SlopeGuard: Towards Generalizable AI-Generated Image Detection via Radial Spectral Slope Guidance
Organizations: School of Computer Science and Technology, Changsha University of Science and Technology · School of Artificial Intelligence, University of Chinese Academy of Sciences · MAIS, Institute of Automation, Chinese Academy of Sciences
Abstract
AI-generated image detectors often overfit to generator-specific artifacts, hindering generalization to unseen synthesis pipelines. Existing frequency-based methods usually learn high-dimensional spectral patterns that may preserve generator-dependent variations. We explore Radial Spectral Slope, a compact descriptor of global spectral decay that captures transferable spectral trends while suppressing fine-grained variations. Our analysis shows that real and generated images exhibit systematic but generator-dependent slope differences, indicating that slope provides useful forensic evidence but is insufficient as a standalone decision criterion. This motivates combining spectral slope with semantic representations for more robust detection. Based on this insight, we propose SlopeGuard, which couples slope-aware data processing with slope guided feature learning to construct informative patch-based views, integrate Radial Spectral Slope with semantic features, and adapt prior preservation according to spectral atypicality. Extensive experiments demonstrate that SlopeGuard achieves state-of-the-art mean accuracies of 97.5% on AIGCDetectBench, 97.8% on GenImage, and 85.1% on AIGIBench, while maintaining competitive performance under blur, cropping, JPEG compression, and additive noise.
Figures & tables
| Method | GAN-based | Diffusion-based | Other | mAcc. | ||||||||||||||
| Big- GAN | Pro- GAN | Star- GAN | Style- GAN | Style- GAN2 | Cycle- GAN | Gau- GAN | Mid- journey | Wu- kong | DALLE2 | SD1.5 | SD1.4 | SDXL | ADM | VQDM | GLIDE | WFIR | ||
| CNNSpot | 50.6 | 97.6 | 60.5 | 68.2 | 61.0 | 59.2 | 51.2 | 52.5 | 63.7 | 52.1 | 77.7 | 77.7 | 51.8 | 50.7 | 50.2 | 52.3 | 50.3 | 60.4 |
| Gram-Net | 54.0 | 98.5 | 94.8 | 75.4 | 69.3 | 56.9 | 50.0 | 67.1 | 96.4 | 66.8 | 99.2 | 99.4 | 72.4 | 57.3 | 53.3 | 76.4 | 46.7 | 72.6 |
| UniFD | 85.9 | 98.4 | 97.5 | 75.4 | 72.7 | 88.0 | 92.1 | 78.3 | 90.4 | 78.1 | 95.7 | 96.1 | 84.9 | 53.9 | 70.8 | 85.5 | 70.4 | 83.2 |
