cs.CVFeb 27, 2026
SaveA Difference-in-Difference Approach to Detecting AI-Generated Images
Organizations: The London School of Economics and Political Science · University of Birmingham
Abstract
Diffusion models are able to produce AI-generated images that are almost indistinguishable from real ones. This raises concerns about their potential misuse and poses substantial challenges for detecting them. Many existing detectors rely on reconstruction error -- the difference between the input image and its reconstructed version -- as the basis for distinguishing real from fake images. However, these detectors become less effective as modern AI-generated images become increasingly similar to real ones. To address this challenge, we propose a novel difference-in-difference method. Instead of directly using the reconstruction error (a first-order difference), we compute the difference in reconstruction error -- a second-order difference -- for variance reduction and improving detection accuracy. Extensive experiments demonstrate that our method achieves strong generalization performance, enabling reliable detection of AI-generated images in the era of generative AI.
Figures & tables
Figure 1: Left: The real image and its distribution on the manifold are shown in blue, whereas the synthetic image and its distribution are shown in orange. Right: Reconstructions of the real and synthetic images. and denote the reconstruction errors of synthetic and real images, respectively, and represents the projection of the real image onto . Top: The generator is weak, resulting in a large discrepancy between the real and synthetic image distributions. Consequently, the reconstruction error of the synthetic image, , is much larger than that of the real image, , making detection easier. Bottom: The generator is strong, so the real and synthetic image distributions are close. Here, is similar to , making AI-generated images difficult to detect.
Figure 2: Visualizations of our DID detector. denotes the differencing operator that computes the pixel-wise difference between two images.
| Train set & Gen. model | ImageNet & ADM | |||||
|---|---|---|---|---|---|---|
| Eval set | Gen. model | DID | DIRE | LaRE 2 | AERO. | UFD |
| LSUN-B | ADM | 99.5 | 99.5 | 50.0 | 52.3 | 63.9 |
| PNDM | 97.6 | 97.6 | 55.4 | 49.5 | 91.8 | |
| DDPM | 99.5 | 99.8 | 56.6 | 56.7 | 76.5 | |
| iDDPM | 99.4 | 99.6 | 50.0 | 50.3 | 85.1 | |
| SDv1 | 99.7 | 99.9 | 100 | 50.6 | 94.5 | |
Table 1: Accuracy (%) of different detectors when trained on ImageNet & ADM . The best results are highlighted in bold , while the second best results are underlined . All averages are computed from the raw values and then rounded to two decimals.
Figure 3: Visualizations of (a) input image ; (b) its first-order reconstruction error ; (c) the reconstruction error of the reconstructed image ; and (d) the second-order reconstruction error . Compared to the first-order error , the second-order error better distinguishes real from fake images: real images produce more stable and brighter responses, while fake images exhibit noticeably weaker signals, resulting in a larger gap between the two categories.
| Train set & Gen. model | LAION & Kan.3 | Imp. | LAION & SDXL | Imp. | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Eval set | Gen. model | DID | DIRE | LaRE 2 | AERO. | UFD | DID | DIRE | LaRE 2 | AERO. | UFD | ||
| LSUN-B | ADM | 90.7 | 88.1 | 50.3 | 50.0 | 50.8 | 21.85% | 95.0 | 94.3 | 50.2 | 50.0 | 50.3 | 13.04% |
| PNDM | 78.2 | 73.5 | 71.2 | 49.8 | 51.5 | 17.55% | 81.7 | 74.6 | 51.6 | 49.8 | 50.9 | 27.95% | |
| DDPM | 88.7 | 85.2 | 56.5 | 56.4 | 51.5 | 24.05% | 95.1 | 93.3 | 57.1 | 56.4 | 50.5 | 27.73% | |
| iDDPM | 80.1 | 77.5 | 50.0 | 49.8 | 54.2 | 11.33% | 89.0 | 86.8 | 50.0 | 49.8 | 53.3 | 16.29% | |
| SDv1 | 98.6 | 97.8 | 99.9 | 49.8 | 49.7 | - | 98.4 | 98.3 | 99.7 | 49.8 | 50.5 | - | |
Table 2: Accuracy (%) of different detectors when trained on LAION & Kan.3 (left block) and LAION & SDXL (right block). The best results are highlighted in bold , while the second best results are underlined . All averages are computed from the raw values and then rounded to two decimals. The last column of each block reports the improvement of DID over the second-best baseline.
