our method only involves adding noise without the denoising process. Compared with previous methods, FIND avoids explicit reconstruction and omits the dependence on DM.
Figure 2
Figure 4: Cross-validation results across different training and testing subsets. We report the accuracy for DIRE, LaRE 2 and Ours on all 8 generators, where each row represents a corresponding training subset. Our method shows consistent and superior performance on different training sets and test sets.
Method
Backbone
Test diffusion generators
Average
Midjourney
SD V1.4
SD V1.5
ADM
GLIDE
Wukong
VQDM
BigGAN
CNNSpot ( Wang et al. 2020 )
RN50
58.20
70.30
70.20
57.00
57.10
67.70
56.70
56.60
61.73
Spec ( Zhang et al. 2019 )
RN50
56.70
72.40
72.30
57.90
65.40
70.30
61.70
64.30
65.13
F3Net ( Qian et al. 2020 )
RN50
55.10
73.10
73.10
66.50
57.80
72.30
62.10
56.50
64.56
GramNet ( Liu et al. 2020 )
RN50
58.10
72.80
72.70
58.70
65.30
71.30
57.80
61.20
64.74
DIRE ( Wang et al. 2023 )
RN50
65.00
73.73
73.69
61.85
69.16
74.33
63.43
56.74
67.24
Table 1: Comparison of the accuracy of our method with existing methods. Each row in the table shows the average test accuracy of the corresponding method, calculated when trained on 8 subsets and tested on the respective test set. The last column gives the overall average accuracy. The symbol †indicates our reproduction from its source code on a CLIP-ViT-B/16 version. The proposed FIND outperforms the SOTAs on both RN50 and ViT-B/16 backbones.
Method
Time per image (ms)
FPS
Rec
Cls
Total
DIRE
151.6
0.3
151.9
6.6
LaRE 2
50.0
0.4
50.4
19.8
Ours
0.0
0.4
0.4
2500.0
Table 2: Average inference time of different methods. Without reconstruction, our method is significantly faster.
N
R w/ N
S w/ N
Midjourney
SD V1.4
SD V1.5
ADM
GLIDE
Wukong
VQDM
BigGAN
Average
1
73.97
83.59
83.54
76.35
91.45
80.78
83.62
71.30
80.58
2
✓
66.66
78.74
79.00
73.59
79.83
75.82
82.43
71.66
75.97
3
✓
✓
76.03
87.62
87.70
86.11
95.03
83.96
89.58
79.75
85.72
4
✓
64.23
79.29
79.40
71.18
76.98
76.09
82.12
70.45
74.97
5
✓
✓
75.23
86.10
86.31
85.48
96.74
82.99
87.52
75.61
84.50
6
✓
82.89
89.69
89.77
87.42
97.82
87.04
89.25
82.96
88.35
Table 3: Impact of different kinds of noise. N: using Gaussian noise as images; R w/ N: adding Gaussian noise to real images; S w/ N: adding Gaussian noise to synthetic images. The R w/ N plays a dominant role, consistent with our hypothesis.
Figure 5: The impact of ϵ . The accuracy surges as ϵ rises initially and starts fluctuating when ϵ exceeds 20. The quality of the images deteriorates rapidly as ϵ increases. We pick a balanced value of ϵ=50 .
School of Artificial Intelligence, UCAS & NLPR, CASIA Beijing, China · School of Advanced Interdisciplinary Sciences, UCAS & NLPR, CASIA Beijing, China · NLPR, Institute of Automation, Chinese Academy of Sciences Beijing, China