| NPR | 57.9 | 99.4 | 53.3 | 83.6 | 79.3 | 64.6 | 50.7 | 85.1 | 98.5 | 94.5 | 99.6 | 99.7 | 90.6 | 63.0 | 57.4 | 84.3 | 55.2 | 77.5 |
| FreqNet | 72.7 | 99.3 | 94.2 | 86.9 | 83.3 | 78.7 | 51.8 | 86.4 | 94.7 | 82.0 | 95.6 | 95.6 | 90.8 | 64.9 | 62.9 | 85.4 | 61.5 | 81.6 |
| Method | Midjourney | SDv1.4 | SDv1.5 | ADM | GLIDE | Wukong | VQDM | BigGAN | Mean | |||||||||
| Acc. | AP. | Acc. | AP. | Acc. | AP. | Acc. | AP. | Acc. | AP. | Acc. | AP. | Acc. | AP. | Acc. | AP. | Acc. | AP. | |
| CNNSpot | 52.5 | 80.4 | 77.7 | 97.6 | 77.7 | 97.6 | 50.7 | 66.6 | 52.3 | 83.1 | 63.7 | 88.9 | 50.2 | 57.0 | 50.0 | 64.5 | 59.3 | 79.5 |
| Gram-Net | 67.1 | 90.4 | 99.4 | 100.0 | 99.2 | 99.9 | 57.3 | 71.5 | 76.4 | 94.6 | 96.4 | 99.7 | 53.3 | 61.5 | 60.9 | 90.3 | 76.2 | 88.5 |
| LGrad | 80.7 | 91.2 | 95.7 | 99.3 | 95.6 | 99.2 | 60.0 | 68.5 | 88.2 | 96.0 | 92.7 | 97.9 | 59.8 | 69.4 | 94.6 | 98.6 | 83.3 | 90.0 |
| UniFD | 78.3 | 91.7 | 96.1 | 99.3 | 95.7 | 99.1 | 53.9 | 65.2 | 85.5 | 95.3 | 90.4 | 97.4 | 70.8 | 87.2 | 83.1 | 94.4 | 81.7 | 91.2 |
| FreqNet | 86.4 | 94.5 | 95.7 | 99.9 | 95.6 | 99.9 | 64.6 | 74.1 | 85.3 | 94.1 | 94.7 | 99.1 | 62.6 | 74.9 | 94.6 | 99.3 | 84.9 | 92.0 |
| Method | CNNSpot | Gram-net | NPR | FreqNet | LGrad | BSF ♣ | SAFE | AIDE | DFFreq | Effort ♣ | DGS-Net ♠ | Ours | ||||||||||||
| Acc. | AP. | Acc. | AP. | Acc. | AP. | Acc. | AP. | Acc. | AP. | Acc. | AP. | Acc. | AP. | Acc. | AP. | Acc. | AP. | Acc. | AP. | Acc. | AP. | Acc. | AP. | |
| ProGAN | 97.6 | 99.9 | 98.5 | 100.0 | 99.4 | 100.0 | 99.3 | 100.0 | 96.6 | 99.8 | 99.3 | 100.0 | 100.0 | 100.0 | 97.2 | 99.6 | 98.1 | 100.0 | 100.0 | 100.0 | 99.7 | 100.0 | 99.8 | 100.0 |
| R3GAN | 50.4 | 52.7 | 47.9 | 52.5 | 50.8 | 61.1 | 62.3 | 56.8 | 54.4 | 58.7 | 94.7 | 93.9 | 93.9 | 98.2 | 92.9 | 97.1 | 61.7 | 74.0 | 95.3 | 98.2 | 90.5 | 97.0 | 93.8 | 97.6 |
| StyleGAN3 | 55.8 | 73.1 | 65.6 | 77.9 | 78.4 | 91.7 | 83.0 | 92.4 | 70.5 | 80.5 | 93.4 | 95.8 | 89.7 | 97.6 | 88.1 | 91.4 | 90.1 | 96.4 | 95.5 | 98.9 | 90.7 | 98.3 | 94.2 | 97.7 |
| StyleGAN-XL | 52.8 | 64.2 | 72.9 | 83.7 | 60.3 | 75.3 | 79.8 | 84.1 | 65.7 | 74.6 | 94.3 | 97.2 | 93.1 | 97.6 | 88.7 | 93.2 | 73.7 | 83.4 | 89.6 | 97.0 | 89.6 | 97.8 | 93.3 | 96.7 |
| StyleSwim | 52.6 | 76.5 | 75.5 | 86.0 | 85.7 | 94.9 | 80.8 | 91.8 | 81.3 | 90.0 | 94.6 | 98.6 | 97.8 | 99.6 | 83.7 | 89.3 | 84.5 | 92.9 | 96.8 | 99.1 | 89.8 | 99.6 | 93.3 | 95.7 |