| Train set & Gen. model | Eval set | Gen. model | Accuracy (%) | Imp. | ||||
|---|---|---|---|---|---|---|---|---|
| DID | DIRE | LaRE 2 | AERO. | UFD | ||||
| LAION & Kan.3 | LSUN-B | StyleGAN | 95.8 | 94.3 | 50.4 | 49.8 | 50.8 | 26.32% |
| ProjectedGAN | 93.4 | 91.0 | 51.2 | 49.8 | 50.7 | 26.52% | ||
| Diff-ProjectedGAN | 92.7 | 90.2 | 51.9 | 49.8 | 51.1 | 25.51% | ||
| Avg. | 93.95 | 91.82 | 51.12 | 49.75 | 50.85 | 26.07% | ||
| LAION & SDXL | LSUN-B | StyleGAN | 96.8 | 96.3 | 50.7 | 49.8 | 50.9 | 12.16% |
Table 3: Detection accuracy (%) on GAN-generated images under different training settings. Best in bold , second best underlined . All averages are computed from the raw values and then rounded to two decimals. The last column reports the improvement of DID over the second-best baseline.
| Eval set | Gen. model | Accuracy (%) | ||
|---|---|---|---|---|
| DID | DIRE | |||
| LSUN-B | ADM | 86.5 | 69.0 | 84.0 |
| PNDM | 83.0 | 81.5 | 67.5 | |
| DDPM | 82.0 | 78.0 | 71.5 | |
| ImageNet | ADM | 98.9 | 98.4 | 91.0 |
| SDv1 | 98.9 | 98.7 | 93.5 | |
Table 4: Accuracy (%) on various evaluation datasets when trained on LAION & Kan3.
| Gen. model | LAION & Kan.3 (PNG-only training) | LAION & SDXL | ||||
|---|---|---|---|---|---|---|
| PNG | JPG95 | JPG75 | PNG | JPG95 | JPG75 | |
| Kan.3 | 97.2 | 98.9 | 95.1 | 90.6 | 85.4 | 70.4 |
| SDXL | 93.2 | 88.9 | 76.3 | 97.3 | 97.1 | 85.6 |
| Vega | 89.1 | 84.3 | 72.3 | 95.2 | 93.0 | 76.3 |
| Playground v2.5 | 94.9 | 92.9 | 82.7 | 91.6 | 88.0 | 73.4 |
| Stable Cascade | 96.2 | 95.6 | 89.4 | 94.4 | 87.2 | 71.7 |
Table 5: Accuracy (%) of DID trained on the PNG-only variant. Left block: trained on LAION & Kan.3 . Right block: trained on LAION & SDXL . Models are evaluated on PNG, JPEG 95, and JPEG 75 versions of each test set. All averages are computed from raw values and then rounded to two decimals.
| Eval set | Gen. model | DID 0.29 | DID 0.5 | Effort | DDA |
|---|---|---|---|---|---|
| GenImage | ADM | 99.90 / 1.000 / 0.00 | 100.00 / 1.000 / 0.00 | 98.55/ 0.999 /0.40 | 99.30/ 1.000 / 0.20 |
| BigGAN | 99.90 / 1.000 / 0.00 | 99.95 / 1.000 / 0.00 | 95.45/0.995/2.60 | 98.70/ 0.999 / 0.50 | |
| GLIDE | 99.90 / 1.000 / 0.00 | 100.00 / 1.000 / 0.00 | 98.35/ 0.999 / 0.20 | 99.20/ 1.000 / 0.20 | |
| Midjourney | 99.70 / 1.000 / 0.00 | 99.70 / 1.000 / 0.00 | 57.25 /0.744/80.00 | 54.95/ 0.845 / 56.30 | |
| SD-v1.4 | 99.45 / 1.000 / 0.00 | 99.50 / 1.000 / 0.00 | 86.70/ 0.975 / 13.00 | 72.81/0.965/15.60 | |
| SD-v1.5 | 99.50 / 1.000 / 0.00 | 99.40 / 1.000 / 0.00 | 83.95/ 0.975 / 11.30 | 72.65/0.962/16.50 |
Table 6: Acc (%) / AUROC / FPR@TPR95 (%) of DID (with two thresholds) and two additional baselines ( Effort and DDA ). All detectors are trained on GenImage and evaluated on GenImage , GenDet , and our two original test datasets. To avoid excessive 1.00 values, we report AUROC with three decimal places.
| Eval set | Gen. model | DID 0.29 | DID 0.5 | Effort | DDA |
| GenImage | ADM | 99.20 / 1.000 / 0.00 | 99.50 / 1.000 / 0.00 | 86.30/ 0.986 /6.70 | 98.30/ 0.999 / 0.10 |
| BigGAN | 99.10 / 1.000 / 0.00 | 99.40 / 1.000 / 0.00 | 80.15/0.878/ 39.00 | 98.50/ 1.000 / 0.00 | |
| GLIDE | 99.20 / 1.000 / 0.00 | 99.50 / 1.000 / 0.00 | 85.85/0.973/12.80 | 97.50/ 0.998 / 0.90 | |
| Midjourney | 99.20 / 1.000 / 0.00 | 99.50 / 1.000 / 0.00 | 57.00/0.597/ 90.20 | 56.15/ 0.810 / 58.80 | |
| SD-v1.4 | 99.20 / 1.000 / 0.00 | 99.50 / 1.000 / 0.00 | 85.30/0.958/19.10 | 82.91/ 0.969 / 13.30 | |
| SD-v1.5 | 99.20 / 1.000 / 0.00 | 99.50 / 1.000 / 0.00 | 84.75/0.950/22.10 | 83.60/ 0.967 / 17.00 |
Table 7: Acc (%) / AUROC / FPR@TPR95 (%) of DID (with two thresholds) and two additional baselines ( Effort and DDA ). All detectors are trained on ImageNet and evaluated on GenImage , GenDet , and our two original test datasets, LAION and ImageNet . To avoid excessive 1.00 values, we report AUROC with three decimal places.
| Method | Reconstruction | Testing |
|---|---|---|
| Time (Memory) | Time (Memory) | |
| LaRE2 | 0.09 (33.34) | 0.66 (9.74) |
| DIRE | 1.01 (396.48) | 0.34 (21.77) |
| DID | 2.01 (398.39) | 0.45 (22.55) |
Table 8: Per-image inference time (seconds) and memory usage (MB) for LaRE2, DIRE, and DID.
| Train set | Eval set | LAION | ImageNet |
|---|---|---|---|
| ImageNet | DID 0.29 | 99.29/1.00/0.00 | 99.85/1.00/0.00 |
| DID SC | 99.03/1.00/0.00 | 99.45/1.00/0.00 |
Table 9: Acc(%) / AUROC / FPR@TPR95(%) of DID 0.29 and a variant that fits both 1st- and 2nd-order signals into a single classifier (DID SC ).
Figure 4: Comparison of first–order and second–order residuals across real and generated images. The DIRE framework [ 80 ] shows that real images typically produce sharper and more structured first-order residuals , whereas fake images tend to yield weaker patterns. However, in broader evaluation across diverse and strong generative models, first-order residuals are not always stable when the real and synthetic distributions are close. In more challenging cases, fake images can even produce stronger residual signals than real images, leading to potential misclassification. In contrast, our proposed second-order residual module provides more consistent separability between real and fake samples by taking the difference between the reconstruction error of the input image and that of its reconstructed version.
Figure 5: Visualization of real images across the DID pipeline. We show the original image , the first reconstruction , the second reconstruction , and their corresponding residual maps , , and .
Figure 6: Visualization of ADM-generated fake images from the ImageNet subset constructed in the DIRE dataset. For each ImageNet class (“A photo of class”, class from ImageNet [ 24 ] ), we show the ADM-generated image , its first reconstruction , the second reconstruction , and their corresponding residual maps , , and .
Figure 7: Visualization of Kan3-generated images constructed for our dataset. For each sample, we show the original synthesized image , the first reconstruction , the second reconstruction , and their corresponding residual maps , , and . All Kan3 images are generated using prompts collected from Midjourney.
Figure 8: Visualization of Vega-generated images constructed for our dataset. For each sample, we show the original synthesized image , the first reconstruction , the second reconstruction , and their corresponding residual maps , , and . All Vega images are generated using prompts collected from Midjourney.
Figure 9: Visualization of SDXL-generated images constructed for our dataset. For each sample, we show the original synthesized image , the first reconstruction , the second reconstruction , and their corresponding residual maps , , and . All SDXL images are generated using prompts collected from Midjourney.
Figure 10: Visualization of playground-2.5-generated images constructed for our dataset. For each sample, we show the original synthesized image , the first reconstruction , the second reconstruction , and their corresponding residual maps , , and . All playground-2.5 images are generated using prompts collected from Midjourney.
Figure 11: Visualization of Stable Cascade-generated images constructed for our dataset. For each sample, we show the original synthesized image , the first reconstruction , the second reconstruction , and their corresponding residual maps , , and . All Stable Cascad images are generated using prompts collected from Midjourney.