| Method | Ref. | mAcc. | AP. |
| C2P-CLIP ∗ | AAAI’25 | 71.8 | 82.2 |
| VIB-Net | CVPR’25 | 69.3 | 70.9 |
| D 3 | CVPR’25 | 63.7 | 68.9 |
| FerretNet | NeurIPS’25 | 79.4 | 85.8 |
| LOTA | ICCV’25 | 77.4 | 83.1 |
| LTD | CVPR’26 | 74.9 | 77.6 |
| Perturbation | NPR | FreqNet | UniFD | AIDE | SAFE | DFFreq | BSF | Ours | |||||||||||
| Blur | Crop | JPEG | Noise | mAcc. | AP. | mAcc. | AP. | mAcc. | AP. | mAcc. | AP. | mAcc. | AP. | mAcc. | AP. | mAcc. | AP. | mAcc. | AP. |
| 70.8 | 78.0 | 67.7 | 73.5 | 72.4 | 76.2 | 77.2 | 81.7 | 79.6 | 74.1 | 78.6 | 77.7 | 79.1 | 76.9 | 80.8 | 84.2 | ||||
| 68.3 | 74.6 | 65.4 | 69.7 | 71.4 | 74.4 | 75.8 | 82.0 | 77.0 | 81.1 | 76.9 | 81.6 | 78.3 | 82.2 | 83.2 | 88.4 | ||||
| 59.0 | 65.4 | 58.3 | 62.1 | 66.4 | 71.4 | 63.9 | 71.6 | 64.8 | 69.2 | 64.1 | 70.4 | 64.6 | 67.6 | 67.2 | 75.3 | ||||
| 60.5 | 69.3 | 59.5 | 64.2 | 68.9 | 72.4 | 69.2 | 80.2 | 65.3 | 74.7 | 78.2 | 78.8 | 71.5 | 81.9 | 81.7 | 84.3 | ||||
| 64.7 | 72.1 | 62.9 | 67.4 | 69.8 | 73.5 | 71.6 | 79.5 | 71.9 | 73.9 | 74.5 | 77.1 | 73.5 | 76.8 | 78.2 | 83.0 | ||||
| Perturbation | NPR | FreqNet | UniFD | AIDE | SAFE | DFFreq | BSF | Ours | |||||||||||
| Blur | Crop | JPEG | Noise | mAcc. | AP. | mAcc. | AP. | mAcc. | AP. | mAcc. | AP. | mAcc. | AP. | mAcc. | AP. | mAcc. | AP. | mAcc. | AP. |
| 81.1 | 88.2 | 75.4 | 82.4 | 81.9 | 91.7 | 86.5 | 92.7 | 85.1 | 85.7 | 86.0 | 86.4 | 86.6 | 84.6 | 90.7 | 97.0 | ||||
| 77.5 | 91.7 | 80.7 | 88.1 | 83.2 | 92.5 | 88.8 | 94.6 | 86.5 | 94.5 | 91.4 | 94.8 | 93.2 | 95.4 | 95.7 | 99.2 | ||||
| 63.9 | 73.7 | 66.3 | 70.1 | 70.5 | 84.1 | 71.6 | 84.1 | 70.3 | 80.9 | 71.7 | 77.1 | 72.3 | 75.8 | 73.4 | 83.2 | ||||
| 66.7 | 81.2 | 69.9 | 77.2 | 76.8 | 88.3 | 79.4 | 91.6 | 70.7 | 83.4 | 84.0 | 88.5 | 82.3 | 85.2 | 87.6 | 95.3 | ||||
| 72.1 | 82.7 | 72.9 | 78.0 | 78.0 | 88.9 | 81.4 | 90.6 | 77.9 | 84.4 | 83.3 | 85.7 | 77.5 | 89.3 | 86.7 | 93.5 | ||||
| Method | LlamaGen | Janus-Pro | VAR | Infinity | Open-MAGVIT2 | Switti | RAR | Mean | ||||||||
| Acc. | AP | Acc. | AP | Acc. | AP | Acc. | AP | Acc. | AP | Acc. | AP | Acc. | AP | mAcc. | mAP | |
| CNN-Spot [ 36 ] | 50.9 | 72.4 | 63.3 | 91.6 | 50.1 | 61.5 | 63.4 | 94.5 | 50.2 | 58.2 | 54.3 | 81.9 | 50.6 | 68.7 | 54.7 | 75.5 |
| UniFD [ 26 ] | 71.2 | 85.0 | 80.7 | 91.3 | 61.1 | 74.9 | 91.8 | 97.0 | 59.3 | 72.7 | 91.2 | 96.9 | 83.2 | 93.2 | 76.9 | 87.3 |
| FreqNet [ 31 ] | 90.3 | 96.9 | 95.6 | 99.8 | 78.3 | 89.2 | 95.5 | 99.7 | 76.4 | 88.3 | 95.3 | 99.8 | 80.8 | 91.7 | 87.5 | 95.1 |
| NPR [ 32 ] | 83.5 | 97.5 | 97.0 | 99.8 | 81.8 | 98.6 | 89.7 | 99.5 | 55.7 | 82.0 | 99.4 | 100.0 | 67.7 | 93.6 | 82.1 | 95.9 |
| LaDeDa [ 1 ] | 89.4 | 99.6 | 99.6 | 100.0 | 80.5 | 98.7 | 99.8 | 100.0 | 78.0 | 99.0 | 99.8 | 100.0 | 87.1 | 99.6 | 90.6 | 99.5 |
| Statistics | RAcc. | FAcc. | mAcc. |
| Fixed training | 90.3 | 74.6 | 82.5 |
| Batch | 89.4 | 80.8 | 85.1 |
| 0.05 | 0.075 | 0.10 | 0.125 | 0.15 | |
| Uniform | 80.9 | 80.6 | 81.8 | 79.4 | 79.4 |
| Slope-guided | 78.8 | 82.5 | 85.1 | 81.0 | 82.8 |
| Spectral Descriptor | mAcc. |
| None | 78.8 |
| Frequency energy ratio | 80.3 |
| Spectral entropy | 81.8 |
| Radial spectral slope | 83.7 |
| Frequency Diversification | Patch Selection | mAcc. |
| 72.0 | ||
| 73.5 | ||
| 76.9 | ||
| 78.8 |
| Direction | Energy | mAcc. |
| 83.7 | ||
| 81.2 | ||
| 83.8 | ||
| 85.1 |
| Group | Mean | SD | Median | IQR |
| Shared real | -0.02339 | 0.00624 | -0.02292 | 0.00851 |
| ADM | -0.02512 | 0.00621 | -0.02491 | 0.00894 |
| DALL-E 2 | -0.02942 | 0.00292 | -0.02967 | 0.00352 |
| Glide | -0.03958 | 0.00368 | -0.03998 | 0.00453 |
| Midjourney | -0.02726 | 0.00794 | -0.02973 | 0.01229 |
| VQDM | -0.02305 | 0.00619 | -0.02255 | 0.00786 |
| Generator | CvM | KS | ||||
| ADM | 3.867 | 0.132 | 0.00173 | -0.00199 | 0.0001 | 0.0017 |
| DALL-E 2 | 64.564 | 0.554 | 0.00620 | -0.00674 | 0.0001 | 0.0017 |
| Glide | 152.573 | 0.890 | 0.01618 | -0.01706 | 0.0001 | 0.0017 |
| Midjourney | 19.418 | 0.338 | 0.00415 | -0.00681 | 0.0001 | 0.0017 |
| VQDM | 0.216 | 0.042 | 0.00040 | +0.00038 | 0.2441 | 0.2441 |
| BigGAN | 11.104 | 0.248 | 0.00257 | -0.00198 | 0.0001 | 0.0017 |
| Reference | Mean CvM | Mean | 95% CI | Significant | |
| Reconstructed | |||||
| Shared real | 1000/1000 | 36.986 | 0.00458 | [0.00433, 0.00488] | 16/17 |
| JPEG matched | 665/1000 | 57.550 | 0.00600 | [0.00560, 0.00637] | 1/1 |
| PNG matched | 335/1000 | 12.092 | 0.00331 | [0.00320, 0.00351] | 16/16 |
| Directory matched | 1000/1000 | 49.150 | 0.00470 | – | 16/17 |
| Original (native) | |||||
AI-Generated Image Detection
AI-generated image detection identifies synthetic or AI-edited images and develops methods that generalize across generators, domains, and image transformations.
See all 206 papers